Encyclopedia of Computational Neuroscience

Dieter Jaeger
Ranu Jung
Editors
Encyclopedia of
Computational
Neuroscience
1 3Reference
Encyclopedia of Computational
Neuroscience
Dieter Jaeger • Ranu Jung
Editors
Encyclopedia of
Computational
Neuroscience
With 1109 Figures and 71 Tables

Editors
Dieter Jaeger Ranu Jung
Department of Biology Department of Biomedical Engineering
Emory University Florida International University
Atlanta, GA, USA Miami, FL, USA
ISBN 978-1-4614-6674-1 ISBN 978-1-4614-6675-8 (eBook)

ISBN 978-1-4614-6676-5 (print and electronic bundle)
DOI 10.1007/978-1-4614-6675-8
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2014958664
# Springer Science+Business Media New York 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way,
and transmission or information storage and retrieval, electronic adaptation, computer software,
or by similar or dissimilar methodology now known or hereafter developed. Exempted from this
legal reservation are brief excerpts in connection with reviews or scholarly analysis or material
supplied specifically for the purpose of being entered and executed on a computer system, for
exclusive use by the purchaser of the work. Duplication of this publication or parts thereof is
permitted only under the provisions of the Copyright Law of the Publisher’s location, in its
current version, and permission for use must always be obtained from Springer. Permissions for
use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable
to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are
exempt from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal
responsibility for any errors or omissions that may be made. The publisher makes no warranty,
express or implied, with respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Booknotes
This edition comprises 4 Volumes:
• Volume 1: Overview
A–C
• Volume 2: D–L
• Volume 3: M–P
• Volume 4: Q–W
v
Preface
Computational Neuroscience has emerged in the last three decades as an

interdisciplinary research area combing approaches from mathematics, physics,
engineering, computer science, and neurobiology. Combining theoretical and
computational approaches with experimental data has proven to illuminate
neural function from molecular to system levels. A testimony to this success
comes from the emergence of several international conferences in this field of
study, such as the Organization of Computational Neuroscience (http://www.
cnsorg.org/) and Cosyne (http://www.cosyne.org) annual meetings. Specialized
funding sources for this area of research have also made a huge impact, and the
field owes much to the organizers of the NIH/NSF Collaborative Research in
Computational Neuroscience program and the German Bernstein Network.
Importantly, new initiatives in data sharing as well as model sharing based on
modern markup language syntax and semantics will make it possible in the near
future to present an accessible collaborative interface between investigators as
the field matures. These efforts are in part promoted by the field of
Neuroinformatics and the International Neuroinformatics Coordinating Facility
(http://incf.org/).
We hope that the new Springer Encyclopedia of Computational Neurosci-
ence will bring this young field closer to the eye of the general scientific
community and provide a valuable resource in explaining the many angles of
this highly active enterprise. The Encyclopedia highlights achievements and
approaches to describe basic neural function and major brain systems as well
as biomedical applications in more than 570 articles organized into 49 sec-
tions of research. At-depth articles provide comprehensive coverage of
important topics in these disciplines, whereas short articles summarize indi-
vidual concepts and key terms. The interplay between computational and
theoretical approaches and experimental data is highlighted at all levels, from
molecular to cognitive. Available shared database resources are also covered.
While overall alphabetically sorted, an introduction of each section is
presented in articles denoted with “Overview,” which also provide organized
links to section articles.
The level of description in the Encyclopedia is aimed to make the material
accessible to graduate students in the many disciplines that contribute to
computational neuroscience while also providing a valuable reference to
advanced researchers. Cited website links allow access to a more detailed
level of information when needed. For those with institutional access to the
vii
viii Preface
online SpringerReference enterprise, a hot-linked version of this encyclope-

dia is available under http://www.springerreference.com/.
The Editors in Chief are pleased to present this encyclopedia and are
looking forward to readers’ comments that will be taken to further improve
and complete future updates of this work.
Dieter Jaeger
Ranu Jung
Research Interests
James Bednar uses computational modeling to understand how visual cortex

circuitry develops and functions. The goal is to find a small set of develop-
mental and other mechanisms that are sufficient to account for the wide range
of functional properties that have been observed in populations of neurons in
adult animals and humans.
Ulrik Beierholm’s research focuses on developing machine-learning-

inspired models (e.g., based on Bayesian statistics or reinforcement learning)
to understand human choices in perception, cognition, and learning and on
testing them through psychophysics and fMRI.
Dr. Ulrik Beierholm studied physics at the University of Copenhagen
before deciding on a research career in Neuroscience. From 2001 to 2007, he
was a Ph.D. student at the California Institute of Technology (Caltech) in the
Computation and Neural Systems program, being awarded a Fulbright
fellowship. After his studies, he completed further postdoctoral training at
the Gatsby Computational Neuroscience Unit (at UCL) in London, where
he worked with Professor Peter Dayan on modeling learning and decision
making while on a Marie Curie Reintegration grant.
Sliman J. Bensmaia is an assistant professor in the Department of Organis-

mal Biology and Anatomy at the University of Chicago, Illinois, USA, where
he is also a member of the Committees on Neurobiology and on Computa-
tional Neuroscience. He received a B.A. in cognitive science from the Uni-
versity of Virginia in 1995 and a Ph.D. in cognitive psychology from the
University of North Carolina at Chapel Hill, USA, in 2003 under the tutelage
of Mark Hollins. He then joined the laboratory of Kenneth Johnson at the
Johns Hopkins University Krieger Mind/Brain Institute, Baltimore,
Maryland, USA, initially as a postdoctoral fellow and then as an associate
research scientist. The main objectives of Bensmaia’s research are to discover
the neural basis of somatosensory perception using psychophysics, neuro-
physiology, and computational modeling. He also seeks to apply insights
from basic science to develop approaches to convey sensory feedback in
upper-limb neuroprostheses.
Kim T. Blackwell is a Professor in the Department of Molecular Neurosci-

ence at George Mason University. Her research interests are to understand
biophysical and biochemical mechanisms underlying memory storage. Her
ix
x Research Interests
approach interlaces experiments and model development on the network

level, cellular level, and subcellular level in order to identify the mechanisms
whereby particular spatiotemporal patterns of inputs produce changes in
synaptic plasticity and intrinsic excitability in the hippocampus and striatum.
She has developed the software tools Chemesis and NeuroRD for large-scale
dynamical modeling of signaling pathways in neurons underlying memory
storage.
Ingo Bojak has worked in computational neuroscience for the last decade,
with over 30 publications in that field. He is an expert in neural population
models (NPMs) and was the first to use them to simulate an entire human
cortex on a parallel compute cluster with MPI-C. This has progressed to
anatomically realistic neural mass meshes, which allow the inclusion of
experimentally determined brain connectivity. Another main focus of his
research has been modeling general anaesthetic agents. Bojak has also
worked on the effect of axonal diameter distributions on brain dynamics,
the influence of alpha rhythm phase on fMRI BOLD, long-range synchrony in
cortical networks, and orientation sensitivity in the visual cortex. He is an
editor with the journals Neurocomputing and EPJ Nonlinear Biomedical
Physics and has served on the Board of Directors and Program Committee
of the OCNS and as Node Representative of the INCF for the UK and the
Netherlands. He has recently joined the EPSRC Peer Review College.
Alla Borisyuk is an Associate Professor of Mathematics and an associate

faculty member in the Neuroscience Program at the University of Utah. Her
research focuses on the mathematical analysis of neuronal models, particu-
larly on the shaping of neural dynamics by low-level processes such as single-
cell properties, network architecture, and plasticity of connections. She is
working on both theoretical and collaborative (directly related to experi-
ments) types of projects including applications in the auditory system, odor
coding in the olfactory system, emergence of oscillations in noisy networks,
and others.
Steven Bressler is a cognitive neuroscientist at the Center for Complex

Systems and Brain Sciences at FAU who studies the dynamics of neural
information processing during goal-directed behaviors. He has been awarded
grants from the National Science Foundation and the National Institute of
Mental Health to study the dynamics of large-scale cortical circuits in vision.
His research is providing important insights into the interactions of neuronal
assemblies from different brain regions that underlie visual perception, atten-
tion, and working memory.
Robert Butera is a Professor jointly appointed in the School of Electrical and

Computer Engineering at Georgia Tech, the Wallace H. Coulter Dept. of
Biomedical Engineering at Georgia Tech and Emory University, Atlanta,
GA, USA. His research interests include peripheral nerve modulation of
sensation and autonomic/organ function, real-time computing methods for
Research Interests xi
electrophysiology experiments, and general mechanisms of rhythm central

pattern generation.
Carmen Canavier’s major research interest is the nonlinear dynamics of

oscillatory neurons and oscillatory networks of neurons. She uses phase
resetting theory to study synchronization between oscillators and nonlinear
dynamical theory to understand how individual neurons, and midbrain dopa-
mine neurons in particular, produce different firing patterns under different
conditions.
Yoonsuck Choe is a professor and director of the Brain Networks Laboratory

in the Department of Computer Science and Engineering at Texas A&M
University. His research interests are broadly in computational neuroscience,
computational neuroanatomy, neuroinformatics, and neuroevolution. His
work ranges from visual cortical modeling (http://computationalmaps.org)
to sensorimotor learning, temporal aspects of brain function (delay, memory,
and prediction), whole-brain physical sectioning imaging (Knife-Edge Scan-
ning Microscopy), and web-based brain atlas frameworks (http://kesm.org).
Diego Contreras’ main focus of research is on how the nervous system

encodes information. In particular, he is studying (i) the possible role of
oscillatory activity in the gamma band (20–80 Hz) for encoding visual stimuli
in corticothalamic networks and (ii) the role of corticothalamic feedback.
Responses to visual stimuli are recorded from the cortex and thalamus with
combined intracellular and optical recordings using voltage-sensitive dyes
and calcium indicators.
Dr. Contreras also studies the mechanisms by which neural networks
engage in the abnormal, paroxystic activity that characterizes epilepsy.
Sharon Crook holds a joint appointment between the School of Mathemat-

ical and Statistical Sciences and the School of Life Sciences at Arizona State
University. Using mathematical models and computer simulations, her
research focuses on the dynamics of neurons and neuronal networks and the
mechanisms underlying plasticity in neural physiology, morphology, and
behavior due to trauma or learning. Dr. Crook also contributes to the devel-
opment of NeuroML, an international effort to create a common standard for
describing computational models for neuroscience research.
Gennady S. Cymbalyuk is an associate professor in the Neuroscience

Institute and the Physics and Astronomy Department of Georgia State Uni-
versity, Atlanta, GA, USA. He is interested in the mechanisms governing the
dynamics of single neurons and oscillatory neuronal networks. His research is
focused on the dynamics of central pattern generators, neuromodulation,
control of bursting activity, and multistability of activity regimes. In his
research, he combines methods developed in dynamical systems theory
with electrophysiological techniques.
xii Research Interests
Susan Denham’s program has two main aims: (a) the development of a
theoretical view of sensory perception and perceptual organization
(Winkler et al. 2009, 2012; Mill et al. 2012) and (b) the consideration of the
potential of novel findings from this account for more detailed neurocompu-
tational modeling (e.g., Lanyon and Denham 2010) and the development of
artificial systems (e.g., Sheik et al. 2012).
Current research themes include studies aimed at (a) understanding the
dynamics and influences on perceptual switching (Bendixen et al. 2010),
(b) methods for analyzing individual differences in multistability (Denham
et al. 2012), (c) developing computational models which explain perceptual
flexibility (Mill et al. 2013), and (d) applying this work to the development of
novel solutions to problems of tracking speech in noise, especially in the
elderly (in collaboration with I Winkler, Hungary). Underlying all this is the
aim to identify principles of biological computation, which can be used in the
development of novel brain-inspired technology for deployment in real-world
environments.
Alain Destexhe is Research Director at the CNRS in the research campus of

Gif sur Yvette near Paris. He leads a team of computational neuroscientists, in
which his colleagues and he study the nature of brain states and their impact
on neuronal computations. This work operates at the interface between
experimental and theoretical neuroscience, often combining them. He is
also director of the European Institute for Theoretical Neuroscience (www.
eitn.org), recently created in the center of Paris. He has published more than
100 peer-reviewed journal articles and is author of numerous chapters and
edited books and two monographs.
Alexander Dimitrov’s main research interests involve the study of neural

information processing, neural coding, and information representation in
biological systems. In particular, he is interested in understanding
information-processing functions of neural ensemble activity and the biolog-
ical mechanisms through which these functions are implemented. Quantita-
tive tools applied to achieve these research goals come mostly from branches
of applied probability (information theory, signal processing theory, multi-
variate statistics, and stochastic differential equations), dynamical systems
theory, and group theory.
Fabrizio Gabbiani is a Professor in the Department of Neuroscience at

Baylor College of Medicine. His research interest is in the computation of
sensory processing and sensorimotor transformations in the central nervous
system.
Marc-Oliver Gewaltig is codirector of the Neurorobotics subproject in the

recently approved EU FET Flagship “Human Brain Project” and leads the
Neurorobotics Section of the Blue-Brain Project at the EPFL in Lausanne.
In his research, Marc-Oliver Gewaltig investigates the computational prop-
erties of the neocortical column in closed action–perception loops. He also
has a strong interest in computer science for large-scale neural simulations
Research Interests xiii
and is coauthor of the neural simulation tool NEST (www.nest-inititive.org).

Before joining the EPFL in 2011, Marc-Oliver Gewaltig worked for the
Honda Research Institute Europe in Offenbach, Germany, where he served
as Principal Scientist (2003–2011) and Project Leader (1998–2002). In 1999,
Marc-Oliver Gewaltig received his Ph.D. in Physics for his work on activity
propagation in cortical networks.
Padraig Gleeson is a postdoc in the lab of Angus Silver at University College

London (http://www.ucl.ac.uk/silverlab). His initial work toward developing
more anatomically detailed models of the cerebellar cortex in this lab led to
the development of the application neuroConstruct (http://www.
neuroconstruct.org), which has been widely used for creating other detailed
3D network models. This application’s ability to generate code for multiple
neuronal simulators got him involved in the development of the NeuroML
language for model specification in computational neuroscience (http://www.
neuroml.org/), which has had a major update over the past 2 years for the
version 2.0 of the language. He has recently been leading the development of
the Open Source Brain repository (http://www.opensourcebrain.org) for col-
laborative model development, which will be the subject of his presentation.
Joshua Goldberg works on preclinical models of Parkinson’s disease and

other movement disorders. His work focuses on the physiological mecha-
nisms underlying neurodegeneration and the pathophysiological adaptations
that occur both in global brain dynamics and in the autonomous activity of
neurons affected by these disorders.
Sonja Gr€ un’s research goal is to gain an understanding of the relevant

spatiotemporal scale(s) at which the cortex effectively interacts and to con-
tribute to the uncovering of its function. Her work involves the simultaneous
observation of large portions of the network on the single neuron and popu-
lation levels (in collaboration with experimentalists), the development of
statistical tools for the analysis of multichannel data from awake-behaving
animals, the analysis of experiments to extract and condense the relevant
characteristics of the system, the interpretation of the system dynamics by
construction of theoretical (biophysical and functional) models, and the
development of software and workflows for reproducible analysis.
Christian Hauptmann works in the field of computational neuroscience,

neuromodulation, and neurotechnology. He received his physics degree from
Darmstadt Technical University, Germany, in 1997 and worked in the field of
epilepsy research at the University of Cologne, where he received his Ph.D. in
2000. After a postdoc at McGill University, Montréal, Canada (Humboldt
fellowship), with Michael Mackey he started in 2002 to work at the Juelich
Research Center, Germany, in the team of Peter A. Tass. Christian Haupt-
mann contributed to the development of novel stimulation techniques and
developed medical devices needed to test innovative therapeutic concepts. In
2008, he received his habilitation in the field of medical physics from the
xiv Research Interests
University Hospital of Cologne, Germany. He has published more than

40 peer-reviewed articles.
His research interests are the modeling of the effects of neuromodulation
and the development of devices for invasive and noninvasive
neuromodulation.
William Holmes is a Professor in the Department of Biological Sciences at

Ohio University. His research interests include understanding how computa-
tion and synaptic modification take place in individual neurons of the hippo-
campus. To do this, he develops mathematical and computational models at
three levels of organization: the neuron, the synapse, and biochemical reac-
tions within dendritic spines to understand how these levels interact to affect
signaling mechanisms involved with long-term potentiation (LTP).
Dieter Jaeger is a Professor in the Department of Biology at Emory Univer-

sity. His research interest is in studying synaptic integration in and function of
basal ganglia and cerebellar structures and their feedback to the cerebral
cortex. His approaches combine cellular and system-level electrophysiology
with detailed compartmental neuronal modeling.
Devin Jindrich is an Assistant Professor at the California State University,

San Marcos. His research goals are to understand the dynamic interactions
among the nervous system, the musculoskeletal system, behavioral task
demands, and the environment that result in movement. Current research
projects focus on the neuromechanics of unsteady locomotion, the motor
control of upper-extremity function to prevent injuries associated with
human–computer interfaces, and developing strategies to restore motor func-
tion after injury.
Ranu Jung holds the Wallace H. Coulter Eminent Scholars Chair in Bio-
medical Engineering at Florida International University, Miami, FL where
she is Professor and Chair of the Department of Biomedical Engineering. Her
research interests are in neural engineering and computational neuroscience.
She is actively engaged in understanding the neural control of end organs and
in the development of neurotechnology that is inspired by biology, is adap-
tive, and could be used to promote plasticity in the nervous system to
overcome neurological disability or trauma. Of special interest to her are
biomimetic and biohybrid living-hardware systems interfaced at the periph-
eral nerve or spinal cord levels to restore function after amputation or spinal
cord injury.
Jeanette Hellgren Kotaleski’s main focus of research has been to use

mathematical modeling to understand the neural mechanisms underlying
information processing, rhythm generation, and learning in motor systems.
Of specific interest are the basal ganglia, a structure in the brain that is
important for the selection and initiation of motor (and cognitive) actions.
The levels of investigation using computational models range from simula-
tions of large-scale neural networks using both biophysically detailed but also
Research Interests xv
more abstract system-level models down to kinetic models of subcellular

processes (e.g., dopamine-induced cascades). The latter approach is impor-
tant for understanding mechanisms involved in, e.g., synaptic plasticity and
learning.
Michel Lemay obtained is Ph.D. in biomedical engineering from Case

Western Reserve University in 1994 with a specialization in functional
neuromuscular stimulation. He is currently professor of Bioengineering at
Temple University, where he studies the effects of providing exogenous
neurotrophins on locomotor recovery and the spinal circuitry in a large animal
model of spinal cord injury.
Christiane Linster is a Professor in the Department of Neurobiology and

Behavior at Cornell University. Her research interests focus on
neuromodulation and plasticity in sensory processing using the olfactory
system as a model system. Her approaches combine system-level electro-
physiology, behavioral pharmacology, and compartmental neuronal
modeling.
Nigel Lovell received the B.E. (Hons) and Ph.D. degrees from the University
of New South Wales (UNSW), Sydney, Australia. He is currently at the
Graduate School of Biomedical Engineering, UNSW, where he holds a
position of Scientia Professor. He has authored 180+ journal papers and
been awarded over $76 million in R&D and infrastructure funding. His
research work has covered areas of expertise ranging from cardiac modeling,
telehealth technologies, biological signal processing, and visual prosthesis
design. He has commercialized a range of telehealth technologies for man-
aging chronic diseases and falls in the older population. He is also one of the
key researchers leading an R&D program to develop in Australia a retinal
neuroprosthesis or “bionic eye.”
William Lytton is a Professor in Physiology, Pharmacology, and Neurology

at SUNY Downstate and works as a clinical neurologist at Kings County
Hospital, seeing patients with a variety of brain ailments. His research is in
Computational Neuroscience, with a focus on the application of multiscale
modeling to various disorders of the brain including Alzheimer’s, stroke,
Parkinson’s, epilepsy, and schizophrenia. He is the author of a Springer
textbook in the field.
Americo Migliaccio is a vestibular researcher with a broad range of skills,

including basic (animal) science, clinical studies, and biomedical engineer-
ing. He is currently a Senior Research Fellow and Group Leader at Neuro-
science Research Australia, where he has been head of the Balance and Vision
Laboratory since 2008. He is a Conjoint Associate Professor at the University
of New South Wales and an Adjunct Associate Professor at Johns Hopkins
University (USA).
The focus of his laboratory is on vestibular treatment and rehabilitation
that increases/restores vestibular function. His three main objectives include
xvi Research Interests
(1) investigation of the physiological basis for the development of a human

vestibular prosthesis for the ~5 % of vestibular patients that have lost com-
plete (bilateral) vestibular end-organ function, (2) determining the crucial
pathways needed for vestibular adaptation and compensation after peripheral
vestibular end-organ injury using animal models, in particular the alpha9
knockout mouse and the tilted mouse, and (3) investigation of techniques that
maximize vestibular plasticity to increase vestibular responses in patients that
have lost partial vestibular function, i.e., the majority of vestibular patients.
This work has led to the development of a portable vestibular rehabilitation
device presently undergoing clinical trials.
John Milton’s research interest is in the interplay between sensory uncer-

tainty and time delays in shaping the dynamic behaviors of neural feedback
control mechanisms. Current interests include balance control, the develop-
ment of expertise, and epilepsy.
Kendall F. Morris is a Professor in the Department of Molecular Pharma-

cology and Physiology at the Morsani College of Medicine of the University
of South Florida, Tampa, FL. His research interests include the study of
brainstem neural networks that control breathing and cardiovascular function.
His approaches combine in vivo multineuron recording and analysis with
neural network computer simulations.
Farzan Nadim is a Professor of Neurobiology in the Federated Department

of Biology of the New Jersey Institute of Technology and Rutgers Universi-
ty–Newark. He has a joint appointment in the Department of Mathematical
Sciences at NJIT. His area of expertise is Central Pattern Generation. His
research combines electrophysiology with mathematical and computer
modeling to understand the role of synaptic and neuronal dynamics in pro-
ducing network oscillations in the central nervous system.
Theoden I. Netoff is an Associate Professor in the Department of Biomedical

Engineering at the University of Minnesota. His research interest is in epi-
lepsy. In particular, he is interested in how changes in connectivity between
neurons result in pathological network behavior, such as seizures.
Klaus Obermayer received his Ph.D. in 1992 from the Department of

Physics, Technical University of Munich, Germany. From 1992 to 1993, he
was a postdoctoral fellow at the Rockefeller University, New York, and the
Salk Institute for Biological Studies, La Jolla, USA. From 1994 to 1995, he
was member of the Technische Fakultaet, University of Bielefeld, Germany.
He became associate professor in 1995 and full professor in 2001 at the
Department of Electrical Engineering and Computer Science of the Berlin
University of Technology, Germany. He is head of the Neural Information
Processing Group and member of the steering committee of the Bernstein
Center for Computational Neuroscience Berlin. He was member of the
governing board of the International Neural Network Society from 2004 to
2012 and was Vice President of the Organiation for Computational
Research Interests xvii
Neuroscience from 2008 to 2011. From 1999 to 2003, he was one of the
directors of the European Advanced Course of Computational Neuroscience.
His current areas of research are computational neuroscience, artificial neural
networks and machine learning, and the analysis of neural data. He
coauthored more than 250 scientific publications.
Karim G. Oweiss received his B.S. (1993) and M.S. (1996) degrees with
honors in Electrical Engineering from the University of Alexandria, Egypt,
and his Ph.D. degree (2002) in Electrical Engineering and Computer Science
from the University of Michigan, Ann Arbor. He completed a postdoctoral
training in Biomedical Engineering at the University of Michigan, Ann
Arbor, in 2002. In 2003, he joined the department of Electrical and Computer
Engineering and the Neuroscience Program at Michigan State University,
where he is currently an Associate Professor and Director of the Neural
Systems Engineering Laboratory. His research interests span the areas of
statistical signal processing and information theory, neural integration and
coordination in sensorimotor systems, computational neuroscience, and
brain–machine interfaces.
Dr. Oweiss is a senior member of the IEEE and a member of the Society for
Neuroscience. He served as a member of the board of directors of the IEEE
Signal Processing Society on Brain Machine Interfaces and continues to serve
on the technical committees of the IEEE Biomedical Circuits and Systems,
the IEEE Life Sciences, and the IEEE Engineering in Medicine and Biology
societies. He was awarded the Excellence in Neural Engineering award from
the National Science Foundation in 2001. His lab is currently supported
through the Neural Interfaces Program (NIP) and the Repair and Plasticity
Program (RPP) at the National Institute of Neurological Disorders and Stroke
as well as DARPA’s Reliable Central-Nervous-System Interfaces (RCI)
program. He is the editor and coauthor of the book Statistical Signal
Processing for Neuroscience and Neurotechnology published by Academic
Press in 2010.
Astrid Prinz is an Associate Professor in the Department of Biology at

Emory University. Her lab combines experimental and computational
approaches to study rhythmic pattern generation and neuronal homeostasis
in pattern-generating neuronal circuits.
Sylvie Renaud is a Professor at ENSEIRB-MATMECA at the University of

Bordeaux. Her research interest is in analog and mixed neuromorphic and
bioprocessing VLSI; in hybrid systems interfacing electronics and biological
excitable tissues; in smart ASICs for biological signal conditioning, event
detection, and tissue stimulation; in active VLSI implants for neurodegener-
ative diseases and diabetes; and in closed-loop living-artificial systems.
Jorge Riera’s scientific interest is to develop methods for the integration of

neuroimaging multimodalities based on modeling mesoscopic phenomena in
the cerebral cortex. To that end, he focuses attention on the following issues:
signal integration by neurons and astrocytes, microcircuitries and networks of
xviii Research Interests
cells in the cortex, dynamics of neuronal masses and brain connectivity

graphs, biophysics of neurovascular/metabolic coupling, and spatiotemporal
inverse problems in neuroimaging.
Patrick Roberts is an affiliate Associate Professor in the Department of

Biomedical Engineering at Oregon Health & Science University and an
adjunct Associate Professor in the Systems Science Program at Portland
State University. His research interest is to develop appropriate mathematical
methods, both analytical and computational, to study the dynamics of neural
activity patterns and to help understand the relationship between these
dynamics and behavior. He applies these theoretical tools to understand
neural processing and learning in sensory systems and to model the efficacy
of drug actions on symptoms of neurological and psychiatric diseases.
Jonathan Rubin is a Professor of Mathematics at the University of Pitts-

burgh. His interests focus on the dynamic mechanisms yielding bursting and
other complex activities seen in neurons and neuronal networks, in the
generation of rhythms (e.g., in central pattern generators) and oscillations,
and in understanding and treating parkinsonism. Beyond these central
themes, he has diverse interests, mostly relating to dynamics in neural models
but also including synaptic plasticity and parameter estimation.
Ilya A. Rybak received his M.S. in Electrical Engineering from Odessa

Polytechnic University (Odessa, Ukraine) in 1975 and his Ph.D. in Biophys-
ics from St. Petersburg State University (St. Petersburg, Russia) in 1988. He
began his scientific career in 1977 at A.B. Kogan Research Institute for
Neurocybernetics at the Rostov State University. After moving to the USA
in 1991, he worked in the DuPont company as Visiting Scientist.
In 1993–1994, he was Visiting Professor at Le Havre University, France.
Dr. Rybak joined Drexel University in 1999 as Research Professor at the
School of Biomedical Engineering. Since 2006, he has been Professor at
the Department of Neurobiology and Anatomy at Drexel University College
of Medicine. Dr. Rybak published more than 300 papers on the mathematical
and computer modeling of various biological neural networks in different
areas of the brain; neural oscillations; and the neural control of locomotion
and breathing, visual perception, and recognition.
Emilio Salinas grew up in Mexico City. He obtained a B.Sc. degree in

Physics there at the Universidad Nacional Autónoma de México (UNAM).
At that time, he became interested in mathematical models of single neurons
and decided to pursue a doctoral degree. He obtained his Ph.D. from Brandeis
University in 1996 under the supervision of Larry Abbott. During this time, he
worked on various projects related to how populations of neurons encode
sensory and motor information. He then returned to Mexico to join the
laboratory of Ranulfo Romo, where he worked on the analysis and modeling
of neurophysiological data from awake, behaving monkeys. In 1999, he went
to the laboratory of Terry Sejnowski at the Salk Institute in San Diego,
California, for another postdoctoral stay. There, he studied the effects of
Research Interests xix
correlated synaptic inputs on the responses of single model neurons. Emilio

Salinas is currently an Associate Professor in the Department of Neurobiol-
ogy and Anatomy at Wake Forest School of Medicine in Winston-Salem,
North Carolina. His current research interests include constructing model
neural circuits that are able to switch tasks in a context-dependent way,
understanding the optimality of cortical sensory representations, and explor-
ing the neural mechanisms that determine how choices are made.
Fidel Santamaria was born in a small town in Mexico called Mexico City; at
that time, it probably only had a few million people. There, he went to college
at the Universidad Nacional Autonoma de Mexico (UNAM) to study Physics.
His first research experience was in a short-lived space program (PUIDE) in
which he got involved in studying rocket propulsion. It was fun at the time;
however, he had been always interested in robotics, artificial intelligence, and
brain function. He then moved to the lab of Prof. Ismael Espinosa, director of
the Cybernetics Lab, who got his Ph.D. from Gerstein at Penn. He studied and
built artificial neural networks, experienced firsthand extracellular electro-
physiology, and decided that he wanted to learn more.
Since he wanted to do biophysical modeling and experiments to under-
stand how neurons and networks work, he applied to the Ph.D. program in
Computation & Neural Systems at Caltech. There, he joined the lab of
Prof. James M. Bower. His thesis was a combination of network modeling
and electrophysiological recordings to understand how the cerebellar cortex
processes sensory information.
By the end of his Ph.D., he became more interested in understanding the
inner workings of neurons instead of looking at large numbers of cells. He
wrote a letter to Prof. George J. Augustine at Duke University, asking him if
he would be interested in having a person like Fidel Santamaria in his lab.
Luckily, he was. Fidel Santamaria studied how the structure of a neuron
affects the diffusion of soluble molecules. For that, he used ultrafast confocal
microscopy and Monte Carlo simulations of molecular diffusion.
Lars Schwabe aims at understanding cortical information processing in

terms of the underlying principles, the neuronal dynamics, and as to whether
the two can be separated at all. The focus of his research has been on the
dynamic properties of strongly recurrent networks in the visual cortex, their
emergent dynamical properties, and how those are affected by inter-areal
feedback from extra-striate cortical areas. The physiology of so-called con-
textual effects in the visual system continue to serve as the main phenomenon
to be explained. He is also interested in testing if theories and models that
have been developed for the visual system generalize. For example, can we
shed some light into (currently) conceptually more fuzzy fields such as, for
example, ‘the neuroscience of the self’ using theories from neural coding,
learning and inference? Method-wise, he is applying mathematical modeling
and simulations, but he also engaged in standardization efforts for describing
neuronal models using strictly declarative approaches. Continuing that work,
xx Research Interests
he is currently applying core computer science concepts such as process

algebras and model checking to simulated neuronal systems.
Frances Skinner is a Senior Scientist in the Division of Fundamental Neu-

robiology at the Toronto Western Research Institute, University Health
Network, and a Professor at the University of Toronto.
Her research interests are in determining the mechanisms underlying the
dynamic output of neurons and neuronal networks. Her approach involves the
use, development, and analysis of mathematical models that have intimate
links with experimental data and analysis. The present focus is on oscillatory
activities in the hippocampus and inhibitory cells.
Volker Steuber is a Reader (Associate Professor) in Biocomputation and

Head of the Biocomputation Research Group in the Science and Technology
Research Institute at the University of Hertfordshire. His research involves
the development of multiscale models of neurons and neuronal networks and
the application of machine-learning techniques to analyze biological data.
One of his main research interests is synaptic plasticity and information
processing in the cerebellum.
Peter Tass develops stimulation techniques with computational methods. He

studied medicine (M.D., Universities of Ulm and Heidelberg, Germany),
physics (Ph.D., University of Stuttgart, Germany), and mathematics
(diploma, University of Stuttgart, Germany) and made a Habilitation in
physiology (RWTH Aachen University, Germany). He is the director of the
Institute of Neuroscience and Medicine—Neuromodulation at Juelich
Research Center, Juelich, Germany, and full professor in neuromodulation
and head of the research division for neuromodulation at Cologne University,
Germany. Furthermore, Peter Tass is consulting professor at the Department
of Neurosurgery, Stanford University, Stanford, CA, USA.
Matthew Tresch is Associate Professor in Biomedical Engineering, Physical

Medicine and Rehabilitation, and Physiology at Northwestern University. He
is interested in the interaction between the neural control of movement and
peripheral limb mechanics using rodent models and drawing on techniques
from neurophysiology, behavioral analyses, and biomechanics.
Sharmila Venugopal is a Research Faculty in the Department of Integrative

Biology and Physiology at UCLA and a part-time Teaching Faculty in the
College of Engineering at California State University, Long Beach. She has a
bachelor’s degree in Electrical Engineering, a master’s degree in Computer
Engineering, and a Ph.D. in Neuroscience. Her research interests include the
computational modeling of neurological disease dynamics, the experimental
investigation of progressive changes in neural excitability in neurodegener-
ative disease models, and the development of software and hardware signal–
and image-processing tools for neuroscience applications.
Research Interests xxi
Wei Wang is an assistant professor in the Department of Physical Medicine

and Rehabilitation, Department of Bioengineering, and Clinical and Transla-
tional Science Institute at the University of Pittsburgh. Dr. Wang’s primary
research interests include motor system neurophysiology, brain–computer
interfaces, and neurorehabilitation.
Dr. Wang earned his Clinical Medicine degree from Peking University
Health Science Center in 1999 and his Ph.D. degree in Biomedical Engineer-
ing from Washington University in St. Louis in 2006. He then worked as a
senior scientist for St. Jude Medical, Inc., an implantable medical device
company, between 2006 and 2007, during which he had two U.S. patents for
implantable devices. Dr. Wang then joined the University of Pittsburgh in
2007 as an assistant professor in the Department of Physical Medicine and
Rehabilitation. Dr. Wang was an associate scientific advisor to the Science
Translational Medicine magazine in 2011. He is a reviewer for the Journal of
Neural Engineering, Neural Information Processing Systems, IEEE Trans-
actions on Neural Systems and Rehabilitation Engineering, and IEEE Engi-
neering in Medicine and Biology.
Dr. Wang currently holds a 5-year Clinical Research Scholar Career
Award from the National Institutes of Health of the U.S.A. In addition, his
research projects are currently supported by National Institute of Neurolog-
ical Disorders and Stroke (NINDS, part of NIH), Craig H. Neilsen Founda-
tion, and other government agencies. His research is also supported by the
Clinical and Translational Science Institute and the University of Pittsburgh
Medical Center.
Douglas Weber’s general research area is Neural Engineering including

studies of functional electrical stimulation (FES), activity-based neuromotor
rehabilitation, neural coding, and neural control of prosthetic devices. He has
expertise in a range of advanced techniques for neurophysiology and biome-
chanics research including 3D motion analysis, electromyography,
multichannel neural recording and stimulation, human magnetoencephalog-
raphy (MEG), and human electrocorticography (ECoG). He has experience
with a variety of animal models including rats, cats, and nonhuman primates
as well as human subject testing in laboratory and clinical studies. Active
projects in his lab include the development of motor and sensory neural
interfaces for controlling and sensing prosthetic limbs.
Douglas Weber is currently splitting his time (until August 2015) between
the University of Pittsburgh and the Defense Advanced Research Projects
Agency (DARPA), where he is a program manager in the Biological Tech-
nology Office (BTO). At DARPA, he is responsible for managing several
neurotechnology programs including RE-NET, HAPTIX, and ElectRx.
Brian Wodlinger received a B.A.Sc. degree in Engineering Science from the

University of Toronto. He received his Ph.D. in Peripheral Nerve Interfaces in
2010 from Case Western Reserve University in Cleveland, OH, in the Neural
Engineering Center and went on to complete postdoctoral work in
Brain–Computer Interfaces at the University of Pittsburgh Human Rehabil-
itation and Neural Engineering Lab.
Acknowledgements
The last decade of the twentieth century, the “Decade of the Brain,” saw
Computational Neuroscience emerge as an essential complement to experi-
mental neuroscience. The first decade of the “twenty-first century” has seen
the additional emergence of the field of Neural Engineering dedicated to
important practical implementations in Computational Neuroscience such
as neuroprosthetics and clinical neurostimulation. It is thus an opportune
time to bring forth a major dynamic living reference work on “Computational
Neuroscience” to convey the diversity of information needed for conducting
basic and applied computational neuroscience research and design and devel-
opment of novel neurotechnologies.
This seminal Encyclopedia of Computational Neuroscience was initiated
under the pseudonym of “Project Bhaskara.” Bhaskara I (c. 600–c. 680) was
an ancient Indian mathematician. He was apparently the first to write the
numbers in the Hindu-Arabic decimal system and use the symbol for zero (the
language of computers). He used figures to present the positional or place-
value decimal system (Coding) and linked mathematical theory with obser-
vational astronomy. He commented on an ancient treatise, the Aryabhatiya,
and wrote a major eight chapter encyclopedia, the Mahabhaskariya. With this
inspiration, Project Bhaskara has now culminated in a twenty-first-century
collaborative effort bringing together diversity of knowledge in the pursuit
and synthesis of computational neuroscience.
In 2010, Ann Avouris, Senior Editor, Neuroscience at Springer Inc.,
approached Ranu Jung, then President of the Organization for Computational
Neuroscience, to consider editing a major reference work. She and Dieter
Jaeger, then Vice-President of the organization, agreed to shepherd the
project because of the willingness of the members of the Board to support
the Editors-in-Chief in creation of this work. It was only then, with their
enthusiastic support, that an international effort was launched.
The work could not have been compiled and completed without the
extensive and thoughtful actions of many. Fifty-three section editors spent
several months developing the topics to be covered under their sections. They
invited authors from around the world to contribute to each of their sections,
and as of today there are 545 entries. Many biologists, engineers, mathema-
ticians, neuroscientists, and physicists have contributed entries. They have
provided definitions, descriptions, and references. The section editors have
provided overviews to intertwine the entries and provided section overviews
xxiii
xxiv Acknowledgements
to the readers. It is through this collective engagement that this first edition is
being launched.
An effort such as this rests on the discoveries, development of methodol-
ogies, and application of ideas by many. Investment to support the birth and
implementation of these ideas is crucial and rests on the championship of
peoples and institutions with foresight. In the USA, Kenneth Whang, from the
National Science Foundation, and Dennis Glanzman and Yuan Liu, from the
National Institutes of Health, in particular have provided unrelenting support
for computational neuroscience from within the funding agencies. The early
establishment of the “Bernstein Centers” in Germany and the International
Neuroinformatics Coordinating Facility (INCF) with its Secretariat in
Sweden also played a crucial role in globalizing the field. Consequent to the
steadfast investments such as these, as of 2014 there are formal mechanisms
to support global collaborations between USA and France, USA and
Germany, USA and Israel, and the 17 INCF national nodes in countries in
Europe, USA, and Asia in computational neuroscience.
The entire project would not have been possible without the momentous
ongoing support of the Springer Editorial Staff based in the USA and the
Federal Republic of Germany. Besides championing the effort, Ann Avouris
shepherded the project through most of its three-year journey. Early on, Jutta
Jaeger-Hamers, Daniela Graf, and Melanie Thanner helped shape the sec-
tions. Daniela and Melanie have tirelessly provided many an answer to
questions from authors and editors and implemented suggestions and
changes. Taking over for Ann in 2013, Simina Calin has taken the responsi-
bility to bring the encyclopedia to the finish line and pave the groundwork for
the next edition. It has been a pleasure working with the whole team and we
look forward to continued engagement.
We (Dieter and Ranu) are not only sincerely thankful to all for their
confidence in us, but also for affording us the opportunity to develop a
stronger friendship as we have communicated and discussed the birth and
growth of the encyclopedia over the 3 years. We look forward to the dynamic
growth of the project and a second edition to follow.
Miami, Florida and Atlanta, Georgia

31 October 2014
Contributors
James J. Abbas School of Biological and Health Systems Engineering,

Arizona State University, Tempe, AZ, USA
Louise C. Abbott Department of Veterinary Integrative Biosciences, Texas
A&M University, College Station, TX, USA
Mohamed N. Abdelghani Department of Biomedical Engineering, Florida
International University, Miami, FL, USA
Moshe Abeles The Leslie and Susan Gonda(Goldschmied) Multidisiplinary
Brain Research Center, Bar-Ilan University, Ramat-Gan, Israel
Michael Ackermann Stanford University, Stanford, CA, USA
Trevor Agus Equipe Audition, Département d’Études Cognitives, École
Normale Supérieure, Paris, France
Jessica L. Allen The W. H. Coulter Department of Biomedical Engineering,
Emory University and Georgia Institute of Technology, Atlanta, GA, USA
Brendan Z. Allison Electrical and Computer Engineering, Old Dominion
University, Norfolk, VA, USA
Heike Althen Cognitive Neuroscience Research Group, University of
Barcelona, Barcelona, Spain
Shun-ichi Amari Laboratory for Mathematical Neuroscience, RIKEN
Brain Science Institute, Wako-shi, Japan
Thomas J. Anastasio Department of Molecular and Integrative Physiology,
and Beckman Institute, University of Illinois at Urbana-Champaign, Urbana,
IL, USA
Costas A. Anastassiou Allen Institute for Brain Science, Seattle, WA, USA
Sara Arganda Centre de Recherches sur la Cognition Animale, Université
de Toulouse, Toulouse, France
Giorgio Ascoli Center for Neural Informatics, Structures, and Plasticity,
Krasnow Institute for Advanced Study, George Mason University, Fairfax,
VA, USA
xxv
xxvi Contributors
Swee T. Aw Central Clinical School, University of Sydney, Sydney, NSW,

Australia
Institute of Clinical Neuroscience, Royal Prince Alfred Hospital,
Camperdown, Sydney, NSW, Australia
Lauren Ayton Macular Research Unit, Centre for Eye Research Australia,
The University of Melbourne, East Melbourne, VIC, Australia
Islam S. Badreldin Electrical and Computer Engineering, Michigan State

University, East Lansing, MI, USA
Steven M. Baer School of Mathematical and Statistical Sciences, Arizona

State University, Tempe, AZ, USA
Sonya Bahar Center for Neurodynamics, University of Missouri at

St. Louis, St. Louis, MO, USA
Gerold Baier Cell and Developmental Biology, Faculty of Life Sciences,

University College London, London, UK
Sylvain Baillet McConnell Brain Imaging Centre, Montréal Neurological

Institute, McGill University, Montréal, QC, Canada
Wyeth Bair Department of Biological Structure, University of Washington,

Seattle, WA, USA
Rembrandt Bakker Nijmegen and Institute of Neuroscience and Medicine

(INM–6), Donders Institute, Radboud University, Nijmegen, The Netherlands
J€ulich Research Centre, J€ulich, Germany
Pragathi Priyadharsini Balasubramani Department of Biotechnology,

Indian Institute of Technology, Chennai, India
Karthikeyan Balasubramanian Department of Organismal Biology and

Anatomy, University of Chicago, Chicago, IL, USA
Anita Bandrowski NIF Project Lead, University of California, San Diego,

La Jolla, CA, USA
Sergio Barbieri Unità Operativa di Neurofisiopatologia Clinica,

Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policlinico, Milan, Italy
William Barnett The Neuroscience Institute, Georgia State University,

Atlanta, GA, USA
Adam B. Barrett Sackler Centre for Consciousness Science and Depart-

ment of Informatics, University of Sussex, Brighton, UK
John Barrett Institute of Neuroscience, The Medical School, Newcastle

University, Newcastle-upon-Tyne, UK
Thomas M. Bartol Neurobiology Laboratory, Salk Institute for Biological

Studies, La Jolla, CA, USA
Contributors xxvii
Maxim Bazhenov Department of Cell Biology and Neuroscience, Univer-

sity of California, Riverside, CA, USA
Claude Bédard Unit of Neuroscience Information and Complexity (UNIC),
Centre National de la Recherche Scientifique (CNRS), Gif-sur-Yvette, France
James A. Bednar Institute for Adaptive and Neural Computation, School of
Informatics, The University of Edinburgh, Edinburgh, UK
David Beeman University of Colorado, Boulder, CO, USA
Ulrik R. Beierholm Centre for Computational Neuroscience and Cognitive
Robotics, University of Birmingham, Birmingham, UK
John Bekkers Australian National University, Canberra, Australia
Jacques Bélair Département de Mathématiques et de Statistique, Université
de Montréal, Montréal, QC, Canada
Jan Benda Institute for Neurobiology, Eberhard Karls University,
T€ubingen, Germany
Paul R. Benjamin Sussex Neuroscience, School of Life Sciences, Univer-
sity of Sussex, Brighton, UK
Sliman J. Bensmaia Department of Organismal Biology and Anatomy,
University of Chicago, Chicago, IL, USA
Theodore W. Berger Department of Biomedical Engineering, Center for
Neural Engineering, University of Southern California, Los Angeles,
CA, USA
Ari Berkowitz Department of Biology, Cellular & Behavioral Neurobiol-
ogy Graduate Program, University of Oklahoma, Norman, OK, USA
Richard Bertram Department of Mathematics, Florida State University,
Tallahassee, FL, USA
Jason Berwick Department of Psychology, The University of Sheffield,
Sheffield, UK
Matthias Bethge Werner Reichardt Centre for Integrative Neuroscience,
University of T€ubingen and Max Planck Institute for Biological Cybernetics,
Anne Beuter University of Bordeaux, Bordeaux, France
Narendra Bhadra Neural Engineering Center, Department of Biomedical
Engineering, Case Western Reserve University, Cleveland, OH, USA
Niloy Bhadra Case Western Reserve University, Cleveland, OH, USA
Upinder S. Bhalla National Centre for Biological Sciences, Tata Institute of
Fundamental Research, Bangalore, India
Arjun Bharioke Janelia Farm Research Campus, Howard Hughes Medical
Institute, Ashburn, VA, USA
xxviii Contributors
Vincent A. Billock National Research Council, US Air Force Research

Laboratory, Wright Patterson Air Force Base, OH, USA
Marc D. Binder Department of Physiology & Biophysics, School of
Medicine, University of Washington, Seattle, WA, USA
Kim T. Blackwell Molecular Neuroscience Department, Krasnow Institute
for Advanced Study, George Mason University, Fairfax, VA, USA
Brian Blais Department of Science and Technology, Bryant University,
Smithfield, RI, USA
Institute for Brain and Neural Systems, Brown University, Providence, RI, USA
Stefano Boccaletti CNR-Institute of Complex Systems, Sesto Fiorentino,
Florence, Italy
Rafal Bogacz Department of Computer Science, University of Bristol,
Bristol, UK
Ingo Bojak School of Systems Engineering, University of Reading, Read-
ing, UK
Victoria Booth Departments of Mathematics and Anesthesiology, Univer-
sity of Michigan, Ann Arbor, MI, USA
Alla Borisyuk Department of Mathematics, University of Utah, Salt Lake
City, UT, USA
Alexander Borst Max-Planck-Institut f€ur Neurobiologie, Martinsried,
Germany
Amitabha Bose Department of Mathematical Sciences, New Jersey Institute
of Technology, Newark, NJ, USA
Mihail Bota Department of Biological Sciences, University of Southern
California, Los Angeles, CA, USA
Jonathan Bourget-Murray FRQS Groupe de Recherche en Neurobiologie
Comportementale/Center for Studies in Behavioral Neurobiology (CSBN),
Department of Exercise Science, Concordia University, Montréal, QC,
Canada
C.M. Program, Faculty of Medicine, McGill University, Montréal, QC,
Canada
Jean-Marie Bouteiller Department of Biomedical Engineering, University
of Southern California, Los Angeles, CA, USA
Douglas M. Bowden Department of Psychiatry and Behavioral Sciences,
School of Medicine, University of Washington, Seattle, WA, USA
James M. Bower Department of Computer Science, University of Califor-
nia, Santa Cruz, Santa Cruz, CA, USA
J. Braasch School of Architecture, Rensselaer Polytechnic Institute, Troy,
NY, USA
Contributors xxix
Chris Bradley Auckland Bioengineering Institute, University of Auckland,

Auckland, New Zealand
Almut Branner Neurobiology and Anatomy, Drexel University College of

Medicine and School of Biomedical Engineering and Health Systems, Philadel-
phia, PA, USA
Michael Breakspear Queensland Institute of Medical Research, Herston,

QLD, Australia
Royal Brisbane & Women’s Hospital, Herston, QLD, Australia
Steven L. Bressler Cognitive Neurodynamics Laboratory, Center for Com-

plex Systems and Brain Sciences, Department of Psychology, Florida Atlan-
tic University, Boca Raton, FL, USA
Paul Bressloff Department of Mathematics, University of Utah, Salt Lake

City, UT, USA
Romain Brette Institut de la Vision, INSERM, CNRS, Université Pierre et

Marie Curie, Paris, France
Jeroen J. Briaire ENT Department, Leiden University Medical Center,

Leiden, The Netherlands
Alan M. Brichta School of Biomedical Sciences and Pharmacy, Hunter

Medical Research Institute, The University of Newcastle, Callaghan, NSW,
Australia
Randall D. Britten Auckland Bioengineering Institute, University of

Auckland, Auckland, New Zealand
Michael Brosch Leibniz Institute for Neurobiology, Magdeburg, Germany
Emery N. Brown Institute for Medical Engineering and Science, Massa-

chusetts Institute of Technology, Cambridge, MA, USA
Department of Brain and Cognitive Sciences, Massachusetts Institute of
Technology, Cambridge, MA, USA
Department of Anesthesia, Critical Care and Pain Medicine, Massachusetts
General Hospital, Harvard Medical School, Boston, MA, USA
Andrew Brownlee Lenox Hill Hospital/NSLIJ, New York, NY, USA
Nicolas Brunel Departments of Statistics and Neurobiology, University of

Chicago, Chicago, IL, USA
Javier M. Buldú Universidad Rey Juan Carlos, Móstoles, Madrid, Spain
Robert Butera School of Electrical and Computer Engineering, Laboratory

for Neuroengineering, Georgia Institute of Technology, Atlanta, GA, USA
Jean-Marie Cabelguen Neurocentre Magendie, INSERM U 862 –

Bordeaux University, Bordeaux, France
xxx Contributors
Joana Cabral Theoretical and Computational Neuroscience Group, Center

for Brain and Cognition, Universitat Pompeu Fabra, Barcelona, Spain
Juan Luis Cabrera Laboratorio de Dinámica Estocástica, Centro de Fı́sica,
Instituto Venezolano de Investigaciones Cientı́ficas, Caracas, Venezuela
Morven Cameron Department of Anatomy and Cell Biology, School of
Medicine, University of Western Sydney, Sydney, Australia
Sue Ann Campbell Department of Applied Mathematics, University of
Waterloo, Waterloo, ON, Canada
Carmen C. Canavier Department of Cell Biology and Anatomy, LSU
Health Sciences Center, New Orleans, LA, USA
Robert Cannon Textensor Limited, Edinburgh, UK
Matteo Cantarelli Department of Neuroscience, Physiology and Pharma-
cology, University College London, London, UK
Jessica A. Cardin Department of Neurobiology and Kavli Institute, Yale
University, New Haven, CT, USA
Jose M. Carmena Helen Wills Neuroscience Institute, UC Berkeley,
Berkeley, CA, USA
Ted Carnevale Department of Neurobiology, Yale University School of
Medicine, New Haven, CT, USA
Thomas L. Carroll U.S. Naval Research Laboratory, Washington, DC, USA
Annie Castonguay Institut Universitaire en Santé Mentale de Québec,
Québec, Canada
V. Srinivasa Chakravarthy Department of Biotechnology, Indian Institute
of Technology, Chennai, India
Young-Hui Chang School of Applied Physiology, Georgia Institute of
Technology, Comparative Neuromechanics Laboratory, Atlanta, GA, USA
Vijayalakshmi Chelliah European Bioinformatics Institute (EMBL-EBI),
European Molecular Biology Laboratory, Hinxton, Cambridge, UK
Tim T. Chen Faculty of Pharmaceutical Sciences, University of British
Columbia, Vancouver, BC, Canada
Zhe Chen Department of Brain and Cognitive Sciences, Massachusetts
Institute of Technology, Cambridge, MA, USA
Department of Psychiatry, Neuroscience and Physiology, New York Univer-
sity School of Medicine, New York, NY, USA
Kei Cheung Center for Medical Informatics, Yale University School of
VA Connecticut Healthcare System, West Haven, CT, USA
Contributors xxxi
Margaret S. Cheung Department of Physics, University of Houston,

Houston, TX, USA
Elisabetta Chicca Faculty of Technology and Cognitive Interaction
Technology – Center of Excellence, Bielefeld University, Bielefeld, Germany
Dmitri B. Chklovskii Janelia Farm Research Campus, Howard Hughes
Medical Institute, Ashburn, VA, USA
Yoonsuck Choe Department of Computer Science and Engineering, Texas
G. Richard Christie Auckland Bioengineering Institute, University of
Colleen E. Clancy Department of Pharmacology, University of California,
Davis, Davis, CA, USA
Thomas A. Cleland Computational Physiology Lab, Department of
Psychology, Cornell University, Ithaca, NY, USA
Claudia Clopath Department of Bioengineering, Imperial College London,
London, UK
Dana Cohen Gonda Brain Research Center, Bar Ilan University,
Ramat-Gan, Israel
Albert Compte Systems Neuroscience, IDIBAPS, Barcelona, Spain
Diego Contreras Department of Neuroscience, School of Medicine,
University of Pennsylvania, Philadelphia, PA, USA
Hugo Cornelis University of Texas, San Antonio, TX, USA
Richard Courtemanche FRQS Groupe de Recherche en Neurobiologie
Canada
Jack D. Cowan Department of Mathematics, University of Chicago,
Gordon Center for Integrative Science, Chicago, IL, USA
Nelson Cowan Department of Psychological Sciences, University of
Missouri-Columbia, Columbia, MO, USA
Sharon Crook School of Mathematical and Statistical Sciences and School
of Life Sciences, Arizona State University, Tempe, AZ, USA
Anthony Cruickshank Institute of Adaptive and Neural Computation,
University of Edinburgh, Edinburgh, UK
Hermann Cuntz Ernst Str€ungmann Institute (ESI) for Neuroscience in
Cooperation with Max Planck Society, Frankfurt am Main, Germany
Institute of Clinical Neuroanatomy, Goethe University, Frankfurt am Main,
Germany
xxxii Contributors
Ian S. Curthoys School of Psychology, University of Sydney, Sydney,

NSW, Australia
Vassilis Cutsuridis Institute of Molecular Biology and Biotechnology,
Foundation for Research and Technology - Hellas, Heraklion, Crete, Greece
Gennady Cymbalyuk The Neuroscience Institute, Georgia State Univer-
sity, Atlanta, GA, USA
Gislin Dagnelie Johns Hopkins University School of Medicine, Baltimore,
MD, USA
Markus A. Dahlem Department of Physics, Humboldt University of Berlin,
Berlin, Germany
Chenkai Dai Vestibular NeuroEngineering Laboratory, Johns Hopkins Uni-
versity, Baltimore, MD, USA
Sriraman Damodaran Krasnow Institute for Advanced Study, George
Mason University, Fairfax, VA, USA
Simon M. Danner Center of Medical Physics and Biomedical Engineering,
Medical University of Vienna, Vienna, Austria
Institute of Analysis and Scientific Computing, Vienna University of Tech-
nology, Vienna, Austria
Ran Darshan Interdisciplinary Center for Neural Computation (ICNC),
The Hebrew University of Jerusalem, Jerusalem, Israel
Andrew P. Davison Unité de Neurosciences, Information et Complexité
(UNIC), Institut de Neurobiologie Alfred Fessard, Centre National de la
Recherche Scientifique (CNRS), Gif-sur-Yvette, France
Licurgo de Almeida Department of Neurobiology and Behavior, Cornell
University, Ithaca, NY, USA
Yves De Koninck Institut Universitaire en Santé Mentale de Québec and
Department of Psychiatry and Neuroscience, Université Laval, Québec,
Canada
Gonzalo G. de Polavieja Instituto Cajal, Consejo Superior de
Investigaciones Cientı́ficas, Madrid, Spain
Horace T. Deans UTSA Neurosciences Institute, The University of Texas at
San Antonio, San Antonio, TX, USA
Gustavo Deco Theoretical and Computational Neuroscience Group, Center
for Brain and Cognition, Universitat Pompeu Fabra, Barcelona, Spain
ICREA, Institució Catalana de Recerca i Estudis Avançats, Barcelona, Spain
Patrick Degenaar School of Electrical Electronic and Computer Engineer-
ing, Newcastle University, Newcastle-upon-Tyne, UK
Contributors xxxiii
Charles Della Santina Vestibular NeuroEngineering Laboratory, Johns

Hopkins University, Baltimore, MD, USA
Sophie Deneve Group for Neural Theory, École Normale Supérieure Paris,
Paris, France
Susan Denham Cognition Institute and School of Psychology, Plymouth
University, Plymouth, Devon, UK
Evelyne Deplazes School of Chemistry and Molecular Biosciences,
University of Queensland, Brisbane, QLD, Australia
Institute for Molecular Bioscience, University of Queensland, Brisbane,
QLD, Australia
Alain Destexhe Unit of Neuroscience Information and Complexity (UNIC),
Centre National de la Recherche Scientifique (CNRS), Gif-sur-Yvette, France
Annaelle Devergnas Yerkes National Primate Research Center, Emory
University, Atlanta, GA, USA
Udall Center of Excellence in Parkinson’s Disease Research, Emory Univer-
Ian M. Devonshire School of Life Sciences, Queen’s Medical Centre,
Nottingham, UK
Mukesh Dhamala Physics and Astronomy, Neuroscience Institute, Georgia
State University, Atlanta, GA, USA
Markus Diesmann Institute of Neuroscience and Medicine (INM-6) and
Institute for Advanced Simulation (IAS-6), J€ulich Research Centre and
JARA, J€ulich, Germany
Medical Faculty, RWTH Aachen University, Aachen, Germany
Alexander Dimitrov Department of Mathematics, Washington State Uni-
versity, Vancouver, WA, USA
Mingzhou Ding The J. Crayton Pruitt Family Department of Biomedical
Engineering, University of Florida, Gainesville, FL, USA
Markus Dittrich Biomedical Applications Groups, Pittsburgh
Supercomputing Center, Carnegie Mellon University, Pittsburgh, PA, USA
Ramana Dodla University of Texas at San Antonio, San Antonio, TX, USA
Socrates Dokos Graduate School of Biomedical Engineering, University of
New South Wales, Sydney, NSW, Australia
Rodney Douglas Institute of Neuroinformatics, University of Z€urich and
ETH Z€urich, Z€urich, Switzerland
Nicolas Doyon Institut Universitaire en Santé Mentale de Québec and
Department of Mathematics and Statistic, Université Laval, Québec, Canada
xxxiv Contributors
Guillaume Drion Department of Electrical Engineering and Computer Sci-

ence, University of Liège, Liège, Belgium
Laboratory of Pharmacology and GIGA Neurosciences, University of Liège,
Liège, Belgium
Shaul Druckmann Janelia Farm Research Campus, Howard Hughes Medi-
cal Institute, Ashburn, VA, USA
Niraj Dudani National Centre for Biological Sciences, Tata Institute of
Geneviève Dupont Theoretical Chronobiology Unit, Université Libre de
Bruxelles, Brussels, Belgium
Dominique M. Durand Department of Biomedical Engineering, Physiol-
ogy, Biophysics and Neurosciences, Neural Engineering Center, Case
Western Reserve University, Cleveland, OH, USA
urr Bielefeld University, Bielefeld, Germany
Volker D€
Stuart Edelstein Babraham Institute, Cambridge, UK
Victor Reggie Edgerton Department of Integrative Biology and Physiol-
ogy, and Brain Research Institute, University of California, Los Angeles,
CA, USA
Martin Egelhaaf Neurobiology & CITEC, Bielefeld University, Bielefeld,
Germany
Jos J. Eggermont Department of Physiology and Pharmacology, University
of Calgary, Calgary, AB, Canada
Department of Psychology, University of Calgary, Calgary, AB, Canada
Stephen J. Eglen Department of Applied Mathematics and Theoretical
Physics, Cambridge Computational Biology Institute, University of Cam-
bridge, Cambridge, UK
Simon B. Eickhoff Institute of Neuroscience and Medicine (INM-1),
Research Centre J€ulich, and Institute for Clinical Neuroscience and Medical
Psychology, Heinrich-Heine University, D€usseldorf, Germany
Gaute T. Einevoll Department of Mathematical Sciences and Technology,
Norwegian University of Life Sciences, Ås, Norway
Jennifer Stine Elam Department of Anatomy & Neurobiology, School of
Medicine, Washington University, St. Louis, MO, USA
Sherif M. Elbasiouny Departments of Neuroscience, Cell Biology and
Physiology (NCBP) and Biomedical, Industrial and Human Factors Engi-
neering (BIE), Wright State University, Dayton, OH, USA
Jordan Engbers Hotchkiss Brain Institute, University of Calgary, Calgary,
AB, Canada
Contributors xxxv
Dominique Engel GIGA-Neurosciences, University of Liège, Liège,

Belgium
Crystal T. Engineer School of Behavioral and Brain Sciences, The Univer-
sity of Texas at Dallas, Richardson, TX, USA
G. Bard Ermentrout Department of Mathematics, University of Pittsburgh,
Pittsburgh, PA, USA
Udo Ernst Department of Computational Neuroscience, Institute for Theo-
retical Physics, University of Bremen, Bremen, Germany
Carles Escera Institute for Brain, Cognition and Behavior (IR3C), Univer-
sity of Barcelona, Barcelona, Catalonia, Spain
Rebekah Evans Krasnow Institute for Advanced Study, George Mason
University, Fairfax, VA, USA
James R. Faeder Department of Computational and Systems Biology,
University of Pittsburgh School of Medicine, Pittsburgh, PA, USA
Michael Farries Department of Biology, University of Texas at San
Antonio, San Antonio, TX, USA
Sarah Feldt Muldoon Department of Bioengineering, University of
Pennsylvania, Philadelphia, PA, USA
Katie A. Ferguson Toronto Western Research Institute, University Health
Network, Toronto, ON, Canada
Physiology, University of Toronto, Toronto, ON, Canada
Eduardo Fernandez Department of Histology and Institute of Bioengineer-
ing, University Miguel Hernández, Elche (Alicante), Spain
Grant M. Fiddyment Graduate Program for Neuroscience, Boston Univer-
sity, Boston, MA, USA
Lee E. Fisher Department of Physical Medicine & Rehabilitation, Univer-
sity of Pittsburgh, Pittsburgh, PA, USA
Nicolas Fourcaud-Trocmé Center for Research in Neuroscience of Lyon,
CNRS UMR5292 - INSERM U1028, Université Lyon 1, Lyon, France
David M. Fox Department of Biological Sciences, New Jersey Institute of
Technology, Rutgers University Newark, Newark, NJ, USA
P. Mickle Fox Research Imaging Institute, University of Texas Health
Science Center, San Antonio, TX, USA
Peter T. Fox Research Imaging Institute, University of Texas Health
E. Paxon Frady Department of Neurosciences, UC San Diego, La Jolla,
CA, USA
xxxvi Contributors
Mitch Frankel Salt Lake City, UT, USA

Ariana Frederick FRQS Groupe de Recherche en Neurobiologie
Canada
David S. Freedman Department of Electrical and Computer Engineering,
Boston University, Boston, MA, USA
Johan H. M. Frijns ENT Department, Leiden University Medical Center,
Leiden, The Netherlands
Armin Fuchs Center for Complex Systems and Brain Sciences, Florida
Atlantic University, Boca Raton, USA
Romulo Fuentes Edmond and Lily Safra International Neuroscience
Institute, Natal, Brazil
Fabrizio Gabbiani Department of Neuroscience, Baylor College of
Medicine, Houston, TX, USA
Martin Garwicz Department of Experimental Medical Science, Neuronano
Research Center, Lund University, Lund, Sweden
Sonia Gasparini Neuroscience Center, LSUHSC, New Orleans, LA, USA
Michael C. Gastpar University of California, Berkeley, CA, USA
Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
Apostolos P. Georgopoulos Graduate Program in Biomedical Informatics
and Computational Biology, University of Minnesota, Minneapolis, MN, USA
Department of Neuroscience, University of Minnesota, Minneapolis, MN, USA
Yury Gerasimenko Pavlov Institute of Physiology, St. Petersburg, Russia
Richard C. Gerkin School of Life Sciences, Arizona State University,
Tempe, AZ, USA
Sam Gershman Department of Psychology and Neuroscience Institute,
Princeton University, Princeton, NJ, USA
Samuel J. Gershman Department of Brain and Cognitive Sciences,
Massachusetts Institute of Technology, Cambridge, MA, USA
George L. Gerstein Department of Neuroscience, University of
Marc-Oliver Gewaltig Blue Brain Project, École Polytechnique Fédéral de
Lausanne, Lausanne, Switzerland
William Gibson School of Mathematics and Statistics, University of
Sydney, Sydney, NSW, Australia
Contributors xxxvii
Daniel T. Gillespie Dan T Gillespie Consulting, Castaic, CA, USA
Benoı̂t Girard Sorbonne Universités, UPMC Univ Paris 06, Paris, France
CNRS, Paris, France
Simon F. Giszter Neurobiology and Anatomy, Drexel University College of

Medicine and School of Biomedical Engineering and Health Systems,
Philadelphia, PA, USA
Michele Giugliano Theoretical Neurobiology and Neuroengineering

Laboratory, Department of Biomedical Sciences, University of Antwerp,
Wilrijk, Belgium
Department of Computer Science, University of Sheffield, Sheffield, UK
Brain Mind Institute, Swiss Federal Institute of Technology of Lausanne,
Lausanne, Switzerland
Padraig Gleeson Department of Neuroscience, Physiology and

Pharmacology, University College London, London, UK
Jean-Marc Goaillard Inserm, UMR_S 1072, Marseille, France

Aix–Marseille Université, UNIS, Marseille, France
Joshua A. Goldberg Department of Medical Neurobiology, Institute for

Medical Research Israel-Canada, Faculty of Medicine, The Hebrew Univer-
sity of Jerusalem, Jerusalem, Israel
Mark S. Goldman Center for Neuroscience, Department of Neurobiology,

Physiology, & Behavior, Department of Ophthalmology & Vision Science,
University of California, Davis, CA, USA
Jorge Golowasch Federated Department of Biological Sciences, New

Jersey Institute of Technology/Rutgers University, Newark, NJ, USA
Geoffrey J. Goodhill Queensland Brain Institute and School of Mathemat-

ics and Physics, University of Queensland, St Lucia, QLD, Australia
Dan F. M. Goodman Department of Otology and Otolaryngology, Harvard

Medical School, Boston, MA, USA
Andrei Gorea Laboratoire Psychologie de la Perception, Université Paris

Descartes and CNRS, Paris, France
Bruce Graham University of Stirling, Stirling, UK
Joseph Graham Blue Brain Project, École Polytechnique Fédérale de

Thom Griffith Department of Engineering Mathematics, University of

Bristol, Bristol, UK
xxxviii Contributors
Warren M. Grill Department of Biomedical Engineering, Duke University,

Durham, NC, USA
Department of Neurobiology, Duke University Medical Center, Durham,
NC, USA
Department of Surgery, Duke University Medical Center, Durham, NC, USA
Sabine Grimm Cognitive Neuroscience Research Group, University of

Barcelona, Barcelona, Spain
Sonja Gr€un Lab for Statistical Neuroscience, Institute of Neuroscience and

Medicine (INM–6) and Institute for Advanced Simulation (IAS–6), J€ulich
Research Centre and JARA, J€ulich, Germany
Theoretical Systems Neurobiology, RWTH Aachen University, Aachen,
Germany
Raúl Guantes Department of Condensed Matter Physics, Materials Science

Institute ‘Nicolás Cabrera’ and Institute of Condensed Matter Physics
(IFIMAC), Universidad Autónoma de Madrid; Cantoblanco, Madrid, Spain
C. Guerrier École Normale Supérieure, Institute for Biology, IBENS,

INSERM 1024 and CNRS Group of Computational Biology and Applied
Mathematics, Paris, France
University Paris 6, Laboratoire Jacques–Louis Lions, Paris, France
John J. Guinan Eaton Peabody Laboratories, Massachusetts Eye and Ear

Infirmary, Harvard Medical School, Boston, MA, USA
unay Department of Biology, Emory University, Atlanta, GA, USA

Cengiz G€
Nitin Gupta National Institutes of Health, NICHD, Bethesda, MD, USA
Robert G€ utig Theoretical Neuroscience, Max-Planck-Institut f€ur

Experimentelle Medizin, Göttingen, Germany
Boris Gutkin Group for Neural Theory, Laboratoire de Neurosciences Cog-

nitives (LNC), Département d’Études Cognitives, École Normale Supérieure,
Paris, France
National Research University Higher School of Economics, Center for
Cognition and Decision Making, Moscow, Russia
Vincent Hakim Laboratoire de Physique Statistique, Ecole Normale

Supérieure, CNRS, Paris, France
Geir Halnes Department of Mathematical Sciences and Technology,

Norwegian University of Life Sciences, Ås, Norway
Albert W. Hamood Brandeis University, Waltham, MA, USA
Roger Hardie Department of Physiology, Development and Neuroscience,

University of Cambridge, Cambridge, UK
Contributors xxxix
Matthew T. Harrison Division of Applied Mathematics, Brown University,

Providence, RI, USA
Ronald Harris-Warrick Department of Neurobiology and Behavior,
Cornell University, Ithaca, NY, USA
Nicholas G. Hatsopoulos Department of Organismal Biology and
Anatomy, University of Chicago, Chicago, IL, USA
Christian Hauptmann Institute of Neuroscience and Medicine –
Neuromodulation (INM–7), Research Center J€ulich, J€ulich, Germany
Andreas Hauri Institute of Neuroinformatics, ETHZ/UZH, Z€urich,
Switzerland
Christopher K. Hauser Department of Neurobiology and Anatomy, Wake
Forest School of Medicine, Winston-Salem, NC, USA
Biyu J. He National Institute of Neurological Disorders and Stroke,
National Institutes of Health, Bethesda, MD, USA
C. J. Heckman Feinberg School of Medicine, Northwestern University,
Chicago, IL, USA
Peter Heil Systems Physiology of Learning, Leibniz Institute for Neurobi-
ology, Magdeburg, Germany
Stanley Heinze Department of Biology, Lund University, Lund, Sweden
Matthias H. Hennig Institute for Adaptive and Neural Computation, School
of Informatics, University of Edinburgh, Edinburgh, UK
R. Matthias Hennig Department of Biology, Humboldt-Universit€at zu
Berlin, Berlin, Germany
Iain Hepburn Computational Neuroscience Unit, Okinawa Institute of Sci-
ence and Technology, Onna-son, Okinawa, Japan
Theoretical Neurobiology, University of Antwerp, Antwerp, Belgium
Pawel Andrzej Herman Computational Biology, KTH Royal Institute of
Technology, Stockholm, Sweden
Henning Hermjakob European Bioinformatics Institute (EMBL-EBI),
European Molecular Biology Laboratory, Hinxton, Cambridge, UK
Claus C. Hilgetag Department of Computational Neuroscience, University
Medical Center Hamburg–Eppendorf, Hamburg University, Hamburg,
Germany
Department of Health Sciences, Boston University, Boston, USA
Sean Hill Brain Mind Institute, École Polytechnique Fédérale de Lausanne,
Brian Hillen Adaptive Neural Systems Laboratory, Department of
Biomedical Engineering, Florida International University, Miami, FL, USA
xl Contributors
Michael Hines Department of Neurobiology, Yale University School of

Joerg F. Hipp Centre for Integrative Neuroscience, University of T€ubingen,
MEG-Center, University of T€ubingen, T€ubingen, Germany
Ursula S. Hofstötter Center of Medical Physics and Biomedical Engineer-
ing, Medical University of Vienna, Vienna, Austria
D. Holcman École Normale Supérieure, Institute for Biology, IBENS,
INSERM 1024 and CNRS Group of Computational Biology and Applied
Mathematics, Paris, France
University Paris 6, Laboratoire Jacques–Louis Lions, Paris, France
William R. Holmes Department of Biological Sciences, Ohio University,
Athens, OH, USA
Scott Hooper Department of Biological Sciences, Ohio University, Athens,
OH, USA
T. K. Horiuchi University of Maryland, College Park, MD, USA
Conor Houghton Department of Computer Science, University of Bristol,
Bristol, UK
Bryan Howell Department of Biomedical Engineering, Duke University,
Durham, NC, USA
Jan Hrabe Center for Advanced Brain Imaging, Nathan S. Kline Institute
for Psychiatric Research, Orangeburg, NY, USA
Sabina Hrabetova Department of Cell Biology, SUNY Downstate Medical
Center, Brooklyn, NY, USA
Michael Hucka Computing and Mathematical Sciences, California Institute
of Technology, Pasadena, CA, USA
Gemma Huguet Departament de Matemàtica Aplicada I, Universitat
Politècnica de Catalunya, Barcelona, Spain
Alexander C. Huk Departments of Neuroscience and Psychology,
Center for Perceptual Systems, The University of Texas at Austin, Austin,
TX, USA
Mark D. Humphries Faculty of Life Sciences, University of Manchester,
Manchester, UK
Peter Hunter Auckland Bioengineering Institute, University of Auckland,
Contributors xli
Phil Husbands Department of Informatics, Centre for Computational

Neuroscience and Robotics, University of Sussex, Brighton, UK
Axel Hutt Team NEUROSYS, INRIA CR Nancy – Grand Est, Villers-les-
Nancy, France
Quentin J. M. Huys Translational Neuromodeling Unit, Institute of
Biomedical Engineering, ETH Z€urich and University of Z€urich, Z€urich,
Switzerland
Department of Psychiatry, Psychotherapy and Psychosomatics, Hospital of
Psychiatry, University of Z€urich, Z€urich, Switzerland
Aapo Hyv€arinen Department of Computer Science, University of Helsinki,
Finland
Auke Jan Ijspeert EPFL École Polytechnique Fédérale de Lausanne,
Kazuo Imaizumi Department of Comparative Biomedical Sciences, School
of Veterinary Medicine, Louisiana State University, Baton Rouge, LA, USA
Robin A. A. Ince School of Psychology, Institute of Neuroscience and
Psychology, University of Glasgow, Glasgow, UK
Giacomo Indiveri Institute of Neuroinformatics, University of Z€urich and
ETH Z€urich, Z€urich, Switzerland
Junji Ito Institute of Neuroscience and Medicine (INM-6) and Institute for
Advanced Simulation (IAS-6), J€ulich Research Centre and JARA, J€ulich,
Germany
Andrew Jackson Institute of Neuroscience, Newcastle University, Newcas-
tle-upon-Tyne, UK
Jesse Jackson Department of Biology, Columbia University, New York,
NY, USA
Dieter Jaeger Department of Biology, Emory University, Atlanta, GA,
USA
M. Saleet Jafri School of Systems Biology and Krasnow Institute for
Advanced Studies, George Mason University, Fairfax, VA, USA
Nicolas Jaumard Department of Bioengineering, University of Pennsylva-
nia, Philadelphia, PA, USA
Peter Jedlicka Institute of Clinical Neuroanatomy, Goethe-University
Frankfurt, Frankfurt am Main, Germany
Joanna Jedrzejewska-Szmek Molecular Neuroscience Department,
VA, USA
xlii Contributors
Devin L. Jindrich Department of Kinesiology, California State University,

San Marcos, CA, USA
Viktor Jirsa Institut de Neurosciences des Systèmes, Faculté de Médecine,

UMR INSERM 1106, Aix-Marseille Université, Marseille, France
Mathew Jones Department of Neuroscience, University of Wisconsin,

Madison, WI, USA
Stephanie R. Jones Department of Neuroscience, Brown University, Pro-

vidence, RI, USA
Sébastien Joucla CNRS, Institute for Cognitive and Integrative Neurosci-

ence (INCIA), UMR 5287, Talence, France
Univ. Bordeaux, Institute for Cognitive and Integrative Neuroscience
(INCIA), UMR 5287, Talence, France
Ranu Jung Adaptive Neural Systems Laboratory, Department of Biomedi-

cal Engineering, Florida International University, Miami, FL, USA
Nick Juty European Bioinformatics Institute (EMBL-EBI), European

Molecular Biology Laboratory, Hinxton, Cambridge, UK
Mikko Juusola Department of Biomedical Science, University of Sheffield,

Sheffield, UK
Marcus Kaiser School of Computing Science, Newcastle University,

Newcastle-upon-Tyne, UK
Institute of Neuroscience, Newcastle University, Newcastle-upon-Tyne, UK
Tobias Kalenscher Department of Comparative Psychology, Institute of

Experimental Psychology, Heinrich-Heine University D€usseldorf,
D€usseldorf, Germany
Iwao Kanno Molecular Imaging Center, National Institute of Radiological

Sciences, Chiba, Japan
Robert E. Kass Carnegie Mellon University, Pittsburgh, PA, USA
Leor N. Katz Departments of Neuroscience and Psychology, Center for

Perceptual Systems, The University of Texas at Austin, Austin, TX, USA
Mitsuo Kawato Computational Neuroscience Laboratories, ATR Brain

Information Communication Research Laboratory Group, Kyoto, Japan
Leslie M. Kay Department of Psychology, Institute for Mind and Biology,

The University of Chicago, Chicago, IL, USA
David Kennedy Department of Psychiatry, Division of Neuroinformatics,

University of Massachusetts Medical Center, Worcester, MA, USA
Adam Kepecs Cold Spring Harbor Laboratory, Cold Spring Harbor, NY,
USA
Contributors xliii
John Keyser Department of Computer Science and Engineering, Texas

Michael P. Kilgard School of Behavioral and Brain Sciences, The Univer-
sity of Texas at Dallas, Richardson, TX, USA
Zachary P. Kilpatrick Department of Mathematics, University of Houston,
Houston, TX, USA
Sung Soo Kim Janelia Farm Research Campus, Howard Hughes Medical
Institute, Ashburn, VA, USA
Taegyo Kim Neurobiology and Anatomy, Drexel University College
of Medicine and School of Biomedical Engineering and Health Systems,
Frederick A. A. Kingdom Department of Ophthalmology, McGill Vision
Research, McGill University, Montréal, Canada
Tamara Kinzer-Ursem Weldon School of Biomedical Engineering, Purdue
University, West Lafayette, IN, USA
Tara Klassen Faculty of Pharmaceutical Sciences, University of British
Columbia, Vancouver, BC, Canada
Thomas R. Knösche Max Planck Institute for Human Cognitive and Brain
Sciences, Leipzig, Germany
Barbara Knowlton Department of Psychology, UCLA, Los Angeles,
CA, USA
Wonryull Koh School of Computing, College of Science and Technology,
University of Southern Mississippi, Hattiesburg, MS, USA
Stefan Kölsch Department of Educational Science and Psychology, Freie
Universit€at Berlin, Berlin, Germany
E. Korkotian Department of Neurobiology, Weizmann Institute of Science,
Rehovot, Israel
Jeanette Hellgren Kotaleski School of Computer Science and Communi-
cation, KTH Royal Institute of Technology, Stockholm, Sweden
Sonja A. Kotz School of Psychological Sciences, University of Manchester,
Manchester, UK
Max Planck for Human Cognitive and Brain Sciences, Leipzig, Germany
Robert Kozma Department of Mathematical Sciences, University of Mem-
phis, Memphis, TN, USA
Mark Kramer Department of Mathematics and Statistics, Boston Univer-
sity, Boston, MA, USA
Holger G. Krapp Department of Bioengineering, Imperial College London,
London, UK
xliv Contributors
André F. Krause Bielefeld University, Bielefeld, Germany

Giri P. Krishnan Department of Cell Biology and Neuroscience, University
of California, Riverside, CA, USA
William B. Kristan Division of Biology, UC San Diego, La Jolla, CA, USA
Dean Krusienski Electrical and Computer Engineering, Old Dominion
University, Norfolk, VA, USA
Alexey Kuznetsov Department of Mathematical Sciences, Indiana Univer-
sity & Purdue University Indianapolis, Indianapolis, IN, USA
Jaerock Kwon Department of Electrical and Computer Engineering,
Kettering University, Flint, MI, USA
Camille Laibe European Bioinformatics Institute (EMBL-EBI), European
Molecular Biology Laboratory, Hinxton, Cambridge, UK
Angela Marie Richmond Laird Department of Physics, Florida Interna-
tional University, Miami, FL, USA
Ilan Lampl Department of Neurobiology, Weizmann Institute of Science,
Rehovot, Israel
John Langton VisiTrend, Boston, MA, USA
Edward W. Large Department of Psychology, University of Connecticut,
Storrs, CT, USA
Yann Le Franc Theoretical Neurobiology and Neuroengineering, Univer-
sity of Antwerp, Wilrijk, Belgium
Nicolas Le Novère Babraham Institute, Babraham Research Campus,
Cambridge, UK
EMBL European Bioinformatics Institute, Cambridge, UK
Arthur Leblois Centre de Neurophysique, Physiologie et Pathologie (UMR
CNRS 8119), Université Paris Descartes, Paris, France
Charles C. Lee Department of Comparative Biomedical Sciences, School
of Veterinary Medicine, Louisiana State University, Baton Rouge, LA, USA
Robert Legenstein Institute for Theoretical Computer Science, Graz
University of Technology, Graz, Austria
Michel Lemay Department of Bioengineering, Temple University,
Scott Lempka Center for Neurological Restoration, Cleveland Clinic,
Cleveland, OH, USA
Rafael Levi The Whitney Laboratory for Marine Bioscience, University of
Florida, Augustine, FL, USA
Contributors xlv
John E. Lewis Department of Biology, University of Ottawa, Ottawa, ON,

Canada
Matthew Lewis Cornell University, Ithaca, NY, USA
Timothy J. Lewis Department of Mathematics, University of California,

Davis, CA, USA
Hualou Liang School of Biomedical Engineering, Drexel University,

Dawai Li Center for Cognitive Neuroscience, Duke University, Durham,

NC, USA
Guoshi Li Department of Psychology, Cornell University, Ithaca, NY, USA
Justin Lieber Committee on Computational Neuroscience, University of

Chicago, Chicago, IL, USA
David T. J. Liley Brain and Psychological Sciences Research Centre,

Swinburne University of Technology, Hawthorn, VIC, Australia
Rebecca Lim School of Biomedical Sciences and Pharmacy, Hunter Med-

ical Research Institute, The University of Newcastle, Callaghan, NSW,
Australia
Bernabe Linares-Barranco Instituto de Microelectrónica de Sevilla

(IMSE-CNM), CSIC and University of Sevilla, Sevilla, Spain
Daniele Linaro Theoretical Neurobiology and Neuroengineering Labora-

tory, Department of Biomedical Sciences, University of Antwerp, Wilrijk,
Belgium
Christiane Linster Computational Physiology Lab, Department of Neuro-

biology and Behavior, Cornell University, Ithaca, NY, USA
Peter Lipton Department of Neuroscience, University of Wisconsin,

Madison, WI, USA
Shih-Chii Liu University of Z€urich and ETH Z€urich, Z€urich, Switzerland
Gerald E. Loeb Department of Biomedical Engineering, University of

Southern California, Los Angeles, CA, USA
Nikos K. Logothetis Max Planck Institute for Biological Cybernetics,

Division of Imaging Science and Biomedical Engineering, University of
Manchester, Manchester, UK
Enrique A. Lopez-Poveda Instituto de Neurociencias de Castilla y Leon,

Instituto de Investigación Biomédica de Salamanca, Departamento de
Cirugı́a, Facultad de Medicina, University of Salamanca, Salamanca, Spain
xlvi Contributors
Nigel H. Lovell Graduate School of Biomedical Engineering, University of

New South Wales, Sydney, NSW, Australia
William W. Lytton Departments of Physiology & Pharmacology and
Neurology, SUNY Downstate Medical Center, Brooklyn, NY, USA
Department of Neurology, Kings County Hospital, Brooklyn, NY, USA
Hamish G. MacDougall School of Psychology, University of Sydney,
Sydney, NSW, Australia
Andre Machado Center for Neurological Restoration, Cleveland Clinic,
Cleveland, OH, USA
Christian K. Machens Molecular and Cell Biology, University of
California, Berkeley, USA
Jakob H. Macke Max Planck Institute for Biological Cybernetics and
Bernstein Center for Computational Neuroscience, T€ubingen, Germany
Reinoud Maex Department of Cognitive Sciences, École Normale
Supérieure, Paris, France
Margaret Y. Mahan Graduate Program in Biomedical Informatics and
Computational Biology, University of Minnesota, Minneapolis, MN, USA
Manuel S. Malmierca Department of Cellular Biology and Pathology,
Faculty of Medicine, University of Salamanca, Salamanca, Spain
Auditory Neuroscience Laboratory, Institute for Neuroscience of Castilla y
Léon, Salamanca, Spain
Paul B. Manis Otolaryngology/Head and Neck Surgery and Cell and
Molecular Physiology, University of North Carolina, Chapel Hill, NC, USA
Addolorata Marasco Department of Mathematics and Applications, Uni-
versity of Naples Federico II, Complesso Universitario di Monte S. Angelo,
Naples, Italy
Robert E. Marc John A. Moran Eye Center, University of Utah, Salt Lake
City, UT, USA
Sara Marceglia Dipartimento di Elettronica, Informazione e Bioingegneria,
Politecnico di Milano, Milan, Italy
Eve Marder Brandeis University, Waltham, MA, USA
Luis Marenco Center for Medical Informatics, Yale University School of
Toma M. Marinov Department of Biology, University of Texas at San
Vincenzo Marra Department of Cell Physiology and Pharmacology,
University of Leicester, Leicester, UK
Contributors xlvii
Diana Martinez Federated Department of Biological Sciences, New Jersey

Institute of Technology/Rutgers University, Newark, NJ, USA
Maryann E. Martone Department of Neuroscience, University of
California, San Diego, La Jolla, CA, USA
Stefano Masoli Section of Physiologysection, Department of Brain and
Behavioral Sciences, University of Pavia, Pavia, Italy
Paul Masset Watson School of Biological Sciences, Cold Spring Harbor
Laboratory, Cold Spring Harbor, NY, USA
T. M. Massoud Department of Electrical and Computer Engineering, Uni-
versity of Maryland, College Park, MD, USA
Victor Matveev Department of Mathematical Sciences, New Jersey Insti-
tute of Technology, Newark, NJ, USA
David Mayerich Beckman Institute for Advanced Science and Technology,
University of Illinois at Urbana-Champaign, Urbana, IL, USA
Robert A. McDougal Department of Neurobiology, Yale University School
of Medicine, New Haven, CT, USA
Meredith McGee Department of Biomedical Engineering, Duke Univer-
sity, Durham, NC, USA
Cameron McIntyre Departments of Biomedical Engineering, Neurology,
and Neurosurgery, Case Western Reserve University School of Medicine,
Cleveland, OH, USA
Robert McPeek Biological and Vision Sciences, SUNY College of Optom-
etry, State University of New York, New York, NY, USA
Tom McTavish Department of Cell and Developmental Biology, Univer-
sity of Colorado Denver, Aurora, CO, USA
Leonel E. Medina Department of Biomedical Engineering, Duke Univer-
sity, Durham, NC, USA
Ian A. Meinertzhagen Department of Psychology and Neuroscience,
Dalhousie University, Halifax, NS, Canada
Jack Mellor School of Physiology and Pharmacology, University of Bristol,
Bristol, UK
Nima Mesgarani Department of Electrical Engineering, The Fu Foundation
School of Engineering and Applied Science, Columbia University,
New York, NY, USA
Emmanuel A. Michaelides UTSA Neuroscience Institute, The University
of Texas at San Antonio, San Antonio, TX, USA
Americo Migliaccio Balance and Vision Laboratory, Neuroscience
Research Australia, University of New South Wales, Sydney, NSW, Australia
xlviii Contributors
Michele Migliore Department of Neurobiology, Yale University School of

Institute of Biophysics, National Research Council, Palermo, Italy
Shawn Mikula Department of Biomedical Optics, Max-Planck Institute for
Medical Research, Heidelberg, Germany
Robert Mill MRC Institute of Hearing Research, Nottingham, UK
Andrew K. Miller Auckland Bioengineering Institute, University of
John P. Miller Department of Cell Biology and Neuroscience, Montana
State University, Bozeman, MT, USA
Paul Miller Department of Biology and Brandeis University, Waltham,
MA, USA
Volen National Center for Complex Systems, Waltham, MA, USA
Perry Miller Center for Medical Informatics, Yale University School of
John Milton W.M. Keck Science Center, The Claremont Colleges,
Claremont, CA, USA
Karen Minassian Center of Medical Physics and Biomedical Engineering,
Medical University of Vienna, Vienna, Austria
Ashutosh Mohan Department of Physiology and Pharmacology, SUNY
Downstate Medical Center, Brooklyn, NY, USA
Department of Neuroscience, University of Wisconsin, Madison, WI, USA
Namrata Mohapatra Institute of Clinical Neuroanatomy, Goethe-
University Frankfurt, Frankfurt am Main, Germany
Behnam Molaee-Ardekani Clinical Neurophysiology Department,
Salengro Hospital, University of Lille 2, Lille, France
CHRU Salengro Hospital (Clinical Neurophysiology Center), Lille, France
Yaroslav Molkov Department of Mathematical Sciences, Indiana
University – Purdue University Indianapolis, Indianapolis, IN, USA
Gianluigi Mongillo Centre de Neurophysique, Physiologie, Pathologie
(CNPP), Université Paris Descartes, Paris, France
Centre National de la Recherche Scientifique, CNRS UMR 8119, Paris, France
George B. Moody Harvard-MIT Division of Health Sciences and Technology,
Massachusetts Institute of Technology, Cambridge, MA, USA
Contributors xlix
Rosalyn Moran Virginia Tech Carilion Research Institute, Roanoke, VA, USA
Bradley Department of Electrical and Computer Engineering, Virginia Tech,
Blacksburg, VA, USA
Department of Psychiatry and Behavioral Medicine, Virginia Tech Carilion
School of Medicine, Roanoke, VA, USA
Jonathan D. Moreno Tri–Institutional MD–PhD Program, Weill Cornell
Medical College, The Rockefeller University, Sloan–Kettering Cancer Insti-
tute, New York, NY, USA
Department of Physiology and Biophysics, Weill Medical College of Cornell
University, New York, NY, USA
John Morley Department of Physiology, School of Medical Sciences, Uni-
versity of New South Wales, Sydney, Australia
Department of Anatomy and Cell Biology, School of Medicine, University of
Western Sydney, Sydney, Australia
Kendall Morris College of Medicine, Molecular Pharmacology & Physio-
logy, University of South Florida, Tampa, FL, USA
Abigail Morrison Institute of Neuroscience and Medicine (INM-6),
Computational and Systems Neuroscience, J€ulich Research Centre, J€ulich,
Germany
Simulation Laboratory Neuroscience - Bernstein Facility for Simulation and
Database Technology, Institute for Advanced Simulation, Jülich Aachen
Research Alliance, Jülich Research Centre, J€ulich, Germany
Institute of Cognitive Neuroscience, Faculty of Psychology, Ruhr-University
Bochum, Bochum, Germany
Thomas M. Morse Department of Neurobiology, Yale University School of
Adonis K. Moschovakis Department of Basic Sciences, Faculty of Medi-
cine, University of Crete and Institute of Applied and Computational Mathe-
matics, FO.R.T.H, Heraklion, Crete, Greece
Marion Murray Department of Neurobiology and Anatomy, Drexel
University College of Medicine, Philadelphia, PA, USA
Farzan Nadim Federated Department of Biological Sciences, New Jersey
Institute of Technology/Rutgers University, Newark, NJ, USA
Department of Mathematical Sciences, New Jersey Institute of Technology,
Newark, NJ, USA
Tristan Nakagawa Computational Neuroscience Group, Center for Brain
and Cognition, Universitat Pompeu Fabra, Barcelona, Spain
Richard Naud Department of Physics, University of Ottawa, Ottawa, ON,
Canada
l Contributors
Martin Nawrot Neuroinformatik/Theoretische Neurobiologie, Institut f€ur

Biologie, Freie Universit€at Berlin, Berlin, Germany
Israel Nelken Edmond and Lily Safra Center for Brain Sciences,
The Hebrew University of Jerusalem, Jerusalem, Israel
Theoden I. Netoff Department of Biomedical Engineering, University of

Minnesota, Minneapolis, MN, USA
Susana R. Neves Department of Pharmacology and Systems Therapeutics,

Friedman Brain Institute, Icahn School of Medicine at Mount Sinai,
New York, NY, USA
Samuel A. Neymotin Department of Neurobiology, Yale School of Medi-

cine, New Haven, CT, USA
Department of Physiology and Pharmacology, SUNY Downstate Medical
Center, Brooklyn, NY, USA
David P. Nickerson Auckland Bioengineering Institute, University of

Poul M. F. Nielsen Auckland Bioengineering Institute, University of
Matthew F. Nolan Centre for Integrative Physiology, University of

Edinburgh, Edinburgh, UK
Sharon Norman School of Electrical and Computer Engineering, Laboratory

for Neuroengineering, Georgia Institute of Technology, Atlanta, GA, USA
Eva Nowak Institut f€ur Neuroinformatik, Ruhr-Universit€at Bochum,

Bochum, Germany
Thomas Nowotny Centre for Computational Neuroscience and

Robotics, School of Engineering and Informatics, University of Sussex,
Brighton, UK
Klaus Obermayer Neural Information Processing Group, Fakult€at IV,

Technische Universit€at Berlin, Berlin, Germany
Michael J. O’Donovan National Institute of Neurological Disorders and

Stroke, Bethesda, MD, USA
Hiroto Ogawa Department of Biological Science, Faculty of Science,

Hokkaido University, Sapporo, Japan
Michael Okun Institute of Neurology, University College London,

London, UK
Stephen O’Leary University of Melbourne, Melbourne, VIC, Australia

The Royal Victorian Eye and Ear Hospital, Melbourne, VIC, Australia
Contributors li
A. V. Olifer School of Science and Technology, Georgia Gwinnett College,

Lawrenceville, GA, USA
Damián Oliva Departamento de Ciencia y Tecnologı́a, Universidad

Nacional de Quilmes, CONICET, Buenos Aires, Argentina
Megan L. O’Mara School of Chemistry and Molecular Biosciences,

University of Queensland, Brisbane, QLD, Australia
School of Mathematics and Physics, University of Queensland, Brisbane,
QLD, Australia
Sorinel Adrian Oprisan Department of Physics and Astronomy, College of

Charleston, Charleston, SC, USA
Amy L. Orsborn Center for Neural Science, New York University, New
York, NY, USA
UCSF Joint Graduate Program in Bioengineering, UC Berkeley, Berkeley,
CA, USA
Michael O’Shea School of Life Sciences, University of Sussex, Brighton, UK
Ivan Osorio Medical Center and Ralph N. Adams Institute of Bioanalytical

Chemistry, University of Kansas, Kansas City, MO, USA
Karim G. Oweiss Electrical and Computer Engineering, Neuroscience and

Cognitive Science, Michigan State University, East Lansing, MI, USA
Andrew J. Oxenham Department of Psychology, University of Minnesota,

Minneapolis, MN, USA
Tohru Ozaki Institute of Statistical Mathematics, Tokyo, Japan
Daniel Palanker Department of Ophthalmology and Hansen Experimental

Physics Laboratory, Stanford University, Stanford, CA, USA
J. Matias Palva Neuroscience Center, University of Helsinki, Helsinki,

Finland
Stefano Panzeri Center for Neuroscience and Cognitive Systems, Istituto

Italiano di Tecnologia, Rovereto (Tn), Italy
Institute of Neuroscience and Psychology, University of Glasgow, Glasgow, UK
David Papo Center of Biomedical Technology, Madrid, Spain
Ruchi Parekh Center for Neural Informatics, Structures, and Plasticity,

VA, USA
C. Alejandro Parraga Computer Vision Centre/Computer Science Depart-

ment, Universitat Autònoma de Barcelona, Barcelona, Spain
lii Contributors
Roy D. Patterson Department of Physiology Development and Neurosci-

ence, University of Cambridge, Cambridge, UK
Felix Patzelt Institute for Theoretical Physics, University of Bremen,
Bremen, Germany
Louis M. Pecora U.S. Naval Research Laboratory, Washington, DC, USA
Yu-Cheng Pei Chang Gung Memorial Hospital and University, Taoyuan
County, Taiwan
William D. Penny Wellcome Trust Centre for Neuroimaging, University
College, London, UK
Bijan Pesaran Center for Neural Science, New York University, New York,
NY, USA
Jean-Pascal Pfister Department of Physiology, University of Bern, Bern,
Switzerland
Theoretical Neuroscience Group, Institute of Neuroinformatics, University of
Z€urich and ETH Z€urich, Z€urich, Switzerland
Fabian Philippart Laboratory of Pharmacology and GIGA Neurosciences,
University of Liège, Liège, Belgium
Andrew Philippides Department of Informatics, Centre for Computational
Neuroscience and Robotics, University of Sussex, Brighton, UK
Andrew J. K. Phillips Division of Sleep Medicine, Brigham and Women’s
Hospital, Harvard Medical School, Boston, MA, USA
Hans Ekkehard Plesser Department of Mathematical Sciences and
Technology, Norwegian University of Life Sciences, Ås, Norway
Institute of Neuroscience and Medicine (INM-6), Computational and
Systems Neuroscience, J€ulich Research Centre, J€ulich, Germany
Panayiota Poirazi Foundation for Research and Technology-Hellas
(FORTH), Institute of Molecular Biology and Biotechnology (IMBB),
Heraklion, Crete, Greece
Jean Baptiste Poline Henry H. Wheeler, Jr. Brain Imaging Center, Helen
Wills Neuroscience Institute, University of California, Berkeley, CA, USA
Chi-Sang Poon Institute for Medical Engineering and Science, Harvard-
MIT Division of Health Sciences and Technology, Massachusetts Institute of
Technology, Cambridge, MA, USA
Marko Popovic Department of Physiology, Yale University School of
Oleksandr V. Popovych Institute of Neuroscience and Medicine –
Neuromodulation (INM-7), Research Center J€ulich, J€ulich, Germany
Contributors liii
Roland Potthast Department of Mathematics, University of Reading,

Reading, UK
Randy Powers Department of Physiology and Biophysics, University of
Washington, Seattle, WA, USA
Steven A. Prescott Neurosciences and Mental Health, The Hospital for Sick
Children, Toronto, ON, Canada
Department of Physiology, University of Toronto, Toronto, ON, Canada
Daniel Pressnitzer Département d’Études Cognitives, École Normale
Supérieure, Paris, France
Nicholas J. Priebe Section of Neurobiology, The University of Texas at
Austin, Austin, TX, USA
Astrid A. Prinz Department of Biology, Emory University, Atlanta, GA, USA
Alberto Priori Centro Clinico per la Neurostimolazione, le
Neurotecnologie ed i Disordini del Movimento, Fondazione IRCCS Ca’
Granda, Ospedale Maggiore Policlinico, Milan, Italy
Dipartimento di Fisiopatologia Medico–Chirurgica e dei Trapianti,
Università degli Studi di Milano, Milan, Italy
Arthur Prochazka Centre for Neuroscience, School of Molecular and Sys-
tems Medicine, University of Alberta, Edmonton, Canada
Yifat Prut Department of Medical Neurobiology, Hadassah Medical
School, The Hebrew University, Jerusalem, Israel
Gillian Queisser Goethe Center for Scientific Computing, Goethe Univer-
sity Frankfurt, Frankfurt am Main, Germany
Gregor Rainer Department of Medicine, University of Fribourg, Fribourg,
Switzerland
Govindan Rangarajan Department of Mathematics, Indian Institute of
Science, Bangalore, India
Frank Rattay Institute of Analysis and Scientific Computing, Vienna Uni-
versity of Technology, Vienna, Austria
Kimberly Ray Research Imaging Institute, University of Texas Health
Subhasis Ray National Centre for Biological Sciences, Tata Institute of
Jenny Read Institute of Neuroscience, Newcastle University, Newcastle-
upon-Tyne, UK
Michiel W. H. Remme Humboldt-Universit€at zu Berlin, Berlin, Germany
liv Contributors
Qiushi Ren Institute for Laser Medicine and Bio-Photonics, Department of

Biomedical Engineering, Shanghai Jiao-Tong University, Shanghai, People’s
Republic of China
Sylvie Renaud Institut Polytechnique de Bordeaux, Université de
Bordeaux, Talence, France
Barry Richmond Section on Neural Coding and Computation, Laboratory
of Neuropsychology, NIMH/NIH/DHHS, Bethesda, MD, USA
Hermann Riecke Department of Engineering Sciences and Applied Math-
ematics, Northwestern University, Evanston, IL, USA
Michael C. Riedel Research Imaging Institute, University of Texas Health
Jorge Riera Department of Biomedical Engineering, Florida International
University, Miami, FL, USA
John Rinzel Center for Neural Science & Courant Institute of Mathematical
Sciences, New York University, New York, NY, USA
Alan Roberts School of Biological Sciences, University of Bristol, Bristol, UK
Byron N. Roberts Department of Pharmacology, University of California,
Davis, Davis, CA, USA
Patrick D. Roberts Department of Biomedical Engineering, Oregon Health
& Science University, Portland, OR, USA
Mario Romero-Ortega Bioengineering Department, The University of
Texas at Dallas, Richardson, TX, USA
Surgery Department/U.T. Southwestern Med. Center, Dallas, TX, USA
U.T. Arlington Research Institute, Fort Worth, TX, USA
Bernhard Ronacher Department of Biology, Humboldt-Universit€at zu
Berlin, Berlin, Germany
Manuela Rosa Centro Clinico per la Neurostimolazione, le
Neurotecnologie ed i Disordini del Movimento, Fondazione IRCCS Ca’
Granda, Ospedale Maggiore Policlinico, Milan, Italy
Robert Rosenbaum Department of Applied and Computational Mathemat-
ics and Statistics, University of Notre Dame, Notre Dame, IN, USA
Mitchell Roslin Lenox Hill Hospital/NSLIJ, New York, NY, USA
Jason S. Rothman Neuroscience, Physiology and Pharmacology, Univer-
sity College London, London, UK
Horacio G. Rotstein Department of Mathematical Sciences, New Jersey
Institute of Technology, Newark, NJ, USA
Jonathan Rubin Department of Mathematics and Center for the Neural
Basis of Cognition, University of Pittsburgh, Pittsburgh, PA, USA
Contributors lv
Ilya Rybak Department of Neurobiology and Anatomy, Drexel University

College of Medicine, Philadelphia, PA, USA
Hannes Saal Department of Organismal Biology and Anatomy, University
of Chicago, Chicago, IL, USA
Sylvain Saı̈ghi University of Bordeaux, Bordeaux, France
Mesut Sahin Biomedical Engineering, New Jersey Institute of Technology,
University Heights, Newark, NJ, USA
Emilio Salinas Department of Neurobiology and Anatomy, Wake Forest
School of Medicine, Winston-Salem, NC, USA
Adam N. Sanborn Department of Psychology, University of Warwick,
Coventry, UK
Maria Victoria Sanchez-Vives ICREA and Systems Neuroscience,
IDIBAPS, Barcelona, Spain
Fidel Santamaria UTSA Neurosciences Institute, The University of Texas
at San Antonio, San Antonio, TX, USA
Roland Schaette University College London, UCL Ear Institute, London, UK
Louis K. Scheffer Janelia Farm Research Campus, Howard Hughes
Medical Institute, Ashburn, VA, USA
Bjoern Schelter Institute for Complex Systems and Mathematical Biology,
University of Aberdeen, Aberdeen, UK
Johannes Schemmel University of Heidelberg, Heidelberg, Germany
Matthew Schiefer APT Center of Excellence, Louis Stokes Cleveland
Department of Veterans Affairs Medical Center (LSCDVAMC), Cleveland,
OH, USA
Department of Biomedical Engineering, Case Western Reserve University,
Cleveland, OH, USA
Hartmut Schmidt Carl-Ludwig-Institute for Physiology, Leipzig, Germany
Michael Schmuker Sussex Neuroscience, CCNR, Informatics, University
of Sussex, Brighton, UK
Gregor Schöner Institut f€ur Neuroinformatik, Ruhr-Universit€at Bochum,
Bochum, Germany
Jan Schnupp Department of Physiology, Anatomy and Genetics, Medical
Sciences Division, University of Oxford, Oxford, UK
Jan-Mathijs Schoffelen Donders Institute for Brain, Cognition and Behav-
iour, Centre for Cognition, Radboud University Nijmegen, Nijmegen, The
Netherlands
Benjamin Scholl Section of Neurobiology, The University of Texas at
Austin, Austin, TX, USA
lvi Contributors
Michael C. Schubert Department of Otolaryngology Head and Neck Sur-

gery, Johns Hopkins University School of Medicine, Baltimore, MD, USA
Department of Physical Rehabilitation and Medicine, Baltimore, MD, USA
Simon R. Schultz Department of Bioengineering, Imperial College London,
London, UK
David J. Schulz Division of Biological Sciences, University of Missouri,
Columbia, MO, USA
Lars Schwabe Department of Computer Science and Electrical Engineering,
Adaptive and Regenerative Software Systems, Universit€at Rostock, Rostock,
Germany
Philipp Schwedhelm Cognitive Neuroscience Laboratory, German Primate
Center, Göttingen, Germany
Bernstein Center for Computational Neuroscience, Göttingen University,
Göttingen, Germany
Michael A. Schwemmer Mathematical Biosciences Institute, The Ohio
State University, Columbus, OH, USA
Walter Senn Department of Physiology, University of Bern, Bern,
Switzerland
Rodolphe Sepulchre Department of Engineering, University of Cambridge,
Cambridge, UK
Peggy Seriès Institute of Adaptive and Neural Computation, University of
Edinburgh, Edinburgh, UK
Teresa Serrano-Gotarredona Instituto de Microelectrónica de Sevilla
(IMSE-CNM), CSIC and University of Sevilla, Sevilla, Spain
Thomas Serre Institute for Brain Sciences, Brown University, Providence,
RI, USA
Anil K. Seth Sackler Centre for Consciousness Science and Department of
Informatics, University of Sussex, Brighton, UK
Vincent Seutin Laboratory of Pharmacology and GIGA Neurosciences,
University of Liège, Liège, Belgium
Shihab Shamma Electrical and Computer Engineering Department and
Institute for System Research, University of Maryland, College Park,
MD, USA
École Normale Supérieure, Paris, France
Robert V. Shannon Department of Otolaryngology, University of Southern
California, Los Angeles, CA, USA
Andrew Sharott Medical Research Council Anatomical Neuropharmacology
Unit, Department of Pharmacology, University of Oxford, Oxford, UK
Contributors lvii
Gordon M. Shepherd Department of Neurobiology, Yale University

School of Medicine, New Haven, CT, USA
William Erik Sherwood University of Utah, Salt Lake City, UT, USA
Natalia A. Shevtsova Department of Neurobiology and Anatomy, Drexel
University College of Medicine, Philadelphia, PA, USA
Woodrow Shew Department of Physics, University of Arkansas, Fayette-
ville, AR, USA
Miho Shidahara Division of Medical Physics, Tohoku University School of
Medicine, Sendai, Japan
Barbara Shinn-Cunningham Center for Computational Neuroscience and
Neural Technology, Boston University, Boston, MA, USA
Shigeru Shinomoto Department of Physics, Kyoto University, Sakyo-ku,
Kyoto, Japan
Angus Silver Department of Neuroscience, Physiology and Pharmacology,
University College London, London, UK
Frances K. Skinner Toronto Western Research Institute, University Health
Network, Toronto, ON, Canada
Physiology, University of Toronto, Toronto, ON, Canada
Medicine (Neurology), University of Toronto, Toronto, ON, Canada
Jamie Sleigh Waikato Clinical School, University of Auckland, Waikato
Hospital, Hamilton, New Zealand
Jeffrey C. Smith Cellular and Systems Neurobiology Section, NINDS/NIH,
Bethesda, MD, USA
Tomasz G. Smolinski Department of Computer and Information Sciences,
Delaware State University, Dover, DE, USA
Stefania Sokolowski Center for Computational Neuroscience and Neural
Technology, Boston University, Boston, MA, USA
Sergio Solinas Fundation Neurological Institute IRCCS “C. Mondino”,
Pavia, Italy
Haim Sompolinsky Racah Institute of Physics, The Edmond and Lily Safra
Center for Brain Sciences, The Hebrew University of Jerusalem, The Edmond
J. Safra Campus, Jerusalem, Israel
Dong Song Department of Biomedical Engineering, Center for Neural Engi-
neering, University of Southern California, Los Angeles, CA, USA
Zhuoyi Song Department of Biomedical Science, University of Sheffield,
Sheffield, UK
Wafa Soofi Biomedical Engineering Department, Georgia Institute of
Technology/Emory University, Atlanta, GA, USA
lviii Contributors
Michael Spratling Department of Informatics, King’s College London,

London, UK
Terrence R. Stanford Department of Neurobiology & Anatomy, Wake
Forest School of Medicine, Winston-Salem, NC, USA
Wolfgang Stein School of Biological Sciences, Illinois State University,
Normal, IL, USA
David C. Sterratt School of Informatics, University of Edinburgh,
Edinburgh, UK
D. Alistair Steyn-Ross School of Engineering, University of Waikato,
Hamilton, New Zealand
Moira Steyn-Ross School of Engineering, University of Waikato, Hamilton,
New Zealand
Klaus Stiefel The MARCS Institute, University of Western Sydney, Penrith,
Australia
Mark Stopfer National Institutes of Health, NICHD, Bethesda, MD, USA
H. Christiaan Stronks Computer Vision, NICTA, Canberra, ACT,
Australia
G. J. Suaning University of New South Wales, Sydney, NSW, Australia
Sathyaa Subramaniyam Fundation Neurological Institute IRCCS
“C. Mondino”, Pavia, Italy
Clara Suied École Normale Supérieure, Paris, France
Christian J. Sumner MRC Institute of Hearing Research, Nottingham, UK
Nicholas Swindale University of British Columbia, Vancouver, BC, Canada
Joel Tabak Florida State University, Tallahassee, FL, USA
Gregg Tabot Committee on Computational Neuroscience, Somatosensory
Research Lab, University of Chicago, Chicago, IL, USA
Keiko Tanaka-Yamamoto Center for Functional Connectomics, Korea
Institute of Science and Technology, Seoul, Republic of Korea
Peter Alexander Tass Institute of Neuroscience and Medicine –
Neuromodulation (INM–7), Research Center J€ulich, J€ulich, Germany
Department of Neurosurgery, Stanford University, Stanford, CA, USA
Department of Neuromodulation, University of Cologne, Cologne, Germany
Aryeh H. Taub Department of Neurobiology, Weizmann Institute of
Science, Rehovot, Israel
Zachary Taxin Department of Physiology and Pharmacology, SUNY
Downstate Medical Center, Brooklyn, NY, USA
Contributors lix
Bartosz Teleńczuk Unit of Neuroscience, Information and Complexity

(UNIC), Centre National de la Recherche Scientifique (CNRC), Gif-sur-
Yvette, France
David Terman Department of Mathematics, The Ohio State University,
Columbus, OH, USA
Marco Thiel Department of Physics, Institute for Complex Systems and
Mathematical Biology, University of Aberdeen, Aberdeen, UK
Paul Tiesinga Neuroinformatics, Donders Institute, Radboud University,
Nijmegen, The Netherlands
Chung Tin Department of Mechanical and Biomedical Engineering, City
University of Hong Kong, Kowloon Tong, Hong Kong
Lena H. Ting The W. H. Coulter Department of Biomedical Engineering,
Emory University and Georgia Institute of Technology, Atlanta, GA, USA
Natalia Toporikova Department of Biology, Washington and Lee Univer-
sity, Lexington, VA, USA
Benjamin Torben-Nielsen Computational Neuroscience Unit, Okinawa
Institute of Science and Technology, Onna-son, Kunigami-gun, Okinawa,
Japan
Joaquin J. Torres Institute “Carlos I” for Theoretical and Computational
Physics and Department of Electromagnetism and Matter Physics, Facultad
de Ciencias, Universidad de Granada, Granada, Spain
Bruce C. Towe School of Biological and Health Systems Engineering,
Arizona State University, Tempe, AZ, USA
Roger Traub Physical Sciences, IBM T.J. Watson Research Center,
Yorktown Heights, NY, USA
Matthew Tresch Department of Biomedical Engineering and Physical
Medicine and Rehabilitation, Northwestern University, Evanston, IL, USA
Stefan Treue Cognitive Neuroscience Laboratory, German Primate Center,
Faculty of Biology and Psychology, Göttingen University, Göttingen,
Germany
Bernstein Center for Computational Neuroscience, Göttingen University,
Jochen Triesch Frankfurt Institute for Advanced Studies, Frankfurt am
Main, Germany
Shreejoy J. Tripathy Centre for High-Throughput Biology and Department
of Psychiatry, University of British Columbia, BC, Canada
Wilson Truccolo Department of Neuroscience, Brown University, Provi-
dence, RI, USA
lx Contributors
Krasimira Tsaneva-Atanasova Department Mathematics, College of

Engineering, Mathematics and Physical Sciences, University of Exeter,
Exeter, UK
George A. Tsianos Department of Biomedical Engineering, University of
Southern California, Los Angeles, CA, USA
Ray W. Turner Hotchkiss Brain Institute, University of Calgary, Calgary,
AB, Canada
Angela M. Uecker Research Imaging Institute, University of Texas Health
Aman Ullah School of Systems Biology and Krasnow Institute for
Advanced Studies, George Mason University, Fairfax, VA, USA
Shankar Vallabhajosula Citigroup Biomedical Imaging Center, Weill
Cornell Medical College, New York, NY, USA
Wim van Drongelen Department of Pediatrics, The University of Chicago,
Chicago, IL, USA
David Van Essen School of Medicine, Washington University, St. Louis,
MO, USA
Werner Van Geit Blue Brain Project, EPFL, Lausanne, Switzerland
Stephan van Gils Department of Applied Mathematics, University of
Twente, Enschede, The Netherlands
Roemer van der Meij Donders Institute for Brain, Cognition and Behav-
iour, Centre for Cognition, Radboud University Nijmegen, Nijmegen, The
Netherlands
André van Schaik University of Western Sydney, Penrith, NSW, Australia
Marijn van Wingerden Department of Comparative Psychology, Institute
of Experimental Psychology, Heinrich-Heine University D€usseldorf,
D€usseldorf, Germany
Pablo Varona Departamento de Ingenieria Informatica, Universidad
Autónoma de Madrid, Madrid, Spain
Michael Vella Department of Physiology, Development and Neuroscience,
University of Cambridge, Cambridge, USA
Sharmila Venugopal Department of Physiology, David Geffen School of
Medicine, University of California Los Angeles, Los Angeles, CA, USA
Jonathan D. Victor Brain and Mind Research Institute and Department of
Neurology, Weill Cornell Medical College of Cornell University, New York,
NY, USA
Marije ter Wal Neuroinformatics, Donders Institute, Radboud University,
Nijmegen, The Netherlands
Contributors lxi
Dagmar Waltemath Systems Biology and Bioinformatics, University of

Rostock, Rostock, Germany
Rixin Wang Center for Medical Informatics, Yale University School of
Wei Wang Department of Physical Medicine and Rehabilitation, University
of Pittsburgh, Pittsburgh, PA, USA
Matthew Ward Weldon School of Biomedical Engineering, Purdue
University, West Lafayette, IN, USA
Hiroshi Watabe Division of Radiation Protection and Safety Control,
Cyclotron Radioisotope Center, Tohoku University, Sendai, Japan
M. Neal Waxham Department of Neurobiology and Anatomy, The Univer-
sity of Texas Health Science Center at Houston, Houston, TX, USA
Douglas J. Weber Department of Bioengineering, and Department of
Physical Medicine and Rehabilitation, University of Pittsburgh, Pittsburgh,
PA, USA
Franz Weber Molecular and Cell Biology, University of California,
Berkeley, USA
Michael Wehr Institute of Neuroscience, University of Oregon, Eugene,
OR, USA
Wei Wei Center for Neural Science, New York University, New York,
NY, USA
Department of Neurobiology and Kavli Institute for Neuroscience, Yale
University School of Medicine, New Haven, CT, USA
James Weiland University of Southern California, Los Angeles, CA, USA
Thomas P. Wellings School of Biomedical Sciences and Pharmacy, Hunter
Medical Research Institute, The University of Newcastle, Callaghan, NSW,
Australia
Fabrice Wendling INSERM, U1099, Rennes, France
University of Rennes 1, LTSI, Rennes, France
Peter Wenner Emory University, Atlanta, GA, USA
William White Department of Biological Sciences, Ohio University,
Athens, OH, USA
Miles Whittington Hull-York Medical School, Hull/York, UK
Thomas Wichmann Yerkes National Primate Research Center, Emory
Department of Neurology, Emory University, Atlanta, GA, USA
Udall Center of Excellence in Parkinson’s Disease Research, Emory Univer-
lxii Contributors
Alex H. Williams University of California, San Diego, La Jolla, CA, USA

Brandeis University, Waltham, MA, USA
Charles Wilson Department of Biology, University of Texas at San

Hugh R. Wilson Center for Vision Research, York University, Toronto,

ON, Canada
Klaus Wimmer Institut d’Investigacions Biomèdiques August Pi i Sunyer,

Barcelona, Spain
Beth Winkelstein Department of Bioengineering, University of

Istvan Winkler Institute of Cognitive Neuroscience and Psychology,

Research Centre for Natural Sciences, MTA, Budapest, Hungary
Laurenz Wiskott Institut f€ur Neuroinformatik, Ruhr-Universit€at Bochum,

Bochum, Germany
Brian Wodlinger Department of Physical Medicine and Rehabilitation,

University of Pittsburgh, Pittsburgh, PA, USA
Daniel K. Wójcik Department of Neurophysiology, Nencki Institute of

Experimental Biology, Warsaw, Poland
Yan Tat Wong Center for Neural Science, New York University, New
York, NY, USA
Jianhong Wu Department of Mathematics and Statistics, York University,

Toronto, ON, Canada
Huei-Fang Yang Research Center for Information Technology Innovation,

Academia Sinica, Taipei, Taiwan, Republic of China
Jacob L. Yates Departments of Neuroscience and Psychology, Center for

Perceptual Systems, The University of Texas at Austin, Austin, TX, USA
Jeffrey M. Yau Johns Hopkins University, Baltimore, MD, USA
Paul Yoo Institute of Biomaterials and Biomedical Engineering, Depart-

ment of Electrical and Computer Engineering, University of Toronto,
Toronto, Canada
Ken Yoshida Department of Biomedical Engineering, School of Engineering

and Technology, Indiana University-Purdue University Indianapolis (IUPUI),
Indianapolis, IN, USA
Nada Yousif Division of Brain Sciences, Imperial College London,

London, UK
Angela J. Yu Department of Cognitive Science, University of California,

San Diego, La Jolla, CA, USA
Contributors lxiii
Tommy Yu Auckland Bioengineering Institute, University of Auckland,

Blaise Yvert CNRS, Institute for Cognitive and Integrative Neuroscience

Univ. Bordeaux, Institute for Cognitive and Integrative Neuroscience
Inserm, Clinatec, UA01, Grenoble, France
CEA, LETI, Clinatec, UA01, Grenoble, France
Dejan Zecevic Department of Physiology, Yale University School of

Ying Zheng School of Systems Engineering, The University of Reading,

Reading, UK
Eberhart Zrenner Center for Ophthalmology and Werner Reichardt Center

for Integrative Neurosciences, Eberhard Karl University of T€ubingen,
Overview
Auditory Sensing Systems: Overview Possibly due to the complexity of the problem,
many aspects of auditory processing have not yet
Susan Denham been modeled at a detailed neurocomputational
Cognition Institute and School of Psychology, level. In addition, a great deal of processing occurs
Plymouth University, Plymouth, Devon, UK even before the incoming signals reach cortex;
with the result that there are more models related
to subcortical processing, than to higher level,
Detailed Description putatively cortical, functions. The articles commis-
sioned for the auditory sensing systems section
Auditory sensing, or the sense of hearing, is therefore span a range of topics from qualitatively
concerned with detecting and extracting infor- different point of view, with a strong focus on the
mation from pressure waves in the surrounding functionality of the auditory system. Together they
medium, typically air or water. Since waves are provide an interesting and insightful view of cur-
generated by movements or collisions, this pri- rent understanding of auditory processing, with
marily tells the perceiver about things happen- a great deal of useful information for modellers
ing in the environment. In addition, since regarding the neuroanatomy and neurophysiology
pressure waves can be reflected, absorbed, and of the auditory system, methodological approaches
refracted by other objects, these pressure waves to studying the auditory system, and functional
also contain a great deal of contextual informa- requirements and perceptual constraints on
tion about the environment and the objects in auditory processing.
it. A fundamental challenge for the auditory A well-illustrated overview of the anatomy
system is to segregate the contributions of indi- and physiology of the auditory system, including
vidual sound sources to the sound pressure the extensive but poorly understood efferent sys-
waves received by the sensors as these are tem is covered by “▶ Anatomy and Physiology of
made up of a combination of all concurrent the Mammalian Auditory System.” A detailed
sources and their various reflections. Sounds account of efferent control of the auditory sensor,
unfold in time, so modellers of auditory the cochlea, is presented in the article on
processing cannot ignore time and the need to the “▶ Physiology and Function of Cochlear
process signals within time; this becomes espe- Efferents.” It is in the cochlea that the biological
cially challenging when considering the system begins to analyze the pressure waves over
multiscale nature of the information contained multiple time scales and to do so within the
within sounds. time constraints of the ongoing multiscale
D. Jaeger, R. Jung (eds.), Encyclopedia of Computational Neuroscience,

DOI 10.1007/978-1-4614-6675-8, # Springer Science+Business Media New York 2015
2 Auditory Sensing Systems: Overview
information flow. Another point of major trans- models and theories more closely related to
formation occurs at the gateway to the cortex, the biology. These tend either to relate to
described in the article “▶ Auditory very specific functions (“▶ Sound Localization
Thalamocortical Transformations.” Three further in Mammals, Models,” “▶ Stimulus-Specific
overview articles dealing with electrophysiolog- Adaptation, Models,” “▶ Auditory Precedence
ical correlates of auditory perception: “▶ Audi- Effect,” and “▶ Masking and Masking Release”),
tory Event-Related Potentials,” “▶ Auditory or relate to processes at or near the
Evoked Brainstem Responses,” and “▶ Electro- sensor (“▶ Auditory Nerve Response, Afferent
physiological Indices of Speech Processing” doc- Signals,” and “▶ Cochlear Inner Hair Cell,
ument methodological approaches to studying Model”).
high-level perceptual functions.
Specific aspects of the functional neurophysi-
ology of the thalamocortical auditory system are Cross-References
addressed in a number of articles. The articles
“▶ Associations and Rewards in Auditory Cor- ▶ Acoustic Timbre Recognition
tex” and “▶ Context-Dependent Processing in ▶ Anatomy and Physiology of the Mammalian
Auditory Cortex” demonstrate that the auditory Auditory System
cortical code needs to be viewed in more complex ▶ Associations and Rewards in Auditory Cortex
terms than simple acoustic feature representa- ▶ Auditory Event-Related Potentials
tions. For example, neural correlates of reward ▶ Auditory Evoked Brainstem Responses
are found even within primary auditory ▶ Auditory Memory
cortex (“▶ Associations and Rewards in Audi- ▶ Auditory Nerve Response, Afferent Signals
tory Cortex”). In contrast, the articles ▶ Auditory Perceptual Organization
“▶ Spectrotemporal Receptive Fields,” ▶ Auditory Precedence Effect
“▶ Stimulus-Specific Information,” and ▶ Auditory Thalamocortical Transformations
“▶ Stimulus Reconstruction from Cortical ▶ Cochlear Inner Hair Cell, Model
Responses” show how auditory cortical activity ▶ Context-Dependent Processing in Auditory
can in some circumstances be usefully interpreted Cortex
in terms of acoustic feature combinations. ▶ Electrophysiological Indices of Speech
An overview of theories and models of higher Processing
level aspects of auditory perception are presented ▶ Masking and Masking Release
in “▶ Auditory Perceptual Organization,” ▶ Music Processing in the Brain
“▶ Music Processing in the Brain,” “▶ Neural ▶ Neural Coding of Speech Sounds
Coding of Speech Sounds,” “▶ Pulse-Resonance ▶ Physiology and Function of Cochlear Efferents
Sounds,” “▶ Auditory Memory,” “▶ Acoustic ▶ Pitch Perception, Models
Timbre Recognition,” “▶ Pitch Perception, ▶ Pulse-Resonance Sounds
Models,” “▶ Rhythm Perception: Pulse and ▶ Rhythm Perception: Pulse and Meter
Meter,” and “▶ Sound Localization and ▶ Sound Localization and Experience-
Experience-Dependent Plasticity.” These diverse Dependent Plasticity
topics emphasize the different ways in which ▶ Sound Localization in Mammals, Models
sound is interpreted by the brain, while the article ▶ Spectrotemporal Receptive Fields
“▶ Tinnitus, Models” shows how modelling may ▶ Stimulus Reconstruction from Cortical
help advance understanding of a perceptual phe- Responses
nomenon that plagues 10–15 % of the population. ▶ Stimulus-Specific Adaptation, Models
Finally, the section contains articles which ▶ Stimulus-Specific Information
present more detailed neurocomputational ▶ Tinnitus, Models
Basal Ganglia: Overview 3
Basal Ganglia: Overview cortex
Jonathan Rubin
Department of Mathematics and Center for the
Neural Basis of Cognition, University of striatum
hyperdirect
Pittsburgh, Pittsburgh, PA, USA
indirect
direct
Definition SNc
thalamus
GPe
The basal ganglia are a collection of four subcor-
tical nuclei: the striatum, substantia nigra, globus
pallidus, and subthalamic nucleus. They are com- GPi/SNr
STN
ponents of several apparently segregated circuits
that can be classified according to function as
motor, oculomotor, associative, or limbic. Certain inhibition excitation
neurons in the basal ganglia are major sources of
the neurotransmitter dopamine, associated with Basal Ganglia: Overview, Fig. 1 Major structures and
pathways of the basal ganglia. Basal ganglia areas are
reward, while others receive dopaminergic inputs; colored in blue. Connecting arrows and lines represent
thus, the basal ganglia have received considerable the existence of direct synaptic interactions between areas.
attention in the context of learning. Imbalances in A projection from GPe to striatum that was recently dis-
activity across the basal ganglia nuclei within the covered in rats is omitted here
motor circuit are associated with various motor
disorders. Following the nomenclature used in primates, the
striatum, encompassing the caudate nucleus
and the putamen, is the predominant recipient of
Detailed Description cortical inputs. There are several types of neurons
within the striatum, most of which release the neu-
This section focuses on the motor aspects of rotransmitter GABA, which has an inhibitory effect
the basal ganglia, which have been the subject of on its targets (see entry “▶ Striatal Models, Cellular
most computational studies. The articles can be Detail”). Striatal neurons are also major recipients
classified as those concentrating on the activity of dopamine. Different types of dopamine recep-
within particular areas of the basal ganglia, those tors arise in different striatal neurons, and this
considering basal ganglia function, and those variability translates into diverse impacts of dopa-
about the roles of the basal ganglia in motor dis- mine release. In coarse terms, dopamine promotes
orders, predominantly parkinsonism. activity in some striatal neurons and inhibits it in
others, but specific effects may be more compli-
Basal Ganglia Structure cated and may depend on ongoing activity levels.
While the structures of the basal ganglia (Fig. 1) The substantia nigra has two main parts,
appear to be common across a wide range of the substantia nigra pars reticulata and the
vertebrate animals, from primates to lamprey substantia nigra pars compacta (SNc). The
(Stephenson-Jones et al. 2011), there are some- former sends inhibitory output to areas outside
what different naming systems used in different of the basal ganglia that may be related to oculo-
species, and there are some structural differences motor activity (see entry “▶ Basal Ganglia: Con-
across species (e.g., mammals versus birds). trol of Saccades”). The latter releases dopamine
4 Basal Ganglia: Overview
to the striatum and other basal ganglia areas (see ganglia output and hence suppress movement.
entry “▶ Dopaminergic Cell Models”). This view is based on the idea that the cortex
The globus pallidus is also a dichotomous excites striatum and thus inhibits GPe and
structure, with segments that are referred to as disinhibits its target STN, which in turn excites
external (GPe) and internal (GPi). Both parts GPi; the reduction in GPe activity also lowers its
predominantly contain GABAergic neurons that inhibition of GPi, yielding additional enhance-
exhibit high firing rates, although with diverse ment of GPi firing. This reasoning omits the
firing patterns, under baseline conditions (see complication of the feedback excitation that
entry “▶ Globus Pallidus Cellular Models”). STN sends to GPe, however. The hyperdirect
The GPe receives projections from a subset of pathway, like the direct pathway, includes only
striatal neurons as well as from the subthalamic one intermediary between the cortex and GPi:
nucleus (STN), and it innervates both the striatum it refers to the cortical projection to STN together
and the STN in return, although it appears that the with the STN link to GPi. Presumably, this path-
subpopulations of GPe neurons that target these way is called “hyperdirect” because it features
areas are distinct (Mallet et al. 2012). The GPi is excitatory transmission via glutamatergic synap-
subject to inputs from a different subset of striatal ses from STN to GPi that is faster and has more
neurons, from the STN, and from the GPe and in direct effect than the direct pathway’s inhibitory
turn projects to thalamic areas outside of the basal signaling via GABAergic synapses from the stri-
ganglia. atum to GPi. Like the indirect pathway, the
The STN stands out as the only basal ganglia hyperdirect pathway promotes basal ganglia
area that releases the excitatory neurotransmitter output.
glutamate. STN neurons exhibit intrinsic firing, Finally, note that in rodent, which is a useful
albeit with heterogeneous properties (see entry experimental model, the output structure analo-
“▶ Subthalamic Nucleus Cellular Models”). gous to the GPi is called the entopeduncular
Besides participating in a reciprocal loop with nucleus, and the major motor output structure is
GPe (see entry “▶ Subthalamopallidal Loop and thought to be the SNr.
Oscillations”), the STN projects directly to GPi.
The STN also receives direct glutamatergic input Basal Ganglia Function
from the cortex and some dopaminergic input The basal ganglia have been postulated to have
from the SNc. a wide range of functions related to motor learn-
The routes along which activity flows through ing and other steps contributing to goal-directed
the basal ganglia have been classified into the movements (Chakravarthy et al. 2010). Models
direct pathway, the indirect pathway, and the have been developed to represent and make
hyperdirect pathway (Fig. 1). This nomenclature predictions about the mechanisms involved in
refers to the number and types of steps between several of these functions. In the Basal Ganglia
the cortex, which provides input to the basal section of this encyclopedia, articles discuss
ganglia, and the GPi, the primary source of background information and computational
motor outputs from the primate basal ganglia. work on the mechanisms for action selection,
The direct pathway is the stream from cortex to habit formation, decision making, and control of
striatum to GPi. In net, the direct pathway inhibits saccades. Furthermore, additional articles treat
basal ganglia output, since the cortex excites the the role of the basal ganglia as an exploration
striatum, which inhibits GPi. Since GPi inhibits engine and the learning and production of songs
downstream areas, reductions in its output can be by birds.
prokinetic. The indirect pathway goes from the Several lines of evidence point to the basal
cortex through the striatum to GPe and continues ganglia as a major site of action selection (see
on to STN and then GPi; the GPe-GPi connection entry “▶ Basal Ganglia: Mechanisms for Action
may also be included in this pathway. Overall, the Selection”). In particular, the basal ganglia are
indirect pathway is believed to promote basal well positioned anatomically to orchestrate
Basal Ganglia: Overview 5
action selection by integrating inputs from and thereby influences future choices. Computa-
a broad range of cortical areas and providing tional models consider how the activities in the
outputs to downstream motor-related sites. The pathways of the basal ganglia interact with dopa-
basic idea of how action selection might work minergic reward signals to drive and adjust the
rests on the observation that under rest condi- decision-making process, positing various roles
tions, the output nuclei of the basal ganglia, GPi for particular brain areas in evidence accumula-
and SNr, exhibit high activity levels and that tion, stimulus detection, threshold adjustment
these areas release inhibitory neurotransmitters, and decision making, and cancelation. Some
which are well suited to block motor activity. To also treat impairments in parkinsonian conditions
perform a movement, this inhibitory “gate” must and alterations under deep brain stimulation
be opened but only in a way that allows the (see entry “▶ Basal Ganglia: Decision-Making”).
desired action while maintaining a blockade on Unlike other actions to which the basal ganglia
all other movements. Models of action selection contribute, rapid eye movements known as
position the direct pathway as the agent that saccades in primates are gated by outputs from
removes inhibition and suggest that the indirect SNr rather than GPi. Issues under investigation
and hyperdirect pathways can act to suppress relating to saccade generation, which have
behavior and to participate in behavior switching been treated in computational models, include
(see entry “▶ Basal Ganglia: Mechanisms for mechanisms of target selection, of learning of
Action Selection”; Nambu et al. 2002). target priorities, and of interactions of these
Gradually, particular actions can become processes with working memory (see entry
habits, which are actions that are elicited by cer- “▶ Basal Ganglia: Control of Saccades”).
tain stimuli in a way that lacks flexibility. A less thoroughly investigated function of the
Although established habits are less dependent basal ganglia is the generation of variability in
on external rewards than are other learned behav- motor responses. Such variability is certainly
iors, rewarding feedback that leads to striatal observed in the behaviors of many species and
activation and associated plasticity of seems to be essential for optimizing behavior and
corticostriatal synapses are central to their being attaining maximal rewards. There is significant
formed, and the corresponding learning process experimental evidence implicating the basal
has been considered in computational models ganglia as a source of such motor exploration,
(see entry “▶ Basal Ganglia: Habit Formation”). and some computational modeling work to
Dopamine is central to this form of plasticity and instantiate these ideas and study their implica-
learning, and impairments in habit formation and tions has been performed (see entry “▶ Basal
performance are noted in subjects with Ganglia System as an Engine for Exploration”).
Parkinson’s disease. In relevant models, while the direct pathway
A basal ganglia function that is distinct from takes on the traditional role of implementing
pure action selection and habit learning arises in action selection, the indirect pathway, and partic-
perceptual decision making. In this process, in ularly the STN, injects variability into this pro-
abstract terms, it is thought that evidence favor- cess. Learning based on reward signals is critical
ing each of a number of choices accumulates until to such models and helps control the relative
one option becomes sufficiently supported to dominance of direct and indirect pathway signals.
elicit a corresponding behavior. The evidence is One example of the prominence of behavioral
presumably represented by firing of various sets variability arises in the song production by song-
of cortical neurons that can influence basal birds. Songs must be learned, which involves
ganglia activity, and basal ganglia activity is exposure to the song of another bird, trial-and-
involved in setting the selection threshold (see error behavior, and auditory feedback. This
entry “▶ Basal Ganglia: Decision-Making”). In process appears to heavily involve a basal ganglia
this view, the dopamine-based reward induced by analogue that participates in learning, in the
a behavior is involved in threshold adjustment introduction of behavioral variability, and in the
6 Basal Ganglia: Overview
selection of rewarding behaviors. Certain songbird dopamine leads to widespread modifications of

experimental preparations are advantageous for neuronal activity patterns, rather than just changes
studying this combination of processes, and in firing rates, in the basal ganglia. These effects
corresponding models of associated reinforcement include an enhanced prevalence of bursting,
learning and song production have been informed altered oscillation structure, and increased corre-
by observations from these preparations (see entry lation in outputs of neurons within, and across,
“▶ Basal Ganglia: Songbird Models”). basal ganglia areas (Rubin et al. 2012).
One potential source of oscillations within the
Basal Ganglia and the Motor Symptoms of basal ganglia is the reciprocal excitatory-inhibitory
Parkinsonism loop between the GPe and the STN (see entry
The loss of dopaminergic neurons in the SNc “▶ Subthalamopallidal Loop and Oscillations”).
appears to be the critical trigger for a slew of Oscillations in the beta band (broadly taken as
changes in neuronal activity within areas of the 10–30 Hz) appear to be particularly prevalent in
basal ganglia that result in the motor symptoms of parkinsonian conditions and may be effective at
parkinsonism. This loss is typically gradual, and hijacking the activity in thalamic areas down-
the causes appear to involve a complex combina- stream from the basal ganglia, and several different
tion of environmental and genetic factors that are sources for such oscillations have been posited (see
still under intensive investigation. entry “▶ Basal Ganglia: Beta Oscillations”).
Classical models of parkinsonism attribute its While oscillations observed in the activity of
motor signs to an imbalance in direct and indirect basal ganglia neurons seem like a natural candi-
pathway outputs stemming from the loss of dopa- date to contribute to parkinsonian tremor, the link
minergic input to the striatum (Albin et al. 1989). between altered neuronal firing and other parkin-
The loss of dopamine removes a source of exci- sonian symptoms may be less explicit. In partic-
tation to the striatal neurons with D1 receptors, ular, the slowing of movement known as
which project to GPi. Less firing by these striatal bradykinesia may emerge from the interactions
neurons, viewed as a weakening of the direct of all of the motor signaling pathways through the
pathway, results in disinhibition of GPi. Mean- basal ganglia, or it may be necessary to consider
while, diminished dopamine translates to reduc- basal ganglia interactions with corticospinal-
tion in inhibition of the striatal neurons with D2 muscular pathways to explain its source (see
receptors, which in turn inhibit GPe. As a result, entry “▶ Bradykinesia Models”).
GPe output is suppressed, which relieves GPe The predominant treatments for Parkinson’s
inhibition of STN and allows enhanced STN fir- disease in its early stages are pharmacological,
ing. The increased STN activity provides more aimed at replacing lost dopamine. In cases where
drive to GPi, viewed as a strengthening of the pharmacological treatments gradually lose effi-
indirect pathway. Together, these changes tip the cacy or induce substantive side effects, deep
balance of direct and indirect pathway effects in brain stimulation therapy represents a treatment
a way that favors the indirect pathway and pro- option that, while highly invasive, is at least
motes GPi activity. GPi outputs inhibit the partially adjustable (by tuning of stimulation
pallidal recipient areas of the thalamus and are parameters) and reversible (by cessation of stim-
believed to represent an inhibitory gate that shuts ulation). Deep brain stimulation for Parkinson’s
down movement. Thus, pathologically enhanced disease typically is targeted at STN or GPi and
GPi firing could interfere with initiating and car- proves effective for many patients; although the
rying out movements in a way that translates into mechanisms underlying its efficacy are not
the motor signs of parkinsonism. known, theories have honed in on its potential
The complete explanation for the emergence impact on patterns of basal ganglia activity
of parkinsonian motor signs likely involves phe- and outputs and their downstream effects
nomena that are omitted from this classical (see entry “▶ Computational Models of Deep
description. It has become clear that loss of Brain Stimulation (DBS)”; Rubin et al. 2012).
Bayesian Approaches in Computational Neuroscience: Overview 7
Cross-References Gerfen CR, Wilson CJ (1996) The basal ganglia. Handb

Chem Neuroanat 12:371–468
Smith Y, Bevan MD, Shink E, Bolam JP (1998) Microcir-
▶ Basal Ganglia System as an Engine for cuitry of the direct and indirect pathways of the basal
Exploration ganglia. Neuroscience 86:353–387
▶ Basal Ganglia: Beta Oscillations
▶ Basal Ganglia: Control of Saccades
▶ Basal Ganglia: Decision-Making
▶ Basal Ganglia: Habit Formation Bayesian Approaches in
▶ Basal Ganglia: Mechanisms for Action Computational Neuroscience:
Selection Overview
▶ Basal Ganglia: Songbird Models
▶ Bradykinesia Models Ulrik R. Beierholm
▶ Computational Models of Deep Brain Centre for Computational Neuroscience and
Stimulation (DBS) Cognitive Robotics, University of Birmingham,
▶ Dopaminergic Cell Models Birmingham, UK
▶ Globus Pallidus Cellular Models
▶ Striatal Models, Cellular Detail
▶ Subthalamic Nucleus Cellular Models Detailed Description
▶ Subthalamopallidal Loop and Oscillations
Bayesian approaches in computational neurosci-
ence rely on the properties of Bayesian statistics
for performing inference over unknown variables
References given a data set generated through a stochastic
process.
Albin RL, Young AB, Penney JB (1989) The functional
anatomy of basal ganglia disorders. Trends Neurosci Given a set of observed data d1:n, generated
12:366–375 from a stochastic process P(d1:n|X) where X is
Chakravarthy VS, Joseph D, Bapi RS (2010) What do the a set of unobserved variables, the posterior prob-
basal ganglia do? Biol Cybern 103:237–253
ability distribution of X is P(X|d1:n) = P(d1:n|X)
Mallet N, Micklem BR, Henny P, Brown MT, Williams C,
Bolam JP, Nakamura KC, Magill PJ (2012) Dichoto- P(X)/P(d1:n) according to Bayes rule. X can be
mous organization of the external globus pallidus. a set of fixed parameters as well as a series of
Neuron 74:1075–1086 variables of the same size as the data itself X1:n.
Nambu A, Tokuno H, Takada M (2002) Functional sig-
Based on the posterior probability and
nificance of the cortico-subthalamo-pallidal
‘hyperdirect’ pathway. Neurosci Res 43:111–117 a specified utility function, an estimate of X can
Rubin JE, McIntyre CC, Turner RS, Wichmann T (2012) be made that can be shown to be optimal, e.g., by
Basal ganglia activity patterns in parkinsonism and minimizing the expected variance.
computational modeling of their downstream effects.
One common use of this principle within com-
Eur J Neurosci 36:2213–2228
Stephenson-Jones M, Samuelsson E, Ericsson J, putational neuroscience is for inferring
Robertson B, Grillner S (2011) Evolutionary conserva- unobserved properties (hidden variables X)
tion of the basal ganglia as a common vertebrate mech- based on observed data, d. These techniques can
anism for action selection. Curr Biol 21:1081–1091
be used for inference on any data set but has in
neuroscience mostly been used for neurophysio-
Further Reading logical recordings, imaging, and behavioral data.
Alexander GE, Crutcher MD, DeLong MR (1990) Basal
ganglia-thalamocortical circuits: parallel substrates for By their nature electrophysiological record-
motor, oculomotor, ‘prefrontal’ and ‘limbic’ func- ings are noisy and stochastic. Given the stochas-
tions. Prog Brain Res 85:119–146 tic responses of neurons to stimuli, Bayesian
Bergman H, Feingold A, Nini A, Raz A, Slovin H,
methods can be used to infer the underlying
Abeles M, Vaadia E (1998) Physiological aspects of
information processing in the basal ganglia of normal stimuli or activation of the neurons, X1:n
and parkinsonian primates. Trends Neurosci 21:32–38 (▶ Electrophysiology Analysis, Bayesian).
8 Biochemical Signaling Pathways and Diffusion: Overview
For imaging data, an underlying neural

activity X1:n, e.g., firing rate at the level of Biochemical Signaling Pathways and
individual columns of cortex, is assumed to give Diffusion: Overview
rise to measured responses d1:n, e.g., the blood
flow measured by fMRI, through a stochastic Kim T. Blackwell
process. Inverting the process through the Molecular Neuroscience Department, Krasnow
Bayesian inference allows for estimating the Institute for Advanced Study, George Mason
unknown neural activity (▶ Imaging Analysis, University, Fairfax, VA, USA
Bayesian).
The ideas can also be useful for inferring
properties about the behavior of individual Detailed Description
human subjects, X, by the assumption of
a stochastic process, P(d1:n|X), through which Signaling pathways modulate the function of
characteristics of each individual subject leads neurons and neuronal networks through diverse
to individual choices in an experimental task, processes. The most well-known function of sig-
d1:n (▶ Behavioural Analysis, Bayesian). naling pathways is synaptic plasticity, which con-
The abovementioned techniques all use trols neuronal networks via modulation of the
Bayesian inference to infer underlying properties strength of synaptic connections. Signaling path-
of recorded data but is essentially used as very ways also are critical for neuronal development,
effective tools for data analysis. An alternative axon guidance, and regulation of transcription
line of research takes as the working hypothesis and translation. Signaling pathways are activated
that the human brain has evolved to the point by the G protein-coupled transmembrane recep-
of itself approximating an ideal Bayesian tors, such as metabotropic glutamate receptors or
observer (sometimes referred to colloquially as noradrenergic receptors; by the receptor tyrosine
the Bayesian brain hypothesis). Accordingly this kinases; and by calcium influx through NMDA
line of research compares human behavior to receptors or voltage-dependent calcium channels.
the output of such an ideal observer within,
e.g., perceptual (▶ Perception, Bayesian Models Calcium
of) or cognitive tasks (▶ Cognition, Bayesian Due to the importance of calcium, its concentra-
Models of). tion is tightly regulated by buffers and pumps.
A related effort has proposed that the compu- One of these calcium buffers, known as calmod-
tations necessary to perform the steps of Bayesian ulin, is not inert; rather, it can activate diverse
inference can be done through populations of enzymes such as adenylyl cyclase, calcineurin,
biological neurons. phosphodiesterase type 1B, and calcium-
There is a continued effort to translate ideas on calmodulin-dependent protein kinase II. In addi-
Bayesian inference from machine learning and tion to calcium influx through plasma membrane
computer science into computational neurosci- channels, both the mitochondria and the smooth
ence, including recent advances in sampling endoplasmic reticulum (SER) are sources of cal-
techniques for inference. cium. Two types of calcium permeable channels
reside on the SER: the inositol trisphosphate
receptor channel and the ryanodine receptor chan-
Cross-References nel. Calcium-dependent calcium release through
these channels can lead to oscillations or waves of
▶ Behavioural Analysis, Bayesian calcium, depending on various factors.
▶ Cognition, Bayesian Models of
▶ Electrophysiology Analysis, Bayesian Kinases
▶ Imaging Analysis, Bayesian The second messengers activated through trans-
▶ Perception, Bayesian Models of membrane receptors have multiple downstream
Brain Imaging: Overview 9
targets, including ionic channels, kinases, and ▶ Calmodulin, Models of

phosphatases. Of the thousands of kinases and ▶ Cerebellum: Overview
phosphatases in the proteome, several have ▶ Deterministic Reaction-Diffusion Simulators
a demonstrated role in synaptic plasticity, though ▶ Diffusion Equation
their relative importance depends on the brain ▶ Enzyme Kinetics, Modeling of
region and cell type. Protein kinase type C is ▶ Extracellular Signal-Regulated Kinases,
a calcium- and lipid-activated kinase which is Models of
critical for LTP in the cerebellum. Protein kinase ▶ Gillespie Algorithm for Biochemical Reaction
type A is a cAMP-activated kinase, which is Simulation
critical for LTP in the striatum, and for several ▶ High-Voltage-Activated Calcium Channels
long-lasting forms of LTP in the hippocampus. ▶ Metabotropic Receptors (G Protein Coupled
Gene transcription and protein translation are Receptors)
required for both memory and for long-lasting ▶ N-Methyl-D-Aspartate (NMDA) Receptors,
forms of synaptic plasticity. One kinase that Conductance Models
appears to bridge these other kinases and tran- ▶ Particle-Based Stochastic Simulators
scription is the ERK1/2 forms of MAPK ▶ Protein Kinase A, Models of
(mitogen-activated protein kinase). ▶ Protein Kinase C, Models of
▶ Signaling Pathways, Modeling of
Modeling Techniques ▶ Stochastic Simulators
Calcium dynamics and signaling pathways are
modeled as cascades of biochemical reactions,
both bimolecular reactions and enzyme reactions,
and diffusion. Many special purpose simulators
are available to implement these signaling path- Brain Imaging: Overview
ways, either using stochastic techniques or deter-
ministic approaches. Modeling calcium influx Jorge Riera
requires the simulator to have capabilities for Department of Biomedical Engineering, Florida
modeling membrane potential. Due to the diver- International University, Miami, FL, USA
sity in temporal and spatial scale involved in
modeling neuronal electrical activity coupled to
reaction-diffusion pathways, very few models Synonyms
incorporate signaling pathways in entire neurons,
though the number of such models is increasing Functional neuroimaging
as computational power increases.
Definition
Cross-References
Brain imaging constitutes a set of techniques used
▶ Bimolecular Reactions, Modeling of to measure functional activity in networks of
▶ Calcium Buffering: Models of interacting brain cells as well as to reveal the
▶ Calcium Dynamics in Neuronal major underlying structural properties of these
Microdomains: Modeling, Stochastic networks. Functional data obtained with
Simulations, and Data Analysis these techniques reflect local changes in brain
▶ Calcium Pumps, Models of perfusion, metabolism, and extracellular elec-
▶ Calcium Release, Models of tric/magnetic potentials originated from the
▶ Calcium Waves, Models of activity of these brain cell networks. Brain imag-
▶ Calcium-Dependent Exocytosis, Biophysical ing has been traditionally used to explore both
Models of normal and pathological brains in a large variety
10 Brain Imaging: Overview
of species ranging from rodents to primates. population of neurons in extended brain regions.
In principles, brain imaging could be applied for IOS and OCT are employed mainly to record
both in vitro and in vivo situations. changes in blood flow/volume in the brain as
well as alterations in blood oxygen concentration
(see “▶ Voltage Sensitive Dye Imaging, Intrinsic
Detailed Description Optical Signals”).
Anatomical refers to static pictures of the cel-
Techniques for brain imaging have been devel- lular networks directly reflecting structural
oped for two major physical brain scales, i.e., the properties. These images are obtained using dif-
mesoscale and the macroscale. Theoretical ferent modalities of sectioning microscopy
models useful to interpret how these two physical (see “▶ Physical Sectioning Microscopy”).
scales interact have been developed in the past
(see “▶ Multiscale Brain Connectivity”). Brain Neuroimaging at the Macroscale
imaging modalities have been combined with Functional refers to observations reflecting the
specific stimulation techniques to study brain activity of large regions inside the brain. Exam-
activity in both mesoscopic (e.g., uncaging ples of functional neuroimaging for this particu-
methods, optogenetics – see “▶ Optogenetics”) lar scale are functional magnetic resonance
and macroscopic (e.g., transcranial magnetic imaging (fMRI), positron emission tomography
stimulation, TMS) levels. Brain imaging (PET), single-photon emission tomography
constitutes one of the most important building (SPECT), functional near-infrared spectroscopy
blocks in the development of brain-machine (fNIRS), and electro (EEG)/magneto (MEG) –
interfaces (see “▶ Brain-Machine Interface and encephalograms. The spatiotemporal profiles of
Neuroimaging”). these brain imaging techniques are quite different
(Riera and Valdes-Sosa 2010). Deciphering the
Neuroimaging at the Mesoscale origin and nature of signatures in the functional
Functional refers to observations reflecting the neuroimaging imprinted by abnormal brain activ-
activity of populations of cells (e.g., neurons, ity is important while using them for diagnosing,
astrocytes) in a particular brain structure with monitoring, and treating brain diseases and dis-
the resolution of single cells. Examples of func- orders (see “▶ Connectivity Analysis in Normal
tional neuroimaging for this particular scale are and Pathological Brains”). To that end, it is cru-
confocal/multiphoton (MP) microscopy, current cial to understand the physiological mechanisms
source density (CSD) analysis from intracranial underlying these brain imaging techniques (see
recordings using MEA, intrinsic optical signal “▶ Biophysical Models: Neurovascular Cou-
(IOS), optical coherence tomography (OCT), pling, Cortical Microcircuits, and Metabolism,”
and voltage-sensitive dye imaging (VSDI). “▶ Kinetic Models for PET/SPECT Imaging,”
These imaging techniques are based on either and “▶ Forward and Inverse Problems of
in vitro or very invasive in vivo experimental MEG/EEG”). These techniques classify either
protocols. Some of the imaging techniques as slightly invasive (i.e., fMRI, fNIRS,
based on optical phenomena (confocal/MP EEG/MEG) or extremely invasive because of
microscopy, VSDI – see “▶ Voltage Sensitive the use of radioisotopes (i.e., PET, SPECT, see
Dye Imaging, Intrinsic Optical Signals”) provide “▶ Radiopharmaceuticals in Molecular Imag-
the ideal spatial resolution to explore the activity ing”). A variety of methods have been developed
of single cells, but their temporal resolution is in for the preprocessing and analysis of brain imag-
the order of hundreds of milliseconds. In con- ing data, with resulting software and platforms
trasts, images resulting from the CSD analysis currently available (see “▶ Software for Neuro-
could reflect electrical phenomena happening in imaging Data Analysis,” “▶ Statistical Analysis
the order of a few milliseconds, but they are of Neuroimaging Data,” and “▶ Meta-analysis in
associated with the activity of synchronized Neuroimaging”).
Brain-Machine Interface: Overview 11
Anatomical refers to static pictures of the References

entire brain, or part of it, directly reflecting
structural properties. Example of imaging Riera J, Valdes-Sosa P (2010) Mesoscale in neuroimag-
ing: creating bridges between the microscopic and
modalities used to study morphometric charac-
system levels. J Integr Neurosci 9(4):v–vii
teristics of brain tissues are the T1/T2 magnetic
resonance imaging (MRI) and the computed
tomography (CT). Diffusion tensor imaging
(DTI) has been commonly employed to study
the strengths of connections between different Brain-Machine Interface: Overview
brain regions.
Karim G. Oweiss
Electrical and Computer Engineering,
Cross-References Neuroscience and Cognitive Science, Michigan
State University, East Lansing, MI, USA
▶ Applications of Information Theory to
Analysis of Neural Data
▶ Biophysical Models: Neurovascular Coupling, Abbreviations
Cortical Microcircuits, and Metabolism
▶ Brain Atlases AP Action potentials
▶ Brain Extracellular Space: A Compartment for BCI Brain-computer interfaces
Intercellular Communication and Drug BMI Brain-machine interface
Delivery ECoG Electrocorticogram
▶ Brain-Machine Interface and Neuroimaging EEG Electroencephalogram
▶ Connectivity Analysis in Normal and EMG Electromyogram
Pathological Brains FES Functional electrical stimulation
▶ Current Source Density (CSD) Analysis LTD Long-term depression
▶ Electrophysiology Analysis, Bayesian LTP Long-term potentiation
▶ Forward and Inverse Problems of MEG/EEG MIMO Multi-input multi-output
▶ Imaging Analysis, Bayesian NI Neural interfaces
▶ Independent Component Analysis of Images SCI Spinal cord injury
▶ Kinetic Models for PET/SPECT Imaging SISO Single-input single-output
▶ Local Field Potential, Relationship to BOLD
Signal
▶ Local Field Potential, Relationship to Synonyms
Electroencephalogram (EEG) and
Magnetoencephalogram (MEG) Brain-computer interfaces; Neural interfaces
▶ Meta-analysis in Neuroimaging
▶ Multiscale Brain Connectivity
▶ Neuroimaging, Neural Population Models for Definition
▶ Noninvasive Brain-Computer Interfaces
▶ Optogenetics A brain-machine interface (BMI) is a direct com-
▶ Physical Sectioning Microscopy munication pathway between the nervous system
▶ Radiopharmaceuticals in Molecular Imaging and a man-made computing device. This commu-
▶ Reconstruction, Electron Microscopy nication is unidirectional in BMIs that either
▶ Software for Neuroimaging Data Analysis record neural activity in the nervous system to
▶ Statistical Analysis of Neuroimaging Data affect the state of an external device or stimulate
▶ Voltage Sensitive Dye Imaging, Intrinsic neural activity to affect the state of the nervous
Optical Signals system. It can also be bidirectional, such as BMIs
12 Brain-Machine Interface: Overview
that record activity from certain parts of the ner- of many neurons (tens to hundreds) in awake
vous system and use this activity – or features behaving subjects, which eventually paved
extracted from it – in real time to stimulate activ- the way for these arrays to become a key
ity in other parts of that system. This communi- element in BMIs development in subsequent
cation can occur at multiple levels, which may years (Nicolelis 1999).
include muscles, peripheral nerves, spinal cord,
or the brain.
Unidirectional BMIs
Detailed Description Afferent BMIs

Afferent BMIs (ABMIs) rely on transforming
BMIs fundamentally rely on the concept of cau- features of sensory stimuli (e.g., auditory, visual,
sation between electricity and movement or etc.) to electrical pulse trains in order to stimulate
between electricity and cognition. The causal neural activity in the central or the peripheral
link between electrical current injection into the nervous systems to cause artificial sensation to
body and movement of parts of that body was first compensate for some form of sensory loss (e.g.,
established in the late eighteenth century by deafness or blindness) (Fig. 1). Auditory prosthe-
Galvani (1791), Fowler and Galvani (1793), sis, such as cochlear implants (CIs) – dating back
who could evoke muscle twitches in the frog to the first implant in 1957 by Djorno and
legs by direct current injection into the muscle Eyries – is the first example of an afferent BMI
ex vivo. A similar discovery in the central ner- that was approved by the FDA in 1984 (House
vous system was made in 1870 by Fritsch and and Urban 1973). CIs transform sound features
Hitzig (1870), in which electrical stimulation of (e.g., intensity or pitch) recorded through
different areas of the cerebral cortex caused mus- a microphone to pulsatile currents in the spiral
cular contractions of specific parts of a dog’s ganglia. This eventually elicits action potentials
body in vivo. It was not until 1928 that recording from residual hair cells in the auditory nerve of
of the all-or-none electrical discharge of single hearing-impaired subjects.
neurons in the optic nerve of the toad was made Likewise, visual prosthesis rely on
by Adrian (Adrian and Bronk 1928) and was transforming features of the visual scenes
found to be strongly correlated with light stimuli. recorded through a video camera to stimulate
Ever since, countless studies have revealed different parts of the visual pathway in legally
numerous mechanisms of neural coding of sen- blind subjects. The site of stimulation along this
sory stimuli or movement parameters, such as pathway is a function of where neural degenera-
orientation tuning by V1 neurons (Hubel and tion occurs, but the most promising demonstra-
Wiesel 1962), spatial position by hippocampal tion of visual prosthesis thus far has been through
place cells (Ranck 1973; O’Keefe 1979)) and the stimulation of the retinal ganglion cell layer
movement direction by primary motor cortex that provide input to the optic nerve in patients
neurons (Georgopoulos et al. 1986). with retinitis pigmentosa and age-related macular
The early 1990s nonetheless has witnessed degeneration (de Balthasar et al. 2008).
a paradigm shift in the way recording and stimu-
lation of neural activity is achieved. In particular, Efferent BMIs
microwire bundles have been used to record the EBMIs rely on extracting features from recorded
activity of a handful of neurons in awake behaving neural activity in real time that are subsequently
animals (McNaughton et al. 1983). Striking transformed into control signals for actuating an
advances in the microfabrication technology of external device (e.g. a paralyzed limb) (Fig. 2).
high-density microelectrode arrays (HDMEAs) They can be categorized based on the recorded
(Drake et al. 1988; Normann et al. 1999) have signal modality, which is predominantly neural
later permitted large-scale simultaneous recording (but see (Sitaram et al. 2009) for an example of
Brain-Machine Interface: Overview, Fig. 1 Basic ele- stimulate target areas in the afferent pathway of the
ments of an afferent BMI. Features are extracted from the corresponding sensory modality
sensory stimuli and then converted to electrical pulses that
Brain-Machine Interface: Overview, Fig. 2 Basic ele- together within the transform block, aka the
ments of an efferent BMI. The read out of neural activity “decoder” – to actuate an artificial device. The user is
can be a single signal source recorded from a target area able to continuously monitor changes in the state of the
within an efferent pathway – or multiple sources fused actuator in real time, typically through visual feedback
Space
(mm)
EEG
1
−3
10
ECoG
Brain-Machine Interface:
Overview, Fig. 3 Spatial, −6 10 −3 10 −2 10-1
10
1
temporal, and information
1
LFP Time
characteristics of neural (sec)
10
signals recorded from the 2 Action

10 Potentials
brain for efferent BMI
operation (Adapted from Information
Oweiss 2010) Rate (bps)
metabolic activity that utilizes blood-oxygen- provide a glimpse over the synchronized activity
level dependent (BOLD)). of large populations of neurons within a few hun-
In EBMIs, the neural readout may differ in the dred microns from the electrode tip (Mitzdorf Jan
temporal and spatial scales of variations, as well as 1985; Katzner et al. 2009). Macroelectrodes, either
in information content, as shown in Fig. 3. While implanted subdurally to record ECoG or fixed
the 1–2 ms APs elicited by individual neurons are extracranially to record surface EEG, offer another
known to contain large information about behav- less risky alternative for EBMI since they are less-
ioral covariates (typically measured in bits per or noninvasive (Wolpaw et al. 2002; McFarland
second), they tend to be highly variable and can et al. Jan 2005; Thongpang et al. 2011).
only be recorded thus far using penetrating micro- Historically, work in efferent BMIs (EBMIs) is
electrodes (McNaughton et al. 1983; Drake more recent than ABMIs. The first EBMI dates
et al. 1988; Normann et al. 1999; Nicolelis 1999), back to the pioneering work of Fetz (1969) and
or using voltage-sensitive dye-based calcium was based on a single-neuron recording from
imaging at shallow cortical depths (Koester and a monkey’s precentral “motor” cortex. In this set-
Sakmann 2000; Stosiek et al. 2003) – although the ting, the BMI was a single-input/single-output
latter signals have not been used in BMIs. While (SISO) system where the firing rate of a single
local field potentials (LFPs) can also be recorded neuron was integrated over small time intervals
using penetrating electrodes, they contain less and used to control an illuminated meter whose
information than spike trains and are believed to pointer deflection was proportional to the activity
Brain-Machine Interface: Overview, Fig. 4 Basic ele- information back to the nervous system through stimula-
ments of a bidirectional – or recurrent – BMI. Neural tion patterns of relevant areas in the peripheral nerve, the
activity is read out to control an actuator. In turn, changes spinal cord, or the brain
in the actuator state are measured and used to provide
integrator’s output. The animal had to volitionally the input to the transform to compute stimulation
modulate the firing rate of the selected neuron in parameters in real time. As such, they are consid-
order to bring the meter to a predetermined thresh- ered “closed loop” compared to unidirectional
old for reward. This operant conditioning para- BMIs that are “open loop.” This specific feature
digm constitutes the first proof of concept of an allows stimulation to be dynamic and to follow
EBMI, though was not used to actuate any limbs. the dynamics of the neural input to the transform,
This approach has been extended in modern which may be volitionally modulated by the
EBMIs to involve the simultaneous recording subject in certain applications (Fig. 4).
and use of hundreds of neurons that are subse-
quently transformed through a “decoding” filter Sensorimotor BMIs
to generate control signals that actuate multiple This class of bidirectional BMIs combines
degrees of freedom (DOFs) (Hochberg recording and stimulation to restore one or more
et al. 2012). As such, these BMIs are considered functions. For example, motor signals can be
multi-input/multi-output systems (MIMOs). recorded from cortical areas and used to directly
control muscle contraction or hand grasp via FES
Cerebral BMIs (Moritz et al. 2008; Ethier et al. 2012). These
Another class of unidirectional BMIs – neither signals can also be used to stimulate the spinal
designed to directly cause sensation nor produce cord below the injury site using parameters
movements – is designed to stimulate neurons in extracted from cortical signals in real time
the central nervous system to directly inhibit path- (Harkema et al. 2011; Moritz et al. Jan 2007;
ological behavior in subjects with neurological Shanechi et al. 2014) or via stimulation of periph-
impairment. An example of such systems is deep eral nerves (Navarro et al. 2005; Micera and
brain stimulators (DBS), in which an electrode Navarro 2009). In that respect, this is a class of
array is implanted in the subthalamic nucleus bidirectional BMIs that restores motor function
(STN) of a Parkinsonian patient, and macrosti- through FES.
mulation of the STN through a few electrode Another class of bidirectional BMIs combines
leads ameliorates the disease symptoms by reduc- efferent and afferent BMIs to restore movements
ing bradykinesia and tremor (Coffey Mar 2009). as well as sensory feedback in the form of
Similar designs are intended to regulate mood dis- touch and proprioception. The importance of
orders (Mayberg et al. 2005; Hajcak et al. 2010) or such design is the indispensible role that
abate epileptic seizures (Theodore and Fisher 2007; somatosensation plays in the integration and
Halpern et al. 2008; Boon et al. 2009; Jones 2010). coordination of limb movements, particularly
for complex upper limb functions such as
reaching and grasping. There is debate over
Bidirectional BMIs which site(s) to stimulate to restore somatosen-
sory feedback, but the majority of approaches
Bidirectional – or sometimes referred to as have focused on stimulation of peripheral nerves
“recurrent” – BMIs differ from unidirectional (Raspopovic et al. 2014). Peripheral stimulation,
BMIs in that neural measurements are used in however, is not feasible in patients with spinal
cord injury (SCI). Recent reports devised proto- Bidirectional BMIs for Neurological Disorders
cols for these patients based on subcortical (Daly This class of BMIs extends cerebral BMIs to
et al. 2012; Liu et al. 2011) or cortical stimulation include a closed loop design that records neural
(Berg et al. 2013; Tabot et al. 2013). activity and uses features extracted from this
A third class of bidirectional BMIs aims to recordings to dynamically adjust stimulation pat-
induce plastic changes in neural circuits by terns in real time (Rosin et al. 2011) (Carlson
using patterns of activity recorded from one et al. 2013; Afshar et al. 2012). A similar
brain site to stimulate activity in another distal approach is used to reduce or prevent seizures
brain site in real time (Jackson et al. 2006). Key to from occurring in drug-resistant epileptic
induce this plasticity is to ensure stimulation is patients (Nagel and Najm 2009) (Morrell and
delivered within a few milliseconds of recording R. N. S. S. i. E. S. Group 2011). This strategy is
spike events so as to promote LTP and LTD of more advantageous compared to resection of
synaptic connections, consistent with Hebbian the parts of the brain where hypothesized epileptic
plasticity (Hebb 1949). It is contended that foci reside, which has been the standard clinical
this type of plasticity triggers functional as treatment for many years. Other examples use
well as structural changes in the targeted a similar concept to treat brain injury or psychiatric
neural circuits (Lucas and Fetz 2009); diseases, such as traumatic brain injury (Schiff
the longevity of this plasticity, however, is yet et al. 2007), major depression (Malone et al.
to be demonstrated, since neurons return to 2009), obsessive compulsive disorders (Goodman
their original encoding properties shortly after et al. 2010), among others (Mohr 2008).
stimulation is terminated.
Cognitive BMIs
Certain brain areas have specific cytoarchi- The Computing Device
tectonic architecture with known role in cogni-
tive functions, such as the hippocampus role in A key element in BMI design is the computational
the formation and maintenance of long-term epi- capability of the device and the corresponding
sodic memory (Bliss and Collingridge 1993). number of inputs and outputs. For example, in
This knowledge is critical to the ability to restore SISO BMIs, computations can be as simple as
inter-areal as well as intra-areal information the detection of APs presence in a noisy electrode
exchange by means of artificial devices. In this recording followed by a spike count within a fixed
setting, neural activity is recorded from one time interval (usually 50 milliseconds). They can
upstream region (e.g., hippocampus CA3) also consist of complex time varying (Kalman
known to provide feedforward information to 1965) – and sometimes nonlinear (Hampson
downstream regions (e.g., CA1) (Yeckel and et al. 2013) – transformation of multiple electrode
Berger Aug 1990). When the communication recordings in MIMO BMIs. EBMIs in particular
between the two regions is disrupted due to are designed to have the decoder extract neural
malfunctioning neural tissue, eventually causing features (typically the firing rate of individual
memory loss, activity recorded from CA3 is cells) to determine the “state” of the population.
“decoded” to predict the input to the CA1 region. Because this state varies in time, it can be
The BMI then uses this prediction to derive the described by a “trajectory” in a neural state space
stimulation patterns needed to evoke biomimetic (Oweiss 2010; Badreldin et al. 2013; Churchland
activity at the output of the CA1 region. Thus, the et al. 2012). The transform then filters this state
transform model in this BMI design lumps to reduce its dimension to a smaller number of
both a “decoder” of the input neural activity in variables that can be subsequently used to control
CA3 and an “encoder” of the patterns to be the actuator (Gilja et al. 2012). As such, the subject
evoked in CA1 through stimulation in one block is required to learn this transformation with
(Song et al. 2007). extended practice.
Likewise, in ABMIs, the computing device References

extracts features from signals measured in the
outside world to influence the “trajectory” of the Adrian ED, Bronk DW (1928) The discharge of impulses
in motor nerve fibres: part I. Impulses in single fibres
stimulation parameter(s) in a stimulation param-
of the phrenic nerve. J Physiol 66:81–101
eter space. Because the mapping between the Afshar P, Khambhati A, Stanslaski S, Carlson D,
trajectory of natural stimuli in the task space Jensen R, Linde D et al (2012) A translational platform
and the evoked response in the neural state for prototyping closed-loop neuromodulation systems.
Front Neural Circuit 6:117
space – which is typically of much higher
Badreldin I, Sutherland J, Vaidya M, Elerya A,
dimension – is unknown, methods for finding Balasubramanian K, Fagg A, et al. (2013)
the “best” stimulation strategies in “open loop” Unsupervised decoder initialization for brain-machine
BMIs are far from optimal. As such, the experi- interfaces using neural state space dynamics.
Presented at the IEEE international conference on
menter/clinician assumes the role of finding an
neural engineering, San Diego, 2013
optimal “tune-up” of stimulation parameters to Berg JA, Dammann JF 3rd, Tenore FV, Tabot GA, Boback
restore a specific function by trial and error. For JL, Manfredi LR et al (2013) Behavioral demonstra-
example, stimulation parameters (e.g., pulse fre- tion of a somatosensory neuroprosthesis. IEEE Trans
Neural Syst Rehabil Eng 21:500–507
quency) in DBS systems are manually adjusted to
Bliss TV, Collingridge GL (1993) A synaptic model of
fit each patient and reduce tremor (Kuncel memory: long-term potentiation in the hippocampus.
et al. 2006). The lack of selectivity of electrical Nature 361:31–39
stimulation and the absence of knowledge of the Boon P, Raedt R, de Herdt V, Wyckhuys T, Vonck
K (2009) Electrical stimulation for the treatment of
dynamics of the neural circuits affected by stim-
epilepsy. Neurotherapeutics 6:218–227
ulation, however, make this approach a real chal- Carlson D, Linde D, Isaacson B, Afshar P, Bourget D,
lenge. As such, characterizing the dynamics of Stanslaski S et al (2013) A flexible algorithm frame-
stimulation-evoked neural activity in upstream or work for closed-loop neuromodulation research sys-
tems. Conf Proc IEEE Eng Med Biol Soc
downstream circuits becomes important in order
2013:6146–6150
to optimize the stimulation “dose” for restoring Churchland MM, Cunningham JP, Kaufman MT, Foster
a desired function and to minimize any side JD, Nuyujukian P, Ryu SI et al (2012) Neural popula-
effects over the long term (Liu et al. 2011; Kuncel tion dynamics during reaching. Nature 487:51–56
Coffey RJ (2009) Deep brain stimulation devices: a brief
et al. 2008; Liu et al. 2010). Thus, studying the
technical history and review. Artif Organs 33:208–220
neural system being stimulated by an artificial Daly J, Liu J, Aghagolzadeh M, Oweiss K (2012) Optimal
device may offer advantages in clinical BMI space-time precoding of artificial sensory feedback
applications. through mutichannel microstimulation in
bi-directional brain-machine interfaces. J Neural Eng
9:065004
de Balthasar C, Patel S, Roy A, Freda R, Greenwald S,
Cross-References Horsager A et al (2008) Factors affecting perceptual
thresholds in epiretinal prostheses. Invest Ophthalmol
Vis Sci 49:2303–2314
▶ Auditory Prosthesis
Drake KL, Wise KD, Farraye J, Anderson DJ, BeMent SL
▶ Cortical Motor Prosthesis (1988) Performance of planar multisite microprobes in
▶ Decoding Field Potentials recording extracellular single-unit intracortical activ-
▶ Deep Brain Stimulation (Models, Theory, ity. Biomed Eng IEEE Trans 35:719–732
Ethier C, Oby ER, Bauman MJ, Miller LE (2012) Resto-
Techniques): Overview
ration of grasp following paralysis through brain-
▶ Hippocampal Memory Prosthesis controlled stimulation of muscles. Nature
▶ Neural Decoding 485:368–371
▶ Noninvasive Brain-Computer Interfaces Fetz EE (1969) Operant conditioning of cortical unit
activity. Science 163:955–958
▶ Recurrent Brain-Computer Interfaces
Fowler R, Galvani L (1793) Experiments and observations
▶ Somatosensory Prosthesis relative to the influence lately discovered by
▶ Vestibular Prosthesis, Interface M. Galvani and commonly called animal electricity.
▶ Vision Prosthesis Printed for T. Duncan: etc, Edinburgh
Fritsch G, Hitzig E (1870) Ueber die elekrtische Katzner S, Nauhaus I, Benucci A, Bonin V, Ringach DL,
erregbarkeit des gross-hirns. Arch Anat Physiol Carandini M (2009) Local origin of field potentials in
37:300–332 visual cortex. Neuron 61:35–41
Galvani L (1791) D viribus electricitatis in Koester HJ, Sakmann B (2000) Calcium dynamics asso-
motu musculari: Commentarius. Tip. Istituto ciated with action potentials in single nerve terminals
delle Scienze, Bologna, 58 p: 4 tavv. ft; in 4.; of pyramidal cells in layer 2/3 of the young rat neo-
DCC. f. 70, vol 1 cortex. J Physiol 529 Pt 3:625–646
Georgopoulos AP, Schwartz AB, Kettner RE (1986) Neu- Kuncel AM, Cooper SE, Wolgamuth BR, Clyde MA,
ronal population coding of movement direction. Sci- Snyder SA, Montgomery EB Jr et al (2006) Clinical
ence 233:1416–1419 response to varying the stimulus parameters in deep
Gilja V, Nuyujukian P, Chestek CA, Cunningham JP, Yu brain stimulation for essential tremor. Mov Disord
BM, Fan JM et al (2012) A high-performance neural 21:1920–1928
prosthesis enabled by control algorithm design. Nat Kuncel AM, Cooper SE, Grill WM (2008) A method to
Neurosci 15:1752–1757 estimate the spatial extent of activation in thalamic
Goodman WK, Foote KD, Greenberg BD, Ricciuti N, deep brain stimulation. Clin Neurophysiol
Bauer R, Ward H et al (2010) Deep brain stimulation 119:2148–2158
for intractable obsessive compulsive disorder: pilot Liu J, Khalil H, Oweiss K (2010) Feedback control of the
study using a blinded, staggered-onset design. Biol spatiotemporal firing pattern of a basal ganglia micro-
Psychiatry 67:535–542 circuit model. In: Computational neuroscience, San
Hajcak G, Anderson BS, Arana A, Borckardt J, Takacs I, Antonio
George MS et al (2010) Dorsolateral prefrontal cortex Liu JB, Khalil HK, Oweiss KG (2011) Neural feedback for
stimulation modulates electrocortical measures of instantaneous spatiotemporal modulation of afferent
visual attention: evidence from direct bilateral epidu- pathways in bi-directional brain-machine interfaces.
ral cortical stimulation in treatment-resistant mood IEEE Trans Neural Syst Rehabil Eng 19:521–533
disorder. Neuroscience 170:281–288 Lucas TH, Fetz EE (2009) Motor cortex plasticity driven
Halpern CH, Samadani U, Litt B, Jaggi JL, Baltuch GH by artificial feedback from an autonomous, closed-
(2008) Deep brain stimulation for epilepsy. Neurother- loop neural implant. Neurosurgery 65:420–421
apeutics 5:59–67 Malone DA Jr, Dougherty DD, Rezai AR, Carpenter LL,
Hampson RE, Song D, Opris I, Santos LM, Shin DC, Friehs GM, Eskandar EN et al (2009) Deep brain
Gerhardt GA et al (2013) Facilitation of memory stimulation of the ventral capsule/ventral striatum for
encoding in primate hippocampus by a treatment-resistant depression. Biol Psychiatry
neuroprosthesis that promotes task-specific neural fir- 65:267–275
ing. J Neural Eng 10:066013 Mayberg HS, Lozano AM, Voon V, McNeely HE,
Harkema S, Gerasimenko Y, Hodes J, Burdick J, Seminowicz D, Hamani C et al (2005) Deep brain
Angeli C, Chen Y et al (2011) Effect of epidural stimulation for treatment-resistant depression. Neuron
stimulation of the lumbosacral spinal cord on volun- 45:651–660
tary movement, standing, and assisted stepping after McFarland DJ, Sarnacki WA, Vaughan TM, Wolpaw JR
motor complete paraplegia: a case study. Lancet (2005) Brain-computer interface (BCI) operation: sig-
377:1938–1947 nal and noise during early training sessions. Clin
Hebb DO (1949) The organization of behavior; Neurophysiol 116:56–62
a neuropsychological theory. Wiley, New York McNaughton BL, O’Keefe J, Barnes CA (1983) The
Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, stereotrode: a new technique for simultaneous isola-
Simeral JD, Vogel J et al (2012) Reach and grasp by tion of several single units in the central nervous sys-
people with tetraplegia using a neurally controlled tem from multiple unit records. J Neurosci Methods
robotic arm. Nature 485:372–375 8:391–397
House WF, Urban J (1973) Long term results of electrode Micera S, Navarro X (2009) Bidirectional interfaces with
implantation and electronic stimulation of the cochlea the peripheral nervous system. Int Rev Neurobiol
in man. Ann Otol Rhinol Laryngol 82:504–517 86:23–38
Hubel DH, Wiesel TN (1962) Receptive fields, binocular Mitzdorf U (1985) Current source-density method and
interaction and functional architecture in the cat’s application in cat cerebral cortex: investigation of
visual cortex. J Physiol 160:106 evoked potentials and EEG phenomena. Physiol Rev
Jackson A, Mavoori J, Fetz EE (2006) Long-term motor 65:37–100
cortex plasticity induced by an electronic neural Mohr P (2008) Deep brain stimulation in psychiatry.
implant. Nature 444:56–60 Neuro Endocrinol Lett 29(Suppl 1):123–132
Jones N (2010) Epilepsy: DBS reduces seizure frequency Moritz CT, Perlmutter SI, Fetz EE (2008) Direct control of
in refractory epilepsy. Nat Rev Neurol 6:238 paralysed muscles by cortical neurons. Nature
Kalman RE (1965) Irreducible realizations and the 456:639–642
degree of a rational matrix. J Soc Ind Appl Math Moritz CT, Lucas TH, Perlmutter SI, Fetz EE (2007) Fore-
13:520–544 limb movements and muscle responses evoked by
18 Brain-Scale Networks: Overview
microstimulation of cervical spinal cord in sedated Thongpang S, Richner TJ, Brodnick SK, Schendel A,
monkeys. J Neurophysiol 97:110–120 Kim J, Wilson JA et al (2011) A micro-
Morrell MJ, R. N. S. S. i. E. S. Group (2011) Responsive electrocorticography platform and deployment strate-
cortical stimulation for the treatment of medically gies for chronic BCI applications. Clin EEG Neurosci
intractable partial epilepsy. Neurology 77:1295–1304 42:259–265
Nagel SJ, Najm IM (2009) Deep brain stimulation for Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G,
epilepsy. Neuromodulation 12:270–280 Vaughan TM (2002) Brain-computer interfaces for
Navarro X, Krueger TB, Lago N, Micera S, Stieglitz T, communication and control. Clin Neurophysiol
Dario P (2005) A critical review of interfaces with the 113:767–791
peripheral nervous system for the control of Yeckel MF, Berger TW (1990) Feedforward excitation of
neuroprostheses and hybrid bionic systems. the hippocampus by afferents from the entorhinal cor-
J Peripher Nerv Syst 10:229–258 tex: redefinition of the role of the trisynaptic pathway.
Nicolelis MAL (1999) Methods for neural ensemble Proc Natl Acad Sci U S A 87:5832–5836
recordings. CRC Press, Boca Raton
Normann R, Maynard EM, Rousche PJ, Warren DJ
(1999) A neural interface for a cortical vision prosthe-
sis. Vision Res 39:2577–2587
O’Keefe J (1979) A review of the hippocampal place cells. Brain-Scale Networks: Overview
Prog Neurobiol 13:419–439
Oweiss K (2010) Statistical signal processing for neuro- Theoden I. Netoff
science and neurotechnology, 1st edn. Academic/
Elsevier, Burlington
Department of Biomedical Engineering,
Ranck JB Jr (1973) Studies on single neurons in dorsal University of Minnesota, Minneapolis, MN, USA
hippocampal formation and septum in unrestrained
rats: part I. Behavioral correlates and firing repertoires.
Exp Neurol 41:462–531
Raspopovic S, Capogrosso M, Petrini FM, Bonizzato M,
Definition
Rigosa J, Di Pino G et al (2014) Restoring natural
sensory feedback in real-time bidirectional hand pros- The brain is composed of many neurons, func-
theses. Sci Transl Med 6:222ra19 tional areas, and layers. Together these compo-
Rosin B, Slovik M, Mitelman R, Rivlin-Etzion M, Haber
SN, Israel Z et al (2011) Closed-loop deep brain stim-
nents work as a network to produce behavior. At
ulation is superior in ameliorating parkinsonism. Neu- minimum, the network behavior is determined by
ron 72:370–384 four things: (1) the network inputs, (2) the
Schiff N, Giacino J, Kalmar K, Victor J, Baker K, Gerber dynamics of the individual nodes, (3) the cou-
M et al (2007) Behavioural improvements with tha-
lamic stimulation after severe traumatic brain injury.
pling functions between the nodes, and (4) the
Nature 448:600–603 topology. This encyclopedia section will provide
Shanechi MM, Hu RC, Williams ZM (2014) a brief overview on characterizing brain-scale
A cortical–spinal prosthesis for targeted limb move- networks. Often the coupled system will have
ment in paralysed primate avatars. Nat Commun
5:3237
emergent behaviors, behaviors that could not be
Sitaram R, Caria A, Birbaumer N (2009) Hemodynamic predicted from analysis of the individual compo-
brain–computer interfaces for communication and nents alone. Understanding how the brain func-
rehabilitation. Neural Netw 22:1320–1328 tions requires an understanding of how
Song D, Chan RHM, Marmarelis VZ, Hampson RE,
Deadwyler SA, Berger TW (2007) Nonlinear dynamic
components work together in a network. In
modeling of spike train transformations for many diseases the cause cannot be pinpointed to
hippocampal-cortical prostheses. Biomed Eng IEEE dysfunction or failure of a single component,
Trans 54:1053–1066 such as an ion channel mutation. Instead, subtle
Stosiek C, Garaschuk O, Holthoff K, Konnerth A (2003)
In vivo two-photon calcium imaging of neuronal net-
changes in cellular behavior may lead homeo-
works. Proc Natl Acad Sci 100:7319 static mechanisms to alter the coupling between
Tabot GA, Dammann JF, Berg JA, Tenore FV, Boback JL, the neurons and brain areas, resulting in patho-
Vogelstein RJ et al (2013) Restoring the sense of touch logical activity such as synchronous population
with a prosthetic hand through a brain interface. Proc
Natl Acad Sci U S A 110:18279–18284
oscillations or unchecked excitability. The ulti-
Theodore WH, Fisher R (2007) Brain stimulation for mate goal in applying the network theory to
epilepsy. Acta Neurochir Suppl 97:261–272 understanding connections within the brain is to
Brain-Scale Networks: Overview 19
develop a measure that can explain the emer- Detailed Description

gence of pathological behaviors and to perhaps
develop approaches to treating diseases. The goal of these encyclopedia entries is to pro-
But first, we will introduce a few common vide an introduction into the network theory. In
terms. A network is a collection of coupled the first entry (▶ Network Theory in Neurosci-
components. Generally, when the components ence), there is an overview of the network theory
coupled together are different elements, it is and its applications to diseases. In “▶ Functional
referred to as a system. If the components of the Network Observations of Diseased Brain States,”
system are similar and interchangeable, it is there is an introduction to functional networks in
instead called a network. The components in neuroscience. In “▶ Determining Network Struc-
the network are referred to as nodes, which ture from Data: Nonlinear Modeling Methods,”
can be individual neurons or brain regions. The we will introduce methods for reconstructing net-
coupling between the nodes is referred to as works from the data using nonlinear measures. In
edges, which can be synapses or fiber paths. “▶ Master Stability Function for Globally Syn-
Coupling in the nervous system is generally chronized Systems,” we introduce a universal
through chemical synapses which are direc- approach to determining if a network will syn-
tional, where the coupling of neuron or region chronize given the dynamics of the individual
A onto B will be different than B onto A. components and the network topology, through
However, for networks of neurons coupled the analysis of the master stability function. In
through electrical synapses, the coupling can be “▶ Connectionist Models of CPG Networks,” we
undirected. Networks are considered weighted if will provide an introduction to nonsynchronous
the strengths of coupling between nodes have network behaviors, such as seen in central pattern
a distribution and unweighted if they are all generators. In “▶ Neuropathologies and Net-
the same. works,” we will introduce pathological network
Generally, the dynamics of the nodes and cou- function to characterize diseases.
pling function are highly nonlinear. A linear
response is defined as given twice the input, the
output will be twice as strong. However, because
Cross-References
neurons have thresholds and synapses are plastic,
the responses are very nonlinear. Furthermore,
▶ Connectionist Models of CPG Networks
neurons are a mixture of deterministic behavior,
▶ Determining Network Structure from Data:
where its behavior can be determined from its
Nonlinear Modeling Methods
past and its inputs, and stochastic, where the
▶ Functional Network Observations of Diseased
activity is also due to some noise in the system
Brain States
which cannot be accounted for.
▶ Master Stability Function for Globally
The statistics of the coupling within a network
Synchronized Systems
is called the topology. A list of all the connec-
▶ Network Theory in Neuroscience
tions within a network is called the graph. If
▶ Neuropathologies and Networks
nodes have a physical location in space, such as
brain regions, and coupling is dependent on the
distance, the network is considered to have
Further Reading
a geometry.
In summary, neuronal networks are nonlinear, Boccaletti S, Latora V, Moreno Y, Chavez M, Hwang DU
directional weighted graphs with a geometry. (2006) Complex networks: structure and dynamics.
These networks are called complex networks, Phys Rep Rev Sect Phys Lett 424:175
Mirollo RE, Strogatz SH (1990) Synchronization of pulse-
and few tools have been developed to analyze
coupled biological oscillators. SIAM J Appl Math
them. The development of these network analysis 50:1645
tools is at the cutting edge. Sporns O (2011) Networks of the brain. MIT, Cambridge
20 Brainstem Processing: Overview
Strogatz SH (2000) From Kuramoto to Crawford: explor- “▶ Control of Breathing, Integration of Adaptive
ing the onset of synchronization in populations of Reflexes” describes computational models of
coupled oscillators. Phys D: Nonlinear Phenom 143:1
Strogatz SH (2003) Sync: the emerging science of spon- adaptation and plasticity of the respiratory
taneous order, 1st edn. Hyperion, New York rhythm in response to external perturbations.
Watts DJ, Strogatz SH (1998) Collective dynamics of There are also relevant articles in other sections
‘small-world’ networks. Nature 393:440 of this encyclopedia, including “▶ Computa-
tional Models of Mammalian Respiratory CPG”
and “▶ Brainstem Motoneurons, Models of”.
The area of pain processing has received much
Brainstem Processing: Overview less attention by computational neuroscientists.
The article titled “▶ Pain Processing Pathway
Robert Butera1 and Kendall Morris2 Models” summarizes computational efforts in
1
School of Electrical and Computer Engineering, this area.
Laboratory for Neuroengineering, Georgia
Institute of Technology, Atlanta, GA, USA
2
College of Medicine, Molecular Cross-References
Pharmacology & Physiology, University of
South Florida, Tampa, FL, USA ▶ Baroreflex Models
▶ Brainstem Motoneurons, Models of
▶ Computational Models of Mammalian
Detailed Description Respiratory CPG
▶ Control of Breathing, Integration of Adaptive
The brainstem has received considerably less Reflexes
attention by computational neuroscientists than ▶ Pain Processing Pathway Models
other regions of the mammalian brain. This sec- ▶ Pre-Botzinger Complex Rhythm Generation
tion highlights key areas that have received atten-
tion: neural circuitry underlying the generation
and control of the respiratory rhythm, the
baroreflex underlying the regulation of blood Cable Theory: Overview
pressure, and pathways for pain processing.
Entries in this section cover all three of these William R. Holmes
areas. Department of Biological Sciences, Ohio
The article titled “▶ Baroreflex Models” University, Athens, OH, USA
reviews over two decades years of computational
models focused on neural pathways where blood
flows in response to pressure changes in the cir- Detailed Description
culatory system mediated by the baroreflex
pathway. Cable theory has a prominent place in neurosci-
Much attention has been given to the modeling ence as it forms the basis of models of single
of respiratory rhythm generation. The generation neurons, axons, and dendrites and has led to an
and control of breathing is covered by two entries understanding of intraneuronal signaling.
in this section. The entry titled “▶ Pre-Botzinger Starting with the conceptual model of a patch of
Complex Rhythm Generation” surveys computa- membrane as an electric circuit, application of
tional models of rhythmic activity in the basic principles from physics leads to
pre-Bötzinger complex, which underlies the a mathematical equation, the cable equation.
inspiratory component of respiratory rhythm We start this section with an article on the cable
generation and is a central pattern-generating equation, discussing the history of the cable equa-
circuit itself when isolated. The article titled tion in neuroscience, showing the derivation of
Cerebellum: Overview 21
the cable equation, providing solutions for the Almost all of these articles assume membrane is
infinite cylinder, steady-state solutions for passive, but it is clear that both dendrites and
semi-infinite and finite cables and branched den- axons contain voltage-dependent conductances.
dritic trees and transient solutions, concluding Nevertheless, cable theory is important for pro-
with a discussion of insights from cable theory. viding a basic understanding of intraneuronal
Several parameters play key roles in signaling that forms the basis for understanding
intraneuronal signaling including membrane what happens with active conductances.
capacitance, axial resistivity, the space (length)
constant, and the time constant, and there are
articles on each of these topics. From there we Cross-References
have two articles describing work done by
Wilfrid Rall, widely regarded as the father of ▶ Cable Equation
computational neuroscience. The first describes ▶ Capacitance, Membrane
how Rall applied cable theory to dendritic trees ▶ Dendritic Spines: Continuum Theory
and showed that given certain conditions a highly ▶ Electrotonic Length, Formulas and Estimates
branched dendritic tree could be reduced mathe- ▶ Equivalent Cylinder Model (Rall)
matically to an equivalent cylinder. The simplic- ▶ Morphoelectrotonic Transform
ity and elegance of the equivalent cylinder model ▶ Quasi-active Approximation of Nonlinear
permitted mathematical analyses and insights Dendritic Cables
regarding signaling in otherwise morphologically ▶ Resistivity, Axial
complex dendritic trees. The second describes ▶ Space (Length) Constant, Lambda, in
various formulas Rall derived to estimate the Neuronal Signaling
electrotonic length of dendritic trees from simple ▶ Time Constant, Tau, in Neuronal Signaling
voltage transients, formulas that have been
applied in hundreds of studies. Finally, there are
three articles that discuss extensions of cable
theory in various ways. The first presents the Cerebellum: Overview
morphoelectrotonic transform, a procedure that
addresses the issue that attenuation of voltage Fidel Santamaria
from a synaptic input from the dendrite to the UTSA Neurosciences Institute, The University of
soma is not simply given by the electrotonic Texas at San Antonio, San Antonio, TX, USA
distance because of the effect of boundary condi-
tions. The morphoelectrotonic transform pro-
vides a graphical means to visualize voltage Definition
attenuation both toward the soma and away
from the soma, directly and intuitively. The sec- The cerebellum processes sensory and motor
ond introduces the continuum theory for model- information. Structurally is divided in the cere-
ing dendritic spines. While there are various ways bellar cortex network and the deep cerebellar
to incorporate the effect of dendritic spines in nuclei. Inputs to the cerebellar cortex arrive via
models, the continuum theory describes a novel the mossy fibers from multiple sources; mossy
way of doing this within the cable equation. fibers contact granule cells, Golgi cells and uni-
Lastly, insights can be obtained from the quasi- polar brush cells. The second input to the cere-
active approximation of a nonlinear cable. Here bellar cortex comes from the inferior olive in the
classic passive cable theory is extended by line- form of climbing fibers that synapse Purkinje
arizing voltage-dependent conductances around cells. The output of the cerebellar cortex is pro-
a given membrane potential cable allowing the vided by Purkinje cells that integrate the synaptic
effect of voltage-dependent conductances on activity of climbing fibers, granule cells, and
dendritic filtering to be studied mathematically. inhibitory interneurons. Golgi cells provide an
22 Cerebellum: Overview
inhibitory feedback mechanism to the granule prompted a discussion on how the cerebellar cor-
cell layer. The deep cerebellar nuclei are the tex processes information. Several reports have
final integrator of cerebellar information. Cells shown that Purkinje cells are activated only when
in the deep cerebellar nuclei receive input from they are directly above the cluster of stimulated
Purkinje cells, mossy fibers and climbing fibers. granule cells. The Purkinje cells along the paral-
lel fiber path are either inhibited or show no
change in firing rate. Experimental and computa-
Detailed Description tional evidence has shown that this is due to fast
feed-forward inhibition from the inhibitory inter-
Cerebellar Cortex neuron on Purkinje cells concomitantly stimu-
The widely studied cerebellar cortex is com- lated by parallel fibers. In fact, blocking of
posed, mainly, of four types of neurons: Granule inhibition reveals a beam of Purkinje cell activity
cells, Purkinje cells, inhibitory interneurons as predicted by the Beam Hypothesis (Santamaria
(stellate and basket), and Golgi cells. There are et al. 2010). Thus, the emergent understanding is
two inputs to the cerebellar cortex provided by not that different from a center surround recep-
mossy fibers and climbing fibers. Climbing fibers tive field model. Under this perspective, granule
are axons from inferior olivary neurons. Mossy cell synapses are effective in driving Purkinje
fibers are a collection of axons from multiple cells to fire only when the Purkinje cells are
areas that can carry direct sensory or cerebro close to the site of stimulation. Parallel fibers
cortical information. The arrangement of the cer- provide excitatory and feed forward inhibitory
ebellar cortical neurons is regular across the cor- input to Purkinje cells far from the site of stimu-
tex. Mossy fibers contact granule cells and Golgi lation without this stimulation being reflected in
cells. Granule cells send an axon into the molec- changes of the Purkinje cell firing rate. However,
ular layer that bifurcates. The bundle of bifur- the dendritic stimulation results in activation of
cated axons is known as the parallel fibers. voltage sensitive calcium channels and calcium
Parallel fibers contact Golgi cells, inhibitory activated potassium channels, modifying the
interneurons and Purkinje cells as they traverse excitability of the dendrite. This dendritic activa-
the cerebellar cortex. Climbing fibers synapse tion can then modify the response of the Purkinje
onto Purkinje cells. cell to direct granule cell stimulation. Therefore,
The particular arrangement of granule cell-to- the long range effect of parallel fibers is to pro-
Purkinje cell connectivity was very attractive to vide contextual information.
computational scientist early on. The ‘Beam The Purkinje cell was one of the first neurons
Hypothesis’ stated that a focal stimulation of the to be studied using the compartmental modeling
granule cells would be transformed into approach. The large amount of information on
a sequential activation of Purkinje cells along somatic and dendritic conductances permitted to
the path of the parallel fibers (Braitenberg and develop detailed compartmental models. These
Atwood 1958). The evoked inhibitory activity models have been reused and enhanced over the
from parallel fiber to inhibitory interneurons years to study different aspects of cerebellar func-
would reduce the activity of the Purkinje cells tion (De Schutter and Bower 1994; Steuber et al.
on either side of the beam of activated Purkinje 2007). There are also several models of inhibitory
cells. The feedback mechanism provided by interneurons, Golgi cells and granule cells.
Golgi cells on to granule cells would then reduce Recent work suggests that the Purkinje cell
over-excitability. The central mechanism of might be simplified to a perceptron (Clopath et
information processing in this model is the exis- al. 2012).
tence of a beam of activated Purkinje cells fol-
lowing the local stimulation of granule cells. Synaptic Plasticity
The lack of beams of Purkinje cell activity Long term depression (LTD) in the parallel fiber-
reported by multiple teams over the years has to-Purkinje cell synapses is a model of learning
Cerebellum: Overview 23
and memory (Ito 2006; Feil et al. 2003). LTD is The Deep Cerebellar Nuclei
induced by the concomitant activation of the The deep cerebellar nuclei (DCN – dentate,
climbing fiber input and stimulation of a small interpositus, and fastigii) receive inhibitory
bundle of parallel fibers. It is widely assumed that inputs from Purkinje cells and excitatory activity
the climbing fiber stimuli carries a training or from mossy fibers and climbing fibers. The cell
error signal, while the encoding signals, e.g., from the cerebellar nuclei generates the bundles
motor command, is encoded in the parallel fiber of axons that go to other areas of the brain. The
activity. The biochemical transduction cascade is cells of the DCN can be thought as the final and
very well understood and has been modeled using global integrator of the input and output signals to
mass action and Monte Carlo models (Doi et al. the cerebellum. DCN cells integrate inputs from
2005). These models and their experimental tests the brain stem, inferior olive, and spinal cord with
have pointed out to positive feedback biochemi- the Purkinje cell output from the cerebellar cor-
cal loop that triggers and sustains LTD. Briefly, tex. There are excitatory and inhibitory DCN
LTD is induced by the influx of calcium ions neurons with their electrophysiological and com-
through voltage sensitive channels and release putational contributions still being studied. The
from intracellular stores. Calcium ions activate precise spike timing control of DCN spiking by
protein kinase C (PKC) which eventually acti- Purkinje cell is believed to be the mechanism of
vates mitogen associated protein kinase the precise motor skills (Gauck and Jaeger 2000).
(MAPK). MAPK interacts with other substrates
to produce arachidonic acid which in return acti- Summary
vates more PKC. PKC then continues working in Although the function of the cerebellum is still
the feedback loop or phosphorylate AMPA being debated, how it processes information is
receptors. Phosphorylated AMPA receptors are becoming clearer due to vast amount of experi-
then internalized and the synaptic strength is mental information from ultra-structure to elec-
reduced. The activation of the feedback loop is trophysiology and behavior. Altogether, this
not a linear integrator of calcium concentration, large amount of data allows computational scien-
instead it acts as a leaky integrator, thus the total tist to build realistic models at different levels of
amount and the temporal release of calcium detail that can be experimentally tested.
determines the expression of LTD (Tanaka et al.
2007; Xia et al. 2007).
Cross-References
Other Cells in the Cerebellar Cortex
There are other types of cells not traditionally ▶ Deep Cerebellar Nuclei
included in cerebellar models. Lugaro cells are ▶ Long Term Depression in the Granule
interneurons that are located in the granule cell Cell-Purkinje Cell Synapse
layer (Lainé and Axelrad 1996). These neurons ▶ Modeling Cerebellar Learning: Input
receive input from Purkinje cells and could Minimization
be connected among themselves with gap junc- ▶ Multiscale Modeling of Purkinje Cells
tions. These cells seem to be inhibitory and ▶ Olivocerebellar Pathway
hypothesized to work in the burst frequency
response of the granule cell layer. Another type
of neuron, primarily located in the vestibulo cer- References
ebellum, is the unipolar brush cell. These
cells receive inhibitory inputs from Golgi Braitenberg V, Atwood RP (1958) Morphological obser-
cells and excitatory inputs from mossy fibers. vations on the cerebellar cortex. J Comp Neurol
109:1–33
Unipolar brush cells have multiple excitatory
Clopath C, Nadal JP, Brunel N (2012) Storage of corre-
targets in the granule cell layer (Mugnaini and lated patterns in standard and bistable Purkinje cell
Floris 1994). models. PLoS Comput Biol 8:e1002448
24 Computational Neuroanatomy: Overview
De Schutter E, Bower JM (1994) An active membrane used to describe and analyze neuroanatomy data
model of the cerebellar Purkinje cell. I. Simulation of from various sources, typically from microscopy
current clamps in slice. J Neurophysiol 71:375–400
Doi T, Kuroda S, Michikawa T, Kawato M (2005) Inositol instruments and neuroimaging. Description and
1,4,5-trisphosphate-dependent Ca2+ threshold dynam- analysis in computational neuroanatomy can
ics detect spike timing in cerebellar Purkinje cells. span across a vast range of scales, from synapses,
J Neurosci 25:950–961 dendritic spines, dendrites, axons, neurons, local
Feil R, Hartmann J, Luo C, Wolfsgruber W, Schilling K,
Feil S, Barski JJ, Meyer M, Konnerth A, De Zeeuw CI, circuits, cortical maps, and up to the whole brain.
Hofmann F (2003) Impairment of LTD and cerebellar Neuronal morphology, representation and quan-
learning by Purkinje cell-specific ablation of cGMP- tification of morphological properties, connec-
dependent protein kinase I. J Cell Biol 163:295–302 tion topology, wiring principles, growth rules,
Gauck V, Jaeger D (2000) The control of rate and timing
of spikes in the deep cerebellar nuclei by inhibition. and statistical analysis of various structural prop-
J Neurosci 20:3006–3016 erties are common topics of investigation in com-
Ito M (2006) Cerebellar circuitry as a neuronal machine. putational neuroanatomy.
Prog Neurobiol 78:272–303
Lainé J, Axelrad H (1996) Morphology of the Golgi-
impregnated lugaro cell in the rat cerebellar cortex:
a reappraisal with a description of its axon. J Comp
Neurol 375:618–640 Detailed Description
Mugnaini E, Floris A (1994) The unipolar brush cell:
a neglected neuron of the mammalian cerebellar cor-
tex. J Comp Neurol 339:174–180
Computational neuroanatomy deals with struc-
Santamaria F, Tripp PG, Bower JM (2007) Feedforward tural data at multiple scales obtained using mul-
inhibition controls the spread of granule cell-induced tiple data-acquisition modalities. At the
Purkinje cell activity in the cerebellar cortex. nanometer scale, electron microscopy is used to
J Neurophysiol 97:248–263
Steuber V, Mittmann W, Hoebeek FE, Silver RA, De
acquire ultrastructural information of synapses,
Zeeuw CI, H€ausser M, De Schutter E (2007) Cerebel- axon terminals, and dendritic spines. Advances in
lar LTD and pattern recognition by Purkinje cells. serial sectioning electron microscopy have
Neuron 54:121–136 enabled the investigation of whole neurons
Tanaka K, Khiroug L, Santamaria F, Doi T, Ogasawara H,
Ellis-Davies GC, Kawato M, Augustine GJ (2007)
at the nanometer scale. At the micrometer scale,
Ca2+ requirements for cerebellar long-term synaptic optical microscopy (bright field and fluores-
depression: role for a postsynaptic leaky integrator. cence) is used to study neuronal morphology,
Neuron 54:787–800 local circuits, and long-range projections.
Xia J, Chung HJ, Wihler C, Huganir RL, Linden DJ
(2000) Cerebellar long-term depression requires
Physical and optical sectioning is routinely used
PKC-regulated interactions between GluR2/3 and in optical microscopy to obtain 3D volume data.
PDZ domain-containing proteins. Neuron 28:499–510 At the millimeter and higher scales, magnetic
resonance and other noninvasive neuroimaging
techniques are used for mapping cortical
Computational Neuroanatomy: areas, major nuclei, and major fiber tracts in the
Overview brain.
The topics of study in computational neurosci-
Yoonsuck Choe ence are varied, including geometry, geometry
Department of Computer Science and reconstruction algorithms, morphometry and
Engineering, Texas A&M University, College statistics, visualization, connections and
Station, TX, USA connectome (full wiring diagram of the brain),
topological properties (small world, scale-free,
etc.), growth and development models, organiza-
Definition tional and optimization principles (e.g., wiring
length minimization, self-organization, etc.),
Computational neuroanatomy is a subfield of representation and ontology, neuroinformatics,
neuroscience where computational means are comparative analysis of the diseased versus the
Computational Neuroanatomy: Overview 25
healthy state, and inferring function from Techniques, and Validation talks about
structure. Early studies in computational general reconstruction methods and validation
neuroanatomy (especially at the cellular level) techniques.
were strongly motivated by the need for Finally, the fourth group of chapters consists
morphologically detailed multicompartmental of theoretical insights and computational simula-
models for use in neuron simulators such as tion methods that analyze and utilize the quanti-
NEURON, GENESIS, and others (see fied neuroanatomical data: (1) ▶ Networks/
Ascoli 2002). Networks discusses various methods for auto-
mated generation of realistic neurons and neuro-
Overview of Chapters nal circuits for use in computer simulations (also
The computational neuroanatomy section features see ▶ NeuroMorpho.org); and (2) ▶ Wiring Prin-
chapters that touch upon the topics listed above. ciples, Optimization presents optimization prin-
The chapters can be roughly grouped into ciples that could potentially be underlying
four categories: (1) brain atlases and the neuronal morphology and connectivity patterns
connectome, (2) imaging, (3) geometric recon- found in the brain.
struction, and (4) wiring principles and synthetic
models. Other Related Chapters and Resources
The first group includes four chapters: There are several chapters beyond this section (and
(1) ▶ Brain Atlases, a collection of pointers to sometimes whole sections) in this encyclopedia that
other chapters that describe in detail available are directly relevant to computational neuroanat-
brain atlases; (2) ▶ Connectome, Drosophila, omy: (1) Brain Imaging (▶ Connectivity Analysis
a chapter with a detailed survey of the known in Normal and Pathological Brains, ▶ Multiscale
partial connectomes of the Drosophila, with Brain Connectivity); (2) Brian Scale Networks
perspectives on the utility, progress, and future (▶ Network Theory in Neuroscience, ▶ Neuropa-
of Drosophila connectome research; thologies and Networks); (3) Databases in Compu-
(3) ▶ Connectome, Mouse, with a survey of tational Neuroscience (▶ NeuroMorpho.org and
publicly available mouse connectome data; many other chapters in this section are relevant);
and (4) ▶ Connectome, General, a general and (4) Model Reproducibility (▶ NeuroML).
survey of connectomics, its potentials, and its Also see the following edited books and
challenges. monographs on computational neuroanatomy:
The second group includes imaging related Ascoli (2002), Capowski (1989), and
chapters: (1) ▶ Imaging, Electron Microscopy Chung (2012).
discusses the use of electron microscopy for Notable omissions at the moment include
computational neuroanatomy, and (2) ▶ Imaging, (1) dendritic spine morphology and statistics
Specimen Preparation presents methods for (Bourne and Harris 2008), (2) delay (axonal con-
brain specimen preparation for subsequent sec- duction delay and integration time) and delay
tioning and imaging. (There are other imaging statistics (Lamme and Roelfsema 2000; Nowak
modalities that are relevant to computational and Bullier 1997), and (3) how to utilize gene
neuroanatomy such as ▶ Physical Sectioning expression data combined with structural
Microscopy and others discussed in the Brain information to infer function (Toled-Rodriguez
Imaging section.) et al. 2004).
The third group of chapters includes recon-
struction algorithms for the extraction of geomet- Conclusion
rical information from the raw image volumes: Computational neuroanatomy is an important
(1) ▶ Reconstruction, Electron Microscopy subfield in neuroscience that provides a frame-
discusses reconstruction techniques for work for organizing, integrating, analyzing,
densely packed neurons in electron microscopy and validating neuroscience data. As structure
image stacks; and (2) ▶ Reconstruction, is strongly correlated with function in the
26 Cortex: Overview
brain, modeling and analysis of the anatomical

(structural) scaffold is a necessary and important Cortex: Overview
first step toward understanding the brain.
Lars Schwabe
Department of Computer Science and Electrical
Cross-References Engineering, Adaptive and Regenerative
Software Systems, Universit€at Rostock, Rostock,
▶ Brain Atlases Germany
▶ Connectivity Analysis in Normal and
Pathological Brains
▶ Connectome, Drosophila Detailed Description
▶ Connectome, Mouse
▶ GENESIS, the GENeral NEural SImulation Anatomically, the cerebral cortex is the outer-
System most neuronal tissue of the brain, and it is
▶ Imaging, Electron Microscopy believed to play a key role in sensation, percep-
▶ Imaging, Specimen Preparation tion, higher cognitive functions, and motor con-
▶ Multiscale Brain Connectivity trol. It is a layered structure referred to as the gray
▶ Network Theory in Neuroscience matter, because it contains largely cell bodies as
▶ NeuroML compared to the white matter containing largely
▶ NeuroMorpho.org myelinated axons. The evolutionary origin can be
▶ NEURON Simulation Environment traced back to reptiles, but it first appeared as
▶ Neuropathologies and Networks a uniform structure in early mammals. The
▶ Physical Sectioning Microscopy increase in the size of this layered cortical sheet
▶ Reconstruction, Electron Microscopy during evolution is believed to be crucial for the
▶ Reconstruction, Techniques and Validation development of human cognition and ultimately
▶ Networks/Networks human culture during human brain evolution.
▶ Wiring Principles, Optimization Even though many entries in this section on “Cor-
tex: Models and Computation” are applicable to
the [hippocampus] as well, the focus is on the
References phylogenetically younger six-layered neocortex.
Why is it so interesting and important to inves-
Ascoli GA (ed) (2002) Computational neuroanatomy: tigate the cortex using modeling? If the
principles and methods. Humana Press, Totowa co-occurrence of the expansion of the neocortex
Bourne JN, Harris KM (2008) Balancing structure and
during evolution of the emergence of human cog-
function at hippocampal dendritic spines. Annu Rev
Neurosci 31:47–67 nition and culture is more than a suspicious coin-
Capowski JJ (ed) (1989) Computer techniques in neuro- cident, then understanding the cortex is essential to
anatomy. Plenum, New York/London understand the human condition. Identifying brain
Chung MK (2012) Computational neuroanatomy: the
and mind is certainly a too naı̈ve approach, but it is
methods. World Scientific, Singapore
Lamme VAF, Roelfsema PR (2000) The distinct modes of now widely accepted that – whatever the relation
vision offered by feedforward and recurrent between mind and brain is – an understanding of
processing. Trends Neurosci 23:571–579 cortex will at least constrain theories of how the
Nowak LG, Bullier J (1997) The timing of
information transfer in the visual system. Cereb Cortex
mind works. Testable theories and predictive
12:205–241 models are needed to complement conceptual
Toledo-Rodriguez M, Blumenfeld B, Wu C, Luo J, modeling and colloquial talk on that matter. More-
Attali B, Goodman P, Markram H (2004) over, understanding how the cortex operates may
Correlation maps allow neuronal electrical
help to diagnose neurological diseases earlier, to
properties to be predicted from single-cell gene
expression profiles in rat neocortex. Cereb Cortex develop more efficient treatments, and to construct
14:1310–1327 [brain-machine interfaces]. Models in general,
Cortex: Overview 27
and patient-specific models in particular, will be Robert Legenstein (“▶ Recurrent Network
useful to link measurements of macroscopic brain Models, Reservoir Computing”) summarizes the
activity obtained with brain imaging techniques to state of the art in how artificial recurrent neural
the underlying causes. This rather practical justifi- networks (RNNs) address demanding learning
cation for modeling cortex is becoming even more tasks. RNN researchers can be thought of as
relevant today as a “deliverable” of computational being in the luxurious position to investigate
neuroscience. However, accurate and faithful computations in cortex-inspired architectures
modeling has to deal with the complexity of the without the burden to comply with all the con-
neural circuits and the cellular and synaptic het- straints set by experimental neuroscience. The
erogeneity. Multiple large-scale initiatives cur- performance of these architectures and learning
rently address this by collecting massive amounts algorithms can serve as a yardstick for existing
of data to facilitate the development of faithful biologically more faithful models and as
models of the cortex. Here, the role of cortical a guideline for constructing new models. Jochen
models is in organizing and summarizing the Triesch’s contribution on “▶ Cortical Function,
data, and they could even be integrated into Normative Models of” gives a bird’s eye perspec-
(semi)automatic data-driven modeling pipelines tive on how to derive models from computational
with little intervention of a “modeler.” I argue principles. Cortical modeling involves determin-
that modeling cortex is interesting and exciting, ing model structure and parameterization, and for
because despite massive amounts of available data-driven approaches, computational neurosci-
data, the activity of modeling will remain in large ence has developed a [rich repertoire of
parts an art: Finding the right level of abstraction to methods]. The normative approach endorsed by
arrive at insights for the question at hand can Triesch is a complementary addition for doing this,
hardly be automated. which applies in particular to explaining cortical
On the one hand, the neocortex is an umbrella networks in terms of their genesis via [learning
for a set of distinct structures that differ, for mechanisms]. The contributions of Sophie
example, in terms of their function, connectivity, Deneve (“▶ Bayesian Inference with Spiking Neu-
and cytoarchitecture. On the other hand, rons”) and Walter Senn and Jean-Pascal Pfister
Mountcastle (1978) suggested that similarities of (“▶ Reinforcement Learning in Cortical Net-
these neocortical regions point to a common com- works”) are instances of this normative approach.
putational machinery in the sense that each region Deneve shows how Bayesian inference can be
has the same basic architecture and operates carried out by spiking neurons. The Bayesian
according to the same computational principles. approach turned out to be very fruitful for under-
This is also the guiding idea of this section, with- standing computations in the cortex as evident by
out accepting it as a truism or dogma. It is indeed a whole section in this encyclopedia being dedi-
conceivable that different regions of the neocortex cated to the [“Bayesian brain”]. It should be noted
operate according to fundamentally different prin- that both Deneve’s and Senn’s and Pfister’s entries
ciples, or that conventional notions of “computa- explicitly address computation with spiking neu-
tion” are not suited as a metaphor to understand rons and thus represent an explicit formal link
the cortex. The philosophy of science is not con- between computational principles and experimen-
clusive regarding a clear distinction of theory tally testable predictions at the level of individual
vs. model. However, as a tentative distinction for neurons. Both of these mutually compatible
modeling cortex, we may adopt the short-hand approaches make explicit a notion of optimality
definition that theory provides meaning to models, as required within the normative approach: The
while models explain data. Computational princi- Bayesian approach is based on a principled way
ples come close to the notion of a theory. With that of conducing logical inference under uncertainty,
reading, the entries assembled in this section cover whereas reinforcement learning is based on Bell-
the whole spectrum between theories and models man optimality, i.e., [decision making] in dynamic
of cortex. and often only partially observable environments.
28 Cortex: Overview
In a similar way, Udo Ernst (“▶ Center- natural consequence of predictive coding in the
Surround Processing, Computational Role of”) hierarchically organized “social brain” (Kilner
addresses the phenomenon of center-surround et al. 2007). Second, it has been applied to
processing (CSP) from a computational point interoception from which predictions about
of view. Even though CSP has been investigated bodily self-consciousness could be derived
mainly in the visual system, where it is exem- (Seth et al. 2011). Third, it has been suggested
plified, e.g., by the phenomenon of end stopping that it may also be applicable to the [motor
already described by Hubel and Wiesen (Hubel system] with the surprising consequence that
and Wiesel 1965), it is also a candidate for the actual motor acts are carried out to fulfill
a canonical cortical computation to be found in predictions about sensory consequences of just
various cortical regions. Ernst links CSP to the these acts (Hawkins and Blakeslee 2005) as
laws formulated by Gestalt psychologists in the compared to being only the output stage of
early twentieth century but also to modern nor- a sensory-to-motor transformation. Future
mative approaches that utilize the statistics of experimental studies will need to further test
natural visual scenes in explaining physiological these predictions. Interestingly, Spratling has
and perceptual phenomena. He points out that shown that predictive coding and the concept
CSP has been successfully implemented in of biased competition can be thought of as
[cortex-inspired artificial vision systems], being just variants of the same mathematical
where it improved object detection and recogni- model. Thus, while the interpretation of models
tion of natural scenes. Such real-world tests of derived from the predictive coding theory in
cortical models are excellent yardsticks mod- terms of “prediction of inputs” may be unusual
elers may want to consider in addition to in some cases, as in the case of the motor sys-
reproducing physiological or perceptual phe- tem, the theory is still a rich framework to sys-
nomena that are usually observed in rather arti- tematically derive mathematical models and
ficial laboratory settings with less naturalistic testable predictions.
stimuli. Michael Spratling (“▶ Predictive Cod- Computation cannot be considered in isola-
ing”) reviews the concept of predictive coding. tion. Communication engineers and designers of
This is an instance of a theory (in the sense processors know this too well. Matthias Bethge
defined above), but models derived from this (“▶ Efficient Population Coding”) considers how
theory can predict CSP as a by-product. The much information is communicated by cortical
distinctive feature of predictive coding is that networks. Bethge introduces the psychometric
downstream areas in the hierarchically orga- and neurometric functions and highlights that
nized cortex continuously predict activity in the information contained in the spiking activity
areas at a lower level of the hierarchy. Given of populations of neurons is often enough to
that downstream areas in sensory cortices inte- predict the behavioral responses of the whole
grate signals from neuronal populations with organisms. This is an empirical finding, but com-
adjacent receptive fields (RFs), the predictions putational neuroscience also needs to ask more
carry not only information about anticipated fundamental questions such as “how much infor-
future inputs but also from the neighboring RFs mation can be transmitted?” Only this allows for
as in CSP. Models derived from predictive cod- assessing how close to optimality the cortical
ing are candidates for a canonical cortical com- circuits are actually operating. To address such
putation. Applications to sensory cortices may questions, he reviews how the concepts of Fisher
be relatively straightforward, but the crucial test and Shannon’s mutual information can be applied
for a theory is its predictions when extrapolated to quantify the information content of population
beyond the postdoc explanations of already activity. Combining the approaches that focus on
known phenomena. Let me point out three computation introduced so far with these studies
selected such extrapolations: First, it has been of communication and information content could
applied to explain mirror neuron activity as the be a very fruitful direction for future studies, in
Cortex: Overview 29
particular when factoring in limitations due to a similar approach and address the question of
fiber bottlenecks between cortical areas and how the CSP in V1, as introduced by Ernst,
energy expenditure. may be realized by cortical circuits. More spe-
RNNs are Turing-complete, which means cifically, I review network models of CSP that
that finite RNNs could, in principle, approxi- are distinct in terms of the assumed pathways.
mate any computation. However, to determine Early models emphasized the role of long-range
the computations actually performed by the connections within an area, but later models
cerebral networks, it is imperative to develop came to acknowledge the role of feedback from
mechanistically plausible models that explicitly downstream areas. The cortical operating
respect the anatomical and physiological con- mode in vivo is characterized by strong recurrent
straints set by experimental neuroscience. Sean excitation and balanced inhibition that
Hill (“▶ Cortical Columns, Models of”) pre- affect how single neurons integrate and propa-
sents models of the so-called cortical column, gate signals (reviewed in my second short
which itself is a theoretical concept motivated contribution “▶ Balanced State”). Interestingly,
by early experimental studies that showed more recent modeling studies investigated the
smooth variation of functional properties tan- role of short-range local connections in CSP
gential to the cortical surface but an invariance and found that the properties of strongly
across cortical layers at a given position. The connected recurrent networks in a balanced
notion of a cortical column remains controver- state need to be considered in models of
sial, but for computational neuroscience it is CSP. Adaptation is another phenomenon that
certainly a goal to deliver predictive mechanis- seems to be omnipresent in the cerebral
tic models of signal propagation across cortical cortex. Klaus Wimmer (“▶ Adaptation in
layers within an area. Probably the most prom- Sensory Cortices, Models of”) reviews models
inent example of mechanistic network modeling of adaptation and considers both their role in
to explain a physiologically observed phenom- perception and how plausible mechanisms such
enon is the models for orientation tuning in as short-term synaptic depression mediate them.
primary visual cortex (V1), which are reviewed Since strongly recurrent networks in a balanced
by Nicholas Priebe and Benjamin Scholl state with static synaptic connections may
(“▶ Emergence of Orientation Selectivity in already exhibit counterintuitive phenomena,
the Cerebral Cortex, Modeling”). Hubel and Wimmer argues for systematic modeling studies
Wiesel discovered orientation tuning (Hubel of structured networks with adaptation mecha-
and Wiesel 1959) and formulated a first nisms as an important approach to understand
model, namely, that the tuning derives from adaptation in sensory cortices.
the pattern of afferent connections from the Selected examples of higher cognitive func-
thalamus onto neurons in V1. The subsequently tions are attention and working memory. Corti-
developed models emphasized the role of cal models of these functions are reviewed by
intracortical connections to account for experi- Philipp Schwedhelm and Stefan Treue
mentally observed properties of orientation (“▶ Attentional Top-Down Modulation, Models
tuning such as contrast invariance. Interestingly, of”) and Gianluigi Mongillo (“▶ Working
the original feedforward model by Hubel and Memory, Models of”). Schwedhelm and Treue
Wiesel is still a guiding idea, even though it had review models of attentional top-down modula-
to be refined. This highlights that such more tion. They highlight how phenomenological
informal and conceptual models remain valu- models have guided experiments and how
able today, even though computational neuro- those fed back into refining the models. Some
science has to show explicitly when and how of the models assume the mechanism of gain
models fail as reviewed by Priebe and Scholl. modulation but remain intentionally agnostic
In my first entrie (“▶ Center-Surround regarding the biophysical mechanisms. This
Processing, Network Models of”), I take exemplifies that cortical modeling with
30 Cortex: Overview
a properly chosen level of description could be mechanistically the ongoing low-activity state
integrated closely with experimental investiga- in recurrently connected local networks (Brunel
tions. Working memory has also been studied et al. 2013) and derived models to explain them
experimentally in great detail, but most early as a stable attractor. Joana Cabral and Gustavo
network models of the persistent activity that Deco (“▶ Spontaneous Activity, Models of”)
is characteristic for the physiological correlate review models of the global spontaneous activity
of working memory were variants of attractor that exhibits characteristic temporal properties
networks, where a self-sustained “bump” of and is found in the so-called default mode
activity was identified with the content of work- network. This activity (and the default mode
ing memory. Only more recent modeling studies network) has been studied intensively using
suggested that self-sustained activity may not be functional magnetic resonance imaging (fMRI),
restricted to the spiking activity of groups of but Cabral and Deco correctly point out that
neurons, because the state of synapses with a deeper analysis of the network models is still
short-term dynamics can also be considered as needed to provide insights into the dynamical
an activity variable that could be exploited to properties of the resting state.
store self-sustained activity. The idea that syn- The entries in this section cover computation
aptic variables, which are by multiple orders of and modeling of the cortex using different
magnitude more numerous than single cell state approaches and models at various levels of
variables, may be crucial for cerebral informa- abstraction. Certainly, the cortex cannot be con-
tion processing has been around in the compu- sidered in isolation but needs to be modeled and
tational neuroscience community for a long understood in concert with other structures, such
time. However, explicit formal models need to as the [thalamus] and [basal ganglia]. Will it be
spell this out and show the potential benefit in, possible to understand cortex without modeling
for example, systematic simulation studies even the whole brain or even closed sensory-motor
if the models are speculative and experimentally loops within an “enactive” approach (Noe 2006)
very hard to test as in the case of the synaptic that states that to understand the brain – in our
theory of working memory. This also applies case, only the cortex – one needs to look at more
for models of attention: Given that after than just the brain? This is indeed an open ques-
50 years of the discovery of orientation tuning tion that is of special interest for philosophers of
in V1, there is still no agreement on network science and mind. However, I argue that the cor-
models of even such a basal response property, tex considered as a complex and self-assembled
it may not come as a surprise that the mecha- adaptive structure will remain a challenge for any
nisms of attention remain elusive. While, for kind of modeling conducted by Computational
example, mechanistic models of top-down gain Neuroscientists who are open to empirical find-
modulation via synchronizing the discharges of ings and brave enough to ignore irrelevant details
inhibitory interneurons may be consistent with without throwing out the baby with the bath
the available anatomical, physiological, and water. The reward shall be motivating: to gain
biophysical knowledge, recording multiple an insight into how the cortex works. Relevance
identified inhibitory interneurons in vivo in and irrelevance of details needs to be decided on
attentional demanding tasks remains to be a case-by-case basis, which also depends on the
achieved. taste of the modeler (or theoretician). Unfortu-
Another currently only poorly understood nately, despite a multitude of models and some
phenomenon is how the so-called resting promising candidates for theories of cortical
state of the brain is generated and maintained. function, one needs to attest that we are not yet
Computational Neuroscience research has there: A unified theory of cortical computation
already identified the problem of explaining with associated models still needs to be derived
Databases and Data Repositories in Computational Neuroscience: Overview 31
and thoroughly tested. My own requirement for

accepting such a theory is that it will cover at least Databases and Data Repositories
the topics addressed by the entries in this section. in Computational Neuroscience:
Overview
Cross-References Richard C. Gerkin1, Shreejoy J. Tripathy2,

Sharon Crook3 and Jeanette Hellgren Kotaleski4
▶ Adaptation in Sensory Cortices, Models of 1
School of Life Sciences, Arizona State
▶ Attentional Top-Down Modulation, Models of University, Tempe, AZ, USA
▶ Balanced State 2
Centre for High-Throughput Biology and
▶ Bayesian Inference with Spiking Neurons Department of Psychiatry, University of British
▶ Center-Surround Processing, Computational Columbia, BC, Canada
3
Role of School of Mathematical and Statistical Sciences
▶ Center-Surround Processing, Network Models of and School of Life Sciences, Arizona State
▶ Cortical Columns, Models of University, Tempe, AZ, USA
▶ Cortical Function, Normative Models of 4
School of Computer Science and
▶ Efficient Population Coding Communication, KTH Royal Institute of
▶ Emergence of Orientation Selectivity in the Technology, Stockholm, Sweden
Cerebral Cortex, Modeling
▶ Predictive Coding
▶ Recurrent Network Models, Reservoir Definition
Computing
▶ Reinforcement Learning in Cortical Networks Computational neuroscience research often pro-
▶ Spontaneous Activity, Models of duces models that help explain empirical data and
▶ Working Memory, Models of provide predictions about the biological systems
that produce the data. Empirical and theoretical
researchers can benefit from resources that facil-
References itate model publication and exchange and pro-
vide neuroscience data that informs and
Brunel N, del Giudice P, Fusi S, Parisi G, Tsodyks M (eds) contains these existing and future models.
(2013) Selected papers of Daniel Amit. World Scien-
tific, Hackensack
Hawkins J, Blakeslee S (2004) On intelligence. Times
Books, New York Detailed Description
Hubel DH, Wiesel TN (1959) Receptive fields of single
neurones in the cat’s striate cortex. J Physiol 148:574–591 Databases of Computational Models
Hubel DH, Wiesel TN (1965) Receptive fields and func-
tional architecture in two nonstriate visual areas
One category of databases in computational neu-
(18 and 19) of the cat. J Neurophysiol 28(2):230–289 roscience is those that focus on computational
Kilner JM, Friston KJ, Frith CD (2007) The mirror-neuron models. The BioModels database (Li et al. 2010;
system: a Bayesian perspective. Neuroreport Vijayalakshmi et al. 2013) is a general repository
18(6):619–623
of computational models of biological processes
Mountcastle VB (1978) An organizing principle for cere-
bral function: the unit model and the distributed sys- that includes some models from neuroscience. The
tem. In: Edelman GM, Mountcastle VB (eds) The SenseLab database, ModelDB, (Migliore et al.
mindful brain. MIT Press, Cambridge, MA 2003) provides a resource for published neurosci-
Noe A (2006) Action in perception. Bradford Book
ence models in a variety of formats. Many of these
Seth AK, Suzuki K, Critchley HD (2011) An interoceptive
predictive coding model of conscious presence. Front were developed specifically for simulating the
Psychol 2:395 electrophysiological and neurochemical
32 Databases and Data Repositories in Computational Neuroscience: Overview
properties of single neurons and networks of neu- provide some widely used examples. BrainInfo
rons, for example, with multicompartment Hodg- provides general information about brain areas,
kin-Huxley type conductance-based models including what they do, where they are located,
(Hodgkin and Huxley 1952). The Physiome and what they contain. The Allen Institute for
Model Repository (Yu et al. 2011) is a software Brain Science provides brain-wide gene expres-
suite that facilitates storage and management of sion atlases, where the expression of each of the
models, focusing on those described in CellML, genes in the mammalian genome has been sys-
a standard markup language for biological models. tematically quantified throughout the brain for
Similarly, OpenSourceBrain (Gleeson et al. a number of animal species and across stages of
2010b) is a community platform for collaborative neural development (Lein et al. 2007, http://
development of computational neuron and net- brain-map.org). The Allen Institute also provides
work models that utilizes open standards such as information on the anatomical connectivity of
NeuroML (Gleeson et al. 2010a) to facilitate inter- different brain regions.
operability and visualization of neuroscience Parallel to this effort is The Human
models developed by different researchers. Connectome Project (Marcus et al. 2013),
a large-scale effort to map complete structural
Databases of Neuron Structure and functional neural connections in vivo in indi-
and Properties vidual humans. BrainMap (Laird et al. 2004)
A second category of databases that are important consists of a database and related software to
for the computational neuroscience community search published functional and structural
includes those that contain structured information human neuroimaging experiments. In contrast,
specific to neurons and their properties. For CoCoMac (Stephan et al. 2001; Bakker et al.
example, information on the detailed shapes of 2012) is focused on the primate brain, containing
neurons (morphology) is compiled by records of tracing studies in the macaque. The
NeuroMorpho (Ascoli et al. 2007) which contains Brain Architecture Management System
user-submitted neuron morphological recon- (BAMS) contains neural connectivity informa-
structions made, for example, using the tion across species that has been manually
NeuroLucida application (Glaser and Glaser curated from the existing research literature
1990). Measured electrophysiological properties (Bota et al. 2005). Finally, the Cell Centered
of neuron types, and the metadata associated with Database (Martone et al. 2003, 2009) focuses on
these measurements, are cataloged in the images from light and electron microscopy, rang-
NeuroElectro Project (Tripathy et al. 2014). Sim- ing from whole brain areas to subcellular
ilarly, other SenseLab databases, including compartments.
NeuronDB and CellPropDB (Crasto et al. 2007),
contain information on the ionic currents and Other Resources
neurotransmitters expressed by each neuron and In addition to these neuroscience domain-specific
how these are distributed with respect to neuronal databases are federated databases that provide
morphology. Detailed information on ion chan- linking facilities for cross-resource data integra-
nel subtypes, including voltage and temporal tion. For example, NeuroLex (Larson and
dynamics, genetic homology, and corresponding Martone 2013) provides a platform for commu-
literature references, is available at Channelpedia nity annotation of neuron types on the basis of
(Ranjan et al. 2011, http://channelpedia.net), morphological, neurochemical, or electrophysio-
a subproject within the Blue Brain Project logical properties. Given this wealth of neurosci-
(Markram 2006). ence resources, the Neuroscience Information
Framework (NIF) provides tools for semantic
Resources for Brain Connectivity search across these diverse databases (Gardner
Another category of databases focuses on the et al. 2008) through the development and incor-
anatomical organization of the brain. Here we poration of neuroscience domain-specific
Databases and Data Repositories in Computational Neuroscience: Overview 33
ontologies (Bug et al. 2008; Larson and Martone Bug WJ, Ascoli GA, Grethe JS, Gupta A, Fennema-
2009; Hamilton et al. 2012; Imam et al. 2012). Notestine C, Laird AR, Larson SD, Rubin D, Shepherd
GM, Turner JA, Martone ME (2008) The NIFSTD and
For example, in NIF, the search query “mitral BIRNLex vocabularies: building comprehensive ontol-
cell” returns a number of database records includ- ogies for neuroscience. Neuroinformatics 6(3):175–194
ing relevant research literature from PubMed, Crasto CJ, Marenco LN, Liu N, Morse TM, Cheung K-H, Lai
computational models from ModelDB, and con- PC, Bahl G, Masiar P, Lam HYK, Lim E, Chen H,
Nadkarni P, Migliore M, Miller PL, Shepherd GM
nectivity information from BAMS. (2007) SenseLab: new developments in disseminating
The number and size of these databases con- neuroscience information. Brief Bioinform 8(3):150–162
tinue to grow with the collection and contribution Gardner D et al (2008) The neuroscience information
of ever more models and data. These databases framework: a data and knowledge environment for
neuroscience. Neuroinformatics 6(3):149–165
give computational neuroscientists a powerful Glaser JR, Glaser EM (1990) Neuron imaging with
tool for constraining data-driven models and neurolucida – a PC-based system for image combining
serve as an alternative to conventional literature microscopy. Comput Med Imaging Graph
searches. 14(5):307–317
Gleeson P, Crook S, Cannon RC, Hines ML, Billings GO,
Farinella M, Morse TM, Davison AP, Ray S, Bhalla
US, Barnes SR, Dimitrova YD, Silver RA (2010a)
Cross-References NeuroML: a language for describing data driven
models of neurons and networks with a high degree
of biological detail. PLoS Comput Biol 6(6):e1000815
▶ BioModels Database: a Public Repository for Gleeson P, Piasini E, Crook S, Cannon R, Steuber V,
Sharing Models of Biological Processes Jaeger D, Solinas S, DAngelo E, Silver RA (2010b)
▶ Brain Architecture Management System The open source brain initiative: enabling collabora-
(BAMS) tive modelling in computational neuroscience. BMC
Neurosci 13(1):1–2
▶ BrainInfo Hamilton DJ, Shepherd GM, Martone ME, Ascoli GA
▶ BrainMap (2012) An ontological approach to describing neurons
▶ Cell Centered Database and their relationships. Front Neuroinform 6:15
▶ Collations of Connectivity Data on the Hodgkin AL, Huxley AF (1952) A quantitative description
of membrane current and its application to conduction
Macaque Brain (CoCoMac) and excitation in nerve. J Physiol 117(4):500–544
▶ Human Connectome Project Imam FT, Larson SD, Bandrowski A, Grethe JS, Gupta A,
▶ ModelDB Martone ME (2012) Development and use of ontol-
▶ NeuroElectro Project ogies inside the neuroscience information framework:
a practical approach. Front Genet 3:111
▶ NeuroMorpho.org Laird AR, Lancaster JL, Fox PT (2004) Brainmap: the
▶ Neuroscience Information Framework (NIF) social evolution of a functional neuroimaging data-
▶ Open Source Brain base. Neuroinformatics 3:65–78
▶ Physiome Repository Larson SD, Martone ME (2009) Ontologies for neurosci-
ence: what are they and what are they good for? Front
▶ PhysioNet Neurosci 3:1
▶ SenseLab: Integration of Multidisciplinary Larson SD, Martone ME (2013) NeuroLex.org: an online
Neuroscience Data framework for neuroscience knowledge. Front
Neuroinform 7:18
Lein ES et al (2007) Genome-wide atlas of gene expression
in the adult mouse brain. Nature 445(7124):168–176
References Li C, Donizelli M, Rodriguez N, Dharuri H, Endler L,
Chelliah V, Li L, He E, Henry A, Stefan MI, Snoep JL,
Ascoli GA, Donohue DE, Halavi M (2007) Neuromorpho. Hucka M, Le Novere N, Laibe C (2010) BioModels
org: a central resource for neuronal morphologies. database: an enhanced, curated and annotated resource
J Neurosci 27(35):9247–9251 for published quantitative kinetic models. BMC Syst
Bakker R, Wachtler T, Diesmann M (2012) Cocomac 2.0 Biol 4:92
and the future of tract-tracing databases. Front Marcus DS, Harms MP, Snyder AZ, Jenkinson M, Wilson
Neuroinform 6:30 JA, Glasser MF, Barch DM, Archie KA, Burgess GC,
Bota M, Dong H-W, Swanson L (2005) Brain architecture Ramaratnam M, Hodge M, Horton W, Herrick R,
management system. Neuroinformatics 3(1):15–48 Olsen T, McKay M, House M, Hileman M, Reid E,
34 Decision-Making: Overview
Harwell J, Coalson T, Schindler T, Elam JS, Curtiss The field itself sprang from work in the 1950s
SW, WU-Minn HCP Consortium, Van Essen DC aimed at uncovering the biophysical mechanisms
(2013) Human connectome project informatics: qual-
ity control, database services, and data visualization. underlying spike generation (Hodgkin and Hux-
Neuroimage 80:202–219 ley 1952), and other classic studies focus on the
Markram H (2006) The blue brain project. Nat Rev computational capabilities of dendritic trees and
Neurosci 7(2):153–160 on how neural activity encodes sensory stimuli or
Martone ME, Zhang S, Gupta A, Qian X, He H, Price DL,
Wong M, Santini S, Ellisman MH (2003) The cell- statistical information about the world, to men-
centered database: a database for multiscale structural tion a couple of well-known examples (Dayan
and protein localization data from light and electron and Abbott 2001). A different point of view, how-
microscopy. Neuroinformatics 1(4):375–379 ever, is one in which neurons are the intermedi-
Martone ME, Tran J, Wong WW, Sargis J, Fong L, Larson
SD, Lamont SP, Gupta A, Ellisman MH (2009) Cell aries between a subject and its environment. As
centered database project: an update on building com- the engines of behavior, neurons need to be com-
munity resources for managing and sharing 3D imag- putationally powerful for the express purpose of
ing data. J Struct Biol 161(3):220–231 giving the subject an advantage, and hence
Migliore M, Morse TM, Davison AP, Marenco L, Shep-
herd GM, Hines M (2003) ModelDB. their efficiency or performance should be mea-
Neuroinformatics 1(1):135–139 sured with respect to the subject’s success. So
Ranjan R, Khazen G, Gambazzi L, Ramaswamy S, Hill the computational neuroscience of decision mak-
SL, Schrmann F, Markram H (2011) Channelpedia: an ing is computational neuroscience in this context;
integrative and interactive database for ion channels.
Front Neuroinform 5:36 it is the quest to understand how single-neuron
Stephan KE, Kamper L, Bozkurt A, Burns GA, Young activity contributes to nonreflexive actions,
MP, Ktter R (2001) Advanced database methodology actions that cannot be predicted from a subject’s
for the collation of connectivity data on the macaque history alone.
brain (CoCoMac). Philos Trans R Soc Lond B Biol Sci
356(1412):1159–1186 The idea of quantitatively explaining a person’s
Tripathy SJ, Savitskaya J, Burton SD, Urban NN, Gerkin subjective experiences based on the responses of
RC (2014) NeuroElectro: a window to the world’s specific groups of neurons goes back to the 1960s,
Neuron electrophysiology data. Frontiers in with the work of Vernon Mountcastle and col-
Neuroinformatics 8:40
Vijayalakshmi C, Laibe C, Le Novere N (2013) leagues in the somatosensory modality. They
Biomodels database: a repository of mathematical found, for instance, that when a small probe, sim-
models of biological processes. Methods Mol Biol ilar to the tip of a ballpoint pen, vibrates at
1021:189–199 a frequency of 2–40 Hz, the intensity of the evoked
Yu T, Lloyd CM, Nickerson DP, Cooling MT, Miller AK,
Garny A, Terk-ildsen JR, Lawson J, Britten RD, sensation reported by a person is directly related to
Hunter PJ, Nielsen PM (2011) The physiome model the neural responses of a particular type of
repository 2. Bioinformatics 27(5):743–744 mechanosensory receptor under the skin, now
known as a Meissner corpuscle (Talbot et al.
1968). Many design elements and analytical tech-
niques used in contemporary experiments were
Decision-Making: Overview already laid out in those pioneering studies: they
applied information theory to determine the coding
Emilio Salinas capacity of sensory neurons (Werner and
Department of Neurobiology and Anatomy, Mountcastle 1965) very much as is done today,
Wake Forest School of Medicine, except for the rudimentary computers, and they
Winston-Salem, NC, USA recognized neuronal variability, as well as rate
versus temporal coding, as fundamental issues for
neural computation (Werner and Mountcastle
Detailed Description 1963). This early work was limited, however,
because it compared psychophysical performance
Much of computational neuroscience begins and in one system (behaving humans) with neural
ends with the responses of individual neurons. activity in a different system (e.g., anesthetized
Decision-Making: Overview 35
monkeys). It was later, with studies like those of name a few. Selection of a particular behavioral
William Newsome and colleagues (Salzman task thus determines both the specific cognitive
et al. 1992), that vision became the more function and the corresponding neural circuits
popular modality and that both behavioral and under investigation. In this context, it has
neuronal responses were simultaneously recorded become clear that individual choices are often
in the same subjects during the performance influenced by a variety of subtle factors that
of perceptually based tasks. That became the superficially may appear inconsequential
standard and a cornerstone for investigating the (“▶ Decision-Making, Bias”). Such sources
neural basis of decision making (Parker and of bias are likely to attract increasingly more
Newsome 1998). attention in the future, as our ability to
Thus, decision making involves the study of correlate neuronal activity and behavior becomes
neurons and neural circuits as a subject captures more sophisticated.
information about the sensory world; analyzes it; Most decision-making tasks have at least two
combines it with other stored information about components: a perceptual step during which cur-
current goals, priorities, and possible courses of rent sensory information is analyzed (e.g., is that
action; and makes a response. In the laboratory, spot red or green?) and a motor report whereby
researchers attempt to simplify this as much as the result of the perceptual judgment is indicated
possible while still maintaining the essence of the (e.g., push a left or a right button). The rest of the
process – an internal evaluation that is not articles in the section focus either on perception
fully predictable – and while measuring three quan- (“▶ Accumulation of Evidence in Decision-
tities as accurately as possible: the sensory stimuli, Making”), motor planning (“▶ Decision-Making,
the relevant neuronal activity, and the subject’s Motor Planning”), or their interface and interac-
behavior (Parker and Newsome 1998). So even tion. Although some neurons clearly relate to
the simplest possible decision-making task either perceptual or motor processing, it is often
requires various types of neural computations, the case that the cells that correlate most strongly
and indeed, one common strategy in the field with a subject’s choices have elements of both
has been to try to break down the problem into (“▶ Perceptual Decision-Making”). In fact,
temporally discrete steps (e.g., fixation, stimulus determining the degree to which a neural
presentation, response selection, reward delivery), response encodes perceptual information versus
to investigate how neurons in various areas par- a motor plan turns out to be surprisingly
ticipate in specific aspects of the decision-making difficult (“▶ Target Selection vs. Response
process (Romo and Salinas 2003). The articles Selection”); the ambiguity arises even at the psy-
that comprise this section of the Springer chophysical level (“▶ Perceptual-Motor Dissoci-
Encyclopedia of Computational Neuroscience ation”). In spite of this, a number of principles
review key components of any such decision- describing how neurons participate in the gener-
making process. ation of perceptual judgments and ultimately pro-
Three of the entries (“▶ Decision-Making duce choices have been identified (“▶ Perceptual
Tasks”; “▶ Choice Behavior”; “▶ Categorical Decision-Making”). Furthermore, relatively
Decisions”) emphasize the repertoire of tasks simple quantitative models have been highly
and associated mathematical models that have successful at reproducing not only the
been successfully used to characterize and quan- traditional behavioral metrics of performance
tify behavior. This provides a foundation for and reaction time but also key aspects of
understanding how primary factors, such as neuronal activity (“▶ Accumulation of Evidence
reward availability or the quality of sensory infor- in Decision-Making”; “▶ Decision-Making,
mation, drive a subject’s actions and how differ- Models”). Because of this, there is a relatively
ent tasks place demands on different cognitive thorough understanding of at least some of the
functions, such as perceptual discrimination, con- computations that are key to perceptual
flict resolution, or working memory capacity, to decision making (“▶ Accumulation of
36 Deep Brain Stimulation (Models, Theory, Techniques): Overview
Evidence in Decision-Making”; “▶ Speed-

Accuracy Tradeoff”; “▶ Decision-Making, Deep Brain Stimulation (Models,
Threshold”). This trend is expected to continue. Theory, Techniques): Overview
Peter Alexander Tass1,2,3 and Christian Hauptmann1

1
Institute of Neuroscience and Medicine –
Cross-References Neuromodulation (INM–7), Research Center
J€ulich, J€ulich, Germany
▶ Accumulation of Evidence in Decision-Making 2
Department of Neurosurgery, Stanford
▶ Categorical Decisions University, Stanford, CA, USA
▶ Choice Behavior 3
Department of Neuromodulation, University of
▶ Decision-Making Tasks Cologne, Cologne, Germany
▶ Decision-Making, Bias
▶ Decision-Making, Models
▶ Decision-Making, Motor Planning Detailed Description
▶ Decision-Making, Threshold
▶ Perceptual Decision-Making Deep brain stimulation (DBS) of the subthalamic
▶ Perceptual-Motor Dissociation nucleus (STN) is a well-established treatment for
▶ Speed-Accuracy Tradeoff medically refractory patients with advanced
▶ Target Selection vs. Response Selection Parkinson’s disease (PD) (Benabid et al. 1991;
Blond et al. 1992; Benabid et al. 2002; Deuschl
et al. 2006) as well as for patients with early
References motor complications (Deuschl et al. 2006;
Schuepbach et al. 2013). Several neurological
Dayan P, Abbott LF (2001) Theoretical neuroscience.
MIT Press, Cambridge, MA
diseases, such as Parkinson’s disease (PD) or
Hodgkin AL, Huxley AF (1952) A quantitative essential tremor, are characterized by pathologi-
description of membrane current and its application cal synchronization (Nini et al. 1995; Brown
to conduction and excitation in nerve. J Physiol et al. 2001). Parkinsonian resting tremor, for
117:500–544
example, seems to origin from a pacemaker-like
Parker AJ, Newsome WT (1998) Sense and the single
neuron: probing the physiology of perception. Annu population of neurons of the basal ganglia firing
Rev Neurosci 21:227–277 in a synchronized and oscillatory manner
Romo R, Salinas E (2003) Flutter discrimination: neural (Hutchison et al. 1997; Hurtado et al. 1999;
codes, perception, memory and decision making. Nat
Magill et al 2001; Trottenberg et al. 2007). In
Rev Neurosci 4:203–218
Salzman CD, Murasugi CM, Britten KH, Newsome WT contrast, under healthy conditions these neurons
(1992) Microstimulation in visual area MT: effects on are active in an uncorrelated and desynchronized
direction discrimination performance. J Neurosci manner (Nini et al. 1995; Magill et al. 2001).
12:2331–2355
The standard DBS protocol employs perma-
Talbot WH, Darian-Smith I, Kornhuber HH, Mountcastle
VB (1968) The sense of flutter-vibration: comparison nent high-frequency (>100 Hz) pulse train stim-
of the human capacity with response patterns of mech- ulation (Benabid et al. 1991; Blond et al. 1992;
anoreceptive afferents from the monkey hand. Benabid et al. 2002). Symptom suppression by
DBS is strongly dependent on stimulation
Werner G, Mountcastle V (1963) The variability of central
neural activity in a sensory system, and its implications frequency – with only high frequencies
for the central reflection of sensory events. (>100 Hz) being effective and effects being rap-
J Neurophysiol 26:958–977 idly reversible (Birdno and Grill 2008). High-
Werner G, Mountcastle VB (1965) Neural activity in mech-
anoreceptive cutaneous afferents: stimulus–response
frequency DBS was developed empirically,
relations, Weber functions, and information transmis- mainly based on clinical observations and exper-
sion. J Neurophysiol 28:359–397 imental results (Volkmann et al. 2006), and the
Deep Brain Stimulation (Models, Theory, Techniques): Overview 37
mechanism of high-frequency DBS is still pulses on an approximately one for one basis
a matter of debate (Benabid et al. 2005). (McIntyre et al. 2004). The various reactions of
Experimental observations indicate that dur- neurons toward stimulation on the network level
ing high-frequency DBS, a regular bursting mode further add complexity that is important for the
is induced (Beurrier et al. 2002), and after creation of a sound model: cells responding differ-
a reduction of stimulation artifacts, robust burst- ently to external inputs, such as somatosensory
ing activity in STN neurons was observed in slice stimulation or stimulation owing to active move-
experiments (Beurrier et al. 2001). In the same ments, are present in the target tissue together with
experiments, the offset of stimulation was so-called no-response cells (Lenz et al. 1994).
followed by a blockade of activity, i.e., Therefore, high-frequency stimulation has
a depolarization blockade (Beurrier et al. 2001). a complex impact on these structures (Benabid
These observations were made in anesthetized et al. 2002; Shen et al. 2003). However, surpris-
animals and are contradicted by measurements ingly, even single STN model neurons – lacking
in awake behaving primates (Anderson synaptic dynamics, neural circuitry, and contribu-
et al. 2003; Hashimoto et al. 2003; Dorval tions of glial cells – subjected to high-frequency
et al. 2008) and rats (McConnell et al. 2012). stimulation reproduce clinically observed
Other groups argue that high-frequency DBS response characteristics (Pyragas et al. 2013).
blocks neuronal activity in relevant target areas To study another aspect of DBS, several groups
during stimulation and therefore mimics the use physical models based on Maxwell’s equa-
effect of tissue lesioning (Benabid et al. 2002). tions to investigate the neuronal activation profile
In 2005, Benabid and coworkers summarized dif- depending on electrode geometry and stimulation
ferent hypothetical mechanisms: membrane inhibi- parameters (Butson and McIntyre 2005, 2006;
tion, excitation of excitatory and inhibitory Miocinovic et al. 2009; Yousif et al. 2008;
afferents, jamming, excitation of efferents, and Chaturvedi et al. 2010; Buhlmann et al. 2011).
plasticity (Benabid et al. 2005). Novel experimental While the standard DBS protocol was devel-
techniques, such as optogenetics, enabled to further oped empirically (Volkmann et al. 2006), novel
reveal the mechanism of DBS and, in particular, of stimulation approaches are based on electrophysi-
the stimulation of afferent axons projecting to the ological as well as computational concepts. Per-
target region (Gradinaru et al. 2009). sonalizing and optimizing high-frequency
Spatially extended single- and multi- stimulation in real time by demand-controlled,
compartment neuron models were used to adaptive DBS might constitute a superior high-
evaluate the contribution of these different mech- frequency stimulation mode, as shown in an
anisms (Grill & McIntyre 2001; Terman acute study in externalized patients (Little
et al. 2002; Rubin and Terman 2004). For exam- et al. 2013). In addition, modeling studies are
ple, Grill and McIntyre (2001) showed that used to further develop the stimulation algorithm,
depending on the stimulation amplitude and the beyond standard high-frequency DBS, in order to
shape of the stimulation pulses, cells were either finally establish superior stimulation mechanisms.
activated directly or fibers mediating excitatory For example, coordinated reset (CR), a patterned
or strong inhibitory action were activated. The stimulation protocol specifically targeting the
activation of a larger number of structures takes reduction of synchronized activity, was developed
place on the single-neuron level with different by means of mathematical models (Tass 2003) and
and possibly conflicting impacts on single- essentially aims at an unlearning of both abnormal
neuron dynamics (Grill and McIntyre 2001). For synaptic connectivity and synchrony (Tass and
example, in the same neuron, the cell body Majtanik 2006). CR was successfully tested in
(soma) is inhibited as a result of activation of a preclinical study, where 5 days of low-dose CR
presynaptic axons and GABA release, while the stimulation induced long-lasting therapeutic
efferent axon is activated by the stimulation effects for 30 days (Tass et al. 2012). Another
38 Deep Brain Stimulation (Models, Theory, Techniques): Overview
example is a closed-loop approach, which was Beurrier C, Garcia L, Bioulac B, Hammond C (2002)
controlled by the extent of oscillatory beta-band Subthalamic nucleus: a clock inside basal ganglia?
Thalamus Relat Syst 2:1–8
activity. During stimulus delivery, this approach Birdno MJ, Grill WM (2008) Mechanisms of deep brain
resulted in a better reduction of akinesia as well as stimulation in movement disorders as revealed by
pallidal firing rates as compared to classical DBS changes in stimulus frequency. Neurotherapeutics
in parkinsonian nonhuman MPTP-treated pri- 5(1):14–25
Blond S, Caparros-Lefebvre D, Parker F et al (1992) Con-
mates (Rosin et al. 2011). Finally, modifications trol of tremor and involuntary movement disorders by
of the standard HF protocol might offer a novel chronic stereotactic stimulation of the ventral interme-
approach to improve the efficacy of deep brain diate thalamic nucleus. J Neurosurg 77:62–68
stimulation (Brocker et al. 2013; Hess et al. 2013). Brocker DT, Swan BD, Turner DA, Gross RE, Tatter SB,
Miller KM, Bronte-Stewart H, Grill WM
The combination of all these modeling and (2013) Improved efficacy of temporally non-regular
experimental and technological approaches deep brain stimulation in Parkinson’s disease. Exp
plays a vital role in shaping our understanding Neurol 239:60–67
and helps to improve the promising therapeutic Brown P, Oliviero A, Mazzone P, Insola A, Tonali P, Di
Lazzaro V (2001) Dopamine dependency of oscilla-
intervention DBS. tions between subthalamic nucleus and pallidum in
Parkinson’s disease. J Neurosci 21:1033–1038
Buhlmann J, Hofmann L, Tass PA, Hauptmann C (2011)
Cross-References Modeling of a segmented electrode for desynchronizing
deep brain stimulation. Front Neuroeng 4:1–8
Butson CR, McIntyre CC (2005) Tissue and electrode capac-
▶ Computational Model-Based Development of itance reduce neural activation volumes during deep
Novel Stimulation Algorithms brain stimulation. Clin Neurophysiol 116:2490–2500
▶ Computational Models of Deep Brain Butson CR, McIntyre CC (2006) Role of electrode design
on the volume of tissue activated during deep brain
Stimulation (DBS) stimulation. J Neural Eng 3:1–8
▶ Computational Models Supporting Parameter Chaturvedi A, Butson CR, Lempka SF, Cooper SE,
Finding for Deep Brain Stimulation McIntyre CC (2010) Patient-specific models of deep
▶ Computational Models to Optimize the brain stimulation: influence of field model complexity
on neural activation predictions. Brain Stimul 3:65–67
Electrodes and Waveforms for Deep Brain Deuschl G, Schade-Brittinger C, Krack P et al (2006)
Stimulation A randomized trial of deep-brain stimulation for
Parkinson’s disease. New Engl J Med 355:896–908
Dorval AD, Russo GS, Hashimoto T, Xu W, Grill WM,
References Vitek JL (2008) Deep brain stimulation reduces neu-
ronal entropy in the MPTP-primate model of
Anderson ME, Postupna N, Ruffo M (2003) Effects of Parkinson’s disease. J Neurophysiol 100:2807–2818
high-frequency stimulation in the internal globus Gradinaru V, Mogri M, Thompson KR, Henderson JM,
pallidus on the activity of thalamic neurons in the Deisseroth K (2009) Optical deconstruction of parkin-
awake monkey. J Neurophysiol 89:1150–1160 sonian neural circuitry. Science 324:354–359
Benabid AL, Pollak P, Gervason C, Hoffmann D, Gao Grill WM, McIntyre CC (2001) Extracellular excitation of
DM, Hommel M, Perret JE, de Rougemount J (1991) central neurons: implications for the mechanisms of
Long-term suppression of tremor by chronic stimula- deep brain stimulation. Thalamus Relat Syst
tion of ventral intermediate thalamic nucleus. Lancet 1:269–277
337:403–406 Hashimoto T, Elder CM, Okun MS, Patrick SK, Vitek JL
Benabid AL, Benazzous A, Pollak P (2002) Mechanisms (2003) Stimulation of the subthalamic nucleus changes
of deep brain stimulation. Mov Disord 17:73–74 the firing pattern of pallidal neurons. J Neurosci
Benabid AL, Wallace B, Mitrofanis J et al (2005) 23(5):1916–1923
A putative generalized model of the effects and mech- Hess CW, Vaillancourt DE, Okun MS (2013) The tempo-
anism of action of high frequency electrical stimula- ral pattern of stimulation may be important to the
tion of the central nervous system. Acta Neurol Belg mechanism of deep brain stimulation. Exp Neurol
105:149–157 247:296–302
Beurrier C, Bioulac B, Audin J, Hammond C (2001) High- Hurtado JM, Gray CM, Tamas TB et al (1999) Dynamic of
frequency stimulation produces a transient blockade of tremor-related oscillations in the human globus
voltage-gated currents in subthalamic neurons. pallidus: a single case study. PNAS USA
J Neurophysiol 85(4):1351–1356 96:1674–1679
Dynamical Systems: Overview 39
Hutchison WD, Lozano AM, Tasker RR, Lang AE, Terman D, Rubin JE, Yew AC, Wilson CJ (2002) Activity
Dostrovsky JO (1997) Identification and characteriza- patterns in a model for the subthalamopallidal network
tion of neurons with tremor-frequency activity in of the basal ganglia. J Neurosci 22:2963–2976
human globus pallidus. Exp Brain Res 113:557–563 Trottenberg T, Kupsch A, Schneider GH, Brown P, Kuehn
Lenz FA, Kwan HC, Martin RL et al (1994) Single unit AA (2007) Frequency-dependent distribution of local
analysis of the human ventral thalamic nuclear group field potential activity within the subthalamic nucleus
Tremor-related activity in functionally identified cells. in Parkinson’s disease. Exp Neurol 205:287–291
Brain 117:531–543 Volkmann J, Moro E, Pahwa R (2006) Basic algorithms
Little S, Pogosyan A, Neal S et al (2013) Adaptive deep for the programming of deep brain stimulation in
brain stimulation in advanced Parkinson disease. Ann Parkinson’s disease. Mov Disord 21(Suppl 14):
Neurol 74:449–457 S284–S289
Magill PJ, Bolam JP, Bevan MD (2001) Dopamine regu- Yousif N, Bayford R, Liu X (2008) The influence of
lates the impact of the cerebral cortex on the reactivity of the electrode-brain interface on the cross-
subthalamic nucleus-globus pallidus network. Neuro- ing electric current in therapeutic deep brain stimula-
science 106(2):313–320 tion. Neuroscience 156:597–606
McConnell GC, So RQ, Hilliard JD, Lopomo P, Grill WM
(2012) Effective deep brain stimulation suppresses
low-frequency network oscillations in the basal
ganglia by regularizing neural firing patterns.
J Neurosci 32(45):15657–15668 Dynamical Systems: Overview
McIntyre CC, Grill WM, Sherman DL, Thakor NV
(2004) Cellular effects of deep brain stimulation:
model-based analysis of activation and inhibition.
Alla Borisyuk
J Neurophys 91:1457–1469 Department of Mathematics, University of Utah,
Miocinovic S, Lempka SF, Russo GS et al (2009) Exper- Salt Lake City, UT, USA
imental and theoretical characterization of the voltage
distribution generated by deep brain stimulation. Exp
Neurol 216:166–176
Nini A, Feingold A, Slovin H, Bergmann H (1995) Neu- Definition
rons in the globus pallidus do not show correlated
activity in the normal monkey, but phase-locked oscil- Many models of computational neuroscience are
lations appear in the MPTP model of Parkinsonism.
formulated in terms of nonlinear dynamical system
Pyragas K, Novicenko V, Tass PA (2013) Mechanism of (sometimes the dynamical system with noise) and
suppression of spontaneous low-frequency oscillations include a large number of parameters. Therefore, it
in high-frequency stimulated neurons. Biol Cybern is difficult to analyze dynamics and find correspon-
107:669–684
dence between parameter values and dynamical
Rosin B, Slovik M, Mitelman R et al (2011) Closed-loop
deep brain stimulation is superior in ameliorating par- mode. Mathematical theory of dynamical systems
kinsonism. Neuron 72:370–384 and bifurcations provide a valuable tool for a qual-
Rubin D, Terman J (2004) High frequency stimulation of itative study and finding regions in parameter space
the subthalamic nucleus eliminates pathological tha-
corresponding to different dynamical modes (e.g.,
lamic rhythmicity in a computational model. J Comput
Neurosci 16:211–235 oscillations or bistability). Thus, the dynamical sys-
Schuepbach WMM, Rau J, Knudsen K et al (2013) tems (or nonlinear dynamics) approach to analysis
Neurostimulation for Parkinson’s disease with early of neural systems has played a central role for
motor complications. New Engl J Med 368:610–622
computational neuroscience for many years
Shen K, Zhu Z, Munhall A, Johnson SW (2003) Synaptic
plasticity in rat subthalamic nucleus induced by high- (summarized, e.g., in recent textbooks, Izhikevich
frequency stimulation. Synapse 50:314–319 (2007) and Ermentrout and Terman (2010)).
Tass PA (2003) A model of desynchronizing deep brain
stimulation with a demand-controlled coordinated
reset of neural subpopulations. Biol Cybern 89:81–88
Tass PA, Majtanik M (2006) Long-term anti-kindling Detailed Description
effects of desynchronizing brain stimulation:
a theoretical study. Biol Cybern 94:58–66
Theory of dynamical systems and bifurcations
Tass PA, Qin L, Hauptmann C et al (2012) Coordinated
reset neuromodulation has sustained after-effects in provides a list of universal scenarios of dynamics
parkinsonian monkeys. Ann Neurol 72:816–820 changes under parameter variation. For example,
40 Dynamical Systems: Overview
one scenario explaining the onset of oscillations is represented by simple mathematical models.
is described by Andronov-Hopf bifurcation. In This approach to modelling of spiking times is
terms of neuroscience, this theory provides fruitful for implementation in the large neural
insight into the mechanisms underlying different network and when mathematical analysis
neural response properties and firing patterns. (in addition to numerical simulations) is desired.
Even more importantly, it allowed to elucidate Somewhat more intricate features of single
the underlying mathematical structure that might cell dynamics are described in articles on model-
be common to whole classes of firing behaviors. ling of “▶ Postinhibitory Rebound and Facilita-
Thus, conclusions of such studies can be wide tion” and “▶ Spike-Frequency Adaptation”
ranging, even if specific biophysical details of phenomena. Finally, the fast-slow dissection
implementation may differ from case to case. and bifurcation analysis really shine in the
A common theme in all articles in this section is description of bursting behavior. The bursting
that they use concepts from nonlinear dynamics, dynamics (of individual cells or of population
especially geometrical methods like phase planes activity) is dissected into active and silent phases
and bifurcation diagrams. Many exploit time scale when trajectories are restricted to lower dimen-
differences to reduce dimensionality by dissecting sional manifolds, and transitions between these
the dynamics using fast-slow analysis, i.e., to sepa- phases correspond to reaching the manifold’s
rately understand the behaviors on the different boundary and jumping to a different manifold.
time scales and then patch the behaviors together. While most of these examples are for single-cell
The articles in this section for the most part dynamics, the qualitative mathematical study is
exploit the idealized model of neuron localized at also applicable to the dynamics of neuronal net-
one point (i.e., electrically compact neuron), works and structures, especially in the mean-field
focusing on the nonlinearities of spiking dynam- approximations. One such example for network-
ics, and using biophysically minimal but biolog- generated rhythms is presented in the article on
ically plausible description of neural dynamics. “▶ Spike-Frequency Adaptation”.
An article on “▶ Fitzhugh-Nagumo Model” Of course, the models presented in this section
describes one of the first examples where dynam- are rather idealized and simplified for mathemati-
ical systems approach has been applied to analy- cal study compared to many other neuronal models
sis of neural dynamics (Fitzhugh 1955). Although that are designed to investigate the biophysical
“Fitzhugh-Nagumo Model” is not biologically details of action potential generation: interaction
grounded and formulated in terms of the cubic of many known ionic currents, or the spatial prop-
nonlinearity, the model is still considered as one agation of activity, etc. Such minimalistic models
of the prototype models for excitable systems. however are invaluable when the problem or ques-
An article on “▶ Morris-Lecar Model” tion at hand require only qualitative or semiquan-
describes the rich dynamic repertoire of the titative characterizations of spiking activity. This
two-variable Morris-Lecar model, as its biophys- is especially important in studies of large networks
ical parameters are varied. The original detailed of interacting cells.
analysis of this model (Rinzel and Ermentrout
1998) has laid ground for similar approaches in
many different contexts and provided Cross-References
a theoretical justification for influential
Hodgkin’s classification of “excitability types” ▶ Excitability: Types I, II, and III
(Hodgkin 1948, described in “▶ Excitability: ▶ Fitzhugh–Nagumo Model
Types I, II, and III”). ▶ Integrate and Fire Models, Deterministic
The articles on “▶ Integrate and Fire Models, ▶ Morris–Lecar Model
Deterministic” and “▶ Theta Neuron Model” ▶ Postinhibitory Rebound and Facilitation
models target the most idealized end of the ▶ Spike-Frequency Adaptation
modelling spectrum where the spiking activity ▶ Theta Neuron Model
Dynamics of Disease States: Overview 41
References physiological dynamics from healthy to unhealthy

with changes in underlying control parameters
Ermentrout GB, Terman D (2010) Mathematical founda- (Mackey and Glass 1977). Subsequently, this con-
tions of neuroscience. Springer, New York
cept of a dynamical disease was extended to that of
Fitzhugh R (1955) Mathematical models of threshold phe-
nomena in the nerve membrane. Bull Math Biophys a dynamic disease to account for the possibility
17(4):257–278 that mechanisms other than those associated with
Hodgkin AL (1948) The local electric changes associated changes in parameters may be involved
with repetitive action in a non-medullated axon.
(▶ Dynamic Diseases of the Brain).
J Physiol Lond 107:165–181
Izhikevich EM (2007) Dynamical systems in neurosci- Computational neuroscience extends dynami-
ence: the geometry of excitability and bursting. MIT cal approaches to neurological disease in two
Press, Cambridge, MA ways. First, by making it possible to include
Rinzel J, Ermentrout B (1998) Analysis of neural excit-
anatomical, physiological, and molecular details,
ability and oscillations. In: Koch C, Segev I (eds)
Methods in neuronal modeling: from ions to networks, computational models provide insights into how
2nd edn. MIT Press, Cambridge, MA, pp 251–291 mechanisms acting at the level of molecules and
individual neurons translate into phenomena
manifested clinically at the bedside (▶ Epilepsy:
Abnormal Ion Channels; ▶ Modeling of
Dynamics of Disease States: Disease – Molecular Level: Overview). Histori-
Overview cally the two neurological diseases which pro-
vided the most insight into cortical function
John Milton were temporal lobe epilepsy and classical
W.M. Keck Science Center, The Claremont migraine. Indeed the study of the geometry of
Colleges, Claremont, CA, USA migraine, drug- and flicker-induced visual forti-
fication patterns (▶ Stochastic Neural Field
Theory; ▶ Visual Hallucinations and Migraine
Synonyms Auras; ▶ Flicker-Induced Phosphenes), and
their propagation (▶ Migraines and Cortical
Dynamical diseases; Periodic diseases Spreading Depression) has provided deep insights
into the functional architecture of the visual cortex.
Surgical approaches for the treatment of patients
Detailed Description with medically intractable epilepsy provided the
impetus to directly record from the cortex of awake
The evolution of an illness is one of the clues that humans. This, in turn, motivated studies into the
a bedside physician uses to arrive at a diagnosis ability of large populations of neurons to synchro-
and treatment strategy for diseases that affect the nize and generate seizures (▶ LFP Analysis: Over-
nervous system. Is the onset acute or subacute? Is view; ▶ Transfer Function, Electrocortical;
the clinical course self-limited, relapsing- ▶ Epilepsy: Computational Models; ▶ Epilepsy,
remitting, cyclic, and chronic progressive? The Neural Population Models of). Moreover, compu-
impetus for studying disease dynamics comes tational models are now making it even possible to
from the mathematics and physics communities: gain insights into the nature of cognitive, func-
their long experience has shown that insights into tional, and psychiatric diseases of the nervous sys-
mechanism often derive from examining how tem that up to now have largely remained
dynamics change. Consequently, the time evolu- mysterious (▶ Computational Psychiatry).
tion of a disease is modeled as differential equa- Second, by having “a disease in a computer
tions, and disease processes are described in terms model,” it is possible to efficiently evaluate and
of the origin, stability, and bifurcations of the refine treatment strategies in silico before apply-
model’s dynamical behaviors. In 1977, Michael ing them to humans. Computational challenges
Mackey and Leon Glass associated changes in arise because of the presence of multistability
42 Dynamics of Disease States: Overview
(▶ Multistability in Seizure Dynamics; (▶ Cortical Motor Prosthesis; ▶ Functional Neu-

▶ Multistability: Stopping Events with Single roscience: Cortical Control of Limb Prosthesis).
Pulses) and time delays (▶ Time-Delayed Neural The frontier for dynamic disease is to under-
Networks: Stability and Oscillations). Indeed stand the collective behaviors of the nervous sys-
time-delayed, multistable dynamical systems tem that emerge over the timescale of years. Can
have a tendency to generate transient oscillations the development of an epileptic focus
that can be easily mistaken for limit cycle oscil- (epileptogenesis) be halted early so that an individ-
lations, thus causing confusion (Pakdaman ual at risk never experiences a seizure? Can the rate
et al. 1998). Nonetheless computational of learning of a complex voluntary skill, such as
approaches have already proven useful to the golf swing, by a patient with a robotic or stem
develop therapeutic strategies. cell-derived limb prosthesis be sped up to the point
Although the ultimate goal of medicine is that the individual could enjoy the use of these
cure, one cannot overlook the need to improve limbs throughout a lifetime? Large, complex phys-
the patient’s quality of life when cure cannot be ical systems tend to self-organize dissipative struc-
achieved. The electrical properties of neurons tures, namely, dynamical entities whose existence
make it possible to use electrical stimuli as is maintained far from equilibrium by a supply of
a treatment modality. Applications range from energy. Already dynamical signatures of this self-
aborting seizures with electrical stimuli organization, including power laws, have been
(▶ Multistability: Stopping Events with Single observed in the bursting propagating activities of
Pulses) to improving the quality of movements living neural populations (▶ Neuronal Ava-
of patients with Parkinson’s disease with deep lanches) and the dynamics of human balance con-
brain stimulation (▶ Parkinson’s Disease: Deep trol (▶ Human Balancing Tasks: Power Laws,
Brain Stimulation; ▶ Computational Models of Intermittency, and Lévy Flights).
Deep Brain Stimulation (DBS); ▶ Deep Brain Computational neuroscience provides the
Stimulation (Models, Theory, Techniques): tools for understanding how the nervous system
Overview). Even noisy stimuli can be beneficial learns to exert control, thereby bringing to many
by enhancing the detection of weak signals by the tangible hope of a better life.
sensory nervous system in order to improve the
function of cochlear implants in the hearing
impaired or balance control in those with periph- Cross-References
eral neuropathies (▶ Stochastic Resonance:
Balance Control and Cochlear Implants). ▶ Computational Models of Deep Brain
As insight increases in our understanding of Stimulation (DBS)
neural encoding, it is becoming possible to ▶ Computational Psychiatry
replace broken parts with electronic ones that ▶ Cortical Motor Prosthesis
perform the same function. A large number of ▶ Deep Brain Stimulation (Models, Theory,
articles in this Encyclopedia point to the current Techniques): Overview
enthusiasm in this therapeutic approach. Appli- ▶ Dynamic Diseases of the Brain
cations include restoring vision to the visually ▶ Epilepsy, Neural Population Models of
impaired (▶ Vision Prosthesis), hearing to those ▶ Epilepsy: Abnormal Ion Channels
who cannot hear (▶ Peripheral Nerve Interface ▶ Epilepsy: Computational Models
Applications, Cochlear Implants), continence to ▶ Flicker-Induced Phosphenes
those incontinent, and relief from pain to those ▶ Functional Neuroscience: Cortical Control of
who suffer (▶ Peripheral Nerve Interface Appli- Limb Prosthesis
cations, Neuropathic Pain). Dramatically, it has ▶ Human Balancing Tasks: Power Laws,
become possible to interface the brain directly Intermittency, and Lévy Flights
with electronic devices and make it possible for ▶ Large-Scale Neural Networks: Vision
a patient to move robotic limbs by thought alone ▶ LFP Analysis: Overview
Gamma and Theta Oscillations, Hippocampus: Overview 43
▶ Migraines and Cortical Spreading Depression and it generates several population activities that
▶ Modeling of Disease - Molecular Level: include theta and gamma rhythms. Many model-
Overview ing studies have focused on these oscillatory
▶ Multistability in Seizure Dynamics activities, and the entries in this section detail
▶ Multistability: Stopping Events with Single much of this work.
Pulses To develop and build models of any biological
▶ Neuronal Avalanches system, an in-depth appreciation of the biological
▶ Parkinson’s Disease: Deep Brain Stimulation and physiological basis of what is being modeled
▶ Peripheral Nerve Interface Applications, is required. This is provided in the “▶ Hippocam-
Cochlear Implants pus, Theta, Gamma, and Cross-Frequency Cou-
▶ Peripheral Nerve Interface Applications, pling” entry, where functional and experimental
Neuropathic Pain aspects are described. In developing models, the
▶ Stochastic Neural Field Theory cellular units and how they are connected need to
▶ Stochastic Resonance: Balance Control and be examined. Three entries on “▶ Hippocampus,
Cochlear Implants Model Inhibitory Cells,” “▶ Hippocampus, Model
▶ Time-Delayed Neural Networks: Stability and Excitatory Cells,” and “▶ Hippocampus,
Oscillations Model Network Architecture” provide these
▶ Transfer Function, Electrocortical details, and the challenges and complexities in
▶ Vision Prosthesis these aspects are laid bare. Two theoretical
▶ Visual Hallucinations and Migraine Auras entries on “▶ Subthreshold Amplitude and
Phase Resonance in Single Neurons” and
“▶ Mixed-Mode Oscillations in Single Neurons”
References provide the reader with the types of theoretical
approaches that can be used to help understand
Mackey MC, Glass L (1977) Oscillation and chaos in how these subthreshold and mixed-mode activi-
physiological control systems. Science 197:287–289
ties that are present in hippocampal cells may
Pakdaman K, Grotta-Ragazzo C, Malta CP (1998) Tran-
sient regime dynamics in continuous-time neural net- contribute to theta and gamma rhythms. Gamma
works with delays. Phys Rev E 58:3623–3627 rhythms have been extensively studied both the-
oretically and experimentally, and as such, there
are well-developed mechanistic understandings
for their generation. These mechanisms are
Gamma and Theta Oscillations, described and illustrated in the entry “▶ Hippo-
Hippocampus: Overview campal Oscillations, Mechanisms (PING, ING,
Sparse).” This is not the case for theta rhythms
Frances K. Skinner or nested theta/gamma rhythms. However, net-
Toronto Western Research Institute, University work models have been developed and these are
Health Network, Toronto, ON, Canada described along with the many considerations
Physiology, University of Toronto, Toronto, that arise in developing such network models in
ON, Canada “▶ Hippocampal Theta, Gamma, and Theta/
Medicine (Neurology), University of Toronto, Gamma Network Models.”
Toronto, ON, Canada In combination, the entries in this section pro-
vide the reader with a solid basis to learn about
theta and gamma oscillations in hippocampus
The brain expresses several rhythms that are considering mathematical modeling and
associated with normal and pathological states. experimental perspectives. An appreciation of
The hippocampus is arguably the most heavily the many aspects that are involved, both general
studied structure in the brain for many reasons and specific, can be gained from reading these
including its importance in learning and memory, entries.
44 Information Theory: Overview
Cross-References Very soon after Shannon’s initial publication

(1948), a small number of manuscripts provided
▶ Hippocampal Oscillations, Mechanisms the foundations for the current use of information
(PING, ING, Sparse) theory in neuroscience. MacKay and McCulloch
▶ Hippocampal Theta, Gamma, and (1952) applied the concept of information to
Theta/Gamma Network Models propose limits of the transmission capacity of
▶ Hippocampus, Model Excitatory Cells a nerve cell. This work foreshadowed future
▶ Hippocampus, Model Inhibitory Cells work on what can be termed “Neural Information
▶ Hippocampus, Model Network Architecture Flow” – how much information moves through the
▶ Hippocampus, Theta, Gamma, and nervous system, and the constraints that informa-
Cross-Frequency Coupling tion theory imposes on the capabilities of neural
▶ Mixed-Mode Oscillations in Single Neurons systems for communication, computation and
▶ Subthreshold Amplitude and Phase Resonance behavior. A second set of manuscripts, by
in Single Neurons Attneave (1954) and Barlow (1961) discussed
information as a constraint on neural system struc-
ture and function, proposing that neural structure
in sensory systems is matched to statistical struc-
Information Theory: Overview ture of the sensory environment, in a way to opti-
mize information transmission. This is the main
Alexander Dimitrov idea behind the “Structure from Information” line
Department of Mathematics, Washington State of research that is still very active today. A third
University, Vancouver, WA, USA thread, “Information Estimates,” started with a -
forward-looking article (Miller 1955), which
pointed out many potential pitfalls of extracting
Definition information quantities from observations. Its sig-
nificance penetrated the mainstream neuroscience
Information Theory started with Shannon’s semi- research later; once information-theoretic analysis
nal paper “A Mathematical Theory of Communi- became more widespread, the biases noted by
cation” (Shannon 1948). Because its importance Miller were rediscovered, and corrected.
and flexibility were quickly recognized, there were
numerous attempts to apply it to diverse field out-
side of its original scope. The entries in this section Subsequent Developments
provide an overview of the current state of Infor-
mation Theory in Neuroscience. The theme that arguably has had the widest influ-
ence on the neuroscience community is that of
“Neural Information Flow”. The initial works of
Detailed Description MacKay and McCulloch (1952) showed that neu-
rons are in principle able to relay large quantities of
When discussing a field, it is useful to review the information. That research also started the first
basic concepts and their properties. That has been major controversy in the field, which still resonates
provided in multiple articles for Information The- today: the debate about timing versus frequency
ory. To assist the reader, the entry “▶ Summary of codes (Stein 1967). A steady stream of articles
Information-Theoretic Quantities” provides followed, both discussing these hypothesis and
a brief summary of the main information-theoretic attempting to clarify the type of information
quantities. For a more thorough investigation, we relayed by nerve cells (Abeles and Lass 1975;
would direct interested readers to the most excel- Eagles and Purple 1974; Eckhorn and Pöpel 1974;
lent introduction to Information Theory by Cover Harvey 1978; Norwich 1977; Poussart 1971;
and Thomas (2006), now in its second edition. Stark et al. 1969; Taylor 1975; Walloe 1970).
Information Theory: Overview 45
After the initial rise in interest, the application developments in Information Theory.
of Information Theory to neuroscience was “▶ Directed Information Flow and Causality in
extended to a few more systems and questions Neural Systems” discusses developments based
but did not spread too broadly. This was presum- on the natural ideas of information flow in
ably because, despite strong theoretical advances a causal direction. Surprisingly, this idea has
in Information Theory, its applicability was ham- taken quite some time to develop in the commu-
pered by difficulty in measuring and interpreting nication literature due to various technical diffi-
information-theoretic quantities. culties. The works originates form the ideas of
The work of de Ruyter van Steveninck and Granger (1969) on causal interactions. The
Bialek (1988) started what could be called the information-theoretic perspective of (Massey
modern era of information-theoretic analysis in 1990; Massey and Massey 2005) removed
neuroscience, in which Information Theory is see- Granger’s linearity assumptions.
ing more and more refined applications. Their And lastly, applications in neuroscience are
work advanced the conceptual aspects of the appli- also pushing the development of new tools in
cation of information theory to neuroscience and Information Theory. One of the examples
provided an impetus to removing biases in infor- presented here, “▶ Information Measures of
mation estimates, discussed in detail in the entry Redundancy and Synergy in Neural Activity,”
“▶ Estimating Information-Theoretic Quantities.” summarizes the progress made in that direction,
for which the original definitions provided by
Shannon proved inadequate.
Current State In conclusion, Information Theory is thriving in
the neuroscience community, and the long efforts
Information Theory found applications in the are bearing fruit, as diverse research questions are
study of neural processing as a theoretical and being approached with more elaborate and refined
practical system for the analysis of communi- tools. As demonstrated by several recent thematic
cated signals. A variety of information-theoretic journal issues (Dimitrov et al. 2011; Milenkovic
quantities are in use in Neuroscience. The entry et al. 2010), Information Theory is firmly inte-
“▶ Applications of Information Theory to Anal- grated in the fabric of neuroscience research, and
ysis of Neural Data” provides a general overview a progressively wider range of biological research
of current applications of Information Theory, as in general, and will continue to play an important
an analysis tool of neural information flow. role in these disciplines. Conversely, neuroscience
The structure of neural activity often intro- is starting to serve as a driver for further research in
duces challenges that have been of marginal Information Theory, opening interesting new
interest to the engineering community. The directions of inquiry.
entry “▶ Metric Space Analysis of Neural Infor-
mation Flow” exemplifies one such case,
metrization. In a metric-space approach to ana- Cross-References
lyzing neural response, the data is reduced from
a set of complicated objects, spike trains, to ▶ Applications of Information Theory to
a much simpler object, the matrix of distances Analysis of Neural Data
between the spike trains. When applied to neural ▶ Directed Information Flow and Causality in
information, this matrix is used to estimate infor- Neural Systems
mation theory quantities for the corresponding ▶ Estimating Information-Theoretic Quantities
spike train data. The tools developed in that direc- ▶ Information Measures of Redundancy and
tion combine metric properties of spike trains Synergy in Neural Activity
with their information transmission function. ▶ Metric Space Analysis of Neural Information
Several entries discuss specific applications in Flow
neuroscience stemming from more recent ▶ Summary of Information-Theoretic Quantities
46 Invertebrate Pattern Generation: Overview
References Stark L, Negrete-Martinze J, Yankelevich G, Theodoridis

G (1969) Experiments on information coding in nerve
Abeles M, Lass Y (1975) Transmission of information by the impulse trains. Math Biosci 4(3–4):451–485
axon: II. The channel capacity. Biol Cybern 19:121–125 Stein RB (1967) The information capacity of nerve cells
Attneave F (1954) Some information aspects of visual using a frequency code. Biophys J 7(6):797–826
perception. Psychol Rev 61:183–193 Taylor RC (1975) Integration in the crayfish antennal
Barlow HB (1961) Possible principles underlying the trans- neuropile: topographic representation and multiple-
formation of sensory messages. In: Rosenblith WA (ed) channel coding of mechanoreceptive submodalities.
Sensory communications. MIT Press, Cambridge, MA Dev Neurobiol 6(5):475–499
Cover T, Thomas J (2006) Elements of information Walloe L (1970) On the transmission of information
theory, 2nd edn, Wiley series in communication and through sensory neurons. Biophys J 10(8):745–763
signal processing. Wiley, Hoboken
de Ruyter van Steveninck R, Bialek W (1988) Real-time
performance of a movement- sensitive neuron in the
blowfly visual system: coding and information transfer Invertebrate Pattern Generation:
in short spike sequences. Proc Royal Soc Ser B Biol
Sci 234(1277):379–414
Overview
Dimitrov AG, Lazar AA, Victor JD (2011) Information
theory in neuroscience. J Comput Neurosci 30(1):1–5 Farzan Nadim
Eagles JP, Purple RL (1974) Afferent fibers with multiple Federated Department of Biological Sciences,
encoding sites. Brain Res 77(2):187–193
New Jersey Institute of Technology/Rutgers
Eckhorn R, Pöpel B (1974) Rigorous and extended applica-
tion of information theory to the afferent visual system University, Newark, NJ, USA
of the cat. I. Basic concepts. Biol Cybern 16:191–200 Department of Mathematical Sciences, New
Granger C (1969) Investigating causal relations by econo- Jersey Institute of Technology, Newark, NJ, USA
metric models and cross-spectral methods.
Econometrica 37(3):424–438
Harvey R (1978) Patterns of output firing generated by
a many-input neuronal model for different model Detailed Description
parameters and patterns of synaptic drive. Brain Res
150(2):259–276
MacKay DM, McCulloch WS (1952) The limiting infor-
Central pattern generators (CPGs) are networks
mation capacity of a neuronal link. Bull Math Biophys of neurons in the central nervous system (CNS)
14:127–135 that produce patterned activity, usually as coher-
Massey J (1990) Causality, feedback and directed infor- ent oscillations, in the absence of external timing
mation. In: Proceedings of the international sympo-
cues. CPGs provide timing input to motor neu-
sium on information theory applications (ISITA-90),
Yokohama National University, Yokohama, Japan, rons whose discharge dictates movements of
pp 303–305 muscles that control rhythmic behavior such as
Massey J, Massey P (2005) Conservation of mutual and respiration or locomotion (Marder and Calabrese
directed information. In: Proceedings of the interna-
1996). The long-held debate between scientists
tional symposium on information theory (ISIT 2005),
IEEE Publishing, New York, NY, pp 157–158 who believed half a century ago that rhythmic
Milenkovic O, Alterovitz G, Battail G, Coleman TP, motor activity is generated by reflex chains and
Hagenauer J, Meyn SP, Price N, Ramoni MF, those who held the more radical view of centrally
Shmulevich I, Szpankowski W (2010) Introduction to
generated rhythms was resolved conclusively, in
the special issue on information theory in molecular
biology and neuroscience. IEEE Trans Inf Theory favor of the latter group, by demonstrating that
56(2):649–652 neural networks generating rhythmic motor
Miller GA (1955) Note on the bias of information esti- activity can do so in the isolated nervous system,
mates. In: Information theory in psychology: problems
in the absence of the body and therefore sensory
and methods. In: Henry Q (ed.), vol II-B. Free Press,
Glenco, IL, pp 95–100 feedback. In the early 1960s, these “fictive”
Norwich K (1977) On the information received by sensory motor patterns were first demonstrated to govern
receptors. Bull Math Biol 39:453–461 rhythmic activation in two invertebrate model
Poussart DJM (1971) Membrane current noise in lobster
systems: the movement of flight wings in locusts
axon under voltage clamp. Biophys J 11(2):211–234
Shannon CE (1948) A mathematical theory of communi- and the beating of swimmerets in crayfish
cation. Bell Syst Tech J 27:623–656 (Wilson 1961; Ikeda and Wiersma 1964).
Invertebrate Pattern Generation: Overview 47
The demonstration of fictive motor activity led to CPG neurons (Plant and Kim 1976), paved
the development of in vitro preparations in search the way for the complete mathematical analysis
of the underlying CPG circuits. Although CPG of bursting mechanisms in neurons and other
networks have been traditionally described in the excitable cells (Rinzel 1987; Rinzel and
context of controlling motor activity, a more con- Lee 1987). This section of the Encyclopedia of
temporary viewpoint of CPGs includes networks Computational Neuroscience provides an over-
that subserve brain oscillations connected to sen- view of the contributions of invertebrate pattern
sory, cognitive, and memory tasks (Yuste generators in the context of computational models.
et al. 2005).
CPGs have historically led systems neurosci-
ence in the understanding of neural circuit inter- Cross-References
actions, partly because of the ease of
identification of neurons and networks whose ▶ Automated Parameter Search in Small
activity correlates with a rhythmic motor activity. Network Central Pattern Generators
The identification of neural circuits has been ▶ Bifurcations Dynamics of Single Neurons and
more successful in invertebrates where the num- Small Networks
ber of neurons involved in neural processing is ▶ Bursting in Neurons and Small Networks
lower, sometimes by orders of magnitude, and the ▶ Gap Junctions in Small Networks
ability of identifying synaptic connections among ▶ Neuromodulation in Small Networks
neurons is facilitated by dual recordings. Verte- ▶ Sensory Input to Central Pattern Generators
brate and especially mammalian neural circuit ▶ Short-Term Synaptic Plasticity in Central
analysis was, until recently, performed with Pattern Generators
cruder techniques such as lesions, but in the past ▶ Stability and Homeostasis in Small Network
decade or so, genetic tools and the identification Central Pattern Generators
of molecular markers have allowed for more pre-
cise circuit analysis in the large networks of ver-
tebrate systems (Han 2012; Arrenberg and
Driever 2013). However, neural circuits have References
been identified only in a few vertebrate model
Arrenberg AB, Driever W (2013) Integrating anatomy and
systems and simple behaviors (Issa et al. 2011;
function for zebrafish circuit analysis. Front Neural
Cangiano and Grillner 2005). The knowledge of Circ 7:74
the neural circuits and the ability to record func- Cangiano L, Grillner S (2005) Mechanisms of rhythm
tionally identified neurons in invertebrates have generation in a spinal locomotor network deprived of
crossed connections: the lamprey hemicord.
allowed for the in-depth analysis of the mecha-
J Neurosci 25(4):923–935
nisms underlying circuit dynamics and plasticity Han X (2012) In vivo application of optogenetics for
(Marder et al. 2005) as well as a rigorous descrip- neural circuit analysis. ACS Chem Neurosci
tion of the computations performed by the neural 3(8):577–584
Ikeda K, Wiersma CA (1964) Autogenic rhythmicity in
circuits (Selverston 2010).
the abdominal ganglia of the crayfish: the control of
The computational description of pattern- swimmeret movements. Comp Biochem Physiol
generating circuits evolved in parallel with the 12:107–115
in vitro experimental studies of cellular and syn- Issa FA, O’Brien G, Kettunen P, Sagasti A, Glanzman DL,
Papazian DM (2011) Neural circuit activity in freely
aptic mechanisms. The complexities of circuit behaving zebrafish (Danio rerio). J Exp Biol
analysis of invertebrate CPGs have led to numer- 214(Pt 6):1028–1038
ous modeling studies, some of which have been Marder E, Calabrese RL (1996) Principles of
influential in shaping our conceptual understand- rhythmic motor pattern generation. Physiol Rev
76(3):687–717
ing of both single-neuron and network operations.
Marder E, Bucher D, Schulz DJ, Taylor AL (2005) Inver-
For example, the computational description of tebrate central pattern generation moves along. Curr
bursting oscillations, led by models of invertebrate Biol 15(17):R685–R699
48 Invertebrate Sensory Systems: Overview
Plant RE, Kim M (1976) Mathematical description of the section on their motor systems. It presents an
a bursting pacemaker neuron by a modification of the overview of computational modeling emphasiz-
Hodgkin-Huxley equations. Biophys J 16(3):227–244
Rinzel J (1987) A formal classification of bursting mech- ing recent progress in the field. While the topics
anisms in excitable systems. Mathematical topics in covered by individual articles are far from
populations biology, morphogenesis. Lecture notes in exhaustive, it is our hope that they will provide
biomathematics, vol 71. Springer, Berlin the reader with up-to-date information and access
Rinzel J, Lee YS (1987) Dissection of a model for neuro-
nal parabolic bursting. J Math Biol 25(6):653–675 to further references covering a broad range of
Selverston AI (2010) Invertebrate central pattern genera- related topics.
tor circuits. Philos Trans R Soc Lond B Biol Sci Among the different sensory systems, vision
365(1551):2329–2345 has arguably been the most intensely investigated
Wilson DM (1961) The central nervous control of flight in
a locust. J Exp Biol 38:471–490 in invertebrates and vertebrates alike. The chap-
Yuste R, MacLean JN, Smith J, Lansner A (2005) The ter on “▶ Phototransduction Biophysics” sum-
cortex as a central pattern generator. Nat Rev Neurosci marizes our current understanding of the
6(6):477–483 mechanisms by which invertebrate photorecep-
tors translate light signals into membrane poten-
tial changes. Most of the work on this topic has
been carried out in flies and particularly the
Invertebrate Sensory Systems: genetically tractable fruit fly, Drosophila
Overview melanogaster. The biochemical cascade-
mediating phototransduction has been studied in
Fabrizio Gabbiani great detail and has recently led to a better under-
Department of Neuroscience, Baylor College of standing of the implications for neural coding.
Medicine, Houston, TX, USA Another classical topic of insect vision has been
the study of the mechanisms underlying
directionally selective motion detection. This
Detailed Description work originated in beetles and was later pursued
in flies, leading to the correlation model of
Invertebrate neurobiology has long been at the Hassenstein and Reichardt (1956). The neural
forefront of computational neuroscience, starting circuits implementing this model, which is
with the mathematical model of the biophysical closely related to the motion-energy model of
basis of the action potential by Hodgkin and Hux- mammalian motion detection (van Santen and
ley in the squid giant axon almost 50 years ago Sperling 1985), have long been proven difficult
(Hogdkin and Huxley 1952). Many invertebrate to investigate due to difficulties in tracing the
systems are highly suitable for detailed modeling anatomical and physiological connections
because of their relatively compact size and the between the neurons that perform the computa-
fact that their neurons can often be uniquely iden- tions of the correlation model.
tified. This feature allows the formulation of pre- Recent progress has been made by combining
cise descriptions of their behaviors and simplifies new anatomical, electrophysiological, imaging,
the interpretation of experimental results. Yet, as and genetic techniques in Drosophila that are sum-
Hodgkin and Huxley’s seminal analysis of action marized in the chapter on “▶ Visual Motion
potential mechanisms has proved, invertebrate Detection in Drosophila.” Although Drosophila
models possess characteristics similar to those of has recently helped better understand the circuitry
vertebrates and the computational principles underlying motion detection, a large body of work
derived from either type of nervous systems on the topic was carried out in bigger flies, leading
have been found universally applicable, even if to detailed biophysical models of a class of neu-
details of the mechanistic implementations differ. rons involved in motion detection. These models
This entry of the Encyclopedia focuses on the are introduced in the chapter on “▶ Fly Lobula
sensory systems of invertebrates, complementing Plate Tangential Cells (LPTCs), Models of.”
Invertebrate Sensory Systems: Overview 49
Local motion detection is only the first step in studied auditory model systems, face physical
processing visual information generated by an constraints quite different from those applying
animal moving in its environment. Such informa- to vertebrates. Their study has led to a wealth of
tion, often called “optic flow,” is particularly information on how auditory information is
important for flying animals. The strategies and processed, both at the level of the auditory
constraints on the processing of such information periphery and in the brain. These results are sum-
are summarized in the two chapters on “▶ Optic marized in the chapter on “▶ Auditory
Flow Processing” and “▶ Visual Processing in Processing in Insects.”
Free Flight.” Wind and air current sensing is carried out by
Another function of the visual system of crit- a specialized sensory system called the cercal
ical importance to animals is predator evasion system in crickets and other orthopteran insects.
and collision avoidance. Our understanding of This system provides a fascinating example of
the neural mechanisms of collision avoidance how information is processed by populations of
has rapidly progressed over the past decade in sensory neurons, as summarized in the chapter
a variety of invertebrate model systems, includ- entitled “▶ Cercal System.” More broadly, the
ing locusts, crabs, and fruit flies. Similar collision chapters on “▶ Leech Local Bend: Neural Cod-
avoidance mechanisms have been documented in ing of Touch Location” and “▶ Computation
vertebrate systems such as goldfish and pigeons. with Population Codes” explain the role such
The article on “▶ Collision Avoidance Models, neural codes play in a variety of other animals.
Visually Guided” summarizes our current under- The leech has proven to be one of the best models
standing of those systems. Animal navigation and to study these topics because of the compact size
migration relies on an internal representation of of its nervous system, well-understood natural
the external world allowing orientation over long behaviors, and the applicability of large-scale
distances. In insects, particularly in monarch but- imaging of neuronal activity using voltage-
terflies, ants, and locusts, the detection of polar- sensitive dyes and other electrophysiological tech-
ized patterns of light in the sky provides an niques. Finally, invertebrate somatosensation is
internal compass for navigation. The detection introduced in the chapter on “▶ Tactile Sensing
and processing of polarized light patterns leads in Insects.”
to a sort of “cognitive map” in the central nervous Adaptation is a fundamental property of neu-
system akin to those studied in the rodent hippo- ronal responses observed throughout sensory sys-
campus. The chapter on “▶ Polarization Vision” tems. The chapter entitled “▶ Biophysics of
provides a summary of our current understanding Adaptation in a Computational Model of the
of this fascinating and exotic sensory modality. Leech T Neuron” details the ionic mechanisms
Olfaction is a prominent and important sense of adaptation in one type of sensory neuron and
in invertebrates, particularly in insects. The its consequences for the processing of sensory
mechanisms of insect olfaction and their close stimuli.
relation to vertebrate olfaction have been inves- In addition to the above mentioned chapters,
tigated intensely over the past 30 years, culmi- the section includes one chapter – entitled
nating in a detailed understanding in several “▶ Nitric Oxide Neuromodulation” – on
organisms, including moths, locusts, and fruit a somewhat exotic and fascinating volume
flies. Examples of models of olfactory coding neuromodulator: nitric oxide. The computational
are presented in the chapter entitled “▶ Insect properties of nitric oxide neuromodulation have
Olfaction: a Model System for Neural Circuit been investigated in detail in several invertebrate
Modeling.” Another primary sense, audition, is systems, leading to models of its role in sensory
used by animals to locate preys or predators and processing and learning. Similar regulation of
for communication with conspecifics. Because of neuronal circuits by volume neurotransmission
their small size, invertebrates and insects such as has recently been described in vertebrates as
crickets and locusts, which are among the best- well (Oláh et al. 2009).
50 Learning Rules: Overview
The last topic covered in this entry is the effi- bei der Bewegungsperzeption des R€ usselk€afers
cient representation of neural information in the Chlorophanus. Z Naturforsch 11b:513–524
Hogdkin AL, Huxley AF (1952) A quantitative
chapter entitled “▶ Sensory Coding, Efficiency.” description of membrane current and its application
The possibility that sensory neural codes may be to conduction and excitation in nerve. J Physiol
efficient has been first raised in the 1960s (Barlow 117:500–544
1961). It has since then been investigated in several Oláh S, F€
ule M, Komlósi G, Varga C, Báldi R, Barzó P,
Tamás G (2009) Regulation of cortical microcircuits
systems, including the visual system of inverte- by unitary GABA-mediated volume transmission.
brates. The chapter includes background informa- Nature 461:1278–1281
tion and a summary of more recent results on van Santen JPH, Sperling G (1985) Elaborated Reichardt
efficient coding in these systems. detectors. J Opt Soc Am A 2:300–321
Cross-References
Learning Rules: Overview
▶ Auditory Processing in Insects
▶ Biophysics of Adaptation in a Computational Klaus Obermayer
Model of the Leech T Neuron Neural Information Processing Group,
▶ Cercal System Fakult€at IV, Technische Universit€at Berlin,
▶ Collision Avoidance Models, Visually Guided Berlin, Germany
▶ Computation with Population Codes
▶ Fly Lobula Plate Tangential Cells (LPTCs),
Models of Detailed Description
▶ Hodgkin-Huxley Model
▶ Insect Olfaction: a Model System for Neural Learning is the most important way by which
Circuit Modeling the environment alters the behavior in animals
▶ Leech Local Bend: Neural Coding of Touch and humans. Neurobiological as well as compu-
Location tational studies have stressed the significance of
▶ Nitric Oxide Neuromodulation synaptic plasticity in the learning process:
▶ Optic Flow Processing Experience and training leads to modifications
▶ Phototransduction Biophysics of synapses, which in turn lead to changes in
▶ Polarization Vision neuronal firing patterns, causing changes in
▶ Sensory Coding, Efficiency behavior. A full understanding of the processes
▶ Sensory Input to Central Pattern Generators underlying learning, therefore, requires an
▶ Spike-Frequency Adaptation understanding on the neural as well as on the
▶ Tactile Sensing in Insects systemic level.
▶ Visual Motion Detection in Drosophila In the behavioral literature, learning is stud-
▶ Visual Processing in Free Flight ied by exposing animals and humans to sensory
stimuli and having them interact with the envi-
ronment. The paradigms which have been used
References gave rise to a categorization into associative and
nonassociative learning. In nonassociative
Barlow HB (1961) Possible principles underlying learning an animal is repeatedly exposed to
the transformation of sensory messages. In: a single type of stimulus, and learning is
Rosenblith W (ed) Sensory communication. MIT
described by the strength of its impact leading,
Press, Cambridge, MA, pp 217–234, Chapter 13
Hassenstein B, Reichardt W (1956) Systemtheoretische Ana- for example, to the phenomena of habituation
lyse der Zeit-Reihenfolgen- und Vorzeichenauswertung and sensitization. In associative learning
Learning Rules: Overview 51
animals learn predictive relationships in the widely believed to be the most important feature
presence of a reinforcement signal. Historically, underlying learning. It also plays an important
classical conditioning (where reinforcements are role during neural development, where the inter-
delivered independent of any action taken) is play of neural activity with the so-called intrinsic
differentiated from operant conditioning processes shapes the circuitry and the neural
(where reinforcements depend on the action response properties. Learning at the neuronal
and where behavior is associated with its level is characterized by how the activation his-
outcome). tory of pre- and postsynaptic neurons relates to
In the neural network literature, learning is the efficacy of the synaptic connection
often classified according to what kind of between them. Different stimulation protocols
information is available in a particular learning of neurons in (mostly) in vitro preparations
scenario. In the so-called supervised learning, an uncovered a wealth of phenomena, the most
agent (e.g., a neural network) forms associations prominent ones being long-term potentiation,
between two sets of events with the goal to long-term depression, or spike-timing-dependent
predict one by the other. Learning is based on plasticity.
a given set of correct pairings. Pairings between “Learning rules,” i.e., the title of this section,
events are also formed in the so-called reinforce- refer to the quantification of the effects of expe-
ment learning; however, evaluative feedback riences, training, and stimulation on brain and
about network performance is only provided in behavior on all abovementioned levels of
the form of reinforcements, for example, reward description. Learning rules inserted into compu-
or punishment signals. Correct pairings imply tational models then help to explore the conse-
high rewards; hence, learning is driven by max- quences of the observed plasticity. Thus, one can
imizing returns. Reinforcement learning bears explore, for example, how activity-dependent
similarities with the abovementioned associative synaptic plasticity interacts with network
learning paradigms, and algorithms which were responses which eventually induce behavior,
originally developed by the neural network com- how a network can adapt its parameters such
munity are now widely applied to quantify asso- that a desired function can be performed, or
ciative learning in animals and humans. In the how interindividual differences in learning
so-called unsupervised learning, responses of an behavior can be related to interindividual differ-
agent are modified when stimuli are presented ences in their objectives. Thus, learning rules
but in a nonassociative way. In an artificial agent serve as a computational tool to link experience-
setting, supervised learning is often applied for induced changes observed on a system level to
solving pattern recognition problems, and the mechanisms operating on the level of neurons
reinforcement learning is used to learn action and networks and vice versa.
sequences. Unsupervised learning, on the At the core of this section are entries for
other hand, is applied to learn internal represen- activity-dependent learning rules, which quantify
tations of the outside world. Statistical regulari- how pre- and postsynaptic activity changes
ties are extracted and are used to build the strength of a synapse (see entries “▶ Hebbian
representations of stimuli and actions which Learning”, “▶ Anti-Hebbian Learning”, “▶ Spike-
are better suited for recognition, planning, Timing Dependent Plasticity, Learning Rules”,
memorization, decision making, and other cog- “▶ Tempotron Learning”). Here, the learning
nitive tasks. rules are formulated in a more abstract setting,
In many areas of neuroscience where “real” and the biophysical foundations are addressed
neurons and networks are studied, learning is in section Synaptic Dynamics (cf. Long-
linked to changes of neuronal and synaptic prop- Term Depression, Long-Term Plasticity, Long-
erties. Activity-dependent synaptic plasticity is Term Potentiation, Short-Term Depression,
52 LFP Analysis: Overview
Short-Term Plasticity, Spike-Timing-Dependent

Plasticity). The effects of learning rules in terms LFP Analysis: Overview
of network dynamics and computation are then
addressed in a neural network setting (see entries Alain Destexhe1 and Joshua A. Goldberg2
1
“▶ Boltzmann Machine”, “▶ Hopfield Network”, Unit of Neuroscience Information and
“▶ Perceptron Learning”, “▶ Slow Feature Anal- Complexity (UNIC), Centre National de la
ysis”) covering supervised and unsupervised learn- Recherche Scientifique (CNRS), Gif-sur-Yvette,
ing paradigms. They are also addressed in a neural France
2
modeling setting, where the computational models Department of Medical Neurobiology, Institute
are used to understand some of the mechanisms for Medical Research Israel-Canada, Faculty of
underlying neural development (see entries Medicine, The Hebrew University of Jerusalem,
“▶ Cortical Maps, Activity-Dependent Develop- Jerusalem, Israel
ment” and “▶ Cortical Maps, Intrinsic Processes”).
The section is complemented by one entry on
reinforcement learning (Reward-Based Learning), Detailed Description
which summarizes current computational
approaches for quantifying behavior for reward- Local field potentials (LFPs) are low-frequency
based, associative learning paradigms. Possible electrical potentials recorded with micro- or
neural implementations of reinforcement macro-electrodes throughout the brain. The upper
learning are discussed in section “Cortex” frequency cutoff of the LFP is often considered to be
(cf. “▶ Reinforcement Learning in Cortical around 100 Hz but may be as high as 500 Hz in some
Networks”). studies. The normal amplitude of the LFP can range
from a few microvolts (e.g., in the basal ganglia)
(Goldberg et al. 2004) to hundreds of microvolts in
Cross-References the cortex. Perhaps the earliest recording of an LFP
may be attributed to Renshaw, Forbes, and Morison
▶ Anti-Hebbian Learning in 1940, who described these potentials in the cortex
▶ Boltzmann Machine and the hippocampus of a cat under various anes-
▶ Cortex: Overview thetics (Renshaw et al. 1940). Extensive electro-
▶ Cortical Maps, Activity-Dependent physiological studies have formed the current
Development view that cortical LFPs result from synaptic activ-
▶ Cortical Maps, Intrinsic Processes ity (Eccles 1951; Creutzfeldt et al. 1966; Elul
▶ Hebbian Learning 1971; Klee and Rall 1977; Mitzdorf 1985; Bedard
▶ Hopfield Network et al. 2004). Correspondingly, the LFP fluctuations
▶ Long Term Plasticity, Biophysical Models tend to have a strict phase relationship to cortical
▶ Perceptron Learning discharge: negative deflections in the LFP coin-
▶ Reinforcement Learning in Cortical Networks cide with increases in the instantaneous firing rates
▶ Reward-Based Learning, Model-Based and of cortical neurons in both superficial and deep
Model-Free layers (Lass 1968; Gray and Singer 1989; Murthy
▶ Short-Term Plasticity, Biophysical Models and Fetz 1996a; Donoghue et al. 1998; Destexhe
▶ Slow Feature Analysis et al. 1999). In addition, these fluctuations are
▶ Spike-Timing Dependent Plasticity (STDP), highly correlated across distances of several mil-
Biophysical Models limeters in the cortex (Eckhorn and Obermueller
▶ Spike-Timing Dependent Plasticity, Learning 1993; Sanes and Donoghue 1993; Murthy and
Rules Fetz 1996b; Contreras et al. 1997; Bullock 1999;
▶ Synaptic Dynamics: Overview Destexhe et al. 1999), indicating that LFPs are not
▶ Tempotron Learning always strictly local.
LFP Analysis: Overview 53
Theoretical attempts to account for the LFPs of EEG and MEG signals and how such signals
and model their generation date back to Lorente relate to the LFP. The “inverse problem” of esti-
de Nó (1947) and include the seminal works of mating neuronal sources from EEG and MEG
Ulla Mitzdorf (1985), who implemented the tech- signals is also reviewed.
nique of current source density (CSD), and The relationship to unit activity is further
Wilfrid Rall, who calculated the potentials gen- described in detail in “▶ Local Field Potential,
erated by various spatial organization of electri- Relationship to Unit Activity” by Bartosz
cal dipoles (Klee and Rall 1977; Goldberg Teleńczuk and Alain Destexhe. This entry
et al. 2004). These theoretical accounts relied on reviews the relation between extracellularly
the large-scale repetitive columnar structure of recorded unit activity and LFP, as a function of
the cortical circuitry to generate large potentials the frequency of the different rhythmical activi-
due to the summed contribution of many aligned ties found in LFPs. It is shown that unit activity
neuronal dipoles (Elul 1971; Klee and Rall 1977; correlates with the high-frequency (>200 Hz)
Abeles 1982; Mitzdorf 1985; Eggermont and components of the LFP and, to a lesser extent,
Smith 1995). with the low-frequency components (<10 Hz),
This “LFP analysis” section, which has contri- whereas the intermediate-frequency bands have
butions from several leading authorities, covers a rather weak correlation with unit activity.
the following topics: (a) methods for measuring In the entry “▶ Local Field Potential, Rela-
the LFP; (b) modern accounts of how LFPs are tionship to Membrane Synaptic Potentials,”
generated; (c) the relationship between the LFP Aryeh Taub, Ilan Lampl, and Michael Okun go
and other common measures of collective brain a bit more deeper in the neuron and review the
activity such as the EEG, the MEG, and the fMRI, relation between the intracellularly recorded
as well as its relationship to and possible influence membrane potential and the LFP signal. The
on single-cell membrane potentials; (d) the power entry considers different types of brain activity,
spectral structure of the LFP, which often displays such as up-down states, and discusses the role of
a 1/f structure, or marked oscillation peaks under inhibition.
various physiological and pathophysiological con- In the entry “▶ Local Field Potentials and
ditions (e) LFP in the context of particular brain Ephaptic Coupling,” Costas Anastassiou dis-
functions such as vision and olfaction; and cusses the hypothesis that the LFP is not merely
(f) current modeling methods of the LFP. These an epiphenomenon of electrical brain activity but
various topics are overviewed below. may actually serve as a global slow signal that
can couple to other neural elements (e.g., axons,
Neurophysiological Basis of LFPs synapses). This ephaptic coupling results from
A first important aspect of the LFP is the relation the fact that the membrane voltages in these ele-
between this signal and other well-known signals ments can be altered by fluctuations in the LFP
of the brain. In the entry “▶ Local Field Potential, simply because they are the difference between
Relationship to BOLD Signal,” Nikos Logothetis the intracellular potential and the extracellular
and Stefano Panzeri review the relation between (local field) potential.
the LFP signals and the BOLD (blood oxygen Finally, in the entry “▶ Local Field Potentials
level dependent) signal, as recorded by MRI tech- (LFP),” Alain Destexhe and Claude Bedard
niques. The entry also discusses multiunit activ- review general properties of LFPs, such as their
ity in relation with the BOLD signal. spatial coherence, namely, how LFPs recorded by
A similar large-scale approach is followed by neighboring electrodes relate to each other.
Stephanie Jones in “▶ Local Field Potential, They also review temporal (spectral) properties
Relationship to Electroencephalogram (EEG) of LFPs, such as their typical 1/f structure. The
and Magnetoencephalogram (MEG).” This entry also overviews different approaches for
entry reviews the biophysical bases of the genesis modeling LFPs.
54 LFP Analysis: Overview
Oscillatory Properties of LFPs The modeling of LFPs is further developed in

A fundamental property of LFP signals is their the entry “▶ Local Field Potential Interaction
propensity to display oscillations. In the entry with the Extracellular Medium” by Claude
“▶ Local Field Potential and Deep Brain Stimu- Bedard and Alain Destexhe. This entry reviews
lation (DBS),” Manuela Rosa, Sara Marceglia, models of LFPs by staying as general as possible
Sergio Barbieri, and Alberto Priori review the and includes the electrical nature of an extracel-
use of LFPs in the treatment of DBS in humans. lular medium. Both microscopic and macro-
It describes how DBS provides LFP recordings in scopic (mean-field) models of the LFP are
humans, how they are related to oscillatory activ- considered. It is shown that the frequency-
ity, and how DBS may interfere with these filtering properties of the extracellular medium
(sometimes pathological) oscillations. can fully explain the spectral properties of LFPs
Pathological oscillations are further investi- (such as power law or 1/f scaling).
gated in the entry “▶ Local Field Potential and
Movement Disorders” by Annaelle Devergnas Methodological Aspects of the LFP
and Thomas Wichmann. They review the use of Different methods to record LFPs are described
LFP recordings in movement disorders and show in the entry “▶ Local Field Potential, Methods of
that LFPs display various types of oscillations, at Recording” by Andrew Sharott. This entry
different frequencies, for different pathologies describes important factors such as the type of
associated to movement disorders. electrode, the role of the reference, and the
In the entry “▶ Local Field Potential in Olfac- recording conditions. The entry discusses how
tion,” Leslie Kay review the LFP signals recorded such factors can be determined to correctly inter-
in the olfactory system (olfactory bulb and pyriform pret the LFP signal.
cortex). Their review emphasizes the emergence of In the entry “▶ Resistivity/Conductivity of
beta and gamma oscillations in olfactory structures. Extracellular Medium,” Scott Lempka and Cam-
The occurrence of LFP oscillations in the visual eron McIntyre review another set of important
system is investigated in the entry “▶ Local Field biophysical properties that contribute to LFP gen-
Potential in the Visual System” by Gregor Rainer. esis: the electrical properties of neural tissue,
This entry reviews the occurrence of LFPs in the associated measurements of conductivity,
visual cortex, with an emphasis on visually evoked the frequency-filtering properties, as well as the
potentials, and visually evoked oscillations in the influence of volume conduction.
gamma frequency band. Finally, in the entry “▶ Current Source Density
Finally, in the entry “▶ Local Field Potential, (CSD) Analysis,” Daniel Wójcik reviews the
Synchrony of,” Ariana Frederick, Jonathan physical bases of the CSD analysis, which consists
Bourget-Murray, and Richard Courtemanche of estimating neuronal current sources from LFP
review how LFPs can reveal the presence of net- measurements performed by electrodes located at
work synchrony. Synchrony is here defined as the equidistant points in space. Different variants of
synchronization between different networks, which the CSD analysis are presented and discussed.
indicates a possible interaction between them, for
example, using oscillations at different frequencies.
Cross-References
Modeling the Spectral Structure of LFPs
The modeling of LFP signals is first considered ▶ Current Source Density (CSD) Analysis
by Biyu Jade He in the entry “▶ Electrocortico- ▶ Electrocorticogram (ECoG)
gram (ECoG).” This entry reviews models of the ▶ Local Field Potential and Deep Brain
ECoG and how this surface signal relates to the Stimulation (DBS)
LFP recorded in depth. The entry also discusses ▶ Local Field Potential and Movement Disorders
the spectral structure of these signals and their ▶ Local Field Potential in Olfaction
power-law frequency scaling. ▶ Local Field Potential in the Visual System
Low Frequency Oscillations (Anesthesia and Sleep): Overview 55
▶ Local Field Potential Interaction with the to periodicity coding in primary auditory cortex.
Extracellular Medium J Neurophysiol 73:227–245
Elul R (1971) The genesis of the EEG. Int Rev Neurobiol
▶ Local Field Potential, Ephaptic Interactions 15:227–272
▶ Local Field Potential, Methods of Recording Goldberg JA, Rokni U, Boraud T, Vaadia E, Bergman
▶ Local Field Potential, Relationship to BOLD H (2004) Spike synchronization in the cortex/basal-
Signal ganglia networks of Parkinsonian primates reflects
global dynamics of the local field potential.
▶ Local Field Potential, Relationship to J Neurosci 24(26):6003–6010
Electroencephalogram (EEG) and Gray CM, Singer W (1989) Stimulus-specific neuronal
Magnetoencephalogram (MEG) oscillations in orientation columns of cat visual cortex.
▶ Local Field Potential, Relationship to Proc Natl Acad Sci U S A 86:1698–1702
Klee M, Rall W (1977) Computed potentials of cortically
Membrane Synaptic Potentials arranged populations of neurons. J Neurophysiol
▶ Local Field Potential, Relationship to Unit 40:647–666
Activity Lass Y (1968) A quantitative approach to the correlation
▶ Local Field Potential, Synchrony of of slow wave and unit electrical activity in the
cerebral cortex of the cat. Electroencephalogr Clin
▶ Local Field Potentials (LFP) Neurophysiol 25:503–506
▶ Resistivity/Conductivity of Extracellular Lorente de Nó R (1947) Analysis of the distribution of
Medium action currents of nerve in volume conductors. Stud
Rockefeller Inst Med Res 132:384–477
Mitzdorf U (1985) Current source-density method and appli-
cation in cat cerebral cortex: investigation of evoked
potentials and EEG phenomena. Physiol Rev 65:37–100
References Murthy VN, Fetz EE (1996a) Oscillatory activity in sen-
sorimotor cortex of awake monkeys: synchronization
Abeles M (1982) Local cortical circuits: an electrophysi- of local field potentials and relation to behavior.
ological study. In: Braitenberg V, Barlow HB, J Neurophysiol 76:3949–3967
Bullock H, Florey E, Grusser O-J, Peters A (eds). Murthy VN, Fetz EE (1996b) Synchronization of neurons
Springer, New York during local field potential oscillations in sensorimotor
Bedard C, Kroger H, Destexhe A (2004) Modeling extra- cortex of awake monkeys. J Neurophysiol 76:3968–3982
cellular field potentials and the frequency-filtering prop- Renshaw B, Forbes A, Morison BR (1940) Activity of
erties of extracellular space. Biophys J 86:1829–1842 isocortex and hippocampus: electrical studies with
Bullock TH (1999) Slow potentials in the brain: still little micro-electrodes. J Neurophysiol 3(1):74–105
understood but gradually getting analytical attention. Sanes JN, Donoghue JP (1993) Oscillations in local field
Brain Res Bull 50:315–316 potentials of the primate motor cortex during voluntary
Contreras D, Destexhe A, Sejnowski TJ, Steriade M movement. Proc Natl Acad Sci U S A 90:4470–4474
(1997) Spatiotemporal patterns of spindle oscillations
in cortex and thalamus. J Neurosci 17:1179–1196
Creutzfeldt OD, Watanabe S, Lux HD (1966) Relations
between EEG phenomena and potentials of single cor-
tical cells. II. Spontaneous and convulsoid activity.
Electroencephalogr Clin Neurophysiol 20:19–37 Low Frequency Oscillations
Destexhe A, Contreras D, Steriade M (1999) Spatiotem- (Anesthesia and Sleep): Overview
poral analysis of local field potentials and unit dis-
charges in cat cerebral cortex during natural wake
Diego Contreras
and sleep states. J Neurosci 19:4595–4608
Donoghue JP, Sanes JN, Hatsopoulos NG, Gaal G (1998) Department of Neuroscience, School of
Neural discharge and local field potential oscillations Medicine, University of Pennsylvania,
in primate motor cortex during voluntary movements. Philadelphia, PA, USA
Eccles JC (1951) Interpretation of action potentials
evoked in the cerebral cortex. J Neurophysiol
3:449–464 Synonyms
Eckhorn R, Obermueller A (1993) Single neurons are
differently involved in stimulus-specific oscillations
Delta oscillations; Relay mode; Sleep; Sleep
in cat visual cortex. Exp Brain Res 95:177–182
Eggermont JJ, Smith GM (1995) Synchrony between oscillations; Sleep spindles; Slow oscillations;
single-unit activity and local field potentials in relation Slow rhythms; Thalamic bursting
56 Low Frequency Oscillations (Anesthesia and Sleep): Overview
Definition
The transition from waking to sleep or to anes-

thesia is characterized by an increase in the
amplitude and a decrease in the frequency of the
electrical activity recorded in the electroenceph-
alogram (EEG). The spectral composition of the
EEG changes from one dominated by
low-amplitude fast frequencies in the beta
gamma range to one dominated by the frequency
ranges of slow (0.1–1 Hz), delta (1–4 Hz), and
sigma (7–15 Hz, which corresponds with sleep
spindles) oscillations (Silva and Schomer 2011).
The dramatic changes in the EEG during the
transition from waking to sleep correlate with
the deafferentation of the forebrain from the
external world and the suppression of conscious-
ness. This section describes cellular mechanisms
and computer models of these oscillatory pro-
cesses and their functional consequences. In this
overview entry, some critical points are discussed
Low Frequency Oscillations (Anesthesia and Sleep):
about the organizations of these rhythms in slow Overview, Fig. 1 A summary of the basic thalamo-
wave sleep and anesthesia. cortico-thalamic circuit: 1. Reciprocal excitatory
(glutamatergic) connections between thalamocortical
(TC) and cortical (Cx) neurons. 2. Both ascending and
descending axons leave collaterals in the reticular nucleus
Detailed Description (RE). 3. RE neurons are exclusively GABAergic and
project topographically only back into the dorsal thalamus
Background (not to cortex). The three cells used for the schematic were
recorded intracellularly in vivo and filled with biocytin
Slow sleep rhythms are generated within the vast
networks connecting thalamus and cortex
(Steriade et al. 1994; McCormick and Bal 1997). feedback inhibition combined with various
The main building blocks of thalamocortical cir- degrees of convergence and divergence are essen-
cuits (Steriade 2003; Jones 2007), which are essential to the generation, distribution, and synchroni-
tial to understanding the basic principles of rhythm zation of the rhythms that characterize sleep and
generation (Fig. 1), are (i) reciprocal excitatory anesthesia (Jones 2001).
glutamatergic connections between cortical pyra-
midal cells and thalamocortical neurons; Cellular Mechanisms
(ii) collaterals of excitatory thalamocortical and The single most important cellular mechanism
corticothalamic axons onto reticular thalamic underlying the dramatic transformation in brain
cells, which are established as these axons cross state during the transition from waking to sleep,
the reticular nucleus on their way in and out of the or to anesthesia, is the change in firing mode of
thalamus; and (iii) inhibitory connections from thalamocortical cells (Fig. 2) from tonic to bursting
reticular thalamic cells onto thalamocortical neu- (Llinas and Jahnsen 1982; Jahnsen and Llinas
rons. Reticular thalamic neurons surround the thal- 1984c; Steriade and Llinas 1988; Steriade
amus in its dorsal and lateral surfaces and do et al. 1994; McCormick and Bal 1997). At
not project to the cerebral cortex but project only depolarized membrane potentials (Vm) during acti-
back into all other thalamic nuclei (Ramon vated brain states, such as waking, thalamocortical
y Cajal 1911). The reciprocal excitation and neurons respond to inputs with spike trains that
Low Frequency Oscillations (Anesthesia and Sleep): Hyperpolarization switches firing mode from tonic to
Overview, Fig. 2 This figure shows a thalamocortical bursting in both cases. Bottom trace shows a rebound
neuron, filled with biocytin (right) in vivo, responding to burst, a burst generated at the termination of
synaptic activation (top three traces) and direct current a hyperpolarizing pulse (From Contreras and Steriade
injection (bottom four traces) through the microelectrode. 1995)
reflect the amplitude and time of the inputs The amplitude of the LTS and, therefore, of the
(Deschenes et al. 1984; McCormick and Feeser resulting burst of sodium action potentials
1990; McCormick and Huguenard 1992). The depend on the proportion of T-channels that are
membrane hyperpolarization that results from the de-inactivated (Deschenes et al. 1982, 1984; Coul-
reduction in cholinergic and adrenergic input dur- ter et al. 1989; McCormick and Huguenard 1992),
ing the transition from waking to sleep (Oakson which reaches 100 % at the hyperpolarized Vm of
and Steriade 1982; Steriade et al. 1982, 1990; Lu late stages of sleep and deep anesthesia.
et al. 1993; Steriade 1993) removes inactivation of The mechanism by which LTS bursts are
T-type calcium channels (Jahnsen and Llinas triggered changes throughout the progressive
1984a; Nowycky et al. 1985; Coulter et al. 1989; stages of sleep. This is due to the progressive
McCormick and Huguenard 1992; Gutierrez hyperpolarization of thalamocortical cells as
et al. 2001). When fully de-inactivated, sleep (Hirsch et al. 1983) or anesthesia
T-channels are capable of generating an all-or- (Contreras and Steriade 1997a) deepens and
none calcium spike, known as a low-threshold EEG activity becomes larger and slower
spike (LTS) because it is triggered at thresholds (Fig. 3). At early sleep stages, when sleep spin-
10 mV lower than the sodium action potential dles are the most conspicuous feature of the
(Deschenes et al. 1982; Llinas and Jahnsen 1982). EEG, thalamocortical cells are still relatively
The LTS in turn generates sufficient depolarization depolarized and T-channels are not fully
to trigger a high-frequency (100–300 Hz) burst of de-inactivated. De-inactivation depends on the
sodium action potentials (Llinas and Jahnsen 1982; phasic strong chloride-dependent GABAA
Deschenes et al. 1984; Jahnsen and Llinas 1984b). IPSPs generated by the inhibitory input from
Low Frequency Oscillations (Anesthesia and Sleep): mode of firing is associated with the presence of global
Overview, Fig. 3 The transition from waking to sleep is slow oscillations that are synchronized and widespread
characterized in the EEG by a transition from high- over large cortical and thalamic territories (the figure
frequency (~40 Hz) low-amplitude activity to shows how activity is synchronized only during sleep
low-frequency (<15 Hz) high-amplitude oscillations. TC between areas 4 and 5, for example). The high-amplitude,
cells recorded simultaneously with the EEG show the synchronized oscillatory activity together with the tha-
transition from tonic, relay mode of firing during the lamic bursting mode of firing disconnect the cortex from
waking state to the bursting, all-or-none mode of firing, the outside world during sleep (From Contreras and
incompatible with information transfer to the cerebral Steridade 1997a)
cortex. In addition, as shown in the figure, the bursting
reticular thalamic cells during sleep spindles oscillations, the EEG is also dominated by
(von Krosigk et al. 1993; Bal et al. 1995; large amplitude rhythmic waves at the delta
McCormick and Bal 1997). During spindles, frequency. However, it is not clear what the
LTS bursts are generated as rebound bursts at relationship is, if any, between the intrinsically
the offset of these large GABAA IPSPs. At generated delta oscillations of thalamocortical
later, deeper (defined by the higher threshold cells and the cortical generators of delta oscil-
to wake up) stages of sleep, thalamocortical lations when this rhythm is the predominant
cells are further hyperpolarized and T-channels feature of the EEG (Amzica and Steriade
are fully de-inactivated. At this stage, 1998; Silva and Schomer 2011; Carracedo
T-channels are activated by the depolarization et al. 2013).
caused by the cationic inward rectifier current A fundamental consequence of the progres-
IH (McCormick and Pape 1990). IH is voltage sive hyperpolarization of the Vm of
dependent and activated at the hyperpolarized thalamocortical cells throughout sleep (Fig. 3)
levels of deep sleep and deep anesthesia. is that, at the deep sleep stages when delta
This IH-dependent depolarization, which is oscillations dominate the EEG (Fig. 3), the Vm
a ubiquitous property of thalamocortical cells is at or below the reversal potential for chloride,
(Dossi et al. 1992), triggers an LTS and and therefore sleep spindles cannot occur.
a spike burst, at the end of which, the mem- This leads to an incompatibility of sleep and
brane returns to its hyperpolarized level and delta rhythms in thalamocortical cells, with
reactivates IH, thus restarting a rhythmic sleep spindles characterizing early stages
cycle at the delta frequency range (McCormick when the Vm is still relatively depolarized and
and Pape 1990; Dossi et al. 1992; Nunez delta oscillations characterizing late stages
et al. 1992a). At the deep stage of sleep when the Vm is very hyperpolarized (Nunez
where thalamocortical cells generate delta et al. 1992b).
Functional Consequences of Slow Rhythms state are transformed into networks that are
Three main consequences result from the transi- slow and unreliable in responding and
tion in firing mode from tonic to bursting during are engaged mostly in slow and synchronized
the transition from waking to sleep or anesthesia: oscillations. These oscillations are incompatible
1. Unreliability of Responses. Rather than reli- with information processing and ultimately
able and predictable responses to sensory, determine the deafferentation of the forebrain
motor, or cortical input, thalamic cells respond during slow wave sleep and anesthesia
to inputs with all-or-none bursts of spikes (Steriade 2000).
which bear little information about the incom-
ing input. Furthermore, because voltage- Anesthesia and Slow Waves
dependent inactivation of T-channels lasts In contrast with the organized and sequential
tens of milliseconds, bursting responses to arrangement of oscillations during sleep, anes-
depolarizing inputs can only occur at low fre- thetics mimic and exaggerate one or more
quencies and therefore represent a strong fre- rhythmic patterns of sleep in different propor-
quency filter to incoming inputs (McCormick tions. For example, during barbiturate anesthe-
and Feeser 1990). sia, the EEG shows spindles most prominently
2. Rhythmicity. The transition to burst firing and virtually no delta or slow oscillations. Fur-
brings about slow rhythmic activity in thermore, under barbiturate anesthesia, spindle
thalamocortical networks. Three main slow frequency is reduced and duration increased
rhythms characterize sleep and anesthesia: with respect to naturally occurring sleep spin-
sleep spindles (7–15 Hz), delta oscillations dles. In contrast, ketamine-xylazine generates
(1–4 Hz), and slow oscillations (0.1–1 Hz). a strong and exaggerated slow oscillatory pat-
Such oscillatory activity requires participation tern on top of which rapid spindles may be
of burst firing by thalamocortical cells and triggered. In addition, the slow rhythm during
therefore extends the role of thalamic burst ketamine-xylazine may reach frequencies above
firing to the deafferentation of the forebrain. 1 Hz and may be confused with EEG delta
Indeed, in addition to the bursting properties waves. Urethane anesthesia generates a very
described above, most thalamic spike bursts stable pattern of slow oscillations with
occur NOT in response to stimuli but, instead, prolonged depolarizing and hyperpolarizing
as part of large oscillatory networks, further phases but very little spindle or delta oscilla-
hindering thalamocortical relay of inputs tions. The combination of propofol and fentanyl
(Steriade et al. 1994; Steriade 2000). produces a rich combination of rhythms includ-
3. Synchronization. An inevitable consequence ing slow oscillations, occasional spindles, and
of thalamic burst firing and slow rhythm gen- delta waves. Detailed descriptions of oscillatory
eration is the broad synchronization of patterns under different anesthetics are far
thalamocortical networks (Contreras and beyond this section.
Steriade 1997a, b). By virtue of the divergence
and convergence of connections between
reticular thalamic and thalamocortical cells, Cross-References
between thalamocortical and cortical neurons
and within cortical networks, the strong burst- ▶ Corticothalamic Feedback: Large-Scale
ing activity entrains large networks into the Synchrony
slow oscillations of sleep and anesthesia ▶ Delta Rhythms: Models and Physiology
(Amzica and Steriade 1995). ▶ Slow Oscillations and Epilepsy: Network
As a result of the above three factors, Models
the reliably responding, rapidly engaged ▶ Slow Oscillations: Models
thalamocortical networks that sustain informa- ▶ Slow Oscillations: Physiology
tion processing and consciousness in the awake ▶ Spindle Oscillations: Models
References Jahnsen H, Llinas R (1984b) Voltage-dependent burst-to-

tonic switching of thalamic cell activity: an in vitro
Amzica F, Steriade M (1995) Short- and long-range neu- study. Arch Ital Biol 122:73–82
ronal synchronization of the slow (<1 Hz) cortical Jahnsen H, Llinas R (1984c) Electrophysiological proper-
oscillation. J Neurophysiol 73:20–38 ties of guinea-pig thalamic neurones: an in vitro study.
Amzica F, Steriade M (1998) Electrophysiological corre- J Physiol 349:205–226. PMCID: PMC1199334
lates of sleep delta waves. Electroencephalogr Clin Jones EG (2001) The thalamic matrix and thalamocortical
Neurophysiol 107:69–83 synchrony. Trends Neurosci 24:595–601
Bal T, von Krosigk M, McCormick DA (1995) Role of the Jones EG (2007) The thalamus. Cambridge University
ferret perigeniculate nucleus in the generation of syn- Press, Cambridge
chronized oscillations in vitro. J Physiol Llinas R, Jahnsen H (1982) Electrophysiology of mam-
483(Pt 3):665–685. PMCID: PMC1157809 malian thalamic neurones in vitro. Nature
Carracedo LM, Kjeldsen H, Cunnington L, Jenkins A, 297:406–408
Schofield I, Cunningham MO, Davies CH, Traub RD, Lu SM, Guido W, Sherman SM (1993) The brain-stem
Whittington MA (2013) A neocortical delta rhythm parabrachial region controls mode of response to
facilitates reciprocal interlaminar interactions via visual stimulation of neurons in the cat’s lateral genic-
nested theta rhythms. J Neurosci 33:10750–10761. ulate nucleus. Vis Neurosci 10:631–642
PMCID: PMC3693056 McCormick DA, Bal T (1997) Sleep and arousal:
Contreras D, Steriade M (1995) Cellular basis of EEG thalamocortical mechanisms. Annu Rev Neurosci
slow rhythms: a study of dynamic corticothalamic 20:185–215
relationships. J Neurosci 15:604–622 McCormick DA, Feeser HR (1990) Functional implica-
Contreras D, Steriade M (1997a) State-dependent fluctu- tions of burst firing and single spike activity in lateral
ations of low-frequency rhythms in corticothalamic geniculate relay neurons. Neuroscience 39:103–113
networks. Neuroscience 76:25–38 McCormick DA, Huguenard JR (1992) A model of the
Contreras D, Steriade M (1997b) Synchronization of electrophysiological properties of thalamocortical
low-frequency rhythms in corticothalamic networks. relay neurons. J Neurophysiol 68:1384–1400
Neuroscience 76:11–24 McCormick DA, Pape HC (1990) Properties of a -
Coulter DA, Huguenard JR, Prince DA (1989) Calcium hyperpolarization-activated cation current and its role
currents in rat thalamocortical relay neurones: kinetic in rhythmic oscillation in thalamic relay neurones.
properties of the transient, low-threshold current. J Physiol 431:291–318. PMCID: PMC1181775
J Physiol 414:587–604. PMCID: PMC1189159 Nowycky MC, Fox AP, Tsien RW (1985) Three types of
da Silva FL, Schomer DL (eds) (2011) Niedermeyer’s neuronal calcium channel with different calcium ago-
Electroencephalography: basic principles, clinical nist sensitivity. Nature 316:440–443
applications, and related fields. Lippincott Williams Nunez A, Amzica F, Steriade M (1992a) Intrinsic and
& Wilkins, Philadelphia synaptically generated delta (1–4 Hz) rhythms in dor-
Deschenes M, Roy JP, Steriade M (1982) Thalamic sal lateral geniculate neurons and their modulation by
bursting mechanism: an inward slow current revealed light-induced fast (30–70 Hz) events. Neuroscience
by membrane hyperpolarization. Brain Res 51:269–284
239:289–293 Nunez A, Curro Dossi R, Contreras D, Steriade M (1992b)
Deschenes M, Paradis M, Roy JP, Steriade M (1984) Intracellular evidence for incompatibility between
Electrophysiology of neurons of lateral thalamic spindle and delta oscillations in thalamocortical neu-
nuclei in cat: resting properties and burst discharges. rons of cat. Neuroscience 48:75–85
J Neurophysiol 51:1196–1219 Oakson G, Steriade M (1982) Slow rhythmic rate fluctu-
Dossi RC, Nunez A, Steriade M (1992) Electrophysiology ations of cat midbrain reticular neurons in synchro-
of a slow (0.5–4 Hz) intrinsic oscillation of cat nized sleep and waking. Brain Res 247:277–288
thalamocortical neurones in vivo. J Physiol Ramon y Cajal S (1911) Histologie du systeme nerveux de
447:215–234. PMCID: PMC1176033 l’homme et des vertebres. Malione, Paris
Gutierrez C, Cox C, Rinzel J, Sherman S (2001) Dynamics Steriade M (1993) Cholinergic blockage of network- and
of low-threshold spike activation in relay neurons of intrinsically generated slow oscillations promotes
the cat lateral geniculate nucleus. J Neurosci waking and REM sleep activity patterns in thalamic
21:1022–1032 and cortical neurons. Prog Brain Res 98:345–355
Hirsch JC, Fourment A, Marc ME (1983) Sleep-related Steriade M (2000) Corticothalamic resonance, states of
variations of membrane potential in the lateral genic- vigilance and mentation. Neuroscience 101:243–276
ulate body relay neurons of the cat. Brain Res Steriade M (2003) Thalamus. Wiley
259:308–312 Steriade M, Llinas RR (1988) The functional states of the
Jahnsen H, Llinas R (1984a) Ionic basis for the electro- thalamus and the associated neuronal interplay.
responsiveness and oscillatory properties of guinea- Physiol Rev 68:649–742
pig thalamic neurones in vitro. J Physiol Steriade M, Oakson G, Ropert N (1982) Firing rates and
349:227–247. PMCID: PMC1199335 patterns of midbrain reticular neurons during steady
Model Reproducibility: Overview 61
and transitional states of the sleep-waking cycle. Exp • Replicability, where the same code and tools
Brain Res 46:37–51 are used to reproduce results
Steriade M, Datta S, Pare D, Oakson G, Curro Dossi RC
(1990) Neuronal activities in brain-stem • Cross-replicability, where the same model is
cholinergic nuclei related to tonic activation used but implemented using different software
processes in thalamocortical systems. J Neurosci • Reproducibility, where a new model is
10:2541–2559 implemented and produces the same scientific
Steriade M, Contreras D, Amzica F (1994) Synchronized
sleep oscillations and their paroxysmal developments. result
Trends Neurosci 17:199–208 As noted in Crook et al. (2013), the boundary
von Krosigk M, Bal T, McCormick DA (1993) Cellular between cross-replicability and reproducibility is
mechanisms of a synchronized oscillation in the thal- not always clear, but all efforts along this contin-
amus. Science 261:361–364
uum of reproducible research contribute to
advances in the computational sciences.
Code Sharing
Model Reproducibility: Overview Code sharing provides a means for promoting
replicability and transparency in the computa-
Sharon Crook tional sciences. For many reasons, the best infra-
School of Mathematical and Statistical Sciences structure for code sharing involves curated model
and School of Life Sciences, Arizona State repositories such as ▶ ModelDB (Migliore
University, Tempe, AZ, USA et al. 2003), the ▶ BioModels Database: a Public
Repository for Sharing Models of Biological Pro-
cesses (Li et al. 2010; Vijayalakshmi et al. 2013),
Definition the ▶ Physiome Repository (Yu et al. 2011), and
▶ Open Source Brain (Gleeson et al. 2010b).
The ability to reproduce an experimental result is
the foundation of scientific inquiry; however, Structured, Declarative Descriptions
computational scientists find it difficult to repro- of Models and Simulations
duce many published results. Here we provide an There are a number of ongoing efforts in support
overview of efforts to support model reproduc- of cross-replicability with the aim of providing
ibility in computational neuroscience. declarative descriptions of models that are simu-
lator independent. Many of these efforts focus on
the use of Extensible Markup Language (XML)
Detailed Description technology (Bray et al. 1998) as an ideal repre-
sentation for complex models.
Reproducing the simulation results of computa- The ▶ Systems Biology Markup Language
tional models and establishing the provenance of (SBML) (Hucka et al. 2003), and ▶ CellML
results should be straightforward given that (Hedley et al. 2000) are two popular languages
computational studies do not suffer from the for describing systems of interacting biomole-
measurement errors seen in the experimental cules that comprise models that are often used
sciences. However, computational science has in systems biology. Both languages also can be
its own challenges for reproducibility, which are used for describing more generic dynamical
described well by Crook et al. (2013). In particu- models, including those in neuroscience.
lar, issues such as the sensitivity of a model to ▶ FieldML follows a similar approach but
numerics or the publication of models that are focuses on multivariate spatiotemporal models
computationally under-specified lead to the need and models based on the finite element method.
for criteria for successful model reproduction in ▶ NeuroML (Goddard et al. 2001; Crook
many cases. These authors also make distinctions et al. 2007; Gleeson et al. 2010a) differs from
among: these languages in that it is domain specific,
62 Model Reproducibility: Overview
where neuroscience concepts like cells, ion chan- ▶ Simulation Experiment Description Markup
nels, and synapses form the basis for objects Language (SED-ML)
described in the language. ▶ NineML is another ▶ Systems Biology Markup Language (SBML)
domain-specific language that focuses on
descriptions of spiking neural networks. Addi-
tionally, the ▶ Simulation Experiment Descrip- References
tion Markup Language (SED-ML), (Köhn and Le
Novère 2008), provides a language for describing Bray T, Paoli J, Sperberg-McQueen C (1998) Extensible
the details of simulation experiments. These dif- markup language (XML) 1.0. http://www.w3.org/TR/
REC-xml
ferent markup languages are complementary, and Crook S, Gleeson P, Howell F, Svitak J, Silver RA
together they cover the different spatial scales for (2007) MorphML: level 1 of the NeuroML standards
the majority of models in neuroscience. for neuronal morphology data and model specification.
Neuroinformatics 5:96–104
Crook S, Davison AP, Plesser HE (2013) Learning from the
Improving Research Reporting past: approaches for reproducibility in computational
In the systems biology community, MIRIAM neuroscience. In: Bower JM (ed) 20 Years of computa-
(Minimum Information Required in the Annota- tional neuroscience, vol 9, Springer series in computa-
tion of Models, http://co.mbine.org/standards/ tional neuroscience. Springer, New York, pp 73–102
Gleeson P, Crook S, Cannon RC, Hines ML, Billings GO,
miriam) is a community-level effort to provide Farinella M, Morse TM, Davison AP, Ray S, Bhalla
a set of guidelines for use with any structured US, Barnes SR, Dimitrova YD, Silver RA (2010a)
format for sharing models. The idea of using NeuroML: a language for describing data driven
best practices and some required metadata for models of neurons and networks with a high degree
of biological detail. PLoS Comput Biol 6(6):e1000815
sharing and publishing models is outlined by Gleeson P, Piasini E, Crook S, Cannon R, Steuber V,
Novère et al. (2005). All of these requirements Jaeger D, Solinas S, D’Angelo E, Silver RA (2010b)
can be adapted readily to models in neuroscience. The open source brain initiative: enabling collabora-
Nordlie et al. (2009) conducted a survey of neu- tive modelling in computational neuroscience. BMC
ronal network models in the literature and found Goddard N, Hucka M, Howell F, Cornelis H, Shankar K,
that current approaches are diverse and inade- Beeman D (2001) NeuroML: model description
quate. These authors propose the adoption of methods for collaborative modeling in neuroscience.
best practices for model descriptions in publica- Philos Trans R Soc Lond B Biol Sci 356:1209–1228
Lloyd CM, Halstead MDB, Nielsen PF (2004) CellML: its
tions, recommending the inclusion of the hypoth- future, present and past. Progress in Biophysics and
esis, model derivation, model description, Molecular Biology 85(2–3):433–450
implementation details, model analysis, and Hucka M, Finney A, Sauro H, Bolouri H, Doyle J,
model justification. They also provide a concise Kitano H, Arkin A (2003) The systems biology
markup language (SMBL): a medium for representa-
tabular format for summarizing network models tion and exchange of biochemical network models.
in publications. Publication standards such as Bioinformatics 19:524–531
those discussed by Le Novère et al. and Nordlie Köhn D, Le Novère N (2008) SED-ML – an XML format
et al. ensure that all possible model details are for the implementation of the MIASE guidelines. In:
Heiner M, Uhrmacher A (eds) Computational methods
provided; however, it is important to be aware of in systems biology, vol 5307, Lecture notes in com-
the limits of replicability and the impact on puter science. Springer, Berlin, pp 176–190
reproducibility. Le Novère N, Finney A, Hucka M, Bhalla US,
Campagne F, Callado-Vides J, Crampin E, Halstead
M et al (2005) Minimum information requested in the
annotation of biochemical models (MIRIAM). Nat
Cross-References Biotechnol 23(12):1509–1515
Li C, Donizelli M, Rodriguez N, Dharuri H, Endler L,
▶ CellML Chelliah V, Li L, He E, Henry A, Stefan MI, Snoep JL,
Hucka M, Le Novere N, Laibe C (2010) BioModels
▶ FieldML database: an enhanced, curated and annotated resource
▶ NeuroML for published quantitative kinetic models. BMC Syst
▶ NineML Biol 4:92
Modeling of Disease - Molecular Level: Overview 63
Migliore M, Morse TM, Davison AP, Marenco L, Shep- Detailed Description

herd GM, Hines M (2003) ModelDB.
Neuroinformatics 1(1):135–139
Nordlie E, Gewaltig MO, Plesser HE (2009) Towards This section includes entries of three types:
reproducible descriptions of neuronal network models. Techniques: These simulations are performed
PLoS Comput Biol 5(8):e1000456 largely using techniques used elsewhere in
Vijayalakshmi C, Laibe C, Le Novere N (2013) Biomodels computational neuroscience. However, some
database: a repository of mathematical models of bio-
logical processes. Methods Mol Biol 1021:189–199 techniques have been introduced or expanded
Yu T, Lloyd CM, Nickerson DP, Cooling MT, Miller AK, in the context of their use in translational
Garny A, Terk-ildsen JR, Lawson J, Britten RD, studies.
Hunter PJ, Nielsen PM (2011) The physiome model Locations: Partitioning of the nervous system
repository 2. Bioinformatics 27(5):743–744
can be done at many scales, from organelle
to major brain area.
Diseases: We make note of a few neurological
Modeling of Disease - Molecular and psychiatric illnesses that have been exam-
Level: Overview ined at this scale.
William W. Lytton Techniques

Departments of Physiology & Pharmacology and Standard techniques of molecular and cellular
Neurology, SUNY Downstate Medical Center, computational neuroscience are utilized for clin-
Brooklyn, NY, USA ical and translational modeling. The techniques
Department of Neurology, Kings County that are applicable to simulation at the molecular
Hospital, Brooklyn, NY, USA and cellular level are listed here by their coverage
in entries of this encyclopedia (loosely grouped
from micro to macro):
Definition ▶ Stochastic Simulators
▶ Gillespie Algorithm for Biochemical Reaction
Though molecular and cellular level modeling is Simulation
the first step in the long ladder of multiscale ▶ Deterministic Reaction-Diffusion Simulators
modeling for clinical translational application, it ▶ Bimolecular Reactions, Modeling of
may be regarded as the critical scale: “Where the ▶ Enzyme Kinetics, Modeling of
rubber meets the road.” The idiom from automo- ▶ Modeling Ion Concentrations
tive engineering refers to the critical endpoint ▶ Cable Equation
where all the ancillary engineering – fuel system, ▶ Equivalent Cylinder Model (Rall)
chassis, pistons, etc. – is finally tested. In the ▶ Modeling Synapses
realm of translation, this point comes when we ▶ Axon, Modeling
translate basic research results to the clinical set- ▶ Numerical Integration Methods
ting. The physician (though not the surgeon) ▶ Action Potential Initiation
intervenes at the molecular scale, using pharma- In this section, we add two additional methods
cological agents to alter physiological activity that have been applied to clinical translational
only observed at far higher scales: the macro- problems.
scopic realms of clinical tests and of signs and “▶ Polypeptide and Protein Modeling for
symptoms. Given the vast scale gap between Drug Design” takes us down to the lowest of the
molecular treatments and macroscopic outcome, multiple scales used in computational neurosci-
there is enormous difficulty in understanding ence. Polypeptide modeling largely utilizes ball-
causal relations, so as to design better drugs and and-spring modeling, staying above the level of
better drug combinations. Multiscale modeling quantum mechanical representations. Individual
can help make these connections in order to pro- atoms or atomic groups (e.g., amino acids) are
vide a basis for rational pharmacotherapeutics. represented as moving in a force field largely
64 Modeling of Disease - Molecular Level: Overview
determined by the chemical bonds connecting are covered elsewhere in these volumes: see
them. This research has direct importance for “▶ Synaptic Dynamics: Overview,” and
clinical translation because the lock-and-key fit “▶ Dendritic Spines.” It is to be expected that
of a ligand (a drug) to a target is dependent on the disorders of spine shape, and of proteins involved
physical relations of the interacting species. in synaptic plasticity, will produce neurological
“▶ Application of Declarative Programming or psychiatric disorders. In this section, we con-
in Neurobiology” demonstrates the novel use of sider the modeling of pharmacotherapeutic
a Declarative Programming Language to address manipulation of ion channels in “▶ Neurophar-
problems in Alzheimer’s disease and in fear con- macological Modeling, Pharmacogenomics and
ditioning. Declarative Programming is distin- Ion Channel Modulation.”
guished from the familiar Imperative A literal route to clinical application is
Programming languages which we commonly described here in “▶ Brain Extracellular Space:
utilize. Imperative programs implement an algo- a Compartment for Intercellular Communication
rithm. Declarative programs describes relation- and Drug Delivery.” The blood–brain barrier and
ships between elements, defining what the extracellular space represent the major path-
computations are possible amount these ways for delivery of pharmacological agents.
elements. Due to therapeutic index (therapeutic concentra-
tion/toxic concentration ratio), it is important to
Locations consider how a particular drug can be delivered
Disease strikes at many locations and at many so as to arrive at adequate concentrations at the
scales. Parsimony and clinical efficacy depend site of disease, without being present at excessive
on finding the proper scale and proper locus for concentration at other locations.
clinical investigation and for clinical At the cellular level, neurological disease has
intervention – in some cases these are not the substantial overlap with disease in other excitable
same location. A recent example from oncology tissues. It is particularly valuable to consider the
is illustrative: a renal cancer and a leukemia were extensive work that has been done in cardiology,
previously classified according to organ system here represented by “▶ Cardiac Excitable Tissue
and therefore treated by different subspecialists Pathology (Ion Channels)” and “▶ Cardiac Excit-
with different therapeutic approaches. These can- able Tissue Pathology (Ischemia).” Cardiology
cers have now been found to share the same models are generally more mature than those in
tumor mutation and to benefit from similar treat- neurology. They also have direct analogies to
ments. As in this oncology example, protean neu- diseases of the brain: e.g., myocardial infarction
rological diseases will also express widely. For and stroke; arrhythmias and epilepsies. Similarly,
example, diseases of mitochondria such as excitable tissue is important in the islets of the
MELAS (mitochondrial encephalomyopathy, lac- pancreas and in other endocrine systems, where
tic acidosis, and stroke-like episodes) and production and release of hormone is directly
MERRF (myoclonic epilepsy with ragged red analogous to production and release of neuro-
fibers; the ragged fibers being broken mitochon- transmitters: see entry on “▶ Endocrine Cell
dria) produce neurological disorders across many Function and Dysfunction.” A broader physio-
neural systems, beyond the already broad scope logical perspective, relevant to many of these
of their acronyms. cell types, is presented here in “▶ Neuropharma-
In addition to connoting locations in the brain cological Modeling, Alterations in Ionic
and body (macroscale), disease localization also Homeostasis.”
needs to consider the locus in the cell: particular Many neurological diseases present outside of
organelles such as mitochondria, particular clas- the brain or involve the brain in its association
ses of molecules such as ion channels, and par- with the body. In particular, the enteric nervous
ticular specialized subcellular spaces such as system (ENS) and autonomic nervous system
synapses and spines. Some of these locations (ANS) are prone to disorders and will be
Modeling of Disease - Molecular Level: Overview 65
important areas for modeling in the future. We highlight here four diseases, although there
In the present section, we feature an entry on are several others where some cellular/molecular
“▶ Pathological Changes in Peripheral Nerve modeling progress has been made, and many
Excitability,” to consider disorders in the periph- more where such progress is greatly needed.
eral nervous system (PNS). It is important to note A model of molecular dependencies in
that myopathies (muscle diseases) also fall within Alzheimer’s is presented in “▶ Application of
the bailiwick of neurology and will benefit from Declarative Programming in Neurobiology.”
modeling, some of which can be borrowed from As a lens through which to view computa-
the cardiology literature. tional neuroscience, disease modeling presents
At the subcellular level, signaling occurs a number of difficulties. First, diseases are gen-
through second and third through nth messengers. erally multiorgan, both in production and in
There has been considerable modeling at this expression. For example, we present here an
scale, but not as much looking at dysfunction. entry on “▶ Brain Ischemia and Stroke.” Stroke
These topics are covered in detail in the section is most often caused by atherosclerosis, a disease
on “▶ Biochemical Signaling Pathways and Dif- of blood vessels. A translational model would
fusion: Overview.” Particular note should be ideally include the cellular changes of the arterial
made of the major role of calcium as a second wall and the changes in hemodynamics, along
messenger, covered in eight entries in that sec- with factors that include the blood–brain barrier
tion. Certainly, dysregulation of calcium control (which breaks) and the role of a broad dispersion
will play a major role in the production and of harmful neurohumoral agents. Stroke frequently
expression of disease. co-occurs with myocardial infarction (MI). These
All disease is to a greater or lesser extent mul- coincident disorders, in brain, heart, and local
tifactorial. Patient susceptibility is determined in microvasculature, affect one another, e.g., MI
substantial part by the genome and its control, as reduces cardiac output and therefore increasing
well as by the epigenome and metagenome, the extent of an incipient stroke. Another
representing additional genomes that are involved difficulty in providing a unified model of
(mitochondrial, bacterial – commensal or patho- a disease is that the disease may involve many
logical, viral). In a neuron, communication disparate variants: hemorrhagic vs bland stroke,
between genome and synapse is particularly com- thrombotic vs embolic stroke, and cortical stroke
plex, insofar as this is information flow between (grey matter) vs internal capsule (white matter)
the organism’s two primary loci of information stroke.
storage. Synaptic changes depend on elaboration, As mentioned in the “Definition” section
transport, and insertion of membrane proteins, above, modeling at molecular and cellular levels
which start with signals at the synapse that trigger is only the lowest rung in the multiscale ladder of
changes in transcription. In the present section, we both downward and upward causality. Disparate
cover the role of second messengers in providing viewpoints at different scales represent
communication with the cell nucleus: transcrip- a challenge but are a necessary stage in the devel-
tional control dysfunction, modeling, and evaluat- opment of future unified multiscale views, as can
ing potential pathology in the linkage between the be illustrated by the case of epilepsy – contrast
cell and its nucleus at the molecular and organelle the entry in the section: “▶ Epilepsy: Abnormal
scales. Ion Channels” with the perspective presented in
entries on “▶ Epilepsy: Computational Models,”
Diseases “▶ Slow Oscillations, and Epilepsy: Network
Evaluation of individual diseases is the natural Models,” “▶ Epilepsy, Neural Population
approach for clinical translation – as an end goal Models of.”
we want to know how to treat a particular disease How are we to build multiscale models out of
or even a particular patient (personalized medi- these disparate pieces? Ideally it would be possi-
cine) or subgroup of patients (precision medicine). ble to embed one scale within the model for the
66 Motoneurons and Neuromuscular Systems: Overview
higher scale, thereby creating a single unified Cross-References

multiscale model, In many cases this is not
possible, so that the models remain separate, ▶ Application of Declarative Programming in
with each level informing the other via emergent Neurobiology
properties discovered at the lower level. In the ▶ Brain Extracellular Space: a Compartment for
case of epilepsy research, emergences from the Intercellular Communication and Drug
effect of a drug on a detailed model of an ion Delivery
channel can be used to change the parameteriza- ▶ Brain Ischemia and Stroke
tions of an ion channel, which could then be ▶ Cardiac Excitable Tissue Pathology (Ion
directly embedded in a compartmental model. Channels)
Emergences from the compartment model can ▶ Cardiac Excitable Tissue Pathology (Ischemia)
then be used in a reduced model or to modify an ▶ Endocrine Cell Function and Dysfunction
integrate-and-fire model. Either of these models ▶ Epilepsy: Abnormal Ion Channels
(or both as in a hybrid network) can be directly ▶ Neuropharmacological Modeling Alterations
embedded to produce large network models. in Ionic Homeostasis
Emergences from the network can then be used ▶ Neuropharmacological Modeling,
to modify a mean-field model. While this upward Pharmacogenomics and Ion Channel
sequence is difficult, the downward sequence is Modulation
even more difficult: we want to follow the scales ▶ Pathological Changes in Peripheral Nerve
back down in order to apply lessons learned Excitability
from the population model in order to eventually ▶ Polypeptide and Protein Modeling for Drug
suggest changes to drug effect at the polypeptide Design
level. ▶ Transcriptional Control Dysfunction,
Modeling
Conclusions
These initial efforts aimed at understanding the

base scale of molecular and cellular interactions Motoneurons and Neuromuscular
for understanding brain disease should be con- Systems: Overview
sidered in the context of the wider field of
computational systems biology, a field Sharmila Venugopal
concerned with multiscale modeling across Department of Physiology, David Geffen School
organ systems and medical specialties. Systems of Medicine, University of California Los
biology has been catapulted forward with the Angeles, Los Angeles, CA, USA
realization that the massive amounts of data
being collected as genomes, proteomes, and
other omes will not be interpretable without Definition
contexts, contexts that can only be provided
by sophisticated computational models. Further Motoneuron models are mathematical represen-
future hope springs from the potential of com- tations of motoneuron structure and function.
puter simulation to provide personalized or pre- Neuromuscular models represent computational
cision medicine; physiological data from models of neural components integrated with
a particular patient would be used to adapt muscle models at various levels of detail and
simulation parameters for simulation, just as complexities. These models provide a computa-
anatomical data for a particular patient is cur- tional framework for investigating neural control
rently used to provide specifics of beam of movement. Together, motoneuron and neuro-
targeting in radiation therapy. muscular models can guide the development of
Motoneurons and Neuromuscular Systems: Overview 67
biomimetic neuromuscular control systems and models (▶ Algorithmic Reconstruction of

neural prostheses. Motoneuron Morphology), more complex multi-
compartment models (▶ Brainstem Motoneu-
rons, Models of), and morphologically realistic
Detailed Description models (▶ Compartmental Models of Spinal
Motoneurons). Among vertebrate motoneurons,
Motor neurons (or motoneurons) are the we explicitly address models of spinal
final common pathway for all neural drive to the (▶ Brainstem Motoneurons, Models of) and
skeletal musculature. Each motoneuron inner- brainstem motoneurons (▶ Computational
vates one or more muscle fibers and the Models of Motor Pools). We further describe
cell bodies of motoneurons are somatotopically novel algorithms that can generate realistic moto-
organized in both vertebrates and invertebrates. neuron morphologies given a simple set of exper-
Motoneurons were among the first neurons to imentally derived morphometric parameters
be studied in great detail by experimentalists (▶ Conductance-Based Models of Nonlinear
largely due to their unique position as output Dynamics in Vertebrate Motoneurons). At the
neurons of the nervous system. The rich tradition population level, we describe motor pool models
of the experimental work dates back to the (▶ Mammalian Motor Nerve Fibers, Models of).
early 1900s beginning with the historical work We also describe the computational principles
by Sir John Eccles (1903–1997) and Sir underlying models of the muscles
Charles Sherrington (1857–1952), followed by (▶ Morphologically Detailed Motoneuron
a host of physiologists to date whose work has Models). Since motor fibers traverse long dis-
laid the foundations for realistic models of tances to provide electrochemical signals for
motoneurons. these effectors, models of motor nerve fibers
Like many neurons in the nervous system, and propagation of electrical signals along them
motoneurons show repetitive firing behavior in are described (▶ Neuromuscular Control Sys-
response to excitation. However, like no other tems, Models of). Lastly, systems-level models
neurons in the nervous system, their frequency of the neural control systems that enable us to
of spike discharge directly dictates the degree of perform tasks by providing suitable input to
muscle contraction. In this way, they mediate motoneurons are described (▶ Physiology and
control of complex behaviors. At the cellular Computational Principles of Muscle Force
level, they possess complex morphologies that Generation).
strongly influence their integrative function.
They house a rich assortment of intrinsic
membrane-bound ion channels that dynamically Cross-References
interact with a wide range of synaptic inputs to
produce unique nonlinearities in their membrane ▶ Algorithmic Reconstruction of Motoneuron
properties. The heterogeneity of cellular proper- Morphology
ties within a motor pool further confers complex ▶ Brainstem Motoneurons, Models of
population dynamics. Taken together, motoneu- ▶ Compartmental Models of Spinal
rons are an important class of neurons in the Motoneurons
nervous system that are attractive candidates for ▶ Computational Models of Motor Pools
modeling in order to gain a deeper understanding ▶ Conductance-Based Models of Nonlinear
of the neural basis of motor control. Dynamics in Vertebrate Motoneurons
In this section, we discuss principles and ▶ Mammalian Motor Nerve Fibers, Models of
methodologies used to develop models of moto- ▶ Morphologically Detailed Motoneuron Models
neurons and motor effectors at various levels of ▶ Neuromuscular Control Systems, Models of
complexities and scales. At the single cell level, ▶ Physiology and Computational Principles of
articles presented describe two-compartment Muscle Force Generation
68 Multistability in Neurodynamics: Overview
motor control, multistability could be a major

Multistability in Neurodynamics: mechanism of operation of multifunctional cen-
Overview tral pattern generators (Getting 1989; Venugopal
et al. 2007; Briggman and Kristan 2008). On the
Gennady Cymbalyuk other hand, multistability can be a pathological
The Neuroscience Institute, Georgia State feature; for example, seizure regimes can coexist
University, Atlanta, GA, USA with normal regimes (Ziburkus et al. 2006;
Cressman et al. 2009; Frohlich et al. 2010). For
the latter, the electrogenic pump has been impli-
Synonyms cated to play a crucial role (Cressman et al. 2009;
Krishnan and Bazhenov 2011).
Bistability; Coexistence; Fold bifurcation; Multistability can be categorized in terms of
Hysteresis; Subcritical bifurcation the regimes of activity which coexist, e.g., coex-
istence of tonic spiking and silence (Rinzel 1978;
Guttman et al. 1980; Forger and Paydarfar 2004;
Definition Hahn and Durand 2001), coexistence of bursting
and silence (Malashchenko et al. 2011a, b), coex-
Multistability in neurodynamics is the coexis- istence of different bursting regimes (Wang
tence of two or more observable regimes of activ- 1994; Butera 1998; Newman and Butera 2010),
ity, i.e., attractors, in the phase space of and coexistence of bursting activity and tonic
a neuronal system. In the absence of noise or spiking (Canavier et al. 1994; Frohlich
perturbation, the neuronal system permanently et al. 2010; Cymbalyuk et al. 2002; Shilnikov
exhibits one of the regimes. Multistability sug- et al. 2005; Malashchenko et al. 2011a). These
gests that by an appropriate choice of perturba- regimes can coexist in the phase space of
tion or by resetting the initial state of the system, a neuronal system if some unstable regime cre-
one could induce a switch from one regime into ates a barrier, i.e., a separatrix demarcating
another. a border between the basins of attraction. The
mechanism underlying multistability can also be
described and classified in terms of the unstable
Detailed Description regime that forms the barrier (Rinzel 1978;
Guttman et al. 1980). One of the key ingredients
Multistable neuronal system can exhibit two or of such descriptions is the identification of the
more regimes of activity, depending on its initial transitions, i.e., bifurcations, at which the unsta-
state. Both isolated neurons and neuronal net- ble regimes appear and disappear. Ubiquitously,
works can exhibit coexistence of several activity a separating regime is either a saddle equilibrium
regimes. The coexistence of silence, subthreshold or a saddle orbit (Rinzel 1978; Malashchenko
oscillations, tonic spiking, and bursting regimes et al. 2011a; Barnett et al. 2013; Marin et al.
with each other has been observed in a number 2013). For example, the stable manifold of the
of theoretical and experimental studies. saddle equilibrium separates tonic spiking and
A multistable neuronal system can show long- silence observed in the simplified model of the
lasting responses to short transient signals. Such cerebellar Purkinje neurons (Loewenstein
systems also may respond to perturbations in et al. 2005; Fernandez et al. 2007), and the stable
a state-dependent fashion, demonstrating hyster- manifold of the saddle periodic orbit is a cause of
esis, which is the hallmark of (Nadim et al. 2008). bistability of spiking and silence in the giant
Multistability has clear implications for dynami- squid axon (Rinzel 1978; Hahn and
cal memory and information processing in neu- Durand 2001).
rons (Marder et al. 1996; Egorov et al. 2002; The methods developed in the bifurcation the-
Durstewitz and Seamans 2006). In the area of ory allow one to investigate stability of, evolution
Multistability in Neurodynamics: Overview 69
of, and transitions between the silent and oscilla- cell-autonomous and network-based mechanisms.
tory regimes of neuronal models in response to J Neurosci 22:10580–10592
Durstewitz D, Seamans JK (2006) Beyond bistability:
variation of the system’s parameters (Kuznetsov biophysics and temporal dynamics of working mem-
2004; Terman and Ermentrout 2010; Izhikevich ory. Neuroscience 139:119–133
2010). Application of these methods identifies Egorov AV, Hamam BN, Fransen E, Hasselmo ME,
and explains the mechanisms supporting Alonso AA (2002) Graded persistent activity in ento-
rhinal cortex neurons. Nature 420:173–178
multistability under normal and pathological con- Fernandez FR, Engbers JD and Turner RW (2007) Firing
ditions (Rinzel 1978; Guttman et al. 1980; Forger dynamics of cerebellar purkinje cells. J Neurophysiol
and Paydarfar 2004; Gutkin et al. 2009; Hahn and 98:278–294. doi:10.1152/jn.00306.2007
Durand 2001; Krishnan and Bazhenov 2011; Forger DB, Paydarfar D (2004) Starting, stopping, and
resetting biological oscillators: in search of optimum
Shilnikov et al. 2005; Fröhlich and Bazhenov perturbations. J Theor Biol 230:521–532
2006; Cressman et al. 2009; Cymbalyuk and Fröhlich F, Bazhenov M (2006) Coexistence of tonic
Shilnikov 2005; Malashchenko et al. 2011a, b). firing and bursting in cortical neurons. Phys Rev
E 74:031922
Frohlich F, Sejnowski TJ, Bazhenov M (2010) Network
Acknowledgment This work was supported by National bistability mediates spontaneous transitions between
Science Foundation grant PHY-0750456. normal and pathological brain states. J Neurosci
30:10734–10743
Getting PA (1989) Emerging principles governing the
Cross-References operation of neural networks. Annu Rev Neurosci
12:185–204
Gutkin BS, Jost J, Tuckwell HC (2009) Inhibition of
▶ Coexistence of Bursting Regimes rhythmic neural spiking by noise: the occurrence of
▶ Multistability Arising from Synaptic a minimum in activity with increasing noise.
Dynamics Naturwissenschaften 96:1091–1097
Guttman R, Lewis S, Rinzel J (1980) Control of repetitive
▶ Multistability in Perception Dynamics firing in squid axon membrane as a model for
▶ Multistability in Seizure Dynamics a neuroneoscillator. J Physiol 305:377–395
▶ Multistability of Coupled Neuronal Oscillators Hahn PJ, Durand DM (2001) Bistability dynamics in sim-
ulations of neural activity in high-extracellular-
potassium conditions. J Comput Neurosci 11:5–18
Izhikevich EM (2010) Dynamical systems in neurosci-
References ence: the geometry of excitability and bursting. The
MIT Press, Cambridge, MA, London, England
Barnett WH, O’Brien G, Cymbalyuk GS (2013) Krishnan GP, Bazhenov M (2011) Ionic dynamics medi-
Bistability of silence and seizure-like bursting. J ate spontaneous termination of seizures and postictal
Neurosci Methods 220(2):179–189 depression state. J Neurosci 31:8870–8882
Briggman KL, Kristan WB (2008) Multifunctional Kuznetsov YA (2004) Elements of applied bifurcation
pattern-generating circuits. Annu Rev Neurosci theory. Springer, New York
31:271–294 Loewenstein Y, Mahon S, Chadderton P, Kitamura K,
Butera RJ (1998) Multirhythmic bursting. Chaos Sompolinsky H, Yarom Y, Häusser M (2005)
8:274–284 Bistability of cerebellar Purkinje cells modulated by
Canavier CC, Baxter DA, Clark JW, Byrne JH (1994) Mul- sensory stimulation. Nat Neurosci 8:202–11.
tiple modes of activity in a model neuron suggest Malashchenko T, Shilnikov A, Cymbalyuk G (2011a) Six
a novel mechanism for the effects of neuromodulators. types of multistability in a neuronal model based on
J Neurophysiol 72:872–882 slow calcium current. PLoS One 6:e21782
Cressman JR Jr, Ullah G, Ziburkus J, Schiff SJ, Barreto Malashchenko T, Shilnikov A, Cymbalyuk G (2011b)
E (2009) The influence of sodium and potassium Bistability of bursting and silence regimes in a model
dynamics on excitability, seizures, and the stability of of a leech heart interneuron. Phys Rev E 84:041910
persistent states: I. Single neuron dynamics. J Comput Marder E, Abbott LF, Turrigiano GG, Liu Z, Golowasch
Neurosci 26:159–170 J (1996) Memory from the dynamics of intrinsic
Cymbalyuk G, Shilnikov A (2005) Coexistence of tonic membrane currents. Proc Natl Acad Sci U S A
spiking oscillations in a leech neuron model. J Comput 93:13481–13486
Neurosci 18:255–263 Marin BM, Barnett WH, Doloc-Mihu A, Calabrese RL,
Cymbalyuk GS, Gaudry Q, Masino MA, Calabrese RL Cymbalyuk GS (2013) High prevalence of
(2002) Bursting in leech heart interneurons: multistability of rest states and bursting in a database
70 Neural Population Models and Cortical Field Theory: Overview
of a model neuron. PLOS Computational Biology 9(3): models together. In contrast, NPMs employ col-
e1002930. doi:10.1371/journal.pcbi.1002930 lective state variables, typically defined as aver-
Nadim F, Brezina V, Destexhe A, Linster C (2008) State
dependence of network output: modeling and experi- ages over the group of cells, in order to describe
ments. J Neurosci 28:11806–11813 the population activity directly in a single model.
Newman JP, Butera RJ (2010) Mechanism, dynamics, and The strength and the weakness of his approach
biological existence of multistability in a large class of are hence one and the same: simplification by
bursting neurons. Chaos 20:023118
Rinzel J (1978) On repetitive activity in nerve. Fed Proc bulk. Is this justified and indeed useful, or does
37:2793–2802 it lead to oversimplification which fails to capture
Shilnikov A, Calabrese RL, Cymbalyuk G (2005) Mech- the phenomena of interest?
anism of bistability: tonic spiking and bursting in In “Mesoscopic Anatomy and Neural Popula-
a neuron model. Phys Rev E Stat Nonlin Soft Matter
Phys 71:056214 tion Models,” Liley explains that the anatomy of
Terman DH, Ermentrout GB (2010) Mathematical foun- the brain is organized into connected modules at
dations of neuroscience. Springer, New York, Dor- the mesoscopic level, which can serve as
drecht, Heidelberg, London a biological substrate for NPMs. In “Neuroper-
Venugopal S, Travers JB, Terman DH
(2007) A computational model for motor pattern colation and Neural Population Models” and
switching between taste-induced ingestion and rejec- “Neural Mass Action,” Kozma describes how
tion oromotor behaviors. J Comput Neurosci this undergirds coherent neural activity. Brunel
22:223–238 and Hakim demonstrate in “Population Density
Wang XJ (1994) Multiple dynamical modes of thalamic
relay neurons: rhythmic bursting and intermittent Model” and “Fokker-Planck Equation” how
phase-locking. Neuroscience 59:21–31 NPMs can be derived mathematically from
Ziburkus J, Cressman JR, Barreto E, Schiff SJ existing descriptions of individual cells, and in
(2006) Interneuron and pyramidal cell interplay during “Stochastic Neural Field Theory,” Bressloff
in vitro seizure-like events. J Neurophysiol
95:3948–3954 shows this for the CFT variant as well. Liley
then introduces the general NPM approach in
“Neural Population Model,” and Hutt the CFT var-
iant in “Neural Field Model, Continuum.” Some
historically important and/or particularly popular
Neural Population Models and NPMs are showcased in “Amari Model” by
Cortical Field Theory: Overview Potthast, “Down Under Neural Population Models”
by Liley, “Jansen-Rit Model” by Knösche, and
Ingo Bojak “Wilson-Cowan Model” by Kilpatrick.
School of Systems Engineering, University of The modelling of connectivity in NPMs has
Reading, Reading, UK been a topic of considerable development over
the years and is covered in “Propagator, Axonal”
and “Synaptic Connectivity in Neural Population
The term neural population models (NPMs) is Models” by Jirsa, as well as in “Gap Junctions,
used here as catchall for a wide range of Neural Population Models and” by Sleigh and
approaches that have been variously called neural the Steyn-Rosses. The dynamical properties
mass models, mean field models, neural field of NPMs are key for their application to
models, bulk models, and so forth. All NPMs various brain phenomena. The quasi-linear and
attempt to describe the collective action of neural noise-driven mode is described in “Transfer
assemblies directly. Some NPMs treat the Function, Electrocortical” by Molaee-Ardekani
densely populated tissue of cortex as an excitable and Wendling. Transitions in dynamics, for
medium, leading to spatially continuous cortical example, to self-sustained oscillations, are
field theories (CFTs). An indirect approach discussed in “Bifurcations, Neural Population
would start by modelling individual cells and Models and” by Knösche and “Phase Transitions,
then would explain the collective action of Neural Population Models and” by Sleigh and the
a group of cells by coupling many individual Steyn-Rosses. Whereas the emergence of spatial
Neural Population Models and Cortical Field Theory: Overview 71
structure is presented in “Pattern Formation in and interactions of so many brain cells in

Neural Population Models” by Hutt and the cha- a biologically meaningful manner and cannot
otic regime in “Chaos, Neural Population Models generate actual human insight into principles of
and” by Knösche. function from the plethora of individual cell
A fundamental strength of NPMs is their abil- activities. Once the computational power
ity to predict the data generated by neuroimaging becomes readily available, this realization will
modalities like the electro- and magnetoence- be unavoidable. The way forward will be multi-
phalogram (EEG/MEG) and functional magnetic scale descriptions wherein higher levels discard
resonance imaging (fMRI), as described in “Neu- irrelevant detail of lower levels for an effective
roimaging, Neural Population Models for” by and efficient description that remains accessible
Bojak and Breakspear. This allows the compari- to the human mind. And given their intimate
son of model predictions with experimental data, connection to (noninvasive) neuroimaging, we
be it for normal brain function as in “Gamma expect that NPMs will play a privileged role in
Rhythm, Neural Population Models of the” by such a future scheme. This section provides
Bojak and “Sleep, Neural Population Models a mere snapshot of a field that is currently
of” by Phillips, under the influence of drugs as growing rapidly and will continue to do so in
in “Anaesthesia, Neural Population Models of” the foreseeable future.
by Sleigh and the Steyn-Rosses or in disease as in
“Epilepsy, Neural Population Models of” by
Wendling and Molaee-Ardekani. While many of
the methods used to predict the recorded signals Cross-References
are sound and even sophisticated, some of the
basics still require substantial improvements, as ▶ Amari Model
Einevoll reminds us in “Extracellular Potentials, ▶ Anesthesia, Neural Population Models of
Forward Modelling of.” Much of the work so far ▶ Bifurcations, Neural Population Models and
has been in forward prediction with NPMs; how- ▶ Chaos, Neural Population Models and
ever, increasingly it is important to estimate the ▶ Coordination Dynamics
NPM state and parameters directly from the data. ▶ Down Under Neural Population Models
This is elucidated by Moran in “Dynamic Causal ▶ Dynamic Causal Modelling with Neural
Modelling with Neural Population Models” and Population Models
by Potthast in “Inverse Problems in Neural Pop- ▶ Embodied Cognition, Dynamic Field Theory
ulation Models.” Finally, it is exciting that of
NPMs, and CFTs in particular, can be used to ▶ Epilepsy, Neural Population Models of
model perception and motor control, as Schöner ▶ Extracellular Potentials, Forward Modeling of
describes in “Embodied Cognition, Dynamic ▶ Fokker-Planck Equation
Field Theory of,” with additional remarks by ▶ Gamma Rhythm, Neural Population Models of
Schöner and Nowak given in “Coordination the
Dynamics.” ▶ Gap Junctions, Neural Population Models and
To answer the question that was posed in the ▶ Inverse Problems in Neural Population Models
beginning, not only is the NPM approach justifi- ▶ Jansen-Rit Model
able and has shown itself to be useful; we expect ▶ Mesoscopic Anatomy and Neural Population
that within a decade or two, NPMs will take their Models
rightful place as the primary means for describing ▶ Neural Field Model, Continuum
mesoscopic brain activity. The central problem ▶ Neural Mass Action
we face at this description level will not be to ▶ Neural Population Model
simulate millions of brain cells in a reasonable ▶ Neuroimaging, Neural Population Models for
amount of time; the central problem is now, and ▶ Neuropercolation and Neural Population
will be then, that we cannot specify the properties Models
72 Neuromodulation: Overview
▶ Pattern Formation in Neural Population growing interest in systems neuroscience, the

Models computational role of neuromodulation in infor-
▶ Phase Transitions, Neural Population Models mation processing has helped elucidate
and neuromodulatory function, and predictive theories
▶ Population Density Model have arisen from computational approaches.
▶ Propagator, Axonal Neuromodulation can be described by its spa-
▶ Sleep, Neural Population Models of tial and temporal characteristics, as well as the
▶ Stochastic Neural Field Theory specific computational function ascribed to
▶ Synaptic Connectivity in Neural Population it. Spatial characteristics include extrinsic, origi-
Models nating from an area extrinsic to the network under
▶ Transfer Function, Electrocortical study, or intrinsic, originating from processes
▶ Wilson-Cowan Model within the area under study. The computational
functions of extrinsic neuromodulation, such as
ACh, NE, 5HT, and DA, are usually considered
somewhat global, because they modulate many
Neuromodulation: Overview areas of the brain simultaneously. Classically,
ACh has been associated with attentional pro-
Christiane Linster cesses, NE with signal-to-noise modulation, DA
Computational Physiology Lab, Department of with reward learning, and 5HT with sleep-wake
Neurobiology and Behavior, Cornell University, transitions. In other cases, modulation is specific
Ithaca, NY, USA to the network under investigation and an integral
part of the computations performed within. Sec-
ond messenger systems, plasticity processes, and
Definition gene regulation are examples of such intrinsic
modulation. From a functional point of view,
Neuromodulation refers to the regulation of neu- neuromodulation is often regulatory, for exam-
ral and synaptic function by regulatory extrinsic ple, in the cases of second messenger systems or
or intrinsic substances. activity-dependent regulation of conductances. In
the sensory systems, neuromodulation is often
linked to tuning of receptive fields (ACh) and
Detailed Description regulation of signal-to-noise ratio. A third highly
important function of most neuromodulators is
Computational modeling of neural substrates pro- the regulation of plasticity, via excitability of
vides an excellent theoretical framework for the neurons, synaptic plasticity, and broader modu-
understanding of the computational roles of lation of network dynamics.
neuromodulation. Neuromodulation can be Exactly how neuromodulation is integrated in
defined as biophysical processes that serve to computational studies depends widely on the
modify – or modulate – the computation details of implementation of the computational
performed by a neuron or network as a function model itself. Effects of neuromodulators can be
of task demands and behavioral state of the animal. implemented from the detailed biophysical level,
These modulatory effects often involve substances to broader regulation of network parameters in
such as acetylcholine (ACh), norepinephrine (NE), case of more abstract large-scale models. For
histamine, serotonin (5-HT), dopamine (DA), and example, specific effects on voltage-gated chan-
a variety of neuropeptides. Modulatory effects are nels may be implemented in a biophysical model
difficult to define, because they often originate by changing channel parameters, in a simplified
from different structures and have different spatial integrate and fire model by changing a related
distributions and time courses of action. Because parameter such as firing threshold. Each
of the wider use of modeling techniques and study chooses the level of detail appropriate to
Neuromorphic Engineering: Overview 73
the question asked and data available. For a computational architecture, shaped to model bio-
comprehensive review of levels of implementa- logical neural systems: neuromorphic engineer-
tion of neuromodulation in computational ing is by essence strongly linked to computational
models, we refer the reader to Fellous and Linster neuroscience.
(1998). The chapters in this section cover over-
views of neuromodulatory computation divided
by substance, nature of task, as well as overviews Detailed Description
of types of network implementations and specific
examples. Neuromorphic engineering offers an interesting
alternative to computer simulations of neural net-
works: while main components of computers are
Cross-References high-precision high-speed digital hardware with
high-power dissipation, neural components are
▶ Computation with Dopaminergic Modulation slow asynchronous computational elements
▶ Computation with Serotonergic Modulation which use a combination of analog and digital
▶ Computational Models of Modulation of signal representations. Neuromorphic systems
Oscillatory Dynamics are hardware implementations that operate in
▶ Computational Models of Neuromodulation physical time; they are inspired by the structure,
function, and plasticity of biological nervous sys-
tems. Their design is facilitated by similarities
References between VLSI (very large-scale integrated cir-
cuits) hardware and neural bioware.
Fellous JM, Linster C (1998) Computational models of Since decades, researchers have been
neuromodulation. Neural Comput 10(4):771–805
implementing electronic models of neural cir-
cuits, following the long history of artificial neu-
Further Reading
ral networks that started in the 1950s. Original
Dayan P (2012) Twenty-five lessons from computational
neuromodulation. Neuron 76(1):240–256. neuromorphic circuits (Mead 1989) were VLSI
doi:10.1016/j.neuron.2012.09.027 systems that used the nonlinear and graded prop-
erties of transistors to emulate protein channels in
neurons. More recently the term neuromorphic
has been used to describe analog, digital, and
Neuromorphic Engineering: mixed analog-digital VLSI as models of neural
Overview computation.
Neuron model computation can be either ana-
Sylvie Renaud log, with a continuous-time and continuous-value
Institut Polytechnique de Bordeaux, Université representation, or digital using numerical algo-
de Bordeaux, Talence, France rithms. In both cases, neuromorphic devices clas-
sically use spiking (asynchronous binary signal
analogs to neural spikes) representation for neu-
Definition ronal communication and learning mechanisms.
At the neural network level, synaptic modifica-
Neuromorphic engineering is a recent interdisci- tion mechanisms are usually introduced to model
plinary field involving biologists, physicists, learning and adaptation in neural systems.
mathematicians, computer scientists, and engi- Furthermore, similar to robustness of neural
neers to design hardware/physical models of neu- networks to changes in the environment, dynamic
ral systems. It aims at designing silicon-based synaptic modification in neuromorphic
neural systems for computational or biomedical systems allows gaining a remarkable robustness
purposes. The term “neuromorphic” relates to the against transistor-level variations. Whatever the
74 Neuronal Model Optimization: Overview
technological support of the neuromorphic elec- ▶ Neuromorphic Sensors, Cochlea

tronics, real-time computation is mandatory for ▶ Neuromorphic Sensors, Head Direction
systems interacting with a biological environ- ▶ Neuromorphic Sensors, Olfaction
ment (sensory or motor). ▶ Neuromorphic Sensors, Vision
To date, widely known neuromorphic systems ▶ Neuromorphic Technologies, Memristors
are highly integrated VLSI that emulate sensory ▶ Spike-Timing Dependent Plasticity, Learning
transduction (vision, olfaction, audition) or more Rules
sophisticated and multimodal neural systems like ▶ Spiking Network Models and Theory:
the head-direction system. Larger neuromorphic Overview
systems emerged in the last decade, emulating ▶ Synaptic Dynamics: Overview
complex or cognitive biological functions and
suggesting new research directions for robotics
by the means of biomorphic engineering. References
According to the VLSI technology road map,
promising low-power and high-density nanoscale Mead C (1989) Analog VLSI and neural systems.
Addison-Wesley, Reading
technologies, including organic electronics,
memristors, or 3D technologies, are expected to
fulfill the need for large-scale neuromorphic sys-
tems, currently implemented using standard
CMOS VLSI technologies. Large neuromorphic Neuronal Model Optimization:
systems have performance measures (high den- Overview
sity, high speed, low power) that are good enough
to enable the study of large-scale and biologically Astrid A. Prinz
inspired neural networks. Such large networks Department of Biology, Emory University,
represent a challenge for computational neurosci- Atlanta, GA, USA
ence. Researchers have recently started investi-
gating cognitive neuromorphic systems that are
able to perform adaptive behaviors and that Synonyms
merge neural fields and single-neuron represen-
tations courtesy of the increased computational Neuronal Model Tuning
power of microelectronic devices.
Entries in this section review state-of-the-art
neuromorphic engineering computational princi- Definition
ples, technologies, and applications. This trans-
disciplinary research field, at its heart, combines Neuronal Model Parameter Optimization is the
the physics of electronics and biology and algo- process of adjusting the parameters of
rithms and then engineers it for solving tasks. a computational model of a neuron or neuronal
Despite the many challenges, neuromorphic network in order to achieve model activity that
engineering may hold great promises for a new mimics that of the living neuron or network being
generation of medicine or industry technologies. modeled.
Cross-References Detailed Description
▶ Brain-Machine Interface: Overview This section contains entries that explain the need
▶ Dynamical Systems: Overview for neuron and network model parameter optimi-
▶ Neuromorphic Cognition zation, discuss various optimization methods,
▶ Neuromorphic Hardware, Large-Scale describe existing software tools for optimization
Neuronal Model Optimization: Overview 75
and visualization of the model databases that is “▶ Bifurcation Analysis,” i.e., the use of
result from some of the optimization methods, dynamical systems analysis to determine
and discuss related concepts that have emerged critical parameters that cause transitions from
from model optimization and exploration efforts one behavior to another. Bifurcation analysis is
over the last few years. most powerful for models of low dimensionality.
For models with many dynamic variables and
Why do Neuronal Model Parameters Need to parameters, the use of bifurcation analysis may
Be Optimized? therefore require prior model simplification
Computational models of neurons and neuronal through “▶ Neuronal Model Reduction,” ideally
circuits are important investigative tools that without the loss of essential features of model
allow the study of neuronal signaling and infor- activity.
mation processing mechanisms that would A more agnostic approach to studying the
not be experimentally accessible. However, dependence of neuronal model activity on the
constructing a model neuron or network requires underlying model parameters is “▶ Neuronal
the specification of numerous parameters, such as Parameter Space Exploration,” i.e., the brute-
maximal conductances, half-activation voltages, force computational exploration of many – often
and time constants of ionic membrane currents, tens of millions – of parameter combinations that
as well as the strength and dynamical properties cover the model’s high-dimensional parameter
of synapses. A complete set of parameters to space. Characteristics of the simulated model
describe a given neuron or network is virtually activity for each parameter combination are then
never available from experimental data, because stored in “▶ Neuronal Model Databases” that can
some of the parameters cannot be directly mea- be mined for model versions with the desired
sured. Combining parameters measured from dif- activity, and to address general questions about
ferent animals is also problematic because almost the relationship between model parameters and
all parameters vary between animals, and the output.
distributions in the space of parameter sets Less exploratory, more targeted model opti-
obtained from different animals are often mization methods are based on “▶ Evolutionary
non-convex, leading to “▶ Failure of Averag- Algorithms.” They require the definition of
ing,” i.e., the problem that the set of parameter a “▶ Neuronal Model Output Fitness Function,”
averages obtained from a distribution of parame- which is a measure for how well a neuron or
ter sets can fall outside that distribution in param- network model’s activity replicates that of the
eter space. living system to be modeled. “▶ Multi-objective
Neuronal models initially constructed from Evolutionary Algorithms” constitute a class of
incomplete and averaged parameter sets therefore evolutionary algorithms that simultaneously opti-
usually generate activity that does not match that mize models to meet several different objectives,
of the living system to be modeled. Such dys- rather than maximizing a single fitness measure.
functional models then require parameter An implementation of evolutionary neuronal
optimization, either through “▶ Neuronal Model model optimization is available in the software
Hand-Tuning” by an expert employing educated tool “▶ Neurofitter.”
guessing and trial-and-error parameter adjust- Finally, “▶ Hybrid Parameter Optimization
ments, or through one of the optimization Methods” combine several of the above methods
methods described in this section. of neuronal model optimization to benefit from
their respective advantages.
Neuronal Model Optimization Methods
One approach that allows the identification of Visualization of Neuronal Model Parameter
neuronal model parameter regimes that produce Spaces
qualitatively different model behaviors (for Several of the model optimization methods
example, spiking versus bursting versus silence) described in this section produce long lists of
76 Olfaction: Overview
parameter sets and their corresponding neuronal Cross-References

model activity that are often collected in “▶ Neu-
ronal Model Databases.” Because the underlying ▶ Bifurcation Analysis
models and their parameter spaces are high- ▶ Evolutionary Algorithms
dimensional, it is difficult to analyze how the ▶ Failure of Averaging
activity of a model varies across parameter ▶ Hybrid Parameter Optimization Methods
space. This problem is addressed by recently ▶ Multi-objective Evolutionary Algorithms
developed methods of “▶ Neuronal Parameter ▶ NDVis–Neuro
Space Visualization,” including the visualization ▶ Neurofitter
tool “▶ NDVis-Neuro,” which is designed for the ▶ Neuronal Model Databases
visualization of model databases that cover ▶ Neuronal Model Hand-Tuning
parameter space with a regular grid of ▶ Neuronal Model Output Fitness Function
parameter sets. ▶ Neuronal Model Reduction
▶ Neuronal Parameter Co-regulation
Beyond Optimization ▶ Neuronal Parameter Non-uniqueness
A number of recent neuronal model optimization ▶ Neuronal Parameter Sensitivity
efforts, as well as a growing body of experimental ▶ Neuronal Parameter Space Exploration
data, have revealed concepts that go beyond ▶ Neuronal Parameter Space Visualization
model optimization in the sense of identifying
a single “optimal” model parameter set.
Neuronal models and the living systems they
mimic display “▶ Neuronal Parameter Olfaction: Overview
Non-uniqueness,” meaning that different param-
eter sets can produce similar and functional neu- Christiane Linster
ron or network activity. The subset of parameter Computational Physiology Lab, Department of
space that contains such functional parameter sets Neurobiology and Behavior, Cornell University,
is called the “Neuronal Solution Space.” Analysis Ithaca, NY, USA
of the structure of neuronal solution spaces has
revealed that parameters do not necessarily
vary independently within the solution space, Definition
but often show “▶ Neuronal Parameter
Co-regulation.” Olfaction or the olfactory systems refer to the
Apart from asking whether a given parameter sense of smell.
set does or does not support functional model
activity, parameter space exploration also allows
for the analysis of “▶ Neuronal Parameter Sensi- Detailed Description
tivity,” that is, in which direction and how
strongly activity characteristics such as spike fre- In natural environments, airborne chemical stimuli
quency depend on model parameters. can occur unpredictably in time and space, and
Together, the realization of “▶ Neuronal odorants from many sources intermix. The olfac-
Parameter Non-uniqueness” and the connected tory system of animals is capable of detecting
concept of “Neuronal Solution Space” have led stimuli of interest to the animal within these chem-
to the strategy of “Ensemble Modeling,” i.e., the ically noisy environments. To achieve this, the
approach of analyzing an entire ensemble of system has to extract relevant signals, compare
parameter sets and model versions that produce the extracted stimulus to previously experienced
functional activity, rather than focusing on odors, differentiate relevant from irrelevant stim-
a single “optimal” model version. uli, and cue an appropriate response. Many of the
Olfaction: Overview 77
Olfaction: Overview,
Fig. 1 Schematic
description of olfactory
structures. The olfactory
sensory neurons (OSN)
expressing a common
receptor converge onto the
common olfactory
glomeruli, where they
make excitatory synapses
with the mitral (Mi),
periglomerular (PG) and
external tufted (ET) cells.
ET, PG, and SA cells
interconnect in a local
network of interneurons.
The high convergence
between the OSNs and
glomerular cells can
increase signal-to-noise
ratio in the system (Linster
and Cleland 2009). PG cells
make inhibitory
connections with Mi cell
primary dendrites, allowing
for contrast enhancement of
the olfactory signal before
the spike generation
mechanism of the Mi cell
(Cleland and Sethupathy
2006)
neural circuit elements comprising the olfactory been proposed to create an associative memory
system are thought to contribute to these pro- for odor storage (Hasselmo et al. 1990).
cesses. Computational models of olfactory Computational models of olfaction have pro-
processing have at the forefront of attributing vided both mechanistic and functional insights
function to experimentally described neurons and into olfactory perception and learning.
networks (Fig. 1). Mechanistical models have long focused on the
The SA, ET, and PG network has been pro- creation and maintenance of well-described oscil-
posed to perform concentration invariant compu- latory processes in the olfactory system, especially
tations (Cleland 2010). In the next layer, Mi and oscillations in the gamma range (see, e.g., Free-
granule (Gr) cells form a reciprocal excitatory/ man 1974; Lagier et al. 2004; Li and Cleland
inhibitory loop, assumed to be the neural sub- 2013). Functional models have focused on com-
strate for fast oscillations and synchrony putations such as signal-to-noise ratio modulation
(Freeman 1974; Brea et al. 2009). Mi cell axons (e.g., Hasselmo et al. 1997; Linster et al. 2011),
project to secondary olfactory structures, includ- contrast enhancement (e.g., Linster and Hasselmo
ing the piriform cortex (PC) in a topographically 1997; Urban 2002; Cleland and Sethupathy 2006),
nonorganized fashion. These synapses undergo plasticity, and associative learning (e.g., Hasselmo
rapid adaptation proposed to underlie short-term et al. 1990; Wilson and Linster 2008). Many
odor memory (Wilson and Linster 2008). The models have focused on linking neural and per-
extensive excitatory connections between Pyrs, ceptual processes (among others, Mandairon
known to undergo synaptic plasticity, have long et al. 2006); others have stayed abstract and
78 Peripheral Nerve Interfaces: Overview
predictive (e.g., Brea et al. 2009). Overall, com- Lagier S, Carleton A, Lledo PM (2004) Interplay between
putational approaches have greatly contributed to local GABAergic interneurons and relay neurons gen-
erates gamma oscillations in the rat olfactory bulb.
our understanding of olfactory computation and J Neurosci 24:4382–4392
have uniquely advanced the field beyond the mere Li G, Cleland TA (2013) A two-layer biophysical model
collection of data. The section on olfaction covers of cholinergic neuromodulation in olfactory bulb.
different types of modeling approaches ranging J Neurosci 33:3037–3058
Linster C, Cleland TA (2009) Olfactory microcircuits.
from detailed biophysical models of individual or Neural Netw 22(8):1169–1173
small groups of neurons to large-scale models of Linster C, Hasselmo M (1997) Modulation of inhibition in
associative learning in the olfactory cortex. a model of olfactory bulb reduces overlap in the neural
representation of olfactory stimuli. Behav Brain Res
84:117–127
Linster C, Nai Q, Ennis M (2011) Nonlinear effects of
Cross-References noradrenergic modulation of olfactory bulb function in
adult rodents. J Neurophysiol 105:1432–1443
▶ Biophysical Models of Olfactory Mitral and Mandairon N, Ferretti CJ, Stack CM, Rubin DB, Cleland
Granule Cells TA, Linster C (2006) Cholinergic modulation in the
olfactory bulb influences spontaneous olfactory dis-
▶ Computational Modeling of Olfactory crimination in adult rats. Eur J Neurosci 24:3234–3244
Behavior Urban NN (2002) Lateral inhibition in the olfactory bulb
▶ Computational Olfaction and in olfaction. Physiol Behav 77:607–612
▶ Olfactory Computation and Adult Wilson DA, Linster C (2008) Neurobiology of a simple
memory. J Neurophysiol 100:2–7
Neurogenesis
▶ Olfactory Computation in Antennal Lobe and Further Reading
Mushroom Bodies Cleland TA, Linster C (2005) Computation in the olfac-
▶ Olfactory Computation in Glomerular tory system. Chem Senses 30(9):801–813
Crasto CJ (2009) Computational biology of olfactory
Microcircuits receptors. Curr Bioinform 4(1):8–15
▶ Olfactory Computation in Mitral-Granule Cell Laurent G, Stopfer M, Friedrich RW, Rabinovich MI,
Networks Volkovskii A, Abarbanel HD (2001) Odor encoding
▶ Olfactory Cortical Associative Memory as an active, dynamical process: experiments, compu-
tation and theory. Annu Rev Neurosci 24:263–297
Models Schoppa NE, Urban NN (2003) Dendritic processing within
olfactory bulb circuits. Trends Neurosci 26(9):501–506
References
Brea JN, Kay LM, Kopell NJ (2009) Biophysical model

Peripheral Nerve Interfaces:
for gamma rhythms in the olfactory bulb via subthresh- Overview
old oscillations. Proc Natl Acad Sci U S A
106:21954–21959 Douglas J. Weber1, Brian Wodlinger2 and
Cleland TA (2010) Early transformations in odor repre-
sentation. Trends Neurosci 33(3):130–139
Wei Wang2
1
Cleland TA, Sethupathy P (2006) Non-topographical contrast Department of Bioengineering, and Department
enhancement in the olfactory bulb. BMC Neurosci 7:7 of Physical Medicine and Rehabilitation,
Freeman WJ (1974) Dynamic patterns of brain cell assem- University of Pittsburgh, Pittsburgh, PA, USA
blies. IV. Mixed systems. Oscillating fields and pulse 2
distributions. Pulse-wave problems. Neurosci Res Pro-
Department of Physical Medicine and
gram Bull 12:102–107 Rehabilitation, University of Pittsburgh,
Hasselmo ME, Wilson MA, Anderson BP, Bower JM (1990) Pittsburgh, PA, USA
Associative memory function in piriform (olfactory)
cortex: computational modeling and neuropharmacol-
ogy. Cold Spring Harb Symp Quant Biol 55:599–610 Definition
Hasselmo ME, Linster C, Patil M, Ma D, Cekic M (1997)
Noradrenergic suppression of synaptic transmission
may influence cortical signal-to-noise ratio. The peripheral nervous system (PNS) comprises
J Neurophysiol 77:3326–3339 nerve fibers and ganglia located outside of the
Peripheral Nerve Interfaces: Overview 79
brain and spinal cord, including 11 pairs of cra- however, through the pioneering modeling work
nial nerves and 31 pairs of spinal nerves. The of researchers like Hodgkin and Huxley, and
spinal (or somatic) nerves originate as dorsal those that have expanded and continued their
and ventral spinal nerve rootlets at their junction work, have we begun to understand the PNS
with the spinal cord. The dorsal spinal nerve roots well enough to design safe and effective inter-
contain afferent (sensory) nerve fibers and their faces to it.
cell bodies, which are located in the dorsal root A wide variety of interface technologies has
ganglia (DRG). The ventral spinal nerve roots been developed to record or stimulate activity in
contain efferent (motor) fibers whose cell bodies peripheral nerves. Most interfaces rely on elec-
are located in the ventral horn of the spinal cord. trodes placed on the nerve surface (epineural) or
The dorsal and ventral roots merge to form the penetrate the interior of the epineurium
spinal nerve, which exits the spinal column (intraneural) of the nerve (▶ Peripheral Nerve
through the intervertebral foramen. Upon exiting Interface, Intraneural Electrode; ▶ Peripheral
the spine, the nerve divides into dorsal and ven- Nerve Interface, Epineural Electrode; ▶ Periph-
tral rami, which may branch further and mix with eral Nerve Signal Processing, Multipole Cuff
other nerve fibers to form plexuses (e.g., brachial Methods; ▶ Peripheral Nerve Interface,
plexus) and eventually form distinct peripheral Regenerative).
nerves and their branches. Peripheral nerve interface technologies have
been applied to the diagnosis or treatment of
a wide range of medical conditions. Examples
Detailed Description of diagnostic applications include electromyog-
raphy and nerve conduction tests used to identify
This section begins with background information and localize nerve damage caused by carpal tun-
on the anatomy and physiology of the peripheral nel syndromes, diabetic neuropathies, and other
nervous system (▶ Peripheral Nerves, Anatomy injuries or diseases (▶ Peripheral Nerve Interface
and Physiology of) and mathematical models for Applications, EMG/ENG).
simulating the biophysical properties of periph- Electrical stimulation of peripheral nerve
eral nerves (▶ Peripheral Nerve Models). Subse- fibers is used to evoke activity in sensory,
quent entries describe devices and signal motor, and autonomic fibers. The cochlear
processing methods for acquiring and analyzing implant is a highly successful example of
electrical signals from peripheral nerves. Tech- a sensory prosthesis that restores hearing to peo-
niques for stimulating activity in peripheral nerve ple with profound hearing loss due to damage to
fibers are also described. This section also the hair cells in the cochlea (▶ Peripheral Nerve
describes several research and clinical applica- Interface Applications, Cochlear Implants). Elec-
tions based on peripheral nerve recording and trical stimulation of cutaneous nerve fibers with
stimulation techniques. TENS/PENS devices is used to treat a variety of
pain syndromes (▶ Peripheral Nerve Interface
Peripheral Nerve Interface Methods and Applications, Neuropathic Pain). Other types of
Applications sensory prostheses are being developed
Interfaces to the peripheral nervous system have (▶ Peripheral Nerve Interface Applications, Sen-
long been considered of medical value, beginning sory Restoration), for example, to provide tactile
in ancient Egypt where the pain-relieving prop- and proprioceptive sensations to users of pros-
erties of the electric catfish were discussed and thetic limbs.
continuing with the Mediterranean electric ray
and torpedo fish. The nineteenth century saw Summary
renewed interest in electrical stimulation of the The highly structured nature of the peripheral
PNS due to dramatic advances in the understand- nervous system makes it amenable to computa-
ing of electricity. Only relatively recently, tional modeling. Advances in these interfaces
80 Phase Response Curves: Overview
have already made a profound impact on modern ▶ Peripheral Nerve Interface Applications, Sleep
medical technology including cochlear implants Apnea
(▶ Peripheral Nerve Interface Applications, ▶ Peripheral Nerve Interface Applications, Vagal
Cochlear Implants); techniques to support blad- Nerve Stimulation
der, bowel, and sexual function; and vagal nerve ▶ Peripheral Nerve Interface, Epineural
stimulation (▶ Peripheral Nerve Interface Appli- Electrode
cations, Vagal Nerve Stimulation). New tech- ▶ Peripheral Nerve Interface, Intraneural
niques currently being investigated will address Electrode
a variety of disorders throughout the body includ- ▶ Peripheral Nerve Interface, Regenerative
ing obesity (▶ Peripheral Nerve Interface Appli- ▶ Peripheral Nerve Models
cations, Obesity), neuropathic pain (▶ Peripheral ▶ Peripheral Nerve Signal Processing, Denoising
Nerve Interface Applications, Neuropathic Pain), ▶ Peripheral Nerve Signal Processing, Multipole
sleep apnea (▶ Peripheral Nerve Interface Appli- Cuff Methods
cations, Sleep Apnea), and sensory restoration in ▶ Peripheral Nerve Signal Processing, Source
amputees (▶ Peripheral Nerve Interface Applica- Localization
tions, Sensory Restoration). The interfaces them- ▶ Peripheral Nerve Stimulation Technique,
selves come in a diverse array of forms Nerve Block
from simple cuff electrodes (▶ Peripheral Nerve ▶ Peripheral Nerves, Anatomy and Physiology of
Interface, Epineural Electrode) which wrap
around the nerve to more complex regenerative
electrode arrays (▶ Peripheral Nerve Interface,
Regenerative), and utilize many advanced Phase Response Curves: Overview
stimulation and signal processing techniques
(▶ Peripheral Nerve Signal Processing, Carmen C. Canavier
Denoising). Department of Cell Biology and Anatomy, LSU
Peripheral nerve interface technologies hold Health Sciences Center, New Orleans, LA, USA
great promise in the diagnosis and treatment of
a wide range of neurological disorders. Research
in this area requires significant collaboration Detailed Description
between researchers in computational modeling,
neuroscience, biomedical engineering medicine, Phase response curves (PRCs, alternatively
and physiology. phase-resetting curves) are a powerful way of
characterizing and explaining the behavior of
nonlinear oscillators without knowing anything
Cross-References about their specific internal dynamics. The phase
response curve represents the shortening or
▶ Peripheral Nerve Interface Applications, lengthening of the cycle period caused by an
Cochlear Implants input depending upon at what point (phase)
▶ Peripheral Nerve Interface Applications, within the cycle an input is received. In contrast,
EMG/ENG for a linear time invariant system, the effect of an
▶ Peripheral Nerve Interface Applications, input is independent of when it is applied. No
Neuropathic Pain matter whether the oscillations represent the
▶ Peripheral Nerve Interface Applications, flashing of fireflies, a pendulum-based clock,
Obesity a cardiac cell, or a neural oscillator, the phase
▶ Peripheral Nerve Interface Applications, response curve predicts the phasic relationship of
Respiratory Pacing the oscillator to periodic forcing or to coupling
▶ Peripheral Nerve Interface Applications, within a network of other oscillators. Each oscil-
Sensory Restoration lator can be reduced to a phase oscillator whose
Phase Response Curves: Overview 81
angular velocity on a circle is constant, except brief pulse rescales linearly with the height of
when it receives an input that resets (advances or the pulse. Thus, a complicated input like
delays) its phase on the circle. In a neural context, a postsynaptic current waveform can be concep-
there is usually a threshold event, often the action tually broken up into a series of shifted and scaled
potential or spike, which is used to demarcate the pulses (Dirac delta functions). The entry on
boundaries of a cycle. If an input shortens the “▶ Weak Coupling Theory” explains the connec-
cycle, the next spike occurs sooner than it other- tion between the iPRC and the general PRC. That
wise would have, and so the next spike is is, if the coupling is sufficiently weak, the general
advanced. Conversely, if the next spike occurs phase resetting due to any arbitrary input wave-
later than it otherwise would have, the spike is form can be computed by summing the phase
delayed. resetting due to each individual pulse: the timing
The phase-resetting curve is sometimes of each pulse gives the phase, and the height of
presented as the response to a specific input. We the pulse gives the scale factor for the resetting.
have called this the general PRC. One example This amounts to convolving the iPRC with the
might be to perturb the oscillator underlying the input waveform. Since the coupling is weak,
circadian rhythm by exposing an animal to the assumption is that the relative phases of
a period of light during its usual period of dark- the oscillators change very slowly. Weak
ness; another example might be to stimulate coupling can be used to predict the synchroniza-
a particular synaptic input or set of synaptic tion tendencies of networks of oscillators. The
inputs to a neural oscillator. The “▶ Spike Time entry on “▶ Phase Models, Noisy” explains the
Response Curve” is a way to plot the phase effect of noise under the assumption of weak
response to a specific input, that is, a spike in coupling.
the presynaptic neuron, in a way that preserves Many factors, including the underlying bifur-
information about time intervals. Phase-resetting cation structure and the particular set of conduc-
theory is generally applied to neural circuits tances expressed by the neuron, influence the
under a simplifying assumption; two common PRC shape as described in the entry on
assumptions are that of pulsatile (brief) coupling “▶ Phase Response Curve, Measurement and
or weak coupling. The entry on “▶ Pulse- Shape of General.” The influence of the underly-
Coupled Oscillators” uses the information in the ing bifurcation structure described in that entry
spike-time response curve (or the information in led to the classification of all PRCs with a single
the general phase-resetting curve combined with sign (all advances or all delays) as type 1 and
the intrinsic period information) to predict the those with both advances and delays as type 2. A
response of an oscillator to periodic forcing or completely separate, classification scheme into
mutual coupling, under the pulse-coupled type 0 and type 1 PRCs is described in the entry
assumption that the effect of each input dissipates on “▶ Phase Response Curve, Topology of.”
quickly compared to the cycle period. The two schemes are a potential source of
A second use of the term phase-resetting curve confusion because of the similar terminology,
is to represent the response of the oscillator to an but these schemes are unrelated. In the
infinitesimal input, in other words as a change in topology-based scheme, which was developed
the period per unit of the input, called the infini- first, weak resetting leads to type 1 phase reset-
tesimal PRC or iPRC. The entry on “▶ Phase ting, in which the phase immediately after an
Response, Measurement of the Infinitesimal” input can take on any value, but strong resetting
explains how to measure the infinitesimal phase leads to type 0 resetting in which only a limited
response curve (iPRC). The iPRC is useful set of phase values are allowed immediately after
mostly in the context of the weak-coupling a strong input. This approach to classifying
assumption, which assumes that the phase reset- phase-resetting curves makes minimal assump-
ting due to two brief, sequential pulses summates tions, including pulse coupling, and is based on
linearly, and the phase resetting due to a single simple topological considerations.
82 Retinal/Visual Interfaces (Models, Theory, Techniques): Overview
In summary, phase response curves have been be included under the broad umbrella of a retinal/
defined for both infinitesimal and general inputs. visual interface. These include optogenetic
PRCs have been classified using separate, methods that use tissue engineering techniques
unrelated schemes based on bifurcation analysis to transfect remaining pathways in the visual
or topology. Finally, PRCs have been used to system with photosensitive properties. Simulated
analyze synchronization under assumptions of prosthetic vision, methods to assess prosthetic
either weak or pulsatile coupling. vision, and computation models of the neural
retina play important roles in increasing our
understanding of how such approaches function
Cross-References and can best be optimized.
▶ Phase Models, Noisy

▶ Phase Response Curve, Measurement and Detailed Description
Shape of General
▶ Phase Response Curve, Measurement of the Vision is arguably the most feature rich and com-
Infinitessimal plex of the senses, with visual cues being critical
▶ Phase Response Curve, Topology of to most activities of daily living. Vision impair-
▶ Pulse-Coupled Oscillators ment results in momentous personal and eco-
▶ Spike Time Response Curve nomic burdens to individuals and to society with
▶ Weak Coupling Theory global estimates of 285 million persons impacted
(Pascolini and Mariotti 2012).
Figure 1 illustrates the human visual path-
ways. Pathology or trauma to various elements
Retinal/Visual Interfaces (Models, of the visual pathway results in vision impairment
Theory, Techniques): Overview or profound vision loss. In retinal degenerative
diseases, such as retinitis pigmentosa (RP) and
Nigel H. Lovell age-related macular degeneration (AMD), the
Graduate School of Biomedical Engineering, photoreceptors in the retina progressively die. In
University of New South Wales, Sydney, NSW, response to such diseases, there can be large-
Australia scale reorganization of the retina with substantial
gliosis (Jones et al. 2003). Despite this, studies
have shown that human retinal ganglion cells
Synonyms (RGCs) maintain their viability after the onset
of these degenerative diseases, that the surviving
Artificial vision; Prosthetic vision; Visual retinal neurons are capable of being electrically
neuroprosthesis stimulated (Humayun et al. 1996), and that rudi-
mentary phosphene vision is achievable in
humans with advanced retinal degeneration.
Definition The etiology and/or site of insult to the visual
pathway will dictate the range of possible target
Retinal and visual interfaces encompass a range locations for a visual prosthesis (Guenther
of approaches and technologies with the most et al. 2012; Lovell et al. 2010). Possible sites for
common being that of a visual prosthesis, which introducing a visual prosthesis include retinal
is a subclass of sensory neuroprostheses. Such (epiretinal (Humayun et al. 2012), subretinal
technologies can be used as a device therapeutic (Zrenner et al. 2011; Palanker et al. 2005),
to restore some form of patterned vision to those suprachoroidal (Matteucci et al. 2013), trans-retinal
suffering from profound vision impairment. (Fujikado et al. 2012)), optic nerve (Veraart
Other approaches to vision restoration can also et al. 2003), and cortical (Fernandez et al. 2005)
Retinal/Visual Interfaces (Models, Theory, Techniques): Overview 83
Retinal/Visual Interfaces (Models, Theory, Tech- possible intervention sites for artificial vision are catego-
niques): Overview, Fig. 1 Right panel: an axial section rized by retinal, optic nerve, and cortical placement loca-
of the human brain outlining a trace of the visual pathway tions (Adapted from Lovell et al. (2010))
from the retina to the primary visual cortex. Other panels:
locations (Fig. 1). In cases of trauma and diseases targets with the implanted components remaining
such as glaucoma, the RGCs can also be securely fixed, even in the presence of rapid head
destroyed. As the RGC axonal processes form and eye movements performed by the recipient.
the optic nerve, in these cases electrical stimula- The device should maintain its position, integrity,
tion of visual pathways distal to the lateral genic- and functionality for several decades of everyday
ulate nucleus (LGN) of the thalamus is usage, cause minimal injury, inflammation, or
ineffective. risk of infection to nearby tissue and cause min-
In the case of retinal prostheses, to effectively imal discomfort to the recipient. Also important
replace vision with the same resolution as that of in visual prosthesis design are aspects of size and
normally sighted humans would require an image battery life in the case of external componentry.
capturing device to replace the function of the A typical vision prosthesis system comprises
photoreceptor cells, of which there are approxi- an external unit which using a camera and micro-
mately 125 million in each eye, converging to processor performs the image capturing and
some one million RGCs. Current retinal devices, processing and an implanted unit consisting of
depending on placement and design, have electrode the microelectronic stimulator and the electrode
numbers from tens to several thousand at most. array. Power and communication between the
The desired outcome is to have the electrode external and implanted units are normally facili-
array placed in close proximity to the neural tated by a transcutaneous radio-frequency
84 Software Tools for Modeling in Computational Neuroscience: Overview
(RF) link. The exceptions to this approach are Guenther T, Lovell NH, Suaning GJ (2012) Bionic vision:
typically subretinal devices that usually have system architectures – a review. Expert Rev Med
Devices 9(1):33–48
photodiodes designed into the implantable com- Humayun MA, de Juan E, Dagnelie G, Greenberg RJ,
ponent and thus have no need for an external Propst RH, Phillips DH (1996) Visual perception
camera. elicited by electrical stimulation of the retina in blind
Other more experimental approaches, in terms humans. Arch Ophthalmol 1141:40–46
Humayun MS et al (2012) Interim results from the inter-
of readiness for human trials, are based around national trial of Second Sight’s visual prosthesis. Oph-
optogenetic techniques (Degenaar et al. 2009). thalmology 119(4):779–788
This involves the use of viral transfection of Jones BW, Watt CB, Frederick JM, Baehr W, Chen CK,
rhodopsins to target various remaining cells in Levine EM, Milam AH, LaVail MM, Marc RE
(2003) Retinal remodeling triggered by photoreceptor
the visual pathway, making them photosensitive. degenerations. J Comp Neurol 464:1–16
Common to all device therapies and target Lovell NH, Morley JW, Chen SC, Hallum LE, Suaning GJ
locations in the visual pathway are a list of chal- (2010) Biological-machine systems integration: engi-
lenges that must be overcome to improve efficacy neering the neural interface. Proc IEEE 98(3):418–431
Matteucci PB, Chen SC, Tsai D, Dodds CW, Dokos S,
and ensure device safety and longevity. These Morley JW, Lovell NH, Suaning GJ (2013) Current
include maintaining a viable and stable neural steering in retinal stimulation via a quasimonopolar
interface over the long-term, device hermeticity stimulation paradigm. Invest Ophthalmol Vis Sci
and safe stimulation paradigms that allow con- 54(6):4307–4320
Palanker D et al (2005) Design of a high resolution opto-
current stimulation at numerous electrode sites. electronic retinal prosthesis. J Neural Eng 2:
S105–S120
Pascolini D, Mariotti SP (2012) Global estimates of visual
Cross-References impairment: 2010. Br J Ophthalmol 96(5):614–618
Veraart C, Wanet-Defalgue M, Gerard B, Vanlierde A,
Delbeke J (2003) Pattern recognition with the optic
▶ Computational Models of Neural Retina nerve visual prosthesis. Artif Organs 27(11):996–1002
▶ Prosthetic Vision, Assessment Zrenner E et al (2011) Subretinal electronic chips allow
▶ Prosthetic Vision, Perceptual Effects blind patients to read letters and combine them to
words. Proc Biol Sci 278(1711):1489–1497
▶ Retinal Disease and Remodeling
▶ Retinal Neurophysiology
▶ Retinal Prosthesis
▶ Visual Prosthesis, Cortical Devices
▶ Visual Prosthesis, Epiretinal Devices Software Tools for Modeling in
▶ Visual Prosthesis, Optic Nerve Approaches Computational Neuroscience:
▶ Visual Prosthesis, Optogenetic Approaches Overview
▶ Visual Prosthesis, Subretinal Devices
▶ Visual Prosthesis, Suprachoroidal and Trans- Padraig Gleeson
retinal Devices Department of Neuroscience, Physiology and
Pharmacology, University College London,
London, UK
References
Degenaar P, Grossman N, Memon MA, Burrone J, Detailed Description

Dawson M, Drakakis E, Neil M, Nikolic K (2009)
Optobionic vision: a new genetically enhanced light
on retinal prosthesis. J Neural Eng 6(3):035007 Modelling in computational neuroscience is
Fernandez E et al (2005) Development of a cortical visual becoming a crucial tool for understanding how
neuroprosthesis for the blind: the relevance of experimentally observed properties of neural sys-
neuroplasticity. J Neural Eng 2(4):R1–R12
Fujikado T et al (2012) Clinical trial of chronic implanta-
tems emerge from lower-level biophysical pro-
tion of suprachoroidal-transretinal stimulation system cesses. In the same way that processing of
for retinal prosthesis. Sensor Mater 24(4):181–187 information happens at multiple physical scales
Software Tools for Modeling in Computational Neuroscience: Overview 85
in the nervous system, software applications have transformations on inputs, including complex
been created which specialize in modelling dif- morphological structures and a host of
ferent aspects of neurons and networks. active membrane conductances. A number of
This overview deals primarily with simulators simulators have been developed which allow
of spiking neural networks, but other applications researchers to construct neuron models at
for modelling at lower levels (e.g., for biochem- this level of detail and potentially link such
ical signalling networks or reaction–diffusion cells together into networks inspired by
within spines or whole cells) or higher scale anatomical circuits. NEURON (http://www.
(e.g., models of cognitive processes) have also neuron.yale.edu/neuron) is a general purpose
been developed. There is also a focus on freely neuronal simulation environment, which allows
available, open-source applications. simulation of networks of neurons of varying
levels of detail, from artificial/abstract cells to
complex, morphologically detailed neurons.
Simulators A key feature of NEURON is its extension
language NMODL, which allows new ion
Abstract Neuron Simulations channel and synapse models to be defined. An
The emergent properties of networks are fre- important recent addition to NEURON’s func-
quently studied using highly simplified neurons tionality has been the ability to run networks
with complex network connectivity and synaptic across multiple processors. GENESIS (http://
dynamics. NEST (http://www.nest-initiative.org) genesis-sim.org) has also been widely used cre-
is a widely used simulator for large neuronal ating and analyzing detailed models of single
models (e.g., 100,000 neurons and millions of cells and networks. MOOSE (http://moose.
synapses). It is highly optimized to run in sourceforge.net) was initially based on GENESIS
a parallel computing environment. Another pop- version 2 but has been further developed with a
ular simulator of abstract neuronal networks is Python-based scripting interface, a new graphical
Brian (http://www.briansimulator.org), which is user interface, and greater support for interacting
Python based and allows definition of new neuron with standardized modelling languages
and synapse models by writing the equations for like SBML and NeuroML. PSICS (http://www.
their dynamics in a simple string-based format. psics.org) can simulate multicompartmental,
Topographica (http://topographica.org) is a sim- conductance-based cell models. Its particular
ulator which deals primarily with neural maps. It strength is allowing efficient simulation of sto-
facilitates investigation of the transformation of chastic ion channels using a kinetic scheme-
information between layers of topographic maps based approach.
as are found in sensory and motor systems. XPP
(http://www.math.pitt.edu/~bard/xpp/xpp.html) Reaction–Diffusion Modelling
is a widely used tool for simulation and analysis Modelling of the diffusion and reaction of
of dynamical systems and has frequently been biochemical substances in complex 3D
used to build and investigate neuronal systems. structures can be useful for helping to understand
Nengo (http://nengo.ca) is a simulator for large- the physical processes by which inputs are
scale neuronal systems. It has been used for cre- transformed in neurites and at synapses.
ating the Spaun (Semantic Pointer Architecture Some of the applications available to create
Unified Network) simulation (Eliasmith such models include MCell (http://www.mcell.
et al. 2012). cnl.salk.edu), STEPS (http://steps.sourceforge.
net) and NeuroRD (http://krasnow1.gmu.edu/
Conductance-Based, Multicompartmental CENlab/software.html). The NEURON
Simulators simulator has also recently added support for
Single neurons possess a range of features reaction–diffusion simulations (McDougal
enabling them to perform computational et al. 2013).
86 Software Tools for Modeling in Computational Neuroscience: Overview
Interoperability Frameworks Detailed neuronal morphologies can also be

automatically generated, based on data obtained
A diversity of simulation tools is beneficial from real neurons, for use in neuronal simula-
from the point of view of offering greater choice tions. TREES Toolbox (http://www.
to users and encouraging diverse approaches treestoolbox.org), CX3D (http://www.ini.uzh.
to neuronal simulation. There are a number of ch/~amw/seco/cx3d), NETMORPH (http://
initiatives which aim to make it easier to use netmorph.org), and NeuGen (http://atlas.gcsc.
computational models across these simulators. uni-frankfurt.de/~neugen) are just some of the
PyNN (http://neuralensemble.org/PyNN) is applications which can be used for this purpose.
a Python library for building neuronal networks, More details on initiatives for generating
which was inspired by an increasing number of artificial neurons which could be used in
neuronal simulators using the Python language neuronal simulations can be found in the entry
as a scripting interface. A neuronal network ▶ “Synthetic Neuronal Circuits/Networks”.
can be created using PyNN and then can be
simulated on multiple simulators. NeuroML
(http://www.neuroml.org) is an XML-based lan- Useful Resources
guage for describing models in computational
neuroscience. This has been mainly used to Other useful resources where models, experimen-
encode 3D networks of multicompartmental, tal data, and other software tools for computa-
conductance-based neurons, but the recently tional neuroscience can be obtained include:
developed version 2.0 increases its scope to • ModelDB (http://senselab.med.yale.edu/
more abstract neuron models. NineML (http:// modeldb): an archive of simulation scripts
software.incf.org/software/nineml) is another for published models in the format originally
initiative to create an XML language for describ- used by the model developers.
ing spiking neural networks. neuroConstruct • Open Source Brain (http://www.
(http://www.neuroconstruct.org) is a graphical opensourcebrain.org): a resource for sharing
application which allows the construction of and collaboratively developing neuronal
complex 3D networks of biophysically detailed models. Models reside in open-source reposi-
neurons. Simulation scripts can then be automat- tories, and reusing, modifying, and converting
ically generated to execute the model in different the models to standardized formats such as
simulation environments. NeuroML and PyNN are actively encouraged.
• Channelpedia (http://channelpedia.epfl.ch)
is a resource developed by the Blue Brain
Neuronal Morphology Databasing Project which provides structured information
and Generation on ion channels, and many of its entries have
downloadable computational models of the
Many modelling studies use detailed neuronal channels.
morphologies to examine how information is • NeuroElectro (http://www.neuroelectro.org)
processed and transformed by single neurons, provides structured information on electro-
through active membrane conductances and physiological properties of neurons obtained
integration of synaptic input. Neuromorpho from the literature.
(http://neuromorpho.org) is the primary resource • NeuralEnsemble (http://neuralensemble.org) is
for obtaining neuronal morphologies which have a resource which aims to promote open, collab-
been digitally reconstructed. It has contributions orative software development in computational
from many labs worldwide and covers a wide neuroscience and hosts a number of related
range of species and neuron types. software packages (e.g., Brian and PyNN).
Software Tools for Modeling in Computational Neuroscience: Overview 87
Other Software for Neuronal Simulation (http://www.ebi.ac.uk/biomodels-main) and the

CellML Model Repository (http://models.
Cognitive Process Modelling cellml.org).
While most of the simulation packages men-
tioned already use spiking neuron models,
a significant amount of modelling work takes Future Developments
place in computational neuroscience using
population-based models and mean-field The following are a number of recently developed
approaches to model large-scale cognitive pro- simulation platforms which promise new features
cesses. Examples of software packages which and possibilities for neuronal simulation. Geppetto
allow this type of modelling are MIIND (http:// (http://www.geppetto.org) is a web-based multi-
miind.sourceforge.net), a simulator for high-level algorithm, multi-scale simulation platform
cognitive modelling, and The Virtual Brain designed to support the simulation of complex
(http://www.thevirtualbrain.org), which provides biological systems and their surrounding environ-
a simulator and a web-based interface for ment. It is open source and is being developed as
constructing neural population models. The part of the OpenWorm (http://www.openworm.
Brain Operation Database System (http://bodb. org) project to create an in silico model of the
usc.edu/bodb) provides structured information nematode C. elegans, though through its support
on a number of published models in the areas of for NeuroML, it should be suitable for many other
cognitive and systems neuroscience. types of neuronal models. The Neural Tissue Sim-
ulator (Kozloski and Wagner 2011) is a novel
Hardware-Based Modelling Solutions approach to large-scale neuronal simulation,
In addition to software-based neuronal simula- focusing on simulating a whole block of neural
tors, many groups are investigating hardware- tissue by efficiently segmenting the 3D regions
based solutions to enable faster and larger-scale into sections, each of which is simulated on
neuronal simulations. Many of these use off-the- a separate processor in a parallel environment. It
shelf hardware like Graphics Processing Units promises efficient simulation of networks based on
(GPUs), e.g., NeMo (http://nemosim. data from connectomics initiatives.
sourceforge.net) and GeNN (http://sourceforge.
net/projects/genn), but other initiatives are devel-
oping new hardware, customized to simulate Conclusions
large-scale neuronal networks, e.g., SpiNNaker
(http://apt.cs.man.ac.uk/projects/SpiNNaker). The development of tools for modelling in com-
putational neuroscience is a dynamic field. The
Systems Biology/Bioinformatics tools overview here has no doubt left out a number of
The Systems Biology community has been simulators and applications which would be of use
actively developing applications and model to the wider community. The author encourages
description languages in recent years to facilitate developers to get in contact with details of their
building and exchanging models of biochemical work for inclusion in future versions of this article.
reactions, signalling pathways, and gene regula-
tory networks. SBML (Systems Biology Markup
Language, http://sbml.org) and CellML (https:// References
www.cellml.org) are two widely used standards
Eliasmith C, Stewart TC, Choo X, Bekolay T, DeWolf T,
in this area, and many models in these Tang Y, Rasmussen D (2012) A large-scale model of
formats, including neuronal models, can the functioning brain. Science 338(6111):1202–1205.
be obtained from the BioModels database doi:10.1126/science.1225266
88 Somatosensory System: Overview
Kozloski J, Wagner J (2011) An ultrascalable solution Thermoreception

to large-scale neural tissue simulation.
Front Neuroinform 5:15. doi:10.3389/
fninf.2011.00015 Our ability to sense whether an object is warm or
McDougal R, Hines M, Lytton W (2013) Reaction–diffusion cold relies on two types of thermoreceptive
in the NEURON simulator. Front Neuroinform 7:28. fibers – so-called “cold” and “warm”
doi:10.3389/fninf.2013.00028 fibers – embedded in the skin. As their names
suggest, “cold” fibers are activated when the skin
is cooled and “warm” fibers are activated when the
skin is warmed. In contrast to thermosensitive
nociceptive fibers, which respond at extreme tem-
Somatosensory System: Overview peratures, thermoreceptive fibers only respond at
intermediate, non-noxious temperatures (with the
Sliman J. Bensmaia exception of the paradoxical response of some
Department of Organismal Biology and cold fibers to high temperatures).
Anatomy, University of Chicago, Chicago,
IL, USA
Touch
Detailed Description The sense of touch plays a critical role in our
ability to grasp and manipulate objects. Indeed,
Somatosensation includes multiple senses: pain cutaneous signals provide information about the
(nociception), temperature (thermoreception), forces we exert on objects and whether these are
touch, and the sense of our limb position in slipping from our grasp. Without these signals,
space (proprioception). Each submodality of we would routinely crush or drop objects. Touch
somatosensation relies on different types of also plays an important role in emotional com-
receptors embedded in the skin, muscle, and munication: We touch the people we care about
joints and involves different structures in the and wish to be touched by them. Finally, our
spinal cord and in the brain. sense of touch plays a key role in embodiment,
making our body feel as a part of us.
The skin is innervated by several types of
Pain mechanoreceptive afferents, each of which con-
veys different information about skin deforma-
Pain is arguably one of the most vital senses as it tions (link: sensory innervation of the skin) and
signals when our body is liable to being damaged. conveys different types of information about
There are many different types of receptors in the events impinging upon the skin. Merkel cells
skin that signal a potentially harmful stimulus. convey information about the shape of objects
Some receptors respond to intense mechanical grasped in the hand (link: cutaneous mechanore-
deformations of the skin, others to extreme tem- ceptive afferents: neural coding of shape),
peratures, and still others to different kinds of Meissner corpuscles about motion of objects
chemicals. Pain comprises a sensory discrimina- across the skin, and Pacinian corpuscles about
tive component, which provides information surface texture (link: cutaneous mechanorecep-
about the location, duration, intensity, and quality tive afferents: neural coding of texture). Affer-
of the pain; an affective one, which signals its ents produce highly repeatable and temporally
unpleasantness; and a cognitive-evaluative one, patterned responses to skin stimulation (link:
which is associated with cognitive variables such somatosensory neurons: spike timing), and
as attention, which can modulate the sensory models have been developed that predict with
experience. high accuracy the responses of somatosensory
Spectral Methods in Neural Data Analysis: Overview 89
neurons to spatiotemporal skin deformations ▶ Somatosensory Cortex: Neural Coding of

(link: mechanotransduction: models). Shape
When we palpate an object, we obtain infor- ▶ Somatosensory Cortex: Organization
mation about its shape (link: somatosensory cor- ▶ Somatosensory Neurons: Spike-Timing
tex: neural coding of shape), its texture (link:
cutaneous mechanoreceptive afferents: neural
coding of texture), and its motion across the
skin (link: somatosensory cortex: neural coding Spectral Methods in Neural Data
of motion). Analysis: Overview
Steven L. Bressler
Proprioception Cognitive Neurodynamics Laboratory, Center for
Complex Systems and Brain Sciences,
Proprioception (link: proprioception) plays Department of Psychology, Florida Atlantic
a critical role in guiding motor behavior. Individ- University, Boca Raton, FL, USA
uals with intact motor systems but compromised
proprioception have difficulty planning and exe-
cuting movements, almost as if they had a motor Detailed Description
impairment. Proprioception, like touch, is also
important for our sense of embodiment. Spectral analysis is a powerful and widely used
There are several types of proprioceptive approach to the study of time series data (Warner
receptors, located in muscles, in the skin, and in 1998; Bloomfield 2000). It provides a useful
joint capsules. Two types of muscle propriocep- complement to other types of analysis in compu-
tors, muscle spindles and Golgi tendon organs, tational neuroscience. Spectral analysis refers to
are thought to be the primary contributors a host of techniques relating to transformed time
to proprioception. One population of receptors series in the frequency domain. A spectral repre-
in the skin is sensitive to skin stretch and can sentation of a time series is a function of fre-
convey information about joint angle. Another quency, where frequency is expressed in units
type of proprioceptor, the joint capsule receptor, of cycles per second, or hertz (Hz). Although
fires at the extreme ends of the joint’s range and spectral analysis is applicable to deterministic
may be involved in preventing overextension of time functions, neural data is typically stochastic
the joint. and thus requires statistical spectral analysis
Proprioceptive and cutaneous signals are then (Brillinger 2001; Bendat and Piersol 2010). Neu-
processed in the dorsal column nuclei, then in the ral data types that are subjected to spectral anal-
ventroposterior lateral nucleus in the thalamus, ysis commonly include continuous time series
and then in primary and secondary somatosen- such as the electroencephalogram (EEG),
sory cortices (link: somatosensory cortex: magnetoencephalogram (MEG), electrocortico-
organization). gram (ECoG), and local field potential (LFP)
but may also include point process time series
Cross-References such as single-unit and multiunit spiking activity
(Glaser and Ruchkin 1976; Dumermuth and
▶ Cutaneous Mechanoreceptive Afferents: Molinari 1987; Hesselmann 1991).
Neural Coding of Texture Spectral methods in neural data analysis pro-
▶ Mechanotransduction, Models vide a frequency-based representation of neural
▶ Proprioception time series data. They may be used to identify
▶ Somatosensory Cortex: Neural Coding of discrete, narrowband oscillations in time
Motion series or to decompose a broadband time series
90 Spectral Methods in Neural Data Analysis: Overview
into frequency-specific components. Spectral of use of the complex Cartesian representation

methods are also employed in the frequency- make complex algebra a useful mathematical
based analysis of neural data arrayed in space tool in spectral analysis. However, complex
rather than time (see “▶ Spatial Spectral algebra is not necessary. The polar and complex
Analysis”). Spatial spectra are representations Cartesian representations are equivalent in
of spatial data, and spatial frequency is expressed providing a complete description of time series
in units of cycles per unit distance. data, and Fourier synthesis can reconstruct the
The traditional approach to spectral analysis is time series from either one.
based on the pioneering work of Joseph Fourier Fourier analysis is often used to quantify
(1768–1830), the French mathematician who is interdependency relations between time series
credited with having first introduced the repre- that are derived from different sources (see
sentation of a mathematical function as a sum of “▶ Spectral Interdependency Methods”).
trigonometric functions. Nowadays, Fourier Frequency-based interdependency measures
methods play a major role in a multitude of appli- may be derived from the complex-valued
cations in mathematics, science, and engineering. cross-spectrum. The cross-spectrum yields two
Fourier analysis refers to the linear decomposi- important real-valued spectra. First is the rela-
tion of a function into elemental trigonometric tive phase (i.e., difference in phase) spectrum,
basis functions, whereas Fourier synthesis is the given by the arctan of the ratio of imaginary to
operation of rebuilding the original function real components of the cross-spectrum. Second
from the component basis functions. Although is the normalized modulus of the cross-
functions subjected to Fourier analysis are spectrum, the coherence spectrum, which is
often continuous and potentially infinite in the frequency domain equivalent of the time
length, they may also be discrete and finite domain cross-correlation function. The relative
in length. The wide availability of methods for phase spectrum reflects the mean, and the
Fourier analysis of digitized finite-length time coherence spectrum reflects the variance, of
series on digital computers has made spectral the distribution of relative phase values of two
data analysis practical, rapid, and inexpensive time series. Coherence is roughly equivalent to
(Marple 1987). As a result, spectral analysis of the phase-locking value, defined as one minus
neural data has become increasingly popular in the circular variance of relative phase, in that
modern times, with a growing list of applications they both reflect relative phase variation. The
in theoretical, experimental, and clinical difference is that coherence additionally reflects
neuroscience. (to a lesser degree) amplitude covariation. Spec-
Fourier analysis of a time series assigns tral methods are used to examine neural
values of amplitude and (absolute) phase to interdependency relations between continuous
the trigonometric basis functions (sines and signals (e.g., between LFPs), or discrete
cosines) of the decomposition. For univariate signals (e.g., spiking activities), or between con-
time series, i.e., from a single recording chan- tinuous and discrete signals (e.g., spike-field
nel, Fourier analysis specifies a range of fre- coherence).
quencies of the component basis functions, Neural time series data are commonly multi-
along with amplitude (or power, squared ampli- variate, being derived from multiple sources, and
tude) and phase values for each frequency com- their spectral analysis often involves the assess-
ponent. This (polar) representation in terms of ment of pairwise interdependency relations in
amplitude and phase has an equivalent the frequency domain. Some interdependency
(Cartesian) representation in terms of the real relations between time series are directional,
and imaginary components of a complex meaning that the interdependency is directed
quantity. The simplicity, symmetry, and facility from one time series to another. Directed spectral
Spectral Methods in Neural Data Analysis: Overview 91
methods quantify directed interdependency in the spectral methods are usually required for neural
frequency domain (see “▶ Directed Spectral time series analysis (Jenkins and Watts 1968;
Methods”). Percival and Walden 1993). Statistical
Because neural time series data commonly spectral analysis considers time series data to be
evolve over time, time-frequency, or spectro- generated by stationary stochastic (random) pro-
graphic, representations are often useful cesses, and the various spectral quantities are
(see “▶ Time-Frequency Analysis”). This type treated as data-derived statistical estimates of
of representation, which allows one to indepen- unknown population variables, rather than as
dently examine the temporal evolution of differ- deterministic quantities. Unlike nonparametric
ent frequency components, may be used to track spectral analysis, which computes spectra
the time-frequency evolution of a range of directly from time series data by Fourier analysis,
time-varying neural processes, including sleep, parametric spectral analysis is an approach that
sensory, motor, and cognitive processing. derives spectral quantities from a statistical
Spectrographic representations of the electroen- model of the time series (see “▶ Parametric Spec-
cephalogram (EEG) are heavily used in neuro- tral Analysis”). In the model, the variable at
pharmacology to assess the actions of a particular time is expressed by statistical rela-
neuroactive drugs. A variety of different time- tions with variables from past times. The paramet-
frequency methods, such as the short-time ric model is typically autoregressive, meaning that
Fourier transform, Wigner-Ville distribution, each time series value is modeled as a weighted
Cohen’s class distribution, or wavelet transform sum of past values (the weights being considered
(see “▶ Wavelet Analysis”), have been applied in as the parameters of the model) (Kay 1988;
neural data analysis. Chatfield 2004). Although parametric models
One important application of spectral methods may be nonlinear, those used in neuroscience are
in neural data analysis is filtering, the transforma- typically linear. Parametric modeling has some
tion of neural time series by the selection of certain distinct advantages in neural data analysis: it
frequency components and the exclusion of others allows a precise time-frequency representation of
(see “▶ Digital Filtering”). Digital filtering, which time-varying neural time series (Ding et al. 2000),
refers to filtering operations performed on dis- and it serves as a theoretically sound basis for
cretely sampled data in a digital computer, has directed spectral analysis (Ding et al. 2006).
the advantage of allowing data at both past and Spectral methods of neural data analysis uti-
future time points to be used in determining the lize frequency-based representations. Since
filtered value of a current time point, unlike analog rhythmic activity, mostly in time but also in
filtering, where the filtering operation depends space, is ubiquitous in neuroscience, frequency-
only on past time points. Although digital filtering based techniques are very important tools in com-
may be performed in the time domain, it is more putational neuroscience. Given that many good
easily performed in the frequency domain by the software options exist for performing spectral
selection of a range of frequencies (the passband) data analysis, these techniques are readily avail-
and exclusion of another range (the stopband). able to neuroscience researchers working with
Data smoothing is an important operation in neural many different types of neural data.
data analysis that is accomplished by digital
(low-pass) filtering. Digital filtering is also used
to separate spike and field signals recorded from Cross-References
a common microelectrode.
Neural time series are rarely deterministic, ▶ Digital Filtering
where each time series value is exactly deter- ▶ Directed Spectral Methods
mined by past values. For this reason, statistical ▶ Parametric Spectral Analysis
92 Spike Train Analysis: Overview
▶ Phase-Locking Methods
▶ Spatial Spectral Analysis Spike Train Analysis: Overview
▶ Spectral Interdependency Methods
▶ Time-Frequency Analysis Sonja Gr€un
▶ Wavelet Analysis Lab for Statistical Neuroscience, Institute of
Neuroscience and Medicine (INM–6) and
Institute for Advanced Simulation (IAS–6), J€ulich
References Research Centre and JARA, J€ulich, Germany
Theoretical Systems Neurobiology, RWTH
Bendat JS, Piersol AG (2010) Random data: analysis and Aachen University, Aachen, Germany
measurement procedures, 4th edn. Wiley, Hoboken
Bloomfield P (2000) Fourier analysis of time series.
Wiley, New York
Brillinger D (2001) Time series: data analysis and theory. Detailed Description
SIAM, Philadelphia
Chatfield C (2004) The analysis of time series: an intro- The brain is composed of billions of neurons, the
duction. Chapman & Hall/CRC, Boca Raton
Ding M, Bressler SL, Yang W, Liang H (2000)
elementary units of neuronal information
Short-window spectral analysis of cortical event- processing. The neocortex, which is critical to
related potentials by adaptive multivariate most higher brain functions, is a highly complex
autoregressive modeling: data preprocessing, model network of neurons each of which receives sig-
validation, and variability assessment. Biol Cybern
83:34–45
nals from thousands of other neurons and projects
Ding M, Chen Y, Bressler SL (2006) Granger causality: its own output via sequences of spikes (“▶ Spike
basic theory and application to neuroscience. Train”) to thousands of other neurons
In: Schelter B, Winterhalder M, Timmer J (eds) (Braitenberg and Sch€uz 2009). In order to
Handbook of time series analysis: recent theoretical
developments and applications. Wiley-VCH,
observe neuronal activity in the active brain,
Weinheim a large variety of recording techniques are being
Dumermuth G, Molinari L (1987) Spectral analysis of employed, ranging from recordings of individual
EEG background activity. In: Gevins AS, Remond A neurons (intra- or extracellularly), to recordings
(eds) Methods of analysis of brain electrical and mag-
netic signals. Handbook of electroencephalography
of neuronal populations on mesoscopic or mac-
and clinical neurophysiology, revised series, roscopic scales. Any particular choice of the
vol 1. Elsevier, Amsterdam recording technique reflects the hypothesis the
Glaser EM, Ruchkin DS (1976) Principles of researcher has in mind about the mechanisms of
neurobiological signal analysis. Academic, New York
Hesselmann NL (1991) The fundamentals of discrete
neuronal processing. The focus on spike record-
Fourier analysis. In: Weitkunat R (ed) Digital ings from individual neurons implies that one
biosignal processing. Elsevier, Amsterdam/New York strives to understand the elementary units of neu-
Jenkins GM, Watts DG (1968) Spectral analysis and its ronal processing. However, approaching the rela-
applications. Holden-Day, San Francisco
Kay SM (1988) Modern spectral estimation: theory and
tionship of different signal types may provide
application. Prentice-Hall, Englewood Cliffs a means of relating processing on different spatial
Marple SL (1987) Digital spectral analysis with applica- scales (“▶ Spike Triggered Average”).
tions. Prentice-Hall, Englewood Cliffs Early electrophysiological experiments were
Percival DB, Walden AT (1993) Spectral analysis for
physical applications: multitaper and conventional
bound to record from single neurons only. The
univariate techniques. Cambridge University Press, resulting insights are now the basis for the “clas-
Cambridge sical” view of sensory coding: firing rates are
Warner RM (1998) Spectral analysis of time-series data. modulated in a feed-forward hierarchy of
The Guilford Press, New York
processing steps. Signals from sensory epithelia
are assumed to eventually converge to cortical
Further Reading
Wikipedia
detectors for certain combinations of stimulus
Stochastic (random) process http://en.wikipedia.org/wiki/ features (“▶ Neural Coding”) or finally trans-
Stochastic_process ferred to motor output. Specific percepts or
Spike Train Analysis: Overview 93
motor actions would be represented by the ele- concepts for the analysis of multiple parallel pro-
vated firing of a single nerve cell (“▶ Estimation cesses (“▶ Gravity Analysis of Parallel Spike
of Neuronal Firing Rate”) or by changes of firing Trains”; “▶ Unitary Event Analysis”; “▶ Infor-
rates of groups of neurons which can be assessed mation Geometry as Applied to Neural Spike
by different conceptual approaches (“▶ Neuronal Data”; “▶ Spatial Temporal Spike Pattern
Population Vector”; “▶ Population Encoding/ Analysis”; “▶ Generalized Linear Models for
Decoding”; “▶ State-Space Models for the Point Process Analyses of Neural Spiking
Analysis of Neural Spike Train and Behavioral Activity”). The evaluation of significance
Data”). (“▶ Significance Evaluation”) is a basic element
An alternative concept of neuronal processing in these analyses, in some cases based on nonpara-
by groups of neurons was suggested by Donald metric methods (“▶ Surrogate Data for Evaluation
Hebb (1949) who first demonstrated the concep- of Spike Correlation”). Analysis of parallel spike
tual power of a brain theory based on cell assem- trains (Gr€un and Rotter 2010; Kass et al. 2014) is
blies. Inspired by Hebb and driven by more recent the logical next step to improve our understanding
physiological and anatomical evidence in favor of the neuronal mechanisms underlying informa-
of a distributed network hypothesis, brain theo- tion processing in the brain.
rists constructed models that rely on groups of
neurons, rather than single nerve cells, as the
functional building blocks for representation Cross-References
and processing of information. Despite concep-
tual similarities, such concepts of neuronal ▶ Correlation Analysis of Parallel Spike Trains
cooperativity differ in their detailed assumptions ▶ Estimation of Neuronal Firing Rate
with respect to the spatiotemporal organization of ▶ Generalized Linear Models for Point Process
the neuronal activity. Analyses of Neural Spiking Activity
To understand the principles of coordinated ▶ Gravity Analysis of Parallel Spike Trains
neuronal activity and its spatiotemporal scales, ▶ Information Geometry as Applied to Neural
it is obligatory to observe the activity of multiple Spike Data
single neurons simultaneously. Due to recent ▶ Joint Peri Stimulus Time Histogram (JPSTH)
technological developments in recording meth- ▶ Neural Coding
odology, this can regularly be done. Coordinated ▶ Neuronal Population Vector
activity of neurons is only visible in (wide-sense) ▶ Population Encoding/Decoding
correlations of their respective spike trains, ▶ Significance Evaluation
which typically admit no simple interpretation ▶ Spatial Temporal Spike Pattern Analysis
in terms of fixed synaptic wiring diagrams. ▶ Spike Train
Rather, it became evident that the correlation ▶ Spike Train Distance
dynamics apparent in time-resolved multiple- ▶ Spike Triggered Average
channel measurements reflect variable and ▶ State-Space Models for the Analysis of Neural
context-dependent coalitions among neurons Spike Train and Behavioral Data
and groups of neurons. Thus, the analysis of ▶ Surrogate Data for Evaluation of Spike
simultaneously recorded spike trains allows us Correlation
to relate concerted activity of ensembles of neu- ▶ Unitary Event Analysis
rons to behavior and cognition. Different analysis
approaches are thereby relevant to distinguish
different or even complementary spatiotemporal
References
scales based on methods ranging from pairwise
analysis (“▶ Spike Train Distance”; “▶ Correla- Braitenberg V, Sch€uz A (2009) Cortex: statistics and
tion Analysis of Parallel Spike Trains”; “▶ Joint geometry of neuronal connectivity, 2nd edn. Springer,
Peri Stimulus Time Histogram (JPSTH)”) to New York. ISBN-13: 978-3540638162
94 Spiking Network Models and Theory: Overview
Gr€un S, Rotter S (2010) Analysis of parallel spike trains. to explain the occurrence of action potentials
In: Gr€un S, Rotter S (eds) Springer series in computa- (Lapicque 1907; Tuckwell 1988). Lapicque’s
tional neuroscience. Springer, New York. ISBN 978-
1-4419-5674-3, e-ISBN 978-1-4419-5675-0 model can therefore be seen as the first integrate
Hebb D (1949) The organization of behavior. Wiley, and fire model. In 1936, Arthur Hill extended
New York Lapicque’s model by adding an adaptive thresh-
Kass RE, Eden U, Brown EN (2014) Analysis of neural old (Hill 1936). These models are phenomeno-
data Kass RE, Eden U, Brown EN (eds) Springer series
in statistics. Springer, New York. ISBN-13: 978- logical since they did not explain the biophysical
1461496014 mechanisms producing action potentials. In the
1950s Hodgkin and Huxley explained with a
series of experiments that action potentials are
caused by ionic currents which result from ion
channels with voltage-dependent conductances
Spiking Network Models and Theory: (Hodgkin and Huxley 1952). But because of its
Overview simplicity, Lapicque’s original threshold model
is still used by experimental and theoretical
Marc-Oliver Gewaltig neuroscientists.
Blue Brain Project, École Polytechnique Fédéral Theoretical network models initially focused
de Lausanne, Lausanne, Switzerland on explaining experimentally observed inter-
spike interval distributions of individual neurons
as well as the origin of the observed high vari-
Definition ability of neuronal responses. Recent theoretical
research mainly focuses on understanding
Spiking neuronal networks are a type of neural low-rate spontaneous activity as well as
network model where the neurons interact by multistability in the context of learning and
sending and receiving the so-called spikes, short memory.
pulses that are only defined by their time of Theoretical analysis of spiking neuronal net-
occurrence. Biologically, spikes correspond to works is only possible if the network model is
the action potentials of neurons. sufficiently simple. In recent years, computer
Neuron models that produce spikes are called simulations have become increasingly important
spiking neuron models. Examples are the ▶ Inte- to support and complement theoretical analysis,
grate and Fire Models, Deterministic; the because they allow the inclusion of more biolog-
Izhikevich model; and the ▶ Hodgkin-Huxley ical detail and the analysis of heterogeneous net-
Model. work architectures.
The term spiking network was introduced to
distinguish these models from formal neuron Spiking Neuron Models
models which have graded activation functions. Theoretical network models typically resort to
simplified neuron models, such as the leaky inte-
grate and fire model (LIF), some variant of it, or
Detailed Description more abstract models such as pulse-coupled
oscillators.
Historical Background All of these neuron models have in common
The first spiking neuron models were developed that they interact by sending and receiving the
at the beginning of the twentieth century and so-called spikes which are abstractions of the
focused on explaining the electrical behavior of action potentials produced by excitable cells
isolated neurons. In 1907, Louis Lapicque pro- such as neurons.
posed an electrical circuit model to describe the Spikes are temporal point events, character-
change in membrane potential after applying a ized by their time of occurrence ^t and mathemat-
current step. He assumed a fixed firing threshold ically expressed by the Dirac delta function d(t).
Spiking Network Models and Theory: Overview 95
n o
A sequence of spikes S ¼ ^t , ^t , . . . is called
1 2
Immediately after a spike, neurons are unable
spike train and can then be written as to produce another spike for a short period of time
tr, called refractory period. Many models imple-
X
sðtÞ ¼ d t ^t ;
k
(1) ment the refractory period by holding the mem-
k brane potential at its reset value, such that
V(t) = Vreset for t (t0 , t0 + tr].
where the index k runs over all spikes in S.
Threshold Adaptation
Integrate and Fire Models Many neurons adapt to sustained input by reduc-
The leaky integrate and fire (LIF) model consti- ing their firing rate. In integrate and fire models,
tutes a whole class of models consisting of two this is often captured by a mechanism called
parts. The first part consists of one or more dif- threshold adaptation.
ferential equations, describing the subthreshold The spike threshold Vy is then no longer con-
behavior of the neuron’s membrane potential. stant, but increases with every spike by an
The second part converts the continuous mem- amount a and decays back to its resting value y
brane potential into discrete spike events. Here with a time constant ty:
we will present the LIF model in its most basic
form and then discuss a number of common d X
d t t0
k
variations. ty Vy ¼ y Vy þ a (4)
dt k
Subthreshold Dynamics
The subthreshold membrane potential V is where the sum runs over all endogenous spikes t0 k
described by the differential equation of the neuron.
d Synaptic Input
t V ðtÞ ¼ V 0 V ðtÞ þ RI ðtÞ (2) Synaptic input enters the membrane potential in
dt
the form of the synaptic current Isyn(t).
where V0 is the resting potential, t the membrane A presynaptic spike at time ^t causes a conduc-
time constant, and R the membrane resistance. tance change at the postsynaptic membrane,
I(t) = Isyn(t) + Iext(t) + . . . is the total current which in turn results in a postsynaptic current of
across the cell membrane and comprises all exter- the form
nal influences on the neuron, such as synaptic

currents Isyn and externally applied currents Iext. I syn ðt, V Þ ¼ Esyn V gsyn ðt ^tÞ (5)
Spike Generation where Esyn is the reversal potential of the synapse

Spikes are generated, when the membrane poten- and gsyn(t) the time course of the conductance
tial V crosses a threshold value Vy from below. change.
The time of spike is then given by the relation Many models implicitly assume that the syn-
aptic conductance changes much faster that the
d membrane potential V. Then V in Eq. 5 can be
V ðt0 Þ ¼ V y and V ðt0 Þ > 0 (3) treated as constant, and the synaptic current no
dt
longer depends on the membrane potential which
In the following, we use t0 to denote the times of greatly simplifies analytical and numerical
endogenous spikes and ^t to denote input spikes analysis.
from other neurons. In recent years, the terms COBA and CUBA
After a spike is generated, the membrane have become popular to distinguish the two
potential is reset to Vreset, a value which is often cases. COBA refers to the original Eq. 5 and
chosen to match the resting potential V0. stands for conductance based, while CUBA
stands for current based and refers to simplified For theoretical analysis of large spiking net-
synaptic currents of the form work, the spike-response model is more conve-
nient than the original LIF model, since it
I syn ðtÞ ¼ pscðt ^tÞ (6) captures the effects of incoming as well as self-
generated action potentials in a closed mathemat-
where the postsynaptic current psc(t) is only a ical form (Gerstner and Kistler 2002).
function of time (Vogels and Abbot 2005).
Since Eq. 2 is a linear differential equation Stochastic Spike Generation and Escape Noise
with constant coefficients, its solution for arbi- For most neuron models with a deterministic
trary postsynaptic currents psc(t) is given by spike threshold, it is also possible to use a sto-
chastic spike generation mechanism by adding
ð1
ðt sÞ the so-called escape noise (Plesser and Gerstner
pspðtÞ ¼ V 0 þ Hðt sÞexp pscðsÞds
0 t 2000). This is done by relating the variable V(t) to
(7) the probability density for observing an
action potential in the infinitesimal time interval
where H(t) is the Heaviside step function and (t, t + dt] (Plesser and Gerstner 2000; Mensi
psp(t) is the so-called postsynaptic potential. et al. 2012). In this picture, the spike generation
In the case where the postsynaptic current is a is a random point process with a conditional
delta function pscðtÞ :¼ dðt ^tÞ, Eq. 7 reduces to density function

ðt ^tÞ V ðtÞ V y
pspðtÞ ¼ V 0 þ exp : (8) lðtjV, V y Þ ¼ l0 exp ; (11)
t DV
In response to a spike train s(t), we obtain where l0 is a scaling factor with unit s1, Vy is the
firing threshold, and DV is a factor that deter-
0 1
mines the steepness of the exponential function
X t ^tk
H t ^t exp@ A:
k
pspðtÞ ¼ V 0 þ and therefore also the sensitivity of the firing
k
t threshold to small fluctuations.
(9) The spike-response model with escape noise
belongs to the larger class of generalized linear
Spike-Response Model models (GLM) which have recently been used
The spike-response model (SRM) is a generali- successfully also in other areas of neuroscience
zation of the LIF model (Gerstner et al. 1996a; (see, e.g., Pillow et al. (2008), Truccolo
Gerstner and Kistler 2002). It exploits the linear- et al. (2011)).
ity of Eq. 2 and expresses the membrane potential
as convolution: Stochastic Models of Neural Activity
ð1 In many regions of the brain, neurons fire con-
stantly at a low rate, even if they are not directly
V ðtÞ ¼ ðt ^tÞ þ kðt, t0 ÞI syn ðt t0 Þdt0 (10)
0 stimulated. In the cortex, this spontaneous activ-
ity is very irregular. Moreover, even if a cortical
where (t t0 ) is a kernel that describes the action neuron in vitro or in vivo is stimulated repeatedly,
potential as well as the after hyperpolarization its response will differ for each stimulus
of the last spike of the neuron at t ¼ t0 . kðt, ^t Þ is presentation.
a kernel that describes the response of the mem- One of the earliest approaches to explain this
brane potential to a presynaptic spike at time ^t . variability of neuronal firing is to consider the
Like in the LIF model, a spike is generated when fluctuation of the membrane potential of a neu-
the membrane potential V(t) crosses a threshold ron, caused by randomly arriving spikes from the
value Vy from below. surrounding network. In the simplest case, each
ð þ1
neuron in the surrounding network will fire at
some constant rate n. The network, thus, gener- Var½V t ¼ nðt sÞpsp2 ðsÞds (16)
0
ates a noisy background activity which then influ-
ences the firing probability of our neuron. When an embedding into a cortical region is
The number of excitatory and inhibitory considered, random input comes from excitatory
spikes which arrive at neuron i during the interval and inhibitory populations. Due to the linearity of
(0, t] can be written as the equations, the contributions from both
populations superimpose, and the mean and var-
X ðt iance of the combined membrane potential are
nE ðtÞ ¼ sj ðt0 dÞdt0 ,
0 given by
X ðt
jEi
nI ðtÞ ¼ sj ðt0 tÞdt0 E½V t ¼ E½V E t þ E½V I t (17)

jI i 0
and
where Ei and Ii denote the sets of all excitatory
and inhibitory neurons projecting to neuron i,
respectively. The membrane potential of a Var½V t ¼ Var½V E t þ Var½V I t0 (18)
neuron can then be written as (Stein 1965;
Tuckwell 1988) respectively.
dV 1 dnE dnI Connections

¼ V þ JE þ JI : (12)
dt t dt dt A network consists of a set of N neurons and
their connections. In a connection between two
For sufficiently large networks, we can assume neurons, the sending neuron is called presynaptic
that nE and nI are Poisson processes with rates nE and the receiving neuron is called postsynaptic.
and nI, respectively. We write the set of presynaptic neurons of
In the absence of a threshold, and for constant neuron i as Ni.
firing rates, the mean and variance of the mem-
brane depolarization are (Tuckwell 1988; Stein Static Connections
1965) In the simplest case, the synaptic current simply
adds the spikes of all presynaptic cells:
E½V ¼ tðJ E nE þ J I nI Þ (13)
X
I syn, i ðtÞ ¼ J ij sj t d ij (19)
1 jN i
Var½V ¼ t J 2E nE þ J 2I nI : (14)
2 X X
d t d ij ^tj
k
¼ J ij (20)
More generally, we can describe the randomly jN i k¼1
arriving spikes as a shot noise process (Papoulis
and Pillai 2002). This allows us to compute the where dij is the propagation delay between neu-
ron j and neuron i and ^tj the kth spike of neuron
k
mean and variance of the resulting membrane
potential as a function of the background firing j. Jij is the weight or efficacy of the connection
rate n(t) and the synaptic response kernel psp(t). between neurons j and i which is in many cases
The mean membrane potential is then assumed to be constant.
ð þ1
Dale’s Principle
E½V t ¼ nðt sÞpspðsÞds (15)
0 Dale’s principle states that the efferent synapses
of a neuron are all of the same type. Thus, if one
with variance efferent synapse of a neuron is excitatory, we
know that all other efferent synapses of this neu- An ¼ Aun Rn ; (21)
ron will also be excitatory. The same applies to
inhibition. with initial values
Dale’s principle introduces correlations into
the connectivity matrix of the network, which
ui :¼ U; (22)
are visible in the eigenvalue distribution of the
matrix in the complex plane (Rajan and Abbot
2006). Moreover, Dale’s principle also affects R1 :¼ 1: (23)
the correlation structure of random spiking
networks. In random networks without Dale’s u and R are then iteratively updated according to
principle, correlations between neurons will
vanish in the limit of infinitely large networks. Dtn
unþ1 ¼ U þ un ð1 U Þexp ; (24)
In such networks, Dale’s principle prevents F
the correlations from disappearing (Kriener
et al. 2008). Dtn
Rnþ1 ¼ 1 þ ðRn Rn un 1Þexp ; (25)
D
Short-Term Plasticity
where Dtn :¼ ^t ^t :
n n1
The weight of a connection can depend on the
spike history of the presynaptic neuron such that The relation of the time constants D and
the weight decreases or increases if several pre- F determines whether a synapse will be depress-
synaptic spikes arrive in short succession. These ing, facilitating, or a combination of the two.
effects are called short-term depression (STD) However, closer analysis of the equations as
and short-term facilitation (STF), respectively. well as experimental results shows that the
Both can be described by a system of kinetic detailed behavior of dynamic synapses depends
equations that model the depletion and also on the spike frequency of the presynaptic
replenishing of synaptic resources (Tsodyks neuron (Fuhrmann et al. 2002).
et al. 1998, 2000). Dynamic synapses have a strong effect on the
Assume that there is a finite amount of behavior of a network. Networks with facilitating
resources available and that each spike uses a synapses can show collective synchronization of
certain fraction of these resources which are all neurons, called population bursts (Tsodyks
then unavailable for a certain amount of time. et al. 2000). By contrast, in networks where the
Depleted resources are replenished at a facilitating and depressing synapses are distrib-
constant rate D. The second factor in the uted as indicated by experimental results, activity
model is the ability of a synapse to use is more stable as the synapses exert some gain
its resources, expressed by a variable u(t). In control on the network (Sussillo et al. 2007).
stochastic models of synaptic transmission, Short-term plasticity has also been linked to
u corresponds to the release probability of a the persistence of network states with elevated
synaptic release site (Fuhrmann et al. 2002). If firing rates. These are thought to play an impor-
u is a constant, the synapse will be depressing. tant role in working memory (Mongillo
For facilitating synapses, u increases with each et al. 2008).
spike up to a maximum and decays at a constant Short-term plasticity is not related to Hebbian
rate F. plasticity, since it depends only on the activity of
In the context of the STP literature, the max- the presynaptic neuron rather than on the activity
imal synaptic efficacy is usually denoted by of both the pre- and postsynaptic neurons.
A which corresponds to the synaptic weight J,
used throughout the rest of this entry. The effec- Spike-Timing-Dependent Plasticity
tive weight for the nth spike can then be Spike-timing-dependent plasticity (STDP) is a
expressed as form of synaptic plasticity where the synaptic
efficacy changes according to the relative timing precedes the spike of the postsynaptic neuron,
of pre- and postsynaptic action potentials. Dt is positive and Eq. 27 will increase the weight
A typical experimental protocol for spike- W by a value that is largest for short and smallest
timing-dependent plasticity involves two neurons for large intervals. If the postsynaptic neuron
with a synaptic connection. Action potentials are spikes first, Dt is negative and Eq. 28 will
induced in both neurons in a defined temporal decrease W.
sequence by injecting current pulses. At the There are a number of variations to this model
same time, the strength of the synaptic potential as well as a number of alternative models which
is measured in the postsynaptic neuron. This pro- incorporate hypothesized biophysical mecha-
cedure is called pairing. After a number of nisms, underlying STDP (Morrison et al. 2008;
pairings, each with the same timing relation Sjöström and Gerstner 2010). But so far,
between pre- and postsynaptic neurons, a change the experimental evidence does not allow to
in the amplitude of the PSP can be observed. identify one of the models as authoritative
Systematic variation of the relative timing (Feldman 2012).
between pre- and postsynaptic action potentials Spike-timing-dependent plasticity is a mecha-
then reveals that the change in PSP amplitude nism that acts in addition to the synaptic short-
depends on the timing relation. term plasticity, described in the previous section.
The classical results of Markram et al. (1997) To obtain the final synaptic efficacy which
and Bi and Poo (1998) show that potentiation is combines both short-term and spike-timing-
largest if the postsynaptic neuron spikes shortly dependent plasticity, we replace A in Eq. 21
after the presynaptic neuron, while the synapse with J + DJ:
became depressed if the postsynaptic neuron fired
shortly before the presynaptic neuron. This is An ¼ un Rn ðJ þ DJ Þ (30)
captured by the following model (Gerstner
et al. 1996b): where J is the equilibrium weight.
XX
DJ ij ¼ W ^t ^t
i j
(26) Network Topologies
^t Sj ^t Si
j i In a spiking network, the weighted, directed
graph of the connections defines the topology of
the network. Common network topologies are
where the sums run over all pre- and postsynaptic
feedforward networks, recurrent networks, and
spikes, respectively. W(Dt) is the so-called learn-
networks with spatial topologies.
ing window, with
Feedforward Networks
Dt
W ðDtÞ ¼ Aþ exp for Dt > 0; (27) Feedforward networks can be broken down into
tþ
disjunct groups or layers of neurons G1, G2, . . .
where each neuron in group Gi is only connected
Dt
W ðDtÞ ¼ A_exp for Dt < 0; (28) to neurons in group Gj with j i.
t_
The propagation of spiking activity in
feedforward networks is directed and in the direc-
and
tion of ascending group numbers. As a result, the
total number of spikes in feedforward networks is
Dt :¼ tpost tpre : (29) limited, and at any given time, only a fraction of
the neurons are active (Griffith 1963).
The learning window is modeled as an exponen-
tial function which scales the degree to which the Synfire Chains
weight changes, depending on the time interval Examples of feedforward networks are the com-
Dt. If the spike of the presynaptic neuron plete transmission line, also known as synfire
chain (Griffith 1963; Abeles 1991). It consists of a reaching task. Thus, neighboring locations in
l mutually exclusive groups Gi, with 1 i, Gl, stimulus or response coordinates are mapped to
each containing w neurons. Each neuron in group neighboring neurons.
i projects to a certain number of neurons in group The simplest examples of topological net-
i + 1, thus forming a chain of groups of neurons. If works are line or ring networks, where the neu-
sufficiently many neurons in the first group fire rons are aligned on a one-dimensional axis. If the
near simultaneously, they will ignite the neurons two ends of the axis are connected back to each
in the second group and so on. The spikes of the other, the line turns into a circle. This is useful for
first group will therefore travel along the chain representing periodic coordinates such as orien-
until either the activity disperses or the chain tation angle or simply to avoid boundary effects
comes to an end. Theoretically, this can be at the end of the line.
described by the so-called pulse packets, propagat- Models of visual processing often use
ing from one group to the next (Diesmann two-dimensional sheets or layers of neurons.
et al. 1999). In the ideal case, that is, given a Here the coordinates can correspond to visual
sufficient number of spikes with a sufficiently nar- space, e.g., the center of the neuron’s receptive
row temporal spread, the pulse packet will propa- field or cortical space, that is, the actual position
gate unchanged from one group to the next. And of the neuron in the brain. To ameliorate
due to the divergent-convergent connectivity, processing at the boundaries of the sheet, the
small deviations from this invariant shape are two pairs of opposing sides are often connected
then automatically repaired. However, depending such that the rectangular sheet turns into a torus.
on the number of neurons in a group and the Typical models of visual processing will com-
connections between the groups, the pulse packet bine several sheets into a functional architecture
may also win or lose spikes or change its temporal (e.g., Masquelier and Thorpe 2007; Grossberg
precision (Diesmann et al. 1999; Câteau and and Versace 2008). Topological networks can
Fukai 2001). be feedforward, recurrent, or a combination of
the two.
Recurrent Networks
Networks with feedback connections are called Network Dynamics
recurrent. Recurrent networks are often random Spiking neural network model are an important
with mixed excitation and inhibition. In random tool to study the dynamics of cortical network
networks with uniform connection probability, activity and to understand physiologically
each neuron has an equal probability to be observed phenomena such as spontaneous activ-
connected to another neuron. In other models, ity, neuronal synchronization, and network oscil-
the connection probability between neurons lations. These are macroscopic network
i and j depends on their distance. phenomena which we call the macro state of the
network. Each macro state usually comprises a
Neighborhood Preserving Topologies large ensemble of micro states, defined by the
Networks in which neurons have a well-defined individual states of all neurons and synapses.
position are called topology networks. In such Many macro states, observed in cortical net-
networks, the connection probability of two neu- works, can be understood, using a simple network
rons usually depends on their distance. For exam- with randomly connected excitatory and inhibi-
ple, the closer two neurons are, the higher is their tory neurons (Brunel 2000). The model consists
probability to be connected. The neuronal posi- of three populations of neurons: one population
tions may correspond to the actual positions of of excitatory neurons, one population of inhibi-
the neurons within a brain region, but they may tory neurons, and a third population of excitatory
also correspond to positions in some abstract neurons, representing the long-range input from
space, for example, the orientation angle of a other brain regions. The excitatory and inhibitory
visual stimulus or the direction of movement in populations are mutually coupled, and both
receive input from the background population. train. Then the spikes of all neurons will occur
We can now study the activity patterns which in synchrony, but the intervals between the spikes
emerge, depending on the ratio of inhibition to will be irregular. However, this state is very
excitation and the strength of the background unlikely to occur in random networks because it
population. requires a well-chosen connectivity.
If the amount of excitation and inhibition is
roughly equal so that on average the respective Acknowledgments This work was supported by the Blue
synaptic currents cancel each other, we speak Brain Project and EU grant FP7-269921 (BrainScaleS).
of balanced network. In this regime, network
activity is highly irregular, since spikes are gen-
erated by the difference in fluctuations of the Cross-References
excitatory and inhibitory currents (Tsodyks and
Sejnowski 1995). ▶ Bayesian Inference with Spiking Neurons
If all neurons fire independently of each other ▶ Excitability: Types I, II, and III
and their spike patterns are essentially random, ▶ Fitzhugh–Nagumo Model
we speak of asynchronous-irregular activity. ▶ Hodgkin-Huxley Model
This state is of particular interest, because it cor- ▶ Integrate and Fire Models, Deterministic
responds best to the state of low-rate spontaneous ▶ Morris–Lecar Model
activity, observed in cortical networks. ▶ Multistability in Neurodynamics: Overview
Networks in the asynchronous-irregular state ▶ Neuronal Avalanches
exhibit chaotic activity (van Vreeswijk and ▶ Pulse-Coupled Oscillators
Sompolinsky 1996). Thus, even small perturba- ▶ Recurrent Network Models, Reservoir
tion such as adding or removing one spike will Computing
quickly result in a completely different micro ▶ Spike-Frequency Adaptation
state. Whether this critical dependence on the ▶ Spike-Timing Dependent Plasticity, Learning
initial conditions is useful or detrimental for Rules
spike-based processing is still under debate ▶ Spontaneous Activity, Models of
(Izhikevich 2006; London et al. 2010). ▶ Theta Neuron Model
If the neurons fire independently, but each at
regular intervals, we speak of asynchronous-
regular activity. In this state, each neuron oscil- References
lates at a different frequency, and occasionally,
large parts of the network will fire synchronously. Abeles M (1991) Corticonics: neural circuits of the cere-
bral cortex. Cambridge University Press, Cambridge
The intervals of these population events depend
Bi GQ, Poo MM (1998) Synaptic modifications in cul-
on the common multiples of the individual firing tured hippocampal neurons: dependence on spike
periods. It is also possible that all neurons fire at timing, synaptic strength, and postsynaptic cell type.
the same frequency, but out of phase. In this case, J Neurosci 18(24):10464–10472
Brunel N (2000) Dynamics of sparsely connected net-
neurons will not be able to occasionally synchro-
works of excitatory and inhibitory spiking neurons.
nize their spikes. J Comput Neurosci 8(3):183–208
If all neurons fire synchronously with the same Câteau H, Fukai T (2001) Characteristic futures of pulse
frequency, the network essentially oscillates and packet propagation in synfire chain revealed by
Fokker-Planck method. Neural Netw 14:675–685
we speak of synchronous-regular activity. Oscil- Diesmann M, Gewaltig MO, Aertsen A (1999) Stable
latory states are also of interest since they can propagation of synchronous spiking in cortical neural
serve as models for epileptic activity or for networks. Nature 402(6761):529–533
stimulus-evoked synchronization. Feldman DE (2012) The spike-timing dependence of plas-
ticity. Neuron 75(4):556–571
Theoretically, there is also a state of
Fuhrmann G, Segev I, Markram H, Tsodyks MV
synchronous-irregular activity, a state where (2002) Coding of temporal information by activity-
each neuron produces the same irregular spike dependent synapses. J Neurophysiol 87(1):140–148
Gerstner W, Kistler WM (2002) Spiking neuron models: Sjöström J, Gerstner W (2010) Spike-timing dependent
single neurons, populations, plasticity. Cambridge plasticity. Scholarpedia 5(2):1362
University Press, Cambridge Stein R (1965) A theoretical analysis of neuronal variabil-
Gerstner W, Hemmen JV, Cowan J (1996a) What matters ity. Biophys J 5(2):173–194
in neuronal locking? Neural Comput 8:1653–1676 Sussillo D, Toyoizumi T, Maass W (2007) Self-tuning of
Gerstner W, Kempter R, van Hemmen J, Wagner H neural circuits through short-term synaptic plasticity.
(1996b) A neuronal learning rule for sub-millisecond J Neurophysiol 97(6):4079–4095
temporal coding. Nature 383(6595):76–78 Truccolo W, Donoghue JA, Hochberg LR, Eskandar EN,
Griffith JS (1963) On the stability of brain-like structures. Madsen JR, Anderson WS, Brown EN, Halgren E,
Biophys J 3:299–308 Cash SS (2011) Single-neuron dynamics in human
Grossberg S, Versace M (2008) Spikes, synchrony, and focal epilepsy. Nat Neurosci 14(5):635–641
attentive learning by laminar thalamocortical circuits. Tsodyks MV, Sejnowski T (1995) Rapid state switching in
Brain Res 1218:278–312 balanced cortical network models. Netw Comput Neu-
Hill A (1936) Excitation and accommodation in nerve. ral Syst 6(2):111–124
Proc R Soc Lond B 814:305–355 Tsodyks M, Pawelzik K, Markram H (1998) Neural net-
Hodgkin A, Huxley A (1952) A quantitative description of works with dynamic synapses. Neural Comput
membrane current and its application to conduction 10(4):821–835
and excitation in nerve. J Physiol 117(4):500–544 Tsodyks MV, Markram H, Uziel A (2000) Synchrony
Izhikevich EM (2006) Polychronization: computation generation in recurrent networks with frequency-
with spikes. Neural Comput 18(2):245–282 dependent synapses. J Neurosci 20(1):50
Kriener B, Tetzlaff T, Aertsen A, Diesmann M, Rotter Tuckwell HC (1988) Introduction to theoretical neurobi-
S (2008) Correlations and population dynamics in ology: volume 2 nonlinear and stochastic theories.
cortical networks. Neural Comput 20(9):2185–2226 Cambridge University Press, Cambridge, UK
Lapicque L (1907) Recherches quantitatives sur l’excita- van Vreeswijk C, Sompolinsky H (1996) Chaos in neuro-
tion electrique des nerfs traitee comme une polariza- nal networks with balanced excitatory and inhibitory
tion. J Physiol Pathol Gen 9:620–635 activity. Science 274(5293):1724–1726
London M, Roth A, Beeren L, Hausser M, Latham PE Vogels T, Abbot LF (2005) Signal propagation and logic
(2010) Sensitivity to perturbations in vivo implies high gating in networks of integrate-and-fire neurons.
noise and suggests rate coding in cortex. Nature J Neurosci 25(46):10786–10795
466(7302):123–127
Markram H, L€ubke J, Frotscher M, Sakmann B (1997)
Regulation of synaptic efficacy by coincidence of post- Further Reading
synaptic APs and EPSPs. Science 275:213–215 A thorough introduction to the theory of spiking networks
Masquelier T, Thorpe SJ (2007) Unsupervised learning of can be found in the somewhat dated but still highly
visual features through spike timing dependent plas- valuable textbooks Introduction to theoretical neuro-
ticity. PLoS Comput Biol 3(2):e31 biology by Tuckwell (1988). An equally thorough and
Mensi S, Naud R, Pozzorini C, Avermann M, Petersen more recent reference is the book Spiking neuron
CCH, Gerstner W (2012) Parameter extraction and models: Single neurons, populations, plasticity by
classification of three cortical neuron types reveals Gerstner and Kistler (2002) which also contains exten-
two distinct adaptation mechanisms. J Neurophysiol sive treatment of learning and plasticity in spiking
107(6):1756–1775 networks. A broader overview is given in the
Mongillo G, Barak O, Tsodyks M (2008) Synaptic textbook Theoretical Neuroscience by Dayan and
theory of working memory. Science (NY) Abbot (2001)
319(5869):1543–1546 Burkitt AN (2006) A review of the integrate-and-fire neu-
Morrison A, Diesmann M, Gerstner W (2008) Phenome- ron model: I. Homogeneous synaptic input. Biol
nological models of synaptic plasticity based on spike Cybern 95(1):1–19
timing. Biol Cybern 98(6):459–478 Burkitt AN (2006) A review of the integrate-and-fire neu-
Papoulis A, Pillai SU (2002) Probability, random vari- ron model: II. Inhomogeneous synaptic input and net-
ables and stochastic, 4th edn. McGraw-Hill, Boston work properties. Biol Cybern 95(2):97–112
Pillow JW, Shlens J, Paninski L, Sher A, Litke AM, Burkitt (2006a, b) has written a set of comprehensive
Chichilnisky EJ, Simoncelli EP (2008) Spatio- reviews of the integrate and fire neuron. The dynamic
temporal correlations and visual signalling in a com- properties of spiking networks have been reviewed by
plete neuronal population. Nature 454(7207):995–999 Vogels et al. (2005)
Plesser HE, Gerstner W (2000) Noise in integrate-and-fire Dayan P, Abbot LF (2001) Theoretical neuroscience:
neurons: from stochastic input to escape rates. Neural computational and mathematical modeling of neural
Comput 12(2):367–384 systems. MIT Press, Cambridge, MA
Rajan K, Abbot LF (2006) Eigenvalue spectra of random Feldman (2012) provides an extensive review of spike-
matrices for neural networks. Phys Rev Lett timing-dependent plasticity, with possible underlying
97(18):188104 synaptic and cellular mechanisms, as well as its
Spinal and Neuromechanical Integration: Overview 103
potential role in learning. The reviews of Morrison In an alternate perspective, peripheral systems
et al. (2008) and Sjöström and Gerstner (2010) give might simplify for motor control. The complexi-
good overview over theoretical models of spike-
timing-dependent plasticity ties of spinal systems or nonlinear properties of
Vogels TP, Rajan K, Abbot LF (2005) Neural network the musculoskeletal system might reflect adapta-
dynamics. Annu Rev Neurosci 28:357–376 tions that allow simplified control by descending
systems. For example, passive mechanics can be
used to assist movements(Collins et al. 2005),
muscle properties can contribute to stability
Spinal and Neuromechanical (Nichols and Houk 1973), basic reflexes allow
Integration: Overview for rapid control (Loeb et al. 1999), and defining
adaptive muscle coordination patterns can poten-
Matthew Tresch1 and Devin L. Jindrich2 tially simplify movement (Tresch and Jarc 2009).
1
Department of Biomedical Engineering and In this perspective, energetic costs associated
Physical Medicine and Rehabilitation, with inefficient, complex control might lead to
Northwestern University, Evanston, IL, USA evolutionary adaptations that simplify control
2
Department of Kinesiology, California State and neural processing.
University, San Marcos, CA, USA These two perspectives are not mutually
exclusive, and both emphasize the importance
of understanding the contribution of peripheral
Detailed Description systems to motor control and brain function.
Entries in other sections of this Encyclopedia
To interact with the environment or other organ- and the Encyclopedia of Neuroscience describe
isms, the nervous system must move. Whether it is the basic physiological properties of muscles (see
a fundamental protective reflex, a stabilizing pos- Heckman et al. 2009), motor units (see Burke
tural adjustment, a rapid prey strike, or an expres- 2009), reflexes (see Grey and Nielsen 2009),
sive dance gesture, the motor repertoire of an and pattern generating spinal networks (see
organism defines the nature of its environmental ▶ Vertebrate Pattern Generation: Overview).
interactions. Structures in the periphery of the The entries in this section expand on those
motor system, the musculoskeletal system and spi- descriptions, characterizing how these systems
nal cord, most directly mediate these environmen- function and how they interact with descending
tal interactions. The actions of central brain regions systems. Spinal interneuronal systems have been
such as cortex, cerebellum, or basal ganglia all studied extensively for more than a century, and
ultimately have to pass through these peripheral although much remains unknown about their role
structures. Understanding the properties of these in behavior, these studies have provided basic
peripheral systems is therefore critical for our information that can be used to guide computa-
understanding of the neural control of movement. tional studies of motor control (▶ Spinal Cord,
How should we consider these peripheral Integrated (Non CPG) Models of). Propriocep-
components of the motor system? In one common tors provide the nervous system with information
perspective, these systems are problems that the about body state – the position of the limb, the
CNS must overcome. In this perspective, the forces produced by muscles, and the interactions
complex properties of muscles, limb mechanics, with the environment. Understanding how these
motor neurons, and spinal circuits require that state variables are encoded in the activity of pro-
central motor systems develop strategies that prioceptor afferents is critical in understanding
invert, bypass, or suppress these complexities. what information the nervous system has
Complexity could therefore lead to greater com- access to when interpreting and interacting with
plexity: evolutionary changes to the periphery the environment (▶ Proprioceptor Models).
could require the co-evolution of more complex Although spinal systems are capable of a great
mechanisms to maintain performance. deal of motor coordination on their own,
104 Spinal Interfaces: Overview
including central pattern generators for basic pro- References

tective reflexes and for locomotion, voluntary
behaviors are accomplished by descending path- Burke RE (2009) Motor units. In: Binder MD, Hirokawa
N, Windhorst U (eds) Encyclopedia of neuroscience.
ways from the brain including the cortex. How
Springer, Berlin Heidelberg, pp 2443–2446
these systems interact with one another to pro- Collins S, Ruina A, Tedrake R, Wisse M (2005) Efficient
duce movement, however, remains poorly under- bipedal robots based on passive-dynamic walkers. Sci-
stood. One central computational issue in these ence 307:1082–1085
Georgopoulos AP, Grillner S (1989) Visuomotor coordi-
interactions is how information about task goals
nation in reaching and locomotion. Science
(e.g., the location of food to be grasped) is trans- 245:1209–1210
lated into motor commands (e.g., the muscle acti- Grey MJ, Nielsen JB (2009) Integration of Spinal
vations that result in movement of the arm and Reflexes. In: Binder MD, Hirokawa N, Windhorst U
(eds) Encyclopedia of neuroscience. Springer, Berlin
hand to the food) (Coordinate Transformations,
Heidelberg, pp 1985–1988
Role of Spinal Circuitry in). Heckman CJ, Perreault E, Sandercock T, Maas H (2009)
Ultimately motor control is the result of Muscle. In: Binder MD, Hirokawa N, Windhorst U
a single dynamic system including both the (eds) Encyclopedia of neuroscience. Springer, Berlin
Heidelberg, pp 2479–2487
nervous system and body, coupled together
Loeb GE, Brown IE, Cheng EJ (1999) A hierarchical
through intrinsic sensory feedback also through foundation for models of sensorimotor control. Exp
dynamic interactions with the environment. Brain Res 126:1–18
Neuromechanics is the field of study that seeks to Nichols TR, Houk JC (1973) Reflex compensation for
variations in the mechanical properties of a muscle.
understand how these two systems work together
Science 181:182–184
to produce behavior. Entries in this section Tresch MC, Jarc A (2009) The case for and against muscle
examine the neuromechanics of postural control synergies. Curr Opin Neurobiol 19:601–607
(Neuromechanics of Postural Control), character-
izing how neural and musculoskeletal systems
interact to stabilize the body to maintain upright
stance, and the neuromechanics of joint coordina- Spinal Interfaces: Overview
tion during locomotion (Neuromechanics of Joint
Coordination), characterizing how task goals are Michel Lemay
accomplished through dynamic coordination of Department of Bioengineering, Temple
low level execution variables. University, Philadelphia, PA, USA
Cross-References Definition
▶ Coordinate Transformations, Role of Spinal Spinal stimulation refers to the activation of the
Circuitry in spinal cord tissue (white matter axonal tracts
▶ Decision-Making, Motor Planning and/or gray matter neurons) or associated spinal
▶ General Overview of Spinal Anatomy and roots (see “▶ General Overview of Spinal Anat-
Physiology Organization omy and Physiology Organization” in this Ency-
▶ Motoneurons and Neuromuscular Systems: clopedia) using various stimulation technologies
Overview to restore function lost to neurological impair-
▶ Neuromechanics of Joint Coordination ment or study spinal circuitry organization.
▶ Neuromechanics of Postural Control Methodologies to activate spinal tissue can be
▶ Neuromuscular Control Systems, Models of divided on the basis of applications (bladder,
▶ Proprioception movement restoration, pain reduction, etc.),
▶ Proprioceptor Models stimulation or electrode technology employed
▶ Spinal Cord, Integrated (Non CPG) Models of (electrical, optical, etc.), electrode position (e.g.,
▶ Vertebrate Pattern Generation: Overview epidural, intradural, intraspinal, etc.), or
Spinal Interfaces: Overview 105
a number of other characteristics. In this entry, actually placed within the central nervous system
spinal cord activation techniques are presented or in the cerebrospinal fluid (CSF). Epidural stim-
based on the degree of invasiveness, and the ulation has been employed and is still currently
similar technological methods used to deliver employed in a number of applications. Sacral
stimulation to the neural tissue are highlighted. anterior root stimulation combined with posterior
In increasing order of invasiveness, spinal cord sacral rhizotomy has been used extensively to
stimulation techniques can be divided into restore bladder and bowel continence and voiding
(1) transcutaneous, (2) epidural, (3) intradural or following spinal cord injury (Brindley 1988).
intrathecal, and (4) intraspinal. Although most electrodes are implanted
intradurally around the sacral roots, extradural
electrode leads are sometimes used to activate
Detailed Description the anterior sacral roots (typically S2, S3, and
S4) (Egon et al. 1998). The anterior root stimula-
Transcutaneous Stimulation tor takes advantage of the different contraction
Transcutaneous stimulation refers to the activa- and relaxation times of the sphincter and bladder
tion of spinal tissue using the same principle as muscles. While the muscle fibers that constitute
electrical stimulation but through a magnetic field the sphincter are mainly “fast-twitch” muscle
which induces ionic currents within the spinal fibers that fatigue easily and relax very quickly,
tissue (see “▶ Paraspinal Magnetic and Transcu- the fibers that compose the detrusor muscles are
taneous Electrical Stimulation” entry in this “slow-twitch” fibers that contract gradually and
Encyclopedia). The basic methodology involves maintain force for a longer period of time. With
inducing depolarization via a time-varying mag- intermittent short bursts of high-frequency elec-
netic field delivered transcutaneously by a coil trical pulses to the sacral nerve roots, bladder and
applied to the skin surface. The principal applica- sphincter muscles contract, but while the sphinc-
tions have been for the initiation of locomotion ter relaxes rapidly at the end of each stimulus
following spinal cord injury or in parkinsonian burst, the pressure in the bladder is still suffi-
patients and for the treatment of pain when stim- ciently high to produce urination. Activation of
ulation is delivered over the dorsal columns (see the S2–S4 roots with longer stimulus trains is
“▶ Paraspinal Magnetic and Transcutaneous used to obtain defecation in a number of patients.
Electrical Stimulation” entry in this Encyclope- Variations of the basic principles are still
dia). Finite element model studies of the spinal employed today, although advances in our
cord indicate that the greatest site of activation is knowledge of urogenital anatomy and electrode
within the tracts due to the lower threshold of design have led to more sophisticated stimulation
activation of myelinated fibers of the white matter schemes (see “▶ Methodologies for the Restora-
versus the cells of the gray matter (see “▶ Finite tion of Bladder and Bowel Functions as well as
Element Models of Transcutaneous Spinal Cord ▶ Intraspinal Stimulation” in this Encyclopedia).
Stimulation” entry in this Encyclopedia). Since Epidural stimulation has also been used exten-
activation is typically delivered dorsally, the larg- sively for the restoration of locomotion following
est site of depolarization typically involves poste- spinal cord injury in various animal models and
rior roots and dorsal columns, although the humans (see “▶ Epidural Stimulation” entry in
anterior motor roots are also hot spots of this Encyclopedia). In all species, stimulation is
depolarization. delivered over the dorsal columns with the most
efficacious segments for initiating locomotion
Epidural Stimulation varying between species. The stimulus train is
Epidural stimulation refers to electrical stimula- typically a low-frequency (40 Hz) train of
tion delivered via leads inserted within the verte- pulses delivered continuously over the dorsal
bral canal but outside the dura. Surgery is columns at lumbar levels to produce activation
involved, but no stimulating electrodes are of the locomotor circuitry. The exact mechanisms
106 Spinal Interfaces: Overview
of activation or even the neuronal composition of technique in terms of the potential damage it
the locomotor centers remains unknown, but the may cause to the nervous system. Stimulation is
technique has been used clinically to restore delivered within the spinal cord, typically via
standing and locomotor movements in two spinal microwires implanted within the white and gray
cord injury individuals with complete motor matter, although new optogenetic techniques are
paralysis below the level of injury (Edgerton being developed in animal models for spinal cord
and Harkema 2011; Harkema et al. 2011). Epidu- applications (Alilain et al. 2008). Applications of
ral stimulation has also been employed to facili- intraspinal stimulation have included locomo-
tate the initiation of locomotion in parkinsonian tion, reaching and grasping, bladder and bowel
patients (Fuentes et al. 2010; also see “▶ Spinal functions, and respiration (see “▶ Intraspinal
Stimulation for Parkinson Treatment” in this Stimulation” in this Encyclopedia).
Encyclopedia).
Finally, epidural stimulation of the dorsal col- Electrode Technologies: Issues and Potential
umns has been applied for the treatment of Solutions
chronic and intractable pain arising from condi- The electrodes used to deliver stimulation to the
tions such as failed back surgery syndrome, phan- spinal cord are for the most part similar to the
tom limb pain, spinal cord injury, diabetic ones used to deliver electrical stimulation to
neuropathy, ischemic limb pain, etc. (Kumar peripheral nerves. They include cuff electrodes
et al. 1998; Ramasubbu et al. 2013). Stimulation that use metalized contacts (typically platinum)
is delivered through electrode leads placed over encapsulated within silicone or other materials.
the dorsal columns in an attempt to provide anal- These electrodes are typically used to stimulate
gesia for a number of various chronic pain con- roots (Brindley-Finetech anterior root stimulator
ditions. Success rate in the long-term relief of electrode, and see Xiao et al. 2012 for more
pain varies depending on the studies ranging advanced designs). Metal disk or lead wires are
from about 20 % to 80 % (Kumar et al. 1998). typically used to deliver stimulation epidurally,
and microwires are used to deliver stimulation
Intradural Stimulation intraspinally. New electrodes being developed
Up in the degree of invasiveness is intradural or and tested in animal models aim to increase the
intrathecal stimulation where leads are placed number of contacts and stimulation sites which
within the dural sac. Surgical and infection risks would allow for greater selectivity in the delivery
increase as the implant is now placed within the of stimulation. These include electrode designs
central nervous system space and CSF, thereby for intraspinal microstimulation (Snow
breaching the blood-brain barrier, with dural clo- et al. 2006) and multi-contact surface electrodes
sure remaining a surgical difficulty. Closer prox- for epidural stimulation (Gad et al. 2013).
imity to the cord provides greater access to Intraspinal microstimulation remains one of the
specific spinal structures allowing stimulation to toughest applications for spinal cord injuries as
be delivered to particular roots, and it is no sur- the stiff metal- or silicone-based electrode pre-
prise that the Brindley-Finetech (described in sents a mechanical impedance mismatch com-
Epidural Stimulation above) sacral anterior root pared to the higher compliance of the spinal
electrodes are mostly implanted intradurally over tissue, leading to tissue damage. In addition, the
the divided (anterior separated from posterior) spinal cord shifts significantly within the verte-
sacral roots to restore bladder and bowel bral canal during movements (Ranger
functions. et al. 2008), and the tension it produces in the
cable attaching electrode to associated connector
Intraspinal Stimulation may cause further damage to the tissue, unless the
Finally, intraspinal stimulation provides the cable’s flexibility is sufficient to dissociate the
greatest access to the spinal cord but is also the motions of the vertebral canal and spinal cord.
most invasive and thus riskiest stimulation A promising approach to minimizing mismatches
Spinal Interfaces: Overview 107
in tissue and electrode material impedances is the Wells et al. 2005a, b; Xia et al. 2014). The meth-
braided electrode design of the Giszter group (see odology is termed optical stimulation, and action
“▶ Braided Electrodes” entry in this Encyclope- potential initiation appears to be caused by the
dia). By braiding extremely fine wires, a very induction of thermal transient within the axons or
high overall compliance electrode and cabling neurons (Wells et al. 2007). Depth of penetration
technology is produced that can be inserted of the light is a limiting factor attempting to
within the spinal cord with the help of activate deep tissue, and to date the technology
a removable cannula guide and produce minimal has not been used in the spinal cord. All these
histological damage to the central nervous tissue stimulation methods offer the advantage of pro-
after long-term implantation (Kim et al. 2013a). ducing no electrical stimulation artifacts which
Wireless stimulators may also offer freedom facilitates neural recordings as the amplifiers, and
from the cable tethering problem and potentially analog-to-digital converters used in recordings
dura resealing. Miniaturized stimulators can be are no longer saturated by the large electrical
implanted close to the neural target and con- signal generated by the stimulus pulses.
trolled via wireless approaches (radio frequency,
optical, or acoustic waves, etc.). The approach
can have the benefits of a cable-free approach
Cross-References
like the transcutaneous stimulators but will
allow much more precise targeting of the stimu-
▶ Biomechanical Model of Low Back Pain
lation delivery (see “▶ Wireless Microsti-
▶ Braided Electrodes
mulators” entry in this Encyclopedia).
▶ Epidural Stimulation
Other stimulation techniques have been used
▶ Finite Element Modeling for Extracellular
experimentally to activate the spinal cord, but to
Stimulation
date no clinical application of these techniques
▶ Finite Element Modeling of Electrical
has been pursued. Bizzi and colleagues have used
Stimulation Using Microelectrodes
iontophoresis to deliver an excitatory neurotrans-
▶ Finite Element Models of Transcutaneous
mitter locally within the gray matter (Saltiel
Spinal Cord Stimulation
et al. 1998), but since electrodes are pulled glass
▶ General Overview of Spinal Anatomy and
micropipettes, applications are limited to cases
Physiology Organization
where the spinal cord can be securely
▶ Intraspinal Stimulation
immobilized. Optogenetic methods have also
▶ Methodologies for the Restoration of Bladder
been used to activate spinal tissue in the rat ani-
and Bowel Functions
mal model (see “▶ Intraspinal Stimulation” entry ▶ Methodologies for the Treatment of Pain
in this Encyclopedia). Stimulation is usually
▶ Micro-Wires
achieved with delivery of light from the surface,
▶ Paraspinal Magnetic and Transcutaneous
but delivery of light deep into tissue is possible Electrical Stimulation
through pulled optic fibers inserted into neural
▶ Spinal Stimulation for Parkinson Treatment
tissue (Gradinaru et al. 2009) and the develop-
▶ Wireless Microstimulators
ment of new miniaturized optoelectronics (Kim
et al. 2013b). However, the optical fibers used
suffer from the same mechanical impedance
References
mismatch issues as conventional implanted elec-
trodes, since most optical fibers and cladding Alilain WJ, Li X, Horn KP, Dhingra R, Dick TE,
have much higher mechanical elastic modulus Herlitze S, Silver J (2008) Light-induced rescue of
than neural tissues. Finally, infrared light from breathing after spinal cord injury. J Neurosci
28:11862–11870
lasers has also been used to elicit action potentials
Brindley GS (1988) The Ferrier lecture, 1986. The actions
in peripheral nerve and cochlea of small (rodents of parasympathetic and sympathetic nerves in human
and guinea pig) animal models (Izzo et al. 2006; micturition, erection and seminal emission, and their
108 Synaptic Dynamics: Overview
restoration in paraplegic patients by implanted electri- Wells J, Kao C, Jansen ED, Konrad P, Mahadevan-Jansen
cal stimulators. Proc R Soc Lond B Biol Sci A (2005a) Application of infrared light for in vivo
235:111–120 neural stimulation. J Biomed Opt 10:064003
Edgerton VR, Harkema S (2011) Epidural stimulation of Wells J, Kao C, Mariappan K, Albea J, Jansen ED,
the spinal cord in spinal cord injury: current status and Konrad P, Mahadevan-Jansen A (2005b) Optical stim-
future challenges. Expert Rev Neurother 11:1351–1353 ulation of neural tissue in vivo. Opt Lett 30:504–506
Egon G, Barat M, Colombel P, Visentin C, Isambert JL, Wells J, Kao C, Konrad P, Milner T, Kim J, Mahadevan-
Guerin J (1998) Implantation of anterior sacral root Jansen A, Jansen ED (2007) Biophysical mechanisms
stimulators combined with posterior sacral rhizotomy of transient optical stimulation of peripheral nerve.
in spinal injury patients. World J Urol 16:342–349 Biophys J 93:2567–2580
Fuentes R, Petersson P, Nicolelis MA (2010) Restoration Xia N, Wu XY, Wang X, Mou ZX, Wang MQ, Gu X,
of locomotive function in Parkinson’s disease by spi- Zheng XL, Hou WS (2014) Pulsed 808-nm infrared
nal cord stimulation: mechanistic approach. Eur laser stimulation of the auditory nerve in guinea pig
J Neurosci 32:1100–1108 cochlea. Lasers Med Sci 29:343–349
Gad P, Choe J, Nandra MS, Zhong H, Roy RR, Tai YC, Xiao L, Demosthenous A, Vanhoestenberghe A, Dai J,
Edgerton VR (2013) Development of a multi-electrode Donaldson N (2012) Active books: the design of an
array for spinal cord epidural stimulation to facilitate implantable stimulator that minimizes cable count
stepping and standing after a complete spinal cord using integrated circuits very close to electrodes.
injury in adult rats. J Neuroeng Rehabil 10:2 IEEE Trans Biomed Networks Syst 6:216–227
Gradinaru V, Mogri M, Thompson KR, Henderson JM,
Deisseroth K (2009) Optical deconstruction of parkin-
sonian neural circuitry. Science 324:354–359
Harkema S, Gerasimenko Y, Hodes J, Burdick J, Angeli C,
Chen Y, Ferreira C, Willhite A, Rejc E, Grossman RG, Synaptic Dynamics: Overview
Edgerton VR (2011) Effect of epidural stimulation of
the lumbosacral spinal cord on voluntary movement, Patrick D. Roberts
standing, and assisted stepping after motor complete
paraplegia: a case study. Lancet 377:1938–1947 Department of Biomedical Engineering, Oregon
Izzo AD, Richter CP, Jansen ED, Walsh JT Jr (2006) Laser Health & Science University, Portland, OR, USA
stimulation of the auditory nerve. Lasers Surg Med
38:745–753
Kim T, Branner A, Gulati T, Giszter SF (2013a) Braided
multi-electrode probes: mechanical compliance char- Definition
acteristics and recordings from spinal cords. J Neural
Eng 10:045001 Synaptic dynamics describes the time-dependent
Kim TI, McCall JG, Jung YH, Huang X, Siuda ER, Li Y, changes in synaptic currents that alter the
Song J, Song YM, Pao HA, Kim RH, Lu C, Lee SD,
Song IS, Shin G, Al-Hasani R, Kim S, Tan MP, Huang Y, strength of coupling between neurons. Various
Omenetto FG, Rogers JA, Bruchas MR (2013b) Inject- mechanisms, both pre- and postsynaptic, contrib-
able, cellular-scale optoelectronics with applications for ute to ongoing changes of synaptic currents and
wireless optogenetics. Science 340:211–216 modulate the overall network activity. These
Kumar K, Toth C, Nath RK, Laing P (1998) Epidural
spinal cord stimulation for treatment of chronic pain- mechanisms operate on various time scales
some predictors of success. A 15-year experience. (milliseconds to years) and can lead to immediate
Surg Neurol 50:110–120, discussion 120–111 changes in neuronal activity, ongoing adaptation
Ramasubbu C, Flagg A 2, Williams K (2013) Principles of of neuronal responses to changing inputs, and
electrical stimulation and dorsal column mapping as it
relates to spinal cord stimulation: an overview. Curr long-term learning and memory.
Pain Headache Rep 17:315
Ranger MR, Irwin GJ, Bunbury KM, Peutrell JM
(2008) Changing body position alters the location of Detailed Description
the spinal cord within the vertebral canal: a magnetic
resonance imaging study. Br J Anaesth 101:804–809
Saltiel P, Tresch MC, Bizzi E (1998) Spinal cord modular Connections between neurons form network cir-
organization and rhythm generation: an NMDA ionto- cuitry, and these connections are not static, but
phoretic study in the frog. J Neurophysiol 80:2323–2339 change in amplitude and timing. When a spike
Snow S, Horch KW, Mushahwar VK (2006) Intraspinal
microstimulation using cylindrical multielectrodes. reaches a presynaptic terminal, processes are ini-
IEEE Trans Biomed Eng 53:311–319 tiated that result in release of neurotransmitters
Synaptic Dynamics: Overview 109
that diffuse across the synaptic cleft to reach the The time-dependent synaptic conductance
postsynaptic receptors. Changes in these release may be based on double exponential functions
dynamics can lead to short-term or lasting (or similarly shaped functions for t 0).
changes in the amount and timing of transmitter

released. In addition, surrounding cells such as gs ðtÞ ¼ gs et=t1 þ et=t2 (2)
glia can influence the local synaptic dynamics in
tripartite synapse. Neurotransmitters bind to post-
synaptic receptors that may induce current to flow where t1 is the onset time constant and t2 is the
that changes the postsynaptic neurons membrane decay time constant. More detailed representa-
potential or initiates signaling pathways that tions of changes in synaptic conductances are
affect various properties of the postsynaptic developed in kinetic models of channels.
neuron. Two important types of excitatory currents are
mediated by AMPA glutamate receptor and
Fast Dynamics N-methyl-D-aspartate (NMDA) receptors that
Changes of synaptic currents following a single are excitatory in most neurons and have time
synaptic transmitter release event are usually on constants for a risetime of a few milliseconds.
the time scale of milliseconds to second and They differ in two respects, time constants and
controlled by the dynamics of postsynaptic recep- dependence on postsynaptic voltage, but are
tors. These fast dynamics follow a time course often colocalized at the same synaptic zones.
determined by transmitter diffusion, conforma- AMPA receptors usually have a larger peak con-
tional changes in postsynaptic receptors that ductance and decay with a few milliseconds
allows current to flow though the receptor pore, (Trussell et al. 1993). NMDA receptors have
or changes that result in triggering second mes- a longer decay (Jahr and Stevens 1990), and the
sengers that modify membrane currents. The peak current is dependent on the postsynaptic
resulting current flow is characterized by a sharp membrane potential. NMDA receptors pass cal-
rise followed by a slower decay as the receptors cium ions that can initiate signaling cascades that
inactivate and neurotransmitters unbind. The may have a secondary effect on synaptic dynam-
resulting currents can either excite or inhibit the ics on a longer time scale such as long-term
activity of the postsynaptic neuron, depending on plasticity (Collingridge and Bliss 1987; Lisman
the selectivity of membrane channels for 1989). Another important current is mediated by
specific ions. gamma-aminobutyric acid (GABA) receptors
Ionotropic receptors. Receptors that allow and is usually inhibitory in adult neurons.
direct current flow though a pore are called Metabotropic receptors. Receptors that trigger
ionotropic receptors (Eccles and McGeer 1979) second messengers to modify membrane currents
and provide the fastest synaptic dynamics. Math- are called metabotropic receptors (Eccles and
ematical representations of a synaptic current, McGeer 1979). These receptors are typically
Is(t), can be represented by a conductance gs(t) slower in their effect on the membrane potential
that regulates current flow across the membrane than ionotropic receptors and can result in sec-
(Gerstner and Kistler 2002; Koch 2004), ondary long-lasting effects on the excitability of
neurons. They can also help to initiate signaling
I s ðtÞ ¼ gs ðtÞ ðV ðtÞ Es Þ, (1) cascades that lead to long-term synaptic plastic-
ity. Their fast dynamics are mediated though
where V(t) is the membrane potential and Es is the changes in membrane currents, such as potassium
reversal potential of the current. If Es is greater currents, and the effects on membrane excitabil-
than the membrane potential, then the synaptic ity can be either excitatory or inhibitory
excites the neuronal activity, and if Es is less than depending on the type of metabotropic receptor.
the membrane potential, then the synapse inhibits In the simplest mathematical representations of
the neuronal activity. these receptors, the conductance can be
110 Synaptic Dynamics: Overview
represented by a double exponential function plasticity may be dependent on the activity of

(Eq. 2) with a long onset time constant (t1) of pre- and postsynaptic neurons, and when the rel-
several milliseconds. ative differences of timing of spikes between the
neurons determine whether the change is poten-
Short-Term Dynamics tiation or depression, then it is referred to as
The synaptic dynamics caused by a spike may be spike-timing-dependent plasticity (STDP)
affected by the history of previous spikes due to (Abbott and Nelson 2000). The processes of
short-term plasticity (Zucker and Regehr 2002; long-term depression and potentiation are bal-
Blitz et al. 2004). If the second of two spikes has anced by synaptic scaling that adjusts the global
a larger postsynaptic current than the first spikes, input to neurons to maintain stable activity.
then the plasticity is called facilitation, and if the Long-term plasticity is believed to form the
second spike has a smaller postsynaptic current, basis of learning and memory as learned sensory
then it is called short-term depression. The cues, and behaviors are encoded in the strengths
amount of facilitation and depression on further of synapse in neural circuitry (Hebb 1949).
spikes in a series changes over time and may
saturate and last up to several seconds. The dura-
tion of facilitation and depression may occur on Cross-References
different time scales leading to complex
sequences of postsynaptic currents that could be ▶ AMPA Glutamate Receptor (AMPA
unique to the pattern of spikes. Receptor), Conductance Models
Both pre- and postsynaptic mechanisms can ▶ Anti-Hebbian Learning
be responsible for short-term plasticity. An ▶ Biochemical Signaling Pathways and
important presynaptic mechanism is based on Diffusion: Overview
the presynaptic probability of neurotransmitter ▶ Bimolecular Reactions, Modeling of
release for each spikes combined with the rate ▶ Enzyme Kinetics, Modeling of
of replenishment of presynaptic vesicles ▶ Facilitation, Biophysical Models
containing neurotransmitters. If the probability ▶ Gamma-Aminobutyric Acid Type-A
of release is high, then synaptic zones may not (GABA-A) Receptors, Kinetic Models
yet be prepared for the next spike leading to ▶ Gap Junctions in Small Networks
a reduced peak synaptic current observed in ▶ Gap Junctions, Neural Population Models and
short-term depression. In contrast, if the release ▶ Hebbian Learning
probability is low, then the first spike may have ▶ Ionotropic Receptors Dynamics, Conductance
helped to prime the release so that the release Models
probability is higher for the second spike. This ▶ Kinetic Models of Postsynaptic Currents
increased release probability at each synaptic ▶ Learning Rules: Overview
zone leads to an increase in peak synaptic current ▶ Long Term Depression in the Granule Cell-
observed in facilitation. Purkinje Cell Synapse
▶ Long Term Plasticity, Biophysical Models
Long-Term Dynamics ▶ Metabotropic Receptors (G Protein Coupled
Changes in synaptic currents that last longer than Receptors)
several minutes are called long-term plasticity. ▶ Metabotropic Receptors Dynamics,
Several mechanisms, both pre- and postsynaptic, Conductance Models
may be responsible for these lasting changes that ▶ N-Methyl-D-Aspartate (NMDA) Receptors,
may increase or decrease the strength of synaptic Conductance Models
currents. Increases in synaptic currents are called ▶ Olfactory Computation in Glomerular
long-term potentiation (LTP) (Bliss and Lomo Microcircuits
1973), and decreases are called long-term depres- ▶ SenseLab: Integration of Multidisciplinary
sion (LTD) (Dudek and Bear 1993). Long-term Neuroscience Data
Vertebrate Pattern Generation: Overview 111
▶ Short-Term Plasticity, Biophysical Models Further Reading

▶ Short-Term Synaptic Plasticity in Central Dayan P, Abbott LF, Abbott L (2001) Theoretical neuro-
science: computational and mathematical modeling of
Pattern Generators
neural systems. Taylor & Francis, Cambridge, MA
▶ Signaling Pathways, Modeling of
▶ Spike-Timing Dependent Plasticity (STDP),
Biophysical Models
▶ Spike-Timing Dependent Plasticity, Learning
Rules Vertebrate Pattern Generation:
▶ Stability and Homeostasis in Small Network Overview
Central Pattern Generators
▶ Tripartite Synapse (Neuron–Astrocyte Ilya Rybak
Interactions), Conductance Models Department of Neurobiology and Anatomy,
▶ Working Memory, Models of Drexel University College of Medicine,
References Synonyms
Abbott LF, Nelson SB (2000) Synaptic plasticity: taming
the beast. Nat Neurosci 3((suppl)):1178–1183 CPG; Neural oscillator, Central pattern genera-
Bliss TV, Lomo T (1973) Long-lasting potentiation of tor; Rhythm generator
synaptic transmission in the dentate area of the
anaesthetized rabbit following stimulation of the
perforant path. J Physiol 232(2):331–356
Blitz DM, Foster KA, Regehr WG (2004) Short-term Definition
synaptic plasticity: a comparison of two synapses.
Nat Rev Neurosci 5(8):630–640 Central pattern generator (CPG) is a limited neu-
Collingridge GL, Bliss TVP (1987) NMDA receptors-
ral network that can produce an organized rhyth-
their role in long-term potentiation. Trends Neurosci
10(7):288–293 mic motor output in the absence of sensory and
Dudek SM, Bear MF (1993) Bidirectional long-term descending inputs from other parts of the nervous
modification of synaptic effectiveness in the system (Marder and Calabrese 1996).
adult and immature hippocampus. J Neurosci
13:2910–2918
Eccles JC, McGeer PL (1979) Ionotropic and
metabotropic neurotransmission. Trends Neurosci Detailed Description
2:39–40
Gerstner W, Kistler WM (2002) Spiking neuron models:
single neurons, populations, plasticity. Cambridge
Vertebrate Central Pattern Generators
University Press, Cambridge The central and peripheral nervous systems in
Hebb DO (1949) The organization of behavior. Wiley, vertebrates contain many types of central pattern
New York generators (CPGs) that generate and control var-
Jahr CE, Stevens CF (1990) A quantitative description of
ious rhythmic movements. These CPGs control
NMDA receptor-channel kinetic behavior. J Neurosci
10(6):1830–1837 many important functions, including different
Koch C (2004) Biophysics of computation: information forms of locomotion, such as swimming, walk-
processing in single neurons. Oxford University Press, ing, running, and flying, and non-locomotor pro-
New York
Lisman J (1989) A mechanism for the Hebb and the anti-
cesses and behaviors, such as breathing,
Hebb processes underlying learning and memory. Proc swallowing, chewing, mastication, scratching,
Natl Acad Sci U S A 86(23):9574–9578 whisking (in rodents), and singing (in birds).
Trussell LO, Zhang S, Ramant IM (1993) Desensitization This review focuses on computational models of
of AMPA receptors upon multiquantal neurotransmit-
ter release. Neuron 10(6):1185–1196
locomotion and breathing in mammals, which are
Zucker RS, Regehr WG (2002) Short-term synaptic plas- briefly observed below. Other known models
ticity. Annu Rev Physiol 64(1):355–405 of vertebrate CPGs include CPGs controlling
112 Vertebrate Pattern Generation: Overview
swimming in Xenopus tadpoles (Wolf et al. 2009; similar to that observed during normal locomo-
▶ Rhythm Generation in Young Xenopus Tad- tion in the intact animal (Rossignol 1996).
poles), zebra fish (Hill et al. 2005; Knudsen Similar pattern of locomotor activity could be
et al. 2006), and lamprey (Grillner et al. 2007); also produced by a systemic administration of
CPGs controlling swimming and walking in sal- the noradrenergic precursor, L-DOPA (e.g.,
amander (Ijspeert 2001; ▶ Control of Aquatic Jankowska et al. 1967a, b). The demonstrated
and Terrestrial Gaits in Salamander); CPGs con- possibility to produce either MLR-evoked or
trolling locomotion and scratching in turtles drug-evoked fictive locomotion in the cat spinal
(Hao et al. 2011; ▶ Control of Locomotion and cord has provided a strong evidence for the exis-
Scratching in Turtles); and CPG responsible for tence of locomotor CPG in the mammalian
rhythm generation in embryonic chick spinal spinal cord.
cord (Tabak et al. 2001; ▶ Rhythm Generation
in Embryonic Chick Spinal Cord). The Unit Burst Generators Concept
The half-center models of the CPG can be theo-
Locomotor CPG in Mammals retically analyzed and classified using such cate-
The Half-Center Concept of the Locomotor CPG gories as intrinsic escape, intrinsic release,
The ability of the spinal cord to generate coordi- synaptic escape, and synaptic release depending
nated locomotor-like rhythmic activity in the on which half-center and which process associ-
absence of rhythmic supraspinal and afferent ated with this half-center, control the transitions
inputs provides strong evidence for both the exis- between activity states (Wang and Rinzel 1992;
tence of locomotor CPGs in vertebrates including Skinner et al. 1994; ▶ Comparative Analysis of
mammals and their location in the spinal cord Half-Center Central Pattern Generators (CPGs)).
(Grillner 1981; Rossignol 1996; Orlovsky At the same time, the classical half-center con-
et al. 1999; Stuart and Hultborn 2008; Guertin cept can only represent a simplified CPG organi-
2009). The first schematic of locomotor CPG, zation and cannot reproduce and explain many
called the “half-center” model, was proposed by features of the real locomotor pattern observed in
Graham Brown (1914, see also Stuart and the mammalian spinal cord. Specifically, the real
Hultborn 2008; Guertin 2009). This model locomotor activity does not exhibit strictly alter-
consisted of two (flexor and extensor) half- nating flexor and extensor activities (with all
centers that reciprocally inhibited each other. motoneurons clearly belonging to one of these
The mutual inhibitory interactions between the two groups) as suggested by the classical half-
half-centers were mediated by the corresponding center model. The real pattern is more complex
inhibitory interneurons, which ensured that only and includes motoneuron pools (such as those
one half-center could be active at a time. The controlling biarticular muscles) generating bursts
activity of the currently active half-center was during both the flexion and extension phases of
reducing gradually because of some fatigue or the step cycle or during only a part of one phase.
adaptation mechanisms leading to the activation There are also noticeable differences in the
of the antagonistic half-center, inhibiting the cur- timing of burst onset and/or offset between dif-
rently active half-center, and switching the loco- ferent flexor and/or between different extensor
motor phase. It was suggested that the flexor and pools. To overcome these and other limitations
extensor half-centers project to and activate the of the classical half-center model, Grillner (1981)
flexor and extensor motoneurons, respectively. proposed a unit burst generator (UBG) concept
Studies on the immobilized decerebrate cat dem- suggesting the existence of several separate
onstrated that continuous electrical stimulation of rhythmogenic modules (or unit burst generators)
the midbrain locomotor region (MLR) can pro- controlling each joint of the limb and interacting
duce “fictive locomotion” – the rhythmic pattern with each other as coupled neural oscillators.
of motoneuron activity characterized by alternat- A simplified mathematical model of the mamma-
ing activation of flexor and extensor motoneurons lian locomotor CPG based on the UBG concept
was proposed and analyzed by Sherwood synergies at the PF level. This CPG model was
et al. (2011; ▶ Computational Analysis of the used in several studies focused on the develop-
Rodent Spinal CPG). Despite many obvious ment of CPG-controlled legged robots
advantages, this model showed a very slow (Chen et al. 2007; Maeda 2008; Amrollah and
recovery after applied perturbations. The analysis Henaff 2010).
of subnetworks of this CPG revealed that the
endogenous bursting properties of coupled neu- Rhythmic Activity in the Isolated Spinal Cord
rons dominated over the phasing and synchroni- Preparations of Rodents
zation properties of the CPG network, leading to The mammalian locomotor CPG was also studied
the slow phase resetting and rhythm stabilization in vitro using the isolated spinal cord preparations
in response to applied perturbations. of rodents (Smith and Feldman 1987, ▶ Locomo-
tor Pattern Generation in the Rodent Spinal
The Two-Level Model of the Locomotor CPG Cord). The locomotor CPG in these preparations
An important extension of the classical half- can be activated by several ways, including phar-
center model was suggested based on the analysis macological manipulations [by using solutions
of (a) spontaneous deletions (missing bursts) in with specific combinations of N-methyl-
the rhythmic motoneuron activities and D-aspartate (NMDA) and serotonin], stimulations
(b) effects of afferent stimulations on the loco- of sensory afferents or dorsal roots, and brainstem
motor pattern during fictive locomotion in the cat stimulation. The elicited motor output shows
(Rybak et al. 2006a, b; McCrea and Rybak 2007, alternation of activity in the left and right ventral
2008; ▶ Two-Level Model of Mammalian Loco- roots (e.g., left L2 and right L2) combined with
motor CPG). This analysis led to the suggestion alternation of ipsilateral flexor (L2) and extensor
that the locomotor CPG has a two-level architec- (L5) activities on each side of the cord. Signifi-
ture with a separate top-level, half-center rhythm cant progress in the understanding of functional
generator (RG) and a pattern formation network and structural organization of spinal circuits has
(PF). While the RG defines the locomotor fre- been achieved using special combinations of
quency, the PF network is controlled by the RG genetic, molecular, and developmental
and defines and coordinates the more compli- approaches. Several classes of spinal interneu-
cated firing patterns of different motoneuron rons were defined in the embryonic and early
pools. The two-level model was able to reproduce postnatal spinal cord based on the dynamic
a number of features of the locomotor pattern expression pattern of transcription factors
observed during cat locomotion, including (Goulding 2009; Whelan 2010; Gosgnach 2011;
phase maintenance of motoneuron activities fol- Kiehn 2011; ▶ Locomotor Pattern Generation in
lowing spontaneous deletions or brief stimulation the Rodent Spinal Cord). The new genetic tech-
of some afferents, which would be difficult or niques allow visualization of identified neurons
even impossible to explain in the framework of as well as manipulation of their activity and their
the classical half-center architecture (see ▶ Two- selective elimination. Several classes of inhibi-
Level Model of Mammalian Locomotor CPG). tory (CINi) and excitatory (CINe) commissural
Some aspects of supraspinal and afferent control interneurons (CINs) have been identified. The
of the two-level locomotor CPG were analyzed axons of these cells cross the midline, project to
using computational modeling approaches the opposite side of the cord, and mediate inter-
(Rybak et al. 2006b; Markin et al. 2010; Spardy actions between left and right circuits. The role of
et al. 2011a, b). different CINs in locomotion has been identified
One advantage of the two-level model is by using transgenic mice with mutations in, or
a possibility to perform independent control knockout (KO) of, specific genes resulting in
(by supraspinal and afferent inputs) of the step various abnormal locomotor phenotypes
cycle duration (locomotor speed) at the RG level (Kullander et al. 2003; Lanuza et al. 2004;
and the activity of motoneurons and motor Akay et al. 2006; Lundfald et al. 2007;
Crone et al. 2008, 2009; Zhang et al. 2008; Rabe KO circuits showed switching from the left–right
et al. 2009; Zagoraiou et al. 2009; Restrepo alternating pattern to a synchronized hopping
et al. 2011; Talpalar et al. 2013). A two-level pattern, and the DCC KO network exhibited
computational model of neural circuits in the uncoordinated left–right activity. The model
spinal cord with left and right rhythm-generating was able to reproduce multiple experimental
populations interacting via inhibitory and excit- data on the effects of above genetic transforma-
atory CINs has been developed based on the tions on the locomotor pattern, providing impor-
analysis of spontaneous deletions in the isolated tant insights into the organization of the spinal
spinal cord preparation with rhythmic activity locomotor network.
evoked by administration of NMDA and seroto-
nin (Zhong et al. 2012; ▶ Locomotor Pattern Respiratory CPG in Mammals
Generation in the Rodent Spinal Cord). Respiratory Pattern and Spatial Organization of
the Respiratory CPG
Modeling the Effects of Genetic The respiratory motor pattern observed during
Transformations on the Locomotor Pattern quiet breathing in mammals (“eupnea”) consists
A representative example of how genetic manip- of three phases: inspiration (I), post-inspiration
ulations can change the phase relationships (post-I or E1), and late expiration (E2), which can
between ipsi- and contralateral neurons involved be recognized in the integrated activity of the
in control of locomotion in mice is based on phrenic and cranial nerves. This pattern origi-
the genetic ablation of special molecules nates within a bilateral column of neurons, called
involved in guidance of CIN axons, such as the ventral respiratory column (VRC), located in
Netrin-1 and DCC. Also, the spinal cord contains the ventrolateral medulla and controlled by
glutamatergic interneurons with ipsilateral pro- inputs from other medullary (retrotrapezoid
jections, whose axon guidance involves the nucleus, RTN, raphé, etc.) and pontine regions.
EphA4 receptor. In EphA4 knockout (KO) and The VRC includes several compartments
Netrin-1 KO mice, the normal left–right alternat- arranged in the rostrocaudal direction: Bötzinger
ing pattern is replaced with a synchronized hop- complex (BötC), pre-Bötzinger complex
ping gait, and the spinal cord of DCC KO mice (pre-BötC), and rostral (rVRG) and caudal
exhibits uncoordinated left and right oscillations (cVRG) subregions of the ventral respiratory
(Rabe et al. 2009; Restrepo et al. 2011; Rabe group (VRG). Respiratory neurons in these com-
Bernhardt et al. 2012; Vallstedt and Kullander partments are usually classified based on their
2013). To investigate the effects of these genetic firing pattern (e.g., decrementing, augmenting)
transformations, Rybak et al. (2013) developed and phase of activity relative to the breathing
a computational model of the spinal circuits cycle, such as early-inspiratory (early-I or
containing the left and right rhythm-generating I-DEC), i.e., inspiratory neurons with a
neuron populations (RGs), each with decrementing discharge pattern; ramp-inspira-
a subpopulation of EphA4-positive neurons, and tory (ramp-I or I-AUG), i.e., inspiratory neurons
the CINi and CINe populations mediating respec- with an augmenting firing pattern; post-
tively mutual inhibition and mutual excitation inspiratory neurons (post-I or E-DEC), i.e.,
between the left and right RGs. In the EphA4 neurons with a decrementing activity during
KO circuits, half of the axons of EphA4-positive expiration; augmenting or stage II expiratory
cells crossed the midline and excited the contra- (aug-E or E-AUG or E2), i.e., expiratory neurons
lateral RG neurons. In the Netrin-1 KO model, with an augmenting pattern; and pre-inspiratory/
the number of contralateral CINi projections was inspiratory neurons (pre-I/I), the neurons whose
significantly reduced, while in the DCC KO activity starts before the onset of inspiration and
model, the numbers of connections from both continues throughout inspiration. The BötC, with
CINi and CINe were reduced. In the simulations predominately post-I and aug-E neurons, is con-
performed, the models of EphA4 and Netrin-1 sidered to be a major source of expiratory
activity. The adjacent, more caudal pre-BötC population bursting. Importantly, generation of
contains neural circuits essential for generating this rhythm does not require inhibitory interac-
inspiratory activity (Cohen 1979; Bianchi tions, which can explain the persistence of the
et al. 1995; Richter 1996; Smith et al. 2013). in vitro oscillations after inhibitory synaptic
Computational models of the respiratory net- transmission was blocked. It was also shown
work have been in development for several that even a small fraction of intrinsically bursting
decades. Early computational models of the cells (5–10 %) can produce a synchronized burst-
respiratory CPG focused on the network interac- ing activity of the entire population. Moreover,
tions between different types of respiratory neu- a synchronized population activity may occur
rons and did not consider intrinsic biophysical even if none of the cells operates in the intrinsic
rhythmogenic properties of single respiratory bursting state. Elevation of tonic drive to the
neurons and their possible contribution to population reduces burst duration, increases
rhythmogenesis. Generation of the respiratory burst frequency, and, finally, switches population
rhythm in these models was based on the half- activity from population bursting to a regime of
center concept suggesting that the respiratory sustained asynchronous activity (Butera
oscillations result from sequential phase switching et al. 1999b).
resulting from the reciprocal inhibitory interac-
tions between different types of respiratory neu- Network-Based Versus Pacemaker-Driven
rons (Botros and Bruce 1990; Duffin 1991; Ogilvie Mechanisms for Respiratory Rhythmogenesis
et al. 1992; Balis et al. 1994; Rybak et al. 1997; and Hybrid Pacemaker-Network Models
reviewed by Lindsey et al. 2012) The early network models where able to generate
a realistic respiratory motor pattern and
The Pre-Bötzinger Complex and Rhythm simulate various alterations of this pattern under
Generation In Vitro different conditions. However, these models
A fundamentally distinct concept of respiratory failed to reproduce some characteristic behaviors
rhythm generation was derived from the neonatal observed in the reduced in vitro preparations and,
in vitro studies. The important discovery has been specifically, the maintenance of the respiratory
that a subregion of the VRC, called the rhythm following inhibition blockade. Alterna-
pre-Bötzinger complex, contains a population of tively, the pacemaker-based models, developed
excitatory interneurons that can intrinsically gen- to fit to in vitro data, could not explain many
erate an inspiratory-like rhythm (Smith respiratory behaviors observed in vivo, such as
et al. 1991). This rhythm was shown to persist the Hering–Breuer and other respiratory
after blockade of synaptic inhibition, indicating reflexes and independent control of the duration
that the pre-BötC may contain cells with intrinsic of each respiratory phase. Neither could
bursting properties. Butera et al. (1999a, b) devel- these models reproduce apneusis, a well-known
oped and analyzed a series of computational abnormal breathing pattern characterized
models of bursting pacemaker neurons and by a significantly prolonged inspiration (up to
populations of these neurons with mutual excit- several seconds) alternating with short
atory connections. In these models, the intrinsic expiratory intervals. Moreover, the rhythmic
bursting was based on a slowly inactivating per- activity generated in the reduced in vitro
sistent sodium current (INaP) as the essential preparations and reproduced by the pacemaker-
burst-generating, inward cationic current. The driven models is characterized by a decrementing
rhythmic bursting cycle in these models was con- shape of inspiratory discharges that clearly
trolled by the slow kinetics of inactivation and differs from the augmenting shape of
recovery from inactivation of INaP. Simulations phrenic discharges generated during eupneic
performed have shown that the excitatory synap- breathing in vivo; these discharges rather
tic interactions coupled with INaP activation can resemble the decrementing bursts observed dur-
readily synchronize cellular bursts and produce ing gasping.
The above contradictions between the transforms to “one-phase” inspiratory oscilla-

network-based and the pacemaker-based con- tions originating within the pre-BötC, without
cepts and models can be resolved by postulating critical involvement of phasic expiratory inhibi-
that (i) the pre-BötC, while capable of intrinsic tion (Rybak et al. 2007; Smith et al. 2007). These
bursting when isolated and under some special results led to the conclusion that (i) generation of
conditions, is normally embedded in the the normal three-phase rhythmic pattern requires
larger brainstem respiratory network which the presence of the pons (i.e., excitatory drive
makes its behavior dependent on its interactions from pontine neurons to the VRC);
with other brainstem structures and (ii) the (ii) generation of the two-phase pattern is intrin-
respiratory rhythmogenesis is state dependent, sic to reciprocal inhibitory synaptic interactions
and therefore the respiratory oscillations may between the BötC and the pre-BötC and may also
be generated by either a network-based or involve the RTN to provide excitatory drive to
pacemaker-driven mechanisms or their specific generate stable behavior; and (iii) the one-phase
combinations depending on the conditions inspiratory oscillations are generated within the
(Smith et al. 2000; Rybak et al. 2002, 2004, pre-BötC and rely on intrinsic cellular mecha-
2007; Smith et al. 2007, 2013; Lindsey nisms operating in the context of the pre-BötC
et al. 2012; ▶ Computational Models of Mamma- excitatory network. These data allowed the con-
lian Respiratory CPG). clusion that there is a spatial and dynamical hier-
archy of interacting pontine, BötC and pre-BötC
The Respiratory CPG and Spatial and Functional circuits, each of which controls different aspects
Hierarchy of Multiple Rhythmogenic Mechanisms of respiratory rhythm generation and pattern for-
A significant progress in understanding of spatial mation, which can be revealed as the network is
and functional organization of the respiratory progressively reduced. The expression of each
CPG in the mammalian brainstem has been rhythmogenic mechanism is state dependent and
achieved by using an arterially perfused in situ produces specific motor patterns likely to under-
rat brainstem–spinal cord preparation (Paton pin distinct motor behaviors (Smith et al. 2007,
1996) and applying sequential rostral-to-caudal 2013; ▶ Computational Models of Mammalian
microtransections through the brainstem while Respiratory CPG).
recording cranial and spinal motor outflow to A minimal network configuration proposed to
observe transformations of network behavior represent the above multiple rhythmic states, and
(Rybak et al. 2007; Smith et al. 2007). This their transformations included (i) a mutually
approach reviled the existence of a rostral-to- inhibitory circuit with a ringlike architecture
caudal stacking of network building blocks composed of post-I and aug-E neurons of the
subserving distinct circuit functions, which are BötC compartment and early-I neurons
fully integrated in the intact system, but can be within the pre-BötC and (ii) a pre-BötC kernel
revealed as particular compartments are removed of excitatory pre-I/I neurons, with intrinsic INaP-
or under special metabolic conditions (hypoxia, dependent bursting properties. The latter partici-
hypercapnia). The major results obtained from pate dynamically in the expiratory–inspiratory
these studies are that sequential reduction of the phase transition and generation of the inspiratory
network progressively reorganizes network phase. The detailed description and computa-
dynamics, such that new rhythmogenic mecha- tional model of this core network and
nisms emerge. Specifically, starting from the circuit elements participating in each
a transection at the pontine–medullary junction, rhythmic state can be found in a series of related
the normal three-phase respiratory pattern is publications (Rybak et al. 2007; Smith
transformed to a two-phase rhythmic pattern et al. 2007; Rubin et al. 2009b, for review see
lacking the post-I phase. With more caudal tran- also Lindsey et al. 2012; Smith et al. 2013;
sections made close to or at the rostral boundary ▶ Computational Models of Mammalian Respi-
of the pre-BötC, the respiratory activity ratory CPG).
Respiratory Rhythm Generation and Coupled suppression of RTN/pFRG inhibition by BötC/

Oscillators pre-BötC. The proposed interactions between the
As described above, the respiratory CPG is con- BötC/pre-BötC and RTN/pFRG oscillators allow
sidered to comprise several interacting the model to reproduce several experimentally
populations of respiratory neurons located in the observed behaviors, including quantal accelera-
pre-BötC and BötC circuits within the VR- tion of abdominal late-E oscillations with progres-
C. A distinct site of neural oscillations was sive hypercapnia and quantal slowing of phrenic
identified in vitro in the isolated neonatal rat activity with progressive suppression of pre-BötC
brainstem–spinal cord preparation (Onimaru and excitability, as well as to predict a release of late-E
Homma 1987, 2003; Onimaru et al. 1988). This oscillations by disinhibition of RTN/pFRG under
additional oscillator, called the parafacial respi- normal conditions (Molkov et al. 2010; ▶ Coupled
ratory group (pFRG), seems to reside within, or Oscillations in Neural Control of Breathing).
overlap with, the RTN. It was also found that Rubin et al. (2011) performed thorough qualitative
RTN/pFRG oscillations in vivo drive abdominal analysis of the reduced model and explained the
motor activity, expressing late-expiratory (late-E regimes of quantal acceleration and quantal
or pre-inspiratory) bursts or biphasic discharges slowing in terms of synchronization of the BötC/
(with pre-I and post-I components) in the abdom- pre-BötC and RTN/pFRG oscillators. They have
inal motor output when the system operates in the shown that the dynamics of each oscillator can be
active (forced) expiration regime (Feldman and represented by a stable limit cycle in some phase
Del Negro 2006; Janczewski and Feldman 2006; space. The phase space of a system of two coupled
Janczewski et al. 2002; Abdala et al. 2009). Sev- oscillators is a Cartesian product of the phase
eral competing concepts concerning the physio- spaces of each oscillator. The behavior of this
logical role of RTN/pFRG oscillations have been system can be represented by a trajectory on 2D
suggested, including the dual oscillator concept invariant torus. If the ratio of oscillation frequen-
that considers the RTN/pFRG as an independent cies of the two oscillators is rational (i.e., equal to
expiratory rhythm generator coupled with N/M, for some integers N and M), then this trajec-
a distinct inspiratory rhythm generator in the tory is closed, indicating N:M synchronization
pre-BötC (Janczewski and Feldman 2006). How- between oscillators, where N and M represent
ever, the exact physiological role of pFRG oscil- numbers of rotations around two orthogonal cir-
lations, the specific conditions for their cles that together span the torus (Rubin
emergence, and the nature and mechanisms of et al. 2011). These modeling studies provide
the interactions between the BötC/pre-BötC and a mechanistic explanations for the emergence of
RTN/pFRG oscillators are not yet known. RTN/pFRG oscillations and their interaction with
Molkov et al. (2010; ▶ Coupled Oscillations the BötC/pre-BötC circuits representing the core
in Neural Control of Breathing) extended the of the respiratory CPG.
previous respiratory CPG model (Rybak
et al. 2007; Smith et al. 2007) to include the Intrinsic Neuronal Properties Involved in
RTN/pFRG oscillator and consider interactions Rhythmic Bursting in the Brainstem and
between the BötC/pre-BötC and RTN/pFRG Spinal Cord
oscillators. The extended model incorporates an The mechanisms generating neural oscillations in
additional late-E population in the RTN/pFRG the mammalian brainstem, particularly in the
compartment, representing a source of late-E pre-Bötzinger complex involved in control of
oscillatory activity. In the proposed model, under respiration, and in the spinal cord (locomotor
normal metabolic conditions, the RTN/pFRG CPG), that persist after blockade of synaptic inhi-
oscillator is inhibited by the BötC/pre-BötC cir- bition, remain poorly understood. Experimental
cuits, and the late-E oscillations can be only studies in medullary slices from neonatal
released either by hypercapnia-evoked activation rodents containing the pre-BötC identified two
of RTN/pFRG or by hypoxia-dependent mechanisms that could potentially contribute to
generation of rhythmic bursting: one based on the and modeling have shown that the emergence of
persistent or slowly inactivating INaP (Butera this rhythmic activity critically involves a [Ca2+]
et al. 1999a, b; Koizumi and Smith 2008) and o-dependent activation of the persistent sodium
the other involving the calcium-activated current (INaP). These results suggest that the loco-
nonspecific cation current (ICAN) activated by motor oscillations in the spinal cord may relay on
intracellular Ca2+ accumulated from extracellular the INaP-dependent pacemaker properties in spi-
(e.g., calcium currents, ICa) and/or intracellular nal interneurons, which can be turned on and off
sources (Pace et al. 2007). The involvement and by activity-dependent changes in [Ca2+]o and
relative roles of these mechanisms in rhythmic [K+]o (Brocard et al. 2013).
bursting are still under debate. Several related
models combining the above two mechanisms
have been developed (Rubin et al. 2009a; Cross-References
Dunmyre et al. 2011; Toporikova and Butera
2011; Jasinski et al. 2013). Jasinski et al. (2013) ▶ Comparative Analysis of Half-Center Central
investigated Na+- and Ca2+-dependent bursting Pattern Generators (CPGs)
generated in single cells and in a heterogeneous ▶ Computational Analysis of Rodent Spinal CPG
population of synaptically interconnected excit- ▶ Computational Models of Mammalian
atory neurons with INaP and ICa randomly distrib- Respiratory CPG
uted within the population. They analyzed the ▶ Control of Aquatic and Terrestrial Gaits in
possible roles of network connections, ionotropic Salamander
and metabotropic synaptic mechanisms, intracel- ▶ Control of Locomotion and Scratching in
lular Ca2+ release, and the Na+/K+ pump in rhyth- Turtles
mic bursting activity generated under different ▶ Coupled Oscillations in Neural Control of
conditions. They showed that heterogeneous Breathing
populations of excitatory neurons with these prop- ▶ Locomotor Pattern Generation in the Rodent
erties can operate in different oscillatory regimes Spinal Cord
with bursting dependent on INaP and/or on ICAN or ▶ Rhythm Generation in Embryonic Chick
independent of both. The exact oscillatory regime Spinal Cord
and the operating bursting mechanism may ▶ Rhythm Generation in Young Xenopus
depend on neuronal excitability, synaptic interac- Tadpoles
tions, and relative expression of particular ionic ▶ Two-Level Model of Mammalian Locomotor
currents. The existence of multiple oscillatory CPG
regimes and their state dependency may provide
explanations for different rhythmic activities
observed in the brainstem and spinal cord under
different experimental conditions. References
Interestingly, the Na+- and Ca2+-dependent
Abdala AP, Rybak IA, Smith JC, Paton JF (2009) Abdom-
bursting mechanisms can be unexpectedly
inal expiratory activity in the rat brain stem-spinal cord
connected. Brocard et al. (2013) have recently in situ: patterns, origins and implications for respira-
found that the evoked locomotor-like activity tory rhythm generation. J Physiol 587:3539–3559
in the isolated neonatal rodent spinal cord Akay T, Acharya HJ, Fouad K, Pearson KG (2006) Behav-
ioral and electromyographic characterization of mice
reduces the extracellular calcium concentration lacking EphA4 receptors. J Neurophysiol 96:642–651
([Ca2+]o]) to 0.9 mM and increases the extracel- Amrollah E, Henaff P (2010) On the role of sensory
lular potassium concentration ([K+]o) to 6 mM. feedbacks in Rowat–Selverston CPG to improve
Such changes in [Ca2+]o and [K+]o trigger robot legged locomotion. Front Neurorobot 4:113
Balis UJ, Morris KF, Koleski J, Lindsey BG (1994) Simu-
a rhythmic bursting activity in spinal interneu-
lations of a ventrolateral medullary neural network for
rons that may represent elements of the respira- respiratory rhythmogenesis inferred from spike train
tory CPG. The performed experimental studies cross-correlation. Biol Cybern 70:311–327
Bianchi AL, Denavitsaubie M, Champagnat J (1995) Cen- (eds) Handbook of physiology. The nervous system.
tral control of breathing in mammals: neuronal cir- Motor control, Sect 1, vol II. American Physiological
cuitry, membrane properties, and neurotransmitters. Society, Bethesda, pp 1179–1236
Physiol Rev 75:1–45 Grillner S, Kozlov A, Dario P, Stefanini C, Menciassi A,
Botros SM, Bruce EN (1990) Neural network implemen- Lansner A, Hellgren KJ (2007) Modeling
tation of a three-phase model of respiratory rhythm a vertebrate motor system: pattern generation, steering
generation. Biol Cybern 63:143–153 and control of body orientation. Prog Brain Res
Brocard F, Shevtsova NA, Bouhadfane M, Tazerart S, 165:221–234
Heinemann U, Rybak IA, Vinay L (2013) Activity- Guertin PA (2009) The mammalian central pattern gener-
dependent changes in extracellular Ca2+ and K+ reveal ator for locomotion. Brain Res Rev 62:45–56
pacemakers in the spinal locomotor-related network. Hao ZZ, Spardy LE, Nguyen EB, Rubin JE, Berkowitz
Neuron 77:1047–1054 A (2011) Strong interactions between spinal cord net-
Butera RJ, Rinzel J, Smith JC (1999a) Models of respira- works for locomotion and scratching. J Neurophysiol
tory rhythm generation in the pre-Bötzinger complex. 106:1766–1781
II. Populations of coupled pacemaker neurons. Hill SA, Liu X-P, Borla MA, José JV, O’Malley DM
J Neurophysiol 82:398–415 (2005) Neurokinematic modeling of complex swim-
Butera RJ, Rinzel J, Smith JC (1999b) Models of respira- ming patterns of the larval zebrafish. Neurocomputing
tory rhythm generation in the pre-Bötzinger complex. 65–66:61–68
I. Bursting pacemaker neurons. J Neurophysiol Ijspeert AJ (2001) A connectionist central pattern genera-
82:382–397 tor for the aquatic and terrestrial gaits of a simulated
Chen Y, Bauer C, Burmeister O, Rupp R, Mikut R (2007) salamander. Biol Cybern 84:331–348
First steps to future applications of spinal neural circuit Janczewski WA, Feldman JL (2006) Distinct rhythm gen-
models in neuroprostheses and humanoid robots. In: erators for inspiration and expiration in the juvenile
Proceedings of 17 workshop computational intelli- rat. J Physiol 570:407–420
gence, Dortmund Janczewski WA, Onimaru H, Homma I, Feldman JL
Cohen MI (1979) Neurogenesis of respiratory rhythm in (2002) Opioid-resistant respiratory pathway from the
the mammal. Physiol Rev 59:1105–1173 preinspiratory neurones to abdominal muscles: in vivo
Crone SA, Quinlan KA, Zagoraiou L, Droho S, Restrepo and in vitro study in the newborn rat. J Physiol
CE, Lundfald L, Endo T, Setlak J, Jessell TM, 545:1017–1026
Kiehn O, Sharma K (2008) Genetic ablation of V2a Jankowska E, Jukes MGM, Lund S, Lundberg A (1967a)
ipsilateral interneurons disrupts left-right locomotor The effect of DOPA on the spinal cord: V. Reciprocal
coordination in mammalian spinal cord. Neuron organization of pathways transmitting excitatory
60:70–83 action to alpha motoneurones of flexors and extensors.
Crone SA, Zhong G, Harris-Warrick R, Sharma K (2009) Acta Physiol Scand 70:369–388
In mice lacking V2a interneurons, gait depends on Jankowska E, Jukes MGM, Lund S, Lundberg A (1967b)
speed of locomotion. J Neurosci 29:7098–7109 The effect of DOPA on the spinal cord: VI. Half-centre
Duffin J (1991) A model of respiratory rhythm generation. organization of interneurones transmitting effects
Neuroreport 2:623–626 from the flexor reflex afferents. Acta Physiol Scand
Dunmyre JR, Del Negro CA, Rubin JE (2011) Interactions 70:389–402
of persistent sodium and calcium-activated Jasinski PE, Molkov YI, Shevtsova NA, Smith JC, Rybak
nonspecific cationic currents yield dynamically dis- IA (2013) Sodium and calcium mechanisms of rhyth-
tinct bursting regimes in a model of respiratory neu- mic bursting in excitatory neural networks of the
rons. J Comput Neurosci 31:305–328 pre-Bötzinger complex: a computational modelling
Feldman JL, Del Negro CA (2006) Looking for inspira- study. Eur J Neurosci 37:212–230
tion: new perspectives on respiratory rhythm. Nat Rev Kiehn O (2011) Development and functional organization
Neurosci 7:232–242 of spinal locomotor circuits. Curr Opin Neurobiol
Gosgnach S (2011) The role of genetically-defined Inter- 21:100–109
neurons in generating the mammalian locomotor Knudsen DP, Arsenault JT, Hill SA, O’Malley DM, José
rhythm. Integr Comp Biol 51:903–912 JV (2006) Locomotor network modeling based on
Goulding M (2009) Networks controlling vertebrate loco- identified zebrafish neurons. Neurocomputing
motion: moving in a new direction. Nat Rev Neurosci 69:1169–1174
10:507–518 Koizumi H, Smith JC (2008) Persistent Na+ and
Graham Brown T (1914) On the fundamental activity of K+dominated leak currents contribute to respiratory
the nervous centres: together with an analysis of the rhythm generation in the pre-Bötzinger complex
conditioning of rhythmic activity in progression, and in vitro. J Neurosci 28:1773–1785
a theory of the evolution of function in the nervous Kullander K, Butt SJ, Lebret JM, Lundfald L, Restrepo
system. J Physiol 48:18–41 CE, Rydstrom A, Klein R, Kiehn O (2003) Role of
Grillner S (1981) Control of locomotion in bipeds, tetra- EphA4 and EphrinB3 in local neuronal circuits that
pods, and fish. In: Brookhart JM, Mountcastle VB control walking. Science 299:1889–1892
Lanuza GM, Gosgnach S, Pierani A, Jessell TM, Goulding locomotor central pattern generator function. Dev Biol
M (2004) Genetic identification of spinal interneurons 366:279–289
that coordinate left-right locomotor activity necessary Rabe N, Gezelius H, Vallstedt A, Memic F, Kullander
for walking movements. Neuron 42:375–386 K (2009) Netrin-1-dependent spinal interneuron sub-
Lindsey BG, Rybak IA, Smith JC (2012) Computational types are required for the formation of left-right alter-
models and emergent properties of respiratory neural nating locomotor circuitry. J Neurosci
networks. Compr Physiol 2:1619–1670 29:15642–15649
Lundfald L, Restrepo CE, Butt SJ, Peng CY, Droho S, Restrepo CE, Margaryan G, Borgius L, Lundfald L,
Endo T, Zeilhofer HU, Sharma K, Kiehn O (2007) Sargsyan D, Kiehn O (2011) Change in the balance
Phenotype of V2-derived interneurons and their rela- of excitatory and inhibitory midline fiber crossing as
tionship to the axon guidance molecule EphA4 in the an explanation for the hopping phenotype in EphA4
developing mouse spinal cord. Eur J Neurosci knockout mice. Eur J Neurosci 34:1102–1112
26:2989–3002 Richter DW (1996) Neural regulation of respiration:
Maeda Y (2008) A hardware neuronal network model of rhythmogenesis and afferent control. In: Gregor R,
two-level central pattern generator. Trans JPN Soc Windhorst U (eds) Comprehensive human physiology,
Med Biol Eng 46:496–504 vol 2. Springer, Berlin, pp 2079–2095
Marder E, Calabrese RL (1996) Principles of rhythmic Rossignol S (1996) Neural control of stereotypic limb
motor pattern generation. Physiol Rev 76:687–717 movements. In: Rowell LB, Shepherd J (eds) Hand-
Markin SN, Klishko AN, Shevtsova NA, Lemay MA, book of physiology, Sect 12. American Physiological
Prilutsky BI, Rybak IA (2010) Afferent control of Society, Bethesda, pp 173–216
locomotor CPG: insights from a simple neuro- Rubin JE, Hayes JA, Mendenhall JL, Del Negro CA
mechanical model. Ann NY Acad Sci 1198:21–34 (2009a) Calcium-activated nonspecific cation current
McCrea DA, Rybak IA (2007) Modeling the mammalian and synaptic depression promote network-dependent
locomotor CPG: insights from mistakes and perturba- burst oscillations. Proc Natl Acad Sci
tions. Prog Brain Res 165:235–253 USA 106:2939–2944
McCrea DA, Rybak IA (2008) Organization of mamma- Rubin JE, Shevtsova NA, Ermentrout GB, Smith JC,
lian locomotor rhythm and pattern generation. Brain Rybak IA (2009b) Multiple rhythmic states in
Res Rev 57:134–146 a model of the respiratory CPG. J Neurophysiol
Molkov YI, Abdala AP, Bacak BJ, Smith JC, Paton JFR, 101:2146–2165
Rybak IA (2010) Late-expiratory activity: emergence Rubin JE, Bacak BJ, Molkov YI, Shevtsova NA, Smith
and interactions with the respiratory CPG. J JC, Rybak IA (2011) Interacting oscillations in neural
Neurophysiol 104:2713–2729 control of breathing: modeling and qualitative analy-
Ogilvie MD, Gottschalk A, Anders K, Richter DW, Pack sis. J Comput Neurosci 30:607–6322
AI (1992) A network model of respiratory Rybak IA, Paton JFR, Schwaber JS (1997) Modeling neu-
rhythmogenesis. Am J Physiol Regul Integr Comp ral mechanisms for genesis of respiratory rhythm and
Physiol 263:R962–R975 pattern. II. Network models of the central respiratory
Onimaru H, Homma I (1987) Respiratory rhythm genera- pattern generator. J Neurophysiol 77:2007–2026
tor neurons in medulla of brain stem-spinal cord prep- Rybak IA, Paton JFR, Rogers RF, St. John WM
aration from newborn rat. Brain Res 403:380–384 (2002) Generation of the respiratory rhythm: state-
Onimaru H, Arata A, Homma I (1988) Primary respiratory dependency and switching. Neurocomputing
rhythm generator in the medulla of brain stem-spinal 44–46:603–612
cord preparation from newborn rat. Brain Res Rybak IA, Shevtsova NA, Paton JFR, Dick TE, St. John
445:314–324 WM, Morschel M, Dutschmann M (2004) Modeling
Onimaru H, Homma I (2003) A novel functional neuron the ponto-medullary respiratory network. Respir
group for respiratory rhythm generation in the ventral Physiol Neurobiol 143:307–319
medulla. J Neurosci 23:1478–1486 Rybak IA, Shevtsova NA, Lafreniere-Roula M, McCrea
Orlovsky GN, Deliagina T, Grillner S (1999) Neuronal DA (2006a) Modelling spinal circuitry involved in
control of locomotion: from mollusc to man. Oxford locomotor pattern generation: insights from deletions
University Press, New York during fictive locomotion. J Physiol 577:617–639
Pace RW, Mackay DD, Feldman JL, Del Negro CA Rybak IA, Stecina K, Shevtsova NA, McCrea DA (2006b)
(2007) Inspiratory bursts in the preBötzinger complex Modelling spinal circuitry involved in locomotor pat-
depend on a calcium-activated non-specific cation cur- tern generation: insights from the effects of afferent
rent linked to glutamate receptors in neonatal mice. stimulation. J Physiol 577:641–658
J Physiol 582:113–125 Rybak IA, Abdala APL, Markin SN, Paton JFR, Smith JC
Paton JFR (1996) A working heart-brainstem preparation (2007) Spatial organization and state-dependent mech-
of the mouse. J Neurosci Methods 65:63–68 anisms for respiratory rhythm and pattern generation.
Rabe Bernhardt N, Memic F, Gezelius H, Thiebes AL, Prog Brain Res 165:201–220
Vallstedt A, Kullander K (2012) DCC mediated axon Rybak IA, Shevtsova NA, Kiehn O (2013) Modelling
guidance of spinal interneurons is essential for normal genetic reorganizations in the mouse spinal cord
Vestibular System: Overview 121
affecting left-right coordination during locomotion. Wang X-J, Rinzel J (1992) Alternating and synchronous
J Physiol 591:5491–5508 rhythms in reciprocally inhibitory model neurons.
Sherwood WE, Harris-Warrick R, Guckenheimer J (2011) Neural Comput 4:84–97
Synaptic patterning of left-right alternation in Whelan PJ (2010) Shining light into the black box of
a computational model of the rodent hindlimb central spinal locomotor networks. Philos Trans R Soc Lond
pattern generator. J Comput Neurosci 30:323–360 B Biol Sci 365:2383–2395
Skinner FK, Kopell N, Marder E (1994) Mechanisms for Wolf E, Soffe S, Roberts A (2009) Longitudinal neuronal
oscillation and frequency control in reciprocally inhib- organization and coordination in a simple vertebrate:
itory model neural networks. J Comput Neurosci a continuous, semi-quantitative computer model of the
1:69–87 central pattern generator for swimming in young frog
Smith JC, Feldman JL (1987) In vitro brainstem-spinal tadpoles. J Comput Neurosci 27:291–308
cord preparations for study of motor systems for mam- Zagoraiou L, Akay T, Martin JF, Brownstone RM, Jessell
malian respiration and locomotion. J Neurosci TM, Miles GB (2009) A cluster of cholinergic
Methods 21:321–333 premotor interneurons modulates mouse locomotor
Smith JC, Ellenberger HH, Ballanyi K, Richter DW, activity. Neuron 64:645–662
Feldman JL (1991) Pre-Bötzinger complex: a brain Zhang Y, Narayan S, Geiman E, Lanuza GM,
stem region that may generate respiratory rhythm in Velasquez T, Shanks B, Akay T, Dyck J, Pearson K,
mammals. Science 254:726–729 Gosgnach S, Fan CM, Goulding M (2008) V3 spinal
Smith JC, Butera RJ, Koshiya N, Del Negro C, Wilson neurons establish a robust and balanced locomotor
CG, Johnson SM (2000) Respiratory rhythm genera- rhythm during walking. Neuron 60:84–96
tion in neonatal and adult mammals: the hybrid Zhong G, Shevtsova NA, Rybak IA, Harris-Warrick RM
pacemaker-network model. Respir Physiol (2012) Neuronal activity in the isolated mouse spinal
122:131–147 cord during spontaneous deletions in fictive locomo-
Smith JC, Abdala AP, Koizumi H, Rybak IA, Paton JF tion: insights into locomotor central pattern generator
(2007) Spatial and functional architecture of the mam- organization. J Physiol 590:4735–4759
malian brain stem respiratory network: a hierarchy of
three oscillatory mechanisms. J Neurophysiol
98:3370–3387
Smith JC, Abdala AP, Borgmann A, Rybak IA, Paton JF
(2013) Brainstem respiratory networks: building Vestibular System: Overview
blocks and microcircuits. Trends Neurosci
36:152–162
Spardy LE, Markin SN, Shevtsova NA, Prilutsky BI, Americo Migliaccio
Rybak IA, Rubin JE (2011a) A dynamical systems Balance and Vision Laboratory, Neuroscience
analysis of afferent control in a neuromechanical Research Australia, University of New South
model of locomotion: I. Rhythm generation. J Neural
Eng 8:065003
Wales, Sydney, NSW, Australia
Spardy LE, Markin SN, Shevtsova NA, Prilutsky BI,
Rybak IA, Rubin JE (2011b) A dynamical systems
analysis of afferent control in a neuromechanical Detailed Description
model of locomotion: II. Phase asymmetry. J Neural
Eng 8:065004
Stuart DG, Hultborn H (2008) Thomas Graham Brown The vestibular system is our sixth sense. Not only
(1882–1965), Anders Lundberg (1920–), and the neu- does it contribute to our sense of orientation and
ral control of stepping. Brain Res Rev 59:74–95 movement in space but, via the vestibulo-ocular
Tabak J, Rinzel J, O’Donovan MJ (2001) The role of
activity-dependent network depression in the expres-
reflex (VOR), it stabilizes our vision, especially
sion and self-regulation of spontaneous activity in the during rapid unpredictable head movements.
developing spinal cord. J Neurosci 21:8966–8978 Without the VOR our visual world would appear
Talpalar AE, Bouvier J, Borgius L, Fortin G, Pierani A, to move every time we undertook an activity that
Kiehn O (2013) Dual mode operation of neuronal
networks involved in left-right alternation. Nature
involves head movement, such as during walk-
500:85–88 ing, running, climbing, or driving on a bumpy
Toporikova N, Butera RJ (2011) Two types of indepen- road.
dent bursting mechanisms in inspiratory neurons: an The vestibular organ is located in the inner ear
integrative model. J Comput Neurosci 30:515–528
Vallstedt A, Kullander K (2013) Dorsally derived spinal
next to the cochlear. As our head movements
interneurons in locomotor circuits. Ann NY Acad Sci consist of rotations (nodding in affirmation) and
1279:32–42 translations (jumping up and down when happy),
122 Vestibular System: Overview
the vestibular organ comprises two sensor sys- damage to the vestibular end organ and associ-
tems. The semicircular canals sense angular head ated nerves. The most common of these include
rotations while the otolith organs sense linear age-related cell loss, tumors, head trauma, oto-
head translations (See ▶ Vestibular Otoliths, toxic injury (e.g., by aminoglycoside drugs),
Response to Vibration and Sound). Both the Meniere’s disease, infection (meningitis and ves-
semicircular canals and otolith organs rely on tibular neuritis), and cochlear implantation
specialized hair cells and the inertia of fluid (See ▶ Vestibular Function After Cochlear
(canals) or a gelatinous mass (otoliths) to bend Implantation). About 1 % of adults under 65 and
the sensory hairs. Because there are two types of 25 % over 65 will have an injury to the vestibular
hair cells and because individual vestibular affer- organs. Other patients will have normal function-
ents receive input from different combinations of ing vestibular organs and nerves but suffer from
these hair cells, the signals coming from the ves- abnormalities of the vestibular organs that result
tibular periphery contain a wealth of sensory in conditions such as benign paroxysmal posi-
information that includes head-rotation velocity tional vertigo or superior canal dehiscence syn-
and acceleration, head linear acceleration, or tilt drome. Because the vestibular system operates
angle (See ▶ Peripheral Vestibular Signal over a large range of head-rotation frequencies
Processing). This information is encoded in the (0 to ~15 Hz) and comprises the canals and oto-
signals coming from the peripheral vestibular lith organs, a range of diagnostic tests have been
organs but further processing occurs centrally developed to test for the different conditions
(See ▶ Central Vestibular Signal Processing) to above including the following: head impulse test-
generate the appropriate three-dimensional ing (See ▶ Vestibular, Canal Testing: the Head
(horizontal, vertical, and torsional) eye move- Impulse Test), video (See ▶ Vestibular, Eye
ments required to stabilize vision and the muscle Movement Testing), imaging, and vestibular
responses that equilibrate balance. Some of this evoked myogenic potential testing. For example,
processing is very rapid. For example, the direct the head impulse test is used to determine the
pathway of the VOR consists of a three-neuron function of each of the six semicircular
arc that has a transmission delay of ~7 ms in canals and their associated nerves during physio-
primates. In fact, the VOR is our fastest neural logically relevant rapid head rotations. For
reflex. patients with complete vestibular organ loss, ves-
The central vestibular neural circuitry, located tibular function can be restored by a vestibular
in the vestibular nuclei of the brainstem, receives prosthesis that senses head movements and trans-
input from the vestibulo-cerebellum to sustain, mits an appropriately encoded electrical signal to
modify, and train the vestibular response (See drive the vestibular nerves (See ▶ Vestibular
▶ Vestibular Adaptation and Compensation). Prosthesis, Interface). For prosthesis recipients
Thus, the vestibular system is an exquisitely and also patients with incomplete vestibular
adaptable system during normal function. loss, vestibular rehabilitation is essential for max-
For example, the VOR response increases imum recovery (See ▶ Vestibular, Rehabilita-
almost immediately when needed to view close tion). Ideally, vestibular rehabilitation exercises
objects or scenes, as if that context had been are tailored for each patient so that compensatory
preprogrammed and can be trained to increase mechanisms are best enabled. The accepted tech-
or decrease when wearing magnifying or minify- niques developed to diagnose, cure, and rehabil-
ing lenses. The vestibular system also has the itate vestibular injury have helped many patients
ability to recalibrate to compensate for perma- because they are based on our fundamental
nent injury to one (complete injury) or both understanding of vestibular signals and their
(incomplete injury) of the vestibular organs and processing. This is the focus of the articles in
their associated nerves. There are many causes of this section.
Vestibular System: Overview 123
Cross-References ▶ Vestibular Otoliths, Response to Vibration and

Sound
▶ Central Vestibular Signal Processing ▶ Vestibular Prosthesis, Interface
▶ Peripheral Vestibular Signal Processing ▶ Vestibular, Canal Testing: the Head Impulse
▶ Vestibular Adaptation and Compensation Test
▶ Vestibular Function After Cochlear ▶ Vestibular, Eye Movement Testing
Implantation ▶ Vestibular, Rehabilitation
A
Absorbing Boundary models of how evidence could (and should) be

accumulated are available and have a rich history
▶ Decision-Making, Threshold of accounting for performance in laboratory
tasks; and (3) there is an apparent disconnect
between the hundreds of milliseconds over
which animals can integrate evidence and the
Accumulation of Evidence in individual computing elements of the brain, neu-
Decision-Making rons, which integrate their inputs over a small
number of milliseconds.
Alexander C. Huk, Leor N. Katz and Although accumulating evidence over time is
Jacob L. Yates a central component of both cognitive function
Departments of Neuroscience and Psychology, and many statistical models for decision making,
Center for Perceptual Systems, The University of the current emphasis on this topic is likely driven
Texas at Austin, Austin, TX, USA by the simple fact that remarkably direct neural
correlates of such temporal accumulation appear
to have been found in the spiking responses of
Definition neurons in brain areas such as the posterior pari-
etal cortex, prefrontal cortex, and other areas with
Accumulation of evidence in decision making is premotor functions (Gold and Shadlen 2007).
the process by which noisy sensory information is
sequentially sampled until sufficient evidence has Mathematical Foundations
accrued to favor one decision over another or The computational neuroscience of evidence
others. accumulation starts with Bloch’s Law,
a fundamental principle of sensory processing
(Bloch 1885). Bloch’s Law is a description of
Detailed Description temporal summation, such that if more time
results in more signal, simple behaviors such as
The accumulation of evidence over time is detection will show increases in accuracy that
a central topic in computational neuroscience depend on the product of stimulus intensity and
spanning behavior, brain, and theory (Huk and time. At longer durations, a breakdown in
Meister 2012; Shadlen et al. 2006; Usher and Bloch’s Law is interpreted as the limit of the
McClelland 2001): (1) it is a fundamental aspect temporal summation properties of the transducer.
of tractable forms of cognition, such as simple Despite its historical significance, Bloch’s Law is
forms of decision making; (2) mathematical rarely applied to modern decision-making tasks,

A 126 Accumulation of Evidence in Decision-Making
as it focuses on “sufficiently short” durations and psychological mechanisms: the number of sam-
does not explicitly model noise, a critical tool for ples required to reach a bound reflects decision
manipulating difficulty in many decision-making time, and the identity of the bound reached
paradigms. reflects the decision. The weight of sampled evi-
As opposed to the linear dependence of accu- dence is a function of stimulus strength such that
racy on time of Bloch’s Law, statistical models a strong stimulus in favor of H1 will require
within the “sequential sampling” framework fewer samples to reach the H1 bound, resulting
assume that significant noise is present and that in faster reaction times and higher accuracy. For
by accumulating many samples, accuracy can be choices where H1 and H2 are equally likely, the
improved. Under the simplest conditions of inde- bounds are equidistant from the starting point, the
pendent additive noise at each instant in time, magnitude of which primarily reflects the tight
accuracy based on ideal temporal accumulation coupling between decision time and accuracy:
will improve as a function of the square root of fast decisions come at the cost of accuracy, and
time. This root-time improvement in accuracy is high accuracy comes at the cost of time (the
directly related to the square root (n) term that speed-accuracy trade-off). A shift in starting
many students will recognize from basic point position towards H1 may reflect an
statistics. a priori bias, resulting in a higher proportion of
This statistical point is the basis of a large H1 choices over H2, with faster decision times.
family of models in the “sequential sampling” The first major distinction between model
framework, a set of models built and adapted types is whether they posit a single accumulator
within statistics, physics, mathematical psychol- or multiple, racing accumulators. If there is
ogy, and neuroscience (Stone 1960; Watson a single accumulation process, models posit
1979; Luce 1986; Link 1992; Ditterich 2006). In a pair of decision bounds (for a two-alternative
essence, these models assume that noisy informa- paradigm), and the accumulation process starts in
tion about a stimulus is sampled sequentially, between these upper and lower limits. Thus,
until sufficient evidence has been accrued to incoming evidence is “signed” with positive evi-
favor a decision. For binary choices, the accumu- dence pushing the accumulator towards the upper
lation follows a noisy trajectory, where informa- bound and negative evidence pushing the accu-
tion is represented as a single quantity in which mulator towards the lower bound. In typical
input supporting one hypothesis is evidence “race” models, a pair of parallel and competing
against another (i.e., H1 vs. H2). Mathematically, mechanisms each accumulates evidence in favor
sampled evidence is weighted to support one of of their particular outcome, and the decision is
the two hypotheses and may be expressed as the determined by the first accumulator to reach its
ratio of likelihoods, the probability of observing respective bound (Usher and McClelland 2001).
the evidence (“e”) given the hypothesis p(e|H1)/ If the signal and noise available to both accumu-
p(e|H2). The logarithm of this quantity may be lators are identical, such perfectly correlated rac-
summed over time to represent the degree of ing accumulators make the same predictions as
evidence accumulated in favor of one of the a single accumulator, despite the difference in
hypotheses over the other, termed the “log likeli- hypothesized architecture. On the other hand,
hood ratio.” This value can be used to determine racing mechanisms can exhibit increasingly com-
the optimal stopping rule for an accumulation plex behaviors as additional parameters between
process as implemented in the sequential proba- them become statistically uncorrelated. Indepen-
bility ratio test (Wald 1947). dent noise changes the behavior of racing versus
single-variable accumulators, and more baroque
Evidence Accumulation in Models of Decision models can behave quite differently than
Making a standard single accumulator (Ditterich 2006).
This family of models has been extensively used The second major distinction within sequen-
to account for behavioral data and linked to tial sampling models is whether they model time
Accumulation of Evidence in Decision-Making 127 A
as discrete or continuous. Discrete models are trial-to-trial variability in drift rate and residual
more straightforward to implement in computer time can capture phenomena like delayed reac-
code, while continuous models avail themselves tion times for incorrect trials (Ratcliff 1978; A
to closed-form mathematics. While it is of course Ratcliff and Rouder 1998; Ratcliff and McKoon
true that in the limit (i.e., as discrete time steps go 2008).
to infinitesimals) discrete models become contin-
uous, this requires some care in practice. It is Evidence Accumulation in Neurons
critical that discrete approximations of continu- The presence or implementation of diffusion-like
ous models be implemented using appropriately algorithms in the brain is still a matter of signifi-
fine time scale steps. Otherwise, small rounding cant interest and contention. The first direct evi-
artifacts can manifest themselves in peculiar dence for a putative diffusion process in the brain
behaviors of the model. It is therefore advisable was observed in the spike rates of single neurons in
that discrete models be implemented with checks the lateral intraparietal sulcus (LIP) of macaque
relative to known properties of the continuous- monkeys during a stochastic motion discrimina-
math model. Modern computers make the pros- tion task where the experimenter can control the
pect of fine time steps less of a practical (speed) amount of motion (the “evidence”) on a single trial
concern, but counterintuitive or quirky predic- (Shadlen and Newsome 2001; Roitman and
tions of a discrete modeling exercise should be Shadlen 2002). Neurons in the LIP have a mean
interpreted with caution. spike rate that ramps up for choices that result in
an eye movement into their response field
The Drift-Diffusion Model (RF) and ramps down for choices out of their
Perhaps the most common model of evidence RF. Moreover, the slope of the ramp is steeper
accumulation is the (continuous time) diffusion for trials with more evidence, mirroring the drift
model, increasingly referred to as the “drift- rate of a diffusion process. Time-varying motion
diffusion to bound” model (DDM) (Shadlen stimuli have further supported the notion that LIP
et al. 2006; Ratcliff 1978; Ratcliff and Rouder neurons reflect temporal accumulation
1998; Palmer et al. 2005; Ratcliff and McKoon (integration) of relevant sensory signals (Huk and
2008). It stands as a hub because it is mathemat- Shadlen 2005; Kiani et al. 2008). Other work has
ically tractable (i.e., the psychometric and chro- shown that the spike rates of single neurons in
nometric functions can be derived analytically) a range of cortical and subcortical brain areas
and can be flexibly extended to capture a wide also exhibit correlates of a diffusion process
range of behavioral data. The simplest diffusion (Ratcliff et al. 2007; Ding and Gold 2012a, b).
models are composed of three parameters: a drift To date, these early investigations shed little
rate, bound height, and accumulator noise. The light on exactly how neurons might implement
drift rate relates the stimulus units to a rate of drift time integration, yet understanding the biophysi-
of the diffusion process. The bound height is the cal mechanisms may provide significant insights
stopping point (for two alternatives, e.g., H1 and into the cognitive constraints of evidence accu-
H2, symmetrically above and below an unbiased mulation (Huk and Meister 2012; Wong and
start point). Accumulator noise is invariant across Wang 2006). Likewise, the current focus on
stimulus conditions and describes the noise in the responses as averaged over neurons, trials, and
accumulation process. This version of the model sessions paints a categorical and potentially inac-
is overdetermined and can be rewritten as a - curate picture of what occurs on the time scale of
two-parameter model with bound height or individual trials and decisions (Churchland
noise fixed. The diffusion model can be easily et al. 2011). The next generation of work on this
extended. For example, an additive nondecision topic will hopefully unpack how the theoretical
time may be included to account for mechanisms of evidence accumulation are
nondecisional sensory transduction and motor implemented by real neural circuits (Bollimunta
execution delays. Parameters added to model et al. 2012).
A 128 Acoustic Memory
References Shadlen MN, Newsome WT (2001) Neural basis of

a perceptual decision in the parietal cortex (area LIP)
Bloch AM (1885) Expériences sur la vision. C R Seance of the rhesus monkey. J Neurophysiol 86:1916–1936
Soc Biol Paris 37:493–495 Shadlen MN, Hanks TD, Churchland AK, Kiani R, Yang
Bollimunta A, Totten D, Ditterich J (2012) Neural dynam- T (2006) The speed and accuracy of a simple percep-
ics of choice: single-trial analysis of decision-related tual decision: a mathematical primer. In: Doya K,
activity in parietal cortex. J Neurosci Ishii S, Rao R, Pouget A (eds) Bayesian brain: proba-
32(37):12684–12701 bilistic approaches to neural coding. MIT Press, Cam-
Churchland AK, Kiani R, Chaudhuri R, Wang XJ, bridge, pp 209–237
Pouget A, Shadlen MN (2011) Variance as Stone M (1960) Models for choice-reaction time.
a signature of neural computations during decision Psychometrika 25:251–260
making. Neuron 69(4):818–831 Usher M, McClelland JL (2001) The time course of per-
Ding L, Gold JI (2012a) Separate, causal roles of the ceptual choice: the leaky, competing accumulator
caudate in saccadic choice and execution in model. Psychol Rev 108(3):550–592
a perceptual decision task. Neuron 75:865–874 Wald A (1947) Sequential analysis. Wiley, New York
Ding L, Gold JI (2012b) Neural correlates of perceptual Watson AB (1979) Probability summation over time.
decision making before, during, and after decision Vision Res 19(5):515–522
commitment in monkey frontal eye field. Cereb Cortex Wong K-F, Wang XJ (2006) A recurrent network mecha-
22:1052–1067 nism of time integration in perceptual decisions.
Ditterich J (2006) Stochastic models of decisions about J Neurosci 26(4):1314–1328
motion direction: behavior and physiology. Neural
Netw 19(8):981–1012
Gold JI, Shadlen MN (2007) The neural basis of decision
making. Annu Rev Neurosci 30:535–574
Huk AC, Meister MLR (2012) Neural correlates and neu- Acoustic Memory
ral computations in posterior parietal cortex during
perceptual decision-making. Front Integr Neurosci ▶ Auditory Memory
6:86
Huk AC, Shadlen MN (2005) Neural activity in macaque
parietal cortex reflects temporal integration of visual
motion signals during perceptual decision making.
J Neurosci 25(45):10420–10436 Acoustic Timbre Recognition
Kiani R, Hanks TD, Shadlen MN (2008) Bounded inte-
gration in parietal cortex underlies decisions even
when viewing duration is dictated by the environment. Daniel Pressnitzer1, Trevor Agus2 and
J Neurosci 28(12):3017–3029 Clara Suied3
Link SW (1992) The wave theory of difference and sim- 1
Département d’Études Cognitives, École
ilarity. Erlbaum, Hillsdale
Luce RD (1986) Response times. Oxford University Press, Normale Supérieure, Paris, France
2
New York Equipe Audition, Département d’Études
Palmer J, Huk AC, Shadlen MN (2005) The effect of Cognitives, École Normale Supérieure, Paris,
stimulus strength on the speed and accuracy of France
a perceptual decision. J Vis 5(5):376–404 3
Ratcliff R (1978) A theory of memory retrieval. Psychol École Normale Supérieure, Paris, France
Rev 85(2):59–108
Ratcliff R, McKoon G (2008) The diffusion decision
model: theory and data for two-choice decision tasks.
Neural Comput 20(4):873–922 Synonyms
Ratcliff R, Rouder JN (1998) Modeling response times for
two-choice decisions. Psychol Sci 9(5):347–356 Auditory recognition; Sound source
Ratcliff R, Hasegawa YT, Hasegawa RP, Smith PL, identification
Segraves MA (2007) Dual diffusion model for single-
cell recording data from the superior colliculus in a -
brightness-discrimination task. J Neurophysiol
97(2):1756–1774 Definition
Roitman JD, Shadlen MN (2002) Response of neurons in
the lateral intraparietal area (LIP) during a combined
visual discrimination reaction time task. J Neurosci Timbre is what allows a listener to distinguish
22:9475–9489 two sounds that have otherwise the same
Acoustic Timbre Recognition 129 A
Acoustic Timbre Recognition, Fig. 1 Visual represen- is the average energy over time. The two instruments
tations of four sounds with the same duration, loudness, illustrate classic dimensions of timbre: depending on the
and pitch, only differing by timbre. Each panel displays sound source and how it is excited, the attack time can be
a time-frequency analysis derived from an auditory model fast (piano) or slow (trombone); the spectral center of
(see Agus et al. 2012 for details). Briefly, color indicates mass can be high (piano) or low (trombone). The two
the pattern of energy within frequency channels (y-axis) as vowels illustrate that other possibly more complex fea-
it evolves over time (x-axis). The top trace is the tures may also be used to distinguish, e.g., vowels from
corresponding pressure waveform. The right-hand trace instruments or vowels from each other
subjective pitch, loudness, location, and duration. potential cue for sound source identification.
For instance, when orchestral musicians tune at Then we put forward two possible approaches to
the beginning of a concert, they all play the same timbre, which we follow into the fields of acous-
note, but one can still tell the difference between tics, perception, neural mechanisms, and compu-
instruments. This is largely because of timbre. tational applications.
Why Do Different Sound Sources Produce

Detailed Description Different Timbres?
Sound sources are physical objects that come in
The standard definition of timbre has several all shapes and sizes. Sound is produced
shortcomings. First, it says what timbre is not, when some energy makes the object vibrate.
rather than what it is. Second, it relates to the The vibrations spread around the source, which
comparison between two sound tokens, whereas then propagate to the air and reach the ear
a more useful function for hearing is to associate of a listener in the form of pressure waves
a single timbre directly with a sound source (the (Fig. 1). Simple physics shows that the wave
timbre of the piano, the timbre of the voice of pattern at the ear can contain a lot of information
a friend). Perhaps as a consequence, there is still about what happened at the source (Helmholtz
a lively debate about the acoustic features, mental 1877). For instance, if the energy input was
representations, and neural mechanisms underly- brief, such as a door knock, the chances are
ing timbre recognition. Here, we first outline the that the sound itself will be brief and have
basic principles that make timbre such a powerful most of its energy concentrated around the
A 130 Acoustic Timbre Recognition
time of the knock. After the knock, the way the

door continues to vibrate is closely related to
its geometry, because some wave patterns are
consistent with some geometries and some are
not. One such rule is that waves with low fre-
quency and thus a long wavelength are not stable
within small objects. Thus, the proportions of
different frequency components that combine to
make the sound of a door knock will be
constrained by the size of the door. Other, more
complex rules apply, depending on the shape of
the object, the nature of the materials involved,
Acoustic Timbre Recognition, Fig. 2 Schematic rep-
and so on.
resentation of the dimensions approach versus the features
Being able to decode the intricate links approach for timbre. (a) For the dimensions approach, all
between wave patterns and sound sources is different timbres can be projected in a low-dimensional
extremely useful for humans and other animals. space of continuous dimensions. (b) For the features
approach, each timbre is defined by a set of distinctive
It allows the auditory system to serve as
features among a very large and unordered set of possible
a warning sense, for instance, to identify sound- features
producing objects that are out of sight. For peo-
ple, it is also the very basis of spoken language:
vowels and consonants are produced by modulat- It is likely that a full account of timbre will lie
ing the shape of the vocal apparatus, resulting in somewhat in between these two simplified
changes in timbre that are the building blocks of hypotheses. However, for clarity, we continue to
oral communication. contrast each approach for different aspects of
timbre research.
Dimensions Versus Features
There is no consensus on what makes timbre Sound Representations
recognition possible for human listeners. To out- To investigate timbre, it is useful to represent
line current controversies, it is useful to consider sound visually. Classically, this has been done
two opposite viewpoints (Fig. 2). A first view is with tools such as the trace of the pressure wave-
that timbre is composed of a reasonably small form over time; the spectral analysis of compo-
number of perceptual dimensions, which are sub- nent frequencies through, e.g., Fourier analysis;
jective descriptions of sound just as pitch or loud- or spectro-temporal transformations such as the
ness. Such dimensions must be metameric, in that short-term Fourier transform or wavelet analyses.
several different sounds may project to the same More recently, computational models that aim to
point on the dimension. mimic peripheral or central auditory processing
A second view is that timbre recognition relies have been suggested (e.g., Patil et al. 2012).
on the distinctive features of a given sound In the “dimensions” approach, summary sta-
source, learned through experience and selected tistics are computed on sound representations to
among a very large space of potential features. define what are referred to as descriptors of tim-
The grain of a friend’s voice may be unique, bre. For instance, the center of mass of all fre-
which is what allows us to recognize her quency components of a sound produces a single
instantly. Such features would be conceptually number that is correlated with the apparent
different from dimensions in that a feature does “brightness” of a sound (McAdams et al. 1995).
not necessarily apply to all possible sound In the “features” approach, the tendency is rather
sources; in fact, it is precisely because it is unique to maximize the richness of the representation, by
to only a few sources (or even a single source) including complex spectro-temporal selectivities.
that it could be efficient for recognition. Such a feature-based representation need not be
Acoustic Timbre Recognition 131 A
orderly. It can be over-complete with thousands known as mismatch negativity, it has been
of partially overlapping features, or sparse, in the found that timbre dimensions such as brightness
sense that a given sound would only activate or onset time could each be represented A
a small number of features within that large pos- separately within auditory cortex (Caclin
sible space (Hromadka and Zador 2009). et al. 2006).
From the features perspective, single-unit
Perceptual Data recordings have uncovered a rich variety of selec-
The basic aim of the dimensions approach is to tivities, at many levels of the auditory system,
uncover the nature and number of the perceptual often without any obvious ordering principle
dimensions underlying timbre. To this effect, sta- (other than by frequency). Using linear analysis
tistical techniques based on multidimensional techniques such as reverse correlation, spectro-
scaling have been used: a pair of sounds is temporal receptive fields have been derived. Var-
presented to the listener, who has to rate how ious spectral and temporal modulation prefer-
similar to each other the two sounds seem. This ences have been observed, e.g., in the primary
is repeated for all possible pairs within a given auditory cortex (Depireux et al. 2001). Adding
sound set. Then, the similarity judgments are a nonlinear component to the analysis adds
treated as perceptual distances and used to obtain another layer of complexity (Machens
the dimensionality and geometry of the et al. 2004). Furthermore, the neural encoding
corresponding mental representation. For musi- of timbre may interact with supposedly indepen-
cal instruments, classic studies point toward two dent sound characteristics, such as pitch or loca-
to three main dimensions: one related to the tion (Bizley et al. 2009).
attack time, one related to the spectral center of A further question is whether the identity of
mass, and one additional dimension that is less a source will be encoded by the activity of a wide
consistently observed (Grey 1977; McAdams network shared by many sound sources or by the
et al. 1995). More recent investigations, using activity of only a small network specifically
both multidimensional scaling and verbal tuned to that source category. Evidence has
descriptions, suggest five main dimensions with been put forward for both models. Using fMRI,
more complex interpretations (Elliott et al. 2013). the identity of a sound source can be inferred
In the features approach, the focus is not on from distributed activity (Staeren et al. 2009).
similarity but rather on the recognition of the At the same time, there are clear indications of
sound source. Again, using musical instruments, localized brain areas specialized for familiar
fast recognition times have been observed (Agus sound sources such as the human voice (Belin
et al. 2012), and recognition was found to be 2006).
preserved even for severely impoverished signals
(Suied et al. 2013). Moreover, recognition was Timbre Recognition by Machines
faster and more robust for highly familiar sources There are several applications for acoustic timbre
such as the human voice, an observation that recognition, such as speaker identification or
could not be traced back to simple acoustic music information retrieval. Even though
dimensions (Agus et al. 2012). These results the techniques used are fast evolving and
strongly suggest the existence of diagnostic fea- a detailed description is beyond the scope of
tures that were learned by listeners, through expe- this section, it is interesting to note that the
rience, to recognize, e.g., voices in a robust and dimensions versus features contrast can also be
efficient manner. seen in the architectures of the computational
systems.
Neural Bases Automatic speech recognition, which can to
Neural correlates of generic timbre dimensions some extent be viewed as a timbre-decoding
have been investigated with brain imaging. Using exercise, has a long tradition of performing clas-
an EEG paradigm to probe sensory memory sification on a small number of generic
A 132 Acoustic Timbre Recognition
coefficients (e.g., mel-frequency cepstrum coef- References

ficients and their variants (Hermansky 1990)).
For musical instruments, a descriptor-based Agus TR, Suied C, Thorpe SJ, Pressnitzer D (2012) Fast
recognition of musical sounds based on timbre.
approach has been directly inspired by the per-
J Acoust Soc Am 131:4124–4133
ceptual dimensions of multidimensional studies, Belin P (2006) Voice processing in human and non-human
with a reasonably small number of explicit primates. Philos Trans R Soc Lond SerB Biol Sci
descriptors (Peeters et al. 2011). However, other 361:2091–2107
Bizley JK, Walker KM, Silverman BW, King AJ,
systems exist that are based on feature generation
Schnupp JW (2009) Interdependent encoding of
from a huge potential feature space, followed by pitch, timbre, and spatial location in auditory cortex.
ad hoc selection for a given classification task J Neurosci Off J Soc Neurosci 29:2064–2075
(Coath and Denham 2005; Pachet and Roy Caclin A, Brattico E, Tervaniemi M, Naatanen R,
Morlet D, Giard MH, McAdams S (2006) Separate
2009). For musical instrument classification,
neural processing of timbre dimensions in auditory
machine-learning algorithms applied on a high- sensory memory. J Cognit Neurosci 18:1959–1972
dimensional auditory model representation have Coath M, Denham SL (2005) Robust sound classification
also been successfully demonstrated (Patil through the representation of similarity using response
fields derived from stimuli during early experience.
et al. 2012).
Biol Cybern 93:22–30
Depireux DA, Simon JZ, Klein DJ, Shamma SA (2001)
Perspectives Spectro-temporal response field characterization with
The outstanding issues for timbre research will dynamic ripples in ferret primary auditory cortex.
probably benefit from considering the various
Elliott TM, Hamilton LS, Theunissen FE (2013) Acoustic
strategies available to a listener. For instance, structure of the five perceptual dimensions of timbre in
when asked for subjective distance judgments, orchestral instrument tones. J Acoust Soc Am
the most reasonable thing to do may be to 133:389–404
Formisano E, De Martino F, Bonte M, Goebel R (2008)
abstract common dimensions to a sound set, and
“Who” is saying “what”? Brain-based decoding of
then use those for the comparisons. However, human voice and speech. Science 322:970–973
when asked to recognize a source as fast as pos- Fritz J, Shamma S, Elhilali M, Klein D (2003) Rapid task-
sible, the mere presence of a diagnostic feature related plasticity of spectrotemporal receptive fields in
primary auditory cortex. Nature Neurosci 6:1216–1223
may be sufficient. The set of useful timbre dimen-
Grey JM (1977) Multidimensional perceptual scaling of
sions or features can also depend on the task: for musical timbres. J Acoust Soc Am 61:1270–1277
a same set of spoken words, different strategies Helmholtz H (1877) On the sensations of tone. Dover,
are used if listeners are asked to identify the New York
Hermansky H (1990) Perceptual linear predictive (PLP)
speaker or report the word content (Formisano
analysis of speech. J Acoust Soc Am 87:1738–1752
et al. 2008). Finally, the very neural representa- Hromadka T, Zador AM (2009) Representations in audi-
tion of timbre may be dynamically tuned to the tory cortex. Curr Opin Neurobiol 19:430–433
immediate acoustic context, through rapid plas- Machens CK, Wehr MS, Zador AM (2004) Linearity of
cortical receptive fields measured with natural sounds.
ticity (Fritz et al. 2003). A fundamental reason
J Neurosci Off J Soc Neurosci 24:1089–1100
that makes timbre so elusive may therefore be McAdams S, Winsberg S, Donnadieu S, De Soete G,
that timbre recognition is a profoundly adaptive Krimphoff J (1995) Perceptual scaling of synthesized
mechanism, able to create and use opportunistic musical timbres: common dimensions, specificities,
and latent subject classes. Psychol Res 58:177–192
strategies that depend on the sounds and task
Pachet F, Roy P (2009) Analytical features: a knowledge-
at hand. based approach to audio feature generation. EURASIP
J Audio Speech Music Process 2009
Patil K, Pressnitzer D, Shamma S, Elhilali M (2012)
Music in our ears: the biological bases of musical
Cross-References timbre perception. PLoS Comput Biol 8:e1002759
Peeters G, Giordano BL, Susini P, Misdariis N, McAdams
S (2011) The timbre toolbox: extracting audio descrip-
▶ Auditory Event-Related Potentials tors from musical signals. J Acoust Soc Am
▶ Pulse-Resonance Sounds 130:2902–2916
Action Potential Backpropagation 133 A
Staeren N, Renvall H, De Martino F, Goebel R, Formisano Detailed Description
E (2009) Sound categories are represented as distrib-
uted patterns in the human auditory cortex. Curr Biol
19:498–502 Simultaneous recordings from dendrites, soma, and A
Suied C, Agus TR, Thorpe S, Pressnitzer D (2013) axon have shown that action potentials are typically
Processing of short auditory stimuli: the rapid audio generated in the region with the lowest threshold
sequential presentation paradigm (RASP). In: Moore for AP initiation, the axon initial segment (Stuart
BCJ, Patterson RD, Winter IM, Carlyon RP, Gockel
HE (eds) Basic aspects of hearing: physiology and et al. 1997). In addition to the canonical forward
perception. Springer, New York propagation along the axon to the presynaptic ter-
minals, APs rapidly invade the soma and propagate
back into the dendrites, where voltage-dependent
channels actively support the depolarization.
As opposed to axonal action potentials, back-
Action Planning propagating action potentials (bAPs) are not
all-or-none events; their amplitude can be modu-
▶ Decision-Making, Motor Planning lated and generally decreases along the dendrites
with the distance from the soma. The extent of
backpropagation varies widely in the different
neuronal types investigated, ranging from
non-decremental to almost passive (Fig. 1).
Action Potential Backpropagation
Determinants of Action Potential
Sonia Gasparini1 and Michele Migliore2,3 Backpropagation
1 The main factors that determine the extent of action
Neuroscience Center, LSUHSC, New Orleans,
LA, USA potential backpropagation are (a) density and prop-
2 erties of dendritic voltage-dependent channels,
Department of Neurobiology, Yale University
School of Medicine, New Haven, CT, USA (b) neuronal morphology, and (c) neuronal activity
3 (firing history and incoming inputs).
Institute of Biophysics, National Research
Council, Palermo, Italy
Dendritic Voltage-Dependent Channels
Different neuronal types have distinct distributions
of dendritic voltage-dependent ion channels
Synonyms (Migliore and Shepherd 2002, 2005; Magee
2008). AP backpropagation is sustained by den-
Backpropagating action potentials; Backpropagating
dritic voltage-dependent Na+ channels and acti-
spikes vates Ca2+ channels (Johnston et al. 1996).
Dendritic K+ channels can modulate the amplitude
and extent of backpropagation; their role has been
Definition most extensively studied in hippocampal CA1
pyramidal neurons, where the density of A-type
Action potential (AP) backpropagation, as K+ currents increases along the apical dendrites
opposed to forward propagation along the axon, with the distance from the soma (Johnston
consists of the active conduction along the neu- et al. 2000). The amplitude and extent of
ronal dendrites of the depolarization produced by backpropagation along the dendrites
the initiation of an axonal AP. The amplitude of therefore depend on the relative density of dendritic
the depolarization generally decreases along the Na+ and K+ channels (see Fig. 2). Thus, modulation
dendrites with increasing distance from the soma; of dendritic ion channels by neurotransmitters or
the degree of attenuation is highly variable plasticity will alter the extent of propagation
depending on the neuronal type. (Johnston et al. 1999; Magee and Johnston 2005).
A 134 Action Potential Backpropagation
Action Potential Backpropagation, Fig. 1 Amplitude areas. The values were normalized to the amplitude of the
of bAPs along the apical dendrites, plotted as a function of somatic AP (Reproduced with permission from Waters
the distance from the soma, in neurons from various brain et al. 2005)
Morphology
It is almost impossible to discern the relative
contribution of dendritic morphology and ion
channel properties to action potential back-
propagation with experimental approaches; how-
ever, computational simulations have shown
that some differences in propagation can be
attributed exclusively to differences in dendritic
morphologies. When the same active and passive
parameters are inserted in compartmental models
with reconstructed morphologies of various neu-
ronal types, there is a strong correlation between
neuronal geometry (branch points and relative
impedance) and backpropagation (Vetter et al.
2001; Spruston et al. 2008, see Fig. 3). These
results indicate that dendritic morphology and
branching patterns have a fundamental role in
shaping backpropagation.
Neuronal Activity
Dendritic excitability and AP backpropagation
are highly dependent on the membrane potential
and therefore on concurrent neuronal activity.
Hippocampal CA1, entorhinal layer V, and neo-
cortical layer V (but not layer II/III) pyramidal
neurons show a significant activity-dependent
decrease in the amplitude of bAPs along the api-
cal dendrites during high-frequency trains
Action Potential Backpropagation, Fig. 2 Effect of (Johnston et al. 1999; Gasparini 2011). In CA1
different densities of dendritic Na+ and A-type K+ currents and entorhinal layer V pyramidal neurons, this
on AP backpropagation in CA1 pyramidal neurons. An
behavior is mostly due to a slow, cumulative
interactive example, where channel densities can be mod-
ified, is available at http://senselab.med.yale.edu/ inactivation of dendritic Na+ channels and can
ModelDB/ShowModel.asp?model=148646 be reduced by neurotransmitters activating
Action Potential Backpropagation 135 A
Action Potential
Backpropagation,
Fig. 3 Effect of dendritic
morphology on AP A
backpropagation, in
computational simulations
with identical dendritic
passive properties and
channel types and densities.
(a–c) color-coded
representation of dendritic
AP amplitude in
reconstructed
morphologies from
a substantia nigra
dopaminergic neuron (a),
neocortical layer V neuron
(b), and cerebellar Purkinje
neuron (c). (d–f) bAP
amplitude as a function of
the distance from the soma
for the cells in (a–c)
(Reproduced with
permission from Vetter
et al. 2001)
protein kinase C (Johnston et al. 1999; Magee neuron has generated an output. This feedback
2008). In addition, the amplitude of bAPs at distal signal has important consequences on neuronal
locations can be boosted by appropriately timed firing and synaptic plasticity. In neocortical layer
synaptic inputs, which promote backpropagation V neurons, pairing bAPs with a distal synaptic
by inactivating A-type K+ channels or by facili- input initiates a dendritic Ca2+ spike
tating Na+ channel activation (Migliore (backpropagation-activated Ca2+ or BAC spike),
et al. 1999; Spruston 2008; Fig. 4). which results in a burst of somatic APs (Spruston
2008). In many neurons, the appropriate timing of
Functional Role bAPs and EPSPs can cause long-term changes in
The depolarization and Ca2+ influx associated synaptic efficacy (spike-timing-dependent plas-
with bAPs provide feedback to the region that ticity or STDP). The depolarization associated
received the synaptic input, signaling that the with the bAP provides the coincidence detection
A 136 Action Potential Backpropagation
have advantages and limitations (Waters

et al. 2005; Scanziani and H€ausser 2009). For
this reason, the investigation of AP
backpropagation and its functional roles has
greatly benefitted from computational models
that use biophysically and morphologically accu-
rate implementations. These models have
supported, explained, and predicted several
experimental findings on AP backpropagation
(Schaefer et al. 2003; Watanabe et al. 2002).
Cross-References
▶ Action Potential Initiation

▶ Delayed Rectifier and A-Type Potassium
Action Potential Backpropagation, Fig. 4 Effect of
Channels
pairing a synaptic input on AP backpropagation in CA1
pyramidal neurons. If the timing is appropriate, bAP is ▶ Dopaminergic Cell Models
boosted by the synaptic input (red trace). An interactive ▶ High-Voltage-Activated Calcium Channels
example, where the timing between AP and synaptic input ▶ Long Term Plasticity, Biophysical Models
can be modified, is available at http://senselab.med.yale.
▶ Low-Voltage-Activated Calcium Channels
edu/ModelDB/ShowModel.asp?model=148646
▶ Multiscale Modeling of Purkinje Cells
▶ N-Methyl-D-Aspartate (NMDA) Receptors,
needed for the removal of Mg2+ from NMDA Conductance Models
glutamatergic receptors, triggering plasticity ▶ Neuromodulation: Overview
phenomena (Mainen and Abbott 2008). The ▶ NEURON Simulation Environment
Ca2+ influx caused by bAP has also been ▶ Patch Clamp Technique
suggested to mediate dendritic release of neuro- ▶ Reduced Morphology Models
transmitters (Ludwig and Pittman 2003). ▶ Short-Term Plasticity, Biophysical Models
▶ Sodium Channels
Techniques to Study Backpropagation ▶ Spike-Timing Dependent Plasticity (STDP),
The interaction of experimental and computa- Biophysical Models
tional methods has been essential to understanding ▶ Voltage Sensitive Dye Imaging, Intrinsic
the mechanisms underlying AP backpropagation, Optical Signals
allowing intrinsic limitations of the individual
approaches to be overcome.
Direct electrophysiological recordings with References
the patch-clamp technique in different configura-
tions have been performed at various distances Gasparini S (2011) Distance- and activity-dependent mod-
from the soma mostly on larger-diameter den- ulation of spike back-propagation in layer V pyramidal
drites (Gurkiewicz and Korngreen 2006), provid- neurons of the medial entorhinal cortex.
ing an idea of channel distributions (Magee 2008) Gurkiewicz M, Korngreen A (2006) Recording, analysis,
and backpropagation (Waters et al. 2005). Fluo- and function of dendritic voltage-gated channels.
rescence imaging techniques (using Ca2+- or Pflugers Arch 453:283–292
voltage-sensitive dyes) have been used to assess Johnston D, Magee JC, Colbert CM, Cristie BR
(1996) Active properties of neuronal dendrites. Annu
smaller dendrites and compartments and/or to Rev Neurosci 19:165–186
obtain data from many dendritic locations in Johnston D, Hoffman DA, Colbert CM, Magee JC
a single neuron. These experimental techniques (1999) Regulation of back-propagating action
Action Potential Initiation 137 A
potentials in hippocampal neurons. Curr Opin Further Reading
Neurobiol 9:288–292 Davie JT, Kole MH, Letzkus JJ, Rancz EA, Spruston N,
Johnston D, Hoffman DA, Magee JC, Poolos NP, Stuart GJ, H€ausser M (2006) Dendritic patch-clamp
Watanabe S, Colbert CM, Migliore M (2000) Den- recording. Nat Protoc 1:1235–1247 A
dritic potassium channels in hippocampal pyramidal Spruston N, H€ausser M, Stuart G (2013) Information
neurons. J Physiol 525:75–81 processing in dendrites and spines. In: Squire LR
Ludwig M, Pittman QJ (2003) Talking back: et al (eds) Fundamental Neuroscience, 4th edn.
dendritic neurotransmitter release. Trends Neurosci Elsevier, Amsterdam, pp 231–260
26:255–261
Magee JC (2008) Voltage-gated ion channels in
dendrites. In: Stuart G, Spruston N, Ha€ usser M (eds)
Dendrites. Oxford University Press, New York,
pp 225–250
Magee JC, Johnston D (2005) Plasticity of dendritic func- Action Potential Initiation
tion. Curr Opin Neurobiol 15:334–342
Mainen ZF, Abbott LF (2008) Functional plasticity at Dejan Zecevic and Marko Popovic
dendritic synapses. In: Stuart G, Spruston N, Ha€ usser
Department of Physiology, Yale University
M (eds) Dendrites. Oxford University Press, New
York, pp 465–498 School of Medicine, New Haven, CT, USA
Migliore M, Shepherd GM (2002) Emerging rules for the
distributions of active dendritic conductances. Nat Rev
Neurosci 3:362–370
Migliore M, Shepherd GM (2005) Opinion: an integrated
Synonyms
approach to classifying neuronal phenotypes. Nat Rev
Neurosci 6:810–818 Spike initiation
Migliore M, Hoffman DA, Magee JC, Johnston D (1999)
Role of an A-type K+ conductance in the back-
propagation of action potentials in the dendrites of
hippocampal pyramidal neurons. J Comput Neurosci Definition
7:5–15
Scanziani M, H€ausser M (2009) Electrophysiology in the Electrical impulses (action potentials (APs) or
age of light. Nature 461:930–939
Schaefer AT, Larkum ME, Sakmann B, Roth A (2003) spikes) which encode and transmit information
Coincidence detection in pyramidal neurons is tuned in the nervous system are initiated in the proximal
by their dendritic branching pattern. J Neurophysiol anatomical region of the axon termed axon initial
89:3143–3154 segment (AIS). The voltage threshold for spike
Spruston N (2008) Pyramidal neurons: dendritic
structure and synaptic integration. Nat Rev Neurosci initiation and the exact location and length of
9:206–221 the spike trigger zone (TZ) within AIS, as well
Spruston N, Stuart G, H€ausser M (2008) Dendritic inte- as the amplitude and waveform of the action
gration. In: Stuart G, Spruston N, Ha€ usser M (eds) potential in different neuronal classes, depend
Dendrites. Oxford University Press, New York,
pp 351–399 on the geometry and passive electrical properties
Stuart G, Spruston N, Sakmann B, H€ausser M (1997) of a neuron as well as on the type, spatial distri-
Action potential initiation and backpropagation in neu- bution, and density of a variety of voltage-
rons of the mammalian CNS. Trends Neurosci sensitive ionic channels.
20:125–131
Vetter P, Roth A, H€ausser M (2001) Propagation of action
potentials in dendrites depends on dendritic morphol-
ogy. J Neurophysiol 85:926–937 Detailed Description
Watanabe S, Hoffman DA, Migliore M, Johnston
D (2002) Dendritic K+ channels contribute to spike-
timing dependent long-term potentiation in hippocam- There is little disagreement over attributing
pal pyramidal neurons. Proc Natl Acad Sci U S A action potential initiation to a site in the axon
99:8366–8371 initial segment (AIS) under most circumstances.
Waters J, Schaefer A, Sakmann B (2005) The question of the exact location and length,
Backpropagating action potentials in neurones: mea-
surement, mechanisms and potential functions. Prog however, of the spike trigger zone (TZ) in the
Biophys Mol Biol 87:145–170 axon, as defined in functional terms, is less clear.
A 138 Action Potential Initiation
In most studies, the spike TZ was characterized The measuring technique and the analysis of
by a single parameter, the distance from the data used to determine the location and length
soma, implying a point of initiation. The length of the spike trigger zone are shown in Fig. 1.
of the initiation site, however, is fundamentally A typical experimental measurement used to
important because successful initiation and prop- determine the location and length of the spike TZ
agation of the action potential wave requires that is illustrated in Fig. 2. These two parameters are
a certain length of an axon is brought to the obtained directly from multisite optical recording
threshold for excitation (Rushton 1937). Besides of the membrane potential transients (Fig. 2a–c)
the fundamental importance of characterizing the either by investigating spike latencies at the
action potential (AP) initiation site, the spike TZ soma/axon hillock and more distal axonal record-
location and length have a recently discovered ing locations or by the inspection of the spatial
specific role in tuning neuronal computation distribution of membrane potential as a function
underlying a well-defined function in auditory of time. The soma–axon latency is plotted against
neurons which mediate sound source localization recording distance from the edge of the soma in
(Carr and Boudreau 1993; Kuba et al. 2006; Kuba Fig. 2d. An alternative way to derive the same
and Ohmori 2009). Moreover, subsequent studies information from the data is to analyze a time
of Kuba et al. (2010) and Grubb and Burrone sequence of color-coded frames showing the spa-
(2010) reported a novel finding that the structure tial maps of AP amplitude (Fig. 2e). The result of
of the spike TZ mediates an intrinsic plasticity of this analysis is a temporal series of individual
the axon and regulates the final stage of integra- frames separated by 10 ms; each showing the
tion of synaptic inputs. This places a great signif- spatial map of membrane potential at one point
icance on our ability to directly probe the location in time. In Fig. 2e, four frames from this series
and length of the spike TZ under different were selected to illustrate characteristic regions
conditions. The available information regarding along the axon that can be clearly identified dur-
TZ plasticity is based on structural data (Grubb ing AP initiation. The red region closest to the
and Burrone 2010; Kuba et al. 2010). Molecular soma was the first to cross the threshold value and
structure of the spike TZ, however, is indirectly reach 50 % amplitude (time point 0 ms) and was
correlated with function in a way that is not identified as the AP TZ. The more distal red
fully understood (Fleidervish et al. 2010; region appearing with a delay (45 ms time point)
Johnston 2010). Thus, the anatomical data is likely to be the first node of Ranvier,
require functional confirmation. The location corresponding to the issuance of an axon collat-
and length of the spike TZ have been difficult to eral, as indicated in the high-resolution image
measure directly using electrodes because extra- (Fig. 2a). The same data are shown in Fig. 2f as
cellular recordings cannot be interpreted with AP signals scaled to the same height and com-
sufficient accuracy and intracellular recordings pared on an expanded time scale. The two red
lack the necessary spatial resolution (e.g., traces show AP signals from the two red areas in
Meeks and Mennerick 2007). Membrane poten- Fig. 2e corresponding to the spike TZ and the first
tial imaging (Vm imaging) offers a unique node. The green dashed trace is the AP signal
advantage of high spatial resolution compared from the axon hillock. The green trace is the AP
with electrical recordings and has been used to signal from the first internodal region. The mean
directly measure the location of action potential length of the TZ determined from the type of data
initiation in invertebrate neurons (Zecevic 1996; shown in Fig. 2 was 16.5 1.1 mm with the
Antic et al. 2000) and mammalian axonal arbors mean center located at 28.9 1.0 mm from the
(Palmer and Stuart 2006; Palmer et al. 2010). edge of the soma.
This technique was recently improved and uti- Vm imaging can be used to analyze the spatial
lized to characterize functionally relevant param- pattern of AP propagation as revealed
eters of the spike TZ in layer 5 pyramidal neurons by monitoring transmembrane potential over lon-
of the cerebral cortex (Popovic et al. 2011). ger sections of individual myelinated axons.
SYNAPTIC STIMULATION
a
AIS Node A
Axon
20 µm
Soma electrical AIS optical Node optical

b
Temporal jitter
Aligned
Average of 9 trials
Bleach correction
Interpolation
One pass binomial
2 ms
smoothing
1 SOMA STIMULATION
c
Axon
Soma
20 µm
2 3 1
d
2
2 ms
Action Potential Initiation, Fig. 1 Signal processing. of traces: averaged signal. Fourth row of traces: bleach
(a) Synaptic stimulation: Upper image – high-resolution correction. Bottom traces: cubic spline interpolation with
confocal image of a stained neuron with axon in recording one pass of temporal smoothing. (c) Somatic stimulation:
position. Recording electrode attached to soma and stim- Upper image – high-resolution confocal image of another
ulating electrode next to basal dendrite shown schemati- neuron with axon in recording position. Lower
cally. Lower image – low spatial resolution fluorescence image – low spatial resolution fluorescence image of the
image of the axon obtained by CCD used for Vm imaging. axon obtained by CCD used for Vm imaging. Traces on
(b) Electrode recordings from soma and optical recordings left: AP transients from three locations: 1, electrode
from spike TZ (red) and from node of Ranvier (green). recording from soma; 2, optical recording from axon
Top traces: raw data from nine trials showing temporal hillock; and 3, optical recording from the first node of
jitter in AP initiation following synaptic activation. Sec- Ranvier. Bottom traces: Superimposed signal from the
ond row of traces: temporally aligned signals. Third row same three locations
A 140 Action Potential Initiation
Action Potential
Initiation, Fig. 2
Measurement of the spatial
distribution of membrane
potential as a function of
time along the proximal
axon during AP initiation.
(a) High resolution
confocal image of the axon
in recording position. (b)
Low spatial resolution
fluorescence image of the
axon obtained by CCD used
for Vm imaging. (c) AP
signals from 10 locations
indicated by yellow
rectangles, each 10 μm in
length. (d) Soma–axon
latency to 30% (grey) and
50% (black) AP amplitude
as a function of distance
from the cell body. The first
minimum identifies the
location and length of the
spike TZ. (e) Time
sequence of frames
showing spatial profile of
colour coded relative Vm
amplitude in the axon at
four characteristic time
points: 0 μs – AP initiation
at TZ; 45 μs and 80 μs –
invasion of the first node;
240 μs – peak
depolarization. (f)
Comparison of AP signals
from four characteristic
locations on an expanded
time scale. The measured
data points and cubic spline
interpolation curves are
shown. Red traces – TZ and
first node; green dashed
trace – axon hillock; green
trace – 1st internodal
region. Membrane potential
colour scale shown on left
Previously, this information was not available for approximately 300 mm clearly identified the posi-
any neuron. A representative experiment (well- tion of the spike TZ and putative nodes of
stained neuron characterized by long axons in one Ranvier; all characterized by localized reduction
plane of focus close to the surface of the slice) is in soma–axon latency typical for saltatory con-
illustrated in Fig. 3. The spatial plot of the duction (Fig. 3a). The spatial plot of AP latency
soma–axon latency along an axonal section of provides a functional readout for the position of
Action Potential
Initiation, Fig. 3 Spatial
pattern of AP initiation and
propagation in an A
individual axon. (a)
Soma–axon latency to 50 %
AP amplitude as a function
of distance from the soma.
(b) High-resolution image
of an axon in recording
position. (c) A color-coded
spatial distribution of
relative Vm amplitude in
the axon at one
characteristic point in time
showing correlation
between the positions of
functionally determined
nodes of Ranvier and
axonal branch points in
panel B
the nodes of Ranvier. Figure 3c shows a color- function in tuning neuronal computation under-
coded spatial map of the depolarizing AP wave at lying a well-defined function (sound localization)
one characteristic point in time which indicates in auditory neurons (Carr and Boudreau 1993;
the position of the spike TZ as well as nodes of Kuba et al. 2006; Kuba and Ohmori 2009).
Ranvier. These findings advocate that the size and position
The voltage-sensitive dye imaging approach of the spike TZ might be cell specific in other
allows determination of the location and length central neurons, depending on their function.
of the spike TZ in the AIS of pyramidal neurons, Additionally, new data (Kuba et al. 2010; Grubb
as defined in functional terms. In addition, it is and Burrone 2010) show that the structure (and,
possible to characterize the AP propagation pat- by extrapolation, the function) of the spike TZ
tern in the main axon and collaterals. It is plausi- participates in neuronal plasticity and might, in
ble to predict that this approach will facilitate fact, be one of the key factors controlling neuro-
the analysis of signal interactions underlying nal excitability and computation (Grubb and
input–output transformations carried out in neu- Burrone 2010; Kuba et al. 2010; Rasband 2010).
ronal processes of different classes of nerve cells.
The question of TZ dimensions is of fundamental
importance because the size of the initiation site References
is a critical functional parameter. Successful ini-
Antic S, Wuskell JP, Loew L, Zecevic D (2000) Func-
tiation and propagation of the action potential tional profile of the giant metacerebral neuron of Helix
wave requires that a certain length of an axon is aspersa: temporal and spatial dynamics of electrical
brought to threshold for excitation to generate an activity in situ. J Physiol 527:55–69
action current large enough to propagate. This Carr CE, Boudreau RE (1993) An axon with a myelinated
initial segment in the bird auditory system. Brain Res
follows from intuitive considerations, from the 628:330–334
classical theory (Rushton 1937), as well as from Colbert CM, Pan E (2002) Ion channel properties under-
more recent experimental studies (Colbert and lying axonal action potential initiation in pyramidal
Pan 2002; Meeks and Mennerick 2007). Addi- neurons. Nat Neurosci 5:533–538
Fleidervish IA, Lasser-Ross N, Gutnick MJ, Ross WN
tionally, Na+ channel clustering in the axon, (2010) Na+ imaging reveals little difference in action
which is critical in determining the spike TZ potential–evoked Na+ influx between axon and soma.
location and length, serves an important specific Nat Neurosci 13:852–860
A 142 Action Potential Model
Grubb MS, Burrone J (2010) Activity-dependent reloca-

tion of the axon initial segment fine-tunes neuronal Action Selection
excitability. Nature 465:1070–1074
Johnston D (2010) The Na+ channel conundrum:
axon structure versus function. Nat Neurosci ▶ Decision-Making, Motor Planning
13:784–785
Kuba H, Ohmori H (2009) Roles of axonal sodium chan-
nels in precise auditory time coding at nucleus
magnocellularis of the chick. J Physiol 587:87–100
Kuba H, Ishii TM, Ohmori H (2006) Axonal site of spike Adaptation
initiation enhances auditory coincidence detection.
Nature 444:1069–1072 ▶ Visual Aftereffects, Models of
Kuba H, Oichi Y, Ohmori H (2010) Presynaptic activity
regulates Na1 channel distribution at the axon initial
segment. Nature 465:1075–1078
Meeks JP, Mennerick S (2007) Action potential initiation
and propagation in CA3 pyramidal axons. Adaptation in Firing Rate Models
Palmer LM, Stuart GJ (2006) Site of action potential
initiation in layer V pyramidal neurons. J Neurosci ▶ Spike-Frequency Adaptation
26:1854–1863
Palmer LM, Clark BA, Grundemann J, Roth A, Stuart G,
Hausser M (2010) Initiation of simple and complex
spikes in cerebellar Purkinje cells. J Physiol
588:1709–1717 Adaptation in Sensory Cortices,
Popovic M, Foust AJ, McCormick DA, Zecevic D (2011) Models of
The spatio-temporal characteristics of action potential
initiation in layer 5 pyramidal neurons: a voltage imag-
ing study. J Physiol 589:4167–4187 Klaus Wimmer
Rasband MN (2010) The axon initial segment and the Institut d’Investigacions Biomèdiques August Pi
maintenance of neuronal polarity. Nat Rev Neurosci i Sunyer, Barcelona, Spain
11:552–562
Rushton WAH (1937) Initiation of the propagated distur-
bance. Proc R Soc Lond B Biol Sci 124:210–243
Zecevic D (1996) Multiple spike-initiation zones in single Synonyms
neurons revealed by voltage-sensitive dyes. Nature
381:322–325 Models of pattern adaptation; Models of sensory
adaptation
Further Reading
Bender KJ, Trussell LO (2012) The physiology of the
axon initial segment. Annu Rev Neurosci 35:249–265
Buffington SA, Rasband MN (2011) The axon initial seg- Definition
ment in nervous system disease and injury. Eur
J Neurosci 34:1609–1619 Models of adaptation in sensory cortices provide
Debanne D, Campanac E, Bialowas A, Carlier E, Alcaraz
G (2011) Axon physiology. Physiol Rev 91:555–602 a functional and/or mechanistic description of the
Kole MH, Stuart GJ (2012) Signal processing in the axon changes in neural responses and perception
initial segment. Neuron 73:235–247 caused by sensory stimuli observed in the recent
Kuba H (2012) Structural tuning and plasticity of the axon past. Sensory systems compute dynamic repre-
initial segment in auditory neurons. J Physiol
590:5571–5579 sentations of the environment: cortical neurons
typically adapt their “code” according to the
recently received sensory input. This continuous
recalibration is reflected in changes in neuronal
response properties and has been interpreted as an
Action Potential Model adjustment of the limited dynamical range to
compensate changes in the environment or
▶ Hodgkin-Huxley Model changes in the observer. Functional models of
Adaptation in Sensory Cortices, Models of 143 A
sensory adaptation have linked these findings to which – after prolonged viewing of oblique
optimal coding. Moreover, adaptation has also lines – vertical lines appear briefly as if they
been studied in biophysical and network models, were tilted in the opposite direction (see ▶ Visual A
with the goal of understanding the mechanisms Aftereffects, Models of).
that give rise to adaptation in biological cortical Here, we focus on (1) functional models of
circuits. the computational principles that may underlie
the adaptive encoding of sensory information
and (2) models of the neuronal and synaptic
Detailed Description mechanisms giving rise to adaptation in cortical
circuits.
Sensory adaptation refers to the changes in neuro-
nal responses and perception caused by Adaptation as Optimal Coding
a prolonged exposure to sensory stimuli. Adapta- Sensory systems must encode natural stimuli
tion is a rapid form of plasticity that has a reversible which change over many orders of magnitude
effect on neuronal selectivity: responses adapt with the limited dynamical range of neuronal
(on short time scales) and recover to their pre- firing rates. Adaptation has been proposed to
adapted state when the source of adaptation is serve for adjusting the sensory representation to
removed. It is found ubiquitously across different the current statistics of the ever changing envi-
sensory modalities (visual system (Kohn 2007; ronment (Fig. 1). This was formalized in the
Clifford et al. 2007); auditory system (King efficient coding hypothesis (grounded in informa-
et al. 2011); whisker system (Petersen et al. 2009)). tion theory), which states that sensory systems
In psychophysical experiments, it can be seek to provide an efficient representation of the
shown that adaptation alters perception: natural environment by maximizing their infor-
prolonged exposure to a stimulus typically leads mation transmission capacity (Barlow 1961;
to “repulsive” aftereffects, which means stimuli Laughlin 1981; Wark et al. 2007). Given the
similar to the adapting stimulus appear to be more statistics of the environment, this hypothesis
different from the adapting stimulus than they predicts the optimal input–output behavior of
actually are. An example is the tilt aftereffect, in a single neuron (Fig. 1) and how it should adapt
Adaptation in Sensory
Cortices, Models of,
Fig. 1 Adaptation to the
mean and the variance of
incoming sensory stimuli.
According to the efficient
coding hypothesis,
a change in the stimulus
distribution (top) should
yield an adaptive change of
a neuron’s transfer function
(bottom)
A 144 Adaptation in Sensory Cortices, Models of
to the mean and variance of the current stimulus individual neuronal tuning curves in the early
intensity distribution. visual system (Dragoi et al. 2000) (see Fig. 2).
Experimental evidence for efficient coding A potential mechanism underlying this adapta-
has been found across a wide range of sensory tion phenomenon, in addition to spike-frequency
modalities and species: contrast adaptation in the adaptation, is short-term synaptic depression.
visual system (Kohn 2007; Clifford et al. 2007), This form of rapid and reversible plasticity
fly visual system (Laughlin 1981; Fairhall leads to a decrease of synaptic efficacy lasting
et al. 2001), midbrain of guinea pigs (Dean from milliseconds to several seconds, due to
et al. 2005), inferior colliculus of cats (Kvale prolonged presynaptic activity. This timescale
and Schreiner 2004), songbird auditory forebrain is comparable to the timescale of orientation
(Nagel and Doupe 2006), and rat barrel cortex adaptation, but synaptic depression has mostly
(Maravall et al. 2007). been studied in vitro, and direct experimental
In a similar spirit, adaptation has also been evidence demonstrating that it is involved in
linked to Bayesian inference, tackling questions orientation adaptation is lacking. Synaptic
such as how to optimally combine sensory obser- depression can be described by a change in trans-
vations with prior knowledge about the stimulus mitter release probability in a phenomenological
distribution (Doya et al. 2007). Adaptation could model (Abbott et al. 1997; Tsodyks and
be interpreted as changing the prior distribution Markram 1997). Computational studies have
within the Bayesian theory of perception, or even investigated how synaptic depression, spike-
the likelihood model could be affected by frequency adaptation, and other mechanisms
adapting stimuli (Stocker and Simoncelli 2006). contribute to adaptive changes of visual cortical
How cortical circuits might implement Bayesian neurons. The different studies do however reach
inference is an active research topic. different conclusions about which particular
intrinsic or circuit mechanism is the most likely
Modeling Underlying Neural Mechanisms origin of orientation adaptation (Bednar and
A different class of models aims at understanding Miikkulainen 2000; Chelaru and Dragoi 2008;
the neuronal mechanisms that underlie adapta- Cortes et al. 2011). We argue that for a full
tion. In some cases the mechanisms have been understanding of adaptation mechanisms in
studied in detail experimentally, an example early visual areas, it is necessary to study how
is contrast adaptation in the visual system. adaptation interplays with the dynamics of local
Adaptation shifts the contrast response curve of circuits, which are dominated by strong, approx-
individual neurons toward higher contrast imately balanced, excitation and inhibition
levels, and this can be well explained by the (Stimberg et al. 2009).
hyperpolarization of individual neurons caused On the other hand, computational network
by the activation of intrinsic channels (Sanchez- models have elucidated the link between orienta-
Vives et al. 2000). This mechanism accounts for a tion adaptation and the psychophysically
decrease in a neuron’s firing rate due to sustained observed tilt aftereffect, which has long been
stimulation and is usually called spike-frequency a challenge (Teich and Qian 2003). Computa-
adaptation; for a phenomenological model, see tional models showed how adaptive changes in
Benda and Herz (2003). the location, the width, and the magnitude of
Often, however, the mechanisms underlying single neuron tuning curves give rise to charac-
adaptation are not well characterized experimen- teristic changes in population responses (Clifford
tally. Models can serve to integrate existing data et al. 2000; Compte and Wang 2006). In fact, the
and to create testable predictions in order to relationship between changes in experimentally
discern different potential mechanisms. An measured single neuron tuning curves and
example is orientation adaptation, in which the response of a population of neurons to
the prolonged exposure to a visual stimulus a single stimulus (which is thought to underlie
of a particular orientation yields changes of perception) can be counterintuitive (Fig. 2).
Adaptation in Sensory Cortices, Models of 145 A
Adaptation in Sensory Cortices, Models of, stimulus). Right: The repulsive tuning curve shifts corre-
Fig. 2 Example of the relationship between adaptive spond to a shift of the population activity (response of
changes in single neuron tuning curves and changes in a population of neurons to a particular stimulus) toward
the population activity, which is thought to underlie per- the adapting stimulus. The neuron label is determined by
ception. (a) Adaptation shifts response curves of individ- each neuron’s preferred orientation. (b) Same as (a), but
ual neurons away from the adapting stimulus. Shown are now adaptation suppresses responses of individual neu-
the tuning curves (responses of a neuron to visual stimuli rons with preferred direction close to the adapting stimu-
of different orientation) of three neurons with preferred lus. Right: This corresponds to a shift of the population
orientation 0 , 20 , and 40 (thin lines before adaptation, activity away from the adapting stimulus
thick lines after adaptation; triangle indicates the adapting
For example, when neuronal tuning curves shift adaptation (Seriès et al. 2009). The predicted
away from the adapting orientation (as typically perceptual effect also depends on the hypotheti-
observed experimentally), the corresponding cal link between population activity and percep-
population activity shifts toward the adapting tion, that is, on the “readout” strategy (such as
orientation (Fig. 2a). If, on the other hand, adap- peak activity, population vector, or maximum
tation only causes a suppression of neuronal likelihood decoding).
responses close to the adapting stimulus More recently, there has been increasing inter-
(Fig. 2b), the result is a shift of the population est in studying high-level adaptation effects,
activity away from the adapting stimulus including face adaptation (Webster 2011) and
(as observed in the tilt aftereffect). Considered adaptation to the perception of causal
together, the combination of adaptive changes interactions (Rolfs et al. 2013). Computational
observed at single neuron level in visual cortex modeling will be helpful in answering
is indeed consistent with the observed changes in interesting questions related to these phenomena,
perception (Jin et al. 2005). Note that in this such as to what degree high-level effects can be
framework it is commonly assumed that adapta- explained by adaptation to low-level features or
tion alters encoding in sensory cortical neurons, how they arise across the processing hierarchy
whereas downstream areas “reading out” the due to recurrent interactions between cortical
stimulus-related information are unaware of areas.
A 146 Adaptation in Sensory Cortices, Models of
Cross-References in visual cortex. J Neurophysiol 94:4038–4050.

doi:10.1152/jn.00571.2004
King AJ, Dahmen JC, Keating P et al (2011) Neural
▶ Perception, Bayesian Models of circuits underlying adaptation and learning in the per-
▶ Perceptual Decision-Making ception of auditory space. Neurosci Biobehav Rev
▶ Short-Term Plasticity, Biophysical Models 35:2129–2139. doi:10.1016/j.neubiorev.2011.03.008
▶ Spike-Frequency Adaptation Kohn A (2007) Visual adaptation: physiology, mecha-
nisms, and functional benefits. J Neurophysiol
▶ Visual Aftereffects, Models of 97:3155–3164. doi:10.1152/jn.00086.2007
Kvale MN, Schreiner CE (2004) Short-term adaptation of
auditory receptive fields to dynamic stimuli.
J Neurophysiol 91:604–612. doi:10.1152/
References jn.00484.2003
Laughlin S (1981) A simple coding procedure enhances
Abbott LF, Varela JA, Sen K, Nelson SB (1997) Synaptic a neuron’s information capacity. Z Naturforsch
depression and cortical gain control. Science C 36:910–912
275:220–224 Maravall M, Petersen RS, Fairhall AL et al (2007) Shifts
Barlow HB (1961) Possible principles underlying the in coding properties and maintenance of information
transformation of sensory messages. In: Rosenblith transmission during adaptation in barrel cortex. PLoS
WA (ed) Sensor communication. MIT Press, Cam- Biol 5:e19. doi:10.1371/journal.pbio.0050019
bridge, MA, pp 217–234 Nagel KI, Doupe AJ (2006) Temporal processing and
Bednar JA, Miikkulainen R (2000) Tilt aftereffects in adaptation in the songbird auditory forebrain. Neuron
a self-organizing model of the primary visual cortex. 51:845–859. doi:10.1016/j.neuron.2006.08.030
Neural Comput 12:1721–1740 Petersen RS, Panzeri S, Maravall M (2009) Neural coding
Benda J, Herz AVM (2003) A universal model for spike- and contextual influences in the whisker system.
frequency adaptation. Neural Comput 15:2523–2564. Biol Cybern 100:427–446. doi:10.1007/s00422-008-
doi:10.1162/089976603322385063 0290-5
Chelaru MI, Dragoi V (2008) Asymmetric synaptic Rolfs M, Dambacher M, Cavanagh P (2013) Visual adap-
depression in cortical networks. Cereb Cortex tation of the perception of causality. Curr Biol CB
18:771–788. doi:10.1093/cercor/bhm119 23:250–254. doi:10.1016/j.cub.2012.12.017
Clifford CW, Wenderoth P, Spehar B (2000) A functional Sanchez-Vives MV, Nowak LG, McCormick DA
angle on some after-effects in cortical vision. Proc Biol (2000) Membrane mechanisms underlying contrast
Sci 267:1705–1710. doi:10.1098/rspb.2000.1198 adaptation in cat area 17 in vivo. J Neurosci
Clifford CW, Webster MA, Stanley GB et al (2007) Visual 20:4267–4285
adaptation: neural, psychological and computational Seriès P, Stocker AA, Simoncelli EP (2009) Is the homun-
aspects. Vis Res 47:3125–3131. doi:10.1016/j. culus “aware” of sensory adaptation? Neural Comput
visres.2007.08.023 21:3271–3304. doi:10.1162/neco.2009.09-08-869
Compte A, Wang X-J (2006) Tuning curve shift by atten- Stimberg M, Wimmer K, Martin R et al (2009) The oper-
tion modulation in cortical neurons: a computational ating regime of local computations in primary visual
study of its mechanisms. Cereb Cortex 16:761–778. cortex. Cereb Cortex 19:2166–2180. doi:10.1093/
doi:10.1093/cercor/bhj021 cercor/bhn240
Cortes JM, Marinazzo D, Series P et al (2011) The effect Stocker A, Simoncelli E (2006) Sensory adaptation within
of neural adaptation on population coding accuracy. a Bayesian framework for perception. In: Weiss Y,
J Comput Neurosci. doi:10.1007/s10827-011-0358-4 Schölkopf B, Platt J (eds) Advances neural informa-
Dean I, Harper NS, McAlpine D (2005) Neural population tion process system, vol 18. MIT Press, Cambridge,
coding of sound level adapts to stimulus statistics. Nat MA, pp 1289–1296
Neurosci 8:1684–1689. doi:10.1038/nn1541 Teich AF, Qian N (2003) Learning and adaptation in
Doya K, Ishii S, Pouget A, Rao RPN (2007) Bayesian a recurrent model of V1 orientation selectivity.
brain: probabilistic approaches to neural coding. MIT J Neurophysiol 89:2086–2100. doi:10.1152/
Press, Cambridge, MA jn.00970.2002
Dragoi V, Sharma J, Sur M (2000) Adaptation-induced Tsodyks MV, Markram H (1997) The neural code
plasticity of orientation tuning in adult visual cortex. between neocortical pyramidal neurons depends on
Neuron 28:287–298 neurotransmitter release probability. Proc Natl Acad
Fairhall AL, Lewen GD, Bialek W, de Ruyter Van Sci U S A 94:719–723
Steveninck RR (2001) Efficiency and ambiguity in an Wark B, Lundstrom BN, Fairhall A (2007) Sensory adap-
adaptive neural code. Nature 412:787–792. tation. Curr Opin Neurobiol 17:423–429. doi:10.1016/
doi:10.1038/35090500 j.conb.2007.07.001
Jin DZ, Dragoi V, Sur M, Seung HS (2005) Tilt aftereffect Webster MA (2011) Adaptation and visual coding. J Vis
and adaptation-induced changes in orientation tuning 11:1–23. doi:10.1167/11.5.3
Algorithmic Reconstruction of Motoneuron Morphology 147 A
branching patterns, quantifying the parameters
Adaptive Control for a given population of experimentally
reconstructed neurons, and then feeding the data A
▶ Spinal Cord, Integrated (Non CPG) Models of into an algorithm which computationally gener-
ates populations of “virtual” neurons.
Afferent Feedback to Neural Detailed Description

Oscillators
Background
▶ Sensory Input to Central Pattern Generators
Digitization of neuronal morphology is important
for computational neuroscience because neuronal
morphology affects synaptic integration and fir-
ing behavior within individual neurons as well as
Afferent Input to Rhythm Generating
determining potential connectivity with other
Networks
neurons (Ascoli 2002). However, experimental
reconstruction techniques are still largely manual
▶ Sensory Input to Central Pattern Generators
or, at best, semiautomated, requiring time and
skill to accurately capture neuronal morphology.
As computational models of the nervous system
grow in scale with computational power, there is
Algorithmic Generation of
an increasing need for large numbers of digitized
Motoneuron Morphology
(“virtual”) morphologies. Algorithmic genera-
tion of virtual morphologies has the potential to
▶ Algorithmic Reconstruction of Motoneuron
fulfill this need.
Morphology
Motoneurons, as the nervous system’s
“final common pathway” of motor control
(Sherrington 1906), have long been studied.
Algorithmic Reconstruction of These neurons exhibit extensive dendritic arbor-
Motoneuron Morphology izations that may stretch over several millimeters
in the spinal cord (Cullheim et al. 1987a, b) which
Joseph Graham makes their reconstruction particularly challeng-
Blue Brain Project, École Polytechnique ing. As such, the earliest forays into algorithmic
Fédérale de Lausanne, Lausanne, Switzerland generation of neuronal morphology occurred
in motoneurons. In particular, one set of
reconstructed motoneurons from Cullheim
Synonyms et al. (1987a, b) has been the basis of algorithmic
generation research by multiple groups (these
Algorithmic generation of motoneuron morphol- morphologies are available for download at
ogy; Computational synthesis of motoneuron neuromorpho.org).
morphology; Computer generation of motoneu- The general process of algorithmic generation
ron morphology begins by parameterizing the neurite branching
patterns. These parameters are statistical distri-
butions which are then quantified from experi-
Definition mental reconstructions. Generation then proceeds
from the soma outwards, first generating primary
Algorithmic reconstruction of neuronal morphol- neurites which go on to branch or terminate
ogy is the process of parameterizing neurite according to the chosen algorithm. This process
A 148 Algorithmic Reconstruction of Motoneuron Morphology
continues until all branches have terminated. The (how much a neurite’s direction changes as
algorithmic reconstructions are then compared to a branch extends outwards) and some measure
the experimental reconstructions, whereupon of “tropism” (a force inducing a directionality in
persistent differences are explored to improve neurite behavior, such as an apical dendrite
the parameterization and algorithm. As this pro- extending towards the pia).
cess is repeated, the algorithmic reconstructions Slightly different parameter sets have since
become more and more similar to the experimen- been proposed and explored based on Hillman’s
tal. The ultimate goal of this line of research is the original fundamental parameters.
capability of generating populations of unique
morphologies which are statistically indistin- Algorithms
guishable from the experimental population Burke et al. (1992) were the first to realize that
upon which they are based and which capture parameterization of neuronal morphology could
the natural variability of such a population. be used to algorithmically generate virtual
neurite trees. Using experimentally reconstructed
Parameterization motoneurons (Cullheim et al. 1987a, b), their
Hillman (1979) was the first to propose that a set strategy was to “(1) devise a model system that
of “fundamental” parameters could be used to can simulate dendritic trees, (2) derive the required
completely and parsimoniously describe neuro- model parameters directly from measurements of
nal morphology. He recognized that there are two real dendrites, and (3) refine the parameter deriva-
separable aspects of neuronal morphology that tions or basic model assumptions, based on the
must be parameterized: the branching patterns degree of congruence between real and simulated
and the space-filling behavior. It is important to dendrites” (Burke et al. 1992). Using this method-
remember that these parameters are not scalar ology, Burke et al. found correlations between their
values, but rather statistical distributions, and parameters and local diameter and path length from
also that these parameters may be intercorrelated the soma. Marks and Burke later extended this
or correlated with other local properties of the work (2007a, b) to include space-filling parameters
arborization. and the generation of entire neurons.
To capture the branching patterns, Hillman Ascoli and Krichmar (2000) were the first to
proposed that every neurite tree begins with apply the methodology in order to create entire
a “stem diameter” and grows for a certain “seg- neurons, by adding somatic and neurite trunk
ment length” while changing diameter by “seg- parameters to a software package called
ment taper.” If the final diameter of a branch is L-Neuron. This software package could replicate
larger than the “terminal diameter,” the branch Burke et al.’s (1992) algorithm, as well as several
bifurcates into two daughter branches, whose algorithms derived more closely from Hillman’s
diameters are related to that of their parent by (1979) parameterization. Ascoli et al. (2001) then
the “branch power” and whose relative diameters went on to explore these algorithms using exper-
are determined by the “daughter ratio.” Hillman imental reconstructions of several different types
proposed that quantifying these parameters was of neurons, including the motoneurons from
sufficient to completely describe neurite Cullheim et al. (1987a, b). Donohue and Ascoli
branching patterns. (2008) later extended this work to allow parame-
To capture the space-filling patterns, Hillman ters to be correlated with local neurite properties
proposed that one must measure the initial direc- (branch order, diameter, and path length from the
tion of neurite trees and the branching angles (the soma). In an extensive analysis, they explored
angles that daughter branches make with regard which local properties are most important for
to the parent branch). Since that time, it has been generating realistic morphologies.
recognized that to capture space filling, more Many groups have been exploring algorithmic
parameters need to be measured. Proposed addi- generation of neuronal morphology, but rela-
tions include some measure of “meander” tively few have explicitly explored motoneurons.
a1I 149 A
Torben-Nielsen et al. (2008) developed a very Burke RE, Marks WB, Ulfhake B (1992) A parsimonious
different sort of algorithm which they used to description of motoneuron dendritic morphology using
computer simulation. J Neurosci 12(6):2403–2416
explore motoneuron morphology. In this algo- Cullheim S, Fleshman JW, Glenn LL, Burke RE (1987a) A
rithm, parameters are not quantified in advance, Membrane area and dendritic structure in type-
but rather the entire parameter space is explored identified triceps surae alpha motoneurons. J Comp
in an evolutionary algorithm which can converge Neurol 255:68–81
Cullheim S, Fleshman JW, Glenn LL, Burke RE (1987b)
on parameter values which produce realistic Three-dimensional architecture of dendritic trees in type-
morphologies. identified alpha motoneurons. J Comp Neurol 255:82–96
Donohue DE, Ascoli GA (2008) A comparative computer
Summary simulation of dendritic morphology. PLoS Comput
Biol 4(5):e1000089
Algorithmic generation of motoneuron morphol- Hillman DE (1979) Neuronal shape parameters and sub-
ogy is a subset of general algorithmic generation structures as a basis of neuronal form. In: Schmitt FO,
of neuronal morphology. The methodology offers Worden FG (eds) The neurosciences: fourth study
the promise of a deeper understanding of neuro- program. MIT Press, Cambridge, MA, pp 477–498
Marks WB, Burke RE (2007a) Simulation of motoneuron
nal morphology and may eventually fulfill the morphology in three dimensions. I. Building individ-
need of computational neuroscientists for large ual dendritic trees. J Comp Neurol 503:685–700
numbers of realistic morphologies for use in sim- Marks WB, Burke RE (2007b) Simulation of motoneuron
ulations. While no algorithm has yet been capable morphology in three dimensions. II. Building com-
plete neurons. J Comp Neurol 503:701–716
of producing morphologies which are statistically Sherrington CS (1906) Integrative actions of the nervous
indistinguishable from experimental reconstruc- system. Yale University Press, New Haven
tions across all morphometrics, much progress Torben-Nielsen B, Tuyls K, Postma E (2008) EvOL-
has been made. The study of motoneurons is Neuron: neuronal morphology generation.
Neurocomputing 71:963–972
especially useful for algorithm development
because several different groups have explored
the same data set (Cullheim et al. 1987a, b), thus
providing a basis for further improvement. Algorithms for Generating Neuronal
Morphologies
Cross-References ▶ Synthetic Neuronal Circuits/Networks

▶ Synthetic Neuronal Morphology
▶ Compartmental Models of Spinal
Motoneurons
▶ Morphologically Detailed Motoneuron Models
▶ Reconstruction, Techniques and Validation a1G
▶ Synthetic Neuronal Circuits/Networks
▶ Synthetic Neuronal Morphology ▶ Low-Voltage-Activated Calcium Channels
References
a1H
Ascoli GA (ed) (2002) Computational neuroanatomy:
principles and methods. Humana Press, Totowa
Ascoli GA, Krichmar JL (2000) L-neuron: a modeling tool
for the efficient generation and parsimonious descrip-
tion of dendritic morphology. Neurocomputing
32–33:1003–1011
Ascoli GA, Krichmar JL, Scorcioni R, Nasuto SJ, Senft SL
(2001) Computer generation and quantitative morpho-
a1I
metric analysis of virtual neurons. Anat Embryol
204:283–301 ▶ Low-Voltage-Activated Calcium Channels
A 150 Amari Model
@u
¼ u þ w f ðuÞ: (3)
Amari Model @t
Roland Potthast The Amari equation provides a continuous ana-

Department of Mathematics, University of logue or continuous description of neural net-
Reading, Reading, UK works, which has become widely used in the
engineering community.
Definition
Historic Background and Applications
The Amari neural field model (cf. (Amari 1975,
1977)) provides a simple field-theoretic approach The earliest field models for describing and study-
to the dynamics of neural activity in the brain. ing neural activity dynamics go back to Beurle
The model uses excitations and inhibitions over (1956), investigating the proportion of active
some distance as an effective model of mixed neurons in randomly connected networks,
inhibitory and excitatory neurons with typical followed by work of Griffith (1963, 1965). The
cortical connectivities. The model is a scalar basis of modern field dynamical models has been
dynamical equation for the voltage or activity the work of Cowan, Nunez, and Amari in the
u(x, t) of the form 1970s (see Wilson and Cowan 1972, 1973;
Nunez 1974; Amari 1977). Cowan proposed an
@u activity-based model with two distinct
ðx, tÞ ¼ uðx, tÞ
@t ð
populations of excitatory and inhibitory subpop-
ulations. Amari suggested a more condensed sca-
þ wðx, yÞ f ðuðy, tÞÞdy, xB, t 0 ,
B
lar model with a Mexican hat-type connectivity
(1) function, where excitation and inhibition are
reflected by the changing sign of the connectivity
where initial conditions u(x, 0) = u0(x), x B are kernel w(x y).
given. Here, B is our brain, i.e., some domain The Amari neural field model has been
where the neural activity takes place; f is the applied, e.g., to autonomous robotic behavior
local activation function or firing rate function; (Erlhagen and Bicho 2006), embodied cognition
and w is the connectivity function which models (Schöner and Dineva 2007), dynamic causal
the strength of the connectivity or signal propa- modeling (Daunizeau et al. 2009), and language
gation from y B to the point x. processing (beim Graben et al. 2008; beim Gra-
A common choice for the activation function ben and Potthast 2012); for further details, we
has sigmoidal shape refer to the recent tutorial by Coombes
et al. (2013).
1
f ðsÞ :¼ , s ℝ; (2)
1 þ ebðshÞ
Theory
which is monotonously growing from
f(1) = 0 to saturation f(1) = 1 for large s, The problem (4) is an integrodifferential equa-
where h is the threshold and b is a steepness tion, establishing an evolutionary dynamical sys-
parameter. Often, the kernel w is chosen to be tem with initial condition u(x, 0) = u0(x), x B,
homogeneous, i.e., w(x, y) = w(x y). In this at time t = 0. For a Lipschitz continuous activa-
case, the integral becomes a convolution integral, tion function f, existence and uniqueness of
such that the Amari equation can be written in the a solution u of Eq. 4 for all times and in one or
form several dimensions are obtained under quite
Amari Model 151 A
general conditions based on elementary argu- heterogeneity (primarily using simulations) can
ments and the fixed-point theorem (compare be found in Brackley and Turner (2007),
Potthast and beim Graben 2010). Over time, Bressloff (2001), Schmidt et al. (2009), and A
there has been significant activity to study the Coombes et al. (2012) and functional analytic
existence and uniqueness of bumps and waves in results in Faugeras et al. (2008), and Potthast
one spatial dimension when a particular homoge- and beim Graben (2010). The inverse problems
neous kernel w(x y) is given (see Kishimoto perspective for either homogeneous or
and Amari 1979; Ermentrout and McLeod 1993) nonhomogeneous kernels has been investigated
where smooth sigmoidal firing rates are used (see by Potthast and beim Graben (2009) and beim
also Coombes and Schmidt 2010; Oleynik Graben and Potthast (2009). More recently, sto-
et al. 2013; Coombes and Owen 2004 for more chastic neural field equations have become very
general but still rather restricted classes of popular; compare the review by Bressloff (2012).
functions f ).
Delay Neural Field Equation and Advantages of the Approach

Homogeneous Kernels
Geometric singular perturbation analysis investi- The Amari neural field model provides a very
gates the stability of static or transient solutions, concise scalar equation to model neural activity
and numerical bifurcation techniques investigate and dynamics. In contrast to microscopic models,
the existence and properties of bifurcation points it summarizes the activity of neurons into the
in state or parameter spaces (compare Pinto and activity function u(x, t), which can be used to
Ermentrout 2001a, b; Laing and Troy 2003b). For reduce the computational complexity signifi-
studies in two or more dimensions, we refer to cantly. The field-theoretic approach opens the
Taylor (1999), Laing and Troy (2003a), Folias dynamic system to mathematical analysis and
and Bressloff (2004), Laing (2005), Owen beyond the range of discrete network models.
et al. (2007), Kilpatrick and Bressloff (2010a),
and Coombes et al. (2012). Often, the Eq. 4 is
complemented by a delay term Limitations
@u The strength of the Amari model is its simplicity,

ðx, tÞ ¼ uðx, tÞ
@t which is at the same time its strongest limitation. It
ð
D ðx, yÞ does not take into account the complex chemical
þ wðx, yÞf u y, t dy,
B v and physiological processes which take place in
(4) addition to electrical dynamics in neural tissue, nor
is it capable to include the different temporal scales
x B, t 0, where D(x, y) is the length of the on which these processes work and propagate.
fiber between x and y and n is the finite propaga-
tion speed of signals. The above analysis for
neural field equations has been extended to Cross-References
delay equations (see Nunez 1974; Jirsa and
Haken 1997; Coombes et al. 2003; Hutt 2004; ▶ Bifurcations, Neural Population Models and
Venkov et al. 2007; Grindrod and Pinotsis ▶ Chaos, Neural Population Models and
2010), including dendritic processing (Bressloff ▶ Inverse Problems in Neural Population Models
and Coombes 1997) and synaptic depression ▶ Neural Field Model, Continuum
(Kilpatrick and Bressloff 2010b). ▶ Neural Population Model
Most of the above work uses homogeneous ▶ Pattern Formation in Neural Population
kernels w(x y). More recent work that tackles Models
A 152 Amari Model
▶ Phase Transitions, Neural Population Folias SE, Bressloff PC (2004) Breathing pulses in an
Models and excitatory neural network. SIAM J Appl Dyn Syst
3:378–407
▶ Stochastic Neural Field Theory Griffith JS (1963) A field theory of neural nets: I: deriva-
▶ Wilson-Cowan Model tion of field equations. Bull Math Biophys 25:111–120
Griffith JS (1965) A field theory of neural nets: II: proper-
ties of field equations. Bull Math Biophys 27:187–195
References Grindrod P, Pinotsis D (2010) On the spectra of certain
integro-differential-delay problems with applications
Amari S (1975) Homogeneous nets of neuron-like ele- in neurodynamics. Phys D Nonlinear Phenom
ments. Biol Cybern 17:211–220 240(1):13–20. doi:10.1016/j.physd.2010.08.002,
Amari S (1977) Dynamics of pattesrn formation in lateral- ISSN 0167–2789
inhibition type neural fields. Biol Cybern 27:77–87 Hutt A (2004) Effects of nonlocal feedback on traveling
beim Graben P, Potthast R (2009) Inverse problems in fronts in neural fields subject to transmission delay.
dynamic cognitive modeling. Chaos 19(1):015103 Phys Rev E 60(1–4):052902
beim Graben P, Potthast R (2012) A dynamic field account Jirsa VK, Haken H (1997) A derivation of a macroscopic
to language-related brain potentials. In: Rabinovich M, field theory of the brain from the quasi-microscopic
Friston K, Varona P (eds) Principles of brain dynamics: neural dynamics. Phys D 99:503–526
global state interactions. MIT Press, Cambridge, MA Kilpatrick ZP, Bressloff PC (2010a) Effects of synaptic
beim Graben P, Pinotsis D, Saddy D, Potthast R (2008) depression and adaptation on spatiotemporal dynamics
Language processing with dynamic fields. Cogn of an excitatory neuronal network. Phys
Neurodyn 2(2):79–88 D 239:547–560
Beurle RL (1956) Properties of a mass of cells capable of Kilpatrick ZP, Bressloff PC (2010b) Spatially structured
regenerating pulses. Phil Trans R Soc Lond oscillations in a two-dimensional excitatory neuronal
B 240:55–94 network with synaptic depression. J Comput Neurosci
Brackley CA, Turner MS (2007) Random fluctuations of 28:193–209
the firing rate function in a continuum neural field Kishimoto K, Amari S (1979) Existence and stability of
model. Phys Rev E 75:041913 local excitations in homogeneous neural fields. J Math
Bressloff PC (2012) Spatiotemporal dynamics of contin- Biol 7:303–318
uum neural fields. J Phys A 45:033,001 Laing CR (2005) Spiral waves in nonlocal equations.
Bressloff PC, Coombes S (1997) Physics of the extended SIAM J Appl Dyn Syst 4:588–606
neuron. Int J Mod Phys B 11:2343–2392 Laing CR, Troy WC (2003a) PDE methods for nonlocal
Coombes S, Owen MR (2004) Evans functions for integral models. SIAM J Appl Dyn Syst 2:487–516
neural field equations with Heaviside firing rate func- Laing CR, Troy WC (2003b) Two bump solutions of
tion. SIAM J Appl Dyn Syst 34:574–600 Amari-type models of working memory. Phys
Coombes S, Schmidt H (2010) Neural fields with sigmoi- D 178:190–218
dal firing rates: approximate solutions. Discret Contin Nunez PL (1974) The brain wave equation: a model for the
Dyn Syst Ser A 28:1369–1379 EEG. Math Biosci 21:279–297
Coombes S, Lord GJ, Owen MR (2003) Waves and bumps Oleynik A, Posnov A, Wyller J (2013) On the properties of
in neuronal networks with axo-dendritic synaptic nonlinear nonlocal operators arising in neural field
interactions. Phys D 178:219–241 models. Journal of Mathematical Analysis and Appli-
Coombes S, Schmidt H, Bojak I (2012) Interface dynam- cations 398:335–351
ics in planar neural field models. J Math Neurosci 2:9 Owen MR, Laing CR, Coombes S (2007) Bumps and rings
Coombes S, beim Graben P, Potthast R et al (2013) Tuto- in a two-dimensional neural field: splitting and rota-
rial on neural field theory. In: Wright J, Potthast R, tional instabilities. New J Phys 9:378
Coombes S, beim Graben P (eds) Neural fields. Theory Pinto DJ, Ermentrout GB (2001a) Spatially structured
and applications. Springer, Berlin activity in synaptically coupled neuronal networks:
Daunizeau J, Kiebel SJ, Friston KJ (2009) Dynamic causal I. Travelling fronts and pulses. SIAM J Appl Math
modelling of distributed electromagnetic responses. 62:206–225
Neuroimage 47:590–601 Pinto DJ, Ermentrout GB (2001b) Spatially structured
Erlhagen W, Bicho E (2006) The dynamic neural field activity in synaptically coupled neuronal networks:
approach to cognitive robotics. J Neural Eng 3: II. Lateral inhibition and standing pulses. SIAM
R36–R54 J Appl Math 62:226–243
Ermentrout GB, McLeod JB (1993) Existence and unique- Potthast R, beim Graben P (2009) Inverse problems in
ness of travelling waves for a neural network. Proc neural field theory. SIAM J Appl Dyn Syst
Roy Soc Edinb 123A:461–478 8(4):1405–1433
Faugeras O, Grimbert F, Slotine JJ (2008) Absolute sta- Potthast R, beim Graben P (2010) Existence and proper-
bility and complete synchronization in a class of neural ties of solutions for neural field equations. Math
fields models. SIAM J Appl Math 69:205–250 Methods Appl Sci 33(8):935–949
AMPA Glutamate Receptor (AMPA Receptor), Conductance Models 153 A
Schmidt H, Hutt A, Schimansky-Geier L (2009) Wave Rabinovich M, Friston K, Varona P (eds) (2012) Principles
fronts in inhomogeneous neural field models. Phys of brain dynamics: global state interactions. MIT
D 238:1101–1112 Press, Cambridge, MA
Schöner G, Dineva E (2007) Dynamic instabilities as Tass P (1995) Cortical pattern formation during visual A
mechanisms for emergence. Dev Sci 10:69–74 hallucinations. J Biol Phys 21:177–210
Taylor JG (1999) Neural ‘bubble’ dynamics in two dimen-
sions: foundations. Biol Cybern 80:393–409, Structure
Venkov NA, Coombes S, Matthews PC (2007) Dynamic
instabilities in scalar neural field equations with space-
dependent delays. Phys D 232:1–15
Wilson HR, Cowan JD (1972) Excitatory and inhibitory Ambiguous Stimuli
interactions in localized populations of model neurons.
Biophys J 12:1–24
Wilson HR, Cowan JD (1973) A mathematical theory of ▶ Multistability in Perception Dynamics
the functional dynamics of cortical and thalamic ner-
vous tissue. Kybernetik 13:55–80
Further Reading
Ben-Yishai R, Bar-Or L, Sompolinsky H (1995) Theory of AMPA Glutamate Receptor (AMPA
orientation tuning in visual cortex. Proc Natl Acad Sci Receptor), Conductance Models
U S A 92:3844–3848
Berger H (1929) Über das Elektroenkephalogramm des
Menschen. Archiv Psychiatr 87:527–570 Patrick D. Roberts
Bressloff PC (2001) Traveling fronts and wave propaga- Department of Biomedical Engineering, Oregon
tion failure in an in-homogeneous neural network. Health & Science University, Portland, OR, USA
Phys D 155:83–100
Bressloff PC, Cowan JD, Golubitsky M, Thomas PJ,
Wiener M (2001) Geometric visual hallucinations,
Euclidean symmetry and the functional architecture Definition
of striate cortex. Phil Trans R Soc Lond B 40:299–330
Ermentrout GB, Cowan JD (1979) A mathematical theory
A glutamate receptor that is permeable to sodium
of visual hallucination patterns. Biol Cybern
34:137–150 ions and carries an excitatory synaptic current
Geise MA (1999) Neural field theory for motion percep- following the binding of glutamate. AMPA
tion. Kluwer, Boston receptors are regulated to control the maximum
Jirsa VKV, Jantzen KJ, Fuchs A, Kelso JAS (2001) Infor-
synaptic current by processes of synaptic
mation processing in medical imaging, chap. Neural
field dynamics on the folded three-dimensional corti- plasticity and they co-localize with NMDA
cal sheet and its forward EEG and MEG. Springer, receptors.
Berlin, pp 286–299
Jirsa VK, Jantzen KJ, Fuchs A, Kelso JAS (2002) Spatio-
temporal forward solution of the EEG and MEG using
network modeling. IEEE Trans Med Imaging Detailed Description
21(5):493–504
Kilpatrick ZP, Bressloff PC (2010) Binocular rivalry in AMPA receptors are named for the
a competitive neural network with synaptic depres-
sion. SIAM J Appl Dyn Syst 9:1303–1347
selective agonist (a-amino-3-hydroxy-5-methyl-
Laing CR, Troy WC, Gutkin B, Ermentrout GB 4-isoxazolepropionic acid) that does not bind
(2002) Multiple bumps in a neuronal model of working well to other glutamate receptors. The receptor
memory. SIAM J Appl Math 63:62–97 is permeable to cations and can allow Na+, K+,
Liley DTJ, Cadusch PJ, Dafilis MP (2002) A spatially
and Ca2+ to cross the membrane and has an equi-
continuous mean field theory of electrocortical activ-
ity. Network 13:67–113 librium potential EAMPA = 0 mV. The perme-
Liley DTJ, Foster BL, Bojak I (2011) Sleep and anesthe- ability to Ca2+ is small and is not considered
sia, chap. A mesoscopic modelling approach to anaes- important for initiating signaling cascades.
thetic action on brain electrical activity. Springer, New
AMPA receptors are composed of four types of
York, pp 139–166
Nunez PL (1995) Neocortical dynamics and human EEG subunits, and the combination of subunits deter-
rhythms. Oxford University Press, New York mines the kinetics and permeability to cations.
A 154 AMPA Glutamate Receptor (AMPA Receptor), Conductance Models
Kinetics of AMPA receptors population of independent asynchronous inputs

Mathematical representations of AMPA currents to deliver a constant depolarization to a neuron.
include the time-dependent synaptic conduc- A single AMPA synapse can deliver a fast syn-
tance, gAMPA(t), in the current–voltage equation: aptic current that is filtered by cable properties of
the dendrite where the postsynaptic terminal is
I AMPA ðtÞ ¼ gAMPA ðtÞðV ðtÞ EAMPA Þ (1) located (Rall 1967). But there are temporal limi-
tations to how much filtering is possible to deliver
The time-dependent synaptic conductance the excitatory effect to the spike-generating zone
represents the opening and closing kinetics of of the postsynaptic neuron. Multiple AMPA cur-
the receptor channels and may be based on rents located on the postsynaptic neurons can
a double exponential function (or a similarly overcome this constraint if the rate of incoming
shaped function). spikes is high enough and well distributed in time
to overlap. Thus, the synaptic dynamics of
gs ðtÞ ¼ gs et=t1 et=t2 (2) a single synaptic terminal may not have a great
influence on the postsynaptic activity.
The strength of the synaptic current is not the
where t1 is the onset time constant (t2 < 0.1 ms)
only determinant of the efficacy of the synaptic
and t2 is the decay time constant (t2 < 3 ms)
input to generate a spike in the postsynaptic neu-
(Hestrin et al. 1990; Trussell et al. 1993; Jonas
ron. Spikes are triggered by changes in mem-
et al. 1993). AMPA currents have a faster decay
brane potential, and the filtering properties of
than NMDA currents (Jonas and Spruston 1994)
the neuron are critical in how the dynamics of
and typically have a substantially larger peak
the individual synapses are transformed into
current.
changes in postsynaptic activity.
Because AMPA currents are usually the domi-
nant excitatory synaptic currents, simple represen-
AMPA synaptic dynamics
tations of these currents may be the only synaptic
In addition to the channel kinetics, AMPA cur-
currents included in the practice of modeling neu-
rents exhibit short-term plasticity that can result
ral circuits. In models with low time resolution,
from both pre- and postsynaptic mechanisms
AMPA currents may be represented as increased
(Zucker and Regehr 2002; Blitz et al. 2004).
currents in a single time bin without short-term
Presynaptic mechanisms affect the release of glu-
temporal details. However, these simple models
tamate and can influence the peak of the current on
may contain long-term plasticity that changes the
subsequent presynaptic spikes. Postsynaptic mech-
circuit dynamics over time by changing the
anisms affect the response of AMPA receptors to
strength of connections between neurons.
the concentration of glutamate in the synaptic cleft.
AMPA currents are also involved in long-term
Computational functions of AMPA receptors plasticity and can be the main component of
The excitatory characteristics of AMPA currents changes in synaptic strength. AMPA currents do
can be the main driver of activity in a network of not play a direct role in inducing long-term plas-
neurons and for communication across long dis- ticity but reflect the changes in presynaptic
tances such as from the periphery to the central release and their own response to glutamate
nervous system. Due to the short time constants caused by other mechanisms.
of AMPA kinetics, signal transmission can carry
high temporal precision. An example of this pre-
cision is found in the auditory pathway where the Cross-References
interaural time difference can be discerned to
extreme precision (Konishi 1990). ▶ Kinetic Models of Postsynaptic Currents
Another consequence of the short duration of ▶ N-Methyl-D-Aspartate (NMDA) Receptors,
AMPA currents is that they require a large Conductance Models
Anatomy and Physiology of the Mammalian Auditory System 155 A
References AN Auditory nerve
AVCN Anteroventral cochlear nucleus
Blitz DM, Foster KA, Regehr WG (2004) Short-term CNC Cochlear nucleus complex
synaptic plasticity: a comparison of two synapses. A
Nat Rev Neurosci 5(8):630–640
CNIC Central nucleus of the inferior
Hestrin S, Sah P, Nicoll RA (1990) Mechanisms generat- colliculus
ing the time course of dual component excitatory syn- DAS Dorsal acoustic stria
aptic currents recorded in hippocampal slices. Neuron DCIC Dorsal nucleus of the inferior
5(3):247–253
Jonas P, Spruston N (1994) Mechanisms shaping
colliculus
glutamate-mediated excitatory postsynaptic currents DCN Dorsal cochlear nucleus
in the CNS. Curr Opin Neurobiol 4(3):366–72 DLL Lateral lemniscus dorsal nucleus
Jonas P, Major G, Sakmann B (1993) Quantal components of GABA g-Aminobutyric acid
unitary EPSCs at the mossy fibre synapse on CA3 pyra-
midal cells of rat hippocampus. J Physiol 472(1):615–663
IAS Intermediate acoustic stria
Konishi M (1990) The neural algorithm for sound locali- IC Inferior colliculus
zation in the owl. Harvey Lect 86:47 IHC Inner hair cell
Rall W (1967) Distinguishing theoretical synaptic poten- LCIC Lateral cortex of the inferior
tials computed for different soma-dendritic distribu-
tions of synaptic input. J Neurophysiol 30(5):1138
colliculus
Trussell LO, Zhang S, Ramant IM (1993) Desensitization LOC Lateral olivocochlear neurons/system
of AMPA receptors upon multiquantal neurotransmit- LSO Lateral superior olive
ter release. Neuron 10(6):1185–1196 MGB Medial geniculate body
Zucker RS, Regehr WG (2002) Short-term synaptic plas-
ticity. Annu Rev Physiol 64(1):355–405
MGD Dorsal division of the medial
geniculate body
Further Reading MGM Medial division of the medial
Dayan P, Abbott LF, Abbott L (2001) Theoretical neuro- geniculate body
science: computational and mathematical modeling of MGV Ventral division of the medial
neural systems. Taylor & Francis, Cambridge, MA geniculate body
Koch C (2004) Biophysics of computation: information
processing in single neurons. Oxford university press,
MNTB Medial nucleus of the trapezoid body
New York MOC Medial olivocochlear neurons/system
MSO Medial superior olive
NLL Nuclei of the lateral lemniscus
NMDA N-Methyl-D-aspartate
Anatomy and Physiology of the OHC Outer hair cell
Mammalian Auditory System PO Periolivary nuclei
SOC Superior olivary complex
Manuel S. Malmierca SPO Superior paraolivary nucleus
Department of Cellular Biology and Pathology, VAS Ventral acoustic stria
Faculty of Medicine, University of Salamanca, VCN Ventral cochlear nucleus
Salamanca, Spain VLL Ventral nucleus of the lateral
Auditory Neuroscience Laboratory, Institute for lemniscus
Neuroscience of Castilla y Léon, Salamanca, VNTB Ventral nucleus of the trapezoid body
Spain
An auditory system is found in all classes of

List of Abbreviations vertebrates, including fish, amphibians, reptiles
and birds, and mammals. Although there are
A1 Primary auditory cortex important similarities across classes, the system
AC Auditory cortex has evolved differently in the different groups.
AMPA a-amino-3-hydroxy-5-methyl-4- Even within the class of mammals, there are
isoxazolepropionic acid notable specializations, especially in
A 156 Anatomy and Physiology of the Mammalian Auditory System
Anatomy and Physiology of the Mammalian Audi- DCN. The efferent MOC innervate the OHCs, and the
tory System, Fig. 1 Afferent and efferent innervation of efferent LOC innervate IHCs (cf. Fig. 19) (Modified
the cochlear epithelium. Several type I afferent fibers after Brown et al. 1988). Abbreviations in the figure: DC
converge onto single IHCs, while a single type II afferent Dorsal cochlear nucleus, LTz Lateral nucleus of the trap-
fibers innervate several OHCs. Type I fibers terminate in ezoid body, VCA Anteroventral cochlear nucleus, VCP
the AVCN, and type II fibers terminate on the granule cell Posteroventral cochlear nucleus
regions (GrC) and marginal shell areas of the VCN and
echolocating mammals such as cetaceans and Sound waves are transmitted mechanically
bats. Because one major objective in hearing through the outer and middle ear to the sensory
research is to understand the structure and physi- hair cells of the organ of Corti, in the cochlear
ology of the human auditory system, this entry is partition of the inner ear. Auditory nerve fibers
restricted to an overview of the general plan of transmit information about receptor potentials
organization of the mammalian system. Insights generated by the sensory hair cells to the
gained from research in animals should aid in brainstem (Fig. 1). In contrast to the minimum
identifying the causes of hearing impairments in of two relay stations between the periphery and
humans and represent an important step toward cerebral cortex in the visual and somatosensory
developing effective treatments. systems, there is a minimum of three relays in the
The specific auditory stimulus consists of auditory system (Fig. 2), with several stages of
pressure waves arriving at the ear within convergence and divergence, and at least seven
a certain frequency range. This audible frequency levels of crossing fibers (Figs. 2 and 3), making
range varies among species (e.g., humans, the auditory system uniquely complicated. In the
0.02–20 kHz; rat, 0.25 to 70 kHz; mouse, first relay center, the cochlear nuclear complex
2–70 kHz; guinea pig, 0.2–45 kHz; and cat, (CNC), signals carried by the cochlear nerve are
0.125–60 kHz) (reviewed in Malmierca 2003; channeled into a number of parallel ascending
Fay 1988). Within their audible range, some spe- tracts (Figs 1 and 2), each with a particular course
cies, such as echolocating bats, are tuned to par- and destination and each presumably serving
ticular frequencies of special importance for their a different function (Fig. 3). Some of these ter-
behavior (reviewed in Fay and Popper 1994) and minate in a collection of nuclei in the pons known
are considered to be “auditory specialists” as the superior olivary complex (SOC). Ascend-
(Echteler et al. 1994). ing auditory tracts from both the CNC and the
Anatomy and Physiology of the Mammalian Audi- l Low-frequency region, ll lateral lemniscus, LTz Lateral
tory System, Fig. 2 Schematic wiring diagram of the nucleus of the trapezoid body, MG Medial geniculate
ascending auditory pathway (Modified after Brodal 1981, body, MTz Medial nucleus of the trapezoid body, PIL
AC is from Herbert et al. 1991). Abbreviations in the Posterior intralaminar nucleus, PP Peripeduncular
figure: bic Brachium of the inferior colliculus, cc Corpus nucleus, Rt Auditory sector of the reticular thalamic
callosum, CIC Central nucleus of the inferior colliculus, nucleus, SPO Superior paraolivary nucleus, Te1 Temporal
cic Commissure of the inferior colliculus, cll Commissure area 1, Te2 Temporal area 2, Te3 Temporal area 3, tz
of the lateral lemniscus (Prosbt), das Dorsal acoustic stria, Trapezoid body (or ventral acoustic stria), VC Ventral
h High-frequency region, DC Dorsal cochlear nucleus, cochlear nucleus
SOC converge on the inferior colliculus in The Auditory Nerve

the midbrain (Figs. 2, 3, and 8). From the mid-
brain upward, the auditory pathway is often The auditory nerve (AN) is made of both afferent
divided into core or “lemniscal” projections and efferent fibers (for review, see Slepecky
with a clear tonotopic organization and belt or 1996). The afferent fibers transmit impulses
“nonlemniscal” projections where tonotopy is from the organ of Corti to the cochlear nuclear
less sharp or absent (for review, see Malmierca complex, while the efferent fibers convey
2003; Malmierca and Hackett 2010). impulses from the superior olivary complex to
Anatomy and Physiology of the Mammalian Audi- of the inferior colliculus, cic Commissure of the inferior
tory System, Fig. 3 Projecting cell types of the CNC colliculus, ll lateral lemniscus, MTz Medial nucleus of the
and their corresponding physiological responses trapezoid body, PnC Pontine reticular nucleus, caudalis,
(Modified after Moore and Osen 1979). For abbreviations, SPO Superior paraolivary nucleus
see list. Abbreviations in the figure: CIC Central nucleus
the organ of Corti (Fig. 1). There are two sub- (type II) fibers are thinner and arise from small,
types of afferent fibers: myelinated and unmy- pseudounipolar spiral ganglion cells that inner-
elinated fibers (Fig. 1). The myelinated (type I) vate the outer hair cells (OHCs). About 90–95 %
fibers are relatively thick and originate from of auditory nerve fibers are type I (Fig. 1); each
bipolar spiral ganglion cells that innervate the type I fiber terminates on a single IHC. The type
inner hair cells (IHCs). The unmyelinated II fibers constitute only about 5 % of all auditory
nerve fibers (Fig. 1). As opposed to type I, type II by fibers of glycinergic commissural neurons as
fibers are highly branched, and a single fiber described below (for review, see Cant and Ben-
forms synapses with many (6–100) OHCs. son 2003). In addition to its ascending inputs A
Three types of type I fiber have been charac- from the AN, the CNC receives descending pro-
terized based on morphological features that cor- jections from the auditory cortex, the inferior
relate with their spontaneous activity and colliculus; the ventral complex of the lateral lem-
threshold sensitivity (Liberman et al. 1990). niscus; and the superior olivary complex (see
Fibers with a low threshold and a high spontane- Malmierca 2003). A large proportion of the latter
ous firing rate have the larger diameter and ter- fibers may be inhibitory, glycine and/or GABA
minate on the pillar side of the IHC, whereas being the transmitters, but there are also excit-
fibers with a high threshold and a low spontane- atory descending fibers, e.g., collaterals of the
ous firing rate are thinner and terminate on the cholinergic olivocochlear bundle (Osen et al.
modiolar side. 1984). The CNC also receives some projections
The efferent fibers of the olivocochlear system from nonauditory brain structures.
belong to the descending auditory pathways
(Fig. 1). They can be divided into two groups: Primary Afferents
the lateral efferent system (lateral olivocochlear Each AN fiber bifurcates into an ascending
system) that innervates auditory nerve fibers near branch, which supplies the AVCN, and
their synapses with IHCs and the medial efferent a descending branch, which supplies the PVCN
system (medial olivocochlear system) that inner- and DCN (Fig. 1). The anatomical distribution of
vates the OHCs (Warr 1992). the primary fibers forms the basis for the laminar
tonotopic organization of the three subnuclei
observed electrophysiologically (for review, see
The Cochlear Nuclear Complex Ryugo and Parks 2003). Type I fibers supply all
parts of the CNC except the superficial granule
The cochlear nucleus complex is situated later- cell areas and the molecular layer of the DCN
ally and superficially in the brainstem (CNC, (Fig. 1). Two basic types of terminals are found:
Figs. 1–5) and is the first relay center for ascend- large, axosomatic endings called “endbulbs of
ing auditory information. It is the site of termina- Held” and small boutons. The endbulbs of Held
tion of all auditory nerve (AN) fibers (Ryugo and arise from the ascending branches, while small
Parks 2003). The CNC consists of a ventral boutons arise from loosely ramifying collaterals
cochlear nucleus (VCN) and a dorsal cochlear of both ascending and descending branches. The
nucleus (DCN). The VCN is subdivided by the type II fibers innervate areas rich in granule cells
cochlear nerve root into anteroventral (AVCN) and appear to supply the marginal shell of the
and posteroventral (PVCN) parts. The DCN VCN (Fig. 1) (for review, see Ryugo and Parks
curves around the inferior cerebellar peduncle in 2003).
the floor of the lateral recess of the fourth ventri-
cle. The axons of CNC projection neurons leave Ventral Cochlear Nucleus
the complex via the three primary pathways to Five main neuronal types are recognized in the
reach higher auditory structures: the dorsal, inter- VCN based on patterns of Nissl staining, den-
mediate, and ventral acoustic striae (DAS, IAS, dritic arborization, and main axonal projections:
and VAS, respectively). The VAS is usually spherical bushy, globular bushy, octopus, multi-
referred to as the trapezoid body (Figs. 2 and 3). polar (or stellate), and small cells (Figs. 3–5;
The projections are largely tonotopically orga- Osen 1969; Brawer et al. 1974).
nized, and neurons within an isofrequency lamina The spherical bushy cells are found rostrally in
of the CNC project to a corresponding the AVCN, the globular bushy cells lie centrally
isofrequency lamina in higher-order centers. on both sides of the nerve root in the caudal
The right and left CNC are also interconnected AVCN and the rostral PVCN, and the octopus
cells are found caudally in the PVCN. The spher- The multipolar and small cells are present
ical bushy, globular bushy, and octopus neurons throughout the VCN (Figs. 3 and 4). The small
(Fig. 3) have non-tapering dendrites that end in cells are most abundant around the peripheral
bushy-like formations, but they differ with regard margins of the nucleus deep to the superficial
to the appearance of the terminal bush, the num- granule cell layer. A large collection of small
ber of root segments, and the relative length of cells located dorsolaterally in a superficial loca-
the stem dendrites. The spherical bushy cells tion forms the small cell cap of the VCN (Osen
receive a small number of large axosomatic ter- 1969). This is particularly conspicuous in cat. The
minals, the endbulbs of Held (for review, see multipolar cells (Figs. 3 and 6) possess moder-
Ryugo and Parks 2003), and have the so-called ately branched, tapering dendrites which are
primary-like responses to pure tone stimulation, contacted by small boutons from many primary
similar to those of the auditory nerve fibers afferent fibers. Two types of multipolar cell have
(Young et al. 1988; Young and Davis 2002). been described (Fig. 3): type I, also referred to as
The spherical bushy cells (Fig. 3) project bilater- T-stellate (mouse) and planar (rat), and type II,
ally to the medial superior olive and to the ipsi- also referred to as D-stellate (mouse) or radiate
lateral lateral superior olive (Cant and Benson (rat).
2003). Globular bushy cells receive a larger num- Multipolar type I cells (Figs. 3 and 4) have
ber of distinct inputs than do the spherical bushy oriented dendritic arbors and project to the
cells, including small (or “modified”) endbulbs. periolivary region of the superior olivary com-
In response to pure tone stimulation, they exhibit plex via the trapezoid body, the nuclei of the
a firing pattern known as “primary-like with lateral lemniscus, and the central nucleus of the
notch.” The globular bushy cells (Fig. 3) inferior colliculus through the lateral lemniscus
project to the ipsilateral lateral nucleus of the (Adams 1979; Cant and Benson 2003; Malmierca
trapezoid body and the contralateral et al. 2005). They also give rise to frequency-
medial nucleus of the trapezoid body. Axons specific collaterals within the VCN and DCN
from both globular and spherical bushy cells (Lorente de Nó 1981). These multipolar neurons
course ventrally in the trapezoid body. The pat- exhibit a “chopper” response to tone bursts with
terns of their AN input, along with unique regularly repeated firing. They may be special-
cell membrane properties, make these cells capa- ized for conveying frequency-specific excitatory
ble of transmitting precise temporal information information about complex acoustic stimuli such
necessary for both high- and low-frequency as speech.
sound localization (Young et al. 1988; Young Multipolar type II cells (Figs. 3 and 4), also
and Davis 2002). The octopus cells (Fig. 6) known as commissural neurons, have
receive small boutons from collaterals arising non-oriented dendritic arbors. They project to
from the descending branch of the AN. They the contralateral CNC via the dorsal acoustic
respond to a tone burst with a single spike and stria (Smith and Rhode 1989; Oertel et al. 1990;
so have been called onset units (Young Doucet and Ryugo 1997). They also emit exten-
et al. 1988; Young and Davis 2002). Their main sive (probably broadly tuned) collaterals within
projection is to the superior paraolivary nucleus the ipsilateral VCN and DCN. They are the only
on both sides and to the contralateral ventral known inhibitory (glycinergic) projection neu-
complex of the lateral lemniscus, and their rons of the CNC (Osen et al. 1990; Doucet
axons course dorsally, looping over the restiform et al. 1999) and respond to pure tone stimulation
body in the intermediate acoustic stria with an “on-chop” pattern (Smith and Rhode
(Wickesberg and Oertel 1988; for review, see 1989).
Cant and Benson 2003). Their function is still The small cells are abundant in the marginal
unclear, but it has been suggested that they shell of the VCN, which is composed of the
encode the pitch period in their temporal firing “granule cell layer” and the subjacent “cap
patterns (Oertel 1999). area”. The granule cell layer is continuous over
Anatomy and
Physiology of the
Mammalian Auditory
System, Fig. 4 Diagram A
of the planar (T-stellate)
and radiate (D-stellate) cells
in the VCN and their
projections to the DCN.
Planar cells have
frequency-specific
projections, while radiate
cells have across frequency
projections (Figure kindly
provided by
Dr. D. K. Ryugo.
Reproduced from Doucet
and Ryugo 1997). For
abbreviations, see list.
Abbreviations in the figure:
AN auditory nerve, DC
Dorsal cochlear nucleus,
LTz Lateral nucleus of the
trapezoid body, VCA
Anteroventral cochlear
nucleus, VCP
Posteroventral cochlear
nucleus
the free surface of the CNC and forms a lamina (Adams 1979; Malmierca et al. 2002, 2005; Ye
partly separating the VCN and DCN (Mugnaini et al. 2000). The available electrophysiological
et al. 1980a, b; for review, see Cant and Benson studies suggest that they form part of a feedback
2003). In the DCN, the granule cell layer is gain control system made up of the cochlea,
covered superficially by a molecular layer. The cochlear nuclear complex, medial olivocochlear
granule cell axons project as parallel fibers system, and outer hair cells (Ye et al. 2000).
(Fig. 5) to the molecular DCN layer (Mugnaini Cochlear Root Neurons (Fig. 3). The CNC of
et al. 1980a, b). The cap area is small but still rodents contains a population of large cells
distinguishable due to its contingent of small scattered in the cochlear nerve root, between the
cells, many of which show glycine- and/or main body of the VCN and the glial Schwann-cell
GABA-like immunoreactivity (for review, see border of the AN. It has been suggested that these
Cant and Benson 2003). The cap is supplied by root neurons participate in the acoustic startle
both type I and type II fibers (Fig. 1), and at least reflex (Sinex et al. 2001).
in cat, nearly all type I auditory nerve fibers that
innervate the cap have low spontaneous rates (for Dorsal Cochlear Nucleus
review, see Ryugo and Parks 2003). The marginal The DCN varies from being markedly laminated
shell also receives descending input. Its cells in rodents and carnivores, where it resembles the
show a wide dynamic range (Ye et al. 2000) cerebellar cortex, to appearing nonlaminated in
and have a large diversity of projections humans (but see Rubio et al. 2008) and some bat
Anatomy and
Physiology of the
Mammalian Auditory
System, Fig. 5 Synaptic
endings containing
glutamate, glycine, and
GABA in the rat dorsal
cochlear nucleus. Synaptic
endings and of the known
(solid lines) and putative
(dashed and doted lines)
neuronal sources of
excitatory and inhibitory
endings onto pyramidal
(FC), vertical (VC), and
cartwheel (CWC) cells
(Figure kindly provided by
Dr. M. E. Rubio.
Reproduced from Rubio
and Juiz 2004)
species; it is virtually absent in some cetaceans. basal arbor, which is supplied by primary afferent
The three superficial layers of the DCN are fibers in a strictly tonotopic manner.
related to the morphology of the principal pyra- The pyramidal cells are the main projection
midal (fusiform) cells (Figs. 3 and 5). The spiny neurons of the DCN, supplying fibers to the con-
apical dendritic arbor of pyramidal cells occupies tralateral IC via the DAS (Fig. 3). In addition,
layer 1 together with granule cell axons and sev- some have a direct projection to the medial divi-
eral other types of interneurons (Fig. 5, see sion of the medial geniculate body (Malmierca
below). Pyramidal cell bodies define layer 2, et al. 2002). The deepest layer of the DCN con-
and their aspinous basal dendritic arbors com- tains two categories of cells based on their size,
prise layer 3. Pyramidal cell dendritic arbors are the giant cells which project to the contralateral
flattened across the long, frequency gradient axis IC through the DAS and smaller glycinergic
of the DCN (see, e.g., Cant and Benson 2003; tuberculoventral interneurons (Fig. 5). Pyramidal
Malmierca 2003). The highest degree of flatness and giant cell excitatory responses are more
and mutually parallel orientation is found in the strongly influenced by their inhibitory inputs
than are those of other projection neurons in the glycinergic cartwheel cells (Fig. 5; Oertel and
CNC and have been classified as types III and IV Young 2004; Rubio and Juiz 2004). The granule
(Oertel and Young 2004). The type IV units have cells receive direct excitatory input from many A
been found to be sensitive to spectral notches sources including the somatosensory system and
created by the pinna, which may be important inhibitory inputs via the Golgi cells. The granule
cues for localizing sounds. DCN projection neu- cell axons contribute parallel fibers to the molec-
rons receive and respond not only to auditory ular layer. The unipolar brush cells seem to rep-
information but also to somatosensory inputs resent a device for feedforward excitation to the
from muscle proprioceptors in and around the mossy fiber pathways, while the stellate cells and
pinna. This innervation has led to speculation cartwheel cells provide feedforward inhibition to
that the DCN may be involved in coordinating the pyramidal cells (Rubio and Juiz 2004; Fig. 5).
pinna orientation with localization cues found in
the different spectra of sounds located at different
points in space. In fact, bilateral lesions of the The Superior Olivary Complex
DAS in cats result in reduced accuracy in head
orientation responses to broadband sounds, par- The superior olivary complex (SOC) comprises
ticularly in elevation (reviewed in Young and a group of nuclei in the caudal pons (Figs. 2
Davis 2002). and 6). Three main nuclei are consistently iden-
DCN possesses a large number of interneurons tified: the lateral superior olive (LSO), the medial
that may be divided into two systems: the superior olive (MSO), and the medial nucleus
tuberculoventral system and the granule cell sys- of the trapezoid body (MNTB). These three
tem (Fig. 5). The tuberculoventral system recip- nuclei are surrounded by more diffusely orga-
rocally interconnects the DCN and VCN. It nized cellular areas, collectively referred to as
contains both frequency-specific and diffuse pro- the periolivary region (PO) (Adams 1983; Osen
jections (Malmierca 2003). The frequency- et al. 1984; Schofield and Cant 1991). The LSO
specific projection from the DCN to the VCN and MNTB are well developed in both the rat and
originates from small interneurons, a subset of cat, while they are relatively small in human. In
the glycinergic “vertical cells” (Fig. 5). A sepa- contrast, the MSO is small in the rat but is prom-
rate set of vertical cells with only local collaterals inent in both the cat and human (Fig. 6). These
contain both GABA and glycine, the relative differences appear to be related to the ability to
amounts of which vary among species. They are use specific frequency cues for directional hear-
located between the basal pyramidal cell den- ing. The MSO extracts the information about
drites in layer 3, have flattened dendritic arbors, interaural timing differences that is available in
and provide the DCN and VCN with low-frequency sounds. Together, the LSO and
a tonotopically organized inhibition. One projec- MNTB detect interaural intensity differences
tion from the VCN to the DCN is made up of the generated by high-frequency sounds.
collaterals of type I multipolar cells described
above and probably is excitatory and Lateral Superior Olive
tonotopic. An inhibitory projection arises from The LSO consists of layers of flattened multipolar
axonal collaterals of the glycinergic commissural neurons (principal cells) with their dendrites ori-
(type II) cells. The vertical cells of the DCN ented perpendicular to its long axis, which is
provide inhibition over a narrow frequency curved into an S-shape. The LSO is tonotopically
range, whereas the on-chop, type II stellate cells organized with low frequencies represented lat-
generate inhibition over a wide frequency range. erally and high frequencies, medially. In addition
The granule cell system includes two types of to the principal cells, other, less abundant, neuro-
excitatory cells: granule cells and unipolar brush nal types are also present (Rietzel and Friauf
cells and three types of inhibitory cells: the 1998). In some species, neurons of the lateral
GABAergic Golgi and stellate cells and the olivocochlear system lie within the LSO.
Anatomy and
Physiology of the
Mammalian Auditory
System, Fig. 6 (a)
Comparison of the superior
olivary complex in the rat
and cat (Redrawn after
Osen et al. 1984). Note the
relative size of of the
LSO–MSO in the two
species and the existence of
a distinct SPO in the rat.
(b), Camera lucida drawing
of a section showing
calbindin-positive neurons
in the MTz and processes in
the rat SOC (Redrawn after
Friauf 1993). (c) Confocal
image illustrating
VGLUT1-ir in the SOC of
adult rats. VGLUT1-ir is
green, and MAP2-ir is red.
For abbreviations, see list
(Data from Blaesse
et al. 2005. Figure kindly
provided by Dr. E. Friauf).
DMPO Dorsomedial
periolivary region, LPO
Lateral periolivary zone,
LTz Lateral nucleus of the
trapezoid body, MTz
Medial nucleus of the
trapezoid body, PO
Periolivary regions, SPO
Superior paraolivary
nucleus, tz Trapezoid body
(or ventral acoustic stria);
VTz Ventral nucleus of the
trapezoid body
The LSO receives direct input from the AVCN sends an inhibitory (glycinergic) input to the
on the ipsilateral side and indirect input from the LSO (Fig. 6). Multipolar type I neurons from
AVCN on the contralateral side (Fig. 6). The the AVCN also innervate the LSO (reviewed in
ipsilateral input derives from spherical bushy Helfert and Aschoff 1997; Thompson and Scho-
cells and is excitatory. Globular bushy cells in field 2000).
the contralateral AVCN project to the MNTB The LSO projects to the central nucleus of the
ipsilateral to the LSO. The MNTB, in turn, inferior colliculus bilaterally (Figs. 2 and 6).
Most of the ipsilaterally projecting cells are The multipolar non-principal cells are scattered
inhibitory (glycinergic), while the contralaterally among these dendrites and are much less numer-
projecting cells are glycine-negative and presum- ous (Smith 1995). The MSO is tonotopically A
ably excitatory. The LSO also innervates the dor- organized with low-frequency tones represented
sal nucleus of the lateral lemniscus bilaterally dorsally and high-frequency tones ventrally,
(Fig. 2). although most of the nucleus appears to be
devoted to low frequencies (Guinan et al. 1972).
Medial Nucleus of the Trapezoid Body The MSO receives direct, excitatory input
Cells of the MNTB are situated among fascicles from spherical bushy cells in the AVCN bilater-
of fibers in the trapezoid body (Figs. 2 and 6). The ally (Fig. 6; for review, see Malmierca 2003), but
principal cells resemble the globular bushy cells these inputs remain segregated on the cell sur-
of the AVCN, whereas non-principal (marginal) face, such that the lateral dendrites receive input
cells have a multipolar appearance (Fig. 6; from the ipsilateral side, while the medial den-
Morest 1968; Banks and Smith 1992; Sommer drites receive input from the contralateral side
et al. 1993). (Smith 1995). A rostrocaudal organization of the
The MNTB receives input from the globular axons is reminiscent of the required input configu-
bushy cells in the contralateral VCN (Fig. 2). ration in the Jeffress model for sound localization.
These cells give rise to thick axons that terminate The direct bilateral input suggests that the MSO
on the principal cells in the MNTB as large neurons are ideally suited to measure interaural
axosomatic calyces of Held (1893) in a one-to- phase or time differences (Joris et al. 1998).
one relationship. These calyces provide a fast and The MSO also receives inhibitory inputs,
secure relay of information from the CNC to the mostly glycinergic, from the medial and lateral
MNTB and from there to the LSO. They consti- nuclei of the trapezoid body on the same side.
tute the largest synaptic terminals in the mamma- The latter may also provide GABAergic inhibi-
lian brain. The responses of cells in the MNTB to tory input (Smith 1995). The principal cells pro-
acoustic stimuli show a sharp onset spike that is ject to the inferior colliculus and to the ipsilateral
also characteristic of the globular bushy cells. dorsal nucleus of the lateral lemniscus (Fig. 8).
In addition to its projection to the LSO, the
MNTB projects to the ipsilateral MSO as well as Superior Paraolivary Nucleus
other parts of the ipsilateral SOC and the dorsal A fourth distinct nucleus in the SOC of rodents,
portion of the ventral complex of the lateral lem- the so-called superior paraolivary nucleus (SPO),
niscus, providing a source of widespread is found in the dorsomedial part of the complex
glycinergic inhibition. (Fig. 6; Osen et al. 1984; Schofield and Cant
The microcircuitry and neurochemistry of the 1991; Schofield 1995). It consists of GABAergic
LSO and MNTB suggest that a major function of multipolar cells, which are the largest in the SOC,
the MNTB is to transform excitatory contralateral and receives inputs from octopus and multipolar
input into ipsilateral inhibition, allowing the LSO cells in the contralateral VCN, from multipolar
to faithfully encode interaural intensity differ- cells in the ipsilateral VCN, and a substantial
ences in the high-frequency range of audition glycinergic input from the MNTB on the same
(for review, see Kopp-Scheinpflug et al. 2008). side. SPO projects to the ipsilateral IC (Schofield
1995) and may represent a hyperdevelopment of
Medial Superior Olive periolivary cells with a similar projection, present
The MSO lies between the LSO and the MNTB in smaller numbers in other mammals (Adams
(Fig. 6) and is populated by two types of cells: 1983). It appears that the SPO neurons are well
principal and non-principal or marginal cells. The suited for the analysis of temporal features of
bipolar principal cells are organized into complex sounds and stimulus features across
a transversely oriented row, with their dendrites broad frequency ranges (Dehmel et al. 2002;
extending in the medial and lateral directions. Kulesza et al. 2003).
Periolivary Nuclei ear via the contralateral VCN and ipsilateral

The PO contains several distinct types of neurons MNTB (Fig. 2). VLL neurons exhibit a variety
with different projection patterns (Adams 1983; of shapes and sizes in Nissl-stained sections, and
Osen et al. 1984). Neurons in the PO areas most of them project to the ipsilateral IC. The
receive input from VCN bilaterally, the lateral majority of cells in the ventral part of the complex
areas, from the ipsilateral side, and the medial are glycine and/or GABA (Riquelme et al. 2001).
areas, from both sides (Fig. 6). Certain parts of In vitro, some VLL neurons show an onset
the PO also receive input from the ipsilateral firing pattern and a nonlinear current-voltage
MNTB (probably inhibitory), from the ipsilateral relationship, while others exhibit a linear
inferior colliculus (probably excitatory), and current-voltage relationship and other firing pat-
from the dorsal nucleus of the lateral lemniscus terns (Wu 1999; Zhao and Wu (2001)). Similar
(probably inhibitory). differences have also been found in in vivo stud-
PO cells project either to the cochlea, the ies (Zhang and Kelly 2006a, b). The VLL neurons
CNC, or the IC, but individual cells do not appear are suitable for encoding temporal events (Covey
to project to all three structures (Adams 1983). and Casseday 1991).
The ventral nucleus of the trapezoid body
(VNTB) is a PO region situated ventral to the The Dorsal Nucleus of the Lateral Lemniscus:
MNTB and is of particular interest because it The Binaural System
may be involved in the activation of the The DLL receives input from both ears, and it
olivocochlear neurons (Rajan 1990). It is strate- projects both to ICs and to its counterpart on the
gically situated at the intersection of ascending opposite side through the commissure of Probst
projections from the CNC and descending pro- (Fig. 2). The DLL plays an important role in
jections from the IC. The VNTB receives major functions dependent on binaural processing such
afferent projections from the contralateral VCN, as sound localization.
the ipsilateral PVCN (Smith et al. 1991; Thomp- Several neuronal types have been described,
son and Schofield 2000), and the marginal shell in depending on the species and the criteria used for
the AVCN from both sides (Ye et al. 2000). It is cell classification. Regardless of the morphological
the major target in the SOC for the descending type, all cells have similar membrane properties
projections from the IC. with a sustained series of regular action potentials
produced by the injection of positive current.
Generally speaking, the DLL receives collat-
The Nuclei of the Lateral Lemniscus erals from afferent fibers that also innervate the
IC (Fig. 8). Thus, DLL receives contralateral
The nuclei of the lateral lemniscus (NLL) is made inputs from the ventral cochlear nucleus and
is of two distinct functional systems (Fig. 2), DLL, ipsilateral input from the medial superior
a monaural ventral (VLL) and a binaural dorsal olive, superior paraolivary nucleus and VLL, and
(DLL) system (Covey and Casseday 1991). There bilateral inputs from the lateral superior olive.
are connectional, neurochemical, and physiolog- The DLL projection to the IC is laminar and
ical properties which are unique to each system bilateral, with a predominant projection to the
(Covey and Casseday 1991; Malmierca contralateral IC. Most DLL cells are GABAergic
et al. 1998). and therefore have an inhibitory influence on
the IC.
The Ventral Nucleus of the Lateral Lemniscus:
The Monaural System
The VLL consists of groups of neurons embed- The Inferior Colliculus
ded within the part of the lateral lemniscus,
located between the SOC and DLL (Fig. 2). It The inferior colliculus (IC) is the principal audi-
receives its inputs mainly from the contralateral tory nucleus in the midbrain and is characterized
strongly influenced by the descending pathways
and commissural inputs. The LCIC (and probably
the RCIC) receives both auditory and A
nonauditory (e.g., somatosensory) inputs.
In addition, to its intrinsic and commissural
fiber systems (Malmierca et al. 1995; Saldaña and
Merchán 1992), the IC provides the major
ascending projections to the MGB as well as
descending projections to the SOC and the CNC
(Oliver et al. 1999; Malmierca et al. 1996;
Peruzzi et al. 1997; Ito and Oliver 2012; Cant
and Benson 2007).
The Central Nucleus of the Inferior Colliculus

The CNIC is distinguished by layers of cells and
fibers organized into “fibrodendritic laminae”
(Oliver and Morest 1984; Faye-lund and Osen
1985; Malmierca et al. 1993). These consist of
a parallel organization of afferent lemniscal
fibers and neurons with flattened dendritic arbors
and constitute the structural basis for the
tonotopic organization of the IC (Schreiner and
Langner 1997; Malmierca et al. 2008; Fig. 9).
The CNIC is composed of two main neuronal
types: disk-shaped or flat neurons and stellate or
less flat neurons (Oliver and Morest 1984;
Anatomy and Physiology of the Mammalian Audi-
tory System, Fig. 7 Subdivisions of the inferior Malmierca et al. 1993). These neurons differ in
colliculus in the rat and cat. The low-frequency represen- several respects, including the thickness,
tation in the central nucleus (CIC) is expanded in cat and branching pattern and orientation of their den-
contracted in rats, while the size of the ventrolateral dritic arbor, and their location with regard to the
nucleus (VLN) (layer 3 of LCIC) is expanded in rat and
contracted in cats (Redrawn from Loftus et al. 2008) laminae (Fig. 10). The flat neurons lie parallel to
the ascending fibers and form laminae that are
mostly one cell thick, about 40–70 mm (Faye-
by a massive convergence of inputs from lower Lund and Osen 1985; Malmierca et al. 1993).
and higher auditory centers as well as from These laminae are separated by interlaminar
nonauditory structures (Figs. 2 and 7; Irvine compartments that are populated by the stellate
1992; Malmierca 2003; Casseday et al. 2002; neurons. IC neurons show complex functional
Loftus et al. 2008). The IC is divided into properties, and there is no simple correlation
a central nucleus (CNIC) surrounded by cortical between physiological response properties and
regions (Fig. 7). These collicular cortices include morphological classifications (Sivaramakrishnan
a dorsal cortex (DCIC) that covers the CNIC and Oliver 2001).
dorsally and caudally, a lateral cortex (LCIC) The CNIC receives ascending afferent projec-
that covers it laterally, and a rostral cortex tions (Fig. 8) that originate in the AVCN, PVCN,
(RCIC) that covers it rostrally (Loftus and DCN bilaterally (although mainly
et al. 2008; Fig. 8). contralaterally); the VLL and MSO ipsilaterally;
Neurons in the CNIC tend to be most strongly and the DLL and LSO bilaterally (Malmierca
influenced by lower auditory centers, while neu- 2003; Cant and Benson 2006). These inputs are
rons in the DCIC and RCIC tend to be most tonotopically organized and tend to show
Anatomy and
Physiology of the
Mammalian Auditory
System, Fig. 8 Summary
of major sources of inputs
to the inferior colliculus on
one side of the brain (left)
from auditory brainstem
(ascending inputs) and
auditory cortex
(descending inputs).
Connection strength is
denoted by the thickness of
the arrows. Red lines
represent excitatory
connections, and blue/
dashed lines represent
inhibitory connections.
DC Dorsal cochlear
nucleus, LSO Lateral
superior olive, MSO Medial
superior olive, SPO
Superior paraolivary
nucleus, VC Ventral
cochlear nucleus, VTz
Ventral nucleus of the
trapezoid body
a “banded” pattern of projections with dense demonstrated that the AMPA receptors regulate
bands (about 200 mm thick) parallel to the the onset response. Both AMPA and NMDA are
isofrequency fibrodendritic laminae. The termi- involved in the maintenance of the response for
nal fields of the various ascending projections the duration of the stimulus, and both GABA and
may also vary in their distribution along the glycine inhibit IC neurons and shape different
main axis of the IC. For example, in the temporal, spectral, and binaural properties (e.g.,
dorsomedial parts of the laminae, the axons LeBeau et al. 2001; reviewed in Kelly and
from the DCN do not overlap with afferents Caspary 2005). Neurons in the IC show
from the LSO. The CNIC projects to the ventral a variety of frequency response areas that include
division of the MGB in a tonotopic manner both V-shaped and non-V shaped types (LeBeau
(Fig. 10), largely to the ipsilateral side but also et al. 2001; Palombi and Caspary 1996). Binaural
with a crossed component. A majority of neurons responses are generally classified as suppression,
use glutamate as the neurotransmitter, but about summation, or mixed. The laminae of the CNIC
a quarter of the cells in the CNIC are GABAergic contain a highly organized representation of both
(Merchán et al. 2005), and some of these project spectral and temporal parameters (for review, see
to the MGB, making short-latency, monosynaptic Schreiner and Langner 1988).
inputs (Peruzzi et al. 1997).
IC neurons possess NMDA and AMPA recep- The Dorsal Cortex of the Inferior Colliculus
tors as well as GABAA, GABAB, and glycine The DCIC covers the dorsomedial and caudal
receptors (reviewed in Kelly and Caspary 2005). aspects of the CNIC (Figs. 7 and 9b) and is
Studies using microiontophoresis in vivo have made of three layers (Faye-Lund and Osen
Anatomy and Physiology of the Mammalian Audi- intervals. Frequency increases as a function of depth in
tory System, Fig. 9 Summary of the tonotopic organi- a stepwise manner. (c) 3-D reconstruction of three axonal
zation of the inferior colliculus. (a) A single electrode laminae (yellow, 1. 7 kHz lamina; red, 1. 8 kHz; green,
penetration downward (blue trace) and upward (red 4. 5 kHz) (Data from Malmierca et al. 2008). (d) A frontal
trace) along the same electrode track (Tr) along the CIC view of the same 3-D reconstruction seen in (c) after 90
in which FRA obtained from multiunit clusters were rotation. (e and f) FRAs recorded at every 50 mm in the
recorded at every 50 mm. (b) Frequency representation in electrode penetration illustrated in (a) at four different
the CIC obtained in a sagittal section after Nissl staining steps (step A–step D). D Depth (Redrawn from Malmierca
showing an electrode track (black line), three electrolytic et al. 2008)
lesions (circles), and best frequency recorded at 50 mm
1985). Layer 1 (the most superficial layer) is slightly thicker and contains mostly small- and
a thin fibrocellular capsule that wraps the entire medium-sized multipolar neurons. Layer 3
surface of the IC including the LCIC. It contains includes a heterogeneous group of multipolar
scattered, small, flattened neurons. Layer 2 is neurons that vary in soma size and shape and in
Anatomy and Physiology of the Mammalian Audi- to MGB. Glutamatergic neurons (3, 4) project to the MGB
tory System, Fig. 10 Schematic diagram of the basic and lack the dense VGLUT2 axosomatic inputs. Red
circuitry in the IC. Large GABAergic neurons (1) puncta indicate excitatory glutamatergic terminals. Blue
receive excitatory inputs on their somata, send their puncta indicate inhibitory (GABAergic and glycinergic)
axons to the MGB, and inhibit tufted or stellate neurons terminals (Figure kindly provided by Dr. D. L. Oliver.
in the MGB. Small GABAergic neurons (2) do not project Reproduced from Ito and Oliver 2012)
their dendritic orientation. A distinct feature of Loftus et al. 2008). Layer 1 is a continuation of
neurons in DCIC (and also LCIC) is that they the fibrodendritic capsule of the DCIC. Layer 2 is
contain the neuromodulator nitric oxide, a fact composed of small- and medium-sized neurons
that may explain some of the long-term potentia- that are partly aggregated into dense clusters rich
tion and neuronal plasticity observed in some IC in acetylcholinesterase and GABA. Layer 3 of the
neurons (Zhang and Wu 2000). ECIC constitutes its largest part and appears to
The ascending input from lower auditory cen- continue also into the non-stratified rostral cor-
ters to the CNIC encroaches upon the DCIC, as tex. In addition to auditory inputs, the LCIC
do the intrinsic projections (Saldaña and Merchán receives somatosensory input from spinal cord,
1992; Malmierca et al. 1995). In addition, the dorsal column nuclei, and spinal trigeminal
DCIC receives descending input bilaterally that nuclei in the cat (Zhou and Shore 2006). Neurons
originates largely from the primary AC. This pro- in the LCIC appear to have relatively broad
jection may generate long-lasting changes in the somatosensory receptive fields in addition to
neuronal responses of the IC. The DCIC projects auditory responses, which are also broadly
to the dorsal division of the MGB (Winer 1985). tuned (Aitkin et al. 1978). The multisensory inte-
gration in the LCIC mirrors similar types of func-
The Lateral Cortex of the Inferior Colliculus tion at higher levels of the “extralemniscal”
The LCIC covers the CNIC laterally (Fig. 7) and auditory pathway. Although the functions of the
ventrally (Loftus et al. 2008) and is made up LCIC are not known, it may be a major source of
of three layers (Faye-Lund and Osen 1985; binaural cues for gaze control in the superior
colliculus (King et al. 1998), and it is very likely The main source of ascending projections to
that LCIC plays an important role in providing the MGB is the IC (Figs. 2 and 10), but other
auditory input to visual-motor areas that direct inputs include the thalamic reticular nucleus and A
the head and eye movements involved in gaze auditory subcollicular nuclei, including the SOC,
initiation. A second role of the LCIC may be in NLL, and CNC, challenging the idea that the IC is
multisensory integration, distinct from the “clas- an obligatory relay station (Malmierca
sical” auditory role of the central nucleus. et al. 2002). Ascending fibers enter the structure
Finally, the CNIC and LCIC differentially acti- medially through the brachium of the IC and
vate the medial and lateral olivocochlear system terminate among neurons in each subdivision.
(Ota et al. 2004). Connections with auditory cortex pass through
the posterior limb of the internal capsule, and,
The Rostral Cortex of the Inferior Colliculus for the most part, these connections are
The rostral cortex (RCIC) is adjacent to the ante- reciprocal.
rior CNIC (Fig. 9b) and includes the largest mul-
tipolar cells in the IC. There are also small- and The Ventral Division of the Medial Geniculate
medium-sized multipolar neurons (Faye-Lund Complex
and Osen 1985; Malmierca et al. 1993). Like the The MGV contains bi-tufted principal neurons
LCIC and DCIC, the RCIC receives input from that tend to be arranged in rows with oriented
the cerebral cortex as well as from nonauditory dendritic fields (Fig. 12a–c; Winer et al. 1999).
structures (for review, see Malmierca 2003). The These neurons are glutamatergic, and most of
RCIC projects to the dorsal and medial divisions them are immunoreactive for parvalbumin
of the MG (Fig. 2). The functional role of RCIC (Jones 2007). In the cat, a quarter of the neurons
neurons is still unknown, although recent studies in the MGV are GABAergic interneurons, but
have demonstrated that many neurons in this these are absent in rodents and bats (Winer and
region (and in other cortical regions) may be Larue 1988). The MGV neurons are well tuned to
specialized for detecting novel sounds frequency, exhibiting low-threshold short-
(Malmierca et al. 2009). latency responses to tones and complex sounds.
They are sensitive to both interaural time and
intensity differences.
The Medial Geniculate Body The MGV receives its ascending inputs pri-
marily from the ipsilateral CNIC (Fig. 10),
The principal auditory nucleus in the thalamus is (Malmierca 2003; Jones 2007). These projections
the medial geniculate body (MGB), which is also arise from both glutamatergic and GABAergic
the last opportunity for subcortical processing in neurons (Bartlett and Smith 1999; Ito and Oliver
the ascending auditory pathways (Fig. 2). The 2012). MGV neurons project mainly to layers III
MGB is divided into three major divisions and IV of the primary (core) areas of the auditory
named relative to their locations within the com- cortex.
plex: the ventral (MGV), dorsal (MGD), and
medial (or magnocellular) (MGM) divisions The Dorsal Division of the Medial Geniculate
(Figs. 11 and 12). Additional subdivisions are Complex
also recognized in some species, usually The MGD is generally characterized by a lower
representing smaller domains within each of the cell-packing density than the MGV, and it lacks
major subdivisions. Physiological studies have a laminar organization. At least two subdivisions
shown that the caudal part of the reticular tha- are clearly recognized on the basis of architec-
lamic nucleus (Fig. 2) is also a part of the auditory tonic variations and connections in most species
pathway (Cotillon et al. 1999; Yu et al. 2009). It (Burton and Jones 1976; Winer 1985). The most
provides the MGB with an inhibitory GABAergic common neuronal type is the radiate cell
input (Bartlett and Smith 1999; Yu et al. 2009). (Fig. 12f–g). These have radially symmetric
Anatomy and Physiology of the Mammalian Audi- immunostained for calbindin, visualized with Cy-2. (e)
tory System, Fig. 11 Distribution of calbindin and Overlay images from (a and c). (f) Overlay images from
calretinin areas at two different rostrocaudal levels of the (b and d). Scale bar = 500 mm (Figure kindly provided by
MGB in mouse. (a and b) Two sections from same mouse Dr. Daniel Llano. Adapted from Lu et al. 2009). Abbrevi-
immunostained for calretinin, visualized with Alexa Fluor ations in the figure: PIL Posterior intralaminar nucleus, PP
594. (c and d) Identical sections from (a and b), Peripeduncular nucleus
dendritic fields with a simple branching pattern. MGV, parvalbumin immunoreactivity is reduced
Tufted cells (Fig. 12a–c) are also present and tend in the MGD, but there is a relative increase in the
to form thin sheets. Finally, there is a small pop- numbers of calbindin-positive cells (Jones 2007).
ulation of small stellate cells. Compared to the GABAergic interneurons are also abundant in the
Anatomy and
Physiology of the
Mammalian Auditory
System, Fig. 12 Camera A
lucida drawings of
neurobiotin-labeled cells of
the rat MG. Tufted cells
populate the MGV. Cells
with two distinct
morphologies, tufted and
stellate, populate
MGD. Cells (d and e) and
cells (f and g) are typical
examples of MGD tufted
and stellate cells,
respectively. The dendritic
trees of these cells appear
similar regardless of the
plane of section.
Magnocellular neurons
populate the MGM
(Figure kindly provided by
Dr. Philip Smith (some
parts reproduced from
Bartlett and Smith 1999)).
SC suprageniculate nucleus
MGD. Most MGD neurons respond over a wide heterogeneous with respect to cell types and con-
range of latencies, typically longer than MGV nections. Several different types of cells have
neurons, and exhibit broader frequency tuning, been identified including the magnocellular
so a clear tonotopic organization is not obvious, type, the largest neurons in the entire MGB
if it is present at all. (Fig. 12h–i; Winer and Morest 1983). Some of
The main ascending inputs to the MGD arise the neurons are calbindin reactive and project
from the DCIC and LCIC (Malmierca 2003; mainly to layers I and II of cortex. Projections
Jones 2007) and may be either glutamatergic or to the middle layers of cortex arise from both
GABAergic in nature (Bartlett and Smith 1999). calbindin- and parvalbumin-positive neurons
The MGD projects to the nonprimary (belt) areas (Jones 2007). Some neurons are narrowly tuned
of auditory cortex. to frequency and respond robustly at short laten-
cies, similar to MGV neurons, while others are
The Medial Division of the Medial Geniculate broadly tuned and have longer latencies. Some
Complex authors define a separate division referred to as
The MGM is located medial to the MGV and suprageniculate nucleus (Fig. 12j–k)
MGD and stretches from the rostral to caudal MGM receives inputs from both auditory and
poles of the MGB (Fig. 11). The MGM is rather nonauditory sources. The main auditory inputs
Anatomy and Physiology of the Mammalian Audi- areas are shaded. Tonotopic gradients are indicated by
tory System, Fig. 13 Schematics of auditory cortex H (high) and L (low) frequency (Reproduced from
models in several mammals. Primary (core) auditory Malmierca and Hackett 2010)
arise from the cortical regions of the IC, as well as some studies of humans (Fig. 13). Species differ-
the CNC, SOC, and VLL (Anderson et al. 2006; ences include the number of areas present, their
Malmierca et al. 2002). Nonauditory inputs relative position and arrangement, cell density,
include the deep layers of the superior colliculus connections, and tonotopic organization, and it is
and other nuclei that appear to drive responses to likely that species differences will also be
somatic, vestibular, visual, and nociceptive stim- reflected in the organization of auditory cortex
uli in some species (see Jones 2007). In addition in other ways. However, a common theme is that
to auditory cortex, the MGM also projects to the a central primary region, or core, is surrounded
striatum and amygdala (Doron and Ledoux by a variable number of secondary, or belt, areas
1999). in all mammals studied (Fig. 13). In nonhuman
primates, the core–belt scheme has been
extended to include a third region, known as the
The Auditory Cortex parabelt (Kaas and Hackett 1998; Winer and
Schreiner 2011).
The auditory cortex (AC) is located in the tem-
poral lobe of the cerebral hemisphere and repre- The Core Region of Auditory Cortex
sents the site of termination for fibers ascending The core region is made of the primary field
from the auditory thalamus (Fig. 2). The most (A1) and two other tonotopically organized
obvious species differences observed in the audi- areas (Fig. 13). Areas in the core are character-
tory brain are those seen in the AC (Fig. 13). For ized by high cell density in a thick layer IV, dense
example, the number of areas identified in the AC myelination across laminae, and relatively high
ranges from 5 to 6 in mice and rats, 6–9 in cats expression of several markers in the horizontal
and ferrets, 10–12 in primates, and over 30 in band involving layers III and IV, compared to
secondary areas in the belt (Jones 2003; Kaas and
Hackett 1998). Several different classes of pyra-
midal and nonpyramidal cells are found across A
the six layers of auditory cortex (Winer 1992).
Pyramidal cells tend to be glutamatergic and are
concentrated in layers III and V, while many
nonpyramidal cells are GABAergic and account
for about one-fourth of the neurons in most
layers, except layer I, where they constitute
more than 90 %. Layer II contains both pyramidal
and nonpyramidal neurons. The smaller cells are
located superficially in layer II, while the larger
pyramidal cells predominate near the border
with layer III. Layer III is populated by several
types of pyramidal and nonpyramidal cells. Layer
IV is mainly populated by small tufted cells,
which have radially oriented dendritic fields and
are involved mainly in local columnar projec-
tions. Layer V contains both pyramidal and
nonpyramidal neurons. The somata of the con-
spicuous large pyramidal neurons located in layer
Vb have apical dendrites that extend to layer I,
with several branches along the way. Other pyra-
midal cells in layer V are smaller and more
evenly distributed. Layer V is of particular inter-
est because its cells form part of the projection to
the thalamus and other subcortical areas (Fig. 14;
Games and Winer 1988; Hefti and Smith 2000;
Malmierca and Ryugo 2011). As in other cortical
regions, two distinct types of pyramidal cells,
“intrinsically bursting” and “regular spiking”
can be distinguished on the basis of their corre- Anatomy and Physiology of the Mammalian Audi-
lated morphology and physiology (Fig. 14; tory System, Fig. 14 Top panel, examples of morphol-
Kawaguchi 1993; Kasper et al. 1994). The intrin- ogy of layer V and layer VI corticothalamic neurons
biocytin-filled corticothalamic neurons. Scale bar
sically bursting pyramidal cells (Fig. 14) have
100 mm; Bottom panels, Llano and Sherman model of
large cell bodies and long, thick apical dendrites differences of synaptic input and response properties
that branch extensively in layer I. Their axons between auditory layer V and layer VI corticothalamic
arborize locally in the infragranular cortical neurons. In response to excitatory input, layer
V corticothalamic neurons fire a burst of action potentials
layers and project into subcortical white matter.
at low threshold, and these neurons receive excitatory
In contrast, the regular-spiking pyramidal cells input from neurons in layers II/III, IV, and V (gray) and
(Fig. 14) have smaller cell bodies and a thinner GABAA-mediated inhibitory input mostly from lower
apical dendrite that seldom extend to layer I. layer V with a smaller contribution from layer II/III
(red). In contrast, in response to excitatory input, these
Their axons also project to the white matter and
neurons fire a regular train of individual action potentials,
arborize locally in the supragranular cortex. In and these neurons receive excitatory input primarily from
slice preparations, the intrinsically bursting neu- layer VI and inhibitory input from adjacent areas in layer
rons exhibit a characteristic firing pattern with VI (Figure kindly provided by Dr. Daniel Llano. Adapted
from Llano and Sherman 2009)
a burst of action potentials followed by either
additional bursts or single spikes, whereas the
Anatomy and Physiology of the Mammalian Audi- region, ll Lateral lemniscus, LTz Lateral nucleus of the
tory System, Fig. 15 Schematic wiring diagram of the trapezoid body, MG Medial geniculate body, MTz Medial
descending auditory pathway of the rat (Modified after nucleus of the trapezoid body, ocb Olivocochlear bundle,
Brodal 1981, AC is from Herbert et al. 1991). Abbrevia- PIL Posterior intralaminar nucleus, PP Peripeduncular
tions in the figure: bic Brachium of the inferior colliculus, nucleus, Rt Auditory sector of the reticular thalamic
cc Corpus callosum, CIC Central nucleus of the inferior nucleus, SOC superior olivary nucleus, SPO Superior
colliculus, cic Commissure of the inferior colliculus, cll paraolivary nucleus, Te1 Temporal area 1, Te2 Temporal
Commissure of the lateral lemniscus (Prosbt), das Dorsal area 2, Te3 Temporal area 3, tz Trapezoid body (or ventral
acoustic stria, DC Dorsal cochlear nucleus, h High- acoustic stria), VC Ventral cochlear nucleus, 8cn Cochlear
frequency region, IC Inferior colliculus, l Low-frequency root of the vestibulocochlear nerve
regular spiking neurons tend to fire single spikes neurons are strongly inhibited and may provide
with a variable degree of adaptation (Hefti and less robust but perhaps more specific, information
Smith 2000, 2003). Intrinsically bursting cells to their inputs. Finally, layer VI has the widest
make up the majority of layer V’s input to sub- variety of cell types, including several classes of
cortical targets (Fig. 15; Games and Winer 1988; pyramidal cells and multipolar, bipolar, and hor-
Winer 1992; Weedman and Ryugo 1996a, b; izontal cells.
Saldaña et al. 1996) and are capable of providing The main source of ascending inputs to A1
a robust input to postsynaptic neurons (Hefti and and other core areas is the MGV (Fig. 2;
Smith 2000). In contrast, most regular spiking Jones 2007; Winer 1985; Winer and Lee 2007).
The thalamocortical termination is concentrated belt regions before reaching the parabelt (Kaas
in layers III and IV. The organization of these and Hackett 1998).
connections is topographic, reflecting tonotopic A
organization within the MGV, as well as the areas
to which it projects. Additional inputs to the core The Descending Auditory Pathways
include MGM, which projects broadly to all areas
of auditory cortex. The intracortical connections In parallel to the ascending auditory pathways,
of the core mainly include other areas of auditory there are stepwise, descending projections from
cortex ipsilaterally and sparse connections with the auditory cortex to the organ of Corti (Fig. 15).
areas beyond auditory cortex. Tonotopically The descending auditory pathway could be con-
matched sites are more densely interconnected sidered to consist of both (1) a descending chain
than non-matched sites (Lee and Winer 2005). of connections and (2) a series of regional feed-
Connections with auditory cortex in the opposite back loops (Warr 1992).
hemisphere are concentrated in the homotopic The descending chain comprises three main
(matching) area, (Wallace and Harper 1997) and levels. The first level originates in the AC and
heterotopic connections are relatively weak. The includes corticothalamic, corticotectal, and
main callosal projections arise from both pyrami- corticopontine projections (Fig. 16). The
dal and nonpyramidal cells in layers III and corticothalamic projection forms reciprocal and
V. Layers V and VI represent the main source nonreciprocal connections between the AC and
of descending projections to the MGB and IC and MGB. The corticotectal projection terminates
brainstem (Fig. 15). in the IC and subcollicular nuclei (Feliciano
and Potashner 1995; Meltzer and Ryugo 2006).
The Belt Region of Auditory Cortex The second level of the chain originates in the
Areas that lie outside of the core region are often IC, whose descending fibers form colliculo-
referred to as the nonlemniscal or belt areas olivary and colliculo-cochleonuclear projections
(Fig. 13). These areas are anatomically and phys- (Fig. 15). The colliculo-olivary fibers terminate
iologically distinct from the core fields and from on the periolivary medial olivocochlear cells
one another. Architectonically, each of the belt which supply the OHCs (Caicedo and Herbert
areas is distinct, but compared to the core region, 1993; Faye-Lund 1985; Vetter et al. 1993).
cell density and myelination are generally They may also terminate on lateral olivocochlear
reduced, as is the expression of cytochrome oxi- cells, which supply the IHCs (Feliciano and
dase, acetylcholinesterase, and parvalbumin Potashner 1995). The last and third level of the
(Jones 2003). chain consists of the olivocochlear system that
In addition to the inputs from the core, the provides efferent innervation to the cochlea
main source of projections to most of the belt (Figs. 15 and 19).
areas is the MGD (Jones 2007; Winer 1985; The regional feedback loops consist of a series
Winer and Lee 2007). Additional connections of cortical projections to subcortical nuclei that
include the MGM and thalamic nuclei that sur- project back to cortex, directly or indirectly,
round the MGB. Belt areas tend to differ with allowing AC to control its inputs from lower
respect to the balance of inputs from different centers.
thalamic nuclei. In nonhuman primates, an addi-
tional group of areas has been identified that The Corticofugal Pathways
surround the belt areas adjacent to the core. This The major AC projections target the MGB and IC
region is known as the parabelt (Kaas and (Figs. 15–18), but there are also AC projections
Hackett 1998). The parabelt region receives to subcollicular nuclei, including the nucleus
inputs from the belt region and MGD, but not sagulum, paralemniscal regions, superior olivary
the core. Thus, some ascending information complex, cochlear nuclear complex, and pontine
appears to pass serially through the core and nuclei. The AC also supplies the amygdala, basal
Anatomy and Physiology of the Mammalian Audi- however, the cortical neurons projecting to more distant
tory System, Fig. 16 Diagrammatic summary of the targets are more narrowly distributed and centered in
laminar organization of cortical cells projecting to the deeper regions of layer V (Figure kindly provided by
inferior colliculus (IC), superior olivary complex (SOC), Dr. D. K. Ryugo. Reproduced from Doucet et al. 2003)
and cochlear nucleus (CN). All three distributions overlap;
ganglia, striatum, superior colliculus, and central types of terminal synaptic boutons (I and II) aris-
gray (reviewed in Malmierca and Ryugo 2011, ing from the core of the AC (Fig. 17; Bartlett
2012), suggesting that the AC has important et al. 2000; Llano and Sherman 2009). Type
roles in addition to auditory sensory processing I terminals are small (<1 mm in long-axis diam-
(Winer 2006). eter), synapse on small caliber dendrites of the
In the auditory thalamus, the MGV receives its MGV and MGD, and arise from the pyramidal
heaviest source of input from the primary AC, cells of layer VI. Type II terminals are large
while the MGM the least and the MGD receives (>2 mm in long-axis diameter), mostly synapse
an intermediate amount. A prominent feature of in the MGD and occasionally in the MGM
the corticothalamic projection is reciprocity, in (Bartlett et al. 2000; Llano and Sherman 2009)
which a cortical region projects to the part of the and arise from pyramidal neurons from layer V.
thalamus from which it receives input. However, Sherman and Guillery (1996) first proposed
there are also nonoverlapping regions (Llano and the notion of “drivers” and “modulators” of tha-
Sherman 2008). In the MGB, there are two main lamic neurons in the visual and somatosensory
As in the MGB, AC may modulate the
processing of sounds in the IC through the acti-
vation of local inhibitory connections within the A
IC. Several studies have shown direct neocortical
projections to regions surrounding the lateral
lemniscus, including the nucleus sagulum ipsilat-
erally and the SOC and CNC bilaterally
(Feliciano and Potashner 1995; Weedman
et al. 1996; Weedman and Ryugo 1996a, b).
The Colliculofugal Pathways

Studies in several species have shown that the IC
has descending projections to the LL, SOC, and
CNC (Fig. 15; reviewed in Malmierca and
Anatomy and Physiology of the Mammalian Audi-
tory System, Fig. 17 Diagram of a cortico- Ryugo 2011, 2012). Perhaps, the most interesting
thalamocortical model of cortical processing. Information projections are the colliculo-olivary projections
reaches the lemniscal AC via a projection from the MGV that originate in the CNIC and LCIC because
to layer IV of either the AAF or the AI. From here, a layer they terminate on the VNTB, the site of origin
5 pyramidal projects to the MGD, where a thalamocortical
relay cell projects to layer IV of the nonlemniscal AC of the MOC (White and Warr 1983), suggesting
(DP or AII). From the nonlemniscal AC, a layer VI pro- that the IC influences MOC neurons (Vetter
jection is sent to the MGB, adhering to the principle that et al. 1993). The IC also projects to nonauditory
all thalamocortical projections, whether coming from the nuclei including the pontine nuclei, lateral
first- or higher-order thalamus, receive a modulator, recip-
rocal projection from layer VI (Figure kindly provided by paragigantocellular nucleus, gigantocellular
Dr. Daniel Llano. Adapted from Llano and Sherman 2008) reticular nucleus, ventrolateral tegmental
nucleus, and caudal pontine reticular nucleus
(Caicedo and Herbert 1993).
thalamus, but this hypothesis has been applied to
the auditory system as well (Llano and Sherman The Olivocochlear System
2008). According to this theory, type I terminals The olivocochlear bundle provides the organ of
play a modulatory role in the first-order thalamic Corti with efferent innervation (Figs. 15 and 19;
nuclei, such as the MGV (Fig. 17). Thus, the Rasmussen 1946) and plays a critical role in
corticothalamic inputs converge with ascending maintaining the normal operation of the cochlea
inputs on thalamic neurons such that the ascend- (Figs. 15 and 19). It may introduce nonlinear
ing inputs drive the thalamic neurons and the dynamics into the auditory system (Eggermont
cortical inputs modulate them. In contrast, in 2001). There are two systems of olivocochlear
the “higher-order” thalamic nuclei, such as the neurons, medial (MOC) and lateral (LOC). The
MGD, the “driver” inputs arise from the large MOC neurons are located medial and ventral to
type II axons and terminals that originate from the main nuclei of the SOC in the PO nucleus
the cortex and interact with other ascending input known as the VNTB and project mainly to the
from the IC. contralateral cochlea, whereas the LOC neurons
The corticofugal projection is glutamatergic are located within or near the LSO and project to
and modulates the MGB responses to sound the ipsilateral cochlea (Warr 1992; White and
through a direct excitatory pathway, but the AC Warr 1983).
can also provide the MGB with an inhibitory The MOC neurons constitute a homogeneous
influence (Bartlett et al. 2000) via its projections population of cholinergic cells and innervate the
to the auditory sector of the thalamic reticular OHCs (Figs. 15 and 19). About 75 % of them
nucleus which, in turn, projects to the MGB. originate on the contralateral side with the
Anatomy and Physiology of the Mammalian Audi- laminar organization of cortical cells (cortical surface is
tory System, Fig. 18 Retrograde labeling in area Te1 towards the top) projecting to the CNC (blue) vs those
for a rat that received an injection of Fast blue (FB) into targeting the IC (yellow) within layer V. (d and e) Exam-
the CNC and Diamidino yellow (DiY) into the IC (top). (a) ples of labeled cortical cells. Cortical cells that contained
Sagittal view of the brain, the gray line through AC both dyes (blue and yellow arrow in d) were observed
indicates the position of the cells along the rostral/caudal much less frequently (Figure kindly provided by Dr. D. K.
axis display of the location of labeled cortical cells shown Ryugo). Reproduced from Doucet et al. 2003). Abbrevia-
in panels (b–e). (b) Photomontage of layer V with few tions in the figure: CN Cochlear nucleus, CIC Central
FB-labeled cells located in deep layer V whereas the nucleus of the inferior colliculus, D Dorsal, M Medial,
DiY-labeled neurons are distributed more broadly. (c) Te1 Temporal area 1, Te2 Temporal area 2, Te3 Temporal
Higher magnification photomontage illustrating the area 3
Anatomy and Physiology of the Mammalian Audi- CNC. In rodents, two types of LOC cells occur: intrinsic
tory System, Fig. 19 Scheme of the LOC and MOC located inside the LSO and shell located in the margins of
neurons, their projections to the IHCs and OHCs, respec- the LSO (Figure kindly provided by Dr. Bruce Warr.
tively, and afferent fiber types I and II projecting to the Adapted from Warr 1992)
remainder originating ipsilaterally (Warr 1992; The functional role of the MOC neurons
Brown and Levine 2008). They probably receive may be to enhance transduction or signal detec-
descending input from the ipsilateral IC and tion through an unmasking effect, thus regulating
ascending input bilaterally from the VCN. the slow motility of the OHCs and thereby the
In rodents, the LOC neurons (Figs. 15 and 19) stiffness of the basilar membrane (Eggermont
consist of two distinct types of neurons: intrinsic 2001). Another function may be to protect the
and shell neurons (Vetter and Mugnaini 1992; inner ear from acoustic injury (Taranda
Warr et al. 1997). Intrinsic neurons are confined et al. 2009). Although the functional role of the
to the ipsilateral LSO and constitute about 85 % of LOC is uncertain, several studies (Safieddine
all LOC neurons (Brown and Levine 2008; Vetter et al. 1997; Safieddine and Eybalin 1992) have
and Mugnaini 1992; Warr et al. 1997). Most shell shown that these neurons may provide
neurons surround the ipsilateral LSO and consti- a modulatory effect to the afferent fibers that
tute the remaining 15 % (Vetter and Mugnaini contact the IHCs.
1992; Brown and Levine 2008; Warr et al. 1997).
(In contrast, in the cat the LOC neurons surround Acknowledgments I am most thankful and indebted to
the LSO, and two populations have not been dis- Dr. Nell Cant for her suggestions and constructive criti-
tinguished.) The LOC neurons innervate the type cisms on a previous version of the manuscript.
Financial support was provided by Spanish MINECO
I primary afferent fibers near the region where they (BFU2009-07286) and EU (EUI2009-04083, in the frame-
contact the IHCs. Virtually all of them originate on work of the ERA-NET Network of European Funding for
the ipsilateral side (Brown and Levine 2008). Neuroscience Research).
Cross-References auditory and spinal somatosensory afferents and their

interactions. J Neurophysiol 41:837–847
Anderson LA, Malmierca MS, Wallace MN, Palmer AR
▶ Acoustic Timbre Recognition (2006) Evidence for a direct, short latency projection
▶ Associations and Rewards in Auditory Cortex from the dorsal cochlear nucleus to the auditory thal-
▶ Attentional Top-Down Modulation, Models of amus in the guinea pig. Eur J Neurosci 24:491–498
▶ Auditory Event-Related Potentials Banks MI, Smith PH (1992) Intracellular recordings from
neurobiotin-labeled cells in brain slices of the rat
▶ Auditory Evoked Brainstem Responses medial nucleus of the trapezoid body. J Neurosci
▶ Auditory Memory 12:2819–2837
▶ Auditory Nerve Response, Afferent Signals Bartlett EL, Smith PH (1999) Anatomic, intrinsic, and
▶ Auditory Perceptual Organization synaptic properties of dorsal and ventral division neu-
rons in rat medial geniculate body. J Neurophysiol
▶ Auditory Precedence Effect 81:1999–2016
▶ Auditory Processing in Insects Bartlett EL, Stark JM, Guillery RW, Smith PH
▶ Auditory Prosthesis (2000) Comparison of the fine structure of cortical
▶ Auditory Sensing Systems: Overview and collicular terminals in the rat medial geniculate
body. Neuroscience 100:811–828
▶ Auditory Thalamocortical Transformations Blaesse P, Ehrhardt S, Friauf E, Nothwang HG
▶ Brainstem Processing: Overview (2005) Developmental pattern of three vesicular glu-
▶ Cochlear Inner Hair Cell, Model tamate transporters in the rat superior olivary complex.
▶ Context-Dependent Processing in Auditory Cell Tissue Res 320:33–50
Brawer JR, Morest DK, Kane EC (1974) The neuronal
Cortex architecture of the cochlear nucleus of the cat. J Comp
▶ Cortical Maps, Activity-Dependent Neurol 155:251–300
Development Brodal A (1981) Neurological anatomy in relation to
▶ Local Field Potentials (LFP) clinical medicine, 3rd edn. Oxford University Press,
Oxford
▶ Music Processing in the Brain Brown MC, Levine JL (2008) Dendrites of medial
▶ Neural Coding of Speech Sounds olivocochlear neurons in mouse. Neuroscience
▶ Neuromorphic Sensors, Cochlea 154:147–159
▶ Peripheral Nerve Interface Applications, Brown MC, Liberman MC, Benson TE, Ryugo DK
(1988) Brainstem branches from olivocochlear axons
Cochlear Implants in cats and rodents. J Comp Neurol 278:591–603
▶ Pitch Perception, Models Burton H, Jones EG (1976) The posterior thalamic region
▶ Receptive Field Modeling and its cortical projection in New World and Old
▶ Sensory Coding, Efficiency World monkeys. J Comp Neurol 168:249–301
Caicedo A, Herbert H (1993) Topography of descending
▶ Sound Localization and Experience- projections from the inferior colliculus to auditory
Dependent Plasticity brainstem nuclei in the rat. J Comp Neurol
▶ Sound Localization in Mammals, Models 328:377–392
▶ Spectrotemporal Receptive Fields Cant NB, Benson CG (2003) Parallel auditory pathways:
projection patterns of the different neuronal
▶ Spike-Frequency Adaptation populations in the dorsal and ventral cochlear nuclei.
▶ Stimulus-Specific Adaptation, Models Brain Res Bull 60:457–474
▶ Stimulus-Specific Information Cant NB, Benson CG (2006) Organization of the inferior
▶ Tinnitus, Models colliculus of the gerbil (Meriones unguiculatus): dif-
ferences in distribution of projections from the
cochlear nuclei and the superior olivary complex.
J Comp Neurol 495:511–528
References Cant NB, Benson CG (2007) Multiple topographically
organized projections connect the central nucleus of
Adams JC (1979) Ascending projections to the inferior the inferior colliculus to the ventral division of the
colliculus. J Comp Neurol 183:519–538 medial geniculate nucleus in the gerbil, Meriones
Adams JC (1983) Cytology of periolivary cells and the unguiculatus. J Comp Neurol 503:432–453
organization of their projections in the cat. J Comp Casseday JH, Fremouw T, Covey E (2002) The inferior
Neurol 215:2752–2789 colliculus, a hub for the central audiotory system. In:
Aitkin LM, Dickhaus H, Schult W, Zimmermann Oertel D, Popper AN, Fay RR (eds) Springer handbook
M (1978) External nucleus of inferior colliculus, of auditory research. Springer, New York
Cotillon N, Nafati M, Edeline JM (1999) Characteristics Hefti BJ, Smith PH (2003) Distribution and kinetic proper-
of reliable tone-evoke oscillations in the rat thalamo- ties of GABAergic inputs to layer V pyramidal cells in
cortical auditory system. Hear Res 142:113–130 rat auditory cortex. J Assoc Res Otolaryngol 4:106–121
Covey E, Casseday JH (1991) The monoaural nuclei of the Held H (1893) Die centralem Bahnen des Nervus A
lateral lemniscus in an echolocating bat, parallel path- acusticus bei der. Katz Arch Anat Abtheil 15:190–271
ways for analyzing temporal features of sound. Helfert RH, Aschoff A (1997) Superior olivary complex
J Neurosci 11:3456–3470 and nuclei of the lateral lemniscus. In: Ehret G,
Dehmel S, Kopp-Scheinpflug C, Dörrscheidt GJ, Romand R (eds) Anatomical and functional aspects
R€ubsamen R (2002) Electrophysiological characteri- of the cochlear nucleus. Oxford University Press,
zation of the superior paraolivary nucleus in the Mon- Oxford, pp 193–257
golian gerbil. Hear Res 172:18–36 Herbert H, Aschoff A, Ostwald J (1991) Topography of
Doron NN, Ledoux JE (1999) Organization of projections projections from the auditory cortex to the inferior
to the lateral amygdala from auditory and visual areas colliculus in the rat. J Comp Neurol 304:103–122
of the thalamus in the rat. J Comp Neurol 412:383–409 Irvine DRF (1992) Physiology of the auditory brainstem.
Doucet JR, Ryugo DK (1997) Projections from the ventral In: Popper AN, Fay RR (eds) Springer handbook of
cochlear nucleus to the dorsal cochlear nucleus in rats. auditory pathway, neurophysiology. Springer, New
J Comp Neurol 385:245–264 York, pp 153–231
Doucet JR, Ross AT, Gillespie MB, Ryugo DK (1999) Gly- Ito T, Oliver DL (2012) Patterns of synaptic organization
cine immunoreactivity of multipolar neurons in the send different messages to the thalamus. Front Neural
ventral cochlear nucleus which project to the dorsal Circuits 6:48. doi:103389/fncir201200048
cochlear nucleus. J Comp Neurol 408:515–531 Jones EG (2003) Chemically defined parallel pathways in
Doucet JR, Molavi DL, Ryugo DK (2003) The source of the monkey auditory system. Ann N Y Acad Sci
corticocollicular and corticobulbar projections in area 999:218–233
Te1 of the rat. Exp Brain Res 4:461–466 Jones EG (2007) The thalamus, vol II, 2nd edn. Cam-
Echteler SM, Fay RR, Popper AN (1994) Structure of the bridge University Press, Cambridge, pp 875–923
mammalian cochlea. In: Fay RR, Popper AN (eds) Joris PX, Smith PH, Yin TCT (1998) Coincidence detec-
Comparative hearing, mammals. Springer, Berlin, tion in the auditory system, 50 years after Jeffress.
pp 134–171 Neuron 21:1235–1238
Eggermont JJ (2001) Between sound and perception, Kaas JH, Hackett TA (1998) Subdivisions of auditory
reviewing the search for a neural code. Hear Res cortex and levels of processing in primates. Audiol
157:1–42 Neurootol 3:73–85
Fay RR (1988) Hearing in vertebrates: a psychophysics Kasper EM, Larkman AU, Lubke J, Blakemore C (1994)
databook. Hill-Fay Associates, Winnetka Pyramidal neurons in layer 5 of the rat visual cortex.
Fay RR, Popper AN (1994) Comparative hearing in mam- I. Correlation among cell morphology, intrinsic elec-
mals. In: Fay RR and Popper AN (Eds), Springer trophysiological properties, and axon targets. J Comp
handbook of auditory research. Springer, New York Neurol 339:459–474
Faye-Lund H (1985) The neocortical projection to the Kawaguchi Y (1993) Groupings of nonpyramidal and
inferior colliculus in the albino rat. Anat Embryol pyramidal cells with specific physiological and mor-
(Berl) 173:53–70 phological characteristics in rat frontal cortex.
Faye-Lund H, Osen KK (1985) Anatomic of the inferior J Neurophysiol 69:416–431
colliculus in rat. Anat Embryol 175:35–52 Kelly JB, Caspary DM (2005) Pharmacology of the infe-
Feliciano M, Potashner SJ (1995) Evidence for rior colliculus. In: Winer JA, Schreiner C (eds) The
a glutamatergic pathway from the guinea pig auditory inferior colliculus. Springer, New York, pp 248–281
cortex to the inferior colliculus. J Neurochem King AJ, Jiang ZD, Moore DR (1998) Auditory brainstem
65:1348–1357 projections to the ferret superior colliculus: anatomical
Friauf E (1993) Transient appearance of calbindin-D28k- contribution to the neural coding of sound azimuth.
positive neurons in the superior olivary complex of J Comp Neurol 390:342–365
developing rats. J Comp Neurol 334:59–74 Kopp-Scheinpflug C, Tolnai S, Malmierca MS, R€ ubsamen
Games KD, Winer JA (1988) Layer V in rat auditory R (2008) The medial nucleus of the trapezoid body:
cortex, projections to the inferior colliculus and con- comparative physiology. Neuroscience 154:60–170
tralateral cortex. Hear Res 34:1–25 Kulesza RJ Jr, Spirou GA, Berrebi AS (2003) Physiologi-
Guinan JJ, Norris BE, Guinan SS (1972) Single auditory cal response properties of neurons in the superior
units in the superior olivary complex, II. Locations of paraolivary nucleus of the rat. J Neurophysiol
unit categories and tonotopic organization. Int 89:2299–2312
J Neurosci 4:147–166 LeBeau FEN, Malmierca MS, Rees A (2001) Iontophore-
Hefti BJ, Smith PH (2000) Anatomy, physiology, and sis in vivo demonstrates a key role for GABAA- and
synaptic responses of rat layer V auditory cortical glycinergic inhibition in shaping frequency response
cells and effects of intracellular GABA(A)blockade. areas in the inferior colliculus of guinea pig. J Neurosci
J Neurophysiol 83:2626–2638 21:7303–7312
Lee CC, Winer JA (2005) Principles governing auditory Malmierca MS, Merchán M, Henkel CK, Oliver DL
cortex connections. Cereb Cortex 15:1804–1814 (2002) Direct projections from the dorsal cochlear
Liberman MC, Dodds LW, Pierce S (1990) Afferent and nucleus to the auditory thalamus in rat. J Neurosci
efferent innervation of the cat cochlea, quantitative 22:10891–10897
analysis with light and electron microscopy. J Comp Malmierca MS, Saint Marie RL, Merchan MA, Oliver DL
Neurol 301:443–460 (2005) Laminar inputs from dorsal cochlear nucleus
Llano DA, Sherman SM (2008) Evidence for and ventral cochlear nucleus to the central nucleus of
nonreciprocal organization of the mouse auditory the inferior colliculus: two patterns of convergence.
thalamocortical-corticothalamic projection systems. Neuroscience 136:883–894
J Comp Neurol 507:1209–1227 Malmierca MS, Izquierdo MA, Cristaudo S, Hernández O,
Llano DA, Sherman SM (2009) Differences in intrinsic Pérez-González D, Covey E, Oliver DL
properties and local network connectivity of identified (2008) A discontinuous tonotopic organization in
layer 5 and layer 6 adult mouse auditory the inferior colliculus of the rat. J Neurosci
corticothalamic neurons support a dual 28:4767–4776
corticothalamic projection hypothesis. Cereb Cortex Malmierca MS, Cristaudo S, Pérez-González D, Covey
19:2810–2826 E (2009) Stimulus-specific adaptation in the inferior
Loftus B, Malmierca MS, Oliver DL (2008) The colliculus of the anesthetized rat. J Neurosci
Cytoarchitecture of the inferior colliculus revisited: 29:5483–5493
a common organization of the lateral cortex in rat Meltzer NE, Ryugo DK (2006) Projections from auditory
and cat. Neuroscience 154:196–205 cortex to cochlear nucleus: a comparative analysis of
Lorente de Nó R (1981) The primary acoustic nuclei. rat and mouse. Anat Rec A Discov Mol Cell Evol Biol
Raven Press, New York 288:397–408
Lu E, Llano DA, Sherman SM (2009) Different distribu- Merchán M, Aguilar LA, Lopez-Poveda EA, Malmierca
tions of calbindin and calretinin immunostaining MS (2005) The inferior colliculus of the rat: quantita-
across the medial and dorsal divisions of the mouse tive immunocytochemical study of GABA and gly-
medial geniculate body. Hear Res 257:16–23 cine. Neuroscience 136:907–925
Malmierca MS (2003) The structure and physiology of the Moore JK, Osen KK (1979) The human cochlear nuclei.
rat auditory system: an overview. Int Rev Neurobiol In: Creutzfeld O, Scheich H, Schreiner C (eds) Exp
56:147–211 Brain Res, Suppementum II, pp 36–44
Malmierca MS, Hackett TA (2010) Structural organiza- Morest DK (1968) The collateral system of the medial
tion of the ascending auditory pathway. In: Moore DR nucleus of the trapezoid body of the cat, its neuronal
(ed) The Oxford handbook of auditory science: the architecture and relation to the olivo-cochlear bundle.
auditory brain. Oxford University Press, New York, Brain Res 9:288–311
pp 9–41 Mugnaini E, Osen KK, Dahl A-L, Friedrich VL Jr, Korte
Malmierca MS, Ryugo DK (2011) Descending connec- G (1980a) Fine structure of granule cells and related
tions of auditory cortex to the midbrain and brainstem. interneurons (termed Golgi cells) in the cochlear
In: Winer JA, Schreiner CE (eds) The auditory cortex. nuclear complex of cat, rat and mouse. J Neurocytol
Springer, New York, pp 189–208 9:537–570
Malmierca MS, Ryugo DK (2012) Auditory system. Mugnaini E, Warr WB, Osen KK (1980b) Distribution and
In: Watson C, Paxinos G, Puelles L (eds) The light microscopic features of granule cells in the
mouse nervous system. Academic, Amsterdam, cochlear nuclei of cat, rat, and mouse. J Comp Neurol
pp 607–645 191:581–606
Malmierca MS, Blackstad TW, Osen KK, Karag€ ulle T, Oertel D (1999) The role of timing in the brain stem
Molowny RL (1993) The central nucleus of the infe- auditory nuclei of vertebrates. Annu Rev Physiol
rior colliculus in rat, a Golgi and computer reconstruc- 61:497–519
tion study of neuronal and laminar structure. J Comp Oertel D, Young ED (2004) What’s a cerebellar circuit
Neurol 333:1–27 doing in the auditory system? Trends Neurosci
Malmierca MS, Rees A, LeBeau FEN, Bajaalie JG 27:104–110
(1995) Laminar organization of frequency-defined Oertel D, Wu SH, Garb MW, Dizack C (1990) Morphol-
local axons within and between the inferior colliculi ogy and physiology of cells in slice preparations of the
of the guinea pig. J Comp Neurol 357:124–144 posteroventral cochlear nucleus of mice. J Comp
Malmierca MS, LeBeau FEN, Rees A (1996) The topo- Neurol 295:136–154
graphical organization of descending projections from Oliver DL, Morest DK (1984) The central nucleus of the
the central nucleus of the inferior colliculus in guinea inferior colliculus in the cat. J Comp Neurol
pig. Hear Res 93:167–180 222:237–264
Malmierca MS, Leergard TB, Bajo VM, Bjaalie JG Oliver DL, Ostapoff EM, Beckius GE (1999) Direct inner-
(1998) Anatomic evidence of a 3-D mosaic pattern of vation of identified tectothalamic neurons in the infe-
tonotopic organization in the ventral complex of the rior colliculus by axons from the cochlear nucleus.
lateral lemniscus in cat. J Neurosci 19:10603–10618 Neuroscience 93:643–658
Osen KK (1969) Cytoarchitecture of the cochlear nuclei in Saldaña E, Merchán MA (1992) Intrinsic and commissural
the cat. J Comp Neurol 136:453–483 connections of the rat inferior colliculus. J Comp
Osen KK, Mugnaini E, Dahl AL, Christiansen AH Neurol 319:417–437
(1984) Histochemical localization of acetylcholines- Saldaña E, Feliciano M, Mugnaini E (1996) Distribution A
terase in the cochlear and superior olivary nuclei. of descending projections from primary auditory neo-
A reappraisal with emphasis on the cochlear granule cortex to inferior colliculus mimics the topography of
cell system. Arch Ital Biol 122:169–212 intracollicular projections. J Comp Neurol 371:15–40
Osen KK, Ottersen OP, Størm-Mathisen J (1990) Schofield BR (1995) Projections from the cochlear
Colocalization of glicine-like and GABA-like nucleus to the superior paraolivary nucleus in guinea
inmunoreactivities, a semiquantitative study of indi- pigs. J Comp Neurol 360:135–149
vidual neurons in the dorsal cochlear nucleus of cat. In: Schofield BR, Cant NB (1991) Organization of the supe-
Ottersen OP, Størm-Mathissen J (eds) Glycine neuro- rior olivary complex in the guinea pig.
transmission. Wiley, Chichester, pp 417–451 I. Cytoarchitecture, cytochrome oxidase histochemis-
Ota Y, Oliver DL, Dolan DF (2004) Frequency-specific try, and dendritic morphology. J Comp Neurol
effects on cochlear responses during activation of the 314:645–670
inferior colliculus in the Guinea pig. J Neurophysiol Schreiner CE, Langner G (1988) Coding of temporal
91:2185–2193 patterns in the central auditory nervous system. In:
Palombi PS, Caspary DM (1996) Responses of young and Edelman GM, Gall WE, Cowan WM (eds) Auditory
aged Fischer 344 rat inferior colliculus neurons to function. Wiley, New York, pp 337–340
binaural tonal stimuli. Hear Res 100:59–67 Schreiner CE, Langner G (1997) Laminar fine structure of
Peruzzi D, Bartlett E, Smith PH, Oliver DL frequency organization in auditory midbrain. Nature
(1997) A monosynaptic GABAergic input from the 388:383–386
inferior colliculus to the medial geniculate body in Sherman SM, Guillery RW (1996) Functional organiza-
rat. J Neurosci 17:3766–3777 tion of thalamocortical relays. J Neurophysiol
Rajan R (1990) Electrical stimulation of the inferior 76:1367–1395
colliculus at low rates protects the cochlea from audi- Sinex DG, Lopez DE, Warr WB (2001) Electrophysiolog-
tory desensitization. Brain Res 506:192–204 ical responses of cochlear root neurons. Hear Res
Rasmussen GL (1946) The olivary peduncle and other 370:1–11
fiber projections of the superior olivary complex. Sivaramakrishnan S, Oliver DL (2001) Distinct K+ cur-
J Comp Neurol 99:61–74 rents result in physiologically distinct cell types in the
Rietzel HJ, Friauf E (1998) Neuron types in the rat lateral inferior colliculus of the rat. J Neurosci 21:2861–2877
superior olive and developmental changes in the com- Slepecky NB (1996) Structure of the mammalian cochlea.
plexity of their dendritic arbors. J Comp Neurol In: Dallos P, Popper AN, Fay RR (eds) The cochlea,
390:20–40 Springer handbook of auditory research. Springer,
Riquelme R, Saldaña E, Osen KK, Ottersen OP, Merchán New York, pp 44–129
MA (2001) Colocalization of GABA and Glycine in Smith PH (1995) Structural and functional differences
the ventral nucleus of the lateral lemniscus in rat, an in distinguish principal from nonprincipal cells in the
situ hybridization and semiquantitative inmunocyto- guinea pig MSO slice. J Neurophysiol 73:1653–1667
chemical study. J Comp Neurol 432:409–424 Smith PH, Rhode WS (1989) Structural and functional
Rubio ME, Juiz JM (2004) Differential distribution of properties distinguish two types of multipolar cells in
synaptic endings containing glutamate, glycine, and the ventral cochlear nucleus. J Comp Neurol
GABA in the rat dorsal cochlear nucleus. J Comp 282:595–616
Neurol 477:253–272 Smith PH, Joris PX, Carney LH, Yin TCT (1991) Projec-
Rubio ME, Gudsnuk KA, Smith Y, Ryugo DK tions of physiologically characterized globular bushy
(2008) Revealing the molecular layer of the primate cell axons from the cochlear nucleus of the cat. J Comp
dorsal cochlear nucleus. Neuroscience 154:99–113 Neurol 304:387–407
Ryugo DK, Parks TN (2003) Primary innervation of the Sommer I, Lingenhöhl K, Friauf E (1993) Principal cells
avian and mammalian cochlear nucleus. Brain Res of the rat medial nucleus of the trapezoid body, an
Bull 60:435–456 intracellular in vivo study of their physiology and
Safieddine S, Eybalin M (1992) Triple immunofluores- morphology. Exp Brain Res 95:223–239
cence evidence for the coexistence of acetylcholine, Taranda J, Maison SF, Ballestero JA, Katz E, Savino J,
enkephalins, and calcitonin gene-related peptide Vetter DE, Boulter J, Liberman MC, Fuchs PA,
within efferent olivocochlear neurons in rats and Elgoyhen AB (2009) A point mutation in the hair cell
guinea-pigs. Eur J Neurosci 4:981–992 nicotinic cholinergic receptor prolongs cochlear inhi-
Safieddine S, Prior AM, Eybalin M (1997) Choline bition and enhances noise protection. PLoS Biol. doi:
acetyltransferase, glutamate decarboxylase, tyrosine 101371/journalpbio1000018
hydroxylase, calcitonin gene-related peptide and opi- Thompson AM, Schofield BR (2000) Afferent projections
oid peptides coexist in lateral efferent neurons of rat of the superior olivary complex. Microsc Res Tech
and guinea-pig. Eur J Neurosci 9:356–367 51:330–354
A 186 Anesthesia, Neural Population Models of
Vetter DE, Mugnaini E (1992) Distribution and dendritic Ye Y, Machado DG, Kim DO (2000) Projection of the
features of three groups of rat olivocochlear neurons. marginal shell of the anteroventral cochlear nucleus to
A study with two retrograde cholera toxin tracers. Anat olivocochlear neurons in the cat. J Comp Neurol
Embryol 185:1–16 420:127–138
Vetter DE, Saldaña E, Mugnaini E (1993) Input from the Young ED, Davis KA (2002) Circuitry and function of
inferior colliculus to medial olivocochlear neurons in the dorsal cochlear nucleus, chapter 5. In: Oertel D,
the rat, a double label study with PHA-L and cholera Fay RR, Popper AN (eds) Springer handbook of audi-
toxin. Hear Res 70:173–186 tory research. Integrative functions in the mammalian
Wallace MN, Harper MS (1997) Callosal connections of auditory pathway, vol 15. Springer, New York,
the ferret primary auditory cortex. Exp Brain Res pp 160–206
116:367–374 Young ED, Shofner WP, White JA, Robert JM, Voigt HF
Warr WB (1992) Organization of olivocochlear efferent (1988) Response properties of cochlear nucleus neu-
systems in mammals. In: Webster DB, Popper AN, Fay rons in relationship to physiological mechanisms. In:
RR (eds) The mammalian auditory pathway, neuro- Edelman GM, Gall WE, Cowan WM (eds) Auditory
anatomy. Springer, Berlin, pp 410–448 function, neurobiological bases of hearing. Wiley,
Warr WB, Boche JB, Neely ST (1997) Efferent innervation New York, pp 277–312
of the inner hair cell region, origins and terminations of Yu XJ, Xu XX, He S, He J (2009) Change detection by
two lateral olivocochlear systems. Hear Res 108:89–111 thalamic reticular neurons. Nat Neurosci
Weedman DL, Ryugo DK (1996a) Projections from audi- 12:1165–1170
tory cortex to the cochlear nucleus in rats, synapses on Zhang H, Kelly JB (2006a) Responses of neurons in the
granule cell dendrites. J Comp Neurol 371:311–324 rat’s ventral nucleus of the lateral lemniscus to mon-
Weedman DL, Ryugo DK (1996b) Pyramidal cells in aural and binaural tone bursts. J Neurophysiol
primary auditory cortex project to cochlear nucleus in 95:2501–2512
rat. Brain Res 706:97–102 Zhang H, Kelly JB (2006b) Responses of neurons in the
Weedman DL, Pongstaporn T, Ryugo DK (1996) Ultra- rat’s ventral nucleus of the lateral lemniscus to
structural study of the granule cell domain of the amplitude-modulated tones. J Neurophysiol
cochlear nucleus in rats: mossy fiber endings and 96:2905–2914
their targets. J Comp Neurol 369:345–360 Zhang Y, Wu SH (2000) Long-term potentiation in the
White JS, Warr WB (1983) The dual origins of the inferior colliculus studied in rat brain slice. Hear Res
olivocochlear bundle in the albino rat. J Comp Neurol 1(47):92–103
219:203–214 Zhao M, Wu SH (2001) Morphology and physiology of
Wickesberg RE, Oertel D (1988) Tonotopic projection neurons in the ventral nucleus of the lateral lemniscus
from the dorsal to the anteroventral cochlear nucleus in rat brain slices. J Comp Neurol 433:255–271
of mice. J Comp Neurol 268:389–399 Zhou J, Shore S (2006) Convergence of spinal trigeminal
Winer JA (1985) The medial geniculate body of the cat. and cochlear nucleus projections in the inferior
Adv Anat Embryol Cell Biol 86:1–97 colliculus of the guinea pig. J Comp Neurol
Winer JA (1992) The functional architecture of the medial 495:100–112
geniculate body and the primary auditory cortex. In:
Webster DB, Popper AN, Fay RR (eds) The mamma-
lian auditory pathway, neuroanatomy. Springer, New
York, pp 222–409 Anesthesia, Neural Population
Winer JA (2006) Decoding the auditory corticofugal sys-
tems. Hear Res 212:1–8 Models of
Winer JA, Larue DT (1988) Anatomy of glutamic acid decar-
boxylase immunoreactive neurons and axons in the rat D. Alistair Steyn-Ross1, Moira Steyn-Ross1 and
medial geniculate body. J Comp Neurol 278:47–68 Jamie Sleigh2
Winer JA, Lee CC (2007) The distributed auditory cortex. 1
Hear Res 229:3–13 School of Engineering, University of Waikato,
Winer JA, Morest DK (1983) The medial division of the Hamilton, New Zealand
medial geniculate body of the cat: implications for 2
Waikato Clinical School, University of Auckland,
thalamic organization. J Neurosci 3:2629–2651 Waikato Hospital, Hamilton, New Zealand
Winer JA, Schreiner CE (2011) The auditory cortex.
Springer, New York
Winer JA, Kelly JB, Larue DT (1999) Neural architecture
of the rat medial geniculate body. Hear Res 31:19–41 Definition
Wu SH (1999) Physiological properties of neurons in the
ventral nucleus of the lateral lemniscus of the rat,
intrinsic membrane properties and synaptic responses. General anesthesia is a reversible, drug-induced
J Neurophysiol 81:2862–2874 state of unconsciousness characterized by lack of
Anesthesia, Neural Population Models of 187 A
awareness of surroundings, lack of responsive- pattern (bursting activity alternating with relative
ness to painful stimuli (nociception), and inabil- silence), and finally collapses into a flat-line trace
ity to form memories (amnesia). The change in at the deepest levels of comatose anesthesia. A
brain state from wakeful to unconscious produces This sequence is reversed as the anesthetic
alterations in cortical electrical activity that can drug is eliminated naturally from the body,
be monitored with electrodes placed on the scalp allowing the patient to return to consciousness.
(electroencephalogram (EEG)) or on the surface However, the fact that the recovery of respon-
of the cortex (electrocorticogram (ECoG)). The siveness generally occurs at a lower drug concen-
goal of neural modelers is to develop equations tration (as measured in the blood) than that
that describe the gross behavior of spatially aver- required to induce unresponsiveness suggests
aged populations of neurons during both induc- a hysteresis separation between induction and
tion of and recovery from general anesthesia. recovery trajectories. Part of this hysteresis can
be explained in terms of the time required for the
drug to diffuse across the blood–brain barrier
Detailed Description (Voss et al. 2007) and so can be compensated
using pharmacokinetics models (Roberts 2007),
Classes of General Anesthesia but such compensations are typically only par-
There are two broad classes of anesthetic drugs: tially successful (Ludbrook et al. 1999; Coppens
inductive agents (such as propofol, etomidate, et al. 2010). The remaining hysteresis may be
isoflurane) that produce a slowed sleeplike EEG a consequence of a recently proposed “neural
and dissociative agents (e.g., ketamine, nitrous inertia” that resists transitions between conscious
oxide) that induce a dissociated state with an and unconscious states (Friedman et al. 2010);
activated EEG similar to that of REM sleep. such distinct induction/recovery paths arise natu-
Most commonly used intravenous and volatile rally if the brain has access to multiple steady
agents – such as propofol or sevoflurane – boost states as suggested by the modeling of Steyn-
inhibition by increasing the influx of chloride Ross et al. (1999, 2004).
ions at gamma-aminobutyric acid (GABA) recep-
tors on postsynaptic membranes (Weir 2006), Cellular Effects of General Anesthetic Drugs
causing the postsynaptic neuron to become Studies of propofol, halothane, and isoflurane
hyperpolarized. In contrast, dissociative drugs are have shown that, at drug concentrations render-
believed to disrupt excitatory synaptic transmis- ing human subjects unresponsive, cerebral blood
sion. In both cases, the excitatory–inhibitory flow and metabolism are reduced by about 50 %
balance required for normal brain function has (Antkowiak 2002) as a result of global reductions
been shifted to favor inhibition. in cortical activity. This is consistent both with
in vivo investigations in rat cortex – where
The Induction–Recovery Trajectory sedative-level concentrations were found to sup-
At low concentrations, most GABAergic agents press neural firing rates by 50–70 % (Gaese and
(e.g., propofol, sevoflurane, etomidate) cause Ostwald 2001) – and with cultured brain-slice
a paradoxical boost in cortical activity (called studies in which low concentrations of general
the “biphasic effect”) across most EEG frequency anesthetics (GABAergic agonists propofol, halo-
bands (Kuizenga et al. 2001), with the biphasic thane, isoflurane, enflurane, sevoflurane,
peak appearing first in the high beta frequencies etomidate, ethanol, and pentobarbital and the
(24–28 Hz), then sliding smoothly towards lower non-GABAergic agent ketamine) significantly
frequencies in time (e.g., see Fig. 3 of Koskinen decreased mean firing rates (Antkowiak 2002).
et al. (2005)). With further increase in concentra- All anesthetic drugs influence cellular func-
tion, the EEG slows as large-amplitude delta- tion in a number of different ways, but the
band oscillations (1–4 Hz) become dominant, major mechanism for GABAergic suppression
then changes to an intermittent burst–suppression of firing rates is believed to be the prolongation
of the opening of chloride channels on the post- Subsequent work by Bojak and Liley (2005)
synaptic neuron, thus causing a substantial on isoflurane anesthesia showed that, for suitable
increase in negative charge transfer (by a factor choices of cortical parameters, a smooth descent
of 2–4 times control at clinically relevant con- into unconsciousness can also generate a biphasic
centrations (Kitamura et al. 2003; Banks and drug response. Using an alternative mean-field
Pearce 1999)) during the inhibitory postsynaptic model, Hutt and colleagues (Hutt and
current (IPSC) pulse. Schimansky-Geier 2008; Hutt and Longtin
Dissociative drugs reduce excitatory transmis- 2010) predicted that biphasic power surges can
sion by blocking N-methyl-D-aspartate (NMDA) be expected for both the bistable (jump transition)
glutamate channels, which probably has and monostable (smooth) inductions of anesthesia.
a significant role in producing the characteristic General anesthetic agents are widely used to
dissociated anesthetic state (Petrenko et al. treat seizures, but paradoxically, some anesthetics
2013); however, these drugs also have other (e.g., enflurane) can also provoke cortical seizures
effects such as inhibition of hyperpolarization- when the patient is deeply anesthetized. Liley and
activated cyclic nucleotide-gated (HCN1) chan- Bojak (2005) and Wilson et al. (2006) used mean-
nels (Chen et al. 2009) or increased potassium field modeling to show that subtle changes in the
channel opening (Gruss et al. 2004). shape and duration of the drug-induced inhibitory
postsynaptic response can explain why enflurane,
Modeling Anesthetic Effects but not isoflurane, is seizurogenic.
The challenge for anesthesia modelers is to An important part of general anesthesia
bridge the scales from the microscopic cellular is the suppression of noxious stimuli. A
drug effects to the consequent macroscopic pop- practical index of antinociception has been
ulation behaviors detected with scalp or cortical developed from a mean-field model (Liley
electrodes. By considering spatially averaged et al. 2010) that informed construction of an
(“mean-field”) properties of cortical tissue, we autoregressive–moving-average (ARMA) noise-
can avoid the need (and computational expense) driven filter whose output approximates the
of attempting to explicitly represent myriads of scalp-recorded EEG. The mean filter frequency
individual neurons (as is done in neural net- tracks the level of propofol-induced hypnosis
works). There is a steadily growing interest in (“cortical state”), while the decrease in required
applying mean-field methods to the challenge of noise intensity (“cortical input”) tracks the con-
understanding anesthesia; see Foster et al. (2008) centration of a coadministered analgesic agent
and Steyn-Ross et al. (2011) for reviews. (remifentanil). This computed “cortical input”
The notion of neural fields dates from signal is presumed to be a measure of cortical
foundation work by Wilson and Cowan (1972) stimulus, both noxious and normal, entering from
that modeled the brain as homogenous the thalamus, and potentially allows differentia-
populations of excitatory and inhibitory neu- tion between hypnotic and analgesic drug effects.
rons. The first attempt at modeling propofol The unconscious state of anesthesia and of
anesthesia by Steyn-Ross et al. (1999) incorpo- deepest natural sleep are both characterized by
rated prolongation of inhibitory response into large-amplitude, slow (0.5–4 Hz) delta waves of
the mean-field neural model of Liley et al. EEG activity. The source of these slow waves is
(1999); it predicted the possibility of multiple unknown but is generally supposed to originate
steady states with distinct first-order phase tran- from gradual alternations in depolarizing and
sitions between activated (“conscious”) and hyperpolarizing ionic currents. By introducing
inactivated (“unconscious”) states and provided a slow ionic gating variable into a mean-field
a possible explanation for the hysteretically sep- model for desflurane anesthesia, Molaee-
arated biphasic power surges observed at loss Ardekani et al. (2007) demonstrated emergence
and recovery of consciousness (Kuizenga of realistic slow waves. A quite different
et al. 2001). slow-wave mechanism has been proposed by
Anesthesia, Neural Population Models of 189 A
Steyn-Ross et al. (2013): if inhibitory gap junc- References
tions are included in the two-dimensional cortical
sheet, then a Turing (pattern-forming) instability Antkowiak B (2002) In vitro networks: cortical mecha-
nisms of anaesthetic action. Br J Anaesth A
can interact with a weakly damped low-frequency
89(1):102–111
Hopf instability to produce turbulent slow-wave Banks MI, Pearce RA (1999) Dual actions of volatile
activity across the cortex. Anesthetic-induced clo- anesthetics on GABA(A) IPSCs: dissociation of
sure of inhibitory gap junctions (Wentlandt blocking and prolonging effects. Anesthesiology
90(1):120–134
et al. 2006) is predicted to weaken the Turing
Bojak I, Liley DT (2005) Modeling the effects of anesthe-
instability in favor of the Hopf oscillation. sia on the electroencephalogram. Phys Rev E 71(4 Pt
There has been interest in modeling some of 1):041902, URL: http://www.ncbi.nlm.nih.gov/
the specific details of EEG spectral changes pubmed/15903696
Bojak I, Day HC, Liley DT (2013) Ketamine, propofol,
caused by various general anesthetic drugs, for
and the EEG: a neural field analysis of HCN1-
example, the displacement in alpha peak fre- mediated interactions. Front Comput Neurosci 7:22.
quency induced by ketamine (Bojak et al. 2013) doi:10.3389/fncom.2013.00022, URL: http://www.
or propofol (Hindriks and van Putten 2012; Hutt ncbi.nlm.nih.gov/pubmed/23576979
Boly M, Moran R, Murphy M, Boveroux P, Bruno MA,
2013) and the burst–suppression pattern of deep
Noirhomme Q, Ledoux D, Bonhomme V, Brichant JF,
anesthesia (Liley and Walsh 2013). Tononi G, Laureys S, Friston K (2012) Connectivity
Increasingly there has been a realization that changes underlying spectral EEG changes during
general anesthesia may disrupt neuronal net- propofol-induced loss of consciousness. J Neurosci
32(20):7082–7090. doi10.1523/JNEUROSCI.3769-
works in an anatomically specific fashion
11.2012
(Kuhlmann et al. 2013; Lee et al. 2013) and that Chen X, Shu S, Bayliss DA (2009) HCN1 channel sub-
the current homogenous and isotropic neuronal units are a molecular substrate for hypnotic actions of
population models might need to include aspects ketamine. J Neurosci 29(3):600–609. doi:10.1523/
JNEUROSCI.3481-08.2009
of network topology. This has led to attempts to
Coppens M, Van Limmen JGM, Schnider T, Wyler B,
link EEG patterns probabilistically with underly- Bonte S, Dewaele F, Struys MMRF, Vereecke HEM
ing anesthetic effects on inhibitory and excitatory (2010) Study of the time course of the clinical effect of
neuronal groups – this should provide propofol compared with the time course of the
predicted effect-site concentration: performance of
a quantitative basis for the estimation of model
three pharmacokinetic-dynamic models. Br J Anaesth
parameters. At an abstract level, dynamic causal- 104(4):452–458. doi:10.1093/bja/aeq028
modeling methods have been employed (Moran Dadok VM, Kirsch HE, Sleigh JW, Lopour BA, Szeri AJ
et al. 2011; Boly et al. 2012), but a more direct (2013) A probabilistic framework for a physiological
representation of dynamically evolving sleep state.
Bayesian approach – which has been used for
J Comput Neurosci. doi:10.1007/s10827-013-0489-x
natural sleep (Dadok et al. 2013) – could be Foster BL, Bojak I, Liley DTJ (2008) Population based
applied to anesthesia EEG. models of cortical drug response: insights from anaes-
thesia. Cogn Neurodyn 2(4):283–296. doi:10.1007/
s11571-008-9063-z
Friedman EB, Sun Y, Moore JT, Hung HT, Meng QC,
Cross-References Perera P, Joiner WJ, Thomas SA, Eckenho RG,
Sehgal A, Kelz MB (2010) A conserved behavioral
▶ Bifurcations, Neural Population Models and state barrier impedes transitions between anesthetic-
induced unconsciousness and wakefulness: evidence
▶ Epilepsy, Neural Population Models of
for neural inertia. PLoS One 5(7):e11903.
▶ Gap Junctions in Small Networks doi:10.1371/journal.pone.0011903
▶ Gap Junctions, Neural Population Models and Gaese BH, Ostwald J (2001) Anesthesia changes fre-
▶ Neural Population Model quency tuning of neurons in the rat primary auditory
cortex. J Neurophysiol 86(2):1062–1066
▶ Pattern Formation in Neural Population
Gruss M, Bushell TJ, Bright DP, Lieb WR, Mathie A,
Models Franks NP (2004) Two-pore-domain K+ channels are
▶ Phase Transitions, Neural Population a novel target for the anesthetic gases xenon, nitrous
Models and oxide, and cyclopropane. Mol Pharmacol
65(2):443–452. doi:10.1124/mol.65.2.443
▶ Sleep, Neural Population Models of
Hindriks R, van Putten MJAM (2012) Meanfield modeling Ludbrook GL, Upton RN, Grant C, Martinez A (1999)
of propofol-induced changes in spontaneous EEG Prolonged dysequilibrium between blood and brain
rhythms. Neuroimage 60(4):2323–2334. doi:10.1016/ concentrations of propofol during infusions in sheep.
j.neuroimage.2012.02.042 Acta Anaesthesiol Scand 43(2):206–211
Hutt A (2013) The anesthetic propofol shifts the frequency Molaee-Ardekani B, Senhadji L, Shamsollahi MB,
of maximum spectral power in EEG during general Vosoughi-Vahdat B, Wodey E (2007) Brain activity
anesthesia: analytical insights from a linear model. modeling in general anesthesia: enhancing local mean-
Front Comput Neurosci 7:2. doi:10.3389/ field models using a slow adaptive firing rate. Phys
fncom.2013.00002, URL: http://www.ncbi.nlm.nih. Rev E 76(4 Pt 1):041911, URL: http://www.ncbi.nlm.
gov/pubmed/23386826 nih.gov/pubmed/17995030
Hutt A, Longtin A (2010) Effects of the anesthetic agent Moran RJ, Jung F, Kumagai T, Endepols H, Graf R, Dolan
propofol on neural populations. Cogn Neurodyn RJ, Friston KJ, Stephan KE, Tittge-meyer M (2011)
4(1):37–59. doi:10.1007/s11571-009-9092-2 Dynamic causal models and physiological inference:
Hutt A, Schimansky-Geier L (2008) Anesthetic-induced a validation study using isoflurane anaesthesia in
transitions by propofol modeled by nonlocal rodents. PLoS One 6(8):e22790. doi:10.1371/journal.
neural populations involving two neuron types. J Biol pone.0022790, URL: http://www.ncbi.nlm.nih.gov/
Phys 34(3–4):433–440. doi:10.1007/s10867-008- pubmed/21829652
9065-4 Petrenko AB, Yamakura T, Sakimura K, Baba H (2013)
Kitamura A, Marszalec W, Yeh JZ, Narahashi T (2003) Defining the role of NMDA receptors in anesthesia:
Effects of halothane and propofol on excitatory and are we there yet? Eur J Pharmacol 723C:29–37.
inhibitory synaptic transmission in rat cortical neu- doi:10.1016/j.ejphar.2013.11.039
rons. J Pharmacol Exp Ther 304(1):162–171. Roberts F (2007) Pharmacokinetics and anaesthesia.
doi:10.1124/jpet.102.043273 Contin Educ Anaesth Crit Care Pain 7(1):25–29.
Koskinen M, Mustola S, Sepp€anen T (2005) Relation of doi:10.1093/bjaceaccp/mkl058
EEG spectrum progression to loss of responsiveness Steyn-Ross ML, Steyn-Ross DA, Sleigh JW, Liley DTJ
during induction of anesthesia with propofol. Clin (1999) Theoretical electroencephalogram stationary
Neurophysiol 116(9):2069–2076. doi:10.1016/j. spectrum for a white-noise-driven cortex: evidence
clinph.2005.06.004 for a general anesthetic-induced phase transition.
Kuhlmann L, Foster BL, Liley DT (2013) Modulation of Phys Rev E 60(6 Pt B):7299–7311, URL: http://
functional EEG networks by the NMDA antagonist www.ncbi.nlm.nih.gov/pubmed/11970675
nitrous oxide. PLoS One 8(2):e56434. doi:10.1371/ Steyn-Ross ML, Steyn-Ross DA, Sleigh JW
journal.pone.0056434, URL: http://www.ncbi.nlm. (2004) Modelling general anaesthesia as a first-order
nih.gov/pubmed/23457568 phase transition in the cortex. Prog Biophys Mol Biol
Kuizenga K, Wierda JM, Kalkman CJ (2001) Biphasic 85(2–3):369–385. doi:10.1016/j.pbiomolbio.
EEG changes in relation to loss of consciousness dur- 2004.02.001
ing induction with thiopental, propofol, etomidate, Steyn-Ross DA, Steyn-Ross ML, Sleigh JW, Wilson MT
midazolam or sevoflurane. Br J Anaesth (2011) Progress in modeling EEG effects of general
86(3):354–360 anesthesia: biphasic response and hysteresis. In: Hutt
Lee U, Ku S, Noh G, Baek S, Choi B, Mashour GA A (ed) Sleep and anesthesia: neural correlates in theory
(2013) Disruption of frontal-parietal communication and experiment, chapter 8, vol 15, Springer series in
by ketamine, propofol, and sevoflurane. Anesthesiol- computational neuroscience. Springer, New York,
ogy 118(6):1264–1275. doi:10.1097/ALN. pp 167–194. doi:10.1007/978-1-4614-0173-5n 8
0b013e31829103f5 Steyn-Ross ML, Steyn-Ross DA, Sleigh JW
Liley DTJ, Bojak I (2005) Understanding the transition to (2013) Interacting Turing-Hopf instabilities drive
seizure by modeling the epileptiform activity of gen- symmetry-breaking transitions in a mean-field model
eral anesthetic agents. Clin Neurophysiol of the cortex: a mechanism for the slow oscillation.
22(5):300–313 Phys Rev X 3(2):021005. doi:10.1103/
Liley DT, Walsh M (2013) The mesoscopic modeling of PhysRevX.3.021005, URL: http://link.aps.org/doi/
burst suppression during anesthesia. Front Comput 10.1103/PhysRevX.3.021005
Neurosci 7:46. doi:10.3389/fncom.2013.00046, URL: Voss LJ, Ludbrook G, Grant C, Upton R, Sleigh JW
http://www.ncbi.nlm.nih.gov/pubmed/23641211 (2007) A comparison of pharmacokinetic/pharmaco-
Liley DTJ, Cadusch PJ, Wright JJ (1999) A continuum dynamic versus mass-balance measurement of brain
theory of electro-cortical activity. Neurocomputing concentrations of intra-venous anesthetics in sheep.
26–27:795–800 Anesth Analg 104(6):1440–1446. doi:10.1213/01.
Liley DT, Sinclair NC, Lipping T, Heyse B, Vereecke HE, ane.0000263274. 62303.1a
Struys MM (2010) Propofol and remifentanil differen- Weir CJ (2006) The molecular mechanisms of general
tially modulate frontal electroencephalographic activ- anaesthesia: dissecting the GABAA receptor. Contin
ity. Anesthesiology 113(2):292–304. doi:10.1097/ Educ Anaesth Crit Care Pain 6(2):49–53. doi:10.1093/
ALN.0b013e3181e3d8a6 bjaceaccp/mki068
Anti-Hebbian Learning 191 A
Wentlandt K, Samoilova M, Carlen PL, El Beheiry learning is commonly defined as follows: corre-
H (2006) General anesthetics inhibit gap junction com- lated activation in the pre- and postsynaptic neu-
munication in cultured organotypic hippocampal
slices. Anesth Analg 102(6):1692–1698. doi:10.1213/ rons leading to the strengthening of the A
01.ane.0000202472.41103.78 connection between the two neurons. However,
Wilson HR, Cowan JD (1972) Excitatory and inhibitory the original definition offered in Hebb (1949)
interactions in localized populations of model neurons. talks about the increase in the presynaptic neu-
Biophys J 12:1–24
Wilson MT, Sleigh JW, Steyn-Ross DA, Steyn-Ross ML ron’s efficiency of eliciting activity in the post-
(2006) General anesthetic-induced seizures can be synaptic neuron, under the same correlated firing
explained by a mean-field model of cortical dynamics. condition. These two definitions (strengthening
Anesthesiology 104:588–593 of connection vs. increased efficiency) are com-
patible only when the presynaptic neuron is excit-
atory. They are contradictory when the
presynaptic neuron is inhibitory: increased con-
nection strength (weight, efficacy) corresponds to
Animal Calls decreased efficiency in eliciting response.
Assuming the original definition of Hebbian
▶ Pulse-Resonance Sounds learning, we can define anti-Hebbian learning as
a form of synaptic plasticity where correlated
activation in the pre- and postsynaptic neurons
leads to the reduction in the efficiency of the
presynaptic neuron’s ability to elicit activation
Animal Models of Pain of the postsynaptic neuron.
▶ Biomechanical Model of Low Back Pain

Detailed Description
In this entry, we will review computational for-

Anomalous Rectifier Potassium mulations of anti-Hebbian learning and their the-
Channels oretical implications. We will also briefly touch
upon neurobiological correlates of the learning
▶ Inward Rectifier Potassium Channels mechanism. The literature on anti-Hebbian learn-
ing is not as rich as that on Hebbian learning.
A review of Hebbian learning by Frégnac
(2003) includes a brief discussion on the neuro-
Anti-Hebbian Learning biological bases of anti-Hebbian learning. See
Földiák (1990) for a computational treatment of
Yoonsuck Choe the subject and Palmieri et al. (1993) for a review.
Department of Computer Science and
Engineering, Texas A&M University,
College Station, TX, USA Basic Formulation
Consider a simple configuration with a single

Definition presynaptic neuron x connecting to a single post-
synaptic neuron y with synaptic weight w as
Anti-Hebbian learning is a form of activity- shown above (Fig. 1). Anti-Hebbian learning is
dependent synaptic plasticity that is defined as the same as Hebbian learning, except for the flip
the opposite of Hebbian learning. Hebbian of the sign
A 192 Anti-Hebbian Learning
w v11
x y x1 y1
v21 w21
Anti-Hebbian Learning, Fig. 1 A pair of neurons v12 w12

x2 y2
forming an anti-Hebbian connection v22
Anti-Hebbian Learning, Fig. 2 Neurons connected

dw
¼ xy, (1) with Hebbian (arrows) and anti-Hebbian synapses
dt (discs) (Simplified from Földiák 1990)
where the learning rate is a small, fixed, positive

value. The neuronal activities x and y are nor-
mally assumed to be positive (or zero), but for x1
purely computational purposes, they can be neg-
y1
ative values as well. In Hebbian learning, adap-
tive threshold or normalization approaches are x2
used to check unbounded growth in the synaptic y2
weight. However, in anti-Hebbian learning, such
a check is unnecessary since the mechanism is y3
inherently stable (Földiák 1990). At first glance,
it appears that the weight w can tend toward 1,
Anti-Hebbian Learning, Fig. 3 A hierarchy of neurons
but this will never happen because high levels of
yi with Hebbian afferent connections (arrows) and anti-
inhibition will shut off y after a certain point and Hebbian connections in a cascade (discs) (Adapted from
thus w will not change any further thereafter. Carlson 1990)
When x and y are allowed to be negative, w can
increase under the same learning rule. Some for- components that make up the input mixtures. The
mulations prevent this from causing infinite model included an anti-Hebbian rule similar to
growth by imposing a limit (w is assumed to be Eq. 1 with an added threshold (wij), a Hebbian
negative and it is reset to 0 as soon as it becomes rule with weight decay (v ij), and a dynamic neural
positive; Földiák 1990), but others do not include activation equation dy dt with adaptive threshold
i
such a limit (e.g., Girolami and Fyfe 1996). and sigmoid nonlinearity (for details, see Földiák
1990).
Now consider a hierarchical network shown
Theoretical Perspectives above (Fig. 3). Hebbian learning is known to
extract the first principal component of the inputs
Anti-Hebbian learning is usually combined with (Oja 1982), so neuron y1 would serve this role.
Hebbian learning to produce interesting theoreti- Using a hierarchical network like Fig. 3 with anti-
cal and practical results. Fig. 2 below shows such Hebbian connections arranged in a cascade, sec-
an example (adapted from Földiák 1990). In this ond, third, and subsequent principal components
figure, two downstream neurons y1 and y2 receive of the input can be found (see Carlson 1990 for
afferent input from x1 and x2 through Hebbian syn- a review and details). The main idea is that once
apses (with weights vij) and exchange activations via yi learned the i-th principal component, and yi+1 is
anti-Hebbian lateral connections (with weights wij). decorrelated with y1, y2, . . ., yi through anti-
(The subscripts ij on the weights indicate the target Hebbian connections (note that y3 receives anti-
[i] and source neuron index [j], respectively.) Hebbian connections from both y1 and y2), yi+1
Földiák showed that the above network can will find the i + 1-th principal component.
learn to decorrelate the output neurons’ activity For a review of more complicated network
(sparse coding in neurons yi), which results in the topologies that involve anti-Hebbian learning,
learning of afferent representations (vij) that are see Palmieri et al. (1993).
Anti-Hebbian Learning 193 A
Neurobiological Underpinnings Fyfe (1996) used anti-Hebbian rule to learn finite
impulse response (FIR) filter coefficients and
As Hebbian learning is usually associated with applied the technique to blind source separation A
long-term potentiation (LTP), anti-Hebbian in the speech recognition domain. Schraudolph
learning is typically explained by long-term and Sejnowski (1992) used anti-Hebbian rule to
depression (LTD). LTP is a phenomenon where learn invariances in disparity tuning for stereo
high frequency stimulation of the presynaptic vision. Földiák (1990) combined Hebbian and
neuron leads to a prolonged increase in synaptic anti-Hebbian rule to learn sparse representations
efficacy (Bliss and Collingridge 1993). LTD is from overlapping visual inputs.
the opposite, where low-frequency stimulation
causes a prolonged decrease in synaptic efficacy
(Dudek and Bear 1992); thus, it fits the anti- Cross-References
Hebbian profile.
Examples of LTD associated with anti- ▶ Hebbian Learning
Hebbian mechanisms have been found in the ▶ Spike-Timing Dependent Plasticity, Learning
mammalian cerebellum (e.g., mouse and rat) Rules
and also in the electrosensory lobe (ELL) in the
teleost electric fishes, and these LTD mecha-
nisms are thought to be playing an anti-Hebbian References
role (see Frégnac 2003 for a brief review).
In some cases, LTP can also be associated with Bliss TVP, Collingridge GL (1993) A synaptic model of
memory: long-term potentiation in the hippocampus.
a form of anti-Hebbian learning, when LTP is Nature 361:31–39
induced not by correlated activation but by pre- Caporale N, Dan Y (2008) Spike timing-dependent plastic-
synaptic firing not being met by or negated by the ity: a Hebbian learning rule. Ann Rev Neurosci 31:25–46
postsynaptic activity. Such a phenomenon has Carlson A (1990) Anti-Hebbian learning in a non-linear
neural network. Biol Cybern 64:171–176
been observed in the hippocampus and has been
Dudek SM, Bear MF (1992) Homosynaptic long-term
dubbed anti-Hebbian LTP. Kullmann and Lamsa depression in area CA1 of hippocampus and effects
(2007) provide an extensive review on this topic of N-Methyl-D-Aspartate receptor blockade. Proce
and provide a comparison of anti-Hebbian LTD Natl Acad Sci USA 89:4363–4367
Földiák P (1990) Forming sparse representations by local
and anti-Hebbian LTP.
anti-Hebbian learning. Biol Cybern 64:165–170
Finally, spike timing-dependent plasticity Frégnac Y (2003) Hebbian synaptic plasticity. In: Arbib MA
(STDP; see Caporale and Dan 2008 for (ed) The handbook of brain theory and neural networks,
a review) has also been linked to anti-Hebbian vol 2. MIT Press, Cambridge, MA, pp 515–522
Girolami M, Fyfe C (1996) A temporal model of linear
learning. STDP is a short-term synaptic plasticity anti-Hebbian learning. Neural Process Lett 4:139–148
mechanism where depending on the ordering of Hebb DO (1949) The organization of behavior:
pre- and postsynaptic events, either LTP or LTD a neuropsychological theory. Wiley, New York
can entail. LTP is induced when the presynaptic Kullmann DM, Lamsa KP (2007) Long-term synaptic
plasticity in hippocampal interneurons. Nat Rev
activity precedes postsynaptic activity and LTD
when the ordering is reversed. Some views the Nelson SB (2004) Hebb and anti-Hebb meet in the
LTD part of STDP as implementing an anti- brainstem. Nat Neurosci 7:687–688
Hebbian rule (see Nelson 2004 for a discussion). Oja E (1982) A simplified neuron model as a principal
component analyzer. J Math Biol 15:267–273
Palmieri F, Zhu J, Chang C (1993) Anti-Hebbian learning
in topologically constrained linear networks:
Applications of Anti-Hebbian Learning A tutorial. IEEE Trans Neural Netw 4:748–761
Schraudolph NN, Sejnowski TJ (1992) Competitive anti-
Hebbian learning of invariants. In: Moody JE, Hanson
Anti-Hebbian learning has been applied to sev-
SJ, Lippmann RP (eds) Advances in neural informa-
eral signal and data processing tasks including tion processing systems. Morgan Kaufmann, San
vision and speech processing. Girolami and Mateo, pp 1017–1024
A 194 Aperture Problem
may be considered as a system specification.

Aperture Problem Many declarative programming languages utilize
term-rewriting logic: the declarations are rewrite
▶ Somatosensory Cortex: Neural Coding of laws that specify how one term should be
Motion replaced by another. Abstract rewriting systems
are written with arrows, specifying that the left-
hand side (LHS) is to be replaced by the term on
the right-hand side (RHS). Other notations, such
Application of Declarative as Backus-Naur Form, utilize some variant of an
Programming in Neurobiology equal sign to specify that RHS is replaced
by LHS.
Thomas J. Anastasio Crucially, because the declarations are
Department of Molecular and Integrative descriptions of system properties, the specifica-
Physiology, and Beckman Institute, University of tion can be used not only for simulation but also
Illinois at Urbana-Champaign, Urbana, IL, USA for analyses such as state-space search and
temporal-logic model checking (explained
below; for a general reference see Huth and
Definition Ryan 2004). Term-rewriting declarative lan-
guages are readily used for expressing data-
The main technique in Computational Neurosci- driven models constructed directly from experi-
ence is imperative programming, which is often mental observations. These models do not require
used to implement simulations of dynamics, and an overarching conceptual framework or
describes how a computation is performed. preconceived hypothesis, allowing the enormous
A complement to this approach is declarative amounts of data generated by neurobiological
programming. Declarations provide descriptions experiments to be directly entered. This allows
of relationships between elements, effectively the data to “speak for itself” and provides a model
describing what the computation should accom- that can be used as a tool to help develop subse-
plish. The use of declarative programming for quent hypotheses.
modeling biological processes is still in its The first term-rewriting language was the
infancy (Fisher and Henzinger 2007) yet has l-calculus, developed in the 1930s by Alonzo
shown itself to be a valuable first step for analysis Church (Cardone and Hindley 2006). Since then
of the seemingly impenetrable complexity of several other term-rewriting, declarative pro-
molecular interactomics: the interplay of the gramming languages have been developed.
myriad proteins and signaling species in the Some well-known examples are Scheme (http://
cell. Declarative programming can also be used www.r6rs.org/), Alloy (http://alloy.mit.edu/
at the connectomic level of understanding con- alloy/index.html), Simile (http://www.
nections among neurons or among brain areas. simulistics.com/tour/declarative.htm), and
ECLiPSe (http://eclipseclp.org/index.html). One
declarative programming language that has been
Detailed Description used effectively for modeling molecular interac-
tions and neurobiological processes is Maude
Declarative Programming (http://maude.cs.uiuc.edu/; Clavel et al. 2007).
The declarations of a declarative program A specification in Maude is based on an under-
describe the relationships between system ele- lying algebra, which is defined by sorts (data
ments. Because of this descriptive nature, types) and by operations that are allowed for
a model implemented in a declarative program each sort (e.g., natural numbers – sort, can be
Application of Declarative Programming in Neurobiology 195 A
added together – operation). Preset algebras such and avionics among others (Meseguer 2012).
as the natural numbers are available, but new Maude is also used to model other programming
algebras can be defined by the user. Declarations languages, and analysis of a Maude model can A
in Maude are written in terms of the underlying verify that a language behaves as intended. This
algebra and can be expressed either as equations type of analytical capability suggests how Maude
or as rules. Equations in Maude specify func- can be used to assess interactomic “programs,” so
tional relationships that simplify the state of the as to determine where they have vulnerabilities,
system being modeled, while rules specify tran- predicting sites for potential failure which would
sitional relationships that change the state of the produce disease, as well as sites where interven-
model system. For example, an equation could tion could compensate for these failures,
specify that 2 apples plus 4 apples plus 3 apples predicting sites where pharmacological interven-
equals 9 apples, while a rule could specify that tion would be of value.
9 apples can transition to an apple pie.
Equations and rules can be unconditional (eq Declarative Models of Fear Conditioning
and rl, respectively) or conditional (ceq and and Alzheimer Disease (AD)
crl). For example, a conditional rule could spec- Maude has been used to model several biological
ify that 9 apples can transition to an apple pie, but processes (Eker et al. 2002; Talcott 2008).
only if mom is available to bake it. While an A recent neurobiological application explores
unconditional equation or rule applies whenever the synaptic plasticity thought to underlie fear
its LHS is present, a conditional rule requires conditioning and its extinction (Anastasio
both the LHS and the required conditions. 2013b). In that microconnectomic model, rules
All rewrite systems have rules (a.k.a. syntax) specified synaptic weight changes while equa-
that transition the system state and evaluations tions specified changes in neural responses
(a.k.a. semantics) that can reduce the state, in caused by weight changes. Thus, each state of
whole or in part, to its underlying value at arbi- the model system had a different one of the
trary times (Kain 1972). Another characteristic of many possible configurations of synaptic
Maude is that equations (i.e., evaluations) must weights. This particular model had only nine
execute whenever they apply while applicable synapses, which were allowed only a limited
rules may execute or not. Rules, by executing, number of discrete weights, so the state space
can transition the system to a state in which was large but still small enough to permit exhaus-
equations become applicable that were not so in tive search. Analysis via state-space search pro-
the previous state. In using Maude for simulation, vided testable hypotheses by revealing which of
all applicable rules would ultimately execute these very many synaptic weight configurations
according to a fairness criterion: every applicable were compatible with fear conditioning followed
rule must execute once before any other rule can by extinction.
execute twice. In using Maude for analyses, such Maude models of some of the molecular inter-
as state-space search or logical model checking, actions that underlie AD have also recently
the rules execute in all possible combinations and appeared (Anastasio 2011, 2013a). These studies
orders. Because rules cause state transitions, were based on the amyloid hypothesis, which
Maude thereby elaborates the state transition posits that AD results from buildup of the
tree implied by the specification and later amyloid-beta peptide (Hardy and Selkoe 2002).
searches this tree for specific states or to verify A subset of the molecular and cellular interac-
certain temporal-logic propositions. tions believed to underlie the regulation of
Maude has been used to specify and analyze amyloid-beta is diagrammed in Fig. 1.
a broad range of complex engineered systems Space constraints prohibit description of all of
used for computer network security, encryption, these interactions. They are based directly on
A
196
Application of Declarative Programming in Neurobiology, Fig. 1 Schematic diagram of many of the cellular and molecular interactions that regulate the level of the
peptide amyloid-beta
Application of Declarative Programming in Neurobiology
Application of Declarative Programming in Neurobiology 197 A
findings from the primary literature and described Application of Declarative Programming in Neuro-
in detail in Anastasio (2011). Each of 36 declara- biology, Table 1 Modeling the effects of nonsteroidal
anti-inflammatory drugs (NSAIDs) and blockers of
tions specified how the level of biological activity hypoxia-inducible factor and caspase 3 (HIF-blocker and A
of one model element, representing a molecular casp-blocker) on the level of amyloid-beta in the presence
species, was determined through interaction of incipient cerebrovascular disease (CVD)
among the others. Many model elements could HIF- Casp- Amyloid-
assume multiple levels, so the entire state space CVD NSAID blocker blocker beta
was too large to be managed on a desktop com- 0 0 0 0 4
puter. As an alternative to high-performance 1 0 0 0 8
computing, the declarations, first specified as 1 1 0 0 5
equations, were then divided into subsets that 1 0 1 0 5
could be converted to rules. In this way, the 1 0 0 1 7
space of configurations involving a subset of 1 1 1 0 4
1 1 0 1 5
interactions could be analyzed using state-space
1 0 1 1 5
search and temporal logical, while the rest of the
1 1 1 1 4
interactions took place in the background.
An example of an equation in the Maude
model of amyloid-beta regulation is
ceq [PPARexpress] : PPARgene(G) more effective in reducing the rise of amyloid-
NSAID(X) cytokine(Y) PPAR(Z) = beta in the presence of CVD in the model.
PPARgene(G) NSAID(X) cytokine In the model, amyloid-beta rose from its nor-
(Y) PPAR(max(0, G + (X - Y)) * G) mative level of 4 to the pathological level of 8 in
if Z =/= max(0, G + (X - Y)) * G . the presence of CVD (Table 1). NSAIDs, known
This equation, labeled PPARexpress, to increase PPAR levels in vitro (Sastre
describes the interaction that determines the et al. 2006), and HIF-blocker each separately
level of the peroxisome proliferator-activated held the rise of amyloid-beta to 5 while casp3-
receptor (PPAR). The operator PPAR(Z)assigns blocker only held it to 7. NSAID and HIF-blocker
integer level Z to PPAR, and the equation together held amyloid-beta to its normative level
describes how that level is determined by the of 4 but casp3-blocker provided no further benefit
presence of its gene PPARgene and the levels in combination with NSAID or HIF-blocker or
of cytokines and nonsteroidal anti-inflammatory both.
drugs (NSAIDs). This equation is conditional These simulations showed that a HIF-blocker
(ceq) and executes only if doing so changes the was more effective than a casp3-blocker in reduc-
level assigned by PPAR. To convert this condi- ing the rise of amyloid-beta in the face of
tional equation to a conditional rule, it would be CVD. We then used temporal-logic analysis to
relabeled crl and = would be replaced with =>. show why this effect occurs. Temporal-logic
The model depicted in Fig. 1 incorporates the analysis provides answers to questions such as
hypothesis that mild cerebrovascular disease whether a particular property is always true,
(CVD) can trigger amyloid-beta accumulation never true, or only true after some other property
(Scheibel et al. 1989; de la Torre 2009). Simula- becomes true. For the AD model, temporal-logic
tions and analysis were focused on compounds analysis was used to determine the value of prop-
known to modulate the activities of model ele- ositions of the form ~ hasAPO U cytACT,
ments. Specifically, cilnidipine blocks hypoxia- where hasAPO is the property that apoptosis
inducible factor (HIF) (Oda et al. 2009) while has occurred, cytACT is the property that cyto-
ifenprodil blocks caspase 3 (casp3) (Dave kines have been activated, and ~ and U are the
et al. 2003). It was of interest to see whether temporal-logic operators “not” and “until,”
a HIF-blocker or a casp3-blocker would be respectively. Maude demonstrated that this
A 198 Application of Declarative Programming in Neurobiology
proposition is true, meaning that apoptosis will Anastasio TJ (2013b) Computational search for hypothe-
not occur until cytokines are activated in the ses concerning the endocannabinoid contribution to
the extinction of fear conditioning. Front Comput
model. That is important because apoptosis acti- Neurosci 7:74
vated caspase 3 in the model. Similar analyses Cardone F, Hindley JR (2006) History of Lambda-
revealed that cytokines were not activated if calculus and combinatory logic. In: Gabbay DM,
HIF-blocker was present. Taken together, the Woods J (eds) Handbook of the history of
logic. Elsevier, Amsterdam
temporal-logic analyses show that casp3-blocker Clavel M, Durán R, Eker S, Lincoln P, Martı́-Oliet N,
was not effective in combination with Meseguer J, Talcott C (2007) All about Maude:
HIF-blocker because HIF-blocker already a high-performance logical framework: how to spec-
prevented caspase 3 activation. ify, program, and verify systems in rewriting
logic. Springer, Berlin
These analyses made several testable Dave JR, Williams AJ, Moffett JR, Koenig ML,
predictions: Tortella FC (2003) Studies on neuronal apoptosis in
1. A HIF-blocker will be effective in reducing primary forebrain cultures: neuroprotective/anti-
amyloid-beta levels in the presence of apoptotic action of NR2B NMDA antagonists.
Neurotox Res 5(4):255–264
incipient CVD. de la Torre JC (2009) Cerebrovascular and cardiovascular
2. A casp-blocker will be less effective than pathology in Alzheimer’s disease. Int Rev Neurobiol
a HIF-blocker. 84:35–48
3. A casp-blocker will provide no further benefit Eker S, Knapp M, Laderoute K, Lincoln P, Meseguer J,
Sonmez K (2002) Pathway logic: symbolic analysis of
if administered in conjunction with biological signaling. Pac Symp Biocomput
a HIF-blocker. 2002:400–412
4. NSAIDs will be more effective in combina- Fisher J, Henzinger TA (2007) Executable cell biology.
tion with a HIF-blocker than alone. Nat Biotechnol 25(11):1239–1249
Hardy J, Selkoe DJ (2002) The amyloid hypothesis
of Alzheimer’s disease: progress and problems
on the road to therapeutics. Science
Conclusions 297(5580):353–356
Huth M, Ryan M (2004) Logic in computer science:
modelling and reasoning about systems. Cambridge
Declarative programming can be applied in the University Press, Cambridge
simulation and temporal-logic analysis of complex Kain RY (1972) Automata theory: machines and lan-
neurobiological processes, whether normal or guages. McGraw-Hill, New York
pathological. The phenomenon at issue in the Meseguer J (2012) Twenty years of rewriting logic. J Log
Alg Prog 81(7–8):721–781
above example, the regulation of amyloid-beta, Oda S, Oda T, Takabuchi S, Nishi K, Wakamatsu T,
occurs predominantly on the molecular level. Tanaka T, Adachi T, Fukuda K, Nohara R, Hirota
However, term-rewriting and declarative pro- K (2009) The calcium channel blocker cilnidipine
gramming can also be used to create and analyze selectively suppresses hypoxia-inducible factor
1 activity in vascular cells. Eur J Pharmacol
models of neurobiological phenomena at all 606(1–3):130–136
scales, from molecular to cellular to network to Sastre M, Dewachter I, Rossner S, Bogdanovic N,
whole brain, or to create multiscale models in Rosen E, Borghgraef P, Evert BO,
order to understand phenomena arising from inter- Dumitrescu-Ozimek L, Thal DR, Landreth G,
Walter J, Klockgether T, van Leuven F,
actions across multiple scales. Heneka MT (2006) Nonsteroidal anti-
inflammatory drugs repress beta-secretase
gene promoter activity by the activation of
PPARgamma. Proc Natl Acad Sci U S
References A 103(2):443–448
Scheibel AB, Duong TH, Jacobs R (1989) Alzheimer’s
Anastasio TJ (2011) Data-driven modeling of Alzheimer disease as a capillary dementia. Ann Med
disease pathogenesis. J Theor Biol 290:60–72 21(2):103–107
Anastasio TJ (2013a) Exploring the contribution of estrogen Talcott C (2008) Pathway logic. In: Bernardo M,
to amyloid-beta regulation: a novel multifactorial com- Degano P, Zavattaro G (eds) Lecture notes in com-
putational modeling approach. Front Pharmacol 4:16 puter science. Springer, Berlin, pp 21–53
Applications of Information Theory to Analysis of Neural Data 199 A
A typical application of mutual information to
Applications of Information Theory spike train analysis is to use it to compare the
to Analysis of Neural Data information content of different representations A
of neural responses that can be extracted from
Simon R. Schultz1, Robin A. A. Ince2 and spike trains. The neural code used by a neuron
Stefano Panzeri3,4 is often defined operationally as the smallest set
1
Department of Bioengineering, Imperial of response variables that carries all (or almost
College London, London, UK all) the information contained in the spike train of
2
School of Psychology, Institute of Neuroscience the neuron. Mutual information is used to quan-
and Psychology, University of Glasgow, tify the information content of increasingly com-
Glasgow, UK plex representations of the neural response, and
3
Center for Neuroscience and Cognitive Systems, the simplest representation that carries the most
Istituto Italiano di Tecnologia, Rovereto (Tn), information is chosen as the putative neural code.
Italy An example of this general approach is the
4
Institute of Neuroscience and Psychology, investigation of the role of spike times in
University of Glasgow, Glasgow, UK encoding information. The most established
hypothesis on how sensory information is
represented in the brain is the spike count-coding
Definition hypothesis (Adrian 1928) which suggests that
neurons represent information by the number of
Information theory is a practical and theoretical spikes discharged over some relevant time win-
framework developed for the study of communica- dow. Another hypothesis is the spike timing
tion over noisy channels. Its probabilistic basis and encoding hypothesis, which suggests that the
capacity to relate statistical structure to function timing of spikes may add important information
make it ideally suited for studying information to that already carried by spike counts (Rieke
flow in the nervous system. It has a number of useful et al. 1997; Panzeri et al. 2001). Information
properties: it is a general measure sensitive to any theory can be used to understand the role of
relationship, not only linear effects; it has meaning- spike times in carrying sensory information, by
ful units which in many cases allow direct compar- using it to characterize the temporal resolution
ison between different experiments; and it can be needed to read out the information carried by
used to study how much information can be gained spike trains. This can be performed by sampling
by observing neural responses in single trials, rather the spike train at different temporal precisions,
than in averages over multiple trials. A variety of Dt, (Fig. 1a) and computing the information para-
information-theoretic quantities are commonly used metrically as a function of Dt (Ruyter et al. 1997).
in neuroscience – (see entry “Definitions of The temporal precision required to read the tem-
Information-Theoretic Quantities”). In this entry poral code then can be defined as the largest Dt
we review some applications of information theory that still provides the full information obtained at
in neuroscience to study encoding of information in higher resolutions. If this precision is equal to the
both single neurons and neuronal populations. overall length of the window over which neurons
carry information, then information is carried
only by the number of spikes. As an example,
Detailed Description we carried out this type of analysis on the
responses of neurons from the VPm thalamic
Information Analysis of Spike Trains to nucleus of rats whose whiskers were stimulated
Investigate the Role of Spike Times in by fast white noise deflections (Montemurro
Sensory Coding et al. 2007). We found that the temporal precision
Mutual information is a widely used tool to study Dt at which neurons transmitted information
how spike trains encode sensory variables. about whisker deflections was finer than 1 ms
A 200 Applications of Information Theory to Analysis of Neural Data
b
a
2 3
1 1 2 1
1 0 1 0 1 1 0 1
2ms
Applications of Information Theory to Analysis of sequence. (b) The information rate (information per unit
Neural Data, Fig. 1 Effect of temporal resolution of time) about whisker deflections carried by VPm thalamic
spike times on information. (a) The response of a neuron neurons as a function of bin width, Dt, used to bin neural
is initially recorded as a series of spike times. To investi- responses. Information rate increased with bin resolution
gate the temporal resolution at which spike times carry up to 0.5 ms, the limit of the experimental setup. This
information, the spike train is binned at a variety of dif- shows that a very fine temporal resolution is needed to
ferent time resolutions, by labeling the response at each read out the sensory messages carried by these thalamic
time with the number of spikes occurring within that bin, spike trains (Figure reprinted with permission from Ince
thereby transforming the response into a discrete integer et al. (2010))
(Fig. 1b), suggesting that these neurons use high- spectrum containing fluctuations over a wide
precision spike timing to carry information. range of frequencies, from <1 to 100 Hz or
so. Given that the power of oscillatory activity
Information Analysis of Local Field Potentials typically increases during the presentation of
to Examine the Information Content of a sensory stimulus, many authors have speculated
Network Oscillations that this oscillatory activity plays a role in brain
Information analysis in neuroscience is not lim- communication and in particular in sensory-
ited only to spike train analysis, but it has been related computations. However, understanding
used also to study the measure of massed popu- the function of these oscillations has remained
lation activity, such as local field potentials elusive and controversial. To gain insights into
(LFPs) (Buzsáki et al. 2012). LFPs are operation- the function of oscillations in sensory encoding, it
ally defined as the low-pass filtered extracellular is important to understand how they contribute to
potential measured by an extracellular intracra- the representation of the natural sensory
nial electrode. There are at least three reasons environment.
why LFPs are widely used in neuroscience. The This problem can be addressed by quantifying
first is that they are more easily and stably the oscillation power in any given trial in
recorded in chronic settings than is the spiking response to different stimuli and then computing
activity of individual neurons. The second is that the information gained by the power at each
the LFP captures key integrative synaptic pro- frequency. Since the power is a continuous vari-
cesses and aspects of subthreshold neural activity able, the computation of its information is poten-
that cannot be measured by observing the spiking tially more difficult than the one based on discrete
activity of a few neurons alone (Einevoll variables like the spike train ones described
et al. 2013). The third is that LFPs are more above. There are at least two ways to solve this
sensitive to network oscillations than measures problem. The first is to discretize the power in
of spiking activity from small populations. LFPs a number of equi-populated bins and to use bias
from a sensory area typically show a power corrections to eliminate the bias. The second is to
Applications of Information Theory to Analysis of Neural Data 201 A
fit the data to a parametric distribution. In this to trial variability for a fixed stimulus). This
case, it is worth reminding that the power com- finding has several implications. First, it shows
puted with most spectral methods follows a - that, despite the broadband spectrum, only a small A
chi-square distribution, and thus, its square or number of privileged frequency scales are
third can reasonably be well approximated by involved in stimulus coding. Second, it suggests
a Gaussian distribution (Magri et al. 2009). This that high-frequency and low-frequency oscilla-
makes the computation of information relatively tions are generated by different stimulus-
straightforward. The third potential approach is processing neural pathways. Third, the finding
to use binless methods such as Nearest Neighbors that different frequency bands code different sen-
approaches (Kraskov et al. 2004). We tried out sory features in separate, truly independent infor-
these methods extensively on computation of mation channels reinforces the concept of
information in power of LFPs, obtaining very “cortical multiplexing” that we proposed above.
similar results with all approaches (see, e.g.,
Magri et al. 2009). Information Analysis of Imaging Data to
We applied this method to recordings from Study Neural Population Coding or Coupling
primary visual cortex of anesthetized macaques Between Different Neural Signals
during stimulation with naturalistic color movies The analysis tools that we have described above
(Belitski et al. 2008; Magri et al. 2012a). This have, to date, largely been applied to spike train
revealed, for the first time, how information about and time series data recorded using electrophys-
the naturalistic sensory environment is spread iological techniques. However, in recent years,
over the wide range of frequencies expressed by imaging technologies have been developed
cortical activity. Although the broadband nature which are capable of resolving neural signaling
of the spectrum suggests a contribution to coding at systems’ cellular and subcellular resolutions on
from many frequency regions, we found that only a single trial basis (Denk et al. 1990, 1994;
two separate frequency regions contribute to cod- Stosiek et al. 2003; Chen et al. 2013). One way
ing: the low-frequency range and the gamma to apply information-theoretic tools to the analy-
range (Belitski et al. 2008; see Fig. 2 below). sis of such imaging data is to convert the data to
Interestingly, low- and high-frequency ranges a “spike train,” for instance, by applying an algo-
act as perfectly complementary or “orthogonal” rithm for the detection of action potential-evoked
information channels: they share neither signal calcium transients to calcium imaging data
(i.e., stimulus information) nor “noise” (i.e., trial (Oñativia et al. 2013). Such an approach has
a LFP Power Spectrum b LFP Information

40
Movie 0.2
Information (bits)
Spontaneous
Power (dB)
30
20 0.1
10
0
0 50 100 150 0 50 100 150 200
Frequency (Hz) Frequency (Hz)
Applications of Information Theory to Analysis of during the presentation of a color movie stimulus (solid
Neural Data, Fig. 2 The visual information carried by line). (b) Information about the movie stimulus carried by
LFP power at different frequencies. (a) LFP power spec- LFP power at different frequencies. The area indicates the
trum of V1 recordings anesthetized macaques either dur- SEM (Reproduced from Magri et al. (2012a))
ing spontaneous activity in the dark (dashed line) or
A 202 Applications of Information Theory to Analysis of Neural Data
been used to perform information-theoretic anal- analogue brain signals is crucially dependent
ysis of simultaneously recorded populations of upon the development of appropriate regulariza-
cerebellar Purkinje cell complex spikes extracted tion and dimensionality reduction algorithms.
from in vivo calcium imaging movies (Schultz These might stem from simple yet efficient
et al. 2009). However, the use of imaging data discretization algorithms (Belitski et al. 2008;
may also allow a wider set of questions to be Magri et al. 2009), Nearest Neighbors regulari-
approached than can be examined electrophysio- zation algorithms (Kraskov et al. 2004), the use
logically, by directly examining patterns of pixel of manifold learning techniques for nonlinear
intensities. dimensionality reduction (Roweis and Saul
Another interesting application of information 2000; Seung and Lee 2000; Gan 2006), and/or
theory to neuroimaging data regards its use for the evaluation of information through
understanding the nature of the coupling between a decoding step (Quian Quiroga and Panzeri
neural activity and fMRI responses. In fact, 2009).
although there is evidence that fMRI BOLD
responses reflect neural activity, it is not clear
Acknowledgements Research is supported by the
whether the BOLD signal reflects only the total
SI-CODE (FET-Open, FP7-284533) project and by the
power of massed neural activity, or only the ABC and NETT (People Programme Marie Curie Actions
power in a given band, or rather the relationships PITN-GA-2011-290011 and PITN-GA-2011-289146)
between powers of neural activity in different projects of the European Union’s Seventh Framework
Programme FP7 2007–2013.
frequency bands. This problem can be cast theo-
retically into quantifying whether more informa-
tion about BOLD can be gained from
simultaneously observing the power of neural References
activity in two or more bands of neural activity
than the information gained by observing either Adrian ED (1928) The basis of sensation. Norton, New
York. http://psycnet.apa.org/psycinfo/1928-01753-
band alone. Because mutual information captures
000. Accessed 17 Jan 2014
all the ways a signal may statistically relate to Belitski A, Gretton A, Magri C, Marayama Y,
another, finding that another signal carries extra Montemurro MA, Logothetis NK, Panzeri S (2008)
information demonstrates that this signal truly Low-frequency local field potentials and spikes in
primary visual cortex convey independent visual infor-
provides some information that cannot be possi-
mation. J Neurosci 28:5696–5709
bly obtained from the first one. This does not Buzsáki G, Anastassiou CA, Koch C (2012) The origin of
necessarily hold when using methods that capture extracellular fields and currents – EEG, ECoG, LFP
only specific relationships between signals. For and spikes. Nat Rev Neurosci 13:407–420
Chen T-W, Wardill TJ, Sun Y, Pulver SR, Renninger SL,
example, an increase in predictability based on
Baohan A, Schreiter ER, Kerr RA, Orger MB,
linear models may reflect both additional infor- Jayaraman V, Looger LL, Svoboda K, Kim DS
mation from the second regressor as well as infor- (2013) Ultrasensitive fluorescent proteins for imaging
mation that was already present in the first neuronal activity. Nature 499:295–300
Denk W, Strickler JH, Webb WW (1990) Two-photon
regressor but was not captured by the linear
laser scanning fluorescence microscopy. Science
assumption. Application of this idea to simulta- 248:73–76
neous recording of LFPs and fMRI BOLD in Denk W, Delaney KR, Gelperin A, Kleinfeld D,
primary visual cortex showed that the beta and Strowbridge BW, Tank DW, Yuste R (1994) Anatom-
ical and functional imaging of neurons using 2-photon
alpha bands carry information about BOLD that laser scanning microscopy. J Neurosci Methods
complements that carried by the gamma band, the 54:151–162
band that most correlates to the BOLD signal Einevoll GT, Kayser C, Logothetis NK, Panzeri S (2013)
(Magri et al. 2012b). Modelling and analysis of local field potentials for
studying the function of cortical circuits. Nat Rev
Since imaging signals such as fMRI have an
analogue rather than discrete nature, the practi- Gan JQ (2006) Feature dimensionality reduction by man-
cality of application of information theory to ifold learning in brain-computer interface design.
Associations and Rewards in Auditory Cortex 203 A
In: Proceedings of the 3rd international workshop on
brain-computer interfaces, Graz, Austria, pp 28–29. Artificial Retina
http://cswww.essex.ac.uk/Research/BCIs/BCI06_
GAN1.pdf. Accessed 17 Jan 2014 A
Ince RAA, Mazzoni A, Petersen RS, Panzeri S (2010) ▶ Vision Prosthesis
Open source tools for the information theoretic analy- ▶ Visual Prosthesis, Epiretinal Devices
sis of neural data. Front Neurosci 4:62–70 ▶ Visual Prosthesis, Subretinal Devices
Kraskov A, Stögbauer H, Grassberger P (2004) Estimating
mutual information. Phys Rev E 69:66138
Magri C, Whittingstall K, Singh V, Logothetis NK,
Panzeri S (2009) A toolbox for the fast information
analysis of multiple-site LFP, EEG and spike train Artificial Silicon Retina (ASR)
recordings. BMC Neurosci 10:81
Magri C, Mazzoni A, Logothetis NK, Panzeri S (2012a)
Optimal band separation of extracellular field poten- ▶ Visual Prosthesis, Subretinal Devices
tials. J Neurosci Methods 210:66–78
Magri C, Schridde U, Murayama Y, Panzeri S, Logothetis
NK (2012b) The amplitude and timing of the BOLD
signal reflects the relationship between local field
potential power at different frequencies. J Neurosci Artificial Vision
32:1395–1407
Montemurro MA, Panzeri S, Maravall M, Alenda A, Bale
MR, Brambilla M, Petersen RS (2007) Role of precise ▶ Prosthetic Vision, Perceptual Effects
spike timing in coding of dynamic vibrissa stimuli in ▶ Retinal/Visual Interfaces (Models, Theory,
somatosensory thalamus. J Neurophysiol Techniques): Overview
98:1871–1882
Oñativia J, Schultz SR, Dragotti PL (2013) A finite rate of ▶ Vision Prosthesis
innovation algorithm for fast and accurate spike detec- ▶ Visual Prosthesis, Epiretinal Devices
tion from two-photon calcium imaging. J Neural Eng
10:046017
Panzeri S, Petersen RS, Schultz SR, Lebedev M, Diamond
ME (2001) The role of spike timing in the coding of
stimulus location in rat somatosensory cortex. Neuron Associations and Rewards
29:769–777 in Auditory Cortex
Quian Quiroga R, Panzeri S (2009) Extracting information
from neuronal populations: information theory
and decoding approaches. Nat Rev Neurosci Michael Brosch
10:173–185 Leibniz Institute for Neurobiology, Magdeburg,
Rieke F, Bialek W, Warland D, de Ruyter van Steveninck Germany
RR (1997) Spikes: exploring the neural code. Bradford
Book
Roweis ST, Saul LK (2000) Nonlinear dimensionality
reduction by locally linear embedding. Science Definition
290:2323–2326
Ruyter D, van Steveninck RR, Lewen GD, Strong SP, In psychology, the term association refers to
Koberle R, Bialek W (1997) Reproducibility and
variability in neural spike trains. Science a connection between different elementary men-
275:1805–1808 tal entities (sensations, thoughts, feelings) (Dudai
Schultz SR, Kitamura K, Post-Uiterweer A, Krupic J, 2004). Aside from innate, reflex-like associa-
H€ausser M (2009) Spatial pattern coding of sensory tions, novel associations are typically acquired
information by climbing fiber-evoked calcium signals
in networks of neighboring cerebellar purkinje cells. during learning. In Pavlovian/classical condition-
J Neurosci 29:8005–8015 ing, a stimulus-stimulus association is formed by
Seung HS, Lee DD (2000) The manifold ways of percep- repetitively pairing an initially neutral condi-
tion. Science 290:2268–2269 tioned stimulus with a biologically significant
Stosiek C, Garaschuk O, Holthoff K, Konnerth
A (2003) In vivo two-photon calcium imaging of unconditioned stimulus that triggers an uncondi-
neuronal networks. Proc Natl Acad Sci U S A tioned behavioral response. In instrumental/oper-
100:7319–7324 ant conditioning, a stimulus–response association
A 204 Associations and Rewards in Auditory Cortex
is formed under the presence of a reinforcer. the auditory cortex, their response strength and
Reinforcers can be either positive, resulting in response latency, as well as interneuronal syn-
an increase in the probability of a response to chrony (Scheich and Brosch 2013; Shepard
the stimulus. Positive reinforcers are rewards, et al. 2013). In detection tasks, the direction of
such as water, food, money, and brain stimulation the receptive field change at the frequency of the
reward. Reinforcers can also be negative (e.g., conditioned tone depends on the “meaning” of
footshocks, airpuffs, money loss), decreasing the tone (Scheich et al. 2011); if the meaning is
the probability of a response to the stimulus. If negative (associated with punishment), the
the reinforcer is removed, the learned association response to the tone is increased (receptive field
risks extinction. is sensitized at the conditioned tone frequency); if
it is positive (associated with reward), the
response is decreased (receptive field is
Detailed Description suppressed at the conditioned tone frequency).
In frequency discrimination tasks, slopes of spec-
Classical View of Brain Structures Reflecting tral tuning curves become sharper around the
Associations and Reinforcement conditioned frequencies. In categorization tasks,
Traditionally associative functions are assigned stimuli of the same category evoke (spatiotempo-
to the so-called association cortex (Creutzfeldt ral) neuronal activity patterns that are more sim-
1983). Inspired by “associationism” (the philo- ilar to each other than those evoked by stimuli of
sophical doctrine that the mind learns and con- other categories.
strues the world bottom up by associating mental These changes coincide with or may even
entities) and based on anatomical considerations, form the basis of changes in representation
Flechsig originally defined the association cortex maps (e. g., tonotopic frequency map) in auditory
as that part of the cerebral cortex that appeared to cortex that occur, at least transiently, after learn-
lack direct afferences from the senses and ing (Shepard et al. 2013). They may also underlie
efferences to peripheral motor structures. He pro- plasticity of neuronal mass action, as revealed in
posed that the association cortex provides the electro- and magnetoencephalography (EEG,
substrate for the fusion of primary sensations to MEG) or functional magnetic resonance imaging
obtain ideas of objects as a whole. This idea has (fMRI) (R€usseler et al. 2005).
been strongly challenged more recently by stud-
ies demonstrating associative functions already Neuromodulators
in primary sensory cortices. In addition, the sen- The formation of novel associations in the audi-
sory cortex is affected by reinforcement (Shuler tory cortex requires the involvement of
and Bear 2006; Brosch et al. 2011a; Arsenault neuromodulators (Shepard et al. 2013). Some of
et al. 2013), which generally is thought to involve the described changes can also be mimicked by
the limbic system, including the hypothalamus, repetitively pairing auditory stimuli with electri-
amygdala, hippocampus, septal nuclei, ventral cal stimulation of neuromodulatory systems, such
tegmental area, and anterior cingulate gyrus. as the ventral tegmental area (dopamine), the
These findings put into question the existence of nucleus basalis of Meynert (acetylcholine), the
unisensory cortical areas at all (Ghazanfar and locus coeruleus (norepinephrine), or the vagus
Schroeder 2006). nerve (triggering widespread release of
neuromodulators). Formation of associations in
Learning the auditory cortex also involves cognitive fac-
Learning novel associations between stimuli, tors: changes in the auditory cortex and learning
between stimuli and behavioral responses, or do not occur when an animal is passively exposed
between stimuli and reinforcers changes auditory to the stimulus-reward pairings of another animal
cortex, such as the feature sensitivity while it was instrumentally conditioned (Blake
(spectrotemporal receptive field) of neurons in et al. 2006).
Associations and Rewards in Auditory Cortex 205 A
Neuronal Correlates of Association Acknowledgment Supported by the Deutsche
Sustained firing and slow firing changes may Forschungsgemeinschaft (SFB 779) and the Europ€aischer
Fond f€
ur regionale Entwicklung (EFRE 2007–2013).
provide neuronal correlates of associations A
between mental entities (Brosch et al. 2011a).
The main condition necessary for the emergence
of slow firing changes is that subjects have learnt References
that conditioned associations are contingent on
reinforcers. When the reinforcer is removed, the Arsenault J, Nelissen K, Jarraya B, Vanduffel W (2013)
slow firing changes disappear within a few trials, Dopaminergic reward signals selectively decrease
fMRI activity in primate visual cortex. Neuron
concomitantly with behavioral changes. These 77(6):1174–1186
firings may be related to the contingent negative Blake DT, Heiser MA, Caywood M, Merzenich MM
variation (Walter et al. 1964), an event-related (2006) Experience-dependent adult cortical plasticity
potential that can be obtained in electroencepha- requires cognitive association between sensation and
reward. Neuron 52:371–381
lography in humans. It is considered to reflect the Brosch M, Selezneva E, Scheich H (2011a) Formation of
contingency and the contiguity of two events that associations in auditory cortex by slow changes of
have a motivational “value.” tonic firing. Hear Res 271:66–73
Different types of events may be associated Brosch M, Selezneva E, Scheich H (2011b) Representa-
tion of reward feedback in primate auditory cortex.
through event-related slow firing changes Front Syst Neurosci 5:5
(Brosch et al. 2011a). Events that trigger such Creutzfeldt OD (1983) Cortex Cerebri: Leistung,
changes can be auditory (and possibly visual) Strukturelle und Funktionelle Organisation der
stimuli, reinforcers, and behavioral responses, Hirnrinde. Springer, Berlin/Heidelberg/New York
Dudai Y (2004) Memory from A to Z. Oxford University
like grasping. An event with which slow Press, Oxford
firing changes end is the reinforcer. Slow firing Ghazanfar AA, Schroeder CE (2006) Is neocortex
changes have been observed between (1) an audi- essentially multisensory? Trends Cogn Sci
tory stimulus and a reinforcer, (2) a behavioral 10:278–285
R€
usseler J, Nager W, Möbes J, M€ unte TF (2005) Cognitive
response and an auditory stimulus, and adaptations and neuroplasticity: lessons from event-
(3) a behavioral response and another behavioral related brain potentials. In: König R, Heil P, Budinger
response. E, Scheich H (eds) Auditory cortex: towards
The association between behaviorally signifi- a synthesis of human and animal research. Lawrence
Erlbaum, Mahwah, pp 467–484
cant events provided by slow firing changes Scheich H, Brosch M (2013) Task-related activation of the
might even be directed in some cases, that is, auditory cortex. In: Cohen YE, Popper AN, Fay RR
this type of firing might provide some sort of (eds) Neural correlates of auditory cognition, Springer
either prospective coding of an upcoming event handbook of auditory research. Springer, New York/
Heidelberg/Dordrecht/London
or retrospective coding about a preceding event. Scheich H, Brechmann A, Brosch M, Budinger E, Ohl
FW, Selezneva E, Stark H, Tischmeyer W, Wetzel
Neuronal Correlates of Rewards W (2011) Behavioral semantics of learning and
In animals actively engaged in listening to audi- crossmodal processing in auditory cortex: the semantic
processor concept. Hear Res 271:3–15
tory stimuli, direct reflections of rewards have Shepard KN, Kilgard MP, Liu RC (2013) Experience-
been demonstrated in the auditory cortex. Neuro- dependent plasticity and the auditory cortex. In:
nal activity in the auditory cortex varied with the Cohen YE, Popper AN, Fay RR (eds) Neural correlates
size of the delivered reward and the size of the of auditory cognition, Springer handbook of auditory
research. Springer, New York/Heidelberg/Dordrecht/
reward that was expected to be earned in a future London
auditory task, as well as the magnitude of the Shuler MG, Bear MF (2006) Reward timing in the primary
mismatch between the expected and delivered visual cortex. Science 311:1606–1609
reward (the reward prediction error) (Brosch Walter WG, Cooper R, Aldridge VJ, McCallum WC,
Winter AL (1964) Contingent negative
et al. 2011b). Reflections of rewards are also variation: an electric sign of sensorimotor association
present in other sensory cortices (Shuler and and expectancy in the human brain. Nature
Bear 2006; Arsenault et al. 2013). 203:380–383
A 206 Attentional Top-Down Modulation, Models of
and to focus processing resources (and ultimately

Attentional Top-Down Modulation, perception) onto this small fraction. This process
Models of is called attention and for the purpose of this entry
can be defined as the selective modulation of
Philipp Schwedhelm1,3 and Stefan Treue1,2,3 sensory information based on its assumed behav-
1
Cognitive Neuroscience Laboratory, German ioral relevance.
Primate Center, Göttingen, Germany
2
Faculty of Biology and Psychology, Göttingen Bottom-Up Versus Top-Down
University, Göttingen, Germany The selection processes underlying attention
3
Bernstein Center for Computational need to fulfill two requirements: on the one
Neuroscience, Göttingen University, Göttingen, hand their ubiquitous (central and incessant)
Germany role in the continuous stream of perceptual deci-
sions requires that they operate efficiently and as
fast as possible. At the same time, the selection
Definition processes’ purpose of dynamically identifying
the most relevant components of the sensory
Attention – the ability of a sensory system to input demands harnessing as much of the cogni-
facilitate the processing of specific information tive power of the species’ central nervous system
at the expense of disregarding the remainder. as possible.
Bottom-up processes – information These seemingly incompatible demands, effi-
processing in the nervous system that operates cient and fast vs. computationally demanding and
in a feedforward way, advancing from sensory thus slow, have created two flavors of selection:
organs or areas at a low level of the cortical 1. A bottom-up (automatic, exogenous) atten-
processing hierarchy. tional selection that exploits the realization
Top-down influence – modulatory signals in that the most informative aspects of our sen-
the nervous system that originate from areas at sory environments are those where one stimu-
a high level of the cortical processing hierarchy, lus differs from their neighbors in space and
influencing information processing in lower time. This local saliency can be identified and
areas. enhanced by simple feedforward filter mecha-
Saliency – a measure of the magnitude of the nisms embedded throughout the processing of
difference of a stimulus from its neighbors in sensory signals in the nervous system.
space and time. 2. A top-down (voluntary, endogenous) atten-
tional selection that integrates any information
available to the organism about the current
Detailed Description situation to make the most informed decision
about which sensory input component repre-
The Case for Attention sents the most relevant information in the
Evolution has provided humans and other highly given situation.
evolved species with powerful sensory systems. In the visual domain, this distinction is well
While our cortical processing capacity has also illustrated with visual search tasks: If we are
evolved and grown impressively, the torrent of confronted with a fairly homogenous visual
information provided by our sensors far outstrips scene, any outlier will be identified, enhanced,
our ability to process it all. In addition, most of and selected by the continuous parallel computa-
the sensory information picked up at any moment tion of local saliency, creating the perceptual
has little importance for our survival. Complex “pop-out” characteristic of simple search tasks
nervous systems faced with this challenge have where the features of the target stimulus differ
developed sophisticated selection mechanism to substantially from the distribution of features of
identify the most relevant incoming information the distractors. Conversely, a target stimulus,
Attentional Top-Down Modulation, Models of 207 A
which is less distinct, either because it is defined this serial integration process is accomplished by
as a conjunction of more than one feature or means of a top-down, spatial “spotlight” of
because it does not differ substantially from the attention. A
distribution of distractor features, does not pop An alternative account for the pattern of reac-
out, but rather requires a more demanding and tion times in search experiments is offered by the
correspondingly slower selection process. guided search theory (GST, Wolfe 1994), which
does not assume an attentional spotlight. Instead,
Taking a Computational Approach to the top-down attentional signal changes the
Attention weight of activation maps before they are com-
Here we illustrate how the attentional modulat- bined to create a ranking of all present stimuli
ion of sensory information processing is based on their likelihood to represent a target.
implemented in computational models. Due to The selection of stimuli is then again performed
the brevity of the entry, we focus on a few exam- serially, from high to low probability, until the
ples of models of top-down attentional modula- target stimulus is detected.
tion in the visual system of man and other While the FIT and the GST emphasize the role
primates. of feature maps in attentional selection, the the-
One of the most influential computational ory of visual attention (TVA; Bundesen 1990)
models of visual attention is the feature integra- takes a different approach. Here the selection of
tion theory (FIT; Treisman and Gelade 1980). In stimuli is dependent on their processing speed.
the FIT, information about different features of Before a stimulus can be encoded in visual short-
stimulus, such as its shape, color, orientation, and term memory and thus enter awareness, it needs
movement, is extracted in parallel, automatically to compete in a computational race with other
and effortlessly through a system of feature maps, stimuli. In the TVA top-down attention speeds
which topographically represent the spatial dis- up the processing of certain items, making them
tribution of specific features in the visual scene. likely to win the race.
This process detects and locates a target stimulus While the FIT, GST, and TVA have been
defined by a single unique feature value (such as developed to account for the perceptual data
the color red) because it is represented by available at the time, more recent models of
a unique hotspot in a single feature map (with attention have been developed to capture data
each distractor represented by a hotspot in its from single-cell recordings from monkey visual
corresponding feature map, such as the one for cortex. Two early conceptual models attempted
the color blue). This target detection is very to account for the enhanced neuronal response to
quick and is unaffected by the numerosity of attended stimuli and the reduced response to
distractor stimuli, matching the experimental unattended stimuli. The biased competition
observation that human reaction times in such model of attention (Desimone and Duncan
simple search tasks are independent of the 1995) envisages a competition between the stim-
number of distractor items. If the target ulus representation of attended and unattended
stimulus is not defined by a single feature alone, stimuli that can be biased by a top-down atten-
but by a conjunction of multiple features, tional signal in favor of the attended stimulus’
information from different feature maps needs representation. The feature similarity gain
to be integrated to detect and localize a target. model of attention (Treue and Martinez-Trujillo
This requires a serial process that actively inte- 1999) alternatively proposes that the enhance-
grates information from different maps to detect ment of neural responses by attention reflects
the target’s unique feature conjunction at one a process where top-down attentional signals
topographical location, matching the linear enhance the gain of those neurons whose pre-
increase in reaction time observed with an ferred features match the current attentional
increase in the number of distracters in state of the organism, independent of the stimulus
a conjunctive search task. The FIT proposes that that currently activates a neuron.
A 208 Attentional Top-Down Modulation, Models of
These two conceptual models have inspired attended ones. While such an integrated saliency
a large number of computational models. The map. Is consistent with a number of perceptual
most prominent of those are models that empha- phenomena and is ideally suited to guide eye
size an interaction of top-down attention with the movements across a visual scene, it is a matter
normalization process that creates the sigmoidal of some debate which of the many topographi-
contrast response functions typical for neurons cally organized areas in the visual cortex repre-
throughout sensory cortex. Multiple varieties of sents this map or whether multiple such maps
such normalization models of attention have exist.
been proposed (Ghose and Maunsell 2008; Similarly, while functional imaging and
Boynton 2009; Ghose 2009; Lee and Maunsell single-cell recording studies have implicated
2009, 2010; Reynolds and Heeger 2009). They all a network of frontoparietal areas in the guidance
emphasize the similarity, in perception, as well as process that is necessary to appropriately allocate
in the neural encoding and also in the central role processing resources (Kastner and Ungerleider
of the response normalization process between 2001; Corbetta and Shulman 2002), such ana-
two influences on the strength a neural stimulus tomic specificity is rarely included in current
representation. One is the physical (bottom-up) computational models of attention.
strength of the stimulus (most directly
represented by its contrast) and the other is the Conclusion
attentional weight (implemented as a kind of In conclusion, in the last decade, a large number
sensory prior) assigned to them through a top- of computational models of top-down attention
down attentional signal. have been developed that can account for a large
Beyond models that emphasize response nor- variety of perceptual and physiological aspects of
malization, there have been numerous other the attentional modulation of sensory information
approaches to model the attentional modulation processing. These models emphasize several core
of sensory information processing. They include issues, such as the response normalization in cor-
the selective tuning model (Tsotsos et al. 2005) tical networks, the multistage nature of cortical
that proposes a layered network architecture information processing, and the concept of an
(representing the hierarchy of cortical areas) to integrated saliency map. Despite this progress
implement a spatial “spotlight of attention” that much more work is needed to achieve
endows certain regions of the visual scene with a complete computational description of
enhanced processing. The spiking network top-down attentional modulation.
model (Deco and Rolls 2005; Deco and Thiele
2011) places much more emphasis than any of the
models discussed above on building its approach Cross-References
on biological components, such as spiking neu-
rons and specific neurotransmitters. ▶ Hierarchical Models of the Visual System
▶ Working Memory, Models of
The Integrated Saliency Map
It should be noted that almost all models of
attention incorporate the concept of an inte-
References
grated saliency map (Treue 2003), that is,
a topographic representation of the stimuli in Boynton GM (2009) A framework for describing the
the current visual scene that combines their rela- effects of attention on visual responses. Vision Res
tive physical strength and their assumed behav- 49:1129–1143
ioral relevance. This combination implements Bundesen C (1990) A theory of visual attention. Psychol
Rev 97:523–547
a weighing of bottom-up and top-down aspects Corbetta M, Shulman GL (2002) Control of goal-directed
of a stimulus, providing processing resources to and stimulus-driven attention in the brain. Nat Rev
strong unattended stimuli as well as to weak Neurosci 3:201–215
Auditory Event-Related Potentials 209 A
Deco G, Rolls ET (2005) Neurodynamics of biased com-
petition and cooperation for attention: a model with Auditory Brainstem Responses
spiking neurons. J Neurophysiol 94:295–313
Deco G, Thiele A (2011) Cholinergic control of cortical A
network interactions enables feedback-mediated atten- ▶ Auditory Evoked Brainstem Responses
tional modulation. Eur J Neurosci 34:146–157
Desimone R, Duncan J (1995) Neural mechanisms of
selective visual attention. Annual Review of Neurosci-
ence 18(1):193–222 Auditory Event-Related Potential
Ghose GM (2009) Attentional modulation of visual
responses by flexible input gain. J Neurophysiol
(AERP)
101:2089–2106
Ghose GM, Maunsell JH (2008) Spatial summation can ▶ Auditory Event-Related Potentials
explain the attentional modulation of neuronal responses
to multiple stimuli in area V4. J Neurosci 28:5115–5126
Kastner S, Ungerleider LG (2001) The neural basis of
biased competition in human visual cortex. Neuropsy- Auditory Event-Related Potentials
chologia 39:1263–1276
Lee J, Maunsell JH (2009) A normalization model of
attentional modulation of single unit responses. PLoS Istvan Winkler1, Susan Denham2 and
One 4:e4651 Carles Escera3
Lee J, Maunsell JH (2010) Attentional modulation of MT 1
Institute of Cognitive Neuroscience and
neurons with single or multiple stimuli in their recep-
tive fields. J Neurosci 30:3058–3066 Psychology, Research Centre for Natural
Reynolds J, Heeger DJ (2009) The normalization model of Sciences, MTA, Budapest, Hungary
attention. Neuron 61:168–185 2
Cognition Institute and School of Psychology,
Treisman A, Gelade G (1980) A feature-integration theory Plymouth University, Plymouth, Devon, UK
of attention. Cog Psychol 12:97–136 3
Treue S (2003) Visual attention: the where, what, how and Institute for Brain, Cognition and Behavior
why of saliency. Curr Opin Neurobiol 13:428–432 (IR3C), University of Barcelona, Barcelona,
Treue S, Trujillo JCM (1999) Feature-based attention Catalonia, Spain
influences motion processing gain in macaque visual
cortex. Nature 399(6736):575–579
Tsotsos JK, Liu Y, Martinez-Trujillo JC, Pomplun M,
Simine E, Zhou K (2005) Attending to visual motion. Synonyms
Compu Vis Image Underst 100:3–40
Wolfe JM (1994) Guided search 2.0: a revised model of Auditory event-related potential (AERP); Audi-
visual search. Psychon Bull Rev 1:202–238
tory evoked field (AEF); Auditory evoked poten-
tial (AEP)
Attractor Neural Network

Definition
▶ Hopfield Network
Auditory event-related potentials are electric poten-
tials (AERP, AEP) and magnetic fields (AEF) gen-
erated by the synchronous activity of large neural
Auditory Brainstem Evoked populations in the brain, which are time-locked to
Potentials some actual or expected sound event.
▶ Auditory Evoked Brainstem Responses
Auditory Brainstem Potentials Measurement and Derivation of AERPs/AEFs

AERPs are derived from the continuous electro-/
▶ Auditory Evoked Brainstem Responses magnetoencephalogram (EEG/MEG) by
A 210 Auditory Event-Related Potentials
extracting segments of the signal (epochs) time- EEG/MEG signals can contain components up
locked to some actual or expected acoustic event. to a few kHz with the faster components mainly
AERPs were first recorded by Hallowell and Pau- originating from lower levels of the auditory sys-
line A. Davis in 1935–1936 (Davis 1939; Davis tem (cf. “▶ Anatomy and Physiology of the
et al. 1939). Because EEG/MEG is typically Mammalian Auditory System” and “▶ Auditory
recorded non-invasively (outside the brain, e.g., Evoked Brainstem Responses”). Cortical contri-
from/around the scalp), these measures only reflect butions are much slower, up to a few tens of
synchronous activity of large neural populations Hz. Unless one is specifically looking for very
(for measuring methods and instrumentation). slow (Vanhatalo et al. 2005) or fast responses
Consequently, the acoustic events eliciting detect- (Curio 2005), AERP recordings are usually
able AERPs consist of relatively large changes of made with bandpass filter settings of
spectral energy occurring within a relatively short 0.01–50 Hz (or 250 Hz for extracting the
time period, such as abrupt sound onsets, offsets, Middle-Latency Response, see below). AERP
and changes within a continuous sound, because amplitudes are typically below 10 mV with the
large acoustic changes affect many neurons within reference (zero) level set to a baseline voltage
the auditory system and the short transition period (unless direct current is recorded), which is usu-
synchronizes the responses of individual neurons ally the average signal amplitude in a time inter-
(Nunez and Srinivasan 2006; cf. “▶ Anatomy and val preceding the AERP-eliciting event.
Physiology of the Mammalian Auditory System”). Although in general, there is no unique solution
Furthermore, the expectation of such changes in to the inverse problem of finding the origins (the
the auditory input can elicit AERP responses even neural generators) of electromagnetic potentials
in the absence of actual stimulation (cf. the Omitted measured outside the brain, by utilizing anatomy/
Stimulus Response in “Long-Latency AERP physiology based constraints, the generators of
Responses”, below). AERPs can be located with reasonable accuracy
The EEG/MEG signal mixes together in the brain (Nunez and Srinivasan 2006; see also
on-going (spontaneous) neuroelectric activity “▶ Brain Imaging: Overview”). Due to the
with that elicited by the event. In order to better underlying physics, magnetic AEFs provide
estimate the brain activity evoked by the event, it more accurate source localization compared
is usually repeated several times (typically with electric AERPs (Nunez and Srinivasan
50–200 trials/sweeps, but up to 2,000 times for 2006; Nagarajan et al. 2012). On the other hand,
“▶ Auditory Evoked Brainstem Responses”) and MEG only allows one to measure the tangential
the EEG/MEG segments are entered into some components of the electromagnetic activity in the
mathematical algorithm extracting the common brain, whereas EEG represents the full activity
part of the single-trial epochs. The most com- (Hansen et al. 2010). For AEFs, however, this
monly used method for extracting AERPs aligns limitation of the MEG signal is less severe than
the single-trial epochs by their common onset and for other sensory/cognitive systems (Picton 2010;
averages them point by point (the averaging Nagarajan et al. 2012). This is because a large
method; Alain and Winkler 2012). There are part of the human auditory system in the cortex is
many other algorithms for extracting ERPs from located in the Sylvian fissure (see “▶ Anatomy
EEG/MEG, each based on different assumptions and Physiology of the Mammalian Auditory
regarding the properties of the event-related System”), thus mostly producing magnetic sig-
response and the spontaneous EEG/MEG activity nals which can be picked up by the MEG device.
(for a general primer, see Luck 2005; for detailed
discussion of ERPs, see Handy 2005; Fabiani AERP Waves, Components, Naming
et al. 2007; for special considerations of Conventions
MEG/AEFs, see Hansen et al. 2010; Nagarajan Figure 1 illustrates the progression of
et al. 2012; for AERPs, see Picton 2010; Alain stimulus-related neuronal activity through the
and Winkler 2012). auditory system and the corresponding series of
Auditory Event-Related Potentials, Fig. 1 The nuclei (including the cochlear nucleus, the superior
human Auditory Event-Related Potential (AERP), its olivary nucleus, the nucleus of the lateral lemniscus);
main waveforms and its generators in the brain. The “IC”, inferior colliculus; “MGB”, the medial geniculate
human AERP is composed of three groups of waveforms body in the thalamus; “AC”, auditory cortex. The main
in three different latency ranges: the Auditory Brainstem assumed brain sources of the different AERPs are marked
Response (ABR) elicited within the first 8–10 ms from by colored circles: the ascending auditory pathway of the
sound onset (green, bottom panel); the Middle-Latency brainstem for ABRs (green); the thalamo-cortical loops
Response (MLR), elicited within the 12–50 ms interval and parts of auditory cortex for MLRs (blue); the auditory
from sound onset (blue, central panel), and the Long- cortex for LLRs (red). AERPs can be broken down into
Latency Responses (LLR) emerging after 50 ms (red, top a series of waves (see the naming convention in the main
panel). The anatomical inset (left panel) highlights the text)
main stages of the auditory pathway: “bn”, brainstem
positive and negative waveforms observable in The waveforms following are called Long-
the AERP response. The earliest detectable Latency Responses (LLR) and they originate
responses (< ca. 10 ms after the acoustic event) largely from auditory cortex, but may also include
originate from subcortical brain structures and contributions from parietal and frontal areas.
are termed the Auditory Brainstem Response ABRs are referred to by Roman numerals set
(ABR; cf. “▶ Auditory Evoked Brainstem in the order of their elicitation. MLR waveforms
Responses”). These are followed by AERP are usually denoted by their polarity at the vertex
responses of thalamo-cortical origin (mainly (approximately the top of the head); P for positive
from the primary auditory cortex), termed the and N for negative polarity waves, and a letter or
Middle-Latency Response (MLR), elicited during a number (see Fig. 1). There are two conventions
the ca. 10-50 ms post-event latency range. for the numbers in referring to LLRs: They either
denote the serial order of the response starting by N€a€at€anen and Picton’s (1987) definition of an
with the first detected response (Davis 1939; ERP component: ‘. . . we define an EP “compo-
Davis et al. 1939), termed N1, or they denote nent” as the contribution to the recorded wave-
the typical peak latency of the waveform, such form of a particular generator process, such as the
as P50 (the same as Pb or P1). However, as more activation of a localized area of cerebral cortex by
and more responses elicited with the same polar- a specific pattern of input’ (p. 376). Thus
ity and in overlapping latency ranges have been a component is defined by two criteria: (1) it
discovered, both notations have become equivo- should have a specific generator structure (e.g.,
cal. Therefore, some recently discovered AERP secondary auditory and frontal cortices) and (2) it
responses are denoted by acronyms referring to should be specific to some experimentally defin-
their functional aspects, such as ORN (Object able stimulus configuration (such as stimulus
Related Negativity) or MMN (Mismatch Nega- change after several stimulus repetitions). One
tivity) (for a detailed description of the variety of could amend this definition with the person’s
ERP responses, see Luck and Kappenman 2012; task/goals/knowledge regarding the given stimu-
for AERPs, Picton 2010; Alain and Winkler lus configuration (e.g., instructed to respond to
2012). Magnetic response fields are usually the given stimulus event). However, the criteria
marked by the letter ‘m’ appended to the name set up by the above definition are seldom met in
of the corresponding (A)ERP (e.g., N1m or ERP research. This is partly due to limitations in
N100m). separating generators (i.e., they are usually dis-
Beyond the categorization based on the ERP tributed over an area in the brain and concurrently
peak latency there are two other typical distinc- active processes often occupy areas very close,
tions in use. ERPs are termed obligatory or exog- possibly even overlapping each other) as well as
enous if they are elicited by each event not knowing what stimulus configurations are
irrespective of its relation to preceding or concur- handled by the same processes in the brain (are
rent events or the person’s task, motivations, all expectation violations processed in the same
knowledge, etc. ERP components elicited only way? – probably not). Thus in practice, the
when there is a certain relation between the majority of ERP research reports use the terms
event and other events or some aspect of the “waveform” and “component” interchangeably,
person’s mental state are termed endogenous. sometimes linking the effects of multiple manip-
Another distinction refers to the person’s volun- ulations to the same waveform, while at other
tary activity with respect to the given stimulus times, attempting to separate the specific genera-
event. ERP responses only elicited when the per- tor process affected by a given stimulus or state
son has some explicit task involving the event variable.
(task-relevant even) are termed “active” ERP There are many different processes, which can
responses, while those elicited irrespective of be reflected in AERPs. Early, obligatory
the person’s task (task-irrelevant) are termed responses typically reflect processes extracting
“passive” ERP responses. auditory features, such as pitch, intensity, loca-
However, waveforms (peaks and dips) are not tion, etc. Most AERP responses are sensitive to
the true building blocks of ERP responses. The the amount of sound energy change and also to
brain is a massively parallel processing instru- some aspects of the sound presentation rate or the
ment. Therefore, at any given moment of time, ratio between sound and silence in time. These
multiple processes may contribute to the observ- attributes of auditory stimuli belong to the pri-
able waveform. For a neurophysiologically and mary descriptors of sound events as studied in
functionally more meaningful decomposition of psychoacoustics (Zwicker and Fastl 1990). There
the complex neuroelectric response, one should are also AERP responses indicating the presence
be able to delineate how each of the concurrent of automatic memory for sounds (Cowan 1984;
processes contributed to the observed Demany and Semal 2007) and predictive
neuroelectric activity. This objective is reflected processing of the auditory input (Friston and
Kiebel 2009; Winkler et al. 2009). Further, some spectro-temporal characteristics, recording the
AERP responses reflect processes involved in MLR requires (a) averaging across close to
auditory scene analysis (Bregman 1990), the sep- 1,000 responses, (b) appropriate filter settings A
aration of concurrently active sound sources in (15–200 Hz; Bell et al. 2004), and (c) careful
the environment and the formation of auditory removal of electromagnetic interference from
perceptual objects (Griffiths and Warren 2004; power supplies and lines, as a large part of the
Winkler et al. 2009). Many AERP responses are power of the MLR responses falls into the
also sensitive to attentional manipulations, 50–60 Hz range. It is also important to avoid
including the active storage of sounds, selective artefacts stemming from the myoelectric activity
attention, and target identification (Cowan 1988; of the postauricular muscle (PAM), which lies
N€a€at€anen 1990). AERP responses specific to behind the ear and is activated by loud sounds.
music and speech perception are described in This is usually achieved by placing the reference
the corresponding entries (“▶ Music Processing electrode on the neck or the sternum (Bell
in the Brain” and “▶ Electrophysiological Indi- et al. 2004). Optimal sounds for eliciting clear
ces of Speech Processing”). Therefore, AERPs MLRs are chirps and clicks, which have sharp
have been extensively used to test theories of onsets and a broad spectrum. Pure tones elicit
perception (e.g., Bregman 1990; Friston 2005), MLRs of somewhat different morphology and
memory (e.g., Broadbent 1958; Baddeley and smaller amplitude (Borgmann et al. 2001). How-
Hitch 1974; Cowan 2001), and attention (e.g., ever, MLRs can be obtained even with
Broadbent 1958; Lavie 1995) and in recent low-intensity tone bursts and relatively indepen-
years they have received increased interest from dently of the arousal level (Jones and Baxter
computational modelling (e.g., Garrido et al. 1988).
2009; May and Tiitinen 2010; Wacongne et al. No hemispheric asymmetry was found for
2011) as well as from clinical applications (e.g., MLRs as a function of the stimulated ear (Starr
Picton 2010; N€a€at€anen et al. 2012). and Don 1988). Based on precise structural maps
In the following, we shall describe the most of individual brains, the spatiotemporal pattern of
important middle- and long-latency AERP neural activation giving rise to MLRs has been
responses (for the auditory brainstem responses, identified in supratemporal auditory areas using
see “▶ Auditory Evoked Brainstem Responses”). either current estimates derived from intracere-
bral recordings (Yvert et al. 2005) or equivalent
Middle-Latency AERP Responses dipole source modelling of scalp-recorded elec-
Discrete auditory stimuli elicit a sequence of very tric brain potentials (Yvert et al. 2001). These
small (<1 mV) negative and positive waveforms studies localized the earliest cortical activity
in the 10–50 ms post-stimulus latency range, (P0) at 16–19 ms from sound onset in the medial
termed the Middle Latency Response (MLR). portions of Heschl’s sulcus (HS) and Heschl’s
These responses can usually be best seen on sig- gyrus (HG), which likely correspond to primary
nals recorded from the vertex with a mastoid or auditory cortex (PAC). Na generation resulted
neck electrode as reference. The names and typ- from activity in more posterior regions of the
ical latencies of MLRs when elicited by click same HS and HG areas. During the Pa/Pb com-
stimuli are: N0 (10 ms), P0 (15 ms), plex, which includes also the Nb, the electric
Na (20 ms), Pa (30 ms), and Nb (~40 ms) (see brain activity propagates in postero-anterior and
Picton 2010). An additional later waveform, the medio-lateral directions in HG to the Planum
Pb, which peaks at about 50 ms from sound onset, Temporale (PT) and then to more anterior parts
is not always included amongst the MLR compo- of the Superior Temporal Gyrus (STG), which
nents, because it can also be obtained as the P50 correspond to secondary auditory areas. Also,
or P1 with the filter bandwidth optimised for frontal and parietal brain regions contribute
measuring LLRs (Regan 1989; see below). as early as 30 ms from sound onset (the P30m
Because of their small amplitude and specific AEF response) to MLR (Itoh et al. 2000).
Animal studies have suggested that MLRs involve probability in a feature-specific manner (Grimm
parallel thalamocortical activation of areas 41 and Escera 2012) with infrequent frequency
(PAC), and 36 (parahippocampal gyrus), while changes enhancing the Pa (Slabu et al. 2010)
human lesion studies have implicated contribu- and Nb (Grimm et al. 2012; Alho et al. 2012),
tions from thalamic projections to Pa (Kraus whereas location changes enhance the Na
et al. 1982) and Na (Kaseda et al. 1991), (Sonnadara et al. 2006; Grimm et al. 2012;
supporting a thalamo-cortical interaction in MLR Cornella et al. 2012). These results suggest that
generation. the MLR components reflect processes
With increasing sound intensity, MLR com- subserving higher-order sensory/cognitive
ponent latencies decrease while the amplitudes functions.
increase, although these effects may not uni-
formly apply to each component (e.g., Na, but Long-Latency AERP Responses
not Pa; Seki et al. 1993; Althen et al. 2011). The auditory P1 (P50, Pb; Fig. 1) component is at
Galambos et al. (1981) found a systematic the border between MLR and LLR. In fact, when
reversed U-shaped relationship between the recorded and analysed with the filter setting most
MLR amplitudes and stimulus presentation rate. useful for deriving MLRs it is termed the Pb (see
At slow rates (10 Hz), peak-to-trough ampli- “Middle-Latency AERP Responses”, above).
tudes are rather small (0.4 mV) and they reach the Using the parameters better suited for assessing
maximum of 1 mV by about 40 Hz presentation LLRs, it typically peaks at about 50 ms from
rate. This twofold increase in amplitude is due to stimulus onset, appearing with positive polarity
superimposition of MLRs elicited by successive at the vertex and with reversed (negative) polarity
sounds. In contrast, at stimulation rates below and at electrodes placed on the other side of the
above 40 Hz out-of-phase responses to succes- Sylvian fissure (e.g., electrodes placed over
sive MLR responses cancel out each other. Some the mastoid apophysis). P1 is the first wave of
authors interpret this finding in terms of the the P1-N1-P2 obligatory exogenous AERP com-
“steady state” potentials (oscillatory activity gen- plex. It is thought to be generated bilaterally in
erated in sensory cortical areas that is time- primary auditory cortex, somewhat larger contra-
locked to the periodicity of stimulus presentation; than ipsilaterally for pure tones (Godey et al.
typically measured from visual and somatosen- 2001) and for other types of pitch-evoking sounds
sory cortical areas; Rees et al. 1986). Other (Butler and Trainor 2012), with some spreading
authors assume that this phenomenon reflects of the neuroelectric activity over its time course
the contribution of transient early evoked (Yvert et al. 2005). P1 is often used as a landmark
gamma-band oscillations to the auditory MLR for primary auditory cortex in AERP and AEF
(Basar et al. 1987; Pantev et al. 1991; M€uller studies aimed at localizing the AERP compo-
et al. 2001). Based on the stimulus-driven prop- nents. Similarly to other obligatory AERP
erties outlined above, MLRs have been consid- responses, P1 is highly sensitive to stimulus fea-
ered exogenous AERP components. However, tures and presentation rate (fully recovering
this view has been challenged by studies showing within a few hundred milliseconds) as well as to
that MLRs are enhanced by strongly focused attentional manipulations (Picton 2010). The P1
attention as early as 20 ms from sound onset was initially assumed to reflect neural activity
(Woldorff and Hillyard 1991; Woldorff involved in extracting auditory features
et al. 1993; cf. “Attention-Related AERP (e.g. N€a€at€anen and Winkler 1999). Recent evi-
Responses” below), and that MLR amplitudes dence also links this response with the automatic
are modulated as early as 50 ms from sound separation of auditory streams (Gutschalk
onset by task difficulty and whether or not et al. 2005; Snyder et al. 2006; Szalárdy
a motor response is required (Ninomiya et al. 2013; cf. “▶ Auditory Perceptual Organiza-
et al. 1997). Further, a recent series of studies tion”): The amplitude of the P1 component has
has shown that MLRs are sensitive to stimulus been found to be modulated by whether
a sequence with two interleaved sounds (e.g., with auditory stream segregation (Gutschalk
ABABAB. . ., where ‘A’ and ‘B’ denote two dif- et al. 2005; Snyder et al. 2006; Szalárdy
ferent sounds) was perceived as a single coherent et al. 2013). For example, the length of the silent A
stream or in terms of two concurrent streams of period after which an N1 with maximal ampli-
sound (one made up of the ‘A’ and the other by tude is elicited by a sound is in good correspon-
the ‘B’ sounds). dence with the behaviourally measurable
The auditory N1 (N100; Fig. 1) wave was the duration of auditory sensory memory traces
first AERP response discovered historically (Cowan 1984). When sounds are presented in
(Davis et al. 1939) as it is the most prominent a train with <10 s silent intervals between them,
deflection at the vertex. It is elicited by abrupt the N1 amplitude decreases sharply within the
changes in sound energy, such as sound onsets first few presentations, reaching an asymptote
and offsets (N€a€at€anen and Picton 1987). N1 typ- within 5–10 presentations (e.g., Cowan
ically peaks with negative polarity over the ver- et al. 1993). Based on this finding, some authors
tex ca. 100 ms after the eliciting event. It is also argue that through adaptation (see “▶ Adaptation
the most widely studied AERP response, having in Sensory Cortices, Models of”), the neurons
been linked with virtually any and all assumed underlying the N1 response may retain all sound
auditory processing steps. The N1 wave has information and thus provide the basis for
a complex generator (and thus subcomponent) detecting violations of auditory regularities
structure (N€a€at€anen and Picton 1987). The (May and Tiitinen 2010). However, this hypoth-
subcomponent most tightly related to auditory esis is debated in the literature (e.g., N€a€at€anen
processes (the supratemporal N1) is mostly et al. 2011). The sensitivity of the auditory N1
located in secondary auditory areas (Godey wave to selective attention initially suggested
et al. 2001), but it also overlaps the areas active that the difference between the N1 responses
during the P1 component (Yvert et al. 2005). elicited by task-relevant (attended) and task-
Similarly to the P1, N1 is larger contralaterally irrelevant (unattended) sounds (the Nd; Hillyard
to the ear of stimulation and it is highly sensitive et al. 1973) may reflect an orientation to the
to stimulus features, presentation rate, and atten- attended auditory features and/or maintenance
tional manipulations. However, unlike the P1, the of the memory trace of the target sound. How-
N1 recovery is much slower, extending beyond ever, others argued that the differential response
10 s (Cowan et al. 1993). Further, N1 is sensitive is separate from the N1, with the early part
to perceived sound features (e.g., pitch), as overlapping the N1 (termed Nde) assumed to
opposed to raw spectral parameters (such as the reflect feature processing, and the later part
harmonic frequencies of a complex tone; Pantev (Ndl, also termed the Processing Negativity, PN;
et al. 1989b), although feature extraction is not N€a€at€anen 1982; see PN in “Attention-Related
yet complete at the time the N1 wave is elicited AERP Responses”) the maintenance of the atten-
(Winkler et al. 1997). The supratemporal N1 also tional trace (Koch et al. 2005; N€a€at€anen
shows both tonotopic (Pantev et al. 1988) and et al. 2011).
ampliotopic organization (Pantev et al. 1989a); Little is known about the auditory P2 (P175,
that is, the location of its generator varies with the P200; Fig. 1) AERP response. It has been mostly
frequency and amplitude of pure tones. However, studied within the P1-N1-P2 or N1-P2 complex.
the N1 generators are not sensitive to combina- P2 typically peaks between 175 and 200 ms from
tions of sound features (i.e., feature the event onset with positive polarity over the
conjunctions). vertex, inverting polarity over the Sylvian fissure.
The processes reflected by N1 have been The generators of P2 lie anterior to those of the
linked with onset and acoustic change detection N1 in secondary auditory areas (M€akel€a
(N€a€at€anen 1992), feature extraction, sensory et al. 1988; Bosnyak et al. 2004). Lesion
memory (L€ u et al. 1992; at least for sound fea- (Woods et al. 1993) and maturation studies
tures, N€a€at€anen and Winkler 1999) and, recently, (Ponton et al. 2000) suggest that P2 may reflect
the output of the mesencephalic reticular activat- AERP responses (Arnott et al. 2001). Some stud-
ing system (see “▶ Anatomy and Physiology of ies have indicated the existence of two indepen-
the Mammalian Auditory System”). Only a few dent lateralized generator processes, since
studies have attempted to distinguish P2 from the although ORN is elicited even when most tones
N1 wave. The P2 amplitude was found to be more in the sequence have been mistuned, the proba-
sensitive to perturbing the feedback of one’s bility of mistuned sounds within the sequence
voice than the N1 (Behroozmand et al. 2009) as differentially affected the ORN generators in the
well as to training with specific types of sounds two hemispheres (Bendixen et al. 2010). If the
(e.g., speech: Tremblay et al. 2001; music: listener is instructed to respond when he/she
Bosnyak et al. 2004; or frequency discrimination: hears two concurrent sounds, a late positive
Tong et al. 2009). There are several speculations response (P400) is elicited in addition to the
regarding the functions of the processes reflected ORN (Fig. 2, panel B right; Alain et al. 2001).
by P2. Based on its assumed neural origin, P2 has The auditory N2 (N200; Fig. 1) wave covers at
been suggested to be generated by a pre-attentive least three (N2a or MMN, N2b, N2c; see Pritchard
alerting mechanism (Tremblay and Kraus 2002). et al. 1991), possibly more AERP components
Other suggestions include P2 reflecting stimulus (Folstein and Van Petten 2008) appearing partly
classification (Crowley and Colrain 2004), mod- overlapping in time between 150 and 300 ms
ulating the threshold for conscious perception from the eliciting event. The somewhat earlier
(Melara et al. 2002), protecting against interfer- N2a or MMN does not require attention to be
ence from irrelevant stimuli (Garcia-Larrea focused on the event (cf. MMN), whereas the
et al. 1992), and the accuracy of memory traces later components are related to attentive monitor-
in short-term memory (Atienza et al. 2002). ing of the acoustic input and they are not specific
The Object Related Negativity (ORN) is to sounds.
elicited when more than one sound are simulta- The Mismatch Negativity (MMN, N2a) is an
neously heard (Alain et al. 2001). Thus ORN AERP component elicited by violations of audi-
reflects the outcome of the analysis of simulta- tory regularities (Winkler 2007; N€a€at€anen
neous (concurrent or vertical) auditory grouping et al. 2011; Fig. 3). MMN typically emerges
cues (cf. “▶ Auditory Perceptual Organization”). between 100 and 200 ms from the onset of devi-
Components of sounds emitted by a single source ation with frontocentrally dominant negative
usually commence at the same time, they origi- polarity that is inverted over the Sylvian fissure.
nate from the same spatial location and, if com- MMN generators are located bilaterally in
posed of discrete frequencies, they consist of secondary-auditory and frontal areas (Alho
harmonics derived from the same base (i.e., inte- 1995). Although traditionally regarded as
ger multiples of the same frequency). When the a component reflecting auditory change detec-
acoustic input does not meet these criteria, one tion, technically, MMN does not reflect acoustic
usually experiences it as two or more concurrent change, as for example, an alternating sequence
sounds and ORN is elicited. ORN is typically of sounds does not elicit the MMN, whereas
recorded by presenting complex tones with one repeating a sound within such a sequence does
harmonic mistuned by 4 % or more (Fig. 2, (Horváth et al. 2001). MMN is derived by
panel A) and derived by subtracting the response subtracting from the response elicited by the
to the one-sound stimulus (e.g., tuned tone) from regularity-violating sound (termed “deviant”)
that to the multiple-sound stimulus (e.g., the response elicited by a control sound. Opti-
mistuned tone). ORN peaks between 140 and mally, the control sound is either identical or very
180 ms from sound onset, with the largest ampli- similar to the deviant sound but does not violate
tude over the fronto-central region of the scalp any auditory regularity (for a detailed discussion
(Fig. 2, panel B left). ORN has bilateral neural of selecting the correct control, see Kujala
generators in auditory cortex, which are separate et al. 2007). MMN is elicited even when the
from those of the previously described obligatory sounds are task-irrelevant, although it can be
Auditory Event-Related Potentials, Fig. 2 Object active condition (participants judged whether they heard
Related Negativity (ORN) (a) Complex tones with the one or two concurrent tones). Mistuned-minus-tuned dif-
second of five harmonics tuned (green) or mistuned ference waveforms (black) show a negative waveform
upwards by 8 % (red) were presented equiprobably in appearing between 100 and 200 ms from sound onset in
a sequence. (b) Group-averaged (N = 20, left; N = both task conditions. This is the ORN response (the range
23, right) AERP responses elicited by tuned and mistuned is marked by grey shading). The positive difference wave-
complex tones recorded at the vertex, separately in the form observed in the 300-500 ms latency range in the
passive (participants disregarded the sounds) and the Active Condition is termed the P400
suppressed by strongly focusing attention on it plays a role in auditory stream segregation

a parallel auditory channel and/or by contextual (cf. “▶ Auditory Perceptual Organization”) by
information (Sussman 2007). Initially discovered maintaining the predictive models underlying
within the oddball paradigm (N€a€at€anen auditory perceptual objects (Winkler et al. 2009).
et al. 1978), MMN has since been observed for The Repetition Positivity (RP) appears as a -
violations of a large variety of abstract and com- fronto-central amplitude modulation of the P50,
plex regularities (N€a€at€anen et al. 2001). In paral- N1 and P2 AERP responses (Fig. 4); all three of
lel, its interpretation shifted from MMN being an them overlap the slow positive RP waveform so
AERP correlate of auditory sensory memory that the P50 and P2 become more positive and the
(N€a€at€anen and Winkler 1999; Cowan 1984) N1 less negative with increasing number of rep-
tasked with detecting potentially relevant events etitions of the eliciting sound (Haenschel
in the auditory environment (N€a€at€anen 1992) et al. 2005; Costa-Faidella et al. 2011a, b). Sim-
towards the compatible but more general notion ilar stimulus repetition effects have been
of representing a process that updates the observed even at shorter latencies, during the
detected auditory regularities when their MLR latency range (Dyson et al. 2005). The RP
predictions are not met by the incoming sound was first observed by Baldeweg et al. (2004) and
(Winkler 2007). The latter interpretation links characterized by Haenschel et al. (2005) in
MMN with predictive coding theories (Friston a study that aimed at investigating the neural
2005; Winkler and Czigler 2012) and posits that correlates of the sensory memory trace
Auditory Event-Related Potentials, Fig. 3 The sound (dashed grey line) measured from the frontal mid-
Mismatch Negativity (MMN). (a) The experimental line electrode. The MMN component, derived by
setup. Participants watched and listened to a movie subtracting the control response from that to the deviant
presented on a TV screen directly in front of them. (difference: black line) is marked with yellow-orange fill
A series of footsteps perceived as moving from left to in the MMN latency range. The results illustrate that
right (Test Sequence; upper arrow) or right to left (1) MMN is only elicited when a sound violates
(Control Sequence; lower arrow) were delivered by a detected regularity, as the regular progression of foot-
a pair of loudspeakers placed symmetrically on two steps needed to be detected and represented by the brain
sides, slightly behind the participant’s head. Ten out of before it could be violated (which could not happen if only
the 11 different digitized natural footstep sounds (marked one “regular” footstep sound preceded the different one);
as black footprints on the blue arrows) could be perceived (2) regularities can be extracted from acoustic variance as
as a coherent sequence produced by someone walking all regular footstep sounds were acoustically different;
across a room. The 10th footstep of the Test Sequence (3) regularities are separately maintained for concurrent
(“deviant”) and the 2nd footstep of the Control Sequence auditory streams, as MMN was elicited for deviation in the
(“control”) however sounded as if the person stepped on footstep stream despite the presence of two other active
a different surface (marked by the white footprint on the sound sources; and (4) MMN elicitation does not require
blue arrows). Street noise was delivered through one to attend the stream in which a regularity has been
a loudspeaker placed directly behind the participant. (b) violated, as participants in this experiment attended the
Group-averaged (N = 8) AERP responses elicited by the movie, not the footsteps (Adapted from Winkler
deviant (continuous grey line) and the identical control et al. 2003)
implicated in the generation of the MMN. It preceding the deviant (e.g., Sams et al. 1983;
was argued that the MMN amplitude dependence Javitt et al. 1998) provides only an indirect
on the number of standard-stimulus repetitions measure of the strength of the underlying
Auditory Event-Related Potentials, Fig. 4 The Repe- pronounced by the 12th repetition. Right: Difference
tition Positivity (RP). Left: Group-averaged (N = 14) waveforms resulting from subtracting the response to the
frontal midline (marked on the schematic head drawing 3rd repetition from that to the 12th repetition under two
at the top right corner) AERPs elicited by pure tones in conditions: Predictable Timing (PT: isochronous presen-
a roving-standard stimulus paradigm (see in the text). The tation, blue) and Unpredictable Timing (UT: the within-
panel shows AERPs (averaged across different frequen- train inter-onset interval was varied, red). Note that the
cies) elicited for the 3rd (blue), 6th (red) and 12th (green) onset of RP is earlier (ca. 70 ms post-stimulus) for the
repetition of the same tone. Note that the positivity cov- predictable than for the unpredictable timing condition
ering the latency range of the P50-N1-P2 waveform com- (ca. 170 ms) (Adapted from Costa-Faidella et al. 2011a)
plex emerges at the 6th repetition and becomes more
memory trace. The AERP elicited by the standard et al. 2011a, b). The generator structure of the RP
sound was expected to show effects of repetition has not yet been fully characterized, but its early
suppression (Desimone 1996), as was observed onset latency (commencing during the P50) and
for individual neurons in the primary auditory its long duration (ending during the P2) suggest
cortex of the cat (Ulanovsky et al. 2003), and that it may involve a distributed cortical network
this could provide a more direct measure of the spanning from PAC up to auditory association
strength of standard-stimulus memory trace. The areas (Baldeweg 2007). The RP has been shown
typical paradigm use for obtaining the RP is to simultaneously encode repetitions over multi-
called the “roving-standard” paradigm ple time scales (Costa-Faidella et al. 2011b; Coo-
(introduced by Cowan et al. 1993), as the classi- per et al. 2013) similarly to single neurons
cal oddball paradigm yields less clear results observed in the cat’s PAC (Ulanovsky
(Cooper et al. 2013). In the roving-standard par- et al. 2004). In addition to stimulus repetition,
adigm, short trains of a repeating sound are deliv- the RP is also sensitive to temporal regularities,
ered without a break with each train delivering such as whether the sounds are presented isochro-
a different sound (e.g., pure tones with different nously or with random timing: Costa-Faidella
frequencies). The number of sound repetitions et al. (2011a) found earlier and larger RP’s for
can also vary from train to train. To separate the isochronous as compared with randomly timed
RP from other concurrent AERP components, tones in the trains. The latter result supports the
the average response elicited by the second or predictive coding view of auditory deviance
the third sound of the train is subtracted from detection (Winkler 2007; Winkler and Czigler
that elicited by the last tone of the train. The 2012), according to which detection of
response to the first sound of the train is not a regularity helps to encode the sensory memory
used in the subtraction, because, due to the trace of upcoming stimuli. Thus higher levels in
sound change between the trains, it should elicit the auditory processing hierarchy feed back to
the MMN (Haenschel et al. 2005; Costa-Faidella lower processing levels (Baldeweg 2006).
Auditory brain responses can also be elicited (described above), or at least its frontal compo-
without hearing sounds. By omitting sounds from nent (Giard et al. 1990; Deouell et al. 1998;
an isochronous sequence, one can record poten- Escera et al. 2000a; Deouell 2007), has been
tials time-locked to the moment when the associated with involuntary attention (N€a€at€anen
sequence would have continued in a regular man- and Michie 1979; N€a€at€anen 1990, 1992). Some
ner. The responses are termed the Omitted Stim- studies have also related the activation of the
ulus Response (OSR). Some of them are elicited supratemporal MMN generator with behavioural
even when listeners don’t focus on the sounds, correlates of involuntary attention, i.e., delayed
thus demonstrating a basic tendency of the audi- response times to target stimuli on a primary task
tory system to generate predictions for incoming (Yago et al. 2001). N€a€at€anen and Michie (1979)
sounds (Friston 2005; Winkler et al. 2009). It has proposed that the process generating MMN may
been shown that when all features of the upcom- issue a call for focal attention (Öhman 1979)
ing sound can be predicted from the preceding upon the detection of an unexpected change in
sound sequence, the OSR elicited by sound omis- the acoustic environment. Initial supportive evi-
sion during the first 50 ms does not differ from the dence was provided by Schröger (1996; Schröger
AERP elicited by the sound itself; however, when and Wolff 1998a) and Escera et al. (1998), who
only the timing of the sound can be predicted, but introduced new auditory-auditory and auditory-
not its features, the OSR starts to differ from the visual distraction paradigms (for a more recent
corresponding AERP at an earlier time (Bendixen design, see Horváth and Winkler 2010). In these
et al. 2009). When sounds are predictably caused paradigms, participants are instructed to perform
by some action of the listener, occasionally omit- a primary auditory or visual task while ignoring
ting one elicits an AERP that is initially (up to rare task-irrelevant violations of an auditory reg-
ca. 100 ms) morphologically similar to that ularity. Several studies have shown that these rare
elicited by the corresponding self-initiated deviations prolong the reaction time and reduce
sound; although the brain generators underlying the hit rate to target stimuli in the primary task
the two responses partly differ from each other (Escera and Corral 2007), thus demonstrating
(SanMiguel et al. 2013). There is also an involuntary attention switching to the task-
MMN-like OSR (Yabe et al. 1999). Elicitation irrelevant deviations.
of these responses is limited to inter-onset- Following the MMN, AERPs recorded in the
intervals (IOI) shorter than ca. 200 ms (Horváth distraction paradigm display a fronto-central pos-
et al. 2007), except when the omitted sound is itive deflection ca. 250–350 ms from stimulus
part of a pattern (Salisbury 2012). With longer onset, termed the P3a or novelty-P3. P3a was
IOIs, an early posterior negative (180 –280 ms) first described by Squires et al. (1975) as an
response and a later anterior positive wave have earlier and more frontal positive deflection com-
been obtained (Busse and Woldorff 2003). Fur- pared to the later and more posterior P3b compo-
ther, ERP responses can also be elicited by men- nent (for a review on P3b, see Donchin and Coles
tal imagery of sounds, although the results vary 1988). Whereas P3a is elicited by rare task-
somewhat with the procedure employed (Meyer irrelevant sounds, P3b is elicited by target sounds
et al. 2007; Cebrian and Janata 2010; Wu (for a detailed comparison between the P3a and
et al. 2011). P3b, see Polich 2007). P3a is also elicited by
widely different and “novel” (unique, categori-
Attention-Related AERP Responses cally different from the context) sounds (Knight
Attention-related AERPs include two distinct 1984), hence it is sometimes referred to as the
groups of responses: those related to involuntary novelty-P3 (for a discussion of whether the P3a
(passive or exogenous) attention, and those and the novelty-P3 can be considered as the same
related to voluntary, mainly selective attention. ERP component, see Simons et al. 2001). Com-
Regarding involuntary attention, at least three pelling evidence linking the novelty-P3 to the
components have to be considered. The MMN orienting reflex (OR; Sokolov 1963) was
obtained by Knight (1996), who found strong (but not MMN) are eliminated or at least strongly
correlation between the novelty-P3 and one of diminished when the task-irrelevant deviant is
the well-known autonomic components of OR, predicted by a visual cue (Sussman et al. 2003; A
the galvanic skin response (GSR). The P3a is Horváth and Bendixen 2012). Cues that provide
composed of two subcomponents distinctly dif- more specific information about the distracting
fering in latency (early and late), scalp distribu- stimulus are more effective in preventing distrac-
tion, and sensitivity to attentional manipulations tion and the elicitation of P3a and RON (Horváth
(Escera et al. 2000a; Yago et al. 2003). Source et al. 2011).
modelling of the magnetic counterpart of P3a Selective attention related AERPs have been
(P3am) elicited by auditory deviants and novel traditionally studied in the context of the classical
sounds has revealed a genuine auditory cortical “cocktail-party” situation described by Cherry
contribution to the early part of P3a (Alho (1953). In the simplified dichotic listening
et al. 1998). Whereas the early part of the model of this situation, participants are exposed
novelty-P3 appears to be insensitive to attentional to two concurrent messages (one to each ear).
manipulations (Escera et al. 1998), the later part Using this paradigm, many studies attempted to
is modulated by working memory (SanMiguel decide between the “early” (Treisman 1964;
et al. 2008) and emotional load (Domı́nguez- Treisman 1998; Broadbent 1970) versus “late”
Borràs et al. 2008). The early P3a is sensitive to selection theories of attention (Deutsch and
stimulus-specific information predicting task- Deutsch 1963; Norman 1968). These theories of
irrelevant auditory deviance, whereas the late attention primarily differ from each other in the
P3a appears to be more closely correlated with placement of a selective filter within the chain of
distraction (Horváth et al. 2011). P3a is widely information processing (Broadbent 1958):
regarded as a correlate of attention switching whereas early selection theories suggest that
(Escera et al. 2000a; Friedman et al. 2001). How- stimuli are selected for elaborate processing
ever, some recent studies suggested that although based on simple sensory features (such as pitch)
P3a is probably an antecedent of attention and unattended stimuli do not receive processing
switching it can be elicited without beyond extracting these sensory features, late
a corresponding shift in the focus of attention selection theories propose that all stimuli receive
(Rinne et al. 2006; Horváth et al. 2008b; Horváth elaborate processing and stimuli can therefore be
and Winkler 2010; Hölig and Berti 2010). selected on the basis of higher-order properties.
The third involuntary attention related AERP (Note that more recent theories of attention do not
component is the so-called Reorienting Negativ- posit a single selective filter; see e.g., Lavie
ity (RON), first described by Schröger and Wolff 1995.) The seminal observation by Hillyard
(1998b). RON is observed as a negative deflec- et al. (1973) that selective attention enhances
tion following the P3a (Escera and Corral 2007). the N1 amplitude for stimuli presented in the
RON has been suggested to reflect processes of to-be-attended channel favoured the early filter-
reorientation (restoring the task set of the primary ing view. However, the findings of N€a€atanen
task) after a distracting stimulus. RON is com- et al. (1978) of a long-lasting negativity elicited
posed of two subcomponents (Escera et al. 2001; by all attended stimuli, the Processing Negativity
Munka and Berti 2006; Berti 2008) the functional (PN; N€a€at€anen 1982) challenged this interpreta-
characterization of which are still debated tion providing support to late-selection theories.
(Escera et al. 2001; Berti 2008). The cortical Subsequent studies confirmed both of these
generators of RON are not well known. Horváth effects (Okita 1979; Hansen and Hillyard 1980;
et al. (2008a) found contributions from primary N€a€at€anen et al. 1980) and proposed subtraction of
motor areas to RON, suggesting that action- the AERP elicited by the non-attended stimuli
selection related activity plays a role in the from that elicited by the attended stimuli as the
reorientation process. Both P3a and RON as method to reveal the Negative Difference (Nd)
well as behavioural correlates of distraction potential to isolate the AERP correlates of
selective attention (Nd; Hansen and Hillyard from that elicited by words, which are semanti-
1980). The Nd is composed of two parts: the cally incongruent with respect to the preceding
early one, termed Nde, associated with a gating context (D’Arcy et al. 2004; Van den Brink and
mechanism preferentially processing the task- Hagoort 2004). Violating speech syntax can lead
relevant stimulus features, and a later part (Ndl) to the elicitation of the Early Left Anterior Neg-
related to the maintenance of the attentional trace ativity (ELAN) in the 150-200 or the Left Anterior
(correspond to the PN). The functional distinc- Negativity (LAN) in the 300-500 ms latency
tion between the Nd and PN has been debated in range, depending on the type of violation,
detail (Alho et al. 1986a; Alho et al. 1986b; Alho whereas potentially correct but syntactically
et al. 1994; Teder et al. 1993). Studies showing complex sentences elicit the Syntactic Positive
very early selective attention effects, e.g., at the Shift (SPS or P600) (for reviews, see Friederici
latency range of the MLR (Woldorff et al. 1987; 2002; Hagoort 2008). Violating semantic expec-
Woldorff and Hillyard 1991) and possibly even tations in speech elicits the N400 component
earlier, at the level of the cochlea (Giard (Kutas and Federmeier 2011). Musical syntax
et al. 1994) support the interpretation of the Nde violations elicit an ELAN-like but predominantly
as a correlate of gating by simple stimulus fea- right-hemispheric response, the Early Right
tures. On the other hand, the fact, that the more Anterior Negativity (ERAN) in the 180-200 ms
similar the stimulus to the target the longer the or the Right Anterior-Temporal Negativity
corresponding PN, supports the notion of (RATN) in the 200-400 ms latency range and
a comparison with the attentional trace. The fron- N400 has been also be observed in musical
tal scalp distribution of Ndl (Woods and models of semantic incongruence (Koelsch and
Clayworth 1987) and the cerebral sources of PN Siebel 2005). For a more detailed discussion of
(Giard et al. 1988) are also compatible with the speech- and music-related ERPs, see “▶ Electro-
memory-based interpretation of Ndl. There are physiological Indices of Speech Processing” and
several further ERP components related to vari- “▶ Music Processing in the Brain”.
ous facets of attention. However, these are not
specific to the auditory modality and thus fall Development of AERPs
outside the scope of this entry. Previous sections described the AERP responses
elicited in adults. Although AERPs can be
AERPs Reflecting Speech and Music recorded immediately after birth and even in foe-
processing tuses within the womb (Draganova et al. 2005),
The sounds of speech and music may elicit any their morphology and functional characteristics
and all the AERP responses described above. widely differ from the adult responses. Further,
There are, however, also some ERP responses, different AERP components become mature at
which arise from events that can be defined in different times and they often undergo several
syntactic or semantic terms. It should be noted intermediate phases before reaching adult-like
that most speech-related ERPs can also be characteristics. As this topic would require a full
elicited through reading. Most AERP responses entry of its own, here we point the reader to some
specific to speech and music have been obtained of the existing literature. The most complete
in paradigms, in which the expectation for the reviews of the maturation of AERPs from infancy
most likely (or simplest) continuation of to adolescence were provided by Wunderlich
a sequence of words has been violated. For exam- et al. (2006) and Coch and Gullick (2012). The
ple, violating the expectation for the first pho- early infantile development of the AERP compo-
neme of the upcoming word elicits a negative nents has been summarized by Kushnerenko
shift in the 150-350 ms latency range, termed (2003); for the maturation of the AERPs
the Phonological Mismatch Negativity (PMN; reflecting auditory change detection, see Jing
Connolly and Phillips 1994). It is, however, and Benasich (2006), for large deviations, see
debated, whether this response can be separated Kushnerenko et al. (2013). The maturation of
obligatory AERP components from 5 to 20 years Adaptation was modelled by the exponential
of age is covered in Ponton et al. (2000, 2002). decay and recovery of the internal state variable
AERP maturation during adolescence is associated with each stimulus frequency, and the A
described in Bishop et al. (2007). Summarizing response function as a read out of the internal
these works, one can conclude that the adult state corresponding to the incoming stimulus.
AERP morphology characterizes humans from The internal state for the prediction error, novelty
17/18 years onward and remains more or less detection, and model adjustment accounts was
unchanged through ageing. There are, however modelled as a Bayesian observer’s belief in the
several findings of differences between elderly tone category of the stimulus, with the evolution
and young adults in specific tasks (for a review, of tone category modelled according to
see Friedman 2012). a transition matrix derived incrementally from
the data according to the ‘free-energy-
Modelling AERP’s: Some General Principles minimisation principle’ (Friston 2005). Two pre-
Theories that seek to explain some of the LLRs diction error response functions were modelled:
have also been explored using more tightly prediction errors with respect to sensory input
constrained mathematical and computational and internal state, respectively. Novelty response
models. Here we focus on models of the functions were modelled as surprise about sen-
mismatch negativity (MMN) component, as it sory input and temporal structure (tone category),
has arguably received the most widespread atten- respectively. Model adjustment response func-
tion. Theoretically, MMN has been variously tions were modelled in terms of adjustments to
associated with change detection, adaptation, the parameters of the internal model, e.g. mean
prediction error, novelty detection, and model frequency of a category, expected sequence
adjustment, although for some years, there has length, transition probabilities between catego-
been controversy as to whether anything more ries. Simulations showed that, at least at this
than adaptation is required to explain the experi- level of detail, prediction error (with respect to
mental data (e.g., see May and Tiitinen 2010 tone category) and model adjustment models
vs. N€a€at€anen et al. 2011). (change in expected sequence length, change in
Using a modelling framework in which exem- transition probabilities between categories),
plars of each of the competing explanations, accounted best for the data (Lieder et al. 2013).
listed above, were expressed as mathematical On the other hand, May and Tiitinen (2010)
functions of stimulation-induced changes in an have argued strongly that their neural model
unobservable ‘internal state’ and resulting which includes adaptation on the inputs can
observable (EEG) responses, Lieder explain all MMN data to date; the key mechanism
et al. (2013) investigated the ability of each being the activation of fresh afferents by stimuli
model to explain empirical MMN responses on that deviate in some way from the standards. In
a trial-by-trial basis. The models were expressed this account, MMN is seen as a modulation of the
in a rather abstract way, as summarized below, N1 component rather than as a separate compo-
with simple expressions for internal state and nent in its own right. The model, consisting of
response functions (intended to predict a bank of neural oscillators driven via adapting
stimulus-evoked MMN amplitudes), that cap- input synapses, can account for the latency as
tured a range of possibilities for each of the cat- well as the amplitude of the MMN (May
egories. Change detection was modelled with the et al. 1999). In addition, extending the model to
internal state simply a record of the log frequency include local inhibitory feedback circuits, results
of the previous tone in the sequence, and response in a set of non-homogeneous band-pass temporal
functions as: a) a flag, set if a difference was filters that can also support the topographic rep-
detected, b) the signed and c) absolute difference resentation of stimulus presentation rate (May
between the frequency of the incoming and pre- and Tiitinen 2001). Ringing in these filters is
vious tone; giving three change detection models. argued to account for the MMN elicited by
a missing expected sound. Diverse receptive population, and the progress of activity through
fields, e.g. to frequency modulations, also allow this population explicitly represents the timing of
the model to simulate MMN responses elicited by the previous event, up to 400 ms. Separate delay
violations of some abstract rules, such as lines are used for each tone frequency modelled,
a repeated tone in a random pattern of ascending thereby also recording their identity. The model
tone pairs. However, although adaptation is simulates MMN by means of prediction errors.
claimed to be the key to MMN, the model Through exposure to tone sequences it learns to
responses also depend upon the amplification of generate a prediction of the next tone (both its
recurrent excitation, lateral inhibition, and the timing and identity) in a repeating pattern. These
connectivity of the network. The model thus predictions are compared with the incoming stim-
essentially contains within its changing pattern uli in the prediction error units, where mis-
of adaptation and inhibition, a memory trace of matches result in a larger signal than matches.
recent activation, and in this sense, contains The model learns transition probabilities between
a memory component embedded within it. successive events, as long as they fit within its
Building on their previous work on a brain- fixed memory span. In contrast to the adaptation
inspired architecture for learning long-term rep- account of MMN, the model relies exclusively on
resentations of action-perception associations, prediction errors. An experiment designed to dis-
Garagnani and Pulverm€uller (2011) proposed tinguish between these two explanations for
a similar model in which, in addition to adapta- MMN found evidence in favour of a predictive
tion and inhibition, spreading activation through error model of MMN (Wacongne et al. 2012),
circuits strengthened by learning (long term a result compatible with the findings of Lieder
memory) caused MMN responses to familiar et al. (2013).
deviants to be larger than that to unfamiliar devi- A predictive coding account of MMN has also
ants. They pointed out that only through some been modelled at a more abstract level using
form of long term memory mechanism could this a Kalman filtering (Kalman 1960) approach
differential sensitivity of MMN to familiarity/ (Kaya and Elhilali 2013). In this case the timing
unfamiliarity be explained. By modelling multi- of events is modelled using a separate filter from
ple auditory areas they also provided a novel the one used to model feature distributions. The
explanation for differences between the N1 and advantage of the Kalman filter is that it provides
MMN generators, with N1 being generated in a well-understood way to recursively estimate the
primary auditory areas subject to strong adapta- system state, refined through analysis of prediction
tion, and MMN in addition to adaptation also errors, and has been shown to be implementable in
being influenced by reverberating excitation the form of a neural network (Szirtes et al. 2005).
within distributed memory circuits. However, The model adapts to the variance in observations
the model processes sequences of static patterns, and, with time, as its predictions improve so its
and as presented, it is not able to account for the tolerance decreases, making it more sensitive to
sensitivity of MMN to unexpected changes in the outliers. Deviants are detected as events not
timing of sequences, such as the omission MMN predicted by any existing filter, and trigger the
(Yabe et al. 1997). creation of a new set of Kalman filters intended
A model that explicitly includes a separate to model a potentially new sound source, making
memory module to keep track of the short term this an interesting framework for more general
history of activation and simulates MMN at auditory scene analysis problems, e.g.
a finer level of granularity, i.e. at the level of (Chakrabarty and Elhilali 2013).
spiking neurons, was proposed by Wacongne In summary, computational models of the the-
et al. (2012). Memory in the model is oretical accounts of MMN have begun to be
implemented using a set of neurons organised explored. However, so far they have either only
into a delay line, i.e. their connectivity ensures been implemented at a rather abstract level; e.g.
that activity passes in one direction across the (Garrido et al. 2009; Lieder et al. 2013; Kaya and
Elhilali 2013), focus exclusively on a single used to determine conduction times along the
mechanism for explaining MMN; e.g. (May and auditory pathway and to localize the anatomical
Tiitinen 2001; Wacongne et al. 2012) or account locus of the brain damage with the help of the A
only for MMN responses to unexpected within- known origin of the different ABR waveforms
event properties (Garagnani and Pulverm€uller (see reviews in Baloh 1997; Chiappa 1997;
2011; Garagnani et al. 2008). The finding, using Lustig et al. 2003). In addition, ABRs are used
dynamic causal modelling, that modifications to in combination with evoked potentials from other
both feed forward and feedback connections are modalities to monitor coma prognosis (Guérit
required (Garrido et al. 2009), and evidence in 2005; Fischer et al. 2006; see below), or in isola-
auditory cortex for adaptation, short term and tion to corroborate brain death (Machado
long term plasticity, recurrent excitation and inhi- et al. 1991). In the surgical theatre, MLR is used
bition suggests that MMN in the brain may actu- to monitor the depth of anaesthesia in adults (Bell
ally depend on the combination of all these et al. 2004) and children (Kuhnle et al. 2013). It
factors. Furthermore, while the learning of tran- has been recently shown that, compared with the
sition probabilities may be sufficient for some traditional clinical assessment of depth of anaes-
scenarios, in the short term at least, people thesia, MLR monitoring led to a reduction in
become sensitive to specific tone patterns; it is (a) the amount anaesthetic drug requirement,
unclear whether any of the models discussed here (b) the use of vasopressors to manage hypoten-
could respond differentially to violations of more sion, and (c) consequential cognitive impairment
extended pattern sequences or more abstract (Jildenstål et al. 2011).
rules. Regarding cognitive AEPRs, MMN (see
above) has shown great promise for potential
Utility of AERP for Clinical Practice clinical applications (N€a€at€anen and Escera
Clinical applications of AERPs range from rou- 2000). Part of this expectation stems from the
tine practice in audiology, neurotology, neurol- fact that MMN indexes auditory discrimination
ogy, and neurosurgery by ABRs and MLRs accuracy without the requirement to perform
(Picton 2010) to highly promising tools for cog- some task (i.e., it can be recorded without the
nitive assessment by some long-latency endoge- patient’s collaboration and even in newborn
nous components, of which MMN is a prime infants; see Alho et al. 1990) and that it can be
example. In audiology, ABRs are used univer- elicited very reliably, compared with other cog-
sally for hearing screening in neonates failing the nitive event-related potentials (Escera and Grau
Otoacoustic Emission test (OAEs; Robinette and 1996; Escera et al. 2000b). Yet, after two decades
Glatkke 2007). Currently, about 97 % of infants of clinical research (see N€a€at€anen et al. 2012),
are screened for hearing impairment in the USA except for coma monitoring and prognosis no
(Gaffney et al. 2010). ABRs, elicited by click routine clinical application has emerged for the
stimuli, are used as a tool for objective audiome- MMN. As for coma monitoring, it has been dem-
try, and ABRs elicited by pure tones can also be onstrated that the presence of MMN in
used for assessing frequency-specific thresholds a comatose patient is associated with the return
in infants (Stapells and Oates 1997; Stapells of consciousness (Kane et al. 1993; Fischer
et al. 1993). In neurotology and neurology, et al. 1999), and that as part of a battery of
AERPs are combined with the patient’s medical physiological indicators of brain activity, MMN
history and with an extensive battery of tests for can be used in the decision tree for estimating
evaluating the anatomy and functional properties awakening from coma (Fischer et al. 2006).
of the ear-brain relationship (Picton 2010) in Given the large variety of disorders and clinical
search for an extensive range of disorders of the conditions in which impaired MMNs have been
ear and the auditory pathway, such as Ménière’s observed, it has been suggested that, rather than
disease and demyelinating lesions such as Multi- providing a specific diagnostic measure for any
ple Sclerosis. In these applications, AERPs are particular disease, the MMN provides an
objective index of dysfunction of N-metyl- time-frequency analysis of the EEG, provide bet-
D-aspartate (NMDA) receptor-mediated cogni- ter information about these types of processes.
tive functions (N€a€at€anen et al. 2011). In general, AERPs are usually smaller than their visual coun-
due to their high variability and complex func- terparts. Consequently the signal to noise ratio,
tional and anatomical origin, endogenous AERPs where activity not time-locked to the sound onset
can only be employed within large test batteries is regarded as noise, is quite low. This forces one
for diagnostic and monitoring purposes. How- to present many trials to the participant and rely
ever, some of these responses provided new on assumptions which are not fully met by the
insights into the cognitive and emotional aspects EEG signal (such as the independence of the
of various neurological and psychiatric disorders signal from the noise, ergodicity, etc.). Further,
(e.g., for schizophrenia research using MMN, see the accuracy of localizing the sources of
Mondragón-Maya et al. 2011). neuroelectric activity is limited by the quality of
constraints (e.g., anatomical knowledge) required
AERPs: Advantages and Limitations to solve the inverse problem, and the dispersion
(A)ERPs provide information about sound- of the electrical fields. Although magnetoenceph-
elicited neural activity with millisecond accu- alography provides a better spatial resolution, as
racy. Thus they are ideally suited for breaking was already mentioned, AEFs only reflect tan-
down the steps of auditory information gential sources, but not radial ones, thus
processing in the brain in the empiricist tradition. restricting their general usefulness. In terms of
It is thus understandable that some of the most spatial accuracy, other neuroimaging methods,
recent theoretical developments in the field (e.g., such as fMRI, provide a superior alternative
predictive coding theories; Friston 2005) trace (at the cost of much lower temporal resolution).
back their roots to Helmholtz’ (1860/1962) theo- Further, the correspondence between perception
ries of perception. High temporal resolution and AERP responses is often not straightforward,
coupled with the possibility of finding the neural as can be gleaned from the often controversial
generators of the various ERP responses is also psychological interpretations mentioned in the
appealing to neurologists and medical doctors, in main text of this entry. Few AERP components
general. By finding correlations between AERPs can be consistently observed across different
and conscious perception on the one hand (such stimulus paradigms, thus limiting the validity of
as the link between ORN and the perception of most process-based interpretations. Efforts to dis-
two concurrent sounds; Alain et al. 2001), and cover the neural bases of ERP responses must
discovering the neural mechanisms underlying overcome many obstacles. One of the most diffi-
the observed AERP waveforms on the other cult problems is that whereas individual neurons
hand (e.g., linking SSA and the deviance- can mainly be studied in animal models due to the
detection responses observed in the MLR latency invasive nature of such investigations, it is often
range; Slabu et al. 2010; Grimm et al. 2011; for difficult to assess how well findings in various
a review, see Ayala and Malmierca 2013), species can be extended to characterizing the
AERPs can provide a crucial link in understand- human brain. Finally, the biggest issue for clini-
ing the neural mechanisms of perception. cal applications is, as was already mentioned, the
However, there are a number of limitations to large inter- and even intra-individual variability
the utility of (A)ERPs for research and applica- of AERPs.
tions. Firstly, they only reflect a part of the infor- In summary, AERPs can potentially provide
mation processing in the brain. When the number much information about sound processing in the
of neurons involved in some process is relatively brain, but for extracting this information, better
small, or the neurons are distributed over a large theories and more tightly constrained models,
area in the brain, or the neural activation is not which can integrate information from the diverse
fully time-locked to the given auditory event, no fields of anatomy, neuroscience, and psychology,
ERP can be measured. Other methods, such as are required.
Cross-References Atienza M, Cantero JL, Dominguez-Marin E (2002) The
time course of neural changes underlying auditory
perceptual learning. Learn Mem 9:138–150
▶ Adaptation in Sensory Cortices, Models of Ayala YA, Malmierca MS (2013) Stimulus-specific adap- A
▶ Anatomy and Physiology of the Mammalian tation and deviance detection in the inferior colliculus.
Auditory System Front Neural Circuits 6:89
▶ Auditory Evoked Brainstem Responses Baddeley AD, Hitch GJ (1974) Working memory. In:
Bower GA (ed) Recent advances in learning and moti-
▶ Auditory Perceptual Organization vation, vol 8. Academic, New York, pp 47–90
▶ Brain Imaging: Overview Baldeweg T (2006) Repetition effects to sounds: evidence
▶ Electrophysiological Indices of Speech for predictive coding in the auditory system. Trends
Processing Cogn Sci 10:93–94
Baldeweg T (2007) ERP repetition effects and mismatch
▶ Music Processing in the Brain negativity generation: a predictive coding perspective.
J Psychophysiol 21:204–213
Baldeweg T, Klugman A, Gruzelier J, Hirsch SR
References (2004) Mismatch negativity potentials and cognitive
impairment in schizophrenia. Schizophr Res
Alain C, Winkler I (2012) Recording event-related brain 69:203–217
potentials: application to study auditory perception. In: Baloh RW (1997) Dizziness, hearing loss, and tinnitus.
Poeppel D, Overath T, Popper AN, Fay RR (eds) The Oxford University Press, New York
human auditory cortex. Springer handbook of auditory Başar E, Rosen B, Başar-Eroglu C, Greitschus F (1987)
research, vol 43. Springer, New York, pp 69–96 The associations between 40 Hz-EEG and the middle
Alain C, Arnott SR, Picton TW (2001) Bottom-up and latency response of the auditory evoked potential. Int
top-down infl uences on auditory scene analysis: evi- J Neurosci 33:103–117
dence from event-related brain potentials. J Exp Behroozmand R, Karvelis L, Liu H, Larson CR (2009)
Psychol Hum Percept Perform 27:1072–1089 Vocalization-induced enhancement of the
Alho K (1995) Cerebral generators of mismatch negativity auditory cortex responsiveness during voice F0
(MMN) and its magnetic counterpart (MMNm) feedback perturbation. Clin Neurophysiol
elicited by sound changes. Ear Hear 16:38–51 120:1303–1312
Alho K, Paavilainen P, Reinikainen K, Sams M, N€a€at€anen Bell SL, Smith DC, Allen R, Lutman ME (2004) Record-
R (1986a) Separability of different negative compo- ing the middle latency response of the auditory evoked
nents of the event-related potential associated with potential as a measure of depth of anaesthesia.
auditory stimulus processing. Psychophysiology A technical note. Br J Anaesth 92:442–445
23:613–623 Bendixen A, Schröger E, Winkler I (2009) I heard that
Alho K, Sams M, Paavilainen P, N€a€at€anen R (1986b) coming: event-related potential evidence for stimulus-
Small pitch separation and the selective attention driven prediction in the auditory system. J Neurosci
effect on the ERP. Psychophysiology 23:189–197 29:8447–8451
Alho K, Sainio K, Sajaniemi N, Reinikainen K, N€a€at€anen Bendixen A, Jones SJ, Klump G, Winkler I (2010) Prob-
R (1990) Event-related brain potential of human new- ability dependence and functional separation of the
borns to pitch change of an acoustic stimulus. object-related and mismatch negativity event-related
Electroencephalogr Clin Neurophysiol 77:151–155 potential components. Neuroimage 50:285–290
Alho K, Teder W, Lavikainen J, N€a€at€anen R (1994) Berti S (2008) Cognitive control after distraction: event-
Strongly focused attention and auditory event-related related brain potentials (ERPs) dissociate between
potentials. Biol Psychol 38:73–90 different processes of attentional allocation. Psycho-
Alho K, Winkler I, Escera C, Huotilainen M, Virtanen J, physiology 45:608–620
J€a€askel€ainen I, Pekkonen E, Ilmoniemi R (1998) Bishop DVM, Hardiman M, Uwer R, von Suchodoletz
Processing of novel sounds and frequency changes in W (2007) Maturation of the long-latency auditory
the human auditory cortex: magnetoencephalographic ERP: step function changes at start and end of adoles-
recordings. Psychophysiology 35:211–224 cence. Devel Sci 10:565–575
Alho K, Grimm S, Mateo-León S, Costa-Faidella J, Escera Borgmann C, Ross B, Draganova R, Pantev C (2001)
C (2012) Early processing of pitch in the human audi- Human auditory middle latency responses:
tory system. Eur J Neurosci 36:2972–2978 influence of stimulus type and intensity. Hear Res
Althen H, Grimm S, Escera C (2011) Fast detection of 158:57–64
unexpected sound intensity decrements as revealed by Bosnyak DJ, Eaton RA, Roberts LE (2004) Distributed
human evoked potentials. PLoS ONE 6:e28522 auditory cortical representations are modified when
Arnott SR, Bardouille T, Ross B, Alain C (2001) Neural non-musicians are trained at pitch discrimination
generators underlying concurrent sound segregation. with 40 Hz amplitude modulated tones. Cereb Cortex
Brain Res 1387:116–124 14:1088–1099
Bregman AS (1990) Auditory scene analysis. The percep- within the human information processing system.
tual organization of sound. MIT Press, Cambridge, Psychol Bull 104:163–191
MA Cowan N (2001) The magical number 4 in short-term
Broadbent DE (1958) Perception and communication. memory: a reconsideration of mental storage capacity.
Pergamon Press, New York Behav Brain Sci 24:87–114
Broadbent DE (1970) Stimulus set and response set: two Cowan N, Winkler I, Teder W, N€a€at€anen R (1993) Short-
kinds of selective attention. In: Mostofsky DI and long-term prerequisites of the mismatch negativity
(ed) Attention, contemporary theory and analysis. in the auditory event related potential (ERP). J Exp
Appleton Century Crofts, New York, pp 51–60 Psychol Learn Mem Cogn 19:909–921
Busse L, Woldorff MG (2003) The ERP omitted stimulus Crowley KE, Colrain IM (2004) A review of the evidence
response to “no-stim” events and its implications for for P2 being an independent component process: age,
fast-rate event-related fMRI designs. Neuroimage sleep and modality. Clin Neurophysiol 115:732–744
18:856–864 Curio G (2005) Ultrafast EEG activities. In:
Butler BE, Trainor LJ (2012) Sequencing the cortical Niedermeyer E, Lopes da Silva FH (eds) Electroen-
processing of pitch-evoking stimuli using EEG analy- cephalography. Basic principles, clinical applications,
sis and source estimation. Front Psychol 3:180 and related fields, 5th edn. Lippincott Williams &
Cebrian AN, Janata P (2010) Electrophysiological corre- Wilkins, Philadelphia, pp 495–504
lates of accurate mental image formation in auditory D’Arcy RCN, Ryner L, Richter W, Service E, Connolly JF
perception and imagery tasks. Brain Res 1342:39–54 (2004) The fan effect in fMRI: left hemisphere spe-
Chakrabarty D, Elhilali M (2013) Predictive analysis of cialization in verbal working memory. Neuroreport
two tone stream segregation via extended Kalman 15:1851–1855
filter. In: Proceedings of 2013 47th annual conference Davis PA (1939) Effects of acoustic stimuli on the waking
information science system (CISS), doi:10.1109/ human brain. J Neurophysiol 2:494–499
CISS.2013.6552279 Davis H, Davis PA, Loomis AL, Harvey EN, Hobart
Cherry EC (1953) Some experiments on the recognition of G (1939) Electrical reactions of the human brain to
speech, with one and with two ears. J Acoust Soc Am auditory stimulation during sleep. J Neurophysiol
25:975–979 2:500–514
Chiappa KH (1997) Evoked potentials in clinical medi- Demany L, Semal C (2007) The role of memory in audi-
cine. Lippincott Williams & Wilkins, Philadelphia tory perception. In: Yost WA, Popper AN, Fay RA
Coch M, Gullick MM (2012) Event-related potentials and (eds) Auditory perception of sound sources. Springer
development. In: Luck SJ, Kappenman ES (eds) The handbook of auditory research, vol 29. Springer, New
Oxford handbook of event-related potential compo- York, pp 77–113
nents. Oxford University Press, New York, Deouell LY (2007) The frontal generator of the mismatch
pp 475–511 negativity revisited. J Psychophysiol 21:188–203
Connolly JF, Phillips NA (1994) Event-related potential Deouell LY, Bentin S, Giard MH (1998) Mismatch nega-
components reflect phonological and semantic tivity in dichotic listening: evidence for
processing of the terminal words of spoken sentences. interhemispheric differences and multiple generators.
J Cogn Neurosci 6:256–266 Psychophysiology 35:355–365
Cooper RJ, Atkinson RJ, Clark RA, Michie PT Desimone R (1996) Neural mechanisms for visual mem-
(2013) Event-related potentials reveal modelling of ory and their role in attention. Proc Natl Acad Sci U S
auditory repetition in the brain. Int J Psychophysiol A 93:13494–13499
88:74–81 Deutsch JA, Deutsch D (1963) Attention: some theoretical
Cornella M, Leung S, Grimm S, Escera C (2012) Detec- considerations. Psychol Rev 70:80–90
tion of simple and pattern regularity violations occurs Domı́nguez-Borràs J, Garcia-Garcia M, Escera C (2008)
at different levels of the auditory hierarchy. PLoS One Negative emotional context enhances auditory novelty
7:e43604 processing. Neuroreport 19:503–507
Costa-Faidella J, Baldeweg T, Grimm S, Escera C (2011a) Donchin E, Coles MGH (1988) Is the P300 component
Interactions between “what” and “when” in the audi- a manifestation of context updating? Behav Brain Sci
tory system: temporal predictability enhances repeti- 11:357–374
tion suppression. J Neurosci 31:18590–18597 Draganova R, Eswaran H, Murphy P, Huotilainen M,
Costa-Faidella J, Grimm S, Slabu LM, Dı́az-Santaella F, Lowery C, Preissl H (2005) Sound frequency change
Escera C (2011b) Multiple time scales of adaptation in detection in fetuses and newborns,
the auditory system as revealed by human evoked a magnetoencephalographic study. Neuroimage
potentials. Psychophysiology 48:774–783 28:354–361
Cowan N (1984) On short and long auditory stores. Dyson BJ, Alain C, He Y (2005) I’ve heard it all before:
Psychol Bull 96:341–370 perceptual invariance represented by early cortical
Cowan N (1988) Evolving conceptions of memory stor- auditory-evoked responses. Brain Res Cogn Brain
age, selective attention, and their mutual constraints Res 23:457–460
Escera C, Corral MJ (2007) Role of mismatch negativity Garagnani M, Wennekers T, Pulverm€ uller F (2008)
and novelty-P3 in involuntary auditory attention. A neuroanatomically grounded Hebbian-learning
J Psychophysiol 21:251–264 model of attention-language interactions in the
Escera C, Grau C (1996) Short-term replicability of the human brain. Eur J Neurosci 27:492–513 A
mismatch negativity. Electroencephalogr Clin Garagnani M, Pulverm€ uller F (2011) From sounds to
Neurophysiol 100:549–554 words: a neurocomputational model of adaptation,
Escera C, Alho K, Winkler I, N€a€at€anen R (1998) Neural inhibition and memory processes in auditory change
mechanisms of involuntary attention to acoustic nov- detection. Neuroimage 54:170–181, Erratum in:
elty and change. J Cogn Neurosci 10:590–604 Garagnani M, Pulverm€ uller F (2011) Neuroimage
Escera C, Alho K, Schröger E, Winkler I (2000a) Invol- 55:435–436
untary attention and distractibility as evaluated with Garcia-Larrea L, Lukaszewicz AC, Mauguiere F (1992)
event-related brain potentials. Audiol Neurootol Revisiting the oddball paradigm. Non-target
5:151–166 vs. neutral stimuli and the evaluation of ERP atten-
Escera C, Yago E, Polo MD, Grau C (2000b) The individ- tional effects. Neuropsychology 30:723–741
ual replicability of mismatch negativity at short and Garrido MI, Kilner JM, Stephan KE, Friston KJ
long inter-stimulus intervals. Clin Neurophysiol (2009) The mismatch negativity: a review of underly-
111:546–551 ing mechanisms. Clin Neurophysiol 120:453–463
Escera C, Yago E, Alho K (2001) Electrical responses Giard MH, Perrin F, Pernier J, Peronnet F (1988) Several
reveal the temporal dynamics of brain events during attention-related waveforms in auditory areas:
involuntary attention switching. Eur J Neurosci a tropographical study. Electroencephalogr Clin
14:877–883 Neurophysiol 69:371–384
Fabiani M, Gratton G, Federmeier KD (2007) Event- Giard MH, Perrin F, Pernier J, Bouchet P (1990) Brain
related brain potentials: methods, theory, and applica- generators implicated in processing of auditory stimu-
tions. In: Cacioppo JT, Tassinary LG, Berntson GG lus deviance. A topographic ERP study. Psychophys-
(eds) Handbook of psychophysiology, 3rd edn. Cam- iology 27:627–640
bridge University Press, Cambridge, pp 85–119 Giard MH, Collet L, Bouchet P, Pernier J (1994) Auditory
Fischer C, Morlet D, Bouchet P, Luaute J, Jourdan C, selective attention in the human cochlea. Brain Res
Salord F (1999) Mismatch negativity and late auditory 633:353–356
evoked potentials in comatose patients. Clin Godey B, Schwartz D, de Graaf JB, Chauvel P, Liegeois-
Neurophysiol 110:1601–1610 Chauvel C (2001) Neuromagnetic source
Fischer C, Luauté J, Némoz C, Morlet D, Kirkorian G, localizationof auditoryevokedfields and intracerebral
Mauguière F (2006) Improved prediction of awaken- evoked potentials: a comparison of data in the same
ing or non-awakening from severe anoxic coma using patients. Clin Neurophysiol 112:1850–1859
tree-based classification analysis. Crit Care Med Griffiths TD, Warren JD (2004) What is an auditory
34:1–5 object? Nat Rev Neurosci 5:887–892
Folstein JR, Van Petten C (2008) Influence of cognitive Grimm S, Escera C, Slabu LM, Costa-Faidella J (2011)
control and mismatch on the N2 component of the Electrophysiological evidence for the hierarchical
ERP: a review. Psychophysiology 45:152–170 organization of auditory change detection in the
Friederici AD (2002) Towards a neural basis of auditory human brain. Psychophysiology 48:377–384
sentence processing. Trends Cogn Sci 6:78–84 Grimm S, Escera C (2012) Auditory deviance detection
Friedman D (2012) The components of aging. In: Luck SJ, revisited: evidence for a hierarchical novelty system.
Kappenman ES (eds) The Oxford handbook of event- Int J Psychophysiol 85:88–92
related potential components. Oxford University Grimm S, Recasens M, Althen H, Escera C (2012)
Press, New York, p 512 Ultrafast tracking of sound location changes as
Friedman D, Cycowicz YM, Gaeta H (2001) The novelty revealed by human auditory evoked potentials. Biol
P3: an event-related brain potential (ERP) sign of the Psychol 89:232–239
brain’s evaluation of novelty. Neurosci Biobehav Rev Guérit JM (2005) Evoked potentials in severe brain injury.
25:355–373 Prog Brain Res 150:415–426
Friston K (2005) A theory of cortical responses. Phil Trans Gutschalk A, Micheyl C, Melcher JR, Rupp A, Scherg M,
R Soc Lond Ser B Biol Sci 360:815–836 Oxenham AJ (2005) Neuromagnetic correlates of
Friston K, Kiebel S (2009) Cortical circuits for perceptual streaming in human auditory cortex. J Neurosci
inference. Neur Netw 22:1093–1104 25:5382–5388
Gaffney M, Eichwald J, Grosse SD, Mason CA (2010) Iden- Haenschel C, Vernon DJ, Dwivedi P, Gruzelier JH,
tifying infants with hearing loss -United States, Baldeweg T (2005) Event-related brain potential cor-
1999–2007. Morb Mortal Wkly Rep 59:220–223 relates of human auditory sensory memory-trace for-
Galambos R, Makeig S, Talmachoff PJ (1981) A 40-Hz mation. J Neurosci 25:10494–10501
auditory potential recorded from the human scalp. Hagoort P (2008) The fractionation of spoken language
Proc Natl Acad Sci U S A 78:2643–2647 understanding by measuring electrical and magnetic
brain signals. Phil Trans R Soc Lond B Biol Sci Jones LA, Baxter RJ (1988) Changes in the auditory
363:1055–1069 middle latency responses during all-night sleep record-
Handy TC (2005) Event related potentials: a methods ing. Brit J Audiol 22:279–285
handbook. Bradford/MIT Press, Cambridge, MA Kalman RE (1960) A new approach to linear filtering and
Hansen JC, Hillyard SA (1980) Endogenous brain poten- prediction problems. J Basic Eng 82:35–45
tials associated with selective auditory attention. Kane NM, Curry SH, Butler SR, Cummings BH
Electroencephalogr Clin Neurophysiol 49:277–290 (1993) Electrophysiological indicator of awakening
Hansen PC, Kringelbach ML, Salmelin R (eds) from coma. Lancet 341:688
(2010) MEG: an introduction to methods. Oxford Kaseda Y, Tobimatsu S, Morioka T, Kato M (1991) Audi-
University Press, New York tory middle-latency responses in patients with local-
Helmholtz H (1860/1962) Handbuch der Physiologischen ized and non-localized lesions of the central nervous
Optik. Southall JPC (ed) English Translation system. J Neurol 238:427–432
vol 3. Dover, New York Kaya E, Elhilali M (2013) A model of auditory
Hillyard SA, Hink RF, Schwent VL, Picton TW deviance detection. In: Proceedings of 2013 47th
(1973) Electrical signs of selective attention in the Annual Conference on Information Science and Sys-
human brain. Science 182:177–180 tem (CISS), Baltimore, MD, doi:10.1109/
Hölig C, Berti S (2010) To switch or not to switch: brain CISS.2013.6552254
potential indices of attentional control after task- Knight RT (1984) Decreased response to novel stimuli
relevant and task-irrelevant changes of stimulus fea- after prefrontal lesions in man. Electroencephalogr
tures. Brain Res 1345:164–175 Clin Neurophysiol 59:9–20
Horváth J, Bendixen A (2012) Preventing distraction by Knight RT (1996) Contribution of human hippocampal
probabilistic cueing. Int J Psychophysiol 83:342–347 region to novelty detection. Nature 383:256–259
Horváth J, Winkler I (2010) Distraction in a continuous- Koch D, Sanders LD, Neville HJ (2005) An event-related
stimulation detection task. Biol Psychol 83:229–238 potential study of selective auditory attention in chil-
Horváth J, Czigler I, Sussman E, Winkler I (2001) Simul- dren and adults. J Cogn Neurosci 17:605–622
taneously active pre-attentive representations of local Koelsch S, Siebel WA (2005) Towards a neural basis of
and global rules for sound sequences. Cogn Brain Res music perception. Trends Cogn Sci 9:578–584
12:131–144 Kraus N, Ozdamar O, Hier D, Stein L (1982) Auditory
Horváth J, Czigler I, Winkler I, Teder-S€alej€arvi WA middle latency responses (MLRs) in patients with cor-
(2007) The temporal window of integration in elderly tical lesions. Electroencephalogr Clin Neurophysiol
and young adults. Neurobiol Aging 28:964–975 54:275–287
Horváth J, Maess B, Berti S, Schröger E (2008a) Primary Kuhnle GE, Hornuss C, Lenk M, Salam AP, Wiepcke D,
motor area contribution to attentional reorienting after Edelmann-Gahr V, Flake G, Daunderer M,
distraction. Neuroreport 19:443–446 Oberhauser M, M€ uller HH, Feuerecker M (2013)
Horváth J, Winkler I, Bendixen A (2008b) Do N1/MMN, Impact of propofol on mid-latency auditory-evoked
P3a, and RON form a strongly coupled chain reflecting potentials in children. Br J Anaesth 110:1001–1009
the three stages of auditory distraction? Biol Psychol Kujala T, Tervaniemi M, Schröger E (2007) The
79:139–147 mismatch negativity in cognitive and clinical neuro-
Horváth J, Sussman E, Winkler I, Schröger E (2011) science: theoretical and methodological consider-
Preventing distraction: assessing stimulus-specific ations. Biol Psychol 74:1–19
and general effects of the predictive cueing of deviant Kushnerenko EV (2003) Maturation of the cortical audi-
auditory events. Biol Psychol 87:35–48 tory event-related brain potentials in infancy. Doctoral
Itoh K, Yumoto M, Uno A, Kurauchi T, Kaga K (2000) dissertation, University of Helsinki, Helsinki. https://
Temporal stream of cortical representation for audi- helda.helsinki.fi/handle/10138/19818
tory spatial localization in human hemispheres. Kushnerenko E, Van den Bergh BRH, Winkler I (2013)
Neurosci Lett 292:215–219 Separating acoustic deviance from novelty during the
Javitt DC, Grochowski S, Shelley AM, Ritter W (1998) first year of life: a review of event-related potential
Impaired mismatch negativity (MMN) generation in evidence. Front Psychol 4:595
schizophrenia as a function of stimulus deviance, prob- Kutas M, Federmeier KD (2011) Thirty years and
ability, and interstimulus/interdeviant interval. counting: finding meaning in the N400 component of
Electroencephalogr Clin Neurophysiol 108:143–153 the event-related brain potential (ERP). Ann Rev
Jildenstål PK, Hallén JL, Rawal N, Gupta A, Berggren Psychol 62:621–647
L (2011) Effect of auditory evoked potential-guided Lavie N (1995) Perceptual load as a necessary condition
anesthesia on consumption of anesthetics and early for selective attention. J Exp Psychol Hum Percept
postoperative cognitive dysfunction: a randomized Perform 21:451–468
controlled trial. Eur J Anaesth 28:213–219 Lieder F, Daunizeau J, Garrido MJ, Friston KJ,
Jing H, Benasich AA (2006) Brain responses to tonal Stephan KE (2013) Modelling trial-by-trial changes
changes in the first two years of life. Brain Dev in the mismatch negativity. PLoS Comput Biol 9:
28:247–256 e1002911
L€u ZL, Williamson SJ, Kaufman L (1992) Behavioral N€a€at€anen R, Escera C (2000) Mismatch negativity
lifetime of human auditory sensory memory predicted (MMN): clinical and other applications. Audiol
by physiological measures. Science 258:1668–1670 Neurootol 5:105–110
Luck SJ (2005) An introduction to the event -related N€a€at€anen R, Michie PT (1979) Early selective attention A
potential technique. MIT Press, Cambridge, MA effects on the evoked potential. A critical review and
Luck SJ, Kappenman ES (eds) (2012) Oxford handbook of reinterpretation. Biol Psychol 8:81–136
event-related potential components. Oxford Univer- N€a€at€anen R, Picton TW (1987) The N1 wave of the human
sity Press, New York electric and magnetic response to sound: a review and
Lustig LR, Niparko J, Minor LB, Zee DS (eds) an analysis of the component structure. Psychophysi-
(2003) Clinical neurotology: diagnosing and manag- ology 24:375–425
ing disorders of hearing, balance and the facial nerve. N€a€at€anen R, Winkler I (1999) The concept of auditory
Martin Dunitz, London stimulus representation in cognitive neuroscience.
Machado C, Valdés-Sosa P, Garcı́a-Tigera J, Virues T, Psychol Bull 125:826–859
Biscay R, Miranda J, Coutin P, Román J, Garcı́a N€a€at€anen R, Gaillard AWK, M€antysalo S (1978) Early
O (1991) Brain-stem auditory evoked potentials and selective attention effect on evoked potential
brain death. Electronecephalogr Clin Neurophysiol reinterpreted. Acta Psychol 42:313–329
80:392–398 N€a€at€anen R, Gaillard AWK, M€antysalo S (1980) Brain
M€akel€a JP, Hari R, Leinonen L (1988) Magnetic responses potential correlates of voluntary and involuntary atten-
of the human auditory cortex to noise/square wave tion. In: Kornhuber AHM, Deecke L (eds) Motivation,
transitions. Electroencephalogr Clin Neurophysiol motor and sensory processes of the brain: electrical
69:423–430 potentials, behavior and clinical use. Elsevier, Amster-
May P, Tiitinen H (2001) Human cortical processing of dam, pp 343–348
auditory events over time. Neuroreport 12:573–577 N€a€at€anen R, Tervaniemi M, Sussman E, Paavilainen P,
May PJC, Tiitinen H (2010) Mismatch negativity (MMN), Winkler I (2001) ‘Primitive intelligence’ in the audi-
the deviance-elicited auditory deflection, explained. tory cortex. Trends Neurosci 24:283–288
Psychophysiology 47:66–122 N€a€at€anen R, Kujala T, Winkler I (2011) Auditory
May P, Tiitinen H, Ilmoniemi RJ, Nyman G, Taylor JG, processing that leads to conscious perception:
N€a€at€anen R (1999) Frequency change detection in a unique window to central auditory processing
human auditory cortex. J Comput Neurosci 6:99–120 opened by the mismatch negativity and related
Melara RD, Rao A, Tong Y (2002) The duality of selec- responses. Psychophysiology 48:4–22
tion: excitatory and inhibitory processes in auditory N€a€at€anen R, Kujala T, Escera C, Baldeweg T,
selective attention. J Exp Psychol Hum Percept Per- Kreegipuu K, Carlson S, Ponton C (2012) The
form 28:279–306 Mismatch Negativity (MMN) – a unique window to
Meyer M, Elmer S, Baumann S, Jancke L (2007) Short- disturbed central auditory processing in ageing and
term plasticity in the auditory system: differential neu- different clinical conditions. Clin Neurophysiol
ral responses to perception and imagery of speech and 123:424–458
music. Restor Neurol Neurosci 25:411–431 Nagarajan S, Gabriel RA, Herman A (2012) Magnetoen-
Mondragón-Maya A, Bernal-Hernández J, Yáñez-T- cephalography. In: Poeppel D, Overath T, Popper AN,
éllez G, Rodrı́guez-Agudelo Y (2011) Mismatch neg- Fay RR (eds) The human auditory cortex. Springer
ativity (MMN) and schizophrenia: a revision. Actas handbook of auditory research, vol 43. Springer, New
Esp Psiquiatr 39:363–373 York, pp 97–128
M€ uller MM, Keil A, Kissler J, Gruber T (2001) Suppres- Ninomiya H, Onitsuka T, Chen CH, Kinukawa N (1997)
sion of the auditory middle-latency response and Possible overlapping potentials of the auditory P50 in
evoked gamma-band response in a paired-click para- humans: factor analysis of middle latency auditory
digm. Exp Brain Res 136:474–479 evoked potentials. Electroencephalogr Clin
Munka L, Berti S (2006) Examining task-dependencies of Neurophysiol 104:23–30
different attentional processes as reflected in the P3a Norman DA (1968) Toward a theory of memory and
and reorienting negativity component of the human attention. Psychol Rev 75:522–536
event-related brain potential. Neurosci Lett Nunez PL, Srinivasan R (2006) Electric fields of the brain:
396:177–181 the neurophysics of EEG. Oxford University Press,
N€a€at€anen R (1982) Processing negativity: an evoked- Oxford
potential reflection of selective attention. Psychophys- Öhman A (1979) The orienting response, attention and
iology 24:375–425 learning: an information processing perspective. In:
N€a€at€anen R (1990) The role of attention in auditory infor- Kimmel HD, van Olst EH, Orlebeke JF (eds) The
mation processing as revealed by event related poten- orienting reflex in humans. Erlbaum, Hillsdale,
tials and other brain measures of cognitive function. pp 443–471
Behav Brain Sci 13:201–288 Okita T (1979) Event-related potentials and selective
N€a€at€anen R (1992) Attention and brain function. Law- attention to auditory stimuli varying in pitch and local-
rence Erlbaum Associates, Hillsdale ization. Biol Psychol 9:271–284
Pantev C, Hoke M, Lehnertz K, L€ utkenhöner B, specific brain activity patterns. J Neurosci

Anogianakis G, Wittkowski W (1988) Tonotopic orga- 33:8633–8639
nization of the human auditory cortex revealed by Schröger E (1996) Neural mechanism for involuntary
transient auditory evoked magnetic fields. Electroen- attention shifts to changes in auditory stimulation.
cephalogr Clin Neurophysiol 69:160–170 J Cogn Neurosci 8:527–539
Pantev C, Hoke M, Lehnertz K, L€ utkenhöner B (1989a) Schröger E, Wolff C (1998a) Behavioral and electrophys-
Neuromagnetic evidence of an ampliotopic organiza- iological effects of task-irrelevant sound
tion of the human auditory cortex. Electroencephalogr change: a new distraction paradigm. Cogn Brain Res
Clin Neurophysiol 72:225–231 71:71–87
Pantev C, Hoke M, L€ utkenhöner B, Lehnertz K (1989b) Schröger E, Wolff C (1998b) Attentional orienting and
Tonotopic organization of the auditory cortex: pitch re-orienting is indicated by human event-related brain
versus frequency representation. Science 246:486–488 potentials. Neuroreport 9:3355–3358
Pantev C, Makeig S, Hoke M, Galambos R, Hampson S, Seki H, Kimura I, Ohnuma A, Saso S, Kogure K (1993)
Gallen C (1991) Human auditory evoked gamma-band The auditory evoked middle-latency responses
magnetic fields. Proc Natl Acad Sci U S (MLRs): their normative variation and generators.
A 88:8996–9000 Tohoku J Exp Med 170:157–167
Picton TW (2010) Human auditory evoked potentials. Simons RF, Graham FK, Miles MA, Chen X (2001) On
Plural Publishing, San Diego the relationship of P3a and the Novelty-P3. Biol
Polich J (2007) Updating P300: an integrative theory of Psychol 56:207–218
P3a and P3b. Clin Neurophysiol 118:2128–2148 Slabu LM, Escera C, Grimm S, Costa-Faidella J (2010)
Ponton CW, Eggermont JJ, Kwong B, Don M (2000) Early change detection in humans as revealed by audi-
Maturation of human central auditory system activity: tory brainstem and middle-latency evoked potentials.
evidence from multi-channel evoked potentials. Clin Eur J Neurosci 32:859–865
Neurophysiol 111:220–236 Snyder JS, Alain C, Picton TW (2006) Effects of attention
Ponton C, Eggermont JJ, Khosla D, Kwong B, Don on neuroelectric correlates of auditory stream segrega-
M (2002) Maturation of human central auditory sys- tion. J Cogn Neurosci 18:1–13
tem activity: separating auditory evoked potentials by Sokolov EN (1963) Higher nervous functions: the
dipole source modeling. Clin Neurophysiol orienting reflex. Ann Rev Physiol 25:545–580
113:407–420 Sonnadara RR, Alain C, Trainor LJ (2006) Occasional
Pritchard WS, Shappell SA, Brandt ME (1991) Psycho- changes in sound location enhance middle latency
physiology of N200/N400: a review and classification evoked responses. Brain Res 1076:187–192
scheme. In: Jennings JR, Ackles PK (eds) Advances in Squires NK, Squires KC, Hillyard SA (1975) Two varie-
psychophysiology: a research annual, vol 4. Jessica ties of long-latency positive waves by unpredictable
Kingsley, London, pp 43–106 auditory stimuli in man. Electroencephalogr Clin
Rees A, Green GG, Kay RH (1986) Steady-state evoked Neurophysiol 38:387–401
responses to sinusoidally amplitude-modulated sounds Stapells DR, Oates P (1997) Estimation of the pure-tone
recorded in man. Hear Res 23:123–133 audiogram by the auditory brainstem response:
Regan D (1989) Human brain electrophysiology. Elsevier, a review. Audiol Neurootol 2:257–280
New York Stapells D, Picton T, Durieux-Smith A (1993) Electro-
Rinne T, Sarkka A, Degerman A, Schröger E, Alho physiologic measures of frequency-specific auditory
K (2006) Two separate mechanisms underlie auditory function. In: Jacobson JT (ed) Principles and applica-
change detection and involuntary control of attention. tions of auditory evoked potentials. Allyn & Bacon,
Brain Res 1077:135–143 New York, pp 251–283
Robinette MS, Glatkke TJ (2007) Otoacustic emissions: Starr A, Don M (1988) Brain potentials evoked by audi-
clinical applications, 3rd edn. Thieme, New York tory stimuli. In: Picton TW (ed) Handbook of electro-
Salisbury DF (2012) Finding the missing stimulus encephalography and clinical neurophysiology.
mismatch negativity (MMN): emitted MMN to viola- Elsevier, Amsterdam
tions of an auditory gestalt. Psychophysiology Sussman E (2007) A new view on the MMN and attention
49:544–548 debate: the role of context in processing auditory
Sams M, Alho K, N€a€at€anen R (1983) Sequential effects in events. J Psychophysiol 21:164–175
the ERP in discriminating two stimuli. Biol Psychol Sussman E, Winkler I, Schröger E (2003) Top-down con-
17:41–58 trol over involuntary attention-switching in the audi-
SanMiguel I, Corral MJ, Escera C (2008) When loading tory modality. Psychon Bull Rev 10:630–637
working memory reduces distraction: behavioral and Szalárdy O, Bőhm T, Bendixen A, Winkler I (2013)
electrophysiological evidence from an auditory-visual Event-related potential correlates of sound organiza-
distraction paradigm. J Cogn Neurosci 20:1131–1145 tion: early sensory and late cognitive effects. Biol
SanMiguel I, Widmann A, Bendixen A, Trujillo-Barreto- Psychol 93:97–104
N, Schröger E (2013) Hearing silences: human audi- Szirtes G, Páczos B, Lörincz A (2005) Neural Kalman
tory processing relies on pre-activation of sound- filter. Neurocomputing 65:349–355
Teder W, Alho K, Reinikainen K, N€a€at€anen R (1993) Woldorff MG, Hillyard SA (1991) Modulation of early
Interstimulus interval and the selective-attention effect auditory processing during selective listening to rap-
on auditory ERPs: “N1 enhancement” versus idly presented tones. Electroencephalogr Clin
processing negativity. Psychophysiology 30:71–81 Neurophysiol 79:170–191 A
Tong Y, Melara RD, Rao A (2009) P2 enhancement from Woldorff MG, Hansen JC, Hillyard SA (1987) Evidence
auditory discrimination training is associated with for effects of selective attention in the mid-latency
improved reaction times. Brain Res 1297:80–88 range of the human auditory event-related potential.
Treisman AM (1964) Selective attention in man. Br Med In: Johnson R Jr, Rohrbaugh JW, Parasuraman R (eds)
Bull 20:12–16 Current trends in event-related potential research.
Treisman A (1998) Feature binding, attention and object Elsevier, Amsterdam, pp 146–154
perception. Philos Trans R Soc Lond B Biol Sci Woldorff MG, Gallen CC, Hampson SA, Hillyard SA,
353:1295–1306 Pantev C, Sobel D, Bloom FE (1993) Modulation of
Tremblay KL, Kraus N (2002) Auditory training induces early sensory processing in human auditory cortex
asymmetrical changes in cortical neural activity. during auditory selective attention. Proc Natl Acad
J Speech Lang Hear Res 45:564–572 Sci U S A 90:8722–8726
Tremblay K, Kraus N, McGee T, Ponton C, Otis B (2001) Woods DL, Clayworth CC (1987) Scalp topography dis-
Central auditory plasticity: changes in the N1-P2 com- sociate N1 and Nd components during selective atten-
plex after speech-sound training. Ear Hear 22:79–90 tion. In: Johnson R Jr, Rohrbaugh JW, Parasuraman
Ulanovsky N, Las L, Nelken I (2003) Processing of R (eds) Current trends in event-related potential
low-probability sounds by cortical neurons. Nat research. Elsevier, Amsterdam, pp 155–160
Neurosci 6:391–398 Woods DL, Knight RT, Scabini D (1993) Anatomical
Ulanovsky N, Las L, Farkas D, Nelken I (2004) Multiple substrates of auditory selective attention: behavioral
time scales of adaptation in auditory cortex neurons. and electrophysiological effects of posterior associa-
J Neurosci 24:10440–10453 tion cortex lesions. Cogn Brain Res 1:227–240
Van den Brink D, Hagoort P (2004) Influence of semantic Wu J, Yu Z, Mai X, Wei J, Luo Y (2011) Pitch and
and syntactic context constraints on lexical selection loudness information encoded in auditory imagery as
and integration in spoken-word comprehension as revealed by event-related potentials. Psychophysiol
revealed by ERPs. J Cogn Neurosci 16:1068–1084 48:415–419
Vanhatalo S, Voipio J, Kaila K (2005) Infraslow EEG Wunderlich JL, Cone-Wesson BK, Shepherd R (2006)
activity. In: Electroencephalography. Basic principles, Maturation of the cortical auditory evoked potential
clinical applications, and related fields, 5th edn. in infants and young children. Hear Res 212:185–202
Lippincott Williams & Wilkins, Philadelphia, Yabe H, Tervaniemi M, Reinikainen K, N€a€at€anen
pp 489–494 R (1997) Temporal window of integration revealed
Wacongne C, Labyt E, van Wassenhove V, by MMN to sound omission. Neuroreport
Bekinschtein T, Naccache L, Dehaene S (2011) Evi- 8:1971–1974
dence for a hierarchy of predictions and prediction Yabe H, Sato Y, Sutoh T, Hiruma T, Shinozaki N,
errors in human cortex. Proc Natl Acad Sci Nashida T, Saito F, Kaneko S (1999) The duration of
U S A 108:20754–20759 the integrating windows in auditory sensory memory.
Wacongne C, Changeux JP, Dehaene S (2012) A neuronal Electroencephalogr Clin Neurophysiol Evoked Pot
model of predictive coding accounting for the Magn Fields Suppl 49:166–169
mismatch negativity. J Neurosci 32:3665–3678 Yago E, Escera C, Alho K, Giard MH (2001) Cerebral
Winkler I (2007) Interpreting the mismatch negativity. mechanisms underlying orienting of attention towards
J Psychophysiol 21:147–163 auditory frequency changes. Neuroreport
Winkler I, Czigler I (2012) Evidence from auditory and 12:2583–2587
visual event-related potential (ERP) studies of devi- Yago E, Escera C, Alho K, Giard MH, Serra-Grabulosa
ance detection (MMN and vMMN) linking predictive JM (2003) Spatiotemporal dynamics of the auditory
coding theories and perceptual object representations. novelty-P3 event-related brain potential. Cogn Brain
Int J Psychophysiol 83:132–143 Res 16:383–390
Winkler I, Tervaniemi M, N€a€at€anen R (1997) Two sepa- Yvert B, Crouzeix A, Bertrand O, Seither-Preisler A,
rate codes for missing fundamental pitch in the audi- Pantev C (2001) Multiple supratemporal sources
tory cortex. J Acoust Soc Am 102:1072–1082 of magnetic and electric auditory evoked middle
Winkler I, Teder-S€alej€arvi WA, Horváth J, N€a€at€anen R, latency components in humans. Cereb Cortex
Sussman E (2003) Human auditory cortex tracks 11:411–423
task-irrelevant sound sources. Neuroreport Yvert B, Fischer C, Bertrand O, Pernier J (2005) Locali-
14:2053–2056 zation of human supratemporal auditory areas from
Winkler I, Denham SL, Nelken I (2009) Modeling intracerebral auditory evoked potentials using distrib-
the auditory scene: predictive regularity representa- uted source models. Neuroimage 28:140–153
tions and perceptual objects. Trends Cog Sci Zwicker E, Fastl H (1990) Psychoacoustics. Facts and
13:532–540 models. Springer, Berlin
A 234 Auditory Evoked Brainstem Responses
seven positive waves, which are generated in the

Auditory Evoked Brainstem first 10 ms after stimulus onset and are spaced at
Responses intervals of about 1 ms (Pratt 2012). They are
labeled by the Roman numerals I to VII, as intro-
Carles Escera1, Heike Althen2 and duced by Jewett and Williston (1971). The ampli-
Sabine Grimm2 tude of ABR components are in the order of
1
Institute for Brain, Cognition and Behavior tenths of a microvolt (Pratt 2012), smaller than
(IR3C), University of Barcelona, Barcelona, 0.5 mv (Møller 2006), and the largest component
Catalonia, Spain is wave V (Picton 2011). Some authors prefer to
2
Cognitive Neuroscience Research Group, plot the ABR with the positive Waves as an
University of Barcelona, Barcelona, Spain upward deflection, whereas others display the
ABR with the positive waves as a downward
deflection (Møller 2006).
Synonyms The auditory stimulus used to elicit ABR is
normally a brief duration click, presented at an
Auditory Brainstem Responses (ABR; has been intensity of 70 dB nHL (normalized Hearing
suggested as the standard abbreviation; Davis Level; normal Hearing Level. Sound intensity is
1979); Brainstem Auditory Evoked Response objectively measured in dB SPL (Sound Pressure
(BAER – used consistently in neurology; Hall Level), defined as 0.0002 dyn/cm2 or
2007); Brain Stem Auditory Evoked Potentials; 20 microPascals (20 mPa). However, a common
Auditory Brainstem Potentials; Auditory convention in clinical settings is to define inten-
Brainstem Evoked Potentials sity with a biological or behavioral reference, in
dB relative to the hearing threshold level of
a particular stimulus for a group of ten to fifteen
Definition normal-hearing young adults, which is then indi-
cated in dB nHL (normal Hearing Level; Hall
ABRs are far-field evoked potentials, generated 2007).) and at a rate close to 10/s (Picton 2011)
by the brainstem auditory pathways. They are the or 20/s (Hall 2007). When measuring ABR, the
earliest auditory evoked potentials that can be stimulus polarity must also be taken into account
extracted from the electroencephalogram, (Hall 2007). The ABR can be best recorded with
recorded from the scalp with electrodes placed an electrode at the vertex, referenced to an elec-
on the vertex and referenced to the earlobe or trode in the vicinity of the stimulated ear (mastoid
mastoid. There are two types of ABRs, the or earlobe; Pratt 2012). Generally, 2,000 trials are
transient-evoked ABRs, normally evoked by averaged for showing the ABR (Picton 2011). For
high-intensity clicks presented at a high rate and extracting the ABR from the auditory evoked
the steady-state frequency following potentials potential, a frequency filter is necessary. The
(sustained phase-locked responses), which can ABR has a high frequency and therefore a -
be evoked by tones of a single polarity and low-pass filter of between 2,000 and 3,000 Hz
phase (Picton 2011; Pratt 2012). and a high-pass filter of between 5 and 30 Hz are
typically applied (Picton 2011). The sampling
rate of the electroencephalographic signal should
Detailed Description not be lower than 20 kHz (Pratt 2012).
Waves I and II are generated by the auditory
Transient-Evoked ABRs nerve (Gelfand 2010). More precisely, Wave
Human transient-evoked ABRs were first system- I originates from the ipsilateral distal eight
atically described by Jewett, Romano and nerve and Wave II from the ipsilateral proximal
Williston in 1970 (Jewett et al. 1970; Jewett and eight nerve (Stone et al. 2009) in the vicinity of its
Williston 1971). They consist of a series of six or entry into the brainstem (Pratt 2012). Wave III
Auditory Evoked Brainstem Responses 235 A
has been attributed to the ipsilateral cochlear and/or to the periodicity in the envelope of an
nucleus and the superior olivary complex (Stone acoustic stimulus. The scalp-recorded FFR can
et al. 2009). Wave IV is not always identified in be evoked by frequencies of up to A
human subjects and is usually partially merged 1,500–2,000 Hz (Picton 2011; Gelfand 2010).
with Wave V (Pratt 2012). It has bilateral multi- According to the general consensus, it reflects
ple brainstem origins (Stone et al. 2009). The synchronized activity of neural elements in mul-
sources of Wave V are the contralateral distal tiple nuclei of the brainstem, primarily the rostral
lateral lemniscus and inferior colliculus (Stone brainstem (Starr and Don 1988; Krishnan 2006).
et al. 2009). Wave VI has been attributed to the This assumption is confirmed by the typically
medial geniculate body (Gelfand 2010). How- observed delay of 4–10 ms between the phase of
ever, there are divergent opinions as to the exact the response relative to the stimulus phase (Starr
origins of the ABR waves and it is known that and Don 1988; Skoe and Kraus 2010). The
multiple generators in the auditory brainstem recording protocol is essentially similar to the
contribute to the waves beyond Wave II transient evoked ABR recording with a few mod-
(Gelfand 2010). ifications, e.g. typically a longer rise time is used
Non-pathological factors which affect the to minimize the onset response (Krishnan 2006).
transient-evoked ARBs are amongst others the Headphone transducers should be magnetically
subject’s age, body temperature, and gender. shielded or placed some distance from the record-
Moreover, stimulus factors like frequency coming sites by using insert earphones with long
position, intensity, presentation rate, and tubes to minimize stimulus artifacts which other-
envelope have an influence (Pratt 2012). High- wise overlap with the FFR response (Skoe and
frequency stimuli and stimuli with a short rise Kraus 2010). Further, the neural FFR should be
time evoke ABRs with larger amplitudes and separated from the cochlear microphonic (CM) –
shorter latencies (Pratt 2012). Also, with increas- a receptor potential generated by the hair
ing stimulus intensity, the ABR peak latency cells – with an appreciably shorter delay
shortens (Pratt 2012). Anesthetics have little or (<2 ms). To minimize the contribution of stimu-
no effect on the ABR (Picton 2011). Regarding lus artifacts and CM, a contralateral recording
attention, the general consensus is that the ABR is technique and/or adding responses to stimuli of
not susceptible to whether the evoking stimulus is reversed polarities are effective, even though the
attended or ignored (Picton 2011). Even during later can also diminish the FFR amplitude and
sleep the ABR does not change significantly leads to a doubling of the FFR frequency
(Picton 2011). (Krishnan 2006). Steady-state FFR is commonly
Fields of clinical application of ABR record- analysed by performing a Fast Fourier Transfor-
ings are, amongst others, newborn infant auditory mation of the time domain response. For FFR in
screening, the estimation of auditory sensitivity response to time-varying complex stimuli, time-
in infants and difficult-to-test children, including frequency analysis or short-term autocorrelation
frequency specific information at audiometric algorithms can be applied to extract measures
frequencies, neurodiagnosis of eighth nerve or such as strength and accuracy of the periodicity
auditory brainstem dysfunction, monitoring tracking (Skoe and Kraus 2010; Krishnan 2006).
eighth nerve and auditory brainstem status Due to its high threshold of 30–40 dB nHL
intraoperatively during surgery potentially (Picton 2011), the clinical use of FFR has been
affecting the auditory system and the diagnosis considered marginal. In recent years there is
of auditory neuropathy (Hall 2007). renewed interest in the FFR since it carries infor-
mation about the encoding of complex sounds,
Frequency-Following ABRs such as speech and music, and can give insight
The frequency-following response (FFR) is into the processing of pitch-relevant information
a sustained evoked potential that is phase-locked and binaural cues at the level of the brainstem.
to the individual cycles of the stimulus waveform Applied to clinical and healthy populations
A 236 Auditory Evoked Field (AEF)
research on FFR can yield information on how

hearing impairment, experience, and training can Auditory Evoked Potential (AEP)
alter the neural representation of complex sounds
as measured at the brainstem level (Krishnan ▶ Auditory Event-Related Potentials
2006). New studies widely confirm the short-
and long-term plasticity of the brainstem FFR
and its relation to learning impairments, prob-
lems of hearing in noise, or reading difficulties Auditory Memory
(Skoe and Kraus 2010).
Dawai Li1 and Nelson Cowan2
1
Center for Cognitive Neuroscience, Duke
References University, Durham, NC, USA
2
Department of Psychological Sciences,
Davis H (1979) United States-Japan seminar on auditory University of Missouri-Columbia, Columbia,
responses from the brain stem. Laryngoscope
MO, USA
89:1336–1339
Gelfand SA (2010) Hearing: an introduction to psychological
and physiological acoustics. Informa Healthcare, London
Hall JW (2007) New handbook of auditory evoked Synonyms
responses. Pearson Education, Boston
Jewett DL, Williston JS (1971) Auditory-evoked far
fields averaged from the scalp of humans. Brain Acoustic memory; Echoic memory
94:681–696
Jewett DL, Romano MN, Williston JS (1970) Human
auditory evoked potentials: possible brain stem com-
ponents detected on the scalp. Science 167:1517–1518
Definition
Krishnan A (2006) Frequency-following response. In:
Burkard RF, Don M, Eggermont JJ (eds) Auditory Auditory memory is the storage of information
evoked potentials: basic principles and clinical appli- about sounds, including both acoustic features
cation. Lippincott, Williams, and Wilkins, Philadel-
(sensory memory) and categorical information
phia, pp 313–334
Møller AR (2006) Hearing: anatomy, physiology and dis- about sound categories and multi-sound
orders of the auditory system. Elsevier, Amsterdam structure.
Picton TW (2011) Human auditory evoked potentials.
Plural Publishing, San Diego
Pratt H (2012) Sensory ERP components. In: Luck SJ,
Kappenman ES (eds) Oxford handbook of event- Detailed Description
related potential components. Oxford University
Press, Oxford, pp 89–114 Auditory memory plays a critical role in various
Skoe E, Kraus N (2010) Auditory brainstem response to
aspects of human activities, such as music, verbal
complex sounds: a tutorial. Ear Hear 31:302–324
Starr A, Don M (1988) Brain potentials evoked by acous- learning, and communication. For example,
tic stimuli. In: Picton TW (ed) Human event-related when a person says, “I said ‘rice,’ not ‘lice,’”
potentials. EEG handbook, vol 3. Elsevier, Amster- the listener must keep the word “rice” in auditory
dam, pp 97–157
memory to compare it with the word “lice”
Stone JL, Calderon-Arnulphi M, Watson KS, Patel K,
Mander NS, Suss N, Fino J, Hughes JR (2009) Brainstem afterwards.
auditory evoked potentials–a review and modified stud- It is widely accepted that auditory memory can
ies in healthy subjects. J Clin Neurophysiol 26:167–175 be partitioned into three components (Cowan
1984; Crowder 1976; Massaro 1975; Neisser
1967), which Massaro terms preperceptual audi-
tory storage (known also as echoic memory),
Auditory Evoked Field (AEF) synthesized auditory memory, and generated
abstract memory. Figure 1 illustrates these com-
▶ Auditory Event-Related Potentials ponents and their relationships.
Auditory Memory 237 A
and aspects of timbre are analyzed and combined
into integrated auditory representations. The
duration of the synthesized auditory memory A
appears to vary from less than 1 s up to 30 s,
depending on how it is measured, but it is most
often found to be several seconds (Cowan 1984).
The distinction between preperceptual audi-
tory storage and synthesized auditory memory is
supported by several lines of research, including
backward masking, dichotic listening, and the
suffix effect (Cowan 1984). One of the most
convincing sources of evidence comes from
a backward masking study by Kallman and
Massaro (1979). Backward masking refers to
the phenomenon that when two sounds are
presented sequentially with a very short interval
between them, the processing of the first sound
Auditory Memory, Fig. 1 Three phases of auditory (target) sustains interference from the second one
memory according to Massaro (1975) (mask). Kallman and Massaro (1979) used two
types of sound sequence: (1) standard tone, target
Preperceptual Auditory Storage tone, and mask (referred to as mask third or
Preperceptual auditory storage retains the “M3”) and (2) standard tone, mask, and target
uncategorized representations of auditory inputs tone (referred to as mask second or “M2”). The
that have not yet been fully processed (Massaro participants needed to judge whether the target
1975) and is also referred to as short auditory tone had a higher or lower frequency than the
storage (Cowan 1984). It is the auditory counter- standard tone. In each type of sequence, the inter-
part of what is thought of as iconic memory in the val between the mask and its preceding tone
visual domain. Preperceptual auditory storage is (stimulus onset asynchrony or SOA) was varied,
the first step in auditory processing and starts and the mask was either similar to the preceding
right after an auditory stimulus enters perception. tone or quite different from it (it was then a white
The duration of preperceptual auditory storage is noise). A prediction can be made on the basis of
very short. Most researchers agree that it lasts less two forms of memory, preperceptual auditory
than 300 ms. One compelling source of evidence storage and synthesized auditory memory.
for the duration of preperceptual auditory storage These two forms can be separately interfered
comes from the finding that when a sound is very with. In both types of trials, the comparison
short (e.g., less than 100 ms), it is still perceived between the target and standard tones should be
as lasting for about a quarter of a second, which is impaired by target-mask similarity at a very short
considered to be the duration of preperceptual SOA because at short SOAs, the similar mask
auditory storage (for a review see Cowan 1984). interferes with preperceptual auditory storage of
the preceding target tone. Additionally, in the M2
Synthesized Auditory Memory trials only, it is expected that the comparison is
The auditory features stored in preperceptual always impaired by a similar mask, regardless of
auditory storage can be further analyzed to form the SOA. The reason is that the mask in this
integrated representations of sound. These inte- procedure comes between the standard and target
grated representations are considered to be stored tones and therefore can interfere with synthesized
in synthesized auditory memory (Massaro 1975). auditory memory of the standard tone. These
The term “synthesized” refers to the process in expectations exactly match what was found; the
which auditory features such as pitch, loudness, target-mask similarity mattered only at short
A 238 Auditory Nerve Response, Afferent Signals
SOAs in the M3 condition, but it mattered at all which all three phases of auditory memory work
SOAs in the M2 condition. This finding supports together to serve the auditory processing
the distinction between preperceptual auditory involved in social interactions.
storage and synthesized auditory memory.
Generated Abstract Memory References

The integrated representations in synthesized
auditory memory can be further processed to Cowan N (1984) On short and long auditory stores.
Psychol Bull 96:341–370
form abstract representations in generated
Cowan N (2001) The magical number 4 in short-term
abstract memory (Massaro 1975). The abstract memory: a reconsideration of mental storage capacity.
representations are considered to be domain Behav Brain Sci 24:87–114
general, meaning that they do not carry informa- Crowder R (1976) Principles of learning and memory.
Erlbaum, Hillsdale
tion about specific sensory details. Thus, abstract
Kallman HJ, Massaro DW (1979) Similarity effects in
representations generated from each sensory backward recognition masking. J Exp Psychol
domain (hearing, vision, touch, and so on) are 5:110–128
all stored together in the generated abstract Massaro DW (1975) Experimental psychology and infor-
mation processing. Rand McNally, Chicago
memory.
Neisser U (1967) Cognitive psychology. Appleton, New
In more recent literature, generated abstract York
memory is often referred to as “the focus of
attention” and is reported to have a core capacity
of three to five items when various memory strat-
egies are controlled (Cowan 2001). It is thought Auditory Nerve Response, Afferent
that information must be saved in generated Signals
abstract memory before high-level thinking
about it can occur. Peter Heil
Systems Physiology of Learning, Leibniz
How Auditory Memory Is Used Institute for Neurobiology, Magdeburg, Germany
Although auditory memory is usually partitioned
into three phases, all three phases can be used in
parallel to process auditory information. Suppose Definition
that you are sitting in a noisy airport reading and
a stranger asks you what time it is. Even though Sequences of action potentials (spikes) of indi-
you did not catch the words immediately, you can vidual auditory nerve fibers (ANFs), the primary
still extract the raw auditory information from the auditory afferents, in response to sounds imping-
preperceptual auditory storage, except for the ing upon the ears.
very last sounds that were masked by someone
else nearby talking immediately afterwards. The
extracted information is then integrated into syn- Detailed Description
thesized auditory memory, which can save the
auditory information long enough for you to Anatomical Foundations
turn your attention away from the reading and Acoustic information relayed from the inner ear
toward the sounds. When your attention is to the central nervous system is encoded in the
focused on the sounds, you can analyze the sequences of spikes produced by (type I) ANFs.
sounds based on their memory, using your In mammals, each ANF contacts a single receptor
existing language knowledge. You form cell (inner hair cell, IHC). ANF spikes are initi-
a generated abstract memory of what the stranger ated by neurotransmitter release from usually
meant, and you can then respond with the correct only a single active zone per afferent, the “rib-
time if you have it. This is a typical scenario in bon,” in the IHC, a unique arrangement
Auditory Nerve Response, Afferent Signals 239 A
(Ashmore 2010; Rutherford et al. 2012). Each 10 kHz; Pickles 2012) but can reach exception-
IHC is contacted by 5–30 ANFs, depending ally high values (>200) in behaviorally relevant
upon species and cochlear location, which thus frequency ranges in species such as echo- A
share some, but not all, functional response locating bats.
properties.
Dynamic Range
Spontaneous Activity With increasing sound level, the mean spike rate
ANFs produce spikes in the absence of external of a given ANF increases before saturating at
sound (spontaneous activity). The mean sponta- several hundred spikes per second at higher
neous rate varies between fibers, in mammals sound levels. The range of sound levels over
from near zero up to more than 100 spikes per which the spike rate increases (the dynamic
second (Pickles 2012). The timing of the spikes range, DR) varies between ANFs. DR is inversely
of a given fiber is highly variable. The distribu- related to the spontaneous rate of an ANF which
tion of inter-spike intervals during spontaneous in turn covaries with the fiber’s sensitivity. These
activity can be described as the result of interdependencies are predicted by a recent
a homogeneous stochastic process generating model of spike rate versus level functions,
excitatory transmitter release events in combina- according to which the spike rate follows a Hill
tion with the fiber’s refractory properties equation, with a Hill coefficient of 3, and the
(Mountain and Hubbard 1996; Heil et al. 2007). independent parameter is the sum of the sound
amplitude and a baseline (Heil et al. 2011). For
Driven Activity a given ANF, DR varies with sound frequency,
Sounds impinging on the ipsilateral ear, when of being largest at CF due to compressive growth of
appropriate spectral composition and amplitude, mechanical responses in the inner ear with sound
affect the spiking behavior of ANFs, most often level. DR and maximum spike rate also adapt to
increasing the spike rate. A threshold sound level stimulus statistics (Wen et al. 2009).
may be defined at which the driven rate of a given
ANF exceeds its spontaneous rate by some crite- Adaptation
rion. The compound action potential (CAP), Adaptation is also manifest as a decrease in spike
a gross stimulus-evoked potential which reflects rate over time in response to sounds of constant
a weighted sum of ANF responses, can be amplitude. Within a few milliseconds, the spike
recorded in or near the cochlea, e.g., at the rate drops rapidly and then more gradually. The
round window (Pickles 2012). decrease is often modeled as the sum of multiple
exponential decays with different time constants
Frequency Tuning or as a fractional power law (Zilaney et al. 2009).
Each ANF is tuned to sound frequency and is Upon cessation of the sound, the spike rate tem-
most sensitive, i.e., threshold is lowest, at porarily decreases below the spontaneous rate
a particular frequency (the characteristic fre- before recovering over tens to hundreds of
quency; CF) which is determined by the position milliseconds.
along the cochlear partition of the IHC which it
contacts (cochleotopy). Threshold versus fre- Phase-Locking
quency curves (tuning curves) are approximately In response to low-frequency sounds or broad-
V shaped, but those of high-CF ANFs exhibit band sounds containing low frequencies, ANFs
low-frequency tails. The sharpness of tuning is exhibit phase-locking, i.e., spikes are
often quantified by the Q value, defined as the CF nonrandomly distributed across the period of a -
divided by the bandwidth of the tuning curve at low-frequency component, and tend to occur at
some level (e.g., 10 dB) above threshold at a particular phase (Michelet et al. 2012; Pickles
CF. Q10 values increase with increasing CF 2012). For a given ANF, this phase varies
(in cats from about 1–10 for CFs from 0.2 to with frequency and sound level. The degree of
A 240 Auditory Pathway
phase-locking is often quantified by the measure Michelet D, Kovačić P, Joris PX (2012) Ongoing
of vector strength. Vector strength also varies temporal coding of a stochastic stimulus as a function
of intensity: time-intensity trading. J Neurosci
with frequency and sound level. Across ANFs, 32:9517–9527
maximum vector strength decreases with increas- Mountain DC, Hubbard AE (1996) Computational analysis
ing frequency in a low-pass fashion, with cutoffs of hair cell and auditory nerve processes. In: Hawkins
of a few kilohertz, depending on species. Phase- HL, McMullen TA, Popper AN, Fay RR (eds) Auditory
computation. Springer, New York, pp 121–156
locking is also seen in the responses of low-CF Pickles JO (2012) An introduction to the physiology of
ANFs to acoustic clicks, elicited by the ringing of hearing, 4th edn. Emerald Group Publishing, Bingley
the basilar membrane caused by these brief Rutherford MA, Chapochnikov NM, Moser T (2012) Spike
broadband sounds. ANFs also phase-lock to the encoding of neurotransmitter release timing by spiral gan-
glion neurons of the cochlea. J Neurosci 32:4773–4789
modulation envelope of sinusoidal amplitude- Wen B, Wang GI, Dean I, Delgutte B (2009) Dynamic
modulated tones and noise. range adaptation to sound level statistics in the audi-
Several of these properties can be altered by tory nerve. J Neurosci 29:13797–13808
sensorineural hearing loss. Cochlear implants Zilaney MSA, Bruce IC, Nelson PC, Carney LH
(2009) A phenomenological model of the synapse
function by evoking spiking activity in ANFs. between the inner hair cell and auditory nerve: long-
term adaptation with power-law dynamics. J Acoust
Soc Am 126:2390–2412
Cross-References
▶ Action Potential Initiation

▶ Anatomy and Physiology of the Mammalian Auditory Pathway
Auditory System
▶ Auditory Evoked Brainstem Responses ▶ Auditory Processing in Insects
▶ Auditory Processing in Insects
▶ Auditory Prosthesis
▶ Calcium-Dependent Exocytosis, Biophysical
Models of
▶ Cochlear Inner Hair Cell, Model
Auditory Perceptual Grouping
▶ Mechanotransduction, Models
▶ Auditory Perceptual Organization
▶ Peripheral Nerve Interface Applications,
Cochlear Implants
▶ Sound Localization in Mammals, Models
▶ Spike-Frequency Adaptation
▶ Spike Train Auditory Perceptual Organization
▶ Tinnitus, Models
Susan Denham1 and Istvan Winkler2
1
Cognition Institute and School of Psychology,
References Plymouth University, Plymouth, Devon, UK
2
Institute of Cognitive Neuroscience and
Ashmore J (2010) The afferent synapse. In: Fuchs PA, Psychology, Research Centre for Natural
Moore DR (eds) The Oxford handbook of auditory Sciences, MTA, Budapest, Hungary
science: the ear. Oxford University Press, Oxford,
pp 259–282
Heil P, Neubauer H, Irvine DRF, Brown M (2007) Spon-
taneous activity of auditory-nerve fibers: insights into Synonyms
stochastic processes at ribbon synapses. J Neurosci
27:8457–8474
Heil P, Neubauer H, Irvine DRF (2011) An improved
Auditory perceptual grouping; Auditory scene
model for the rate-level functions of auditory-nerve analysis; Cocktail party problem; Sound source
fibers. J Neurosci 31:15424–15437 separation
Auditory Perceptual Organization 241 A
Definition provide meaningful information about individual
objects in the environment (e.g., potential mates
The process of extracting acoustic features from or aggressors), it needs to partition the acoustic A
sound waves and partitioning them into meaning- features into meaningful groups, a process known
ful groups. as auditory perceptual organization or auditory
scene analysis (Bregman 1990).
Detailed Description Grouping Principles

Auditory Events
Introduction Natural environments typically contain many
Traveling pressure waves (i.e., sounds) are pro- concurrent sound sources, and even isolated
duced by the movements or actions of objects. So sounds can be rather complex, e.g., animal vocal-
sounds primarily convey information about what izations contain many different frequency com-
is happening in the environment. In addition, ponents, and both the frequencies of the
some information about the structure of the envi- components and their amplitudes can vary within
ronment and the surface features of objects can be a single sound. The problem for the auditory
extracted by determining how the original (self- system is to find some way of correctly associat-
generated or exogenous) sounds are filtered or ing the features which originate from the same
distorted by the environment (e.g., the notion of sound source. The classical view of this process is
“acoustic daylight,” Fay 2009). In this entry we that the cochlea decomposes the incoming com-
consider how the auditory systems process sound posite sound waveform into its spectral compo-
signals to extract information about the environ- nents, generating a topographically organized
ment and the objects within it. array of signals which sets up the cochleotopic
The auditory system faces a number of spe- (or tonotopic) organization found throughout
cific challenges which need to be considered in most of the auditory system, up to and including
any account of perceptual organization: the primary auditory cortex (Zwicker and Fastl
(1) sounds unfold in time; we can’t (normally) 1999). Other low-level features such as onsets,
go back to reexamine them. Therefore, informa- amplitude and frequency modulations, and bin-
tion must be extracted and perceptual decisions aural differences are extracted subcortically and
made in a timely manner. (2) The information largely independently within each frequency
contained within sounds generally requires channel (Oertel et al. 2002). These acoustic fea-
processing over many timescales in order to tures are bound together to form auditory events
extract their meaning (Nelken 2008). For exam- (Bertrand and Tallon-Baudry 2000; Zhuo and Yu
ple, a brief impulsive sound may tell the listener 2011) or tokens (Shamma et al. 2011), i.e., dis-
that two objects have been in collision, but crete sounds that are localized in time and per-
a series of such sounds is needed in order for the ceived as originating from a single sound source
listener to know that someone is clapping rather (Ciocca 2008). Events are subsequently grouped
than walking. (3) Many objects of interest gener- sequentially into patterns, streams, or perceptual
ate sounds intermittently. Therefore, some means objects.
for associating temporally discontiguous events
are required. (4) Sound pressure waves are addi- Gestalt Grouping Principles
tive; what the ear receives is a combination of all Perceptual decisions regarding the causes of the
concurrently active sound sources and their signals received by the sensors must in general be
reflections off any hard surfaces. Many animals made with incomplete information (Brunswik
and birds communicate acoustically in large 1955). Therefore, potential solutions need to be
social groups, making the problem of source sep- constrained in some way, e.g., by knowledge
aration particularly tricky (Bee 2012). Despite about likely sound sources (Bar 2007) or by
these challenges, if the auditory system is to expectations arising from the recent context
A 242 Auditory Perceptual Organization
(Winkler et al. 2012). In his seminal book, sound component can contribute to the per-
Bregman (1990) pointed out that many such conception of a complex sound as well as being
straints had already been identified by the Gestalt heard separately (Rand 1974; Fowler and
school of psychology (Köhler 1947) early in the Rosenblum 1990).
twentieth century. The core observation of (e) Closure: Objects tend to be perceived as
Gestalt psychology was that individual features whole even if they are not complete, e.g.,
form larger perceptual units, which have proper- a glide continuing through a masking noise
ties not present in the separate components (von if the glide offset is masked (Miller and
Ehrenfels 1890), and, conversely, that the percep- Licklider 1950; Riecke et al. 2008). This
tion of the components is influenced by the over- applies more generally to the perception of
all perceptual structure (Wertheimer 1912). global patterns (or “Gestalts”), e.g., individ-
Focusing primarily on visual stimuli, the Gestalt ual notes are subsumed into a melodic pattern
psychologists described the following grouping (McDermott and Oxenham 2008) and pre-
principles (laws of perception), here discussed in dictable individual speech sounds are per-
terms of auditory grouping. ceived as present even if they are masked or
(a) Good continuation: Smooth continuous missing (Warren et al. 1988). The auditory
changes in perceptual attributes favor group- system is extraordinarily sensitive to repeat-
ing, while abrupt discontinuities are per- ing patterns and appears to readily use this
ceived as the start of something new. This cue to parse complex scenes (Winkler 2007;
principle can operate both within and McDermott et al. 2011).
between individual events. An important concept that emerges from the
(b) Similarity: Similarity between the percep- idea of a “Gestalt” as a pattern is that of predict-
tual attributes of successive events (e.g., ability. In the case of auditory perception, this
pitch, timbre, location) promotes grouping refers to expectancies about sound events that
(Bregman 1990; Moore and Gockel 2002, have not yet occurred. By detecting patterns
2012). Similar to the perception of visual (or feature regularities) in the acoustic input, the
motion (Weiss et al. 2002), it appears that it brain can construct representations that allow it to
is not so much the raw difference that is anticipate or “explain away” (Pearl 1988) future
important, but rather the rate of change; the events. In this way Gestalt theory connects to the
slower the rate of change between successive ideas of unconscious inference (Helmholtz 1885)
sounds, the more similar they are judged and perception as hypothesis formation (Gregory
(Winkler et al. 2012). In other words, in the 1980).
auditory modality, similarity and good con-
tinuation may be equivalent. Auditory Objects
(c) Common fate: Correlated changes in fea- While visual objects are widely accepted as fun-
tures promote grouping, recently formalized damental representational units, the notion of an
as temporal coherence, i.e., feature correla- auditory object is less well established, and there
tions within time windows that span periods is as yet no universal agreement on how they
longer than individual events (Elhilali should be defined, e.g., see Kubovy and Van
et al. 2009; Shamma et al. 2011). Valkenburg (2001), Griffiths and Warren
(d) Disjoint allocation (or belongingness): (2004), Winkler et al. (2006), Shinn-Cunningham
Refers to the principle that each element of (2008). Based on the Gestalt principles and ideas
the sensory input is only assigned to one of perceptual inference, outlined above, Winkler
perceptual object, e.g., exclusive border et al. (2009) proposed a definition of an auditory
assignment in Rubin’s face-vase illusion. perceptual object as a predictive representation,
However, although generally true, this prin- constructed from feature regularities extracted
ciple is sometimes violated in auditory per- from the incoming sounds. These object repre-
ception, e.g., in duplex perception, the same sentations are temporally persistent and encode
distributions over featural and temporal patterns, outlined above, these two are not really distinct
determined by the current context. The consoli- (simultaneous cues are influenced by prior
dated object representation therefore refers to sequential grouping, e.g., Darwin et al. (1995) A
patterns of sound events; individual sound events and Bendixen et al. (2010b), just as sequential
are processed within the context of the whole to grouping is influenced by the perceptual qualities
which they belong. This definition of an auditory of individual events (simultaneous grouping)
perceptual object is compatible with the defini- (Bregman 1990); nevertheless, they provide
tion of an auditory stream, as a coherent sequence a useful starting point for models of auditory
of sounds separable from other concurrent or scene analysis.
intermittent sounds (Bregman 1990). However,
whereas the term “auditory stream” refers to Simultaneous Grouping
a phenomenological unit of sound organization, In the absence of sequential grouping cues, there
with separability as its primary property, the def- are some features which automatically trigger the
inition proposed by Winkler et al. (2009) empha- formation of individual sound events; for reviews
sizes the extraction and representation of the unit see Darwin and Carlyon (1995) and Ciocca
as a pattern with predictable components (2008). Common onsets and offsets form clear
(Winkler et al. 2012). While the usage of the temporal boundaries, and the strategy adopted by
term object is not universally accepted within the auditory system is to match onsets to offsets
the auditory domain, we will use it in this entry (including similarities between features and tem-
as defined by Winkler et al. (2009). poral proximity) in order to segregate perceptual
events (Nakajima et al. 2000). Harmonicity (i.e.,
Auditory Scene Analysis the presence of frequency components which are
In order to determine the perceptual qualities of integer multiples of a common fundamental fre-
individual sound events, the brain must first bind quency) is another important grouping cue
their component features even though the number (Darwin and Carlyon 1995). For example, when
of concurrent auditory objects and which features one component of a complex harmonic tone is
belong to each is unknown a priori; this must be mistuned, listeners perceive two concurrent
inferred incrementally from the ongoing sensory sounds, a complex tone consisting of the harmon-
input. Therefore, it is clear that the auditory sys- ically related components and a pure tone,
tem needs to use (top-down) contextual informa- corresponding to the mistuned component
tion to guide its grouping decisions and some (Moore et al. 1986). However, not all acoustic
means for evaluating these decisions and revising features trigger concurrent grouping, e.g.,
them in the event that they prove to be incorrect. a location cue (common interaural time differ-
In the currently most widely accepted framework ences) between a subset of frequency compo-
describing perceptual sound organization, audi- nents within a single sound event does not
tory scene analysis, Bregman (1990) proposes generate a similar segregation of component sub-
two separable processing stages. The first stage sets within individual sound events (Culling and
is suggested to be concerned with partitioning Summerfield 1995).
sound events into potential groups based primar- Another important strategy for segregating
ily on featural similarities and differences. The sound events is template matching. If people
second stage, within which prior knowledge and have prior knowledge of events, then it is possible
task demands exert their influence, is to hear them out. This effect was exploited in the
a competitive process between candidate organi- many double-vowel experiments used to test the
zations that determines which one is perceived. influence of different acoustic features, e.g.,
Within this framework there are two types of Assmann and Summerfield (1990) and
grouping: simultaneous grouping based on con- Summerfield and Assmann (1991), and even in
current cues and sequential grouping based on the absence of featural differences, it was shown
contextual temporal cues. For the reasons that known vowel sounds can be identified well
above chance (Assmann and Summerfield 1989). based on some representation of the preceding
This template-matching phenomenon appears to sounds; for reviews, see (Moore and Gockel
be rather general and applies to any sound that is 2002; Carlyon 2004; Haykin and Chen 2005;
repeated. The auditory system is very sensitive to Snyder and Alain 2007; Ciocca 2008; Shamma
repetition (Teki et al. 2011). If a previously and Micheyl 2010; Shamma et al. 2011; Moore
unheard sound is repeated against a different and Gockel 2012). Most studies of this class of
background, then it can be segregated and iden- grouping have used sequences of discrete sound
tified significantly above chance, even with only events to investigate the influences of acoustic
a single repetition, and even if many of usual features and temporal structure. In the most
grouping cues are absent (McDermott widely used experimental approach (termed the
et al. 2011). Similarly, arbitrary repeated noise auditory streaming paradigm), sequences of
segments can be rapidly learnt within a few trials alternating sound events differing in some fea-
(Agus et al. 2010). ture(s) are presented to listeners (van Noorden
1975). When the feature separation is small
Models of Event Formation and/or they are delivered at a slow pace, listeners
Many models have been developed to investigate predominantly hear a single integrated stream
simultaneous grouping and the segregation of containing all the sounds. With large feature sep-
perceptual events, e.g., see models described in aration and/or fast presentation rates, listeners
Wang and Brown (2006). A model of auditory report hearing the sequence separate out into
saliency which used low-level cues of spectral two segregated streams. In this there is a cue
and temporal contrast to highlight salient events trade-off: smaller feature differences can be com-
in continuous noisy soundscapes predicted pensated with higher presentation rates and vice
human event detection very well (Kayser versa (van Noorden 1975). Differences in various
et al. 2005). Temporal contrasts effectively high- auditory features, including frequency, pitch,
light onsets and offsets, while spectral peaks loudness, location, timbre, and amplitude modu-
carry information about the resonances of sound lation, have been shown to support auditory
sources and to some extent their identity (von stream segregation (Vliegen and Oxenham
Kriegstein et al. 2007). The segregation of 1999; Grimault et al. 2002; Roberts et al. 2002).
overlapping events using pitch cues has been Thus it appears that sequential grouping is based
widely explored (c.f. Pitch Perception, Models), on perceptual similarity, rather than on specific
e.g., for explaining enhanced double-vowel seg- low-level auditory features (Moore and Gockel
regation (de Cheveigne et al. 1995). The segre- 2002, 2012). Temporal structure has also been
gation of events using repetition was shown to be suggested as a key factor in segregating streams
possible in principle by using a combination of either by guiding attentive grouping processes
cross-correlation and averaging to incrementally (Jones 1976; Jones et al. 1981; Large and Jones
build a representation of the repeated target 1999) or through temporal coherence that binds
(McDermott et al. 2011). Because of the impor- correlated component features in the auditory
tance of longer-term context on grouping, none of input (Elhilali et al. 2009; Shamma and Micheyl
these models provide general solutions to the 2010; Shamma et al. 2011, 2013).
problem of auditory scene analysis; nevertheless,
they provide important building blocks in this Models of Auditory Streaming
process. Early models of auditory streaming, e.g.,
Beauvois and Meddis (1991), focused on the
Sequential Grouping relationship between frequency differences and
Sequential grouping generally conforms to the event rate and the proposal that streaming could
Gestalt principles of similarity/good continuation be explained almost exclusively by peripheral
and common fate. In contrast to concurrent channeling mechanisms (Hartmann and Johnson
grouping, sequential grouping is necessarily 1991) or the degree of overlap between neural
responses to each of the alternating tones, e.g., grouping process) to refine its representation.
McCabe and Denham (1997). In these models the These correlation-based models overcome
perceptual decision was represented by levels of a problem faced by the population separation A
activation across a spatial array of neurons; see account of streaming (Micheyl et al. 2005) that
also Micheyl et al. (2005) for a similar interpre- predicted widely separated components would be
tation of neural activity in primary auditory cor- segregated even if they overlapped in time, which
tex. A different approach in which grouping is is not the case (Elhilali et al. 2009). They also
signaled by temporal correlations within network provide a means for binding the component fea-
responses was proposed by Wang, Brown, and tures of an event, not considered in the earlier
colleagues (Brown and Wang 2006; Wang and models. Later refinements to the temporal coher-
Chang 2008). For example, the model proposed ence account of streaming (Shamma et al. 2011,
by Wang and Chang (2008) consists of a 2013), included the strong claims that (a) feature
2-dimensional array of oscillators with one binding occurs only with attention, i.e., attention
dimension representing frequency and the other is responsible for grouping features that belong to
external time. Units are connected by local excit- the foreground object, c.f. (Treisman 1998), and
atory connections and by global inhibition. Char- (b) all other features remain ungrouped in an
acteristic results of classical auditory streaming undifferentiated background. However, the pro-
experiments (van Noorden 1975) are simulated posed role of attention in feature binding has long
by including strong local excitatory connections been debated in the visual domain, e.g., Duncan
(encouraging synchronization) and weaker long- and Humphreys (1989), and it is not consistent
range connections (which are easily overcome by with the results of experiments testing feature
inhibition and therefore encourage desynchro- binding in the absence of attention by recording
nization). Sensitivity to event rate is modeled by auditory event-related potentials (AERP) in
dynamic weight adjustments. However, while the response to rare feature combinations (Takegata
representation of grouping is different from the et al. 2005; Winkler et al. 2005a).
models previously outlined, this model also
depends on peripheral channeling and the degree Competition and Selection
of overlap in the incoming activity patterns to The models described above all conform to the
determine its grouping decision. assumptions that in response to alternating
A similar focus on temporal coherence (in this two-tone sequences, (a) auditory perception
case the average correlation within a sliding win- always starts from the integrated organization
dow 50–500 ms in duration) is seen in the model and (b) that eventually a stable final perceptual
of streaming proposed by Elhilali and colleagues, decision is reached (Bregman 1990). However, it
e.g., Elhilali and Shamma (2008) and Shamma has been found, when listeners report their per-
et al. (2011) (Note, Figs. 6 and 9 in this entry have cepts continuously while listening to such
incorrect colour scale labels (0 % and 100 %, sequences for long periods, that perception fluc-
interchanged; Shamma and Elhilali (2013)). The tuates between different perceptual organizations
computational model developed by Elhilali and (Winkler et al. 2005b; Pressnitzer and Hupe
Shamma (2008) extracts multiple features from 2006). Perceptual switching occurs in all listeners
the incoming acoustic input including frequency, and for all combinations of stimulus parameters
pitch, direction, and spectral shape and assigns tested (Anstis and Saida 1985; Roberts
the resulting activity patterns to one of two clus- et al. 2002; Denham and Winkler 2006;
ters which come to represent the properties of the Pressnitzer and Hupe 2006; Schadwinkel and
events in each stream. The temporal coherence Gutschalk 2011; Denham et al. 2012), even com-
measure is used to determine which components binations very far from the ambiguous region
should be grouped. The clusters compete to identified by van Noorden (1975). Furthermore,
incorporate each event, and the winning cluster for stimuli with parameters that strongly promote
uses the event features (as determined by the segregation, participants often report hearing
segregation first (Deike et al. 2012; Denham to account for the observed stochasticity of per-
et al. 2012). It has also been found that perceptual ceptual switching (successive phase durations are
organizations other than the classic integrated largely uncorrelated, and the distribution of phase
and segregated categories may be reported durations resembles a gamma or log-normal dis-
(Bendixen et al. 2010a, 2012; Bőhm et al. 2012; tribution) (Levelt 1968). The questions for audi-
Denham et al. 2012; Szalárdy et al. 2012), show- tory multistability are what are the competing
ing that auditory perceptual organization in entities, and what form does this competition
response to alternating two-tone sequences is take in order to explain dynamic nature of per-
multistable (Schwartz et al. 2012). ceptual awareness reported by listeners.
The notion of perceptual multistability is chal- The computational model of auditory
lenged by everyday subjective experience of multistability proposed by Mill et al. (2013) is
a world perceived as stable and continuous and based on the idea that auditory perceptual orga-
by experimental results obtained by averaging nization rests on the discovery of recurring pat-
over the reports of different listeners, which gen- terns embedded within the stimulus, constructed
erally show that within the initial 5–15 s of by forming associations (links) between incom-
two-tone sequence, the probability of reporting ing sound events and recognizing when
segregation monotonically increases (termed the a previously discovered sequence recurs and can
buildup of auditory streaming) (but see Deike thus be used to predict future events. These pre-
et al. (2012)). For these reasons it has been dictive representations, or proto-objects (Rensink
suggested that perceptual multistability observed 2000; Winkler et al. 2012), compete for domi-
in the auditory streaming paradigm may be sim- nance with any other proto-objects which predict
ply a consequence of the artificial stimulation the same event (a form of local competition) and
protocol used. However, there is a growing are the candidate set of representations that have
body of experimental data supporting the exis- the potential to become the perceptual objects of
tence of multistability and just as visual conscious awareness. This model accounts for the
multistability has provided new insights into emergence of, and switching between, alternative
visual processing, e.g., Kovacs et al. (1996); it organizations; the influence of stimulus parame-
seems likely that understanding spontaneous ters on perceptual dominance, switching rate, and
changes in the perception of unchanging sound perceptual phase durations; and the buildup of
sequences will help throw new light on auditory auditory streaming. In a new sound scene, the
perception. proto-object that is the easiest to discover deter-
mines the initial percept. Since the time needed
Modeling Multistability in Auditory Streaming for discovering a proto-object depends largely on
Multistability of auditory perceptual organization the stimulus parameters (i.e., to what extent suc-
cannot be explained by any of the theories cessive sound events satisfy/violate the similar-
or models outlined above, which all have essen- ity/good continuation principle), the first percept
tially one fixed attractor. Models of visual strongly depends on stimulus parameters. How-
multistability have a longer history, e.g., Laing ever, the duration of the first perceptual phase is
and Chow (2002); Shpiro et al. (2009); van Ee independent of the percept (Hupe and Pressnitzer
(2009). These models typically contain three 2012), since it depends on how long it takes for
essential components (Leopold and Logothetis other proto-objects to be discovered (Winkler
1999): (a) mutual inhibition between competing et al. 2012). The model also accounts for the
stimuli to ensure exclusivity (i.e., perceptual different influences of similarity and closure on
awareness generally switches between the differ- perception; the rate of perceptual change
ent alternatives rather than fusing them), (similarity/good continuation) determines how
(b) adaptation to ensure the observed inevitability easy it is to form the links between the events
of perceptual switching (the dominant percept that make up a proto-object, while predictability
cannot remain dominant forever), and (c) noise (closure) does not affect the discovery of
proto-objects, but can increase the competitive- auditory event. Because they can be recorded
ness (salience) of a proto-object once it has been noninvasively from the human scalp, they have
discovered (Bendixen et al. 2010a). been widely used to study the brain responses A
accompanying auditory stream segregation;
Neural Correlates of Perceptual Organization c.f. auditory event-related potentials, especially
Neural responses to individual sounds are prolong-latency AERP responses. Three AERP com-
foundly influenced by the context in which they ponents are of particularly relevance in this
appear (Bar-Yosef et al. 2002). The question is to regard: (a) the “object-related negativity”
what extent the contextual influences on neural (ORN) which signals the automatic segregation
responses reflect the current state of perceptual of concurrent auditory objects (Alain et al. 2002),
organization. This question has been addressed (b) the amplitude of the auditory P1 and N1
by a number of studies ranging in focus from the which varies depending on whether the same
single neuron level (c.f. stimulus-specific adapta- sounds are perceived as part of an integrated or
tion) to large-scale brain responses (c.f. auditory segregated organization (Gutschalk et al. 2005;
evoked potentials), and the results provide impor- Szalárdy et al. 2013), and c) the mismatch nega-
tant clues about the processing strategies adopted tivity (MMN; N€a€at€anen et al. 1978) which has
by the auditory system. been used as an indirect index of auditory stream
Studies investigating single neuron responses segregation, e.g., Sussman et al. (1999); Nager
to alternating tone sequences, e.g., Fishman et al. (2003); Winkler et al. (2003a); Gutschalk
et al. (2004), Bee and Klump (2005), Micheyl et al. (2005).
et al. (2005)), and Micheyl et al. (2007), have The detection and representation of regulari-
shown an effect called differential suppression, ties by the brain, as indexed by the MMN, pro-
i.e., at the start of the sequence, the neuron vided the basis for the definition of an auditory
responds to both tones, but with time the response object proposed by Winkler et al. (2009). Using
to one of the tones (typically corresponding to the evidence from a series of MMN studies, they
best frequency of the cell) remains relatively defined an auditory object as a perceptual repre-
strong, while the response to the other tone sentation of a possible sound source, derived
diminishes. Since neuronal sensitivity to fre- from regularities in the sensory input (Winkler
quency difference and presentation rate was 2007, 2010) that has temporal persistence
found to be consistent with the classical van (Winkler and Cowan 2005) and can link events
Noorden (1975) parameter space, it was claimed separated in time (N€a€at€anen and Winkler 1999).
that differential suppression was a neural corre- This representation forms a separable unit
late of perceptual segregation (Fishman (Winkler et al. 2006a) that generalizes across
et al. 2004). This was supported by the finding natural variations in the sounds (Winkler
that spike counts from neurons in primary audi- et al. 2003b) and generates expectations of parts
tory cortex predict an initial integration/segrega- of the object not yet available (Bendixen
tion decision closely matching human perception et al. 2009).
(Micheyl et al. 2005; Bee et al. 2010). However, It should be pointed out that while traditional
differential suppression does not account for psychological accounts of auditory perceptual
perceptual multistability or for the perception of organization implicitly or explicitly refer to rep-
overlapping tone sequences (Elhilali et al. 2009); resentations of objects, there are models of audi-
therefore, while differential suppression may be tory perception which are not concerned with
a necessary component of the auditory streaming positing a representation directly corresponding
process, it does not provide a complete auditory objects. The hierarchical predictive cod-
explanation. ing model of perception, e.g., Friston and Kiebel
Auditory event-related brain potentials (2009), includes predictive memory representa-
(AERPs) represent the synchronized activity of tions, which are in many ways compatible with
large neuronal populations, time locked to some the notion of auditory object representations
(Winkler and Czigler 2012), but no explicit con- perceptual switching and thalamus (medial
nection with object representations is made. geniculate) decreasing (Kashino and Kondo
Shamma and colleagues’ temporal coherence 2012). Consistent with these findings,
model of auditory stream segregation (Elhilali Schadwinkel and Gutschalk (2011), using
and Shamma 2008; Elhilali et al. 2009; Shamma a different stimulus paradigm which allowed
et al. 2011, 2013) provides another way to avoid them to influence the timing of perceptual
the assumption that object representations are switching, found transient auditory cortical acti-
necessary for determining sound organization; vation associated with perceptual switching and
instead it is proposed that objects are essentially a further transient activation in inferior
whatever occupies the perceptual foreground and colliculus, although whether the inferior
exist only insofar as they do occupy the fore- colliculus is responsible for triggering switching
ground. In summary, there is currently little con- or simply reflects the transient switching activa-
sensus on the role of auditory object tion in cortex is not clear. In summary, neural
representations in perceptual organization, and correlates of auditory streaming have been found
the importance placed on object representations in many areas within the auditory system and
by the various models and theories differs beyond, suggesting that creating and switching
markedly. between alternative perceptual organizations
fMRI studies of auditory streaming have involve a broadly distributed network within the
found neural correlates in a number of brain brain.
regions. In one of the earliest studies, Cusack
(2005) failed to find differential activity in audi- Conclusions and Open Questions
tory cortex corresponding to perceptual organi- The Gestalt principles and their application to
zation into one or two streams, but he did find auditory perception instantiated in Bregman’s
such activity in the intraparietal sulcus, an area (1990) two-stage auditory scene analysis frame-
associated with cross-modal processing and work provided the initial basis for understanding
object numerosity. Shortly afterwards Wilson auditory perceptual organization, and recent pro-
et al. (2007) showed that auditory cortical activity posals have extended this framework in interest-
increased with increasing frequency difference ing ways. Nevertheless, there remain many
and that as the frequency difference increased, unanswered questions and there have been few,
the cortical response changed from being rather if any, attempts to build neuro-computational
phasic (i.e., far stronger at the onset of the models capable of dealing with the complexity
sequence) towards a more sustained response of real auditory scenes in which grouping and
throughout the stimulus sequence. Taking categorization cues are not immediately avail-
a closer look at the dynamics of cortical activity able; however, see (Yildiz and Kiebel 2011).
associated with perceptual switching, Kondo and Feedback connections are pervasive within the
Kashino (2009) showed that both auditory cortex auditory system, including all stages of the sub-
and thalamus are involved, with an increase in cortical system, yet to our knowledge no models
thalamic activity preceding that in cortex associ- include such connections. Although fMRI results
ated with a switch from the nondominant to the are useful for identifying regional involvement,
dominant percept and, conversely, an increase in detailed understanding of the neural circuitry
cortical activity preceding that in thalamus asso- involved in auditory perceptual organization is
ciated with a switch from the dominant to the sketchy, and the neural representations of audi-
nondominant percept. They also found differen- tory objects and perceptual organization are
tial activation in posterior insular cortex and in unknown. Even the role of primary auditory cor-
the cerebellum. Interestingly, activations in the tex remains something of a mystery, e.g., see
cerebellum and thalamus are negatively corre- Nelken et al. (2003) and Griffiths et al. (2004);
lated in auditory streaming, with the left cerebel- perhaps studying the switching of perceptual
lar activation level increasing with the rate of awareness between different representations in
awake behaving animals will help to elucidate the Bendixen A, Denham SL, Gyimesi K, Winkler I (2010a)
representations and processing strategies adopted Regular patterns stabilize auditory streams. J Acoust
Soc Am 128(6):3658–3666
by cortex. Bendixen A, Jones SJ, Klump G, Winkler I (2010b) Prob- A
ability dependence and functional separation of the
object-related and mismatch negativity event-
Cross-References related potential components. Neuroimage
50(1):285–290
Bendixen A, Bőhm TM, Szalárdy O, Mill R, Denham SL,
▶ Auditory Event-Related Potentials Winkler I (2012) Different roles of similarity and
▶ Stimulus-Specific Adaptation, Models predictability in auditory stream segregation. J Learn
Percept (in press)
Bertrand O, Tallon-Baudry C (2000) Oscillatory gamma
activity in humans: a possible role for object represen-
References tation. Int J Psychophysiol 38(3):211–223
Bőhm TM, Shestopalova L, Bendixen A, Andreou AG,
Agus TR, Thorpe SJ, Pressnitzer D (2010) Rapid forma- Georgiou J, Garreau G, Pouliquen P, Cassidy A,
tion of robust auditory memories: insights from noise. Denham SL, Winkler I (2012) Spatial location of
Neuron 66(4):610–618 sound sources biases auditory stream segregation but
Alain C, Schuler BM, McDonald KL (2002) Neural activ- their motion does not. J Learn Percept (in press)
ity associated with distinguishing concurrent auditory Bregman AS (1990) Auditory scene analysis: the percep-
objects. J Acoust Soc Am 111(2):990–995 tual organization of sound. MIT, Cambridge, MA
Anstis S, Saida S (1985) Adaptation to auditory streaming Brown GJ, Wang DL (eds) (2006) Neural and perceptual
of frequency-modulated tones. J Exp Psychol Hum modelling. Computational auditory scene analysis:
Percept Perform 11:257–271 principles, algorithms, and applications. Wiley/IEEE
Assmann PF, Summerfield Q (1989) Modeling the per- Press, Chichester
ception of concurrent vowels: vowels with the same Brunswik E (1955) Representative design and probabilis-
fundamental frequency. J Acoust Soc Am tic theory in a functional psychology. Psychol Rev
85(1):327–338 62(3):193–217
Assmann PF, Summerfield Q (1990) Modeling the per- Carlyon RP (2004) How the brain separates sounds.
ception of concurrent vowels: vowels with different Trends Cogn Sci 8(10):465–471
fundamental frequencies. J Acoust Soc Am Ciocca V (2008) The auditory organization of complex
88(2):680–697 sounds. Front Biosci 13:148–169
Bar M (2007) The proactive brain: using analogies and Culling JF, Summerfield Q (1995) Perceptual separation
associations to generate predictions. Trends Cogn Sci of concurrent speech sounds: absence of across-
11(7):280–289 frequency grouping by common interaural delay.
Bar-Yosef O, Rotman Y, Nelken I (2002) Responses of J Acoust Soc Am 98(2 Pt 1):785–797
neurons in cat primary auditory cortex to bird chirps: Cusack R (2005) The intraparietal sulcus and perceptual
effects of temporal and spectral context. J Neurosci organization. J Cogn Neurosci 17(4):641–651
22(19):8619–8632 Darwin CJ, Carlyon RP (1995) Auditory grouping. In:
Beauvois MW, Meddis R (1991) A computer model of Moore BCJ (ed) The handbook of perception and
auditory stream segregation. Q J Exp Psychol cognition: hearing, vol 6. Academic, London,
A 43(3):517–541 pp 387–424
Bee MA (2012) Sound source perception in anuran Darwin CJ, Hukin RW, al-Khatib BY (1995) Grouping in
amphibians. Curr Opin Neurobiol 22(2):301–310 pitch perception: evidence for sequential constraints.
Bee MA, Klump GM (2005) Auditory stream segregation J Acoust Soc Am 98(2 Pt 1):880–885
in the songbird forebrain: effects of time intervals on de Cheveigne A, McAdams S, Laroche J, Rosenberg
responses to interleaved tone sequences. Brain Behav M (1995) Identification of concurrent harmonic and
Evol 66(3):197–214 inharmonic vowels: a test of the theory of harmonic
Bee MA, Micheyl C, Oxenham AJ, Klump GM cancellation and enhancement. J Acoust Soc Am
(2010) Neural adaptation to tone sequences in the 97(6):3736–3748
songbird forebrain: patterns, determinants, and rela- Deike S, Heil P, Böckmann-Barthel M, Brechmann
tion to the build-up of auditory streaming. J Comp A (2012) The build-up of auditory stream segregation:
Physiol A Neuroethol Sens Neural Behav Physiol a different perspective. Front Psychol 3:461
196(8):543–557 Denham SL, Winkler I (2006) The role of predictive
Bendixen A, Schröger E, Winkler I (2009) I heard that models in the formation of auditory streams.
coming: event-related potential evidence for stimulus- J Physiol Paris 100(1–3):154–170
driven prediction in the auditory system. J Neurosci Denham SL, Gymesi K, Stefanics G, Winkler I (2012)
29(26):8447–8451 Multistability in auditory stream segregation: the role
of stimulus features in perceptual organisation. J Learn Köhler W (1947) Gestalt psychology: an introduction to
Percept (in press) new concepts in modern psychology. Liveright Pub-
Duncan J, Humphreys G (1989) Visual search and stimu- lishing Corporation, New York
lus similarity. Psychol Rev 96:433–458 Kondo HM, Kashino M (2009) Involvement of the
Elhilali M, Shamma SA (2008) A cocktail party with thalamocortical loop in the spontaneous switching of
a cortical twist: how cortical mechanisms contribute percepts in auditory streaming. J Neurosci
to sound segregation. J Acoust Soc Am 29(40):12695–12701
124(6):3751–3771 Kovacs I, Papathomas TV, Yang M, Feher A (1996) When
Elhilali M, Ma L, Micheyl C, Oxenham AJ, Shamma SA the brain changes its mind: interocular grouping during
(2009) Temporal coherence in the perceptual organi- binocular rivalry. Proc Natl Acad Sci
zation and cortical representation of auditory scenes. USA 93(26):15508–15511
Neuron 61(2):317–329 Kubovy M, Van Valkenburg D (2001) Auditory and visual
Fay R (2009) Soundscapes and the sense of hearing of objects. Cognition 80(1–2):97–126
fishes. Integr Zool 4(1):26–32 Laing CR, Chow CC (2002) A spiking neuron model for
Fishman YI, Arezzo JC, Steinschneider M (2004) Audi- binocular rivalry. J Comput Neurosci 12(1):39–53
tory stream segregation in monkey auditory cortex: Large EW, Jones MR (1999) The dynamics of attending:
effects of frequency separation, presentation rate, and how people track time-varying events. Psychol Rev
tone duration. J Acoust Soc Am 116(3):1656–1670 106:119–159
Fowler CA, Rosenblum LD (1990) Duplex perception: Leopold DA, Logothetis NK (1999) Multistable phenom-
a comparison of monosyllables and slamming doors. ena: changing views in perception. Trends Cogn Sci
J Exp Psychol Hum Percept Perform 16(4):742–754 3(7):254–264
Friston K, Kiebel S (2009) Predictive coding under the Levelt WJM (1968) On binocular rivalry. Mouton, Paris
free-energy principle. Philos Trans R Soc Lond B Biol McCabe SL, Denham MJ (1997) A model of auditory
Sci 364(1521):1211–1221 streaming. J Acoust Soc Am 101(3):1611–1621
Griffiths TD, Warren JD (2004) What is an auditory McDermott JH, Oxenham AJ (2008) Music perception,
object? Nat Rev Neurosci 5(11):887–892 pitch, and the auditory system. Curr Opin Neurobiol
Griffiths TD, Warren JD, Scott SK, Nelken I, King AJ 18(4):452–463
(2004) Cortical processing of complex sound: a way McDermott JH, Wrobleski D, Oxenham AJ (2011) Recov-
forward? Trends Neurosci 27(4):181–185 ering sound sources from embedded repetition. Proc
Grimault N, Bacon SP, Micheyl C (2002) Auditory stream Natl Acad Sci USA 108(3):1188–1193
segregation on the basis of amplitude-modulation rate. Micheyl C, Tian B, Carlyon RP, Rauschecker JP
J Acoust Soc Am 111(3):1340–1348 (2005) Perceptual organization of tone sequences in
Gutschalk A, Micheyl C, Melcher JR, Rupp A, Scherg M, the auditory cortex of awake macaques. Neuron
Oxenham AJ (2005) Neuromagnetic correlates of 48(1):139–148
streaming in human auditory cortex. J Neurosci Micheyl C, Carlyon RP, Gutschalk A, Melcher JR,
25(22):5382–5388 Oxenham AJ, Rauschecker JP, Tian B, Courtenay
Hartmann WM, Johnson D (1991) Stream segregation and Wilson E (2007) The role of auditory cortex in the
peripheral channeling. Music Percept 9(2):153–183 formation of auditory streams. Hear Res
Haykin S, Chen Z (2005) The cocktail party problem. 229(1–2):116–131
Neural Comput 17(9):1875–1902 Mill R, Bőhm T, Bendixen A, Winkler I, Denham SL
Helmholtz H (1885) On the sensations of tone as (2013) Competition and cooperation between frag-
a physiological basis for the theory of mentary event predictors in a model of auditory
music. Longmans, Green, London scene analysis. PLoS Comput Biol (in press)
Hupe JM, Pressnitzer D (2012) The initial phase of audi- Miller GA, Licklider JCR (1950) The intelligibility of
tory and visual scene analysis. Philos Trans R Soc interrupted speech. J Acoust Soc Am 22:167–173
Lond B Biol Sci 367(1591):942–953 Moore BCJ, Gockel HE (2002) Factors influencing
Jones MR (1976) Time, our lost dimension: toward a new sequential stream segregation. Acta Acust 88:320–333
theory of perception, attention, and memory. Psychol Moore BC, Gockel HE (2012) Properties of auditory
Rev 83:323–355 stream formation. Philos Trans R Soc Lond B Biol
Jones MR, Kidd G, Wetzel R (1981) Evidence for rhyth- Sci 367(1591):919–931
mic attention. J Exp Psychol Hum Percept Perform Moore BC, Glasberg BR, Peters RW (1986) Thresholds
7:1059–1073 for hearing mistuned partials as separate tones in har-
Kashino M, Kondo HM (2012) Functional brain networks monic complexes. J Acoust Soc Am 80(2):479–483
underlying perceptual switching: auditory streaming N€a€at€anen R, Winkler I (1999) The concept of auditory
and verbal transformations. Philos Trans R Soc Lond stimulus representation in cognitive neuroscience.
B Biol Sci 367(1591):977–987 Psychol Bull 125(6):826–859
Kayser C, Petkov CI, Lippert M, Logothetis NK N€a€at€anen R, Gaillard AWK, M€antysalo S (1978) Early
(2005) Mechanisms for allocating auditory attention: selective attention effect on evoked potential
an auditory saliency map. Curr Biol 15(21):1943–1947 reinterpreted. Acta Psychol 42:313–329
Nager W, Teder-S€alej€arvi W, Kunze S, M€ unte TF models of perceptual bistability. J Comput Neurosci
(2003) Preattentive evaluation of multiple perceptual 27(1):37–54
streams in human audition. Neuroreport Snyder JS, Alain C (2007) Toward a neurophysiological
14(6):871–874 theory of auditory stream segregation. Psychol Bull A
Nakajima Y, Sasaki T, Kanafuka K, Miyamoto A, 133(5):780–799
Remijn G, ten Hoopen G (2000) Illusory recouplings Summerfield Q, Assmann PF (1991) Perception of con-
of onsets and terminations of glide tone components. current vowels: effects of harmonic misalignment and
Percept Psychophys 62(7):1413–1425 pitch-period asynchrony. J Acoust Soc Am
Nelken I (2008) Processing of complex sounds in the 89(3):1364–1377
auditory system. Curr Opin Neurobiol 18(4):413–417 Sussman ES, Ritter W, Vaughan HG Jr (1999) An inves-
Nelken I, Fishbach A, Las L, Ulanovsky N, Farkas tigation of the auditory streaming effect using event-
D (2003) Primary auditory cortex of cats: feature related brain potentials. Psychophysiology
detection or something else? Biol Cybern 36(1):22–34
89(5):397–406 Szalárdy O, Bendixen A, Tóth D, Denham SL, Winkler
Oertel D, Fay RR, Popper AN (2002) Integrative functions I (2012) Modulation-frequency acts as a primary cue
in the mammalian auditory pathway. Springer, New for auditory stream segregation. J Learn Percept
York (in press)
Pearl J (1988) Probabilistic reasoning in intelligent sys- Szalárdy O, Bőhm T, Bendixen A, Winkler I (2013) Per-
tems: networks of plausible inference. Morgan ceptual organization affects the processing of incom-
Kaufmann, San Mateo ing sounds: an ERP study. Biol Psychol (in press)
Pressnitzer D, Hupe JM (2006) Temporal dynamics of Takegata R, Brattico E, Tervaniemi M, Varyagina O,
auditory and visual bistability reveal common princi- Naatanen R, Winkler I (2005) Preattentive representa-
ples of perceptual organization. Curr Biol tion of feature conjunctions for concurrent spatially
16(13):1351–1357 distributed auditory objects. Brain Res Cogn Brain
Rand TC (1974) Letter: dichotic release from masking for Res 25(1):169–179
speech. J Acoust Soc Am 55(3):678–680 Teki S, Chait M, Kumar S, von Kriegstein K, Griffiths TD
Rensink RA (2000) Seeing, sensing, and scrutinizing. (2011) Brain bases for auditory stimulus-driven figure-
Vision Res 40(10–12):1469–1487 ground segregation. J Neurosci 31(1):164–171
Riecke L, Van Opstal AJ, Formisano E (2008) The Treisman A (1998) Feature binding, attention and object
auditory continuity illusion: a parametric perception. Philos Trans R Soc Lond B Biol Sci
investigation and filter model. Percept Psychophys 353:1295–1306
70(1):1–12 van Ee R (2009) Stochastic variations in sensory aware-
Roberts B, Glasberg BR, Moore BC (2002) Primitive ness are driven by noisy neuronal adaptation: evidence
stream segregation of tone sequences without differ- from serial correlations in perceptual bistability. J Opt
ences in fundamental frequency or passband. J Acoust Soc Am A Opt Image Sci Vis 26(12):2612–2622
Soc Am 112(5 Pt 1):2074–2085 van Noorden LPAS (1975) Temporal coherence in the
Schadwinkel S, Gutschalk A (2011) Transient bold activ- perception of tone sequences. Doctoral dissertation,
ity locked to perceptual reversals of auditory streaming Technical University Eindhoven
in human auditory cortex and inferior colliculus. Vliegen J, Oxenham AJ (1999) Sequential stream segre-
J Neurophysiol 105(5):1977–1983 gation in the absence of spectral cues. J Acoust Soc
Schwartz JL, Grimault N, Hupe JM, Moore BC, Am 105(1):339–346
Pressnitzer D (2012) Multistability in perception: von Ehrenfels C (1890) Über Gestaltqualit€aten (English
binding sensory modalities, an overview. Philos “On the qualities of form”). Vierteljahrsschr Wiss
Trans R Soc Lond B Biol Sci 367(1591):896–905 Philos 14:249–292
Shamma SA, Elhilali M (2013) von Kriegstein K, Smith DR, Patterson RD, Ives DT,
Shamma SA, Micheyl C (2010) Behind the scenes of Griffiths TD (2007) Neural representation of auditory
auditory perception. Curr Opin Neurobiol size in the human voice and in sounds from other
20(3):361–366 resonant sources. Curr Biol 17(13):1123–1128
Shamma SA, Elhilali M, Micheyl C (2011) Temporal Wang DL, Brown GJ (2006) Computational auditory
coherence and attention in auditory scene analysis. scene analysis: principles, algorithms, and applica-
Trends Neurosci 34(3):114–123 tions. Wiley/IEEE Press, New York
Shamma S, Elhilali M, Ma L, Micheyl C, Oxenham AJ, Wang DL, Chang PS (2008) An oscillatory
Pressnitzer D, Yin P, Xu Y (2013) Temporal coher- correlation model of auditory streaming. Cogn
ence and the streaming of complex sounds. Adv Exp Neurodyn 2:7–19
Med Biol 787:535–543 Warren RM, Wrightson JM, Puretz J (1988) Illusory con-
Shinn-Cunningham BG (2008) Object-based auditory and tinuity of tonal and infratonal periodic sounds.
visual attention. Trends Cogn Sci 12(5):182–186 J Acoust Soc Am 84(4):1338–1342
Shpiro A, Moreno-Bote R, Rubin N, Rinzel J (2009) Bal- Weiss Y, Simoncelli EP, Adelson EH (2002) Motion illu-
ance between noise and adaptation in competition sions as optimal percepts. Nat Neurosci 5(6):598–604
A 252 Auditory Precedence Effect
Wertheimer M (1912) Experimentelle Studien u €ber das

Sehen von Bewegung. Z Psychol 60 Auditory Precedence Effect
Wilson EC, Melcher JR, Micheyl C, Gutschalk A,
Oxenham AJ (2007) Cortical FMRI activation to
sequences of tones alternating in frequency: relation- Barbara Shinn-Cunningham
ship to perceived rate and streaming. J Neurophysiol Center for Computational Neuroscience and
97(3):2230–2238 Neural Technology, Boston University, Boston,
Winkler I (2007) Interpreting the mismatch negativity.
J Psychophysiol 21:147–163 MA, USA
Winkler I (2010) In search for auditory object representa-
tions. In: Winkle I, Czigler I (eds) Unconscious mem-
ory representations in perception: processes and Synonyms
mechanisms in the brain. John Benjamins, Amster-
dam, pp 71–106
Winkler I, Cowan N (2005) From sensory to long-term Law of the first wavefront
memory: evidence from auditory memory reactivation
studies. Exp Psychol 52(1):3–20
Winkler I, Czigler I (2012) Evidence from auditory
and visual event-related potential (ERP) Definition
studies of deviance detection (MMN and vMMN)
linking predictive coding theories and perceptual The precedence effect is a well-studied phenom-
object representations. Int J Psychophysiol enon in spatial hearing that is related to how we
83(2):132–143
Winkler I, Sussman E, Tervaniemi M, Horváth J, Ritter W, localize sounds accurately in everyday settings.
N€a€at€anen R (2003a) Preattentive auditory context Specifically, when two sound sources reach
effects. Cogn Affect Behav Neurosci 3(1):57–77 a listener close together in time, listeners often
Winkler I, Teder-Salejarvi WA, Horváth J, N€a€at€anen R, hear a single “fused” image whose perceived
Sussman E (2003b) Human auditory cortex tracks
task-irrelevant sound sources. Neuroreport direction is near the location of the first-arriving
14(16):2053–2056 sound.
Winkler I, Czigler I, Sussman E, Horvath J, Balazs
L (2005a) Preattentive binding of auditory and
visual stimulus features. J Cogn Neurosci
17(2):320–339 Detailed Description
Winkler I, Takegata R, Sussman E (2005b) Event-related
brain potentials reveal multiple stages in the percep- The Effects of Room Acoustics on Auditory
tual organization of sound. Brain Res Cogn Brain Res
25(1):291–299 Spatial Cues
Winkler I, van Zuijen TL, Sussman E, Horvath J, The signals reaching the listener’s ears directly
Naatanen R (2006) Object representation in from a sound source convey information about
the human auditory system. Eur J Neurosci the source’s location (Blauert 1997;
24(2):625–634
Winkler I, Denham SL, Nelken I (2009) Modeling the Schnupp et al. 2010). However, in ordinary set-
auditory scene: predictive regularity representations tings, soon after the direct sound reaches the
and perceptual objects. Trends Cogn Sci listener, reflected sound arrives from random
13(12):532–540 directions, coming off of walls, floors, and other
Winkler I, Denham S, Mill R, Bohm TM, Bendixen
A (2012) Multistability in auditory stream segregation: reflective surfaces. This reflected sound energy
a predictive coding view. Philos Trans R Soc Lond adds acoustically to the direct sound before enter-
B Biol Sci 367(1591):1001–1012 ing each ear, changing the total signal reaching
Yildiz IB, Kiebel SJ (2011) A hierarchical neuronal model the ear (e.g., see Allen and Berkley 1979). The
for generation and online recognition of birdsongs.
PLoS Comput Biol 7(12):e1002303 content of this reflected sound energy depends on
Zhuo G, Yu X (2011) Auditory feature binding and its the geometry of the listening space, the positions
hierarchical computational model. In: Third interna- of objects in the space, and the location and
tional conference on artificial intelligence and compu- orientation of the listener. Moreover, the reflected
tational intelligence. Springer
Zwicker E, Fastl H (1999) Psychoacoustics. Facts and sound energy differs at the left and right ears. As
models. Springer, Heidelberg/New York a result, this late-arriving sound distorts the
Auditory Precedence Effect 253 A
spatial information conveyed by the direct sound, a sound. This peripheral adaptation helps explain
degrading the spatial cues in the total signal. As why a lagging sound that arrives within a few
a result, the acoustic location information in the milliseconds of a leading sound does not convey A
signals a listener hears in everyday settings is less strong spatial cues (Hartung and Trahiotis 2001).
reliable after an initial “clean glimpse” of the However, there are numerous studies that show
direct sound that happens at sound onset, before that the perceptual dominance of the leading
the reflected energy arrives. sound extends beyond very brief lead-lag delays
that can be fully explained by peripheral adapta-
The Precedence Effect tion. It is likely that microcircuitry in the
Perceptually, judgments of the direction of brainstem contributes to the precedence effect at
a sound source depend strongly on spatial infor- longer lead-lag delays through some type of inhi-
mation in the onset of sound and relatively bition triggered by the leading sound (Xia and
weakly on spatial information in later-arriving Shinn-Cunningham 2011).
portions of sound (e.g., see Brown and Stecker
2010). This phenomenon is known as the prece-
dence effect (Wallach et al. 1949; Zurek 1987; Cross-References
Litovsky et al. 1999). Given that in ordinary
listening spaces the onset of the sound has the ▶ Sound Localization and Experience-
most reliable information about source direction, Dependent Plasticity
the precedence effect is thought to be one of the ▶ Sound Localization in Mammals, Models
reasons why listeners are relatively good at judg-
ing source location even when listening in rooms.
The precedence effect has been studied exten- References
sively with pairs of clicks (one leading and one
lagging); for such brief stimuli, the precedence Allen JB, Berkley DA (1979) Image method for efficiently
simulating small-room acoustics. J Acoust Soc Am
effect is strongest when the leading click pre-
65:943–950
cedes the lagging click by 1–5 ms and then rap- Blauert J (1997) Spatial hearing (2e). MIT Press, Cam-
idly becomes weaker (so that listeners start to bridge, MA, pp 201–287
hear the second click as a separate event and Brown AD, Stecker GC (2010) Temporal weighting of
interaural time and level differences in high-rate
then begin to localize it with increasing accuracy;
click trains. J Acoust Soc Am 128:283–292
see Blauert 1997). For more “natural” sounds, Hartung K, Trahiotis C (2001) Peripheral auditory
like speech or music, the precedence effect per- processing and investigations of the “precedence
sists for tens of ms. Many researchers argue that effect” which utilize successive transient stimuli.
the precedence effect has a longer time course for
Litovsky RY, Colburn HS, Yost WA, Guzman SJ
ongoing signals because they can be thought of as (1999) The precedence effect. J Acoust Soc Am
containing multiple “onsets” due to local energy 106:1633–1654
fluctuations, each of which can add to the prece- Schnupp J, Nelken I, King A (2010) Auditory neurosci-
ence. MIT Press, Cambridge, MA, pp 177–222
dence effect (e.g., see Zurek 1980).
Wallach H, Newman EB, Rosenzweig MR (1949) The
precedence effect in sound localization. Am
Mechanisms of the Precedence Effect J Psychol 52:315–336
Although the precedence effect is often discussed Xia J, Shinn-Cunningham BG (2011) Isolating mecha-
nisms that influence measures of the precedence effect:
as a single psychophysical phenomenon, many theoretical predictions and behavioral tests. J Acoust
different mechanisms likely contribute to the Soc Am 130:866–882
dominance of early spatial information on later- Zurek PM (1980) The precedence effect and its possible
arriving information. For instance, the most role in the avoidance of interaural ambiguities.
peripheral portion of the auditory system, the
Zurek PM (1987) The precedence effect. In: Yost WA,
auditory nerve, responds more vigorously at the Gourevitch GA (eds) Directional hearing. Springer,
onset of a sound than to later-arriving portions of New York, pp 85–105
A 254 Auditory Processing in Insects
a different context (proprioception by

Auditory Processing in Insects chordotonal organs: Meier and Reichert 1990;
van Staaden and Römer 1998). In insects, two
R. Matthias Hennig and Bernhard Ronacher types of ears are known that are sensitive to two
Department of Biology, Humboldt-Universit€at zu different physical attributes of sound, that is, the
Berlin, Berlin, Germany movement of particles in an elastic medium:
tympanal ears that are specialized to sense
changes in sound pressure and antennae or fili-
Synonyms form hairs that are sensitive to particle velocity
(Fig. 1b–d, Michelsen 1979).
Auditory pathway; Hearing
Goals of Hearing and Auditory Processing
The ability to hear sound and to extract relevant
Definition information from an acoustic signal has evolved in
three functional contexts. Notably, the production of
Auditory processing in insects serves to extract sound alone is not a sufficient indicator of hearing
relevant information from acoustic signals about ability as many insects produce defensive sounds
identity and location of acoustic objects, usually when threatened (e.g., Bura et al. 2011). Conversely,
in the context of mate attraction and predator numerous insects can hear sound but are themselves
avoidance. For this goal, insects process spectral mute (Riede 1987; Riede et al. 1990).
information from the carrier frequency of a signal
and obtain temporal information from the Acoustic Communication
sound’s amplitude modulation pattern (the enve- In many species the perception of acoustic signals
lope). Auditory processing in insects is occurs in the context of acoustic communication
constrained by size in several aspects: first, the for mate recognition and mate localization. The
signal often contains ultrasonic frequencies due most prominent examples stem from grasshop-
to small sender size; second, for localization, pers, crickets, and bushcrickets (orthoptera), but
insects have only poor directional cues because cicadas and moths have also evolved elaborate
of the small distance between their ears; and acoustic signals. A major goal of intersexual
third, due to their small brains, the auditory acoustic communication is to discriminate con-
processing capacity is limited to a small number specific from heterospecific signals which helps
of neurons. to avoid fitness losses (Fig. 2). Sexual selection
by female choice for song signals of particularly
attractive mating partners is also well known (von
Detailed Description Helversen and von Helversen 1994; Andersson
and Simmons 2006). Insects employ very stereo-
Overview and Background typed signals for acoustic communication, the
Large Diversity of Hearing Insects and Their Ears production and recognition of which has a strict
The sense of hearing has evolved in vertebrates innate basis (Bentley and Hoy 1972, von
and arthropods, and hearing organs are known Helversen and von Helversen 1975a, b, 1987).
from several orders of insects. Ears have evolved For this reason, the acoustic communication of
in rather different locations of their body, not grasshoppers, crickets, and bushcrickets has
only on the head but also on the thorax, abdomen, served as a model system to study the mecha-
and even wings and legs (Fig. 1a, Fullard and nisms of neuronal processing within the auditory
Yack 1993). Consequently, numerous neuronal pathway by carefully designed behavioral exper-
substrates for the processing of acoustic informa- iments and recordings of neuronal activity of
tion are known that evolved from single cells (Schildberger and Elsner 1994;
a mechanosensory modality employed in Huber et al. 1989; Gerhardt and Huber 2002).
Auditory Processing in Insects 255 A
Auditory Processing in
Insects, Fig. 1 Insect ears
and sound transduction. (a)
Evolution of insect ears in A
different body locations
(b–d). Types of ears: (b)
sound pressure receiver
(mammals and humans);
(c) pressure difference
receiver (insects); (d)
sound velocity receiver
(filiform hair of insects); S
sensory cells, p sound
pressure, pd sound pressure
difference, and v particle
velocity. (e) Sensory
transduction at a tympanic
ear: 1, sound wave; 2,
vibration of membrane and
movement of
mechanosensitive ion
channels due to sound
pressure; and 3–4,
membrane potential and
elicited action potential in
sensory cell. (f, g)
Intensity-response curves
of an auditory neuron in
crickets for different
background intensities and
different carrier
frequencies (f, 3 kHz; g, 16
kHz). At increasing
background intensities as
indicated by the top
symbols in (f, g), the
response curves shift to
higher intensities ((a)
Modified from Fullard and
Yack 1993, with
permission; see also there
for explanation of numbers.
(b–d) From Penzlin 2005,
with permission. (e) From
Gollisch and Herz 2005,
with permission. (f, g)
From Hildebrandt et al.
2011, with permission)
Since the nervous system of insects is small, “external world” can be reduced to
all computations must be performed by relatively a manageable set of relevant stimuli, the stimulus
few neurons. This likely imposes a strong pres- space. This rather small set of highly relevant
sure for an efficient and sparse representation of stimuli allows one to characterize how the stimuli
the external world. A big advantage of investi- are represented within the nervous system and
gating nervous system processes via the study of how these representations are transformed at dif-
communication signals is that an animal’s ferent stages of processing.
Auditory Processing in Insects, Fig. 2 Acoustic com- (e) Males prefer signals with only low-frequency compo-
munication in grasshoppers and crickets. (a, f) A sender, nents as they are typical for a female song (as in c). (g) The
the male, produces a specific sound signal that is trans- spectral content of a song of a cricket as in (f) is narrowly
mitted and perceived by a receiver, the female. (a) In some tuned to a peak frequency (5.0 kHz). (h) Auditory tuning
grasshoppers, the female responds to the male with her for the specific carrier frequency of a conspecific song for
own song (upper trace). Songs of males and females have different species of crickets (frequency axis was normal-
different pulse shapes (lower traces). (b, c) The power ized to the respective best frequency; P.p. Paroecanthus
spectra differ between males (b) and females (c) in rela- podagrosus, G.b. Gryllus bimaculatus, G.c. G.
tive content of low (L)- and high (H)- frequency compo- campestris). Note the differences in the sharpness of
nents. (d, e) Females and males respond selectively to the tuning ((a–e) From von Helversen and von Helversen
relative content of low- and high-frequency components 1997; (f) modified from Huber 1992; (g) modified from
in a signal. (d) Females prefer the combination of low and Montealegre-Z et al. 2011; and (h) from Schmidt et al.
high components as they occur in the male song (as in b). 2011, with permission)
To allow for successful communication, sig- Robertson and Money 2012) and in general
nal properties must be matched to the sensory insects’ body temperatures vary directly with
characteristics of receivers. A straightforward ambient temperature, a signal’s temporal pattern
example is the frequency tuning of a cricket will vary with ambient temperature. This creates
female’s ears to the narrowband signals of the a recognition problem for the receiver’s auditory
males (see Fig. 2g, h). However, since in many system, in particular, if sender and receiver differ
species the temporal pattern of communication in their body temperatures (von Helversen and
signals conveys the species-specific information, von Helversen 1987; Gerhardt and Huber 2002;
sender and receiver should be matched not only Ronacher et al. 2004)
for carrier frequency but also for temporal char-
acteristics. Here, a new problem arises: tempera- Detection and Avoidance of Predators
ture. Since the function of neurons and muscles is Although hearing in insects evolved more than
strongly temperature-dependent (Janssen 1992; 200 million years ago (MYA), the appearance of
bats about 60 MYA sparked an explosion of Two Types of Ears in Insects and Their Constraints
insect species with ears (Fig. 1a, Fullard and by Size and Signals
Yack 1993; Hoy et al. 1998; Stumpner and Generally, two principal types of ears are known: A
von Helversen 2001). While some species mod- particle velocity receivers and sound pressure
ified already-existing ears for the detection of receivers (Fig. 1b–d, Michelsen 1979; Faure
ultrasound, numerous other insects that had et al. 2009). (1) Particle velocity receivers exploit
previously lacked hearing evolved ultrasonic the vector component of sound particles close to
ears at this time (Miller and Surlykke 2001). the sound source. The antennae (arista) of flies,
For moths this acquired sensory ability especially of Drosophila and mosquitoes, are
may even have served as a starting point for a well-investigated model system for sound per-
the evolution of intraspecific acoustic commu- ception, auditory transduction, and active sensing
nication. Within their auditory pathways, insects (Robert and Göpfert 2002). Many insects, as well
employ different neuronal substrates for the as arthropods in general, employ filiform hairs
processing of intraspecific signals and (located on the abdomen, cerci, legs, or other
sounds from predators (see below: categorical parts of the body, Fig. 1d) of different length to
perception in crickets and parallel processing of detect sounds of predators and prey (Barth 2002).
information). Some species perform stream A general property of particle velocity receivers
segregation within individual neurons by spec- is their limitation to lower frequency ranges
tral and temporal cues (Schul and Sheridan (<500 Hz). Low frequency signals from small
2006). Numerous species of grasshoppers have senders commonly have low amplitude and thus
ears and hearing abilities, but produce no sound small range (e.g., wing movements by Drosoph-
during mate attraction. For these species, ila), and the perception of particle velocity is then
hearing likely serves for predator detection usually limited to the near field of the sender
only, and predation was the selection pressure (a distance of a few wavelengths). Since the
under which ears evolved or were conserved vector component of sound is perceived, these
(Riede 1987; Riede et al. 1990; Lehmann ears also provide directional information
et al. 2007). (Michelsen 1979; Faure et al. 2009). Their sensi-
tivity matches or even surpasses that of tympanal
Host Finding ears (Robert and Göpfert 2002). (2) The percep-
Hearing has evolved independently at least tion of sound pressure is mediated by tympanal
twice in parasitic flies in the context of finding ears and follows the same principles as in verte-
hosts for their eggs. The performance of flies in brates including humans (Fig. 1b,c, Montealegre
localizing their host by its sound signal is et al. 2012). Tympanal ears in insects arose
most impressive, in view of their tiny ears from the cuticular surface of their exoskeleton
located underneath the head (Robert et al. under which large air sacs derived from tracheal
1996; Lakes-Harlan et al. 1999). With these tubes, the respiratory system of insects, were
tympanal ears flies can detect and localize located. Since insects possess an abundance of
not only the broadband sounds of bushcrickets mechanosensory proprioreceptors for monitoring
and cicadas but also the pure tone signals the strain and movement of their cuticle, auditory
emitted by crickets. For that goal flies possess organs were prone to evolve from chordotonal
sharply tuned hearing, as is evident from the organs in almost any part of the body (e.g., the
sensitivity of their sensory receptors and legs, thorax, abdomen, and wings, Fig. 1a,
auditory interneurons (Stumpner and Lakes- Fullard and Yack 1993; van Staaden and Römer
Harlan 1996; Robert and Hoy 1998; Robert and 1998). Tympana are usually small, and the sensi-
Göpfert 2002). Another example of host finding tivity of the ears often extends into the ultrasonic
via auditory cues is a bloodsucking corethrellid frequency range. Sound localization via
fly that is attracted by frog calls (Bernal tympanal ears in larger vertebrates (Fig. 1b)
et al. 2006). requires the computation of interaural intensity
and time differences as pressure receivers do not However, insects face constraints on the com-
respond to the vector component of sound putational power provided by their small brains.
(Brown 1994; Yost 2000). Due to the small size The concept of identified neurons, in which indi-
of insects, intensity differences and in particular vidual neurons could be identified by their mor-
interaural time differences are too small to be phology and physiology, arose from
exploited directly. Insects circumvent this prob- neurobiological research in insects and other
lem by using tympanic pressure difference arthropods (Huber and Markl 1983). Computa-
receivers, in which an internal connection by tions performed by thousands of neurons in mam-
tracheal tubes exists between both ears (Fig. 1c). mals may find their counterpart in a single
Small frogs, lizards, and birds face similar prob- identifiable neuron of an insect, which illustrates
lems and also have developed pressure difference an impressive compression of function (e.g., con-
receivers. In these ears, the vibration amplitude tralateral inhibition for directional hearing is
of the tympanum is determined not only by the mediated by the lateral superior olive in mam-
sound pressure at the outer side but also by the mals and by a single local interneuron, ON1, in
sound traveling through the body to the inside of crickets and bushcrickets (Grothe 2000;
the tympanum (Autrum 1942). The resulting Selverston et al. 1985; Römer and Krusch 2000).
pressure difference between the inside and out- Notably, the processing and coding capacity
side will then determine the tympanal vibration. of insect nervous systems is restricted to relevant
The amplitude of these vibrations depends on tasks. Peripheral filters and computations for
sound direction, because the phase angle of instance aid in reducing the required processing
a given sound frequency at both sides of the power. Therefore, the ears of insects and their
tympanum is also dependent on the direction of auditory pathways are by no means all-purpose
the incident sound wave (Michelsen et al. 1994). devices, and processing is rather specific to func-
tion. Examples include crickets that distinguish
Themes of Auditory Processing in Insects only 2 categories of sound (mate signals and
A major challenge in summarizing the capacities predators, Hoy 1989) and the tympanic ear of
for auditory processing of insects results from the a moth that is equipped with only 2 sensory
overwhelming diversity of hearing species, of cells for bat detection (Boyan and Fullard
functional ears, and of the different designs of 1988). Generally, insects also employ simple
auditory pathways. The subchapters below there- algorithms for processing at the cost of acuity,
fore give only a brief overview of the computa- for example, by computing acoustic hemispheres
tional capabilities of insects for different tasks rather than localizing the angle of a sound source
and under different constraints. (von Helversen 1997; Römer and Krusch 2000).
For auditory processing, insects exploit Nevertheless, insects are by no means imprecise.
numerous general principles of sensory At least in some species their performance for
processing that are well known from other modal- temporal resolution in a gap detection task has
ities and from vertebrates, including mammals. a precision in the millisecond range, rivaling that
Among these are the capacity for sound fre- of humans (von Helversen 1972; Prinz and
quency analysis by a traveling wave, tonotopic Ronacher 2002).
representations and formation of internal neuro-
nal maps, parallel processing of information, the Basic Steps of Auditory Processing:
timing and balance of excitation and inhibition Transduction and Information Coding by
for feature extraction, lateral and contralateral Sensory Neurons
inhibition for contrast enhancement, transforma- In both tympanal ears and particle velocity
tion of coding from a temporal code to a place receivers, sound induces the vibration of
code, burst coding, resonant properties of neu- a structure (a thin membrane or a lever on
rons, selective attention, and even stream a flexible pivot, Michelsen 1979; Robert and
segregation. Göpfert 2002). The mechanical oscillation
distorts the dendrite of a scolopidial cell, which is respects (see below), the frequency resolution of
attached to the lever, the tympanum, or a tracheal insects is generally inferior to that of vertebrates.
tube (Fig. 1e1, e2). This distortion opens In crickets we find a kind of “categorical A
mechanosensitive ion channels, which are not response,” by which the frequency scale is seg-
yet fully characterized, and the resulting current mented into two parts: sound pulses with carrier
transduces the movement into a change of the frequencies above 15–20 kHz (up to 100 kHz)
cell’s membrane potential (Fig. 1e3; Gollisch evoke an avoidance response, whereas sound
and Herz 2005). Evidently, the membrane poten- pulses with frequencies between 3 and 10 kHz
tial of sensory neurons cannot follow the fast are attractive and evoke a positive steering
vibrations of the tympanal membrane in the response when flying (Moiseff et al. 1978; Hoy
kilo-Hertz range; rather, the membrane potential 1989). Remarkably, however, in an interneuron
depends on the instantaneous sound pressure of a bushcricket, a sharpening of the broader
level. Similar to vertebrates, frequency discrimi- frequency tuning of auditory receptors by fre-
nation occurs according to a frequency-place quency-dependent “lateral” inhibition similar to
transformation (Montealegre et al. 2012). Either vertebrates has been observed (Stumpner 1997).
in the sensory neuron itself or in a downstream The sharpness of frequency tuning of auditory
neuron, the membrane depolarization is then neurons is commonly described by the Q10dB
translated into a series of action potentials, score. This gives the ratio between the character-
a spike train, which encodes the sound envelope istic frequency (i.e., the frequency of the neuron’s
by modulations of the spike rate (Fig. 1e4). Spike lowest threshold) and the width of the tuning
rates of auditory afferents can be high, up to curve 10 dB above the lowest threshold. Typical
400 Hz or more (Römer 1976). values for insects are between 0.5 and 2.5 and
Although insect and vertebrate ears obviously only rarely extend to 3.5 (Hennig et al. 2004),
evolved independently, and the insect mechano- whereas in vertebrates, we find much higher
receptors (scolopidia) differ from auditory hair values, between 1 and 25 (and up to 400 in the
cells of vertebrates, in the last decade many unex- acoustic foveae of bats, Suga et al. 1997).
pected commonalities were detected. Recently, The dependency of spike rate on sound inten-
active processes were found in insect ears that sity is commonly depicted by the f-I curve (firing
serve to attain and adjust their formidable sensi- rate vs. intensity, also rate-intensity curve;
tivity, similar to the function of outer hair cells in Fig. 1f, g). The dynamic range between threshold
the mammalian cochlea (Robert and Göpfert and saturation indicates the region of discrimina-
2002; Nadrowski et al. 2011). In both locusts ble sound intensities. The steepness of the f-I
and bushcrickets, traveling waves were observed curve determines how well small sound pressure
to play an essential role in frequency discrimina- differences can be discriminated – which is
tion. As in the mammalian cochlea, traveling important for directional hearing and when fine
waves exhibit peaks at different locations, modulation details of a signal have to be assessed
depending on sound frequency (Windmill et al. (compare Fig. 1f, g). A steep f-I curve, however,
2005; Hummel et al. 2011; Montealegre et al. has the disadvantage that it covers only a small
2012). In addition, in central projections of audi- part of the relevant sound intensity range which
tory afferents, there is a tonotopic representation may extend to 100–120 dB SPL (Fig. 1f). (Note
of frequencies (Römer 1983; Römer et al. 1988; that the dB SPL scale is a logarithmic scale that
Stumpner 1996; Stölting and Stumpner 1998). expresses sound pressure relative to 20 mPa; a 100
Finally, the development of the auditory organ dB sound has a 105 larger sound pressure com-
of Drosophila depends on similar genes, e.g., of pared to the human hearing threshold at 1 kHz.)
the atonal family, as the vertebrate ear (Senthilan As in other sensory systems, the intensity range
et al. 2012). problem may be solved by adaptation, by which
Although the temporal resolution of insect the f-I curve can be adjusted to the average ambi-
ears can match that of vertebrates in some ent sound pressure levels (Benda and Herz 2003;
Benda and Hennig 2008). Spike frequency adap- Processing of Signal Envelopes Within the
tation as early as the level of sensory neurons Auditory Pathway
helps to attain a certain degree of intensity invari- Temporal Resolution and Temporal Integration
ance that is important for object identification and The communication signals of many species con-
behavioral decisions (Benda and Hennig 2008). tain fast amplitude modulations that are evalu-
However, in different neuron types we may find ated by females to assess the attractiveness of
different biophysical realizations, ranging from potential mates. Hence, we must ask whether
cell-intrinsic spike-triggered adaptation currents there are neuronal constraints that determine the
to inhibitory inputs and presynaptic adaptation behavioral limits of temporal resolution. Two
mechanisms (Gollisch and Herz 2004; widely used paradigms to investigate these limits
Hildebrandt et al. 2009). A complementary way are modulation transfer functions (MTF) and gap
to cope with the large range of encountered sound detection (for reviews see Green 1985; de Boer
intensities may be a kind of “range fractionation.” 1985; Michelsen 1985; Viemeister and Plack
In locusts and bushcrickets, for example, we find 1993; Joris et al. 2004). In the MTF paradigm
auditory receptors with a rather limited dynamic either random modulations in a certain frequency
range of 20–30 dB. Since different receptors band or sine wave modulations are used. The
exhibit a similarly broad frequency tuning but latter paradigm can be applied in neurophysio-
very different thresholds, intensity discrimination logical as well as in behavioral experiments.
is possible over a broad range (Römer 1976; Stimuli with appropriate carrier frequencies are
Römer et al. 1998). presented that exhibit sinusoidal amplitude mod-
Obviously, the spike trains of sensory neurons ulations of different frequencies and reveal what
are the only information any central nervous sys- range of modulation frequencies the system is
tem has about events in the external world. In able to represent, and if there are specific modu-
other words, the brain has to infer the structure lation frequencies to which a system responds
of the outer world from the spike trains arriving particularly well, see Fig. 3a for an example.
from various sense organs. The very successful The black dots in this diagram show the attraction
stimulus reconstruction methods seek to under- of cricket females to a 4.5 kHz tone that was
stand signal processing from the viewpoint of the amplitude modulated at frequencies between 1
central nervous system and to infer information and 50 Hz; two regions of enhanced attractive-
about an external stimulus from the spike trains ness around 3 and 30 Hz are obvious (Wendler
of sensory neurons (Rieke et al. 1997). The basic 1989; Hennig 2009). In spike train recordings one
idea is to reconstruct the stimulus envelope from can determine the average spike count (rate,
spike trains that have been recorded in response r-MTF) or evaluate how well the spikes are
to that stimulus. By this procedure we can esti- locked to a period of the stimulus envelope (tem-
mate how much information the CNS can obtain poral, t-MTF). r-MTF reveals a neuron’s filter
about a sensory stimulus and what aspects of the properties, e.g., high-pass, low-pass, band-pass,
stimulus are lost. For example, from the spike or band-reject features for sound pulse rates. The
trains of the locusts’ auditory afferents, stimuli t-MTF, in contrast, indicates how well fast mod-
with large modulation amplitudes can be ulations can be resolved by a neuron. One should
reconstructed more accurately than stimuli with be aware, though, that the construction principle
small modulation depths (Machens et al. 2001). of MTFs is based on a large amount of averaging,
Remarkably, grasshopper songs seem to be and therefore, the information provided by single
matched to this feature of the receiver’s auditory spike trains may be lower than that suggested by
periphery. This investigation has further shown a MTF (Wohlgemuth et al. 2011).
that in the very periphery of the auditory path- The second paradigm, gap detection, has
way, a single sensory neuron transmits a high been applied in behavioral experiments to
amount of information, up to 180 bits/s (Machens grasshoppers and crickets (von Helversen 1972;
et al. 2001). von Helversen and von Helversen 1997;
Auditory Processing in Insects, Fig. 3 Temporal res- syllables. This neuron responds to sound onset first with
olution of amplitude modulations in grasshoppers and a deep inhibition, an IPSP (arrows), followed by excita-
crickets. (a) Behavioral modulation transfer function of tion and spikes (upper traces). In the interrupted stimuli,
crickets (filled symbols, stimuli as in 1 and 2 at right). Best each onset after a gap triggers the IPSP anew which leads
responses are obtained if lower and higher modulation to an effective suppression of spiking (lower traces,
frequencies for pulse and chirp (i.e., groups of pulses) adapted from Ronacher & Stumpner 1988). (d) Gap detec-
are combined in one stimulus as in patterns 3 and 4 at tion in crickets (behavioral data) ((a) From Hennig 2009;
right (open symbols). (b) Gap detection in a grasshopper (b) modified from von Helversen 1972 and Franz and
measured in behavioral experiments (black curve) and Ronacher 2002; (c) modified from Ronacher and
neuronal response (red curve, AN4). (c) Response of the Stumpner 1988; and (d) from Schneider and Hennig
AN4 neuron to uninterrupted and gap containing sound 2012, with permission)
Schneider and Hennig 2012). Grasshoppers may stem from the males’ sound production sys-
detect gaps of 2–3 ms duration and in this respect tem: cricket songs are produced by a resonant
are not inferior to vertebrates (Fig. 3b, Prinz and mechanism which precludes very fast amplitude
Ronacher 2002). This high resolution of gaps changes (Bennet-Clark 1998). Hence, on the
seems to be mediated by the specific interactions receiver’s side there is no need to push the tem-
between inhibition and excitation in one identi- poral resolution to extremes. In addition, com-
fied interneuron (Fig. 3c; Ronacher and Stumpner pared to broadband signals, which are typical in
1988). The temporal resolution of cricket ears is many grasshoppers and bushcrickets, pure tone
lower compared to grasshoppers; minimal detect- signals, such as those produced by many crickets,
able gap widths are between 6 and 8 ms (Fig. 3d, tend to be more strongly affected by random
Schneider and Hennig 2012). One reason for this amplitude fluctuations when traveling through
Auditory Processing in Insects, Fig. 4 Transformation spike rates at different processing levels. (d) Variability of
of coding along the auditory pathway. (a) Scheme of interspike intervals (CV variation coefficient). (e) Vari-
a grasshopper’s auditory pathway (AFF auditory affer- ability of spike count (expressed as Fano factor FF). Axis
ents, LN local neurons, AN neurons whose axons ascend in (c–e) AFF afferents, LN local neurons, and AN ascend-
to the brain). Numbers indicate approximate numbers of ing neurons; numbers indicate sample size ((a) From
neurons at the respective levels. (b) Response (numbers of Ronacher 2013; (b–e) modified from Vogel et al. 2005,
action potentials at right) of an ascending neuron (AN1) to with permission)
five presentations of an identical stimulus. (c) Maximal
the habitat, which also sets limits for temporal Wakefield 1991; Tougaard 1998; Pohl et al.
resolution (Römer and Lewald 1992). 2013).
Temporal integration refers to the time-
intensity trading paradigm. In these experiments Transformation of Coding Along the Auditory
the minimal audible threshold was found to Pathway
depend on the duration of the stimuli used. To In insects and other arthropods, many neurons
detect very short stimuli, e.g., of 5 ms duration, can be uniquely identified on the basis of
higher sound intensities are necessary than for their characteristic morphology. The peripheral
longer, e.g., 100 ms, stimuli. The product of stage of a grasshopper’s auditory pathway com-
sound intensity and stimulus duration determines prises sensory neurons (afferents), local neurons
the threshold up to durations around 200–300 ms, whose processes are confined to the thoracic
whereas for longer stimuli, the threshold stays ganglia, and ascending neurons, whose axons
constant (Green 1985). Hence, we are confronted reach the brain. Present knowledge indicates
with an apparent discrepancy, the temporal inte- that this corresponds to a feedforward network
gration-resolution paradox (de Boer 1985): gap (Fig. 4a, Vogel et al. 2005; Vogel and Ronacher
detection and MTF paradigms yield time con- 2007).
stants in the order of 1–6 ms, whereas from the Auditory afferents exhibit high firing rates, up
time-intensity trading paradigm, we are left with to several hundred Hertz (Fig. 4c). Their tonic
time constants in the range of 150–300 ms spike responses represent the amplitude modula-
(for a discussion of this paradox and possible tion patterns of auditory stimuli by a modulation
solutions, see de Boer 1985; Viemeister and of the firing rate. The variability of their
responses is rather low (see Fig. 4d, e). The pre- The envelopes of the communication signals
cise responses of sensory neurons allow for of many insect species have a highly regular and
a good discrimination and classification of audi- repetitive structure and consist of a series of ste- A
tory stimulus ensembles (Machens et al. 2003; reotyped subunits. Remarkably, several experi-
Wohlgemuth and Ronacher 2007). ments have shown that crickets and
Along the auditory pathway the maximal grasshoppers accept song signals with random-
spike rates decrease, whereas the spike train ized or shuffled patterns as conspecific (Fig. 5,
variability increases (both spike count and inter- Pollack and Hoy 1979; von Helversen and von
spike-interval variability, Fig. 4c–e). In accor- Helversen 1998; Schmidt et al. 2008). This
dance with the larger variability of higher-order suggested that central processing may be
neurons, the temporal resolution and the classifi- restricted to the frequency domain and serve to
cation success for similar stimuli decrease mark- compute an amplitude spectrum of the song enve-
edly among the ascending neurons. At the level of lope. A crucial experiment to determine whether
afferents and among primary-like local neurons, a signal is processed in the frequency or time
we find a high classification success that depends domain is to present reversed or inverted versions
almost exclusively on the timing of spikes. In of an attractive signal (see Fig. 5e, f). Inverted
contrast, among ascending neurons the classifica- variants exhibit the same amplitude spectrum as
tion success based on a single neuron’s responses the original but differ in their phase components
decreases, and spike count differences between and thus temporal qualities. If such modified song
stimuli become more important (Wohlgemuth models show the same attractiveness as the orig-
and Ronacher 2007). Among ascending neurons, inal, a spectral analysis of the signal envelope is
the information appears to be distributed among very likely as opposed to temporal processing.
several neurons and to be represented as However, evidence from experiments as shown
a labeled-line population code (Clemens et al. in Fig. 5e, f and others revealed large differences
2011, 2012). A similar reduction in spike rates in attractiveness which suggests that insects pro-
from ascending to brain neurons is observed cess the amplitude modulations of sound stimuli
within the auditory pathways of crickets in the time domain (von Helversen and von
(Schildberger 1984; Kostarakos and Hedwig Helversen 1998; Schmidt et al. 2008; Hennig
2012). 2009).
Central Processing in the Frequency or Time Global Algorithms of Coding

Domain? Which global algorithms of coding are
With Fourier analysis, a signal’s temporal struc- implemented in the auditory pathways of insects
ture can be broken down into sine waves, each to achieve the goals of hearing (see Overview and
having a particular amplitude and phase (see, Background)? In the context of acoustic commu-
e.g., Yost 2000). If both the resulting spectra for nication as well as predator avoidance, insects
amplitude and phase are known, the original sig- have to recognize specific sound signals for
nal can be fully reconstructed. Therefore, there which they possess an innate, internal represen-
are two principal means of processing a periodic tation. Presently there is no evidence for
signal: an analysis in the frequency domain, i.e., a maplike neural representation of specific fea-
of the amplitude spectrum without phase and thus tures in the auditory pathways of insects, except
without temporal information, and an analysis in for the tonotopic frequency maps in the periphery
the time domain by the computation of temporal (Hildebrandt 2014). For predator avoidance ultra-
parameters such as durations and periods of sonic cues and the detection of strong onsets as
events. For instance, the processing of the sound a typical effect of intense bat calls appear to be
carrier by a traveling wave is equivalent to the most important. Burst coding in the auditory
computation of an amplitude spectrum for pathways of crickets and grasshoppers shows
a frequency analysis. how onsets are detected (see below, Marsat and
Auditory Processing in Insects, Fig. 5 Regular song is randomized. The song pattern in (d) from the sibling
patterns and the importance of temporal order for recog- species is not attractive. Although this experiment
nition. (a) Regular calling song of the cricket Teleogryllus suggested central processing of the signal envelope in
oceanicus. (b–d) Song models (top) for tests of preference the frequency domain, the bulk of experimental evidence
of females and relative frequency of different pulse inter- demonstrates processing in the time domain by crickets
vals (bottom). Note the differences in the interval distri- and other Insects. (e, f) Envelopes of reversed song models
butions between (b, c, and d). (b) Song model of T. for behavioral tests with grasshoppers. The pattern in (e) is
oceanicus. (c) Shuffled song model with the same relative very attractive, but the reversed version in (f) is rejected
frequency of pulse intervals as in (a). (d) Song model of T. (von Helversen and von Helversen 1998) ((a–d) From
commodus. Song models of (b) and (c) are attractive for Pollack and Hoy 1979 with permission; (e, f) from von
females of T. oceanicus, although the temporal order in c Helversen and von Helversen 1998, with permission)
Pollack 2006; Creutzig et al. 2009; see also Krahe von Helversen 1994)). In a recent alternative
and Gabbiani 2004). For the recognition of con- approach, the recognition of sound signals was
specific signals, several global schemes were pro- examined using linear-nonlinear (LN) models
posed: autocorrelation, cross-correlation with adapted from computational neuroscience
a template, or combinations of high-, low-, and (Clemens and Hennig 2013; Clemens and
band-pass filters. All these schemes are derived Ronacher 2013). For crickets as well as grasshop-
from Fourier transformation and represent math- pers, Gabor functions emerged as filters that
ematical and technical solutions, for which evi- responded best to particular subunit shapes com-
dence from several investigations exists mon to the conspecific song signal (combinations
(Schildberger 1994; Weber and Thorson 1989; of pulse and pause or pairs of pulses, Fig. 6d, e).
Hennig 2003). However, only a few studies Gabor functions (a sine wave multiplied with
have been able to reproduce the preferences a Gaussian function) are well known from sen-
exhibited by female insects over a wider range sory pathways in vertebrates (visual, Daugman
of signals (see Fig. 6a–c for response profiles 1984; Simoncelli and Olshausen 2001; Priebe
from several species of crickets, bushcrickets, and Ferster 2012; auditory, Smith and Lewicki
and grasshoppers upon presentation of stereo- 2006). However, the most notable property of the
typed sound patterns composed of subunits built proposed LN model was the independence of the
by regular pulses and pauses (von Helversen and computation from the exact timing of occurrence
of the template within a larger time window. influence sound waves on their way from sender
Therefore, this LN-model approach offers an ele- to receiver. Hence, as a rule, receivers have to
gant solution to the attractiveness of shuffled and cope with signals that are masked and degraded A
irregular song signals (see Fig. 5 and subchapter: in their temporal structure (e.g., Michelsen and
processing in the frequency or time domain). Larsen 1983; Römer et al. 1989; Römer 2001;
Small modifications of Gabor functions are capa- Schmidt et al. 2011; see also Brumm and
ble of reproducing response profiles known from Slabbekoorn 2005; Wiley 2006).
several species of insects (Fig. 6a–c, Clemens and Crickets with their pure tone songs have found
Hennig 2013). Although these filters by default a solution to reduce the impact of external noise.
describe the output of the whole recognition sys- The females’ hearing system is tuned to the car-
tem, present evidence suggests that certain iden- rier frequency of male songs and becomes
tified neurons in crickets may represent increasingly less sensitive to frequencies farther
a neuronal correlate of specific LN features (com- from the carrier frequency (Fig. 2g, h). The sharp-
pare Fig. 6e–g, Zorović and Hedwig 2011; ness of the tuning curve may depend on ecolog-
Kostarakos and Hedwig 2012; Clemens and ical conditions (Schmidt et al. 2011). A cricket
Hennig 2013). species (Paroecanthus podagrosus) living in
very noisy tropical rainforests exhibits an excep-
Local Mechanisms of Coding tionally narrow tuning (Q10dB ~4) that allows for
For insects there exist several prominent exam- an efficient suppression of ambient noise
ples of how global algorithms of coding are at (Figs. 2h, 8a). Thus, in this species we find
least in part implemented by identified neurons. a narrow peripheral filter that is perfectly
For instance, the detection of bat predators by matched to the carrier frequency of conspecific
crickets is mediated by a specific auditory neuron signals and dismisses signals from other species,
(AN2), whose bursting is crucial for a behavioral at the expense of reducing the range of perceiv-
response (Fig. 7, Nolen and Hoy 1984; Marsat able sounds (Fig. 8a). Many bushcrickets use
and Pollack 2006). Similarly, the AN12 in grass- broadband communication signals which yield
hoppers codes for a specific song feature by a twofold advantage. First, broadband signals
bursts (Creutzig et al. 2009). Specific combina- are less likely to be degraded in the biotope com-
tions of excitation and inhibition account for fea- pared to narrowband signals (Michelsen and
ture extraction in grasshoppers (gap detection by Larsen 1983; Römer and Lewald 1992). In addi-
AN4; Ronacher and Stumpner 1988 – see tempo- tion, the receiver’s nervous system can compare
ral resolution) and serve as a basis for pulse rate the neuronal signals from differently tuned audi-
detection by the identified neuron B-LI4 in tory receptors and thereby reduce intrinsic neu-
crickets (Kostarakos and Hedwig 2012). Simi- ronal noise (see below). Note that the hearing
larly, resonant properties within the auditory systems of vertebrates (mammals, owls) also
pathway of bushcrickets mediate the detection use broadband signals to reduce noise or to
of specific pulse rates and can be modeled as resolve localization ambiguities (Konishi 1990).
a property of single neurons (Bush and Schul In addition to external noise, auditory
2006; Webb et al. 2007). Short time constants of systems face a second noise problem. Neuronal
single identified neurons allow them to act as signals are inherently noisy due to the stochastic
feature detectors for ultrasonic pulses in the audi- opening and closing of ion channels. This intrin-
tory pathway of moths (Boyan and Fullard 1988). sic noise becomes evident as trial-to-trial vari-
ability of spike trains in response to repeated
Processing Under the Constraints of Noise: presentations of an identical stimulus (see
A Twofold Problem Fig. 4b). Animals with a large nervous system
Noise poses unavoidable problems for all sensory may alleviate this problem by averaging
systems. There are two classes of noise, external responses from many neurons with similar prop-
and internal. Various types of external noise erties. However, due to size constraints insects
Auditory Processing in Insects, Fig. 6 Preference pro- filters (left panels) and the respective nonlinearities
files of female insects for song signals and LN models. (central panel) that predict behavioral responses of female
(a–c) Preference profiles for song signals in the time crickets (right panel). Note that the duration of the linear
domain by crickets, bush crickets, and grasshoppers (A: filters is only 64 ms. The red dot in the right panel of (d)
oce oceanicus, com commodus; B: cau caudata, can refers to the pattern in (e). (e) Output of the LN models (d)
cantans, vir viridissima; C: br brunneus, big biguttulus, in response to a song model (upper trace). Response of the
mo mollis). (d–e): LN models account for preference linear filters to the song model before (left panels) and
functions of the cricket Gryllus bimaculatus. (d) Two after passing through the nonlinearity (right panels). The
LN models (red, green) that predict the preferences for upper filter (red in d and e) responds like an onset detector;
pulse patterns by female crickets (G. bimaculatus): linear the lower filter resembles a Gabor-function and selectively
probably cannot afford this solution. Under some ascending neurons information is distributed
conditions noise may play a beneficial role and according to a labeled-line code (Clemens et al.
improve neural computations, for example, by 2011, 2012; see also Clemens and Ronacher A
stochastic resonance (for review see McDonnell 2013).
and Ward 2011).
To quantify the trial-to-trial variability or to Directional Hearing
compare neuronal responses to different stimuli, The biophysical qualities of the pressure differ-
one can apply a spike train metric (see, e.g., van ence receivers equip the ears of insects with
Rossum 2001). This metric describes the similar- a directional dependence of the vibrational
ity of two spike trains by a single number, which amplitudes of their tympana (Michelsen 1998).
is in line with our intuition of distance: small Sensory neurons reflect this dependence in spike
values indicate high similarity. The metric uses numbers, and for some species, this difference in
an adjustable parameter to take into account both response strength is also translated into timing
differences in the spike count as well as in the differences (enhanced by ramps Krahe and
timing of spikes. This metric approach was used Ronacher 1993; Ronacher and Krahe 2000). The
to determine the relative impacts of external and contrast in response magnitude of sensory neu-
intrinsic noise on the encoding of envelope- rons for left and right differences is enlarged by
degraded stimuli by auditory neurons of grass- contralateral inhibition from local interneurons
hoppers (Neuhofer et al. 2011). Unexpectedly, (ON neurons in crickets, Selverston et al. 1985;
the contribution of external signal degradation bushcrickets, Römer and Krusch 2000; LN in
to the overall spike train distances was low: grasshoppers, Marquart 1985). The representa-
even for the highest degradation level, its amount tion of sound from one side is therefore enhanced,
did not exceed that of intrinsic noise (Fig. 8b). because the responses to sound from the opposite
As long as rather different signals have to be side are suppressed, and this takes place already
analyzed, neuronal noise may not be a very seri- at early levels of auditory processing of sound
ous problem. However, with communication sig- direction, usually at the first synapse after sensory
nals that serve to attract mates, it will be receptors. This mechanism of contralateral inhi-
necessary to discriminate between similar signals bition in auditory pathways transforms the acous-
and to detect small deviations from a species- tic environment of insects into acoustic
specific pattern. This is especially the case if hemispheres. These hemispheres enable a rather
quality cues from the sound signals of potential accurate distinction of left and right sound signals
partners are to be extracted. Remarkably enough, at the cost of accuracy in determining the angle of
the spike trains of even a single auditory afferent the sound source (lateralization in grasshoppers,
allow for an almost perfect discrimination of von Helversen and Rheinlaender 1988).
songs of different males of one species (Machens A corollary of the computation of acoustic hemi-
et al. 2003). At higher stages in the auditory spheres is that sound signals are selectively
pathway, however, the discrimination deterio- represented on one side of the insect or the
rates and intrinsic noise cannot be neglected as other. This computation resembles the phenome-
a limiting factor. Grasshoppers appear to have non of selective attention (Pollack 1988) and
circumvented this intrinsic noise problem by allows insects to discriminate sound sources
changing the coding scheme at a rather peripheral in much the same way as the cocktail party
stage of processing (see Fig. 4a): among the effect well known from humans (Fig. 9,
ä
Auditory Processing in Insects, Fig. 6 (continued) responses to sound patterns with different pulse rates
responds to pairs of pulses (green in d and e). (f–g) The ((a–c) From Hennig et al. 2004; (d, e) modified from
response pattern of an auditory interneuron in the cricket Clemens and Hennig 2013; and (f, g) from Zorović and
brain resembles the output of the onset detector (red in Hedwig 2011, with permission)
d, e). The neuronal discharges in (g) illustrate onset
Auditory Processing in Insects, Fig. 7 Burst coding by and behavioral response (bottom trace: abdominal move-
an identified interneuron (AN2) and avoidance behavior in ments away from the sound source). (c) Amplitude of
crickets to bat like stimuli. (a) Bursts in AN2 in response abdominal movements after isolated action potentials
to amplitude-modulated stimuli (carrier frequency: 30 (gray) or bursts (black). Positive values indicate abdomen
kHz); burst spikes are marked as black dots in raster flexion away from the sound source (From Marsat and
plot. (b) Response of AN2 to a sound stimulus (as in a) Pollack 2006, with permission)
Auditory Processing in Insects, Fig. 8 Processing dissimilarities of three representative neurons, assessed
under the constraints of noise: a twofold problem. (a) with the van Rossum metric and corrected for spike rate
Song of a tropical cricket Paroecanthus podagrosus differences. The black arrow at “orig” indicates the aver-
(upper trace). Lower traces. 1: Song envelope. 2: Song age spike train distance found for repeated presentation of
envelope under ambient noise levels as recorded in the the original song pattern, i.e., the result of trial-to-trial
habitat. 3: Song plus noise less efficiently filtered with the variability. The open arrow indicates the additional dis-
broader tuning curve of a European cricket (Gryllus tance caused by the most strongly degraded signal. AFF
bimaculatus). 4: Song plus noise filtered with the narrow sensory neuron, TN1, and SN1 two primary-like local
tuning curve of P. podagrosus, compare Fig. 2h (Adapted neurons. ((a) Modified from Schmidt et al. 2011. (b)
from Schmidt et al. 2011). (b) Effects of intrinsic noise Modified from Neuhofer et al. 2011 and Ronacher 2013,
and external signal degradation on spike train with permission)
Römer and Krusch 2000). Notably, contralateral computations for enhanced directional responses
inhibition and acoustic hemispheres can exert can also affect the singing behavior of males and
a considerable influence on mate choice as promote synchronization and thus chorusing
females of bushcrickets prefer males that take among males (Greenfield and Roizen 1993;
a leader role among singing males, since the Greenfield 1994; Hartbauer et al. 2005). Due to
representation of the song signal of a follower is the computational conflict between representation
suppressed in their auditory pathway (Hartbauer of sound pattern and sound source, grasshoppers
et al. 2005; Siegert et al. 2011). These local split the sensory pathways for processing of cues
Auditory Processing in Insects, Fig. 9 Acoustic hemi- from either side. (c) Both interneurons selectively repre-
spheres and selective representation of sound patterns in sent only the sound pattern from one side (inset: experi-
bush crickets. (a, b) Left and right specimens of the local mental arrangement, correlation coefficient of the spike
interneuron ON1 were recorded simultaneously, while train with the sound pattern as a measure for copying
different sound signals (pattern 1 and 2) were presented fidelity) (From Römer and Krusch 2000, with permission)
for pattern and direction by parallel processing (von are often performed by single neurons or very
Helversen 1984; Ronacher et al. 1986). However, small populations of identifiable neurons. This
in the evolutionarily older communication systems size constraint and the focus on a few relevant
of crickets and bushcrickets, serial processing of tasks facilitate experimental approaches.
pattern and direction appears to prevail (Wendler
1989; Stabel et al. 1989; von Helversen and von Acknowledgment We want to thank the members of our
Helversen 1995; Schul et al. 1998). lab who contributed to several of the cited studies. Special
thanks are due to Dr. Michael Reichert who gave helpful
advice on the English style and substantially improved the
manuscript, as well as to an anonymous reviewer whom
Conclusions we owe many helpful suggestions.
Auditory processing in insects is constrained by

small body size and a relatively small number of References
neurons. Thus, the hearing capacities are focused
on highly relevant tasks, such as predator detec- Andersson M, Simmons LW (2006) Sexual selection and
tion, mate attraction, and, in some cases, mate choice. Trends Ecol Evol 21:296–302
Autrum HJ (1942) Schallempfang bei Mensch und Tier.
eavesdropping on prey or host signals. Insect ears Naturwissenschaften 5:69–85
evolved from cuticular mechanosensory proprior- Barth FG (2002) A spider’s world: senses and behavior.
eceptors and may respond to different aspects of Springer, Berlin/Heidelberg/New York
sound waves – particle velocity or sound pressure; Benda J, Hennig RM (2008) Spike-frequency
adaptation generates intensity invariance in
ears can be found in almost any part of the body, a primary auditory interneuron. J Comput Neurosci
and as a consequence, different neural structures 24:113–136
are involved in auditory processing. In the audi- Benda J, Herz AVM (2003) A universal model for
tory pathways, we find similar computational prin- spike-frequency adaptation. Neural Comput
15:2523–2564
ciples and transformations of sensory Bennet-Clark HC (1998) Size and scale effects as con-
representations as in vertebrates, however, with straints in insect sound communication. Phil Trans
an important difference: complex computations R Soc Lond B 353:407–419
Bentley DR, Hoy RR (1972) The genetic control of the Gollisch T, Herz AVM (2004) Input-driven components
neuronal network generating cricket (Teleogryllus of spike-frequency adaptation can be unmasked
gryllus) song pattern. Anim Behav 20:478–492 in vivo. J Neurosci 24:7435–7444
Bernal XE, Rand AS, Ryan MJ (2006) Acoustic prefer- Gollisch T, Herz AVM (2005) Disentangling sub-
ences and localization performance of blood-sucking millisecond processes within an auditory transduction
flies (Corethrella Coquillett) to tungara frog calls. chain. PLoS Biol 3(e8):1–11
Behav Ecol 17:709–715 Green DM (1985) Temporal factors in psychoacoustics.
Boyan GS, Fullard JH (1988) Information processing at In: Michelsen A (ed) Time resolution in
a central synapse suggests a noise filter in the auditory auditory systems. Springer, Berlin/Heidelberg,
pathway of the noctuid moth. J Comp Physiol pp 122–140
A 164:251–258 Greenfield MD (1994) Synchronous and alternating cho-
Brown CH (1994) Sound localization. In: Fay RR, Popper ruses in insects and anurans: common mechanisms and
AN (eds) Comparative hearing: mammals. Springer, diverse functions. Am Zool 34:605–615
New York/Berlin, pp 57–96 Greenfield MD, Roizen I (1993) Katydid synchronous
Brumm H, Slabbekoorn H (2005) Acoustic communica- chorusing is an evolutionary stable outcome of female
tion in noise. Adv Stud Behav 35:151–209 choice. Nature 364:618–620
Bura VL, Rower VG, Martin PR, Yack JE (2011) Whis- Grothe B (2000) The evolution of temporal processing in
tling in caterpillars (Amorpha juglandis, the medial superior olive, an auditory brainstem struc-
Bombycoidea): sound-producing mechanism and ture. Prog Neurobiol 61:581–610
function. J Exp Biol 214:30–37 Hartbauer M, Kratzer S, Steiner K, Römer H (2005)
Bush SL, Schul J (2006) Pulse-rate recognition in an Mechanisms for synchrony and alternation in song
insect: evidence of a role for oscillatory neurons. interactions of the bushcricket Mecopoda elongata
J Comp Physiol A 192:113–121 (Tettigoniidae, Orthoptera). J Comp Physiol
Clemens J, Hennig RM (2013) Computational principles A 191:175–188
underlying the recognition of acoustic signals in Hennig RM (2003) Acoustic feature extraction by cross-
insects. J Comput Neurosci 35:75–85 correlation in crickets? J Comp Physiol
Clemens J, Ronacher B (2013) Feature extraction and A 189:589–598
integration underlying perceptual decision making Hennig RM (2009) Walking in Fourier’s space: algo-
during courtship in grasshoppers. J Neurosci rithms for the computation of periodicities in song
33:12136–12145 patterns by the cricket Gryllus bimaculatus. J Comp
Clemens J, Kutzki O, Ronacher B, Schreiber S, Physiol A 195:971–987
Wohlgemuth S (2011) Efficient transformation of an Hennig RM, Franz A, Stumpner A (2004) Processing of
auditory population code in a small sensory system. auditory information in insects. Microsc Res Tech
Proc Natl Acad Sci USA 108:13812–13817 63:351–374
Clemens J, Wohlgemuth S, Ronacher B (2012) Nonlinear Hildebrandt KJ (2014) Neural maps in insect versus ver-
computations underlying temporal and population tebrate auditory systems. Curr Opin Neurobiol
sparseness in the auditory system of the grasshopper. 24:82–87
J Neurosci 32:10053–10062 Hildebrandt KJ, Benda J, Hennig RM (2009) The origin of
Creutzig F, Wohlgemuth S, Stumpner A, Benda J, adaptation in the auditory pathway of locusts is spe-
Ronacher B, Herz AVM (2009) Time-scale invariant cific to cell type and function. J Neurosci
representation of acoustic communication signals by 29:2626–2636
a bursting neuron. J Neurosci 29:2575–2580 Hildebrandt KJ, Benda J, Hennig RM (2011) Multiple
Daugman JG (1984) Spatial visual channels in the Fourier arithmetic operations in a single neuron: the recruit-
plane. Vision Res 24:891–910 ment of adaptation processes in the cricket auditory
de Boer E (1985) Auditory time constants: a paradox? In: pathway depends on sensory context. J Neurosci
Michelsen A (ed) Time resolution in auditory systems. 31:14142–14150
Springer, Berlin/Heidelberg, pp 141–157 Hoy RR (1989) Startle, categorical response, and attention
Faure PA, Mason AC, Yack JE (2009) Invertebrate ears in acoustic behavior of insects. Ann Rev Neurosci
and hearing. In: Binder MD, Hirokawa N, Windhorst 12:355–375
U, Hirsch MC (eds) Encyclopedia of neuroscience. Hoy RR, Popper AN, Fay RR (eds) (1998) Comparative
Springer, Berlin, pp 2035–2042 hearing: insects. Springer, New York
Franz A, Ronacher B (2002) Temperature dependence of Huber F (1992) Verhalten und Neurobiologie von
temporal resolution in an insect nervous system. J stimmbegabten Insekten. Naturwissenschaften
Comp Physiol A 188:261–271 79:393–406
Fullard JH, Yack JE (1993) The evolutionary biology of Huber F, Markl H (1983) Neuroethology and behavioural
insect hearing. Trends Ecol Evol 8:248–252 physiology: roots and growing points. Springer, Hei-
Gerhardt HC, Huber F (2002) Acoustic communication in delberg/New York
insects and anurans. University of Chicago Press, Huber F, Moore TE, Loher W (eds) (1989) Cricket behav-
Chicago ior and neurobiology. Cornell University Press, Ithaca
Hummel J, Kössl M, Nowotny M (2011) Sound-induced Michelsen A, Larsen ON (1983) Strategies for acoustic
tympanal membrane motion in bushcrickets and its communication in complex environments. In: Huber
relationship to sensory output. J Exp Biol F, Markl H (eds) Neuroethology and behavioural phys-
214:3596–3604 iology. Springer, Berlin/Heidelberg, pp 321–331 A
Janssen R (1992) Thermal influences on nervous system Michelsen A, Popov A, Lewis B (1994) Physics of direc-
function. Neurosci Biobehav Rev 16:399–413 tional hearing in the cricket Gryllus bimaculatus.
Joris PX, Schreiner CE, Rees A (2004) Neural processing J Comp Physiol A 175:153–164
of amplitude-modulated sounds. Physiol Rev Miller LA, Surlykke A (2001) How some insects detect
84:541–577 and avoid being eaten by bats: tactics and
Konishi M (1990) Similar algorithms in different sensory countertactics of prey and predator. BioScience
systems and animals. Cold Spring Harb Symp Quant 51:571–582
Biol 55:575–584 Moiseff A, Pollack G, Hoy R (1978) Steering responses of
Kostarakos K, Hedwig B (2012) Calling song recognition flying crickets to sound and ultrasound: mate attraction
in female crickets: temporal tuning of identified brain and predator avoidance. Proc Natl Acad Sci USA
neurons matches behavior. J Neurosci 32:9601–9612 75:4052–4056
Krahe R, Gabbiani F (2004) Burst firing in sensory sys- Montealegre-Z F, Jonsson T, Robert D (2011) Sound
tems. Nat Rev Neurosci 5:13–24 radiation and wing mechanics in stridulating field
Krahe R, Ronacher B (1993) Long rise times of sound crickets (Orthoptera: Gryllidae). J Exp Biol 214:
pulses in grasshopper songs improve the directionality 2105–2117
cues received by the CNS from auditory receptors. Montealegre-Z F, Jonsson T, Robson-Brown KA, Postles
J Comp Physiol A 173:425–434 M, Robert D (2012) Convergent evolution
Lakes-Harlan R, Stölting H, Stumpner A (1999) Conver- between insect and mammalian audition. Science
gent evolution of insect hearing organs from 338:968–971
a preadaptive structure. Proc R Soc Lond Nadrowski B, Effertz T, Senthilan PR, Göpfert MC
B 266:1161–1167 (2011) Antennal hearing in insects – new findings,
Lehmann GUC, Strauß J, Lakes-Harlan R (2007) Listen- new questions. Hearing Res 273:7–13
ing when there is no sexual signalling? Maintenance of Neuhofer D, Stemmler M, Ronacher B (2011) Neuronal
hearing in the asexual bushcricket Poecilimon precision and the limits for acoustic signal recognition
intermedius. J Comp Physiol A 193:537–545 in a small neuronal network. J Comp Physiol
Machens CK, Stemmler MB, Prinz P, Krahe R, Ronacher A 197:251–265
B, Herz AVM (2001) Representation of acoustic com- Nolen TG, Hoy RR (1984) Initiation of behavior by single
munication signals by insect auditory receptor neu- neurons: the role of behavioral context. Science
rons. J Neurosci 21:3215–3227 226:992–994
Machens CK, Sch€utze H, Franz A, Stemmler MB, Penzlin H (2005) Lehrbuch der Tierphysiologie. Elsevier,
Ronacher B, Herz AVM (2003) Auditory receptor M€ unchen
neurons preserve characteristic differences between Pohl NU, Slabbekoorn H, Neubauer H, Heil P, Klump
conspecific communication signals. Nat Neurosci GM, Langemann U (2013) Why longer song elements
6:341–342 are easier to detect: threshold level-duration functions
Marquart V (1985) Local interneurons mediating excita- in the Great Tit and comparison with human data.
tion and inhibition onto ascending neurons in the audi- J Comp Physiol A 199:239–252
tory pathway of grasshoppers. Naturwissenschaften Pollack GS (1988) Selective attention in an insect auditory
72:42–44 neuron. J Neurosci 8:2635–2639
Marsat G, Pollack GS (2006) A behavioural role for fea- Pollack GS, Hoy RR (1979) Temporal pattern as a cue for
ture detection by sensory bursts. J Neurosci species-specific calling song recognition in crickets.
26:10542–10547 Science 204:429–432
McDonnell MD, Ward LM (2011) The benefits of noise in Priebe NJ, Ferster D (2012) Mechanisms of neuronal
neural systems: bridging theory and experiment. Nat computation in mammalian visual cortex. Neuron
Rev Neurosci 12:415–425 75:194–208
Meier T, Reichert H (1990) Embryonic development and Prinz P, Ronacher B (2002) Temporal modulation transfer
evolutionary origin of the Orthopteran auditory functions in auditory receptor fibres of the locust
organs. J Neurobiol 21:592–610 (Locusta migratoria L.). J Comp Physiol
Michelsen A (1979) Insect ears as mechanical systems. A 188:577–587
Am Sci 67:696–706 Riede K (1987) A comparative study of mating behaviour
Michelsen A (ed) (1985) Time resolution in auditory sys- in some neotropical grasshoppers (Acridoidea). Ethol-
tems. Springer, Berlin/Heidelberg ogy 76:265–296
Michelsen A (1998) Biophysics of sound localization in Riede K, K€amper G, Höfler I (1990) Tympana, auditory
insects. In: Hoy RR, Popper AN, Fay RR (eds) Com- thresholds, and projection areas of tympanal nerves in
parative hearing: insects. Springer, Berlin/New York, singing and silent grasshoppers (Insects, Acridoidea).
pp 18–62 Zoomorphology 109:223–230
Rieke F, Warland D, de Ruyter van Steveninck R, Bialek directional information in the CNS of a grasshopper,
W (1997) Spikes – exploring the neural code. MIT as revealed by surgical experiments. J Comp Physiol
Press, Cambridge, MA A 158:363–374
Robert D, Göpfert MC (2002) Novel schemes for hearing Ronacher B, Franz A, Wohlgemuth S, Hennig H (2004)
and orientation in insects. Curr Opin Neurobiol Variability of spike trains and the processing of tem-
12:715–720 poral patterns of acoustic signals–problems, con-
Robert D, Hoy RR (1998) The evolutionary innovation of straints, and solutions. J Comp Physiol A 190:257–277
tympanal hearing in Diptera. In: Hoy RR, Popper AN, Schildberger K (1984) Temporal selectivity of identified
Fay RR (eds) Comparative hearing: insects. Springer, auditory neurons in the cricket brain. J Comp Physiol
New York, pp 197–227 A 155:171–185
Robert D, Miles RN, Hoy RR (1996) Directional hearing Schildberger K (1994) The auditory pathway of crickets:
by mechanical coupling in the parasitoid fly Ormia adaptations for intraspecific acoustic communication.
ochracea. J Comp Physiol A 179:29–44 In: Schildberger K, Elsner N (eds) Neural basis of
Robertson RM, Money TG (2012) Temperature and neu- behavioural adaptations. G Fischer, Stuttgart,
ronal circuit function: compensation, tuning and toler- pp 209–225
ance. Curr Opin Neurobiol 22:724–734 Schildberger K, Elsner N (1994) Neural basis of
Römer H (1976) Die Informationsverarbeitung behavioural adaptations. G. Fischer, Stuttgart
tympanaler Rezeptorelemente von Locusta Schmidt A, Ronacher B, Hennig RM (2008) The role of
migratoria. J Comp Physiol A 109:101–122 frequency, phase and time for processing amplitude
Römer H (1983) Tonotopic organization of the auditory modulated signals by grasshoppers. J Comp Physiol
neuropile in the bushcricket Tettigonia viridissima. A 194:221–233
Nature 306:60–62 Schmidt AKD, Riede K, Römer H (2011) High back-
Römer H (2001) Ecological constraints for sound commu- ground noise shapes selective auditory filters in
nication: from grasshoppers to elephants. In: Barth FG, a tropical cricket. J Exp Biol 214:1754–1762
Schmid A (eds) Ecology of sensing. Springer, Berlin/ Schneider E, Hennig RM (2012) Temporal resolution for
Heidelberg/New York, pp 59–77 calling song signals by female crickets, Gryllus
Römer H, Krusch M (2000) A gain-control mechanism for bimaculatus. J Comp Physiol A 198:181–191
processing of chorus sounds in the afferent auditory Schul J, Sheridan RA (2006) Auditory stream segregation
pathway of the bushcricket Tettigonia viridissima in an insect. Neuroscience 138:1–4
(Orthoptera, Tettigoniidae). J Comp Physiol Schul J, von Helversen D, Weber T (1998) Selective
A 186:181–191 phonotaxis in Tettigonia cantans and T. viridissima
Römer H, Lewald J (1992) High-frequency sound trans- in song recognition and discrimination. J Comp
mission in natural habitats: implications for the evolu- Physiol A182:687–694
tion of insect acoustic communication. Behav Ecol Selverston A, Kleindienst H-U, Huber F (1985) Synaptic
Sociobiol 29:437–444 connectivity between cricket auditory interneurons as
Römer H, Marquart V, Hardt M (1988) Organization of studied by selective photoinactivation. J Neurosci
a sensory neuropile in the auditory pathway of two 5:1283–1292
groups of Orthoptera. J Comp Neurol 275:201–215 Senthilan PR, Piepenbrock D, Ovezmyradov G,
Römer H, Bailey WJ, Dadour I (1989) Insect hearing in Nadrowski B, Bechstedt S, Pauls S, Winkler M,
the field: III masking by noise. J Comp Physiol Möbius W, Howard J, Göpfert MC (2012) Drosophila
A 164:609–620 auditory organ genes and genetic hearing defects. Cell
Römer H, Spickermann M, Bailey W (1998) Sensory basis 150:1042–1054
for sound intensity discrimination in the bushcricket Siegert ME, Römer H, Hashim R, Hartbauer M (2011)
Requena verticalis (Tettigoniidae, Orthoptera). Neuronal correlates of a preference for leading signals
J Comp Physiol A 182:595–607 in the synchronizing bushcricket Mecopoda elongata
Ronacher B (2013) Processing of species-specific signals (Orthoptera: Tettigoniidae). J Exp Biol 214:3924–3934
in the auditory pathway of grasshoppers. In: Hedwig Simoncelli E, Olshausen B (2001) Natural image statistics
B (ed) Insect hearing and acoustic communication. and neural representation. Annu Rev Neurosci
Springer, Berlin, Heidelberg, pp. 185–204 24:1193–1216
Ronacher B, Krahe R (2000) Temporal integration vs. Smith EC, Lewicki MS (2006) Efficient auditory coding.
parallel processing: coping with the variability of neu- Nature 439:978–982
ronal messages in directional hearing of insects. Eur Stabel J, Wendler G, Scharstein H (1989) Cricket
J Neurosci 12:2147–2156 phonotaxis: localization depends on recognition of
Ronacher B, Stumpner A (1988) Filtering of the calling song pattern. J Comp Physiol
behaviourally relevant temporal parameters of A 165:165–177
a grasshopper´s song by an auditory interneuron. Stölting H, Stumpner A (1998) Tonotopic organization of
J Comp Physiol A 163:517–523 auditory receptors of the bushcricket Pholidoptera
Ronacher B, von Helversen D, von Helversen O (1986) griseoaptera (Tettigoniidae, Decticinae). Cell Tissue
Routes and stations in the processing of auditory Res 294:377–386
Stumpner A (1996) Tonotopic organization of the hearing von Helversen D, von Helversen O (1975b) Verhaltens-
organ in a bushcricket. Naturwissenschaften 83:81–84 genetische Untersuchungen am akustischen
Stumpner A (1997) An auditory interneurone tuned to the Kommunikationssystem der Feldheuschrecken
male song frequency in the duetting bushcricket (Orthoptera, Acrididae). II. Das Lautschema von A
Ancistrura nigrovittata (Orthoptera, Phaneropteridae). Artbastarden zwischen Chorthippus biguttulus und C.
J Exp Biol 200:1089–1101 mollis. J Comp Physiol 104:301–323
Stumpner A, Lakes-Harlan R (1996) Auditory interneu- von Helversen O, von Helversen D (1987) Innate receiver
rons in a hearing fly (Therobia leonidei, Ormiini, mechanisms in the acoustic communication of orthop-
Tachinidae, Diptera). J Comp Physiol A 178:227–233 teran insects. In: Guthrie DM (ed) Aims and methods
Stumpner A, von Helversen D (2001) Evolution and func- in neuroethology. Manchester Univ Press, Manchester,
tion of auditory systems in insects. Naturwis- pp 104–150
senschaften 88:159–170 von Helversen O, von Helversen D (1994) Forces driving
Suga N, Zhang Y, Yan J (1997) Sharpening of frequency coevolution of song and song recognition in grasshop-
tuning by inhibition in the thalamic auditory nucleus of pers. In: Schildberger K, Elsner N (eds) Neural basis of
the mustached bat. J Neurophysiol 77:2098–2114 behavioural adaptations. G. Fischer, Stuttgart,
Tougaard J (1998) Detection of short pure-tone stimuli in pp 253–284
the noctuid ear: what are temporal integration and von Helversen D, von Helversen O (1995) Acoustic pat-
integration time all about? J Comp Physiol tern recognition and orientation in orthopteran insects:
A 183:563–572 parallel or serial processing. J Comp Physiol
van Rossum MCW (2001) A novel spike distance. Neural A 177:767–774
Comput 13:751–763 von Helversen D, von Helversen O (1997) Recognition of
van Staaden MJ, Römer H (1998) Evolutionary transition sex in the acoustic communication of the grasshopper
from stretch to hearing organs in ancient grasshoppers. Chorthippus biguttulus (Orthoptera, Acrididae).
Nature 394:773–776 J Comp Physiol A 180:373–386
Viemeister NF, Plack CJ (1993) Time analysis. In: Yost von Helversen D, von Helversen O (1998) Acoustic
WA, Popper AN, Fay RR (eds) Human psychophysics. pattern recognition in a grasshopper: processing in
Springer, Berlin/Heidelberg/New York, pp 116–154 the frequency or time domain? Biol Cybern
Viemeister NF, Wakefield GH (1991) Temporal 79:467–476
integration and multiple looks. J Acoust Soc Am Webb B, Wessnitzer J, Bush SL, Schul J, Buchli J, Ijspeert
90:858–865 A (2007) Resonant neurons and bushcricket behav-
Vogel A, Ronacher B (2007) Neural correlations increase iour. J Comp Physiol A 193:285–288
between consecutive processing levels in the auditory Weber T, Thorson J (1989) Phonotactic behavior of walk-
system of locusts. J Neurophysiol 97:3376–3385 ing crickets. In: Huber F, Moore TE, Loher W (eds)
Vogel A, Hennig RM, Ronacher B (2005) Increase of Cricket behavior and neurobiology. Cornell University
neuronal response variability at higher processing Press, Ithaca, pp 310–339
levels as revealed by simultaneous recordings. Wendler G (1989) Acoustic orientation in crickets in the
J Neurophysiol 93:3548–3559 presence of two sound sources. Naturwissenschaften
von Helversen D (1972) Gesang des M€annchens und 76:128–129
Lautschema des Weibchens bei der Feldheuschrecke Wiley RH (2006) Signal detection and animal communi-
Chorthippus biguttulus (Orthoptera, Acrididae). cation. Adv Study Behav 36:217–247
J Comp Physiol 81:381–422 Windmill JFG, Göpfert MC, Robert D (2005) Tympanal
von Helversen D (1984) Parallel processing in auditory travelling waves in migratory locusts. J Exp Biol
pattern recognition and directional analysis by the 208:157–168
grasshopper Chorthippus biguttulus L (Acrididae). Wohlgemuth S, Ronacher B (2007) Auditory discrimina-
J Comp Physiol A 154:837–846 tion of amplitude modulations based on metric
von Helversen D (1997) Acoustic communication and distances of spike trains. J Neurophysiol
orientation in grasshoppers. In: Lehrer M (ed) Orien- 97:3082–3092
tation and communication in arthropods. Birkh€auser, Wohlgemuth S, Vogel A, Ronacher B (2011) Encoding
Basel, pp 301–341 of amplitude modulations by auditory neurons of
von Helversen D, Rheinlaender (1988) Interaural intensity the locust: influence of modulation frequency, rise
and time discrimination in an unrestrained grasshop- time, and modulation depth. J Comp Physiol
per: a tentative behavioural approach. J Comp Physiol A 197:61–74
A 162:333–340 Yost WA (2000) Fundamentals of hearing – an introduc-
von Helversen D, von Helversen O (1975a) Verhaltens- tion. Academic, San Diego/New York
genetische Untersuchungen am akustischen Zorović M, Hedwig B (2011) Processing of species-
Kommunikationssystem der Feldheuschrecken specific auditory patterns in the cricket brain by
(Orthoptera, Acrididae). I. Der Gesang von ascending, local, and descending neurons
Artbastarden zwischen Chorthippus biguttulus und C. during standing and walking. J Neurophysiol
mollis. J Comp Physiol 104:273–299 105:2181–2194
A 274 Auditory Prosthesis
This allows taking advantage of the tonotopic

Auditory Prosthesis organization of the cochlea and the auditory
nerve, where the high frequencies are encoded
Johan H. M. Frijns and Jeroen J. Briaire at the basal end, while the low frequencies
ENT Department, Leiden University Medical are encoded more to the apical end (Ruggero
Center, Leiden, The Netherlands 2009). In this way, each electrode contact of
a multichannel implant aims to stimulate
a different neural population, which physiologi-
Synonyms cally encodes a certain pitch as determined by its
intracochlear position. With current devices,
Bionic ear; Cochlear implant; Inner ear prosthesis however, the wish to encode all spectral informa-
tion relevant for speech understanding leads to
a mismatch between their tonotopic map and the
Definition physiological one.
In 1984, the first cochlear implant obtained
An auditory prosthesis is an implantable device FDA approval for implantation in adults. This
used to (partially) restore the auditory function in was followed by an NIH consensus in 1995 stat-
people with a severe to profound hearing loss by ing that cochlear implantation is a proven and
electrically stimulating the auditory neural path- effective rehabilitation method for deaf children
way. The cochlear implant, stimulating the audi- and deaf adults. Today, over 188,000 people have
tory nerve from within the cochlea, is widely received a cochlear implant (http://report.nih.
accepted as the standard rehabilitation device gov/nihfactsheets/ViewFactSheet.aspx?csid=83).
for this population. An auditory brain stem Initially, the implants provided a signal
implant uses the same technology to stimulate function and an aid in lipreading. Nowadays,
the neurons of the cochlear nucleus in the brain driven by improved electronics and speech coding
stem and is used when the cochlea is not accessi- strategies, better electrodes and changes in inclu-
ble (e.g., due to ossification after meningitis or sion criteria, the majority of the recipients
severe hypoplasia) or the cause of deafness is achieves open set speech understanding and is
found in the internal auditory canal (bilateral able to use the telephone, although this still
acoustic neuroma, aplasia of the auditory nerve). requires an intensive rehabilitation process (see
Fig. 1).
Detailed Description Components and Signal Processing of

Multichannel Implants
Background A cochlear implant consists of an external part
Due to their reduced oral communication skills, and an internal part, as shown in Fig. 2.
severe to profoundly deaf people are restricted in An otolaryngologist surgically implants the
their social functioning. Since the pioneering internal part (the so-called receiver-stimulator
work of Djourno and Eyries (1957) and House package with the electrode array) under general
in the 1970s of the last century (House and Urban anesthesia; the externally worn speech processor
1973), it has become possible to restore some of (body worn or behind the ear) is connected after
the hearing functions through direct electrical several weeks of wound healing. The speech pro-
stimulation of the auditory nerve. Currently cessor captures the incoming sound and, after
used cochlear implants utilize electrode arrays preprocessing (typically noise cancellation and
with 12–22 contacts on a Silastic carrier that are amplitude compression), encodes it into
most commonly inserted into the scala frequency-specific electrical information to be
tympani through either the round window mem- sent to the individual electrode contacts in the
brane or a drilled cochleostomy in its vicinity. cochlea.
Auditory Prosthesis 275 A
The speech coding strategies used in all cur- interaction between neighboring electrode con-
rent implants are extensions of the continuous tacts by stimulating all contacts in a sequential
interleaved sampling (CIS) strategy (Wilson mode rather than simultaneously. A digital filter A
et al. 1991). This strategy tries to avoid electrical bank is used to process the signal into separate
frequency bands. Next, the envelope of each
band, determined by rectification and low-pass
70.0 filtering of the signal, is used to set the amplitude
60.0 of a sequence of nonsimultaneous pulses on the
Words correct (%)
50.0 implanted electrode contacts. The rate of stimu-

40.0
lation is determined by the device brand and by
the patient’s performance, but typically ranges
30.0
between 400 pulses/s and 4,000 pulses/s per
20.0 channel.
10.0 Both the encoded stimulation pattern and the
0.0 energy are transmitted to the implanted receiver-
Pre-op 1 wk 2 wk 1 mo 3 mo 6 mo 1 yr stimulator package through an RF-link. The
Time external and internal coils for this RF-link are
kept aligned with magnets in the center of both
Auditory Prosthesis, Fig. 1 Performance over the first
year of cochlear implant use of 70 consecutive patients coils. The signal is picked up by the electronics in
implanted with a HiRes90K implant with a HiFocus 1 J the receiver-stimulator package, which in turn
electrode array (Advanced Bionics, Valencia, CA). The delivers electrical pulses to the auditory nerve
bars represent the percentage of correctly understood
fibers via electrode contacts in the electrode
Dutch monosyllabic (CVC) words, presented from CD
(65 dB SPL, free-field in quiet). The preoperative data array.
were obtained with the best-fitted hearing aid
Auditory Prosthesis,
Fig. 2 The components of
a cochlear implant with
a behind-the-ear speech
processor (Courtesy of
Advanced Bionics)
A 276 Auditory Prosthesis
Modern cochlear implants also have back auditory nerve neuron based on the classical
telemetry, allowing to record electrically evoked work on amphibian nerve fibers of
compound action potentials (eCAPs) of the audi- Frankenhæuser and Huxley (1964). In order to
tory nerve via the implanted electrode array. fine tune the model to represent physiological
data obtained from single auditory nerve fiber
Computational Modeling experiments in squirrel monkeys, they had to
To provide more insight in the fundamentals of adapt the modeled nerve fiber’s anatomy signifi-
functional electrical stimulation of the auditory cantly. Motz and Rattay (1986) used a single-
nerve, computational models have been devel- node model with the Hodgkin and Huxley
oped. This involves stimulating not only the model of unmyelinated squid giant axon mem-
response of a nerve fiber to an externally applied brane (▶ Hodgkin-Huxley Model) to investigate
potential field but also the calculation of this the time structure of the response of the
potential distribution from the currents on the (myelinated!) auditory nerve to electrical
stimulating electrodes. This is especially intricate stimuli. The gSEF model (Frijns et al. 1995) is
in the case of cochlear implants due to the com- a nonlinear cable model, which represents
plex geometry of the inner ear. essential mammalian nerve fiber properties,
including spike duration and conduction
Electrical Volume Conduction in the Cochlea velocity, refractory behavior, and repetitive fir-
An analytic solution of such a 3D volume con- ing, better than previous models and can deal
duction problem is restricted to geometries that with arbitrary stimulus wave forms. It is based
are much simpler than the cochlea, and many upon voltage clamp measurements in rat and cat
theoretical models on the (actually three- motor nerve fibers at mammalian body tempera-
dimensional) potential pattern set up in the ture performed by Schwarz and Eikhof (1987).
cochlea by the stimulating current sources The gSEF model and its variants have, in con-
assumed an exponential decay of current from junction with electrical volume conduction
its source to the nerve fibers along the cochlea, models, been used not only to predict which
modeled in one dimension (O’Leary et al. 1985), (intact or degenerated) fibers are excited by spe-
while other analytical approaches assume cific patterns of electrical stimulation (Smit
a simplified unrolled anatomy (Goldwyn et al. 2010; Frijns et al. 2009a, 2011) or to explain
et al. 2010). Suessermann and Spelman (1993) the results obtained with psychophysical experi-
used an electrical network as a practical repre- ments (Carlyon et al. 2010; Snel-Bongers
sentation of the electro-anatomy of the cochlea. et al. 2013) but also to calculate the eCAP pro-
Numerical methods nowadays, however, duced on the basis of predicted single fiber action
allow to incorporate much more detailed potentials (Briaire and Frijns 2005; Westen
(electro-)anatomical information (including the et al. 2011).
shape and position of the electrode array) and The abovementioned neural models have in
sometimes even allow for patient-specific model- common that they are deterministic in the way
ing on the basis of CT-scans (Carlyon they treat the neural membrane responses. If the
et al. 2010). The numerical methods that have focus of research is more on the effect of high
been used include the finite difference method stimulation rates or on repetitive near-threshold
(Whiten 2007), the finite element method stimulation, stochastic models come into play.
(Rattay et al. 2001; Hanekom 2001), and the Most models of this type are single-node thresh-
boundary element method, also known as the old models (Bruce et al. 1999), while cable
integral equation method (Frijns et al. 2001). models (Rubinstein et al. 1999; Imennov and
Rubinstein 2009), although computationally
Simulating the Auditory Nerve Fiber Responses very intensive and requiring supercomputers,
Colombo and Parkins (1987) were the first to can give insight in more complex stimulation
develop a cable model of the mammalian patterns.
Auditory Prosthesis 277 A
Auditory Prosthesis, Fig. 3 (a) The structure of a 3D a realistic representation of the electrode array in the
volume conduction model of the implanted human scala tympani. (b) The potential distribution in the neural
cochlea (as developed at the Leiden University Medical compartment due to monopolar stimulation
Center), including the auditory nerve (in yellow) and
Integrated Use of Volume Conduction and Neural position in scala tympani: a model approach. Hear
Models: State of the Art Res 214(1–2):17–27
Bruce IC, Irlicht LS, White MW, O’Leary SJ, Dynes S,
While in the early days of cochlear implantation Javel E, Clark GM (1999) A stochastic model of the
all insights in the mechanisms underlying their electrically stimulated auditory nerve: pulse-train
function had to come either from clinical practice response. IEEE Trans Biomed Eng 46(6):630–637
and associated psychophysics or from animal Carlyon RP, Macherey O, Frijns JHM, Axon PR, Kalkman
RK, Boyle P, Baguley DM, Briggs J, Deeks JM,
experiments, nowadays sophisticated computa- Briaire JJ, Barreau X, Dauman R (2010) Pitch com-
tional models exist, which integrate a model of parisons between electrical stimulation of a cochlear
electrical volume conduction in the cochlea with implant and acoustic stimuli presented to a normal-
active neural models. Such models can not only hearing contralateral ear. J Assoc Res Otolaryngol
11(4):625–640
be used to explain effects of current and future Colombo J, Parkins JW (1987) A model of electrical
electrode designs and stimulation schemes but excitation of the mammalian auditory-nerve neuron.
are also able to predict the of anatomical varia- Hear Res 31:287–312
tions (Frijns et al. 2009b), species Djourno A, Eyries C (1957) Auditory prosthesis by means
of a distant electrical stimulation of the sensory nerve
differences (Frijns et al. 2001), and the effects with the use of an indwelt coiling. Presse Med 65:1417
of neural degeneration (Briaire and Frijns 2006; Frankenhæuser B, Huxley AF (1964) The action potential
Snel-Bongers et al. 2013) (Fig. 3). in the myelinated nerve fiber of Xenopus laevis as
computed on the basis of voltage clamp data.
J Physiol (Lond) 171:302–315
Frijns JHM, de Snoo SL, Schoonhoven R (1995) Potential
Cross-References distributions and neural excitation patterns in
a rotationally symmetric model of the electrically
▶ Hodgkin-Huxley Model stimulated cochlea. Hear Res 87:170–186
Frijns JHM, Briaire JJ, Grote JJ (2001) The importance of
human cochlear anatomy for the results with modiolus
hugging multi-channel cochlear implants. Otol
Neurotol 22(3):340–349
References Frijns JHM, Kalkman RK, Vanpoucke FJ, Bongers JS,
Briaire JJ (2009a) Simultaneous and
Briaire JJ, Frijns JHM (2005) Unravelling the electrically non-simultaneous dual electrode stimulation in
evoked compound action potential. Hear Res cochlear implants: evidence for two neural response
205(1–2):143–156 modalities. Acta Otolaryngol 129(4):433–439
Briaire JJ, Frijns JHM (2006) The consequences of neural Frijns JH, Kalkman RK, Briaire JJ (2009b) Stimulation of
degeneration regarding optimal cochlear implant the facial nerve by intracochlear electrodes in
A 278 Auditory Recognition
otosclerosis: a computer modeling study. Otol Wilson BS, Finley CC, Lawson DT, Wolford RD,
Neurotol 30(8):1168–1174 Eddington DK, Rabinowitz WM (1991) Better speech
Frijns JHM, Dekker DMT, Briaire JJ (2011) Neural exci- recognition with cochlear implants. Nature
tation patterns induced by phased-array stimulation in 352(6332):236–238
the implanted human cochlea. Acta Otolaryngol
131(4):362–370
Further Reading
Goldwyn JH, Bierer SM, Bierer JA (2010) Modeling the
NIH factsheet cochlear implants: http://report.nih.gov/
electrode-neuron interface of cochlear implants:
nihfactsheets/ViewFactSheet.aspx?csid=83
effects of neural survival, electrode placement, and
the partial tripolar configuration. Hear Res
268(1–2):93–104
Hanekom T (2001) Three-dimensional spiraling finite ele-
ment model of the electrically stimulated cochlea. Ear
Hear 22(4):300–315 Auditory Recognition
House WF, Urban J (1973) Long term results of electrode
implantation and electronic stimulation of the cochlea
▶ Acoustic Timbre Recognition
in man. Ann Otol Rhinol Laryngol 82(4):504–517
Imennov NS, Rubinstein JT (2009) Stochastic population
model for electrical stimulation of the auditory nerve.
IEEE Trans Biomed Eng 56(10):2493–2501
Motz H, Rattay F (1986) A study of the application of the
Hodgkin-Huxley and the Frankenhaeuser-Huxley
Auditory Scene Analysis
model for electrostimulation of the acoustic nerve.
Neuroscience 18:699–712 ▶ Auditory Perceptual Organization
O’Leary SJ, Black RC, Clark GM (1985) Current distri-
butions in the cat cochlea: a modelling and electro-
physiological study. Hear Res 18:273–281
Rattay F, Leao RN, Felix H (2001) A model of the elec-
trically excited human cochlear neuron. II. Influence of Auditory Sensory Receptor Cell,
the three-dimensional cochlear structure on neural Model
excitability. Hear Res 153(1–2):64–79
Rubinstein JT, Wilson BS, Finley CC, Abbas PJ
(1999) Pseudospontaneous activity: stochastic inde- ▶ Cochlear Inner Hair Cell, Model
pendence of auditory nerve fibers with electrical stim-
ulation. Hear Res 127(1–2):108–118
Ruggero MA (2009) Cochlea. In: Binder MD, Hirokawa
N, Windhorst U (eds) Encyclopedia of Neuroscience.
Springer, Berlin Heidelberg, pp 765–769 Auditory Thalamocortical
Schwarz JR, Eikhof G (1987) Na currents and action Transformations
potentials in rat myelinated nerve fibres at 20 C and
37 C. Pflugers Arch 409:569–577
Kazuo Imaizumi and Charles C. Lee
Smit JE, Hanekom T, van Wieringen A, Wouters J,
Hanekom JJ (2010) Threshold predictions of different Department of Comparative Biomedical
pulse shapes using a human auditory nerve fibre model Sciences, School of Veterinary Medicine,
containing persistent sodium and slow potassium cur- Louisiana State University, Baton Rouge,
rents. Hear Res 269(1–2):12–22
LA, USA
Snel-Bongers J, Briaire JJ, Van Veen EH, Kalkman RK,
Frijns JHM (2013) Threshold levels of dual electrode
stimulation in cochlear implants. J Assoc Res
Otolaryngol 14:781–790 Definition
Suesserman MF, Spelman FA (1993) Lumped-parameter
model for in vivo cochlear stimulation. IEEE Trans
Biomed Eng 40:237–245 Auditory thalamocortical transformations arise
Westen AA, Dekker DMT, Briaire JJ, Frijns JHM from the ascending neural processing of sponta-
(2011) Stimulus level effects on neural excitation and neous activity as well as external sound-evoked
eCAP amplitude. Hear Res 280:166–176
activity from the auditory thalamus, the medial
Whiten D (2007) Electro-anatomical models of the
cochlear implant. PhD thesis, Massachusetts Institute geniculate body, to the thalamorecipient layers of
of Technology, Cambridge, MA the auditory cortex in mammals.
Auditory Thalamocortical Transformations 279 A
Detailed Description are organized lateromedially in most species
studied, with neurons in the lateral regions
Thalamic and Cortical Organization responding to lower frequencies of sound, while A
The Thalamus neurons in the medial regions respond to higher
The thalamus is the obligate neural station con- frequencies of sound (Imig and Morel 1985).
veying ascending sensory information to the cor- Other physiological properties, such as band-
tex (Sherman and Guillery 2006; Jones 2007). width tuning and aurality, are found interdigi-
Each sensory modality, except for olfaction, is tated among neurons across these isofrequency
represented in defined nuclei of the thalamus laminae (Ehret 1997). The main ascending pro-
through which sensory information must first be jection to the ventral division originates from the
processed before eventually being transmitted to central nucleus of the inferior colliculus (Calford
the respective sensory areas of the neocortex 1983; Wenstrup 2005). These projections pre-
(Sherman and Guillery 2006; Jones 2007). In serve the topographic organization connecting
the auditory system, the medial geniculate body similarly tuned regions in the inferior colliculus
is the main auditory thalamic nucleus (Winer to matched regions in the ventral division of the
1984; Jones 2007). Classically, the medial genic- medial geniculate body (Wenstrup 2005). The
ulate body has been divided into three main sub- main output of the ventral division is to the pri-
divisions, i.e., the ventral division, the dorsal mary auditory cortex. These projections termi-
division, and the medial division (Winer 1984; nate primarily in layer 4 of the primary auditory
Imig and Morel 1985). Each of these thalamic cortex but also have branched projections to layer
nuclei can be identified on the basis of their 6 (Huang and Winer 2000; Smith et al. 2012).
cytoarchitecture, physiological properties, and Similar to the projection from the inferior
connections (Huang and Winer 2000; de la colliculus, these projections are also topographi-
Mothe et al. 2006; Lee and Winer 2008a). The cally organized, such that the low-frequency
thalamocortical transformation is constrained by regions of the ventral division project to the
the neuroanatomical organization of projections lower-frequency regions of the primary auditory
from each of these thalamic nuclei to each of the cortex (Morel and Imig 1987; Morel et al. 1993).
several areas of the auditory cortex (Winer The ventral division also connects with other
et al. 2005; Winer and Lee 2007; Lee and Winer tonotopically organized auditory cortical areas,
2011b). which vary in number in different species, e.g.,
five in the cat (Reale and Imig 1980) and three in
The Ventral Division of the Medial Geniculate the monkey (Kaas and Hackett 2000). These
Body The ventral division of the medial genic- areas also receive topographically organized pro-
ulate body is the principal thalamic nucleus con- jections from the ventral division but mainly
veying ascending auditory information to the from nonoverlapping sectors (Morel and Imig
primary auditory cortex (Huang and Winer 1987). Very few neurons in the ventral division
2000; Smith et al. 2012). A subregion within the send branched projections to multiple auditory
ventral division, the pars ovoidalis, is located in areas, i.e., low-frequency neurons in the ventral
the medial part of the nucleus, bordering the division do not connect with multiple
dorsal division and medial division (Winer low-frequency regions in different tonotopically
1984; Jones 2007). Neurons in the ventral divi- organized areas (Lee et al. 2004, 2011).
sion are sharply tuned to sound frequencies (Imig
and Morel 1985). These neurons are arranged in The Dorsal Division of the Medial Geniculate
laminar rostrocaudal sheets, with their dendritic Body The dorsal division of the medial genicu-
fields aligned in parallel along the sheet. The late body is the part of the auditory thalamus that
neurons in each sheet respond to similarly tuned connects to non-tonotopically organized areas of
frequencies (Imig and Morel 1985). These sheets the auditory cortex (de la Mothe et al. 2006;
A 280 Auditory Thalamocortical Transformations
Lee and Winer 2008a). This nucleus is composed auditory areas, very few neurons within the dor-
of various subnuclei, including the dorsal super- sal division send branched projections to multiple
ficial, deep dorsal, dorsocaudal, and ventrolateral areas (Kishan et al. 2008; Lee and Winer 2008a;
nuclei (Winer et al. 2005; Lee and Winer 2008a). Lee et al. 2011).
Neurons in the dorsal division are broadly tuned
to frequencies, many with multi-peaked and com- The Medial Division of the Medial Geniculate
plex receptive fields (Morel and Imig 1987; Body The medial division of the medial genicu-
Winer et al. 2005). In contrast to the organization late body projects widely across all auditory cor-
of the ventral division of the medial geniculate tical areas, with a pattern of axonal termination
body, neurons in the dorsal division do not that contrasts with the projections from the ven-
exhibit an oriented laminar pattern of organiza- tral division and the dorsal division of the medial
tion (Winer 1984). The dendritic arborizations of geniculate body (Huang and Winer 2000; Jones
neurons in the dorsal division are more 2007). Neurons in the medial division exhibit
isotropically organized. The subdivision of this highly complex receptive fields, which often
nucleus is based primarily on cytoarchitectonic extend beyond purely auditory responses, with
densities and connections with several many neurons responding to stimuli from several
non-tonotopic auditory areas (Lee 2013). Projec- modalities, i.e., visual and somatosensory (Bordi
tions to the dorsal division originate primarily and LeDoux 1994). Neurons in the medial divi-
from the non-tonotopic area of the inferior sion differ from those in the ventral division and
colliculus, i.e., the dorsal cortex (Wenstrup dorsal division, in that they display a wide range
2005). The dorsal division nuclei project broadly of sizes and dendritic arborization patterns
to the non-tonotopic areas of the auditory cortex, (Bartlett and Smith 1999; Smith et al. 2007).
which also include multimodal and limbic- The medial division contains the magnocellular
related areas, whose relation to auditory neurons, which are the largest neurons in the
processing in part is derived from the direct medial geniculate body (Winer 1984). Like the
inputs received from thalamocortical sources dorsal division, the neurons in the medial division
(Lee and Winer 2008a, 2011a). Although do not appear to be organized isotropically along
a functional metric, such as tonotopy, appears to any particular anatomical domain and are loosely
be absent in the nuclei of the dorsal division, their packed compared with the other divisions (Winer
projections to the non-tonotopic areas of the audi- 1984). The medial division receives its primary
tory cortex are highly topographic, similar in input from both the dorsal cortex and external
extent to the topography of projections from the cortex regions of the inferior colliculus, which
ventral division to the tonotopic regions of the also contain non-tonotopic and multimodal
auditory cortex (Lee and Winer 2005; Schreiner responsive neurons (Wenstrup 2005). Every
and Winer 2007). The termination pattern of the area of the auditory cortex receives a projection
dorsal division projections to the non-tonotopic from the medial division (Lee and Winer 2008a),
cortical areas, in particular, the secondary audi- but unlike the ventral division and the dorsal
tory cortical region, is similar to that of the ven- division, the laminar terminations of axons in
tral division projection to the primary auditory these areas are primarily concentrated in layer
cortex, i.e., terminations primarily in layers 4 and 1 (Huang and Winer 2000). While axonal diver-
6 (Huang and Winer 2000). However, the synap- gence from the medial geniculate body is low in
tic terminals of the dorsal division projection to general, the highest proportion of neurons
the secondary auditory cortex are slightly larger projecting to multiple auditory cortical areas is
on average than the projections from the ventral found in the medial division; however, these
division to the primary auditory cortex (Smith comprise on average less than 2 % of neurons in
et al. 2012). Again, although the dorsal division the medial division of the medial geniculate body
projects broadly to several non-tonotopic (Kishan et al. 2008, 2011; Lee et al. 2011).
Inhibitory Circuits in the Thalamus The tha- and containing all of the neuronal cell bodies
lamic reticular nucleus, although not a specific and the latter residing beneath the gray matter
constituent of the medial geniculate body, is inti- and composed of the axonal fiber tracts of affer- A
mately intertwined with the operations of both ent and efferent projections (Nieuwenhuys 2013).
the thalamus and cortex and thus is an essential Although regional variations exist, the gray
structural component of the thalamocortical matter of the cerebral cortex has a laminar orga-
transformation (Winer and Larue 1996; Crabtree nization divided into cytoarchitectonically distin-
et al. 1998; Pinault 2004; Sherman and Guillery guishable layers, such that neuronal cell bodies
2006). The thalamic reticular nucleus is com- are situated in structural and functional groups
posed of inhibitory GABAergic neurons that relative to their location along the pial to white
form a shell surrounding the thalamus, roughly matter axis (Mountcastle 1997). These layers of
located along the lateral border of the thalamus neuronal cell bodies have specific afferent and
and extending rostrocaudally along nearly its efferent connections with other cortical regions
entire length (Pinault 2004; Lam and Sherman and with subcortical structures, in particular the
2005, 2007, 2010). In general, each nucleus of thalamus (Sherman and Guillery 2006). In total,
the thalamus innervates a specific sector of the there are six classically defined layers of the
thalamic reticular nucleus and receives reciprocal cortex (Mountcastle 1997). Of these, layer 4 of
topographic inhibitory feedback projections from the cerebral cortex is the main recipient layer for
that region of the thalamic reticular nucleus (Lam sensory information ascending from the
and Sherman 2005). In addition, feedback pro- primary sensory thalamic nuclei (Sherman and
jections from layer 6 of the neocortex en route to Guillery 2006). In addition, layer 6 receives
the thalamus branch to innervate the thalamic branched projection from these primary sensory
reticular nucleus (Lam and Sherman 2010), thalamic nuclei (Huang and Winer 2000; Smith
which establishes an extended thalamocorti- et al. 2012; Lee and Imaizumi 2013), and layer 1
cothalamic inhibitory feedback loop (Pinault is the recipient of thalamocortical inputs
2004). The thalamic reticular nucleus and the from nonspecific thalamic nuclei, e.g., the medial
inferior colliculus are the main sources of inhibi- division of the medial geniculate body (Huang and
tion in the medial geniculate body of rodents, Winer 2000; Jones 2007). In addition, layer 6 is the
which contain very few local inhibitory neurons source of neurons that send feedback projections
(Winer and Larue 1996). In humans, local inhib- to the thalamic nucleus that provides its main
itory neurons in the medial geniculate body com- thalamocortical input in layers 4 and 6, which
prise ~20 % of the total number of neurons, which establishes a thalamocorticothalamic feedback
is accompanied by a proportional reduction in loop (Sherman and Guillery 2006). Cortical layer
inhibitory projections from the thalamic reticular 5 is the source of feedforward nonreciprocal
nucleus, compared with rodents and other spe- thalamocortical projections, which serve as a con-
cies, which have fewer than 1 % of local inhibi- duit for communication between cortical areas via
tory neurons in the medial geniculate body a corticothalamocortical route (Sherman and
(Winer and Larue 1996). Guillery 2006).
The surface of the cerebral cortex is regionally
The Cerebral Cortex specified into distinct functional areas that are
The cerebral cortex is phylogenetically the involved in processing sensory and motor infor-
newest structure in the mammalian brain, respon- mation (Kaas 2008). The boundaries of these
sible for the higher-order processing of sensory, cortical areas, including those involved with
motor, and limbic information (Kaas 2008). audition, are broadly defined based on their
Broadly, the cerebral cortex is composed of cytoarchitectural organization, connections with
regions of “gray” matter and “white” matter, the other structures, and physiological responses of
former residing near the outer cortical surface constituent neurons, although precise borders and
definitions for many cortical areas in different resemble those to the primary auditory cortex
organisms still remain elusive (Kaas and (Huang and Winer 2000; Smith et al. 2012),
Hackett 2000; Hackett 2011). Although all mam- which could serve as an anatomical basis for
mals have cortical regions devoted to the common thalamocortical transformations across
processing of auditory information, the number the expanse of auditory cortical areas (Lee and
of auditory cortical areas varies widely among Sherman 2008, 2011).
different species (Lee and Winer 2008b, 2011b).
Despite this heterogeneity, all mammals studied Receptive Fields and the Thalamocortical
thus far have an identifiable auditory cortical Transformation
region that receives direct input from the ventral Spectral Receptive Field
division of the medial geniculate body (Lee and An important physiological parameter in the
Winer 2011b). This primary auditory cortical auditory thalamocortical transformation is the
area is also defined on the basis of a clearly iden- spectral receptive field. The spectral receptive
tifiable tonotopic map, which reflects the fre- field is, in general, measured based on a -
quency segregation established in the cochlea frequency-threshold tuning curve (response to
and propagated along the auditory pathway sound level as a function of sound frequency).
(Ehret 1997). A common measure is the Q factor by which
Although the primary auditory cortical area is characteristic frequency is divided by a linear
the most conserved across species, constellations measure of bandwidth at a given sound level
of other cortical regions devoted to the processing above threshold (e.g., Q10; Q value at 10 dB
of sound exist among the cortices of different above threshold) (Imaizumi et al. 2004). Because
mammalian species (Kaas 2008). However, the Q value is a normalized measure, the larger
their physiological and anatomical organization the Q value, the more sharply tuned are the neu-
is generally less well understood relative to that rons. Another measure is the square root
of the primary auditory cortex. Nevertheless, transformation:
these other areas can be broadly grouped into pffiffiffiffiffiffi pffiffiffiffiffiffi
distinct categories: tonotopic, non-tonotopic, F2 F1
multimodal, and limbic related (Lee and Winer
2011a). The tonotopic regions, like the primary where F2 and F1 are the highest and lowest edges
auditory cortex, contain identifiable maps of fre- of bandwidth at a given sound level (Rouiller
quency across their surface, generally with fre- et al. 1981; Calford 1983). Unlike the Q value,
quency reversals at their borders, and receive the smaller the square root transformation value,
strong projections from the ventral division of the more sharply tuned are the neurons.
the medial geniculate body (Reale and Imig The auditory thalamocortical transformation
1980; Hackett et al. 1998; Hackett et al. 2011). is mediated by the excitatory neurotransmitter,
The non-tonotopic areas contain disordered rep- glutamate, from the medial geniculate body to
resentation of frequency and generally receive the auditory cortex. Therefore, only the excit-
more prominent inputs from the dorsal division atory receptive field is transformed in
of the medial geniculate body (Schreiner and thalamocortical projections. A problem describ-
Cynader 1984; Smith et al. 2012). Multimodal ing spectral receptive fields in the
and limbic areas reside at the limits of the classi- thalamocortical transformation is the availability
cal auditory areas and receive inputs from multi- of data utilizing the same recording method (e.g.,
modal and limbic thalamic nuclei and areas and local field potentials, single- or multiunit extra-
have complex responses reflective of these con- cellular recording, whole-cell recording, etc.)
vergent inputs (Bowman and Olson 1988; Clarey under similar recording conditions (e.g., anesthe-
and Irvine 1990; Clascá et al. 1997; de la Mothe sia: isoflurane, ketamine, or pentobarbital; depth
et al. 2006). As discussed above, features of the of anesthesia; awake states; restricted or freely
thalamocortical inputs to these other areas moving) in the same animal species. This creates
some controversy and readers should use caution as sharp as was thought (Liu et al. 2007). These
for the following section. examples above are based on studies in rodents.
Functional organization of the spectral Spectral receptive fields, however, vary widely A
receptive field has been well appreciated in the across species. In the human auditory cortex,
auditory cortices of several species. The most neurons are extremely sharply tuned (Bitterman
well-known example is the primary auditory cor- et al. 2008), which suggests that spectral recep-
tex of the mustached bat. A large area, called the tive fields may become sharper through the
DSCF (Doppler-shifted constant frequency) area, thalamocortical transformation. How does this
of the primary auditory cortex is devoted to opposite trend occur in thalamocortical transfor-
a particular frequency range (60.6–62.3 kHz) mation? Clear evidence is available in the awake
for their prey-hunting behavior (Suga 1994). mustached bat. Based on Q10, Q30, and Q50
Neurons in the DSCF area are extremely sharply values, DSCF neurons in the primary auditory
tuned to characteristic frequency (Q50 values cortex are more sharply tuned than those in the
range from ~10 to 500 or higher) (Suga and medial geniculate body (Suga et al. 1997). Simi-
Manabe 1982; Suga et al. 1997). Neurons in the lar evidence is also available in the awake mar-
anterior and posterior parts of the DSCF area are moset (Bartlett et al. 2011). Thus, the sharpening
more broadly tuned. Similarly organized clusters of spectral receptive fields through the
of sharply or broadly tuned neurons are also thalamocortical transformation may relate to
found in the auditory cortex of carnivores and behavioral and/or ethological functions.
primates (Recanzone et al. 1999; Cheung et al. For thalamocortical transformation of spectral
2001; Read et al. 2001; Imaizumi et al. 2004; receptive fields, limited experimental cases are
Philibert et al. 2005; Imaizumi and Schreiner available using in vivo single-unit recordings in
2007). In particular, the cat primary auditory cor- awake guinea pigs (Creutzfeldt et al. 1980) and
tex has an interesting functional organization of anesthetized cats (Miller et al. 2001, 2002). For
spectral receptive fields. Sharply or broadly tuned secure functional transformation, medial genicu-
neurons are clustered alternatively along the dor- late body and auditory cortex neurons require an
soventral axis only in the mid-frequency range alignment of less than 1/3 octave difference in
(5–20 kHz) (Imaizumi and Schreiner 2007). best frequency (sound frequency evoked the best
Unlike those animal species, functional organiza- response in a neuron at a given sound level)
tion of spectral receptive fields in the rodent audi- (Miller et al. 2001). To fully activate a neuron
tory cortex is not clear (Polley et al. 2007). in the auditory cortex, synaptic convergence from
Whereas rich information of functional organiza- 20 to 25 neurons in the medial geniculate body is
tion of spectral receptive fields is available in the required (Miller et al. 2001). Using a ripple noise
auditory cortex, only limited information is avail- stimulus and computational analytical approach
able in the medial geniculate body. Neurons in the of reverse correlation technique, Miller
ventral division are, in general, more sharply tuned et al. (2002) estimated spectral modulation rates
than in the dorsal division (Rouiller et al. 1981; through thalamocortical transformation. Both
Calford 1983; Edeline et al. 1999). However, no thalamic and cortical neurons show lower spec-
clear spatial organization is available due to the tral modulation rates. Whereas cortical neurons
deep location in the brain. have significantly lower spectral modulation
In general, it is believed that spectral receptive rates than thalamic neurons based on the best
fields become broader through the spectral modulation rates, the overall spectral
thalamocortical transformation (Kaur et al. filter properties between thalamic and cortical
2004, 2005). However, more recent studies neurons are similar.
using in vivo whole-cell recordings combined Overall, spectral receptive (modulation) fields
with pharmacological manipulation (blocking in the thalamocortical transformation are not sim-
cortical input by muscimol and SCH50911) ple. Depending on animals and the behavioral
have revealed that thalamocortical input is not significance of sound frequency, spectral
receptive fields become broader or sharper where VSpp is the phase-projected vector strength
through the thalamocortical transformation. per trial, VSt is the vector strength per trial, and ft
and fc are the trial-by-trial and mean phase angle
Temporal Receptive Field in radians, calculated for each stimulus condition,
An important function of the central auditory and
system is to decode species-specific communica-
Xn
tions and human speech sounds. Periodic modu- sin yi
lations are ubiquitous temporal features of f ¼ arctan2 Xni¼1
cos yi
species-specific communications and human i¼1
speech sounds (Rosen 1992; Joris et al. 2004).
Measuring repetition-rate transfer functions or where n is the number of spikes per trial (for ft)
modulation transfer functions captures response or across all trials (for fc) and arctan2 is
characterization to assess information for a modified version of the arctangent that deter-
a temporal range that corresponds to periodicities mines the correct quadrant of the output based on
in communication sounds (Eggermont 2001; the signs of the sine and cosine inputs. Whereas
Joris et al. 2004). Two commonly used measures vector strength value ranges from 1 (all spikes
to characterize repetition-rate transfer functions occur at the same stimulus phase) to 0 (spikes
are firing rate and vector strength (VS). Firing occur randomly), phase-projected vector strength
rate estimates overall response magnitude in value ranges from 1 (all spikes in phase with
a particular time window. Vector strength esti- population mean phase) to 1 (all spikes 180
mates spike-timing precision to a particular phase out of phase with population mean phase) with
of repetition or modulation stimulus: 0 corresponding to randomly occurring spikes
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (Yin et al. 2011; Niwa et al. 2012). Another way
ðS cos yÞ2 þ ðS sin yÞ2 to overcome a shortcoming of vector strength is
VS ¼ the product of two measures (vector strength and
n
firing rate), i.e., phase-locked rate or synchro-
t nized rate (Eggermont 1998b; Joris et al. 2004;
y ¼ 2p
T Imaizumi et al. 2011). The synchronized rate
measure incorporates both timing and response
where n is the total number of spikes, t is time of strength measures.
spike occurrence, and T is the interstimulus inter- Two coding schemes for temporal receptive
val (Goldberg and Brown 1969). Significance of fields have been proposed: precise spike timing
synchronization to the stimulus is examined by estimated by vector strength codes slow repeti-
a Rayleigh test, >13.8 (p < 0.001) (Mardia tion rates (or modulation frequencies), while fir-
1972). Because vector strength does not incorpo- ing rate codes faster repetition rates
rate response strength, it may give rise to high (De Ribaupierre et al. 1972; Bieser and Muller-
vector strength values when the response strength Preuss 1996; Schulze and Langner 1997; Lu and
is low. To overcome a shortcoming of the VS Wang 2000; Lu et al. 2001; Joris et al. 2004).
measure, phase-projected vector strength (VSpp) However, more recent studies have proposed dif-
is used (Yin et al. 2011; Niwa et al. 2012). VSpp ferent coding schemes, as will be described later.
compares the mean phase angle for each trial with A problem describing temporal receptive
the mean phase angle of all trials and penalizes fields in the thalamocortical transformation is
single-trial vector strength values if they are not also the availability of data utilizing the same
in phase with the global response. VSpp is com- stimulus (e.g., nature of stimulus: click or white
puted on a trial-by-trial basis as follows: noise; modulation carrier: pure tone or noise;
VSpp ¼ VSt cos ðft fc Þ modulation depth; and modulation shape:
rectangular or sinusoidal), stimulus presentation and awake states (Ter-Mikaelian et al. 2007).
method (free field or sealed earphones), and Clear evidence of an effect of anesthesia is avail-
recording method (single- or multiunit extracel- able in the rat primary auditory cortex (Rennaker A
lular recording) under similar recording condi- et al. 2007). Ketamine anesthesia significantly
tions (e.g., anesthesia: isoflurane, ketamine, or decreased cutoff repetition rates (higher border
pentobarbital; depth of anesthesia; awake states: of repetition-rate transfer function) to <20 Hz
restricted, freely moving, or engaged behavior) in from 80 to 90 Hz in the awake state. A more
the same animal species. In many cases, the ana- recent study using awake state (head-fixed) rats
lytical criteria are also different. These create examined repetition-rate transfer function using
difficulty for direct comparisons and some dis- click trains in the primary auditory cortex and the
crepancy. Thus, the readers should treat the fol- anterior auditory field (Ma et al. 2013). Both core
lowing section with caution. fields show high best repetition rates (up to
The vast majority of studies regarding tempo- 32–64 Hz) and cutoff repetition rates (up to
ral receptive fields have focused on the primary 256 Hz), although neurons in the anterior audi-
auditory cortex. However, functional magnetic tory field prefer significantly higher repetition
resonance imaging or positron emission tomog- rates than those in the primary auditory cortex.
raphy in humans and macaques suggested that the These best and cutoff repetition rates in the
superior temporal plane is specific to human awake rat auditory cortex are, in general, higher
speech or macaque species-specific calls over than those in the anesthetized rat (Kilgard and
nonspecific calls or other sounds (Belin Merzenich 1999; Chang et al. 2005). Attention
et al. 2000; Poremba et al. 2004; Petkov or engaged behaviors also alter temporal recep-
et al. 2008). These fields are located anterior to tive fields. By presenting at 15 Hz repetition rate
the primary core fields and may correspond to the (with various carrier frequencies) paired with
rostral field in primates. Reversible lesion exper- electrical stimulation of the nucleus basalis for
iments in cats also show that the anterior auditory 20–25 days, neurons in the rat primary auditory
field is specific to temporal pattern discrimination cortex are capable of following higher repetition
(Lomber and Malhotra 2008). Thus, the rostral rates by rate coding than those in control (Kilgard
field in primates and the anterior auditory field in and Merzenich 1998). Training in a sound maze
carnivores (and rodents) may be more important in which rats use sound source location for food
for temporal pattern processing. rewards based on auditory cues (noise repetition
rates increased with decreasing distance between
Rodent Auditory Cortex and the rat and target location) also enhanced tempo-
Thalamus Under anesthesia of pentobarbital, ral receptive fields (Bao et al. 2004). Compared to
neurons in the rat primary auditory cortex, in studies in the rodent auditory cortex, a small
general, show a low-pass filter property of number of studies are available in rodent thala-
repetition-rate transfer functions by rate coding mus. In awake guinea pigs, thalamic neurons
(firing rate) or band-pass filter property by tem- show more robust responses to higher modulation
poral coding (vector strength) (Kilgard and frequencies than cortical neurons (Creutzfeldt
Merzenich 1999; Chang et al. 2005). et al. 1980).
Ter-Mikaelian et al. (2007) examined the effects
of anesthesia (a combination of pentobarbital and Cat Auditory Cortex and Thalamus The cat
ketamine) and awake (head-fixed) states in the auditory cortex has been a focus of studies of
gerbil primary auditory cortex. While anesthesia temporal pattern processing. In particular, the
certainly affects modulation transfer functions in primary auditory cortex and the anterior auditory
individual neurons, overall, the population field are readily identifiable based on their loca-
response seems to be similar between anesthesia tion relative to sulcal patterns (Knight 1977;
Imaizumi et al. 2004; Lee et al. 2004) and are in the cat thalamus. In nitrous oxide-anesthetized
often compared in temporal pattern processing cats, approximately half of the thalamic units
(Schreiner and Urbas 1988; Eggermont 2000; showed precisely time-locked responses to click
Joris et al. 2004). Under anesthesia, neurons in trains (Rouiller et al. 1981). Half of these units
the anterior auditory field prefer higher modula- had cutoff repetition rates (higher border of
tion frequencies (and repetition rates) than those repetition-rate transfer function) up to 100 Hz.
in the primary auditory cortex (Schreiner and Local field potentials in the auditory cortex
Urbas 1988; Eggermont 1998b). Anesthesia also record subthreshold responses potentially from
reduces temporal receptive fields by half in neu- thalamocortical fibers (and corticocortical fibers),
rons of the cat primary auditory cortex (Goldstein thus indicating thalamic responses. Both best
et al. 1959). In the cat anterior auditory field modulation frequencies and cutoff modulation
under ketamine anesthesia, Imaizumi frequencies are generally higher in local field
et al. (2010) proposed different coding schemes potentials than unit recordings in the primary
using a combination of an in vivo high-resolution auditory cortex and anterior auditory field
cortical mapping technique with information the- (Eggermont 1998b). Under ketamine anesthesia
ory. Because this study was made using multiunit in the primary auditory cortex and the ventral
recordings, they computed discriminability of six division of the medial geniculate body of cats,
different low repetition rates (1–30 Hz) Miller et al. (2002) conducted in vivo simulta-
(Imaizumi et al. 2010). Unlike the traditional neous single-unit recordings from thalamus and
coding scheme (precise spike timing codes low cortical neurons. They found that temporal mod-
repetition rates), inter-spike intervals can carry ulation transfer functions are significantly deteri-
much more information than timing and rate orated through thalamocortical projections.
codes: some multiunits carried an information
value >2 bits that is close to a maximum of Primate Auditory Cortex and Thalamus A
~2.58 bits (=log2(6)). Furthermore, spatial distri- majority of studies of temporal receptive fields
bution of normalized firing rate to six different in primates are conducted in the awake state. In
repetition rates differs across the stimuli, which the awake squirrel monkey, neurons in the pri-
provides a potential coding scheme in the view of mary auditory cortex and the rostral field show
an ideal observer. These results suggest concur- both temporal and rate codes to amplitude mod-
rent coding schemes of temporal pattern ulation frequencies up to 64 or 128 Hz (Bieser
processing by inter-spike intervals, firing rate, and Muller-Preuss 1996). Neurons in the primary
and a map (Imaizumi et al. 2010). Using behav- auditory cortex showed higher best modulation
iorally trained cats, Dong et al. (2011) compared frequencies than those in the rostral field. In the
neurometric (neural responses to six repetition awake macaque, neurons in the primary auditory
rates from 12.5 to 200 Hz by in vivo single-unit cortex have higher best modulation frequencies
recordings) with psychometric (Go/No-Go by both temporal (means are 13 and 4.8 Hz) and
behavioral responses to the same six repetition rate (means are 45 and 19 Hz) codes and higher
rates) functions. Their recordings were focused vector strength than those in the rostral field
on the relatively low-frequency locations (Malone et al. 2010; Scott et al. 2011). These
<16 kHz of the primary auditory cortex. Preva- examples from the awake squirrel monkey and
lence is found of more synchronized units in the macaque have suggested the opposite trend of
behaviorally engaged primary auditory cortex temporal receptive fields in the core auditory
than in an anesthetized one (Lu and Wang 2000; fields to rodents and cats (Schreiner and Urbas
Dong et al. 2011). However, rate coding by 1988; Eggermont 1998b; Joris et al. 2004) despite
non-synchronized units correlates well with psy- the similar cortical locations (relative to the pri-
chometric functions. mary auditory cortex) of the anterior auditory
Compared to the studies in the cat auditory field in rodents and cats and the rostral field in
cortex, a small number of studies are available the primates. In the awake primate auditory
cortex, neurons may carry both temporal and by both temporal and rate coding than cortical
rate codes for lower and higher repetition neurons in the awake marmoset (Bartlett and
rates (or modulation frequencies) (Bieser and Wang 2007). A
Muller-Preuss 1996; Lu et al. 2001; Liang Overall, temporal receptive fields change
et al. 2002; Malone et al. 2007; Yin et al. 2011). through the thalamocortical transformation. In
However, a proportion of synchronized (using general, thalamic neurons are more capable of
vector strength or similar measures) and following higher repetition rates (modulation fre-
non-synchronized (using firing rate) neurons are quencies) than cortical neurons. However, neu-
different among the studies. These discrepancies rons in the auditory cortex may employ different
may be caused by the stimulus, the range of coding schemes to follow different ranges of rep-
repetition rates, and/or analytical criterion (Yin etition rates (this is not restricted only to high
et al. 2011). There is an interesting proposal of repetition rates but also low to mid repetition
low to mid range of repetition rates (10–45 Hz) rates) either through the thalamocortical transfor-
corresponding to flutter perception by two differ- mation or at the cortical level.
ent populations of neurons in the awake marmo-
set auditory cortex. One population of neurons Latency
increases firing rate with increasing repetition First-spike latency (hereafter latency) is another
rates, while the other population decreases firing important physiological parameter (Eggermont
rate with increasing repetition rates (Bendor and 2001). However, because only the experimenter
Wang 2007). All examples reviewed above are knows the stimulus onset (the brain and neurons
based on studies of primates passively listening to do not know it), relative latencies might be a good
stimuli in awake state. However, active engage- candidate for neural coding of temporal patterns
ment of behaviors (discriminating modulated (Eggermont 1998b; Schreiner and Raggio 1996;
sounds, 2.5–500 Hz, from unmodulated sounds) Lu and Wang 2000; Liang et al. 2002;
improves both temporal and rate codes in single Ter-Mikaelian et al. 2007; Imaizumi
neurons of the macaque primary auditory cortex et al. 2011), vocalizations (Wang et al. 1995;
(Niwa et al. 2012). Nagarajan et al. 2002), sound localization
Compared to the studies in primate auditory (Eggermont 1998a; Furukawa et al. 2000), and
cortex, a small number of studies are available in auditory scene (Dear et al. 1993). Latency, in
primate thalamus. In the awake squirrel monkey, general, decreases with increasing sound level
neurons in the thalamus show a similar tendency (Heil 1997). However, neurons in the primary
of temporal and rate coding to best modulation auditory cortex of the little brown bat show
frequencies up to 128 Hz (Preuss and Muller- shorter latencies to lower sound level than higher
Preuss 1990). In the awake marmoset, neurons sound level (Sullivan 1982b). Furthermore, when
in the ventral and the anterodorsal divisions of the the two sounds (higher and lower sound levels)
medial geniculate body show a mixture of tem- are presented by a behaviorally relevant delay
poral and rate coding, while neurons in the between pulse and echo, latencies are facilitated
posterodorsal division show a dominant tendency and become shorter for the echolocating behavior
of rate coding (Bartlett and Wang 2011). These (Sullivan 1982a). In general, neurons in the ante-
examples suggest that separation of temporal and rior auditory field have shorter latencies than
rate coding for low and high repetition rates may those in the primary auditory cortex across
be created within the auditory cortex (Bartlett and many different species (Schreiner and Urbas
Wang 2007). However, other data suggest that 1986, 1988; Linden et al. 2003; Rutkowski
separation from temporal to rate coding for low to et al. 2003; Imaizumi et al. 2004; Bizley
high repetition rates may be completed through et al. 2005). However, neurons in the primate
thalamocortical transformation (Malone primary auditory cortex show shorter latencies
et al. 2007; Yin et al. 2011). Thalamic neurons than those in the rostral field (Scott et al. 2011),
are capable of following higher repetition rates which is related to the fact that neurons in the
primate primary auditory cortex follow higher from the dorsal cochlear nucleus to the auditory thal-
repetition rates than those in the rostral field amus in the guinea pig. Eur J Neurosci 24:491–498
Bao S, Chang EF, Woods J, Merzenich MM (2004) Tem-
(Bieser and Muller-Preuss 1996; Malone poral plasticity in the primary auditory cortex induced
et al. 2010). by operant perceptual learning. Nat Neurosci
Neurons in the ventral division of the medial 7:974–981
geniculate body show shorter latencies than those Bartlett EL, Smith PH (1999) Anatomic, intrinsic, and
synaptic properties of dorsal and ventral division neu-
in other divisions (Rouiller et al. 1981; Calford rons in rat medial geniculate body. J Neurophysiol
1983; Edeline et al. 1999). However, more recent 81:1999–2013
studies in the mouse and guinea pig thalamus Bartlett EL, Wang X (2007) Neural representations of
show that neurons in the medial division have temporally modulated signals in the auditory thalamus
of awake primates. J Neurophysiol 97:1005–1017
shorter latencies than those in the ventral division Bartlett EL, Wang X (2011) Correlation of neural
of the medial geniculate body (Anderson response properties with auditory thalamus subdivi-
et al. 2006; Anderson and Linden 2011). This sions in the awake marmoset. J Neurophysiol
evidence strongly supports the shorter latencies 105:2647–2667
Bartlett EL, Sadagopan S, Wang X (2011) Fine frequency
in neurons of the anterior auditory field than those tuning in monkey auditory cortex and thalamus.
in the primary auditory cortex (Schreiner and J Neurophysiol 106:849–859
Urbas 1986, 1988; Linden et al. 2003; Rutkowski Belin P, Zatorre RJ, Lafaille P, Ahad P, Pike B (2000)
et al. 2003; Imaizumi et al. 2004; Bizley Voice-selective areas in human auditory cortex.
Nature 403:309–312
et al. 2005) because the anterior auditory field Bendor D, Wang X (2007) Differential neural coding of
receives input not only from the ventral division acoustic flutter within primate auditory cortex. Nat
but also from the medial division. Latency Neurosci 10:763–771
through thalamocortical transformation can be Bieser A, Muller-Preuss P (1996) Auditory responsive
cortex in the squirrel monkey: neural responses to
estimated by in vivo simultaneous single-unit amplitude-modulated sounds. Exp Brain Res
recordings and cross-correlation analysis. 108:273–284
A maximum peak in the correlogram is shifted Bitterman Y, Mukamel R, Malach R, Fried I, Nelken
to the expected travel and synaptic delay (e.g., I (2008) Ultra-fine frequency tuning revealed in single
neurons of human auditory cortex. Nature
a few milliseconds) (Miller et al. 2001). Thus, it is 451:197–201
generally believed that latency in thalamocortical Bizley JK, Nodal FR, Nelken I, King AJ (2005) Functional
transformation is inherited from the thalamus to organization of ferret auditory cortex. Cereb Cortex
cortex. However, a recent study using in vivo 15:1637–1653
Bordi F, LeDoux JE (1994) Response properties of single
whole-cell recordings combined with units in areas of rat auditory thalamus that project to
a pharmacological application (to silence cortical the amygdala. II. Cells receiving convergent auditory
activity by a mixture of muscimol and and somatosensory inputs and cells antidromically
SCH50911) in the rat primary auditory cortex activated by amygdala stimulation. Exp Brain Res
98:275–286
unfolds a different story: difference in latency Bowman EM, Olson CR (1988) Visual and auditory asso-
between the thalamus and cortex is generated by ciation areas of the cat’s posterior ectosylvian gyrus:
synaptic integration time by excitation and inhi- cortical afferents. J Comp Neurol 272:30–42
bition through corticocortical interactions (Zhou Calford MB (1983) The parcellation of the medial genic-
ulate body of the cat defined by the auditory response
et al. 2012). properties of single units. J Neurosci 3:2350–2364
Chang EF, Bao S, Imaizumi K, Schreiner CE, Merzenich
MM (2005) Development of spectral and temporal
response selectivity in the auditory cortex. Proc Natl
References Acad Sci U S A 102:16460–16465
Cheung SW, Bedenbaugh PH, Nagarajan SS, Schreiner
Anderson LA, Linden JF (2011) Physiological differences CE (2001) Functional organization of squirrel monkey
between histologically defined subdivisions in the primary auditory cortex: responses to pure tones.
mouse auditory thalamus. Hear Res 274:48–60 J Neurophysiol 85:1732–1749
Anderson LA, Malmierca MS, Wallace MN, Palmer AR Clarey JC, Irvine DRF (1990) The anterior ectosylvian
(2006) Evidence for a direct, short latency projection sulcal auditory field in the cat: I. An
electrophysiological study of its relation to surround- Hackett TA, Stepniewska I, Kaas JH (1998) Subdivisions
ing auditory cortical fields. J Comp Neurol of auditory cortex and ipsilateral cortical connections
301:289–303 of the parabelt auditory cortex in macaque monkeys.
Clascá F, Llamas A, Reinoso-Suárez F (1997) Insular J Comp Neurol 394:475–495 A
cortex and neighboring fields in the cat: a redefinition Hackett TA, Barkat TR, O’Brien BM, Hensch TK, Polley
based on cortical microarchitecture and connections DB (2011) Linking topography to tonotopy in the
with the thalamus. J Comp Neurol 384:456–482 mouse auditory thalamocortical circuit. J Neurosci
Crabtree JW, Collingridge GL, Isaac JT (1998) A new 31:2983–2995
intrathalamic pathway linking modality-related nuclei Heil P (1997) Auditory cortical onset responses
in the dorsal thalamus. Nat Neurosci 1:389–394 revisited. I. First-spike timing. J Neurophysiol
Creutzfeldt O, Hellweg FC, Schreiner C (1980) 77:2616–2641
Thalamocortical transformation of responses to com- Huang CL, Winer JA (2000) Auditory thalamocortical
plex auditory stimuli. Exp Brain Res 39:87–104 projections in the cat: laminar and areal patterns of
de la Mothe LA, Blumell S, Kajikawa Y, Hackett TA input. J Comp Neurol 427:302–331
(2006) Thalamic connections of the auditory cortex Imaizumi K, Schreiner CE (2007) Spatial interaction
in marmoset monkeys: core and medial belt regions. between spectral integration and frequency gradient
J Comp Neurol 496:72–96 in primary auditory cortex. J Neurophysiol
De Ribaupierre F, Goldstein MH Jr, Yeni-Komshian 28:2933–2942
G (1972) Cortical coding of repetitive acoustic pulses. Imaizumi K, Priebe NJ, Crum PAC, Bedenbaugh PH,
Brain Res 48:205–225 Cheung SW, Schreiner CE (2004) Modular functional
Dear SP, Simmons JA, Fritz J (1993) A possible neuronal organization of cat anterior auditory field.
basis for representation of acoustic scenes in auditory J Neurophysiol 92:444–457
cortex of the big brown bat. Nature 364:620–623 Imaizumi K, Priebe NJ, Sharpee TO, Cheung SW,
Dong C, Qin L, Liu Y, Zhang X, Sato Y (2011) Neural Schreiner CE (2010) Encoding of temporal informa-
responses in the primary auditory cortex of freely tion by timing, rate, and place in cat auditory cortex.
behaving cats while discriminating fast and slow PLoS One 5:e11531
click-trains. PLoS One 6:e25895 Imaizumi K, Priebe NJ, Cheung SW, Schreiner CE
Edeline JM, Manunta Y, Nodal FR, Bajo VM (1999) Do (2011) Spatial organization of repetition rate
auditory responses recorded from awake animals processing in cat anterior auditory field. Hear Res
reflect the anatomical parcellation of the auditory thal- 280:70–81
amus? Hear Res 131:135–152 Imig TJ, Morel A (1985) Tonotopic organization in ven-
Eggermont JJ (1998a) Azimuth coding in primary audi- tral nucleus of medial geniculate body in the cat.
tory cortex of the cat. II. Relative latency and J Neurophysiol 53:309–340
interspike interval representation. J Neurophysiol Jones EG (2007) The thalamus. Cambridge University
80:2151–2161 Press, Cambridge
Eggermont JJ (1998b) Representation of spectral and tem- Joris PX, Schreiner CE, Rees A (2004) Neural processing
poral sound features in three cortical fields of the cat. of amplitude-modulated sounds. Physiol Rev
Similarities outweigh differences. J Neurophysiol 84:541–577
80:2743–2764 Kaas JH (2008) The evolution of the complex sensory and
Eggermont JJ (2000) Sound-induced synchronization of motor systems of the human brain. Brain Res Bull
neural activity between and within three auditory cor- 75:384–390
tical areas. J Neurophysiol 83:2708–2722 Kaas JH, Hackett TA (2000) Subdivisions of auditory
Eggermont JJ (2001) Between sound and perception: cortex and processing streams in primates. Proc Natl
reviewing the search for a neural code. Hear Res Acad Sci U S A 97:11793–11799
157:1–42 Kaur S, Lazar R, Metherate R (2004) Intracortical path-
Ehret G (1997) The auditory cortex. J Comp Physiol ways determine breadth of subthreshold frequency
A 181:547–557 receptive fields in primary auditory cortex.
Furukawa S, Xu L, Middlebrooks JC (2000) Coding of J Neurophysiol 91:2551–2567
sound-source location by ensembles of cortical neu- Kaur S, Rose HJ, Lazar R, Liang K, Metherate R (2005)
rons. J Neurosci 20:1216–1228 Spectral integration in primary auditory cortex: lami-
Goldberg JM, Brown PB (1969) Response of binaural nar processing of afferent input, in vivo and in vitro.
neurons of dog superior olivary complex to dichotic Neuroscience 134:1033–1045
tonal stimuli: some physiological mechanisms of Kilgard MP, Merzenich MM (1998) Plasticity of temporal
sound localization. J Neurophysiol 32:613–636 information processing in the primary auditory cortex.
Goldstein MHJ, Kiang NYS, Brown RM (1959) Responses Nat Neurosci 1:727–731
of the auditory cortex to repetitive acoustic stimuli. Kilgard MP, Merzenich MM (1999) Distributed
J Acoust Soc Am 31:356–364 representation of spectral and temporal
Hackett TA (2011) Information flow in the auditory cor- information in rat primary auditory cortex. Hear Res
tical network. Hear Res 27:133–146 134:16–28
Kishan AU, Lee CC, Winer JA (2008) Branched projec- fields in areas AI and AAF of mouse auditory cortex.
tions in the auditory thalamocortical and J Neurophysiol 90:2660–2675
corticocortical systems. Neuroscience 154:283–293 Liu BH, Wu GK, Arbuckle R, Tao HW, Zhang LI
Kishan AU, Lee CC, Winer JA (2011) Patterns of (2007) Defining cortical frequency tuning with recur-
olivocochlear branches. Open J Neurosci 1:1–7 rent excitatory circuitry. Nat Neurosci 10:1594–1600
Knight PL (1977) Representation of the cochlea within the Lomber SG, Malhotra S (2008) Double dissociation of
anterior auditory field (AAF) of the cat. Brain Res ‘what’ and ‘where’ processing in auditory cortex. Nat
130:447–467 Neurosci 11:609–616
Lam YW, Sherman SM (2005) Mapping by laser photosti- Lu T, Wang X (2000) Temporal discharge patterns evoked
mulation of connections between the thalamic reticular by rapid sequences of wide- and narrowband clicks in
and ventral posterior lateral nuclei in the rat. the primary auditory cortex of cat. J Neurophysiol
J Neurophysiol 94:2472–2483 84:236–246
Lam YW, Sherman SM (2007) Different topography of Lu T, Liang L, Wang X (2001) Temporal and rate
the reticulothalamic inputs to first- and higher-order representations of time-varying signals in the
somatosensory thalamic relays revealed using auditory cortex of awake primates. Nat Neurosci
photostimulation. J Neurophysiol 98:2903–2909 4:1131–1138
Lam YW, Sherman SM (2010) Functional organization of Ma L, Tai X, Su L, Shi L, Wang E, Qin L (2013) The
the somatosensory cortical layer 6 feedback to the neuronal responses to repetitive acoustic pulses in
thalamus. Cereb Cortex 20:13–24 different fields of the auditory cortex of awake rats.
Lee CC (2013) Thalamic and cortical pathways PLoS One 8:e64288
supporting auditory processing. Brain Lang 126:22–28 Malone BJ, Scott BH, Semple MN (2007) Dynamic ampli-
Lee CC, Imaizumi K (2013) Functional convergence of tude coding in the auditory cortex of awake rhesus
thalamic and intrinsic inputs in cortical layers 4 and macaques. J Neurophysiol 98:1451–1474
6. Neurophysiology 45:396–406 Malone BJ, Scott BH, Semple MN (2010) Temporal codes
Lee CC, Sherman SM (2008) Synaptic properties of tha- for amplitude contrast in auditory cortex. J Neurosci
lamic and intracortical inputs to layer 4 of the first- and 30:767–784
higher-order cortical areas in the auditory and somato- Mardia KV (1972) Statistics of directional data. Aca-
sensory systems. J Neurophysiol 100:317–326 demic, London
Lee CC, Sherman SM (2011) On the classification of Miller LM, Escabı́ MA, Read HL, Schreiner CE
pathways in the auditory midbrain, thalamus, and cor- (2001) Functional convergence of response properties
tex. Hear Res 276:79–87 in the auditory thalamocortical system. Neuron
Lee CC, Winer JA (2005) Principles governing auditory 32:151–160
cortex connections. Cereb Cortex 15:1804–1814 Miller LM, Escabı́ MA, Read HL, Schreiner CE
Lee CC, Winer JA (2008a) Connections of cat auditory (2002) Spectrotemporal receptive fields in the lemnis-
cortex: I. Thalamocortical system. J Comp Neurol cal auditory thalamus and cortex. J Neurophysiol
507:1879–1900 87:516–527
Lee CC, Winer JA (2008b) Connections of cat auditory Morel A, Imig TJ (1987) Thalamic projections to fields A,
cortex: III. Corticocortical system. J Comp Neurol AI, P, and VP in the cat auditory cortex. J Comp
507:1920–1943 Neurol 265:119–144
Lee CC, Winer JA (2011a) Convergence of thalamic and Morel A, Garraghty PE, Kaas JH (1993) Tonotopic orga-
cortical pathways in cat auditory cortex. Hear Res nization, architectonic fields, and connections of audi-
274:85–94 tory cortex in macaque monkeys. J Comp Neurol
Lee CC, Winer JA (2011b) A synthesis of auditory cortical 335:437–459
connections: thalamocortical, commissural, and Mountcastle VB (1997) The columnar organization of the
corticocortical systems. In: Winer JA, Schreiner CE neocortex. Brain 120:701–722
(eds) The auditory cortex. Springer, New York, Nagarajan SS, Cheung SW, Bedenbaugh P, Beitel RE,
pp 147–170 Schreiner CE, Merzenich MM (2002) Representation
Lee CC, Imaizumi K, Schreiner CE, Winer JA (2004) Con- of spectral and temporal envelope of twitter vocaliza-
current tonotopic processing streams in auditory cor- tions in common marmoset primary auditory cortex.
tex. Cereb Cortex 14:441–451 J Neurophysiol 87:1723–1737
Lee CC, Kishan AU, Winer JA (2011) Wiring of divergent Nieuwenhuys R (2013) The myeloarchitectonic studies on
networks in the central auditory system. Front the human cerebral cortex of the Vogt-Vogt school,
Neuroanat 5:46 and their significance for the interpretation of func-
Liang L, Lu T, Wang X (2002) Neural representations of tional neuroimaging data. Brain Struct Funct
sinusoidal amplitude and frequency modulations in the 218:303–352
primary auditory cortex of awake primates. Niwa M, Johnson JS, O’Connor KN, Sutter ML
J Neurophysiol 87:2237–2261 (2012) Active engagement improves primary auditory
Linden JF, Liu RC, Sahani M, Schreiner CE, Merzenich cortical neurons’ ability to discriminate temporal mod-
MM (2003) Spectrotemporal structure of receptive ulation. J Neurosci 32:9323–9334
Petkov CI, Kayser C, Steudel T, Whittingstall K, II. Comparison between cortical fields. Hear Res
Augath M, Logothetis NK (2008) A voice region in 32:49–64
the monkey brain. Nat Neurosci 11:367–374 Schreiner CE, Winer JA (2007) Auditory cortex mapmak-
Philibert B, Beitel RE, Nagarajan SS, Bonham BH, ing: principles, projections, and plasticity. Neuron A
Schreiner CE, Cheung SW (2005) Functional organi- 56:356–365
zation and hemispheric comparison of primary audi- Schulze H, Langner G (1997) Periodicity coding in the
tory cortex in the common marmoset (Callithrix primary auditory cortex of the Mongolian gerbil
jacchus). J Comp Neurol 487:391–406 (Meriones unguiculatus): two different coding strate-
Pinault D (2004) The thalamic reticular nucleus: structure, gies for pitch and rhythm? J Comp Physiol
function and concept. Brain Res Brain Res Rev A 181:651–663
46:1–31 Scott BH, Malone BJ, Semple MN (2011) Transformation
Polley DB, Read HL, Storace DA, Merzenich MM of temporal processing across auditory cortex of
(2007) Multiparametric auditory receptive field orga- awake macaques. J Neurophysiol 105:712–730
nization across five cortical fields in the albino rat. Sherman SM, Guillery RW (2006) Exploring the thalamus
J Neurophysiol 97:3621–3638 and its role in cortical function. MIT Press, London
Poremba A, Malloy M, Saunders RC, Carson RE, Smith PH, Bartlett EL, Kowalkowski A (2007) Cortical
Herscovitch P, Mishkin M (2004) Species-specific and collicular inputs to cells in the rat paralaminar
calls evoke asymmetric activity in the monkey’s tem- thalamic nuclei adjacent to the medial geniculate
poral poles. Nature 427:448–451 body. J Neurophysiol 98:681–695
Preuss A, Muller-Preuss P (1990) Processing of amplitude Smith PH, Uhlrich DJ, Manning KA, Banks MI
modulated sounds in the medial geniculate body of (2012) Thalamocortical projections to rat auditory cor-
squirrel monkeys. Exp Brain Res 79:207–211 tex from the ventral and dorsal divisions of the medial
Read HL, Winer JA, Schreiner CE (2001) Modular orga- geniculate nucleus. J Comp Neurol 520:34–51
nization of intrinsic connections associated with spec- Suga N (1994) Multi-function theory for cortical
tral tuning in cat auditory cortex. Proc Natl Acad Sci processing of auditory information: implications of
U S A 98:8042–8047 single-unit and lesion data for future research.
Reale RA, Imig TJ (1980) Tonotopic organization in audi- J Comp Physiol A 175:135–144
tory cortex of the cat. J Comp Neurol 182:265–291 Suga N, Manabe T (1982) Neural basis of amplitude-
Recanzone GH, Schreiner CE, Sutter ML, Beitel RE, spectrum representation in auditory cortex of the mus-
Merzenich MM (1999) Functional organization of tached bat. J Neurophysiol 47:225–255
spectral receptive fields in the primary auditory cortex Suga N, Zhang Y, Yan J (1997) Sharpening of frequency
of the owl monkey. J Comp Neurol 415:460–481 tuning by inhibition in the thalamic auditory nucleus of
Rennaker RL, Carey HL, Anderson SE, Sloan AM, the mustached bat. J Neurophysiol 77:2098–2114
Kilgard MP (2007) Anesthesia suppresses Sullivan WE 3rd (1982a) Neural representation of target
nonsynchronous responses to repetitive broadband distance in auditory cortex of the echolocating bat
stimuli. Neuroscience 145:357–369 Myotis lucifugus. J Neurophysiol 48:1011–1032
Rosen S (1992) Temporal information in speech: acoustic, Sullivan WE 3rd (1982b) Possible neural mechanisms of
auditory and linguistic aspects. Philos Trans R Soc target distance coding in auditory system of the
Lond B Biol Sci 336:367–373 echolocating bat Myotis lucifugus. J Neurophysiol
Rouiller E, de Ribaupierre Y, Toros-Morel A, de 48:1033–1047
Ribaupierre F (1981) Neural coding of repetitive clicks Ter-Mikaelian M, Sanes DH, Semple MN (2007) Trans-
in the medial geniculate body of cat. Hear Res formation of temporal properties between auditory
5:81–100 midbrain and cortex in the awake Mongolian gerbil.
Rutkowski RG, Miasnikov AA, Weinberger NM J Neurosci 27:6091–6102
(2003) Characterisation of multiple physiological Wang X, Merzenich MM, Beitel RE, Schreiner CE
fields within the anatomical core of rat auditory cortex. (1995) Representation of a species-specific vocaliza-
Hear Res 181:116–130 tion in the primary auditory cortex of the common
Schreiner CE, Cynader MS (1984) Basic functional orga- marmoset: temporal and spectral characteristics.
nization of second auditory cortical field (AII) of the J Neurophysiol 74:2685–2706
cat. J Neurophysiol 51:1284–1305 Wenstrup JJ (2005) The tectothalamic system. In: Winer
Schreiner CE, Raggio MW (1996) Neuronal responses in JA, Schreiner CE (eds) The inferior colliculus.
cat primary auditory cortex to electrical cochlear stim- Springer, New York
ulation. II. Repetition rate coding. J Neurophysiol Winer JA (1984) The human medial geniculate body. Hear
75:1283–1300 Res 15:225–247
Schreiner CE, Urbas JV (1986) Representation of ampli- Winer JA, Larue DT (1996) Evolution of GABAergic
tude modulation in the auditory cortex of the cat. I. The circuitry in the mammalian medial geniculate body.
anterior auditory field (AAF). Hear Res 21:227–241 Proc Natl Acad Sci U S A 93:3083–3087
Schreiner CE, Urbas JV (1988) Representation of ampli- Winer JA, Lee CC (2007) The distributed auditory cortex.
tude modulation in the auditory cortex of the cat. Hear Res 229:3–13
A 292 Auditory Transducer, Model
Winer JA, Miller LM, Lee CC, Schreiner CE (2005) Audi- parameters that result in physiologically realistic
tory thalamocortical transformation: structure and neuronal models of the CPGs. Such methods
function. Trends Neurosci 28:255–263
Yin P, Johnson JS, O’Connor KN, Sutter ML (2011) Cod- include “brute-force” explorations of predefined
ing of amplitude modulation in primary auditory cor- parameter spaces, as well as various heuristics
tex. J Neurophysiol 105:582–600 (e.g., multi-objective evolutionary algorithms)
Zhou Y, Mesik L, Sun YJ, Liang F, Xiao Z, Tao HW, used to arrive at a single or more of viable
Zhang LI (2012) Generation of spike latency tuning by
thalamocortical circuits in auditory cortex. J Neurosci model parameter combinations.
32:9969–9980
Auditory Transducer, Model Central pattern generators (CPGs) are neural net-
works that produce rhythmically patterned out-
▶ Cochlear Inner Hair Cell, Model puts, without relying on any sensory feedback
(Hooper 2001). CPGs drive such critical rhyth-
mic activity as breathing, chewing, swimming,
walking, heartbeat control, etc. CPGs have been
Augmentation shown to produce rhythmic outputs akin to nor-
mal rhythmic activity patterns, even in isolation
▶ Short-Term Synaptic Plasticity in Central from other parts of the nervous system, which
Pattern Generators makes them popular physiological models. Fur-
thermore, due to their relative simplicity, espe-
cially in such model organisms as lobsters, crabs,
or leeches, CPGs have also become quite wide-
Autoassociative Networks spread in computational modeling studies of cel-
lular and synaptic properties of individual
▶ Olfactory Cortical Associative Memory neurons and small neural networks.
Models While hand-tuning has been traditionally used
in the process of creating CPG neuronal models
(e.g., Soto-Treviño et al. 2005), in light of recent
advances in the computational capabilities of
modern computing systems, which now facilitate
Automated Parameter Search in the use of more complex neuronal models (i.e., in
Small Network Central Pattern terms of the number of compartments or free
Generators parameters), and allow for the exploration of
unprecedentedly large parameter search spaces,
Tomasz G. Smolinski this approach has become virtually obsolete.
Department of Computer and Information Therefore, automated methods for model param-
Sciences, Delaware State University, Dover, eter search have been lately gaining much
DE, USA attention.
There are basically two approaches to the
problem of searching for optimal (i.e., physiolog-
Definition ically realistic) sets of parameter values for
models of small network central pattern genera-
Automated parameter search in small network tors: (1) “brute-force” explorations of predefined
central pattern generators (CPGs) involves the parameter spaces and (2) explorations utilizing
use of any methods other than manual (i.e., heuristic optimization approaches, such as multi-
hand-tuning) to generate or tune sets of objective evolutionary algorithms (MOEAs).
Automated Parameter Search in Small Network Central Pattern Generators 293 A
“Brute-Force” Explorations of Predefined granularity may allow for five to ten possible
Parameter Spaces values, incorporating up to 20-fold variation
In the case of “brute-force” explorations of with hundreds of possible values for each param- A
predefined parameter spaces, the study usually eter is not unheard of in a heuristic approach.
starts with a hand-tuned model of the CPG, Since this approach is not tasked with simulation
which serves as the “center” for the parameter and analysis of all of the possible combinations of
search space that is created around it. Then, phys- values in such created parameter search space, it
iologically realistic ranges for the model param- remains computationally feasible.
eters (e.g., maximal conductances of membrane Another critical step in this approach is the
and synaptic currents) are chosen, along with the definition of one or more measures of the given
granularity for each of the parameters. The gran- CPG’s characteristics that will be used to deter-
ularity determines how many possible values mine physiological adequacy of the models.
each of the parameters can assume and does not While such measures can be virtually identical
have to be the same for all the parameters, as to the ones used in the “brute-force” approach,
some of them will exhibit different sensitivities the difference lies in the fact that they are being
to changes in their values. In some cases, the first used during the process of model generation
step may be omitted and only the parameters, itself, rather than at the end to filter out the
along with their ranges and granularities, are unwanted models. These measures, in essence,
determined. become the fitness functions utilized in the pro-
After such a grid-based parameter search cess of optimizing model parameter values to
space has been constructed, all of the possible drive it toward generating as many as possible
combinations of the parameter values are simu- “good” models which match the biological sys-
lated and tested for their physiological adequacy tem, while limiting the number of “bad”
(possibly under multiple simulation scenarios, solutions.
such as spontaneous activity, response to current After the model parameters and the
injections, removal of neuromodulation, etc.). corresponding ranges and granularities have
Only those models that match the behavior of been determined, and the appropriate fitness
the biological CPG, which is determined by functions (possibly multiple, even conflicting)
means of one or more quantitative or qualitative have been defined, the iterative process of opti-
measures of the CPG’s characteristics (e.g., spike mization of the model parameter values begins
height, inter-spiking interval, burst duration, and ultimately yields a collection of physiologi-
period, preservation of the phase of the rhythm, cally realistic CPG models.
etc.), are retained for further analyses. However,
the rejected models are also sometimes subjected Applications
to examination in order to determine what makes Most of the hitherto applications of the auto-
“bad” models unacceptable. mated parameter searches in small network cen-
tral pattern generators have been performed in
Explorations Utilizing Heuristic Optimization relatively simple invertebrate CPGs. For exam-
Approaches ple, Doloc-Mihu and Calabrese (2011) utilized
In the case of explorations utilizing various heu- the “brute-force” approach to construct a large
ristic approaches, such as multi-objective evolu- (on the order of terabytes) database of
tionary algorithms, the study usually starts with conductance-based models of the half-center
the determination of the model parameters, along oscillator from the leech heartbeat central pattern
with their ranges and granularities, similarly to generator to determine how neuronal parameters
the “brute-force” approach, but often on a much influence the network activity. Using the same
larger scale. In other words, while the range in the approach, G€unay and Prinz (2010) utilized
“brute-force” approach may reflect a 3- to 4-fold a large (20,250,000) database of models of the
variation in the parameter values, and the lobster pyloric network to study calcium sensors
A 294 Axon Model
for network homeostasis. Smolinski, Prinz et al.,

used both the “brute-force” approach and multi- Axon Model
objective evolutionary algorithms to study
the cellular and synaptic properties of the ▶ Peripheral Nerve Models
AB/PD (anterior burster/pyloric dilator) pace-
maker kernel in the lobster pyloric network
(2006, 2009), as well as the conductance correla-
tions involved in the recovery of bursting after Axon, Modeling
neuromodulator deprivation (Shim et al. 2012;
Malik et al. 2013). Bruce Graham
University of Stirling, Stirling, UK
Cross-References
Definition
▶ Neuronal Model Databases
▶ Neuronal Parameter Space Exploration Computational modeling of axons is used to
determine the action potential initiation and prop-
agation properties along these long and highly
branched structures. Issues under investigation
References include the site and threshold of action potential
initiation, propagation speed in unmyelinated and
Doloc-Mihu A, Calabrese R (2011) A database of compu-
tational models of a half-center oscillator for analyzing
myelinated axons, and safety factors of propaga-
how neuronal parameters influence network activity. tion through branch points and other geometrical
J Biol Phys 37:263–283 inhomogeneities, such as presynaptic boutons.
G€
unay C, Prinz AA (2010) Model calcium sensors for Modeling allows exploration of the interaction
network homeostasis: sensor and readout parameter
between axonal morphology (diameters and
analysis from a database of model neuronal networks.
J Neurosci 30(5):1686–169 branching structure) and passive and active mem-
Hooper SL (2001) Central pattern generators. In: Ency- brane properties in determining the speed and
clopedia of life sciences. Wiley Hoboken, NJ reliability of action potential propagation. This
Malik A, Shim K, Prinz AA, Smolinski TG (2013) Multi-
can include the effects of ion channel noise.
objective evolutionary algorithms for analysis of con-
ductance correlations involved in recovery of bursting Modeling is also used to explore axonal develop-
after neuromodulator deprivation in lobster ment, including growth cone guidance and
stomatogastric neuron models. BMC Neurosci 14(1): branching.
P370
Shim K, Prinz AA, Smolinski TG (2012) Analyzing con-
ductance correlations involved in recovery of bursting
after neuromodulator deprivation in lobster Detailed Description
stomatogastric neuron models. BMC Neurosci 13(1):
P37
The axon is the principle communication path-
Smolinski TG, Prinz AA (2009) Computational intelli-
gence in modeling of biological neurons: A case way from one neuron to another. Brief voltage
study of an invertebrate pacemaker neuron. In: Pro- pulses called action potentials (AP) are initiated
ceedings of international joint conference on neural near the cell body and travel along the axon to
networks, Atlanta, GA, pp 2964–2970
Smolinski TG, Soto-Treviño C, Rabbah P, Nadim F, Prinz
reach the many presynaptic boutons. The AP
AA (2006) Analysis of biological neurons via model- causes neurotransmitter release in a bouton in
ing and rule mining. Int J Inf Technol Intell Comput a stochastic manner, resulting in an electrical
1(2):293–302 response in a target neuron. The reality is more
Soto-Treviño C, Rabbah P, Marder E, Nadim F (2005)
complex than this: AP propagation is not always
Computational model of electrically coupled, intrinsi-
cally distinct pacemaker neurons. J Neurophysiol reliable due to geometrical and electrical con-
94(2):590–604 straints in the axon; timing of AP arrival at
Axon, Modeling 295 A
boutons may vary, having implications for
information processing in target nuclei; and AP
shape in boutons will influence transmitter A
release. The experimental data showing these
effects is nicely summarized in Debanne (2004).
How theoretical results and computational
models have been used to aid our understanding
of these processes is described in Segev and
Schneidman (1999), a paper that is still highly
relevant to the field.
Computational modeling is also making
a significant contribution to our understanding Axon, Modeling, Fig. 1 Action potential (AP) in a long,
of how axons develop, from their differentiation uniform axon. (a) AP over time at 25 % along the axon. (b)
from dendrites at initiation to their guidance to AP over space after 20 ms. Simulation performed using
the NEURON simulator
target structures and formation of functional con-
nections. This work is documented by van Ooyen
in a recent review (van Ooyen 2011) and in an process takes place within a few milliseconds,
edited volume (van Ooyen 2003). forming a sharp voltage wave of around
100 mV, which is the action potential (Fig. 1a).
The basic model describes changes in the trans-
Hodgkin-Huxley Model of the Action membrane voltage (Vm) in an isopotential patch
Potential of neuronal membrane as a function of membrane
capacitance (Cm), sodium (INa) and potassium
Computational modeling of the action potential (IK) ionic currents, a nonspecific “leak” current
in axons begins with the work of Hodgkin and (IL), and experimentally injected electrode cur-
Huxley in determining a model based on experi- rent (Ie; to initiate the action potential). The equa-
mental data of the action potential in the squid tions of this model are as follows:
giant axon (Hodgkin and Huxley 1952). Their
model forms the basis of most subsequent models Membrane Voltage
of action potentials and also of models of current
flow through membrane-bound ion channels in dV m
Cm ¼ gL ðV m EL Þ gNa m3 hðV m ENa Þ
general in the nervous system. Based on their dt
experimental data, they formulated models of gK n4 ð V m E K Þ þ I e
the electrical currents generated by the flow of
sodium (Na) and potassium (K) ions across the
membrane. Importantly, these currents are volt- Sodium Current
age sensitive, so that a membrane depolarization I Na ¼ gNa m3 hðV ENa Þ
results in an increase in the sodium current, which
dm
in turn leads to further depolarization and an ¼ am ð1 mÞ bm m
dt
increase in the sodium current through
a positive feedback mechanism. This activation
of the sodium conductance is curtailed by a rapid dh
¼ a h ð 1 h Þ bh h
inactivation. The potassium current also dt
increases with depolarization, but more slowly
than the sodium current. Eventually, the potas-
V þ 40
sium current grows large enough that it, along am ¼ 0:1
with the inactivation of the sodium current, leads 1 expððV þ 40Þ=10Þ
to a repolarization of the membrane. This entire bm ¼ 4expððV þ 65Þ=18Þ
A 296 Axon, Modeling
ah ¼ 0:07expððV þ 65Þ=20Þ electrical load, depending on the diameters of the

1 daughter branches. The work of Rall and col-
bh ¼ leagues (see collected papers in Segev
expððV þ 35Þ=10Þ þ 1
et al. 1995) established the important concept of
Potassium Current the geometrical ratio (GR) between the diameters
of two daughter branches (d1 and d2) to that of the
dn parent branch (dp; section of axon or dendrite
I K ¼ gK n4 ð V EK Þ ¼ a n ð 1 n Þ bn n
dt closer to the cell body), defined as
V þ 55
an ¼ 0:01 3=2
d1 þ d2
3=2
1 expð ðV þ 55Þ=10 GR ¼
dp3=2
bn ¼ 0:125expððV þ 65Þ=80Þ
Given uniform membrane properties, if
Action Potential Propagation
GR = 1, the branch point imposes no change in
electrical impedance, and the branching can be
This model can be extended to include action
collapsed to an equivalent uniform cylinder.
potential propagation along the axon using these
However, if GR > 1, the branch point imposes
membrane active properties (sodium and potas-
an extra electrical load that slows the AP, reduces
sium currents) in a compartmental model of an
its height, and can potentially lead to a failure of
elongated axon (Fig. 1b). The compartmental
AP propagation. In contrast, if GR < 1, AP veloc-
model arises as a spatial discretization of the
ity and its height both increase as the branch point
partial differential equation describing spatially
is approached, and propagation is reliable. Gold-
extensive current flow along an axon:
stein and Rall (1974) provided the first compre-
hensive computational investigation of AP
@V m d @2V
Cm ¼ gL ð V m EL Þ propagation through specified geometrical inho-
@t 4Ra @x2 mogeneities. An example is shown in Fig. 2. Here
gNa m3 hðV m ENa Þ gK n4 ðV m EK Þ there is a single bifurcation in a long axon. The
parent axon (at the left) has a diameter of 1 mm.
For the simulation shown in Fig. 1, the axon is The two daughter branches have an equal diam-
a cylinder 300 mm long, with a uniform diameter eter of 2.52 mm, giving GR = 8. The action
of 476 mm. It is divided into 500 equal-length potential (AP) traces are from the parent axon,
computational compartments to obtain the branch point, and one daughter branch, as illus-
numerical solution. The simulation was trated in the figure. In this configuration (solid
performed using the NEURON simulator lines in time plots), the AP is significantly slowed
(Carnevale and Hines 2006) with code derived and reduced in amplitude at the branch point
from that used in Chapter 3 of Principles of (second trace from left), before quickly recover-
Computational Modelling in Neuroscience ing and increasing in velocity. If the daughter
(Sterratt et al. 2011). diameters are reduced to 0.63 mm, then
GR = 1, and the configuration is equivalent to
Speed and Reliability of Propagation a single, uniform axon of 1 mm diameter. In this
Axons are rarely simply long, uniform cylinders. case, the AP propagates with uniform amplitude
Many axons become highly branched as they near and velocity (dashed lines). Note that though the
their target structures. Diameters can change AP at the branch point is significantly delayed
along the length of the axon, with a general taper- when GR = 8, compared to GR = 1, the larger
ing to thinner diameters toward terminals but daughter branch diameter results in a higher
with large increases in diameter at presynaptic velocity for the AP once it is through the branch
boutons. Branch points may impose a change in point, and it quickly overtakes the AP traveling in
Axon, Modeling, Fig. 2 Action potential (AP) in GR = 8. Dashed lines: GR = 1. Simulations performed
a branched axon. Recording sites illustrated at left. Traces using the NEURON simulator
of APs over time shown at right. Solid lines:
the uniform axon (right-hand solid and dash which has been seen experimentally (Debanne
plots, respectively). Goldstein and Rall (1974) 2004), cannot arise due to the branch point geom-
showed that for axons with the same membrane etry alone (Segev and Schneidman 1999).
(Rm) and axial (Ra) resistivity, AP speed is con- Changes in the active membrane properties are
stant in units of length constant per unit time, also required and may arise due to differences in
irrespective of diameter (d). The passive length ion channel distributions between branches or
constant (l) is differential extracellular ion accumulation during
successive AP propagations leading to alterations
rffiffiffiffiffiffiffiffiffi
dRm in ion current reversal potentials (particularly EK;
l¼ see the Hodgkin-Huxley model current equa-
4Ra
tions) and hence current magnitudes.
This equates to an increase in speed in units of In summary, computational modeling has
physical length per unit time with an increase in been fundamental to improving our understand-
diameter. ing of the characteristics of and constraints to AP
Manor et al. (1991) considered AP propaga- propagation along branched and irregular axons.
tion in realistic axonal branching structures, with
a particular emphasis on the delays imposed by
geometrical irregularities. This study shows that Myelinated Axon
though delays in AP propagation at individual
branch points are at most small (a few tenths of Propagation speeds along axons containing
a millisecond), accumulated delays in a complex, a continuous distribution of ion channels along
branching axon can result in arrival times at their length (so-called unmyelinated axons) are
boutons differing by a few milliseconds. Delays rather modest, on the order of <1 m/s. Most long
or speedups from successive, nearby branch axons in the nervous system are myelinated, with
points do not sum linearly. In particular, succes- glial cells providing a high-resistance membrane
sive summed delays may be supralinear com- sheath around the axons for much of their length.
pared to the delays at the individual, isolated Ion channels are largely restricted to the sites of
branch points. regular breaks in this sheath, known as nodes of
Differential branch point failure, in which the Ranvier. Action potential propagation speeds are
AP fails to propagate along one daughter branch, much greater along these axons than along unmy-
while successfully propagating along the other, elinated fibers, on the order of 10–100 m/s.
source being the stochastic opening and closing

of ion channels in neuronal membrane. The thin-
ness of axons makes them vulnerable to this form
of noise. Computational models in which the
deterministic ion channel models of Hodgkin
and Huxley are replaced by equivalent stochastic
models are used to investigate the impact of ion
channel “noise” on action potentials (Goldwyn
and Shea-Brown 2011).
Axon, Modeling, Fig. 3 Myelinated axon. Structural
The standard approach to turning
details around a node of Ranvier
a deterministic model of the action potential,
Computational modeling contributes to our such as the Hodgkin-Huxley model, into an
understanding of AP propagation along myelin- equivalent stochastic model is to treat the ionic
ated axons, particularly the implications of the conductance in a patch of membrane of
very specific locality of sodium and potassium a particular ionic species as being a function of
ion channels at and around the nodes of Ranvier the number of open ion channels for that species.
(Fig. 3). Sodium channels are located at the node, Firstly, the dynamics of conductance change in
with potassium channels largely restricted to the the deterministic model can be rewritten in the
paranodal region. form of a kinetic scheme, which is interpreted as
In a simplistic approach, myelination may be the concentration of ion channels in different
modeled as an increased passive resistance and states, one of which is an “open” state and pro-
a decreased capacitance of the internodal vides the conductance. An example of such
(myelinated) axonal membrane, along with a scheme for the potassium conductance in the
nonuniform distribution of active sodium and Hodgkin-Huxley model is
potassium channels. However, the narrow extra-
4an , bn 3an , 2bn 2an , 3bn an , 4bn
cellular (periaxonal) space afforded by the ½n0 Ð ½n1 Ð ½n2 Ð ½n3 Ð ½n4
myelination, plus the spatial separation of sodium
and potassium channels, may require the modeling Now the total conductance for potassium in
of ion concentrations and the electro-diffusion of a patch of membrane is
ions intra- and extracellularly, to fully capture the
action potential propagation properties of the mye- gK ðv, tÞ ¼ gK ½n4
linated axon (Nygren and Halter 1999).
Models show that excessive accumulation of where gK is the conductance of a single open
potassium in the periaxonal space following potassium channel.
high-frequency repetitive activity can lead to This becomes a stochastic model if the patch of
action potential failure (Parnas and Segev 1979). membrane is assumed to contain N ion channels
Models are also used to explore the impact of loss and the rates of the kinetic schemes become tran-
of myelin, as occurs in diseases such a multiple sition probabilities between states for individual
sclerosis (Coggan et al. 2010). Loss of myelin channels, resulting in continuous-time Markov
leads to a slowing of action potential propagation chain. The stochastic model approaches the deter-
or even action potential failure altogether. ministic model as N goes to infinity. However, for
finite numbers of all ion channels, N, the stochastic
model can exhibit significantly different behavior
Ion Channel Noise from the deterministic model.
Ion channel numbers in a patch of axon may be
Noise in the nervous system comes in many dif- such that the variation in membrane potential
ferent forms (Faisal et al. 2008), with a major from the opening and closing of individual ion
channels may be apparent, particularly near the
threshold of action potential initiation where the
number of open ion channels is small A
(Schneidman et al. 1998). This can lead to
stochasticity in the generation of action poten-
tials, as illustrated in Fig. 4.
Fluctuations in ion channel opening can
advance or retard action potential initiation and
even lead to failures in propagation, though prop-
agation is more reliable than initiation (Faisal and
Laughlin 2007). Multiple trials of a stochastic
Axon, Modeling, Fig. 4 Spontaneous generation of
model of axon stimulation by a fluctuating cur-
action potentials in a stochastic Markov model of rent are shown in Fig. 5.
a 100 mm2 patch of membrane containing 6,000 sodium Simulating the Markov model for all ionic
and 1,800 potassium channels. Simulation performed species is computationally demanding. Other
using the MATLAB code of Goldwyn and Shea-Brown
(2011) as available in ModelDB (accession number
ways of incorporating membrane noise, which
138950) involve some stochastic description of the noise
Axon, Modeling, Fig. 5 Action potential initiation and 60 repeated trials recorded at equally spaced axonal posi-
propagation due to fluctuating current stimulation in tions. Data is extracted from 10-s trials (Figure 1 from
a stochastic membrane model of a squid-type axon with Faisal and Laughlin (2007), reproduced with permission
0.2 mm diameter. The topmost row shows the stimulus according to the Creative Commons Attribution License)
current. Below, each row contains spike raster plots of
Axon, Modeling,
Fig. 6 Neurite outgrowth
in an external environment
containing three sources of
a diffusible attractive
chemical (Reproduction of
Figure 10.10 from Sterratt
et al (2011), with
permission of the authors)
Axon, Modeling, Fig. 7 Axonal pathfinding. Detection on opposite sides of the cone, the tension may be enough
of chemoattractants in the external environment by to split the cone into two, leading to the formation of
filopodia produces tension on the growth cone in particu- daughter branches (Reproduction of Fig. 3 from Graham
lar directions. The growth cone will turn toward and grow and van Ooyen (2006), with permission according to the
along the dominant direction. If similar forces are exerted Creative Commons Attribution License)
contribution of channel fluctuations, have been Ooyen (2007) provide a suitable mathematical
tried and can provide a good approximation to the description of such an environment in the form
full Markov description (Goldwyn and Shea- of partial differential equations (PDEs) and
Brown 2011). quasi-steady-state approximations solved across
a spatial grid. Axons themselves may not play
a space-filling role in the environment.
Axon Development Models concerned with the details of axon
growth and branching, on the other hand, need
A completely different class of model is used to to include the mechanisms of intracellular signal-
explore the development of axons. A major issue ing, the viscoelastic properties of the axon, and
here is how an axon reaches its target nucleus and the spatial extent of growth cones (Fig. 7).
forms synapses therein. Models of axon guidance Computational simulation environments,
make explicit the interaction between extracellu- including CX3D (Zubler and Douglas 2009) and
lar attractive and repulsive cues and intracellular NETMORPH (Koene et al. 2009), provide facil-
growth mechanisms in the growth cone and ities for modeling aspects of axonal and dendritic
trailing axon. This requires modeling some form development and the subsequent formation of
of extracellular environment that contains physi- neural networks. A review of modeling and com-
cal barriers to axon growth and diffusible or puter simulation techniques for neuronal devel-
membrane-bound molecules acting as attractive opment is provided by Graham and van
and repulsive cues (Fig. 6). Krottje and van Ooyen (2006).
Cross-References generation of large scale neuronal networks with real-
istic neuron morphologies. Neuroinformatics
7(3):195–210
▶ Hodgkin-Huxley Model Krottje J, van Ooyen A (2007) A mathematical framework A
for modelling axon guidance. Bull Math Biol
69(1):3–31
References Manor Y, Koch C, Segev I (1991) Effect of geometrical
irregularities on propagation delay in axonal trees.
Biophys J 60:1424–1437
Carnevale NT, Hines M (2006) The NEURON book. Nygren A, Halter JA (1999) A general approach to model-
Cambridge University Press, Cambridge ing conduction and concentration dynamics in excit-
Coggan JS, Prescott SA, Bartol TM, Sejnowski TJ able cells of concentric cylindrical geometry. J Theor
(2010) Imbalance of ionic conductances contributes Biol 199:329–358
to diverse symptoms of demyelination. Proc Natl Parnas I, Segev I (1979) A mathematical model for con-
Acad Sci USA 107:20602–20609 duction of action potentials along bifurcating axons.
Debanne D (2004) Information processing in the axon. J Physiol (Lond) 295:323–343
Nat Rev Neurosci 5:304–316 Schneidman E, Freedman B, Segev I (1998) Ion channel
Faisal AA, Laughlin SB (2007) Stochastic simulations on stochasticity may be critical in determining the reli-
the reliability of action potential propagation in thin ability and precision in spike timing. Neural Comput
axons. PLoS Comput Biol 3(5):e79 10:1679–1703
Faisal AA, Selen LPJ, Wolpert DM (2008) Noise in the Segev I, Schneidman E (1999) Axons as computing
nervous system. Nat Rev Neurosci 9:292–303 devices: basic insights gained from models. J Physiol
Goldstein SS, Rall W (1974) Changes in action potential Paris 93:263–270
shape and velocity for changing core conductor geom- Segev I, Rinzel J, Shepherd GM (eds) (1995) The theoret-
etry. Biophys J 14:731–757 ical foundation of dendritic function: the collected
Goldwyn JH, Shea-Brown E (2011) The what and where papers of Wilfrid Rall. Bradford Book/MIT Press,
of adding channel noise to the Hodgkin-Huxley equa- Cambridge, MA
tions. PLoS Comput Biol 7(11):e1002247 Sterratt D, Graham B, Gillies A, Willshaw D (2011) Prin-
Graham BP, van Ooyen A (2006) Mathematical modelling ciples of computational modelling in neuroscience.
and numerical simulation of the morphological devel- Cambridge University Press, Cambridge
opment of neurons. BMC Neurosci 7(Suppl 1):S9 van Ooyen A (ed) (2003) Modeling neural development.
Hodgkin A, Huxley A (1952) A quantitative description of Bradford Book/MIT Press, Cambridge, MA
membrane current and its application to conduction van Ooyen A (2011) Using theoretical models to analyse
and excitation in nerve. J Physiol (Lond) 117:500–544 neural development. Nat Rev Neurosci 12:311–326
Koene RA, Tijms B, van Hees P, Postma F, de Ridder A, Zubler F, Douglas R (2009) A framework for modeling the
Ramakers GJ, van Pelt J, van Ooyen A (2009) growth and development of neurons and networks.
NETMORPH: a framework for the stochastic Front Comput Neurosci 3:25
B
Backpropagating Action Potentials Balanced State
▶ Action Potential Backpropagation Lars Schwabe

Department of Computer Science and Electrical
Engineering, Adaptive and Regenerative
Software Systems, Universit€at Rostock,
Rostock, Germany
Backpropagating Spikes
▶ Action Potential Backpropagation

Synonyms
High-conductance state
Baer-Rinzel (BR) Continuum Model

Definition
▶ Dendritic Spines: Continuum Theory
The balanced state of neurons is characterized by
a high value of the total membrane conductance
caused by synaptic inputs. It is the state of corti-
cal neurons in vivo in awake animals, and it is due
Balance Control Dynamics
to massive excitatory and inhibitory inputs that
almost cancel out. In this state, the responses are
▶ Human Balancing Tasks: Power Laws, Inter-
driven by fluctuations (possible also due to coin-
mittency, and Lévy Flights
cident events) in the input, and due to a greatly
reduced effective membrane time constant, the
neuron responds rapidly.
Balance Following Cochlear

Implantation Detailed Description
▶ Vestibular Function After Cochlear The notion of high-conductance states, and the
Implantation fact that neurons could integrate differently in

B 304 Balanced State
such states, was first proposed by modeling stud- of conductance relative to the resting leak con-
ies. A meanwhile classical dichotomy distin- ductance is to assume it as due to synaptic inputs.
guished between two operating modes of
neurons: the integrator mode and the coincidence Consequences for Single-Neuron
detection mode. In the integrator mode, the post- Computations
synaptic neuron is emitting spikes, because it inte- In 1994 Shadlen and Newsome explored the idea
grates over time the excitatory presynaptic inputs. that excitatory and inhibitory synaptic inputs might
In the coincidence detection mode, it fires spikes, be balanced and keep the membrane potential close
because many presynaptic excitatory neurons syn- to, but below, the firing threshold (Shadlen and
chronize their excitatory inputs. By changing the Newsome 1994). In this scenario the neuron can
time constant of the postsynaptic neuron, one still respond, because fluctuations can push the
could switch between these two modes: Coinci- membrane potential above a threshold. Simulation
dence detection calls for a short time constant, studies showed that in this regime, the neural
where the neuron quickly “forgets” previous responses are much more irregular compared to a
inputs and needs coincident spikes so that the non-balanced scenario with little or no inhibition,
membrane potential is elevated enough to emit i.e., when strong excitation drives the responses.
a postsynaptic spike; in the integration mode, the For neuronal models this means that neurons in
time constant should be longer to allow for sum- a balanced state are best described with stochastic
ming inputs over time. Cortical neurons in vivo are formalisms (Tuckwell 1988).
likely to operate in the latter mode, because due to Given that neurons in the balanced state are
massive excitatory and balanced inhibitory inputs, sensitive to fluctuations in the inputs, they could
the effective time constant is very small. This has also detect changes in such fluctuations. Of par-
various consequences for cortical computations at ticular interest have been those brought about by
both the single-neuron and network levels, which changing correlations between presynaptic excit-
are now addressed in greater detail, aver the atory, inhibitory, or between excitatory and
observation of a “high-conductance state” in inhibitory spikes (Salinas and Sejnowski 2000).
recapitulated. Interestingly, such correlations could change
without affecting the average of the inputs. For
Empirical Evidence for a “High-Conductance neuronal modeling this is important, because it
State” calls for stochastic models beyond simplistic
It is well known to physiologists that neurons in rate-base descriptions. The fluctuations of bal-
anesthetized animals often appear during intra- anced inputs have been implicated in changing
cellular recordings as if they operate in two dis- the gain of neurons. Such gain modulations have
tinct states, namely a so-called “up” state and been described as a physiological signature of
a “down” state. In the “up” state, the neurons attentional top-down modulations. Chance
are depolarized and can fire action potentials, et al. (2002) have shown the feasibility of this
which are irregular as observed in extracellular mechanism using modeling and dynamic-clamp
recordings in awake animals. In the “down” state (Sharp et al. 1992), but in vivo tests of this
the neurons are hyperpolarized and emit no hypothesis remain technically too challenging.
spikes. How do these two states emerge? The Another consequence of the balanced state is
detailed mechanisms are still under investigation, that responses to inputs can be very rapid as
but experiments revealed that the “down” state pointed out already in 1975 by Barrett (1975).
can be much more easily manipulated via current This can be demonstrated using the simple RC
injections than the “up” state. The lack of strong circuit model of the subthreshold membrane
effects in the “up” state is consistent with a very potential dynamics: Let us consider a leaky inte-
high membrane conductance. A plausible expla- grator, modeled as an RC circuit, with a time
nation of this approximately 2–3 times increase constant t = RC = C/g, where R is the
Balanced State 305 B
membrane resistance (and g = 1/R the conduc- the soma predicted by Rall and Barrett could be
tance) and C is the membrane capacitance. The balanced by an increased likelihood of evoking
dynamics of membrane potential V are given by a dendritic spike as the input resistance increases
with distance from the soma.
dV
C ¼ gV þ gAMPA ðtÞðEAMPA V Þ B
dt Consequences for Network-Based
þ gGABA ðtÞðEGABA V Þ Computations
While single-neuron modeling studies usually
where gAMPA(t) and gGABA(t) are the excitatory
assume background inputs with certain properties
and inhibitory synaptic conductances (with rever-
such as balanced inputs, network models need to
sal potentials EAMPA and EGABA) due to approx-
explicitly account for them. In 1996 van
imately balanced presynaptic inputs, i.e., the
Vreeswijk and Sompolinsky (1996) showed that
resulting currents cancel each other at a certain
sparsely connected recurrent networks with
subthreshold voltage. These synaptic conduc-
strong but balanced excitation and inhibition nat-
tances are time dependent due to the different
urally lead to the temporally irregular spiking
arrivals of spikes at the synapses and the kinetics
patterns observed in awake animals. One may
of synaptic transmission. Averaging these con-
wonder how such a tight balance between excita-
ductances over time yields gAMPA and gGABA.
tion and inhibition required by network models
dt ¼ 0 and solving for the effective time
Setting dV
could be realized in cortical networks, given that
constant teff shows that
local inhibitory interneurons are far less numer-
ous than the excitatory neurons. Despite their
C C lower absolute number, inhibitory neurons are
teff ¼ < :
g þ gAMPA þ gGABA g often reported to have higher firing rates. More-
over, their synapses onto postsynaptic targets
Thus, with increasing synaptic conductance, show much less short-term depression than excit-
the neuron responds faster. This may be of paratory synapses. In combination, this could make
ticular relevance in sensory systems with high- the recurrent inhibition much stronger than evi-
frequency inputs, because in a non-balanced dent by their lower absolute number.
state, the neuron could not lock to such inputs. It is known that blocking inhibition causes the
Such a high-conductance state also affects the cortex to become epileptic (Dichter and Ayala
integration of dendritic signals. In line with Rall’s 1987), and already in 1975 Sillito showed that
classical works (Rall 1964), Barrett also suggests reducing inhibition lets sensory neurons
a location-dependence of the impact of dendritic (in primary visual cortex, V1) lose their orienta-
inputs onto the somatic membrane potential. He tion selectivity (Sillito 1975). This further supports
emphasized the role of synaptic background the notion that the balanced state is the natural
activity, which would decrease the resistance cortical state of awake animals. Experimental
and further reduce the impact of distal dendritic studies later investigated in greater detail the
inputs. Experiments showed, however, that the potential network mechanisms underlying the
dendritic inputs are approximately location-inde- spike responses in vivo using paired recordings.
pendent (Magee 2000). This observation could be For example, Okun and Lampl showed in 2008 via
a natural consequence of the massive synaptic simultaneous intracellular recordings of pairs of
inputs onto active excitable dendrites in the bal- nearby neurons in rat somatosensory cortex (Okun
anced state, i.e., dendrites that support the local and Lampl 2008) that the excitatory and inhibitory
initiation of spikes (London and H€ausser 2005). inputs covary in strength and time. They found this
Rudolph and Destexhe have shown in simulation covariation during both the spontaneous “resting-
studies (Rudolph and Destexhe 2003) that the state” activity and during sensory stimulation.
decreased impact with increasing distance from Marino et al. (2005) performed intracellular
B 306 Balanced State
recordings in vivo in V1 and showed that such fascinating properties of the balanced state that
a covariation in space could explain the experi- warrant this investment are still to be discovered.
mentally observed invariance of the orientation
tuning with location in the orientation map
of V1. The balanced state as observed in References
single neurons may also have large-scale macro-
scopic correlates, namely, the spontaneously Barrett JN (1975) Motoneuron dendrites: role in synaptic
occurring states reported during resting-state integration. Fed Proc 34:1398–1407
activity in V1 (Kenet et al. 2003). From Blumenfeld B, Bibitchkov D, Tsodyks M (2006) Neural
network model of the primary visual cortex: from
a modeling perspective, it is worth mentioning functional architecture to lateral connectivity and
that such macroscopic phenomena can be simu- back. J Comp Neurosci 20(2):214–241
lated and analyzed using rate model (Blumenfeld Chance FS, Abbott LF, Reyes AD (2002) Gain modulation
et al. 2006). from background synaptic input. Neuron 35:773–782
Dichter MA, Ayala GF (1987) Cellular mechanisms of
epilepsy: a status report. Science 237:157–164
Discussion and Future Directions Kenet T, Bibitchkov D, Tsodyks M, Grinvald A, Arieli A
Meanwhile, the balanced state is accepted in the (2003) Spontaneously emerging cortical representa-
community as the state that characterizes cortical tions of visual attributes. Nature 425(6961):954–956
London M, H€ausser M (2005) Dendritic computation. Ann
neurons in vivo. Many studies have investigated Rev Neurosci 28:503–532
how it affects single-neuron computations. Other Magee JC (2000) Dendritic integration of excitatory syn-
studies have developed self-consistent network aptic input. Nat Rev Neurosci 1(3):181–190
models that reproduce the balanced state. Exper- Mariño J, Schummers J, Lyon DC, Schwabe L, Beck O,
Wiesing P, Obermayer K, Sur M (2005) Invariant
imental tests of model predictions appeared in the computations in local cortical networks with balanced
literature with quite some delay relative to the excitation and inhibition. Nat Neurosci 8(2):194–201
modeling studies. This is not a surprise, because Okun M, Lampl I (2008) Instantaneous correlation of
direct empirical evidence for the balanced state is excitation and inhibition during ongoing and sensory-
evoked activities. Nat Neurosci 11:535–537
technically very hard to obtain, as it calls for Rall W (1964) Theoretical significance of dendritic trees
intracellular multiunit recordings of connected for neuronal input-output relations. In: Reiss R, Alto
neurons in vivo. Future modeling studies should P (eds) Neural theory and modeling. Stanford Univer-
respect this and aim at specific predictions for sity Press, Stanford
Rudolph M, Destexhe A (2003) A fast-conducting, sto-
easier to access macroscopic observables. Future chastic integrative mode for neocortical neurons
modeling studies could also investigate how in vivo. J Neurosci 23:2466–2476
adaptation at various timescales affects the bal- Salinas E, Sejnowski TJ (2000) Impact of correlated syn-
ance between excitation and inhibition. Given aptic input on output firing rate and variability in
simple neuronal models. J Neurosci 20:6193–6209
that cortical networks seem to be exquisitely bal- Shadlen MN, Newsome WT (1994) Noise, neural codes
anced, even small perturbations could have major and cortical organization. Curr Opin Neurobiol
impacts. The role of inhibitory plasticity in 4:569–579
maintaining the balanced state has been Sharp AA, Abbott LF, Marder E (1992) Artificial electri-
cal synapses in oscillatory networks. J Neurophysiol
addressed in first modeling studies (Vogels 67:1691–1694
et al. 2012). Exploring how a certain adaptation Sillito AM (1975) The contribution of inhibitory mecha-
or learning mechanism behaves in the balanced nisms to the receptive field properties of neurones in
state in vivo is now becoming a necessity for all the striate cortex of the cat. J Physiol 250:305–329
Tuckwell HC (1988) Introduction to theoretical neurobi-
modeling studies of adaptation, learning, and cor- ology. Cambridge University Press, Cambridge, UK
tical computation. Finally, one may need to van Vreeswijk CA, Sompolinsky H (1996) Chaos in neu-
address explicitly the energetic costs of the bal- ronal networks with balanced excitatory and inhibitory
anced state. Are the potential benefits, from activity. Science 274:1724–1726
Vogels TP, Sprekeler H, Zenke F, Clopath C, Gerstner
a normative point of view, really worth all the W (2012) Inhibitory plasticity balances excitation and
energy the brain is investing into maintaining this inhibition in sensory pathways and memory networks.
mode of operation? It could be that the most Science 336(6083):802
Baroreflex Models 307 B
From modeling perspective the baroreflex
Baroreceptor Reflex feedback control system consists of three parts
(Fig. 1):
▶ Baroreflex Models • The afferent part where the arterial pressure is
being read out, transduced, and supplied B
• The central nervous system (CNS) where this
input is processed and converted to the sym-
Baroreflex Models pathetic and parasympathetic nerve activities
• The effector organs that respond to sympa-
Yaroslav Molkov thetic and parasympathetic tones by adjusting
Department of Mathematical Sciences, Indiana heart rate, contractility, vascular resistance,
University – Purdue University Indianapolis, and a number of other physiological
Indianapolis, IN, USA parameters
The afferent part is represented by barorecep-
tors, mechanoreceptors located in the great arter-
Synonyms ies, heart, and pulmonary vasculature, that
transduce arterial pressure into action potential
Baroreceptor reflex trains. Baroreceptors provide input to second-
order neurons in nucleus tractus solitarius (NTS)
located in the lower brainstem. Via a network of
Definition interneurons, the second-order barosensitive neu-
rons provide an excitatory input to cardiac vagal
Blood pressure is controlled by several feedback motoneurons (CVM) in nucleus ambiguus
mechanisms. The fastest one baroreceptor reflex (NA) and dorsal motor nucleus of vagus
(baroreflex) can be defined as the biological neu- (DMNX) and inhibit the pre-sympathetic neurons
ral control system responsible for the short-term in rostral ventrolateral medulla (RVLM). CVMs
blood pressure regulation. and pre-sympathetic RVLM neurons define
Baroreflex Models, Fig. 1 Location of baroreceptors dorsal motor nucleus of vagus, CVLM caudal ventrolateral
and neural pathways of baroreflex responses. Abbrevia- medulla, RVLM rostral ventrolateral medulla, IML
tions: CNS central nervous system, para-SNA parasympa- intermediolateral column, PVN paraventricular nucleus,
thetic nerve activity, SNA sympathetic nerve activity, NTS SON supraoptic nucleus of hypothalamus (From Chapleau
nucleus tractus solitarius, NA nucleus ambiguus, DMNX 2011 with permission)
B 308 Baroreflex Models
Baroreflex Models, Fig. 2 Baroreflex negative feedback control of arterial blood pressure (BP) (From Chapleau 2011
with permission)
parasympathetic (para-SNA) and sympathetic called nucleus tractus solitarius (NTS) through
(SNA) nerve activities, respectively. Baroreflex glossopharyngeal nerve (from carotid sinus
models are aimed to quantify para-SNA and SNA baroreceptors) and vagus nerve (from aortic
reflex responses to sharp changes in arterial BP arch) that together are known as the buffer
and, thus, attempt to describe the neural part of nerves. In the NTS this information is
the system. processed by an intricate network of interneu-
Closed-loop models of cardiovascular system rons that control efferent parasympathetic
usually include a baroreflex compartment, and (para-SNA) and sympathetic (SNA) nerve
a number of effector organs controlled by sym- activities and release of antidiuretic peptide
pathetic and parasympathetic activities (Fig. 2). vasopressin (AVP) (see Fig. 1). All these mech-
anisms in response to changes in the arterial BP
cause reflexes (Fig. 2) that ultimately push BP
Detailed Description in the opposite direction providing a negative
feedback control of arterial BP called arterial
Deviations in blood pressure (BP) and blood vol- baroreflex. The CNS pathways of cardiopul-
ume are sensed in cardiovascular system by monary baroreceptors are similar to that of arte-
baroreceptors, mechanosensitive nerve endings rial ones except their activation has little effect
activated by vascular or cardiac distension on heart rate (HR) which is the fastest mecha-
(Fig. 2). Arterial baroreceptors (located in the nism of BP control.
aortic arch and the carotid sinuses, see Fig. 1) Due to complexity of the system, modeling
increase their activity when BP rises, and reduce studies usually focus on a particular aspect
it if BP lowers. Cardiopulmonary barorecep- depending on phenomena considered while
tors in the heart, vena cava and pulmonary vas- using simplified description of other compart-
culature are aimed to sense central blood volume, ments. Several integrative mathematical models
and are sometimes called volume receptors or were used to elucidate the role of baroreflex in
low-pressure receptors. cardiovascular instabilities, such as heart rate
The afferent activity from arterial barorecep- variability and spontaneous fluctuations of the
tors is sent to the medullary brainstem area arterial pressure (e.g., Mayer waves).
Pressure-Electrical Transduction withdrawal from the increased BP, baroreceptors
As a result of a change in arterial pressure, the first exhibit reduced activity (post-excitatory
cross-sectional area of the arteries changes lead- depression) which then adapts back to baseline.
ing to mechanical deformation of artery walls. In short, the properties of the firing rate of the
Changes in the membrane strain resulting from baroreceptors in response to changing arterial B
the artery wall deformation lead to changes in the pressure modeling-wise important can be formu-
conductance of a variety of mechanosensitive ion lated as follows (Ottesen et al. 2004):
channels in the sensory terminals of arterial baro- • Firing rate increases with BP.
receptors. The cumulative effect of the opening • The response exhibits threshold and
of those channels in response to rise in pressure is saturation.
an increased inward (depolarizing) current. If the • Sufficiently fast decreases in pressure cause
resulting depolarization exceeds certain thresh- firing rate to fall below the threshold.
old, baroreceptors start firing action potentials • A step change in pressure causes a step change
with a rate defined by the magnitude of this in firing rate followed by a decay in firing rate
current. (called adaptation or resetting depending on
In mechanoreceptors variations in membrane the timescale).
strain affect protein conformations and, • Response curves are sigmoidal and show
hence, cause changes in the probabilities for asymmetric hysteresis-like behavior.
mechanosensitive membrane channels to be in
open and close states (Morris 1990). Accord- Conductance-Based Baroreceptor Model
ingly, a net inward current flow through Baroreceptor fibers are divided into two different
mechanosensitive ion channels is often modeled types, A-type and C-type, depending on whether
as a nonspecific current with zero reversal poten- they are myelinated or not. Schild et al.
tial and Boltzmann-like dependence of the con- (1994) performed a thorough study of their elec-
ductance on arterial pressure P: trophysiological properties and developed a
conductance-based model of A- and C-type
g cells based on voltage-clamp recordings in rat
I baro ¼ max
ðV Em Þ
1 þ exp K P P1=2 nodose sensory neurons (carotid sinus barorecep-
tors). Their model included two Na+ currents
where gmax is a maximal conductance of this exhibiting fast and slow tetrodotoxin (TTX)-
current, P1/2 is half-activation pressure, insensitive kinetics; low- and high-threshold
K defines the sensitivity and Em is a reversal Ca2+ currents exhibiting transient and long-
potential usually accepted 0. lasting dynamics, respectively; and outward K+
Rose et al. (1995) used a little different currents consisting of a delayed-rectifier K+ cur-
approach. Their baroreceptors had a pressure- rent and Ca2+-activated K+ current. The model
dependent depolarizing current which was linear also includes extra- and intracellular C2+
above a threshold pressure: Ibaro = K(P Pthr) dynamics.
if P > Pthr and Ibaro = 0 otherwise.
Firing Rate-Based Baroreceptor Model
Baroreceptors In Ottesen et al. (2004) a very elegant model is
Baroreceptor response to changes in arterial BP is suggested that accounts for major properties of
very complex, nonlinear, and has multiple time- baroreceptor response. His model describes
scales. Rapid jumps in BP lead to much stronger nonlinear responses of a baroreceptor to changes
changes of baroreceptor activity than gradual in carotid pressure in terms of a single nonlinear
excursions of the same magnitude. Acute hyper- ordinary differential equation:
tension results in a sharp rise in the baroreceptor
dn dPc nðM nÞ n N
activity which, however, declines in time if ¼k
hypertension is maintained. After acute dt dt ðM=2Þ2 t
where n is the firing rate of the baroreceptor, Pc is in the NTS. One of the puzzles of the identified
the carotid pressure, M is the maximal firing rate, baroreflex-related neurons in NTS is that in spite
t is the adaptation time constant, k is a weighting of receiving direct inputs from the first-order
constant, and N denotes the threshold. Instanta- neurons, their activity does not have any fre-
neous rate of change of the firing rate is propor- quency component corresponding to cardiac fre-
tional to the time derivative of carotid pressure quency. Somehow they do not respond to each
and to the firing rate itself. Accordingly, the heartbeat by performing a “low-pass filtration” of
change in firing rate is most sensitive to change their inputs.
in carotid pressure when the value of the firing Rogers et al. (2000) suggested that this prop-
rate is in the middle of it physiological range erty can be concerned either with specific bio-
[0, M] and almost insensitive near the boundary. physical properties of NTS baroreceptors (Fig. 3)
This defines the sigmoidal form of the response or with specific interactions in the NTS network
curve. of second-order interneurons (Fig. 4). By analyz-
Experimental data shows that the adaptation ing the responses of the second-order NTS neu-
has many different timescales from seconds to rons to induced blood pressure pulses, they have
hours or even days. As mentioned, baroreceptor noticed that barosensitive second-order NTS neu-
fibers are divided into different types, fast A and rons respond to blood pressure changes with
slow C fibers, in accordance to whether they are a burst of activity whose frequency is much
myelinated or not. The fast A fibers are further lower than the frequency of cardiac cycle, and
divided into subtypes. It was suggested that dif- the activity of these neurons is inhibited just
ferent types of baroreceptors can be characterized
by significantly different adaptation time con-
stants (Ottesen and Olufsen 2011). In this case
their integrated activity n is calculated as
X
K
n¼Nþ Dni
i¼1
where Dni is a deviation of the firing rate of the

type i baroreceptor from the threshold value N,
driven by a nonlinear differential equation simi-
lar to the above:
d dPc nðM nÞ 1
Dni ¼ ki Dni
dt dt ðM=2Þ2 ti
and K is a number of different baroreceptor types.

Ottesen and Olufsen (2011) show the example Baroreflex Models, Fig. 3 Schematic used by Rogers
where a satisfactory fit of the experimentally et al. (2000) for NTS model based on specific electrophys-
observed responses of the firing rate to a step iological properties. Arterial pressure is transduced by
baroreceptors (nodose) according to their pressure thresh-
input pressure requires K = 3 types of barorecep-
olds, indicated by shading (top cell: highest pressure
tors with time constants t1 0.5, t2 10, and threshold). All NTS neurons receive input from all baro-
t3 3000 s. receptors. Each input to any given NTS cell is equally
weighted. The only difference between the NTS neurons is
the total synaptic weight provided by the baroreceptors,
Nucleus Tractus Solitarius (NTS)
which is varied systematically with the top neuron (A)
Baroreceptors map the spatial pressure distribu- receiving the highest total drive and the bottom neuron
tion onto the inputs to the second-order neurons (E) receiving the lowest total drive
Baroreflex Models, Fig. 4 Architecture of the model input to the excitatory population of second-order NTS
proposed to explain variability of NTS neuronal responses neurons. This population sends excitatory inputs to the
to buffer nerve shock and arterial pressure increases. This population of inhibitory neurons (represented by the
model consists of two populations of baroreceptors, type shaded circle), which in turn inhibits the former
A and type C, that have different pressure thresholds and (Adapted from Rogers et al. 2000 with permission)
sensitivities. The two baroreceptor populations provide
before and after the bursts. Using these observa- buffer nerve shock with excitation (EPSPs), inhi-
tions they developed a series of mathematical bition (IPSP), or an EPSP/IPSP combination.
models of early stages of baroreflex based on Accordingly, they may be excited, inhibited, or
the intrinsic cell properties and network responded with an extinction response to
mechanisms. increases in arterial pressure. To account for
a variety of responses, Rogers et al. (2000) pro-
Intrinsic Properties-Based Model posed a populational model of barosensitive NTS
Individual NTS neurons were modeled using neurons based on specific neural circuitry. In this
a single-compartment description in the case they also used Hodgkin–Huxley-based
Hodgkin–Huxley formalism. The most important description but without Ca2+ and K(Ca2+) con-
feature of their individual neuron model was that ductances. So this model solely relies on the
in addition to ionic channels responsible for spik- network architecture, rather than on the neurons’
ing behavior (like fast sodium and potassium biophysical properties.
delayed rectifier channels), they incorporated
mechanisms for the response adaptation (A-type Barotopic Hypothesis
potassium, calcium-dependent potassium, and Because of the different pressure thresholds,
low threshold calcium currents). This mechanism baroreceptors have a distributed sensitivity to
together with intracellular calcium dynamics BP and its rate of change. “Barotopic” organiza-
allowed them to reproduce several important tion hypothesis suggests that individual barore-
properties of the second-order baroreceptors pre- ceptors’ pressure thresholds (below which they
viously identified experimentally: (1) frequency are silent) are topically distributed in an interval
adaptation in the cell’s response to a step stimu- of the pressures and that each second-order
lus; (2) sensitivity to the rate of change of input; barosensitive neuron receives inputs from the
and (3) the lack of response to each pulse of first-order ones, whose thresholds lie in
a pulsatile stimulus; however, if these pulses a definite small interval of pressure values.
were imposed on a ramp stimulus, the cell could Rose et al. (1995) used this architecture in
respond to each pulse. their model of baroreflex control of heart rate.
The neural portion of their model included baro-
Inhibition-Based Model receptors, neurons of the nucleus tractus
Barosensitive NTS neurons show a variety of solitarius (NTS), and cardiac vagal motoneurons
responses to both buffer nerve shock and arterial (CVM) (Fig. 5) all described in Hodgkin–Huxley
pressure increases. These neurons may respond to style.
Baroreflex Models,
Fig. 5 Architecture of
a neural portion of the
model by Rose et al. (1995)
(Reproduced with
permission)
In addition to basic spiking currents barore- network are unknown. For the sake of simplicity,
ceptors in their model included A-type potassium NTS is often considered as a source of single
current responsible for rapid adaptation and a excitatory drive that directly activates CVMs.
pressure-dependent depolarizing current which The effect on SNA is mediated by GABAergic
was linear above a threshold pressure (see section interneurons in CVLM that in turn inhibit
“Pressure-Electrical Transduction”). Each baro- pre-sympathetic neurons in RVLM. In firing
receptor had its own threshold pressure Pthr, and rate-based approach, the parasympathetic and
the thresholds had a sigmoidal distribution. Each sympathetic activities are usually calculated as
NTS neuron received excitatory inputs from sev- ascending and descending sigmoid functions of
eral baroreceptors with close thresholds and aggregate NTS output, respectively. Ottesen
projected to several CVMs. The NTS neurons et al. (2004), for example, suggest the following
and CVMs had six conductances: sodium, sigmoids:
delayed-rectifier potassium, A-type potassium,
AHP-type potassium, L-type calcium, leakage 1
nsym ðnÞ ¼ , npar ðnÞ ¼ 1 nsym ðnÞ
conductances, and intracellular calcium dynam- 1 þ ðn=n0 Þn
ics including pumping and buffering. The CVMs 1
projected to the sinoatrial (SA) node where their ¼
1 þ ðn=n0 Þn
action potentials activated an acetylcholine
(ACh)-activated inward rectifier K+
where nsym(n) and npar(n) are SNA and para-
hyperpolarizing current which slowed the depo-
SNA, respectively, n is NTS output, n0 is the
larization of the node and hence decreased the
baseline NTS output with fully adapted barore-
heart rate. The effect of vagal impulses on the
ceptor activity, and n defines the steepness of the
hyperpolarizing current was greater for those
curves.
arriving later in the cardiac cycle.
The parasympathetic output is provided
directly by vagal motoneurons, while sympa-
Sympathetic and Parasympathetic Activities thetic activity is formed by a longer chain of
Second-order baroreceptors in the NTS through neural structures and involves several neurotrans-
a network of interneurons modulate parasympa- mitter systems. This difference is often taken into
thetic and sympathetic efferent outputs by excit- account by introducing a time delay in sympa-
ing cardiac vagal motoneurons (CVM) in NA and thetic response which can constitute up to
DMNX and inhibiting pre-sympathetic neurons several seconds. Ottesen and Olufsen (2011), for
of RVLM, respectively. The details of this example, model parasympathetic activity as
proportional to the baroreceptor firing rate the apneustic inspiratory burst. In all cases the
n(t) and SNA is defined by the delayed version sympathetic baroreflex-induced lowering of SNA
of n(t): persisted confirming the presence of a direct
baroreflex pathway which bypasses the respira-
nð t Þ nðt td Þ tory network. B
npar ðtÞ ¼ , nsym ðtÞ ¼ 1 Based on their experimental findings, Baekey
M M
et al. (2010) developed a mathematical model
where M is the maximal baroreceptor firing rate, of baroreceptor-respiratory-sympathetic interac-
and td is the time delay. tions that explained phase-dependent respiratory
response to transient barostimulation. One of the
Respiratory Baroreflex implications of their model was that there is
As mentioned, the classical baroreflex control of indeed a second pathway from baroreceptors to
sympathetic nerve activity (SNA) operates via pre-sympathetic neurons in RVLM. This second
the second-order baroreceptor neurons that pro- pathway is mediated by respiratory circuits, spe-
ject to the caudal ventrolateral medulla region cifically by the post-I neurons of BötC. This
(CVLM). Through this path, the baroreceptor baroreflex pathway is strongly dependent on the
activation provides activation of CVLM neurons respiratory-sympathetic interactions and needs
which in turn inhibit the pre-sympathetic rostral pons to be intact.
ventrolateral medulla (RVLM) neurons hence Another interesting implication of their model
lowering both the RVLM activity and SNA. This is that the sympathetic baroreflex gain depends on
pathway provides a direct negative SNA reflex the respiratory phase, because a subpopulation of
response. NTS second-order baroreceptors is inhibited dur-
Sympathetic activity contains the respiratory ing inspiration (Fig. 6).
modulation, which suggests that sympathetic and
respiratory networks interact. Respiratory activ- The Role of Cardiopulmonary Baroreceptors
ity is also known to be modulated by the barore- A majority of modeling studies only take into
ceptor input. Together these facts imply that there account arterial baroreceptors. However, under
may be another baroreflex pathway mediated by certain conditions blood volume regulation pro-
the respiratory neurons. vided by cardiopulmonary baroreceptors plays
Baekey et al. (2010) have demonstrated that a major role in blood pressure control. To under-
transient pressure pulses perturb the respiratory stand the interplay between different mecha-
pattern in a phase-dependent manner, and those nisms, Ursino (2000) developed a closed-loop
perturbations strongly depend on the integrity of mathematical model which included both arterial
the pons. The stimuli were delivered during inspi- and cardiopulmonary baroreceptors and four
ration, post-inspiration, or late expiration. With effector components: heart period, systemic
pons intact, the applied barostimulation had peripheral resistance, systemic venous unstressed
almost no effect on the amplitude and duration volume, and heart contractility. He simulated the
(i.e., inspiratory period) of the phrenic bursts model to mimic the baroreflex response to mild
even when stimuli were delivered during inspira- and severe acute hemorrhages and found that the
tion. At the same time, these stimuli suppressed cardiopulmonary baroafferents play a significant
or abolished inspiratory modulation of SNA. In role in the control of systemic arterial pressure
contrast, the same stimuli delivered during post- during mild hemorrhages (lower than 3–4 % of
inspiration or late expiration produced an the overall blood volume). Under these condi-
increase in the expiration period and decreased tions arterial pressure could be maintained at its
SNA. The barostimulation-evoked prolongation normal level solely by cardiopulmonary feedback
of expiration was greater if stimulation was loop without the intervention of the arterial baro-
applied later during the expiratory phase. After receptors. During more severe hemorrhages the
pontine transection the barostimulation shortened latter take over the responsibility for pressure
Baroreflex Models, Fig. 6 (a) Conceptual model of RVLM and SNA. Blue dotted arrows represent the effects
interaction between respiratory-related activity of the ven- of VRC and PONS on RVLM providing respiratory mod-
tral respiratory column (VRC), pontine circuits (PONS), ulation of SN activity. (b) The schematic of the computa-
sensory network in the nucleus tractus solitary (NTS), and tional model by Baekey et al. (2010) showing interactions
rostral and caudal ventrolateral medulla (RVLM/CVLM). between different populations of respiratory neurons
The sympathetic baroreceptor reflex operates via two within major brainstem compartments involved in the
pathways (red solid arrows): one direct pathway includes control of breathing and sympathetic motor activity
baroreceptors, 2nd-order barosensitive cells (Baro) in the (pons, BötC, pre-BötC, rVRG, VLM, and NTS). Each
NTS and CVLM, which inhibits RVLM and SNA; the population (shown as a sphere) consists of 20–50 single-
other pathway routes via the Bötzinger complex (BötC) compartment neurons described in the Hodgkin–Huxley
in the VRC, whose post-inspiratory neurons inhibit style (Adapted from Baekey et al. 2010 with permission)
control, and the contribution of cardiopulmonary system with the delayed feedback is unstable for
baroreceptors becomes negligible. sufficiently large time delay values, many differ-
Ursino (2000) noticed that according to their ent models of dynamic arterial pressure control
model the stability region in the parameter space have succeeded in predicting Mayer waves
of the closed-loop system is quite narrow. For (Julien 2006). The magnitude of the time delay
example, an increase in the static gain of the is the determinant factor of the emerging oscilla-
baroreceptors or a decrease in the rate-dependent tion frequency. Mayer wave frequency varies
component may lead to self-sustained oscilla- significantly between species which suggests
tions similar to Mayer waves. that different species should have significantly
different time delays in baroreflex feedback
Baroreflex and Mayer Waves loop. In spite of numerous hypotheses suggested,
Mayer waves are synchronous oscillations in no consistent explanation of this species-to-
SNA and arterial BP slower than respiratory species variability was yet provided. Another
rhythm. The origin and physiological mecha- challenge for the baroreceptor theory is the
nisms of these oscillations are currently dependence of Mayer wave amplitude on the
unknown. The fact that Mayer waves are mean SNA level which the existing models fail
abolished by sinoaortic baroreceptor denervation to explain.
strongly suggests the involvement of arterial
baroreflex in this phenomenon. The baroreflex
theory of Mayer waves explains the emergence References
of self-sustained oscillations in arterial pressure
by the instability caused by fixed time delays in Baekey DM, Molkov YI, Paton JFR, Rybak IA, Dick TE
(2010) Effect of baroreceptor stimulation on the
the baroreflex (presumably sympathetic) feed-
respiratory pattern: insights into respiratory-
back loop (see the review by Julien (2006) and sympathetic interactions. Respir Physiol Neurobiol
the literature therein cited). Since virtually any 174:135–145
Basal Ganglia System as an Engine for Exploration 315 B
Chapleau MW (2011) Baroreceptor reflexes. In: Robert- been identified, the hypothesis that the indirect
son DW, Biaggioni I, Burnstock G, Low PA, Paton pathway of the BG is the subcortical substrate for
JFR (eds) Primer on the autonomic nervous system,
3rd edn. Academic Press, London, UK pp 161–165 exploration leads to explanations for how a range
Julien C (2006) The enigma of Mayer waves: facts and of putative BG functions might be performed.
models. Cardiovasc Res 70(1):12–21 B
Morris CE (1990) Mechanosensitive ion channels.
J Membr Biol 113(2):93–107
Olufsen MS, Larsen JK (2004) Applied mathematical Detailed Description
models in human physiology. Society for Industrial
Mathematics, Philadelphia. Reinforcement Learning and the Basal
Ottesen JT, Olufsen MS (2011) Functionality of the baro- Ganglia
receptor nerves in heart rate regulation. Comput
Method Programs Biomed 101(2):208–219 For nearly a century, a certain “mysteriousness”
Ottesen JT, Olufsen MS, Larsen JK (2004) Applied math- has been attributed to the function of the basal
ematical models in human physiology. Society for ganglia (BG) system – a deep brain circuit of
Industrial Mathematics, Philadelphia multiple interconnected nuclei, with rich connec-
Rogers RF, Rybak IA, Schwaber JS (2000) Computational
modeling of the baroreflex arc: nucleus tractus tions to large parts of the cortex (Kinnier Wilson
solitarius. Brain Res Bull 51:139–150 in his Croonian lectures in 1925, Marsden 1982).
Rose WC, Rybak IA, Schwaber JS (1995) Closed loop The mystique surrounding BG has its roots per-
model of vagally-mediated baroreflex control of heart haps in the multifarious functions of this circuit.
rate. In: Engineering in medicine and biology society,
IEEE 17th annual conference, Montreal, Quebec, Action selection, action gating, sequence gener-
Canada, 20 Sep 1995, pp 1367–1368 ation, motor preparation, reinforcement learning,
Schild JH, Clark JW, Hay M, Mendelowitz D, Andresen timing, working memory, goal-directed behavior,
MC, Kunze DL (1994) A- and C-type rat nodose and exploratory behavior – the list of putative BG
sensory neurons: model interpretations of dynamic
discharge characteristics. J Neurophysiol 71: functions is long and perhaps, by the current state
2338–2358 of knowledge, is also incomplete. Lesions of this
Ursino M (2000) Modeling the interaction among several circuit can show manifestations in many
mechanisms in the short-term arterial pressure control. forms – from simple reaching movements to
In: Ottesen JT, Danielsen M (eds) Mathematical
modeling in medicine. IOS Press, Amsterdam, The handwriting, balancing and gait, speech and lan-
Netherlands, pp 139–162 guage function, eye movements, and force gen-
eration, in addition to cognitive and affective
manifestations. A long line of neurological
(Parkinson’s disease, Huntington’s disease,
Basal Ganglia System as an Engine athetosis, chronic fatigue syndrome) (DeLong
for Exploration 1990) and neuropsychiatric disorders
(schizophrenia, obsessive–compulsive disorder,
V. Srinivasa Chakravarthy and Pragathi ADHD, apathy, abulia, insomnia) (Ring and
Priyadharsini Balasubramani Serra-Mestres 2002) are associated with BG
Department of Biotechnology, Indian Institute of impairment.
Technology, Chennai, India In one of the earliest forays into the mystery of
the BG, Albin et al. (1989) interpreted BG anat-
omy to consist of two pathways (Albin
Definition et al. 1989) with complementary roles in motor
manifestations (Fig. 1). BG consist of several
The basal ganglia (BG) system is a deep brain nuclei – striatum (caudate and putamen), globus
circuit with wide-ranging brain functions. Explo- pallidus externa/interna, subthalamic nucleus,
ration refers to the sampling of a variety of behav- and substantia nigra pars compacta (SNc)/
iors not firmly established within a learned reticulata. The BG circuit as a whole receives
repertoire. While the neural source of variability extensive inputs from the cortex and projects
driving exploration within the subcortex has not back to cortex via thalamus. Cortical signals to
B 316 Basal Ganglia System as an Engine for Exploration
A need to go beyond the simple Go/NoGo

picture of BG function had its seeds in experi-
ments (Houk et al. 1995; Schultz et al. 1997) on
the firing properties of mesencephalic dopamine
cells. Although activities of dopaminergic cells
have been linked to reward sensing for a long
time, experiments by Schultz et al. (1997) specif-
ically showed that dopamine neurons of ventral
tegmental area (VTA) respond to unconditional
rewards (food or juice). When a sensory stimulus
(like a sound or a light flash) consistently pre-
cedes the appearance of reward such that the
stimulus is predictive of the reward, then dopa-
mine cells fire in response to the stimulus and also
the reward. Such findings led to the insight that
dopamine cell activity is analogous to a quantity
known as temporal difference error (TD error)
that appears in reinforcement learning
(RL) theory – a branch of machine learning
(Sutton and Barto. 1998). Recognition of the
analogy between mesencephalic dopamine sig-
nals and TD error signals of RL has inspired
a much larger effort to draw parallels between
other elements of RL theory and anatomical com-
Basal Ganglia System as an Engine for Exploration, ponents of BG. Although the effort to explain
Fig. 1 The schematic showing the direct and indirect various functions of BG using RL concepts is
pathways of basal ganglia a story in the making, it is believed that RL
holds the promise to create a comprehensive the-
BG flow along two pathways: the direct pathway ory of BG in the long term (Chakravarthy
(DP) and the indirect pathway (IP). et al. 2010).
The projections from the striatum, the input port RL theory describes how an agent can learn
to BG, to globus pallidus interna (GPi), one of correct stimulus–response (SR) relationships
the output ports, constitute the DP. The indirect using reward feedback from the environment.
route which connects the striatum to GPi via For a given stimulus, responses that yield rewards
globus pallidus externa (GPe) and subthalamic are reinforced, while those that result in punish-
nucleus (STN) is the IP (Alexander and ment are attenuated. The problem is often com-
Crutcher 1990). Lesions of DP affecting particu- plicated by the fact that rewards come after
larly the projections from the striatum to GPi a delay following the response or even after
are associated with hypokinetic disorders a whole series of responses. The agent then
(distinguished by a paucity of movement), and needs a surrogate to reward, which guides its
lesions of IP produce hyperkinetic disorders like responses in the intervening period. RL theory
chorea and tremor. These findings led to the proposes the value function as such a surrogate;
thinking that activation of the DP facilitates a computational module called the critic com-
movement, and hence it became known as the putes value. (The module that performs actions
“Go” pathway. Contrarily, IP was dubbed the is known as the actor.) Value is defined as the
“NoGo” pathway since its activation typically total discounted future reward that an agent
inhibits movement (Contreras-Vidal and expects to receive from a given state. Thus,
Stelmach 1995). once the value is known, at any instant, the
agent chooses the response or action that brings prefrontal cortex and ACC when subjects were
about the greatest increase in value, a process exploring a maze (Yoshida and Ishii 2006). In the
known as exploitation. Sometimes it is desirable subcortex, it was suggested that the ventral and
for the agent to try out actions that are not opti- dorsal striata correspond to critic and actor,
mal, by way of adapting to the changing reward respectively (O’Doherty et al. 2004). Thus, B
patterns of the real world. This selection of though both cortical and subcortical substrates
suboptimal actions, typically stochastically, is of exploitation have been discovered, no
known as exploration. Thus, exploitation and corresponding subcortical substrates for explora-
exploration are two key complementary pro- tion have been found.
cesses, the yin and yang, of RL theory (Sutton Can there be subcortical substrates for explo-
and Barto. 1998). ration? Stein et al. (1997) showed that
decorticated kittens can exhibit exploratory and
The Indirect Pathway and Exploration goal-oriented behavior (Stein et al. 1997). Rats
In actor/critic (AC) models of BG, the emphasis with damaged STN were shown to exhibit per-
is often on the respective substrates for the actor severative behavior or reduced exploration of
and critic components in the striatum (Joel new options, with persistent selection of older
et al. 2002), and the dopamine signal for its role unrewarding ones (Baunez et al. 2001). When
in training the AC components. Thus, though bicuculline, a GABA antagonist, was injected
exploitation and exploration are complementary into the anterior GPe of primates, the animals
processes, exploitation receives most of the atten- exhibited stereotypic movements, and when it
tion. Such omission is perhaps not surprising as was injected into dorsolateral GPe, the animal
even in the AC framework, the actor and critic are produced hyperactivity that included exploratory
recognized explicitly as modules, while explora- or searching movements for food (Grabli
tion is a mere mechanism obtained by a “noise et al. 2004). Drawing inspiration from the studies
term” in RL equations. Since variability is ubiq- of Usher and Cohen et al. (1999), Doya (2002)
uitous in the brain – either arising due to thermal suggested a link between norepinephrine levels
noise or chaotic neural dynamics – the search for and the “inverse temperature” parameter which
a specific substrate for exploration in BG was felt controls exploration in RL literature. It is note-
unnecessary. worthy that globus pallidus is reported to have
Experimental evidence particularly from high norepinephrine levels (Russell et al. 1992).
functional neuroimaging seems to support this Thus, it appears compelling that the STN–GPe
partial view, with a bias towards cortical subsystem constituting the IP of BG might be the
strates over subcortical ones. In fMRI studies, subcortical substrate for exploratory behavior.
gamblers were asked to choose between slots The STN–GPe system and its intriguing oscilla-
that are expected to give highest rewards tory activity do not seem to occupy a prominent
(“exploit”) and less familiar slots that might place in AC modeling literature. On the other
turn out to be more profitable (“explore”). The hand, there is an entire line of modeling work
areas in the brain that are preferentially activated that presents the STN–GPe system as a
during exploitation or exploration are noted (Daw pacemaker in the brain, in reference to its oscil-
et al. 2006). While substrates for value computa- latory activity (Ring and Serra-Mestres 2002;
tions were found in orbitofrontal cortex (Knutson Willshaw and Li 2002). These oscillations have
et al. 2001), substrates for exploration were found also been linked to Parkinsonian tremor (Hurtado
in anterior frontopolar cortex and intraparietal et al. 1999; Terman et al. 2002). Though the
sulcus (Daw et al. 2006). The anterior cingulate aforementioned STN–GPe models explain the
cortex (ACC) is suggested to be involved in behavioral effects of pathological oscillations,
balancing between exploitation and exploration they attribute no role to the oscillations in the
(Rushworth and Behrens 2008). Studies by RL framework that is thought to govern the pro-
Yoshida and Ishii (2006) found activation in cesses of BG. Under dopamine-deficient or
Parkinsonian conditions, the firing patterns of and D2 receptors of striatal medium spiny
STN and GPe neurons show dramatically neurons
increased correlation without significant increase – Oscillations in the STN–GPe system
in firing rate (Bergman et al. 1994; Brown – Value computations in the striatum
et al. 2001). Since exploration is driven by noise – The classical Go (DP) and NoGo (IP) with the
in RL models, a brain region that drives explora- added “Explore” behavior.
tion is expected to be a source of noise generated
perhaps by complex neural dynamics. Consider- Striatum
ing the low correlation in STN–GPe under nor- The binary action selection problem consists of
mal conditions, with increased correlation or loss choosing between two inputs based on their
of complexity in pathology, it is plausible that “salience.” Thus, the input, a 2D vector, Iext,
the STN–GPe is a subcortical substrate for representing the corticostriatal afferents, is
exploration. presented to the striatum that consists of two 1D
By an extended application of RL concepts, layers of sigmoidal neurons representing medium
a comprehensive model of BG can be built in spiny neurons (MSNs) (Fig. 2). The first layer
which the exploitative dynamics of the DP can represents neurons that express D1-type dopa-
be combined with STN–GPe oscillations that mine receptors (R), whereas the second layer
drive exploratory behavior (Chakravarthy represents D2R-expressing MSNs. The D1R-
et al. 2010). Thus emerges a view that while DP and D2R-MSN layers project to GPi and GPe,
supports exploitation, IP subserves exploration, respectively, and they also receive dopaminergic
differing from the classical Go/NoGo view of projections from SNc. Each component of Iext is
BG. In a recent modeling study, it was shown uniquely connected to one neuron each in D1 and
that the exploitation (DP) versus exploration D2 layers. The dopamine signal (d) controls the
(IP) can be reconciled with Go (DP) versus sigmoidal gain of neurons in D1 and D2 layers.
NoGo (IP) view, by inserting a third regime Whereas the gain of D1 neurons increases with d,
dubbed the Explore regime. This regime would that of D2 neurons decreases with d. With such an
correspond to exploration and resides between arrangement, it can be noticed that the DP is
the classic Go and NoGo regimes (Kalva selected at higher dopamine levels and the IP at
et al. 2012). A series of BG models based on lower levels (Humphries and Gurney 2002;
this view have been developed to account for Humphries and Prescott 2010).
a wide variety of BG-related motor and cognitive
behaviors such as spatial navigation, saccades, The STN–GPe System
reaching, and reward–punishment learning STN and GPe form a loop with excitatory pro-
(Sridharan et al. 2006; Chakravarthy et al. 2010; jections from STN to GPe and inhibitory projec-
Krishnan et al. 2011; Kalva et al. 2012; tions in the reverse direction. Excitatory/
Priyadharsini et al. 2012; Gupta et al. 2013; inhibitory pairs of neuronal pools are known to
Muralidharan et al. 2013). exhibit limit cycle oscillations (Gillies
et al. 2002). The STN and GPe system receives
The Basic Model two inputs: the inhibitory (GABAergic) projec-
The intuitive ideas outlined before can be embod- tion from D2R MSNs of striatum to GPe and the
ied in a simple mathematical model of BG. The excitatory (glutamatergic) cortical input via the
framework explicitly represents striatum, STN, hyperdirect pathway to STN. Several factors
GPe and GPi, and DP and IP, to capture the elicit oscillations in a STN–GPe neuron pair.
structural aspects of BG, and also models: 1. Increased striatal input to GPe: This property
– The nigrostriatal dopamine signal as TD error, is corroborated by electrophysiological data
using it to train corticostriatal connections from Bergman and Wichmann et al. (1994).
– The action of dopamine in switching between Kravitz et al. (2010) observed that increased
DP and IP, via its differential action on the D1 firing of D2R-MSNs in the striatum induce
x1 x2 neuron uniquely connected bidirectionally to
a GPe neuron. Assume that both STN and GPe
W1 W2
ext
I1
ext
I2 have complete internal connectivity with all con-
D1 D2
nections within a nucleus having the same strength:
λD1 λD1 λD2 λD2 es for STN, eg for GPe; the one-to-one common B
#1 #2 #1 #2 Indirect
Dopamine pathway connection strength for STN ! GPe is wsg and
Striatum that of GPe ! STN is wgs. With (e = es = eg)
SNc
#1 #2 GPe and (w = wsg = wgs) and searching the (e , w)
space for finding the oscillatory behavior of the
STN–GPe system, simulations show that for strong
STN
connections between STN and GPe (w) and weak
#1 #2
Direct lateral connections (e), the network exhibits oscil-
pathway lations (Fig. 3) with poor correlation between the
STN–GPe neuronal activity (Fig. 4).
If the STN–GPe system is to serve as a source
#1 #2 GPi of exploration, a high spatiotemporal complexity
Inhibition
Excitation in STN activity manifested in the form of low
pair-wise correlations among neurons is
expected. The Figures 3 and 4 also illustrate the
#1 #2 Thal
ability of es to control correlation within STN and
show that es can be used to control the exploration
Thal
x1 x2
Thal in the BG model.
Basal Ganglia System as an Engine for Exploration, GPi

Fig. 2 The schematic of signal flow in the BG network GPi combines the GABAergic striatal output via
model DP with glutamatergic STN output from IP. There
is evidence that this combination of DP and IP
a state similar to Parkinson’s, with motor outflows in GPi is modulated by dopamine pro-
symptoms like freezing, bradykinesia, and dif- jections to GPi. When D1R in GPi, primarily
ficulty in movement initiation (Kravitz located on the GABAergic striato-pallidal axonal
et al. 2010). projections, is activated, firing levels of GPi neu-
2. Increasing cortical input to STN: Electrophys- rons are reduced (Kliem et al. 2007). Since D1Rs
iological studies show that ablation of cortical are activated at increased dopamine levels, the
areas that project to STN largely abolished facilitation of the DP outflow over IP at higher
low-frequency oscillations in STN–GPe dopamine levels is consistent with the nature of
(Magill et al. 2001). switching facilitated by dopamine in the striatum.
3. Reducing dopamine levels in STN–GPe:
Organotypic culture studies show that the Action Selection in Thalamus
STN–GPe system exhibits low-frequency If the primary function of the BG circuit is action
oscillations under dopamine-deficient condi- selection, where in the circuit is the precise site of
tions as in that of Parkinson’s disease (Plenz such selection? If action salience is computed in
and Kital 1999). STN and GPe oscillations the striatum and STN–GPe provides exploration,
seem to be triggered by the effect of dopamine action selection could be happening downstream
loss on D2R which strengthens the STN–GPe in GPi or in the thalamic nuclei receiving affer-
coupling (Steiner and Tseng 2010). ents from GPi. The competitive dynamics of neu-
Now consider a network model of the rons of thalamic reticular complex makes them
STN–GPe system in which STN and GPe layers ideally suited for implementing action selection
have equal number of neurons with each STN (Humphries and Gurney 2002). During binary
Basal Ganglia System as an Engine for Exploration, Fig. 3 The regions of parameter space (e, w) over which
probability of oscillations in (a) STN and (b) GPe is depicted (averaged over 50 trials)
Basal Ganglia System as an Engine for Exploration, given values of parameter pair (e, w), correlation is calcu-
Fig. 4 Correlations within (a) STN and (b) GPe over the lated only when there is at least 1 oscillating neuron in the
parameter space (e, w) averaged for 50 instances. For case of STN for (a) and GPe for (b)
action selection in the model, the GPi outputs to Simulation Experiments

thalamus converge on two neurons that represent Binary Action Selection
the two action alternatives. These two thalamic Salience-based action selection is considered to
neurons integrate the GPi inputs through time: the be one of the primary functions of BG (Redgrave
one that crosses a preset threshold first wins the et al. 1999; Gurney et al. 2001). Consider a binary
competition, while the second neuron is reset action selection problem. The corticostriatal
immediately. Accordingly, if xThal i (t) > xth for i with i = 1, 2, represents two possible
input, Iext
i (= 1, 2) at time t, then the states of all the actions, and the components of Iext i represent
other thalamic neurons immediately reset when action saliencies. The selected action is denoted
“i”th action being selected is expressed by xThal j (t) by the winning neuron in the thalamus. Due to the
= 0; j 6¼ i. If all xThal
i (t) fail to reach xth, no action complex dynamics of the STN–GPe system, it is
is selected. not necessary that the winning action is always
Basal Ganglia System as a 1 b
an Engine for 0.9
Exploration, Fig. 5 (a)
Probability of Selection
0.8
Probability of selection of EXPLORE
0.7 NoGo Go
Go/Explore/NoGo; (b) Go
NOGO
schematic: 0.6
EXPLORE
B
Go–Explore–NoGo (GEN) 0.5
regime selection for 0.4 δ
higher–intermediate–lower
0.3
values of d, respectively
0.2
0.1
0
–1 –0.5 0 0.5 1
δ
the one with greater saliency. There could be are (1) value computation in the striatum,
three possible outcomes: (2) feedback of previous action to the striatum,
“Go” – winning neuron has greater salience. and (3) resolving the nigrostriatal signal into two
“Explore” – winning neuron has lesser salience. dopamine signals – dTD and dV.
“NoGo” – no winner and therefore no action If x denotes the action selected (xi = 1, if the
selection. ith arm is selected; xj = 0 for j ~ = i), the action
Now consider the effect of d (dopamine) in value is computed as
determining the type of action selection: When
the network depicted in Fig. 2 is simulated with X
n
the STN–GPe system exhibiting uncorrelated V¼ wi xi (1)

i¼1
oscillations, a new Explore regime in addition to
the classical Go and NoGo regimes (Fig. 5b) can where wi denote the corticostriatal weights.
be observed. This is consistent with the classical Instantaneous output error, dTD, is defined as
picture of selecting the Go regime with high
probability for large d and the NoGo for small dTD ¼ r V (2)
d. In addition, the Explore regime is also selected
with the probability maximized for moderate d where r is the reward. Note that dTD in Eq. 2
(Fig. 5a). This Go/Explore/NoGo profile is called above is a special case of the more general tem-
the GEN profile. poral difference (TD) error (Eq. 3) in RL,
The notion that higher correlations (caused by
higher es) among STN neurons result in weaker dTD ¼ r ðtÞ þ gV ðt þ 1Þ V ðtÞ (3)
exploration is depicted by the GEN profiles of
Fig. 6a, b, c. Note the progressive reduction in the that arises when g = 0.
Explore regime with increasing es (Fig. 6). This dTD is used to update the corticostriatal
connections as
Modeling the N-Armed Bandit Problem
In the n-armed bandit problem, a generalization Dwi ¼ dTD xi (4)
of the binary action selection problem, the setup
consists of n slot machines each delivering a fixed In addition to dTD, dv is introduced as a similar
reward (deterministically or probabilistically) on yet novel quantity representing the temporal gra-
selection. The objective is to maximize the total dient of value function and is expressed as
reward received by the agent. Additional features
that need to be added to the binary action selec-
tion for simulating the n-armed bandit problem dV ðtÞ ¼ V ðtÞ V ðt 1Þ (5)
a 1 b 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
–1 –0.5 0 0.5 1 –1 –0.5 0 0.5 1
δ δ
c 1
0.9
0.8
0.7
0.6 Go
0.5 NOGO
EXPLORE
0.4
0.3
0.2
0.1
0
–1 –0.5 0 0.5 1
δ
Basal Ganglia System as an Engine for Exploration, Fig. 6 Probability of selection of Go/Explore/NoGo regimes
as a function of d: (a) for es = 0.001, (b) for es = 0.05, (c) for es = 0.95
The difference between TD error (Eqs. 2 and 3) thereby suppressing action (NoGo). For small
and value gradient (Eq. 5) is as follows. While magnitudes of dv, DP is still reduced, and driven
TD error controls learning of corticostriatal by the complex dynamics of STN, a random
weights (Eq. 4), value gradient controls the gain action is selected next time (Explore).
of D1R- and D2R-MSNs. D1R-MSNs are Figure 7a shows the change in value function
modeled to be activated at higher dV and with time for the above model applied to a five-
D2R-MSNs at lower levels. Thus, dV controls armed bandit problem. The rewards are generated
exploration by determining the relative contribu- by the following distribution: ri = i/n + A*n,
tions of DP or IP. where ri is the reward of the ith arm, n is
A large positive dv implies a large increase in a random variable uniformly distributed over
value, which therefore recommends selection of [0,1], and A = 0.3. Note the effect of the strength
the same action next time (Go), since the contri- of STN lateral connection, es, on the growth of
bution of DP to GPi dominates that of STN. A value function. For small es (= 0.0145), value
large negative dv implies a large reduction in increases rapidly but settles at a lower value.
value, suggesting exploration for new actions. For large es (= 0.0475), value rises slowly, but
Since the IP is selected for strong negative dv, for moderate es (= 0.0305), it rises slower than
IP contribution dominates that of DP at GPi, the previous case but settles at a level higher than
a 1 b 1
0.9 0.9
0.8 0.8
B
0.7 0.7
Average reward
Average reward
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
εs = 0.0145 ε = 0.01
0.2 εs = 0.0305 0.2 ε = 0.1
εs = 0.0475 ε=1
0.1 0.1
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Iterations Iterations
Basal Ganglia System as an Engine for Exploration, different e values using e-greedy policy, for a five-armed
Fig. 7 The mean reward versus iterations obtained with bandit task averaged for 500 trials
(a) different es using the BG network model and (b)
two previous cases. In this respect, the effect of es dopamine: (1) Go regime (“repeat the previous
on action selection seems to be analogous to that action”) for large positive dv, (2) Explore regime
of the e parameter in e-greedy methods used in (“try random actions”) for intermediate values of
RL (Fig. 7b). The parameter e, usually a small dv, and (3) NoGo regime (“no action”) for large
positive number, refers to the probability with negative values of dv. These regimes inspire
which a nonoptimal action is selected (Sutton a simple mechanism of hill-climbing in continu-
and Barto 1998). ous action spaces as follows.
Let V(x) be value function, x be n-dimensional
Climbing Value Gradient Using dV state vector (x S, where S Rn), dv = V(t) –
It can be observed from the network model used V(t 1), and Dx = x(t) x(t 1). Dx is then
for the n-armed bandit problem that the value updated by the following equations:
function increases gradually when the action
selected is iterated through the loop (thalamo- if ðdv > Dhi Þ
striatal) in Fig. 2. Thus, the network dynamics Dxðt þ 1Þ ¼ Dxðt }Go} ða
else ifðdv > Dlo ^ dv ðtÞ Dhi
implements hill-climbing over the value func-
Dxðt þ 1Þ ¼ f }Explore} ðb
tion, which is a form of stochastic hill-climbing, else ðdv Dlo Þ
thanks to the complex dynamics of STN–GPe Dxðt þ 1Þ ¼ 0 }NoGo} ðc
system. In fact, the aforementioned effect of dv
on value change (by selecting the previous/ (6)
random/no action for large positive/moderate/
large negative values of dv, respectively) where Dhi > 0, Dlo < 0, and F is a random vec-
is strongly reminiscent of simulated tor whose each component, Fi, is given by Eq. 7:
annealing – a form of stochastic optimization
(Kirkpatrick, Jr. et al. 1983). It is noted that the
BG model exhibits three behaviors depending on fi ¼ Gð0, 1Þexp dv 2 =s2 (7)
where G(0,1) is a Gaussian random variable with Discussion

mean 0 and standard deviation 1. The last Eq. 6c Application of RL concepts to BG function
seems redundant as there is no state update in that (Houk et al. 1995; Schultz et al. 1997; Hollerman
case. It can be altered by substituting 0 in Eq. 6c and Schultz 1998) sought a revision of the
with –Dx(t). Now in NoGo regime, the state x is Go/NoGo picture as the models may not have
updated in an opposite direction compared to the given sufficient attention to exploration –
previous update. The Eq. 6a, b, c may be the complementary process to exploitation. The
expressed in modified form as follows: Go/NoGo picture of BG can explain how the BG
circuit can learn simple binary action selection
if ðdv > Dhi Þ using RL, but it is inadequate to know how the
Dxðt þ 1Þ ¼ Dxðt }Go} ða
else ifðdv > Dlo ^ dv ðtÞ Dhi
BG circuit can solve more challenging RL prob-
Dxðt þ 1Þ ¼ f }Explore} ðb lems in continuous state and action spaces.
else ðdv Dlo Þ The binary Go/NoGo thinking about BG func-
Dxðt þ 1Þ ¼ Dxðt }NoGo} ðc
tion is supported by a simplistic interpretation of
the functional neurochemistry of the two BG
(8)
pathways (Albin et al. 1989; Contreras-Vidal
In Eq. 8 above, the Go, Explore, and NoGo and Stelmach 1995). But presence of the feed-
regimes are depicted as discrete, disjoint regimes back from STN to GPe allows the possibility of
demarcated by thresholds – Dhi and Dlo. But the complex dynamics in the STN–GPe loop, thereby
regimes observed in the GEN profile obtained introducing an added complication in our func-
from binary action selection problem are not dis- tional understanding of BG pathways. The
joint; their probability distributions as a function STN–GPe loop has also been dubbed as the
of dv overlap (Fig. 5). This inspires a combination “pacemaker” of BG considering its role in gener-
of Eq. 8a, b, c into a single update equation where ating pathological oscillations associated with
the three regimes smoothly overlap as follows: Parkinsonian tremor (Hurtado et al. 1999;
Terman et al. 2002). This STN–GPe system is

Dxðt þ 1Þ ¼ logsigðk3 ðdV ðt Dhi Dxðt an excitatory–inhibitory pair that is capable of
exhibiting oscillations and other forms of com-
þ cexpð dV ðtÞ=sE
2 2
plex dynamics (Brunel 2000). The fact that neu-
logsigðk4 ðdV ðtÞ Dlo Dxðt rons in this system exhibit uncorrelated firing
patterns in normal conditions, and highly corre-
(9) lated and synchronized firing under dopamine-
deficient pathological conditions, seems to offer
where logsig(x) = 1/(1 + exp(x)). Thus, Eq. 9 an important clue to the possible role of this
describes a map between the current state update circuit in exploration. From the above studies, it
(Dx(t)) to the next state update (Dx(t + 1)). Dx can be concluded that the STN–GPe system is in
(t + 1) is nearly in the same direction as Dx(t) for the best position to serve as an explorer, thereby
large positive dV and is nearly in the same direc- supplying the missing piece in the RL-machinery
tion as Dx(t) for large negative dV; Dx(t + 1) is of BG. Behavioral strategies of reaching, sac-
random for the intermediate values of dV. Note cades, spatial navigation, gait, and willed action
that the map of Eq. 9 is stochastic due to the are shown to be better modeled using the above
middle term on the right-hand side. theory as follows.
Equation 9, known as the GEN rule, represents A model of reaching movements highlighting
an abstract, summarized representation of how the role of BG was described in Magdoom and
BG selects actions based on DA signals. The Subramanian et al. (2011), where a neural net-
GEN rule or the approach that it is based on has work representing motor cortex is trained to drive
been applied successfully to model a range of BG a two-joint arm to a target. The output of the BG
functions. that is predominant in early stages of learning is
combined with that of motor cortex whose rela- dip under control conditions and exhibited
tive contribution grows with learning. The BG a significant reduction in velocity close to the
dynamics governed by the earlier described doorway under PD conditions.
GEN rule discovers desired activations which Clinical literature shows that BG has a role in
are used by the motor cortex for learning. When willed action, and its impairment is seen in con- B
the dopamine signal is clamped to reflect dopa- ditions of BG lesions or diseases like Parkinson’s
mine deficiency in Parkinsonian conditions, the disease that affect BG (Mink 2003). Recently it
model exhibited Parkinsonian features in was suggested that the BG circuit amplifies
reaching like bradykinesia, undershoot, and will signals, presumably weak, by a stochastic
tremor. resonance process (Chakravarthy 2013). This
The idea that the DP and IP subserve exploita- study shows that the GEN policy, which is
tion and exploration respectively was used in a combination of deterministic hill-climbing pro-
a model of BG on saccade generation (Krishnan cess and a stochastic process, may be reinterpreted
et al. 2011) when applied to standard visual as a form of stochastic resonance. Applying the
search tasks like feature and conjunction search, model to a simple reaching task, it was shown that
directional saccades, and sequential saccades. the arm reaches the target with probability close to
On simulating Parkinsonian conditions by dimin- unity at optimal noise levels. The arm dynamics
ished BG output, the model exhibited impaired for subthreshold noise is reminiscent of Parkinso-
visual search with longer reaction times – nian akinesia, whereas for superthreshold noise
a characteristic symptom in Parkinson’s disease the arm shows uncontrolled movements resem-
(PD) patients. bling Parkinsonian dyskinesias.
The GEN approach (Eqs. 6, 8, and 9) was used A perspective that the BG circuit is an explo-
to model the relative contributions of BG and ration engine, carved out of the popular RL
hippocampus to spatial navigation (Sukumar approach to BG modeling, can thus be
et al. 2012). The model combines two naviga- substantiated.
tional system: the cue-based system subserved
by BG and the place-based system by hippocam-
pus. The two navigational systems are associated References
with their respective value functions that are
Albin RL, Young AB et al (1989) The functional anatomy
combined by the softmax policy (Sutton and of basal ganglia disorders. Trends Neurosci
Barto. 1998) to select the next move. This 12(10):366–375
model describes the results of an experimental Alexander GE, Crutcher MD (1990) Functional architec-
study that investigates competition between ture of basal ganglia circuits: neural substrates of par-
allel processing. Trends Neurosci 13(7):266–271
cue-based and place-based navigation (Devan Baunez C, Humby T et al (2001) Effects of STN lesions on
and White 1999). Under dopamine-deficient con- simple vs choice reaction time tasks in the rat: pre-
ditions, the model exhibited longer escape laten- served motor readiness, but impaired response selec-
cies similar to PD-model rats (Miyoshi tion. Eur J Neurosci 13(8):1609–1616
Bergman H, Wichmann T et al (1994) The primate
et al. 2002). subthalamic nucleus. II. Neuronal activity in the
The GEN approach to BG was also applied to MPTP model of parkinsonism. J Neurophysiol
model impaired gait patterns in PD 72(2):507–520
(Muralidharan et al. 2013). Studies by Cowie Brown P, Oliviero A et al (2001) Dopamine dependency
of oscillations between subthalamic nucleus and
and Limousin et al. (2010) investigated gait pallidum in Parkinson’s disease. J Neurosci
changes, as PD patients walked through 21(3):1033–1038
a narrow doorway and observed a strong dip in Brunel N (2000) Dynamics of sparsely connected net-
velocity a short distance from the doorway. In the works of excitatory and inhibitory spiking neurons.
J Comput Neurosci 8(3):183–208
model of Muralidharan and Balasubramani Chakravarthy VS (2013) Do basal ganglia amplify willed
et al. (2013), the simulated agent passed through action by stochastic resonance? A model. PLoS One
the doorway without any significant velocity 8(11):e75657
Chakravarthy VS, Joseph D et al (2010) What do the basal basal ganglia outflow in monkeys. J Neurophysiol
ganglia do? A modeling perspective. Biol Cybern 98(3):1489–1500
103(3):237–253 Knutson B, Adams CM et al (2001) Anticipation of
Contreras-Vidal J, Stelmach GE (1995) Effects of Parkin- increasing monetary reward selectively recruits
sonism on motor control. Life Sci 58(3):165–176 nucleus accumbens. J Neurosci 21(16):RC159
Cowie D, Limousin P et al (2010) Insights into the neural Kravitz AV, Freeze BS et al (2010) Regulation of Parkin-
control of locomotion from walking through doorways sonian motor behaviours by optogenetic control of
in Parkinson’s disease. Neuropsychologia basal ganglia circuitry. Nature 466(7306):622–626
48(9):2750–2757 Krishnan R, Ratnadurai S et al (2011) Modeling the role of
Daw ND, O’Doherty JP et al (2006) Cortical substrates for basal ganglia in saccade generation: is the indirect
exploratory decisions in humans. Nature 441(7095): pathway the explorer? Neural Netw 24(8):801–813
876–879 Magdoom KN, Subramanian D et al (2011) Modeling
DeLong MR (1990) Primate models of movement disor- basal ganglia for understanding Parkinsonian reaching
ders of basal ganglia origin. Trends Neurosci movements. Neural Comput 23(2):477–516
13(7):281–285 Magill P, Bolam J et al (2001) Dopamine regulates the
Devan BD, White NM (1999) Parallel information impact of the cerebral cortex on the subthalamic
processing in the dorsal striatum: relation to hippo- nucleus–globus pallidus network. Neuroscience
campal function. J Neurosci 19(7):2789–2798 106(2):313–330
Doya K (2002) Metalearning and neuromodulation. Neu- Marsden C (1982) The mysterious motor function of the
ral Netw 15(4–6):495–506 basal ganglia: the Robert Wartenberg lecture. Neurol-
Gillies A, Willshaw D et al (2002) Functional interactions ogy 32:514–539
within the subthalamic nucleus. The basal ganglia Mink JW (2003) The basal ganglia and involuntary move-
VII. Springer, New York. pp 359–368 ments: impaired inhibition of competing motor pat-
Grabli D, McCairn K et al (2004) Behavioural disorders terns. Arc Neurol 60(10):1365
induced by external globus pallidus dysfunction in pri- Miyoshi E, Wietzikoski S et al (2002) Impaired learning in
mates: I. Behavioural study. Brain 127(9):2039–2054 a spatial working memory version and in a cued ver-
Gupta A, Balasubramani PP et al (2013) Computational sion of the water maze in rats with MPTP-induced
model of precision grip in Parkinson’s disease: a utility mesencephalic dopaminergic lesions. Brain Res Bull
based approach. Front Comput Neurosci 7:172 58(1):41–47
Gurney K, Prescott TJ et al (2001) A computational model Muralidharan V, Balasubramani PP et al (2013)
of action selection in the basal ganglia. I. A new A computational model of altered gait patterns in
functional anatomy. Biol Cybern 84(6):401–410 parkinson’s disease patients negotiating narrow door-
Hollerman JR, Schultz W (1998) Dopamine neurons ways. Front Comput Neurosci 7:190
report an error in the temporal prediction of reward O’Doherty J, Dayan P et al (2004) Dissociable roles of
during learning. Nat Neurosci 1(4):304–309 ventral and dorsal striatum in instrumental condition-
Houk JC, Davis JL et al (1995) Models of information ing. Science 304(5669):452–454
processing in the basal ganglia. The MIT press, Plenz D, Kital ST (1999) A basal ganglia pacemaker
Cambridge, MA formed by the subthalamic nucleus and external globus
Humphries M, Gurney K (2002) The role of intra-thalamic pallidus. Nature 400(6745):677–682
and thalamocortical circuits in action selection. Netw Priyadharsini BP, Ravindran B et al (2012) Understanding
Comput Neural Syst 13(1):131–156 the role of serotonin in basal ganglia through a unified
Humphries MD, Prescott TJ (2010) The ventral basal model. Artificial neural networks and machine
ganglia, a selection mechanism at the crossroads of learning–ICANN 2012, Springer, Berlin, pp 467–473
space, strategy, and reward. Prog Neurobiol Redgrave P, Prescott TJ et al (1999) The basal ganglia:
90(4):385–417 a vertebrate solution to the selection problem? Neuro-
Hurtado JM, Gray CM et al (1999) Dynamics of tremor- science 89(4):1009–1023
related oscillations in the human globus pallidus: Ring H, Serra-Mestres J (2002) Neuropsychiatry of the
a single case study. Proc Natl Acad Sci basal ganglia. J Neurol Neurosurg Psychiatry
96(4):1674–1679 72(1):12–21
Joel D, Niv Y et al (2002) Actor-critic models of the basal Rushworth MF, Behrens TE (2008) Choice, uncertainty
ganglia: new anatomical and computational perspec- and value in prefrontal and cingulate cortex. Nat
tives. Neural Netw 15(4–6):535–547 Neurosci 11(4):389–397
Kalva SK, Rengaswamy M et al (2012) On the neural Russell V, Allin R et al (1992) Regional distribution of
substrates for exploratory dynamics in basal ganglia: monoamines and dopamine D1-and D2-receptors in
a model. Neural Netw 32:65–73 the striatum of the rat. Neurochem Res 17(4):387–395
Kirkpatrick S, Gelatt CD Jr et al (1983) Optimization by Schultz W, Dayan P et al (1997) A neural substrate of
simulated annealing. Science 220(4598):671–680 prediction and reward. Science 275(5306):1593–1599
Kliem MA, Maidment NT et al (2007) Activation of nigral Sridharan D, Prashanth PS et al (2006) The role of the
and pallidal dopamine D1-like receptors modulates basal ganglia in exploration in a neural model based on
Basal Ganglia: Beta Oscillations 327 B
reinforcement learning. Int J Neural Syst
16(2):111–124
Stein PS, Grillner S et al (1997) Neurons, networks, and
behavior. MIT Press, Cambridge, MA
Steiner H, Tseng KY (2010) Handbook of basal ganglia
structure and function: a decade of progress. Access B
online via Elsevier. Academic press, San Diego
Sukumar D, Rengaswamy M et al (2012) Modeling the Basal Ganglia: Beta Oscillations, Fig. 1 One second of
contributions of basal ganglia and hippocampus to local field potentials recorded from internal segment of
spatial navigation using reinforcement learning. globus pallidus (top panel) and subthalamic nucleus
PLoS One 7(10):e47467 (bottom panel) from a patient with Parkinson’s disease
Sutton R, Barto A (1998) Reinforcement learning: an who was off medications (Reproduced from Brown et al.
introduction. Adaptive computations and machine (2001) with permission)
learning. MIT Press/Bradford, Cambridge, MA
Terman D, Rubin J et al (2002) Activity patterns in
a model for the subthalamopallidal network of the suggested that beta oscillations are related to
basal ganglia. J Neurosci 22(7):2963–2976 maintaining the current body posture (Engel and
Usher M, Cohen JD et al (1999) The role of locus
coeruleus in the regulation of cognitive performance. Fries 2010). The causal influence of beta oscilla-
Science 283(5401):549–554 tions on inhibition of movement is suggested by
Willshaw D, Li Z (2002) Subthalamic–pallidal interac- observations that stimulation at beta frequency
tions are critical in determining normal and abnormal with transcranial magnetic stimulation
functioning of the basal ganglia. Proc R Soc Lond Ser
B Biol Sci 269(1491):545–551 (Pogosyan et al. 2009) and with deep-brain stim-
Yoshida W, Ishii S (2006) Resolution of uncertainty in ulation (Chen et al. 2007) results in slower
prefrontal cortex. Neuron 50(5):781–789 movements.
Relationship with Symptoms

The relationship between beta oscillations and
Basal Ganglia: Beta Oscillations inhibition of movements discussed above sug-
gests that the increased beta power in Parkinson’s
Rafal Bogacz disease may directly relate to the impaired move-
Department of Computer Science, University of ments observed in the disease (Brown 2007). This
Bristol, Bristol, UK hypothesis is supported by the observation that
the power of beta oscillations is reduced by the
treatments that alleviate symptoms, namely,
Definition dopaminergic medications and high-frequency
(>100 Hz) deep-brain stimulation of subthalamic
Beta oscillations are defined as the oscillations in nucleus (STN) (Hammond et al. 2007). Further-
neural activity with frequencies between 13 and more, the patients with the best improvement in
30 Hz (Fig. 1). In the basal ganglia, the power of motor skills due to medications also have the
beta oscillations is significantly increased in largest reduction in beta power in STN after tak-
Parkinson’s disease (Hammond et al. 2007). ing the medications (Kuhn et al. 2006).
Models of the Effects of Beta Oscillations

Detailed Description Simulations demonstrated that the exaggerated
oscillatory activity leads to movement difficulties
Relationship to Behavior possibly because the thalamic neurons (see
The power of beta oscillations is reduced prior to Fig. 2) become entrained to the excessively oscil-
and during movements in both the cortex and latory input from basal ganglia and are no longer
basal ganglia (see Engel and Fries 2010 for able to transmit sensorimotor signals from the
review) and increased during tonic contractions cortex (Rubin and Terman 2004). Additionally,
(Baker et al. 1997). Consequently, it has been an analysis of neural activity in the external
B 328 Basal Ganglia: Beta Oscillations
activity, occurring in Parkinson’s disease, trigger

beta oscillations. Furthermore, due to the com-
plex connectivity in the cortico-basal-ganglia-
thalamic circuit (Fig. 2), it is still unclear in
which part of the circuit the beta oscillations are
generated. The circuit contains many feedback
loops, and it is known that neuronal populations
connected in loops can generate oscillations
(Tiesinga and Sejnowski 2009).
One of the parts of the circuit which is thought
to be critical for generation of Parkinsonian beta
oscillations is a sub-circuit composed of STN and
GPe (Bevan et al. 2002). Its role is suggested by
observations that the beta oscillations are promi-
nent in this network (Mallet et al. 2008b), this
sub-circuit was shown to be able to produce
slower (delta) oscillations in vitro (Plenz and
Kital 1999), and blocking connections between
Basal Ganglia: Beta Oscillations, Fig. 2 Subset of STN and GPe abolishes excessive beta oscilla-
connectivity in the cortico-basal-ganglia-thalamic circuit.
tions (Tachibana et al. 2011). Additionally, inter-
Black rectangles denote brain structures, and the rectangle
labeled “Output” denotes the output nuclei of the basal actions between the STN-GPe network and
ganglia comprised of the internal segment of globus cortex are likely to play important role in gener-
pallidus and substantia nigra pars reticulata. Green arrows ating Parkinsonian beta, because the cortex and
and red lines with circles denote excitatory and inhibitory
STN become coherent in the disease (Mallet et al.
connections, respectively. Blue dashed contours indicate
sub-circuits that have been shown to generate beta oscil- 2008b) and blocking connections between the
lations in simulations cortex and STN abolishes excessive beta oscilla-
tions (Tachibana et al. 2011).
segment of globus pallidus (GPe) using informa- Computational Models of Beta Generation
tion theory showed reduced entropy during beta Three sub-circuits containing feedback loops
oscillations in rats, suggesting that the beta (indicated by blue contours in Fig. 2) have been
oscillations limit the amount of information shown to generate beta oscillation in simulations.
transmitted by the neurons in the basal ganglia First, models of the STN-GPe network were
(Cruz et al. 2009). shown to generate beta oscillations (a popula-
tion-level model: Nevado-Holgado et al. 2010;
Origin of Beta Oscillations and a model using integrate-and-fire neurons:
Parkinson’s disease is caused by the death of Kumar et al. 2011). Mathematical analyses of
dopaminergic neurons which project throughout the population-level model revealed conditions
the basal ganglia and modulate synaptic transmis- the parameters of the STN-GPe model need to
sion and plasticity. However, the exaggerated satisfy to generate oscillations (Nevado-Holgado
beta oscillations do not appear immediately et al. 2010; Pavlides et al. 2012; Passillas-Lepine
after the death of these neurons, but only several 2013). Second, a computational model of stria-
days later (Mallet et al. 2008a). This suggests that tum (using Hodgkin-Huxley neurons: McCarty
the beta oscillations result from the adaptations in et al. 2011) has been shown to generate beta
the network following the reduced dopamine oscillations. Third, the beta oscillations have
level. been also observed in simulations of the entire
It is still uncertain which of the changes in cortico-basal-ganglia circuit (a population-level
the connectivity and the levels of neuronal model: van Albada et al. 2009), and the analysis
Basal Ganglia: Beta Oscillations 329 B
of the model suggested that the beta oscillations Brown P (2007) Abnormal oscillatory synchronization in
originate in the corticothalamic loop in this the motor system leads to impaired movement. Curr
Opin Neurol 17:656–664
model. Brown P, Oliviero A, Mazzone P, Insola A, Tonali P, Di
Additionally, beta oscillations have been also Lazzaro V (2001) Dopamine dependency of oscilla-
observed in simulations of the entire cortico- tions between subthalamic nucleus and pallidum in B
basal-ganglia circuit (using integrate-and-fire Parkinson’s disease. J Neurosci 21:1033–1038
Chen CC, Litvak V, Gilbertson T, Kuhn A, Lu CS, Lee ST,
neurons: Humphries et al. 2006). Furthermore, Tsai CH, Tisch S, Limousin P, Hariz M, Brown
oscillations close to the beta range (12 Hz) P (2007) Excessive synchronization of basal ganglia
were generated in a model of a subset of neurons in 20 Hz slows movement in Parkinson’s
cortico-basal-ganglia circuit (in which firing disease. Exp Neurol 2015:214–221
Cruz AV, Mallet N, Magill PJ, Brown P, Averbeck BB
rates of individual neurons were described: (2009) Effects of dopamine depletion on network
Leblois et al. 2006). The analysis of the model entropy in the external globus pallidus.
suggested the oscillations were generated in J Neurophysiol 102:1092–1102
a loop composed of cortex, STN, output nuclei, Engel AE, Fries P (2010) Beta-band
oscillations – signalling the status quo? Curr Opin
and thalamus in this model. Neurobiol 20:156–165
Eusebio A, Pogosyan A, Wang S, Averbeck B, Gaynor
Model-Based Data Analysis LD, Cantiniaux S, Witjas T, Limousin P, Azulay J-P,
Computational models have been used to infer Brown P (2009) Resonance in subthalamo-cortical
circuits in Parkinson’s disease. Brain 132:2139–2150
the changes in network structure occurring in Hammond C, Bergman H, Brown P (2007) Pathological
Parkinson’s disease. Eusebio et al. (2009) fitted synchronization in Parkinson’s disease: networks,
a simple damped oscillator to event-related models and treatments. Trends Neurosci 30:357–464
potentials recorded from the cortex of Humphries MD, Stewart RD, Gurney KN
(2006) A physiologically plausible model of action
Parkinson’s patients after stimulation of STN. selection and oscillatory activity in the basal ganglia.
They observed that the fitted oscillator had J Neurosci 26:12921–12942
a natural frequency in the beta range, and the Kuhn AA, Kupsch A, Schneider GH, Brown P (2006)
data from patients on medications were described Reduction in subthalamic 8–35 Hz oscillatory activity
correlates with clinical improvement in PD. Eur
by a model with higher damping parameter than J Neurosci 23:1956–1960
data from patients off medications. Also, the Kumar A, Cardanobile S, Rotter S, Aertsen A (2011) The
dynamic causal modeling was used to infer from role of inhibition in generating and controlling
local field potentials the differences in effective Parkinson’s disease oscillations in the basal ganglia.
Front Syst Neurosci 5:86
connectivity in cortico-basal-ganglia-thalamic Leblois A, Boraud T, Meissner W, Bergman H, Hansel
circuit between healthy and Parkinsonian rats D (2006) Competition between feedback loops under-
(Moran et al. 2011), and between patients on lies normal and pathological dynamics in the basal
and off medications (Marreiros et al. 2013). ganglia. J Neurosci 26:3567–3583
Mallet N, Pogosyan A, Sharott A, Csicsvari J, Bolam JP,
Both of these studies reported increased effective Brown P, Magill PJ (2008a) Disrupted dopamine
connectivity between the cortex and STN in the transmission and the emergence of exaggerated beta
states associated with higher beta power. oscillations in subthalamic nucleus and cerebral cor-
tex. J Neurosci 28:4795–4806
Mallet N, Pogosyan A, Marton LF, Bolam JP, Brown P,
Magill PJ (2008b) Parkinsonian oscillations in the
References external globus pallidus and their relationship with
subthalamic nucleus activity. J Neurosci
Baker SN, Olivier E, Lemon RN (1997) Coherent oscilla- 28:14245–14258
tions in monkey motor cortex and hand muscle EMG Marreiros AC, Cagan H, Moran RJ, Friston KJ, Brown P
show task-dependent modulation. J Physiol (2013) Basal ganglia-cortical interactions in Parkinso-
501(1):225–241 nian patients. Neuroimage 66:301–310
Bevan MD, Magill PJ, Terman D, Bolam JP, Wilson CJ McCarty MM, Moore-Kochlacs C, Gu X, Boyden ES, Han
(2002) Move to the rhythm: oscillation in the X, Kopell N (2011) Striatal origin of the pathologic
subthalamic nucleus – external globus pallidus net- beta oscillations in Parkinson’s disease. Proc Natl
work. Trends Neurosci 25:525–531 Acad Sci U S A 108:11620–11625
B 330 Basal Ganglia: Control of Saccades
Moran R, Mallet N, Litvak V, Dolan RJ, Magill PJ, Friston these selections towards the most profitable
KJ, Brown P (2011) Alteration in brain connectivity options. One of these loops appears to be special-
underlying beta oscillations in Parkinsonism. PLoS
Comput Biol 7:e1002124 ized in saccadic eye movements. Most of the
Nevado-Holgado AJ, Terry J, Bogacz R (2010) Condi- existing computational models consider that it is
tions for the generation of beta oscillations in the responsible for the choice, among all possible
subthalamic nucleus – globus pallidus network. targets, of the target of the next saccade.
J Neurosci 30:12340–12352
Passillas-Lepine W (2013) Delay-induced oscillations in
Wilson and Cowan’s model: an analysis of the
subthalamo-pallidal feedback loop in healthy and Par- Detailed Description
kinsonian subjects. Biol Cybernet 107: 289–308
Pavlides A, Hogan SJ, Bogacz R (2012) Improved condi-
tions for the generation of beta oscillations in the Basal Ganglia Saccadic Circuitry
subthalamic nucleus-globus pallidus network. Eur The basal ganglia are a set of subcortical nuclei
J Neurosci 36:2229–2239 common to all vertebrates (see “▶ Basal Ganglia:
Plenz D, Kital S (1999) A basal ganglia pacemaker formed Overview”). The basal ganglia are components of
by the subthalamic nucleus and external globus
pallidus. Nature 400:677–682 two main types of loops: cortico-baso-thalamo-
Pogosyan A, Gaynor LD, Eusebio A, Brown P (2009) cortical loops (Alexander et al. 1986) as well as
Boosting cortical activity at beta-band frequencies subcortical ones (McHaffie et al. 2005). Within
slows movement in humans. Curr Biol 19:1637–1641 these two categories, sub-loops can be character-
Rubin JE, Terman D (2004) High frequency stimulation of
the subthalamic nucleus eliminates pathological tha- ized, dedicated to various functions; one cortical
lamic rhythmicity in a computational model. J Comput loop dedicated to eye movements has been iden-
Neurosci 16:211–235 tified, as well as a few subcortical ones, all
Tachibana Y, Iwamuro H, Kita H, Takada M, Nambu projecting to the superior colliculus, where infor-
A (2011) Subthalamo-pallidal interactions underlying
parkinsonian neuronal oscillations in the primate basal mation converges to allow the generation of eye
ganglia. Eur J Neurosci 34:1470–1484 movements. The cortical loop receives contribu-
Tiesinga P, Sejnowski TJ (2009) Cortical enlightenment: tions from the frontal eye fields, the supplemen-
are attentional gamma oscillations driven by ING or tary eye fields, the dorsolateral prefrontal cortex,
PING? Neuron 63:727–732
Van Albada SJ, Gray RT, Drysdale PM, Robinson PA and the parietal cortex and goes through the ven-
(2009) Mean-field modeling of the basal ganglia- tral anterior, ventrolateral, and medial dorsal tha-
thalamocortical system II. Dynamics of Parkinsonian lamic nuclei. The subcortical ones all receive
oscillations. J Theor Biol 257:664–688 inputs from the superior colliculus via the lateral
posterior, the intralaminar, and the pulvinar
nuclei of the thalamus.
The outputs of the basal ganglia are inhibitory
Basal Ganglia: Control of Saccades and tonically active, so that at rest, they maintain
their targets under constant inhibition, preventing
Benoı̂t Girard their activation. The generic role of the basal
Sorbonne Universités, UPMC Univ Paris 06, ganglia is understood as an action selection sys-
Paris, France tem, used to solve resource allocation problems
CNRS, Paris, France (Mink 1996; Redgrave et al. 1999): competing
channels within the basal ganglia represent the
different potential actions, and when simulta-
Definition neous inputs activate different channels, the
basal ganglia resolve the competition and disin-
The basal ganglia are a set of subcortical hibit only one of the outputs, and thus allow the
interconnected nuclei, subdivided in parallel cir- activation of a single action. In the context of
cuits. They form loops with specific cortical and saccadic eye movements, at a given moment,
subcortical regions and play a central role in the the activities caused by multiple points of interest
selection of actions as well as in learning to bias present in the visual field are susceptible of
Basal Ganglia: Control of Saccades 331 B
eliciting multiple saccades, which would allow and the specific mechanisms of the basal ganglia
for the refined analysis by the fovea of these saccadic circuit has been reached. All of them
points of interest. The role of the basal ganglia have been built at the same level of description,
is thus understood as filtering these multiple using variations of leaky-integrator models of
solicitations in order to keep only the one with neurons. B
the highest priority. Nevertheless, all of these models consider that
The basal ganglia also interact with dopami- the neurons that are involved in saccades towards
nergic nuclei (the ventral tegmental area and the visible targets belong to a circuit in charge of
substantia nigra pars compacta): they project to selecting the target of the next saccadic move-
them and receive modulatory projections at ment in a winner-takes-all manner, in line with
the striatal, pallidal, and subthalamic levels. The the current understanding of the general basal
current theory proposes that these interactions are ganglia function. Those models are based on tra-
the neural substrate of reinforcement learning. ditional basal ganglia computational models,
They would allow learning, by trial and error, of adapted to saccade selection: each channel in
which actions are the most profitable and would competition codes for a given saccade metric
bias the selection process accordingly. The dopa- and the selected channel disinhibits the area of
minergic signal is thought to correspond to the superior colliculus specifically encoding that
a measurement of the reward prediction error, metric. The common selection mechanism
which is used to modify the strength of the among those models is based on the focused
cortico-striatal synapses according to temporal- inhibition of the striatum, by which each channel
difference learning algorithms (see Reward- tries to inhibit its output, and the diffuse excita-
Based Learning, Model-Based and Model-Free). tion of the subthalamic nucleus, used to excite the
In the saccadic eye movement context, this would competitors.
be the substrate of the evaluation of the priority of The models which also incorporate reinforce-
the competing targets, allowing the brain to make ment learning (Dominey et al. 1995; Brown
the best decision according to experience. et al. 2004; N’Guyen et al. 2010, 2014; Guthrie
Electrophysiological recording in monkeys et al. 2013) rely on variations of classical
showed that saccade-related activity in the basal temporal-difference reinforcement learning algo-
ganglia of monkey is quite rich (Hikosaka rithms (see “▶ Reinforcement Learning in Corti-
et al. 2000). Input neurons in the striatum exhibit cal Networks”). In these models, derived from
visual and motor saccade-related activity, which machine learning theories, the dopamine released
is also reflected in the output activity of the in the striatum is supposed to correspond to the
substantia nigra. But more interestingly, one reward prediction error signal used by these algo-
third of the neurons have a working memory- rithms, it drives the modifications of the strength
specific activity: some of them have visual or of the cortico-striatal synapses.
motor burst activity only for memory-driven sac- In a few models, a second loop parallel to the
cades, while the rest have sustained activity while “visual” one is dedicated to working memory
the position of a target is kept in working (Dominey and Arbib 2012). It is supposed to
memory. select which of the current potential targets have
to be stored, in order to allow for saccades
Computational Models of the Basal Ganglia towards targets that may disappear from the
Saccadic Circuitry visual field. Such a functionality requires a
Relatively few computational models of the basal remapping capability: as the frontal eye field,
ganglia have specifically targeted the saccadic parietal cortex, or superior colliculus maps,
circuitry (Dominey and Arbib 2012; Dominey which are supposed to store these memories, are
et al. 1995; Brown et al. 2004; Chambers retinotopic, the positions of the next saccades
et al. 2005; N’Guyen et al. 2010, 2014; Guthrie have to be updated after each eye movement.
et al. 2013), thus no final consensus about the role This mechanism can be extended to store the
B 332 Basal Ganglia: Decision-Making
order of the targets, allowing for a more elabo- Dominey P, Arbib M, Joseph JP (1995) A model of
rated sequence generation capability. Note that corticostriatal plasticity for learning oculomotor
associations and sequences. J Cogn Neurosci
models of the role of the basal ganglia in working 7(3):311–336
memory have been proposed in other contexts Guthrie M, Leblois A, Garenne A, Boraud T (2013) Inter-
and could easily be adapted to the saccadic func- action between cognitive and motor cortico-basal
tion (see “▶ Working Memory, Models of”). ganglia loops during decision making: a computational
study. J Neurophysiol 109(12):3025–3040
Concerning the subcortical basal ganglia Hikosaka O, Takikawa Y, Kawagoe R (2000) Role of the
loops through the superior colliculus, no model basal ganglia in the control of purposive saccadic eye
has ever considered the cohabitation and the pos- movements. Physiol Rev 80(3):953–978
sible interactions of the three anatomically McHaffie JG, Stanford TR, Stein BE, Coizet V, Redgrave
P (2005) Subcortical loops through the basal ganglia.
described circuits (McHaffie et al. 2005). Those TINS 28(8):401–407
which considered the basal ganglia-superior Mink JW (1996) The basal ganglia: focused selection and
colliculus loops included only one of them. inhibition of competing motor programs. Prog
They either used it in a slave mode (Brown Neurobiol 50:381–425
N’Guyen S, Pirim P, Meyer JA, Girard B (2010) An
et al. 2004), where it executes the motor decisions integrated neuromimetic model of the saccadic
taken by a master cortical loop or as an additional eye movements for the psikharpax robot. In: From
player operating in parallel to the cortical loops animals to animats 11. Springer, Berlin/Heidelberg,
(Chambers et al. 2005; N’Guyen et al. 2010, pp 114–125
N’Guyen S, Thurat C, Girard B (2014) Saccade learning
2014). with concurrent cortical and subcortical basal ganglia
loops. Frontiers in Computational Neuroscience.
8(00048) doi: 10.3389/fncom.2014.00048
Cross-References Redgrave P, Prescott TJ, Gurney K (1999) The basal
ganglia: a vertebrate solution to the selection problem?
Neuroscience 89:1009–1023
▶ Basal Ganglia: Mechanisms for Action
Selection Further Reading
▶ Basal Ganglia: Overview Girard B, Berthoz A (2005) From brainstem to cortex:
▶ Dopaminergic Cell Models computational models of saccade generation circuitry.
▶ Integrate and Fire Models, Deterministic Prog Neurobiol 77:215–251
▶ Reinforcement Learning in Cortical Networks
▶ Working Memory, Models of
Basal Ganglia: Decision-Making

References
Wei Wei
Alexander GE, DeLong MR, Strick PL (1986) Parallel Center for Neural Science, New York University,
organization of functionally segregated circuits New York, NY, USA
linking basal ganglia and cortex. Annu Rev Neurosci Department of Neurobiology and Kavli Institute
9:357–381
Brown JW, Bullock D, Grossberg S (2004) How laminar
for Neuroscience, Yale University School of
frontal cortex and basal ganglia circuits interact to Medicine, New Haven, CT, USA
control planned and reactive saccades. Neural Netw
17(4):471–510
Chambers JM, Gurney K, Humphries M, Prescott TJ
(2005) Mechanisms of choice in the primate brain:
Definition
a quick look at positive feedback. In: Bryson JJ, Pres-
cott TJ, Seth AK (eds) Modelling natural action selec- The basal ganglia (BG) participate not only in the
tion: proceedings of an international workshop. AISB, selection of motor plans but also in perceptual
Sussex, pp 45–52
Dominey P, Arbib M (2012) A cortico-subcortical model
decision-making. The functional structure of the
for generation of spatially accurate sequential sac- BG and their close interconnections with the cor-
cades. Cereb Cortex 2(2):153–175 tex and dopamine system allow the BG to be
Basal Ganglia: Decision-Making 333 B
actively involved in the perceptual decision-
making processes.
Detailed Description B
The role of the BG in perceptual decision-making
processes has recently attracted much attention
(Ding and Gold 2013). Although the BG were
much earlier proposed to be the central substrate
for action selection and habit learning (Graybiel
1995; Mink 1996; Redgrave et al. 2010), their
active participation in perceptual decision-
making has been investigated only more recently
(Ding and Gold 2010). The BG have close inter-
Basal Ganglia: Decision-Making, Fig. 1 Scheme of the
action with the frontal cortex and the lateral basal ganglia circuit and connections with the cortex, the
intraparietal area (LIP), where are believed to be thalamus, and the superior colliculus (SC). The direct,
the sites of evidence accumulation in decision indirect, and hyper-direct pathways are indicated. GPe
processes (Gold and Shadlen 2007; Huk and external segments of the globus pallidus (GP), GPi inter-
nal segments of the GP, STN subthalamic nucleus, SNr
Shadlen 2005; Kim and Shadlen 1999). The BG substantia nigra pars reticulate, LIP lateral intraparietal
also have prominent interaction with the down- area, FEF frontal eye field, pSMA pre-supplementary
stream subcortical motor-related areas, such as motor area, ACC anterior cingulate cortex
the thalamus and the superior colliculus (SC),
and the midbrain dopamine system (Hikosaka has been observed for human subjects (Frank
et al. 2000). As diagrammed in Fig. 1, informa- et al. 2007). Several models have been suggested
tion from the cortex can be transmitted to the to investigate the roles that the BG play in per-
thalamus and SC through three BG pathways: ceptual decision-making for both normal and par-
(1) the direct pathway, from the striatum directly kinsonian subjects, which will be discussed in the
to the output nuclei GPi/SNr; (2) the indirect following.
pathway, from the striatum through the GPe and
STN to the output nuclei; and (3) the hyper-direct
pathway, from the cortex directly to the STN then Decision-Making Utilizing Go and No-Go
to the output nuclei. The opposing effect of dopa- Pathways
mine signals on striatum medium spiny neurons
that project through the direct and indirect path- Network models including the BG, the cortex,
ways differentially makes the BG suitable for and the thalamus can be constructed to investi-
implementing reward dependence learning dur- gate the role of the BG in perceptual decision-
ing the training process in decision-making tasks making under both normal and PD conditions
(Gerfen and Surmeier 2011). In addition, the (Frank 2006; Frank et al. 2004, 2007). In one
hyper-direct pathway provides a natural substrate such model (Frank et al. 2004), both the direct
for inhibitory control when an evidence accumu- (“go”) and indirect (“no-go”) pathways of the BG
lation process needs to be cancelled. Parkinson’s (but without STN) were included, and the role of
disease (PD), originating from depleted dopa- dopamine on reinforcement learning was
mine levels within the BG, shows significant checked. The striatal neurons projecting through
disruption of BG activity (DeLong and the two pathways to the output nuclei are differ-
Wichmann 2009; Obeso et al. 2008). Impairment entially modulated by positive and negative
in perceptual decision-making performance in outcomes through reward-dependent dopamine
PD, e.g., impulsivity in the face of uncertainty, signals: positive outcomes increase the
B 334 Basal Ganglia: Decision-Making
excitability of direct pathway striatal projection asymptotically optimal decision test for more
neurons and decrease the excitability of indirect than two alternatives, i.e., the multi-hypothesis
pathway projection neurons, while negative out- sequential probability ratio test (MSPRT) (Baum
comes have the opposite effect (Gerfen and and Veeravalli 1994). A model used to explore
Surmeier 2011). It was shown in both simulation this concept (Bogacz and Gurney 2007) was
and experiments that learning from negative out- based on the same BG functional structure as
comes was impaired in patients with PD receiv- employed in a previous work (Gurney etal.
ing dopamine medication. This model was later 2001). In this model, each BG nucleus was
extended to include the STN to investigate effects represented by its average firing rate. The for-
of STN deep brain stimulation (DBS) on impul- mula for detection probability in MSPRT was
sivity in decision-making with high conflict first log-transformed, and then each term in the
(Frank 2006; Frank et al. 2007). These two later expression for the log-transformed probability
works emphasized the role of the cortico-STN was connected with the firing activity of the
projection in threshold adjustment during deci- STN and GP (GPe of primates). This mapping
sion-making in high conflict situations. It was posed requirements for the input–output relation
found that two forms of PD treatment impaired in these two nuclei. Specifically, the STN firing
the performance in decision-making differently: rate was required to be an exponential function of
dopamine medication impaired the ability to the inputs from the cortex and from the GP, while
learn from wrong choices, while DBS treatment the GP firing rate was related to the summation of
increased the impulsivity arising with high STN output for different channels corresponding
conflict. These two independent mechanisms to each alternative action. These requirements for
that lead to impulsive choices for PD patients the STN and GP input–output relationship were
were then supported by experiments using supported by experimental data from in vitro
a probabilistic selection task. measurements of rat BG. The implementation of
The above model has also been extended for MSPRT could also be extended to include the
investigating inhibitory control (Wiecki and indirect pathway. In this model, the interaction
Frank 2013). The stop-signal task is one common between the direct and hyper-direct pathways and
kind of perceptual decision-making paradigm the diffusive innervation from the STN to the GP
that involves cancellation of an evidence accu- and entopeduncular nucleus (EP), which is the
mulation process before decision initiation homologue of GPi of primates, are essential for
(Verbruggen and Logan 2008). The BG provides the BG in implementing the MSPRT.
a natural substrate for implementing the cancel-
lation process through the hyper-direct pathway.
This has been evaluated in a cortical-BG-SC cir- Basal Ganglia Mechanism for Optimal
cuit model (Wiecki and Frank 2013), which Threshold Modulation
extended the authors’ previous model by adding
a frontal control circuit. The model has been A biophysically based spiking network model
shown to reproduce some key behavioral and involving the BG, the cortex, and the SC has
physiological data. been proposed to provide a mechanism for setting
and adjusting the threshold in perceptual deci-
sion-making (Lo and Wang 2006). In this
Basal Ganglia Model for Multi-choice model, integrate-and-fire neuron model was
Decision-Making used to describe individual neurons and tuned to
specific properties in each area. The SC neurons
Another hypothesis that has been considered fired a burst when the input was higher than
computationally is the idea that the direct and a threshold value. The cortical circuit was
hyper-direct pathways in the BG might provide modeled as an attractor network in which
a natural substrate for implementing an evidence accumulation occurred (Wang 2002).
Basal Ganglia: Decision-Making 335 B
The cortical circuit projected to the SC through research directions. For example, in some circuit
two routes: one was direct projection from the models, the STN and striatum received the same
cortex to the SC and the other route through the copy of input from the cortex, which usually
BG to the SC. In the BG, only the direct pathway represents evidence accumulated over time
was included. The neurons in the striatum were (e.g., Bogacz and Gurney 2007). But further B
designed to be silent and fire spikes only when the experiments are needed to clarify the origin of
input was higher than a threshold level. When one cortical inputs to the direct and hyper-direct path-
of the competing populations in the cortex ways of the BG (Mathai and Smith 2011).
reached a value that drove the CD neurons to Ramping of striatal neuronal activity during evi-
spike, the output nucleus SNr was inhibited, dence accumulation has been observed in pri-
which in turn led to a disinhibition of the SC. mates (Ding and Gold 2013) and will curtail the
When the SC was disinhibited, it could fire efficiency of a corticostriatal mechanism for
a burst in response to the excitatory input arriving threshold modulation that includes only the direct
directly from the cortex. This burst sent feedback pathway of the BG (as introduced in Lo and
to the cortex, terminated the evidence accumula- Wang 2006). Therefore, an extension of such
tion there, and set a threshold for perceptual deci- a mechanism to incorporate the new experimen-
sion-making. The SC burst was turned off by tal observation is called for, and experiments are
local recurrent connections within the SC. It needed to test the idea that the corticostriatal
was found that adjusting the strength of the connection strength is modulated during tasks
corticostriatal connection could modulate the requiring a speed-accuracy trade-off.
threshold within a wide range. On the contrary,
adjusting the cortex to SC connection strength
could only modestly modulate the threshold. Cross-References
A large range of threshold variability is beneficial
when a subject needs to make a transition ▶ Basal Ganglia: Habit Formation
between response speed and accuracy (Gold and ▶ Basal Ganglia: Mechanisms for Action
Shadlen 2002). In this model, modulation of the Selection
corticostriatal connection led to an inverted ▶ Decision-Making: Overview
u shape of reward rate, reflecting a trade-off
between speed and accuracy. The peak reward
rate could be shifted according to the task References
difficulty. Physiologically, the modulation in
the corticostriatal connection efficacy could be Baum CW, Veeravalli VV (1994) A sequential procedure
for multihypothesis testing. IEEE Trans Inform The-
realized by a reward-dependent reinforcement
ory 40(6):1994–2007. doi:10.1109/18.340472
learning process through dopaminergic Bogacz R, Gurney K (2007) The basal ganglia and cortex
synapses observed extensively in the striatum implement optimal decision making between alterna-
projection neurons (Gerfen and Surmeier 2011). tive actions. Neural Comput 19(2):442–477.
doi:10.1162/neco.2007.19.2.442
Further experiments are needed to test this
DeLong M, Wichmann T (2009) Update on models of
hypothesis. basal ganglia function and dysfunction. Parkinsonism
Relat Disord 15(Suppl 3):S237–240. doi:10.1016/
S1353-8020(09)70822-3
Ding L, Gold JI (2010) Caudate encodes multiple compu-
Future Directions tations for perceptual decisions. J Neurosci
30(47):15747–15759. doi:10.1523/JNEUROSCI.
Computational modeling represents an important 2894-10.2010
approach in investigating possible roles of the BG Ding L, Gold JI (2013) The Basal Ganglia’s contributions
to perceptual decision making. Neuron 79(4):640–649.
in perceptual decision-making. Such work usu-
doi:10.1016/j.neuron.2013.07.042
ally invokes some assumptions that require fur- Frank MJ (2006) Hold your horses: a dynamic computa-
ther experimental testing and also suggests new tional role for the subthalamic nucleus in decision
B 336 Basal Ganglia: Habit Formation
making. Neural Netw 19(8):1120–1136. doi:10.1016/ implications for Parkinson’s disease. Nat Rev
j.neunet.2006.03.006 Neurosci 11(11):760–772. doi:10.1038/nrn2915
Frank MJ, Seeberger LC, O’Reilly RC (2004) By carrot or Verbruggen F, Logan GD (2008) Response inhibition in
by stick: cognitive reinforcement learning in parkin- the stop-signal paradigm. Trends Cogn Sci
sonism. Science 306(5703):1940–1943. doi:10.1126/ 12(11):418–424. doi:10.1016/j.tics.2008.07.005
science.1102941 Wang XJ (2002) Probabilistic decision making by slow
Frank MJ, Samanta J, Moustafa AA, Sherman SJ reverberation in cortical circuits. Neuron
(2007) Hold your horses: impulsivity, deep brain 36(5):955–968
stimulation, and medication in parkinsonism. Wiecki TV, Frank MJ (2013) A computational model of
Science 318(5854):1309–1312. doi:10.1126/ inhibitory control in frontal cortex and basal ganglia.
science.1146157 Psychol Rev 120(2):329–355. doi:10.1037/a0031542
Gerfen CR, Surmeier DJ (2011) Modulation of striatal
projection systems by dopamine. Annu Rev Neurosci
34:441–466. doi:10.1146/annurev-neuro-061010-
113641 Basal Ganglia: Habit Formation
Gold JI, Shadlen MN (2002) Banburismus and the brain:
decoding the relationship between sensory stimuli,
decisions, and reward. Neuron 36(2):299–308 Barbara Knowlton
Gold JI, Shadlen MN (2007) The neural basis of decision Department of Psychology, UCLA, Los Angeles,
making. Annu Rev Neurosci 30:535–574. CA, USA
doi:10.1146/annurev.neuro.29.051605.113038
Graybiel AM (1995) Building action repertoires: memory
and learning functions of the basal ganglia. Curr Opin
Neurobiol 5(6):733–741 Definition
Gurney K, Prescott TJ, Redgrave P (2001)
A computational model of action selection in the
basal ganglia. I. A new functional anatomy. Biol
In large part, learning is characterized by the
Cybern 84(6):401–410 formation of new associations. Habit learning
Hikosaka O, Takikawa Y, Kawagoe R (2000) Role of the appears to be a form of associative learning that
basal ganglia in the control of purposive saccadic eye can occur in the absence of awareness of what has
movements. Physiol Rev 80(3):953–978
Huk AC, Shadlen MN (2005) Neural activity in macaque
been learned. Behaviors that lead to a reward can
parietal cortex reflects temporal integration of visual become habitual with sufficient training. When
motion signals during perceptual decision making. behaviors are habitual, they can be automatically
J Neurosci 25(45):10420–10436. doi:10.1523/ elicited by antecedent stimuli even when the out-
JNEUROSCI.4684-04.2005
Kim JN, Shadlen MN (1999) Neural correlates of
come of the behavior is no longer attractive.
a decision in the dorsolateral prefrontal cortex of the
macaque. Nat Neurosci 2(2):176–185. doi:10.1038/
5739 Detailed Description
Lo CC, Wang XJ (2006) Cortico-basal ganglia circuit
mechanism for a decision threshold in reaction time
tasks. Nat Neurosci 9(7):956–963. doi:10.1038/ The Representation of Habits
nn1722 Fundamentally, habits are associations between
Mathai A, Smith Y (2011) The corticostriatal and stimuli and responses, or S-R associations
corticosubthalamic pathways: two entries, one target.
So what? Front Syst Neurosci 5:64. doi:10.3389/
(Graybiel 2008; Mishkin and Petrie 1984;
fnsys.2011.00064 Packard and Knowlton 2002). For example,
Mink JW (1996) The basal ganglia: focused selection and when you are confronted with an intersection on
inhibition of competing motor programs. Prog your drive to work, this visual stimulus may
Neurobiol 50(4):381–425
Obeso JA, Marin C, Rodriguez-Oroz C, Blesa J, Benitez-
automatically elicit a left-turn response after
Temino B, Mena-Segovia J, Olanow CW (2008) The years of driving the same route. S-R associations
basal ganglia in Parkinson’s disease: current concepts contrast with arbitrary associations between stim-
and unexplained observations. Ann Neurol 64(Suppl uli that we typically think of as declarative learn-
2):S30–S46. doi:10.1002/ana.21481
Redgrave P, Rodriguez M, Smith Y, Rodriguez-Oroz MC,
ing, such as learning associations between word
Lehericy S, Bergman H, Obeso JA (2010) Goal- pairs. However, the distinction between stimulus-
directed and habitual control in the basal ganglia: stimulus (S-S) associations and S-R associations
Basal Ganglia: Habit Formation 337 B
does not fall neatly in the distinction between cue-outcome associations when posed in
declarative and nondeclarative memory. For a different way than what was present during
example, Pavlovian conditioning is the S-S asso- learning. For example, subjects may be asked to
ciation of a conditional stimulus with an uncon- imagine that the outcome is sunny and that two
ditional stimulus, and at least simple Pavlovian cues had been presented, and they should decide B
conditioning can occur in the absence of aware- which cues would they most likely be. While
ness of what has been learned and independently control subjects seem to have flexible knowledge
of medial temporal lobe structures, including the of these associations, amnesic patients do not
hippocampus, that support declarative memory. (Reber et al. 1996). It seems that their knowledge
Thus, both S-S and S-R associations can be of the association between the cues and outcomes
formed implicitly. is encapsulated as a set of stimulus–response
habits, while control subjects have gained declar-
Properties of Habits ative knowledge of these associations. While the
Independence of Awareness control subjects may have an S-R representation
In order to make the case that habit learning is of the cues and the correct responses that are
distinct from other types of learning, it is neces- associated with them, they also appear to have
sary to develop a list of its properties. As stated declarative knowledge of which cues are associ-
above, learning of S-R habits does not appear to ated with each outcome that allows them flexible
require awareness of what has been learned. That access to this knowledge. It appears that the
is, people may be acquiring a habit, and not amnesic patients are only able to solve the task
recognize it until they recognize themselves using S-R habits, such that the cues come to elicit
performing it. A powerful means to examine the correct response (pressing either the sun or
whether learning depends on awareness is to test rain key). These habits may exist independently
amnesic patients. These patients have a selective of awareness and are only observable when the
deficit in declarative memory due to damage in subject is confronted with the eliciting stimulus.
the medial temporal lobe or associated dience-
phalic structures (Squire and Zola-Morgan 1991). Lack of Flexibility
Thus, if these patients are able to learn habits, it Another characteristic of habits is that the knowl-
would suggest that habit learning can proceed edge gained is not readily applied in other con-
independently of awareness. Several studies texts, for example, in a forced-choice task, in
have demonstrated that patients with severe which subjects were presented with a series of
memory problems can acquire associations pairs of items and were required to make a choice
between stimuli and responses normally. For between the items of the pair. One item of the pair
example, in a “weather prediction” task in was consistently rewarded. Patients with dense
which different cues were associated with the amnesia were able to gradually learn to select the
outcome either “rain” or “sun,” amnesic patients correct item across many trials. However, they
were able to learn to select the correct choice at did not have any explicit knowledge of why they
the same rate as healthy control subjects. Over were selecting one item over the other, and they
about 100 trials, both groups were able to achieve attributed their choices to preference rather than
a level of performance significantly above their experience on previous trials (Bayley
chance. However, amnesic patients were unable et al. 2005). In contrast, control subjects learned
to remember much about the testing event, such to pick the correct item in each pair more quickly,
as the layout of the screen or the order of events and they were well aware that their choices were
on each trial (Knowlton et al. 1996). Amnesic based on which item had been rewarded on pre-
patients’ knowledge of the association between vious trials. Their conscious memory for the
the cues and the outcomes also seems to differ rewarded items of each pair enabled them to
from that of the control subjects. Control subjects pick these items out from among all the
were able to answer questions about the items – that is, their memory was flexible in that
they could readily demonstrate their memory in when you meant to go somewhere else. Although
a context that was different than the one that was the outcome (getting to work) was not desired at
present during training. Although the amnesic that time, the habitual response persisted when
patients eventually reached the same level as the the intersection was encountered. Another exam-
control subjects in the original forced-choice ple may be entering a room and reaching to turn
task, they were making their responses based on on a light switch through force of habit. Even
S-R habit. They did not have awareness of which when you have learned that the light has burned
items were rewarded and they were unable to use out or perhaps it is daytime and you do not need
their knowledge flexibly. This lack of awareness any light, the habit persists.
and flexibility of knowledge can be considered In several studies of instrumental learning, the
hallmarks of habit learning. training schedule was shown to be an important
factor as to whether learning is based on an A-O
Insensitivity to Devaluation or S-R association (Dickinson et al. 1983; Yin
One difficulty in determining whether perfor- et al. 2004). When rats were trained to press a bar
mance on a task is based on habit or declarative for food under a ratio schedule, devaluing the
memory is that just about every circumstance in food reinforcer resulted in a decrease in response
which performance could be supported by rate, suggesting that the rats were pressing the bar
a habit, performance also could be supported by in order to get food. In a ratio schedule, reinforce-
declarative memory. If a rat learns to press a lever ment is given each time the animal makes
resulting in a food reward, it may be that it has a certain number of responses – in other words,
developed an S-R habit, with the lever automat- food is given after every third bar press. In con-
ically eliciting a response. However, it may be trast, training rats under an interval schedule
that the rat remembers that food appears as an leads to performance that is insensitive to rein-
outcome of the lever pressing action, an A-O forcer devaluation. In an interval schedule, rein-
association. Similarly, a person may open their forcement is given if a response is made at any
refrigerator door in order to get food that is time during an interval (e.g., 30 s). No more
known to be inside (and A-O association), or reinforcement is given if 1 or 100 responses are
they may be simply opening the door out of made during the interval. A real-world example
“force of habit,” in an S-R fashion. may be like checking your mail. After training
In order to determine the nature of the memory rats in an interval schedule, one can then make
supporting behavior, we need to use probe tasks. the food that was used during training unattrac-
For example, we can test subjects’ awareness and tive to the rats. One way to accomplish this is to
flexibility of knowledge to demonstrate that what give the rats an unlimited amount of the food
was learned differed for control subjects and while they are in their home cage. When they
amnesic patients. However, these tasks did become sated on the food, they are immediately
not probe the structure of the underlying tested to see whether they will continue to
association – whether performance was respond as vigorously as a hungry rat. When
supported by S-R or A-O connections. In studies rats are trained on an interval schedule, devalua-
of animal learning, this has been accomplished by tion of the reinforcer has no effect. This is some
assessing the effect of devaluing the outcome on of the strongest evidence that performance can be
subsequent performance. The logic here is that in supported by S-R associations and that such habit
an S-R association, unlike in an A-O association, learning is distinct from learning the association
the outcome is not represented. Thus, perfor- between responses and their outcomes.
mance of the habit should not change if the out- One reason why training with ratio and inter-
come is no longer desired (Dickinson 1985). For val schedules leads to such different underlying
example, if you are in the habit of turning left at associative representations is that the contin-
a certain intersection on your route to work, you gency between the response and the outcome
may unintentionally make this turn one Saturday differs. Under a ratio schedule, each sequence of
responses is reinforced – if responding decreases, distracted while studying will hamper learning of
the number of reinforcers decreases. However, new facts and that it is more difficult to remember
with an interval schedule, the number of rein- information on a test when a distraction is pre-
forcers received is not nearly so tied to the sent. In contrast, habit learning does not seem to
response rate – under a 30-s schedule – and benefit as much from focused attention. In fact, it B
responding once per second and once per 30 s is often when we are distracted that we produce
would yield the same amount of reward. The a habitual response when we intended otherwise,
level of contingency between an action and its such as when we make the turn to go to work
outcome appears to be a key factor in whether when we in fact had set out for a different
learning results in a habit (Balleine and Dickin- destination.
son 1998). The presence of distraction during learning
Historically, the idea that learning is based on may result in a reliance on the habit learning
the formation of stimulus–response habits has system under circumstances in which one is
been influential. Hull promoted the idea that all likely to form declarative memory representa-
learning could be reduced to a series of S-R habits tions when attention is not divided. In a study
(Hull 1943). The role of reinforcement was to using the probabilistic classification task, sub-
strengthen these S-R bonds. However, this view jects learned two different sets of associations
was challenged by the fact that some types of (Foerde et al. 2006). They were told that they
learning did not seem to be consistent with the would be learning to predict the weather in two
acquisition of an S-R habit. The most striking different cities (Budapest and Prague) using two
example is latent learning described by Tolman different sets of cues. The two different sets of
and Honzik (1930), who showed that animals cues were in different colors to make it clear
learn the spatial layout of an environment even which task they were performing on each trial.
when not given reinforcement. Although it seems The subjects predicted the weather for one city
clear in hindsight that not all learning is habit under single task conditions – the weather pre-
learning, the later acknowledgment that there diction task was the focus of their attention. In
are multiple learning and memory systems made other blocks, subjects predicted the weather for
it clear that habit learning can coexist with other the other city while they concurrently performed
forms of learning. Mishkin and colleagues made another task in which they had to keep a running
this point in drawing the distinction between count of the number of tones that were presented.
habit and cognitive learning, with these two dif- Although subjects performed numerically worse
ferent types of learning depending on different at predicting the weather when they were learn-
brain systems (Mishkin et al. 1984). Differences ing under dual task conditions, performance was
in the neural architecture supporting these differ- similar for the two cities when subjects were later
ent types of learning give rise to differences in tested under single task conditions for both tasks.
their psychological properties. This suggests that while the performance on
a concurrent task may slightly impair perfor-
Automaticity mance of a habit (perhaps by occasionally caus-
Another feature of habit learning that seems to ing the subject to misperceive the cues or make
distinguish it from declarative or “cognitive” the wrong response), it does not seem to impair
learning is that it occurs relatively automatically. the acquisition of the habit.
Declarative learning and the retrieval of informa- One of the most interesting behavioral find-
tion from declarative memory are fairly attention ings of this study was the fact that the represen-
demanding. Memory encoding is poorer when tations underlying single and dual task learning in
performing a concurrent task, and while retrieval these two tasks appeared to be different. For the
from memory may be accurate under dual task city in which the task learned under single task
conditions, it is significantly slower (Craik conditions, subjects had flexible knowledge of
et al. 1996). We certainly understand that being the cue-outcome associations. They were able to
accurately assess which cues were most likely insensitive to reinforcer devaluation, indicating
to be present if it was rainy or sunny, so they that the response has become “habitized”
were able to access their knowledge of the (Dickinson 1985). Overtraining may serve to
cue-outcome associations in a way that was dif- break down the contingency between responding
ferent than that posed in the original task. For the and the outcome, because very few responses are
city in which the weather prediction learning withheld at this point. The animal no longer
occurred under dual task conditions, this flexible experiences the consequences of not making
knowledge was much less apparent, even though a response, so the link between the response and
performance on the task itself was fairly normal. the outcome is less apparent (Dickinson
These results are consistent with the idea that et al. 1983).
habit learning can proceed under conditions in The effect of overtraining on the development
which attentional resources are limited. In addi- of habit learning can also be seen using the cross-
tion, these data suggest that learning while dis- maze task (Tolman et al. 1946). In this task, a rat
tracted can result in the learned knowledge is trained to go to a baited arm that is either to the
restricted to S-R habits rather than flexible left or to the right of the start arm (e.g., for a rat
declarative representations. that starts in the “south” arm, the food reward will
always be at the end of the “east” arm). If rats are
Incremental Learning given a moderate amount of training and are then
An important characteristic of declarative learn- tested using the opposite start arm (in the exam-
ing is that it can occur in a single trial. We have ple, the “north” arm), they go to the location
conscious memories for specific moments in time where they received the reward. However, with
that we will keep in our entire lives. The unique overtraining, rats persist in making the same
structure of the hippocampus is thought to be able motor response as during training – they turn in
to support the rapid learning necessary for the the same direction, and thus when they are started
encoding of episodes (Rolls 2007). Our experi- from the “north” arm, they turn right and end up
ence of habit learning is different in that we think in the “west” arm (Hicks 1964; Packard and
of habits as behaviors that develop over time. McGaugh 1996; Fig. 1). Overtraining results in
Habits are acquired gradually, with feedback this shift to a reliance on a response strategy, in
incrementally strengthening the bond between which the maze elicits a motor response in an S-R
stimulus and response (Packard and Knowlton fashion.
2002). The neural structures that support habit
learning may not be capable of single-trial learn- Dependence on Feedback
ing. It may be that the slow incremental learning Finally, an important characteristic of habit learn-
of habits and the rapid learning of episodes are ing is that rewarding feedback is necessary to
complementary, ensuring that behaviors that strengthen the habit. Declarative learning is
have a long history of reward are automatic and thought to occur continually and incidentally
stable while still maintaining the flexibility to (Stark and Okado 2003). As in the case of the
alter behavior when there is a conditional change latent learning procedure described above, infor-
(Squire 1992). mation is stored in memory even without rein-
The gradual nature of habit learning is exem- forcement. For habit learning, reinforcement is
plified by the fact that overtraining typically thought to be critical for strengthening S-R
results in goal-directed actions becoming habits. bonds, even though the reinforcer is not part of
As discussed above, when rats are trained to the learned representation. One study that specif-
make an instrumental response under a ratio ically examined the role of feedback in habit
schedule, performance is goal directed, in that learning used the probabilistic classification task
responding will decrease when the reinforcer (Shohamy et al. 2004). In the version of the task
used is devalued. However, when training is con- used by these investigators, subjects learned to
tinued, performance will eventually become guess which flavor of ice cream (chocolate or
Training Probe test
B
Goal Response Place
Start
Basal Ganglia: Habit Formation, Fig. 1 The cross to the “place” the reward had been presented (the right
maze used to probe whether a rat is responding based on arm). If the rat is responding based on habit (making the
memory for the location of a reward or based on a habitual right-turn “response” when in the maze), it will go to the
motor response. If the rat chooses based on its memory for left arm in the probe trial
the location of the reward, during the probe test, it will go
vanilla) was preferred by a Mr. Potatohead figure associated with activation in the striatum, while
that was presented on each trial. Unbeknownst to learning through observation results in activation
the subject, the features that could be present or of medial temporal lobe (Poldrack et al. 2001).
absent on the Mr. Potatohead were probabilisti-
cally associated with either a chocolate or Brain Mechanisms of Habit
a vanilla preference. The procedure was similar Mishkin and Petrie (1984) argue that “Hullian”
to that used in the weather prediction version of habit learning and cognitive learning described
the task for some of the subjects. These subjects by Tolman coexist based on the idea that distinct
made a response (pressing either the “chocolate” brain systems support different forms of learning.
or “vanilla” key) and received feedback that they Mishkin and Petrie identify the basal ganglia and
were correct or incorrect on every trial. The other medial temporal lobe structures as supporting
subjects received an “observational” version of habit versus cognitive learning. A great deal of
the task in which the figures were presented support for this idea has come from lesion studies
followed by the correct answer (chocolate or that illustrate double dissociations between the
vanilla). In the observational task, the same infor- effects of damage to the striatum of the basal
mation about the association between cues and ganglia on habits and lesions of the medial tem-
outcomes was present, but the subjects did not poral lobe on cognitive learning. In nonhuman
make a response that was rewarded. It appeared primates, damage to medial temporal lobe struc-
that these two learning situations differ, in that tures including the hippocampus and surrounding
only feedback learning was impaired in patients temporal lobe cortices results in significant
with Parkinson’s disease. This finding suggests impairment in cognitive memory tasks such as
that feedback-based learning is supported by dif- the delayed non-match to sample task (Murray
ferent neural substrates than observational 2000). In this task, the subjects are presented with
learning. a sample item and after a delay (from seconds to
The idea that feedback-based and observa- hours) are presented with the sample item and
tional learning differ in terms of neural mecha- a different item. The subject’s task is to pick the
nisms is also supported by findings from new item. This task requires the monkey to
neuroimaging. Results from a study using the remember an event – the presentation of the sam-
weather prediction version of the probabilistic ple stimulus – in order to make the correct choice.
classification task have shown that when the Learning in this task is unaffected by lesions of
task is learned with feedback, learning is the tail of the caudate nucleus – the region which
receives visual information from cortex. In con- indicates that food is present at the end of the
trast these lesions severely impair performance arm. Unlike in the win-shift version of the task,
on a concurrent discrimination task (Fernandez- the rat does not need to remember any specific
Ruiz et al. 2001). As described above, in this task event. Rather, the rat simply must learn to run
the subject is given pairs of objects, with one down lit arms. Perhaps because it does not tap
element of the pair consistently rewarded. As into a natural foraging strategy, rats learn this task
discussed previously, this task can also be learned gradually over many sessions – keep in mind that
either declaratively or as a set of S-R habits. It the important information learned in the win-shift
appears that when monkeys are given a single task is acquired in single trial on each testing day.
trial per day for each discrimination, learning Packard et al. (1989) found that lesions of the
becomes habitual, perhaps because they are striatum (the lateral caudate nucleus, which
unable to readily recall previous trials with such roughly corresponds to the primate putamen)
a long delay between them. Under these condi- impair learning of the win-stay version of the
tions, lesions of the tail of the caudate nucleus radial arm maze task. Like other habit learning
disrupt learning, while medial temporal lobe tasks, this task is not affected by damage to the
lesions have no effect (Fernandez-Ruiz hippocampal system. Because the win-shift ver-
et al. 2001; Malamut et al. 1984). These results sion of the task was not impaired by lesions of the
indicate that habit learning and declarative learn- striatum, there is a double dissociation between
ing rely on different neural systems. the effects of lesions on the hippocampus and
striatum. This finding provides strong evidence
Double Dissociations Between Memory Systems that these two tasks depend on independent brain
Several studies using rats have demonstrated systems.
clear double dissociations between medial tem- A similar dissociation was shown using the
poral lobe and striatal learning systems. In cross maze described above (Packard and
a seminal study by Packard et al., rats were McGaugh 1996). Rats in which the hippocampus
trained on two different versions of the radial was inactivated by an intracranial injection of
arm maze task (Packard et al. 1989). In the radial lidocaine did not exhibit a preference for the
arm maze, eight arms radiate out from a central location strategy after a moderate amount of
platform, and a food reward is placed at the end of training, unlike rats in the control group which
each arm. In the win-shift version of this task, received saline injections. Hippocampal inactiva-
arms are not rebaited as the rat eats the food, so tion appeared to prevent the expression of the
the rat should not return to these arms during the location response. In contrast, rats in which the
session. The win-shift version of this task taps dorsolateral striatum was inactivated after mod-
into foraging strategies that rats have presumably erate training showed normal expression of the
evolved to efficiently recover food over an area. location response, but when the dorsolateral stri-
Packard et al. (1989) replicated the findings from atum was inactivated after overtraining, the rats
a number of studies that lesions of the hippocam- maintained the location response and did not
pus severely disrupt performance in this task switch to the habit response. Thus, the dorsolat-
(Becker et al. 1980). In order to avoid revisiting eral striatum must be active for expression of the
arms, the rat must remember the locations it has gradually developing habit response.
just visited on that session. This would seem to There is also evidence in human subjects that
require episodic memory, or memory for specific there is a dissociation between striatal-dependent
events. habit learning and learning that depends on the
In contrast to the win-shift version of the task, medial temporal lobe. The probabilistic classifi-
rats can also learn a win-stay discrimination on cation task has been extensively studied in patient
the radial arm maze, in which a cue consistently groups. For example, patients with Alzheimer’s
signals when an arm is baited, for example, the disease, which is accompanied by damage to the
cue may be a light, which when illuminated medial temporal lobe memory system, are able to
learn the S-R associations in the probabilistic performance when declarative memory is not
classification task (Eldridge et al. 2002). How- strong (older subjects, distracted young subjects).
ever, patients with Parkinson’s disease, which Although these dissociations based on lesions
affects dopaminergic input to the striatum, appear strongly suggest that different memory systems
to be impaired on this task (Knowlton exist, these findings do not directly demonstrate B
et al. 1996). In patients with mild Parkinson’s that the striatal-dependent form of learning has the
disease, performance is not different than that of defining features of habit learning. While win-stay
healthy control subjects. However, fMRI showed learning in rats has some of the properties of habit
that the patients and control subjects used differ- learning in that it is gradual and does not depend
ent neural systems during performance of this on brain regions important for declarative learn-
task (Moody et al. 2004). Performing the weather ing, more recent work demonstrated that learning
prediction task was associated with greater acti- in this task does indeed depend on S-R associa-
vation in the putamen in control subjects than in tions (Sage and Knowlton 2000). After training
the patients. The control subjects also showed rats on a win-stay task in which they ran down lit
a decrease in activation in the hippocampus dur- arms to receive food, the food reward used was
ing the weather task, a finding that has been seen devalued by pairing it in the home cage with an
before. It appears that in the controls, the hippo- injection of lithium chloride, which results in
campus is more active in the baseline condition, a learned taste aversion to the reward food. The
in which the subject is performing a very simple key question was when the rats were placed on the
task (it is likely that under such conditions, the radial arm maze after devaluation, would their
mind can wander and the subject spontaneously behavior be modified in any way in light of their
engages in mnemonic activity). In contrast, in the acquired aversion to the food reward. In fact, rats
patients with Parkinson’s disease, hippocampal continued to run down the correct arms just as
activation was greater than the level seen in the rapidly as they had when the food was attractive
baseline condition, consistent with the idea that to them. This suggests that the rats had learned an
these subjects were relying on declarative mem- association between the light and the response to
ory to support performance in this task. Because run down the arm, with the reward not represented
the habit learning system was compromised in in the association. In contrast, rats trained in the
these patients, they appear to have relied on hippocampal-dependent win-shift task slow down
declarative memory to perform this task. significantly when the food reward is devalued.
A similar distinction between habit and They still employ their typical foraging strategy
declarative learning was found in the fMRI data of efficiently traversing the maze, but their slow
of the study described above comparing perfor- responding indicates that their behavior is
mance of the weather prediction task under single influenced by the reward value.
task conditions and during distraction (Foerde Lesions of the dorsolateral striatum have been
et al. 2006). For the associations learned under shown to prevent habit learning in instrumental
single task conditions, performance was actually procedures as well. In one study, rats with dorso-
correlated with hippocampal activity in these lateral striatal lesions and control rats were
healthy young subjects. For the associations trained under an interval schedule to press a bar
learned under distraction, performance correlated for a food reward (Yin et al. 2004). Both the
with activity in the striatum and not the hippo- lesioned group and the control group learned to
campus, suggesting that the different neural sys- make the response at roughly the same rate. How-
tems were responsible for learning under the two ever, when the reward food was devalued by
conditions. These findings also suggest that when pairing it with illness, the two groups behaved
declarative memory is functioning well, as in very differently. The performance of the control
undistracted young subjects, it can support group was as expected for animals trained using
learning in the probabilistic classification task. an interval schedule; they continued to make the
However, habit learning appears to support instrumental response even though the food
reward was no longer attractive to them. In con- of rats with lesions of the dorsolateral caudate
trast, rats with dorsolateral striatal lesions nucleus – this group is able to learn the location
showed a significant decrease in responding. In of the hidden platform, but has difficulty learning
a sense, the lesioned rats are behaving with to swim to a cue that signals the location of the
greater “insight” in that they are avoiding platform.
a response that is associated with undesired out- In another study using the cross maze,
come. These data strongly make the case that the a dissociation was found between lesions of the
dorsolateral striatum is necessary for the devel- dorsolateral and dorsomedial striatum (Yin and
opment of habits. In the absence of the structure, Knowlton 2004). Rats with dorsolateral striatal
habitual responding does not appear to develop lesions did not show the typical tendency to run
and responding is guided by the outcome. down an arm based on a stimulus–response habit
after extensive overtraining. They retained the
Basal Ganglia Loops tendency to go to the rewarded location even
While the distinction between medial-temporal- when starting from a new arm. However, lesions
lobe-dependent memory and striatal-dependent placed more medially produced an effect more
memory is an appealing idea, it is also too simple. similar to that of hippocampal system lesions.
The striatum is organized into fairly discrete These rats did not go to the location where they
loops that include the pallidum and thalamus had been rewarded on previous trials, but rather,
and different cortical regions. A seminal paper they responded based on the motor habit. The
by Alexander et al. (1986) describes the architec- mediodorsal striatum in the rat is similar to the
ture of these loops based on neuroanatomical head of the caudate in the primate. This region is
findings (Fig. 2). Just as different regions of cor- interconnected with prefrontal regions that may
tex have very different functions, these loops be more involved in goal-directed action than in
support rather different cognitive operations. habit learning (Yeterian and Pandya 1991). This
Thus, it does not make sense to think of the loop may in fact be important for implementing
striatum as the substrate of a particular type of actions based on the contents of declarative
memory, but rather that the different loops memory.
may serve to support different memory processes Electrophysiological studies have also
(Yin and Knowlton 2006). Lesions that affect supported the idea that goal-directed actions and
habit learning appear to involve the dorsolateral habits depend on different corticostriatal loops.
striatum in the rat or the putamen in the In nonhuman primates, performance early in
primate, suggesting that the motor loop is training is accompanied by increased firing in
critical for habit learning. The fact that this loop the caudate nucleus. When the slope of the learn-
includes cortical regions that are important for ing curve is steep, firing in the caudate nucleus
motor output is consistent with the idea that was correlated with the slope of the learning
habit learning involves an association with curve. In contrast, firing in the putamen was cor-
a response. related with the level of learning, with firing
Other corticostriatal loops also appear to be highest after extensive training when the mon-
important for behavior dependent on declarative keys were adept at the task (Williams and
learning. Damage to the dorsomedial caudate Eskandar 2006).
nucleus in the rat results in behavior that appears
habitual under conditions in which intact rats Communication Between Loops
perform based on declarative memory. In the Corticostriatal loops are often characterized as
Morris water maze, rats with dorsomedial cau- functioning independently, based to a great
date lesions have difficulty learning the location extent on anatomy. The thalamic targets of these
of a hidden platform and behave similarly to rats loops are the same cortical regions that originated
that have hippocampal lesions (Devan the loop. For example, in the motor loop, cortical
et al. 1999). This contrasts with the performance projections from the motor cortex are directed to
Basal Ganglia: Habit Formation, cortex, ACC anterior cingulate cortex, dl/vm dorsolateral/
Fig. 2 Abbreviations: SMA supplemental motor area, ventromedial, GPI/SNR globus pallidus internal/
DLPFC/OFC dorsolateral prefrontal cortex/orbitofrontal substantia nigra reticulata
the putamen, which projects to relatively discrete A third mechanism by which loops can com-
regions of the globus pallidus that project to the municate is through the projections from the stri-
ventrolateral nucleus of the thalamus. This tha- atum to dopaminergic neurons in the midbrain
lamic nucleus sends its cortical projections back (Joel and Weiner 2000). These dopaminergic
to motor cortex, thus completing the loop. How- neurons in the ventral tegmental area and the
ever, it is also the case that these loops have the substantia nigra are the source of a major projec-
potential to interact based on other features of the tion to the striatum. There is reciprocity in these
anatomy. The thalamocortical projections termi- projections, in that the different striatal regions
nate in multiple cortical layers, including layer project back to the midbrain region that supplied
III, the site of neurons that gives rise to cortico- dopaminergic input to that region. For example,
cortical projections. Thus, the output of one loop the ventral tegmental area is a major source of
is able to influence another through these cortico- dopaminergic input to the ventral striatum, and
cortical projections (McFarland and Haber 2002). the ventral striatum projects back to the ventral
In addition, there are extensive projections from tegmental area. In addition, however, the ventral
the cortex back to the thalamus. For example, tegmental region also projects to other striatal
cortical regions involved in planning actions regions, thereby allowing communication
(pre-supplementary motor area and premotor cor- between the limbic loop and the other striatal
tex) send projections to VL, which is part of the loops. The dopaminergic neurons targeted by
motor loop that projects to primary motor cortex striatal regions of the association loop (i.e., the
and supplementary motor cortex. By these “open caudate nucleus) also give rise to both reciprocal
loop” projections, different cortical regions can projections and projections that target the striatal
influence corticostriatal loops at the level of the regions of the motor loop (i.e., the putamen).
thalamus. Thus, the midbrain dopaminergic system is
another means by which corticostriatal loops may before it. Through this recursive mechanism,
interact. activity in the dopaminergic neurons becomes
An interesting feature of the striatal projec- a predictive signal for reinforcement (Houk
tions to both the thalamus and the dopaminergic et al. 1995).
system is that there seems to be a hierarchical
organization in the potential interactions between Plasticity in the Striatum
loops (Joel and Weiner 2000). Cortical regions in One question that arises in discussing the role of
the limbic loop project to thalamic regions in the neostriatum in learning is where the plastic
loops involved in executive function, and cortical changes that support learning occur, and whether
executive regions project to thalamic regions the physiological mechanisms of this plasticity
involved in the motor loop. A similar hierarchy are similar to or different than those that have
exists for the dopaminergic projections. The ven- been described in the medial temporal lobe sys-
tral striatum (limbic loop) projects to midbrain tem. As in the hippocampus, long-term potentia-
dopaminergic regions supplying all other loops, tion (LTP) occurs in the striatum, and this
striatal regions in the executive loop send pro- phenomenon has been considered a model system
jections to dopaminergic regions supplying the for the formation of habit memories (Reynolds
executive and motor loops, and the motor stria- et al. 2001). In addition, long-term depression
tum primarily sends only reciprocal connections (LTD) can also be induced in the striatum.
back to the dopaminergic neurons that project to High-frequency stimulation of the cerebral cortex
it. Thus, it seems that in both mechanisms of induces LTP or LTD at the corticostriatal
interaction, the limbic loop has the potential for glutamatergic synapses. Dopamine is important
greater influence over the other loops, loops for the induction of plasticity in the striatum. LTP
involved in executive function can influence or LTD cannot be readily induced in slices from
those involved in motor function, while the animals in which dopamine is depleted (Calabresi
motor loops have limited ability to influence the et al. 1992; Kerr and Wickens 2001).
other loops. Unlike in the hippocampus, where LTP has
The reciprocity of connections in the motor been studied more extensively than LTD, striatal
loop has been incorporated into computational LTD has received more attention in the past than
models of learning in the striatum. Dopaminergic LTP. High-frequency stimulation trains that pro-
input plays an important role in providing duce LTP in the hippocampus produce LTD in
a reinforcement signal in these models. The the striatum. In fact, LTP was not initially
actor-critic architecture (Barto et al. 1983) cap- obtained in the striatum unless magnesium was
tures this role of reinforcement, by postulating absent. It appears that the facilitative effects of
a “critic,” in the form of dopaminergic input, the lack of magnesium were mediated by
that provides the “actor,” in the form of striatal dopamine-NMDA receptors, which are normally
neurons, with a reinforcement signal based on blocked by magnesium, and can be activated on
whether success has been obtained. One chal- presynaptic dopamine terminals, leading to an
lenge to these models is the problem of credit increase in dopamine levels (Wickens and
assignment – the consequences of an action are Arbuthnott 2005).
often delayed. These models are able to address More recently, the role of dopamine in striatal
this challenge based on the fact that the dopami- plasticity has been shown to be complex. The
nergic input to each region of the striatum arises different types of dopamine receptor (D1 and
from cells in that same zone that receives input D2) are on different populations of medium
from that striatal region. This reciprocity means spiny neurons in the striatum. Both receive dopa-
that dopaminergic activity will strengthen the minergic input from the substantia nigra, but
response just preceding it. On subsequent trials, these two classes of neurons participate in two
dopaminergic activity will then occur during this different striatal output pathways: D1 receptors
response, thus strengthening the response just are on neurons in the direct pathway, and D2
Basal Ganglia: Habit
Formation, Fig. 3 Direct
and indirect pathways in the
basal ganglia.
Abbreviations: GPI globus
pallidus internal, ctx cortex, B
GPE globus pallidus
external, STN subthalamic
nucleus
receptors are on neurons in the indirect pathway containing D2 receptors, while only LTD is
that includes the globus pallidus external and the enabled in neurons containing D1 receptors.
subthalamic nucleus (Gerfen et al. 1990; Fig. 3). However, when dopamine levels rise, as when
On both of these pathways, glutamate input from a reinforcing stimulus is present, LTP can be
the cortex is excitatory, but the net effect of the induced in neurons containing D1 receptors,
two pathways is different. Exciting the direct while those containing D2 receptors will be
pathway serves to remove tonic inhibition in the biased toward LTD.
thalamus, which thus is excitatory to the cortex. These findings point to an important role for
Exciting the indirect pathway has the net effect of dopamine in the control of plasticity in
strengthening this tonic inhibition. Thus, striatal corticostriatal synapses and suggest that medium
plasticity can lead to either increased or spiny neurons with D1 and D2 receptors play
decreased cortical activity. different roles in learning. When D1 receptors
In a study examining the role of dopamine in are activated, as in the case of a large, unexpected
striatal plasticity, Shen et al. (2008) assessed LTP reward, efficiency will increase in corticostriatal
and LTD in both the direct and indirect pathways. synapses on neurons containing D1 receptors,
On striatal medium spiny neurons that have D1 thus strengthening the direct pathway, leading to
receptors (the direct pathway), LTP can be gen- greater excitation of the cortex. However, when
erated from pairing presynaptic activity with dopamine is at a tonic level, experience would
postsynaptic depolarization. When D1 receptors result in enhancement only in corticostriatal syn-
are blocked, this pairing leads to LTD. In medium apses on neurons containing D2 receptors, thus
spiny neurons with D2 receptors (the indirect strengthening the indirect pathway, leading to
pathway), LTP and LTD can be induced, increased tonic inhibition of cortex (Shen
depending on the timing of the pre- and postsyn- et al. 2008). In addition to differences in the role
aptic stimulation. Blocking D2 receptors disrupts of different corticostriatal loops in behavior, con-
LTD, while increasing D2 receptor activation ditions within each loop (e.g., dopamine levels)
results in LTD under conditions in which LTP is control the net effect of that loop on its target
normally induced. Thus, under low dopamine cortical region.
levels, when only higher-affinity D2 receptors It is tempting to think of these two mecha-
are activated, plasticity is bidirectional in neurons nisms of plasticity of supporting learning of
actions and habits; early in learning, new associ- indirect pathways with behavior is essential for
ations are formed in the direct pathway when understanding the significance of striatal
dopamine signals the presence of reinforcement. architecture.
As practice increases and reinforcement becomes If different memory systems acquire memo-
expected, dopamine levels decrease and perfor- ries in parallel, competition is likely to occur at
mance becomes increasingly controlled by the the level of control for behavioral output. It may
indirect pathway, characterized by a lack of flex- be that the selection is based on the strength of the
ibility and insensitivity to devaluation. learned representation. Declarative memories
The interplay between the direct and indirect may be formed faster and thus control behavior
pathways is incorporated into models of dynamic early in training, while habits accumulate asso-
gating by the basal ganglia (Frank et al. 2001). ciative strength more slowly, but eventually
When the direct pathway is activated, there is become dominant. For example, in the cross-
a disinhibition of cortex, enabling the updating maze task described above, the correct response
of working memory and flexibility. Activity in according to a declarative memory for the loca-
the indirect pathway has the opposite effect – it tion of reward is incompatible with the
increases inhibition in the cortex, thus favoring stimulus–response-based habit to turn in one
maintenance. In this way, the architecture of direction in that these two responses lead to dif-
the cortico-basal ganglia system balances ferent places when a new starting location is used
between the conflicting demands of updating vs. at the opposite end of the maze. Thus, the control
maintaining behavior. Furthermore, the existence of behavioral output may be determined by rel-
of parallel loops with relative independent neu- ative memory strength, which would likely
ronal circuits allows the updating of some infor- change with practice level. If behavioral output
mation, but not others according to these gating is determined by memory strength, the critical
models (O’Reilly and Frank 2006). question is how the “strongest” memory is deter-
mined that will control behavioral output. Of
Interactions Between Learning Systems course, in most situations, declarative memory
What does the anatomy and neurophysiology of and habit would presumably lead to the same
the striatum suggest about how different systems outcome. However, situations such as the cross
for learning interact? The loop architecture sug- maze can bring these two systems into conflict.
gests parallel modes of learning; loops involving This could be seen in the “real-life” example of
the caudate and prefrontal cortex may be more making a turn along a well-learned route when
involved in voluntary action than loops involving in fact you had intended to turn the other way at
the putamen and motor cortex. Learning may the intersection to go to a different destination
proceed in these loops in parallel, with behavior on this day. Responding based on habit and on
controlled by whichever loop is the most effec- declarative memory will lead to opposing
tive at influencing the structures supporting responses.
behavioral output. While the limbic and associa- Given that responses based on declarative
tive loops are able to influence the effectiveness memory and habit learning can differ, it is impor-
of the motor loop, there is substantial tant to determine the conditions that favor the use
independence – one could be learning a goal- of one representation over another. Our awareness
directed action while also strengthening a habit is based on declarative knowledge. Thus, when we
based in the motor loop. By this view, learning a are consciously aware of responding based on
goal-directed action does not impede the learning memory, we are using declarative memory. How-
of a habit or vice versa. On the other hand, the ever, when awareness is otherwise occupied,
antagonistic relationship between plasticity in behavior is more likely to depend on habits due
the direct and indirect pathways suggests that at to their relative insensitivity to working memory
some level there is competition between types of load (Foerde et al. 2007). Conditions that occupy
learning. Linking plasticity in the direct and or reduce working memory resources, such as
anxiety, depression, and fatigue, may bias behav- Conclusion
ior toward the learning and retrieval of habits. Although we are often not aware of our habits,
Individuals suffering from depression and anxiety they arguably have a greater influence on our
exhibit impaired working memory (Rose and ongoing behavior than declarative memories
Ebmeier 2006; Samuelson et al. 2006). In addition, do. Our habits provide a backdrop to our con- B
even mild forms of depression are accompanied scious behavior. Psychiatric disorders such as
by a tendency to ruminate and have intrusive addiction and compulsive behaviors are funda-
thoughts, thus squandering working memory mentally problems with habits that are intrusive
resources (Joormann and Gotlib 2008). Emotional in daily life. Thus, how habits are formed and
factors may thus play a role in the type of repre- how they interact with declarative memories are
sentations that are learned and that control behav- key questions in understanding both normal and
ior. The implications of this idea are that people abnormal behaviors.
under stress or preoccupied may be more likely to
fall back on old habits even when an updated
approach is warranted. Cross-References
Habit learning is a form of instrumental learn-
ing, in that making the response is necessary to ▶ Basal Ganglia: Overview
obtain the outcome during training. In Pavlovian ▶ Hippocampus, Model Network Architecture
conditioning, pairing of the conditional stimulus
(e.g., a buzzer) and an unconditional stimulus
(e.g., food) leads to the conditional stimulus References
eliciting a conditional response (e.g., salivation).
The response has no bearing on whether the Alexander GE, DeLong MR, Strick PL (1986) Parallel
unconditional stimulus is received. Interestingly, organization of functionally segregated circuits
these two different forms of learning interact. linking basal ganglia and cortex. Annu Rev Neurosci
9:357–381
Presenting cues that have been conditional stim- Balleine BW, Dickinson A (1998) Goal-directed instru-
uli in Pavlovian conditioning can act to facilitate mental action: contingency and incentive learning and
the transfer of actions to habits. This phenome- their cortical substrates. Neuropharmacology
non has been discussed most extensively in the 37:407–419
Barto A, Sutton R, Anderson C (1983) Neuron-like ele-
context of addiction, where behaviors such as ments that can solve difficult control problems. IEEE
drug taking become compulsive and persist Trans Syst Man Cybern 13
despite the desire to cease. The transition from Bayley PJ, Frascino JC, Squire LR (2005) Robust
voluntary action to habitual responding is thought habit learning in the absence of awareness and inde-
pendent of the medial temporal lobe. Nature
to characterize, at least in part, the transition from 436:550–553
drug user to drug addict (Everitt et al. 2001). Becker JT, Walker JA, Olton DS (1980) Neuroanatomical
During drug use, Pavlovian conditioning occurs bases of spatial memory. Brain Res 200:307–320
that associates drug-related cues and contexts Calabresi P, Maj R, Mercuri NB, Bernardi G (1992)
Coactivation of D1 and D2 dopamine receptors is
(conditional stimuli) with the physiological required for long-term synaptic depression in the stri-
effects of the drugs (unconditional stimuli). In atum. Neurosci Lett 142:95–99
animal models of addiction, the presence Craik FIM, Govoni R, Naveh-Benjamin M, Anderson ND
of these conditional stimuli leads to faster learn- (1996) The effects of divided attention on encoding
and retrieval processes in human memory. J Exp
ing of an instrumental response (Pavlovian- Psychol Gen 125:159–180
instrumental transfer) but also a faster transition Devan BD, McDonald RJ, White NM (1999) Effects
from voluntary responding to habitual of medial and lateral caudate-putamen lesions on
stimulus–response responding. This effect of place- and cue-guided behaviors in the water maze:
relation to thigmotaxis. Behav Brain Res 100:5–14
Pavlovian conditioning may result from the influ- Dickinson A (1985) Actions and habits: the development
ence of limbic loops of the striatum on other of behavioural autonomy. Philos Trans R Soc Lond
loops (Everitt and Robbins 2005). 308:67–78
Dickinson A, Nicholas DJ, Adams CD (1983) The effect succeed in object discrimination learning despite
of instrumental training schedule on susceptibility of 24-hour intertrial intervals. Behav Neurosci
reinforcer devaluation. Q J Exp Psychol 35B:35–35I 98:759–769
Eldridge LL, Masterman D, Knowlton BJ (2002) Intact McFarland NR, Haber SN (2002) Thalamic relay nuclei of
implicit habit learning in Alzheimer’s disease. Behav the basal ganglia form both reciprocal and
Neurosci 116:722–726 nonreciprocal cortical connections linking multiple
Everitt BJ, Robbins TW (2005) Neural systems for rein- cortical areas. J Neurosci 22:8117–8132
forcement for drug addiction: from actions to habits to Mishkin M, Petrie HL (1984) Memories and habits: some
compulsion. Nat Neurosci 8:1481–1489 implications for the analysis of learning and retention.
Everitt BJ, Dickinson A, Robbins TW (2001) The neuro- In: Squire LR, Butters N (eds) Neuropsychology of
psychological basis of addictive behaviour. Brain Res memory. Guilford Press, New York, pp 287–296
Rev 36:129–138 Mishkin M, Malamut B, Bachevalier J (1984) Memories
Fernandez-Ruiz J, Wang J, Aigner TG, Mishkin M (2001) and habits: two neural systems. In: Lynch G, McGaugh
Visual habit formation in monkeys with neurotoxic JL, Weinberger NW (eds) Neurobiology of learning
lesions of the ventrocaudal neostriatum. Proc Natl and memory. Guilford, New York, pp 65–77
Acad Sci 98:196–201 Moody TD, Bookheimer SY, Vanek Z, Knowlton BJ
Foerde K, Knowlton BJ, Poldrack RA (2006) Modulation (2004) An implicit learning task activates medial tem-
of competing memory systems by distraction. Proc poral lobe in patients with Parkinson’s disease. Behav
Natl Acad Sci USA 103:11778–11783 Neurosci 118:438–442
Foerde K, Poldrack RA, Knowlton BJ (2007) Secondary Murray EA (2000) Memory for objects in nonhuman
task effects on classification learning. Mem Cognit primates. In: Gazzaniga M (ed) The cognitive neuro-
35:864–874 sciences, 2nd edn. MIT Press, Cambridge, MA,
Frank MJ, Loughry B, O’Reilly RC (2001) Interactions pp 753–763
between the frontal cortex and basal ganglia in work- O’Reilly RC, Frank MJ (2006) Making working memory
ing memory: a computational model. Cognit Affect work: a computational model of learning in the frontal
Behav Neurosci 1:137–160 cortex and basal ganglia. Neural Comput 18:283–328
Gerfen CR, Engber TM, Mahan LC, Susel Z, Chase TN, Packard MG, Knowlton BJ (2002) Learning and memory
Monsma FJ Jr, Sibley DR (1990) D1 and D2 dopamine functions of the basal ganglia. Annu Rev Neurosci
receptor-regulated gene expression of striatonigral and 25:563–593
striatopallidal neurons. Science 250:1429–1432 Packard MG, McGaugh JL (1996) Inactivation of hippo-
Graybiel A (2008) Habits, rituals, and the evaluative brain. campus or caudate nucleus with lidocaine differen-
Annu Rev Neurosci 31:359–387 tially affects expression of place and response
Hicks LH (1964) Effects of overtraining on acquisition learning. Neurobiol Learn Mem 65:65–72
and reversal of place and response learning. Psychol Packard MG, Hirsh R, White NM (1989) Differential
Rep 15:459–462 effects of fornix and caudate lesions on two radial
Houk JC, Adams JL, Barto AG (1995) A model of how the maze tasks: evidence for multiple memory systems.
basal ganglia generates and uses neural signals that J Neurosci 9:1465–1472
predict reinforcement. In: Houk JC, Davis J, Beiser Poldrack RA, Clark J, Pare-Blagoev EJ, Shohamy D, Creso
D (eds) Models of information processing in the basal Moyano J, Myers C, Gluck MA (2001) Interactive mem-
ganglia. MIT Press, Cambridge, MA ory systems in the human brain. Nature 414:546–550
Hull CL (1943) Principles of behavior. Appleton, New Reynolds JN, Hyland BI, Wickens JR (2001) A cellular
York mechanism of reward- related learning. Nature
Joel D, Weiner I (2000) The connections of the dopami- 413:67–70
nergic system with the striatum in rats and primates: an Rolls ET (2007) An attractor network in the hippocampus:
analysis with respect to the functional and compart- theory and neurophysiology. Learn Mem 14:714–731
mental organization of the striatum. Neuroscience Rose EJ, Ebmeier KP (2006) Pattern of impaired working
96:451–474 memory during major depression. J Affect Disord
Joormann J, Gotlib IH (2008) Updating the contents of 90:149–161
working memory in depression: interference from Sage JR, Knowlton BJ (2000) Effects of US devaluation
irrelevant negative material. J Abnorm Psychol on win-stay and win-shift radial maze performance in
117:182–192 rats. Behav Neurosci 114:295–306
Kerr JND, Wickens JR (2001) Dopamine D1/D5 receptor Samuelson KW, Meylan TC, Metzler TJ, Lenoci M,
activation is required for long-term potentiation in the Rothlind J, Henn-Haase C, Choucroun G, Weiner
rat neostriatum in vitro. J Neurophysiol 85:117–124 MW, Marmar CR (2006) Neuropsychological func-
Knowlton BJ, Mangels JA, Squire LR (1996) A neostriatal tioning in posttraumatic stress disorder and alcohol
habit learning system in humans. Science abuse. Neuropsychology 20:716–726
273:1399–1402 Shen W, Flajolet M, Greengard P, Surmeier DJ
Malamut BL, Saunders RC, Mishkin M (1984) Monkeys (2008) Dichotomous dopaminergic control of striatal
with combined amygdalo-hippocampal lesions synaptic plasticity. Science 321:848–851
Basal Ganglia: Mechanisms for Action Selection 351 B
Shohamy D, Myers CE, Grossman S, Sage J, Gluck MA, diverged from the main vertebrate line approxi-
Poldrack RA (2004) Cortico-striatal contributions to mately 560 MYA. Moreover, there are analogous
feedback-based learning: converging data from neuro-
imaging and neuropsychology. Brain 127:851–859 structures in birds and bony fish. The preservation
Squire LR (1992) Memory and the hippocampus: of this circuit implies a crucial operation that
a synthesis of from findings with rats, monkeys, and solves a problem common to all vertebrates. B
humans. Psychol Rev 99:195–231 While many functions for the basal ganglia have
Squire LR, Zola-Morgan S (1991) The medial temporal
lobe memory system. Science 253:1380–1386 been proposed, our current best theory to unify
Stark CE, Okado Y (2003) Making memories without these ideas is that this circuit solves the common
trying: medial temporal lobe activity associated with problem of action selection. Informally put, this
incidental memory formation during recognition. is the problem of how to decide what to do next.
J Neurosci 23:6748–6753
Tolman EC, Honzik CH (1930) Degrees of hunger,
reward, and non-reward, and maze learning in rats.
Univ Calif Publ Psychol 4:241–256 Detailed Description
Tolman EC, Ritchie BF, Kalish D (1946) Studies in spatial
learning: I. Orientation and the short-cut. J Exp
Psychol 36:13–24 Why Action Selection?
Wickens JR, Arbuthnott GW (2005) Structural and func- The necessity of a mechanism for action selection
tional interactions in the striatum at the receptor is imposed by the strong constraint that complex
level. In: Dunnett SB, Bentovoglio M, Bjorklund A, animals have a final common motor pathway: the
Hokfelt T (eds) Dopamine, handbook of chemical
neuroanatomy, vol 21. Elsevier, Amsterdam, connections of the spinal cord and the number of
pp 199–236 muscle groups limit the set of actions that can be
Williams ZM, Eskandar EN (2006) Selective enhance- expressed simultaneously. Some mechanism is
ment of associative learning by microstimulation of required to reduce the repertoire of currently
the anterior caudate. Nat Neurosci 9:562–568
Yeterian EH, Pandya DN (1991) Prefrontostriatal connec- available actions to just those actions capable of
tions in relation to cortical architectonic organization simultaneous expression. Given the constraints of
in rhesus monkeys. J Comp Neurol 312:43–67 biological tissue, a central rather than distributed
Yin HH, Knowlton BJ (2004) Contributions of striatal selection mechanism is likely the preferred solu-
subregions to place and response learning. Learn
Mem 11:459–463 tion implemented by a neural substrate (Prescott
Yin HH, Knowlton BJ (2006) The role of the basal ganglia et al. 1999). Briefly, the argument runs as follows.
in habit formation. Nat Rev Neurosci 7:464–476 Typically, distributed selection mechanisms,
Yin HH, Knowlton BJ, Balleine BW (2004) Lesions of formed by reciprocal, inhibitory links between n
dorsolateral striatum preserve outcome expectancy but
disrupt habit formation in instrumental learning. Eur behavior-representing nodes, contain n(n – 1)
J Neurosci 19:181–189 links and grow as 2n for every additional node.
By contrast, a central selection device which is
reciprocally connected with all n nodes (thus
allowing control over the expression of each
Basal Ganglia: Mechanisms for node’s represented behavior) requires just 2n
Action Selection links and grows by 2 for each additional node.
Thus, a central selection device is more econom-
Mark D. Humphries ical in both the number of connections required
Faculty of Life Sciences, University of and the cost of adding nodes. Such economy of
Manchester, Manchester, UK wiring appears to be a priority for the central
nervous system (Cherniak 1994).
The basal ganglia appear to fulfill the criteria
Definition for such a central mechanism (Redgrave et al.
1999). The main input nucleus, the striatum,
The basal ganglia are an interconnected circuit of receives input from every region of cortex, from
fore- and midbrain nuclei common to all extant primary visual, auditory, and somatosensory cor-
vertebrates, including the lamprey, which tex, through motor cortices, to the subregions of
B 352 Basal Ganglia: Mechanisms for Action Selection
prefrontal cortex. It is thus in a position to inte- their targets. A sufficiently strong input signal
grate sensory information, current motor com- from cortex can elicit activity in a population of
mands, planning, and goals into representations projection neurons in the striatum, which in turn
of the relative importance, or salience, of each converge on and temporarily inhibit the firing of
action. The output of the basal ganglia targets neurons in the output nuclei. This pause in tonic
regions in the thalamus and brain stem that under- firing is then interpreted as the signal for
pin the execution of different actions. The basal selection.
ganglia are thus also in a position to directly This disinhibition process has been worked
control the activity in neural systems responsible out in detail for the control of voluntary eye
for executing actions. movements, or saccades, in primates (see
The idea that the basal ganglia perform action Hikosaka et al. 2000). Phasic activity in the cau-
selection is also a potential solution to the long- date region of the striatum precedes saccade ini-
standing paradox of their role in controlling tiation and is time-locked to a decrease in activity
movement (Marsden and Obeso 1994). Activity in a subset of SNr neurons. The SNr projects to
in the basal ganglia is strongly correlated with the intermediate layers of the superior colliculus,
movement onset and parameters of that move- which contains a retinotopic map – neurons are
ment, and stimulation of regions of the striatum laid out in a sheet that represents coordinates of
can evoke discrete, smooth limb movements gaze direction. A buildup of neural activity on
(Alexander and DeLong 1985). Yet lesions of this map drives the brain stem circuitry that con-
basal ganglia nuclei in animal studies or in trols the eye muscles and moves gaze to the
Parkinson’s disease patients to alleviate their location signaled by the map. The pause in SNr
symptoms do not clearly impair the initiation of activity precedes and appears to be required for
movement (Marsden and Obeso 1994). The the- the buildup of activity in the retinotopic map of
ory that the basal ganglia are necessary to control the superior colliculus. Conversely, continuously
the choice of actions explains how movement- inhibiting SNr activity results in the inability to
related activity can arise yet not be needed to suppress saccades, as though every eye move-
directly generate movement. ment motor command was being executed
(Hikosaka and Wurtz 1985). Thus, the direct stri-
How Can the Basal Ganglia Circuit “Select”? atum-SNr pathway likely controls the selection
While the basal ganglia’s inputs and outputs posi- of saccade direction through a process of
tion these nuclei to potentially act as a central disinhibition.
selection mechanism, the key evidence in favor
of this idea is that their internal circuitry is seem- Multiple Mechanisms Contributing to Action
ingly designed to implement a selection process. Selection
Figure 1 illustrates the basic circuitry of the basal Computational models have provided quantita-
ganglia as currently understood. The original pro- tive tests of these ideas, revealed the further nec-
posals for how the internal circuitry of the basal essary mechanisms for successful selection, and
ganglia may implement a selection process shown how these mechanisms are implemented
envisioned a division of labor between the by components of the basal ganglia circuit. The
“direct” and “indirect” pathways (Albin et al. box-and-arrow model accounts of “direct” path-
1989; DeLong 1990; Alexander and Crutcher way control of selection focussed on the flow of
1990). The direct pathway was proposed to be signals necessary to select one input and so omit-
responsible for selection through a process of ted two crucial aspects of a selection process:
disinhibition (Chevalier and Deniau 1990), illus- how to resolve the competition between multiple
trated in Fig. 1. The output nuclei of the basal action-representing inputs and how to switch
ganglia are continuously active, providing from a currently selected action to another,
a source of constant (or “tonic”) inhibition to more salient, action.
Basal Ganglia: Mechanisms for Action Selection, globus pallidus external segment (GPe) projects within
Fig. 1 Basal ganglia circuit and selection. The striatum the basal ganglia; the globus pallidus internal segment
receives input from most of cortex and many thalamic (GPi) and substantia nigra pars reticulata (SNr) project
nuclei; a subset of the same regions project to the to targets in the brain stem and thalamus. The cartoon
subthalamic nucleus (STN). The striatum is split into spike-trains illustrate the basic process of disinhibition in
two populations of projection neurons according to the “direct” pathway: continuous firing from the GPi/SNr
whether they predominantly express the D1 or D2 receptor tonically inhibits their targets; brief phasic activity in
for dopamine: the populations and their targets respec- striatum, driven by coordinated cortical activity, can
tively form the “direct” and “indirect” pathways. The pause that tonic inhibition, allowing the targets to fire
Competition basal ganglia’s intrinsic circuitry, such that

To be able to resolve a competition between there are separate channels for arm, leg, and
actions first requires a substrate for separately face representations (Alexander and Crutcher
representing each of those actions. A long- 1990; Romanelli et al. 2005). Similar topo-
standing concept in basal ganglia research is graphic maps have been proposed for the other
the existence of parallel loops running through macroscopic loops (Alexander and Crutcher
the basal ganglia (Alexander et al. 1986; Mid- 1990). Moreover, within these limb representa-
dleton and Strick 2000). At the macroscopic tions, there may be additional discrete channels
level, independent domains of the basal ganglia corresponding to particular movements,
have been identified by the closed send-and- demonstrated in striatum by microstimulation
return loops originating in broad cortical regions (Alexander and DeLong 1985) and markers for
(Alexander et al. 1986). Each macroscopic metabolic activity during behavior (Brown and
domain is thought to subdivide into a further Sharp 1995), and to repeated representations of
large set of discrete parallel loops, termed chan- body map locations (Brown et al. 1998). Com-
nels here to distinguish them from the macro- putational models of the basal ganglia have pro-
scopic-scale loops. For example, the posed that it is these channels which form the
somatotopic map found within the striatal basis for action representation, and so the reso-
motor domain is maintained throughout the lution of competition is expressed by the
B 354 Basal Ganglia: Mechanisms for Action Selection
selection of a particular channel in the output of

the basal ganglia (Gurney et al. 2001a).
The basal ganglia circuit contains a number of
mechanisms potentially suited to resolving com-
petition between competing channels. One can-
didate mechanism is the inhibitory network
formed by the local axon collaterals of the pro-
jection neurons in striatum, which could underpin
a winner-takes-all competition between input sig-
nals (Alexander and Wickens 1993). A problem
with this idea is that these feedback connections
are weak and asymmetric so that a classical win-
ner-takes-all competition is unlikely to be
supported (Plenz 2003). Even if it is, because of
the range of the local axon collaterals, this inter-
action can only extend over a short distance in the
striatum, on the order of 200–300 mm (Wickens
1997; Humphries et al. 2010). Consequently, any
dynamics within striatum are likely only to
resolve very local competitions between cortical
inputs (Gurney et al. 2001b). Basal Ganglia: Mechanisms for Action Selection,
The competition between a set of channels Fig. 2 Competition resolution and switching by the
most likely takes place at the inputs to the basal basal ganglia. Actions are thought to be represented in
ganglia output nuclei, the SNr and GPi. As parallel loops or channels running throughout the basal
ganglia, with each action’s relative salience computed by
shown in Fig. 2, each population of striatal neu- inputs from cortex. The bars in each box are cartoons of
rons projects to a corresponding population of activity of each population in each structure. Each striatal
SNr or GPi neurons, but the output of each population projects to a corresponding SNr/GPi popula-
population of STN neurons also contacts neigh- tion, providing focussed inhibitory input; STN output is
comparatively diffuse, projecting across neighboring
boring channels in SNr and GPi. This creates an channels and providing diffuse excitatory input. The result
off-center, on-surround network ideal for is an off-center, on-surround network in SNr/GPi, where
resolving competition between action signals the strongest signal from among the striatal populations is
(Mink 1996; Gurney et al. 2001a; Humphries able to inhibit the firing of its target SNr/GPi population,
while at the same time the combined excitatory output of
et al. 2006; Leblois et al. 2006) as illustrated in the STN counteracts the effect of the other inhibitory
Fig. 2. signals from striatum
A key insight from computational models is
that the on-surround effect of the combined STN not encode that action, but rather encodes the
output can lead to SNr/GPi activity in many simultaneous nonselection of another potential
channels increasing above their tonic rate at the action. Moreover, the models show that the num-
time of action selection – a phenomenon termed ber of neurons with such increased activity, time-
contrast enhancement. This makes functional locked to an action, will be in the majority and
sense: an enhanced tonic inhibitory signal being hence will appear more frequently during neural
output from most channels indicates a definite recordings (Humphries et al. 2006).
nonselection of that action. However, when
observed in experimental data, such increases in Switching
activity, time-locked to movement, have been Once action selection has been achieved, then
interpreted as encoding that movement. By con- action switching is required: the sudden availabil-
trast, the models show how an increase in SNr/ ity of a more salient action through changes in the
GPi neuron output time-locked to an action does environment or the animal’s needs should
terminate the ongoing action and then switch to A further mystery was the role of STN feed-
selecting the new, more salient alternative. Com- back to the GPe (Fig. 1). An analysis by Gurney
putational models have again shown how the et al. (2001b) showed that this negative feedback
same off-center, on-surround design naturally loop provided capacity scaling for action selec-
handles this: the new input further increases the tion. Without a loop between STN and GPe, the B
STN output, canceling selection of the current diffuse STN output means that the total excit-
action, while the corresponding new, strong atory input to the SNr/GPi would grow in propor-
inhibitory signal from striatum inhibits its target tion to the number of active inputs to the basal
population in GPi/SNR (Gurney et al. 2001b; ganglia; with a sufficient number of inputs, the
Humphries and Gurney 2002). Embedding these excitatory input would grow large enough to
models in mobile robots has shown how smooth overwrite all focussed inhibition from the stria-
behavioral sequences naturally emerge from the tum and prevent any selection. However, the
changing saliences of each action, without the negative feedback loop ensures that the STN
need for explicit storage of sequences (Khamassi output is scaled to the number of active inputs:
et al. 2005; Prescott et al. 2006). the more the inputs, the more the STN output tries
to grow, but is held in check by the GPe output
The “Indirect” Pathway: Prevention of which is itself a function of the STN output. In
Selection and Scaling of Selection this way, the basal ganglia’s capacity can scale
The contribution of the “indirect” pathway to and so can operate as a selection mechanism
selection has been more difficult to unravel. The independently of the number of currently com-
original box-and-arrow models proposed that this peting channels.
pathway acts to counteract the selection of an
action: increased inhibition of the GPe by its
striatal inputs would lead to enhanced STN out- Cross-References
put to SNr/GPi, thereby counteracting inhibition
they were receiving in the direct pathway ▶ Basal Ganglia: Overview
(Alexander and Crutcher 1990). Consequently, ▶ Decision-Making: Overview
it was proposed that any imbalance between the ▶ Reinforcement Learning in Cortical Networks
relative strengths of the direct and indirect path-
ways could lead to movement disorders (Albin
et al. 1989). Recent experimental studies have References
shown how selectively activating the striatal D1
or D2 projection neuron populations does indeed Albin RL, Young AB, Penney JB (1989) The functional
anatomy of basal ganglia disorders. Trends Neurosci
lead to opposite effects on behavior: stimulation
12:366–375
of D1 projection neurons facilitates behavior, Alexander GE, Crutcher MD (1990) Functional architec-
whereas stimulation of D2 neurons suppresses ture of basal ganglia circuits: neural substrates of par-
ongoing behavior (Kravitz et al. 2010). allel processing. Trends Neurosci 13:266–272
Alexander GE, DeLong MR (1985) Microstimulation of
The computational models of Frank and col-
the primate neostriatum. I. Physiological properties of
leagues (Frank et al. 2004; Frank 2005) have striatal microexcitable zones. J Neurophysiol
extended these ideas by proposing that recruit- 53:1401–1416
ment of these “go” and “no-go” pathways by Alexander ME, Wickens JR (1993) Analysis of striatal
dynamics: the existence of two modes of behaviour.
cortical input during learning would depend on J Theor Biol 163:413–438
the level of available dopamine. They showed Alexander GE, DeLong MR, Strick PL (1986) Parallel
these models could account for a range of learn- organization of functionally segregated circuits
ing biases exhibited by Parkinson’s disease linking basal ganglia and cortex. Annu Rev Neurosci
9:357–381
patients, providing support for the go versus no-
Brown LL, Sharp FR (1995) Metabolic mapping of rat
go competition between the direct and indirect striatum: somatotopic organization of sensorimotor
pathways. activity. Brain Res 686:207–222
B 356 Basal Ganglia: Songbird Models
Brown LL, Smith DM, Goldbloom LM (1998) Organizing Marsden CD, Obeso JA (1994) The functions of the basal
principles of cortical integration in the rat neostriatum: ganglia and the paradox of stereotaxic surgery in
corticostriate map of the body surface is an ordered Parkinson’s disease. Brain 117:877–897
lattice of curved laminae and radial points. J Comp Middleton FA, Strick PL (2000) Basal ganglia and cere-
Neurol 392:468–488 bellar loops: motor and cognitive circuits. Brain Res
Cherniak C (1994) Component placement optimization in Brain Res Rev 31:236–250
the brain. J Neurosci 14:2418–2427 Mink JW (1996) The basal ganglia: focused selection and
Chevalier G, Deniau JM (1990) Disinhibition as a basic inhibition of competing motor programs. Prog
process in the expression of striatal function. Trends Neurobiol 50:381–425
Neurosci 13:277–280 Plenz D (2003) When inhibition goes incognito: feedback
DeLong MR (1990) Primate models of movement disor- interaction between spiny projection neurons in striatal
ders of basal ganglia origin. Trends Neurosci function. Trends Neurosci 26:436–443
13:281–28 Prescott TJ, Redgrave P, Gurney K (1999) Layered control
Frank MJ (2005) Dynamic dopamine modulation in the architectures in robots and vertebrates. Adapt Behav
basal ganglia: a neurocomputational account of cogni- 7:99–127
tive deficits in medicated and nonmedicated Parkin- Prescott TJ, Montes-Gonzalez FM, Gurney K, Humphries
sonism. J Cogn Neurosci 17:51–72 MD, Redgrave P (2006) A robot model of the basal
Frank MJ, Seeberger LC, O’Reilly RC (2004) By carrot or ganglia: behavior and intrinsic processing. Neural
by stick: cognitive reinforcement learning in parkin- Netw 19:31–61
sonism. Science 306:1940–1943 Redgrave P, Prescott TJ, Gurney K (1999) The basal
Gurney K, Prescott TJ, Redgrave P (2001a) ganglia: a vertebrate solution to the selection problem?
A computational model of action selection in the Neuroscience 89:1009–1023
basal ganglia I: a new functional anatomy. Biol Romanelli P, Esposito V, Schaal DW, Heit G (2005)
Cyber 85:401–410 Somatotopy in the basal ganglia: experimental and
Gurney K, Prescott TJ, Redgrave P (2001b) clinical evidence for segregated sensorimotor chan-
A computational model of action selection in the nels. Brain Res Brain Res Rev 48:112–128
basal ganglia II: analysis and simulation of behaviour. Wickens J (1997) Basal ganglia: structure and computa-
Biol Cyber 85:411–423 tions. Netw Comput Neural Syst 8:R77–R109
Hikosaka O, Wurtz RH (1985) Modification of saccadic
eye movements by GABA-related substances. II.
Effects of muscimol in monkey substantia nigra pars
reticulata. J Neurophysiol 53:292–308
Hikosaka O, Takikawa Y, Kawagoe R (2000) Role of the Basal Ganglia: Songbird Models
basal ganglia in the control of purposive saccadic eye
movements. Physiol Rev 80:953–978 Arthur Leblois1 and Ran Darshan2
Humphries MD, Gurney KN (2002) The role of intra- 1
Centre de Neurophysique, Physiologie et
thalamic and thalamocortical circuits in action selec-
tion. Netw Comput Neural Syst 13:131–156 Pathologie (UMR CNRS 8119), Université Paris
Humphries MD, Stewart RD, Gurney KN Descartes, Paris, France
(2006) A physiologically plausible model of action 2
Interdisciplinary Center for Neural Computation
selection and oscillatory activity in the basal ganglia.
(ICNC), The Hebrew University of Jerusalem,
J Neurosci 26:12921–12942
Humphries MD, Wood R, Gurney K (2010) Jerusalem, Israel
Reconstructing the three-dimensional GABAergic
microcircuit of the striatum. PLoS Comput Biol 6:
e1001011
Khamassi M, Lacheze L, Girard B, Berthoz A, Guillot
Definition
A (2005) Actor-critic models of reinforcement learn-
ing in the basal ganglia: from natural to artificial rats. Songbirds produce complex vocalizations,
Adapt Behav 13:131–148 a behavior that depends on the ability of juveniles
Kravitz AV, Freeze BS, Parker PRL, Kay K,
to imitate the song of an adult. Song learning
Thwin MT, Deisseroth K, Kreitzer AC (2010)
Regulation of parkinsonian motor behaviours by relies on a specialized basal ganglia-
optogenetic control of basal ganglia circuitry. Nature thalamocortical loop. Several computational
466:622–626 models have examined the role of this circuit in
Leblois A, Boraud T, Meissner W, Bergman H, Hansel
song learning, shedding light on the neurobiolog-
D (2006) Competition between feedback loops under-
lies normal and pathological dynamics in the basal ical mechanisms underlying sensorimotor
ganglia. J Neurosci 26:3567–3583 learning.
Basal Ganglia: Songbird Models 357 B
Detailed Description X receives dense dopaminergic innervation and
contains neuron types homologous to both the
Songbirds use learned vocalizations to communi- mammalian striatal and pallidal neurons, forming
cate during courtship or aggressive behaviors. similar circuits (Farries and Perkel 2002; Leblois
These vocalizations, called song, require fast et al. 2009). Songbirds and their specialized B
coordination of laryngeal and respiratory mus- BG-thalamocortical loop devoted to a naturally
cles. Songbirds learn their song as juveniles learned complex sensorimotor task offer a unique
through a long process comprising two sequential opportunity to study the function of the BG in skill
phases: the juvenile first listens to and memorizes learning and execution (Doupe et al. 2005).
one or more tutor songs and then uses auditory
feedback to match its song to the memorized BG Functions in Songbirds: Experimental
model through trial and error. Background
While song production is under the control of The BG loop in songbirds is necessary for song
two cortical nuclei, HVC (used as a proper name) learning but not for song production per se. BG
and the robust nucleus of the arcopallium (RA, neurons respond selectively to playbacks of the
Fig. 1a), song learning and plasticity strongly rely bird’s own song (Doupe and Solis 1997) and were
on the basal ganglia (BG, Scharff and initially thought to convey auditory feedback sig-
Notthebohm 1991). Interestingly, songbirds nals to be compared with a stored template of the
display a specialized portion of their BG- tutor song. However, no neuronal correlate of
thalamocortical circuitry devoted to song learning such comparison can be found in the BG-
and plasticity (Fig. 1b; Brainard and Doupe 2002) thalamocortical loop (Leonardo 2004). Auditory
and homologous to the mammalian motor BG- feedback-related activity has however been
thalamocortical loop (Fig. 1c; Reiner et al. 1998). reported in upstream cortical nuclei (HVC),
In particular, the song-related BG nucleus Area where similar neuronal responses can be observed
Basal Ganglia: Songbird Models, Fig. 1 The song- nidopallium; RA, robust nucleus of the arcopallium; VP,
related BG-thalamocortical network in songbirds and its ventral pallidum; VTA, ventral tegmental area; nXIIts,
homologue circuit in mammals. In panels a, b, and c, the supraspinal nucleus. (b) Diagram of the synaptic connec-
nature of synaptic transmission is indicated by the end of tivity in the song-related BG-thalamocortical network of
the arrow: black arrows for excitatory (glutamatergic) con- songbirds. (c) Diagram of the synaptic connectivity in the
nections, black disks for inhibitory (GABAergic) connec- motor loop of the mammalian BG-thalamocortical net-
tions, and gray diamonds for dopaminergic connections. (a) work. GPe, globus pallidus pars externa; GPi, globus
Schematic parasagittal representation of the song system. pallidus pars interna; SNc, substantia nigra pars compacta;
DLM, dorsolateral nucleus of the anterior thalamus; STN, subthalamic nucleus; VA, ventral anterior nucleus of
LMAN, lateral magnocellular nucleus of the anterior the thalamus; VL, ventral lateral nucleus of the thalamus
in response to syllable production or playback

(“mirror neurons,” Prather et al. 2008). During
song production, the BG-thalamocortical loop
introduces motor variability allowing vocal explo-
ration (Kao et al. 2005; Olveczky et al. 2005), as
needed in a reinforcement learning (RL) frame-
work, and can guide adaptive changes in song to
minimize errors (Andalman and Fee 2009).
Reinforcement Learning in Songbirds

Most models of BG-dependent learning in song-
birds are in a RL framework (but see Ganguli and
Hahnloser 2013). In these models, a timing signal
is assumed to be produced in HVC, in which
output is sent to the BG to serve as a clock input
during learning. Based on this clock input, the RL
circuit learns to produce the correct motor ges- Basal Ganglia: Songbird Models, Fig. 2 Architecture
of the basic song RL models (Doya and Sejnowski 1998;
tures at each time step. In the RL framework, the
Fiete et al. 2007). Dashed box represents the search/exper-
agent learns to correlate random motor explora- imenter element. Gray box represents the critic. Plastic
tions with fluctuations in a reward signal in order synapses are displayed in wide black arrows
to select motor patterns that lead to the highest
reward. For example, changes in the song (the
motor output) that increase the amount of expected
reward should be implemented, whereas song results in a plasticity in HVC-RA synapses after
changes leading to lower expected reward should each epoch, i.e., at the end of song rendition.
be discarded. Implementation of such RL frame- Fiete et al. (2007) elaborated on the Doya and
work in the song-related neuronal circuitry must Sejnowski model in a biologically realistic net-
include four components: (1) an “actor” that pro- work of spiking neurons. While the location of
duces the song; (2) a mechanism for exploration, the critic in their model was only speculative, the
implemented in a variability circuit (Fee and other elements of the model were similar to Doya
Goldberg 2011), a searcher (Doya and Sejnowski and Sejnowski (1998): plasticity took place in
1998), or an experimenter (Fiete et al. 2007); HVC-to-RA synapses and LMAN was again the
(3) a comparator circuit or “critic” (Doya and search element, called the “experimenter.”
Sejnowski 1998) that computes the reward signal Besides its realistic nature, their model differs
by evaluating the produced song with respect from previous models by utilizing a new learning
to the memorized template; (4) and a learning algorithm relying on node perturbations, where
mechanism to modify motor output with time. LMAN neurons perturb the activity of RA neu-
rons instead of their synaptic efferences from
Models of Song Learning and the HVC (Fig. 2), and an online reward signal during
Involvement of the BG-Thalamocortical Loop song. Using such a realistic learning algorithm,
In a first RL model for song learning, Doya and the circuit learns a simple song in a biologically
Sejnowski (1998) proposed that nucleus LMAN plausible number of song renditions.
could act as the “searcher,” perturbing each syn- Recently, Fee and Goldberg (2011) have
aptic connection between HVC and RA indepen- published a conceptual model that argues that
dently, while the BG would evaluate each the BG-thalamocortical loop is not a comparator
produced song using tutor-selective neurons circuit but is rather biasing the motor system
(Fig. 2). This weight perturbation algorithm (Fig. 3). Based on ideas taken from the classical
Basal Ganglia: Songbird Models 359 B
Template Comparison, Efference Copy, or
Inverse Model?
In the RL models discussed above, a global
reward signal evaluating auditory feedback qual-
ity is delivered by the critic. Because of delays B
that are inherent in the transformation of neural
activity into vocal gestures and in the auditory
processing of produced sounds, the reward signal
is necessarily delayed with respect to the neuro-
nal activity underlying the rewarded gesture. In
other words, the reward is likely temporally
imprecise. To overcome this delay problem, an
eligibility trace can be used (Doya and Sejnowski
1995; Fiete et al. 2007).
A different solution for the delay problem was
given by Troyer and Doupe (2000, Fig. 4). While
they still assume that the AFP evaluates the song,
they propose that evaluation is based on an
Basal Ganglia: Songbird Models, Fig. 3 Architecture efference copy of the song’s motor program. In
of the more complete song RL model including BG cir- the first stage, a motor-to-sensory mapping is
cuits (Fee and Goldberg 2011). Arrows and color code as
in Fig. 2 built by implementing Hebbian learning at syn-
apses between HVC neurons projecting to RA
(HVCRA), assumed to be motor, and HVC neu-
rons projecting to Area X (HVCX), assumed to be
implementation of RL in the mammalian motor auditory. The bird thereby learns to predict audi-
BG loop (Wickens et al. 2003), they hypothesize tory consequences of its motor gestures. Such
that the song-related BG receive an evaluation a mapping is usually termed a “forward model”
signal from neuromodulators such as dopamine or “efference copy.” After implementing the
and use this signal to select the relevant motor motor-to-sensory transformation (blue arrow in
gestures at the right time. As in mammals, the Fig. 4), the AFP neurons displaying tuned
avian BG-thalamocortical circuit and the responses to tutor song evaluate the motor
descending connections to motoneurons form (efference) copy, whereas auditory feedback to
a topographically organized circuit displaying HVCX is “gated on” by adaptation mechanisms to
segregated “motor channels.” In this model, avoid interference. In this stage the AFP acts like
LMAN again drives variability in song produc- a critic. As in Doya and Sejnowski, synapses
tion and sends a copy of this variability signal to within RA and from HVCRA to RA are modified
striatal neurons. Receiving a convergent timing to optimize song.
signal (or “context”) from HVC and a global The last model, proposed by Ganguli and
dopaminergic reward signal, the striatal neurons Hahnloser (2013), consists of two brain areas,
are considered as coincidence detectors and an auditory area and a motor area, represented
select the right motor variation in the right con- in a very simplified framework. Their model
text in each motor channel. Synaptic plasticity at relies on three already mentioned assumptions:
HVC-to-X synapses then allows the acquisition production of highly variable vocalization dur-
of correct motor biases. Although this model has ing learning, gating mechanisms for specific
not been computationally implemented yet, it is synaptic inputs, and a synaptic mechanism for
algorithmically similar to Fiete et al. (2007), with Hebbian learning. It is not based on a RL frame-
the learning occurring at HVC-to-X synapses work, and no evaluation of the song or compar-
instead of HVC-to-RA synapses. ison with the tutor’s song is needed. Instead, the
Basal Ganglia: Songbird Models, Fig. 4 Schematic represent the motor-auditory transformation in (a) and the
description of the efference copy learning (a) and song “efference copy” of this transformation in (b). Other color
learning (b) in the Troyer and Doupe (2000). Blue arrows code as in Fig. 2
bird learns an inverse model through the map-

ping of sensory (auditory) commands to the
appropriate motor commands, relying solely
on Hebbian learning. First, random vocaliza-
tions induce Hebbian learning between auditory
and motor representation of the produced
sounds (Fig. 5a), creating an inverse model.
Later, the auditory feedback is gated off, while
an alternative input to the auditory area acti-
vates sequentially the neuronal assemblies
representing the tutor song (Fig. 5b). Thanks
to the preceding learning of the inverse model
in the auditory-motor connection, activation of
the song representation in the auditory area Basal Ganglia: Songbird Models, Fig. 5 Architecture
drives the right motor pattern in the postsynap- of the inverse model for song learning (Ganguli and
Hahnloser 2013). Wide blue arrows represent the motor-
tic motor area. Thereby, the tutor song can be auditory transformation. (a) Learning the inverse model
produced instantaneously. (in wide black). (b) Song production. After learning the
inverse model (dashed blue arrow), auditory feedback and
the searcher element are gated, while the representation of
syllables of the tutor song is sequentially activated in the
Conclusion auditory area
In summary, the models we have presented an inverse model, in the BG or elsewhere, may
address the question of how sensorimotor learn- also participate to imitation. Whether these two
ing leads to the faithful copy of a previously types of learning coexist in the same neural cir-
imprinted tutor song. The BG-thalamocortical cuits and how they may be combined to achieve
loop plays the role of the actor-critic reinforce- song learning remains to be elucidated in future
ment circuit in these models, although learning of theoretical work.
Bayesian Inference with Spiking Neurons 361 B
References parts of the zebra finch song system: implications for
vocal learning. J Neurosci 11:2896–2913
Andalman AS, Fee MS (2009) A basal ganglia-forebrain Troyer TW, Doupe AJ (2000) An associational model of
circuit in the songbird biases motor output to avoid birdsong sensorimotor learning I. Efference copy and
vocal errors. Proc Natl Acad Sci the learning of song syllables. J Neurophysiol
USA 106:12518–12523 84:1204–1223 B
Brainard MS, Doupe AJ (2002) What songbirds teach us Wickens JR, Reynolds JN, Hyland BI (2003) Neural
about learning. Nature 417:351–358 mechanisms of reward-related motor learning. Curr
Doupe AJ, Solis MM (1997) Song- and order- Opin Neurobiol 13:685–690
selective neurons develop in the songbird anterior
forebrain during vocal learning. J Neurobiol
33:694–709
Doupe AJ, Perkel DJ, Reiner A, Stern EA (2005) Bird-
brains could teach basal ganglia research a new song. Bayesian Estimation
Trends Neurosci 28:353–363
Doya K, Sejnowski T (1995) A novel reinforcement ▶ Estimation of Neuronal Firing Rate
model of birdsong vocalization learning. Adv Neural
Inf Process Syst 7:101–108.
Doya K, Sejnowski TJ (1998) A computational model of
birdsong learning by auditory experience and auditory
feedback. In: Poon PWF, Brugge JF (eds) Central
auditory processing and neural modeling. Plenum,
Bayesian Inference with Spiking
New York, pp 77–88 Neurons
Farries MA, Perkel DJ (2002) A telencephalic nucleus
essential for song learning contains neurons with phys- Sophie Deneve
iological characteristics of both striatum and globus
pallidus. J Neurosci 22:3776–3787
Group for Neural Theory, École Normale
Fee MS, Goldberg JH (2011) A hypothesis for basal Supérieure Paris, Paris, France
ganglia-dependent reinforcement learning in the song-
bird. Neuroscience 198:152–170
Fiete IR, Fee MS, Seung HS (2007) Model of birdsong
learning based on gradient estimation by dynamic per-
Definition
turbation of neural conductances. J Neurophysiol
98:2038–2057 Neurons respond reliably to external stimuli
Ganguli S, Hahnloser RHR (2013) Bird song learning despite the fact that their input is noisy and the
without reinforcement: the Hebbian self-organization
of sensorimotor circuits. Soc Neurosci Abstr 105.12/
state of the environment changes constantly (e.g.,
YY4 stimuli appear or disappear unpredictably). The
Kao MH, Doupe AJ, Brainard MS (2005) Contributions of Bayesian spiking neuron (BSN) framework inter-
an avian basal ganglia-forebrain circuit to real-time prets neural dynamics as a form of probabilistic
modulation of song. Nature 433:638–643
Leblois A, Bodor AL, Person AL, Perkel DJ (2009) Milli-
inference and describes how neurons and neural
second timescale disinhibition of thalamic neurons by circuits should be optimally tuned to the temporal
relief of single spike basal ganglia GABAergic veto. statistics of their inputs and the level of noise.
J Neurosci 29:15420–15433 Output spike trains communicate probability dis-
Leonardo A (2004) Experimental test of the birdsong
error-correction model. Proc Natl Acad Sci
tributions over stimulus states to downstream
USA 101:16935–16940 neurons using a predictive coding strategy.
Olveczky BP, Andalman AS, Fee MS (2005) Vocal exper-
imentation in the juvenile songbird requires a basal
ganglia circuit. PLoS Biol 3:e153
Prather JF, Peters S, Nowicki S, Mooney R (2008) Precise
auditory-vocal mirroring in neurons for learned vocal
communication. Nature 451:305–310 General Framework
Reiner A, Medina L, Veenman CL (1998) Structural and In BSN formalism, prior assumptions about the
functional evolution of the basal ganglia in vertebrates.
Brain Res Brain Res Rev 28:235–285
input tuning curves, noise, and temporal dynam-
Scharff C, Nottebohm F (1991) A comparative study of ics of the stimulus are treated as the parameters of
the behavioral deficits following lesions of various an internal predictive model (Lochmann and
B 362 Bayesian Inference with Spiking Neurons
c
a b
Time
xt xt xt1 xt2 xt3
Time Time
st1 s2t st1 s2t st1 s2t s3t s4t
L Li Lt1 Lt2 Lt2

−L
−Li
ot i oit
Time Time Time
Bayesian Inference with Spiking Neurons, (b) Case of a single continuous variable evolving with
Fig. 1 Examples of internal predictive models and their Markovian “Brownian motion.” In this case, the log pos-
BSN implementations. (a) Case of single binary stimulus terior is represented by a population of neurons whose
(dashed circle), modulating two input ring rates (dashed spike trains (colored vertical lines) track the stimulus.
arrows and line), generates two input spike trains (vertical Predictive coding requires lateral inhibitory connections
lines). A single spiking neuron (yellow star) receiving (thin arrows). (c) Case of 3 different binary stimuli
these convergent synaptic inputs (plain connections) can co-modulating the synaptic inputs. Explaining away
represent the log probability of stimulus present in its requires divisive presynaptic inhibition between neurons
output spikes, using predictive coding (thin plain arrow). (dashed connections)
Deneve 2011). Such an internal model represents equations, one corresponding to synaptic integra-
how the “hidden” stimulus state causes the tion of inputs (inference) and the other to
“observed” noisy inputs. These assumptions are a dynamic neural threshold dependent on recent
in turn reflected in the structure and dynamics of spike history (encoding).
a corresponding neural network. For example, the Inference consists of computing the probabil-
presence or absence of an external sensory stim- ity distribution over the stimulus state xt, given
ulus can be represented as a binary “hidden” the feed-forward synaptic inputs received so far,
variable modulating the input rates. Such internal s0!t. This can be expressed in terms of the log
model can be implemented in a single spiking posterior probability, L(x, t) = log(p(xt|s0!t).
neuron (Fig. 1a, Deneve 2008a). Using simplifying assumptions such as
In turn, the parameters of the internal model (1) Poisson distributed input spike trains and
fully predict the biophysical properties of the (2) Markovian dynamics for stimulus xt (i.e., xt
spiking neurons, e.g., synaptic time constants, only depends on xt –dt), this equation takes the
adaptation, synaptic weights, and ring dynamics. general form:
Since these parameters also reflect the “natural”
(i.e., expected) statistics of the synaptic inputs,
@Lðx, tÞ X
these optimal biophysical parameters can be ¼ fðLðx, tÞÞ þ wi ðxÞsi ðtÞ (1)
@t i
learned in an unsupervised way (through spike-
time-dependent plasticity rules) (Deneve 2008b;
Mongillo and Deneve 2008). where si(t) is an input spike train while the “syn-
aptic weights” wi(x) depend on the input tuning
Dynamics of Bayesian Spiking Neurons curves. f is a (typically nonlinear) “leak” term
BSNs are similar to adaptive “integrate-and-re” dependent on the dynamics of the hidden
neurons and are described by two dynamic variable (e.g., how quickly the stimulus changes).
Bayesian Inference with Spiking Neurons 363 B
b c wik
a
Membrane potential
RF left surround
Log Probability
Receptive field
RF Low cont
B
i
L
Time L–L
Time RF High cont
Bayesian Inference with Spiking Neurons, multiple overlapping “blob-shaped” stimuli are
Fig. 2 Predictive coding and implication for contextual represented (dashed grey line), the receptive field of the
modulation of receptive fields. (a) The log probability of corresponding neuron is highly dependent on context. For
stimulus, Lt = L(x, t), is tracked over time by an output example, it is narrow and center surround for high-
estimate L^t corresponding to a leaky integration of the contrast stimuli (plain red), wide and without surround
output spike train. (b) This can be implemented by an for low-contrast stimuli (dot-dashed blue), and repulsed in
adaptive integrate-and-re neuron whose membrane poten- the presence of a stimulus in the left surround (dashed
tial is a function of the prediction error Lt L^t . (c) When green)
This equation can correspond to synaptic integra- Extension to Multiple Stimuli

tion in single neuron (for binary xt) or in The framework can be extended to the tracking of
a recurrent neural network (for multinomial or more than one stimulus at a time (Fig. 1c). This is
continuous xt, as in Fig. 1b). particularly relevant for the processing of sensory
Encoding consists of representing an estimate scenes, since the same sensory inputs can usually
^
Lðx, tÞ of the log posterior in the output spike be explained by different combinations of
trains. In order to do this, the BSN uses overlapping features. For example, different
a “greedy” predictive coding scheme, i.e., neu- speech sounds activate highly overlapping sets
rons res whenever their membrane potential, of auditory receptors. The auditory input is
defined as the prediction error V j ðtÞ ¼ explained by either one or the other phoneme,

^
kj L L (where kj is a kernel specific to but not both. Thus, neurons selective to similar
each neuron), exceeds a particular threshold. As phonemes (e.g., “ba” and “da”) will need to com-
a result, each new spike decreases the coding pete: if they did not, neurons would have highly
(prediction) error Lðx, tÞ L^ðx, tÞ: With respect correlated, nonselective responses. This phenom-
to self-consistency, L^ðx, tÞ is assumed to be the enon is called “explaining away.”
result of postsynaptic integration of the output Let us assume that the sensory input is
spike trains in downstream neurons, e.g., caused by K different stimuli xk(t) combining
linearly to cause the sensory scene, i.e., each
input si(t) is a Poisson process with instanta-
@ L^ðx, tÞ X neous firing rate fi(t) = kwikxk(t). In a BSN
t ¼ L^ðx, tÞ þ kj ðxÞoj ðtÞ (2)
@t j
implementation of this internal model, the
encoding is identical, but the inference needs
to implement explaining away. The network
The result can be interpreted as a recurrent equations are as same before, except that the
network of adaptive leaky integrate-and-re neu- synaptic weights are no longer fixed (as in
rons with lateral connections ki * kj (Fig. 2b). By Eq. 1), but divisively gain modulated by the
always maintaining a small prediction error, context. More precisely, the effective weights
BSNs represent the posterior probability effi- w~ ij ðtÞ between input i and feature detector k are
ciently (Fig. 2a). given by
B 364 Bayesian Modelling of Neural Recordings
w interpret spikes as samples from the posterior

w~ ik ðtÞ ¼ X ik (3)
1þ w x^ ðtÞ
j6¼k ij j
probability distribution (rather than a determinis-
tic representation of the log posterior) (Berkes
where x^j ðtÞ is the prediction (i.e., the expected et al. 2011; Buesing et al. 2011). The
value) of feature j at time t. This can be “probabilistic population coding” (PPC)
implemented by lateral divisive inhibition in approach represents log-posteriors in the output
a recurrent neural network (Fig. 1c). Each input ring rates. This can be implemented in a spiking
to each neuron is presynaptically shunted network using random spike generators or large
according to how well this input is predicted by balanced spiking networks (Beck et al. 2008), the
other neurons. underlying assumption being that the added noise
is neglectable compared to the uncertainty
Some Applications of the BSN Framework already present in the input.
One of the most important implications of the
BSN framework is that the “Poisson-like” vari-
ability observed in cortical spike trains is References
a natural consequence of predictive coding. It
reflects sensory noise and redundancy between Beck JM, Ma WJ, Kiani R, Hanks T, Churchland AK,
Roitman J, Shadlen MN, Latham PE, Pouget A (2008)
neurons (e.g., many neurons have similar coding Probabilistic population codes for Bayesian decision
kernels), but not neural noise. The variability making. Neuron 60(6):1142–1152
persists even when the inputs are not Poisson Berkes P, Orban G, Lengyel M, Fiser J (2011) Spontane-
distributed and even when the input is completely ous cortical activity reveals hallmarks of an optimal
internal model of the environment. Science
deterministic (Bourdoukan et al. 2012). 331(6013):83–87
Moreover, this framework assumes highly Boerlin M, Deneve S (2011) Spike-based population cod-
dynamical and adaptive response properties. ing and working memory. PLoS Comput Biol 7(2):
BSNs provide “ideal observer” models of single e1001080
Bourdoukan R, Barrett D, Machens C, Deneve S (2012)
neurons or microcircuit and thus are useful tools Learning optimal spike-based representations. In:
to relate neural spike trains with their hypotheti- Bartlett P, Pereira FCN, Burges CJC, Bottou L, Wein-
cal functions. While those models are highly test- berger KQ (eds) Advances in neural information
able (and disprovable), they largely remain to be processing systems, vol 25, pp 2294–2302
Buesing L, Bill J, Nessler B, Maass W (2011) Neural
tested experimentally. dynamics as sampling: a model for stochastic compu-
More specifically, BSNs have been shown to tation in recurrent networks of spiking neurons. PLoS
achieve quasi-optimal information transmission Comput Biol 7(11):e1002211
in single neurons (Lochmann and Deneve 2008). Deneve S (2008a) Bayesian spiking neurons i: inference.
Neural Comput 20(1):91–117
Populations of BSN neurons can represent slowly Deneve S (2008b) Bayesian spiking neurons ii: learning.
varying variables (such as position or direction of Neural Comput 20(1):118–145
motion) and implement optimal working memo- Lochmann T, Deneve S (2008) Information transmission
ries (Boerlin and Deneve 2011). Finally, BSNs with spiking Bayesian neurons. New J Phys
10:055019, 19pp
representing overlapping stimuli can account for Lochmann T, Deneve S (2011) Neural processing as
a large number of contextual modulation of sen- causal inference. Curr Opin Neurobiol 21(5):774–781
sory responses, such as divisive normalization, Mongillo G, Deneve S (2008) Online learning with hidden
adaptation, surround suppression, and repulsion, Markov models. Neural Comput 20(7):1706–1716
as illustrated in Fig. 2c (Lochmann and
Deneve 2011).
Relation to Other Approaches Bayesian Modelling of Neural

In addition to the BSN, several other frameworks Recordings
have proposed different encoding of uncertainty
with spikes. For example, sampling models ▶ Electrophysiology Analysis, Bayesian
Behavioural Analysis, Bayesian 365 B
and our prior beliefs. Bayes’ rule stipulates how
Bayesian Neural Data Analysis to form the posterior from the prior and
likelihood:
▶ Electrophysiology Analysis, Bayesian
PðDjH ÞPðH Þ B
PðHjDÞ ¼ X
H
PðDjH ÞPðHÞ
Beat The denominator is sometimes known as the
marginal likelihood. When H is continuous, the
▶ Rhythm Perception: Pulse and Meter
summation is replaced by integration. A practical
overview of Bayesian analysis can be found in
Kruschke (2010). Below we discuss some of the
Behavioural Analysis, Bayesian relevant issues that arise in applying Bayesian
analysis to behavioral data.
Samuel J. Gershman
Department of Brain and Cognitive Sciences, Posterior Probabilities Versus P-values
Massachusetts Institute of Technology, The classical approach to behavioral data analy-
Cambridge, MA, USA sis is based on the null hypothesis significance
testing (NHST) framework, in which the signifi-
cance of a summary statistic is evaluated by cal-
Definition culating the probability of obtaining an equally or
more extreme value if the statistic was drawn
Bayesian analysis is a method for reasoning prob- from a null distribution. If the resulting p-value
abilistically about parameters of a model given is below some conventional threshold (e.g., 0.05),
some observed data. Applied to behavioral data, the null hypothesis is rejected and the result is
Bayesian analysis can be used to fit and compare considered significant. This framework is related
models of cognition. This approach to behavioral to Bayesian analysis: the p-value is the tail prob-
data analysis offers a number of advantages ability of the likelihood when H corresponds to
over classical (frequentist) analysis, including the null hypothesis.
a coherent representation of uncertainty, flexibil- In contrast to p-values, posterior probabilities
ity to handle complex models and missing data, quantify the evidential support for all possible
and an avoidance of pathologies inherent to sig- values of H. Often different values of H can plau-
nificance testing based on p-values. sibly have generated the data, and the posterior
captures this uncertainty. In contrast, the confi-
dence intervals of NHST express the range of
Detailed Description H that would not be rejected by a significance test.
Under NHST, when multiple tests are applied
Bayesian analysis starts with the specification of to a data set, the probability of erroneously
a joint distribution on the data D and parameters declaring a null effect significant increases. This
(or hidden variables) H. This joint distribution necessitates some kind of correction for multiple
can be broken down into two components: comparisons. Bayesian analysis does not require
P(D,H) = P(D|H) P(H). The first component, such corrections, since there is a single posterior
P(D|H), is known as the likelihood, and the sec- distribution that does not depend on how many
ond component, P(H), is known as the prior. ways you examine it.
Given some data, we are interested in the condi-
tional distribution, P(H|D), commonly known as Hierarchical Models
the posterior; this distribution specifies every- A powerful use of Bayesian analysis is the
thing we know about H given the observed data construction of hierarchical models in which
B 366 Bidirectional Brain-Machine Interface
uncertainty about the parameters of the prior is large hypothesis spaces. Consequently, most
captured by placing a “hyper-prior” on these algorithms for Bayesian analysis use approxima-
parameters (Lee 2011). An important example tions. Most commonly, these algorithms approx-
of this is modeling individual differences: each imate the posterior with a set of stochastic
subject in a study has a separate set of parameters, samples – the so-called Monte Carlo approxima-
but these parameters are coupled by virtue of tions (Robert and Casella 2004). As the number
being drawn from a common group-level distri- of samples increases, the approximation becomes
bution. One advantage of hierarchical models is increasingly accurate. There are several software
that they allow sharing of “statistical strength” packages that compute Monte Carlo approxima-
between subjects, such that the parameter esti- tions automatically given a probabilistic model
mates of one subject are informative about the (see Kruschke 2010 for an introduction), making
parameter estimates of another subject. At the Bayesian methods widely accessible for behav-
same time, subjects are allowed to express idio- ioral data analysis.
syncratic parameters. Hierarchical models can
thus strike a balance between forcing all subjects
to have the same parameters and modeling each References
subject separately.
Kass RE, Raftery AE (1995) Bayes factors. J Am Stat
Assoc 90:773–795
Model Comparison
Kruschke JK (2010) Doing Bayesian data analysis:
An important problem in cognitive modeling is a tutorial introduction with R and BUGS. Academic,
model comparison. The Bayesian approach to Burlington
this problem centers on calculating the Bayes Lee MD (2011) How cognitive modeling can benefit from
hierarchical Bayesian models. J Math Psychol 55:1–7
factor (Kass and Raftery 1995), which is the
Robert CP, Casella G (2004) Monte Carlo statistical
ratio of marginal likelihoods for two models. methods. Springer, New York
The Bayes factor rewards models that fit the Vanpaemel W (2010) Prior sensitivity in theory testing: an
data better but penalizes model complexity. Intu- apologia for the Bayes factor. J Math Psychol
54:491–498
itively, complex models are able to fit a given
data set better than simple models, but complex
models also spread their probability mass over
many data sets, and consequently the probability
of any given data set will tend to be smaller. Bidirectional Brain-Machine
Whereas the posterior will often be relatively Interface
insensitive to the prior when the amount of data is
large, Bayes factors can be exquisitely sensitive ▶ Recurrent Brain-Computer Interfaces
to the choice of prior (Kass and Raftery 1995).
This has sometimes been considered a drawback
of Bayesian model comparison. However, others
(e.g., Vanpaemel 2010) have argued that prior Bifurcation Analysis
sensitivity is reasonable, since the prior should
be considered part of a fully specified cognitive William Barnett and Gennady Cymbalyuk
model. The Neuroscience Institute, Georgia State
Computation
For all but a small class of models, employing
Bayes’ rule exactly is computationally intracta- Definition
ble. This happens because the marginal likeli-
hood computation involves summing over all Bifurcation analysis provides theoretical guide-
possible hypotheses, which is infeasible for lines and a powerful analytical framework to tune
Bifurcation Analysis 367 B
the activity of neuronal models. A bifurcation is Bifurcation Analysis, Table 1 Dynamical laws arising
a qualitative change in the activity of a dynamical from specific bifurcations. The bifurcation parameter a
takes its critical value for bifurcation at a*
system, such as a transition from silence to spik-
ing. Bifurcation analysis identifies general Bifurcation Dependence
Homoclinic Period / ln(|a a|)
dynamical laws governing transitions between pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi B
SNIC Period / 1= ja a j
activity regimes in response to changes in con- pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Blue Sky catastrophe Period / 1= ja a j
trolling parameters. By using these quantitative pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Andronov-Hopf Amplitude / ja a j
laws, the temporal characteristics of a neuronal
model can be tuned to reproduce the characteris-
tics of experimentally recorded activity by navi- homoclinic bifurcation, the saddle-node bifurca-
gating the parameter space in the vicinity of tion on an invariant circle (for bursting and for
bifurcations. In most of the bifurcations tonic spiking), the blue sky catastrophe, and the
described here, the general shape of the Andronov-Hopf bifurcation (Table 1). We pre-
waveform – for example, the shape of action sent the dynamical laws generated by these bifur-
potentials in a time series – is preserved, while cations in the context of spiking, bursting, and
a specific temporal characteristic changes. transient activity.
Detailed Description Spiking
The knowledge of the bifurcations of a model Three bifurcations of closed periodic orbits are
neuron provides quantitative laws governing the ubiquitously implicated in the control of spiking
temporal characteristics of spiking activities. activity. A saddle-node bifurcation on invariant
Neurons can be envisaged as nonlinear dynami- circle (SNIC) and a homoclinic bifurcation
cal systems. Bifurcation analysis has been describe how the frequency of spiking depends
applied to the study of spiking, bursting, and on a controlling parameter a, e.g., the injected
chaotic neuronal systems (Rinzel 1987; current. A supercritical Andronov-Hopf bifurca-
Guckenheimer et al. 1993, 1997; Rinzel and tion describes the dependence of the amplitude of
Ermentrout 1998; Ermentrout and Terman 2010; oscillations on the bifurcation parameter.
Ghigliazza and Holmes 2004; Rabinovich A SNIC occurs when an equilibrium
et al. 2008). Biophysical parameters are used as possessing the properties of both a saddle and
controlling parameters, and changes in regimes a node – a saddle-node equilibrium – is born on
of activity can be identified concomitantly with a closed periodic orbit and the eigenvectors asso-
bifurcations (Rinzel 1978; Guckenheimer ciated with the half-stable direction of the saddle-
et al. 1993, 1997; Shilnikov and Cymbalyuk node equilibrium are tangent to the periodic orbit.
2005; Shilnikov et al. 2005; Medvedev 2006; This bifurcation describes the transition of spik-
Barnett and Cymbalyuk 2014). The association ing into silence (Gutkin and Ermentrout 1998).
of transitions between regimes of neuronal activ- At the critical value for bifurcation, a*, the
ity with specific bifurcations gives predictions saddle-node equilibrium obstructs the spiking
about the properties of stationary and oscillatory orbit. For values of a that support spiking and
solutions (Booth et al. 1997; Doiron et al. 2003; are close to a*, the frequency ffi of spiking activity is
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Shilnikov and Cymbalyuk 2005; Shilnikov proportional to ja a j: This bifurcation does
et al. 2005; Barnett and Cymbalyuk 2014). This not constrain the amplitude of spiking. This sce-
approach is applicable for tuning the dynamics of nario is an example of the type I spiking scenario
neuronal activity – such as bursting, spiking, and (Gutkin and Ermentrout 1998; Izhikevich 2007).
rest states – with biophysical parameters. We A homoclinic bifurcation of a saddle
provide examples of four bifurcations that deter- equilibrium can terminate a spiking regime
mine quantitative characteristics of activity: the (Izhikevich 2007). It occurs when an orbit is
B 368 Bifurcation Analysis
interrupted by the saddle; as a changes and and burst duration: one parameter, a, controls the
approaches the critical value for bifurcation, a*, transition from bursting to silence as in the SNIC,
the spiking frequency decays according to 1/ln and another parameter, b, controls the transition
(|a a*|). from bursting to spiking as in the blue sky catas-
In a supercritical Andronov-Hopf bifurcation, trophe. At the critical value for both parameters,
a stable focus loses stability, and a stable periodic (a, b), there occurs both the saddle-node equi-
orbit with zero amplitude and nonzero frequency librium and the saddle-node periodic orbit. In this
is born. The frequency is inherited from the res- way, the duration of the interburst interval is con-
onant frequency of the focus. In subthreshold trolled as in the SNIC, and the burst duration is
oscillations or spiking activity associated with controlled as in the blue sky catastrophe (Fig. 1)
this bifurcation, the
pamplitude
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi of the orbit grows (Barnett and Cymbalyuk 2014).
proportionally to ja a j: This scenario is an The Lukyanov-Shilnikov bifurcation is a
example of the type II spiking scenario (Gutkin homoclinic bifurcation of a periodic orbit
and Ermentrout 1998; Izhikevich 2007; Rinzel (Lukyanov and Shilnikov 1978). In bursting,
and Ermentrout 1998). this bifurcation occurs when the unstable mani-
fold of the saddle-orbit leads into the stable man-
ifold of the orbit (Shilnikov et al. 2005). The
Bursting unstable manifold near the saddle-orbit deter-
mines the duration of each burst. Near this bifur-
Four types of bifurcations are crucial for the cation, bursting is commonly chaotic, and on
control of temporal characteristics in bursting average, the burst duration grows proportionally
activity: the SNIC for bursting, the blue sky to ln(|a a|) as the bifurcation parameter, a,
catastrophe, the cornerstone bifurcation, and the approaches the critical value a*.
Lukyanov-Shilnikov bifurcation.
In bursting activity, the SNIC describes the
transition from bursting to silence(Booth Transient Activity
et al. 1997; Doiron et al. 2003; Barnett and
Cymbalyuk 2014). According to this bifurcation, Transient dynamics are critical for information
the interburst interval growsffi smoothly and is pro-
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi processing (Rabinovich et al. 2008). The tempo-
portional to 1= ja a j as the bifurcation ral characteristics of transient activity can simi-
parameter, a, approaches the critical value a* larly be governed by bifurcations. Typically,
(Fig. 1). At the bifurcation, a saddle-node equi- temporal laws are associated with global bifurca-
librium is born on the bursting orbit. tions. However, a local bifurcation can determine
The blue sky catastrophe occurs when a the timescale of transient neuronal responses to
saddle-node periodic orbit is born on the mani- external perturbations, such as a pulse of synaptic
fold that governs spiking in a bursting neuron current (Fig. 2).
(Shilnikov and Cymbalyuk 2005; Barnett and The saddle-node bifurcation for periodic
Cymbalyuk 2014). This bifurcation describes orbits can impose temporal laws on transient
the transition from bursting to tonic spiking. spiking activity in an endogenously silent neuron
The burst duration grows smoothly and is propor-
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (Roa et al. 2007). For values of the bifurcation
tional to 1= jb b j as the bifurcation parame- parameter a which support silence and are close
ter, b, approaches the critical value b* (Fig. 1). to the birth of the spiking regime at a*, the
The cornerstone bifurcation satisfies the con- dynamics of perturbation-induced transient spik-
ditions of both the SNIC and the blue sky catas- ing activity are controlled by the proximity (|a
trophe; it is a global codimension-2 bifurcation a|) to the bifurcation. Far from the bifurcation,
(Shilnikov and Turaev 2000; Barnett and a perturbation may trigger a single action poten-
Cymbalyuk 2014). This bifurcation describes tial, but as the quantity |a a| becomes small,
the independent control of interburst interval the same perturbation induces an increasing
Bifurcation Analysis 369 B
Bifurcation Analysis, Fig. 1 The transition from burst- The parameter b is varied to approach its critical value for
ing to silence and the transition from bursting to tonic bifurcation, b*, and the burst duration increases. (g) Curve
pffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
spiking. (a) Initial values for a and b produce periodic fits for the burst duration took the form b= b b þ c:
bursting activity. (b, c) The parameter a is varied to (f, g) Graphs are plotted in the log-log scale. The interburst
approach its critical value for bifurcation, a*, and the interval and burst duration are depicted as blue dots. The
interburst interval increases. (f) p Curve fits
ffi for the
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi curves fitted to these data are depicted as red curves
interburst interval took the form b= ja a j þ c: (d, e)
number of spikes. The duration of this transient the cornerstone bifurcation (Fig. 2) (Barnett and
activity – under the condition of an invariant Cymbalyuk 2014).
external stimulus – is proportional to In the context of the endogenously tonically
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1= ja a j: A similar scenario is generated by spiking neurons, the cornerstone bifurcation
B 370 Bifurcation Analysis
Bifurcation Analysis, Fig. 2 The inverse-square-root correspond to the burst duration measured at the respec-
laws in transient responses triggered by a pulse of current tive values of a. The red curve is the graph of the curve
pffiffiffiffiffiffiffiffiffiffiffiffiffi
for parameter values that support silent and tonic spiking fitted in the form b= a a þ c: (e–h) Latency to spiking
regimes. (a–d) An individual burst was triggered by was shown by administering an individual pulse of
a hyperpolarizing pulse of injected current. Burst duration injected current. Interburst interval was controlled by the
was controlled by the voltage of half activation of half activation of a hyperpolarization-activated current
a potassium current (a). The duration of individual bursts (b). The delays shown here are (e) 10.3 s, (f) 103.4 s,
were (a) 10.3 s, (b) 103.5 s, and (c) 309.3 s for different and (g) 317.7 s for different values of b. (h) Latency to
values of a. (d) The log-log graph of burst duration of spiking for sampled parameterp values (blue dots) and the
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
individual bursts plotted against the difference between graph of the curve fitted to b= jb bj þ c (red curve)
the voltage of half activation of a potassium current and its were plotted in log-log scale against the sampled values of
critical value for bifurcation (a*). The blue dots b where b* is the critical value for bifurcation
describes a transient silent interval after inhibi- state after inhibition is delayed if a is close in
tion according to the properties imposed by a value to a*, and this delay is proportional to
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
saddle-node bifurcation for equilibria (Barnett 1= ja a j (Fig. 2).
and Cymbalyuk 2014). The return to the spiking
Bifurcations Dynamics of Single Neurons and Small Networks 371 B
Conclusions Guckenheimer J, Harris-Warrick R, Peck J, Willms
A (1997) Bifurcation, bursting, and spike frequency
adaptation. J Comput Neurosci 4:257–277
We have described how the applications of Gutkin BS, Ermentrout GB (1998) Dynamics of mem-
bifurcation analysis to neuronal dynamics gov- brane excitability determine interspike interval vari-
erns characteristics of oscillatory activity. In ability: a link between spike generation mechanisms B
spiking activity, the homoclinic and saddle- and cortical spike train statistics. Neural Comput
10:1047–1065
node bifurcation on invariant circle each control Izhikevich E (2007) Dynamical systems in neuroscience.
spike frequency, and the Andronov-Hopf bifur- MIT Press, Cambridge, MA
cation controls spike amplitude. In bursting Lukyanov V, Shilnikov L (1978) On some bifurcations of
activity, the SNIC, the blue sky catastrophe, dynamical systems with homoclinic structures. Soviet
Math Dokl 19:1314
the cornerstone bifurcation, and the Lukyanov- Medvedev GS (2006) Transition to bursting via determin-
Shilnikov bifurcation each impose temporal istic chaos. Phys Rev Lett 97:048102
laws on the period of bursting activity. In tran- Rabinovich M, Huerta R, Laurent G (2008) Transient
sient dynamics, the saddle-node bifurcation dynamics for neural processing. Science 321:48–50
Rinzel J (1978) On repetitive activity in nerve. Fed Proc
for periodic orbit determines the duration of 37:2793–2802
individual perturbation-evoked bursts in Rinzel J (1987) Mathematical topics in population biol-
endogenously silent neurons, and the saddle- ogy, morphogenesis, and neuroscience,
node bifurcation for equilibria determines the vol 71, Lecture notes in biomathematics. Springer,
Berlin
delay before spiking after inhibition in an Rinzel J, Ermentrout B (1998) Analysis of neural excit-
endogenously spiking neuron. Knowledge of ability and oscillations. In: Koch C, Segev I (eds)
these laws describes mechanisms underlying Methods in neural modeling. The MIT Press, Cam-
modulation and adjustment of stationary and bridge, MA, pp 251–292
Roa MAD, Copelli M, Kinouchi O, Caticha N (2007)
oscillatory regimes and transient responses of Scaling law for the transient behavior of type-II neuron
neurons. models. Phys Rev E 75:021911
Shilnikov AL, Cymbalyuk GS (2005) Transition
between tonic-spiking and bursting in a neuron
model via the blue-sky catastrophe. Phys Rev Lett
Acknowledgment This work was supported by National 94:048101
Science Foundation grant PHY-0750456. Shilnikov L, Turaev D (2000) A new simple bifurcation of
a periodic orbit of blue sky catastrophe type. Am Math
Soc Transl II Ser 200:165–188
Shilnikov AL, Calabrese RL, Cymbalyuk GS (2005)
References Mechanism of bistability: tonic spiking and bursting
in a neuron model. Phys Rev E 71:056214
Barnett W, Cymbalyuk G (2014) A codimension-2 bifur-
cation controlling endogenous bursting activity and
pulse-triggered responses of a neuron model. PLoS
One 9:e85451
Booth V, Carr TW, Erneux T (1997) Near threshold burst- Bifurcations Dynamics of Single
ing is delayed by a slow passage near a limit point. Neurons and Small Networks
SIAM J Appl Math 57:1406–1420
Doiron B, Laing C, Longtin A, Maler L (2003)
Ghostbursting: a novel neuronal burst mechanism. Amitabha Bose
J Comput Neurosci 12:5–25 Department of Mathematical Sciences, New
Ermentrout G, Terman D (2010) Mathematical founda- Jersey Institute of Technology, Newark, NJ, USA
tions of neuroscience, vol 35, Interdisciplinary applied
mathematics. Springer, New York
Ghigliazza R, Holmes P (2004) Minimal models of burst-
ing neurons: how multiple currents, conductances, and Definition
timescales affect bifurcation diagrams. SIAM J Appl
Dyn Syst 3:636–670
Guckenheimer J, Gueron S, Harris-Warrick R (1993)
A bifurcation occurs when a system undergoes
Mapping the dynamics of a bursting neuron. Phil a qualitative change in its output as a result of
Trans R Soc Lond 341:345–359 a change in parameter. Under certain conditions,
B 372 Bifurcations Dynamics of Single Neurons and Small Networks
the voltage of a cell membrane can change from where x is an n-dimensional vector of
being at rest to becoming oscillatory as a result of unknowns, m is an m-dimensional vector of
a bifurcation. Oscillatory properties of small net- parameters, and F is a function that depends
works are often understood using bifurcation on the two. For the time being, assume that
analysis. m = 1 and there is only one parameter of inter-
est. Equation 1 describes how the dependent
variable x evolves in time under the flow dic-
Detailed Description tated by F(x, m). A bifurcation occurs when
a system undergoes a qualitative change in its
What Is a Bifurcation? output as a result of a change in the parameter m.
The word bifurcate is commonly used to denote A bifurcation analysis of Eq. 1 would involve
a split, as in “up ahead, the road bifurcates into systematically characterizing the different
two parts.” The use of the term in this context types of solution x(t) that arise for different
implies that a driver traveling on this road can values of the parameter m.
make a real-time choice of being able to take one In general, a dynamical system can exhibit
or the other branch of the road when the bifurca- local and/or global bifurcations. Local ones are
tion point is reached. In mathematics, and in its more prevalent in neuroscience and also easier to
application to neuroscience, the term bifurcation recognize and understand. The simplest local
has an entirely different contextual meaning. bifurcation involves assessing the existence and
While a mathematical bifurcation is similar in stability of fixed points of Eq. 1. A fixed point x*
that it involves different branches, the decision is a value such that F(x*, m) = 0 for either
on which branch is chosen is largely made a priori a specific value of m or a set of values of m.
and depends on the choice of parameters. A fixed point can be asymptotically stable (all
It is instructive to first gain some insight into nearby initial conditions converge to it), stable
situations that are not bifurcations. Different ini- (initial conditions that start nearby stay nearby it),
tial conditions converging to different solutions or unstable (at least one nearby initial condition
of the same differential equation while all param- leaves a neighborhood of it). Often, and some-
eters are held constant are not examples of bifur- what confusingly, the term stable is used in place
cations. Instead it may be that one initial of asymptotically stable. Both the existence and
condition lies in the basin of attraction of stability of a fixed point x* can depend on the
a particular solution, while the other one does parameter m in which case the possibility of
not. Think, for example, of a double well poten- a local bifurcation exists. To determine if one
tial where each well can trap particles starting occurs, Eq. 1 is linearized about x* and the eigen-
with different initial conditions. A solution to values of the ensuing Jacobian matrix are ana-
a differential equation that is perturbed and then lyzed. The eigenvalues also depend on the
suddenly veers off toward another attracting solu- parameter m. Different bifurcations correspond
tion is also not an example of a bifurcation. In to the eigenvalues satisfying specific mathemat-
these examples, just because a change in behavior ical conditions for different values of m. Simi-
has occurred or a difference in convergence is larly, a periodic solution of Eq. 1 can undergo
observed, a bifurcation has not occurred. To qual- a local bifurcation as a parameter is varied. As
ify as a bifurcation, the change in behavior must discussed later, such bifurcations can often be
be associated with a change in a parameter value. studied using a discrete map rather than
Bifurcations arise in mathematical equations a continuous flow. See Guckenheimer and
called dynamical systems. A dynamical system Holmes (1983) for a detailed analysis of both
can generically be defined by the equation local and global bifurcations.
Bifurcation analysis is a useful tool in mathe-
dx matics because it simplifies the identification of
¼ Fðx, mÞ: (1)
dt different kinds of stable and unstable solutions.
Each bifurcation type has a concomitant set of dv
¼ f ðv, wÞ þ I ext
stable or unstable solutions that are produced dt (2)
(or destroyed) as a result of it. The conditions dw
¼ ϵgðv, wÞ:
for identifying the existence of a bifurcation are dt
often easier to satisfy than proving the existence B
of an actual solution. In neuroscience, bifurca- In Eq. 2, the variable v represents voltage, w a
tions are of importance not just at the single-cell recovery variable, and ϵ a positive number used to
level but also in understanding the dynamics of demarcate the time scales of the two variables.
networks. As described below, often a bifurcation When ϵ is small, then Eq. 2 effectively has a fast
mechanism that has been identified at the single- and a slow time scale; v is called the fast variable
cell level suggests how dynamical changes (not and w the slow. The parameter Iext can be thought
necessarily bifurcations) occur at the network of as an external current that only affects the
level. membrane voltage. The v-nullcline is defined
as the set of points (v, w) that satisfy the equation
Typical Bifurcations in the Dynamics of f(v, w) = 0. For the two models mentioned
Single Neurons above, the v-nullcline is a cubic-shaped curve.
Perhaps the most widely studied bifurcations The w-nullcline, the points (v, w) such that
involving single neurons are those that give rise g(v, w) = 0, is sigmoidal-shaped like a typical
to oscillations. Take a neuron whose voltage is at activation or inactivation curve.
rest. The resting membrane potential represents To understand the SNIC bifurcation, suppose
a stable state. If after the injection of a constant that at Iext = 0, there are two intersections of the
amplitude current, the membrane voltage v- and w-nullclines near the local minimum of the
exhibits rhythmic behavior, then a bifurcation cubic (Fig. 1a). These intersections correspond to
has occurred. Mathematically, there are two typ- fixed points of Eq. 2. Linearizing at them would
ical bifurcations that characterize this change in show that under generic assumptions, the inter-
behavior. The first is called an SNIC (saddle node section on the left branch is a stable node
on an invariant circle) bifurcation and the second (attracting nearby trajectories) and the one on
is a Hopf bifurcation. Both bifurcations arise in the middle branch is a saddle point (unstable
common neuronal models such as the with only one attracting direction and other direc-
two-dimensional Hodgkin-Huxley (Ermentrout tions being repelled). As Iext is increased, the
and Kopell 1998) or Morris-Lecar (Morris and v-nullcline moves up in the v w phase plane
Lecar 1981) equations. and the two fixed points come closer to one
Consider a set of two equations that depend another. At the SNIC bifurcation value ISNIC,
the parameter Iext. the fixed points coincide and become one. For
a b
Bifurcations Dynamics of Single Neurons and Small unstable (yellow square) fixed point along the middle
Networks, Fig. 1 Nullclines of (1). (a) Typical structure branch. The upper v-nullcline for Iext > ISNIC has only
for an SNIC bifurcation. The lower black v-nullcline for the unstable (yellow square) fixed point. (b) Typical struc-
Iext < ISNIC has a stable (black circle) and an unstable ture for an Hopf bifurcation. Here there is only one fixed
(black square) fixed points near the local minima and an point associated with each cubic
a b
Voltage Freq
I1 ISNIC I2 Iext ISNIC Iext
Bifurcations Dynamics of Single Neurons and Small fixed points that meet at ISNIC are destroyed. At this point
Networks, Fig. 2 SNIC bifurcation diagrams. (a) Stable a branch of stable periodic solutions arises. See text for
(solid curves) and unstable (dashed curves) solutions as explanations for I1 and I2. (b) The f I curve shows that
a function of Iext. Black curves denote fixed points and oscillations begin with arbitrarily low frequency
blue curves periodic solutions. As Iext increases, the two
Iext > ISNIC there are no longer any fixed points in increases, the v-nullcline moves up and the fixed
a neighborhood of the local minimum and point moves toward the local minimum and even-
a periodic orbit arises. Figure 2a shows a bifurca- tually occurs on the middle branch. The eigen-
tion diagram associated with this situation. The values depend continuously on Iext and their real
diagram can be interpreted as follows. For each parts become smaller in absolute value. At some
fixed value Iext, the vertical slice at that value point, Iext = IHopf, after the fixed point has moved
shows the steady-state solutions that exist. Solid to the middle branch, the eigenvalues at the line-
curves denote stable solutions and dashed curves arization are pure imaginary. For Iext > IHopf, the
are unstable ones. For instance, when Iext = I1, eigenvalues have positive real part. A Hopf bifur-
there are three fixed points. The one with smallest cation occurs at IHopf resulting in the creation of
voltage is stable, while the other two are unstable. a periodic orbit. There are two distinct types of
At Iext = I2, there is an unstable fixed point and Hopf bifurcations: supercritical which produce
a stable periodic solution. The solid blue curves stable periodic solutions and subcritical which
are the upper and lower boundaries of the voltage produce unstable ones. Figure 3a shows the bifur-
amplitude of the periodic solution. In this and cation diagram associated with a supercritical
subsequent figures, blue curves are associated Hopf bifurcation. The amplitude of oscillations
with periodic solutions and black curves with just after bifurcation is small. The frequency of
fixed points. Oscillations that arise through an oscillations just above the Hopf bifurcation
SNIC bifurcation can have arbitrarily low fre- points is bounded from below resulting in
quencies and typically are of large amplitude. a discontinuous f I curve (Fig. 3b). Neurons
The f I curve associated with an SNIC bifur- that oscillate through a Hopf bifurcation mecha-
cation is continuous at ISNIC (Fig. 2b). Ermentrout nism typically have a type II phase response
(1996) has shown that neurons that oscillate curve that is of both signs.
through an SNIC bifurcation have type I phase In some neuron models, the subcritical Hopf
response curves that are strictly of one sign. bifurcation is often found in conjunction with
A Hopf bifurcation can arise when the v- and another bifurcation called the saddle node of
w-nullclines only intersect once in periodic orbits (SNPO) bifurcation. Figure 4
a neighborhood of the local minimum (Fig 1b). shows such a case where the branch of unstable
Suppose when Iext = 0, this single intersection orbits extends from IHopf for decreasing values of
(fixed point) is on the left branch. This fixed Iext. Surrounding the branch of unstable periodic
point is asymptotically stable. Linearization at it orbits is a branch of stable periodic orbits of
would show that the eigenvalues of the Jacobian larger amplitude. The stable and unstable peri-
matrix are complex with negative real part. As Iext odic branches come together at Iext = ISNPO
a b
Voltage Freq
B
IHopf Iext IHopf Iext
Bifurcations Dynamics of Single Neurons and Small of stable periodic solutions. (b) The f I curve shows that
Networks, Fig. 3 Hopf bifurcation diagrams. (a) oscillations begin with a frequency that is bounded from
A single stable fixed point (black curve) loses stability at below away from zero
a supercritical Hopf bifurcation point spawning the branch
neurons in the suprachiasmatic nuclei undergo

a subcritical Hopf bifurcation as the maximal
conductance of their calcium channel, gCa, is
Voltage increased (Belle et al. 2009). Belle et al. show
that as gCa is increased, an asymptotically stable
rest state undergoes a subcritical Hopf bifurcation
leading to oscillations during only a certain phase
of the circadian rhythm. They also show that
a second Hopf bifurcation can arise during a
ISNPO IHopf Iext phase corresponding to daytime in which
a highly depolarized resting state becomes
Bifurcations Dynamics of Single Neurons and Small stable. Mathematically, the mechanism for the
Networks, Fig. 4 Subcritical Hopf bifurcation diagram.
creation and control of oscillations is exactly the
A single stable fixed point (black curve) loses stability at a
subcritical Hopf bifurcation point spawning the branch of same whether the bifurcation results from
unstable periodic solutions. That unstable branch merges changes in gCa as in this case or from Iext as
at ISNPO with a branch of stable periodic orbits. The inter- shown in Fig. 4.
val ISNPO < Iext < IHopf is a region of bistability in that
Bifurcation analysis is also a primary tool in
both a stable fixed point and a stable periodic solution exist
understanding bursting oscillations where
a neuron exhibits a sequence of (relatively)
which destroys the two orbits. As the bifurcation high-frequency spikes, followed by a (relatively)
diagram shows, for ISNPO < Iext < IHopf, long interburst interval. The main idea here is to
bistability exists between the stable fixed point exploit the difference in time scales between the
on the middle branch and the larger amplitude high-frequency spikes and the long interburst
periodic orbit. Starting at the stable rest state, as interval to define fast and slow variables. Con-
Iext is increased past the Hopf bifurcation point, sider the following set of three first-order
the solution suddenly changes to a large ampli- equations:
tude periodic orbit. Once on this branch of large
amplitude solutions, decreasing Iext shows an x0 ¼ f 1 ðx, y, zÞ
example of hysteresis whereby the stable branch y0 ¼ f 2 ðx, y, zÞ (3)
of periodic solutions exists over a range where the z0 ¼ ϵgðx, y, zÞ:
stable fixed points also do.
Variations in different parameters besides Iext The small positive parameter ϵ implies that
can lead to bifurcations. For example, in a model x and y are fast variables and z is the slow vari-
associated with the circadian rhythm, per1 able. To use bifurcation analysis, set ϵ = 0 in
Eq. 3 and note that now z0 = 0, implying that

z will not change. This means that z will act as
a parameter in the fast equations for x0 and y0 . The Voltage
fast subsystem can now be studied with z acting
as a bifurcation parameter to see what, if any,
bifurcations exist. For each z, the equations f1(x, A.
y, z) = 0 and f2(x, y, z) = 0 are the nullclines of
the fast subsystem. They form null surfaces in the
full three-dimensional phase space. Intersections
of any two null surfaces form curves of fixed zSNPO zHopf zHom z
points of the fast subsystem and their bifurcations
Bifurcations Dynamics of Single Neurons and Small
can readily be studied. The mathematical tech- Networks, Fig. 5 Bifurcation diagram associated with
nique of singular perturbation theory then shows square wave bursting. The variable z acts as a parameter
that solutions to the actual ϵ > 0, but small, prob- for the ϵ = 0 fast subsystem of (3). As z is increased along
the upper branch of fixed points (upper solid black curve
lem lie close to the null surfaces except for times
of the Z-shaped curve), the fixed point undergoes
when they make very fast transitions between a subcritical Hopf bifurcation. The unstable periodic
different null surfaces. The location where these branch merges with a stable branch of periodic solutions
fast transitions occur is computable through at zSNPO. The stable branch of periodic solutions termi-
nates at a homoclinic bifurcation point at zHom. The actual
bifurcation and singular perturbation analysis.
flow of (3) for ϵ small is graphically depicted. The red
Rinzel (1985, 1987) first used these ideas to lines pointing left (right) indicate that the solution closely
mathematically classify different types of burst- follows the lower (upper) branch of the Z-shaped curve as
ing in single-cell models. Figure 5 shows an z slowly decreases (increases). The green lines indicate the
fast transitions between the upper and lower branches of
example of square wave bursting bifurcation dia-
the Z-shaped curve. Magenta arrows between the stable
gram which illustrates this type of analysis. See periodic branches indicate the fast spiking portion of the
(Izhikevich 2000) for a rigorous mathematical burst
exploration of bursting in single neurons or
(Ermentrout and Terman 2010) for an overview
of several important types of bursting arising in In networks, many interesting bifurcations
neuroscience. occur with changes in synaptic parameters. Con-
sider a network of two mutually coupled spiking
Bifurcations in Small Neuronal Networks neurons governed by the integrate and fire model
Just as certain parameters intrinsic to a neuron and connected by alpha synapses gsyna2te at.
can determine whether an individual neuron can Here gsyn governs the strength and a the speed
oscillate, these parameters can also do the same of the synapse. The phase f at which one cell fires
for a coupled network of neurons. A small net- with respect to the other can be used to define
work of cells may not be able to oscillate unless a function whose roots correspond to phase-
some or all of the individual cells can. Thus, locked periodic solutions. For this model of two
whatever bifurcation causes the individual cell identical cells, both the synchronous (f = 0 or 1)
to oscillate may, in turn, cause the network to be and antiphase (f = 0.5) solutions exist for both
rhythmic. Changes in intrinsic parameters having excitatory (gsyn > 0) and inhibitory (gsyn < 0)
nothing to do with the existence of oscillations coupling. Van Vreeswijk et al. (1994) showed
can also lead to bifurcations in network output. that these solutions change stability through
For example, Zhang et al. (2009) show that a pitchfork bifurcation as the parameter a
increases in the maximal conductance of an changes; at the pitchfork bifurcation point, three
A-current in a postsynaptic cell can cause the branches emerge from one. In particular, for
cell to go through a sequence of border collision excitatory synapses, increases in a make the
bifurcations (Nusse et al. 1994), resulting in antiphase solution bifurcate from being stable to
a cascade of different phase-locked solutions. unstable (Fig. 6a). At the bifurcation point, a new
a b
1 1
Φ Φ
0.5 0.5
B
0 0
α α
Bifurcations Dynamics of Single Neurons and Small out-of-phase solutions shows that solutions approach syn-
Networks, Fig. 6 Bifurcation diagram as a function of chrony as the synapse gets faster. (b) For inhibitory syn-
synaptic speed, (Adapted from Van Vreeswijk et al. apses, the synchronous solution is always stable. The
(1994)). (a) For excitatory synapses, as the synaptic antiphase solution stabilizes by increasing a though the
speed increases, the stable antiphase solution at f = 0.5 pitchfork bifurcation point
undergoes a pitchfork bifurcation. The new curve of stable
stable out-of-phase solution arises; by symmetry give rise to saddle-node bifurcations of fixed
there are two such solutions. Further increasing a points of a one-dimensional return map (Fig. 7).
causes these new curves of out-of-phase solutions The fixed points of the map correspond to differ-
to come closer to the curve corresponding to the ent oscillatory solutions of different frequency.
synchronous solution. Note in this case, that Solutions on the upper branch are controlled by
while the synchronous solution is always unsta- parameters that are intrinsic to the excitatory cell
ble, as a increases and the synapse becomes and are of high frequency. The high frequency
faster, the phase of locking f tends to synchrony. causes the inhibitory synapse to remain in
The opposite happens for inhibitory synapses a depressed state and never recover. Solutions
(Fig. 6b). As a increases and the synapse is on the lower branch correspond to
made faster in its rise and decay, the antiphase low-frequency solutions with a long interspike
solution undergoes a pitchfork bifurcation taking interval. Because of this long interval, the
it from unstable to stable. The curves of out-of- depressing synapse has a chance to recover
phase solutions that are born are unstable. The between spikes and the synaptic current to the
synchronous solution is stable for all values of a, excitatory cell at each spike is large. Here the
but its basin of attraction shrinks as a increases. frequency is controlled by a parameter governing
A common way to find bifurcations in the decay of inhibition. Note that this network
a neuronal network is to define a Poincare return also exhibits bistability and hysteresis similar to
map or a time T map. The Poincare map records what was seen to occur in certain single-neuron
the values of relevant variables along a defined models. Perturbations can cause the solution to
hyperplane in the mathematical phase space. The switch between different oscillatory solutions as
time T map records the values of appropriate seen in Fig. 7. The switch is a consequence of the
variables at every time nT, for n = 1, 2, 3, . . .. bifurcation structure, but is not in and of itself
In either case, a fixed point of the map corre- a bifurcation.
sponds to a periodic solution of the network. In As is suggested by the prior example, bifurca-
this way, mathematical tools to study bifurcations tions in networks of neurons are also closely
of fixed points of maps (discrete dynamical sys- related to the issue of what sets of parameters
tems) can be used to study bifurcations of peri- control the oscillatory state. This is especially
odic solutions of networks. For example, in true in central pattern generating (CPG) networks
a model consisting of an excitatory cell recipro- where the existence of oscillations is a given.
cally coupled to an inhibitory cell whose synapse What is more relevant is what controls the
exhibits short-term synaptic depression, Bose existence of oscillations and their frequency.
et al. (2001) showed how changes in the maximal For example, in a network of two cells
synaptic conductance of the inhibitory synapse mutually coupled by fast inhibitory synapses,
in the intrinsic escape or release mode but is

highly relevant in the synaptic escape or release
Freq
mode. Understanding how the synaptic threshold
can play this role relies, in part, on bifurcation
dynamics associated with single cells. For single
cells it was earlier shown that increasing Iext
causes the stable solution to bifurcate from
being a rest point to an oscillatory one by shifting
ginh the v-nullcline up in the phase space and remov-
ing a fixed point on its left branch. In this two-cell
network, the removal of an inhibitory synaptic
current to the postsynaptic cell acts similarly to
increasing Iext in a single cell by shifting the
v-nullcline up in phase space and removing an
intersection of the v- and w-nullclines. Thus,
although there is not a bifurcation in the network
as the result of the application or removal of
a synaptic current, the mechanism by which the
postsynaptic cell fires is readily understood as
a result of single-neuron bifurcation analysis.
The possible application or removal of inhibition
occurs only when the synaptic threshold is
crossed. When the individual cells are modeled
as relaxation oscillators (with cubic-shaped
v-nullclines), then the fast transition between
the silent and firing states effectively occurs at
Bifurcations Dynamics of Single Neurons and Small the local minima and maxima of the cubic
Networks, Fig. 7 Dynamics of an excitatory inhibitory nullcline. Thus, when the synaptic threshold lies
network. Upper panel shows bifurcation diagram as
a function of maximal inhibitory conductance ginh. The between these two extrema, the threshold is not
two branches of stable periodic solutions are connected by relevant for controlling frequency as small
an unstable branch. Lower panel shows voltage trajecto- changes in its value have little effect on the time
ries (E excitatory, I inhibitory) with ginh chosen at the spent above or below it. On the other hand, if the
value of the dotted line in upper panel. The solution starts
off at high frequency along the upper branch. A brief synaptic threshold intersects either the left or
perturbation given to the inhibitory cell temporarily sup- right branch of the cubic, then small changes in
presses its activity allowing its synapse to recover from its value can have an effect since the cell’s tra-
depression and strengthen. In turn, this causes the solution jectory evolves slowly there; that the evolution is
to switch to a lower frequency periodic orbit along the
lower branch slow is a consequence of the assumption of relax-
ation oscillations. It is through the scaffold of
Skinner et al. (1994) describe circumstances single-cell bifurcation analysis and fast/slow
under which the synaptic threshold is of impor- time scale decomposition that Skinner et al. are
tance in controlling the frequency of antiphase, able to define intervals for the synaptic thresholds
half-center oscillations. They define the terms that demarcate intrinsic versus synaptic escape/
intrinsic escape, intrinsic release, synaptic release mechanisms and thereby control the fre-
escape, and synaptic release to indicate different quency of oscillations.
ways in which the oscillators can switch orienta- A more complicated example of using bifur-
tions from being active to silent and vice versa. cation analysis is found in a model of
The synaptic threshold turns out to be irrelevant a respiratory CPG due to Rubin et al. (2009).
The authors consider a four-cell network of excit- But networks have the potential to be much richer
atory and inhibitory cells and are interested in and complicated in their bifurcation structure. An
understanding transitions between oscillatory important concept highlighted above is that bifur-
patterns that correspond to different phasic states cation analysis can reveal how certain parameters
of the respiratory rhythm. By using fast/slow time determine which network elements control B
scale analysis, they show how to identify certain the behavior in the network. Interestingly,
transitional curves associated with the fast a bifurcation occurring as a result of a change in
subsystem. These curves lie on equilibrium sur- an intrinsic parameter may imply that a synaptic
faces (null surfaces) of the fast subsystem. In the mechanism becomes relevant for controlling the
problem considered, there are relevant equilib- network output, or vice versa.
rium surfaces corresponding to cells being either Bifurcations of the type described here are
in the active or inactive state. When the trajectory associated with steady-state solutions (fixed
of the system reaches one of these transitional points or periodic orbits) of differential equa-
curves, there is a switch from inspiration to expi- tions. They are local in the sense that the infor-
ration or vice versa. The transitional curves are mation obtained from the bifurcation only
themselves defined because they satisfy certain provides information in a neighborhood of the
bifurcation conditions for fixed points of the fast bifurcation point. While this neighborhood may
subsystem. Through their analysis, the authors in fact be mathematically large, it is important to
show how changes in the maximal conductance keep in mind that the entire set of possible solu-
of a persistent sodium current in one of the cells tions in the global phase space are not determined
play a major role in controlling the frequency and through local bifurcations. Nonetheless, bifurca-
duty cycle of oscillations within the network. tion analysis is a powerful tool for understanding
They also show how changing the total synaptic the dynamics of neurons and the networks in
drive to individual cells modulates the bifurca- which they participate.
tion diagram again resulting in an understanding
of how frequency and duty cycle change. Acknowledgment This work was supported in part by
The interested reader is referred to (Rubin NSF DMS 1122291.
et al. 2009) to see further details.
References
Concluding Remarks
Belle M, Diekman C, Forger D, Piggins H (2009) Daily
electrical silencing in the mammalian circadian clock.
Bifurcations at the single-neuron level often
Science 326:281–284
focus on understanding how a single fixed point Bose A, Manor Y, Nadim F (2001) Bistable oscillations
is created or destroyed or how it gains or loses arising from synaptic depression. SIAM J Appl Math
stability. Different kinds of bifurcations imply 62:706–727
Ermentrout GB (1996) Type i membranes, phase resetting
the existence of different kinds of oscillatory
curves and synchrony. Neur Comp 8:979–1001
solutions, each possessing different properties. Ermentrout GB, Kopell N (1998) Fine structure of neural
As a single-neuron model becomes more compli- spiking and synchronization in the presence of con-
cated, say for describing a bursting neuron, bifur- duction delays. Proc Natl Acad Sci 95:1259–1264
Ermentrout GB, Terman D (2010) Mathematical founda-
cations occurring in a fast subsystem help to build
the skeleton on which the dynamics of the full Guckenheimer J, Holmes P (1983) Nonlinear oscillations,
system can be understood. In networks, the bifur- dynamical systems and bifurcations of vector fields.
cation parameter can be either intrinsic to a cell or Springer, New York
Izhikevich E (2000) Neural excitability, spiking and burst-
related to one of the synapses. Bifurcations ing. Int J Bifur Chaos 10:1171–1266
within them are often understood by knowing Morris C, Lecar H (1981) Voltage oscillations in the
the bifurcation possibilities of single neurons. barnacle giant muscle fiber. Biophys J 35:193–213
B 380 Bifurcations, Neural Population Models and
Nusse H, Ott E, Yorke J (1994) Border-collision bifurca- Detailed Description

tions: an explanation for observed bifurcation phe-
nomena. Phys Rev E 49:1073–1077
Rinzel J (1985) Bursting oscillations in an excitable mem- Depending on their particular parameter set,
brane model. In: Sleeman B, Jarvis R (eds) Ordinary NPMs experience different asymptotic dynamic
and partial differential equations: proceedings of the behaviors characterized by constant output, har-
8th Dundee conference, vol 1151, Lecture notes in monic (i.e., sinusoidal) or nonharmonic oscilla-
mathematics. Springer, New York
Rinzel J (1987) A formal classification of bursting mech- tions, which can be periodic or quasiperiodic (i.e.,
anisms in excitable systems. In: Teramoto E, the signal repeats itself approximately, but not
Yamaguti M (eds) Mathematical topics in population exactly), and even chaotic behavior (Frascoli
biology, morphogenesis and neurosciences, et al. 2011; Spiegler et al. 2010, 2011). The tran-
vol 71, Lecture notes in biomathematics. Springer,
New York sitions in parameter space between these qualita-
Rubin J, Shevtsova N, Ermentrout GB, Smith J, Rybak tively different states are called bifurcations (see
I (2009) Multiple rhythmic states in a model of the Breakspear and Jirsa 2007). They partition the
respiratory central pattern generator. J Neurophysiol parameter space into regions with different
101:2146–2165
Skinner F, Kopell N, Marder E (1994) Mechanisms for dynamics. Their analysis is of fundamental
oscillation and frequency control in reciprocally inhib- importance for the description of the behavior
itory model neural networks. J Comput Neurosci of neural systems, which often features the tran-
1:69–87 sition of gradual changes in some variables into
van Vreeswijk C, Abbott L, Ermentrout GB (1994) When
inhibition, not excitation synchronizes neural firing. sudden qualitative changes in system behavior.
J Comp Neurosci 1:313–321 Bifurcation diagrams are an important utility for
Zhang Y, Bose A, Nadim F (2009) The influence of the the investigation of bifurcations. They show the
a-current on the dynamics of an oscillator-follower stable and unstable fixed points (i.e., points,
feed-forward inhibitory network. SIAM J Appl Dyn
Syst 8:1564–1590 where the time derivatives of the state variables
are zero; also referred to as equilibrium points)
and limit cycles (i.e., closed trajectories in state
space) as a function of one or more system
parameters (see Fig. 1). In the following, we
Bifurcations, Neural Population will give a brief overview on types of bifurcations
Models and that are relevant for neural population models.
Then we will demonstrate, using the example of
Thomas R. Knösche the Jansen-Rit model (Jansen and Rit 1995), how
Max Planck Institute for Human Cognitive and bifurcation analysis characterizes the dynamic
Brain Sciences, Leipzig, Germany fingerprint of a neural population model.
Types of Bifurcations
Definition There is a great variety of different types of
bifurcations (see also Guckenheimer 2007;
In neural population models (NPMs), as in Breakspear and Jirsa 2007). Depending on
dynamical systems in general, smooth changes whether or not it can be identified just by stability
in system parameters, for example, coupling analysis in a small vicinity of a fixed point,
strength, input levels, and time constants, may a bifurcation is termed local or global. In addi-
result in sudden qualitative changes in system tion, bifurcations are characterized by their
dynamics, such as the occurrence or disappear- codimension, which is the number of indepen-
ance of oscillatory or chaotic behavior. Such phe- dent parameters that need to be varied for the
nomena are referred to as bifurcations and are bifurcation to occur. Consider the bifurcation
interesting because they may account for diagram in Fig. 1, where the asymptotic behavior
switching phenomena, which are common in of the pyramidal cells’ membrane potential is
both brain signals and organisms’ behavior. plotted versus a single system parameter, namely,
Bifurcations, Neural Population Models and 381 B
Bifurcations, Neural Population Models and, black lines for unstable limit cycles. Bifurcations are
Fig. 1 Bifurcation diagram and exemplary solutions of indicated by diamonds, triangles, and circles (see legend
the original configuration of the Jansen-Rit model (1995) in inset). For all Andronov-Hopf bifurcations, the
with no input to the interneurons and time constants eigenfrequencies of the limit cycles are indicated. Differ-
10 and 20 ms for excitatory and inhibitory interneurons, ent extrinsic input levels on pyramidal cells (black and
respectively. Solid lines represent stable and dashed lines green vertical lines) result in different behaviors
unstable states. Branches of limit cycles correspond to depending on the past of a system’s trajectory (blue and
gray regions bounded by solid lines for stable and dashed red arrows) (Modified from Spiegler et al. (2010))
the mean input to the pyramidal cells. It shows population models include Andronov-Hopf
a number of codimension 1 bifurcations (see fig- (AH) as well as saddle-node (SN), saddle-focus
ure legend). (SF), and saddle-saddle (SS) bifurcations
Local (or subtle) bifurcations are character- (Grimbert and Faugeras 2006; Spiegler
ized by a change of asymptotic stability, that is, et al. 2010; Touboul et al. 2011). Further impor-
fixed points and limit cycles suddenly change tant local bifurcations are the pitchfork,
stability upon smooth change of a system param- transcritical, and period-doubling bifurcations
eter. They can be analyzed by looking at the, (see Breakspear and Jirsa 2007).
generally complex, eigenvalues of the system’s AH bifurcations occur if two conjugate com-
local linearization (Jacobian matrix). If the real plex eigenvalues cross the imaginary axis. They
parts of one or several of the eigenvalues change mark the transition between a fixed point (real
sign, the local stability of the system is altered part is negative) and a limit cycle (real part is
and a bifurcation occurs. Types of local bifurca- positive). Depending on the largest Lyapunov
tion that are most often identified in neural exponent, the bifurcation can be supercritical
B 382 Bifurcations, Neural Population Models and
(i.e., both limit cycle and fixed point are stable) or Hopf, GH) bifurcation, which occurs if an AH
subcritical (i.e., limit cycle and fixed point are bifurcation, upon change of a second parameter
unstable). AH bifurcations give rise to oscilla- of the system, turns from supercritical to subcrit-
tions, which are more or less harmonic, ical; (2) the Bogdanov-Takens (BT) bifurcation,
depending on their amplitude and the system’s where AH, SN, and homoclinic curves intersect;
degree of nonlinearity. and (3) the cusp (C) bifurcation, where two
SN/SF bifurcations (also called tangential or branches of saddle-node bifurcation curve meet
fold bifurcations) occur if, upon change of tangentially, giving rise to hysteresis phenomena.
a parameter, two fixed points, one unstable (the For more information, refer to Guckenheimer
saddle) and one stable (the node or focus), collide (2007) and Kuznetsov (2004).
and annihilate each other. When approaching the
bifurcation from the other side, this means that Bifurcations of the Jansen-Rit Model
two fixed points are suddenly created. At the The Jansen-Rit model has been investigated by
critical point itself, the fixed points lose asymp- Touboul and Faugeras (2007) as well as Spiegler
totic stability, as they become non-hyperbolic and colleagues (2010, 2011) by generating
(i.e., the second derivative becomes zero). If codimension one bifurcation diagrams, where
both colliding fixed points are saddles, the bifur- the membrane potential of the pyramidal cells is
cation is an SS bifurcation. plotted as the relevant state variable against the
Global (or catastrophic) bifurcations have to constant input level to the same cell population
be studied by considering the vector field outside (Fig. 1).
the local neighborhood of fixed points. They are For the original parameter configuration of
characterized by a loss of structural stability, that Jansen and Rit, the following dynamics occur,
is, fixed points and limit cycles suddenly occur or while gradually increasing the pyramidal cell
vanish upon smooth change of a system parame- input: At very low input levels, only one stable
ter. Accordingly, the saddle family of bifurca- fixed point (node) exists (Section I in Fig. 1). At
tions (see above) can also be considered global, about 1.2 mV, two unstable fixed points
as the two colliding fixed points vanish. One (saddles) appear in an SS bifurcation
typical example of a global bifurcation is given (Section VIII). The system still remains
by the infinite period bifurcation, where a stable monostable with one stable fixed point. With
fixed point (node or focus) is connected to further increase of the input level to about
a saddle via a particular trajectory, called the 0.5 mV, one of the saddles undergoes a subcrit-
heteroclinic orbit. Upon change of the parameter ical AH bifurcation, resulting in an unstable
toward the critical point, these two fixed points limit cycle and stable focus (section VII). The
get closer and closer, until they collide and van- system is now bistable. At about 3 mV the
ish. Now, the trajectory connecting the former stable focus undergoes a supercritical AH bifur-
fixed points turns into a periodic orbit, giving cation, giving rise to stable harmonic oscilla-
rise to sustained oscillations of the system. If tions (section VI). Further increase to about
approached from the other side, this bifurcation 3.5 mV lets the stable node and the unstable
is characterized by a slowing of the oscillation, saddle collide, and a periodic homoclinic orbit
until at the critical point the period becomes is born in an infinite period bifurcation (section
infinite and the periodic orbit is turned into V). In another global bifurcation, this orbit col-
a heteroclinic orbit connecting two fixed points. lides at about 4.5 mV with unstable limit cycle
Global bifurcations also include homoclinic and originating from the subcritical AH bifurcation.
heteroclinic bifurcations, in which periodic orbits The cycles vanish and the system’s only
collide with one or several saddle points, or col- remaining stable state is the stable limit cycle
lisions between more than one periodic orbit. (section III), which turns into a stable node by
Examples for higher codimension bifurcations another supercritical AH bifurcation at about
include the following: (1) the Bautin (generalized 10.5 mV.
Bimolecular Reactions, Modeling of 383 B
This exemplary bifurcation scheme is subject Kuznetsov Y (2004) Elements of applied bifurcation the-
to drastic change, if the other system parameters ory, 3rd edn. Springer, New York
Spiegler A, Kiebel SJ, Atay FM, Knösche TR
are varied, which has been shown by systematic (2010) Bifurcation analysis of neural mass models:
sampling of the parameter space (Spiegler impact of extrinsic inputs and dendritic time constants.
et al. 2010) and by characterization of higher Neuroimage 52:1041–1058 B
codimension bifurcations (Touboul and Faugeras Spiegler A, Knösche TR, Schwab K, Haueisen J, Atay FM
(2011) Modeling brain resonance phenomena using
2009). In addition to the SN and AH bifurcation a neural mass model. PLoS Comput Biol 7:e1002298
manifolds, Touboul and Faugeras (2009) found Touboul J, Faugeras O (2007) Codimension two bifurca-
some codimension two and three bifurcations for tions and rhythms in neural mass models.
the Jansen-Rit model, namely, C, BT, and GH arXiv:0907.2718 [math.DS]
Touboul J, Wendling F, Chauvel P, Faugeras O (2011)
bifurcations. Neural mass activity, bifurcations, and epilepsy. Neu-
ral Comput 23:3232–3286
Cross-References
▶ Anesthesia, Neural Population Models of

▶ Bifurcation Analysis Bimolecular Reactions, Modeling of
▶ Bifurcations Dynamics of Single Neurons and
Small Networks Stuart Edelstein
▶ Chaos, Neural Population Models and Babraham Institute, Cambridge, UK
▶ Epilepsy, Neural Population Models of
▶ Gamma Rhythm, Neural Population Models of
the Definition
▶ Jansen-Rit Model
▶ Deep Cerebellar Nuclei A bimolecular reaction refers to the chemical
▶ Pattern Formation in Neural Population combination of two molecular entities in
Models a reaction that can be considered either reversible
▶ Phase Transitions, Neural Population Models or irreversible. The reaction can involve two
and chemically distinct molecules, e.g., A + B, or
▶ Sleep, Neural Population Models of two identical molecules, e.g., A + A. The
reaction can be characterized with respect to
association and dissociation rate constants,
References which also define the equilibrium constant. How-
ever, it is also possible to establish the equilib-
Breakspear M, Jirsa VK (2007) Neuronal dynamics and rium constant without direct knowledge of the
brain connectivity. In: Jirsa VK, McIntosh AR (eds) rate constants. For many applications to biologi-
Handbook of brain connectivity, 1st edn. Springer, cal systems, one or both of the components pos-
Berlin/Heidelberg/New York, pp 3–64
Frascoli F, van Veen L, Bojak I, Liley DTJ sess multiple reaction sites and/or distinct
(2011) Metabifurcation analysis of a mean field conformational states, adding levels of complex-
model of the cortex. Physica D 240(11):949–962 ity that must be considered for a complete
Grimbert F, Faugeras O (2006) Bifurcation analysis of description of the system.
Jansen’s neural mass model. Neural Comput
18:3052–3068
Guckenheimer J (2007) Bifurcations. Scholarpedia
2(6):1517. http://www.scholarpedia.org/article/Bifur- Detailed Description
cations. Accessed 21 Mar 2013
Jansen BH, Rit VG (1995) Electroencephalogram and
visual evoked potential generation in For computational neuroscience encompassing
a mathematical model of coupled columns. Biol the molecular level, bimolecular reactions are
Cybern 73:357–366 encountered for virtually all topics considered.
B 384 Bimolecular Reactions, Modeling of
Neurons and their accompanying cells of other appropriate kinetic constants, koff/kon. Because
classes are living entities with a complete array of more than one binding site is involved, statistical
biochemical reactions, most of which are not factors come into play. For example, for a ligand
specific to the nervous system, such as the full binding to an unliganded dimer (n = 2), the on
complement of metabolic enzymes. Enzyme- rate must be increased by a factor of two, since
substrate reactions are one form of bimolecular two sites are available on the dimer. Similar rea-
reactions, but will not be considered here, since soning applies to the dissociation of ligand from
another chapter that included this topic was a fully or partially saturated oligomer. Generally,
recently completed (Endler et al. 2012). For reac- for a protein with n sites and i ligands bound, the
tions in individual neurons, only small numbers binding of an additional ligand is modified by
of molecules are involved in many cases and the statistical factor (ni + 1)/i. The binding
stochastic principles must be applied to consider reactions are illustrated in Fig. 1, along with the
the reactions appropriately. This subject will not
be treated, but general principles are well _
described elsewhere (Steinfeld et al. 1989), Equations for occupancy
_ (Y) and
along with a detailed description for ligand- fraction in R state (R):
gated channels based on terminologies that are
presented here (Edelstein et al. 1996). The focus
for this entry is on ligands binding to multisite
proteins that can exist in more than one function-
ally distinct conformational state, as commonly
occur for allosteric membrane proteins (Edelstein
and Changeux 2010) and other allosteric regula- Reaction cycle for dimer with two-states:
TRk
tory proteins, such as calmodulin (Stefan 2X + T0 0
R0 + 2X
et al. 2008). These phenomena are most effi- RTk
0
L0 = RTk0 / TRk0
ciently represented by the formalism of the Tk 2 Tkon Rk 2 Rkon
off off
Monod-Wyman-Changeux (MWC) model L1 = RTk1 / TRk1
TRk
(Monod et al. 1965; Edelstein et al. 1996; X + T1 1
R1 + X L2 = RTk2 / TRk2
RTk
Changeux and Edelstein 2005). 1
2 Tkoff Tk
on 2 Rkoff Rk
on
KR = Rkoff / Rkon
Formalism for Two Conformational States TRk KT = Tkoff / Tkon

2
T2 R2
The MWC model defines a multisite (oligomeric) RTk
2
protein with two conformational states, T and
R related by the allosteric constant, Li, where Energy diagram for dimer with two-states:
Li = [Ti]/[Ri]. Here “i” refers to the number of
TS0
ligand molecules bound up to the limit of “n”
molecules, usually equivalent to the number of
TS1
subunits in an oligomeric protein. Ligand is R0
represented by “X,” with distinct dissociation
T0 TS2
equilibrium constants for the T and R states: KT R1
and KR. The model is formulated such that T1
the T state is favored in the absence of ligand T2
R2
(L0 1), but the R state binds the ligand more
tightly (KR << KT). Hence, typically for a dimer
with two ligands bound, L2
1. Each reaction
equilibrium constant is the ratio of the Bimolecular Reactions, Modeling of, Fig. 1
Biomechanical Model of Low Back Pain 385 B
equations for occupancy and fraction of mole- PP2B and CaMKII. Proc Natl Acad Sci
cules in the R state. Finally an energy diagram USA 105:10768–10773
Steinfeld JI, Francisco JS, Hase WL (1989) Chemical
is presented in the figure to reveal how the kinetic kinetics and dynamics. Prentice-Hall, Englewood
constants are related to the thermodynamic Cliffs
parameters in terms of activation barriers fixed B
by linear free energy relations (Edelstein
et al. 1996).
Binary Neural Network
Limitations
The concepts presented here do not distinguish ▶ Hopfield Network
between the free and the total concentrations for
the ligand in question. Commonly the total con-
centration is assumed to be effectively indistin-
guishable from the free concentration, but under Binocular/Monocular Rivalry
conditions that approximate the physiological
state, this assumption is often in applicable. In ▶ Multistability in Perception Dynamics
that case, the binding of ligand can cause
a significant change in the free concentration
(a phenomenon known as “ligand depletion”)
and the appropriate corrections must be intro- Biomechanical Model of Low
duced, since the effects may be substantial Back Pain
(Edelstein et al. 2010).
Beth Winkelstein and Nicolas Jaumard
Department of Bioengineering, University of
References Pennsylvania, Philadelphia, PA, USA
Changeux JP, Edelstein SJ (2005) Allosteric

mechanisms of signal transduction. Science
308:1424–1428 Synonyms
Edelstein SJ, Changeux JP (2010) Relationships
between structural dynamics and functional kinetics Animal models of pain; Computational models of
in oligomeric membrane receptors. Biophys painful loading; Finite element models of the
J 98:2045–2052
Edelstein SJ, Schaad O, Henry E, Bertrand D, Changeux spine; Lumbar spine pain; Model systems for
JP (1996) A kinetic mechanism for nicotinic acetyl- spine pain; Spinal biomechanics; Tissue loading
choline receptors based on multiple allosteric transi- and pathology
tions. Biol Cybern 75:361–379
Edelstein SJ, Stefan MI, Le Novere N (2010) Ligand
depletion in vivo modulates the dynamic range and
cooperativity of signal transduction. PLoS One 5: Definition
e8449
Endler L, Stefan MI, Edelstein SJ, Le Novere N (2012) Low back pain can arise from a wide range of
Using chemical kinetics to model neuronal
signalling pathways. In: Le Novere N (ed) Computa- inciting conditions and pathologies in the lumbar
tional systems neurobiology. Springer, Dordrecht, and sacral spines. Symptoms can remain local-
pp 81–117 ized to the axial spine but also can extend to the
Monod J, Wyman J, Changeux J-P (1965) On the nature of torso and upper back, buttocks, legs, and feet,
allosteric transitions: a plausible model. J Mol Biol
12:88–118 depending on which tissues in the spine are
Stefan MI, Edelstein SJ, Le Novère N (2008) An allosteric affected and the extent of the related pathology.
model of calmodulin explains differential activation of The most common spinal tissues involved in
B 386 Biomechanical Model of Low Back Pain
producing low back pain are the intervertebral disc height, and increased motion (Shapiro and
disc, facet joints, nerve roots, and musculature. Risbud 2013).
Many different biomechanical models have been
developed in order to study low back pain. In In Vivo Animal Models
particular, such models mimic the mechanical Despite the potential differences in anatomy,
loading to the spine and/or its isolated tissues scaling, and physiology between animals and
that serve as sources of pain, and define the humans, in vivo models have tremendous utility
local tissue biomechanics for various normal for biomechanical modeling of LBP (DeLeo and
and abnormal spinal loading conditions. Biome- Winkelstein 2002; Winkelstein 2011), especially
chanical models focus on describing and given the ethical and financial advantages they
informing different aspects of the cascades offer while also providing meaningful insight on
related to tissue loading, mechanical responses, the mechanics of spinal tissues and the related
physiological cascades that are initiated and physiological and symptom responses. Animal
maintained, and the symptom response. models of nociception help in understanding neu-
ral mechanisms of pain from the spine, such as
arthritis, discogenic pain, and radicular nerve root
Detailed Description injury, all of which can lead to the development
of acute and chronic LBP (Deleo and Winkelstein
Biomechanical models can be categorized into 2002; McMahon and Koltzenburg 2005; Jaumard
several different types: animal and human cadav- et al. 2011).
eric models, in vivo animal models, and compu- The earliest studies using animal models of
tational and mathematical models. Over the chronic back pain involve direct surgical manip-
years, each of these modeling approaches has ulation of the sciatic nerve including sciatic
incorporated biomechanical techniques and cryoneurolysis (DeLeo et al. 1994); sciatic
together provides a complementary and increas- nerve transection, ligation, or crush (Przewlocka
ingly more detailed insight about the causes, con- et al. 1999; Obara et al. 2003); partial nerve injury
sequences, and complications associated with (Seltzer et al. 1990); chronic constriction injury
low back pain. (Bennett et al. 2003); and L5/L6 spinal nerve
ligation (Colburn et al. 1999). Although those
Cadaveric Models models have been important for defining the cas-
Cadaveric models of both humans and other spe- cades that lead to pain following neural injury in
cies have formed the basis of the early efforts in the spine, they have not standardized the role of
biomechanical modeling. They most often biomechanics. More recent work in the rodent
describe the spine’s kinematic and kinetic has defined the role of tissue loading for nerve
response to loading, as well as the mechanical root and disc injuries (Iatridis et al. 1999; DeLeo
response that individual tissues undergo during and Winkelstein 2002; Winkelstein et al. 2002;
spinal loading. In addition to defining the ana- Jaumard et al. 2011; Showalter et al. 2012). Only
tomical, material, and structural relationships recently, with the advances in measurement tech-
between the various tissues in the spine, these nologies, have local tissue mechanics been able
models have defined how such relationships to be incorporated in such models to relate spe-
vary with various loading configurations in the cific injury and loading environments to tissue
context of normal motions, injury, degeneration responses, neuroimmune cascades regulating
states, and following surgical treatments (White pain, and resulting symptoms of pain.
and Panjabi 1990; Winkelstein 2012). The
main findings have defined injury tolerance for Computational/Mathematical Models
failure of different tissues and the physical con- Due to the complexity of modeling injury
sequences of repetitive loading to certain tissues, and pain from tissue loading, cadaveric and
such as the production of ligament laxity, loss of in vivo modeling have only limited utility for
Biomechanical Model of Low Back Pain 387 B
Biomechanical Model of
Low Back Pain,
Fig. 1 Example of
a lumped-parameter model
of the rodent used to
evaluate spinal kinematics B
and kinetics in whole body
vibration along the spine’s
axis. The three masses (M1,
M2, M3) simulate the mass
of the three body segments.
Springs (K) and dampers
(C) simulate the
viscoelastic interactions
between the masses and the
masses and the vibrating
platform
modeling LBP. As such, computational models of the body to a direct or indirect mechanical
have been developed based on anthropomorphic excitation (Ma et al. 1995).
measurements that permit many loading configu- Finite element (FE) models are the most com-
rations to be tested. Such models extrapolate the plex since they mimic the anatomical and mate-
biomechanics of the lumbar spine for loading rial properties of the spinal tissues and their
conditions that cannot be tested using live sub- relationships based on different loading scenarios
jects, such as injury scenarios from vehicle and motions. Several FE models have been devel-
impacts, sports injury, and exposure to high- oped to evaluate spine biomechanics and simu-
frequency vibrations. This is particularly relevant late clinical procedures. Such models were
when modeling LBP in which ethical consider- designed to understand how changes in the
ations limit the exposures that are allowed. spine, such as disc degeneration, osteophytes,
Biomechanical computational models can be disc arthroplasty, facetectomy, or laminoplasty,
grouped as lumped-parameter, multibody, or affect the local mechanical response of its tissues
finite element models (Liang and Chiang 2006). (bone, ligament, cartilage), as well as the overall
Lumped-parameter models (Fig. 1) are generally spinal biomechanics (Rohlmann et al. 2007;
limited to only one-dimensional analysis, simu- Rundell et al. 2008; Zhang and Teo 2008; Kuo
lating the human body as an assembly of concen- et al. 2010). Other FE models provide utility for
trated masses, representing different parts of the evaluating tissue loading from spinal biomechan-
body, connected by springs and dampers that ics during nonphysiological injurious loading
simulate the viscoelastic properties of the bio- (Sairyo et al. 2005; Kitagawa et al. 2006; El-
logic tissues. The lumbar spine can be Rich et al. 2009). While these approaches have
represented as masses for the vertebrae connected become tremendously intricate, with detailed
by a spring and damper in parallel to simulate the physiological and mechanical processes captured
intervertebral disc (Keller et al. 2002). Multibody in the models, there are currently no such models
models consist of rigid bodies connected by that can simulate the nociceptive processes that
hinges and/or ball-and-socket joints and can be are initiated or modulated by spinal loading as it
used to study the kinematics and kinetic response relates to pain.
B 388 Biomechanical Model of Low Back Pain
Overall, computational models help to under- lumbar spine on facet joint force and intradiscal pres-
stand how the biomechanics of the lumbar spine sure: a finite element study. BMC Musculoskelet
Disord 11:151
and its associated spinal tissues can be altered and Liang CC, Chiang CF (2006) A study on biodynamic
enable inferences about the development of low models of seated human subjects exposed to vertical
back pain. The outcome of these models also vibration. Int J Ind Ergon 36:869–890
provides means to design interventions, devices, Ma D, Obergefell LA, Rizer A (1995) Development of
human articulating joint model parameters for crash
and therapeutic interventions to protect, repair, dynamics simulations. SAE Pap 952726:239–250
and treat the lumbar spine in order to avoid and/or McMahon S, Koltzenburg M (2005) Wall and Melzack’s
reduce the symptoms intrinsic to low back pain. textbook of pain, 5th edn. Churchill Livingstone,
London
Obara I, Mika J, Schafer MK, Przewlocka B (2003)
Cross-References Antagonists of the kappa-opioid receptor enhance
allodynia in rats and mice after sciatic nerve ligation.
Br J Pharmacol 140(3):538–546
▶ Pain Processing Pathway Models Przewłocka B, Mika J, Labuz D, Toth G, Przewłocki
▶ Spinal and Neuromechanical Integration: R (1999) Spinal analgesic action of endomorphins in
Overview acute, inflammatory and neuropathic pain in rats. Eur
J Pharmacol 367(2–3):189–196
Rohlmann A, Burra NK, Zander T, Bergmann G (2007)
Comparison of the effects of bilateral posterior
References dynamic and rigid fixation devices on the loads in the
lumbar spine: a finite element analysis. Eur Spine
Bennett GJ, Chung JM, Honore M, Seltzer Z (2003) J 16(8):1223–1231
Models of neuropathic pain in the rat. Curr Protoc Rundell SA, Auerbach JD, Balderston RA, Kurtz SM
Neurosci Chapter 9: Unit 9.14 (2008) Total disc replacement positioning affects
Colburn RW, Rickman AJ, DeLeo JA (1999) The effect of facet contact forces and vertebral body strains. Spine
site and type of nerve injury on spinal glial activation 33(23):2510–2517
and neuropathic pain behavior. Exp Neurol Sairyo K, Katoh S, Sasa T, Yasui N, Goel VK, Vadapalli
157(2):289–304 S, Masuda A, Biyani A, Ebraheim N (2005) Athletes
DeLeo JA, Winkelstein BA (2002) Physiology of chronic with unilateral spondylolysis are at risk of stress frac-
spinal pain syndromes: from animal models to biome- ture at the contralateral pedicle and pars
chanics. Spine 27(22):2526–2537 interarticularis: a clinical and biomechanical study.
DeLeo JA, Coombs DW, Willenbring S, Colburn RW, Am J Sports Med 33(4):583–590
Fromm C, Wagner R, Twitchell BB (1994) Character- Seltzer Z, Dubner R, Shir Y (1990) A novel behavioral
ization of a neuropathic pain model: sciatic model of neuropathic pain disorders produced in
cryoneurolysis in the rat. Pain 56(1):9–16 rats by partial sciatic nerve injury. Pain 43(2):205–218
El-Rich M, Arnoux PJ, Wagnac E, Brunet C, Aubin CE Shapiro IM, Risbud MV (eds) (2013) The intervertebral
(2009) Finite element investigation of the loading rate disc: molecular and structural studies of the disc in
effect on the spinal load- sharing changes under impact health and disease. Springer, Dordrecht
conditions. J Biomech 42(9):1252–1262 Showalter BL, Beckstein JC, Martin JT, Beattie EE,
Iatridis JC, Mente PL, Stokes IA, Aronsson DD, Alini M Espinoza Orı́as AA, Schaer TP, Vresilovic EJ, Elliott
(1999) Compression-induced changes in intervertebral DM (2012) Comparison of animal discs used in disc
disc properties in a rat tail model. Spine 24(10):996–1002 research to human lumbar disc: torsion mechanics and
Jaumard NV, Welch WC, Winkelstein BA (2011) Spinal collagen content. Spine 37(15):E900–E907
facet joint biomechanics and mechanotransduction in White AA, Panjabi MM (1990) Clinical biomechanics of
normal, injury and degenerative conditions. J Biomech the spine, 2nd edn. Lippincott, New York
Eng 133(7):071010 Winkelstein BA (2011) How can animal models inform on
Keller TS, Colloca CJ, Béliveau JG (2002) Force- the transition to chronic symptoms in whiplash? Spine
deformation response of the lumbar spine: a sagittal 36(25s):S218–S225
plane model of posteroanterior manipulation and Winkelstein BA (2012) Orthopaedic biomechanics. CRC
mobilization. Clin Biomech 17(3):185–196 Press/Taylor & Francis Group, Boca Raton
Kitagawa Y, Yasuki T, Hasegawa J (2006) A study of Winkelstein BA, Weinstein JW, DeLeo JA (2002) The
cervical spine kinematics and joint capsule strain in role of mechanical deformation in lumbar
rear impacts using a human FE model. Stapp Car Crash radiculopathy: an in vivo model. Spine 27(1):27–33
J 50:545–566 Zhang QH, Teo EC (2008) Finite element application in
Kuo CS, Hu HT, Lin RM, Huang KY, Lin PC, Zhong ZC, implant research for treatment of lumbar degenerative
Hseih ML (2010) Biomechanical analysis of the disc disease. Med Eng Phys 30(10):1246–56
BioModels Database: a Public Repository for Sharing Models of Biological Processes 389 B
to facilitate model retrieval. BioModels Database
BioModels can be accessed through a web interface and
programmatically using web services.
▶ BioModels Database: a Public Repository for All models are distributed under the terms of
Sharing Models of Biological Processes the Creative Commons CCO Public Domain B
Dedication License, ensuring that models are
available free for use, modification, and distribu-
tion for any user.
BioModels Database: a Public
Repository for Sharing Models of
Biological Processes Detailed Description
Vijayalakshmi Chelliah1, Nick Juty1, BioModels Database (Li et al. 2010) hosts math-
Camille Laibe1, Henning Hermjakob1 and ematical models of varying complexity, from
Nicolas Le Novère2,3 simple biochemical reaction systems to larger
1
European Bioinformatics Institute and complex dynamic models, metabolic net-
(EMBL-EBI), European Molecular Biology work models, and flux balance analysis (FBA)
Laboratory, Hinxton, Cambridge, UK models. It is a part of the BioModels.net initiative
2
Babraham Institute, Babraham Research (http://biomodels.net; Le Novère 2006), which
Campus, Cambridge, UK aims to help modelers to exchange and integrate
3
EMBL European Bioinformatics Institute, existing models, through the use of established
Cambridge, UK and shared standards.
Model Categories
Synonyms BioModels Database holds two major sets of
models: (A) models described in scientific litera-
BioModels; Mathematical models of biological ture and (B) models automatically generated
processes; Model repository from pathway resources.
A. Models published in peer-reviewed scientific

Definition literature
These models are included in the resource either
BioModels Database is a public repository for directly by a team of curators or after direct
mathematical models of biological processes. It submission by modelers or authors. Models
provides access to models described in peer- have also been contributed through established
reviewed scientific literature, as well as those collaborations with other model repositories,
generated automatically from pathway resources. such as the former SBML model repository
Model components, structure, and behavior of (Caltech, USA), JWS Online (http://jjj.
a large proportion of models in the database are biochem.sun.ac.za), the Database Of Quantita-
thoroughly checked. This manual procedure tive Cellular Signaling (DOQCS) (http://doqcs.
ensures model correspondence with the original ncbs.res.in), and the CellML Repository. More
publication. Furthermore, model components are importantly, several publishers of scientific
semantically enriched with cross-references to journals recommend deposition of models to
external ontologies and other relevant data BioModels Database in their author’s instruction
resources. Models are exported in various for- guidelines. These include Biophysical Journal,
mats, including SBML. Graphical representa- FEBS, PNAS, Nature Publishing Group (NPG),
tions of reaction networks are also provided. Public Library of Science (PLoS), Royal Society
Simple and advanced search options are available of Chemistry (RSC), and BioMed Central (BMC).
B 390 BioModels Database: a Public Repository for Sharing Models of Biological Processes
BioModels Database: a Public Repository for Shar- models. This figure illustrates the sequence of steps that
ing Models of Biological Processes, each model undergoes, from submission to publication
Fig. 1 Production pipeline of the literature-based
Model submission to BioModels Database is networks, and models that provide insufficient
free and open to everyone. Models are accepted quantitative results required for validation.
in two formats, SBML (Systems Biology Markup This branch also temporarily includes models
Language) (Hucka et al. 2003) and CellML that have not yet been curated due to time or
(Lloyd et al. 2004). Once submitted, each model resource constraints.
is assigned a unique and perennial identifier, Following curation and annotation (see sec-
which allows users to access and retrieve it. It tion “Model Annotation” below), models are
then passes through a series of checking steps tagged as ready for publication and become avail-
before it is published in BioModels Database able online with the following release of
(Fig. 1). Models are divided in two subbranches. BioModels Database. Releases are scheduled
Models in both subbranches are fully SBML two to four times a year.
compliant but differ in their curation status: Figure 2 shows the growth in number and size
i. Curated models: Models that satisfy the (complexity) of models, since its origin in 2005.
MIRIAM (Minimum Information Required Figure 3 shows the classification of curated
in the Annotation of Models) guidelines (Le models based on Gene Ontology (GO)
Novère et al. 2005) progress to the curated (Ashburner et al. 2000) terms present in the
branch. These models are thoroughly checked model annotation.
for standard compliance, correspondence with Certain simulation tools support only specific
the reference publication and reproducibility Levels or Versions of SBML. For this reason,
of results. BioModels Database provides models in alterna-
ii. Non-curated models: Models can be in the tive SBML versions, in addition to the version
non-curated branch for several reasons: non- that was used by the curators to check the model.
MIRIAM-compliant models (e.g., they do not It also provides models in a variety of other
reproduce results published in the reference formats such as BioPAX (http://www.biopax.
publication), pathway maps or models of org/), the Virtual Cell Markup Language (VCML)
BioModels Database: a Public Repository for Sharing Models of Biological Processes 391 B
BioModels Database: a Public Repository for Shar- “assignment rules,” and “events.” There has been approx-
ing Models of Biological Processes, Fig. 2 Growth of imately a 20-fold increase in the number of models since
BioModels Database. Number of models (blue) and num- the launch of the resource in 2005, with the average
ber of variable relationships (green). The number of rela- complexity of the models being increased five times dur-
tionships includes SBML’s “reactions,” “rate rules,” ing the same period
BioModels Database: a Public Repository for Sharing Models of Biological Processes, Fig. 3 Classification of
models based on GO annotation of curated models
(http://vcell.org/), XPPAUT (http://www.math.pitt. in the Systems Biology Graphical Notation

edu/bard/xpp/xpp.html), Scilab (http://www. (SBGN) (Le Novère et al. 2009) is available in
scilab.org/), Octave (m-file) (http://www.gnu.org/ PNG and SVG formats.
software/octave/), and PDF (http://www.ra.cs.uni- BioModels Database provides a basic online
tuebingen.de/software/SBML2LaTeX). Further- simulation using SOSlib (Machne et al. 2006).
more, the reaction network of the model encoded The simulation results are returned both in
B 392 BioModels Database: a Public Repository for Sharing Models of Biological Processes
graphical and textual forms. An additional and models are cross-linked with other database
more flexible simulation tool is available for entries and terms from controlled vocabularies.
many models using JWS Online. Some models Annotations are included in the models directly
are described in greater detail in the “Model of using Identifiers.org URIs (Juty et al. 2012). Such
the Month.” This takes the form of a brief article annotations provide a consistent and perennial
discussing the biological background, signifi- way to identify model components, and an anchor
cance, structure, and results of a particular which may be used to merge, or expand the scope
model. Another important feature provided for of models.
curated models is the ability to generate sub-
models from a user-selected set of model ele- Model Search, Browse, and Retrieval
ments. These sub-models can be downloaded in BioModels Database provides a complete list of
SBML and used, for instance, in the modular models available for browsing. Curated models
design of new models or to expand existing can also be browsed based on taxonomy or Gene
models. Ontology (GO) terms used in their annotation.
Users can also search and retrieve models of
B. Models Generated from Pathway Resources interest using the annotations and related
(Path2Models) information.
Driven by the growing number of biochemical
pathways and reconstructions, the Path2Models Web Services
project (B€uchel 2013) targeted the conversion of BioModels Database provides web services
pathway information initially from KEGG (http://www.ebi.ac.uk/biomodels-main/web-services),
(Kanehisa and Goto 2000), BioCarta (Schaefer allowing, for example, direct search and retrieval,
et al. 2009), MetaCyc (Karp et al. 2000), and and the creation of sub-models. A Web Services
SABIO-RK (Wittig et al. 2012) into standard- Description Language (WSDL) file enables soft-
compliant computational models. ware to understand available functions and their
There are currently three major types of usage. The complete list of available methods, as
models in this branch: well as a Java library and the associated docu-
i. Genome-scale metabolic reconstructions: mentation is provided on the BioModels Data-
These models are generated by extracting base website.
pathway data from the KEGG and MetaCyc BioModels Database is developed under the
databases and subjected to FBA. GNU General Public License and the software is
ii. Quantitative, kinetic models of metabolic freely available from its SourceForge repository
pathways: These models are generated (http://www.ebi.ac.uk/biomodels-main/webservices).
based on the metabolic pathways distributed
by KEGG described as processes, in combi-
Cross-References
nation with experimentally determined rate
laws and parameter values from the SABIO-
▶ Systems Biology Markup Language (SBML)
RK database.
iii. Qualitative, logical models of non-metabolic
(primarily signaling) pathways: These
References
models are generated based on the non-
metabolic pathways distributed by KEGG, Ashburner M et al (2000) Gene ontology: tool for the
with additional information from BioCarta unification of biology. The gene ontology consortium.
pathways. Nat Genet 25(1):25–29
B€
uchel Fea (2013) Path2Models: large-scale generation of
computational models from biochemical pathway
Model Annotation maps. BMC Syst Biol 7:116
In order to facilitate efficient search and enhance Hucka M et al (2003) The systems biology markup lan-
interpretability both by users and software tools, guage (SBML): a medium for representation and
Biophysical Models of Olfactory Mitral and Granule Cells 393 B
exchange of biochemical network models. Bioinfor-
matics 19(4):524–531 Biophysical Mechanisms that
Juty N, Le Novère N, Laibe C (2012) Identifiers.org
and MIRIAM registry: community resources to Photoreceptors Use to Sample Light
provide persistent identification. Nucleic Acids Res Information
40(Database issue):D580–D586 B
Kanehisa M, Goto S (2000) KEGG: kyoto encyclopedia of ▶ Phototransduction Biophysics
genes and genomes. Nucleic Acids Res 28(1):27–30
Karp PD et al (2000) The EcoCyc and MetaCyc databases.
Nucleic Acids Res 28(1):56–59
Le Novère N (2006) Model storage, exchange and inte-
gration. BMC Neurosci 7(Suppl 1):S11. Biophysical Models of Olfactory
Le Novère N et al (2005) Minimum information requested
in the annotation of biochemical models (MIRIAM). Mitral and Granule Cells
Nat Biotechnol 23(12):1509–1515
Le Novère N et al (2009) The systems biology graphical Guoshi Li
notation. Nat Biotechnol 27(8):735–741 Department of Psychology, Cornell University,
Li C et al (2010) BioModels database: an enhanced,
curated and annotated resource for published quantita- Ithaca, NY, USA
tive kinetic models. BMC Syst Biol 4:92
Lloyd CM, Halstead MD, Nielsen PF (2004) CellML: its
future, present and past. Prog Biophys Mol Biol Definition
85(2–3):433–450
Machne R et al (2006) The SBML ODE solver library:
a native API for symbolic and fast numerical analysis Biophysical models of olfactory mitral and gran-
of reaction networks. Bioinformatics ule cells are conductance-based compartmental
22(11):1406–1407 models based on the morphological and electro-
Schaefer CF et al (2009) PID: the pathway interaction
database. Nucleic Acids Res 37(Database issue): physiological properties of mitral and granule
D674–D679 cells. Each cellular model incorporates
Wittig U et al (2012) SABIO-RK – database for biochem- a number of interconnected sections that repre-
ical reaction kinetics. Nucleic Acids Res 40(Database sent physical cellular elements such as the soma,
issue):D790–D796
axon, dendrites, and/or specialized compartments
such as spines. Each section, in turn, consists of
one or more isopotential compartments, each of
which contains passive and active ionic channels
Bionic Ear along with other mechanisms such as calcium
buffering, typically modeled using the Hodgkin-
▶ Auditory Prosthesis Huxley formalism (Hodgkin and Huxley 1952).
The parameters of these coupled equations are
then adjusted to reproduce the salient membrane
properties of mitral and granule cells. These bio-
Bionic Eye physical models serve as an important tool to
understand information processing within the
▶ Vision Prosthesis olfactory bulb (OB) quantitatively.
▶ Visual Prosthesis, Epiretinal Devices
▶ Visual Prosthesis, Optogenetic Approaches
Morphological and Electrophysiological

Bionic Vision Properties of Mitral Cells
Mitral cells (MCs) in OB have several salient
▶ Prosthetic Vision, Perceptual Effects morphological characteristics (Shepherd and
▶ Visual Prosthesis, Optogenetic Approaches Greer 1998). The cell body sends out a single
B 394 Biophysical Models of Olfactory Mitral and Granule Cells
primary (apical) dendrite and several secondary Development of a Biophysical MC Model:

(lateral) dendrites. The primary dendrite extends An Example
several hundred micrometers and arborizes in Based on the morphology of mitral cells,
a complex tuft within a glomerulus, while lateral a reduced, oligocompartmental MC model could
dendrites are widely distributed in the external consist of four sections: a soma, an apical den-
plexiform layer (EPL). The input resistance drite, a lateral dendrite, and a glomerular tuft at
ranges from 88 to 280 MΩ, and the mean resting the distal end of the apical dendrite (Fig. 1a).
potential is about 63 mV (Desmaisons Each section is modeled as a cylinder whose
et al. 1999; Cang and Isaacson 2003). In response dimensions are based on morphological measure-
to depolarizing current pulses, MCs generate ments from mitral cells; reasonable examples are
clusters of action potentials interspersed with shown in Table 1 (Li and Cleland 2013). The long
long period of subthreshold oscillations (STOs) apical and lateral dendrite sections can be further
(Chen and Shepherd 1997; Desmaisons divided into multiple compartments to more
et al. 1999; Balu et al. 2004). Inhibitory postsyn- accurately model cable properties (individual
aptic potentials (IPSPs) reset the phase of the compartments are by definition isopotential).
STOs and also can generate rebound action For example, in the Li and Cleland model
potentials (Desmaisons et al. 1999). (Li and Cleland 2013), the apical dendrite com-
Biophysical Models of Olfactory Mitral and Granule circuit of two identical interconnected neural compart-
Cells, Fig. 1 Compartmental model of a mitral cell (MC, ments used to simulate MC cell excitability. Subscripts
Li and Cleland 2013). (a), Schematic representation of the c1..ci denote i different active (variable) conductances and
MC model with distribution of active ionic conductances their associated reversal potentials (GNa, GKS, etc., with
in each of its four sections. (b), Equivalent electrical reversal potentials ENa, EKS, etc.)
Biophysical Models of Olfactory Mitral and Granule Cells, Table 1 Morphological parameter values and passive
electrical properties of an MC model (Li and Cleland 2013)
Diameter (mm) Length (mm) Cm (mF/cm2) Rm (O cm2) Ra (O cm) Em (mV)
Soma 20 25 1.2 30,000 70 60
Apical dend 3.5 370
Tuft 0.5 20
Lateral dend 3.4 500
prises five compartments, and the lateral dendrite dx
¼ fx ðax ðV Þð1 xÞ bx ðV ÞxÞ (3)
comprises seven compartments, whereas the dt
electrotonically compact soma comprises
a single compartment. where ax(V ) and bx(V ) are the voltage-dependent
Generally, each compartment contains rate constants. The parameter values for the rate B
a passive leakage current and a number of active constants ax and bx, steady state x1, and time
ionic currents (Fig. 1a, compartments within constant tx of the gating variables determine the
a section have identical currents). The Li and model’s dynamical properties and are typically
Cleland model contains seven active currents: based on electrophysiological measurements of
fast, spike-generating (INa), and persistent (INaP) mitral cells (e.g., voltage clamp or current
sodium currents; a potassium delayed rectifier clamp). The excitability of a segment of MC
IDR; two transient potassium currents (fast- cell membrane modeled by a compartment of
inactivating IA and slow-inactivating IKS); an the type depicted in Fig. 1b can then be described
L-type high-voltage-activated calcium current by the following equation:
(ICaL); and a Ca2+-dependent potassium current
IKCa, based on previous experimental and model- dV m ðEm V m Þ X
ing studies (Bhalla and Bower 1993; Wang cm ¼ þ Gci ðEci V m Þ
dt Rm i
et al. 1996; Desmaisons et al. 1999; Davison 0 00
et al. 2000; Balu et al. 2004; Rubin and Cleland Vm Vm Vm Vm
þ þ (4)
2006; Balu and Strowbridge 2007). Figure 1b Ra Ra
shows the equivalent electrical circuit of two
basic neural compartments connected with an where V0 m and V00 m are the voltages of the two
axial resistance Ra. Each compartment has adjacent compartments. Solving many coupled
a membrane capacitance Cm, a fixed membrane equations of the form in Eq. 4, with appropriate
resistance Rm, and an equilibrium potential Em parameters, will generate the dynamical behavior
associated with the ohmic leakage current that of a mitral cell. Fortunately, specialized neuronal
flows across Rm (values shown in Table 1 for simulators such as NEURON (Carnevale and
the Li and Cleland model). Hines 2006) and GENESIS (Bower and Beeman
All ionic conductance kinetics in this model 2003) have made the implementation and numer-
are modeled using the Hodgkin-Huxley formal- ical simulation of biophysical models easy and
ism (Hodgkin and Huxley 1952). Specifically, the straightforward.
conductance for channel i, Gci, is modeled as: After implementation of the model, the next
step is to select suitable parameters, so the model
Gci ¼ gci mp hq (1) reasonably reproduces the experimental data
from a particular cell type of interest. This usually
where gci is its maximal conductance density, involves adapting parameter values from previ-
m its activation variable (with exponent p), and ous literature, estimating or inferring them from
h its inactivation variable (with exponent q). The experimental data, and finally performing
kinetic equation for the gating variable x (m or h) a parameterization and fitting procedure. Param-
satisfies a first-order kinetic model, eterization can be performed via exhaustive
searching (brute force), supervised trial and
dx x1 ðV Þ x error, and/or any of a number of more advanced,
¼ fx (2)
dt tx ðV Þ automated parameterization techniques. The Li
and Cleland model, after parameterization, repro-
where fx is a temperature-dependent factor, duces the experimentally observed activation pat-
x1(V) is the voltage-dependent steady state, and terns of mitral cells such as delayed first spike,
tx(V) is the voltage-dependent time constant. clusters of action potentials interspersed with
Equivalently, Eq. 2 can be written as: subthreshold oscillations (STOs), STO phase
reset, and rebound spikes generated by IPSPs implementational details depending on the inves-
(Li and Cleland 2013). tigators’ particular goals.
Other Biophysical MC Models Morphological and Electrophysiological

A number of biophysical MC models incorporat- Properties of Granule Cell
ing different levels of detail and complexity have Granule cells (GCs) differ from mitral cells – e.g.,
been developed (Bhalla and Bower 1993; Shen very different morphologies, channel compo-
et al. 1999; Davison et al. 2000; Migliore nents, and kinetics – and consequently are
et al. 2005; Bathellier et al. 2006; Rubin and modeled differently. GCs have a soma (~8 mm
Cleland 2006; Inoue and Strowbridge 2008). diameter), which connects to a number of basal
The model developed by Bhalla and Bower is dendrites in the vicinity of the cell body, and
a detailed, 286-compartment model based on several thin (approximately 1–2 mm), long
intracellular recordings and morphological (up to about 600 mm) radial dendrites projecting
measurements. It aimed to study the details of to the EPL, where they contact mitral/tufted cell
single-cell morphology and channel distribution secondary dendrites via large dendritic spines
in cellular information processing. This complex (Egger et al. 2003; Pinato and Midtgaard 2003).
model was reduced to two-, three- and four- Two prominent features of GCs are their
compartment models by Davison et al. (2000) extremely high input resistances and Na+ spike
for computational efficiency in larger networks thresholds. Depending on the species, the input
(Davison et al. 2003). Neither model resistance ranges from 0.5 to 2 GΩ, and the spike
implemented STOs or intermittent cluster firing threshold is about 28 mV (Egger et al. 2003;
patterns, however, which are important mem- Pinato and Midtgaard 2003; Pressler and
brane characteristics of MCs (Desmaisons Strowbridge 2006). The resting potential is
et al. 1999; Balu et al. 2004); consequently, about 65 mV for rats (Cang and Isaacson
Rubin and Cleland (2006) adapted the four- 2003; Pressler and Strowbridge 2006), and the
compartment Davison model to replicate these membrane time constant is about 50 ms (Pinato
dynamical properties by adding additional cur- and Midtgaard 2003). In response to positive
rent mechanisms as described by Desmaisons current injection, GCs fire a short burst of Na
et al. (1999). The Shen et al. (1999) and Migliore spikes with a gradual attenuation of spike
et al. (2005) models are detailed, multi- amplitude and slowly activating “ramping”
compartmental biophysical models, but, as they depolarization often terminating in a sustained
sought only to study action potential generation plateau near 20 mV (Wellis and Kauer 1994;
and inter-glomerular synaptic transmission, they Hall and Delaney 2002; Egger et al. 2003; Pinato
included only the fast sodium and delayed potas- and Midtgaard 2003). Biophysical models of
sium channels. Bathellier et al. (2006) developed granule cells should replicate these basic
a single-compartment model with five core cur- properties.
rents capable of reproducing the essential firing
properties of MCs for implementation of an oscil- Development of a Biophysical GC Model:
latory network. This single-compartment model An Example
was very efficient in network simulation, A simple yet functional oligocompartmental GC
although at the price of neglecting the signal model can consist of a soma and a radial dendrite
interactions between the soma and the dendrites connected to a spine shaft and body (Fig. 2; Li
(e.g., forward and back propagation). The model and Cleland 2013). Omitting detailed simulations
by Inoue and Strowbridge (2008) comprised two of multiple spines, for example, this model was
compartments (a soma and a dendrite) and uti- built for efficient integration of synaptic inputs
lized reduced current mechanisms in order to from MCs in a network model. Its morphological
study persistent activity in OB. Thus, biophysical parameter values and passive electrical properties
models of MCs can vary substantially in are shown in Table 2. The GC model cell contains
Biophysical Models of Olfactory Mitral and Granule currents. The spine compartments have the same ionic
Cells, Fig. 2 Schematic representation of the Li and currents as the radial dendrite
Cleland GC model with distribution of active ionic
Biophysical Models of Olfactory Mitral and Granule Cells, Table 2 Morphological parameter values and passive
electrical properties of a GC model (Li and Cleland 2013)
Diameter (mm) Length (mm) Cm (mF/cm2) Rm (O cm2) Ra (O cm) Em (mV)
Soma 8 8 2 30,000 70 60
Dend 1 150
Spine shaft 1 1
Spine body 1 1
the following eight active ionic currents: a fast, Other Biophysical GC Models
spike-generating sodium current INa; a potassium Several other biophysical GC models have been
delayed rectifier IDR; a transient A-type potas- developed, exhibiting different levels of details
sium current IA; a noninactivating muscarinic and complexity (Bhalla and Bower 1993;
potassium current IM; a low-threshold Davison 2001; Inoue and Strowbridge 2008;
inactivating (ICaT) and a high-threshold (ICaP/N) Migliore and Shepherd 2008). The model devel-
calcium currents; a Ca2+-activated nonspecific oped by Bhalla and Bower is a detailed,
cation current ICAN, and a Ca2+-dependent potas- 944-compartment model based on intracellular
sium current IKCa (Li and Cleland 2013). The GC recordings and morphological measurements.
is modeled as an equivalent electrical circuit of That complex model was reduced by Davison
connected compartments as described above for (2001) to a three-compartment model that
the MC. exhibited most of the original model’s physiolog-
After parameterization, the Li and Cleland ical properties and was later implemented in
model reproduced the experimentally observed a network simulation (Davison et al. 2003).
salient properties of GCs, including delayed first A separate GC model developed by Migliore
spike generation, ramping depolarization, and and Shepherd (2008) contained a soma, a main
muscarinic neuromodulatory effects including radial dendrite, and three second-order dendrites
the conversion of afterhyperpolarizations representing the medial and distal dendritic tree.
(AHPs) into afterdepolarizations (ADPs) and This model only included three conductances
conditional persistent firing (Li and Cleland (Na, KDR, KA), omitting all Ca2+ and Ca2+-
2013). This type of model enables the direct dependent currents. To investigate persistent fir-
association of channel properties with cellular ing properties in GCs, Inoue and Strowbridge
responses. For example, the initial spike delay is (2008) developed a two-compartment GC model
caused by an “A”-type potassium current IA, containing low- and high-threshold Ca2+ cur-
whereas blocking the IM and IKCa currents abol- rents, a Ca2+-dependent nonselective cation cur-
ishes AHPs, thereby revealing latent ADPs medi- rent, a sodium current, and two potassium
ated by the nonspecific cation current ICAN currents. The model was able to reproduce Ca2+
(Pressler et al. 2007). spikes and Ca2+- and voltage-dependent ADP
responses as observed experimentally. It also Cang J, Isaacson JS (2003) In vivo whole-cell recording of
suggested that persistent spiking in granule cells odor-evoked synaptic transmission in the rat olfactory
bulb. J Neurosci 23:4108–4116
results from the interaction between a voltage- Carnevale NT, Hines ML (2006) The neuron book. Cam-
dependent after depolarization and a low- bridge University Press, Cambridge
threshold Ca2+ spike. Chen WR, Shepherd GM (1997) Membrane and synaptic
properties of mitral cells in slices of rat olfactory bulb.
Brain Res 745:189–196
Summary Davison AP (2001) Mathematical modeling of informa-
tion processing in the olfactory bulb. Ph.D. thesis.
University of Cambridge, Cambridge
Biophysical models of olfactory mitral and gran- Davison AP, Feng J, Brown D (2000) A reduced compart-
ule cells are conductance-based compartmental mental model of the mitral cell for use in network
models constrained by the specific morphological models of the olfactory bulb. Brain Res Bull
51:393–399
and electrophysiological properties of mitral and
Davison AP, Feng J, Brown D (2003) Dendrodendritic
granule cells. The general procedures for build- inhibition and simulated odor responses in a detailed
ing such models include three steps: (1) represent olfactory bulb network model. J Neurophysiol
the cell morphology with interconnected 90:1921–1935
Desmaisons D, Vincent JD, Lledo PM (1999) Control of
isopotential compartments, (2) decide upon
action potential timing by intrinsic subthreshold oscil-
appropriate ionic channels in each compartment lations in olfactory bulb output neurons. J Neurosci
based on their physiological properties, and 19:10727–10737
(3) select suitable parameters so the model Egger V, Svoboda K, Mainen ZF (2003) Mechanisms of
lateral inhibition in the olfactory bulb: efficiency and
exhibits patterns of activity that correspond to
modulation of spike-evoked calcium influx into gran-
experimental data. Depending on the particular ule cells. J Neurosci 23:7551–7558
goals of interest, different biophysical mitral and Hall BJ, Delaney KR (2002) Contribution of a calcium-
granule cell models incorporating different levels activated non-specific conductance to NMDA
receptor-mediated synaptic potentials in granule cells
of detail and complexity have been developed.
of the frog olfactory bulb. J Physiol 543:819–834
These biophysical models serve as important Hodgkin AL, Huxley AF (1952) A quantitative
tools to integrate and interpret experimental data description of membrane current and its application
and to formulate quantitative working hypotheses to conduction and excitation in nerve. J Physiol
117:500–544
concerning olfactory bulb function.
Inoue T, Strowbridge BW (2008) Transient activity
induces a long-lasting increase in the excitability of
olfactory bulb interneurons. J Neurophysiol
References 99:187–199
Li G, Cleland TA (2013) A two-layer biophysical model
Balu R, Strowbridge BW (2007) Opposing inward and of cholinergic neuromodulation in olfactory bulb.
outward conductances regulate rebound spiking in J Neurosci 33:3037–3058
olfactory bulb mitral cells. J Neurophysiol Migliore M, Shepherd GM (2008) Dendritic action poten-
97:1959–1968 tials connect distributed dendrodendritic microcir-
Balu R, Larimer P, Strowbridge BW (2004) Phasic stimuli cuits. J Comput Neurosci 24:207–221
evoke precisely timed spikes in intermittently Migliore M, Hines ML, Shepherd GM (2005) The role of
discharging mitral cells. J Neurophysiol 92:743–753 distal dendritic gap junctions in synchronization of
Bathellier B, Lagier S, Faure P, Lledo PM (2006) Circuit mitral cell axonal output. J Comput Neurosci
properties generating gamma oscillations in a network 18:151–161
model of the olfactory bulb. J Neurophysiol Pinato G, Midtgaard J (2003) Regulation of granule cell
95:2678–2691 excitability by a low-threshold calcium spike in turtle
Bhalla US, Bower JM (1993) Exploring parameter space olfactory bulb. J Neurophysiol 90:3341–3351
in detailed single neuron models: simulations of the Pressler RT, Strowbridge BW (2006) Blanes cells mediate
mitral and granule cells of the olfactory bulb. persistent feedforward inhibition on granule cells in
J Neurophysiol 69:1948–1965 the olfactory bulb. Neuron 49:889–904
Bower JM, Beeman D (2003) The book of GENESIS: Pressler RT, Inoue T, Strowbridge BW (2007) Muscarinic
exploring realistic neural models with the General receptor activation modulates granule cell excitability
Neural Simulation System. Internet Edition. http:// and potentiates inhibition onto mitral cells in the rat
www.genesis-sim.org/GENESIS olfactory bulb. J Neurosci 27:10969–10981
Biophysical Models: Neurovascular Coupling, Cortical Microcircuits, and Metabolism 399 B
Rubin DB, Cleland TA (2006) Dynamical mechanisms of in functional neuroimaging, a variety of such
odor processing in olfactory bulb mitral cells. extended biophysical models has been consoli-
Shen GY, Chen WR, Midtgaard J, Shepherd GM, Hines dating around three main research topics over the
ML (1999) Computational analysis of action potential last decade: (a) the principles for neurovascular
initiation in mitral cell soma and dendrites based on coupling, (b) the organization of cortical micro- B
dual patch recordings. J Neurophysiol 82:3006–3020 circuits, and (c) the cellular metabolism with
Shepherd GM, Greer CA (1998) Olfactory bulb. In: Shep-
herd GM (ed) Synaptic organization of the brain. emphasis on the nerve-ending particles. After
Oxford University Press, Oxford, pp 159–178 modeling from mesoscopic variables to neuroim-
Wang XY, McKenzie JS, Kemm RE (1996) Whole-cell aging data, inversion of generative models pro-
K+ currents in identified olfactory bulb output neurons vides insight into these underlying mechanisms
of rats. J Physiol 490:63–77
Wellis DP, Kauer JS (1994) GABAergic and of brain function and cognition.
glutamatergic synaptic input to identified granule
cells in salamander olfactory blub. J Physiol
475:419–430 Detailed Description
Neuroimaging data constitute different reflec-

tions, governed by basic principles in physics,
Biophysical Models: Neurovascular of a common cause originated from the activity
Coupling, Cortical Microcircuits, and of connected cellular networks. Therefore, for-
Metabolism mulating biophysical models that embrace both
the underlying causal phenomena and the obser-
Jorge Riera vational principles is essential for (Stephan
Department of Biomedical Engineering, Florida et al. 2006):
International University, Miami, FL, USA (a) Performing statistical inference about the cel-
lular states, exogenous inputs, and network
topology
Synonyms (b) Combining different modalities of functional
neuroimaging data
Forward-generative models of functional neuro- The state-space formalism constitutes one of
imaging data the most general forms for representing these
models. In this formalism, a state-space equation
is used to describe the evolution of the cellular
Definition activity, and an observation equation specifies

!
how these hidden (cellular) states x r , t pro-
Forward-generative models are considered cru- duce discrete neuroimaging data ytk .
cial to interpret and integrate data obtained with The state equation:
different functional neuroimaging modalities
! ! ! !
(Riera 2014). Typically, these models are formu- x r , t ¼ f x r , t½t t, tÞ , i r , t , Y r
lated to represent biological principles at the
(1)
mesoscopic scale, which basically stand for an

average voxel (1 mm3) in any brain imaging ! !
The states x r , t ¼ x1 r , t ,
technique. To this end, biophysical models for
! !
neuronal activity, which are mostly available x2 r , t , , xN r , t are represented by an
from the neuronal computation community, N-dimensional vector field that dynamically
have been modified to incorporate general phys- evolves on time as describedby Eq.
1. An exog-
!
iological mechanisms governing glial cell activ- enous multivariate input i r , t is assumed
ity, vascular dynamics, and metabolism. For the to perturb the system at each time instant.
cerebral cortex, the most relevant brain structure In principle, this input could include both
B 400 Biophysical Models: Neurovascular Coupling, Cortical Microcircuits, and Metabolism

! ! time, it represents a heteroscedastic state-space
deterministic u r , t and stochastic x r , t
model for nonstationary time series. Compart-
!
components. A vector field Y r is included mental GARCH models are very useful for
to represent important physiological parameters dynamic nonstationary signal decomposition.
of the neuronal network (e.g., synaptic strength, For neuroimaging, time series can be assumed
metabolic rate, vascular reactivity). In general, in general stationary (Sltk ¼ Sl ). Function fl can

!
the state vector x r , t depends through the be represented by the Hammerstein, ExpAR, and
function f in Eq. 1 on its present/t-past values. RBF-AR models for nonlinear dynamics. An
Note that since neuronal networks are intrinsi- innovation-based approach is typically used for
cally autonomous systems, the dependency with the characterization of nonlinear and/or
time is exclusively by way of the exogenous non-Gaussian time series. This approach is com-
input. bined with maximum likelihood criteria to define
The observation equation: proper statistical identification methods. Hidden
cellular states can be estimated using Kalman and

! extended Kalman filtering techniques. There
ytk ¼ g x r , tk þ etk (2)
exist useful empirical (e.g., bootstrap, cross-
validation, Jackknife) and theoretical (Akaike
The data vector ytk comprises neuroimaging
n o information criteria “AIC,” Bayesian information
observations ylt,k m from different locations criteria “BIC”) criteria for model selection. We
(subscript l) and modalities (superscript m). can use some version of Granger causal modeling
Function g in Eq. 2 does not depend explicitly (i.e., Wiener–Akaike–Granger–Schweder influ-
on time. The detailed form of this function for ence, WAGS causality) to study the topology of
each particular modality could be found in Toga the cellular networks. Although discrete nonpara-
and Mazziotta (2002). They may depend on metric models are not formulated on pure biolog-
instrument characteristics, sensor specificities, ical assumptions, they are very useful to explore
and experimental conditions. An independent specific hypotheses about working principles of
multivariate white noise etk is typically added to cellular networks and accordingly narrow the
the observations. Figure 1 summarizes family of plausible biophysical models. The fol-
cause–effect mechanisms and mesoscopic states lowing literature is recommended for details
underlying biophysical models in neuroimaging about this particular type of state-space models
as well as typical observation modalities and (Ozaki 2012, 2014; Valdes-Sosa et al. 2011;
techniques. Stephan and Roebroeck 2012). This type of
The state Eq. 1 is written in a general form. model has been extendedly used to describe
There are two relevant cases that are of interest. EEG/MEG and fMRI data. A model by Riera
Discrete Nonparametric Models et al. (2005) constitutes one of the most represen-
tative examples of using a discrete nonparametric
n o
xltk ¼ f l xltkk0 , k0 ¼ 0, , N l , ultk , Yl þ jltk model for the combination of both multimodal
neuroimaging and the statistical inference about
(3a) network dynamics/topology.
Continuous Parametric Models
This constitutes a type of multivariable
autoregressive (MAR) model, which is useful to
dxl ðtÞ ¼ f l xl ðtÞ, ul ðtÞ, Yl dt þ Gl dwl ðtÞ (3b)
characterize discrete cellular state dynamics. For
each voxel, the system comprises a deterministic
This constitutes a stochastic differential equa-
and time-varying external input ultk as well as
tion (SDE), which is useful to represent cellular
a vector of parameter Yl. In this particular case,
states with a continuous dynamics. Equation 3b
the system noise is represented by a multivariate
can be used to represent the states of single cells
white noise jltk 0, Sltk . When Sltk varies with in a population, which are affected by
Biophysical Models: Neurovascular Coupling, effect of the functional hyperemic response. The final
Cortical Microcircuits, and Metabolism, mesoscopic variables are as follows: the primary current
Fig. 1 Cause–effect relationship between the cellular density (PCD), the cerebral metabolic rate of oxygen
states and actual observations. A particular inducing (CMRO2), the deoxyhemoglobin (dHb), the cerebral
event will trigger neuronal (i.e., action potential “AP,” blood flow (CBF), and the cerebral blood volume (CBV).
postsynaptic potential “PSP”) and astrocytic (i.e., calcium There are constitutive relationships between some of these
signaling) activities. These activities will cause three mesoscopic variables. They are either directly or indi-
major responses: (a) a flow of ionic currents across the rectly reconstructed from functional neuroimaging (e.g.,
cortical microcircuits, (b) the metabolizing of oxygen and EEG/MEG, PET/SPECT, NIRS, and MRI) by a variety of
glucose to build the required energy substrates, and (c) the techniques, e.g., current source imaging (CSI); 99 mTc
functional hyperemic response necessary for new supplies hexamethylpropyleneamine oxime (99mTc-HMPAO) by
and residual cleaning. The ionic currents will create vol- SPECT; fluorodeoxyglucose (FDG) and 15O2 by PET;
umetric current source densities (vCSD) within a voxel of the differential path-length factor (DPF) method by
interest. Changes in the D-glucose concentration (Gluc) NIRS; calibrated fMRI, blood-oxygen-level dependent
and the oxygen partial pressure (pO2) will result from the (BOLD) signal, arterial spin labeling (ASL), 13C MRS,
metabolism of these two molecules. A reduction in the and 17O2 by MRI
cerebral vascular resistance (CVR) constitutes the primary
a deterministic external input ul ðtÞ and a stochas- The Fokker–Planck equation describes the
tic thermal noise wl ðtÞ (i.e., Wiener process, time evolution of the probability density function
0
E wl ðt þ tÞ, wl ðtÞ ¼ 2DdðtÞdl, l0 Þ for each pl xl , t of the cellular states xl ðtÞ that follow the
voxel. A vector Yl is also included to represent SDEs Eq. 3b.
biologically meaningful parameters. While deal-
ing with large amount of interacting cells, theo-
dpl xl , t
retical frameworks for assembles are very useful ¼ ∇ f l xl ðtÞ, ul ðtÞ, Yl pl xl , t dt
dt
(summary in Fig. 2). Among them, the 1 X @ 2 l l
Fokker–Planck equation for populations of cells þ H p x ,t
2 ij @xi @xj ij l
having a diffusion model of interactions
0
constitutes one of the most attractive. with Hl = 2GlDGl . (4)
Equation 4 is also known as the Kolmogorov SDEs. The weak methods allow the computation
forward equation. It is assumed that the sample of any smooth functional on the state variables,
trajectories of the particle are continuous func- where the moments are particular cases. Several
tions of time. However, they are nowhere differ- methods have been proposed for weak integration
entiable with respect to time. The high of stochastic differential equations (Kloeden and
dimensionality of the Fokker–Planck equation Platen 1999). Continuous parametric models with
can be reduced
l by
using kthe
lmean-field
approxi- delays have been solved in the past for neuroim-
mation, pl x , t ¼ ∏k pl xk , t . The resulting aging data by transforming them into systems of
low-dimensional Fokker–Planck equations SDEs and algebraic equations. The following
describe the evolution of separable ensembles literature is recommended for details about this
that are coupled by mean-field effects. These particular type of state-space models (Deco
ensembles could represent different neuronal et al. 2008; Deco 2014; Valdes-Sosa
populations of a cortical minicolumn, e.g., layer et al. 2011). The dynamic causal modeling
II/III (or V) pyramidal cells (PCs), spiny stellate (DCM, Friston et al. 2003; Stephan et al. 2007;
cells, and GABAergic interneurons. For example, Daunizeau et al. 2011) constitutes the most
the mean firing rate of large basket GABAergic established type of continuous parametric models
interneurons could be used as a postsynaptic- with variants in almost of blue-highlighted levels
driven force to each individual layer II/III in Fig. 2. The DCM formalism has been exten-
PC. However, these ensembles could also repre- sively used to obtain connectionist diagrams (i.e.,
sent different types of cells (e.g., neurons and effective connectivity) between neuronal
astrocytes). For example, the mean firing rate of populations in different brain locations from neu-
a particular population of nitrergic GABAergic roimaging data. Different variants of the local
interneurons could be the driven vasodilative linearization (LL, Ozaki 2012) method have
force acting on individual smooth muscle cells been applied to EEG and fMRI data covering all
at the same cortical area. In other words, by yellow-highlighted levels in Fig. 2.
coupling the ensembles through the mean-field
approach the flow in the phase space
of one
Neuronal Mass Models
ensemble is f kl xlk ðtÞ, mlk0 ðtÞ, 8k0 6¼k , ul ðtÞ, Yl , The so-called neuronal mass models result from
with the means mlk0 ðtÞ ¼ E xlk0 ðtÞ . In the SDE particular approximations either from the mean-
representation, the mean-field approach is equiv- field models (no diffusion) or from the coupled
alent to a system of coupled SDEs. Still, the main mean-field models through the method of
problem with both approaches is to create effi- moments (non-curved). In general, for the neuro-
cient methods to integrate respective differential nal mass models, the full ensemble density is
equation systems, which have in principle a high replaced by a center of mass at a particular
dimensionality. Circumnavigating the problem of point. Then, a dynamic equation is used to
integrating high-dimensional Fokker–Planck describe the evolution of this center of mass
equation systems for the density in state-space with time. The neural-field models are an impor-
models defined by coupled SDEs is not a new tant generalization of the neural-mass models
topic. There are two main methods for solving by allowing for states to be functionals of
this problem. One of them is based on the position on the cortical sheet. This type of
assumption of Gaussianity for densities pl xl , t model cannot be defined as particular case of
(i.e., the Laplace assumption). Therefore, it general Eq. 1. A general discussion about these
focuses on obtaining ordinary differential equa- two types of models can be found in Deco
tions for the first two moments (i.e., means and et al. (2008). !
variances) of the state variables (Marreiros General mesoscopic states: PCD, J P ;
et al. 2008). However, the most popular method CMRO2, cMRO2 ; CMRGlc, cMRglc; dHb, q;
is based on the so-called weak integrators for CBF, f; CBV, v
Biophysical Models:
Neurovascular Coupling,
Cortical Microcircuits,
and Metabolism,
Fig. 2 Theoretical
frameworks for assembles B
of interacting cells
Examples of Relationships Between dephosphorylated Glc) have been employed to

Mesoscopic Variables describe the relative changes in cMRglc and v.
In general, these mesoscopic states can be In order to understand the origin of the rela-
represented by some of the abovementioned tionships between mesoscopic variables in the
coarse-grained versions of Eq. 1. Biophysical neocortex, three main biological mechanisms
generative models with balanced excitatory must be studied in detail: (a) the neurovascular
and inhibitory neuronal populations with spe- coupling, (b) the cortical microcircuits, and
cific oxygen-to-glucose indexes (OGI) have (c) the cellular metabolism.
been systematically introduced
!
to establish
relationships between J P and metabolic/ Neurovascular Coupling
hemodynamic variables. Some of these models Functional Hyperemia
!
link J P with instantaneous (ensemble average) The cerebral circulatory system is endowed with
postsynaptic membrane depolarizations and/or assorted vascular controlling mechanisms to
presynaptic firing rates of neurons. The facilitate the desirable cerebral perfusion in any
nonlinear balloon model was introduced to particular brain area at a given time. Surface
relate states q, f, and v, which was inspired by vessels in the brain are richly innervated at stra-
the Windkessel theory for the flow–volume tegic positions by both sympathetic and parasym-
relationship and the oxygen limitation pathetic nerves, known as the extrinsic
assumption. Oxygen transport models are com- innervations of the brain circulation. Specific
bined with mass-balance equations to estab- perivascular structures of an intrinsic origin
lish relationships between cMRO2 , f, and q, come into view after arterioles leave the
through some internal states such as the intra-/ Virchow–Robin space to penetrate the cerebral
extravascular oxygen contents and the blood cortex, the latter including subcortical innerva-
oxygen saturation fraction. In some models, tions. Spontaneous vasomotions originate from
neuronal demand of oxygen is defined by active mechanisms existing in cortical vessels,
complex dependencies in the oxygen extraction which to some extent are influenced by these
fraction (OEF). Compartmental models for plasma innervations. However, the perfusion of the
and tissue (with both phosphorylated and brain rapidly increases in areas where the activity
of neuronal networks is prolonged or enhanced. on the vasculature via a cyclic GMP (cGMP)-
In such a case, the CBF is locally intensified, dependent mechanism has been proposed to
a phenomenon denominated functional hyper- be modulatory rather than mediatory. (phasic
emia. It is hypothesized that such a blood flow signaling)
increase is accomplished by increased capillary 3. The release of prostaglandin E2 (PGE2) from
blood velocity rather than capillary recruitment. certain pyramidal cells, which have
From many years now, researchers have been dendritic and axonal profiles containing
trying to elucidate the basic principles for this inducible prostanoid-synthesizing enzyme
phenomenon (Riera and Sumiyoshi 2010). We cyclooxygenase-2 (COX-2). These specific
summarize below six major pathways by means glutamatergic neurons are frequently in close
of which neurons could communicate with the proximity to the penetrating microvessels.
surrounded vasculature. They can be classified Released PGE2 might reach a variety of recep-
into two major groups: (1) a phasic signaling tors on either nearby glial processes or vascu-
(e.g., gases, neurotransmitters, and lipids) issued lar smooth muscle cells (SMCs). COX-2 is
directly from the neurons may reach the micro- peculiarly restricted to presynaptic and post-
vasculature through either fuzzy diffusion or synaptic regions in these pyramidal cells,
local axonal release and b) a tonic signaling showing robust synaptic antilocalization with
(i.e., Ca2+ waves underlying long-term brain the nitric oxide-synthesizing enzyme. Nitric
requirements such as metabolites/oxygen) origi- oxide diffusion into the synaptic sites of
nated from gap-junction-connected astrocytes, these glutamatergic neurons may result in
which is evoked by the release of various neuro- COX-2 activation. Cortical microvessels are
transmitters from nearby neurons during also targets of axonal varicosities of such
activation. pyramidal cells. (phasic signaling)
1. The release of large vasoactive molecules 4. Enhancements of astrocytic Ca2+ signaling
[e.g., neuropeptide Y (NPY), vasointestinal cause the production of arachidonic acid
polypeptide (VIP), somatostatin (SOM), cho- through the action of phospholipase A2
lecystokinin (CCK), and acetylcholine (ACh)] enzymes on membrane phospholipids.
by GABAergic interneurons through direct Arachidonic acid is able to diffuse towards
innervations of neighboring (even distances adjacent cells, where it serves as a substrate
>100 mm away from the cells) microvessels. for two cytochrome P450 enzymes (i.e.,
Some nitrergic GABAergic interneurons with CYP4A subtype in SMCs and CYP2C subtype
specific laminar profiles in the neocortex in astrocytes) to produce 20-hydroxyeicosate-
transmute also neuronal signals from subcor- traenoic (20-HETE) and epoxyeicosatrienoic
tical regions to the vasculature of a given cor- (EET) acids. 20-HETE has a contractive effect
tical area. (phasic signaling) on the SMCs through the activation of the
2. The release of nitric oxide into the extracellu- L-type calcium channels and deactivation
lar space from the axon terminals of some of the calcium-activated potassium
GABAergic interneurons via activation of (BK) channels. EET diffuses into the tissue,
the nNOS. These interneurons could producing vasodilations by hyperpolarizing
release these large vasoactive molecules nearby SMCs through the activation of BK
perivascularly through vessel innervations. channels. The negative and positive effects
This particular population of interneurons is on the activation of BK channels in the
very small and is not homogeneously distrib- SMCs caused by 20-HETE and EET acids,
uted along the cortical layers. As nitrergic respectively, may be in equilibrium with low
glutamatergic neurons are very rare in the levels of nitric oxide, which constitutes an
neocortex, this mechanism might be very spe- inhibitory agent for CYP4A. This equilibrium
cific for inhibitory neuronal networks. The gives way whenever nitric oxide dissemina-
impact of such neuronal nitric oxide signaling tion in the tissue increases. Nitric oxide
signaling from neurons and endothelial cells is between pyramidal cells (e.g., “principal
activates the soluble guanylate cyclase in the cells,” DeFelipe and Fariñas 1992) and interneu-
SMCs, resulting in cGMP-related muscle rons. Interneurons are divided into excitatory
relaxations. (tonic signaling) (spiny stellate cells) and inhibitory types
5. Arachidonic acid increase causes PGE2 pro- (Markram et al. 2004). B
duction from the astrocytic constitutive Pyramidal neurons (PCs) comprise more
COX-1 enzyme. This lipid diffuses in the tis- than 70 % of all cortical neurons. They are the
sue and via PGE2 receptors could stimulate main output source of the neocortical microcir-
cyclic AMP formation in SMCs. The intracel- cuit, projecting to subcortical brain areas such as
lular lactate is used by the astrocytic prosta- the thalamus, basal ganglia, and brainstem but
glandin transporter to take up extracellular also connecting cortical regions, e.g., the contra-
PGE2. However, when extracellular lactate and ipsilateral sections of a sensory area.
levels are increased, this PGE2 clearance Although they appear in a variety of forms, the
mechanism is lost, causing augmented vasodi- following common features are typically
lations. The extracellular lactate rises by reported: (a) an apical dendrite branch which
enhancing the activity of the lactate dehydro- extends from the soma vertically towards the
genase enzyme, for example, increasing the cortical surface with a gradually tapering form
substrate levels (pyruvate and NADH) of this and (b) basal dendrites which originate from the
redox reaction. Indeed, the levels of these sub- soma towards random transversal directions. PC
strates increase under either hyperglycolytic dendrites have highest spine density in the middle
activity or suppression of the tricarboxylic portions of the apical dendrite. Their axons pro-
acid (TCA) cycle. Astrocytic lactate ject vertically downwards into the white matter
transported to the extracellular space by after sending local collaterals within the cortical
monocarboxylate transporters MCT-1 pro- layers. The apical dendrites of layer VI PC pro-
vides a signaling that bring together metabo- ject also to non-vertical angles before climbing
lism and functional hyperemia through the up towards the superficial layers. Some PCs gen-
lactate–PGE2 coupling mechanism. (tonic erally form a dendritic tuft in layer I. Local axon
signaling) collaterals of PCs innervate neighboring cortical
6. External K+ promotes dilation of vessels neurons. Layer V thick-tufted PC constitutes one
through Kir2.1 channels in SMCs. Large- of the best known types of cortical neurons. Their
conductance BK channels have been found in apical dendrite always reaches layer I where it
astrocytic end feet, which might open in bifurcates extensively forming a wide tuft. These
response to significant Ca2+ increases. neurons have broad basal dendrites that span
A variant of this hypothesis, named the K+ radially up to 300 mm around the soma. The
siphoning, was initially proposed as local axonal collaterals of this neuron project
a mechanism for controlling the vascular vertically to all layers and horizontally within
tone through astrocytic Kir4.1, but recently layers V and VI. They project to deeper brain
criticized. (tonic signaling) areas (e.g., superior colliculus and brainstem
The spatiotemporal profiles of these mecha- nuclei). These neurons are considered the micro-
nisms show strong laminar variations (Herman circuit integrators of inputs from neurons in all
et al. 2013; Hirano et al. 2011) as a function of cortical layers, which eventually result in the
the brain state (Maandag et al. 2007; Smith main cortical outputs. These typical cytoarchi-
et al. 2002). tectonic structures for principal cells in occipital
and parietal – and, to a lesser extent,
Cortical Microcircuits temporal – lobes are not conserved in prefrontal
Cortical neurons display a wide variation in their cortices.
morphological, electrophysiological, synaptic, Interneurons consist of ~20 % of neocortical
and molecular properties. The clearest distinction neurons. Interneurons lack of apical dendrites.
Their axons are mostly confined to the same to infragranular layers. They display
functional column. Usually, interneurons do not a regular spacing into microcolumns.
project to other cortical areas via the white mat- • Martinotti cells are interneurons with bitufted
ter, i.e., they are “local circuit neurons.” In terms dendritic morphology and an ovoid soma.
of their dendritic spine density, interneurons are They are primarily found in cortical layers
divided into smooth, sparsely spiny, and spiny II–VI. Their axons ascend to and spread in
cells. These neurons have a very diverse axonal layer I up to a few millimeters laterally from
ramification, with characterization established the somatic location. They inhibit the tuft and
according to the postsynaptic domain (axon, distal dendrites of PCs.
soma, or dendrite) innervated by them. Spiny • Neurogliaform cells are small interneurons
stellate cells are the only excitatory interneurons existing in all cortical layers. They have fine
and they are mostly located in layer IV of primary and multipolar-spanning dendrites. Their
sensory cortices. They constitute the major input axons are thin and extensively branched.
to the cortex from subthalamic nuclei. They dis- They innervate PCs with GABA type
play radially projecting spiny dendrites origi- B synapses.
nated from small somata, some of them
spreading to neighboring layers. Inhibitory inter- Main Microcircuit Principles
neurons have a large diversity of morphological Cortical neurons display a large variety of geo-
and electrophysiological properties, which help metrical and intrinsic electrophysiological prop-
interneuron classification. According to their erties (DeFelipe et al. 2002). These properties
arborizations, inhibitory interneurons are determine the input–output relationships for
subclassified into bipolar, bitufted, and multipo- each particular neuron. They depend on the
lar neurons. expression of ionic channels along the dendrites
• Basket cells (BC) are the most common inhib- as well as on the geometry of dendritic trees.
itory interneurons in the neocortex. They are Cortical neurons are engaged in different time
found in cortical layers II–VI. There are three scales. The single-cell mRT-PCR method allows
main subclasses of BC (i.e., large, small, and the establishment of relationships between
net). Their dendrites extend to different direc- ion-channel expression patterns and electrophys-
tions, with the morphology being multipolar iological profiles of cortical neurons. A widely
and dendrites either smooth or sparsely used classification scheme considers neurons as
spined. (a) regular spiking (RS), (b) fast spiking (FS), and
• Chandelier (axo-axonic) cells are found (c) intrinsic bursting (IB), which rely only on the
throughout layer II to V in different cortical response to long suprathreshold stimuli. With the
areas of mammalians. They have an ovoid-like addition of the “chattering cell” type, this classi-
soma, from which smooth (or sparsely spined) fication scheme has been used to quantitatively
dendrites develop in opposite directions. Their evaluate electrophysiological behaviors in vivo.
dendrites can be found in cortical layers I–VI, Two other electrophysiological types of neurons
and their axons form a candle-like arrange- have been later added to the original scheme:
ment of vertical bouton strings that target PC (a) low-threshold spiking (LTS) and (b) late spik-
axon initial segments, primarily in the ing (LS). The majority of synaptic connections in
supragranular layer. This type of neurons the neocortex are between PCs, most of them in
plays a crucial role in controlling PC activity. layer V (formed on basal, apical, oblique, and tuft
• Double bouquet cells (DBC) are interneurons dendrites). Short-term connections between
present in superficial layers with vertically spiny stellate neurons in layer IV are formed on
oriented ovoid-like somata. Therefore, they dendrites. Interlaminar connections between
have two dendritic bundles that extend from excitatory neurons target also dendrites. The spa-
two opposite directions. Their axons form tial properties of excitatory and inhibitory synap-
tight axonal bundles that descend vertically tic inputs determine the efficacy at which
individual neurons are excited. Synapses are dis- selectively decided which PCs are going to be
tributed at specific distances from the soma (i.e., their final target, for which multiple boutons
Euclidean and electrotonic distances) and with will be created around the axon initial segment.
targeted locations relative to dendritic bifurca- This selection process cannot be explained by
tions and branches. Ionic channel densities random processes and geometrical consider- B
along dendritic trees, soma, and axon, as well as ations, which is far from the characteristic con-
their reversal potential distributions, play an nectivity patterns between pyramidal neurons.
important role in the genesis/propagation of
action potentials. For example, as Metabolism
a consequence of the higher values of chloride Vital functions in neurons and astrocytes
reversal potential at the postsynaptic location requiring cellular work are as follows:
(axon initial segment), chandelier cells have (a) the glutamate/GABA–glutamine cycle;
excitatory effects on PCs. Due to the physical (b) vesicularization, docking, and exocytosis of
proximity, axon- and soma-targeting synapses neurotransmitters from vesicles; (c) the trans-
can immediately affect the action-potential initi- membrane ATPases in both types of cells;
ation; hence, they act as action-potential “edi- (d) actin-filament growth in dendrites;
tors.” Distal dendritic synapses have low (e) phospholipid and glycolipid metabolism; and
temporal precision because signals produced by (f) axoplasmic transport. The isolated nerve-
them are filtered at the soma and initiation zone. ending particle (i.e., synaptosome) preparation
Dendritic spines are the preferable postsynaptic has been used successfully to study many princi-
targets of most excitatory synaptic connections. ples of the metabolic pathways in the brain
They play a crucial role in dendritic integration (Riera et al. 2008). It has been recently demon-
and synaptic plasticity. Although dendritic arbors strated that in spite of the complexity, metabo-
are physically stable, dendritic spines and lism shows certain level of homogeneities across
filopodia are variable elements (i.e., elimination cortical networks of the human brain (Hyder
and growth). These spines are the means by et al. 2013a).
which dynamically postsynaptic target cells con- Abbreviations: Glc, glucose; GLUT, glucose
nect selectively to specific presynaptic boutons. transporter; HK, hexokinase; PGI, phosphoglucose
The selectivity of synaptic contacts between cor- isomerase; F6P, b-Dfructose 6-phosphate; G6P,
tical neurons can be divided into two separate a-D-glucose 6-phosphate; PFK-1, phosphofructo-
questions: (i) selectivity in the postsynaptic ele- kinase-1; F1,6BP, b-Dfructose 1,6-bisphosphate;
ment being connected to and (ii) selectivity in the GADP, D-glyceraldehyde 3-phosphate; PGK,
unique identity of the innervated neuron. phosphoglycerate kinase; PEP, phosphoenolpyr-
(i) Domain selectivity uvate; PK, pyruvate kinase; MCT,
(a) Axon-targeting cells (e.g., chandelier cells) monocarboxylate transporter; LDH, lactate dehy-
(b) Somatic and perisomatic targeting inter- drogenase; PDH, pyruvate dehydrogenase; CS,
neurons (e.g., basket cells) citrate synthase; a-KG, a-Ketoglutarate; SDH, suc-
(c) Dendritic targeting cells (e.g., double cinate dehydrogenase; OAA, oxaloacetate; GDH,
bouquet, neurogliaform, and Martinotti glutamate dehydrogenase; GLT, glutamate trans-
cells) porter; AAT, aspartate aminotransferase; PAG,
(ii) Target cell selectivity phosphate-activated glutaminase; GAD, glutamate
How does a neuron choose its postsynaptic decarboxylase .
partners? The selection of a target pyramidal
neuron by chandelier cells may occur because Postsynaptic and Action Potentials
of a random axodendritic proximity with Neurotransmitters (e.g., Glu, GABA) released
a further proliferation as the result of local inter- into the synaptic cleft in the course of presynaptic
actions. While both basket and chandelier cells neuronal activity will freely diffuse towards neu-
develop connections to neighboring cells, they ronal postsynaptic buttons and nearby astrocytic
processes. They will cause either excitatory produce Pyr and ATP. There are no transmem-
(EPSP) or inhibitory (IPSP) postsynaptic poten- brane transporters for G6P; thus, Glc cannot leak
tials by receptor-specific flows of ions in the out of the cells after it is phosphorylated by HK.
postsynaptic neurons. After the genesis of
EPSP/IPSP, ionic gradients are reestablished by The Glycogen-Shunt Proposition
way of transmembrane ATPases, which consume G6P could have two fates, either being converted
large amounts of ATP. Action potentials are gen- into F6P by PGI, and thus continuing along the
erated as a consequence of the spatiotemporal glycolytic pathway, or otherwise being stored as
integration of these postsynaptic potentials, and glycogen polymer. Glycogen polymers are con-
they propagate along nerves. The genesis and sidered as one of the largest energy reservoirs in
propagation of these action potentials cause also the brain and could be used as an emergency fuel
the breakdown of the ionic balances across cellu- supply during severe energetic stress resulting
lar membranes and, therefore, require the use of not only from a pathological condition (e.g.,
ATP. In the postsynaptic buttons and axons, the hypoglycemia, cerebral ischemia) but also from
required ATP is produced from Pyr through the high neuronal activity in a healthy brain. Glyco-
TCA cycle and electron-transport/oxidative- gen polymer breakdown preferentially occurs in
phosphorylation pathways. the astrocytic compartment, but its activation
would plausibly seem to be linked to the neuronal
The Glycolytic Pathway compartment. The OGI may be an important link
The glycolytic pathway involves breakdown of mediating the balance between the glycolytic
Glc into Pyr. Glc is continuously delivered from pathway and glycogen shunt.
capillaries to the extracellular milieu through
GLUT-1. Cells take up extracellular Glc through The Astrocyte–Neuron Lactate Shuttle
GLUT-1 (astrocytes) and GLUT-3 (neurons). In A fraction of around 15 % of the Glc oxidized
different proportions, Pyr and ATP are produced in neurons is not coupled to the total
from glycolysis in both cell types. Breakdown glutamate–glutamine cycle (Hyder et al. 2006).
steps at different levels can be either impeded or This fraction might represent ATP needed
hastened through the activity of many enzymes in for non-cycling activities. Based on these obser-
this pathway, which constitute checkpoints vations, it has been suggested that the
located in different cellular compartments. The glutamate–glutamine cycle, a two ATP-
preparatory phase comprises four steps: the first consuming dual task (i.e., one ATP is used to
phosphorylation of Glc by enzyme HK (control convert glutamate into glutamine by glutamine
point G-1), an isomerase reaction of G6P by synthetase and the other to reestablish the Na+
enzyme PGI, a phosphorylation of F6P by gradient by the activation of the Na+/K+-
enzyme PFK-1 (control point G-2), and an aldol ATPase), might trigger lactate production by gly-
reaction on F1,6BP by enzyme ALDO that finally colysis in the astrocytes during the course of
produces DHAP and GADP. DHAP and GADP neuronal activity, so stimulating them to take up
are rapidly and reversibly interconverted by TPI, extracellular Glc (Pellerin and Magistretti 2004).
a step essential to produce energy efficiently. The Lactate so produced may be released for oxida-
payoff phase comprises five steps: a redox reaction in adjacent neurons. It is not at all clear
tion on GADP by enzyme GAPDH to form where the oxidation of lactate derived from Glc
1,3BPG, a product with high phosphoryl-transfer and/or glycogen metabolism may occur (Hertz
potential; a substrate-level phosphorylation of et al. 2007). Moreover, during activation,
1,3BPG by enzyme PGK (it is the break-even glutamatergic activity in neurons seems better
point in glycolysis); a mutase reaction on 3PG preserved by Glc metabolism than by lactate
by enzyme PGAM, a hydration of 2PG by metabolism. The observed increase of lactate
enzyme ENO; and a substrate-level phosphoryla- under steady-state sensory stimulation has been
tion of PEP by enzyme PK (control point G-3) to explained on the basis of stimulation evoked
glycolysis and glycogen-shunt pathways TCA cycle by either oxidative deamination cata-
(Shulman et al. 2001a, b). Lactate may be taken lyzed by GDH or transamination via the aspartate
up into neurons through the monocarboxylate aminotransferase (AAT), an enzymatic reaction
transporter MCT-2, which has the highest affinity producing aspartate from OAA.
for its substrate. An extensive overproduction of B
lactate must efflux into the circulating blood. In The Glutamate/GABA–Glutamine Cycle
summary, the astrocyte–neuron lactate shuttle The glutamate (Glu) released by a glutamatergic
may be facilitated by the differential presence of presynaptic terminal is mainly taken up into astro-
LDH-1 and LDH-1&5 in neurons and astrocytes, cytes, although a small portion could flood back
respectively. This shuttle may be favored by into the terminal through the transporter GLT-1b.
the existence of monocarboxylate transporters In the GABAergic synapse, the released GABA is
MCT-2 and MCT-1&4 in neurons and astrocytes, taken up into both the presynaptic terminal and the
respectively. astrocytes, with the former having higher affinity
for the neurotransmitter. Inside the astrocyte,
TCA Cycle and Oxidative Phosphorylation GABA is catabolized to succinate before entering
The majority of ATPs yielded from Glc oxidation the TCA cycle. The carbon skeleton of this amino
are generated by breakdown of carbon skeletons acid exits the TCA cycle as a-KG, which is then
in the TCA cycle and electron donors located transformed to glutamate. The Glu is amidated to
inside the mitochondria. The biochemical mech- glutamine (Gln) by glutamine synthetase, an enzy-
anisms underlying these energy substrates as well matic reaction that benefits adjacent tissues by
as major enzymatic checkpoints inside the consuming and detoxifying free ammonia. The
TCA cycle and electron-transport/oxidative- synthesized glutamine returns to the presynaptic
phosphorylation pathway are detailed below. terminals of both glutamatergic and GABAergic
The PDH complex performs the decarboxylation neurons, with a preference for the former. In both
of Pyr by enzymes pyruvate dehydrogenase (E1), terminals, glutamine is used to regenerate Glu and
dihydrolipoyl transacetylase (E2), and ammonia via phosphate-activated glutaminase
dihydrolipoyl dehydrogenase (E3). In order to (PAG). The regenerated Glu in the GABAergic
sense the cellular energy charge, two enzymes presynaptic terminal is converted back into GABA
(i.e., the PDH kinase and the PDH phosphatase) via Glu decarboxylase (GAD).
are endowed with a variety of allosteric modula-
tors and covalent modifiers. The TCA cycle com- Oxygen Transport
prises eight steps: a condensation of acetyl-CoA Since oxygen, unlike Glc, does not have specific
by enzyme CS, an isomerization of citrate by transporters to enter the extravascular space, oxy-
enzyme (aconitase), an oxidative decarboxyl- gen transport from blood to tissue is mainly
ation of isocitrate by enzyme IDH, an oxidative driven by diffusion. Diffusion-limited models
decarboxylation of a-KG by the multienzyme suggest that blood flow is the only way by
complex a-KGDH, a substrate-level phosphory- which tissue oxygen capacity can be increased
lation of succinyl-CoA by enzyme SCS, (Buxton and Frank 1997; Gjedde 1997). The pre-
a dehydrogenation of succinate by enzyme diction of such diffusion-limited oxygen trans-
SDHA (SDHA is tightly bound to the mitochon- port models is nonlinear coupling between
drion inner membrane through the protein sub- blood flow and oxygen metabolism. Based on
units SDHB, SDHC, and SDHD, which all observed linear relationships between blood
constitute the complex II of the electron-transport flow and oxygen metabolism (across regions,
chain), a hydration of fumarate by enzyme FH, conditions, and/or species), other empirically
and a dehydrogenation of malate by enzyme derived models suggest facilitated diffusion
MDH to produce OAA. Energy in the form of through processes like altering number of per-
ATP and NADH is produced at several points fused capillaries and/or intracapillary stacking
along the entire TCA cycle. Glu can enter the of erythrocytes (Hyder et al. 1998, 2000).
The general assumption in all of these models data: on the infeasibility of the neuronal current-source
that all of the oxygen leaving the capillary is monopoles. J Neurophysiol 109:1681–1682
Grieb P, Forster RE, Strome D, Goodwin CW, Pape PC
metabolized and that there is no oxygen back (1985) O2 exchange between blood and brain tissues
flux into the blood has been experimentally con- studied with 18O2 indicator-dilution technique. J Appl
firmed (Grieb et al. 1985; Kassissia et al. 1995; Physiol 58:1929–1941
Herman et al. 2006). Herman P, Tr€ ubel HK, Hyder F (2006) A multi-parametric
assessment of oxygen efflux from the brain. J Cereb
Blood Flow Metab 26:79–91
Major Controversies Herman P, Sanganahalli BG, Blumenfeld H, Rothman DL,
– What does underlie the brain functional hyper- Hyder F (2013) Quantitative basis for neuroimaging of
emia: subthreshold synaptic activity or spik- cortical laminae with calibrated fMRI. Proc Natl Acad
Sci USA 110:15115–15120
ing (Logothetis et al. 2001; Mukamel Hertz L, Peng L, Dienel GA (2007) Energy metabolism in
et al. 2005)? astrocytes: high rate of oxidative metabolism and spa-
– Are the EEG brain sources purely dipolar in tiotemporal dependence on glycolysis/glycogenolysis.
nature (Riera et al. 2012; Gratiy et al. 2013)? J Cereb Blood Flow Metab 27:219–249
Hirano Y, Stefanovic B, Silva AC (2011) Spatiotemporal
– Does oxidatively derived ATP contribute evolution of the functional magnetic resonance imag-
equally to postsynaptic and action potentials ing response to ultrashort stimuli. J Neurosci
(Attwell and Laughlin 2001) across a range of 31:1440–1447
activity levels and/or species (Hyder Hyder F, Shulman RG, Rothman DL (1998) A model for
the regulation of cerebral oxygen delivery. J Appl
et al. 2013b)? Physiol 85:554–564
Hyder F, Kennan RP, Kida I, Mason GF, Behar KL,
Rothman DL (2000) Dependence of oxygen delivery
References on blood flow in rat brain: a 7 tesla nuclear magnetic
resonance study. J Cereb Blood Flow Metab
Attwell D, Laughlin SB (2001) An energy budget for 20:485–498
signaling in the grey matter of the brain. J Cereb Hyder F, Patel AB, Gjedde A, Rothman DL, Behar KL,
Blood Flow Metab 21(10):1133–1145 Shulman RG (2006) Neuronal-glial glucose oxidation
Buxton RB, Frank LR (1997) A model of the coupling and glutamatergic-GABAergic function. J Cereb
between cerebral blood flow and oxygen metabolism Blood Flow Metab 26:865–877
during neural stimulation. J Cereb Blood Flow Metab Hyder F, Fulbright RK, Shulman RG, Rothman DL
17:64–72 (2013a) Glutamatergic function in the resting awake
Daunizeau J, David O, Stephan KE (2011) Dynamic human brain is supported by uniformly high oxidative
causal modelling: a critical review of the biophysical energy. J Cereb Blood Flow Metab 33:339–347
and statistical foundations. Neuroimage 58:312–322 Hyder F, Rothman DL, Bennett MR (2013b) Cortical
Deco G, Jirsa VK, Robinson PA, Breakspear M, Friston K energy demands of signaling and nonsignaling com-
(2008) The dynamic brain: From spiking neurons to ponents in brain are conserved across mammalian spe-
neural masses and cortical fields. PLoS Comput Biol 4 cies and activity levels. Proc Natl Acad Sci
(8): e1000092 USA 110(9):3549–3554
Deco G (2014) Multi-Scale brain connectivity. Encyclo- Kassissia IG, Goresky CA, Rose CP, Schwab AJ,
pedia of computational neuroscience, Springer Simard A, Huet PM, Bach GG (1995) Tracer oxygen
DeFelipe J, Fariñas I (1992) The pyramidal neuron of the distribution is barrier-limited in the cerebral microcir-
cerebral cortex: morphological and chemical charac- culation. Circ Res 77:1201–11
teristics of the synaptic inputs. Prog Neurobiol Kloeden PE, Platen E (1999) Numerical solution of sto-
39(6):563–607 chastic differential equations, 3rd edn. Springer, Berlin
DeFelipe J, Alonso-Nanclares L, Arellano JI Logothetis NK, Pauls J, Augath M, Trinath T, Oeltermann
(2002) Microstructure of the neocortex: comparative A (2001) Neurophysiological investigation of the basis
aspects. J Neurocytol 31:299–316 of the fMRI signal. Nature 412:150–157
Friston KJ, Harrison L, Penny W (2003) Dynamic causal Maandag NJ, Coman D, Sanganahalli BG, Herman P,
modelling. Neuroimage 19:1273–1302 Smith AJ, Blumenfeld H, Shulman RG, Hyder
Gjedde A (1997) The relation between brain function and F (2007) Energetics of neuronal signaling and fMRI
cerebral blood flow and metabolism. In: Batjer HH activity. Proc Natl Acad Sci USA 104:20546–20551
(ed) Cerebrovascular disease. Lippincott-Raven, Phil- Markram H, Toledo-Rodriguez M, Wang Y, Gupta A,
adelphia, pp 23–40 Silberberg G, Wu C (2004) Interneurons of the neo-
Gratiy SL, Pettersen KH, Einevoll GT, Dale AM cortical inhibitory system. Nat Rev Neurosci
(2013) Pitfalls in the interpretation of multielectrode 5:793–807
Biophysics of Adaptation in a Computational Model of the Leech T Neuron 411 B
Marreiros AC, Kiebel S, Daunizeau J, Harrison L, Friston
KJ (2008) Population dynamics under the Laplace Biophysics of Adaptation in
assumption. Neuroimage 44:701–714
Mukamel R, Gelbard H, Arieli A, Hasson U, Fried I, a Computational Model of the
Malach R (2005) Coupling between neuronal firing, Leech T Neuron
field potentials, and fMRI in human auditory cortex. B
Science 309:951–954 Raúl Guantes1, Sara Arganda2 and
Ozaki T (2012) Time series modeling of neuroscience
data. Chapman & Hall/CRC Interdisciplinary Statis- Gonzalo G. de Polavieja3
1
tics, London Department of Condensed Matter Physics,
Ozaki T (2014) Statistical analysis of neuroimaging data. Materials Science Institute ‘Nicolás Cabrera’ and
Encyclopedia of computational neuroscience, Institute of Condensed Matter Physics
Springer
Pellerin L, Magistretti PJ (2004) Neuroenergetics: calling (IFIMAC), Universidad Autónoma de Madrid;
upon astrocytes to satisfy hungry neurons. Neurosci- Cantoblanco, Madrid, Spain
2
entist 10(1):53–62 Centre de Recherches sur la Cognition Animale,
Riera J, Aubert E, Iwata K, Kawashima R, Wan X, Ozaki Université de Toulouse, Toulouse, France
T (2005) Fusing EEG and fMRI based on a bottom-up 3
model: inferring activation and effective connectivity Instituto Cajal, Consejo Superior de
in neural masses. Phil Trans R Soc B 360, 1025–1041 Investigaciones Cientı́ficas, Madrid, Spain
Riera J, Sumiyoshi A (2010) Brain oscillations: ideal
scenery to understand the neurovascular coupling.
Curr Opin Neurol 23(4):374–381
Riera J (2014) Brain Imaging: Overview. Encyclopedia of Definition
Computational Neuroscience, Springer
Riera J, Schousboe A, Waagepetersen HS, Howarth C, Adaptation is a change over time in the respon-
Hyder F (2008) The micro-architecture of the cerebral siveness to stimuli. In adaptation to stimulus sta-
cortex: functional neuroimaging models and metabo-
lism. Neuroimage 40:1436–1459 tistics, this change is relative to statistical
Riera J, Ogawa T, Goto T, Sumiyoshi A, Nonaka H, properties of the stimulus like the standard devi-
Evans A, Miyakawa H, Kawashima R (2012) Pitfalls ation. Biophysical models of this phenomenon
in the dipolar model for the neocortical EEG sources. are computational models of the components of
J Neurophysiol 108(4):956–975
Shulman RG, Hyder F, Rothman DL (2001a) Lactate a neuron or a circuit to show the mechanistic
efflux and the neuroenergetic basis of brain function. basis of adaptation.
NMR Biomed 14:389–396
Shulman RG, Hyder F, Rothman DL (2001b) Cerebral
energetics and the glycogen shunt: neurochemical Detailed Description
basis of functional imaging. Proc Natl Acad Sci
U S A 98:6417–6422 Computational models are instrumental to inves-
Smith AJ, Blumenfeld H, Behar KL, Rothman DL,
tigate and understand the mechanisms that shape
Shulman RG, Hyder F (2002) Cerebral energetics
and spiking frequency: the neurophysiological the spike pattern of a neuron’s response given its
basis of fMRI. Proc Natl Acad Sci U S A synaptic inputs. They can complement experi-
99:10765–10770 mental approaches by providing insights into
Stephan KE, Roebroeck A (2012) A short history of causal
possible mechanisms. One important feature of
modeling of fMRI data. Neuroimage 62:856–863
Stephan KE, Mattout J, David O, Friston KJ (2006) Models many sensory systems is that they can adapt their
of functional neuroimaging data. Curr Med Imaging response to the levels and context of the stimuli.
Rev 2(1):15–34 Adaptation of the neuronal response is a process
Stephan KE, Harrison LM, Kiebel SJ, David O,
by which the responsiveness of the neuron
Penny WD, Friston KJ (2007) Dynamic causal models
of neural system dynamics: current state and future changes as a result of previous stimulation. In
extensions. J Biosci 32:129–144 the present case, we had experimentally studied
Toga AW, Mazziotta JC (2002) Brain mapping: the adaptation mechanisms using a mechanosensory
methods, 2nd edn. Academic, San Diego
neuron of the leech Hirudo medicinalis known as
Valdes-Sosa PA, Roebroeck A, Daunizeau J, Friston K
(2011) Effective connectivity: influence, causality and the T neuron and built a computational model to
biophysical modeling. Neuroimage 58:339–361 further test some of the proposed mechanisms of
B 412 Biophysics of Adaptation in a Computational Model of the Leech T Neuron
a
Soma Dendrite b
100 mV
gL gNa gCa gK gK,Ca 10 s

c
d
C Na pump
EL ENa ECa EK
Biophysics of Adaptation in a Computational Model pulses followed by an epoch of high-amplitude pulses

of the Leech T Neuron, Fig. 1 (a) Two-compartment before returning to the initial, low-amplitude pulses
model of the T neuron including known conductances and (panel d, in red). (c) Model response after removing the
the Na+ pump. (b) Model response during a stimulation Na+ pump
protocol consisting of a low-amplitude periodic train of
adaptation to stimulus statistics (Arganda neuron’s geometry are more realistic (at least
et al. 2007). Here we review how to use when we know all parameters in some detail),
a simple two-compartment model to get insight the large number of variables involved makes it
into realistic mechanisms of bursting and difficult to isolate and understand the basic mech-
adaptation. anisms responsible for a given response feature.
An important function of adaptation in the An intermediate approach is to consider
T neuron is to mediate adaptive scaling of neuro- a biophysical model including the known exper-
nal responses to stimulus statistics. Sensory sys- imental conductances of the neuron and look for
tems have to deal with stimuli that fluctuate in the minimal ingredients necessary to reproduce
time and even change their scale. A sensory neu- a given behavior.
ron has a limited range of outputs, insufficient to Following this idea, we simplified a
represent all the stimulus values present in detailed multi-compartment model of the
a natural context, which can span several orders mechanosensitive T neuron (Cataldo et al. 2005)
of magnitude. To efficiently encode naturalistic and considered only two compartments (soma/
stimuli, a neuron then needs to adapt its response axon and dendrite) represented as equivalent
to the stimulus variance, adjusting its output electrical circuits and including all known con-
range to the range of inputs (Wark et al. 2007). ductances of the T cell (Fig. 1a). Each compart-
This adaptive scaling to the variance of the stim- ment consists of a membrane capacitance
ulus distribution enhances information transmis- C = 1 mF/cm2 in parallel with the conductances
sion (Brenner et al. 2000; Sharpee et al. 2006). associated to two inward currents (a fast Na+
Very simplified models, such as “integrate and current, INa, and a high-threshold Ca2+ current,
fire,” are useful for neurons whose spiking output ICa); an outward persistent K+ current, IK; a leak
directly reflects the amplitude of the input. More current, IL; and two outward currents solely
complex responses as the ones discussed here regulated by intracellular Na+ and Ca2+ pools,
(bursting and adaptation) require the explicit con- a Ca2+ -activated K+ current, IK,Ca, and the
sideration of specific voltage-gated conductances Na+ pump, Ipump.
and different biophysical mechanisms. While The voltage-dependent currents INa, IK, and
detailed multi-compartment models including ICa are described by Hodgkin-Huxley-type equa-
all known conductances of the cell and the tions. The properties of the Na+ and K+ currents
were adjusted to reproduce the amplitude, dura- Biophysics of Adaptation in a Computational Model
tion, and shape of the action potential of the of the Leech T Neuron, Table 1 Nernst potentials and
maximum conductances of the ionic currents used in the
T cell, as well as the resting membrane potential model
and membrane time constant of the neuron. The
Currents INa IK ICa IL IK,Ca Ipump
conductance of the leakage channel, IL, was cho-
VNernst(mV) 45 62 60 48 62 200
B
sen to obtain values of input resistance in the
g (mS/cm2) 350 90 0.5 0.1 3 0.5
range of the experimental values. These currents
are given by
hand, the Ca2+-dependent K+ current IK,Ca is fast

I iNa ¼ gNa m3Na hðV i V Na Þ
(Wang 1998; Cataldo et al. 2005) and, in combi-
nation with the high-threshold, non-inactivating
I iK ¼ gK n2 ðV i V K Þ
Ca2+ current ICa, is responsible for the bursting
mechanism (Wang 1998; Kepecs et al. 2002).
I iCa ¼ gCa mCa ðV i V Ca Þ The role of the persistent inward current ICa is
to produce a positive feedback on membrane
I iL ¼ gL ðV i V L Þ, potential, thus driving a high firing rate, whereas
the activity-dependent current IK,Ca generates
where the index i = s, d stands for soma or a slower negative feedback that terminates firing,
dendrite. Conductances (in mS/cm2) and thus producing a burst of action potentials. There-
Nernst potentials (in mV) are the same for both fore, the IK,Ca current mainly determines burst
compartments and are given in Table 1. The duration, and its parameters can be tuned to pro-
activating and inactivating gating variables m, duce the burst duration distribution observed
n, and h evolve according to first-order kinetics, experimentally.
dmðV i Þ
dt ¼ ½m1 ðV i Þ mðV i Þ=tm ðV i Þ, where the On the other hand, the Na+ pump acts on
asymptotic values and time constants a much slower time scale and is responsible for
(in milliseconds) are taken from (Cataldo the long refractory times observed after sustained
et al. 2005) activity (Cataldo et al. 2005). As spiking activity
is maintained, the concentration of Na+ ions
1 inside the cell increases, elevating the activity
m1,Na ðV i Þ ¼ tm,Na ¼ 0:1 of the slow Na+ outward current, Ipump. This
1þ eðV i þ35Þ=5
generates a delayed negative feedback on the
1 membrane potential, hyperpolarizing the mem-
h1 ð V i Þ ¼ 2 brane in an activity-dependent manner. When
½1 þ eðV i þ50Þ=9 firing activity stops, the action of the pump
13:8 decreases, allowing a slow recovery of the mem-
th ð V i Þ ¼ þ 0:2
1 þ eðV i þ36Þ=3:5 brane potential.
The two currents modulated by the intracellu-
1
n1 ð V i Þ ¼ lar Ca2+ and Na+ pools are given by
1þ eðV i þ22Þ=9
5
tn ðV i Þ ¼ þ1 I iK,Ca ¼ gK,Ca kCa ðV i V K Þ
1 þ eðV i þ10Þ=10

1 I ipump ¼ gpump kNa V i V pump
m1,Ca ðV i Þ ¼ tm,Ca ¼ 0:6
1 þ eðV i þ10Þ=2:8
where the dynamics of the ion-dependent con-
The two outward currents IK,Ca and Ipump are ductances kCa and kNa and of the ion concentra-
modulated by activity and are therefore responsi- tions is given by first-order kinetics (Yamada
ble for different adaptation processes. On the one et al. 1998):
dkCa ½Ca2þ kCa With these currents, the membrane potential

¼ of both compartments changes in time according
dt tk,Ca
to the equations:
d½Ca2þ ½Ca2þ aCa I Ca
¼
dt ti,Ca
dV s
C ¼ I L þ I Na þ I K þ I Ca þ I sK,Ca þ I spump
þ dt
dkNa ½Na kNa
¼ þ gc ðV s V d Þ=p
dt tk,Na
d½Naþ ½Ca2þ aNa I Na
¼ dV d
dt ti,Na C ¼ I L þ I Na þ I K þ I Ca þ I dK,Ca þ I dpump
dt
Note that the terms aCaICa, aNaINa represent g ðV d V s Þ
þ c I stim
the increase in intracellular ionic concentrations 1p
due to influx of ions through specific channels
(ICa and INa are negative) and that ion removal is where the indices s and d stand for soma and
exponential with decay constants ti,Ca, ti,Na. dendrite, respectively. The electrotonic coupling
Strictly, removal of Na+ ions should depend on between compartments, given by the parameter
the pump activity, Ipump. However, this contribu- gc, and the ratio of soma to total membrane area,
tion is very small compared to the influx of p, also modulate bursting dynamics (Kepecs
sodium due to the fast current, INa (because et al. 2002; Izhikevich 2007). Within our simpli-
Ipump is on the order of few pA, while INa spans fied two-compartment model, we choose these
tenths of nA), and sodium removal can be parameters to mimic experimental burst sizes
modeled using an exponential decay. In the when stimulated by a random stimulus with
model, we do not consider diffusion of ions Gaussian distribution and set gc = 1.5 mS cm2
between compartments. and p = 0.7.
Time constants for removal of Ca2+ are ti, We inject in the dendritic compartment an
Ca = 1,250 ms for soma and 1,000 ms for den- electrical current Istim with the same features
drite. The time constants for Na+ removal are than the mechanical stimulation delivered to
adjusted to reproduce the delay in response after real T cells: a slowly varying Gaussian white
prolonged periods of activity observed in real noise with a frequency cutoff and different stan-
T neurons and are set to ti,Na = 16,000 ms for dard deviations (Arganda et al. 2007). Note, how-
soma and 20,000 ms for dendrite. Note that ever, that the mechano-electrical transduction
these time constants are one order of magnitude processes are unknown and the amplitude and
larger than those for calcium to take into cutoff frequencies of the stimulus current are
account the empirical observation that the not directly comparable to those of the mechani-
decrease in excitability and recovery due to cal stimulation. Moreover, the T cell model
the Na+ pump is much slower than that due to responds with bursts only to upstrokes of the
the IK,Ca (Jansen and Nicholls 1973; Cataldo stimulus (detects only positive slopes) contrary
et al. 2005). to the real T cell.
The parameters aNa and aCa give the propor- First, we tested if our model neuron was able
tion of the ionic currents contributing to to reproduce the behavior observed in experi-
the sodium and calcium pools, respectively, and ments in response to simple stimulation protocols
are set to aNa = 0.016, aCa = 0.1 for the that hyperpolarize the membrane. We thus
somatic compartment and aNa = 0.16, injected a current Istim consisting of a train of
aCa = 0.2 for the dendritic one. Time constants low-amplitude pulses followed by a long period
for concentration-dependent conductances take of high-amplitude periodic stimulation (Fig. 1d),
the values tk,Ca = 10 ms, tk,Na = 100 ms for before returning to the low-amplitude train. With
both compartments. the full model, the somatic membrane potential
a b
20 20
r(v)/rav
r(v)/rav
B
10 10
0 0
0 1 0 1 2
velocity(nA.ms−1) velocity/σvel
c 0.15 d 0.8
0.6
P (v |b=2)
0.1
P (v |b=2)
0.4
0.05
0.2
0 0
0 1 0 1 2 3
velocity(nA.ms−1) velocity/σvel
Biophysics of Adaptation in a Computational Model s = 18 nA (green). (b) Normalized burst rate as in (a) ver-
of the Leech T Neuron, Fig. 2 Adaptive scaling of the sus input velocity rescaled by the standard deviation of the
two-compartment model. (a) Burst rate, r(v), as a function stimulus velocity ensemble. (c) Probability distribution of
of the stimulus velocity, v, in the two-compartment model, velocities before bursts of two spikes, for the same stim-
normalized by mean burst rate, rav. Three random stimuli ulus ensembles. (d) Probability distributions prior to
with different amplitudes (standard deviations s) were bursts of two spikes versus rescaled input velocity
used: s = 4.5 nA (black line), s = 9 nA (orange), and
hyperpolarizes in a way similar to the real T cell random stimuli of different amplitudes. The full
neuron (Fig. 1b), inducing a latency period of no neuron model is able to respond with a large burst
spiking until the basal resting potential is recov- rate (in bursts per second) to moderate stimulus
ered. To see the role of the sodium pump on this velocities (Fig. 2a) and only saturates at large
behavior, we repeated the protocol in a model velocities of the high-amplitude stimulus (green
with the current Ipump absent from the equations line).
(Fig. 1c). This rendered a response similar to the When the input velocity is scaled by the width
T neuron treated with the drug strophanthidin, of the distribution of velocities, we see that burst
a sodium pump blocker (Arganda et al. 2007). rates code for a very similar range of scaled
The small gap that we observe in Fig. 1c is a con- velocities independently of the stimulus statistics
sequence of the recovery time of the after- (Fig. 2b). This adaptive scaling holds for veloci-
hyperpolarization current IK,Ca, with a faster ties as large as twice the standard deviation of the
time scale than that of the sodium pump. stimulus, demonstrating that our simple model is
Once that we had properly modeled the hyper- able to reproduce the experimentally observed
polarization effect of the sodium pump, we could statistical adaptation for burst rates.
check its role on coding and adaptive scaling. Since the neuron also codes information about
First, we probed the dynamic range of the burst- the stimulus velocity in the duration of bursts, we
ing neuron model stimulating the dendrite with checked whether the burst code can also adapt to
a b
6 6
r(v)/rav
r(v)/rav
4 4
2 2
0 0
0 1 0 1 2
velocity(nA/ms) velocity/σvel
c d
0.6
0.1
P (v |b=2)
P (v |b=2)
0.4
0.05
0.2
0 0
−0.5 0.5 1.5 −1 1 3
velocity(nA/ms−1) velocity/σvel
Biophysics of Adaptation in a Computational Model (orange), and s = 18 nA (green). Compare to Fig. 2a.
of the Leech T Neuron, Fig. 3 The sodium pump is (b) Normalized burst rate versus input velocity rescaled by
responsible for adaptive scaling. (a) Burst rate as the standard deviation of the stimulus velocity ensemble.
a function of the stimulus velocity in the (c) Probability distribution of velocities before bursts of
two-compartment model without the sodium pump cur- two spikes in the model without sodium pump. (d) Prob-
rent, Ipump, for three Gaussian stimulus ensembles of dif- ability distributions prior to two spike bursts versus
ferent amplitude: s = 4.5 nA (black line), s = 9 nA rescaled input velocity
the stimulus statistics. We built the probability First, we see that burst rates saturate much earlier
distribution of velocities prior to bursts of two (compare Figs. 3a and 2a) and the probability
spikes, for different stimulus ensembles (Fig. 2c). distribution of velocities before bursts span
The probability range increases with stimulus a narrower range (compare Figs. 3c and 2c).
width (orange and green lines in Fig. 2c), but This suggests that the response dynamics without
when velocities are normalized by the standard the sodium pump is restricted and probably not
deviation, the velocity code spans the same range able to adapt the response to wide stimulus
independently of the stimulus ensemble (Fig. 2d). ensembles. This is indeed what we observe
Next, we investigated the role of the sodium when we plot burst rates and probabilities versus
pump in the neuron coding properties. We the rescaled input velocities (Fig. 3b, d). The
repeated the stimulation protocols with three dif- rescaled plots show a marked difference between
ferent ensembles, but in a model without the Ipump the low- and high-amplitude ensembles espe-
current (Fig. 3). cially at large velocities.
We see that the sodium pump affects both the We thus conclude that slow activity-
dynamic range of the neuron response and the dependent hyperpolarization, as that induced by
adaptive scaling to the stimulus distribution. the sodium pump, is a general mechanism to
Boltzmann Machine 417 B
enlarge the dynamic range of a neuron’s response
and achieve adaptation to the statistical context of Boltzmann Machine
the stimulus, providing the flexibility needed for
changing real-world stimuli. Joaquin J. Torres
Institute “Carlos I” for Theoretical and B
Computational Physics and Department of
References Electromagnetism and Matter Physics, Facultad
de Ciencias, Universidad de Granada, Granada,
Arganda S, Guantes R, de Polavieja G (2007) Sodium Spain
pumps adapt spike bursting to stimulus statistics. Nat
Neurosci 10:1467–1473
Brenner N, Bialek W, de Ruyter van Steveninck R (2000)
Adaptive rescaling maximizes information transmis- Synonyms
sion. Neuron 26(3):695–702
Cataldo E, Brunelli M, Byrne JH, Av-Ron E, Cai Y, Bax-
Noisy neural network; Stochastic Hopfield
ter DA (2005) Computational model of touch sensory
cells (T cells) of the leech: role of the afterhyperpo- network
larization (AHP) in activity-dependent conduction
failure. J Comput Neurosci 18(1):5–24
Izhikevich EM (2007) Dynamical systems in neurosci-
ence: the geometry of excitability and bursting. MIT
Definition
Press, Cambridge, MA
Jansen JK, Nicholls JG (1973) Conductance changes, an Boltzmann machines are usually defined as neu-
electrogenic pump and the hyperpolarization of leech ral networks in which the input–output relation-
neurones following impulses. J Physiol
229(3):635–655
ship is stochastic instead of deterministic (as it is
Kepecs A, Wang XJ, Lisman J (2002) Bursting neurons in the original Hopfield network). Then,
signal input slope. J Neurosci 22(20):9053–9062 Boltzmann machines are normally defined as sto-
Sharpee TO, Sugihara H, Kurgansky AV, Rebrik SP, chastic versions of the Hopfield model or other
Stryker MP, Miller KD (2006) Adaptive filtering
attractor neural networks. A main difference
enhances information transmission in visual cortex.
Nature 439(7079):936–942 between Hopfield networks and Boltzmann
Wang XJ (1998) Calcium coding and adaptive temporal machines is that whereas in Hopfield networks,
computation in cortical pyramidal neurons. the deterministic dynamics brings the state of the
system downhill, toward the stable minima of
Wark B, Lundstrom BN, Fairhall A (2007) Sensory adap-
tation. Curr Opin Neurobiol 17(4):423–429 some energy function related with some informa-
Yamada WM, Koch C, Adams PR (1998) Multiple chan- tion content, in a Boltzmann machine, such pre-
nels and calcium dynamics. In: Koch C, Segev I (eds) scribed states of the system cannot be reached
Methods in neuronal modelling. MIT press,
due to stochastic fluctuations. For a Boltzmann
Cambridge, MA, pp 137–170
machine, then, the steady state of the system is
characterized by an equilibrium probability dis-
tribution to be in such states, which is given as
a function of the energy of such states by the
Bistability Boltzmann distribution. The inherent stochastic
dynamics of Boltzmann machines allows the sys-
▶ Multistability in Neurodynamics: Overview
tem sometimes to go uphill and, therefore, to
▶ Multistability: Stopping Events with Single
have some probability to scape from undesirable
Pulses
local minima and reach a global minimum. In
practice, this can be done using a simulating
annealing procedure, that is, start the network
BK Channels (KCa1.1) dynamics at high temperature and decrease it
until an equilibrium state is reached with energy
▶ Calcium-Dependent Potassium Channels fluctuating around the global minimum.
B 418 Boltzmann Machine
Boltzmann machines are used, for instance, to 1

Probðsi ðt þ 1Þ ¼ sÞ ¼
retrieve or complete patterns distributed 1 þ exp½ð2s 1Þbhi ðtÞ
according to some predefined equilibrium distri- (2)
bution and make use of visible units, which are
clamped to the patterns in the training set, and with s = 1, 0, and b ¼ kB1T where kB is the
hidden units which are free-running variables of Boltzmann constant and T is a temperature
the system. Boltzmann machines were invented parameter that drives the stochastic changes in
by G. Hinton and T. Sejnowski (1983a, b). the neurons. Expression Eq. 2 implements the
so-called Glauber dynamics and can be obtained
from the Hamiltonian Eq. 1 if one assumes that at
Detailed Description each time t, the configurational energy of the
network is given by E(s(t)). Then, a stochastic
As in the Hopfield model, the simplest form of change at time t in the state of neuron i from
a Boltzmann machine is constituted by a network si = s to si = 1 s implies a change in this
of N binary units or “neurons” whose state is configurational energy given by
represented by state variables si = 1 or 0, i = 1,

. . . , N. The main difference with Hopfield net- DEi ðtÞ ¼ E si ðtÞ EðsðtÞÞ
work arises in the fact that Boltzmann machine " #
X
units are stochastic. Also, the Boltzmann ¼ ð2s 1Þ wij sj ðtÞ yi
machine is characterized by an “energy function” j
characterizing a particular configurational state
¼ ð2s 1Þhi ðtÞ:
of the network s = (s1, . . . , sN) and which is
defined as in the Hopfield model, that is,
Here, si(t) represents a network state in which
the state of neuron i is si (t) = 1 s, assuming
1 X X
EðsÞ ¼ wij si sj þ yi s i (1) that the network state s(t) has si (t) = s. Equa-
2
i, j i tion 2 finally is obtained when Prob(si
i 6¼ j (t + 1) = s) is supposed to be given by the
Boltzmann factor, that is,
where wij represents the weight or strength of the
link or “synapse” connecting neuron i and j. wij is ebEðsðtÞÞ
assumed to be symmetric, that is, wij = wji with Probðsi ðt þ 1Þ ¼ sÞ ¼ , (3)
Z
wii = 0, and yi is a “bias” of unit i that represents
an activation neuron threshold. where Z is obtained from normalization of Eq. 3,
In a Boltzmann machine, the state of a given that is,
neuron i is determined by its total input current,
or local field, defined as Probðsi ðt þ 1Þ ¼ sÞ þ Probðsi ðt þ 1Þ ¼ 1 sÞ ¼ 1:
X
hi ð t Þ ¼ wij sj ðtÞ yi , The last then implies
j
Z ¼ ebEðsðtÞÞ þ ebEðs ðtÞÞ ,

i
where the bias term is assumed to be fixed. Then,

the dynamics of the Boltzmann machine is and therefore,
implemented as follows: given the state of the
ebEðsðtÞÞ
network at time t, a neuron is chosen at Probðsi ðt þ 1Þ ¼ sÞ ¼
random at time t + 1 from the population, and ebEðsðtÞÞ þ ebEðsi ðtÞÞ
its state is updated according with the probabilis- 1
¼
tic rule: 1 þ eb½Eðsi ðtÞÞEðsðtÞÞ
Boltzmann Machine 419 B
which has the form given by Eq. 2 using the value of one neuron and c(si ! s) is the transition
definition of DEi(t) above. probability from state si to s. In fact, one has for
The use of Glauber dynamics Eq. 2 to describe a Boltzmann machine
stochastic changes in neurons in a Boltzmann

machine allows to study its equilibrium proper- c si ! s ¼ Probðsi ðt þ 1Þ ¼ sÞ (7) B
ties using the statistical physics techniques. One
can assume that this stochastic dynamics drives when s = (s1; . . ., s, . . ., sN) and si = (s1, . . . , 1 –
the Boltzmann machine state into a steady state s, . . . , sN). Moreover, a sufficient condition to
which is described by a probability distribution reach a steady state in Eq. 6 is
P(s). Let us consider as above two configurations

or states of the network s and si that differ only in c s i ! s P s i ¼ c s ! s i Pð s Þ
the value that the neuron state si takes. Since
stochastic changes in each neuron occur indepen- which is equivalent to Eq. 4 and it is known as
dently from the others and using Eq. 3, one has a detailed balance condition. With the choice
Eq. 7, this condition guarantees that the station-
PðsÞ Probðsi ¼ sÞ bEðsÞ
ary distribution is PðsÞ ¼ e Z :
¼ eb½EðsÞEðs Þ , (4)
i
¼
Pðs Þ Probðsi ¼ 1 sÞ
i
which implies a steady-state distribution Mean-Field Boltzmann Machines
ebEðsÞ One can define the mean firing rate of neuron i as

Pð s Þ ¼ , (5)
Z X
mi ðtÞ hsi it ¼ si Pt ðsÞ,
which has the form of the Boltzmann distribution
X s
with the normalization factor Z ¼ ebEðsÞ
s
and deriving this expression in time leads to
being the so-called partition function. Therefore, X
P(s) as in Eq. 5 characterizes an equilibrium state @t m i ¼ si @t Pt ðsÞ:
s
of the Boltzmann machine described by the Ham-
iltonian E(s). Introducing the master Eq. 6 in the rhs of the
last equation gives

Master Equation Description @t mi ¼ 2 si c s ! si t þ c s ! si t ,
Under the probabilistic rule Eq. 2, stochastic where h it means an average over Pt(s). Using
changes in neurons in a Boltzmann machine Eq. 7 one gets after some algebra
occur as in a Markov chain (van Kampen 1990).

That is, the dynamics of the state of the network 1 1 bhi
can be viewed as a Markovian stochastic process @t mi ¼ mi þ þ tanh :
2 2 2 t
whose dynamics is governed by the master
equation: The steady state of this dynamics leads to
X
@t Pt ðsÞ ¼ c s i ! s Pt s i 1 1 bhi
i mi ¼ þ tanh
2 2 2
c s ! si Pt ðsÞ, (6)
where now the average is performed over the
where s = (s1,. . ., si,. . ., sN) and si = (s1,. . ., 1 – stationary distribution P(s). Finally, using the
si, . . . , sN) are network states which differ in the mean-field assumption si ! mi, one obtains
B 420 Boltzmann Machine
0 X 1
so-called visible units, vi, i = 1, . . . , n with
1 1 b wij m j y i mi
mi ¼ þ tanh@
j A, n < N that are clamped during learning to the
2 2 2 training set of vectors (which now have only
n units) and the “hidden” units ui i = n + 1, . . . ,
which defines the steady-state mean-field equa- N which are free-running variables of the system.
tion for a Boltzmann machine. The network state then is s = (v1,. . ., vn, un+1,. . .,
uN), and the equilibrium steady-state distribution
Eq. 5 can be written now as P(s) = P(v, u). We
Learning in Boltzmann Machines can compute the marginal equilibrium
probability distribution for visible units as
Boltzmann machines are used to retrieve some follows:
particular global states (the training set) distrib-
uted according to some external given probability X X ebEðv, uÞ
Pð v Þ ¼ Pðv, uÞ ¼
distribution over these states. For this purpose it u u
Z
is necessary to train the network such that the
final weights induce global minima in E(s), asso- with
ciated with the training set and with the highest
X
probabilities according with the external proba- Z¼ ebEðv, uÞ :
bility distribution. This task cannot be done in v, u
standard stochastic Hopfield networks with
Hebbian weights since the obtained equilibrium P(v) gives then the probability of visible units in
distribution describes minima of the E(s) related the free-running system. However, we want these
with global states, for instance, s = j, only with units to be distributed according to some desired
second-order correlations among units as maxi- probabilities PðvÞ which are the real probabilities
mum. In fact, in this case the minimization of of the training set or environment. A useful quan-
Eq. 5 through gradient descent gives tity to measure the difference between distribu-
tions P(v) and PðvÞ is the Kullback–Leibler
" #
@lnPðxÞ X divergence or relative entropy (Kullback and
¼ b xi x j PðsÞsi sj : (8) Leibler 1951) defined as
@wij s
X Pð v Þ
This impedes in the practice to the network to G¼ PðvÞln ,
v
Pð v Þ
learn global states with correlations that involve
more than two units. Such higher-order corre-
lated states occur, for instance, in the well- whichever is positive or zero. Precisely, G has the
known XOR problem, where the global states to minimum zero value when the equilibrium prob-
be retrieved involve correlations among three ability P(v) coincides with the desired probability
units. For a Hopfield network with N = 3, for PðvÞ. One can use this fact to find an algorithm
instance, there are 23 = 8 possible global states which learns the training set or environment pat-
each one with a probability 18 to occur without terns using a gradient descent of G as follows:
previous learning. However, only the patterns
@G X PðvÞ@PðvÞ
(110), (101), (011), and (000) solve the XOR Dwij ¼ ¼ (9)
problem which implies some particular three @wij v
PðvÞ@wij
units structure to be learned by the network
which cannot be implemented by Eq. 8. with > 0 being the learning rate and where it is
To consider higher-order correlations to be used that the clamped probability PðvÞ is inde-
learned by Boltzmann machines, it is necessary pendent of wij. After some algebra (see for
to separate units in the network in two sets, the instance Ackley et al. 1985), one finally obtains
Bradykinesia Models 421 B
h i Kullback S, Leibler RA (1951) On information and suffi-
Dwij ¼ b si sj c si sj f (10) ciency. Ann Math Stat 22(1):79–86
van Kampen NG (1990) Stochastic processes in physics
where hsisji f hsisji = sP(s) sisj is the and chemistry. North-Holland, Amsterdam
standard average over the equilibrium joint
B
probability P(s) = P(v,u) of the free-running
system (without being clamping
to anyXexternal
environment state) and si sj c ¼ s
PðsÞsi sj Bradykinesia Models
is the average over the equilibrium joint
probability PðsÞ ¼ Pðv, uÞ when the visible Vassilis Cutsuridis
units are clamped to the vectors of the external Institute of Molecular Biology and
training set (i.e., the standard Hebbian Biotechnology, Foundation for Research and
learning term). It is worth noting that in Technology - Hellas, Heraklion, Crete, Greece
Eq. 10, si can be either a visible or a hidden
unit in the network.
Learning in Boltzmann machines as it is given Synonyms
in Eq. 10 can be very slow in practice when many
hidden layers are considered which increases the Hypokinesia model; Slowness of movement
size of the network. In this case the system needs model
a long time to reach its equilibrium distribution,
particularly when the system is free-running due
to the existence of many local minima. Then, Definition
many different approaches have been developed
to decrease learning time. One of the most suc- Bradykinesia is the cardinal symptom of
cessful of these techniques is to consider some Parkinson’s disease (PD). It is related to an
restrictions in the network topology as in the abnormal slowness of movement. The causes of
so-called restricted Boltzmann machines where PD bradykinesia are not known largely, because
the visible-visible or the hidden-hidden connec- there are multiple brain areas and pathways
tions are not permitted, being connections involved from the neuronal degeneration site
between visible and hidden neurons the only (dopamine neurons in substantia nigra pars
ones allowed. compacta (SNc) and ventral tegmental area
(VTA)) to the muscles. Bradykinesia models are
mathematical and computational constructs
Cross-References attempting to uncover how information is
processed in the affected brain areas and what
▶ Hopfield Network are the biophysical mechanisms giving rise to
the observed slowness of movement in PD
bradykinesia.
References
Ackley DH, Hinton GE, Sejnowski TJ (1985) A learning Detailed Description

algorithm for Boltzmann machines. Cogn Sci
9(1):147–169
Hinton GE, Sejnowski TJ (1983a) Analyzing cooperative Bradykinesia, the hallmark and most disabling
computation. In: Proceedings of the 5th annual con- symptom of PD, refers to an extreme slowness
gress of the cognitive science society, Rochester of movement. In early stages of the disease, PD
Hinton GE, Sejnowski TJ (1983b) Optimal perceptual
patients have difficulties with daily activities
inference. In: Proceedings of the IEEE conference on
computer vision and pattern recognition. IEEE Com- such as walking, speaking, or getting in and out
puter Society, Washington DC, pp 448–453 of chairs (Gibberd 1986). In the later phases of
B 422 Bradykinesia Models
the disease, the entire movement process • Reduction of peak neuronal activity and rate
becomes increasingly slow and occasionally of development of neuronal discharge in the
results in a complete inability to move. Patients primary motor cortex and premotor area
must concentrate intensely to overcome the iner- (Gross et al. 1983; Watts and Mandir 1992).
tia of their limbs even when they are executing • Abnormal oscillatory GP (external and inter-
the simplest motor tasks. Movement is particu- nal) neuronal responses (Tremblay et al.
larly impaired when novel movements are 1989).
attempted (Connor and Abbs 1991) or when sev- • Disinhibition of reciprocally tuned cells
eral movements are combined (Benecke et al. (Doudet et al. 1990). Reciprocally tuned cells
1986; Lazarus and Stelmach 1992). are cells that discharge maximally in one
movement direction but pause their activities
Anatomical Overview in the opposite direction.
Initiation and execution of voluntary movements • Significant increase in mean duration of neu-
involves brain areas and pathways. The “motor ronal discharge in motor cortex preceding and
circuit” originates in the motor areas of the fron- following onset of movement (Gross et al.
tal cortex, which activate motor portions of the 1983; Doudet et al. 1990; Benazzouz et al.
basal ganglia subcortical structures (striatum, 1992).
globus pallidus external (GPe) and internal • Multiple triphasic patterns of muscle activa-
(GPi) segments, subthalamic nucleus (STN), tion (Hallett and Khoshbin 1980; Doudet et al.
substantia nigra pars reticulata (SNr)) and the 1990). Triphasic pattern of muscle activation
thalamus and which in turn project back to the is a characteristic electromyographic (EMG)
frontal motor areas of the cortex. The basal pattern characterized by alternating bursts of
ganglia structures are implicated in the selection agonist and antagonist muscles. The first ago-
of the most appropriate motor command given nist burst provides the impulsive force for the
the current context. Motor commands from the movement, whereas the antagonist activity
frontal motor areas of the cortex (premotor, sup- provides the braking force to halt the limb.
plementary, and primary motor areas) activate Sometimes a second agonist burst is needed
the corresponding motor spinal centers, which to bring the limb to the final position. In PD
then activate the muscles. patients, multiple such patterns are observed
in order for the subjects to complete the
movement.
Physiological and Behavioral • Reduction in the rate of development and peak
Phenomena amplitude of the first agonist burst of EMG
activity (Godaux et al. 1992; Corcos et al.
PD bradykinesia has been linked with the degen- 1996; Hallett and Khoshbin 1980; Doudet
eration of dopamine neurons in SNc and VTA. et al. 1990; Watts and Mandir 1992; Berardelli
Bradykinesia manifests only when 80–90 % of et al. 1986).
dopamine neurons die. All motor cortical and • Co-contraction of muscle activation
subcortical areas are innervated by SNc and (Benazzouz et al. 1992). In PD patients, the
VTA dopamine neurons (Williams and alternating agonist-antagonist-agonist muscle
Goldman-Rakic 1998; Bjorklund and Lindvall activation is disrupted resulting in the
1984; Gerfen et al. 1990). The degeneration of coactivation of opponent muscle groups.
dopamine neurons leads to a number of changes • Increases in electromechanical delay time
relevant to bradykinesia in the neuronal, electro- (time between the onset of modification of
myographic (EMG), and movement parameters agonist EMG activity and the onset of move-
reported in parkinsonian human and animal ment) (Benazzouz et al. 1992; Doudet et al.
brains: 1985, 1990).
• Asymmetric increase in acceleration (time dopamine neurons die, resulting in a reduction
from movement onset to peak velocity) and in transmission through the direct pathway, and
deceleration (time from peak velocity till end an increase in transmission through the indirect
of movement) times of a movement. pathway. Dominance of the indirect pathway
• Decrease in the peak value of the velocity leads to an excessive inhibition of the thalamus B
trace (Godaux et al. 1992; Camarata et al. by the GPi, thus leading to an abnormal slowness
1992; Weiss et al. 1996; Benazzouz et al. of movement (i.e., bradykinesia).
1992; Doudet et al. 1985, 1990; Rand et al. Additional anatomical observations have
2000). suggested a more complex intrinsic organization
• Significant increases in movement time (Rand of BG structures:
et al. 2000; Weiss et al. 1996; Doudet et al. • Both populations of striatal output neurons
1985, 1990; Watts and Mandir 1992; project to the GPe, one exclusively (enkepha-
Benazzouz et al. 1992). lin/D2 neurons) and the other via collaterals
from the fibers innervating the output nuclei
(substance P/D1 neurons) (Parent et al. 2000).
Types of Theoretical Models of • GP neurons make direct contact with the out-
Bradykinesia put nuclei as well as with the STN (Smith et al.
1998).
Theoretical models of bradykinesia fall under • The GP also projects back to the striatum
two major categories: (Bevan et al. 1998).
• Verbal-conceptual models: using informal and • The STN is driven by both cortical and sub-
natural language, describe the brain areas, cortical structures external to the basal
pathways, and interactions leading to parkin- ganglia, now known as the “hyperdirect” path-
sonian bradykinesia. way (Nambu et al. 2002).
• Mathematical and computational models: Connections between the various BG compo-
using mathematical equations as a language, nents are topographically ordered (Mink 1996).
describe the interactions between the various Some projections such as the striatonigral projec-
brain areas involved in movement control and tion are more focused, whereas others such as the
execution in parkinsonian bradykinesia. subthalamo-nigral projection are more diffuse
(Mink 1996). The hyperdirect pathway provides
a widespread excitation of the GPi (Mink 1996)
Verbal-Conceptual Models and is considered to suppress thalamic and corti-
cal areas related to both the selection of the most
An influential model of basal ganglia intrinsic appropriate motor program given the current con-
organization was proposed by Albin and col- text and competing programs before movement
leagues (1989). In their model, motor cortical begins. On the other hand, the direct pathway
areas drive two populations of striatal medium facilitates movement via a more focused disinhi-
spiny output neurons. Neurons containing sub- bition of the thalamus, while the indirect pathway
stance P and D1-type dopamine receptors com- brakes this facilitation (Alexander and Crutcher
prise the “direct” pathway and make contact with 1990).
the basal ganglia output nuclei. At the same time, These experimental observations extend the
striatal neurons containing enkephalin and original Albin and colleagues model (1989) of
D2-type dopamine receptors comprise the “indi- BG information processing. Some of the mathe-
rect” pathway and contact the output nuclei via matical and computational models of the next
relays in the GPe and STN. Basal ganglia output section attempt to explain how dysfunction of
was thought to reflect a balance between these these various interconnected BG structures may
two projections, which is disrupted when lead to bradykinesia.
B 424 Bradykinesia Models
Mathematical and Computational 2010, 2013; Cutsuridis and Perantonis 2006).

Models He suggested that the observed abnormal slow-
ness of movement in PD bradykinesia is due to
Basal Ganglia-Thalamocortical Interactions inadequately activated motor cortical and spinal
In line with the Albin and colleagues (1989) and cord centers because of dopamine reduction not
Nambu et al. (2002) models, Contreras-Vidal only in the basal ganglia but also in cortical and
and Stelmach (1995) introduced a detailed spinal sites. His models, which were population-
population-based model of basal ganglia- based models, were composed of two modules
thalamocortical relations in normal and parkinso- coupled together: (1) the cortical module and
nian movements. The model’s architecture was (2) the spinomuscular module. Both modules
based on the direct, indirect, and hyperdirect and their corresponding neuronal components
pathways schema of the basal ganglia. Activation were modulated by dopamine. The cortical mod-
of the direct pathway resulted in activation of ule computed the motor commands sent to the
thalamocortical motor circuits leading to initia- spinomuscular module. The spinomuscular mod-
tion and modulation of movement, whereas acti- ule was an opponent processing control model of
vation of the indirect pathway led to breaking of how spinal circuits afford independent voluntary
ongoing movement. Activation of the hyperdirect control of joint stiffness and position. Both mod-
pathway facilitated rapid movement switching or ules consisted of all major neuronal populations
the prevention of movement release. Contreras- as they have been reported in the experimental
Vidal and Stelmach showed that loss of striatal literature. The model accounted for all physio-
DA as it occurs in PD leads to an imbalance in the logical and behavioral phenomena as they have
neurotransmitter dynamics in the direct and indi- been described in a previous section. Model sim-
rect pathways, producing smaller-than-normal ulations showed that reduction of DA in cortical
BG output signals. In turn these output signals and subcortical motor areas disrupts, via several
activate insufficiently otherwise normally func- pathways, the rate of development and peak neu-
tioning motor cortical and spinal sites and pro- ronal activity of primary motor cortical cells.
duce weak and slow movements. These changes lead in delays in recruiting the
Moroney and colleagues (2008) extended the appropriate level of muscle force sufficiently
previous model to investigate the factors that fast and in a reduction of the peak muscle force
contribute to the slowness of movements of required to complete the movement. Repetitive
PD patients when they perform simple and com- and sometimes co-contractive patterns of muscle
plex voluntary movements. Excessive dopamine activation are needed to complete the movement.
depletion in the striatum and loss of spatial These disruptions result in an abnormal slowness
segregation of neuronal populations operating of movement.
as functionally independent modules
somatotopically mapping particular body
Cross-References
parts contribute to a slowness of movement and
to a reduced ability to suppress unwanted move-
▶ Basal Ganglia: Mechanisms for Action
ments. They further showed that the therapeutic
Selection
effects of deep brain stimulation (DBS) in STN
▶ Basal Ganglia: Overview
result from stimulation-induced inhibition of
▶ Subthalamic Nucleus Cellular Models
STN, partial synaptic failure of efferent projec-
▶ Subthalamopallidal Loop and Oscillations
tions, or excitation of inhibitory afferent axons.
Cortico-spinomuscular Interactions References

An alternative view to the observed abnormal
slowness of movement in PD bradykinesia was Albin RL, Young AB, Penney JB (1989) The functional
proposed by Cutsuridis (Cutsuridis 2006a, b anatomy of basal ganglia disorders. TINS 12:366–375.
Alexander GE, Crutcher MD (1990) Functional architec- Doudet DJ, Gross C, Lebrun-Grandie P, Bioulac B (1985)
ture of basal ganglia circuits: neural substrates of par- MPTP primate model of Parkinson’s disease:
allel processing. TINS 13(7):266–271 a mechanographic and electromyographic study.
Benazzouz A, Gross C, Dupont J, Bioulac B (1992) Brain Res 335:194–199
MPTP induced hemiparkinsonism in monkeys: Doudet DJ, Gross C, Arluison M, Bioulac B (1990)
behavioral, mechanographic, electromyographic and Modifications of precentral cortex discharge and B
immunohistochemical studies. Exp Brain Res EMG activity in monkeys with MPTP induced
90:116–120 lesions of DA nigral lesions. Exp Brain Res
Benecke R, Rothwell JC, Dick JPR (1986) Performance of 80:177–188
simultaneous movements in patients with Parkinson’s Gerfen CR, Engber TM, Mahan LC, Susel Z, Chase TN,
disease. Brain 109:739–757 Monsma FJ Jr, Sibley DR (1990) D1 and D2 dopamine
Berardelli A, Dick JPR, Rothwell JC, Day BL, Marsden receptor-regulated gene expression of striatonigral and
CD (1986) Scaling of the size of the first agonist EMG striatopallidal neurons. Science 250:1429–1432
burst during rapid wrist movements in patients with Gibberd FB (1986) The management of Parkinson’s dis-
Parkinson’s disease. J Neurol Neurosurg Psychiatry ease. Practitioner 230:139–146
49:1273–1279 Godaux E, Koulischer D, Jacquy J (1992) Parkinsonian
Bevan MD, Booth PAC, Eaton SA, Bolam JP bradykinesia is due to depression in the rate of rise of
(1998) Selective innervation of neostriatal interneu- muscle activity. Ann Neurol 31(1):93–100
rons by a subclass of neuron in the globus pallidus of Gross C, Feger J, Seal J, Haramburu P, Bioulac B (1983)
the rat. J Neurosci 18:9438–9452 Neuronal activity of area 4 and movement parameters
Bjorklund A, Lindvall O (1984) Dopamine containing recorded in trained monkeys after unilateral lesion of
systems in the CNS. In: Bjorklund A, Hokfelt T (eds) the substantia nigra. Exp Brain Res 7:181–193
Handbook of chemical neuroanatomy. Classical trans- Hallett M, Khoshbin S (1980) A physiological mechanism
mitters in the CNS, Part 1, vol 2. Elsevier, Amsterdam, of bradykinesia. Brain 103:301–314
pp 55–121 Lazarus JC, Stelmach GE (1992) Inter-limb coordination
Camarata PJ, Parker RG, Park SK, Haines SJ, Turner DA, in Parkinson’s disease. Mov Disord 7:159–170
Chae H et al (1992) Effects of MPTP induced Mink JW (1996) The basal ganglia: focused selection and
hemiparkinsonism on the kinematics of a two-dimen- inhibition of competing motor programs. Prog
sional, multi-joint arm movement in the rhesus mon- Neurobiol 50:381–425
key. Neuroscience 48(3):607–619 Moroney R, Heida C, Geelen J (2008) Increased
Connor NP, Abbs JH (1991) Task-dependent variations in bradykinesia in Parkinson’s disease with increased
parkinsonian motor impairments. Brain 114:321–332 movement complexity: elbow flexion-extension
Contreras-Vidal JL, Stelmach G (1995) A neural model movements. J Comput Neurosci 25:501–519
of basal ganglia-thalamocortical relations in normal Nambu A, Tokuno H, Takada M (2002) Functional sig-
and Parkinsonian movements. Biol Cybern nificance of the cortico-subthalamo-pallidal
73:467–476 ‘hyperdirect’ pathway. Neurosci Res 43:111–117
Corcos DM, Chen CM, Quinn NP, McAuley J, Rothwell Parent A et al (2000) Organization of the basal ganglia: the
JC (1996) Strength in Parkinson’s disease: relationship importance of axonal collateralization. Trends
to rate of force generation and clinical status. Ann Neurosci 23:S20–S27
Neurol 39(1):79–88 Rand MK, Stelmach GE, Bloedel JR (2000) Movement
Cutsuridis V (2006a) Biologically inspired neural accuracy constraints in Parkinson’s disease patients.
architectures of voluntary movement in normal and Neuropsychologia 38:203–212
disordered states of the brain. Unpublished PhD Smith Y, Bevan MD, Shink E, Bolam JP (1998) Microcir-
dissertation cuitry of the direct and indirect pathways of the basal
Cutsuridis V (2006b) Neural model of dopaminergic con- ganglia. Neuroscience 86:353–387
trol of arm movements in Parkinson’s disease Tremblay L, Filion M, Bedard PJ (1989) Responses of
bradykinesia. In: Kolias S, StafilopatisA, Duch pallidal neurons to striatal stimulation in monkeys with
W (eds) ICANN 2006: artificial neural networks. MPTP-induced parkinsonism. Brain Res 498(1):17–33
LNCS, vol 4131. Springer, Berlin, pp 583–591 Watts RL, Mandir AS (1992) The role of motor cortex in
Cutsuridis V (2010) Neural network modeling of volun- the pathophysiology of voluntary movement deficits
tary single joint movement organization. II. associated with parkinsonism. Neurol Clin
Parkinson’s disease. In: Chaovalitwongse WA, 10(2):451–469
Pardalos P, Xanthopoulos P (eds) Computational neu- Weiss P, Stelmach GE, Adler CH, Waterman C (1996)
roscience. Springer, New York, pp 193–212 Parkinsonian arm movements as altered by task diffi-
Cutsuridis V (2013) Bradykinesia models of Parkinson’s culty. Parkinsonism and Related Disorders 2(4):
disease. Scholarpedia, under review 215–223
Cutsuridis V, Perantonis S (2006) A neural model of Williams SM, Goldman-Rakic PS (1998) Widespread ori-
Parkinson’s disease bradykinesia. Neural Netw gin of the primate mesofrontal dopamine system.
19(4):354–374 Cereb Cortex 8:321–345
B 426 Braided Electrodes
Further Reading The stability of chronic electrode implants is

Scholarpedia influenced among other things by (a) the compli-
Basal Ganglia
ance, (b) the size, and (c) the biocompatibility and
Dopamine Anatomy
Models of Basal Ganglia coatings of the implanted structure (Branner
Models of Parkinson’s Disease Bradykinesia et al. 2004; Hochberg et al. 2006; Misra
Models of Spinal Cord et al. 2013). The compliance of a single wire scales
Models of Thalamocortical System
as the cube of the length over radius to the fourth
Wikipedia
Hypokinesia power (e.g., Head et al. 1989; Kim et al. 2013).
Motor Control This means that the degree of stress and strain in
Motor System the tissue surrounding an implanted electrode is
Parkinson’s Disease
reduced as electrode wires become thinner and
longer. The conductance of the rodlike electrode
members will only reduce as a function of the
second power of radius reductions, so conductance
Braided Electrodes is not an issue as wires get thinner.
Fifty micron wires are routinely used to record
Simon F. Giszter, Taegyo Kim and neural activity. Given the relationships men-
Almut Branner tioned above, creating a neural interface of
Neurobiology and Anatomy, Drexel University a collection of independent 10 mm wires will
College of Medicine and School of Biomedical significantly improve the compliance of the
Engineering and Health Systems, Philadelphia, ensemble and increase recording electrode den-
PA, USA sity while potentially preserving wire conduc-
tance within the operating range of current and
conventional headstages and amplifiers used in
Definition single-unit recording.
Mechanical Testing: Mechanical bending tests
Braided electrodes consist of multiple, indepen- have shown that the implanted portion of braided
dent fine wires that are braided around a stiff core microelectrodes consisting of either 12 or
or needle that is used to aid insertion. 24 ultrafine wires (9.6 um or 12.7 um) had 4–21
times better mechanical compliance than a single
50 um wire. The wired connections between an
Detailed Description implant and a percutaneous connector also play
a vital role in the long-term performance of an
The braids are like a textile or fabric covering implant since these tethers can produce forces on
around a stiff core. This needlelike core is needed the implant. The mechanical compliance of the
to shape the braid initially and can also be used as bundles of ultrafine wires were 6–96 times better
an aid for insertion into tissue. It is removed after than the compliance of a single 50 um wire (Kim
insertion for chronic applications to allow the et al. 2013).
braid to move with the tissue and reduce its foot-
print in the tissue. Each wire in a braid is Histology: Short Summary of Histology
a separate conductor, i.e., microelectrode. These Different Types of Braided Electrodes: Braids are
wires can be used for either recording or stimu- made from various topologies of interwoven
lation depending on the material used and the strands (Ko and colleagues, 4). The mechanical
impedances of the active sites (Figs. 1 and 2). properties of precisely braided composites offer
Rationale for Braided Electrodes: Braided elec- several advantages that are presently utilized in
trodes take advantage of two things: (a) the flexi- textiles and modern composite materials (e.g.,
bility/compliance of ultrafine microwires and various sports equipment and body armor).
(b) the structure and strength of interwoven strands. Examples of topologies for microelectrode braids
Braided Electrodes 427 B
Braided Electrodes, Fig. 1 Examples of braided electrode probe assemblies and springy flexibility of braided probes
after core removal (From Kim et al. 2013 with permission)
Braided Electrodes, Fig. 2 Stages in braided probe Remaining braid/tether probe after deployment
insertion and deployment. (a) Assembly. (b) Insertion of (Reproduced from Kim et al. 2013, with permission)
assembly. (c) Removal of core leaving only braid. (d)
are (a) tubular braids, (b) 3D braids, and (c) Figured braids do not have a symmetric
(c) figured braids: repeating pattern but have systematic alter-
(a) Tubular braids form the simplest assemblies ations in the braid pattern through the struc-
with symmetric repeating patterns. The ture to alter compliance or to assemble wires
recording sites are located at the ends of the in specific groupings and to aid in the place-
wires which roughly puts them in a circular ment of recording sites along the braid.
arrangement at a particular depth below the
tissue surface.
(b) 3D braids are similar to tubular braids as they Special Techniques Enabled by Braiding
consist of symmetric repeating patterns, but
the recording sites for each wire in the braid Combinatorics
are placed at varying locations throughout the The two characteristics of combinatoric probes
braid. This allows a more distributed place- are that (a) each wire has multiple recording sites
ment of recording sites in the tissue. along its length and (b) the recording sites of
B 428 Braided Electrodes
Braided Electrodes, Fig. 3 Combinatorics concept combinatorics probe with three recording points per
examples. (a) A “4 choose 2” combinatorics probe with wire, and four spatial location sites ABC, BCD, ACD,
two recording points per wire, and six spatial location sites and ABD
AB, BC, CD, AC, AD, and BD. (b) A “4 choose 3”
different wires are organized in a patterned braid spatial site in a braid, a neuron in a combinatoric
into unique combinations of electrodes such as electrode will have a “signature.” That is, a unit
stereotrodes or multitrodes. Gerstein and Clark can be spatially assigned based on the unique
showed in 1964 that it was possible to record collection of wires on which it co-occurs. The
multiple classifiable single-unit action potentials combinatoric probe is thus an innovation that
from multiple 5–10 mm exposed recording points exploits principles of stereotrode/tetrode
spaced along a single 10 mm Tungsten microelec- methods (Gray et al. 1995) for multisite wires
trode (Gerstein and Clark 1964). We have also organized into a 3D combinatoric pattern.
observed that two electrically connected elec- Because we know which electrodes picked up
trodes in a Utah array can occasionally still a particular event, cross-correlated patterns of
record single units, but both channels record the recorded spikes can then be separated out into
same activity. Based on their results and our their location of origin on the probe in most
observations, and back-of-the-envelope calcula- instances. Some activity may remain spatially
tions of theoretical electrical networks, signals undefined in cases of limited (thus ambiguous)
from separate recording sites along one such pickup. The improved yield of units in
electrode wire will roughly add (Fig. 3). a combinatoric arrangement of n wires arranged
Under the assumption that most single-unit in groups of k recording points per spatial site
pickup near a spatial site will simultaneously from combinations of recording points on multi-
co-occur on each of the wires that comprise the recording-point wires can potentially be huge.
Brain Architecture Management System (BAMS) 429 B

n nð n 1Þ . . . ð n k þ 1Þ tolerance of the applications and analysis algo-
n Ck or ¼ , rithms using the data.
k kðk 1Þ:::1
n!
i:e:,
k!ðn kÞ!
References B
The yield improvement arises from the pro- Branner A, Stein RB, Fernandez E, Aoyagi Y, Normann
RA (2004) Long-term stimulation and recording with
cess of arranging recording points on wires into
a penetrating microelectrode array in cat sciatic nerve.
many or all of the possible combinations at spa- IEEE Trans Biomed Eng 51:146–157
tial sites. For example, in Fig. 3a, we show four Gerstein GL, Clark WA (1964) Simultaneous studies of
wires with four recording points each. In Fig. 3a, firing patterns in several neurons. Science
143:1325–1327
we obtain 4C2 spatial sites (4 choose
Gray CM, Maldonado PE, Wilson M, McNaughton
2) = 6 stereotrode sites. In Fig. 3b, we employ B (1995) Tetrodes markedly improve the reliability
four wires, each with three recording points, used and yield of multiple single-unit isolation from multi-
in triode groupings, and obtain 4C3 (4 choose unit recordings in cat striate cortex. J Neurosci
Methods 63:43–54
3) = 4 triode sites, compared to a single tetrode
Head AA, Ko FK, Pastore CM (1989) Handbook of indus-
site. Accordingly, a braid of six wires allows 6C4 trial braiding. Atkins and Pearce, Covington, KY, USA
spatial sites, which gives 15 tetrode sites, com- Hochberg LR, Serruya MD, Friehs GM, Mukand JA,
posed using only six multisite wires, each wire Saleh M, Caplan AH, Branner A, Chen D, Penn RD,
Donoghue JP (2006) Neuronal ensemble control of
with ten recording points.
prosthetic devices by a human with tetraplegia. Nature
For this recording strategy to provide unam- 442:164–171
biguous unit data the firing must ideally be Kim T, Branner A, Gulati T, Giszter SF (2013) Braided
sparse, largely desynchronized on a millisecond multi-electrode probes: mechanical compliance char-
acteristics and recordings from spinal cords. J Neural
time scale, and loss of data or spatial precision
Eng 10(4):045001. doi:10.1088/1741-2560/10/4/
due to occasional collisions must be acceptable 045001 (Epub 2013 May 31, PMID: 23723128)
for the use and algorithms applied to recorded Misra A, Kondaveeti P, Nissanov J, Barbee K,
data. One determining factor for collision likeli- Shewokis P, Rioux L, Moxon KA (2013) Preventing
neuronal damage and inflammation in vivo during
hood is the number of recording points per wire.
cortical microelectrode implantation through the use
A 24-count braid organized as tetrodes allows up of Poloxamer P-188. J Neural Eng 10:016011
to 24C4, i.e., 42,504 uniquely coded spatial sites.
The collision likelihood per wire in the probe with
sufficient recording points to fully span the poten-
tial 42,504 separate spatial sites would clearly be Brain Architecture Management
unacceptably large. In contrast the number of spa- System (BAMS)
tial sites that are possible using 4–6 recording
points per wire would still be an improvement Mihail Bota
compared with current conventional designs, Department of Biological Sciences, University of
while the likelihood of the collisions on a wire Southern California, Los Angeles, CA, USA
would probably be acceptable. For example, set-
tling on limiting the design to only four recording
points per wire allows 24 spatially separate tet- Definition
rodes to be very simply constructed within the
24-wire tubular braid, allowing 4 times the unit The Brain Architecture Management System
yield per wire compared with using six separate (BAMS; http://brancusi.usc.edu) is an online
tetrodes. Optimized recording arrangements in publicly accessible neuroinformatic platform
this combinatoric design space will thus depend designed to handle neurobiological information
on the neural density, average firing frequency, at different organization levels of the vertebrate
and spatial firing pattern statistics and the collision central nervous system (CNS).
B 430 Brain Architecture Management System (BAMS)
Detailed Description nomenclatures, hierarchically organized. A

unique feature of this module is the set of classi-
BAMS Structure fication criteria that can be associated to any
The backend database of BAMS is implemented hierarchically organized neuron nomenclature
in the latest MySQL public version, MariaDB (“is-a” relationship). BAMS includes more than
(http://mariadb.org). The backend structure of 100 structural, hodological (inputs and outputs),
BAMS is modular, with the Brain Parts module electrophysiological, and functional criteria and
as the center of the system. Any brain part (gray sub-criteria, which can be used to specify the
matter regions, fiber tracts, and ventricles) are common features and specific differences of neu-
uniquely defined in BAMS by a tuplet made of ronal classes and types (Bota and Swanson 2007).
its name, the nomenclature used to identify and Any neuronal class or type captured in
name it, the nomenclature version, and species. a hierarchically organized manner includes the
A neuroanatomical nomenclature is defined here specific sets of criteria, as well as the relevant
as an internally consistent set of terms used to metadata (textual definitions, associated refer-
name a CNS part. The number of terms included ences, and collators names).
in a nomenclature is variable (at least one), and The Molecules module represents data and
the nomenclature version uniquely specifies the metadata as collated from the literature and
set. Usually, a brain nomenclature is associated pertaining to brain regions or neurons that are
with a neuroanatomical atlas, but this does not recorded in two general classes of physiological
constitute a constraint for inserting it. The con- states of the animals (subjects): “normal” and
straints that are considered are the internal con- “manipulatedQ” Molecules are also classified in
sistency, whether the reference is original two general classes: “cell associated” and
research, and the brain region names are associ- “releasable by neurons.” The “cell associated”
ated with textual definitions, or at least depicted class is further divided in classes accepted by
on a brain atlas (Bota and Swanson 2005). the IUPHAR nomenclature (http://www.iuphar-
The BAMS modules are Connections, Rela- db.org/index.jsp).
tions, Cells, and Molecules. The Connections The Molecules module allows insertion of
module holds data and metadata about macro- quantitative and qualitative attributes of mole-
scopic neuroanatomical projections between cules identified using the most used techniques
gray matter regions. Any neuroanatomical con- used in neuroanatomy: radioactive tracers,
nection is represented in BAMS by a set of more immunohistochemistry, and expression of genes
than 40 attributes (Bota et al. 2005). The Rela- at the regional and neuronal levels. It also
tions module stores qualitative (topological) spa- includes a comprehensive schema for metadata
tial relations between brain regions in the same representation of experiments, from the details of
species and defined in different nomenclatures, as the used antibodies to the plane of sectioning.
well as the supporting references as metadata. This metadata schema allows insertion of data
The Cells module allows insertion of names and and metadata specific to each of the used tech-
definitions of neuron types and classes, as well as niques listed above. A full description of this
relations between them, as collated from the module can be found in Bota and Swanson
published literature. It records neuronal popula- (2006).
tion attributes, such as topographical positions
within brain regions, densities, and patterns of BAMS Interfaces and Inference Engines
staining. It also includes a classification schema, The Web BAMS interfaces are comprehensive
which allows specification of criteria used to and versatile, allowing users to search for infor-
hierarchically organize neuron nomenclatures mation and manipulate inserted data and meta-
(Bota and Swanson 2007). Thus, the Cells mod- data in different way. Both interfaces and
ule allows insertion of different neuronal inference engines are implemented in PHP
Brain Architecture Management System (BAMS), Fig. 1 The general output of search for brain regions in BAMS.
See text for details
(http://www.php.net/), as well as JavaScript. The retrieved results (gray matter regions names)
middleware PHP is continuously updated and will return the region information, recorded in
upgraded, and at present we use its 5.5.x version. BAMS, and links to the associated neurons,
Generally, the outputs of searches are either text, expressed molecules, and to its inputs and outputs
tables, and images (PNG format). Each BAMS (Fig. 1).
module described in the previous section can be BAMS users have several options to create
searched for information in a different way. gray matter networks. First, they can use the
A complete description of search engines input and output interfaces, choosing a certain
implemented in BAMS, as well as examples, set of regions of interest (Bota et al. 2005). The
can be found in its Manual: http://brancusi.usc. connections between these regions will be first
edu/public/files/user_manual.pdf. displayed in tabular formats. The interface that
Two of the simplest ways to search for infor- displays these tables of connections includes
mation in BAMS are by gray matter regions a link for extension of the networks with one
names and abbreviations. The results of these more layer. Users will be asked to choose the
searches will be lists of regions with those regions of interest that are the targets of the affer-
names or abbreviations, respectively, which best ent regions of the initial tables. Once this action is
match the searched strings. These lists are finished, the user-chosen networks will be
ordered in descending order of the relevance, as displayed as extended tables, as well as in graph-
well as by the type of information that is associ- ical formats. Examples of such complex net-
ated in the system. Clicking on any of the works that can be constructed by users can be
B 432 Brain Architecture Management System (BAMS)
found in BAMS’ Manual and at the URL: http:// The Cells module interfaces include a simple
tinyurl.com/mf4jxws. A second option to create browsing function, which lists the names of neu-
gray matter regions networks is automated, and rons, the associated neuron nomenclature, their
the users have to specify the first and the last definitions, and the collators’ names. It also
region, as well as the number of intermediary includes search forms by names of neurons and
layers (Bota et al. 2003, 2005). This engine infers by brain regions. The former is divided in
all possible networks between the user-specified “BAMS Reference” and “Other.” The names of
regions and with the preset number of layers neurons tagged with “BAMS Reference” are the
(intermediary stations), from the connections standard terms of the system, while those tagged
data inserted in BAMS. To reduce the number “Others” are synonyms or partial correspon-
of possible networks, this engine constructs only dences. The criteria for choosing a neuron name
the forward ones and with at least moderate con- as “BAMS Reference” are described and
nections. The third and the newest way to con- discussed in detail in Bota and Swanson (2007).
struct and visualize networks uses a graphical Additional search functionalities include those
interface constructed de novo in the second gen- for references inserted in BAMS and search for
eration of BAMS, BAMS2 (see below). This brain regions by species and for relations between
functionality can be accessed from the brain parts defined in different nomenclatures.
“Connectome” tab of BAMS’ Menu. The net-
works that can be accessed by users are those BAMS2
that can be constructed using the rat Swanson- In 2010 we decided to upgrade BAMS to the
2004 neuroanatomical nomenclature and hierar- second generation. Thus, we constructed de
chy (Swanson 2004) and “custom” matrices, cre- novo the entire infrastructure, structure, and
ated by the curators of the system. The first set of interfaces of the new system, BAMS2.
networks is those of the major divisions of the rat BAMS2 uses both PostgreSql as a classic rela-
CNS. An example of such network is shown in tional database and RDF triples to store the data
Fig. 2. and metadata. The Web framework Django is
Additionally, users can search for connec- used to create its interfaces. BAMS2 (http://
tions of regions, using autocomplete forms brancusi1.usc.edu) is not a simple iteration of
that allow both names and abbreviations. the older system, because it includes new func-
These forms are implemented in BAMS2 (see tionalities, besides the ones already existent in
below). BAMS. Secondly, the BAMS2 public data and
The Molecules module data and metadata can metadata is encoded in RDF triplets and queried
be accessed in two ways: a simple browse func- using SparQL. Finally, we developed in BAMS2
tion of molecules by categories and an engine a comprehensive and versatile Personal
that compares the experiments inserted in Workspace module, which allows rapid insertion
BAMS. BAMS includes both presence and of connections data annotated and collated
absence of molecules in brain regions or neuron; directly from experiments. The Workspace data
therefore, this variable will be always showed. and metadata are stored in PostgreSQL relational
The state of the animal, “normal” or “manipu- database. Besides its extended functionality to
lated,” is also always associated with each handle and represent neuroanatomical data
record. BAMS includes an additional function directly inserted by neuroanatomists, the
that of inference of the expression patterns of Workspace also allows third-party systems anon-
gray matter parts that include subparts, and these ymous access, to create tabular and graphical
are associated with molecules data and meta- representations of their stored connections data.
data. A complete description of the Molecules This novel functionality of BAMS2 Workspace is
module and its interfaces can be found in Bota exemplified to date in the “Connectome” tab of
and Swanson (2006). the classical BAMS (see above).
Brain Architecture Management System (BAMS), nomenclature (Swanson 2004), constructed automatically
Fig. 2 The network representation of connections of the in BAMS
rat somato-motor cortex, as defined in the Swanson-2004
BAMS: Inserted Information References

BAMS includes to date about 1.25 million
details, most of them manually inserted. It stores Bota M, Swanson LW (2006) A new module for on-line
manipulation and display of molecular information in
about 10,000 CNS parts in five species and 33
the Brain Architecture Management System.
nomenclatures; about 70,000 connections Neuroinformatics 4:275–298
reports, mostly in the rat CNS; 341 neuron iden- Bota M, Swanson LW (2007) The neuron classification
tified names identified in more than 200 rat problem. Brain Res Rev 56:79–88
Bota M, Dong H-W, Swanson LW (2003) From gene
regions; and about 9,000 molecule reports in
networks to brain networks. Nat Neurosci 6:795–799
both rat and mouse CNS and related to more Bota M, Dong H-W, Swanson LW (2005) Brain architec-
than 100 molecules expressed in these species ture management system. Neuroinformatics 3:15–48
brains. These data and metadata have been col- Swanson LW (2004) Brain maps: structure of the rat brain.
A laboratory guide with printed and electronic tem-
lated from more than 700 published references,
plates for data, models and schematics, 3rd edn.
most of them original research articles. Elsevier, Amsterdam
B 434 Brain Atlases
Brain Atlases Brain Extracellular Space:

A Compartment for Intercellular
Yoonsuck Choe Communication and Drug Delivery
Engineering, Texas A&M University, College Sabina Hrabetova1 and Jan Hrabe2
1
Station, TX, USA Department of Cell Biology, SUNY Downstate
Medical Center, Brooklyn, NY, USA
2
Center for Advanced Brain Imaging, Nathan S.
Definition Kline Institute for Psychiatric Research,
Orangeburg, NY, USA
An atlas is a collection of maps. Used in
a geographic sense, an atlas refers to
a collection of maps of an entire planetary body Synonyms
such as that of the Earth or the Mars. In an
analogous manner, brain atlases are collections Extracellular compartment; Interstitial space;
of brain image data where each collection typi- Interstitial volume
cally spans across one (or more) whole brain.
This article primarily provides pointers to other
articles dealing with brain atlases. See the Cross- Definition
References/Related Terms below.
The extracellular space (ECS) comprises
a system of contiguous narrow spaces interposed
between the brain cells. It occupies approxi-
Cross-References mately 20 % of the brain volume. Mathematical
modeling quantifies the macroscopic parameters
▶ BrainMap of the ECS from experimental data and tests the
▶ Cell Centered Database hypotheses about the ECS microstructure.
▶ Collations of Connectivity Data on the
Macaque Brain (CoCoMac)
▶ Connectome, Drosophila Detailed Description
▶ Connectome, General
▶ Connectome, Mouse From a perspective of patient care, ECS forms the
▶ Human Connectome Project final segment of a delivery route for all nutrients
and drugs destined for the brain cells. Therapeu-
tic agents enter the ECS across the blood-brain
barrier at a capillary or from the cerebrospinal
fluid, or they are released directly into the ECS.
Brain Computer Interfaces The efficacy of treatment depends on compatibil-
ity between the drug or its carrier and the ECS
▶ Brain-Machine Interface and Neuroimaging microenvironment.
The ECS is also fundamentally important for
the brain function. Instead of merely filling the
void between the brain cells, as its name might
suggest, the ECS exerts a direct influence on
Brain Damage many essential processes in the brain, including
the diffusion of neurotransmitters and
▶ Brain Ischemia and Stroke neuromodulators mediating intercellular
Brain Extracellular Space: A Compartment for Intercellular Communication and Drug Delivery 435 B
communication, the distribution of ionic currents, from a point source, and their subsequent ECS
the neurotrophic effects, the cellular homeostasis, concentration is detected either by special ion-
and the transport of nutrients and metabolites. selective microelectrodes or by optical imaging
The ECS functionally complements neuronal and analyzed as a function of position and time.
and glial networks. In diffusion experiments employing a small extra- B
Once introduced into the ECS, substances are cellular probe molecules, two macroscopic param-
transported predominantly by diffusion because eters of the ECS structure are quantified, namely,
the ECS lacks any active transport mechanism. the volume fraction and the hindrance. Volume
Physiological effects of drugs and other sub- fraction a is a proportion of the tissue volume
stances diffusing in the ECS are governed by occupied by the ECS (a = VECS/Vtissue). Hin-
their concentrations and their arrival times, drance describes how much the diffusion process
which in turn depend on the ECS structure affect- in an ECS environment slows down in comparison
ing the diffusion. Biophysical parameters of the with an obstacle-free medium. It can be quantified
ECS have to be taken into account whenever either by a diffusion permeability parameter
quantitative understanding of any diffusion- y = D*/Dfree, where D* is an effective diffusion
mediated process is desired. coefficient in a complex medium and Dfree is a free
The ECS is filled with ionic solution and mac- diffusion coefficient (Hrabe et al. 2004), or by
romolecules of the extracellular matrix. Its struc- a tortuosity parameter l = y1/2 (Nicholson
ture is determined by the shape of brain cells 2001). Diffusion experiments also yield a direct
(neurons and glia), the width of the pores between measure of substance transport in the form of the
the cells, and the ECM (Fig. 1). These factors effective diffusion coefficient D*. In an isotropic
induce dynamic reversible changes of the ECS brain tissue probed by small molecules, a is about
during brain activity as well as profound and 0.20, y is about 0.39, and l is about 1.6 (Sykova
often irreversible changes in brain pathologies. and Nicholson 2008). The ECS is a dynamic struc-
The results of ECS research are thus important ture, and volume fraction and hindrance change,
both for understanding how the altered ECS struc- often independently, during brain activity and in
ture in neuropathological states disrupts the chem- disease (Hrabetova and Nicholson 2007). For
ical traffic of the brain and for designing effective example, during a stroke, a is as low as 0.07 and
drug delivery strategies for patients suffering from y decreases to 0.23 (l = 2.1), while in malignant
neurological disorders and brain tumors. brain tumors, a expands to 0.48 and y decreases to
0.32 (l = 1.8) (Sykova and Nicholson 2008).
Experimental Approaches Diffusion measurements employing a wide
Research of the ECS is challenging, in part, range of molecular sizes also allowed to estimate
because the ECS microstructure cannot be easily the width of extracellular pores in a living brain
visualized, the interstitial pores being below the tissue. One approach combined diffusion mea-
light microscopy resolution limit. Electron surements with the restricted diffusion theory to
microscopy studies provide indispensable infor- estimate the ECS pore width in the range of
mation on topology, but they require tissue 38 64 nm (Thorne and Nicholson 2006).
processing methods known to diminish the ECS Another approach combined diffusion measure-
volume (van Harreveld 1972; Sykova and Nichol- ments with the scaling theory developed in poly-
son 2008). Electron microscopy also cannot quan- mer physics to obtain an estimate of 31 nm (Xiao
tify parameters of the extracellular matrix or et al. 2008). These studies suggest that the indi-
visualize dynamic changes of the ECS structure. vidual ECS pores of a living brain tissue are much
Currently, the preferred experimental approach wider than the 15 20 nm typically obtained
uses diffusion measurements to indirectly obtain with electron microscopy. The width of extracel-
information on the ECS structure (Nicholson and lular pores is an essential parameter for drug
Sykova 1998; Hrabetova and Nicholson 2007). In delivery strategies because it provides an upper
a diffusion experiment, molecules are released limit for the size of particles (e.g., liposomal and
B 436 Brain Extracellular Space: A Compartment for Intercellular Communication and Drug Delivery
Brain Extracellular
Space: A Compartment
for Intercellular
Communication and
Drug Delivery,
Fig. 1 Schematic of brain
ECS. (a) Brain cells (gray)
surrounded by the ECS
(red). (b) Cell geometry
(gray), pore width (double
arrow), and extracellular
matrix (green) are key
components of the ECS
structure (Modified from
Hrabetova S, 2011)
viral drug carriers) that can uniformly penetrate environment and consequently do not provide
the ECS and access cells in an unrestricted information about the ECS features at
manner. a microscopic level. Very little is known about
the local structure of individual ECS pores, their
Computational Approaches organization, and statistical distribution of their
Computational approaches are indispensable in parameters. Monte Carlo simulations of molecu-
the ECS research. At a macroscopic level, an lar diffusion in simplified three-dimensional ECS
appropriate solution of the diffusion equation is models are used to test various hypotheses about
fitted to the data obtained in a diffusion experi- the microscopic ECS features. One important
ment to extract the ECS parameters. The most simulation tool is the MCell (Monte Carlo cell)
general form of the diffusion equation accounts package (Stiles and Bartol 2001; www.mcell.psc.
for an anisotropic environment and allows both edu). It was designed for biologically oriented
the volume fraction and the diffusion permeabil- users and uses a simple model description lan-
ity (a tensor quantity in this case) to become guage to specify simulation components such as
inhomogeneous (modified from Saghyan et al. the cellular surfaces and reaction mechanisms.
2012): Monte Carlo simulations have been used to
simulate diffusion of molecules in media

!
@c r , t Dfree h i constructed of simplified cell shapes that allow
! ! !
¼ ! ∇ a r u r ∇c r , t tight packing (Tao and Nicholson 2004; Hrabe
@t a r et al. 2004). They revealed that an extracellular
environment with convex cellular elements can-
!
s r ,t
! ! not have diffusion permeability y lower than 2/3,
þ ! k r c r , t ,
a r which is far above the typical experimental value.
This led to the development of a model of hin-

!
where c r , t is the concentration at position drance based on diffusion dwell times (Hrabe
! et al. 2004). More complex shapes were intro-
r ¼ ðx, y, zÞ and time t, Dfree is the diffusion
duced, which included concave cellular geome-
coefficient in obstacle-free medium, a is the
tries. The volume fraction a was divided into the
ECS volume fraction, u is the diffusion perme-
well-connected (ao) and dead-space (ads) parts.
ability tensor, s is the source strength, and k is the
A new relationship between the y and a was
nonspecific linear clearance approximating a loss
obtained, namely,
proportional to the concentration.
The ECS parameters extracted from diffusion
measurements are the result of macroscopic aver- y a0 a0
¼ ¼ :
aging of a geometrically complex microscopic y 0 a 0 þ a ds a
Brain Ischemia and Stroke 437 B
This model of the ECS composed of both related techniques. In: Michael AC, Borland LM
convex and concave cells can explain diffusional (eds) Electrochemical methods for neuroscience.
CRC Press, Taylor Francis Group, Boca Raton,
hindrance measured in the brain under normal as pp 167–204
well as some pathological conditions. It predicts Nicholson C (2001) Diffusion and related transport mech-
that about 40 % of the brain ECS under normal anisms in brain tissue. Rep Prog Phys 64:815–884 B
conditions is formed by dead spaces (cul-de- Nicholson C, Sykova E (1998) Extracellular space struc-
ture revealed by diffusion analysis. Trends Neurosci
sacs), increasing to 60 % in a stroke (Hrabetova 21:207–215
and Nicholson 2004). Saghyan A, Lewis DP, Hrabe J, Hrabetova S (2012) Extra-
cellular diffusion in laminar brain structures exempli-
fied by hippocampus. J Neurosci Methods
Conclusions 205:110–118
Stiles JR, Bartol TM (2001) Monte Carlo methods for
simulating realistic synaptic microphysiology using
Brain ECS facilitates the delivery of drugs,
MCell. In: Schutter ED (ed) Computational neurosci-
intercellular communication, nutrient, and ence: realistic modeling for experimentalists. CRC
metabolite trafficking. ECS research combines Press, London, pp 87–127
diffusion experiments in the brain with computa- Sykova E, Nicholson C (2008) Diffusion in brain extra-
cellular space. Physiol Rev 88:1277–1340
tional modeling approaches to ultimately charac-
Tao L, Nicholson C (2004) Maximum geometrical hin-
terize the microstructure of the ECS and drance to diffusion in brain extracellular space sur-
understand its regulation. The long-term goal of rounding uniformly spaced convex cells. J Theor Biol
the ECS research is to construct a realistic three- 229:59–68
Thorne RG, Nicholson C (2006) In vivo diffusion analysis
dimensional functional model of the brain ECS.
with quantum dots and dextrans predicts the width of
The results are important for understanding the brain extracellular space. Proc Natl Acad Sci U S
neuronal and glial behavior not only under phys- A 103:5567–5572
iological conditions but also in various neuro- van Harreveld A (1972) The extracellular space in the
vertebrate central nervous system. In: Bourne GH
pathological states which often affect the ECS
(ed) The structure and function of nervous tissue. Aca-
properties and result in disrupted traffic of ions, demic, New York, pp 447–511
signaling molecules, and nutrients. They are also Xiao F, Nicholson C, Hrabe J, Hrabetova S (2008) Diffu-
essential for the design of effective drug delivery sion of flexible random-coil dextran polymers mea-
sured in anisotropic brain extracellular space by
strategies in patients suffering from neurological
integrative optical imaging. Biophys J 95:1382–1392
disorders and brain tumors where the ECS struc-
ture is often significantly altered.
Acknowledgments This work was supported by the Brain Ischemia and Stroke
National Institute of Neurological Diseases And Stroke
of the National Institutes of Health under Award Number
R01NS047557. Samuel A. Neymotin1,2, Zachary Taxin2,
Ashutosh Mohan2,3 and Peter Lipton3
1
Department of Neurobiology, Yale School of
References Medicine, New Haven, CT, USA
2
Department of Physiology and Pharmacology,
Hrabe J, Hrabetova S, Segeth K (2004) A model of effec-
SUNY Downstate Medical Center, Brooklyn,
tive diffusion and tortuosity in the extracellular space
of the brain. Biophys J 87:1606–1617 NY, USA
3
Hrabetova S (2011) Extracellular diffusion in brain: dis- Department of Neuroscience, University of
tinct diffusion regimes at different spatial scales. dif- Wisconsin, Madison, WI, USA
fusion-fundamentals.org 16, pp 1–2
Hrabetova S, Nicholson C (2004) Contribution of dead-
space microdomains to tortuosity of brain extracellular Synonyms
space. Neurochem Int 45:467–477
Hrabetova S, Nicholson C (2007) Biophysical properties
of brain extracellular space explored with ion-selective Brain damage; Cerebrovascular disease; Oxygen
microelectrodes, integrative optical imaging and deprivation
B 438 Brain Ischemia and Stroke
Definition of signaling pathways that are potentially

involved in cell death, both intracellular and
Cerebral ischemia, which occurs during stroke, is intercellular, make it useful to study neuronal
insufficient blood flow to the brain and leads to ischemia using multiscale modeling, which can
loss of oxygen and glucose supply. This severely describe the effects across subcellular and cellu-
compromises cellular respiration, causing neuro- lar levels.
nal damage. Cells move towards cell death, Blood enters the brain via a complex arterial
which can be simplified as being apoptotic tree. When arteries or large arterioles are dam-
(controlled and “programmed”) or necrotic aged through occlusion or rupture, blood supply
(uncontrolled). Necroptopic cell death, which is is locally reduced, causing ischemic damage. The
a stereotypical necrotic death pathway, has also degree of damage is highly variable, depending
recently been characterized during stroke on the extent and duration of the insult. Gener-
(Christofferson and Yuan 2010). In addition to ally, we conceptualize two major separable
being both a single-cell and multicell phenome- regions: an ischemic core where blood flow is
non, ischemic insult also involves transcriptional essentially zero and an ischemic penumbra
changes to regulatory networks of genes. Taken where there is some residual blood flow. The
together, modeling the molecular and genetic ischemic penumbra is a region that is considered
aspects of ischemia is fundamental to understand- a target for rescue by appropriate treatment. At
ing how to modulate neuroprotective elements in normal body temperature, full ischemia for
the case of stroke, which is the primary therapeu- 3–5 min is generally lethal to neurons. By con-
tic approach in saving at-risk tissue. Computer- trast, occurrence of long-lasting damage in the
based cellular modeling can assist in understand- penumbra may require several hours of ischemia.
ing the molecular changes involved, the decision The cellular and molecular changes in these
and control points for various cell fates, and the regions have markedly different presentations
potential for recovery with partial damage as well (Lipton 1999). Damage in the penumbra will be
as help define the scope of the genetic and regu- highly variable across multiple timescales. The
latory network changes that occur. duration of oxygen deprivation that triggers the
ischemic cascade and the extent of damage will
likely depend upon genetic and environmental
Detailed Description factors in the individual and in the precise
types of cells involved. In particular, it is hypoth-
The mechanisms by which cells accumulate dam- esized that larger cells, such as Purkinje cells and
age, signal damage to neighboring cells, and cortical pyramidal cells, may be particularly
eventually die all remain unclear. The geometric vulnerable.
complexity of neurons makes it difficult to eluci-
date single-cell diffusive mechanisms in the
ischemic signaling cascade. Further complexity Nutrient Delivery
occurs due to difficulties in disconnecting effects
at the cellular level from effects at the tissue Ultimately, all ischemia results from loss of sub-
level, as well as effects due the synaptic intercon- strates for the cell to generate adenosine triphos-
nectivity of the neurons, which can lead to net- phate (ATP), its primary energy currency.
work cascades in dying tissue. Compounding Computer models of nutrient delivery at the
these difficulties is the fact that many signaling level of blood and of extracellular tissue are
cascades are heavily interdependent on non- therefore useful in providing the setting for
apoptotic, homeostatic mechanisms and have effects on the individual cell. This nutrient
multiple nonlinear dependencies. The intricacies modeling allows us to better define molecular
Brain Ischemia and Stroke 439 B
differences in types of ischemia that will be pro- via calcium entry due to depolarized membrane
duced. Simplified ischemia models at this level potential.
represent brain tissue as a set of voxels which Loss of ion homeostasis may precipitate the
exchange O2, with influx and outflux of blood accumulation of reactive oxygen species and
between the regions and oxygen consumption intracellular calcium dysregulation, both of B
integrated at each simulation time-step (Hudetz which are primary “perpetrators” of cell death
et al. 1982). In a paper published in 2002, Duval (Lipton 1999). Models of this process are similar
et al. (2002) modeled 4 parameters: cerebral blood to the early work done by Magnus and Keizer in
flow, rate of oxygen extraction from blood, meta- pancreatic b-cells, another electrically excitable
bolic rate of oxygen use by tissue, and the apparent cell type (Keizer and Magnus 1989). This model
diffusion coefficient of water. A “survival delay” demonstrated that increased glucose concentra-
variable was also used to quantify the decay of tion could enhance burst spiking, generating
tissue in the penumbra from functional to salvage- elevated intracellular Ca++ and excitotoxicity.
able to necrotic. This modeling suggested that the Later work by Magnus and Keizer focused on
penumbra could take on two major forms: ische- developing more robust models of mitochondrial
mic and edematous. Ischemic penumbra tissue Ca++ signaling (Magnus and Keizer 1998).
was metabolically altered and generally not sal- Similarly, in neuronal cells, loss of ATP due to
vageable, whereas the edematous penumbra tissue poor blood flow triggers a cascade of reactions
appeared to retain greater recovery potential. This leading to apoptosis or necrosis. Four major
model represented a first step towards clarifying mechanisms are recognized (Lipton 1999):
how extracellular nutrient parameters might affect 1. Influx of Ca++ ions through NMDA receptors
the generation of penumbral regions. Other and voltage-gated Ca++ channels.
models have included blood flow parameters in 2. Increased intracellular Na+ causes glutamate
molecular models as a way of developing more efflux, which also eventually leads to
robust models of cellular networks (Dronne et al. increased intracellular Ca++.
2006). 3. Decreased ATP/ADP ratio producing a fall in
intracellular pH.
4. Production of free radicals, particularly near
Initiation of the Ischemic Cascade mitochondrial membranes.
Ionic dysfunction is a hallmark of both
After vascular damage, neuronal uptake of O2 necrotic and apoptotic cellular changes. This dys-
and glucose is reduced, leading to a sequence function is primarily manifest via abnormal Ca++
of metabolic and molecular changes. An early fluctuations within the cell. The generation of
result of nutrient loss is a reduction in ATP. The free radicals and inability to manage pH may
biochemistry of energy production in brain tis- also be tied to Ca++ dysregulation via Ca++ -
sue utilizes the same reactions that are used in binding control proteins such as calmodulin,
other cells of the body. However, neurons are calbindin, and calpain (Saftenku and Friel
run at high metabolic states in order to maintain 2012). The products from these early reactions
the electrochemical gradient that enables neu- are heavily involved in the cell’s final fate.
rons to fire action potentials. They have addi- Due to the importance of Ca++ and Ca++ buff-
tional ATP delivery challenges associated with ering in determining cell fate, much attention in
the presence of long, thin processes with large ischemia research has focused on modeling this
surface to volume ratios. With disruption in ATP ion. Normally, there are very low resting levels of
production, plasma membrane ionic gradients Ca++ in the neuron. This requires stochastic
will begin to decay, preventing neuronal electri- modeling, because small concentrations of ions
cal function and allowing excitotoxic damage cannot be identified as bulk units, which can be
B 440 Brain Ischemia and Stroke
handled deterministically. Such a deterministic wave oscillation interactions with other elements
approach (e.g., rate laws) relies on the Law of in the single cell. Such an approach, however,
Large Numbers; many aspects of Ca++ signaling does not link Ca++ waves to network pathological
do not obey this law. It has been suggested, in processes within and without the cell. Recently,
fact, that the cell uses chemistry as Kazemzadeh et al. used a combined Boolean and
a computational tool and is able to harness sto- ODE approach to model network regulatory
chastic Ca++ fluctuations to generate robust intra- effects of apoptosis in a modified (c.f. “human-
cellular signaling networks. Ischemic cascades ized”) yeast model. Using databases to generate
therefore involve stochastic events as well as robust Boolean Networks that could then be
deterministic events, and both involve consider- solved at time-steps enabled them to calculate
able species diffusion through neurons and extra- steady states of complex apoptotic networks and
cellular tissue space. Modeling allows us to to test the effects of genetic knockouts and vari-
explore Ca++ fluctuations and species diffusion able ischemic insults (Kazemzadeh et al. 2012).
at a very small-scale level. Another set of recent experiments by McDermott
et al. were aimed at using BN formalism to study
regulatory gene networks in response to ischemic
Use of Boolean Networks in Ischemia insult. They modeled clusters of functionally
Modeling related gene elements in steady state and then
perturbed this system by using various types of
The changes in nutrients, in ATP, in Ca++, preconditioning insults, including mild ischemia,
etc. detailed above are only the beginning of and a later more serious ischemic insult. Results
a large array of interconnecting molecular between the various initial states were compared,
changes. One way to make sense of such and the functional significance of genetic contribu-
a complex and large network has been to use tion to neuroprotection was analyzed by looking
a Boolean Network (BN) formalism to model at the robustness of cluster response to
system dynamics. By conceptualizing the mole- preconditioning (McDermott et al. 2012). Most of
cules in the signaling networks as having the results from the computer model were con-
a discrete set of states (ON/OFF or activated/ firmed by previously published experimental data.
repressed), interaction rules can be defined for BN formalism is a reductionist view of network
a particular set of reactions. Mai et al. used BNs signaling, whereas simultaneous ODE solving at
to model aspects of apoptosis (Mai and Liu discrete time-steps gives a larger picture of intra-
2009). They found downstream redundancies, cellular signaling. Attempts to link these two
feedback loops, points of control, and steps lead- modeling methods in studying apoptosis and ische-
ing towards irreversible commitment to this end- mic cascades have been recently gaining popular-
point. Although this approach simplifies reaction ity in computational neuroscience as well as
dynamics to a 0/1 decision, it can provide valu- systems biology (Bogda et al. 2013).
able information on system properties and points
of modulation within the apoptotic pathway.
Therefore, the type of information that can be Future Considerations
gleaned from a model depends very much on how
the model is designed. One alternative to BN Computer modeling affords several advantages
modeling is to use ordinary differential equation in the study of ischemia. It allows us to
(ODE) kinetics to model Ca++ wave propagation accurately measure many reactants simultaneously
and reactivity with a higher degree of precision. in exceedingly complex environments. We can
Within this context, one can also model Ca++ link top-down and bottom-up models of ischemia
wave propagation and define parameters for to create robust multicellular models. We have also
Brain-Machine Interface and Neuroimaging 441 B
begun to understand how genetic regulation
and preconditioning effects may confer Brain-Machine Interface and
neuroprotection. And, importantly, we can identify Neuroimaging
control points for cascade modulation, which are
appropriate targets for therapeutic intervention. Mitsuo Kawato B
Such research will likely provide proof of concept Computational Neuroscience Laboratories,
and system extrapolation for the development of ATR Brain Information Communication
new drugs to encourage recovery or enhance the Research Laboratory Group, Kyoto, Japan
duration of recoverability in the penumbra.
Synonyms
References
Brain computer interfaces; Neural prosthetics
Bogda MN, Hat B, Kocha Czyk M, Lipniacki T (2013)
Levels of pro-apoptotic regulator Bad and anti-
apoptotic regulator Bcl-xL determine the type of the
apoptotic logic gate. BMC Syst Biol 7(1):67
Definition
Christofferson DE, Yuan J (2010) Necroptosis as an alter-
native form of programmed cell death. Curr Opm Cell Brain machine interfaces (BMIs) directly connect
Biol 22(2):263–268 brain activity and devices outside the brain for
Dronne MA, Boissel JP, Grenier E (2006) A mathematical
such purposes as sensory substitution, motor
model of ion movements in grey matter during
a stroke. J Theor Biol 240(4):599–615 compensation and enhancement, neuroreh-
Duval V, Chabaud S, Girard P, Cucherat M, Hommel M, abilitation, suppression of maladaptive neural
Boissel JP (2002) Physiologically based model of circuits, and treatment for brain lesions and neu-
acute ischemic stroke. J Cereb Blood Flow Metab
rological and psychiatric disorders. A BMI usu-
22(8):1010–1018
Hudetz AG, Halsey JH, Horton CR, Conger KA, Reneau ally consists of some of the following electrical,
DD (1982) Mathematical simulation of cerebral blood mechanical, and computing elements: a brain
flow in focal ischemia. Stroke 13(5):693–700 activity measurement system (electrodes, ampli-
Kazemzadeh L, Cvijovic M, Petranovic D (2012) Boolean
model of yeast apoptosis as a tool to study yeast and
fier, AD converter, etc.), a computer for decoding
human apoptotic regulations. Front Physiol 3:446 brain information and/or controlling stimuli to
Keizer J, Magnus G (1989) ATP-sensitive potassium the brain, an effector system (prosthetic limb,
channel and bursting in the pancreatic beta cell. rehabilitation robot, computer cursor, etc.), and
A theoretical study. Biophys J 56(2):229–242
a neural stimulation system.
Lipton P (1999) Ischemic cell death in brain neurons.
Physiol Rev 79(4):1431–1568
Magnus G, Keizer J (1998) Model of beta-cell mitochon-
drial calcium handling and electrical activity. II. Mito- Detailed Description
chondrial variables. Am J Physiol 274(4 Pt 1):
C1174–C1184
Mai Z, Liu H (2009) Boolean network-based analysis of BMIs can be classified from at least the following
the apoptosis network: irreversible apoptosis and sta- three viewpoints. The first classification concerns
ble surviving. J Theor Biol 259(4):760–769 which brain function is substituted, compensated,
McDermott JE, Jarman K, Taylor R, Lancaster M,
enhanced, or cured by BMI. Only three very basic
Shankaran H, Vartanian KB, Stevens SL, Stenzel-
Poore MP, Sanfilippo A (2012) Modeling dynamic functions are listed here: sensory processing, cen-
regulatory processes in stroke. PLoS Comput Biol tral decision-making, and motor control,
8(10):e1002722 which includes communication. The second
Saftenku EE, Friel DD (2012) Chapter 26: Combined
addresses the BMI’s objective: the substitution,
computational and experimental approaches to under-
standing the ca2+ regulatory network in neurons. In: compensation, and enhancement of lost or nor-
Calcium signaling, Springer 569–601 mal brain functions, or the cure, recovery, and
B 442 Brain-Machine Interface and Neuroimaging
rehabilitation of degraded brain functions. The concepts was demonstrated around 2000 in
third classification of BMI is whether it is inva- rodents and monkeys by pioneers such as
sive, i.e., whether it uses thin wire electrodes, Johan Wessberg (Wessberg et al. 2000),
semiconductor electrodes, electrocorticograms Andrew Schwartz (Taylor et al. 2002), and
(ECoG; EEGs directly attached to the cerebral John Donoghue, and the technique was trans-
cortex surface), deep brain electrodes for brain lated into human patients around 2005
activity measurement and/or electrical stimula- (Hochberg et al. 2012). Invasive electrodes
tion; or whether it is noninvasive, using scalp measure brain activity (several tens to a few
EEGs, fMRIs, NIRSs, and so on. Theoretically, hundred neurons), a decoder extracts the
we have 3 2 2 = 12 different BMI classes, motor command information from neural sig-
but only a few have become practical. Below we nals, and finally an effector system is con-
explain some of the roughly two thirds that are trolled by this extracted movement
actively being developed. information. Users observe the results of the
1. An artificial sensory organ or sensory BMI is effector actions, and a loop is closed
an invasive system of sensory-function substi- connecting the brain, the measurement sys-
tution for handicapped people. Artificial sen- tem, the decoder and the effector, and the
sors, such as microphones or CCD cameras, signals are returned to the brain. This is one
measure auditory or visual signals and are of the most actively studied classes of BMI in
transformed as inputs to a neural stimulation the past 15 years. Subdural ECoG electrodes
system with multiple electrodes that electri- have recently attracted attention because they
cally stimulate the peripheral sensory nerves are more stable than the intracortical elec-
or the central nervous system. The oldest, trodes that are inevitably associated with
most successful example, with 30 years of scars. These less invasive BIM systems,
practical usage, is the cochlear implant. Arti- which are not invasive to the brain itself,
ficial retinas have also nearly become practi- might be quite practical because decoding of
cal. Visual cortex stimulation (artificial the three-dimensional arm trajectory of mon-
vision) and artificial vestibular systems are keys using ECoG did not significantly deteri-
under active research and development. orate for a year or two (Chao et al. 2010), and
2. A Brain Computer Interface (BCI) resembles ECoG-BMI can control a multi-joint robot
a BMI in its wider definition, but in its arm and a hand system with exactly the same
narrower usage, it is a noninvasive BMI used decoding parameters for over a week in
to compensate for lost motor functions. An humans (Yanagisawa et al. 2012).
EEG-BCI uses such varied, features as senso- 4. An EEG-BMI rehabilitation system is
rimotor cortex m rhythm, slow cortical poten- a noninvasive BMI used for the recovery of
tial, and the P300 to decode movement motor functions in stroke patients. This has
parameters including gaze direction and to opened up a quite promising newly-emerging
control an interface, including controlling BMI application field. With conventional
a computer mouse or selecting a character for rehabilitation protocols, useful (i.e., grasp-
communication (Birbaumer et al. 1999). Some able) hands are restored in only 15 % of stroke
BCIs are practical and/or commercially avail- patients but that figure increases to around
able. As rare examples of nonmedical BMI 80 % with BMI rehabilitation. Pioneers
applications, inexpensive examples have Jun-ichi Ushiba and Meigen Liu recorded sen-
been developed for gaming. sorimotor cortex m rhythms from severe stroke
3. Neural prosthetics or BMIs in a narrow sense patients whose extensor muscles could not
are an invasive BMIs used to compensate for even generate EMG, and decoded their move-
lost or damaged motor functions. Proof of the ment intention based on the decrease in the m
Brain-Machine Interface and Neuroimaging 443 B
power: event-related EEG desynchronization. users (Fetz 1969) can be understood from com-
When the decoder output was successful, both putational algorithms derived from supervised
visual and kinesthetic feedback was given to and reinforcement learning theories.
patients via a robotic splint that extended
their paralyzed wrists and fingers (Shindo B
et al. 2011).
5. fMRI-decoded neurofeedback (DecNef) can
References
be used as a noninvasive BMI system for
Birbaumer N, Ghanayim N, Hinterberger T, Iversen I,
treating abnormal central decision-making Kotchoubey B, K€ ubler A, Perelmouter J, Taub E,
functions in psychiatric disorders (Kawato Flor H (1999) A spelling device for the paralysed.
WFSBP Plenary talk 2013). Shibata, Nature 398(6725):297–298
Chao ZC, Nagasaka Y, Fujii N (2010) Long-term asyn-
Watanabe, Sasaki, and Kawato developed
chronous decoding of arm motion using electrocorti-
DecNef, by means of which spatial activity cographic signals in monkeys. Front Neuroeng 3:3
patterns in a specific brain area can be nonin- Fetz EE (1969) Operant conditioning of cortical unit
vasively controlled through real-time fMRI activity. Science 163(3870):955–958
Georgopoulos AP, Kalaska JF, Caminiti R, Massey JT
neurofeedback (Shibata et al. 2011). First, (1982) On the relations between the direction of
a decoder is obtained that classifies BOLD two-dimensional arm movements and cell discharge
signal patterns by visual attributes, emotional in primate motor cortex. J Neurosci 2(11):1527–1537
states or normal and pathological states in Hochberg LR, Bacher D, Jarosiewicz B, Masse NY,
Simeral JD, Vogel J, Haddadin S, Liu J, Cash SS,
psychiatric disorders. In the induction stage,
van der Smagt P, Donoghue JP (2012) Reach and
volunteers unconsciously control their brain grasp by people with tetraplegia using a neurally con-
activity so that a desirable pattern is produced trolled robotic arm. Nature 485:372–375
while being guided by a real-time reward sig- Kawato M (2013) Plenary lecture: computational neuro-
science approach to biomarkers and treatments for
nal computed as the decoder output. Thus,
psychiatric diseases. In: WFSBP 2013, Kyoto, 27 June
three important elements of DecNef are Miyawaki Y, Uchida H, Yamashita O, Sato MA,
decoding of information from fMRI brain Morito Y, Tanabe HC, Sadato N, Kamitani Y (2008)
activity, real-time neurofeedback of decoded Visual image reconstruction from human brain activity
using a combination of multiscale local image
information, and unconscious reinforcement decoders. Neuron 60(5):915–929
learning by participants. Shibata K, Watanabe T, Sasaki Y, Kawato M (2011) Per-
In most BMIs, except the sensory substitution ceptual learning incepted by decoded fMRI
BMI, necessary brain information is first neurofeedback without stimulus presentation. Science
334(6061):1413–1415
extracted from brain activity signals using a
Shindo K, Kawashima K, Ushiba J, Ohta N, Ito M, Ota T,
decoder. This information is transformed and Kimura A, Liu M (2011) Effects of neurofeedback
fed back to users in some sensory modality training with an electroencephalogram-based brain
with/without reward signals, and finally BMI computer interface for hand paralysis in patients with
chronic stroke. J Rehabil Med 43(10):951–957
users learn to change their brain networks based
Taylor DM, Tillery SI, Schwartz AB (2002) Direct corti-
on synaptic plasticity. In these three steps, com- cal control of 3D neuroprosthetic devices. Science
putational neuroscience plays a critical role. For 296(5574):1829–1832
example, computational motor control models Wessberg J, Stambaugh CR, Kralik JD, Beck PD,
Laubach M, Chapin JK, Kim J, Biggs SJ, Srinivasan
outline the neural prosthetics paradigm to MA, Nicolelis MA (2000) Real-time prediction of
record from many neurons, and use their firing hand trajectory by ensembles of cortical neurons in
frequencies to decode movement kinematics primates. Nature 408(6810):361–365
(Georgopoulos et al. 1982). Computational Yanagisawa T, Hirata M, Saitoh Y, Kishima H,
Matsushita K, Goto T, Fukuma R, Yokoi H,
models also play important roles in non-invasive
Kamitani Y, Yoshimine T (2012) Electrocorti-
decoding algorithms (Miyawaki et al. 2008). cographic control of a prosthetic arm in paralyzed
Finally, neural operant conditioning in BMI patients. Ann Neurol 71(3):353–361
B 444 Brain Stem Auditory Evoked Potentials
previously bookmarked. Many also enter

Brain Stem Auditory Evoked through other websites that focus on a specific
Potentials aspect of neuroanatomy, such as the brain of
a particular species or the inputs and outputs,
▶ Auditory Evoked Brainstem Responses cell types, or gene expression in structures. Such
sites link to BrainInfo to give users one-click
access to a wider range of information on spe-
cific neural structures (Fig. 1). A guide to use of
Brain-Computer Interfaces the Portal is found at Bowden and Dubach
(2005).
▶ Brain-Machine Interface: Overview
BrainInfo, in turn, links out to pages in some
▶ Functional Neuroscience: Cortical Control of
50 other websites to provide users access to
Limb Prosthesis
information and illustrations of some 2,500
structures and superstructures in the CNS of
the four species most studied by neuroscien-
BrainInfo tists: the human, macaque, rat, and mouse
(Fig. 2, SPECIES). With more than a thousand
Douglas M. Bowden distinct cell groups identified in the mammalian
Department of Psychiatry and Behavioral brain, mastering the neuroanatomy is
Sciences, School of Medicine, University of a challenge. The challenge is magnified several
Washington, Seattle, WA, USA times by variability across species, methods of
identification, and nomenclature (Bowden and
Martin 1997).
Definition The Portal is designed particularly for neu-
roscientists who work in disciplines other than
BrainInfo can be thought of as a Web-based hand- anatomy. It is for those whose knowledge of
book of neuroanatomy. It is an authoritative neuroanatomy is limited and who need rapid
resource for computational neuroscientists and access to information that is scattered
others who need a quick in-depth orientation to under a variety of names and across numerous
specific structures in the central nervous system publications and websites. The computational
(CNS). BrainInfo has three components. The neuroscientist who comes upon an unfamiliar
BrainInfo Portal provides extensive descriptive structure in a model system, or the psychiatrist
information and images of some 2,500 brain struc- who encounters an unfamiliar structure in
tures; NeuroNames is a neuroanatomical ontology reading about the neural mechanism of
that provides standard English names for the struc- a classical disorder, can Search by Name, and
tures plus 13,000 synonyms in eight languages; and BrainInfo will display the Central Directory to
NeuroMaps provides brain atlases of the macaque a wealth of text and images of the structure
and mouse together with tools for mapping data (Fig. 2).
and processing images for publication. The geneticist or molecular biologist who
finds gene expression in an unfamiliar brain
area can Search Atlases, find the area, and
Detailed Description click it, and BrainInfo will display the Central
Directory for the structure together with its
The BrainInfo Portal (http://braininfo.org) standard name, a brief description, illustrations,
Users enter the Portal by several routes. Most kinds of neurons it contains, inputs, outputs, and
enter via Google or via a page that they have so forth.
BrainInfo 445 B
BrainInfo, Fig. 1 Visitors enter BrainInfo either Network (NIF), the Mouse Brain Library (MBL), the Scal-
directly, through Google, or through links from many able Brain Atlas, and Wikipedia
other websites including the Neuroscience Information
NeuroNames (http://braininfo.rprc. All anatomical information in BrainInfo

washington.edu/Nnont.aspx) is indexed according to the standard nomencla-
Neuroanatomical terminology is unusually con- ture of the NeuroNames ontology. The
fusing. Beginning in the late 1800s, scientists in ontology defines the relations between
multiple countries, studying the brains of many 15,500 names (character strings), for
species, and publishing in multiple languages, ~2,500 concepts (definitions) of the entities
contributed an overwhelming array of redundant (structures) they represent in different models
and overlapping terms to the vocabulary (hierarchies or networks) of CNS organization
(Bowden and Martin 1997). The typical brain (Fig. 3).
structure has an average of six English and The NeuroNames ontology allows visitors
Latin names, many of which appear in the text at the input interface of BrainInfo to Search
and labels of today’s publications and websites. by Name using not only the standard term but
The standard nomenclature of NeuroNames any synonym or homonym for the structure
reflects an attempt to relate every name of of interest. It includes all English terms for
a CNS structure in the neuroscience literature to all structures and names for the most com-
a defined structure and to select one of those mon ~1,000 structures in seven other
terms as its standard name. languages: French, German, Indonesian,
B 446 BrainInfo
BrainInfo, Fig. 2 The BrainInfo Portal links out to pages relations and internal features of CNS structures in the
deep in some 50 other neuroscience websites that illustrate human, macaque, rat, and mouse
most accurately, clearly, and thoroughly the anatomical
Italian, Latin, Russian, and Spanish. On the NeuroMaps (http://braininfo.rprc.

output side, when BrainInfo sends a visitor to washington.edu/MapViewData.aspx)
a website that uses a different name than the Neuroanatomy is the common ground of the var-
NeuroNames standard, a message box in the ious neuroscientific disciplines. All data, whether
sidebar informs the visitor, “Look for [the of gene expression, functional MRI activity, or
local name].” neuronal firing patterns, are related through the
Developers of other neuroscience websites use site in the brain from which they are recorded.
NeuroNames to index databases in the standard NeuroMaps provides standard atlases of the
nomenclature and to allow visitors to enter their macaque and mouse brains to which investigators
websites using any synonym for structures can map data of all modalities for comparison,
(Bowden et al. 2012a). The ontology is freely projection, and publication.
downloadable from BrainInfo in either XML or Each NeuroMaps atlas is based on a high-
Excel format. Both versions include the web resolution MRI (magnetic resonance image) ori-
address (URL) of BrainInfo’s Central Directory ented according to standard stereotaxic practice
for each structure, which allows developers to for the species; the left and right hemispheres are
link their visitors to BrainInfo for more informa- identical to ease comparison of data mapped to
tion (Fig. 1). the two hemispheres (Bowden et al. 2011).
BrainInfo 447 B
NeuroMaps tools allow the investigator to slice,
tilt, and rotate the atlas MRI to match the plane of
section of tissue bearing the data. They allow the
designation of unlimited pairs of registration
points to match landmarks in the experimental B
image to landmarks in the atlas. The investigator
draws boundaries around data areas in the exper-
imental image. A mapping tool automatically
transfers them by a nonlinear process to
BrainInfo, Fig. 3 NeuroNames is an informatics tool for corresponding locations in the atlas. Image
resolving neuroanatomical ambiguities (Bowden et al. processing tools enable the investigator to size,
2012b). In neuroanatomy, the same name (left panel) can color, label, and adjust the resolution (dpi = dots
represent multiple concepts (the homonym problem); and per inch) of the merged image to match specifi-
the same structural concept, or model composed of structural
concepts (central panel), can have multiple names (the syn- cations of the journal where it is to be published
onyms problem). A given model includes multiple concepts, (Fig. 4).
and the same concept can occur in multiple models. Different The NeuroMaps image processor can display
scientists can hold different concepts of the same entity (right data in multiple formats. The cross-sectional dis-
panel) and can model a complex entity in different ways. The
NeuroNames ontology resolves naming and conceptual ambi- play of areal data (Fig. 4; BrainInfo 2013a) is
guities by assigning a unique standard name to each concept suitable for mapping fMRI activation, gene
and specifying the relations of all other names (as synonyms expression, lesions, injection sites, and projection
or homonyms) to the standard name and its concept areas. Similar cross-sectional displays are used
BrainInfo, Fig. 4 The NeuroMaps image processor cortex of the mouse in the Allen Brain Atlas (AIBS
enables users to adjust the size, colors, labels, and resolu- 2008) have been mapped to the NeuroMaps symmetrical
tion of images for projection and publication. Here, the version of the Waxholm mouse brain (Bowden et al. 2011)
boundaries of cytoarchitectural areas in the cerebral
B 448 Brain-Machine Interfaces
BrainInfo, Fig. 5 NeuroMaps illustrations: (a) dopami- cortex that produces a top-down attentional deficit in
nergic cells at the gray-white boundary of prefrontal cor- macaques (Rossi et al. 2007)
tex (Dubach 1994) and (b) a lesion of lateral prefrontal
for mapping point data: stimulation sites, record- Oct 2012. http://www.wanprc.org/braininfo-images/
ing sites, and individual neurons stained for spe- nndevelopers.pdf. Accessed 9 Oct 2013
Bowden DM, Song E, Kosheleva J, Dubach MF (2012b)
cific enzymes, receptors, or neurochemistry NeuroNames: an ontology for the BrainInfo portal to
(Fig. 5a). The three-dimensional MRI model is neuroscience on the web. Neuroinformatics 10:97–114
suitable for mapping surface areas, such as archi- BrainInfo (2013a) NeuroMaps mapper: mouse atlas. Uni-
tectonic areas and cortical lesions (Fig. 5b; versity of Washington, Seattle. http://neuromaps.
braininfo.org/cgi-bin/NeuroMaps.py?species=Mouse.
BrainInfo 2013b). Accessed 9 Oct 2013
BrainInfo (2013b) NeuroMaps mapper: macaque atlas.
University of Washington, Seattle. http://neuromaps.
braininfo.org/cgi-bin/NeuroMaps.py?species=Macaque.
References Accessed 9 Oct 2013
Dubach M (1994) Telencephalic dopamine cells in mon-
keys, humans, and rats. In: Smeets WJAJ, Reiner
AIBS (2008) Reference atlas version 1 (mouse). Allen A (eds) Phylogeny and development of catecholamine
Institute for Brain Science, Seattle. http://mouse.bra systems in the CNS of vertebrates. Cambridge Univer-
in-map.org/experiment/siv?id=undefined&imageId sity Press, Cambridge, pp 273–292
=102162306&imageType=atlas&initImage=atlas&s Rossi AF, Bichot NP, Desimone R, Ungerleider LG
howSubImage=y&contrast=0.5,0.5,0,255,4. (2007) Top down attentional deficits in macaques
Accessed 9 Oct 2013 with lesions of lateral prefrontal cortex. J Neurosci
Bowden DM, Dubach M (2005) BrainInfo: a portal 27:11306–11314
to neuroanatomy on the web. BrainInfo document.
http://braininfo.rprc.washington.edu/pdf/For%20Stein-
busch%20on%20Web%20final.pdf. Accessed 8 Oct
2013
Bowden DM, Martin RF (1997) A digital Rosetta stone for
primate brain terminology. In: Bloom FE, Bjorklund Brain-Machine Interfaces
A, Hokfelt T (eds) Handbook of chemical neuroanat-
omy. The primate nervous system, vol 13, part 1.
▶ Functional Neuroscience: Cortical Control of
Elsevier Science, Amsterdam, pp 1–35
Bowden DM, Johnson GA, Zaborsky L, Green WDK, Limb Prosthesis
Moore E, Badea A, Dubach MF, Bookstein FL
(2011) A symmetrical Waxholm canonical mouse
brain for NeuroMaps. J Neurosci Methods
195:170–175
Bowden DM, Song E, Dubach MF (2012a)
Brain-Machine-Brain Interface
NeuroNames for developers. Poster GGG-10, Society
for neuroscience annual meeting, New Orleans, 13–17 ▶ Recurrent Brain-Computer Interfaces
BrainMap 449 B
associated with cognition, perception, and action.
BrainMap Consequently, the volume of functional imaging
publications has been increasing exponentially
Michael C. Riedel1, Kimberly Ray1, since in the early 1990s. Introduced over two
P. Mickle Fox1, Angela M. Uecker1, decades ago, the BrainMap database (http:// B
Simon B. Eickhoff2, Peter T. Fox1 and brainmap.org) was created to archive the results
Angela Marie Richmond Laird3 of the growing number of neuroimaging publica-
1
Research Imaging Institute, University of Texas tions in a manner that leads to optimum accessi-
Health Science Center, San Antonio, TX, USA bility (Fox et al. 1994; Laird et al. 2005, 2009).
2
Institute of Neuroscience and Medicine The BrainMap Project supports a database of
(INM-1), Research Centre J€ulich, and Institute functional neuroimaging experiments, as well as
for Clinical Neuroscience and Medical voxel-based morphometry database (VBM). Cur-
Psychology, Heinrich-Heine University, rently (as of October 2013), the BrainMap Func-
D€ usseldorf, Germany tional Database includes 2,336 publications,
3
Department of Physics, Florida International which reported the results of 11,103 functional
University, Miami, FL, USA activation experiments across a total of 45,188
subjects, while the BrainMap VBM Database
includes 796 publications, which reported the
Definition results of 2,444 structural neuroimaging experi-
ments across a total of 55,222 subjects.
BrainMap is an online community database that
archives the results of published functional and Standardized Coordinates
anatomical neuroimaging experiments and their Neuroimaging experiments report locations of
associated metadata (http://brainmap.org). activation or deactivation as three-dimensional
Experimental results and metadata are archived coordinates (x,y,z) in standard space. This prac-
into the database according to an expert-refined tice allows for more precise quantitative locali-
taxonomy developed to describe the experimen- zation of activation rather than qualitatively
tal procedure employed in each publication describing the structure associated with the acti-
within the database. Users may input publications vation. Due to the variability between each sub-
and their associated neuroimaging results to the jects’ brain morphometry, activation coordinates
database using the Scribe application (http:// are transformed into a common coordinate sys-
brainmap.org/scribe) or search the database for tem for comparison across multiple subjects. The
publications of interest by applying filters in the two best-known reference spaces for the human
Sleuth application (http://brainmap.org/sleuth). brain were developed by Talairach and Tournoux
Coordinates of activation resulting from filtered (Talairach and Tournoux 1988) and the Montreal
searches may be downloaded and exported for Neurological Institute (MNI; Collins et al. 1994).
further meta-analysis using the GingerALE appli-
cation (http://brainmap.org/ale), which uses the BrainMap Coding Scheme
activation likelihood estimation algorithm (ALE; A hierarchical taxonomy (Fig. 1) was developed
Turkeltaub et al. 2002; Eickhoff et al. 2009). for BrainMap such that the results of publications
could be coded into the databases according to
information about the subject population, task,
Detailed Description imaging modality, or behavioral conditions asso-
ciated with the experiment. The cataloguing of
The advent of magnetic resonance imaging and information in this manner not only provides
positron emission tomography to acquire func- consistent labeling of each experiment but also
tional data of the human brain revolutionized promotes rapid retrieval from the database. The
the way scientists examined neural systems publication is first coded according to citation
B 450 BrainMap
BrainMap, Fig. 1 Hierarchy of BrainMap coding encoded. The condition level is composed of information
scheme. Each publication consists of three levels of about the behavioral stimulus, instructions, and response.
descriptors. At the highest level of each publication, gen- The experiment level includes metadata fields that
eral information about citation information, subjects, and describe the contrasted brain images, with a tertiary level
summaries of the experimental design and results are for archiving coordinate locations
information: publication journal and year, insti- identified in Sleuth, it is exported to GingerALE
tution, and authors. Additional metadata fields to perform a coordinate-based meta-analysis
include subject description (e.g., age, gender, using the ALE algorithm. The ALE score deter-
handedness, ethnicity) and diagnosis (e.g., mines the probability that a given location was
healthy or diseased), which are all fully described reported active across the experiments selected.
at the publication level. The published three- A permutation test has been implemented to test
dimensional (x,y,z) coordinates extracted from for statistical significance. The ALE method yields
statistical parametric images are archived from locations of consistent activation patterns across
each experiment. Individual experiments are a set of similar neuroimaging experiments, which
classified according to paradigm class, which offers many benefits to simply reviewing the
describes the specific task being performed, and images on an experiment-by-experiment basis.
behavioral domain, which describes the cognitive
construct being studied. The statistical contrasts
of interest (e.g., “Task–Rest”) producing the sta- References
tistical parametric image are also archived at the
experiment level (Fox and Lancaster 2002). Collins DL, Neelin P, Peters TM, Evans AE (1994) Auto-
When published coordinates are entered into the matic 3D intersubject registration of MR volumetric
data in standardized Talairach space. J Comput Assist
BrainMap database, MNI coordinates are
Tomogr 18:192–205
transformed into Talairach space (Lancaster Eickhoff SB, Laird AR, Grefkes C, Wang LE, Zilles K,
et al. 2007), and vice versa, providing the option Fox PT (2009) Coordinate-based activation likelihood
for coordinate extraction in either coordinate sys- estimation meta-analysis of neuroimaging data:
a random-effects approach based on empirical esti-
tem. Additionally, activation volume extent, sta-
mates of spatial uncertainty. Hum Brain
tistical parameter, and significance level are all Mapp 30:2907–2926
archived with each experiment. Fox PT, Lancaster JL (2002) Mapping context and con-
tent: the BrainMap model. Nat Rev Neurosci
3:319–321
Activation Likelihood Estimation
Fox PT, Mikiten S, Davis G, Lancaster J (1994)
Since the results of neuroimaging experiments BrainMap: a database of human functional brain map-
are typically reported as three-dimensional coor- ping. In: Thatcher RW, Zeffiro T, Huerta M (eds)
dinates in standard space, the primary purpose of Advances in functional neuroimaging: technical foun-
dations. Academic, Orlando, pp 95–106
BrainMap is to support coordinate-based meta-
Laird AR, Lancaster JL, Fox PT (2005) BrainMap: the
analyses of functional and structural neuroimag- social evolution of a functional neuroimaging data-
ing studies. Once a set of experimental results is base. Neuroinformatics 3:65–78
Brainstem Motoneurons, Models of 451 B
Laird AR, Eickhoff SB, Kurth F, Fox PM, Uecker AM, models (Poliakov et al. 1997; Powers et al. 2005)
Turner JA, Robinson JL, Lancaster JL, Fox PT to complete Hodgkin-Huxley-type ionic
(2009) ALE meta-analysis workflows via the
BrainMap database: progress towards a probabilistic conductance-based models with reconstructed
functional brain atlas. Front Neuroinform 3:23 dendritic trees (Kulagina 2011). The choice of
Lancaster JL, Tordesillas-Gutierrez D, Martinez M, Sali- model type has been largely dictated by the ques- B
nas F, Evans A, Zilles K, Mazziotta JC, Fox PT tion at hand, i.e., responses to synaptic inputs,
(2007) Bias between MNI and Talairach coordinates
analyzed using the ICBM-152 brain template. Hum intrinsic properties, aggregate or pool behavior,
Brain Mapp 28:1194–1205 and contribution to neural circuits. In all variants,
Talairach J, Tournoux P (1988) Coplanar stereotaxic atlas these models have proved extremely useful in the
of the human brain. Thieme Medical, New York quantitative analysis of electrophysiological
Turkeltaub PE, Eden GF, Jones KM, Zeffiro TA
(2002) Meta-analysis of functional neuroanatomy of recordings from motoneurons and in the interpre-
single word reading: method and validation. tation of studies of movements controlled by
Neuroimage 16:765–780 brain stem circuits (e.g., eye movements, breath-
ing, and swallowing).
The analytical models typically involve record-
ing the responses of brain stem motoneurons to
Brainstem Auditory Evoked electrical noise injected across the cell membrane
Response through an intracellular electrode. The resultant
spike trains are analyzed using systems identifica-
▶ Auditory Evoked Brainstem Responses tion techniques that typically yield two-stage cas-
cade models, consisting of a linear input filter and
a static nonlinearity (e.g., Poliakov et al. 1997) or
alternatively autoregressive-moving average
Brainstem Motoneurons, Models of models that differentiate the roles of input filtering
and intrinsic spike-generating conductances in
Marc D. Binder governing motoneuron discharge behavior
Department of Physiology & Biophysics, School (Powers et al. 2005). These models are useful in
of Medicine, University of Washington, Seattle, describing how brain stem motoneurons respond
WA, USA to the excitatory and inhibitory synaptic inputs
they receive.
Partial Hodgkin-Huxley type ionic
Definition conductance-based models (Powers et al. 1999;
Pierson et al. 2001; Purvis and Butera 2005) have
Brain stem motoneuron models are mathematical been used to understand the mechanisms under-
representations of the input–output properties of lying spike-frequency adaptation of brain stem
motoneurons located in the different motor nuclei motoneurons and their responses to
of the brain stem. neuromodulators. More recently, complete
Hodgkin-Huxley type ionic conductance-based
models with reconstructed dendritic trees
Detailed Description (Kulagina 2011) have been used to study how
the topographic location of synaptic inputs and
Several classes of mathematical models have the presence of voltage-gated conductances in the
been developed to describe the input–output dendritic membrane shape motoneuron discharge
properties of the resident motoneurons of the behavior.
different motor nuclei of the brain stem (i.e., To study the aggregate behavior of a pool of
abducens, facial, hypoglossal, oculomotor, tri- brain stem motoneurons, the individual cells are
geminal, and trochlear). These models range in typically modeled as integrate-and-fire elements
complexity from simple “black box” analytical that begin discharging when a threshold input
B 452 Brian Spiking Neural Network Simulator
current is encountered and have their firing fre-

quencies modulated in accord with a specified Brian Spiking Neural Network
current-frequency transform. Such models may Simulator
capture some of the variability inherent in
a population of motoneurons by varying the Dan F. M. Goodman1 and Romain Brette2
1
threshold and current-frequency gains and are Department of Otology and Otolaryngology,
useful for determining the relative contributions Harvard Medical School, Boston, MA, USA
2
of intrinsic electrical properties versus the distri- Institut de la Vision, INSERM, CNRS,
butions of specific synaptic input systems to Université Pierre et Marie Curie, Paris,
motor output (Dean 1997). Integrate-and-fire rep- France
resentations of brain stem motoneurons are also
useful in modeling brain stem circuits where the
detailed physiology of the cellular elements is
Definition
sublimated to the goal of understanding the
behavior of the system (e.g., Scudder et al. 2002).
Brian (http://briansimulator.org) is an open
source Python package for developing simula-
tions of networks of spiking neurons (Goodman
Cross-References
and Brette 2008, 2009). The design is aimed at
minimizing users’ development time, with exe-
▶ Hodgkin-Huxley Model
cution speed as secondary goal. Users specify
▶ Neural Population Model
neuron and synapse models by giving their
▶ Spike Train Analysis: Overview
equations in standard mathematical form, cre-
ate groups of neurons, and connect them via
synapses. The intent is to make the process as
References flexible as possible so that researchers are
not restricted to using neuron and synapse
Dean P (1997) Simulated recruitment of medial rectus
motoneurons by abducens internuclear neurons: syn- models already built into the simulator. The
aptic specificity vs. intrinsic motoneuron properties. entire simulator is written in Python, using the
J Neurophysiol 78:1531–1549 NumPy and SciPy numerical and scientific
Kulagina IB (2011) Patterns of impulsation generated by computing packages. Parts of the simulator
abducens motoneurons with active reconstructed den-
drites. Neurophysiology 43:389–398 can optionally be run using C++ code generated
Pierson P, Tribollet E, Raggenbass M (2001) Effect of on the fly (Goodman 2010). Computationally,
vasopressin on the input-output properties of rat facial Brian uses vectorization techniques (Brette and
motoneurons. Eur J Neurosci 14:957–967 Goodman 2011) so that for large numbers of
Poliakov AV, Powers RK, Binder MD (1997) Functional
identification of the input-output transforms of neurons, execution speed is of the same order
motoneurones in the rat and cat. J Physiol (Lond) of magnitude as C++ code (Goodman and
504:401–424 Brette 2008).
Powers RK, Sawczuk A, Musick JR, Binder MD
(1999) Multiple mechanisms of spike-frequency adap-
tation in motoneurons. J Physiol (Paris) 93:101–114
Powers RK, Dai Y, Bell BM, Percival DB, Binder MD
(2005) Contributions of the input signal and prior
activation history on the discharge behavior of rat
motoneurones. J Physiol (Lond) 562(3):707–724 Philosophy
Purvis LK, Butera RJ (2005) Ionic current model of a The aim of Brian is to be, in order of priority:
hypoglossal motoneuron. J Neurophysiol 93:723–733
1. Easy to learn, use, and understand.
Scudder CA, Kaneko CRS, Fuchs AF (2002) The
brainstem burst generator for saccadic eye movements. 2. Expressive and flexible.
A modern synthesis. Exp Brain Res 142:439–462 3. Computationally efficient.
Brian Spiking Neural Network Simulator 453 B
These choices were designed to both encourage Design
correct and reproducible simulations and to mini- Brian is designed around equations: users specify
mize the development time of a simulation study as neuron models in terms of differential equations
this is often much greater than the simulation time. in standard mathematical notation rather than
using predefined neuron types, and event code B
Overview and Example such as threshold condition or reset code is also
A simulation written with Brian is a Python mod- specified in terms of mathematical expressions or
ule (script) that typically involves doing at least statements in standard notation. For example:
the following: eqs = """
1. Import the Brian package. dv/dt = (ge+gi-(v+49*mV))/
2. Define parameters and equations. (20*ms) : volt
3. Create groups of neurons with the dge/dt = -ge/(5*ms) : volt
NeuronGroup object. dgi/dt = -gi/(10*ms) : volt
4. Create synapses with the Connection or Syn- """
apses objects. The text after the colon on each line gives the
5. Specify variables to be recorded using units of the variable being defined (v, ge, gi here).
SpikeMonitor and StateMonitor objects. Equations are required to be dimensionally
6. Run the simulation with the run function. consistent.
7. Analyze and plot the output.
The following example illustrates these steps Features
(the output is shown below) (Fig. 1): Brian has support for the following basic features:
from brian import * • Define neuron models using arbitrary differ-
eqs = """ ential equations, threshold conditions, and
dv/dt = (ge+gi-(v+49*mV))/ reset codes.
(20*ms) : volt • Define synaptic neuron models using arbitrary
dge/dt = -ge/(5*ms) : volt synaptic differential equations and propaga-
dgi/dt = -gi/(10*ms) : volt tion equations.
""" • Define parameter values using SI units.
P = NeuronGroup(4000, eqs, • A library of standard neuron and synapse
threshold=’v>-50*mV’, models.
reset=’v=-60*mV’) • Record neuron activity, both spikes and traces.
P.v = -60*mV+10*mV*rand(len(P)) • Plasticity (STDP, STP, arbitrary user-defined
Pe = P.subgroup(3200) plasticity models).
Pi = P.subgroup(800) • Gap junctions.
Ce = Connection(Pe, P, ’ge’, • Nonlinear synapses.
weight=1.62*mV, sparseness=0.02) In addition, Brian comes with the following
Ci = Connection(Pi, P, ’gi’, extension packages:
weight=-9*mV, sparseness=0.02) • Model Fitting (Rossant et al. 2010, 2011). This
spikemon = SpikeMonitor(P) package fits a user-specified neuron model to
statemon = StateMonitor(P, ’v’, electrophysiological data. Fitting can be
record=range(4)) performed in parallel over multiple CPUs,
run(250*ms) multiple machines, and, if available, using
subplot(211) graphics processing units (GPUs) for a speed
raster_plot(spikemon) improvement up to 60 .
subplot(212) • Brian Hears (Fontaine et al. 2011). This pack-
statemon.plot() age is for simple and efficient modeling of the
show() auditory periphery. It consists of a base library
B 454 Brian Spiking Neural Network Simulator
Brian Spiking Neural

Network Simulator,
Fig. 1 Example output of
Brian simulator
for performing filter-based modeling, as well ▶ PyNN: a Python API for Neural Network
as implementations of many standard filter Modelling
banks used in auditory periphery modeling. ▶ Software Tools for Modeling in Computational
• Electrophysiology. Models of electrodes and Neuroscience: Overview
amplifiers, compensation, and analysis
methods.
The online support for Brian consists of the References
following:
• Extensive documentation, including tutorials, Brette R, Goodman DFM (2011) Vectorized algorithms
examples, a manual, and a reference section. for spiking neural network simulation. Neural Comput
• An online support mailing list, with feedback 23(6):1503–1535
Fontaine B, Goodman DFM, Benichoux V, Brette R
from developers and other users. (2011) Brian hears: online auditory processing using
vectorization over channels. Front Neuroinform 5:9
Goodman DFM (2010) Code generation: a strategy for
Cross-References neural network simulators. Neuroinformatics
8(3):183–196
Goodman DFM, Brette R (2008) Brian: a simulator for
▶ GENESIS, the GEneral NEural SImulation spiking neural networks in Python. Front Neuroinform
System 2:5
▶ MOOSE, the Multiscale Object-Oriented Goodman DFM, Brette R (2009) The Brian simulator.
Front Neurosci 3(2):192–197
Simulation Environment Rossant C, Goodman DFM, Platkiewicz J, Brette R (2010)
▶ NEST: the Neural Simulation Tool Automatic fitting of spiking neuron models to electro-
▶ NEURON Simulation Environment physiological recordings. Front Neuroinform 4:2
Bursting in Neurons and Small Networks 455 B
Rossant C, Goodman DFM, Platkiewicz J, Magnusson neuronal property, resulting solely from the inter-
AK, Brette R (2011) Fitting neuron models to spike action among ionic currents, or whether it is
trains. Front Neurosci 5:9
a network property, emerging from the interac-
tion among ionic and synaptic currents.
Further Reading
Brette R (2012) On the design of script languages for B
neural simulation. Network 23(4):150–156
Brette R, Goodman DFM (2012) Simulating spiking neu- Detailed Description
ral networks on GPU. Network 23(4):167–182
Markram H, Wang Y, Tsodyks M (1998) Differential sig-
naling via the same axon of neocortical pyramidal neu- Introduction and Background
rons. Proc Natl Acad Sci USA 95(9):5323–5328 Bursting patterns consist of episodes of relatively
fast spiking (bursts) separated by intervals of
either quiescence or subthreshold activity such
as subthreshold oscillations (Kispersky
Brute-Force Model Databases et al. 2010; Malashchenko et al. 2011; Desroches
et al. 2013). Bursting is ubiquitous in the nervous
▶ Neuronal Model Databases system and has been shown to play an important
role in the generation of network rhythmic pat-
terns (Nusbaum and Beenhakker 2002). Bursting
neurons are of particular importance in rhythmi-
Bursting in Neurons and Small cally active CPG networks, which control the
Networks ongoing motor behaviors that underlie coordi-
nated activity such as swimming, invertebrate
David M. Fox1, Horacio G. Rotstein3 and heartbeat, feeding, and limbed locomotion
Farzan Nadim2,3 (Friesen and Pearce 1993; Calabrese 1995;
1
Department of Biological Sciences, New Jersey Calabrese et al. 1995; Marder and Calabrese
Institute of Technology, Rutgers University 1996; Smarandache et al. 2009).
Newark, Newark, NJ, USA Extensive research into the mechanisms that
2
Federated Department of Biological Sciences, generate network bursting activity and its role in
New Jersey Institute of Technology/Rutgers motor behavior has been done in invertebrates
University, Newark, NJ, USA (Calabrese 1995; Nusbaum and Beenhakker
3
Department of Mathematical Sciences, New 2002; Selverston 2005). Invertebrate CPG net-
Jersey Institute of Technology, Newark, NJ, USA works are comprised of a relatively small number
of neurons and, in many systems, the intrinsic
properties of the participating neurons, and the
Definition network connectivity is well known. These net-
works produce multiple complex outputs that
Bursting refers to patterns of neural activity control similar behaviors as the much larger
consisting of episodes of relatively fast spiking CPG networks of vertebrate animals (Marder
separated by intervals of quiescence. Bursting and Calabrese 1996).
neurons are ubiquitous in the nervous system Bursting activity can be generated by individ-
and play important roles in the production of ual neurons or through synaptic interactions
motor, sensory, and cognitive behaviors. Because within networks. Endogenous bursting in indi-
bursting is the predominant mode of activity in vidual neurons emerges as the result of the inter-
central pattern generator (CPG) networks that action among participating ionic currents, often
underlie rhythmic motor activity, bursting neu- in the presence of tonic excitatory drive (Harris-
rons have been best characterized in invertebrate Warrick and Flamm 1987; Guckenheimer
CPG networks. Bursting neurons fall into two et al. 1993). Endogenous bursting neurons are
classes, based on whether bursting is an intrinsic often pacemakers of the networks in which they
B 456 Bursting in Neurons and Small Networks
are embedded, which rhythmically drive follower particular, what determines the intra-burst spike
neurons (Selverston 2005). Bursting may also frequency, as well as the dynamical properties of
emerge as a network phenomenon through syn- the inter-burst intervals (IBIs).
aptic interactions of individual neurons that are
not necessarily endogenous bursters (Selverston Classification of Bursting Neurons Using
et al. 2009). In such networks, the rhythmic prop- Bifurcation Theory
erties cannot be ascribed to any individual neuron A classification of mechanisms of spike initiation
but emerge as a result of synaptic organization. and termination has been described in terms of
However, as described below, network bursting bifurcations of dynamical systems (Izhikevich
can also involve endogenous bursters. 2000b, 2007). Bifurcations describe the qualitative
While bursting is a stereotypical mode of neu- structural changes in the dynamics of the system.
ronal activity, there are qualitatively different The dynamical systems analysis of neuron and
types of bursting patterns that differ not only in network dynamics is based on the assumption
the bursting attributes (e.g., burst frequency, that the output of the system is the solution of
number of spikes per burst, inter-burst interval, a set of differential equations, based on the bio-
duty cycle) but also in the mechanisms that gov- physical properties of the neurons and their syn-
ern their generation. These mechanisms can be aptic connections. The number of variables in the
investigated from two different, but complemen- differential equation is often referred to as the
tary, perspectives: membrane biophysics and dimension of the system. From a mathematical
dynamics. The biophysical mechanisms involve viewpoint, these equations result in a stable solu-
the participating ionic and synaptic currents and tion trajectory (not necessarily solvable explicitly)
neuromodulators that interact to generate burst- that describe the time-dependent changes in the
ing activity. The dynamical mechanisms involve dynamical properties of the neurons, including the
nonlinearities, time scales, and bifurcations that voltage trajectories. Using dynamical systems
govern the initiation and termination of the methods, these equations can be analyzed, often
bursts, the duration of the inter-burst intervals, in a reduced form, by exploring the transitions
and additional properties of the bursting patterns. between fast and slow kinetics. The resting state
There is no one-to-one correspondence between of a neuron, for example, may correspond to
these two mechanisms. In fact, different sets of a stable equilibrium point of the dynamical sys-
ionic currents can give rise to qualitatively similar tems equations. Repetitive spiking, in contrast,
bursting patterns by the same dynamical mecha- corresponds to a stable periodic orbit (limit cycle).
nism, and, similarly, the same sets of ionic currents Dynamical systems often evolve in multiple
can give rise to different bursting patterns. This time scales. For instance, the rest state of a neuron
article focuses on the dynamical mechanisms of can, in a slow time scale, gradually drift to a more
bursting and some of the biophysical mechanisms depolarized voltage and then suddenly transition
underlying the dynamical behavior. to a spiking state. In the fast time scale, these
transitions appear as sudden changes in the geo-
Endogenous Bursting in Individual Neurons metric properties of the system: a stable fixed
Bursting patterns can be considered as bursts of point (rest state) suddenly changes into an unsta-
spiking activity in otherwise quiescent neurons ble one, and a stable periodic orbit (spiking)
or, alternatively, as persistent spiking which is appears. These transitions reflect the bifurcations
periodically suppressed. These contrasting of the underlying dynamical system. A thorough
descriptions can be used to describe the main study of the bifurcations underlying the genera-
mechanistic features of bursting: the burst initiation and termination of bursting activity can be
tion and termination. Additionally, a complete found in Izhikevich (2000b).
description of bursting activity must include an In the classification of bursting dynamics, the
analysis of spiking properties within a burst, in underlying bifurcations of the fast dynamics can
be used to describe the overall burst structure. For X0 ¼ FðX, Y Þ
instance, a saddle-node (SN) bifurcation and
a homoclinic (HOM) bifurcation govern burst Y 0 ¼ mGðX, Y Þ,
initiation and termination, respectively, in
square-wave bursters (Fig. 1). In contrast, both where X is a vector of at least 2 fast variables for B
burst initiation and termination in parabolic repetitive spiking. Y is a vector of slow variables
bursters (Fig. 2) are governed by saddle-node on that modulates fast spiking. The small parameter
an invariant circle (SNIC) bifurcation (Rinzel m
1 is a ratio of fast–slow time scales. To
and Lee 1987; Butera et al. 1996; Rinzel and determine the type of bursting activity that the
Ermentrout 1998). Elliptic bursters involve dynamical system will exhibit, the slow variable
a subcritical Andronov–Hopf bifurcation (sub- Y is treated as a constant parameter in the fast
AH) and an SN bifurcation for burst initiation subsystem equation X 0 = F(X, Y). The parame-
and termination, respectively (Izhikevich 2000a) ter Y is used to determine the bifurcations in the
(Fig. 4). To a large extent, these bifurcations also geometric properties of the fast subsystem which
govern the dynamics of both the intra-burst spike correspond to the transitions at the burst onset and
frequency and the IBIs, for example, the shape termination.
and amplitude of spikes, spike-frequency adapta- The number of slow variables necessary to
tion, and bistability (Cymbalyuk et al. 2002). produce bursting in a model neuron depends on
From a dynamical systems viewpoint, a sim- the underlying generic burst mechanisms that
ple model that is able to generate oscillatory involve additional specific mechanisms of burst
behavior is two-dimensional and involves the initiation and termination. A bursting neuron
interaction of voltage and a recovery variable. can be driven by one of the two generic mecha-
Voltage increases due to a positive feedback nisms: a hysteresis loop such as in the square-
effect on a relatively fast time scale, typically wave burster (Fig. 1), or a slow wave, such as in
generated by calcium or sodium. The recovery the parabolic burster (Fig. 2). Hysteresis loop
variable opposes the changes in voltage on refers to the existence of bistable dynamics in
a slower time scale and, in biophysical models, the fast subsystem: if the slow variables are kept
corresponds to a state variable of a voltage-gated fixed at a constant value (within some range), the
ionic current such as a potassium current. fast subsystem will exhibit two attractors
Although the classical Hodgkin–Huxley model (a stable equilibrium point and a periodic orbit)
is four-dimensional (Hodgkin and Huxley simultaneously. Hysteresis-loop bursting
1952), the spiking behavior in this model requires only one slow variable, resulting in
can be reduced to a two-dimensional model three total variables as a minimal model for
(Kepler et al. 1992) without losing significant bursting. The slow variable allows the bistable
information. fast subsystem to transition between its two sta-
A simple way to produce bursting activity in ble states in either of two ways (Rinzel and
a model is to use a two-dimensional system that Ermentrout 1998). For example, when the tra-
produces spiking and add one or two slow vari- jectory is near the stable fixed point of the fast
ables that modulate the spiking activity. The subsystem, the slow variable grows until this
dynamical variables that make up a bursting neu- fixed point becomes unstable (through a fast
ron can then be separated into two subgroups: fast subsystem bifurcation) and the trajectory transi-
(if they evolve on the time scale of a spike) and tions to the stable periodic orbit. The slow var-
slow (if they evolve on the time scale of a burst, iable then decays until the stable periodic orbit
roughly determined by the IBI), thus giving these vanishes or becomes unstable (again through
models their name, fast–slow bursters. The burst- a fast subsystem bifurcation) and the trajectory
ing model can then be written in the following transitions back to the stable fixed point and the
form: cycle repeats.
a b
V, mV
z(t)
0
Ca(t)
−20
V(t)
10 mV −40
−2 −1 0 1
200 msec z=Ca0/(Ca+Ca0)
Bursting in Neurons and Small Networks, by Ca0 = e(m gCam1(V)(V ECa)), Ca0 = 10,
Fig. 1 Square-wave bursting in the modified e = 0:005, m = 0:2. (b) Bifurcation diagram showing
Morris–Lecar model. (a) Slow negative feedback via the region of bistability and bursting. The bifurcation
calcium accumulates during the burst and slowly parameter z is a function of calcium. Arrows indicate the
decays during the silent phase. The slow accumulation direction of the change in z during silent and active phases
of calcium activates IK(Ca), which subsequently of bursting
terminates the burst. Calcium concentration is governed
a b
s
20
0.2
0
V, mV
0.1
−20
−40
0
0.3 0.4 0.5 z 0 3 6
Time, sec
Bursting in Neurons and Small Networks, represents the boundary between equilibria and oscillation
Fig. 2 Parabolic bursting in the modified Morris–Lecar for the fast subsystem when the slow variables are fixed; it
model. Model is the same as in Fig. 1 with an additional is where the SNIC bifurcation occurs. Below the
slow variable s to generate the underlying slow oscilla- low-voltage steady state is the only fast subsystem attrac-
tion: ICa,s = gCa,s(V ECa). (a) Projection of the bursting tor (silent phase), whereas above this curve there is an
trajectory onto the slow-variable plane. Direction of oscillation of the fast subsystem (spiking). (b) Time
movement is indicated with arrows. The straight line course of parabolic bursting
In contrast, if the fast subsystem is monostable In what follows, we will think of a fast–slow
such as in the parabolic burster (Fig. 2), two slow burster as a slow system driving the fast system
variables will be necessary to generate bursting through various bifurcations that govern the ini-
activity, resulting in a total of four variables for tiation and termination of bursting activity. There
a minimal model of bursting. In this case, the are 16 possible pairs of bifurcations in hysteresis-
slow subsystem requires two variables and loop bursters and 8 possible pairs of bifurcations
exhibits an autonomous limit cycle attractor in slow-wave bursters (Izhikevich 2000b). Each
without feedback from voltage changes in the one of these bifurcations corresponds to
fast subsystem. The neuron bursts because the a different topological bursting mode. Although
fast subsystem is driven periodically through not all these bifurcations have been found in
SNIC bifurcations. biological neurons, this description has proved
a b
Circle/circle
bursting
Saddle-Node
Separatrix Loop
Fold/circle
B
Bifurcation
Circle/homoclinic bursting
bursting
Fold/homoclinic
bursting
Bursting in Neurons and Small Networks, such an SN-HOM bifurcation may exhibit four different
Fig. 3 Pairs of bifurcations in fast–slow hysteresis types of fast–slow bursting. Traversing from the red
bursters with the fast subsystem near a saddle-node region to the green region leads to square-wave bursting
homoclinic orbit (SN-HOM) bifurcation. (a) Bifurcation in which burst initiation is through the fold (or SN) bifur-
diagram showing changes in the geometry of the fast cation and the burst termination is through a homoclinic
subsystem as parameters are varied. (b) A system near orbit bifurcation
to be insightful. We discuss some of them in the transition from spiking to quiescence. The
detail below (for additional information, see homoclinic orbit bifurcation results in a spike
Izhikevich (2000b)). frequency that decreases as a function of the
One of the 8 slow-wave bursters can arise distance to the saddle. The proximity of
from a pair of SNIC bifurcations (both burst homoclinic and a sub-AH bifurcation was used
onset and termination) and is also known as the to explain the phenomenon of spike-frequency
circle–circle or parabolic bursting (Fig. 2). The adaptation. The slow adaptation variable will
topological type of the hysteresis-loop burster cause the trajectories of the fast system to track
depends on whether the fast system is close to a family of stable periodic orbits until the
either an SN-homoclinic orbit or a Bautin bifur- homoclinic bifurcation is reached, after which
cation (Figs. 3 and 4). For instance, one of the they become quiescent (Guckenheimer
16 possible hysteresis-loop bursters involves et al. 1997).
sub-AH/fold combination of bifurcations to
obtain an elliptic burster (Figs. 4 and 5). Models of Bursting Neurons
The pairs of bifurcations that determine the Bursting in neurons is a dynamical mechanism
initiation and termination of a burst can also be generated by the interplay of different types of
used to classify bursting neurons based on their ionic currents. There is not a one-to-one mapping
intra-burst spike-frequency content. If the burst between the sets of ionic currents and the stereo-
onset is through a SNIC bifurcation, the burst typical type of bursting patterns. Bursting mech-
shows a progressive increase in spike frequency. anisms have been studied using biophysical
If the burst termination is through a homoclinic (conductance-based) models, which include
orbit bifurcation, there is spike-frequency adap- a detailed description of the participating cur-
tation (a progressive lengthening of inter-spike rents, or reduced models, capturing stereotypical
intervals (ISIs)) near the end of the burst. This is bursting patterns in a relatively simple way, but
a prevalent feature in invertebrate bursting neu- lacking any correspondence to the underlying
rons. Guckenheimer et al. (1997) studied the biophysics. Below, we briefly describe these
scaling properties of ISIs at transitions associated two approaches with a special emphasis on bio-
with a homoclinic bifurcation corresponding to physical models related to invertebrate systems.
Bursting in Neurons and Small Networks, b change. In the upper two quadrants, the bifurcation of
Fig. 4 Pairs of bifurcations in fast–slow hysteresis the rest state takes place via the super-AH and in the lower
bursters with the fast subsystem near a Bautin bifurca- two quadrants via the sub-AH bifurcation. (b) Four possi-
tion. (a) Bifurcation diagram showing changes in the ble topological types of bursting with different bifurca-
geometry of the fast subsystem as parameters a and b in tions of the rest and spiking states. Traversing from the red
the topological normal form z0 = (a + io) + bz|z|2 z|z|4 region to the green region leads to elliptic bursting in
are varied. When the system is near a Bautin which the burst initiation is through the sub-AH bifurca-
(or Andronov–Hopf) bifurcation it may exhibit four tion and the burst termination is through the fold
types of bursting when the bifurcation parameters a and bifurcation
Reduced Models activation. The subthreshold dynamics are, there-

The simplest mechanism to generate bursting fore, two-dimensional. This additional current
activity consists of a sinusoidally forced endows the model with the ability to exhibit
integrate-and-fire (IF) model (Lapicque 1907) post-inhibitory rebound (PIR). This model is
where the subthreshold dynamics are linear. The able to generate bursts of activity when voltage
sinusoidal input brings voltage above threshold is hyperpolarized enough to de-inactivate IT.
for a portion of its period, thus causing persistent Since the threshold for de-inactivation is more
firing during an interval within this period and depolarized than the resting voltage, the neuron
silence outside it. This mechanism can be can respond to a brief input with a burst of spikes.
extended to include models with two- and three- However, the intrinsic dynamics of the model
dimensional linear subthreshold dynamics (with does not produce enough hyperpolarization, and
or without intrinsic subthreshold oscillations) and so periodic bursting requires an additional sinu-
the quadratic integrate-and-fire models (Feng soidal input current.
2001). While the subthreshold dynamics in The FitzHugh–Rinzel (FHR) model (Rinzel
these models can be in principle connected to 1986) is a three-dimensional generalization of
the biophysical properties of neurons, the spike the (two-dimensional) FitzHugh–Nagumo
initiation and termination mechanisms are (FHN) model (Fitzhugh 1961; Nagumo
ad hoc. et al. 1962; Fitzhugh 1969). The classical FHN
The integrate-and-fire-or-burst (IFB) (Smith forms the fast subsystem, which, to first approx-
et al. 2000) model is a generalization of the IF imation, produces oscillations when the
model to include an additional slow variable (unstable) fixed point is located in the middle
representing the de-inactivation of a simplified branch of the cubic V-nullcline. The transition
T-type calcium current (IT) with instantaneous from a stable fixed point to a stable limit cycle
Subcritical Andronov−Hopf
Bifurcation
Fold Limit Cycle B
Bifurcation
W
Spiking
V’=0
w’=0 V
V(t)
Rest Rest
Spiking
Bursting in Neurons and Small Networks, changes in the geometry of the phase plane of the fast
Fig. 5 Elliptic bursting in the Morris–Lecar model subsystem. These changes correspond to bifurcations. In
(variables V and w) with an additional IK(Ca). Green the elliptic burster, burst initiation is through the sub-AH
curves are nullclines V0 = 0 and w0 = 0 of the fast bifurcation, and burst termination is through the fold
subsystem. As the neuron transitions from silence to (or SN) bifurcation
bursting and back again, there is a clear sequence of
occurs through an Andronov-Hopf bifurcation as adaptation variable provides the necessary nega-
the tonic (DC) current is increased. In the FHR tive feedback to generate square-wave bursting.
model, this control parameter is substituted by
a slow variable, which governs the generation of Minimal (Phenomenological) Biophysical Models
elliptic bursting. Modifications to this model The two-dimensional version of the
could, in principle, generate bursting activity by Morris–Lecar (ML) model (Rinzel and
other mechanisms. Ermentrout 1998) describes the generation of
The Hindmarsh–Rose (HR) model oscillatory activity as the result of the interaction
(Hindmarsh and Rose 1984; Hindmarsh and Cor- of a non-inactivating calcium current with instan-
nelius 2005) is also three-dimensional and taneous activation and a delayed-rectifier potas-
slightly more complex than the FHR model. In sium current. The mechanisms responsible for
contrast to the FHN model, the fast subsystem of the generation of oscillations include the
the HR model has a quadratic recovery-like Andronov-Hopf and SNIC bifurcations and
nullcline that creates extra equilibrium points. depend on the parameters describing the activa-
This produces bistability through proximity to a tion/inactivation curves and voltage-dependent
saddle-node homoclinic orbit bifurcation. A slow time scales. The parameters of the basic ML
model can be adjusted to simulate the envelope of bursting oscillations involve the presence of
bursting waveforms, which has been carried out a slow inward ionic current. The current–voltage
in modeling the AB–PD neurons of the pyloric (IV) relationship of this inward current includes
pacemaker in the crustacean stomatogastric gan- a negative slope region and results in an inverted
glion (STG) (Skinner et al. 1994; Kopell bell-shaped structure which allows the inward
et al. 1998). The substitution of a biophysical current to become active regeneratively. Such
parameter (Iapp) by a slow variable endows the currents include the persistent Na+ current (INap)
ML model with the ability to generate bursting and low-threshold inactivating calcium currents
activity. One example is the bursting patterns (ICa). Such currents can work together with other
generated by adding a calcium-activated K+ cur- currents such as the slow calcium-activated
rent, IK(Ca): nonselective cation current (ICAN) and the
hyperpolarization-activated inward current (Ih)

Cap to depolarize the membrane potential and pro-
I KðCaÞ ¼ gKðCaÞ ð V EK Þ
Cap þ 1 duce a burst of action potentials. The activation
of these inward currents, which leads to a burst of
The modified ML model for bursting becomes spikes, is counteracted by the slow activation of
voltage- and calcium-dependent outward cur-

dV I Ca þ I K þ I L þ I KðCaÞ rents which terminate the burst. Moreover, some
¼ inward currents, such as the T-type calcium cur-
dt C
rent, inactivate upon depolarization of the mem-
dw l½w1 w brane potential, and this inactivation may also
¼
dt tð V Þ contribute to the termination of the burst.
Often neurons that do not burst in isolation can
dCa produce bursting activity in a network
¼ eðmI Ca CaÞ (as described in the next section). The pyloric
dt
network LP neuron does not burst endogenously
As the calcium concentration varies, the sys- but, in response to inhibitory input, can produce
tem may become bistable so that the fast post-inhibitory rebound and plateau potentials
subsystem can periodically switch back and that support a burst of action potentials. These
forth between repetitive spiking and resting, properties were described in detailed biophysical
thus producing bursting. The modified ML models of this neuron, first in a single-
model is able to produce either square-wave or compartment model (Buchholtz et al. 1992;
elliptic bursting patterns by changing parameter Golowasch et al. 1992) and, more recently, in a
values that alters the type of bistability in the fast multi-compartment model that was used to
subsystem. Burst initiation by a sub-AH bifurca- explore how different ionic conductance levels
tion produces an elliptic burster, whereas burst can produce similar bursting output in this neuron
initiation by a saddle-node bifurcation produces a (Taylor et al. 2009). In the biological network, the
square-wave burster. To obtain the slow-wave LP neuron receives periodic inhibition from the
oscillation underlying parabolic bursting in the pacemaker neurons AB and PD which result in
ML model, the addition of one more variable is the production of bursting activity out of phase
required to provide a slow positive feedback that with the pacemaker bursts. Plateau potential gen-
counteracts the slow negative feedback of IK(Ca). eration is supported by a similar set of ionic
currents as those that underlie bursting activity.
Biophysical Properties of Bursting Neurons For example, the low-threshold slowly
Bursting in neurons arises from a complex inter- inactivating ICa is substantially inactive at rest,
action between many ionic currents, where the but inactivation can be removed upon inhibitory
actual ionic mechanisms may vary between dif- input, thus promoting the plateau potential, often
ferent bursting neuron types. In most cases, together with a burst of spikes. Zhang
et al. (1995) showed that, in the dorsal gastric A biophysical model of the R15 neuron showed
(DG) motor neuron of the STG, ICAN, which is that IK(Ca) and ICAN are not necessary for
activated by calcium but is carried by Na+ and producing the parabolic-like bursting in this neu-
K+, can underlie the plateau potential that results ron (Canavier et al. 1991). This modeling study
in a long after-depolarization and contributes to also demonstrated the presence of slow oscilla- B
burst firing. The current usually acts alongside a tions in the absence of action potentials,
low-threshold ICa, and the dynamical interaction supporting the parabolic structure (Fig. 2) of the
can cause a strong regenerative current that sup- bursting. The parabolic nature of the R15 burst-
ports plateau potentials. ing was further supported by the nonuniform
The pacemaker AB neuron of the crustacean intra-burst spike frequency, even in the absence
pyloric network is an endogenous burster whose of IK(Ca). Similar to the R15 neuron, the lobster
bursting activity depends on the presence of cardiac ganglion interneurons respond to a steady
extrinsic neuromodulatory inputs (Hooper and depolarizing input with slow-wave oscillations
Marder 1987). Early attempts to dissect the called driver potentials. These driver potentials
ionic mechanism underlying endogenous burst- are mediated by slow activation and calcium-
ing involved a plateau potential which is initiated dependent inactivation of an ICa (Tazaki and
and maintained by INap and ICa and terminated by Cooke 1990).
IK(Ca) (Gola and Selverston 1981). Burst fre-
quency in AB is strongly dependent on calcium Bursting in Networks of Neurons
entry because burst frequency decreases with cal- Principles Underlying Rhythmic Bursting in
cium entry and higher calcium entry results in Networks
a larger potassium-mediated post-burst hyperpo- CPG oscillations can arise either from pacemaker
larization. The AB neuron employs different neuron activity or as a network property. In
mechanisms for bursting under a variety of mod- pacemaker-driven networks, one or more neurons
ulatory conditions (Harris-Warrick and Flamm can be identified that produce bursting oscilla-
1987). A modeling analysis of the AB neuron tions even when synaptically isolated from the
bursting activity examined the mechanisms that network. The rhythm generated by these pace-
generate and control bursting in this neuron and maker neurons is then propagated to the network
its electrically coupled partner, the PD neuron through synaptic interactions. In pacemaker-
(Soto-Trevino et al. 2005). This model demon- driven networks, the properties of the pacemaker
strated that the bursting in the AB neuron was neurons are the primary determinant of the
crucially dependent on the modulator-activated rhythm frequency. In contrast to pacemaker-
inward current IMI, and, as discussed in the next driven CPG networks, in other CPG networks,
section, the PD neuron greatly influences the duty the rhythmic pattern generation is controlled by
cycle and dynamical range of bursting frequency subnetworks of neurons and thus emerges as
in the pacemaker group. a network property. Although network-driven
Bursting can also occur solely due to the slow CPGs may also include endogenously bursting
inactivation of ICa without the need for neurons, the network frequency in these CPGs is
bistability. Two examples of ionic mechanisms primarily determined by parameters that control
underlying burst firing are the R15 neuron, the activity of the rhythm-generating
located in the abdominal ganglion of the mollusk subnetwork.
Aplysia californica, and the lobster cardiac gan-
glion interneurons. The essential ingredients for Half-Center Oscillators
bursting in the R15 neuron are the negative slope Early studies of locomotion in cats by Graham
region of ICa that determines the burst onset, Brown led him to suggest that the alternating
together with the calcium-dependent inactivation activity of flexor and extensor muscles is con-
of this current, which terminates the burst (Adams trolled by a network of two neuron populations
and Levitan 1985; Kramer and Zucker 1985). in the spinal cord, each of which inhibits the other
to produce antiphase alternating activity. He non-oscillatory neurons (Skinner et al. 1994).

termed these putative central networks that pro- Synaptic release refers to the free neuron mem-
duce antiphase motor activity half-center oscilla- brane potential falling below the synaptic thresh-
tors (HCOs) (Graham Brown 1911). Later old, thereby allowing the inhibited neuron to
studies have found reciprocal inhibition to be transition to a burst. In contrast, synaptic escape
a prevalent feature of CPG networks in both refers to the inhibited cell membrane potential
vertebrates and invertebrates (Marder and depolarizing and crossing a (low) synaptic
Calabrese 1996). In invertebrate CPGs, HCOs threshold, thereby inhibiting the free cell. The
are often comprised of two neurons coupled description of these mechanisms assumes rather
through reciprocal inhibition. The two-cell restrictive conditions: (i) synapses with well-
HCOs provide the simplest network mechanism defined synaptic thresholds and no dynamics
that is capable of generating stable bursts of alter- and (ii) HCOs that are active in the relaxation
nating activity. Alternating bursting oscillations regime (very fast transitions for burst onset and
can arise in HCOs as an emergent network prop- termination). When these conditions are relaxed,
erty even when the individual component neu- antiphase patterns may emerge as a result of
rons are not oscillatory (Wang and Rinzel 1992). combinations of escape and release such as in
When the HCO is composed of endogenous the HCO of the leech heartbeat timing network
bursting neurons, mutual inhibition acts to stabi- (Nadim et al. 1995a).
lize the burst period, and bursting occurs over
a much wider range of biophysical parameters The Role of Electrical Coupling
(Cymbalyuk et al. 2002). Examples of CPG net- Computational studies of electrical coupling have
works whose rhythm-generating mechanism provided much insight into the role of gap junc-
involves HCOs include the CPGs that control tions in networks especially because an inability
leech swimming and heartbeat (segmental oscil- to cleanly block gap junctions has led to a paucity
lators), the crustacean gastric mill network, and of experimental data. Electrical coupling is
the swim networks in lamprey and the gastropod known to promote synchrony in a network. How-
mollusk Tritonia. ever, more complex and nonintuitive behaviors
Antiphase patterns in HCOs can be generated can arise from electrically coupled networks
by various mechanisms that have been classified (Sherman and Rinzel 1992; Chow and Kopell
according to whether intrinsic or synaptic prop- 2000). Electrical coupling between an oscillatory
erties are involved and whether the inhibited cell and a bistable neuron can result in a wide variety
transitions to the bursting regime before or after of behaviors that depend on both the intrinsic
the termination of inhibition by the other cell. biophysical properties and the coupling strengths
The terms intrinsic release/escape and synaptic (Sherman and Rinzel 1992; Kopell et al. 1998).
release/escape (Wang and Rinzel 1992; Skinner Weak coupling between two identical neurons
et al. 1994) have been used to describe these can induce antiphase oscillation, which extends
mechanisms. Intrinsic release refers to the far beyond the parameter range for oscillation in
inhibited cell initiating a burst only when the single neurons, thus making for a more robust
free cell reaches the end of its burst and turns oscillator (Sherman and Rinzel 1992). Electrical
off inhibition. Intrinsic escape refers to the coupling can also produce oscillatory activity
inhibited neuron generating a burst while it is when neurons that are silent in isolation are
still inhibited and subsequently shutting off the coupled through gap junctions (Sherman and
free neuron. Post-inhibitory rebound (PIR) has Rinzel 1992; Manor et al. 1997).
been shown to initiate the inhibited cell burst The properties of a non-oscillatory neuron can
when the neuron is non-oscillatory (Wang and regulate features such as frequency and ampli-
Rinzel 1992), although it has also been shown tude of a bursting neuron to which it is electri-
that a PIR-like current is not required to generate cally coupled. However, the influence of a
oscillations in HCOs composed of non-oscillatory through gap junctions on
oscillations of a bursting neuron can be (Soto-Trevino et al. 2005). For large coupling
nonintuitive (Kepler et al. 1990). Kepler et al strengths, the non-oscillatory PD neuron
examined the effect of the PD neuron on the transitioned from tonic spiking to large-amplitude
bursting activity of the AB neuron and showed bursting. Coupling of the PD to the AB neuron
that the electrical coupling between these neu- also increased the frequency of AB oscillation B
rons can serve to either increase or decrease the compared to the isolated AB frequency. Thus,
oscillation frequency depending on the voltage the network possesses a way to regulate bursting
waveform of the bursting AB neuron. This effect oscillations that is anatomically separate from the
was further explored through a mathematical mechanism of the generation of the bursting.
analysis of the effect of a bistable neuron coupled
to an oscillatory neuron which showed how the Segmented Networks of Bursting Neurons
network frequency is influenced by the relative In segmented animals, locomotion and other
voltages of the bistable neuron compared to the motor functions often require that the movement
voltage range of oscillations in the oscillatory of different body parts maintain a constant phase
neuron. This influence is dependent on the relative to adjacent parts, despite changes in fre-
strength of coupling in a nonlinear manner quency. In crayfish and leech, swimming is con-
(Kopell et al. 1998). Using ML neurons in the trolled by the coordinated bursting activity of
relaxation-oscillation regime, the authors showed distributed CPGs in which the undulatory move-
how the electrical coupling can “pin” the voltage ments appear as a traveling wave. This traveling
trajectory to the low or high branches of the wave is generated by segmental oscillators with
voltage nullcline than compared to the uncoupled intersegmental phase lags. The neural basis that
oscillator, thus changing the frequency and shape accounts for the stable intersegmental phase dif-
of the oscillations. Unexpectedly, a coupling with ferences has been modeled as linear chains of
higher strength may not necessarily be more weakly coupled oscillators (Friesen and Pearce
effective at influencing the frequency. 1993; Skinner et al. 1997; Jones et al. 2003).
The effect of electrical coupling has also been The crayfish swimmeret system is a good
investigated in biophysically realistic model neu- example of intersegmental coordination. The
rons. In the STG pyloric network, for example, bursting activity that drives the alternating
electrical coupling between the AB and PD pace- power and return strokes of individual swim-
maker neurons causes these neurons to fire in merets on each segment is driven by a local
synchrony. Of the two, AB is the only neuron CPG module. Bursts of impulses in coordinating
that is capable of endogenous bursting. These interneurons control the relative timings of
two neurons, by virtue of their electrical cou- intersegmental bursts and result in a very stable
pling, fire burst in phase over a wide range of anterior-to-posterior phase progression of bursts
frequencies (Soto-Trevino et al. 2005). The fre- that differ in phase by about 90 between adjacent
quency of the isolated AB neuron can be greater segments (Smarandache et al. 2009; Mulloney
than 2 Hz but limited to 1 Hz in the intact net- and Smarandache 2010).
work, i.e., when coupled to the PD neuron Swimming in the leech is also involves
(Hooper and Marder 1987). Previous modeling intersegmental coordination and is characterized
work using reduced models had suggested that by a wave that travels rostro-caudally along the
the role of the PD neuron in this coupled pair is to animal. The crests and troughs are produced by
regulate the network frequency (Hooper and antiphase contractions in dorsal and ventral lon-
Marder 1987; Kepler et al. 1990; Abbott gitudinal muscles in body wall segments.
et al. 1991). The Soto-Trevino et al. multi- Intersegmental phase lag in body movement is
compartment model of the AB–PD pacemaker stable at ~20 (Friesen and Pearce 1993). Phase
ensemble demonstrated that increasing the cou- differences are also observed between segmental
pling strength caused the slow-wave oscillation oscillators in the timing network of the leech
and spikes to become near synchronized heartbeat CPG.
The leech heartbeat pattern is bilaterally approach for deriving the interaction function
asymmetric: on one side, a rostro-caudal wave H (Hoppensteadt and Izhikevich 1997). The
of muscle contractions causes the peristaltic essential step in deriving H is the infinitesimal
flow of blood through the heart tube, whereas, phase response curve (iPRC) and its relation to
on the contralateral side, all segments contract in the solution to the adjoint equation. The iPRC
synchrony. This asymmetric activity switches quantifies the normalized shift due to an infini-
after 20–50 cycles. Rhythmic activity of the tesimally small perturbation delivered at any
leech heartbeat CPG arises independently in given phase on the limit cycle so that the oscilla-
pairs of HCOs located in the third and fourth tor responds in a linear fashion.
ganglia (Nadim et al. 1995b; Hill et al. 2002). In the crayfish swimmeret system, the cou-
Rhythmic bursting arises from mutual inhibitory pling functions H have been approximated exper-
connections between two such neurons. These imentally to understand how multiple oscillator
rhythm-generating subnetworks are coupled via elements coordinate their activity to produce
coordinating interneurons from other segmental wavelike activity representative of swimmeret
ganglia to produce a stable phase difference in the locomotion (Skinner et al. 1997; Jones
range 15–20 % (Hill et al. 2002). et al. 2003). In this case, the iPRC was used to
The theory of weakly coupled oscillators approximate the H function and the zeroes of the
(Schwemmer and Lewis 2012) has been used to G function coupling two modules for a model
understand intersegmental coordination. This containing two ascending and one descending
theory is used to predict phase differences connection predicted a stable phase difference
between segmental oscillators under of 90 . A good agreement with the experimental
a mathematically described assumption of weak data showed that the effects of both excitatory
coupling. A chain of such oscillators with and inhibitory ascending connections combine to
nearest-neighbor coupling can support a promote a Stable 90 phase lag between power
constant-speed traveling wave. The weak cou- stroke neuron bursts in neighboring ganglia.
pling implies that the intrinsic dynamics are dom- Different strategies are used to achieve the
inant, so that the perturbed system remains close appropriate phase differences between neighbor-
to the limit cycle and the coupling only affects the ing segments. Phase differences generated by
speed with which the neuron moves around its asymmetries in the coupling between segmental
limit cycle. This limit cycle represents the oscil- oscillators, where ascending and descending
latory spiking activity in the case of a tonic spik- interneurons vary in strength and sign (Skinner
ing neuron or the slow-wave oscillations in the and Mulloney 1998; Smarandache et al. 2009).
case of a bursting neuron (as in Jones et al. 2003). On the other hand, the excitability gradient
The dynamics of each neuron can be reduced to a hypothesis states that phase differences arise
single-phase equation that describes the phase of from the intrinsic frequency or excitability differ-
the neuron in the limit cycle. This allows for the ences between segmental oscillators (Grillner
construction of phase models to investigate the et al. 1993). In the crayfish swimmeret CPG,
pattern of timings of individual elements in stable phase differences arise due to the com-
a network and the collective properties of the bined effect of excitation and inhibition from
system. ascending and descending coordinating neurons
The key to reducing a high-dimensional model (Jones et al. 2003). In contrast, in the leech heart-
and exploiting the form of the phase equations is beat timing network, a stable phase difference
the computation of the phase interaction function seems to arise from coupling between segmental
H (Ermentrout and Kopell 1991; Schwemmer oscillators of different inherent frequencies.
and Lewis 2012), which reduces the interaction When two segmental oscillators with different
between the oscillators to a function of their intrinsic frequencies were coupled in a model
phase difference D’. Methods of singular network, the faster one dominated the network
perturbation theory can often be used as an frequency and led in phase. The phase difference
was directly correlated with the difference in inward currents. Even in a two-cell network, the
periods (Hill et al. 2002). This phase difference coupling of a bursting neuron to a non-oscillatory
arises because the faster segmental oscillator neuron can result in phase-locked bursting. In this
bursts before the slower one and thereby termi- case, the properties of the non-oscillatory neuron
nates the activity of the shared coordinating inter- can influence the network frequency and the B
neuron. There is a brief time in each cycle when oscillation amplitude, often in unexpected ways.
the slower segmental oscillator is relieved of Electrical coupling of a non-oscillatory neuron to
inhibition from coordinating interneurons before a bursting neuron, for example, can lead to faster
the faster segmental oscillator ends its burst. or slower bursting oscillations. On the other hand,
Thus, the slow segmental oscillator can be reciprocally inhibitory coupling can lead to either
entrained by the fast segmental oscillator through alternating or synchronized oscillations. In larger
the indirect action of the coordinating networks of bursting neurons, such as those
interneuron. involved in the control of locomotion, the oscil-
latory activity can be analyzed as a chain of
Summary coupled oscillators. These oscillators often repre-
Bursting is a ubiquitous property of neurons in sent the bursting activity of the locomotor net-
many systems. Bursting in single neurons often works as recorded from the segments of the spinal
results from the interaction of voltage-gated ion cord. The theory of weakly coupled oscillators
channels at various time scales. Bursting can be provides a powerful tool for the understanding of
modeled at different levels of complexity, from how phase-locked oscillations arise in chains of
abstract reduced models to minimal two- and segmental oscillators and can be used to predict
three-dimensional models and biophysical realis- which cellular or synaptic parameters control the
tic conductance-based models, often with good frequency of the network and the relative phase
agreement with the experimental activity. of the segmental oscillators.
A bursting neuron is often described by
a system of ordinary differential equations oper- Acknowledgment Supported by grants NIH MH060605
ating at two different time scales, also called (FN), NSF DMS1313861(HGR).
fast–slow bursters. The different types of bursting
in such systems have been classified according to
dynamical systems theory in terms of the transi-
tions between the silent and spiking states. The Cross-References
hysteresis-loop bursting is the simplest form of
bursting that requires only one slow variable in ▶ Bifurcations Dynamics of Single Neurons and
addition to the fast subsystem that can operate in Small Networks
two stable states: quiescent or tonic spiking. On
the other hand, without such fast bistable dynam-
ics, the slow-wave oscillation must be generated References
independent of the fast spiking and requires at
Abbott LF, Marder E, Hooper SL (1991) Oscillating net-
least two slow variables. The classification of works: control of burst duration by electrically coupled
different bursting mechanisms is usually done neurons. Neural Comput 3:487–497
by describing the transitions between the silent Adams WB, Levitan IB (1985) Voltage and ion depen-
dences of the slow currents which mediate bursting in
and active phases of bursting using bifurcation Aplysia neurone R15. J Physiol 360:69–93
theory analysis of dynamical systems. Buchholtz F, Golowasch J, Epstein IR, Marder
Many oscillatory networks include bursting E (1992) Mathematical model of an identified
neurons that are not endogenous bursters. These stomatogastric ganglion neuron. J Neurophysiol
67:332–340
neurons respond to synaptic input with bursts of
Butera RJ Jr, Clark JW Jr, Byrne JH (1996) Dissection and
spikes that are generated by the activation of reduction of a modeled bursting neuron. J Comput
regenerative or hyperpolarization-activated Neurosci 3:199–223
Calabrese RL (1995) Oscillation in motor pattern- Hill AA, Masino MA, Calabrese RL (2002) Model of
generating networks. Current Opin Neurobiol intersegmental coordination in the leech heartbeat neu-
5:816–823 ronal network. J Neurophysiol 87:1586–1602
Calabrese RL, Nadim F, Olsen OH (1995) Heartbeat con- Hindmarsh J, Cornelius P (2005) The development of the
trol in the medicinal leech: a model system for under- Hindmarsh-Rose model for bursting. In: Bursting: the
standing the origin, coordination, and modulation of genesis of rhythm in the nervous system. World Sci-
rhythmic motor patterns. J Neurobiol 27:390–402 entific, Hackensack
Canavier CC, Clark JW, Byrne JH (1991) Simulation of Hindmarsh JL, Rose RM (1984) A model of neuronal
the bursting activity of neuron R15 in Aplysia: role of bursting using three coupled first order differential
ionic currents, calcium balance, and modulatory trans- equations. Proc R Soc Lond B 221:87–102
mitters. J Neurophysiol 66:2107–2124 Hodgkin AL, Huxley AF (1952) A quantitative descrip-
Chow CC, Kopell N (2000) Dynamics of spiking neurons tion of membrane current and its application to con-
with electrical coupling. Neural Comput duction and excitation in nerve. J Physiol 117:500–544
12:1643–1678 Hooper SL, Marder E (1987) Modulation of the lobster
Cymbalyuk GS, Gaudry Q, Masino MA, Calabrese RL pyloric rhythm by the peptide proctolin. J Neurosci
(2002) Bursting in leech heart interneurons: cell- 7:2097–2112
autonomous and network-based mechanisms. Hoppensteadt FC, Izhikevich EM (1997) Weakly
J Neurosci 22:10580–10592 connected neural networks. Springer, New York
Desroches MK, Kaper TJ, Krupa M (2013) Mixed-mode Izhikevich EM (2000a) Subcritical elliptic bursting of
bursting oscillations: dynamics created by a slow pas- Bautin type. SIAM J Appl Math 60:503–535
sage through spike-adding canard explosion in Izhikevich EM (2000b) Neural excitability, spiking, and
a square wave burster. Chaos 23:1–13 bursting. Int J Bifurcat Chaos 10:1171–1266
Ermentrout GB, Kopell N (1991) Multiple pulse interac- Izhikevich EM (2007) Dynamical systems in neurosci-
tions and averaging in systems of coupled neural oscil- ence: the geometry of excitability and bursting. MIT
lators. J Math Biol 29:195–217 Press, Cambridge
Feng HF (2001) Is the integrate-and-fire model good Jones SR, Mulloney B, Kaper TJ, Kopell N (2003) Coor-
enough? A review. Neural Netw 14:955–975 dination of cellular pattern-generating circuits that
Fitzhugh R (1961) Impulses and physiological states in control limb movements: the sources of stable differ-
theoretical models of nerve membrane. Biophys ences in intersegmental phases. J Neurosci
J 1:445–466 23:3457–3468
Fitzhugh R (1969) Mathematical models for excitation Kepler TB, Marder E, Abbott LF (1990) The effect of
and propagation in nerve. McGraw Hill, New York electrical coupling on the frequency of model neuronal
Friesen WOP, Pearce RA (1993) Mechanisms of oscillators. Science 248:83–85
intersegmental coordination in leech locomotion. Kepler TB, Abbott LF, Marder E (1992) Reduction of
Semin Neurosci 5:41–47 conductance-based neuron models. Biol Cybern
Gola M, Selverston A (1981) Ionic requirements for burst- 66:381–387
ing activity in lobster stomatogastric neurons. J Comp Kispersky TW, White JA, Rotstein HG (2010) The mech-
Physiol 145:191–207 anism of abrupt transition between theta and hyper-
Golowasch J, Buchholtz F, Epstein IR, Marder E (1992) excitable spiking activity in medial entorhinal cortex
Contribution of individual ionic currents to activity of layer II stellate cells. PloS One 5:1–21
a model stomatogastric ganglion neuron. Kopell N, Abbott LF, Soto-Trevino C (1998) On the
J Neurophysiol 67:341–349 behavior of a neural oscillator electrically coupled to
Graham Brown TG (1911) The intrinsic factors in the act a bistable element. Phys D 121:367–395
of progression in the mammal. Proc Royal Soc Lond Kramer RH, Zucker RS (1985) Calcium-induced inacti-
B 84:308–319 vation of calcium current causes the inter-burst hyper-
Grillner S, Matsushima T, Wadden T, Tegner J, El polarization of Aplysia bursting neurones. J Physiol
Manira A, Wallen P (1993) The neurophysiological 362:131–160
bases of undulatory locomotion in vertebrates. Semin Lapicque L (1907) Recherches quantitatives sur l’excita-
Neurosci 5:17–27 tion e’lectrique des nerfs traite’e comme une polariza-
Guckenheimer J, Gueron S, Harris-Warrick RM tion. J Physiol Pathol Gen 9:620–635
(1993) Mapping the dynamics of a bursting neuron. Malashchenko T, Shilnikov A, Cymbalyuk G (2011) Six
Philos Trans R Soc Lond B Biol Sci 341:345–359 types of multistability in a neuronal model based on
Guckenheimer J, Harris-Warrick R, Peck J, Willms slow calcium current. PloS One 6:e21782
A (1997) Bifurcation, bursting, and spike frequency Manor Y, Rinzel J, Segev I, Yarom Y (1997)
adaptation. J Comput Neurosci 4:257–277 Low-amplitude oscillations in the inferior olive:
Harris-Warrick RM, Flamm RE (1987) Multiple mecha- a model based on electrical coupling of neurons with
nisms of bursting in a conditional bursting neuron. heterogeneous channel densities. J Neurophysiol
J Neurosci 7:2113–2128 77:2736–2752
Marder E, Calabrese RL (1996) Principles of rhythmic Skinner FK, Kopell N, Marder E (1994) Mechanisms for
motor pattern generation. Physiol Rev 76:687–717 oscillation and frequency control in reciprocally inhib-
Mulloney B, Smarandache C (2010) Fifty years of CPGs: itory model neural networks. J Comput Neurosci
two neuroethological papers that shaped the course of 1:69–87
neuroscience. Front Behav Neurosci 4(45):1–8 Skinner FK, Kopell N, Mulloney B (1997) How does the
Nadim F, Olsen Ø, Schutter E, Calabrese R (1995a) The crayfish swimmeret system work? Insights from B
interplay of intrinsic and synaptic currents in a half- nearest-neighbor coupled oscillator models.
center oscillator. In: Bower J (ed) The neurobiology of J Comput Neurosci 4:151–160
computation. Springer, US, pp 269–274 Smarandache C, Hall WM, Mulloney B (2009) Coordina-
Nadim F, Olsen OH, De Schutter E, Calabrese RL (1995b) tion of rhythmic motor activity by gradients of synap-
Modeling the leech heartbeat elemental oscillator. tic strength in a neural circuit that couples modular
I. Interactions of intrinsic and synaptic currents. neural oscillators. J Neurosci 29:9351–9360
J Comput Neurosci 2:215–235 Smith GD, Cox CL, Sherman SM, Rinzel J (2000) Fourier
Nagumo J, Arimoto S, Yoshizawa S (1962) An active analysis of sinusoidally driven thalamocortical relay
pulse transmission line stimulating nerve axon. Proc neurons and a minimal integrate-and-fire-or-burst
IRE 50:2061–2070 model. J Neurophysiol 83:588–610
Nusbaum MP, Beenhakker MP (2002) A small-systems Soto-Trevino C, Rabbah P, Marder E, Nadim F (2005)
approach to motor pattern generation. Nature Computational model of electrically coupled, intrinsi-
417:343–350 cally distinct pacemaker neurons. J Neurophysiol
Rinzel J (1986) A formal classification of bursting mech- 94:590–604
anisms in excitable systems. In: Proceedings of the Taylor AL, Goaillard JM, Marder E (2009) How multiple
international congress of mathematics, AMS, Provi- conductances determine electrophysiological proper-
dence, pp 1578–1593 ties in a multicompartment model. J Neurosci
Rinzel JE, Ermentrout B (1998) Analysis of neural excit- 29:5573–5586
ability and oscillation. In: Koch CS, Segev I (eds) Tazaki K, Cooke IM (1990) Characterization of Ca
Methods in neuronal modeling: from ions to networks, current underlying burst formation in lobster
2nd edn. MIT Press, Cambridge, pp 251–291 cardiac ganglion motorneurons. J Neurophysiol
Rinzel J, Lee YS (1987) Dissection of a model for neuro- 63:370–384
nal parabolic bursting. J Math Biol 25:653–675 Wang X-J, Rinzel J (1992) Alternating and synchronous
Schwemmer M, Lewis T (2012) The theory of weakly rhythms in reciprocally inhibitory model neurons.
coupled oscillators. In: Schultheiss NW, Prinz AA, Neural Comput 4:84–97
Butera RJ (eds) Phase response curves in neurosci- Zhang B, Wootton JF, Harris-Warrick RM (1995)
ence. Springer, New York, pp 3–31 Calcium-dependent plateau potentials in a crab
Selverston AI (2005) A neural infrastructure for rhythmic stomatogastric ganglion motor neuron. II. Calcium-
motor patterns. Cell Mol Neurobiol 25:223–244 activated slow inward current. J Neurophysiol
Selverston AI, Szucs A, Huerta R, Pinto R, Reyes 74:1938–1946
M (2009) Neural mechanisms underlying the genera-
tion of the lobster gastric mill motor pattern. Front
Neural Circuit 3:12 Further Reading
Sherman A, Rinzel J (1992) Rhythmogenic effects of Coombes S, Bressloff PC (2005) Bursting: the genesis of
weak electrotonic coupling in neuronal models. Proc rhythm in the nervous system. World Scientific,
Natl Acad Sci U S A 89:2471–2474 Hackensack
Skinner FK, Mulloney B (1998) Intersegmental coordina- Ermentrout GB (1992) Stable periodic-solutions to dis-
tion of limb movements during locomotion: mathe- crete and continuum arrays of weakly coupled
matical models predict circuits that drive swimmeret nonlinear oscillators. SIAM J Appl Math
beating. J Neurosci 18:3831–3842 52:1665–1687
C
Cable Equation than the resistance through the axoplasm, over

short distances along the cylinder, electric current
William R. Holmes will preferentially flow down the core. In fact cur-
Department of Biological Sciences, Ohio rent will flow through the core for a considerable
University, Athens, OH, USA distance before the total amount of current that has
leaked out through the membrane becomes signif-
icant. This “core conductor” concept was given
Definition a mathematical framework in the 1870s when
Weber derived and solved steady-state equations
The cable equation is a mathematical equation in cylindrical coordinates for three-dimensional
derived from a circuit model of the membrane current flow in and around an axon. Nevertheless,
and its intracellular and extracellular space to it was not until around 1900 that Hermann and
provide a quantitative description of current others explicitly recognized that the reduction of
flow and voltage change both within and between these equations from three spatial dimensions to
neurons, allowing a quantitative and qualitative one spatial dimension (distance along the axon or
understanding of how neurons function. cable) led to an equation equivalent to Lord
Kelvin’s cable equation.
Quantitative comparison of cable equation
Detailed Description calculations to experimental results requires the
use of single fiber preparations, but such prepa-
Brief History of Cable Theory in Neuroscience rations were not available until the 1930s. Given
The origins of cable theory date back to 1855 to the current emphasis on cable theory for dendritic
work by Lord (William Thomson) Kelvin, “On the neurons, it might seem surprising to many today
Theory of the Electric Telegraph,” which provided to learn that the first applications of cable theory
the mathematical theory necessary for laying the were to determine the passive electrotonic prop-
transatlantic telegraph cable. The utility of this erties of axons. Axons are however the prototyp-
work for neuroscience, however, was not recog- ical cable. Hodgkin and Rushton in their classic
nized until many years later, after nerve axons 1946 paper derived solutions of the cable equa-
came to be regarded as “core conductors.” In the tion for an infinite cable. They noted that weak
core conductor concept, nerve axons are consid- subthreshold currents produce a voltage change
ered to be long, thin, cylindrical tubes of membrane that varies linearly with current, and they applied
filled with an electrically conducting axoplasm and their cable equation solutions to recordings of
immersed in a highly conductive medium. Because such responses to estimate membrane capacity,
the resistance across the membrane is much larger membrane resistivity, and axoplasm and

C 472 Cable Equation
extracellular resistivity in lobster leg axons. cable to be applied to many cell types. Third, axons
Hodgkin and Rushton argued that an analysis of and dendrites play different roles in signaling, so the
this “passive behavior is essential to any theory of questions for cable theory to address are much
excitation,” it “provides an insight into the struc- different. For example, how far synapses are from
ture of the surface membrane,” and “such an the soma electrotonically is a fundamental question
analysis must precede an understanding of the for dendritic neurons; simple electrotonic length
more complicated electrical changes which formulas were derived by Rall (1969) to address
make up the nervous impulse itself.” Several this question, and much work in this area has been
investigators applied cable theory to estimate done by others since (for a review and references,
passive cable parameters in a number of giant see Rall and Agmon-Snir (1998)).
invertebrate axons, and results have been sum- As information accumulates about voltage-
marized by Rall (1977; Table 1). dependent conductances in dendrites, one might
In the 1950s, the advent of glass micropipettes question the relevance of passive cable theory to
made it possible to stimulate and record intracel- the study of dendritic neurons today. On this
lularly from neurons having diameters much point, Rall and Agmon-Snir (1998) echo and
smaller than those of the large invertebrate axons. extend the arguments for the importance of
Some of the early recordings were made from the cable theory for axons given by Hodgkin and
soma of large motoneurons or pyramidal cells. It Rushton, quoted above, stating “the intuition
had been known from the time of Ramón y Cajal and methods given by passive cable theory
that these neurons possessed extensive dendritic enhance our understanding of the integrative
trees, and in the late 1950s it was verified with the mechanisms even in excitable dendrites. The pas-
electron microscope that dendrites and dendritic sive case is an important reference for the excit-
spines were sites of synapses. What role did den- able case and helps us better understand the role
drites play in neuronal excitability, and how did of excitability. Moreover, the passive case is
dendritic properties affect integration of synaptic a useful approximation that, analyzed with pow-
inputs? The early answers to these questions, erful methods, gives rise to general rules
based on interpretations of somatic recordings, concerning the role of the geometry of the den-
turned out to be incorrect. It took the application dritic tree and of passive biophysical properties.”
of cable theory to these recordings by Rall (1957, Some of these general rules will be shown below.
1959, 1960), along with subsequent work by Rall
and others, to provide a correct understanding of The Cable Equation
the role dendrites play in signaling (for a detailed Given that neurons can be represented as core
history, see commentaries and original papers in conductors with a patch of membrane adequately
Segev et al. (1995)). approximated by an electric circuit (Fig. 1), the
The application of cable theory to dendrites had derivation of the cable equation is a straightfor-
to be different from the application to axons. First, ward application of Ohm’s law, Kirchhoff’s law,
while axons are long and can be effectively and the relationship q = CV. The most difficult
modeled as infinite cables, dendrites are much part of the derivation is defining resistances,
shorter, requiring solutions for a finite cable. capacitances, and currents in the proper units
Many such solutions are given by Rall (1959, and knowing how to convert these terms to spe-
1960) and Jack et al. (1975). Second, axons tend cific resistances and capacitance. Given the
to run for long distances without branching, but circuit in Fig. 1, Ohm’s law leads to @Vi/@x =
dendritic trees are highly branched. Solutions for iira and @Ve/@x = iere, Kirchhoff’s law pro-
branched dendrites are given in Rall (1959) and also vides @ii/@x = im and @ie/@x = +im, and
Rall and Rinzel (1973) and Rinzel and Rall (1974). q = CV and Ohm’s law together give membrane
Furthermore, Rall showed that under certain condi- current im = cm @Vm/@t + (Vm Er)/rm. When
tions a branched structure could be reduced to an these equations are combined, the result is the
“equivalent cylinder” allowing solutions for a finite cable equation:
Cable Equation 473 C
Cable Equation, Fig. 1 Electric circuits for a membrane membrane current per unit length mA/cm. Note mem-
patch (a) and a neuron cable (b). rm resistance of a unit brane resistivity Rm = rmpd in O cm2, membrane capacity
length patch of membrane O cm, ra axial resistance per Cm = cm/pd in mF/cm2, axial resistivity Ra = rapd2/4 in
unit length O/cm, cm capacitance per unit length mF/cm, im O cm (also called Ri), and d diameter
rm @2Vm @V m Note that the above cable equation assumes

¼ r m cm þ ðV m Er Þ: current flow in one dimension, down the core of
r a þ r e @x 2 @t
the cable. This is a useful approximation because
This equation can be rewritten as the length of axons and dendrites far exceeds the
diameter making radial and angular flow of cur-
@2V @V rent negligible except in the immediate vicinity
l2 ¼t þ V where l
@x2 @t of a current source. A three-dimensional
rffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rm approach may be warranted when voltage
¼ and t ¼ r m cm
re þ ra changes in a large extracellular volume are of
importance (Eisenberg and Johnson 1970), but
with l known as the space constant or length
in most cases, the one-dimensional cable equa-
constant and t known as the membrane time
tion is entirely appropriate.
constant. In nondimensional form with T = t/t
and X = x/l this equation becomes
Solutions for the Infinite Cable
@V @ V 2 For an axon, the assumption is that current is
¼ V: applied at a middle point x = 0, with the axon
@T @X
cable extending in both directions from this point
We usually neglect re (see Rall 1959) and
to 1, with voltage being bounded.
express l and t in terms of specific resistances
The solution obtained for these conditions by
and capacitance as
Hodgkin and Rushton (1946) and more directly
rffiffiffiffiffiffiffiffiffi
Rm d by Jack et al. (1975) is
l¼ and
4Ra
t ¼ Rm Cm where d is cable diameter:

ra I0 l X pffiffiffi X pffiffiffi
V ðX, T Þ ¼ expðXÞerfc pffiffiffi T expðXÞerfc pffiffiffi þ T
4 2 T 2 T
in nondimensional form, where I0 is a constant stimulation site aspa ffiffiffifunction

of time, V(0,
current and erfc is the complementary error func- T) = (raI0l/2) erfc T , and the steady-state
tion. The solution above can be used to obtain two voltage, V(X,1) = (raI0l/2) exp(X). The for-
useful particular solutions – voltage at the mer shows that voltage rises to 84 % of its final
value at T = t/t = 1 (compared to 63 % for an

1 dV
isopotential sphere), and the latter illustrates the Leaky end: ¼ V ð‘ÞG‘
r a dx x¼‘
usefulness of l as a measure of voltage decay 2
with distance. 1 dV pd
e:g:, ¼ V ð ‘Þ for an end cap:
r dx
a x¼‘ R 4
m
Steady-State Solutions for Semi-infinite and
Finite Cables 1 dV
Current injection: ¼ I0 :
In the steady-state @V/@t = 0, reducing the cable r a dx x¼0
equation to an ordinary differential equation,
Note that the leaky end condition would not
have the minus sign if we considered leak at
d2 V
l2 V ¼ 0: x = 0 because of the direction of the leak current.
dx2 Simulation software packages typically assume
a sealed end condition at terminations because
This equation is easily solved, and the general the leak conductance, G‘, through an end cap of
solution can be expressed in various equivalent membrane with the same membrane resistivity,
forms: Rm, as the dendrite turns out to be negligible.
Application of boundary conditions to the
general solution leads to special solutions that
V ðxÞ ¼ A1 expðx=lÞ þ A2 expðx=lÞ
provide significant insight. First, consider the
V ðxÞ ¼ B1 coshðx=lÞ þ B2 sinhðx=lÞ
V ðxÞ ¼ C1 coshðð‘ xÞ=lÞ þ C2 sinhðð‘ xÞ=lÞ semi-infinite cable where boundary conditions
are V(0) = V0 and V(1) is bounded. These con-
ditions lead to
where ‘ is the length of the cable. The arbitrary
constants depend on the boundary conditions
x
(BC) which specify either the voltage or the cur- V ðxÞ ¼ V 0 exp :
l
rent (derivative of the voltage) at the ends of the
cable: For current input, I0, at x = 0, V0 would be
replaced by raI0l which is a factor of 2 larger than
Voltage BC : V ð0Þ ¼ V o and noted above for the infinite cable. This solution
V ð‘Þ ¼ V L or V is bounded ðnot infiniteÞ: shows the significance of l as a measure of volt-
age decay with distance in that voltage drops 1/e

1 dV 1 dV or 37 % for each l distance. It must be stressed
Current BC : ¼ i and ¼ ii :
r a dx x¼0 r a dx x¼‘ that this result applies strictly only for semi-
i
infinite and infinite cables – for finite cables
Voltage boundary conditions are straightfor- with sealed ends, the decay is less.
ward. The current boundary conditions come As a second example, consider voltage clamp
from Ohm’s law and the convention that current in a finite cable. Boundary conditions are
V(0) = V0 and sealed end at ‘. The solution is
is positive to the right, so one must be careful
with the direction of the injected current, partic-
‘x ‘
ularly at the end of the cable. This explains the V ðxÞ ¼ V 0 cosh cosh :
lack of a minus sign for the current BC at x = ‘. l l
Some commonly used current boundary condi-
tions are: One can use this expression to estimate how
effective voltage clamp at the soma space clamps
distal regions by considering the solution at x = ‘
1 dV
Sealed end: ¼ 0: or V(‘) = V0/cosh(‘/l). For cells with an electro-
r a dx x¼‘ tonic length L = ‘/l = 1, the solution shows that
the voltage change at the end of the cable is only reduction of a branched tree to an equivalent
65 % of the difference between V0 and the resting cylinder. Note that as ‘ ! 1, RN ! ral in
potential. For transient inputs, the effective space agreement with the semi-infinite solution above.
clamp is considerably worse (Rall and Segev Alternatively, we say that the input conductance
1985). of a semi-infinite cylinder, G1 = 1/(ral). It is
A third useful case is the solution for current straightforward to convert the last expression
input in a finite cable. Boundary conditions are above into the following useful equations: C
current input at x = 0 and sealed end at x = ‘.
The solution is
Rm L RN AD tanhðLÞ
r a lI 0 coshðð‘ xÞ=lÞ RN ¼ also Rm ¼
V ðxÞ ¼ : AD tanhðLÞ L
sinhð‘=lÞ
This solution is useful for illustrating the where L = ‘/l is electrotonic length and AD is the
dependence of input resistance, RN, a measure membrane area of the cable. The first of these
of cell excitability, on various electrotonic and expressions shows the dependence of input resis-
morphological factors: tance on membrane area, electrotonic length, and
membrane resistivity, and the second is useful for
r a lI 0 ra l estimating Rm when input resistance, membrane
V ð 0Þ ¼ so RN ¼ area, and electrotonic length are known or alter-
tanhð‘=lÞ tanhð‘=lÞ
pffiffiffiffiffiffiffiffiffiffiffi 3=2 natively for estimating membrane area when
2 R m Ra d
¼ : input resistance, membrane resistivity, and elec-
p tanhð‘=lÞ trotonic length are known. A means to estimate
electrotonic length independently is described
The last expression shows the dependence of below.
input resistance on d3/2, a factor that proves to be Another useful case is current input at x = 0
important for branched case solutions and and leaky end at x = ‘. In this case, the solution is
r a lI 0 ½coshðð‘ xÞ=lÞ þ ðGL =G1 Þsinhðð‘ xÞ=lÞ

V ðxÞ ¼
sinhð‘=lÞ þ ðGL =G1 Þcoshð‘=lÞ
where GL is the leaky end conductance and G1 is extension of not one but two cylinders or perhaps
the conductance of an infinite extension of the flare in dendritic tree diameters; conversely, GL/
cable (semi-infinite cylinder) as defined above. If G1 = 0.5 might represent taper. Sample solu-
GL = 0 (sealed end), we get the solution shown tions for a number of different boundary condi-
earlier for current input into a finite cylinder. If tions are shown in Fig. 2.
GL = G1, then the leak is equivalent to an infi-
nite extension of the cylinder and V(x) = I0/G1 Steady-State Solutions for Branched
exp(x/l) as shown above for the semi-infinite Dendritic Trees
cylinder. If GL = 1, then the solution represents Steady-state solutions for branched dendritic
the case where voltage at the end is clamped to 0. trees can be obtained by letting each dendritic
It is interesting to think about what other values segment be represented by a finite cable and
of GL/G1 might mean. A value of GL/G1 = 2 then connecting these cable segments at the
might correspond to the attachment of an infinite branch points using continuity of voltage and
Cable Equation, V
Fig. 2 Voltage attenuation Vo
with distance. a, Infinite 0.9 g
cylinder; b, c, d, voltage
clamp to rest in cylinders 0.8
with L = 2, 1, and 0.5; e, f,
g, sealed end in cylinders 0.7
f
with L = 2, 1, and 0.5
0.6
(From Rall (1959), Fig. 3
with permission from 0.5
Elsevier)
0.4
0.3
e
0.2
0.1 a
d c b
0 0.5 1 2 x
λ
conservation of current. It is useful to express the conductance of the initial cylinder known, along
resulting equations in terms of the unknown volt- with the leak conductances GL and G1 for all
ages at the nodes – branch points, terminations, three segments, the voltage can then be computed
and the start. For a dendritic tree with n segments, iteratively from the starting segment out to ter-
the solution would require solving n + 1 simul- minating segments. Numerical examples of this
taneous equations for voltages at the nodes. From procedure are given in Rall (1959) and Jack
these voltages, it is then possible to compute the et al. (1975).
voltage anywhere in the dendritic tree. Although the above approaches, along with
Another approach is to use the iterative algo- compartmental modeling approaches, can be
rithm proposed by Rall in 1959. In this algorithm, used to obtain steady-state solutions for cells
additional cylinders at branch points are consid- with arbitrarily branched geometry, these solu-
ered as a leak conductance at the end of the parent tions do not provide the understanding and
cylinder and use is made of the leaky end solution insight that can be obtained with closed form
of the cable equation given just above. The leak mathematical solutions. Rall and Rinzel (1973)
conductance of the initial cylinder equals the sum constructed an idealized branched neuron model
of the input conductances of attached cylinders, with suitable symmetry assumptions and equiva-
which each have a leak conductance equal to the lent cylinder constraints on branch diameters and
sum of the input conductances of their attached were able to apply superposition to obtain math-
cylinders, and so forth iteratively, until termina- ematical solutions for the case when input was
tions are reached. The leak conductance at termi- applied to only one distal branch of the dendritic
nations is assumed to be known, typically zero or tree. Explicit expressions were given for input
sealed end. Given the leak conductance at termi- resistance at the site of input and for voltage
nations, the input conductance of the terminating attenuation from the input location to the soma.
branches can be computed, and the sum of input Numerical illustrations showed how branch input
conductances of terminating branches having the resistance and voltage attenuation, factors impor-
same parent becomes the leak conductance of tant for the effectiveness of synaptic inputs,
the parent branches, which allows the input con- depend on certain morphological and electrotonic
ductance of the parent branches to be computed, properties of the dendritic tree. A result of partic-
iteratively, back to the initial segment. With input ular note is the asymmetric attenuation toward
transients were much too small when dendritic
trees were not considered appropriately. Other
methods of solution make use of Green’s
functions.
Transient solutions for branched dendrites
can become very complicated and not very
insightful without some simplifying assumptions, C
as evidenced by solutions with the Laplace trans-
form approach by several investigators. How-
ever, the development by Rall (1962) of
the equivalent cylinder model allowed the analy-
sis to be simplified significantly (see entry
“▶ Equivalent Cylinder Model (Rall)”). Briefly,
Rall showed that with a certain set of assump-
tions, a highly branched dendritic tree could be
reduced mathematically to a single equivalent
cylinder. Solutions that make use of the equiva-
lent cylinder assumption to provide particularly
insightful results have been obtained by Rall
(1969) and Rinzel and Rall (1974) as we will
Cable Equation, Fig. 3 Asymmetric voltage attenua- now discuss.
tion. Steady-state voltage attenuation with input at loca- Rall (1969) assumed that a neuron’s dendritic
tion I (solid line) or with the same input at the soma (dotted tree could be approximated by an equivalent cyl-
line). Note the steep attenuation from I to the soma and the inder and used separation of variables to solve the
lack of much attenuation from any branch point toward the
periphery. Note also that the soma voltage for input at cable equation, obtaining in nondimensional
I equals the voltage at I for input at the soma. This illus- form:
trates the reciprocity property of linear systems
(Figure reproduced from The Book of Genesis #2003
V ðX, T Þ ¼ ðA sin ðaXÞ þ B cos ðaXÞÞexp 1 þ a2 T :
James M Bower and David Beeman which was based on
Fig. 4 from Rall and Rinzel (1973) reproduced with per-
mission from Elsevier) Boundary conditions and the initial condition
are then used to evaluate the constants A, B, and
a. Applying sealed end boundary conditions at
both ends gives
the soma compared with attenuation toward the
periphery as shown in Fig. 3. X
1

np2
npX
V ðX,T Þ ¼ Bn cos exp 1 þ T
L L
Transient Solutions to the Cable Equation n¼0
Transient solutions to the cable equation have

been derived for various conditions. Mentioned where then Bn have to be obtained from the initial
above is the solution by Hodgkin and Rushton for condition, V(X,0). Regardless of the initial con-
the infinite cable, computed via Laplace trans- dition, the exponential term in the above
forms. A number of investigators have developed equation shows that the voltage will decay back
solutions with the use of Laplace transforms for to rest with an infinite number of time constants
various boundary conditions including Rall defined by
(1960), Jack and Redman (1971a, b), and others.
The analysis from Rall (1960) was particularly " #
1 1 þ ðnp=LÞ2 t0
important for showing that estimates of the mem- ¼ or tn ¼
brane time constant from recordings of voltage
tn t0 1 þ ðnp=LÞ2
where we have replaced T by t/t and then include methods are used. Estimates of L tell us if the cell
the factor 1 + (np/L)2 to define the tn. Note that is electrotonically compact or not. This is impor-
t0 = tm, the membrane time constant, when tant for estimating the effectiveness of synaptic
membrane properties are uniform. Then we can inputs and in particular the strength of distal
simplify our solution to: inputs compared to identical proximal inputs
and also for estimating the extent of space
clamp during voltage clamp. There are
X
1
V ðX, tÞ ¼ Cn expðt=tn Þ: a number of other formulas that can be used to
n¼0 estimate L, although the one given above is con-
sidered to be the most robust. A caveat is that
Rall called the tn “equalizing time constants.” L estimates made with this formula assume that
In a cylinder, the initial voltage decay is fast, as it the cell can be approximated as an equivalent
is governed by the fast time constants which cylinder; deviations from this assumption will
describe how voltage “equalizes” among the den- influence t1 and will cause errors in the
drites, while the final slower decay is governed by L estimate (see entry “▶ Electrotonic Length,
the membrane time constant. A sphere or uni- Formulas and Estimates”).
formly polarized cell has only one decay time Rinzel and Rall (1974) developed a transient
constant, the slow membrane time constant. It solution for a branched dendritic tree using the
was ignoring the fast time constants and trying same idealized dendritic tree and similar symme-
to fit voltage transients with just one exponential try assumptions and superposition methods as
term that led to the significant underestimation of used in their steady-state solutions for a branched
the membrane time constant by some investiga- tree mentioned above. The mathematical expres-
tors in the 1950s and early 1960s. Note that volt- sions derived were used to illustrate the attenua-
age charging is also faster in a cylinder than in tion and delay characteristics of the voltage peak
a sphere. This may seem counterintuitive if the caused by a transient input in a distal branch as it
fast time constants are “equalizing time con- proceeds to the soma. While voltage attenuation
stants.” To explain this paradox, it must be of transient inputs from the input site to the soma
remembered that in a sphere, the capacitance of was considerably more severe than that observed
the whole cell has to be discharged before signif- previously for steady-state inputs, a significant
icant current flows through the membrane, fraction (about half) of the input charge (the
whereas in a cylinder, only local capacitance integral of the voltage change) was nonetheless
has to be discharged before there is local mem- delivered to the soma. Expressed another way,
brane current flow; consequently, voltage charg- the attenuation of the integral of the voltage
ing begins earlier. between the input site and the soma equals the
The usefulness of this transient solution was steady-state attenuation between these two loca-
pointed out by Rall who showed that if t0 and t1 tions. These calculations also showed that the
could be estimated from an experimental tran- time to peak and half-width of the voltage
sient, then the electrotonic length of the dendritic response at the soma depend on the distance of
tree L = ‘/l can be estimated with the following the input from the soma. Some of these results are
formula: illustrated in Fig. 4.
p Insights from Cable Equation Solutions

L ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi :
t0 =t1 1 Cable equation solutions have provided consid-
erable insight leading to a quantitative and
In early studies, the time constants were esti- qualitative understanding of how neurons func-
mated by a procedure called “exponential peel- tion. Insights may have come from solutions for
ing,” but with the advent of sophisticated simplified neuron structures with particular
computers and software, nonlinear curve fitting assumptions about branching and diameters at
10−1 BI BI 10−1 BI
BI
P BS
MA
P GP BS BC-1
SO
10−2 OT GGP 10−2 OT
GP
BC-1
BC-2
−3 GGP
10 10−3 BC-2
SOMA C
OT
−4
10
OT 10−4
10−5 10−5
0 0.25 0.5 0.75 1.0 1.25 0 0.25 0.5 0.75 1.0 1.25 1.5
T T
Cable Equation, Fig. 4 Voltage attenuation for tran- points to the terminals than between branch points. Com-
sient inputs. Input is at BI. Left: voltage attenuation to pare with Fig. 3 (Reproduced from Rinzel and Rall (1974)
the branch points. Right: voltage attenuation to the termi- Figs. 4 and 5 with permission from Elsevier)
nals. Note how much less attenuation is from the branch
branch points and passive membrane with summation of synaptic inputs if they are
(usually) uniform properties, but their applicabil- small. However, it should be noted that syn-
ity is general. The list below is long but is not aptic inputs, being a conductance multiplied
meant to be exhaustive: by a driving force, are not linear, so results
1. Space constant l. The space constant (length should be interpreted with caution.
constant) l is a useful qualitative measure of 4. Reciprocity. In a linear system, the voltage
voltage decay with distance along a cable. response at the soma to current injection at
However, as a quantitative measure, l is the a dendritic location will equal the voltage
distance over which the steady-state voltage response at this dendritic location following
decays to 1/e or 37 % of its value at the origin the same current injection at the soma. This
only in infinite and semi-infinite cables. Elec- result applies generally for any two points in
trotonic length and boundary conditions, as the cell and can be used to determine if
well as the presence of side branches, matter a dendritic tree is passive or if blockers
for steady-state voltage attenuation in finite have made it so.
cables. 5. Input resistance calculations. Cable theory
2. Membrane time constant t. The membrane solutions can be used to calculate the input
time constant t is a useful qualitative mea- resistance in any passive branched dendritic
sure of how voltage decays in time. How- tree. Input resistance can be much higher in
ever, as a quantitative measure, t is the time distal dendrites than at the soma. The smaller
it takes for voltage to decay to 1/e or 37 % of diameter of the dendritic branch produces a
its initial value only in an isopotential or larger voltage response to current injection.
uniformly polarized cell. Additional time Boundary conditions also play a role.
constants affect the early voltage time course 6. Steady-state voltage attenuation depends on
depending on morphology. Neglect of these cylinder electrotonic length and boundary
additional time constants can lead to errone- conditions. As mentioned above, steady-
ous estimates of t in dendritic neurons. state voltage attenuation is described quanti-
3. Summation of inputs. In a linear system, the tatively by l only for infinite and
solution for two current inputs is the sum of semi-infinite cables. For finite cylinders,
the solutions for each input. This is helpful both cylinder electrotonic length and bound-
for understanding spatial and temporal ary conditions affect attenuation. If the
terminal boundary condition is a sealed end, the shape index will depend not only on the
attenuation with distance is less in electro- synaptic location but also on the time course
tonically shorter cylinders than longer ones. of the synaptic conductance and the electro-
If the terminal boundary condition is voltage tonic length of the dendritic tree as well.
clamp to rest, attenuation with distance is 10. The d3/2 and signal attenuation. If dendrite
larger in electrotonically shorter cylinders diameters at branch points satisfy the “3/2
than longer ones. The sealed end and voltage rule,” dp3/2 = dd13/2 + dd23/2, where dp is
clamp to rest conditions represent two the diameter of the parent branch and dd1
extremes that determine voltage attenuation and dd2 are diameters of daughter branches,
(Fig. 2). then there is impedance matching at branch
7. Voltage attenuation is asymmetric. points, and the attenuation with electrotonic
A segment of dendrite within a dendritic distance remains unchanged. However, if
tree will have a certain core resistance, but diameters at branch points do not satisfy
voltage attenuation across this core resis- this condition, attenuation will be steeper or
tance will be asymmetric, depending on the less steep after the branch point depending on
boundary conditions at the ends of the seg- whether the exponent is less than 3/2 or
ment. For an input in a distal branch, the greater than 3/2, respectively. Changes in
boundary condition heading proximally the summed d3/2 along dendrites control sig-
toward the soma is a very leaky end, nal attenuation.
approaching a voltage clamp to rest condi- 11. Consequences for voltage clamp. Attenua-
tion, while the boundary condition heading tion out from soma is relatively small. Recall
distally is a not very leaky end, approaching that the analytic solution for voltage at the
a sealed end condition. Consequently, apply- end of a cylinder given voltage clamp at the
ing the previous point above, there will be origin is V(L) = V(0)/cosh(L). This equation
severe attenuation heading proximally, but allows us to estimate the extent of space
little attenuation heading distally. This clamp in the cell when the soma is voltage
explains the asymmetric attenuation in clamped. Although much of the voltage
Fig. 3. The degree of voltage attenuation change at the soma is felt at the distal den-
across a segment depends on how much the drites, the space clamp is far from perfect at
terminal boundary condition in the direction distal locations, and this can be very impor-
current is flowing that resembles a “sealed tant in interpreting results.
end” or “clamp to rest” condition. 12. Implications for placement of excitatory and
8. Transient voltage attenuation is more severe. inhibitory inputs. The differences in input
Voltage attenuation is more severe for tran- resistance at the soma and distal dendrites,
sient inputs than for steady-state inputs. Nev- the asymmetry of voltage attenuation, and
ertheless, the attenuation of the integral of the solution for voltage clamp at the soma
the voltage change over time, with the inte- have implications for the placement of excit-
gral computed at the input site and at the atory and inhibitory inputs. Regarding inhi-
soma, equals the steady-state voltage attenu- bition, inhibition is most effective when
ation between these two locations. placed at the soma because the effect is felt
9. Dendritic filtering. In passive systems, den- throughout the cell. Regarding excitation,
drites provide low-pass filtering of the volt- soma input will not produce the large
age response to synaptic inputs. As the signal changes in voltage seen with more distal
travels to the soma, the voltage response is input, so excitation at the soma is less likely
attenuated and becomes spread out. A shape to activate voltage-gated conductances,
index which plots EPSP half-width vs. time suggesting that excitation might be better
to peak can be used to predict the electrotonic placed in the dendrites. If there are voltage-
location of the synapse. Note however that gated conductances in the dendrites, the large
distal voltage changes could activate them input resistance is small, and there is much
and amplify the signal. less attenuation. Because of this trade-off
13. Electrotonic length. The electrotonic between input resistance and attenuation,
length, L, of the dendritic tree can be esti- there will be a diameter that will maximize
mated from experimental measurements of the effectiveness inputs delivered at the end
t0 and t1 if the cell can be approximated as of the cable.
an equivalent cylinder. However, one must 17. Modeling cells as an equivalent cylinder. To C
be careful in the interpretation of the time a first approximation, dendrites can be
constants when the cell deviates from an modeled as an equivalent cylinder. This
equivalent cylinder. Most neurons have an approximation is useful for providing
electrotonic length of 1 or less. insights into how voltage changes in den-
14. Dendritic spines. Dendritic spines represent dritic trees.
a special case of asymmetric voltage attenu-
ation. Voltage attenuation will be steep from
the spine head to the dendrite because the Cross-References
boundary condition at the spine-dendrite
junction resembles a very leaky end. How- ▶ Electrotonic Length, Formulas and Estimates
ever, there will be negligible voltage ▶ Equivalent Cylinder Model (Rall)
attenuation from the dendrite to the spine
head because the boundary condition at the
tip of the spine head is practically References
a sealed end.
15. Effect of reconstruction errors. Cable theory Eisenberg RG, Johnson EA (1970) Three-dimensional
can explain effects of reconstruction errors electrical field problems in physiology. Prog Biophys
on parameter value estimates. Dendrite Mol Biol 20:1–65
Hodgkin AL, Rushton WAH (1946) The electrical con-
diameters are notoriously difficult to mea- stants of a crustacean nerve fibre. Proc R Soc Lond
sure accurately because diameters are near B Biol Sci 133:444–479
the optical resolution of the microscope. Jack JJB, Redman SJ (1971a) The propagation of transient
Length values can be inaccurate because of potentials in some linear cable structures. J Physiol
Lond 215:283–320
tissue shrinkage, particularly in the z-axis. Jack JJB, Redman SJ (1971b) An electrical description of
Cable theory says if the reconstruction diam- the motoneurone and its application to the analysis of
eters are uniformly off by a factor of x, then synaptic potentials. J Physiol Lond 215:321–352
the Rm, Cm, and Ra values used in models Jack JJB, Noble D, Tsien RW (1975) Electric current flow
in excitable cells. Oxford University Press, Oxford,
should be equal to the actual Rm multiplied UK
by x, the actual Cm divided by x, and the Rall W (1957) Membrane time constant of motoneurons.
actual Ra multiplied by x2. If reconstruction Science 126:454
lengths are uniformly off by a factor y, then Rall W (1959) Branching dendritic trees and motoneuron
membrane resistivity. Exp Neurol 1:491–527
the Rm, Cm, and Ra values used in models Rall W (1960) Membrane potential transients and mem-
should be equal to the actual Rm multiplied brane time constant of motoneurons. Exp Neurol
by y, the actual Cm divided by y, and the 2:503–532
actual Ra divided by y. Rall W (1962) Theory of physiological properties of den-
drites. Ann N Y Acad Sci 96:1071–1092
16. Optimal diameters. Cable theory calcula- Rall W (1969) Time constants and electrotonic length of
tions suggest that a neuron can optimize the membrane cylinders and neurons. Biophys
effectiveness of distal inputs with its mor- J 9:1483–1508
phology or electrotonic parameters. For Rall W (1977) Core conductor theory and cable properties
of neurons. In: Handbook of physiology. The nervous
a cylinder of given length, if diameter is system. Cellular biology of neurons, Sect 1, vol I, pt
small, input resistance is large, and attenua- 1, chap 3. American Physiological Society, Bethesda,
tion is steep, whereas if diameter is large, pp 39–97
C 482 Cable Model(s)
Rall W, Agmon-Snir H (1998) Cable theory for dendritic are termed Ca2+ sensors and are modeled simi-
neurons. In: Koch C, Segev I (eds) Methods in neuro- larly, except that target reactions are included.
nal modeling, 2nd edn. MIT Press, Cambridge, MA,
pp 27–92 Finally, perturbations by Ca2+ indicator dyes,
Rall W, Rinzel J (1973) Branch input resistance and which are themselves buffers, can be addressed
steady attenuation for input to one branch of similarly.
a dendritic neuron model. Biophys J 13:648–688
Rall W, Segev I (1985) Space-clamp problems when volt-
age clamping branched neurons with intracellular
microelectrodes. In: Smith TG, Lecar H, Redman SJ,
Gage P (eds) Voltage and patch clamping with micro-
electrodes. American Physiological Society, Bethesda,
pp 191–215 Background
Rinzel J, Rall W (1974) Transient response in a dendritic Opening of Ca2+ permeable channels results in
neuron model for current injected at one branch.
steep increases in the intracellular free Ca2+ con-
Biophys J 14:759–790
Segev I, Rinzel J, Shepherd GM (eds) (1995) The theoret- centration ([Ca2+]i) that can reach tens of
ical foundation of dendritic function. Selected papers micromol/liter at the entry site. Ca2+ buffers
of Wilfrid Rall. MIT Press, Cambridge, MA make a major contribution to reducing [Ca2+]i to
lower levels within milliseconds and to limiting
Further Reading the spatial extent of [Ca2+]i. Thus, buffering is
Rall W, Burke RE, Holmes WR, Jack JJB, Redman SJ,
central for maintaining a high temporal respon-
Segev I (1992) Matching dendritic neuron models to
experimental data. Physiol Rev 72:S159–S186 siveness to [Ca2+]i signals, which is eminent for
neuronal coding.
Second-Order Buffer Reaction

Cable Model(s) Chemical Reaction
The simplest reaction is binding of Ca2+ to
▶ Mammalian Motor Nerve Fibers, Models of a buffer (B) with only one binding site:
koff
½Cai þ ½Bf
)

* ½CaB
(1)
kon
Calcium Buffering: Models of
½Bf ¼ ½BT ½CaB (2)
Hartmut Schmidt
Carl-Ludwig-Institute for Physiology, Leipzig, 1 koff ½Bf ½Cai
KD ¼ ¼ ¼ (3)
Germany K A kon ½CaB
[Ca2+]i, [B]f, and [CaB] are the concentrations of

Definitions free Ca2+, free buffer, and Ca2+ buffer complex,
respectively. [B]T is the total buffer concentra-
Neuronal Ca2+ dynamics are tightly controlled by tion. KD and KA are the equilibrium dissociation
endogenous calcium -binding proteins (CaBPs), and association rate constants, respectively. In
which may bind one or more Ca2+ ions either in neuronal modeling, it is convenient to use KD.
a cooperative or noncooperative manner. The Note that a higher KD value implies that the
binding of Ca2+ to buffers can be described kinet- buffer has a lower affinity for Ca2+ and is less
ically by using binding and unbinding rate con- easily saturated. Concentrations and KD are in
stants in ordinary differential equations (ODEs) units of M (mol/liter). kon (unit: M1 s1) and
or by using equilibrium constants. In addition to koff (s1) are the forward and backward binding
buffering, some CaBPs perform specific Ca2+- rate constants, respectively, determining the
dependent actions on downstream targets; these velocity of the reaction.
Calcium Buffering: Models of 483 C
Kinetic Model
The above reactions together with a Ca2+ influx
can be converted to ODEs for each well-mixed
compartment, assuming mass action kinetics
(Markram et al. 1998; Schmidt et al. 2003):
d½Cai ðtÞ C
¼JþR (4)
dt
d½Bf ðtÞ
¼R (5)
dt
d½CaBðtÞ
¼ R (6)
dt
R ¼ kon ½Cai ðtÞ ½Bf ðtÞ þ koff ½CaBðtÞ (7)
where R is the reaction term and J the Ca2+ influx,

both in units of M/s. In order to solve the system Calcium Buffering: Models of, Fig. 1 [Ca2+]i decay
numerically (▶ Deterministic Reaction-Diffusion kinetics (a) and fractions of Ca2+-bound buffers (b) after
Simulators), starting conditions for t = 0 are a 100 ms influx step in the presence of either a rapid (solid
lines, kon = 108 M1 s1, koff = 100 s1) or a slow
determined by assigning [Ca2+]i(0) = [Ca2+]rest (dashed lines kon = 107 M1 s1, koff = 10 s1) buffer,
and by calculating [B]f(0) and [CaB](0) from Eq. 3. both with a KD of 1 mM and [B]T of 100 mM. Dotted lines
From a brief increase, [Ca2+]i relaxes approx- indicate resting levels
imately exponentially to a value determined by
KD, and the fraction of Ca2+ bound buffer
Ca2þ T
¼ Ca2þ i þ ½CaB
increases concomitantly (Fig. 1). The time con-
½BT ½Cai
stant is determined by kon, koff, [Ca2+]i, and [B]f ¼ Ca2þ i þ (9)
but can be approximated as t 1/(kon [B]f(1)) K D þ ½Cai
(Markram et al. 1998). By differentiating [CaB] with respect to [Ca2+]i
in Eq. 8, the “buffer capacity” (k) is defined, which
gives a dimensionless measure of how many Ca2+
Rapid Buffer Approximation (RBA)
ions are bound for each free Ca2+. For example, k
The RBA is a steady-state simplification for bind- of 100 (a typical value for neurons) implies that
ing in well-mixed compartments under the 100 Ca2+ are bound for each free Ca2+:
assumption of quasi-instantaneous equilibration
of all reactants (Neher 1998). Although this d½CaB K D ½BT
assumption is not met during rapid signaling, the k¼ ¼ 2 (10)
d½Cai K D þ ½Ca
RBA is often useful. Combining Eqs. 2 and 3 ([B]f i
is typically unknown, whereas [B]T may be esti-
mated experimentally) and rearranging yields By differentiating [Ca2+]T with respect to
[Ca2+]i in Eq. 9 and inverting the result, the
“buffering factor” (b) is defined, which is a num-
½BT ½Cai
½CaB ¼ (8) ber between zero and one and a measure of the
K D þ ½Cai fraction of Ca2+ remaining free:
Using Eq. 8, the total Ca2+ concentration (free d½Cai 1

b¼ ¼ (11)
and bound Ca2+) is given by d½CaT 1 þ k
C 484 Calcium Buffering: Models of
The usefulness of k and b becomes apparent

when considering the sum of Eqs. 4 and 6 for the
total Ca2+
d½CaT ðtÞ
¼J (12)
dt
and then deriving the change in [Ca2+]i
d½Cai ðtÞ d½Cai ðtÞ d½CaT ðtÞ

¼ ¼b J (13)
dt d½CaT ðtÞ dt
Equation 13 quantifies the degree to which

a very rapid buffer limits the impact of a Ca2+
influx J on [Ca2+]i in a well-mixed compartment.
Although k and b are functions of Ca2+, they can
be regarded as constant if [Ca2+]i<< KD. On the
other hand, if [Ca2+]i >>KD, then k -> 0 and
b -> 1, i.e., the buffer is saturated.
Competition Between Rapid and Slow Buffers Calcium Buffering: Models of, Fig. 2 Competition of
A neuron typically expresses multiple CaBPs fast and slow buffer for Ca2+. (a) [Ca2+]i kinetics as in Fig. 1
(with different kinetics) that compete for Ca2+ but in the presence of both buffers. (b) Ca2+ bound frac-
tions of the rapid (solid line) and the slow buffer (dashed
binding. In such a system, the amount of Ca2+ line) as in Fig. 1 but now during competition. Note the
bound to a given buffer can initially deviate sub- difference in their binding behavior compared to Fig. 1
stantially from the equilibrium (Markram et al.
1998). A rapid buffer will initially bind more reaction, for example, positive cooperativity
Ca2+ than predicted from equilibrium, while the means that kon,2 > kon,1 and/or koff,2 < koff,1,
slow buffer will bind less Ca2+. Later on, Ca2+ such that KD,2 < KD,1. Calretinin is an example
liberated by the rapid buffer will be bound by the for positive cooperativity (Faas et al. 2007):
slow buffer, and both will relax to their equilib- 2 kon, 1
rium occupancy (Fig. 2). This generates 2 ½Cai þ ðBI BII f
) * ½Cai
koff , 1
multiphase [Ca2+]i decay kinetics (Schmidt et al.
kon, 2
2003). þ ½CaðBI BII Þ * ½Ca2 ðBI BII Þ
) (14)
2 koff , 2
Competition Between Ca2+ and Mg2+
The corresponding reaction rate is
CaBPs bind other divalent cations like Mg2+. The
competition is included by adding ODEs for
Rc ¼ 2kon, 1 Ca2þ i ðt ½BI BII f ðt
Mg2+ similar to those for Ca2+. This shifts the

effective KD for Ca2+ towards larger values kon, 2 Ca2þ i ðt CaðB I BII ðt
(Fig. 3, Schmidt 2012). þ 2koff , 2 ½Ca2 ðBI BIIÞðt

þ koff , 1 CaðBI BII ðt (15)
Higher-Order Buffering
Many CaBPs have several binding sites that may Replacing R by Rc in Eqs. 4, 5, and 6 yields the
cooperate with each other, i.e., binding of Ca2+ to ODEs for the kinetic simulation.
one site facilitates (positive cooperativity) or hin-
ders (negative cooperativity) binding of Ca2+ to Spatial Resolution
other sites. This is expressed by different kon and Spatial resolution can be achieved by subdividing
koff for the individual steps. In a two-step a well-mixed compartment into smaller
Calcium Buffering: Models of 485 C
Calcium Buffering: Models of, Fig. 3 Steady-state Ca2+-binding curves for parvalbumin (KD,Ca = 9nM, KD,Mg = 31
mM) in the absence (dotted line) and presence (solid line) of 600 mM [Mg2+]i
compartments that are coupled by diffusion. This The system can be solved numerically by the
is obtained by introducing diffusion flux terms JD method of finite elements, i.e., by breaking down
(M/s) into Eqs. 4, 5, and 6 for each compartment the total geometry into a finite number of
n (Markram et al. 1998; Schmidt et al. 2013): subcompartments (meshing).
An, n1 Limitations

J D ¼ Dx ½Xn1 ðtÞ ½Xn ðtÞ (16)
d Vn Real chemical reactions are discrete events that
depend on the collision of the reactants, i.e., they
where Dx (m2/s) is the diffusion coefficient of are stochastic and noisy. With larger numbers of
species X, A the exchange surface area, d the molecules, the reactions are well described by the
distance between subcompartments, and V their deterministic differential equations. For small
volume. molecule numbers, the effects of noise may
Spatial resolution can also be obtained by become important, and it could be appropriate to
converting Eqs. 4, 5, 6, and 7 to partial differen- use stochastic simulation. Such a Monte Carlo
tial equations (PDEs) with the spatial coordinates simulation is more time-consuming, and it has to
as additional independent variables. The inde- be carefully weighted whether it is really required.
pendent variables are omitted in the following As a rule of thumb, a reaction with less than 100
for clarity: molecules is likely to be more adequately
described stochastically (Bhalla and Wils 2009).
@ ½Cai
¼ DCa ∇2 ½Cai þ R þ J (17)
@t
References
@ ½Bf
¼ DB ∇2 ½Bf þ R (18)
@t Bhalla US, Wils S (2009) Reaction-diffusion modeling.
In: De Schutter E (ed) Computational modeling
@ ½CaB
¼ DCaB ∇2 ½CaB R (19) methods for neuroscientists. The MIT Press,
@t Cambridge, pp 61–91
C 486 Calcium Dynamics in Neuronal Microdomains
Faas GC, Schwaller B, Vergara JL, Mody I (2007) Resolv- added to measure the number of calcium ions can
ing the fast kinetics of cooperative binding: Ca2+ buff- severely perturb the endogenous chemical reac-
ering by calretinin. PLoS Biol 5:2646–2660
Markram H, Roth A, Helmchen F (1998) Competitive tions. Over the years, an alternative approach
calcium binding: implications for dendritic calcium based on modeling, mathematical analysis, and
signaling. J Comput Neurosci 5:331–348 numerical simulations has demonstrated that it
Neher E (1998) Usefulness and limitations of linear can be used to obtain precise quantitative results
approximations to the understanding of Ca++ signals.
Cell Calcium 24:345–357 about the order of magnitude, rate constants, the
Schmidt H (2012) Three functional facets of calbindin role of the cell geometry, and flux regulation across
D-28k. Front Mol Neurosci 5:25 scales from channels to the cell level.
Schmidt H, Stiefel K, Racay P, Schwaller B, Eilers J The aim of this ECN is to present physical
(2003) Mutational analysis of dendritic Ca2+ kinetics
in rodent Purkinje cells: role of parvalbumin and models of calcium ions from the molecular
calbindin D28k. J Physiol (Lond) 551:13–32 description to the concentration level and to pre-
Schmidt H, Brachtendorf S, Arendt O, Hallermann S, sent the mathematical tools used to analyze the
Ishiyama S, Bornschein G, Gall D, Schiffmann SN, model equations. From such analysis, asymptotic
Heckmann M, Eilers J (2013) Nanodomain coupling at
an excitatory cortical synapse. Curr Biol 23:244–249 formulas can be obtained, which are usually valid
for a certain range of parameters. However these
Further Reading formulas allow exploring at low cost the parameter
De Schutter ED (2009) Computational modeling methods space. The methods to analyze these equations are
for neuroscientists. The MIT Press, Cambridge part of the classical analysis of partial differential
Smith GD (2000) Modeling local and global calcium equations and stochastic processes, which will not
signals using reaction- diffusion equations. In:
Computational neuroscience. CRC Press, Boca Raton be reviewed here (see Schuss 1980, 2010a). We
shall present several models related to diffusion,
where formulas can be derived, and we shall spec-
ify how these formulas are used to extract param-
eters from experimental measurements. But in
Calcium Dynamics in Neuronal general models are far too complicated to lead to
Microdomains: Modeling, Stochastic equations that can be analyzed, and most of the
Simulations, and Data Analysis time, numerical simulations have to be built.
Building rational simulations requires discretizing
C. Guerrier1,2, E. Korkotian3 and D. Holcman1,2 the physical equations and bridging the gap
1
École Normale Supérieure, Institute for between the limits of the equations and the physi-
Biology, IBENS, INSERM 1024 and CNRS cal description that they account for. We will pre-
Group of Computational Biology and Applied sent here several stochastic simulations and their
Mathematics, Paris, France rules, limitations, and tricks that have been devel-
2
University Paris 6, Laboratoire Jacques–Louis oped over the years. As we shall see here, any
Lions, Paris, France bottleneck in the equation can lead to heavy simu-
3
Department of Neurobiology, Weizmann lations running for days. In that case, coarse
Institute of Science, Rehovot, Israel graining is a key step to reduce the complexity of
the equation so that some analysis can be obtained
and can be used to check in some limit the validity
Definition of the simulations.
All together physical modeling, mathematical
Calcium is a key but a ubiquitous messenger in cell analysis, numerical simulation, and their applica-
physiology. Yet direct electrophysiological or light tion to the statistical analysis of experimental
imaging measurements are limited by the intrinsic data form an ensemble of approaches that are
small nano- to micrometer space where chemical used today to better understand molecular inter-
reactions occur and also by the small number of action in nano- to microdomains. But the most
molecules. Thus any fluorescence dye molecule striking convergence of these methods is to
Calcium Dynamics in Neuronal Microdomains 487 C
derive physiological laws from their first physical the main driving force for the motion of particles
principles. We shall present (1) stochastic model- such as ions or molecules. Calcium motion has
ing of calcium ions and their trajectories, been well approximated as diffusion, which con-
(2) modeling of local interactions with discussion firms that the ionic charge is screened. This is not
of the rate constants, (3) derivation of asymptotic the case for calcium influx through channels,
formulas for the residence of calcium in where the nanometer space restriction is
microdomains and dendritic spines, (4) modeling a dominant factor for the interaction between C
and simulation of diffusion in a crowded three- the channel charges and the ion (Benazilla 2000;
dimensional dendrite, (5) modeling of calcium in Eisenberg et al. 1995). In cells, long distances are
the spines, (6) modeling the CaM activation path- overcome mostly by the diffusion process. For
way, and (7) modeling of coarse-grained Markov example, free calcium ions diffuse in the den-
chain to estimate the probability that the number dritic spines or any neuronal microdomain in the
of calcium ions bound to key molecule is reached. dendrite or the axons (Korkotian et al. 2004). At
Finally, we shall discuss how these simulations the molecular level, diffusion of ions is described
can be used to study physiological processes such by a random walk. When an ion meets any plasma
as transient calcium ions in a dendritic spine, membrane, it is reflected or translocated inside an
long-term potentiation induction, or what limits organelle at pumps or exchangers. The transition
the spread of calcium following synaptic from one region to another can be modeled by
stimulations. specifying absorbing conditions in specific
boundary regions, such as the fictive separation
between a dendrite and a spine.
Description of Calcium Stochastic Trajectories
Neuronal Microdomains A calcium ion trajectory is described by the
Neurons can be decomposed into several key Smoluchowski limit in the large damping approx-
microdomains involved in specific functions: imation of the Langevin equation. The position
dendrites integrate electrical signals, whereas x at time t satisfies the stochastic equation
dendritic spines are microcompartments receiv-
ing the postsynaptic terminal of excitatory syn- pffiffiffiffiffiffiffi
g½x_ V ðx, tÞ þ FðxÞ ¼ 2egw: (1)
apses (Fig. 1). The presynaptic terminal is also
a key compartment that controls calcium in rela-
tion to vesicular release. Dendrites further Here e = kBT/m, where T is the temperature
decompose into distal and apical dendrites and kB the Boltzmann constant; g = 6pa is the
which seem to have different electrophysiologi- dynamical viscosity, where is the viscosity
cal properties that can be correlated with coefficient per unit mass and a is the radius of
a difference in the structure. The soma and the the ion. w is a random white noise modeling the
axon are also separate compartments. We show in thermal fluctuations, and the electrostatic force is
Fig. 1 different distinct compartments. Finally, it
was found that aspiny neurons can confine cal- FðxÞ ¼ ze∇x U 0 ðxÞ,
cium for a time scale comparable to a dendritic
spine (Goldberg et al. 2003). Can organelles such where U0 is the potential created by the site where
as vesicles or endoplasmic reticulum control cal- the proteins are located. In a first approximation,
cium concentration? each protein creates a localized parabolic poten-
tial, where the depth can be calibrated by using
Diffusion in Neuronal Microdomains: the backward-binding rate and the radius by using
Stochastic Modeling the forward-binding rate (Korkotian and Segal
We start with a Brownian description of calcium 2006). The frictional drag force, g½x_ V ðx, tÞ,
ions. Indeed motion due to thermal fluctuations is is proportional to the relative velocity of the ion
Calcium Dynamics in Neuronal Microdomains: dendrites from the hippocampus. The PSDs of excitatory
Modeling, Stochastic Simulations, and Data Analy- synapses are marked in red and of inhibitory synapses in
sis, Fig. 1 Neuronal microdomains. (a) Electron blue. Filopodia (marked F) and mushroom spines (marked
microscopy of a synapse. The postsynaptic terminal is M) are clearly seen. (c) All possible types of the dendritic
located on a dendritic spine (S) branched in the dendrite spines, for example, a 3D reconstructed hippocampal den-
(D). Located around the axon (A) are the glial cells (G). drite, 3 weeks old in primary culture
(b) Three-dimensional EM reconstruction of two
and the cytoplasmic fluid. The field of fluid V(x, t) space. We consider a given flow field V(x, t) (see
induced by calcium ions will be discussed below. description below) and that ions interact with
a fixed potential of the charges on the proteins,
The Langevin Equations U0(x), and with the variable potential of all other
We shall here explain the physical consideration ions. The variable potential consists of both the
behind the reduction to Eq. 1. For a dendritic electrostatic ion-ion interaction potential, U ii ðx~Þ,
spine containing N ions of different species and the potential of Lennard-Jones-type repul-
(e.g., Ca++, Na+, Cl–, and so on), xi (t) is the sions, ULJ ðx~Þ (that represents the finite size of
displacement vector of the i-th ion, mi is its the ions). The force per unit mass on the i-th ion is
mass, and zi is its valence. x~ ¼ ðx1 , x2 , , xN Þ
is the coordinate of the N ions in configuration Fi ðx~Þ ¼ zi e∇xi ½U 0 ðxi Þ þ U ii ðx~Þ ∇xi U LJ ðx~Þ:
The dynamics of the i-th ion is given by the Specification of the Hydrodynamic Flow
Langevin equation The flow of the incompressible cytoplasmic fluid
in the spines is generated by the local contraction
pffiffiffiffiffiffiffiffiffi
x€i þ gi ½x_ i Vðxi , tÞ þ Fi ðx~Þ ¼ 2ϵi gi w_ i , (2) of actin-myosin complexes saturated by calcium
ions. We assume that the flow field is derived
where ϵi = kBT/mi, T is the temperature, from a potential f(x, t) (see, e.g., Landau and
gi = 6paii is the dynamical viscosity, i is the Lifshitz 1975), C
viscosity coefficient per unit mass, and ai is the
radius of the ion. The frictional drag force, Vðx, tÞ ¼ ∇fðx, tÞ: (5)
g½x_ i Vðxi , tÞ , is proportional to the relative
velocity of the ion and the cytoplasmic fluid. The The incompressibility condition, ∇ V(x, t) = 0,
accelerations w_ i represent the thermal fluctuations reduces to the Laplace equation in the head OH(t) of
of the fluid. The relation between the velocity the spine at time t. The surface of the head, S(t), is
diffusion constant and the friction coefficient, partitioned into the surface SH(t) of the spine head,
that does include the surface common with the neck,
kB T and the cap SN(t) of the surface of the head inside the
Di ¼ , neck, S(t) = SH(t) [ SN(t). The Laplace equation in
m i gi
OH(t) is
is Einstein’s fluctuation-dissipation principle
(Schuss 1980). In the Smoluchowski limit of Dyfðy, tÞ ¼ 0 for yOH ðtÞ, t > 0, (6)
large damping (Schuss 1980) the Langevin equa-
tion (2) reduces to with the boundary conditions
gi ½x_ i Vðxi , tÞ þ Fi ðx~Þ ¼ 2ϵi gi w_ i , (3) @fðy, tÞ @fðy, tÞ
¼ V ð tÞ, ¼ FðV ðtÞÞ,
@n ySH ðtÞ @n ySN ðtÞ
When neglecting the ion-ion interactions, we (7)
set U LJ ðx~Þ ¼ U ii ðx~Þ ¼ 0, so that Eq. 3 becomes
where V(t) is the average velocity induced by the
gi ½x_ i Vðxi , tÞ þ Fðxi Þ ¼ 2ϵi gi w_ i , (4) deformation of the head (see Eq. 8 below)
(Holcman and Schuss 2004), due to the sum of
where all the local contractions, and F(V(t)) is the
induced field velocity at the top of the neck
Fðxi Þ ¼ zi e∇xi U 0 ðxi Þ: SN(t): for a volume displaced per unit time
equal to 4pR2(t)V(t) in dimension 3 and 2pR(t)
Since we are interested in tracing only one V(t), where R(t) is the instantaneous radius of the
species in the spine, namely, the concentration head, then R_ ðtÞ ¼ V ðtÞ. The flux through SN is
of calcium, we assume that gi ¼ gCaþþ , |SN|v(t); hence
mi ¼ mCaþþ , zi ¼ z ¼ 2: Under these assump-
tions, equations in (4) are independent and iden- 8
>
> 4pR2 ðtÞV ðtÞ
tical, so that their transition probability densities < in dimension 3
are identical. We denote the transition probability vðtÞ ¼ FðV ðtÞÞ ¼ jSN j ,
>
> 2pRðtÞV ðtÞ
density function (pdf) of each ion by p(x, t|x0, t0) : in dimension 2
so that the calcium concentration is jSN j
ð
when the field is due to the contraction of
cðx, tÞ ¼ pðx, tjx0 , t0 Þc0 ðx0 Þdx0 ,
Ot myosin after 4 calcium ions are bound. The total
number of sites bound to 4 calcium is S(4)(t)
where c0(x0) is the initial calcium density. and can be obtained by solving a system of
reaction-diffusion equations (Holcman and function G(x, y, t) is the solution (defined up to

Schuss 2004). Finally the velocity at the bound- a constant) of the equation
ary is given by
1
V ðtÞ ¼ vQ Sð4Þ ðtÞ, (8) Dy Gðx, y, tÞ ¼ dðx yÞ for x, yOH ðtÞ
jOt j
@Gðx, y, tÞ
where vQ is a constant velocity. The quantities ¼ 0 for xOH ðtÞ, ySðtÞ:
V(t) and F(V(t)) are stochastic processes, that are @nðyÞ
proportional to the number of saturated proteins (9)
at any given time t. The flow field can be
expressed explicitly in terms of the functions Multiplying Eq. 6 by G(x, y, t) and Eq. 9 by
V(t) and F(V(t)) by Green’s function for the Neu- f(y, t) and integrating with respect to y over the
mann problem for Poisson’s equation in a sphere domain, using Stokes’ theorem and the boundary
(or a disk) through Stokes’ formula. Green’s condition (7), we get
ð ð
@fðy, tÞ @Gðx, y, tÞ
fðx, tÞ ¼ Gðx, y, tÞdSy fðy, tÞdSy
ySðtÞ @n ySðtÞ @n
ð
1
þ fðy, tÞdy
V H O H ð tÞ
ð ð
@fðy, tÞ 1
¼ Gðx, y, tÞdSy þ fðy, tÞdy
ySðtÞ @n V H OH ðtÞ
ð ð
¼ V ðtÞGðx, y, tÞdSy þ FðV ðtÞÞGðx, y, tÞSdy
SH ðtÞ SN ðtÞ
ð
1
þ fðy, tÞdy
VH OH ðtÞ
ð ð
¼ V ðtÞ Gðx, y, tÞdSy þ FðV ðtÞÞ Gðx, y, tÞdSy
SH ðtÞ SN ðtÞ
ð
1
þ fðy, tÞdy:
VH OH ðtÞ
The flow field is given by where k is a unit vector along the axis of the neck.
ð We note that according to Eq. 8, as the number of
saturated proteins increases, the hydrodynamic
∇fðx, tÞ ¼ V ðtÞ ∇x Gðx, yÞdSy
SH flow begins to dominate the diffusion. In
ð Holcman and Schuss (2004), and Holcman
þ FðV ðtÞÞ ∇x Gðx, yÞdSy : et al. (2004), we connected the strength of flow
SN
field to the number of bound myosin molecules
induced by calcium. This results in nonlinear
In the neck, due to the symmetries and the
coupled partial differential equations, that can
uniform initial conditions, we simplify the flow
be solved numerically by stochastic simulations.
field by assuming its velocity is parallel to the
axis of the neck. It is given by
Rate Constants and Molecular Dynamics
We shall now recall how to model chemical reac-
∇fðx, tÞ ¼ Vðx, tÞ ¼ FðV ðtÞÞk, tions, described by the backward- and the
forward-binding rate, which are usually obtained for the initial concentration condition. When the
in aqueous solution, where diffusion is not lim- binding sites are located on the boundary,
ited by space. In confined microdomains where the narrow escape formula should be used
the number of ions involved can be small, the (Holcman and Schuss 2013).
binding rates have to be reinterpreted.
Modeling the Interaction of Calcium Ion with
Backward-Binding Rate The mean time that Surface or Receptors C
two molecules react chemically is modeled as We shall now specify the interaction conditions
the mean time the first molecule stays imprisoned between a calcium ion and a receptor or
in the potential well of the second. The random a membrane, the model at a molecular and popu-
time interval between the binding and the lation level, and the physical laws that can be
reappearance of the binding molecule into the derived from elementary physical principles.
free state is exponentially distributed with a rate
constant equal to the backward-binding reaction. Absorption
The exponentially distributed waiting time for Absorption at surface @O is the process by which
the backward reaction is based on Kramers’ the- a particle is removed after it hits @O, which can
ory of activated barrier crossing, as described in be an artificial interface such as the one between
Matkowsky et al. (1982). Thus for transient the spine and the dendrite or an effective one such
chemical reactions, each molecule reacting with as a channel. Indeed, during a simulation, when
a calcium ion has two consequences: first, the an ion hits such a surface, it disappears. In gen-
molecule can become activated, and second, the eral, absorption on a surface is modeled by killing
time course of calcium is delayed. The unbinding the trajectories when it encounters or passes over
events are modeled as Poissonian processes. Fix- the surface @O. Thus the probability density func-
ing a scale Dt, the probability to unbind is kbDt tion to make a transition from x @O to any
(take a uniform variable and check whether it is point y during time t is zero: p(y, t|y) = 0.
above or below kbDt).
Partial Absorption
Forward-Binding Rate The forward-binding Partial absorption accounts for a probability that
rate Kfor corresponds to the flux of particles to a particle arriving at a surface is reflected with
the binding sites. Contrary to the backward- a probability p or absorbed with probability 1 – p.
binding rate, this rate does not contain local prop- This condition can be calibrated to describe the
erties only, but includes the effect of the global interaction between the moving particle and the
geometry of the domain, where the chemical binding site of a receptor. This condition is for-
reaction occurs. Such a rate has been computed mulated at a molecular level for stochastic simu-
at equilibrium by Smoluchowski and can be lations or with a probability density function to
converted as the effective radius Ra of a ball describe the macroscopic level. This condition is
that mimics the binding site and so that the aver- called also radiation or reactive or Robin bound-
age probability that an ion meets such ball is ary conditions and has been widely used to
equal to the forward rate. The radius is calibrated describe diffusion in a biological cell with chem-
according to the formula ical reactions on its surface (Andrews and Bray
2004; Batsilas et al. 2003; Berezhkovskii
et al. 2004; Erban and Chapman 2007; Monine
K for ¼ 2pRa D Ca2þ , (10)
and Haugh 2005; Lamm and Schulten 1983; Tai
et al. 2003; Zwanzig 1990).
where [Ca2+] is the initial calcium concentration
and D the diffusion constant. This calibration can Schematic Description
also be used to estimate the effective radius of The overdamped Langevin equation can be writ-
a binding molecule in a transient state, calibrated ten as a stochastic equation
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
x_ ¼ aðx, tÞ þ 2sðx, tÞw: _ (11) xDt ðsÞ ¼ x (13)
The process x(t) defined by Eq. 11 with par- in the interval x > 0, for 0 t s T, with
tially absorbing boundaries can be defined as the Dt = T/N, t – s = iT/N (i = 0,1,. . ., N), where for
limit of Markovian jump processes generated by each t the random variables Dw(t, Dt) are nor-
the Euler scheme mally distributed and independent with zero
mean and variance At. In dimension one,
with boundary at 0, the boundary behavior
xDt ðt þ DtÞ ¼ xDt ðt þ aðxDt ðtÞ, tÞDt
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi for the simulated trajectories that cross the
þ 2sðxDt ðtÞ, tÞD þ wðt, DtÞ for t s boundary, identified by xDt ðtÞ þ aðxDt ðtÞ, tÞDtþ
(12) pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2sðxDt ðtÞ, tÞ Dw < 0, is described by
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffi
xDt ðtÞ þ aðxDt ðtÞ, tÞDt þ 2sðxDt ðtÞ, tÞDw w:p:1 P Dt
xDt ðt þ DtÞ ¼ (14)
terminate trajectory otherwise:
Thus the exiting trajectory is normally Robin boundary condition and the absorption
reflected w.p. parameter
P(x) are related by
pffiffiffiffiffi
R ¼ 1 P Dt (15) pffiffiffiffiffiffiffiffiffiffiffiffi
kðx, tÞ ¼ rPðxÞ sn ðtÞ, x1 ¼ 0, (17)
and is otherwise terminated (absorbed). The scal-
pffiffiffiffiffi pffiffiffi
ing of the termination probability with Dt with r ¼ 1= p and sn(t) = nTs(t)n. The relation
reflects the fact that the discrete unidirectional (17) can also be adapted for curved boundaries
diffusion current atpffiffiffiffiffi
any point, including the and applied to the tangent plane at each point of
boundary, is O 1= Dt (Singer et al. 2008). the boundary. Indeed this is due to the fact that
This means that the number of discrete trajecto- a smooth local mapping of the domain to a half
ries hitting or crossing the boundary
pffiffiffiffiffi in any finite space with an orthogonal system of coordinates
time interval increases as 1= Dt: preserves the constant isotropic diffusion matrix,
though the drift changes according to Itô’s for-
Partial Reflecting Condition in Dimension mula. In this case the vector v coincides with the
Larger than 2 normal n.
The scheme (14) is generalized to diffusion with The reflection law and the relation are new for
drift and anisotropic constant diffusion matrix diffusion in higher dimensions. The constant r for
s(t) in the half space, x1 > 0, with partial oblique the Euler scheme is not the same as that for other
reflection. The Robin boundary condition is schemes, e.g., for a discrete prandom
ffiffiffi walk with
recovered if and only if trajectories are reflected radiation boundaries, r ¼ 1= 2: The reflection
in the direction of the unit vector can be constructed explicitly. Indeed, the
d-dimensional stochastic dynamic is
sn
v5 , (16)
ksnk pffiffiffi
x_ ¼ aðx, tÞ þ 2BðtÞw_ (18)
where n is the unit normal to the boundary. The
radiation parameter k(x,t) in the d-dimensional in the half space

O ¼ x ¼ ðx1 , x2 , . . . , xd Þℝd : x1 > 0 The Fokker-Planck equation for the pdf of x(t)
can be written as
where w is a vector of d independent Brownian
motions and when we assume that the diffusion @pðy, tjx, sÞ
¼ ∇y Jðy, tjx, sÞ for all y, xO:
tensor s(t) = B(t)BT(t) is uniformly positive def- @t
inite for all t s. We use henceforward the (21)
abbreviation s(t) = s. The radiation condition C
(35) becomes If x O, but
pffiffiffi
Jðy,tjx, sÞ n ¼ kðy,tÞpðy,tjx,sÞ, for y @O, x O, x0 5 x þ aðx, tÞDt þ 2BðtÞ Dwðt, DtÞ 2
= O,
(19)
the Euler scheme for Eq. 18 with oblique
where the components of the flux vector reflection in @O reflects the point x0 obliquely
J(y, t | x, s) are defined by in the constant direction of v to a point x00 O,
as described below. First, we denote by xB0 the
normal projection of a point x0 on @O, that is,
Jk ðy, tjx, sÞ ¼ ak ðy, tÞpðy, tjx, sÞ
xB0 = x0 (x0 n)n. Then we write the Euler
Xd
@ j, k scheme for Eq. 18 with partially reflecting bound-
þ s pðy, tjx, sÞ : (20)
j¼1
@yj ary as
8 0 0
< x for x O pffiffiffiffiffi
xðt þ DtÞ ¼ x w:p: 1 P x0B Dt, if x0 2
00
= O, (22)
: pffiffiffiffiffi
terminate trajectory w:p: P x0B Dt, if x0 2
= O:
2y0 n
The value of the termination probability
pffiffiffiffiffi y ¼ y0 v (24)
v1
P x0B Dt, that varies continuously in the
boundary, is evaluated at the normal projection
of the point x0 on the boundary. The oblique is the oblique reflection of y0 (see Fig. 2). If the
reflection in the direction of the unit vector scheme described above is not used, a paradox
v(v1 6¼ 0) is defined by can arise (Singer et al. 2008): while the pdf of the
solution of Eqs. 12 and 13 converges to the solu-
2x01 tion of the FPE Eq. 32 and the initial condition
x00 ¼ x0 v: (23) Eq. 34, it does not satisfy the boundary condition
v1
Eq. 35, leading to a boundary layer, due to the
diffusion approximations in the Markovian jump
Note that x100 = x10 guarantees that the
process.
reflected point of a crossing trajectory is inside
the domain O. The fact that the normal compo-
nents of x00 and x0 are of equal lengths makes the Generic Modeling of Calcium Ions at Pumps or
high-dimensional boundary layer analysis similar Exchangers
to that in one dimension. Normal reflection cor- Partial absorbing boundary condition can be used
responds to v = n = (1, 0, . . ., 0). We note that for to model the behavior of calcium ion near
a point y O, we can write Pr{x00 = y} = Pr a channel or a pump. However cooperativity
{x0 = y0 }, where should be implemented for each case at hand.
(GluR2-calcium permeable), and voltage-

sensitive calcium channel (VSCC) for simulating
the flux of calcium inside neuronal cells. The flow
of ions through open channels has been studied
using Langevin equation (Eisenberg et al. 1995;
v y' Roux et al. 1995). The fluxes of ions can be
implemented numerically as follows.
n y
Calcium Influx Through NMDAR Calcium
influx through NMDA channels can be approxi-
x mated by (Koch 1999, p. 99)
et=tN, 1 et=tN, 2
I N ðtÞ ¼ gN ðV m EN Þ,
1 þ 0:33½Mg2þ e0:06V m
(25)
Calcium Dynamics in Neuronal Microdomains: where the conductance is gN = 0.16 nS and the
Modeling, Stochastic Simulations, and Data Analy- Nernst potential En = 0, Vm is the membrane
sis, Fig. 2 Reflection Scheme. A simulated trajectory potential. When the potential Vm is fixed and
can get from x to y in a single time step Dt in two different the fraction of current carried by calcium ions is
ways: (i) directly from x to y, without crossing the bound-
ary, and (ii) by crossing the boundary from x to y0 and 15 %, we can simulate such a flux by injecting
reflection in the oblique direction v with probability particles at random times such that the instanta-
pffiffiffiffiffi
1 P y0B Dt to y. The reflection law Eqs. 22–24 satisfies neous rate is the one obtained from relation
y10 = y1 Eq. 25. Indeed, the entrance is a Poissonian
process with a time-dependent rate
ðtÞDt
lðt þ DtÞ ¼ ICa2e : The number of entering ions
For example, after entering inside an exchanger, I Ca ðtÞDt Ð
an ion takes a certain time to exit: this is modeled is N ðtÞDt ¼ 2e , where ICa(t)Dt = t+Dt t
IN(t)ds.
by changing a partial reflecting boundary condi- An example of stochastic calcium ion entry is
tion to a reflecting one as long as the ion is inside presented in Fig. 4 (discretized at a time step
the exchanger. Dt = 0.1 ms). In some dendritic spine models
Some pumps can work with several ions. (Holcman et al. 2004; Holcman and Schuss
As the intrinsic biophysical mechanism of 2004), the entrance dynamics is neglected and
permeability is not necessarily understood, there ions are initially placed at channels, located
is no consensus for a universal coarse-graining at the top of the dendritic spine Ð 1 head.
scheme of ion extrusion. If two ions are required, Typically, the total charge is QN = 0 I(s)ds;
we propose that once the first ion enters the thus the fraction of calcium is QCa = 0.15 QN
channel, it cannot move before another has hit Coulomb. The number of calcium ions entering is
the binding area. If the second ion enters while N N , Ca ¼ Q2eCa , where the e is the electron charge.
the first one is not returned (after an exponential Using parameters of Table 1, we obtain that
waiting time), then the first one can be extruded. QN = 6.38pC and thus there are about NN,
During that time, no other ions can enter the Ca = 3, 000 ions entering in average inside
channel. a single NMDA receptor (for a fixed mean volt-
age Vm = 65.1 mV).
Modeling Calcium Influx from Channels: NMDA
Receptors, AMPA Receptors, and VSCC Calcium Influx Through AMPAR The sto-
We shall now present examples of three chastic arrival of ions entering through an
classical channels such as NMDA, AMPAR AMPAR is computed exactly with the same
Calcium Dynamics in Neuronal Microdomains: dV
Modeling, Stochastic Simulations, and Data Analy- C ¼ I Na ðV, nÞ I K ðV, n I L ðV
dt
sis, Table 1 Calcium entry into dendritic spines
I Ca ðV, m, h gI N I A
Parameter Description Value
dx
Fraction of NMDAR current 15 % ¼ 0:1ax ð1 xÞ bx x, for x ¼ n, m, h,
carried by Ca2+ dt
Total charge entering IN 6.38
(NMDAR) 10–15 C where the currents are C
Calcium ions entering through 3,000
NMDAR ions I Na ¼ gNa p3 ð0:89 1:1nÞðV ENa
gN NMDAR conductance 0.16 nS I K ¼ gK n4 ð V EK Þ
EN Equilibrium potential 0 mV I Ca ¼ gCa m3 hðV m ECa Þ
(NMDAR) I L ¼ gL ð V EL Þ
tN,1 NMDAR time constant 11.5 ms
tN,2 NMDAR time constant 0.67 ms
with parameters
Fraction of AMPAR current 1.4 %
carried by Ca2+
11.51
ap 1 yk V m
Total charge entering IA p¼ , ak ¼ ,
(NMDAR) 1015 C a p þ bp tk eyk V
tk
m
1
Calcium ions entering through 500 ions V m þ65
NMDAR bk ¼ k e sk
, for k ¼ n, m, h, p:
gA Conductance (AMPAR) 0.3 nS
EA Equilibrium potential 0 mV where g and are summarized in Table 1. The
(AMPAR) total charge becomes QV = 0.6 fC, which leads
tA AMPAR time constant 0.2 ms approximately to 2,000 ions entering
Total charge entering IV 0.64 through VSCC.
(NMDAR) 10–15 C
Calcium ions entering through 2,000
VSCC ions Stochastic Simulation of Calcium
in a Dendritic Spine
The major benefit of stochastic simulations is to
access the total number of biochemical bonds
induced by calcium ions on specific molecules
method as for NMDA, except that the mean flux and to quantify the amount of structural changes
is given by occurring at the spine level. There are many other
consequences such as computing the hydrody-
t t=tA namic component that changes the nature of the
I A ¼ gA e ðV m EA Þ, (26) ion trajectories or distinguishing the periods of
tA
calcium dynamics. Novel coarse-grained equa-
where gA = 0.3 nS and the Nernst potential tions are derived in (Holcman and Schuss 2004).
Ea = 0 mV. The fraction of AMPAR current
carried by calcium ions is 1.4 %; thus we find Space Exploration
that the total charge is QA = 11.51 fC, leading to The geometric characteristics of ionic trajectories
approximately 500 ions. with the hydrodynamic flow are distributed dif-
ferently from pure diffusion (Holcman and
Influx Through VSCC Calcium influx through Schuss 2004) (see Fig. 3). Not only the nature of
VSCC requires computing the changes in the the movement is different, but the hydrodynamic
membrane potential depolarization. One possibil- flow causes the ions to drift in the direction of the
ity is to use the simplified Hodgkin-Huxley neck, and consequently the time they spend in
model (Hille 2001). The voltage change follows the spine head is reduced. As a consequence,
the dynamics the probability of a trajectory to leave through
Calcium Dynamics in Neuronal Microdomains: at the top of the spine head where channels are located and
Modeling, Stochastic Simulations, and Data Analy- continues until it is terminated at the dendritic shaft or at
sis, Fig. 3 The filling of space by five random trajec- an active pump. The parameters for the simulation are
4 1
tories in the spine with no drift (a) and with drift (b) d1 = 0.02 mm, d2 = 0.01 mm, KAM back = 10 s
3 1
Each color corresponds to a trajectory. Proteins are uni- Kback = 2.10 s , R = 0.5 mm, d/2 = 0.21794 mm, l =
cal
formly distributed in the spine head and are represented by 1.5 mm, Npumps = 10
circles and crossed circles, respectively. A trajectory starts
a pump located in the head decreases. Similarly, It was shown that the first period corresponds to
the probability to return to the head from the a fast calcium extrusion, measured in (Majewska
spine neck is reduced if it has to diffuse upstream, et al. 2000) with an exponential decay rate
against the hydrodynamic drag force. Thus the l = 0.14 s1 and is due to the diffusion of satu-
ionic trajectory stays inside the spine a shorter rated buffers, binding kinetics of endogenous
time in the presence of the hydrodynamic flow, as buffers, diffusion of buffers, buffered calcium
compared to the time without it. As discussed in diffusion across the spine neck, and the effect of
(Holcman et al. 2004), the total number of bound the pumps. The first period dynamics is defined
molecules can change as much as 30 % with and by the fast binding to calcium stores (Majewska
without the flow. et al. 2000; Sabatini et al. 2001). On the other
hand, in the simulation resulting from the model
Two Stages of Calcium Concentration Decay (Holcman et al. 2004), based on fast spine motil-
Two very distinct decay rates of the fast extrusion ity, the first time period has an exponential decay
periods have been reported in (Majewska rate, constant lt = 0.16 s1 derived in Holcman
et al. 2000; Sabatini et al. 2001). The second and Schuss (2004). The decay seems to be
decay period is identified as purely driven by a consequence of the dynamics created by the
random movement, and the time constant equals push effect, since stores were neglected. Further
the first eigenvalue of the Laplacian on the spine studies that include large numbers of buffers
domain, with the adequate boundary conditions. should reveal the precise contribution of buffers
to the calcium fast decay rate, as compared to the absorbing sites. Inside a structure that
rate imposed by the spine contraction. contains a high concentration of binding
sites, a different simulation approach is
Multiscale Modeling of Connecting needed: either to derive a homogenized equa-
a Continuum Bath with Single Molecular tion or the dynamics in the wells should not be
Dynamics There are various interesting coarse grained.
multiscale modeling approaches where the goal • Partial absorption. When there are many C
is to connect a discreet description of Brownian absorbing sites located at close proximity,
particles with a continuum (Franz et al. 2013; nonlinear effects should be taken into account.
Flegg et al. 2012). The conservation of fluxes at It is possible to derive homogenized boundary
the connecting interface generates boundary conditions (see for details (Taflia and
layer behavior that needs to be specifically Holcman 2011)).
studied. • Coarse graining a simulation with the nar-
row escape rate. Instead of running a full
Recipe for a Successful Stochastic Simulation Brownian simulation where the position of
• Choosing a time step for a simulation. The each calcium ion and calmodulin molecule is
time step of a simulation is critical: when it is computed with the Euler scheme, it is possible
too small, the simulations take forever, while to coarse grain it into a rate model based on the
if it is too large, then binding to buffers of narrow escape theory (Schuss and Holcman
a certain size a is missed. For a time step Dt 2013; Holcman and Schuss 2013a). The gain
and a diffusion coefficient D, a Brownian par- of this procedure is to avoid lengthy simulation
ticle moving in three dimensions jumps to imposed by the smallest length. In a Brownian
a distance Dx2 = 6DDt. Thus the constraint simulation, due to the small calmodulin-
|Dx| = a/4 provides an estimate for the time binding site of radius RCaM = 2 nm, in order
step. In general the size of the smallest target to ensure that the process does not jump over
defines the condition for the time step Dt. In the site, the MSD formula leads to a time step of
most cases, where target sites are fixed, it is 106 ms (with DCa = 200 mm2 ms 1), which
more interesting to refine the time step when would lead to days of simulations. To circum-
a particle enters its neighborhood. vent this difficulty, it is possible to compute
• Positioning a particle that has unbound. In directly the rate of arrival of calcium ions to
the Smoluchowski limit equation (where there one of the free calmodulin-binding sites. The
is no dynamic for the velocity), the coarse mean first binding time tB of a Brownian par-
graining of a potential well is usually done ticle diffusing with diffusion coefficient D1 in
by freezing the position of the particle at the a spherical domain O, to a small spherical tar-
binding position during an exponential get (radius r) diffusing with coefficient
waiting time, the rate of which is the backward tB ¼ 4pðDj1OþD
j
2 Þr
: Due to the small size r, the
rate (reciprocal of the mean time to escape the
binding times are exponentially distributed
well). After the particle unbinding, where
with a mean l1 ¼ 1=tB : For N-independent
should the particle be positioned? Certainly
calcium ions diffusing within O, the first bind-
not at the absorbing boundary defining the
ing time Tb to a calmodulin site is distributed
well; otherwise the particle is immediately
with rate
absorbed, unless the boundary condition is
changed from absorbing to reflecting during
a certain refractory time. Another possibility is lðN Þ ¼ l1 N: (27)
to position the particle outside the boundary
layer of the binding site: a small distance away Thus, the probability that a binding event
from the site (3–4 radii away). This is possible occurs between time t and t + Dt is ℙ(t
if the sites are not surrounded by other Tb < t + Dt) = l(N)Dt. An application consists
b Ca2+
NMDA Ca2+
a 40
# Ca2+/ms
35 Ca2+ entry Calciumions
(Δ t = 0.1 ms) CaM
30 +
ActivCaM
25
Calciumchannels
# Ca2+
20 Calciumpumps
15 PSD
10
0
0 10 20 30 40 50
Time (ms)
Ca2+
Calcium Dynamics in Neuronal Microdomains: receptor. (b) Schematic representation of the calcium
Modeling, Stochastic Simulations, and Data Analy- and calmodulin (CaM) pathway inside a dendritic spine
sis, Fig. 4 Calcium entry and dynamics. (a) Time course (Guerrier and Holcman 2014a)
of calcium entry inside a spine through an NMDA
in replacing the Brownian movement of cal- Diffusion Laws in Microdomains with Small
cium ions in a spine head by the binding rate Openings
l(N)Dt at each calmodulin site. We thus need to Synaptic input creates calcium transients in sin-
compute at each time step the number N of free gle dendritic spines and dendrites (Fig. 6). We
calcium in the spine. Furthermore, calcium ions shall now present the modeling approach used to
can escape a dendritic spine head Ohead through study calcium transients.
small pumps. To model this calcium escape, we
use a similar method as described above where Probability Density Function of an Ion
we approximate the binding of calcium to The stochastic nature of the calcium motion
a pump by using the mean first passage time requires a probabilistic approach. Indeed, the
of a diffusing ion to a small target located on location of an ion is not certain, and the probabil-
the boundary. When there are N calcium ions, ity p(x,t) to find an ion at time t at a position
the first binding time is x satisfies the standard Fokker-Planck or diffu-
sion equation
4Rpump DCa N
mðN Þ ¼ : (28)
jOjhead @p
ðx, tÞ ¼ DDpðx, tÞ, (29)
@t
It is also possible to account for the com-
petition with escaping through the spine neck where D is the Laplacian operator and D is the
using the narrow escape formula for a spine diffusion constant in the cytoplasm. The solution
(Eq. 45). This procedure accelerates the sim- of Eq. 29 requires specifying initial and boundary
ulations and gives excellent results (Fig. 5). conditions and allows the entire characterization
In Figs. 4, 5, we present a simulation of calcium of a transient regime or the steady state distribu-
entry in a dendritic spine binding to calmodulin tion of a single ion. For a general domain, the
(Guerrier and Holcman 2014a). solution cannot be derived analytically, but it is
a1 a2
3000 3000
2500 2500
Nbr free Ca
2000 Nbr tot of Ca in spine 2000
Nbr of Ca entered
1500 Nbr of Ca out

1500 C
Nbr of Ca bond
Nbre of Ca binding sites
1000 1000
500 500
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Time (ms) Time (ms)
b1 200 b2 200
Free CaM
CaM−Ca
CaM−2Ca
150 CaM−3Ca 150
CaM−4Ca
# CaM
# CaM
100 100
50 50
0 0
0 50 100 150 200 250 300 0 50 100 150 200 250 300
Time (ms) Time (ms)
Calcium Dynamics in Neuronal Microdomains: (dark blue), can be bound to calmodulin (magenta), or
Modeling, Stochastic Simulations, and Data Analy- can leave the spine (green). The curves are compared with
sis, Fig. 5 Calcium dynamics: comparing a Brownian the total number of ions (black) and the total amount of
simulation with a coarse-grained rate model for bind- calcium in the spine (free plus bound) (in red). B1 and B2:
ing to calmodulin. A1 and A2: Time course of calcium for Activation of calmodulin during the Brownian simulation
3,000 ions entering a dendritic spine through NMDA (B1) and the rate model (B2). A calmodulin molecule can
receptors containing 200 calmodulin. The Brownian time have zero, one, two, three, or four calcium ions bound
step is Dt = 3 106 ms (A1). For the rate model, (black, dark blue, green, magenta, and red respectively)
Dt = 104 ms (A2). Ions diffuse freely in the synapse
possible to obtain long- and short-time asymp- where N is the initial number of ions. Ignoring
totic estimations. As we shall see, these expres- at this stage the effect of any chemical
sions provide the dependency with respect to reaction, the microdomain geometry O is the
many geometrical parameters. When necessary, main determinant of the characteristic time
numerical simulations are used to obtain the scale involved in diffusion. When the boundary
missing information. They are usually tedious to decomposes into two parts, @Oa the absorbing
obtain and require careful discretization of the part, made of key fast-binding elements, and
domain, especially when small and large scales @Or the reflective part, then it is usually
are present. a critical aspect to quantify the mean time ðtÞ
For many independent calcium ions, the confor an ion to reach @Oa. The boundary
centration c(x,t) is given by c(x,t) = Np(x,t), conditions are
Calcium Dynamics in
Neuronal Microdomains:
Modeling, Stochastic
Simulations, and Data
Analysis, Fig. 6 Upper:
Isolated calcium
transients in a single
dendritic spine. Synaptic
activity generates
a transient calcium change
in a spine and a whole
dendritic segment,
triggered by a back-
propagating action
potential. Lower:
Spontaneous calcium
activity in neurons.
Synchronized and not
synchronized activity as
well as small (probably
EPSPs), larger (probably
single spike), and very
large events (burst of
several spikes). The decay
phase of calcium is
approximated by
exponentials or sum of
exponentials
@pðx, tÞ diffusion equations can be found in Carslaw and

¼ 0 on @Or ,
@n Jaeger (1959) and Crank (1975). Although
cðx, tÞ ¼ 0 on @Oa : expression (30) justifies fitting a sum of exponen-
tials to experimental data, connecting the eigen-
The general solution of the diffusion equation values with the precise geometry is in general
can be formally expanded as very difficult, except in a few cases where the
geometry contains a narrow passage or small
X
1
hole. This is the case of a dendritic spine or
cðx, tÞ ¼ ck uk ðxÞexpðlk tÞ (30)
k¼1
a narrow domain in dendrites. When |@Oa|
|@Or|, the reciprocal of the mean time to escape

where lk > 0 are the eigenvalues; uk, k = 1, are a domain ðtÞd is the first eigenvalue (Schuss
the eigenfunctions; and ck are the constants. The et al. 2007). Indeed, l10 is usually very large, so that
general explicit computation of the solution of there is a large gap with the rest of the eigenvalue
l0
l1, and thus the solution 30 can be further (Eq. 11) on the positive axis with drift and with
approximated by a single exponential for a time a variable diffusion coefficient is
t l11 ,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
kðtÞ ¼ rP sð0, tÞ, r ¼ pffiffiffi : (36)
cðx, tÞ c0 expðl0 tÞ: (31) p
We conclude that for domains with narrow The relation Eq. 36 is true for diffusion with C
pffiffiffi
neck, the arrival of diffusing particles to the variable coefficients. The value r ¼ 1= p is
small domain is almost Poissonian. This is different than values obtained
pffiffiffifor other schemes,
a consequence of the geometry. e.g., than the value r ¼ 1= 2 , predicted by the
discrete random walk theory of radiation bound-
Transition Probability Density Function with aries (Collins and Kimball 1949). Values of r for
Partial Reflecting Boundary Condition other schemes are given in Erban and
The transition probability density function (pdf) Chapman (2007).
of the limit process 11, p(y, tjx, s) =
Pr{x(t) = yjx(s) = x}, is the solution of the FPE Exit Diffusion Rate from Dendritic Spines
We summarize in this section the approach used
@pðy, tjx, sÞ @ ½aðy, tÞpðy, tjx, sÞ to derive asymptotic formulas for the rate of
¼ diffusional exit from the spines. A dendritic
@t @y
spine with a narrow neck has very degenerate
@ 2 ½sðy, tÞpðy, tjx, sÞ geometry (Andrews and Bray 2004), but the exit
þ , (32)
@y2 time of a diffusing particle, which is the
reciprocal of the first eigenvalue, can be
or equivalently, directly measured from fluorescence imaging.
We shall now recall the main formula for the
@pðy, tjx, sÞ @J ðy, tjx, sÞ exit rate that was obtained in the context of the
¼ for all y, x > 0,
@t @y narrow escape (NET) and dire strait (DST) the-
ory (Holcman and Schuss 2014; Schuss and
where Holcman 2014).
A free Brownian particle moves in a bounded
@ ½sðy, tÞpðy, tjx, sÞ domain D ℝd(d = 2, 3), whose boundary @O is
J ðy, tjx, sÞ ¼ aðy, tÞpðy, tjx, sÞ
@y sufficiently smooth (the analysis in higher dimen-
(33) sions is similar to that for d =3). The Brownian
trajectory x(t) is reflected at the boundary, except
is the flux. The initial condition is for a small hole @Oa, where it is absorbed, as
shown in Fig. 7. The reflecting part of the bound-
pðy, tjx, sÞ ! dðy xÞ as t # s, (34) ary is @Or = @O @Oa. The lifetime in O of
a Brownian trajectory that starts at a point x O
and the radiation boundary condition is is the first passage time t of the trajectory to the
absorbing boundary @Oa. The NET
J ð0, tjx, sÞ ¼ kpð0, tjx, sÞ, (35)
vðxÞ ¼ ½tjxð0Þ ¼ x (37)
where k is a constant related to the constant c and
to the values of the coefficients at the boundary. is finite under quite general conditions (Schuss
The no flux and Dirichlet boundary conditions are 2010b). As the size (e.g., the diameter) of the
recovered if c = 0 and c = 1, respectively. The absorbing hole decreases to zero, but that of the
relation between the reactive “constant” k(t) and domain remains finite, the NET increases indefi-
the absorption parameter P for the dynamics nitely. A measure of smallness can be chosen as
j@Oj1=ðd1Þ
¼ Oð1Þ for e
1: (39)
jOj1=d
The NET v(x) can be obtained by solving the

Pontryagin-Andronov-Vitt (PAV) mixed bound-
ary value problem for Poisson’s equation
(Pontryagin et al. 1933; Pontryagin et al. 1989;
Schuss 2010b)
1
DvðxÞ ¼ for x O (40)
D
Calcium Dynamics in Neuronal Microdomains: vðxÞ ¼ 0 for x @Oa (41)

Modeling, Stochastic Simulations, and Data Analy-
sis, Fig. 7 Brownian escape from a spherical window
@vðxÞ
¼ 0 for x @Or , (42)
@nðxÞ
the ratio between the surface area of the absorb-
ing boundary and that of the entire boundary, for where D is the diffusion coefficient and n(x) is
example, the unit outer normal vector to the boundary
at x @O. For a circular window of radius
j@Oa j 1=ðd1Þ a
j@Oj1/2 (Fig. 7),
e¼
1, (38)
j@Oj
provided that the isoperimetric ratio remains

bounded:
jOj
t ¼ for a
j@Oj1=2 : (43)
Lð0Þ þ N ð0Þ
4aD 1 þ aloga þ oðalogaÞ
2p

1 ‘þ 3=2 V
The MFPT to the absorbing boundary at the t ¼ pffiffiffi ð1 þ oð1ÞÞ for a0
‘þ , (44)
2 a0 ‘þ D
end of the funnel of a solid of revolution
obtained by rotating the symmetric planar
domain composed of a bulky head connected where V = jO0 j is the volume of the domain. The
to a narrow neck formed by two tangent circular NET tx!@Oa of a diffusing particle from a three-
arcs of radius l+ separated by a distance a0 dimensional domain ‘O’ with a bottleneck in the
(see Fig. 8 in Holcman and Schuss (2013)) is form of a narrow circular cylinder of cross-
given by sectional area pa2 is given by
1 0.5
0.8 0.4
0.6 0.3 C
T3D (S)
T3D (S)
0.4 0.2
0.2 0.1
0 0
0.5 1 1.5 2 0 1 2 3 4 5
Radius of spine a (mm) Neck Length L(mm)
Calcium Dynamics in Neuronal Microdomains: spine. Left: mean residence time of calcium t3D as
Modeling, Stochastic Simulations, and Data Analy- a function of the spine radius a. Right: t3D is plotted as
sis, Fig. 8 Residence time of calcium in a dendritic a function of the spine length L
8
> jO1 j a R Oð1Þ L2 jO1 jL
>
> 1þ log þ þ þ
>
> 4aD pR a D 2D pa2 D
>
>
>
> solid spherical head of radius Rsconnected tos
the neck ats a ffiffiffiffiffiffiffiffiffiffiffi
right angle
>
> " 1=2 ffiffiffiffiffiffiffiffiffiffiffi ffiffiffiffiffiffiffiffiffiffiffi !#
>
>
>
> jO1 j Lx þ Rx @Oa j@O1 j j@Oa j j@Oa j
>
> 1þ log þo log
< 4aD 2p p j@O j a j@O j 1 j@O j1
tx!@Oa ¼ (45)
>
> L2
jO1 jL
>
> þ þ
>
> 2D pa2 D
>
> a general head connected to the neck at a right angle
>
>
>
> 1 Rc 3=2 jO1 j L2 jO1 jL
>
>
>
> pffiffiffi ð1 þ oð1ÞÞ þ þ
>
: 2 a Rc D 2D pa2 D
a general head connected smoothly to the neck by a funnel,
where Rc is the curvature at the cusp. The extrusion) was initially presented in Svoboda
asymptotic expression is derived in Holcman et al. (1996). This term accounts for the many
and Schuss (2011). The order 1 term can be returns of an ion between the spine neck and
computed for the sphere using the explicit head (Biess et al. 2007). This term is not present
expression of the Neumann-Green function when an ion cannot return to the head once it
(Cheviakov et al. 2010). When the spine radius enters the neck (Korkotian et al. 2004). Because
is small, the leading order term for the mean exit the other terms in formula 45 diverge to infinity,
time (which is also the rate of diffusion their contribution cannot be neglected and they
affect significantly the residence time of a diffus- when there is cooperativity at pumps, more elab-
ing particle in a dendritic spine (see Fig. 8). We orated boundary conditions are needed (Hille
conclude from these analytical formulas that the 2001) and will be discussed later on. There is no
spine connection determines the rate of extrusion. general solution of such equations. If binding is
For short spines, all terms are significant. fast, since cðx, tÞ þ ½c Bðx, tÞ ¼ N 0 ell t , the
decay rate depends on the unbinding time, such
Remarks The presence of a spine apparatus that for a long enough time (so that binding
inside the spine might affect the extrusion of occurred), the asymptotic decay rate is well
calcium or any other diffusing molecules. In approximated by a sum of two exponentials
addition, molecular binding affects calcium
extrusion, and the rate is no longer Poissonian. cðx, tÞ ¼ Aell t þ Bekb t , (49)
Influence of Calcium Buffers on the Residence where A and B are constants

Time of Calcium in Microdomains such as the
Dendritic Spines Residence Time of Calcium in the Dendritic Spines
The residence of calcium in a spine can be with a Hydrodynamic Flow
influenced by other mechanisms than pure diffu- Dendritic spines can change shape in a few hun-
sion. Binding and unbinding to calcium buffers dred milliseconds (Fischer et al. 1998; Holcman
affect the time course. In that case, the entire sys- et al. 2004), after calcium ions flow in. This fast
tem of partial differential equations (PDE) describ- change of shape decreases the spine head volume.
ing the process should be solved. Only numerical Spine motility was proposed by Blomberg
simulations are available. The generic example is et al. (1977), and the fast twitching movement
of the spine was anticipated by F. Crick in (1982).
kf
Ca þ Bfree
)* Ca B,
(46) Spine fast contraction was attributed to actin-
kb myosin molecules or troponin C, which were
observed inside the spine head. As in muscle
which leads to the PDE system of equations
cells, high concentrations of actin molecules indi-
@c cate that rapid movement can follow the arrival of
ðx, tÞ ¼ DDcðx, tÞ kin cðx, tÞBðx, tÞ calcium ions. It has been proposed in Holcman
@t
et al. (2004) that calcium ions set the spine in
þ kb ½c Bðx, tÞ, (47)
motion by initiating the contraction of actin-
myosin (AM) as they bind at active sites. Each
@ ½c B
ðx, tÞ ¼ DB Dcðx, tÞBðx, tÞ molecule is assumed to give rise to a local con-
@t traction. In a simplified model, all contractions
þ kin cðx, tÞBðx, tÞ kb ½c Bðx, tÞ: add up to achieve a global contraction, neglecting
(48) anisotropic contraction due to a delay interval
between each molecule contraction. Once cal-
cB is the bound calcium that can diffuse with
cium ions enter the spine, they arrive to the bind-
diffusion coefficient DB, but stays confined in the
ing sites by diffusion and can bind there. When
spine head, while the calcium can be extruded at
four calcium ions bind to a single troponin
pumps or at the spine neck, which translates into
C protein, a local contraction of the protein
the following boundary conditions:
occurs. Adding all local contractions at a given
time produces a global contraction and induces
@cðx, tÞ
¼ 0 on @Or , a hydrodynamic movement of the cytoplasmic
@n fluid. Calcium trajectories are no longer pure
cðx, tÞ ¼ 0 on @Oa , Brownian, but contain a drift, and thus the prob-
@ ½c Bðx, tÞ ability to reach the dendritic shaft through the
¼ 0 on @Or :
@n spine neck is increased (Holcman et al. 2004).
The model requires feedback of a flow field v(x,t) interactions are difficult to access directly exper-
on the velocity. The flow field can be computed imentally due to ubiquitous pathways especially
using Green’s function of the spine domain for calcium dynamics. In a reduced model of
(Holcman et al. 2004). Adding the effect of dif- diffusion in dendrites, the one-dimensional effec-
fusion and hydrodynamics, the transient escape tive diffusion equation and an effective diffusion
rate is not necessarily a single exponential. The constant account for the presence of heterogene-
concentration inside the spine follows: ity in the medium. C

v20
cðx, y, z, tÞ ¼ C exp -l1 t þ t , Modeling Diffusion in a Heterogeneous Den-
4D dritic Cytoplasm To characterize diffusion in
where the hydrodynamic decay time is a heterogeneous dendrite, containing various
4D organelles such as mitochondria, spine apparatus,
t¼ , (50) endoplasmic reticulum, and other structures, one
v20
method (Biess et al. 2011) consists of coarse
where v0 is the initial average velocity (Holcman graining a three-dimensional cylindrical dendrite
et al. 2004). Figure 3 illustrates the effect of into a one-dimensional effective diffusion equa-
adding a hydrodynamic drift on pure diffusion tion in the limit where the space between organ-
for the exploration of the spine head. Contrary elles is small. Diffusing ions can still move inside
to the effect of buffer, the hydrodynamic effect a dendritic domain O, and the nature of the
shifts the extrusion rate, but does not lead to motion is not impaired and is well approximated
a sum of exponentials. by the Smoluchowski limit of the Langevin equa-
tion (Schuss 2010a): a particle at position X(t) at
Crowding Model of a Dendrite time t is described by
Molecular crowding reduces diffusion, and this
effect can be estimated by computing the effec- 1 pffiffiffiffiffiffi
X_ þ ∇FðXÞ ¼ 2Dw_ ðtÞ, (51)
tive diffusion coefficient of a Brownian particle g
moving between obstacles in dimension 2. For
obstacles positioned periodically, the effective where F is a potential per unit of mass, g is the
coefficient of diffusion decays nonlinearily with friction coefficient, D is the aqueous diffusion
the density of obstacles (Holcman et al. 2011). constant, and w(t) is the Gaussian white noise.
This decay also depends on the shape of the The potential F represents the effective force on
obstacles. The effective diffusion coefficient can the particle. When a moving molecule hits
be computed asymptotically by conformal mapping impenetrable organelles Oi, it is reflected.
(Holcman et al. 2011). In addition, when obstacles The distribution of independent molecules is
are positioned with additional variabilities, local characterized by the probability density function
narrow passages, funnels, or dead ends can be (pdf) p(x,t) which satisfies the Fokker-Planck
observed that lead to heterogeneity in the diffusion equation
trajectories (Ghosh et al. 2012; Holcman and
Schuss 2012, 2013). However due to the complex-
@p 1
ity of obstacle geometry, there is no achieved theory ¼ DDp þ ∇ ð∇FðXÞÞp (52)
@t g
of diffusion with obstacles in three dimensions.
We present here a local model and numerical
simulations to study calcium diffusion in in the domain Oe ¼ On[i Oi and a zero flux con-
a dendrite. The dendrite cytoplasmic medium is dition on the organelles and the dendritic mem-
highly heterogeneous and filled with many organ- e
brane @ O:
elles. Thus the motion of a diffusing particle is
affected by many interactions with its environ- @p p @F
J n ¼ D þ ¼ 0, (53)
ment. The functional consequences of these @n g @n
0.0 0.6 1.2 x (μm)
d 1 1
1 glass tube x = 0 μm x = 0.6 μm x = 1.2 μm
normalized concentration
0.8 experiment 0.8 0.8

3D-Brownian simulations
0.6 1D effective diffusion equation 0.6 0.6
0.4 0.4
0.4
0.2 0.2
0.2
0 0
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3
time(s) time(s) time(s)
1 1 1
x = 0 μm
dendrite x = 0.6 μm x = 1.2 μm

0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3
1 1 1
dendrite with spine x = 0 μm x = 0.6 μm x = 1.2 μm
0.8 0.8 0.8
0.6 0.6 0.6
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.05 0.1 0.15 0.2 0.25 0.3
Calcium Dynamics in Neuronal Microdomains: experiments. (a) Model glass pipette. Shown is the initial
Modeling, Stochastic Simulations, and Data Analy- particle distribution as taken from the experimental data
sis, Fig. 9 Brownian simulations of uncaging and the sampling volumes (white cylindrical disks) at
where J is the flux and n the outer normal of the @cðx, tÞ @ 2 cðx, tÞ
e ¼ Deff , (55)
domain O. @t @x2
To account for the overall effect of crowding
on diffusion, we adopt an approach based on where Deff ¼ 4laS D , the effective diffusion con-
a compartmentalization of the dendritic domain stant, and V = Sl, with S the cross-sectional area.
and the narrow escape theory (Holcman and The effective diffusion constant depends on two
Schuss 2013), which provides the mean time for parameters: the compartment length l and the size C
a Brownian particle to exit a domain through of the opening a.
a small absorbing opening. The dendrite is The model parameters are determined by
divided into periodic compartments of length (i) measuring the ratio of diffusion constants
l and volume V, separated from their neighbors Deff/D and (ii) a calibration condition of the
by a reflecting cross section, except for a small form l/a = 4. The calibration condition expresses
opening of radius a. This compartment should be that the parameters l and a do not necessarily
large enough so that the organelle density is the have direct physiological meanings, and we can
same in each of them. The small openings allow thus set one parameter arbitrarily within the
diffusing molecules to move across compart- limits of the small hole approximation. Addi-
ments. In contrast to previous models where tional measurements of the diffusion constant
crowding has been described by spherical obsta- will then fix the other parameter. A Brownian
cles (Biess et al. 2011) that pose barriers to dif- simulation of diffusing ions around an obstacle
fusing molecules, crowding is modeled as is presented in Fig. 9. Other applications of reduc-
a sequence of periodic compartments and small ing equations are to compute the mean time for
openings at the boundaries of neighboring com- calcium ions or diffusing molecules to travel
partments. A compartment k starts at position xk along crowded dendrites or axons.
and ends at position xk+1. The number Nk(t) of
particles in compartment k changes according to Calcium Spread Following High-Frequency
the net flux across the small windows. The flux is Stimulation The one-dimensional effective dif-
estimated by the small hole approximation (see fusion equation presented in the last paragraph
Biess et al. (2011) for details). allows analyzing calcium spread originating from
The concentration c(x,t) = N(x,t)/V(x,t) localized inputs such as synapses. At dendritic
satisfies synapses calcium can enter through NMDA
receptors. To estimate calcium spread as
a function of the synaptic input frequency, Ca2+
@cðx, tÞ 4l2 D @ @
¼ aðxÞ cðx, tÞ : (54) influx was simulated in the middle of a dendritic
@t V ðxÞ @x @x segment (Fig. 10) with buffers and pumps (see
Biess et al. (2011) for the reaction-diffusion
Similar equations have been derived in other equation). The different input frequencies are
contexts (Zwanzig 1990; Biess et al. 2007; f = 5, 10, 20, 50, 80 Hz. Interestingly, for input
Berezhkovskii et al. 2004). When the parameters frequencies larger than 20 Hz, the calcium signal
a(x) and V(x) are spatially independent, Eq. 54 in the dendrite reaches a stationary value. For
simplifies to high input frequencies ( 20 Hz) calcium spread
Calcium Dynamics in Neuronal Microdomains: 0.4 mm). (d) Comparison of 3D Brownian simulations
Modeling, Stochastic Simulations, and Data Analy- with the uncaging experiments and the results derived
sis, Fig. 9 (continued) different locations from the from the solutions of the 1D effective diffusion equation.
uncaging spot. (b) Compartmentalized model dendrite. The normalized concentration profiles are shown for the
(c) Compartmentalized model dendrite with attached glass tube (a), the dendrite (b), and the dendrite with
spine (dendrite geometry as in B with spine neck radius attached spine (c) at three locations from the uncaging
0.3 mm, spine neck length 0.2 mm, spine head radius spot (Adapted from Biess et al. (2011))
f = 20 Hz
a 0 e 0 f 0
f = 5 Hz
0
f = 10 Hz
200 200 200
time in ms
400 400
time (ms)
400
600 600
5
600 800 800
time in ms
1000 1000
800 −5 2.5 0 2.5 5 −5 2.5 0 2.5 5
f = 50 Hz f = 80 Hz
1000 0 0
10 −5 2.5 0 2.5 5
−5 0 5 200 200
distance from peak (μm) 400
400
0
b 600 600
800 800
time (ms)
1000 1000
−5 2.5 0 2.5 5 −5 2.5 0 2.5 5
5
distance from NMDAR inμm
5
10 f = 20 Hz 4.5
−5 0 5 0
4
distance from peak (μm)
FWHM in μm
3.5
200
c 0 3
2.5
time in ms
400
2
1.5 5Hz
time (ms)
600 10Hz
1 20Hz
5 50Hz
0.5 80Hz
800
0
0 0.2 0.4 0.6 0.8 1
1000 time in s
−5 2.5 0 2.5 5 2 x 103
10
−5 0 5 distance from NMDAR inμm 10 20 30 40 50 60
distance from peak (μm) # particles /μm
d 0
400
# particles/μm
time (ms)
300
5
200
100
10 0
−5 0 5
distance from peak (μm)
Calcium Dynamics in Neuronal Microdomains: the NMDAR in the middle of the dendritic segment as
Modeling, Stochastic Simulations, and Data Analy- shown in the upper and middle panel. The resulting spa-
sis, Fig. 10 Lateral extent of calcium driven by high- tiotemporal profile in the dendrite is shown in the lower
frequency stimulation. (a) Calcium diffusion in an aque- panel. (f) Spatiotemporal profiles in the dendrite for dif-
ous solution contained in a pipette. (b) Calcium diffusion ferent influx frequencies at the location of the NMDAR.
in a crowded dendrite. The initial concentration is equal to (g) Corresponding calcium spread in the dendrite as mea-
about 600 particles and evaluates to about 470 particles sured by the full width at half maximum (FWHM) of the
per micron for a dendrite. (c) Same settings as in (a) but calcium signal (Adapted from Biess et al. (2011)). Fig-
with additional buffers (medium buffer concentration) and ure 11 shows the plot of tirrev
T for several values of the
pumps. (d) Same settings as in (b) but with additional threshold T, compared to Brownian simulations in
buffers (medium buffer concentration) and pumps. (e) a circular disk O = D(R) with reflecting boundary, except
Calcium was injected at 20 Hz for 1 s at the location of at the targets
1 35
0.8
30
0.6
Mean first time to T

0.4 25
0.2
20
0
C
−0.2 15
−0.4 10
−0.6
5
−0.8
−1 0
−1 −0.5 0 0.5 1 4 5 6 7 8 9 10 11
T
Calcium Dynamics in Neuronal Microdomains: simulations (dotted blue line, variance in black), the the-
Modeling, Stochastic Simulations, and Data Analy- oretical formula 74 (dotted black line), and its approxima-
sis, Fig. 11 The MFTT. Left: trajectories of diffusing tion 75 (continuous blue line) for a circular disk in the
molecules in a microdomain containing five binding sites irreversible case (k1 = 0). The other parameters are
on the boundary. Right: the time tirrev
T is plotted as S0 = 15, M0 = 10, e = 0.05, D = 0.1 mm2s1, and the
a function of the threshold T. We present the Brownian radius of the disk R = 1 mm (200 runs)
does not exceed 2.5 mm (= 0.5 FWHM) as Dirichlet conditions on @OA = [ Ni=1Ai. If the
measured from the input source. Buffers and holes Ai are sufficiently far apart, the smallest
pumps limit calcium spread to a few micrometers eigenvalue lA1 is asymptotically the sum of the
(Biess et al. 2011). eigenvalues lA1 i of the Laplace operator in O with
mixed Neumann-Dirichlet boundary conditions
Calcium Extrusion along a Cylinder: on @O Ai and Ai, respectively, which can be
Homogenization of Hole into a Killing Rate calculated from the narrow escape theory
Computing the final distribution of calcium (Holcman and Schuss 2013). For circular holes
ions between two possible fates is a generic with fixed radius e, we have the formula
problem. It can be the proportion of bound cal-
cium ions versus the number extruded or the 4eD 4eDN
fraction that reached the dendrite versus the lA1 i , lA1 ¼ : (56)
jO j jOj
fraction that got pumped. We present a general
approach based on homogenization to reduce the
complexity of this computation. The pdf of the Brownian motion in the perfo-
rated domain relaxes to a quasi-steady state and
Homogenization of Perforated by Partially can be described by a single exponential decay in
Reflecting Boundary time with rate lA1 and a uniform quasi-steady state
We approximate the flux through a reflecting distribution in O, except in boundary layers near
boundary, perforated by many small independent Ai. The total absorption probability flux on the
absorbing holes (Berg and Purcell 1977; boundary, lA1 , is partitioned among the holes with
Ai
l
Berezhkovskii et al. 2004). For diffusion, the probabilities Pi ¼ XN1 : If the holes are
problem can be solved using the eigenvalues of lA1 i i¼1
the Laplace equation in the domain Q with mixed distributed with surface density n(x) on @O, the
Neumann boundary conditions on @O @OA and normal absorption flux density is
4eDnðxÞ For long-time asymptotic, the solution of

JðxÞ nðxÞ ¼ : (57)
jO j Eq. 61 is approximated by the first eigenfunction:

The homogenization procedure consists in pðx, tÞ ¼ po ðxÞclo t þ O clo t , (63)
replacing the perforated holes by a radiation con-
dition that preserves asymptotically the same normalized by
quasi-steady state distribution. We replace the
leading eigenfunction and eigenvalue of the ðL
Robin problem by a Laplace equation with poðxÞdx ¼ 1: (64)
0
a small radiation function k(x):
DDpðxÞ ¼ lpðxÞ for xO (58) When the killing term k is a constant, the
eigenvalue problem
@pðxÞ
D ¼ kðxÞpðxÞ for x@O: (59) @ 2 pð x Þ
@n D kðxÞpðx ¼ lpðx
@x 2
(65)
Matching the flux density in Eqs. 57 and 59, @pð0Þ
pðLÞ ¼ 0, ¼ 0:
we obtain in dimensional coordinates @x
2 In case that k(x) = k = const, the first

e e
kðxÞ ¼ 4eDpo ðxÞnðxÞ þ O 2 log : (60) eigenfunction and eigenvalue are
L L
The quasi-steady state pdf is cos ðOxÞ p p2

p o ð xÞ ¼ O , O¼ , l0 ¼ D 2 þ k:
sin ðOLÞ 2DL 4L

1 þ o k~o
po ðxÞ ¼ ,
jOj
The total flux J of killed diffusing particles is
except in boundary layers near Ai. given by
After the first homogenization which consists ðL
in replacing a three-dimensional diffusion with J¼ kðxÞpo ðxÞdx ¼ k:
total and/or partial absorption by a reduced dif- 0
fusion equation in dimension one, with a killing
rate k, the case of a cylindrical geometry can be
treated immediately. The one-dimensional case Particles Splitting: Absorbed Versus Arrived
of a diffusion with uniform killing inside an To quantify how ions split between pumps
interval (Holcman et al. 2005) with reflecting located on the surface of a cylinder and the ones
and absorbing endpoints has a survival pdf of that arrive at the end, we use the ratio
a particle satisfying the equation
ð
@pðx, tÞ JðxjyÞ:nðxÞdSx
¼ ∇ J ðx, tÞ kpðx, tÞ @Oa
@t (61) Rs ¼ ð
pðL, tÞ ¼ 0, pðx, 0Þ ¼ rðxÞ, kðxÞpðxjyÞdx
O
ð ð
where the probability flux density is given by FðxÞdSx kðxÞpðxjyÞdx
¼ @Oi ð O
(66)
@ kðxÞpðxjyÞdx
J ðx, tÞ ¼ D pðx, tÞ þ bðxÞpðx, tÞ: (62)
@x O
that can be computed in the limit of a very thin Probability to Bind a Fixed Number of Molecules
cylinder. The flux and the killing term are During a Transient Process
computed from solving the steady state using To illustrate the need of a multiscale approach to
Fokker-Planck bridge the molecular to the cellular scale, we
recall that during long-term potentiation
(LTP) induction, a transient calcium signal is
0 ¼ DDpðxjyÞ kðxÞpðxjyÞ for x, yO: (67)
converted into a long-term change in the synaptic C
properties (Bliss and Collingridge 1993; Malenka
The boundary conditions are
and Nicoll 1999; Lee et al. 2009; Lisman
pðxjyÞ ¼0 forx@O, yOa
et al. 2012). This process specifically involves
JðxjyÞ nðx ¼ 0 for x@O a class of kinases (CaMKII) that have complex
@Oa @Oi , yO, t > 0:
JðxjyÞ nðx ¼ Fðx for x@Oi : local cooperative binding sites organized into
a ring structure.
The time-independent flux is F(x) 0. The However, the first problem is to define the
external steady-state flux of absorbed particles is meaning of activation. For example, it can either
be a single binding by a CaM(Ca)3 or CaM(Ca)4
ð molecules, several bonds, or one to six phosphor-
Ja ¼ JðxjyÞ nðxÞdSx : (68) ylations. Once a criterion is chosen, it becomes
@Oa
a computational question to estimate the proba-
The total inward flux is bility Pk that k CaMKII molecules are activated
and also compute statistical quantities such as the
ð ð mean number <Nact> of CaMKII that are acti-
Ji ¼ JðxjyÞ nðxÞdSx ¼ FðxÞdSx : vated following a transient calcium entry.
@Oi @Oi
More complicated calcium patterns are also
(69)
generated where calcium ions are flowing inside
If the killing measure is uniform in the interval a synapse. In that case, it is worthwhile to com-
qffiffiffiffiffiffiffi pute the number of bound molecules before an
[0, L], then RS ðLÞ ¼ coshð1bLÞ1 , where b ¼ Nw D : arbitrary time t or the probability Pact(t) and the
The case where the killing is a Dirac mean number <Nact(t)> of activations before
k(x) = kd(x x1), located at a single point x1, time t. These quantities are in practice difficult
and k is a constant has been treated in Holcman to compute and depend on many parameters such
et al. (2005). Interestingly, changing the distribu- as the dendritic spine geometry, the intrinsic
tion of killing from uniform to concentrated property rate constants, the binding site interac-
at one point has a drastic effect on the final tion forces, their localization, and so on. There
repartition of ions (see Fig.2, Holcman are two complementary approaches:
et al. (2005)). 1. Coarse-grained Markov models: this
approach is based on the narrow escape meth-
Calcium Cascade Initiating Cellular Activation odology, where instead of accounting for the
The molecular implementations that describe entire time dynamics within the complex
the transformation of a transient signal to geometry organization of the microdomain,
cell activation are still unclear. We shall here a Poissonian rate of arrival to small targets is
present a threshold-based model: when the used to approximate the target search and
number of bound molecules equals a given finding. Various quantities of interest such as
number, we suppose that a cellular change is the activation probability and the statistic for
initiated. We present a Markov chain model the number of bound CaMKII can be esti-
that reduces the geometrical complexity based mated (Holcman and Schuss 2005; Dao Duc
on the narrow escape rate formula (Schuss and Holcman 2010, 2012; Holcman
et al. 2007). et al. 2014).
2. Brownian simulations. Contrary to the previ- where kf is the forward-binding rate constant, kb
ous approach, it is not possible to obtain the is the backward-binding rate constant, and Sfree is
exact dependency of the probability; rather the unbound substrate. We assume in our model
this approach is tedious and allows estimating of the reaction that the M molecules diffuse in O
any moment of interest for a given set of independently and, when bound, are released
parameters. independently of each other at exponential
For a microstructure such as a dendritic spine, waiting times with rate k1.
there is no equilibrium, because the steady state is To calculate the average number of unbound
zero. Thus the relaxation time is the time for (or bound) sites in the steady state, the following
a diffusing particle to be extruded by diffusion, reduced model is used. The number k(t) of
which is in the range of tens of ms (see Holcman unbound receptors at time t is a Markovian
and Schuss 2011). The probability for N CaMKII birth-death process with states 0,1, 2,. . ., min
to be activated and the time to induction is the {M, S} and transition rates lk!k+1 = lk,
mean time for this to happen during a repetitive lk!k1 = m = k1. The boundary conditions are
stimulation. Because the calcium to CaMKII lS!S+1 = 0 and l0!1 = 0. Setting Pk(t) = Pr
pathways requires CaM intermediates, there is {k(t) = k}, the Kolmogorov equations for the
no coarse-grained model yet (Guerrier and transition probabilities are given by (Holcman
Holcman 2014a). To present the threshold and Schuss 2005)
method, we shall now present a simplified

model where a diffusing molecule can bind to P_ k ðtÞ ¼ ½lk þ k1 ðS kÞPk ðt
a ligand. The probability to reach a threshold
þ lkþ1 Pkþ1 ðt þ k1 ðS k þ 1 Pk1 ðt
has been developed in Dao Duc and Holcman
(2010) and is reviewed here briefly. for k ¼ ðS MÞþ þ 1, . . . , S 1
(71)
Coarse-Grained Markov Models Traditional
chemical kinetics, based on mass-action laws or with the boundary equations
reaction-diffusion equations, give an inappropri-
ate description of the stochastic chemical reac- P_ ðSMÞþ ðtÞ ¼ k1 SPðSMÞþ ðt þ l1 PðSMÞþ þ1 ðt

tions in microdomains, where only a small P_ S ðtÞ ¼ lS PS ðt þ k1 PS1 ðt
number of substrate and reactant molecules is
involved. A reduced Markovian description of and initial condition Pk,q(0) = dk,Sdq,0. In the limit
the stochastic dynamics of the binding and t ! 1 to the model (71) gives the average
unbinding of molecules is given in Holcman number
and Schuss (2005) and applied in Dao Duc and
Holcman (2010, 2012). Specifically, consider X
S
two finite species, the mobile reactant M that hk 1 i ¼ jPj ,
diffuses in a bounded domain O and the station- j¼ðSMÞþ
ary substrate S (e.g., a protein) that binds M. The
boundary @O of the domain O is partitioned into where Pj = limt!1 Pj(t). Similarly, the stationary
an absorbing part @Oa (e.g., pumps, exchangers, variance of the number of unbound sites is
another substrate that forms permanent bonds s2(M, S) = hk21i hk1i2, where
with M, and so on) and a reflecting part @Or 2 XS
(e.g., a cell membrane). In this model, the volume k1 ¼ j¼ðSMÞþ
j2 Pj : The rates lk are
of M is neglected. In terms of traditional chemical modeled as follows. For a single diffusing mole-
kinetics, the binding of M to S follows the law cule, the time to binding is the first passage time
to reach a small absorbing portion @Oa of the
kf
M þ Sfree Ð MS, (70) boundary, which represents the active surface of
kb the receptor, whereas the remaining part of @O is
reflecting. Due to the small target and to the deep be used to characterize the stability of chemical
binding potential well, the binding and unbinding processes, especially when they underlie
of M to S are rare events on the time scale of a biological function. Using the above Markov
diffusion (Schuss et al. 2007). This implies that chain description, the MFTT can be expressed in
the probability distribution of binding times is terms of fundamental parameters, such as the
approximately exponential (Schuss 2010b) with number of molecules and ligands and the
rate l1 ¼ 1=t1 , where the NET t1 is the MFPT forward- and backward-binding rates. It turns C
to @Oa. When there are S binding sites, k(t) of out that the MFTT depends nonlinearly on the
which are unbound, there are N = [M S + k]+ threshold T. Specifically consider M Brownian
free diffusing molecules in O, where x+ = max{0, molecules that can bind to immobile targets
x}. The arrival time of a molecule to the next S inside a microdomain, modeled generically by
unbound site is well approximated by an expo- Eq. 70. The first time the number [MS](t) of MS
nential law with state-dependent instantaneous molecules at time t reaches the threshold is
rate (see discussion in Holcman and Schuss defined as
(2005))
tT ¼ inf ft > 0 : ½MjSðtÞ ¼ T g, (73)
þ
Nk k ðM S þ k Þ
lk ¼ ¼ : and its expected value is tT : Consider the case of
t1 t1
an ensemble of the targets initially free
The results of the Markovian model (71) are and distributed on the surface of a closed
microdomain and assume that the backward rate
1 vanishes (k1 = 0) and kf > 0. The dynamical
PS ¼ system for the transition probabilities of the
SMÞþ
SðX
∏Si¼Skþ1 iðM S þ iÞþ
1þ Markov process MS(t) is similar to that above,
k¼1 k!ðt1 k1 Þk but for the absorbing boundary condition at the
þ
XÞ
ðSM
∏Si¼Skþ1 iðM S þ iÞþ threshold T, which gives Eq. 71 (Dao Duc and
hk1 i ¼ PS ðS k Þþ
k!ðt1 k1 Þk Holcman 2010). When the binding is irreversible
k¼S1
XÞ
ðSM þ ðk1 ¼ 0Þ, tT is the sum of the forward rates
2 2 ∏Si¼Skþ1 iðM S þ iÞþ
k 1 ¼ PS ðS k Þþ
k!ðt1 k1 Þk 1 1 1
k¼S1
2 tirrev ¼ þ þ ... þ
sS ðMÞ ¼ k1 hk1 i
2 2 T
l0 l1 lT1
(72) 1X
T 1
1 (74)
¼ :
l k¼0 ðM0 kÞðS0 kÞ
(see Holcman and Schuss (2005) for further
details). In particular, when M0 = S0 and M0 1,
These formulas are used to estimate the frac- Eq. 74 becomes asymptotically tirrev T/l
T
tion of bound receptors in photoreceptor outer M0(M0 T). In addition, when the number of
segments and also to interpret the channel noise diffusing molecules greatly exceeds the number
measurement variance in Holcman and Schuss of targets (M0 S0, T), (74) gives the asymptotic
(2005). In Holcman and Triller (2006) this analy- formulas
sis was used to estimate the number of bound
8
AMPA receptors in the postsynaptic density. > 1 S0
>
> lM log S T for M0 S0 , T
A similar gated Markovian model was proposed >
>
< 1 0 0
M0
in Bressloff and Earnshaw (2009). tT
irrev
log for S0 M0 , T
The reduced Markovian model is used for the >
> lS M 0T
>
>
0
calculation of the mean time of the number of >
: T for M0 , S0 T:
bound molecules to reach a given threshold lM0 S0
T (MFTT). In a cellular context, the MFTT can (75)
Discussion and Conclusion and the plasma membrane (Neher 2010). Analyz-
We have summarized here biophysical models at ing diffusion in cusps can be studied by mapping
a molecular level, mathematical analysis, and locally the cusp conformally to a nonsingular
coarse-graining models to study calcium dynam- domain or to use analytical computation
ics in cellular microdomains. The present (Holcman and Schuss 2012; Guerrier and
approach can be implemented to better under- Holcman 2014b). Another difficulty is to account
stand how calcium dynamics can induce sophis- for various varicosities, leading to divergences,
ticated processes such as long-term potentiation and a multiscale analysis should be developed.
or depression, cellular process that are responsi- Modeling approaches are already shedding a new
ble for long-lasting changes in physiological light on the unsolved debate about the role of the
properties. Yet what calcium is doing at synapses dendritic spines in regulating electrical versus
still remains unclear: where calcium is accumu- chemical activity. It is certainly time to revisit
lating, what is the number of activated CaMKII the electrical properties of the spines based on
(it is not exactly clear what is the meaning of solving the full Poisson-Nernst-Planck equation.
being activated; it can be that a single site is Chemical properties should consider surface
phosphorylated), and what other calcium- receptors trafficking as a chemical reservoir.
activated molecules are important. We presented
as an example of complex calcium feedback how
spine twitching can be described at a molecular References
level. We also described Brownian and simplified
simulation of calcium and the CaMKII pathway. Andrews S, Bray D (2004) Stochastic simulation of chem-
ical reactions with spatial resolution and single mole-
The CaMKII molecule can be modeled as cule detail. Phys Biol 1(3-4):137–151
a simplified ring made of six balls. A calibration Batsilas L, Berezhkovskii AM, Shvartsman SY
procedure is then use to determine the radius a of (2003) Stochastic model of autocrine and
a sphere or disk in that matches the paracrine signals in cell culture assays. Biophys
J 85:3659–3665
Smoluchowski formula. We have not reviewed Benazilla F (2000) The voltage sensor in voltage-
the presynaptic terminal, which is a critical dependent ion channels. Physiol Rev 80(2):555–592
microdomain where calcium modulates the Berezhkovskii AM, Makhnovskii YA, Monine MI,
release of vesicles. Vesicular release is triggered Zitserman VY, Shvartsman SY (2004) Boundary
homogenization for trapping by patchy surfaces.
by calcium entrance at specific calcium channels J Chem Phys 121(22):11390–11394
(Holderith et al. 2013). The exact steps from the Berg HC, Purcell EM (1977) Physics of chemoreception.
calcium entrance to the vesicular release are still Biophys J 20(2):193–219
under investigation. The steps of calcium diffu- Biess A, Korkotian E, Holcman D (2007) Diffusion in
a dendritic spine: the role of geometry. Phys Rev
sion have been investigated both experimentally E 76(2 Pt 1):021922
and numerically. In particular the distance from Biess A, Korkotian E, Holcman D (2011) Barriers to
the channel to the vesicle is a key parameter. diffusion on dendrites and estimation of calcium
Channel clustering also would be interesting to spread following synaptic inputs. PLoS Comput Biol
7(10):e1002182
investigate. The possibility that channels are not Bliss TVP, Collingridge GL (1993) A synaptic model of
fixed and the membrane is constantly remodeled memory: long-term potentiation in the hippocampus.
shows that this process is quite complex. Nature 361(6407):31–39
Interestingly, the release probability can be mod- Blomberg F, Cohen RS, Siekevitz P (1977) The structure
of postsynaptic densities isolated from dog cerebral
ulated by six orders of magnitude (Kochubey cortex. II. Characterization and arrangement of some
et al. 2011; Schneggenburger et al. 2012). Simu- of the major proteins within the structure. J Cell Biol
lations of presynaptic calcium are based on 74(1):204–225
numerically solving partial differential equations Bressloff PC, Earnshaw BA (2009) A dynamical corral
model of protein trafficking in spines. Biophys
(Matveev et al. 2004; Zucker and Regehr 2002; J 96:1786–1802
Zucker 1993); however, it is hard to account for Carslaw HS, Jaeger JC (1959) Conduction of heat in
the cusp nature of the region between the vesicles solids. Oxford University Press, New York
Cheviakov AF, Ward MJ, Straube R (2010) An asymptotic Holcman D, Triller A (2006) Modeling synaptic dynamics
analysis of the mean first passage time for narrow driven by receptor lateral diffusion. Biophys
escape problems: Part II: the sphere. Multiscale J 91(7):2405–2415
Model Simulat 8(3):836–870 Holcman D, Schuss Z, Korkotian E (2004) Calcium
Collins FC, Kimball GE (1949) Diffusion-controlled reac- dynamics in dendritic spines and spine motility.
tion rates. J Colloid Sci 4(7–8):425–437 Biophys J 87:81–91
Crank J (1975) The mathematics of diffusion, 2nd edn. Holcman D, Marchewka A, Schuss Z (2005) The survival
Oxford University Press, London probability of diffusion with trapping in cellular biol-
Crick F (1982) Do dendritic spines twitch? Trends ogy. Phys Rev E, Stat Nonlin Soft Matter Phys
C
Neurosci 5:44–46 72(3):031910
Dao Duc K, Holcman D (2010) Threshold activation for Holcman D, Hoze N, Schuss Z (2011) Narrow escape
stochastic chemical reactions in microdomains. Phys through a funnel and effective diffusion on
Rev E 81(4 Pt 1):041107 a crowded membrane. Phys Rev E 84:021906
Dao Duc K, Holcman D (2012) Using default constraints Holcman D, Daoduc K, Burrage K (2014) Successful
of the spindle assembly checkpoints to estimate the delivery of pten in the cytoplasm escaping from
associate chemical rates. BMC Biophys 5:1 micrornas degradation (pre-print)
Eisenberg RS, Klosek MM, Schuss Z (1995) Diffusion as Holderith N, Lorincz A, Katona G, Rózsa B, Kulik A,
a chemical reaction: stochastic trajectories Masahoki W, Nusser Z (2013) Release probability
between fixed concentrations. J Chem Phys of hippocampal glutamatergic terminals scales with
102:1767–1780 the size of the active zone. Nat Neurosci
Erban R, Chapman J (2007) Reactive boundary conditions 15(7):988–997
for stochastic simulations of reaction diffusion pro- Koch C (1999) Biophysics of computation, information
cesses. Phys Biol 4:16–28 processing in single neurons. Oxford University Press,
Fischer M, Kaech S, Knutti D, Matus A (1998) Rapid New York
actin-based plasticity in dendritic spines. Neuron Kochubey O, Lou X, Schneggenburger R (2011) Regula-
20(5):847–854 tion of transmitter release by ca2+ and synaptotagmin:
Flegg M, Chapman J, Erban R (2012) Two regime method insight from large synapse. Trends Neuro
for optimizing stochastic reaction-diffusion simula- 34(5):237–46
tions. J R Soc Interface 9(70):859–868 Korkotian E, Segal M (2006) Spatially confined diffusion
Franz B, Flegg M, Chapman J, Erban R (2013) Multiscale of calcium in dendrites of hippocampal neurons
reaction-diffusion algorithms: PDE-assisted revealed by flash photolysis of caged calcium. Cell
brownian dynamics. SIAM J Appl Math Calcium 40(5–6):441–449
73(3):1224–1247 Korkotian E, Holcman D, Segal M (2004) Dynamic regu-
Ghosh PK, Hanggi P, Marchesoni F, Nori F, Schmid lation of spine-dendrite coupling in cultured hippo-
G (2012) Brownian transport in corrugated channels campal neurons. Eur J Neuroscience
with inertia. Phys Rev E 86(2):021112 20(10):2649–2663
Goldberg JH, Tamas G, Aronov D, Yuste R (2003) Cal- Lamm G, Schulten K (1983) Extended brownian dynam-
cium microdomains in aspiny dendrites. Neuron ics. II. Reactive, nonlinear diffusion. J Chem Phys
40(4):807–821 78(5):2713–2734
Guerrier C, Holcman D (2014a) Activation of CaMKII in Landau LD, Lifshitz EM (1975) Fluid mechanics.
dendritic spines: a stochastic model (in preparation) Pergamon Press, Elmsford
Guerrier C, Holcman D (2014b) The dire strait time for Lee SR, Escobedo-Lozoya J, Szatmari EM, Yasuda
hidden targets (in preparation) R (2009) Activation of CaMKII in single dendritic
Hille B (2001) Ion channels of excitable membranes, spines during long-term potentiation. Nature
3rd edn. Sinauer Associates, Massachusetts 458(7236):299
Holcman D, Schuss Z (2004) Modeling calcium dynamics Lisman J, Yasuda R, Raghavachari S (2012) Mechanisms
in dendritic spines. SIAM J Appl Math of CaMKII action in long-term potentiation. Nat Rev
65(3):1006–1026 Neurosci 13(3):169–182
Holcman D, Schuss Z (2005) Stochastic chemical reac- Majewska A, Tashiro A, Yuste R (2000) Regulation of
tions in microdomains. J Chem Phys 122:114710 spine calcium dynamics by rapid spine motility.
Holcman D, Schuss Z (2011) Diffusion laws in dendritic J Neurosci 20(22):8262–8268
spines. J Math Neuroscience 1:10 Malenka RC, Nicoll RA (1999) Long-term potentiation–a
Holcman D, Schuss Z (2012) Brownian motion in dire decade of progress? Science 285(5435):1870–1874
straits. SIAM J Multiscale Model Simulat 1:10 Matkowsky BJ, Schuss Z, Ben-Jacob E (1982) A singular
Holcman D, Schuss Z (2013) Control of flux by narrow perturbation approach to Kramers’ diffusion problem.
passages and hidden targets in cellular biology. Phys SIAM J Appl Math 42(4):835–849
Progr Report 76(7) 074601 Matveev V, Zucker RS, Sherman A (2004) Facilitation
Holcman D, Schuss Z (2014) The narrow escape problem. through buffer saturation: constraints on endogenous
SIAM Rev 56(2): 213–257 buffering properties. Biophys J 86(5):2691–2709
C 516 Calcium Pumps, Models of
Monine MI, Haugh JM (2005) Reactions on cell mem-

branes: comparison of continuum theory and Brownian Calcium Pumps, Models of
dynamics simulations. J Chem Phys 123(7):074908
Neher E (2010) Complexin: does it deserve its name ?
Neuron Prev 68(5):803–806 Kim T. Blackwell
Pontryagin LS, Andronovn AA, Vitt AA (1933) On the Molecular Neuroscience Department, Krasnow
statistical treatment of dynamical systems. J Theor Institute for Advanced Study, George Mason
Exper Phys (Russian) 3:165–180
Pontryagin LS, Andronov AA, Vitt AA (1989) On the University, Fairfax, VA, USA
statistical treatment of dynamical systems. Noise
Nonlinear Dynamics 1:329–340
Roux B, Prod’hom B, Karplus M (1995) Ion transport in the Definition
gramicidin channel: molecular dynamics study of single
and double occupancy. Biophys J 68(3):876–892
Sabatini BL, Maravall M, Svoboda K (2001) Ca2+ signal- Calcium pump is the general term that refers to
ling in dendritic spines. Curr Opin Neurobiol molecules that extrude or remove calcium from
11(3):349–356 the cytosol. The calcium may be pumped into the
Schneggenburger R, Han Y, Kochubey O (2012) Ca2+
channels and transmitter release at active zone. Cell extracellular space, or pumped into organelles
Calcium 52(3-4):199–207 such as the smooth endoplasmic reticulum or
Schuss Z (1980) Theory and applications of stochastic mitochondria, both of which function as calcium
differential equations, Wiley series in probability and stores. Some of these pumps directly utilize ATP
statistics. Wiley, New York
Schuss Z (2010a) Diffusion and stochastic processes: an as an energy source and are known as calcium
analytical approach. Springer, New York ATPases. Other pumps rely on concentration gra-
Schuss Z (2010b) Theory and applications of stochastic dients and thus are exchangers.
processes, an analytical approach, vol 170, Springer
series on applied mathematical sciences. Springer,
New York
Schuss Z, Holcman D (2013) The narrow escape problem Detailed Description
and its applications in cellular and molecular biology.
SIAM Rev, SIREV So far only three different types of calcium
Schuss Z, Holcman D (2014) Time scales of Diffusion for
Molecular and Cellular processes, J.Phys A: Mathe- pumps have been used in computational models.
matical and Theoretical 47(17), 173001 Two of them are calcium ATPases: the smooth
Schuss Z, Singer A, Holcman D (2007) The narrow escape endoplasmic reticulum ATPase (SERCA) pump
problem for diffusion in cellular microdomains. Proc and the plasma membrane calcium ATPase
Natl Acad Sci USA 104(41):16098–16103
Singer A, Schuss Z, Osipov A, Holcman D (2008) Par- (PMCA) pump (e.g., Morgans et al. 1998). Both
tially reflected diffusion. SIAM J Appl Math of these are typically modeled as an enzyme
68:98–108 reaction, e.g.,
Svoboda K, Tank DW, Denk W (1996) Direct measure-
ment of coupling between dendritic spines and shafts.
Science 272(5262):716–719 nCai þPcyt ÐCan Pcyt ÐCan Pec ÐPec þnCao ÐPcyt
Taflia A, Holcman D (2011) Estimating the synaptic cur- (1)
rent in a multiconductance ampa receptor model.
Biophys J 101:781–792
Tai K, Bond SD, MacMillan HR, Baker NA, Holst MJ, This indicates that n calcium molecules
McCammon JA (2003) Finite element simulations of from the cytosol bind to the pump when the
acetylcholine diffusion in neuromuscular junctions. binding sites are facing the cytosol, i.e., in the
Biophys J 84(4):2234–2241
Zucker RS (1993) Calcium and transmitter release.
Pcyt conformation. Next, the conformation
J Physiol Paris 87(1):25–36 changes so that binding sites are facing the
Zucker RS, Regehr WG (2002) Short-term synaptic plas- extracellular space or lumen of the organelle
ticity. Annu Rev Physiol 64:355–405 (Pec conformation), and then the calcium
Zwanzig R (1990) Diffusion-controlled ligand binding to
spheres partially covered by receptors: an effective
molecules dissociate to the extracellular space
medium treatment. Proc Natl Acad Sci U S A or internal organelles. This reaction scheme can
87:5856–5857 be modeled as explained for enzyme reactions.
Calcium Pumps, Models of 517 C
Reaction Scheme 1 already simplifies the enzyme (or the extracellular space for the PMCA) is
reaction which involves ATP hydrolysis and thus zero at the desired basal calcium value. This
involves additional states. A 12 state model formulation is widely used, both in models of
includes more detail and varying degrees of neurons (e.g., Blackwell 2002; Amini
cooperativity among calcium ions (Tran et al. 1999) and cardiac myocytes (e.g., Jafri
et al. 2009). Alternatively, reaction scheme can et al. 1998; Korhonen et al. 2008).
be simplified further by assuming that the confor- The other pump is the sodium-calcium C
mational changes are instantaneous: exchanger (NCX), which depends on both inter-
nal and external sodium and calcium concentra-
nCai þ PÐCan P!P þ nCao (2) tion. Because this exchanger derives energy from
the sodium concentration gradient, it exhibits
This enzyme reaction can be modeled a dependence on membrane voltage. Several dif-
either stochastically or deterministically as ferent formulations of this exchanger have been
a system of three elementary reactions, but usu- proposed in an attempt to capture all dependen-
ally they are simplified further using the cies on ion concentration. One formulation (Gall
Michaelis-Menton formulation, which gives et al. 1999) calculates dependence on membrane
calcium flux, F: potential from the reversal potential for calcium
and sodium and also has an additional term
Vmax ½Caþþ
n
F5 (3) describing the sigmoid dependence on internal
½Caþþ n þ KnD calcium concentration:
For the smooth endoplasmic reticulum cal-

½Caþþ
n
VNCX
cium ATPase (SERCA) pump, KD is 0.1 mM FNCX 5 area ðVM VNaCa Þ
F ½Caþþ n þKNCX
(Li and Rinzel 1994), and n = 2. For the plasma
where
membrane calcium ATPase (PMCA) pump, KD is 2þ !
1.0 mM (Enyedi et al. 1994), and n = 1. There is RT ½Naþ o Ca o
VNaCa 5 m ln þ
ln
some evidence that the pumps can be modified by F ½Na i ½Ca2þ i
calcium. For both pumps, Vmax is typically (5)
adjusted to effect and is proportional to mem-
brane surface area. The factor of m in the equation for FNCX
This formulation results in a basal calcium represents a stoichiometry of m Na+ to 1 Ca2+,
concentration of 0, thus a compensatory leak which is typically 3 or 4 (Fujioka
back into the cytosol is typically included: et al. 2000; Blackwell 2004; Kang and
Hilgemann 2004). VNCX is the exchanger capac-
Vmax ½Caþþ
n
ity, VM is membrane potential, T is temperature,
F5 JL Caþþ
ec ½Ca
þþ
(4)
½Caþþ n þ KDn R is the ideal gas constant, and F is Faraday’s
constant.
The second term on the left hand side is A different formulation was derived to
a compensatory leak; JL is adjusted such that account for the currents observed in bull-frog
net calcium flux from the organelle lumen cardiac atrial cells (Campbell et al. 1988):

g ðr2ÞVM F ð1g Þðr2ÞVM F
½Naþ i ½Caþþ o exp RT ½Naþ o ½Caþþ i exp
3 3
RT
INCX 5VNCX
(6)
1 þ dNa, Ca ½Naþ 3i ½Caþþ o þ ½Naþ 3o ½Caþþ i
C 518 Calcium Pumps, Models of
This formulation has been used in several Due to the importance of calcium for
models (Amini et al. 1999; Rasmusson excitation-contraction coupling, many models of
et al. 1990). Since the driving potential for cal- cardiac myocytes have even more elaborate
cium depends on sodium, these models often models of the NCX that are based on an influen-
include equations for the sodium-potassium tial model (DiFrancesco and Noble 1985):
exchanger.

gVM F ð1gÞVM F
½Naþ i ½Caþþ o exp RT ½Naþ o ½Caþþ i exp RT
3 3
INCX 5VNCX

ð1gÞVM F
(7)
K3m, Na þ ½Naþ 3o Km, Ca þ ½Caþþ o 1 þ ksat exp RT
In addition to dependence on membrane oscillations in midbrain dopaminergic neurons: a com-

putational study. J Neurophysiol 82(5):2249–2261
potential and concentration gradients, this equa- Blackwell KT (2002) Calcium waves and closure of leak
tion also has terms representing saturation of the potassium currents caused by GABA stimulation of
exchanger by external sodium concentration and the type B photoreceptor of Hermissenda crassicornis.
calcium concentration. This equation has been J Neurophysiol 87:776–792
Blackwell KT (2004) Paired turbulence and light do not
used in models to show that enhanced activity
produce a supralinear calcium increase in
of the NCX contributes to prolongation of the Hermissenda. J Comput Neurosci 17:81–99
cardiac AP (Priebe and Beuckelmann 1998), to Campbell DL, Giles WR, Robinson K, Shibata EF
show that genetic alterations of NCX can indeed (1988) Studies of the sodium-calcium exchanger in
bull-frog atrial myocytes. J Physiol 403:317–340
account for alternations in excitation contraction
DiFrancesco D, Noble D (1985) A model of cardiac elec-
coupling (Korhonen et al. 2008), and to demon- trical activity incorporating ionic pumps and concen-
strate how force generation is altered by changes tration changes. Philos Trans R Soc Lond B Biol Sci
in heart rate (Jafri et al. 1998). 307:353–398
Enyedi A, Verma AK, Heim R, damo HP, Filoteo AG,
An even more elaborate equation was devel- Strehler EE, Penniston JT (1994) The Ca2+ affinity of
oped (Weber et al. 2001) to account for allosteric the Plasma Membrane Ca2+ pump is controlled by
modulation of NCX. That equation also has been alternative splicing. J Biol Chem 269:41–43
used in several models of cardiac action poten- Fujioka Y, Komeda M, Matsuoka S (2000) Stoichiometry
of Na+-Ca2+ exchange in inside-out patches excised
tials (Rovetti et al. 2010; Hund and Rudy 2004).
from guinea-pig ventricular myocytes. J Physiol
A different approach is to use Markov kinetic 523(Pt 2):339–351
models to describe activity of the exchanger Gall D, Gromada J, Susa I, Rorsman P, Herchuelz A,
(Omelchenko and Hryshko 1996; Haase and Har- Bokvist K (1999) Significance of Na/Ca exchange for
Ca2+ buffering and electrical activity in mouse pancre-
tung 2009; Kim and Matsuoka 2008). The goal of
atic beta-cells. Biophys J 76:2018–2028
this approach is to understand state transitions Haase A, Hartung K (2009) Pre-steady-state kinetics of
of the molecule and to completely capture effects Ba-Ca exchange reveals a second electrogenic step
of voltage and ion concentration on exchanger involved in Ca2+ translocation by the Na-Ca
exchanger. Biophys J 96(11):4571–4580
current for a wide range of conditions.
Hund TJ, Rudy Y (2004) Rate dependence and regulation
In summary, the large number of studies and of action potential and calcium transient in a canine
models reflects the importance of control of cal- cardiac ventricular cell model. Circulation
cium concentration for cell function. 110(20):3168–3174
Jafri MS, Rice JJ, Winslow RL (1998) Cardiac Ca2+
dynamics: the roles of ryanodine receptor adaptation
and sarcoplasmic reticulum load. Biophys
References J 74(3):1149–1168
Kang TM, Hilgemann DW (2004) Multiple transport
Amini B, Clark JW Jr, Canavier CC (1999) Calcium modes of the cardiac Na+/Ca2+ exchanger. Nature
dynamics underlying pacemaker-like and burst firing 427(6974):544–548
Calcium Release, Models of 519 C
Kim B, Matsuoka S (2008) Cytoplasmic Na+-dependent other electrically excitable cells, Ca2+ changes
modulation of mitochondrial Ca2+ via electrogenic are mainly carried by Ca2+ entry from the exte-
mitochondrial Na+–Ca2+ exchange. J Physiol
586(6):1683–1697 rior, through voltage-dependent Ca2+ channels.
Korhonen T, Rapila R, Tavi P (2008) Mathematical model These channels are able to generate large
of mouse embryonic cardiomyocyte excitation- increases in cytosolic Ca2+ in a few milliseconds.
contraction coupling. J Gen Physiol 132(4):407–419 In contrast, in electrically non-excitable cells, the
Li YX, Rinzel J. (1994) Equations for InsP3 receptor-
mediated [Ca2+]i oscillations derived from a detailed main source for Ca2+ is the endoplasmic reticu- C
kinetic model: a Hodgkin-Huxley like formalism. J lum (ER). Release of Ca2+ from the ER occurs
Theor Biol. 166(4):461–73 through the inositol 1,4,5-trisphosphate receptor
Morgans CW, El Far O, Berntson A, W€assle H, Taylor (InsP3R) or through the ryanodine receptor (RyR)
WR. (1998) Calcium extrusion from mammalian pho-
toreceptor terminals. J Neurosci. 18(7):2467–74 in response to a hormonal stimulation (Fig. 1).
Omelchenko A, Hryshko LV (1996) Current–voltage rela- However, also in neurons, release of ER Ca2+
tions and steady-state characteristics of Na+-Ca2+ can be triggered during an action potential and
exchange: characterization of the eight-state consecu- thereby modulate Ca2+-activated plasma mem-
tive transport model. Biophys J 71(4):1751–1763
Priebe L, Beuckelmann DJ (1998) Simulation study of brane channels and affect neuronal excitability.
cellular electric properties in heart failure. Circ Res Thus, Ca2+ exchanges between intracellular organ-
82(11):1206–1223 elles and the cytoplasm need to be accurately
Rasmusson RL, Clark JW, Giles WR, Shibata EF, Camp- modelled when describing the signaling activity
bell DL (1990) A mathematical model of a bullfrog
cardiac pacemaker cell. Am J Physiol 259(2 Pt 2): of neurons. Of particular importance is the fact that
H352–H369 Ca2+ can be released from the ER in a regenerative
Rovetti R, Cui X, Garfinkel A, Weiss JN, Qu Z (2010) manner, through the so-called Ca2+-induced Ca2+
Spark-induced sparks as a mechanism of intracellular release regulation. This autocatalytic regulation,
calcium alternans in cardiac myocytes. Circ Res
106(10):1582–1591 mediated by either the InsP3R or the RyR, allows
Tran K1, Smith NP, Loiselle DS, Crampin EJ for an excitable process additional to plasma mem-
(2009) A thermodynamic model of the cardiac sarco- brane electrical activity. Other factors control
plasmic/endoplasmic Ca(2+) (SERCA) pump. Biophys cytoplasmic Ca2+ changes in a way that is inde-
J 96(5):2029–2042
Weber CR, Ginsburg KS, Philipson KD, Shannon TR, pendent from the electrical properties of the
Bers DM (2001) Allosteric regulation of Na/Ca plasma membrane, such as buffers, ATP-driven
exchange current by cytosolic Ca in intact cardiac Ca2+ pumps, Ca2+ exchangers, InsP3 metabolism,
myocytes. J Gen Physiol 117:119–131 mitochondria, or store-operated Ca2+ entry
(Keener and Sneyd 2009). Factors that control
Ca2+ dynamics in a voltage-independent way are
Calcium Release, Models of reviewed here below, both from a biological and
a modelling point of view.
Geneviève Dupont
Theoretical Chronobiology Unit, Université
Libre de Bruxelles, Brussels, Belgium Detailed Description
Calcium Channels
Definition At rest, Ca2+ concentration is low (30–100 nM) in
the cytoplasm and mitochondria and high
Intracellular Ca2+ levels control nearly all cell (0.5–1 mM) in the extracellular medium and
functions. To adequately fulfill its physiological ER. These gradients allow for a rapid increase
roles, cytosolic Ca2+ concentration is thus in cytosolic Ca2+ following the opening of spe-
dynamically regulated by channels mediating cific Ca2+ channels when required for the signal-
Ca2+ exchanges with the extracellular medium ing function of the cell. On the other hand,
or with intracellular stores, by proteins acting as maintenance of these gradients occurs through
Ca2+ buffers and by diffusion. In neurons, as in active, ATP-dependent pumping.
C 520 Calcium Release, Models of
Calcium Release, Models

of, Fig. 1 Schematic
representation of the main
processes regulating
cytosolic Ca2+
concentration in
electrically non - excitable
cells
InsP3 Receptor is the subject of interest, such models can be

The InsP3 receptor is a tetrameric protein located simplified on the basis of the different time scales
in the membrane of the ER, although the presence of the activatory and inhibitory processes
of this receptor has also been reported in the (Dupont et al. 2011). The evolution of the frac-
perinuclear or plasma membrane (Parys and De tion of inhibited channels can be described by
Smedt 2012). This receptor is a Ca2+ channel that, a unique differential equation:
when opened, lets Ca2+ flow along its gradient. Its
Ca2+-releasing activity is not only regulated by dRi 1 Ri
¼ kþ Cni
na k Ri
InsP3 but also by the Ca2+ concentration at the dt 1þ C K act
cytoplasmic terminal: Ca2+ concentrations up to
~300 nM tend to open the channel, while higher
concentrations inhibit Ca2+ release, albeit on where k+ and k are the kinetic constants of Ca2+
a slower time scale. The activation, acting as an association to and dissociation from the inhibi-
autocatalytic regulation through which Ca2+ acti- tory Ca2+ binding site of the receptor, C repre-
vates its own release in the cytoplasm, plays sents the concentration of cytosolic Ca2+, and
a major role in the occurrence of the well- Kact is the dissociation constant of Ca2+ binding
known Ca2+ oscillations observed in most to the activating Ca2+ binding site of the InsP3
non-excitable cell types (Berridge 2012). receptor. Thus, Ca2+ can be released from the
ER through the InsP3 receptor when it is active
Deterministic Model Models can describe (i.e., open); the fraction of open channel is related
sequential activation and inhibition by assuming to the fraction of inhibited receptor by the
that the InsP3 receptor possesses 3 types of bind- following algebraic equation:
ing sites: one for InsP3, one for Ca2+ with an
activatory effect, and one for Ca2+ with an inhib- Cna IP3
IRa ¼ ð1 Ri Þ
itory effect, characterized by a lower affinity for ðK act Þna þ Cna ðK IP Þ3 þ IP3
Ca2+. These detailed models for the receptor (see
Thul et al. 2008, for review) can accurately Even at basal InsP3 concentration, the InsP3R/
describe the properties of the channel and are Ca channel can release Ca2+ through its auto-
2+
easily implemented to perform stochastic simu- catalytic activation and amplify Ca2+ entry
lations (Falcke 2004). When the cellular behavior through voltage-gated Ca2+ channels.
A more accurate description of InsP3-induced dRi 1 Ri ðK IP Þ3
2+
Ca fluxes can be obtained when taking into ¼ kþ Cni
na k Ri
dt 1þ C ðK IP Þ3 þ IP3
account the existence of the isoforms of this K act
receptor/channel. Three isoforms have been iden-
tified (InsP3R1, InsP3R2, and InsP3R3), and
their levels of expression are largely tissue Stochastic Model Accounting for Mode
specific. Experiments where the levels of expres- Changes Deterministic models do a very good C
sion of these isoforms have been modified job in simulating global Ca2+ changes and in
(overexpress or knockdown) clearly indicate reproducing most characteristics of signal-
that their proportions significantly affect the induced Ca2+ oscillations. However, in order to
time course of cytosolic Ca2+ concentration understand measurements of time-dependent
(Hattori et al. 2004). Although all the InsP3 single-channel currents through an InsP3R
receptor subtypes display similar permeation (Ionescu et al. 2007; Mak et al. 2007), stochastic
properties, they differ significantly in their reg- models describing the transitions between many
ulatory properties due to different interactions channel states need to be considered.
with accessory proteins, various modes of regu- \In particular, recent experiments revealed that
lation by ATP, or phosphorylation (Parys and De the InsP3R exhibits mode changes: the average
Smedt 2012). The characteristics of one specific open probability indeed jumps between two levels
receptor subtype can be considered in a model even when ligand (Ca2+, InsP3, or ATP) concentra-
by assigning the appropriate values to the disso- tions are kept constant. These modes are referred to
ciation constants for IP3 binding, Ca2+ binding as drive (highly active) and park (inactive) modes
to the activatory site, and Ca2+ binding to the by Siekmann et al. (2012). In each of these modes,
inhibitory site. Although still controversial, the receptor can exist in different open or closed
putative values are summarized in the table states. By comparing fits of Markov models with
below: different numbers of states and different topologies,
it was found that the drive mode is best represented
Type 1(mM) Type 2(mM) Type 3(mM) by a model with one open and three closed states,
KIP 0.50 0.20 4.00 while the best model for the park mode only has one
Kact 0.50 0.50 0.50 open and one closed state (Fig. 2).
Kinh 0.50 0.25 1.15 Inside each mode, rate constants are ligand inde-
pendent. Some of these constants are different for
In this table, the dissociation constant for Ca2+
the type I and the type II InsP3R. Only the transi-
on the inhibitory site corresponds to
tion rates between both modes, q24 and q42, are
1=ni fitted by different values depending on the concen-
k trations of Ca2+, IP3, and ATP. These values are
K inh ¼
kþ given for a large set of conditions in Siekmann
et al. (2012). Interestingly, the park and drive
As another difference, the evolution equation times depend only on the rates exiting the modes:
for the fraction of inhibited receptor has to be
slightly modified to reproduce the behavior of
q45 1
the InsP3R1: it is assumed that InsP3 inhibits tpark ¼ 1þ
q54 q42
Ca2+ binding to the inhibitory site, which is
necessary to account for the observation that
for type 1 receptors, inhibition occurs for higher which can be approximated by q1 since the ratio
42
q45
concentrations of cytosolic Ca2+ when the con- q is low for both types I and II InsP3R.
54
centration of InsP3 increases (Dupont and

Combettes 2006). The appropriate equation q21 q23 q26 1
tdrive ¼ 1þ þ þ
is then q12 q32 q62 q24
channel, in the same way as IP3 for the IP3

receptor.
The RyR can be modelled the same way as the
InsP3 receptor, although simpler kinetic expres-
sions are also adequate. Modelling studies indeed
suggest that inactivation does not play
a significant role for RyR-mediated Ca2+ release.
Indeed, describing this flux by a very simple Hill-
type expression, the following sigmoidal function
of cytosolic Ca2+
!
C3
V RyR ¼ k1 þ k2 ðCER CÞ
ðK d Þ3 þ C 3
gives excellent agreement with observations in

bullfrog sympathetic neurons (Friel 1995). How-
ever, simulations of the RyR should take
into account the prominent opposition of RyR
with plasma membrane Ca2+ channels, and
more especially L-type Ca2+ channels, as geom-
etry is crucial for the Ca2+ dynamics (Thul
et al. 2012).
Calcium Release, Models of, Fig. 2 Drive and Park
Markov model for the InsP3R developed by Siekmann et Stochastic Model Including Regulation by
al. (2012) Ca2+ in the ER A striking property of RyRs is
that their activity spontaneously terminates
An important point to remember at this point is after a few tens of milliseconds. This strongly
that in such type of models, the channel states and suggests the existence of an inactivated state.
connections between them are not related to any Experimental evidences also show that high
underlying reaction scheme. Rather, they repre- levels of Ca2+ in the ER increase the open prob-
sent the most complex model that can be built ability of the RyR (Shannon and Bers 2000).
unambiguously from statistical analysis of A plausible molecular mechanism could
single-channel data. involve an interaction between calsequestrin,
a widespread ER Ca2+ buffer, and the RyR
Ryanodine Receptor (Györke and Terentyev 2008).
Deterministic Model The ryanodine receptor This dual regulation of the RyR by cytosolic
(RyR) much resembles the InsP3 receptor. It is and by Ca2+/calsequestrin is taken into account in
expressed in all excitable cells. It is responsible the model of Restrepo et al. (2008) that assumes
for excitation–contraction coupling in muscle the existence of 2 open (O) and 2 closed (C) states
cells. The main activator is Ca2+ itself, and, as (Fig. 3):
for the InsP3 receptor, excess Ca2+ has an inhib- Transition from C1 to O1 and from C2 to O2
itory effect on the RyR, although the level of Ca2+ (Fig. 3) is activated by cytosolic Ca2+. Thus,
required for inhibition much depends on the
receptor subtype. Other factors that modulate k12 ¼ K u C2
the activity of the RyR are caffeine, ryanodine,
and the second messenger cyclic ADP ribose, and k43 = KbC2. Calsequestrin regulates the pas-
which may be the endogenous activator of the sage from C1 to C2 and from O1 to O2. In fact,
possible way by assuming that the rate of Ca2+
entry increases linearly with the level of cell
stimulation (Keener and Sneyd 2009).
In most situations, the major molecular com-
ponents of SOCs are STIM and Orai. STIM
resides in the ER membrane, not far from the
plasma membrane where Orai channels are C
expressed. STIM is able to sense the filling state
of the ER (through an EF-hand motif in its
N-terminus). Upon Ca2+ dissociation, it changes
its conformation, which allows self-association
Calcium Release, Models of, Fig. 3 Model for the RyR of STIM molecules. When aggregated, STIM
and its regulation by calsequestrin developed by Restrepo molecules interact with Orai at plasma membrane
et al. (2008) junctions. Orai then allows the entry of Ca2+ in
the cytosol.
A simple model for Ca2+ entry based on the
STIM–Orai interaction has been proposed by Liu
only the calsequestrin monomers interact with et al. (2010). Ca2+ flux through SOC channels is
RyR. Calsequestrin dimerization is favored by given by
high ER Ca2+. Thus, C2 and O2 states are pre-
dominant when ER Ca2+ is low. Assuming a fast Cex
V IN ¼ V SOC ½SO
equilibrium between calsequestrin monomers K ex þ Cex
and dimers, Restrepo et al. (2008) have derived
an algebraic equation for transition rates: where [SO] represents the fraction of open SOCs
and VSOC the maximal rate of Ca2+ entry of
k23 ¼ k14 ¼ MðCER Þt1
b
extracellular Ca2+ (Cex) through SOC channels.
Open SOCs correspond to STIM–Orai com-
where M is an increasing function of CER. plexes. Thus, [SO] must be computed as the
Other rates are constant. Importantly, k12 > > ratio between the concentration of STIM–Orai
k43, which means that unbinding of calsequestrin complexes and the total concentration of
induced by dimer formation at high ER Ca2+ Orai. Its time evolution is simply given by the
leads to an increase in the RyR open probability, balance between STIM–Orai binding and
as observed experimentally. unbinding, i.e.,
Influx d ½SO
In non-excitable cells, influx of Ca2+ is thought to ¼ f 0 ð1 ½SOÞ½STIM b0 ½SO
dt
occur through two possible mechanisms,
depending on the filling state of the endoplasmic where [STIM] represents the concentration of ER
reticulum. Such an influx is thus said to occur Ca2+-free STIM. Thus,
through “store-operated Ca2+ channels (SOCs).”
When the ER is only slightly depleted, the rate of d ½STIM
influx is small and proportional to the level of ¼ f S ½STIMðCER Þns þ bS ½STIMtot ½STIM
dt
agonist stimulation through a mechanism that
may involve arachidonic acid (Shuttleworth When solved at steady state, this simple
2009). Arachidonic acid is produced by model gives an excellent agreement with
diacylglycerol, a membrane compound that is the experimental data about store-operated
produced by the same reaction as InsP3. Thus, Ca2+ entry in human Jurkat T cells (Luik
this stimulated influx is modelled in the simplest et al. 2008).
Inositol 1,4,5-Trisphosphate between Ca2+ and InsP3 concentrations in an

Metabolism intuitive manner. Models taking these regulations
into account have proven useful to understand
Inositol 1,4,5-trisphosphate metabolism is over- experimental results and make theoretical predic-
looked in most models for intracellular Ca2+ tions (Dupont and Erneux 1997; Dupont
dynamics, although it is well known that the et al. 2003). Using enzyme kinetics, the evolution
level of this messenger has a profound influence of the concentrations of InsP3 and InsP4 can be
on the level of cytosolic Ca2+. It makes no doubt described as
that variations in the concentration of InsP3 will
modify the Ca2+ flux of Ca2+ from the ER through dIP3 IP3 C4
¼ g:V PLC V k
the InsP3 receptor. dt K k þ IP3 ðK d Þ4 þ C4
InsP3 originates from the receptor-activated IP3
hydrolysis of phosphoinositides through well- V P1
IP4
characterized signaling cascades. The ultimate K P1 1þ þ IP3
K P2
enzyme of this cascade is membrane-bound phos-
pholipase C. This enzyme possesses several
dIP4 IP3 C4
isoforms (b, g, d, e, and z) differing by the sig- ¼ Vk
dt K k þ IP3 ðK d Þ4 þ C4
naling pathway in which they are involved and
by their Ca2+ sensitivity. PLCb is activated by V P2
IP4
k:IP4
protein G-coupled receptors, PLCg by tyrosine IP3
K P2 1 þ þ IP4
kinases, PLCd by PIP2 and Ca2+, PLCe by Ras, K P1
and PLCz by Ca2+. Concerning its degradation,
InsP3 can be either dephosphorylated by a
5-phosphatase into IP2 or phosphorylated by a
3-kinase into InsP4. The 3-kinase is stimulated Mitochondria
by an increase in Ca2+ as binding of Ca2+ and
calmodulin enhances its activity. The stimula- Mitochondria also affect cytoplasmic Ca2+ sig-
tion factor varies from 2 to 10 depending on the nals. They can both buffer cytosolic Ca2+ changes
isoform (Dupont and Erneux 1997). Thus the and release Ca2+. At rest, intramitochondrial and
InsP3-induced Ca2+ release leads to a decrease cytosolic Ca2+ concentration are similar, of the
in InsP3 concentration. Another level of com- order of 100 nM. Ca2+ entry in the mitochondria
plexity comes from the fact that the product of occurs through a multistep mechanism. By
InsP3 phosphorylation by 3-kinase, InsP4, is extruding protons out of the mitochondria, the
a competitive inhibitor of the other InsP3- respiratory chain creates a large inside-negative
metabolizing enzyme, the 5-phosphatase potential difference across the inner mitochondrial
(Fig. 4). membrane. This DCm, which is harnessed by the
Because of this network of intertwined regu- ATP synthase in the production of ATP, allows
lations, it is difficult to apprehend the interplay a uniporter to transport Ca2+ inside the
Calcium Release, Models

of, Fig. 4 Inositol
1,4,5-Trisphosphate
(Ins 1,4,5-P3) metabolism
mitochondria when cytosolic Ca2+ locally reaches Friel D (1995) [Ca2+]i oscillations in sympathetic neurons:
high concentrations, which arises in the close an experimental test of a theoretical model. Biophys
J 68:1752–1766
vicinity of ER Ca2+ release sites. Ca2+ entry then Györke S, Terentyev D (2008) Modulation of ryanodine
depolarizes the mitochondria, thus reducing its receptor by luminal calcium and accessory proteins in
own driving force. When cytosolic Ca2+ returns health and cardiac disease. Cardiovasc Res 77:245–255
to its basal value, extrusion of Ca2+ out of the Hattori M, Suzuki A, Higo T, Miyauchi H, Michikawa T,
Nakamura T, Inoue T, Mikoshiba K (2004) Distinct roles
mitochondria occurs mainly through the rather of inositol 1,4,5-trisphosphate receptor types 1 and 3 in
C
slow Na+–Ca2+ exchanger (Santo-Domingo and Ca2+ signalling. J Biol Chem 279:11967–11975
Demaurex 2010; Rizzuto et al. 2012). In addition Ionescu L, White C, Cheung K, Shuai J, Parker I, Pearson J,
to the uniporter/exchanger pathway for Ca2+ Foskett J (2007) Mode switching is the major mecha-
nism of ligand regulation of InsP3 receptor calcium
cycling between the cytosol and the mitochondria, release channels. J Gen Physiol 130:631–645
an important increase in matrix Ca2+ can also lead Keener J, Sneyd J (2009) Mathematical physiology,
to opening of the permeability transition pore 2nd edn. Springer, New-York
(PTP), a voltage-dependent, high-conductance Liu W, Tang F, Chen J (2010) Designing dynamical out-
put feedback controllers for store-operated Ca2+ entry.
channel behaving as a large pore allowing solutes Math Biosci 228:110–118
with a molecular weight less than about 1,500 kDa Luik R, Wang B, Prakriya M, Wu A, Lewis R (2008)
to equilibrate across the inner membrane. Oligomerization of STIM1 couples ER calcium deple-
The level of Ca2+ in the mitochondria thus tion to CRAC channel activation. Nature 454:538–541
Mak D, Pearson J, Loong K, Datta S, Fernandez-Mongil-
results from the balance between entry through M, Foskett J (2007) Rapid ligand-regulated gating
the uniporter and efflux via the Na+–Ca2+ kinetics of single inositol 1,4,5-trisphosphate receptors
exchanger and possibly the permeability transi- Ca2+ release channels. EMBO Rep 8:1044–1051
tion pore: Oster A, Thomas B, Terman D, Fall C (2011) The low
conductance mitochondrial permeability transition
dCm pore confers excitability and CICR wave propagation
¼ J Na, ex þ J uni þ J PTP in a computational model. J Theor Biol 273:216–231
dt
Parys J, De Smedt H (2012) Inositol 1,4,5-trisphosphate
All fluxes depend on cytosolic Ca2+, mito- and its receptors. In: Islam MS (ed) Calcium signaling,
chondrial Ca2+, and mitochondrial potential in advances in experimental medicine and biology 740.
Springer, Dordrecht, pp 255–279
a complex manner. Detailed, accurate kinetic
Restrepo J, Weiss J, Karma A (2008) Calsequestrin-
expressions can be found in the literature (Oster mediated mechanism for cellular calcium transient
et al. 2011). alternans. Biophys J 95:3767–3789
Rizzuto R, De Stefani D, Raffaello A, Mammucari
C (2012) Mitochondria as sensors and regulators of
References calcium signalling. Nat Cell Biol 13:566–578
Santo-Domingo J, Demaurex N (2010) Calcium uptake
Berridge MJ (2012) Cell signalling biology. doi:10.1042/ mechanisms of mitochondria. Biochim Biophys Acta
csb0001002. www.cellsignallingbiology.org 1797:907–912
Dupont G, Combettes L (2006) Modelling the effect of Shannon T, Bers M (2000) Potentiation of fractional sar-
specific inositol 1,4,5-trisphosphate receptor isoforms coplasmic reticulum calcium release by total and free
on cellular Ca2+ signals. Biol Cell 98:171–182 intra-sarcoplasmic reticulum calcium concentration.
Dupont G, Erneux C (1997) Simulations of the effects of Biophys J 78:334–343
inositol 1,4,5-trisphosphate 3-kinase and 5- Shuttleworth T (2009) Arachidonic acid, ARC channels
phosphatase activities on Ca2+ oscillations. Cell Cal- and Orai proteins. Cell Calcium 45:602–610
cium 22:321–331 Siekmann I, Wagner L, Yule D, Crampin E, Sneyd
Dupont G, Koukoui O, Clair C, Erneux C, Swillens S, J (2012) A kinetic model for type I and II IP3R
Combettes L (2003) Ca2+ oscillations in hepatocytes accounting for mode changes. Biophys J 103:658–668
do not require the modulation of InsP3 3-kinase activ- Thul R, Bellamy T, Roderick H, Bootman M, Coombes S
ity by Ca2+. FEBS Lett 534:101–105 (2008) Calcium oscillations. Adv Exp Med Biol
Dupont G, Combettes L, Bird G, Putney J (2011) Calcium 641:1–27
oscillations. Cold Spring Harb Perspect Biol 1:3, pii: Thul R, Coombes S, Roderick H, Bootman M (2012)
a004226 Subcellular calcium dynamics in a whole-cell model
Falcke M (2004) Reading the pattern in living cells – the of an atrial myocyte. Proc Natl Acad Sci U S A
physics of Ca2+ signaling. Adv Phys 53:255–440 109:2150–2155
C 526 Calcium Waves, Models of
ryanodine receptor (RyR) that is found in the

Calcium Waves, Models of sarcoplasmic reticulum, the internal calcium
store in muscle, release occurs in the form of
M. Saleet Jafri and Aman Ullah quarks (a fundamental opening event) and cal-
School of Systems Biology and Krasnow cium sparks which are the elementary unit of
Institute for Advanced Studies, George Mason calcium release which sum up to the calcium
University, Fairfax, VA, USA transient. The calcium transient can propagate
to carry a signal from one part of the cell to
another in a calcium wave.
Definition
Basic Mechanism of Ca2+ Signaling
A calcium wave is a transient rise in cytosolic In the cell cytosol, Ca2+ can enter from two
calcium concentration that propagates spatially sources: (1) from extracellular sources through
within the cell. It has the property that it is regen- a variety of plasmalemmal Ca2+ channels such
erative and is followed by a refractory period. It is as voltage-operated channels (VOCs), receptor-
believed that the wave carries a signal from one operated channels (ROCs), or store-operated
part of the cell to another. channels (SOCs) or (2) from intracellular Ca2+
stores such as the endoplasmic reticulum (ER)
and sarcoplasmic reticulum (SR). The endoplas-
Detailed Description mic reticulum (ER) comprises up to 20 % of the
cell; it is the major Ca2+ storage organelle in
Calcium ions play an important role in a variety many cell types (Foskett et al. 2007). The extra-
of cellular functions that impact nearly all aspects cellular calcium concentration is in the mM
of cellular life. In fact, the calcium signal is one range, whereas the internal concentration in ER
of the most versatile and universal signaling is of the order of a few hundred mM and may
agents of the human body that controls almost reach up to mM (Falcke 2004). This provides
everything we do – Ca2+ ions contribute to egg a large driving force for calcium entry into the
activation upon fertilization, muscle contraction, cytosol which is much lower. This low concen-
enzyme secretion, neurotransmitter release, how tration of Ca2+ is maintained by pumps and other
our brains process information and store memo- Ca2+ transporters located on the ER membrane
ries, wound healing, and apoptosis. Calcium ions and plasma membrane (Foskett et al. 2007) as
play such an important role in so many biological well as intracellular calcium-binding proteins
processes that it has been branded as the “life and which serve as Ca2+ buffers. Hence, a finely
death signal” (Berridge et al. 1998). tuned mechanism regulates cytosolic calcium.
At rest, intracellular calcium concentration
remains low, about 0.1 mM. Transient increases Mechanisms of Calcium Waves
of intracellular Ca2+ concentration influence The specific mechanism of calcium waves can
different cellular processes across a variety of vary depending upon cell type and condition.
different cells. This transient rise can peak at Mathematically they all have some common fea-
0.5–3.0 mM depending upon the cell type and tures. First, the system must be excitable which
signal. The change in Ca2+ concentration trans- means that an increase in calcium above
mits information through an organized set of a threshold will trigger additional calcium
spatiotemporal changes (Adkins and Taylor release. This requires positive feedback of cal-
1999). When considering the inositol 1,4,5- cium release (such as calcium-induced calcium
trisphosphate (IP3) receptor channel that is release) that is possible with systems that contain
found on the endoplasmic reticulum, release the IP3 receptor or the RyR. These can be
occurs as localized blips, puffs, transients, oscil- described as components of a system of coupled
lations, and waves (Fig. 1). When considering the differential equations giving the ability to
Calcium Waves, Models of 527 C
Calcium Waves, Models
of, Fig. 1 Calcium blips
(or quarks in the case of
RyR clusters) are
considered to be openings
of one or a few IP3 receptor
channels (top). Calcium
puffs (or sparks in the case C
of RyR clusters) reflect the
opening of a cluster of IP3
receptor channels (middle).
When calcium release
propagates from one cluster
to another, a calcium wave
results (bottom)
analyze the mechanism of calcium waves. arrow on right) until if falls off the nullcline and
Mathematically a system of differential equations moves left until it finds the other stable part of the
can be studied two variables at a time using n-shaped nullcline (left pointing red arrow). The
a phase plane diagram that explores the values trajectory will then follow the nullcline back to
of the derivatives at different values of the vari- the steady state (thick red arrow on left). Detail of
ables. The phase plane considers the nullclines this mechanism can be found in the model devel-
(the curves where each derivative is zero), and the oped by Sneyd and coworkers (Girard et al.
values of the derivatives at other points (vector 1992). If the “straight” nullcline is shifted so
field) define the trajectories of possible solutions. that it crosses the unstable portion of the n-shaped
The system can be characterized by an n-shaped nullcline, oscillations occur (Fig. 2b). Any small
nullcline and a second nullcline that is relatively perturbation off the unstable steady state (solid
straight that crosses the other at a stable steady black circle) will move to the stable portion of the
state (Fig. 2a). A steady state is the point where n-shaped nullcline, and the red trajectory will be
the derivatives are all equal to zero. The dashed followed in an oscillatory fashion, unable to
portion of the n-shaped nullcline is the unstable return to the stable point. In models of agonist-
portion, and the solid portions are the stable por- induced calcium wave and oscillations, changing
tions. A trajectory will move toward the stable the IP3 concentration would shift the “straight”
portions of the nullcline and away from the unsta- nullcline. Hence for a range of values of IP3
ble portion. At rest, the system will be at its stable (when the “straight nullcline crosses the unstable
steady state shown by the solid black circle. If portion of the n-nullcline (dashed)), oscillations
the perturbation is small and does not cross the will occur.
dashed portion, the trajectory will return to the Classification of calcium waves into five cat-
steady state. A sufficient perturbation (green egories can be accomplished by the mechanism
arrow) will move the trajectory past the unstable giving rise to the waves: (1) membrane depolar-
portion of the n-shaped nullcline. The trajectory ization-triggered calcium waves, (2) agonist-
will then proceed to the right to the stable part of induced calcium waves, (3) fire-diffuse-fire
the nullcline (red arrow). The trajectory will then calcium waves, (4) calcium phase waves, and
track along the n-shaped nullcline (thick red (5) kinematic calcium waves. A common feature
stores resulting in calcium waves. For example,

membrane depolarization, as seen during an
action potential, results in the opening of volt-
age-gated calcium channels. The wave of depo-
larization triggers calcium release from the
intracellular stores in two ways. In the case of
ryanodine receptor type 1 (RyR1) in skeletal
muscle, the conformational change in the volt-
age-gated L-type calcium channel during open-
ing triggers a conformational change in the
adjacent RyR1 causing calcium release from the
sarcoplasmic reticulum. On the other hand, in
cardiac myocytes which contain ryanodine recep-
tor type 2 (RyR2), the calcium entry through
these plasmalemmal calcium channels can trig-
ger calcium release from the internal stores
through a calcium-induced calcium-release
mechanism. These waves are quite fast as in an
action potentials the depolarization propagates
very rapidly across the plasmalemma. In net-
works of neurons, the rapid spread of depolariza-
tion between neurons or astrocytes can trigger
a wave of calcium elevation in these cells. In
other cells, this can involve either ryanodine
receptors, IP3 receptors, or both.
Agonist-Triggered Calcium Waves – A wave
of IP3 has been observed to trigger a wave of
Calcium Waves, Models of, Fig. 2 The dynamics of the
system of differential equations can be studied with calcium release from the endoplasmic reticulum
a phase plane diagram that characterizes the trajectories in oocytes during fertilization. This occurs as IP3
of solutions for two system variables using the vector field binds to the IP3 receptor in the ER-releasing
(values of derivatives for different values of the variables). calcium (Fig. 3a). The calcium activates IP3 pro-
(a) An excitable system has this characteristic n-shaped
nullcline crossed by a second “straight nullcline” duction allowing the wave of IP3 to propagate
establishing a stable steady state (black circle) so that (Fall et al. 2004). Another example is the calcium
sufficient perturbations from the stable steady state waves observed in airway epithelium cells where
(green arrow) cross the unstable portion of the n-shaped calcium waves depend upon IP3 diffusion (Sneyd
nullcline. The trajectory then follows the red arrows gen-
erating a transient rise in the independent variable of the et al. 1995). The speed of these waves is governed
plot. (b) When the nullclines cross, yielding an unstable by the propagation speed of the wave of IP3
steady state (black circle), perturbations off the steady which is controlled by the IP3 diffusion rate and
state result in an oscillatory trajectory its dynamics.
Fire-Diffuse-Fire Calcium Waves – In this
of these mechanisms is that they all involve cal- type of calcium wave (Fig. 3b), calcium released
cium release from the endoplasmic reticulum from the endoplasmic reticulum through chan-
which is an intracellular calcium store. These nels, such as the IP3 receptor or ryanodine recep-
calcium waves can be found in single cells as tor, diffuses to adjacent channels and triggers
well as multicellular systems. release from these sites through calcium-induced
Membrane Depolarization-Triggered Cal- calcium release (Atri et al. 1993; Jafri and Keizer
cium Waves – In this type of wave, a propagating 1995). The wave speed (c) of this type of waves is
signal triggers release of calcium from internal proportional to the square root of the effective
Calcium Waves, Models
of, Fig. 3 Schematic
diagram of mechanisms of
calcium waves. (a)
Agonist-triggered calcium
wave. (b) Fire-diffuse-fire
calcium wave
C
diffusion constant, Deff, for calcium divided by of calcium (and no net movement of calcium).
a characteristic time, T1: Instead, there are localized cycles of uptake and
release that give the appearance of a wave
(Fig. 4). For these waves, the wave speed is
i:e:;c2 ¼ Deff =T1 :
equal to the ratio of the wavelength to the period
of oscillation (Jafri and Keizer 1994; Jafri and
Calcium Phase Waves – Phase waves result Keizer 1995).
from a series of spatially adjacent oscillators in Kinematic Calcium Waves – Kinematic waves
which there is a slight phase delay between oscil- are related to phase waves but actually have weak
lators as one moves down the chain of oscillators. diffusion of calcium. Murray provides
For example, during receptor stimulation by an a mathematical description of kinematic waves
agonist, IP3 is produced which induces calcium (Murray 1993). As calcium is highly buffered,
oscillations. The initial gradient of IP3 can initiate calcium diffusion in cells is slow and can be
calcium oscillations sequentially from sites mov- considered weak. Hence, with oscillatory cal-
ing down the initial gradient. As these sites fire in cium dynamics, kinematic waves are observed.
sequence, it appears that a wave of calcium actu- Waves propagate as a result of local cycles
ally propagates even though there is no diffusion of uptake and release with little diffusion.
Calcium Waves, Models of, Fig. 4 The spatial length release. As the phase progresses, the waves propagate to
between two waves is the wavelength (l), and the time the right as shown by the purple arrows even though there
between two successive transients is the period between is no calcium diffusion. As the phase advances through
oscillations (t). The phase of the oscillation for the spatial one period (from top to bottom on right), the wave
point shown by the red dashed line is shown by the clocks advances to the right by one wavelength (left). Hence the
on the right. When the arrow is pointing up, a calcium wave speed is the ratio of wavelength do period, i.e.,
transient occurs as a cycle of local calcium uptake and c = l/t
This was first suggested for calcium waves by produces fire-diffuse-fire calcium waves, phase
Jafri and Keizer (Jafri and Keizer 1994; Jafri waves, and kinematic waves. The model consists
and Keizer 1995, 1997). This type of wave has of five differential equations for the dynamic
been observed experimentally in astrocytes by variables: cytosolic calcium concentration
Russell and coworkers and has been described ([Ca2+]cyt), endoplasmic reticulum calcium con-
by a model (Roth et al. 1995). As with phase centration ([Ca2+]ER), cytosolic IP3 concentration
waves, the wave speed is equal to the ratio of ([IP3]), and the fraction of channels in the open
the wavelength to the period of oscillation and closed states of the IP3 receptor (X100 and
(Fig. 5). Experimental data from oocytes under X101, respectively). The dynamics of the cyto-
certain conditions also support this relation solic calcium concentration is described by the
(Lechleiter et al. 1998). contributions of diffusion (Jdiffusion), ER calcium
leak (Jleak), calcium release from the IP3 receptor
Example of Deterministic Model for Ca2+ (Jrelease), and sequestration back into the ER by
Waves the ER calcium pump (Jpump). These fluxes are
The Jafri-Keizer model provides an example of buffered by native and exogenous molecules
a computational model for calcium waves that (fluorescent indicator dyes):
where [Bi]cyt is the total concentration of cyto-
solic calcium buffer i and Kcyt
i is its disassociation
constant for calcium with the subscripts i = s, m,
e represent native stationary buffer, native mobile
buffer, and exogenous buffers, respectively. The
diffusion of calcium (Jdiffusion) and the contribu-
C
tions to calcium diffusive flux due to calcium
bound to mobile buffers (native, m, and exoge-
nous, e) are described by
2 2þ
J diffusion ¼ Dcyt
Ca þ gm Dm þ ge De
cyt cyt cyt cyt
∇ ½Ca cyt
!
cyt cyt
gcyt
m Dm gcyt
e De
2 cyt þ ∇½Cacyt
K m þ ½Cacyt K cyt
e þ ½Cacyt
∇½Cacyt ,
(3)
Calcium Waves, Models of, Fig. 5 The wave speed is
equal to the oscillation wavelength divided by the oscilla-
where Dcyt
i is the diffusion constant for calcium,
tion period (c = l/t). Changing the level of IP3 alters native mobile buffer, and exogenous buffers
the oscillation frequency (1/ t) and alters the slope of the (i = Ca, m, e, respectively) and gcyt
i is (i = m, e,
line. Shown are data points from simulations. Normal respectively)
diffusion with [IP3] = 0.52 mM (red squares). No diffu-
sion with [IP3] = 0.52 mM (blue diamonds). Normal dif-
fusion with [IP3] = 0.26 mM (purple circles). No ½B1 cyt K cyt
1
diffusion with [IP3] = 0.26 mM (green triangles). Lines gcyt
i ¼
2 : (4)
represent linear curves fitted for data going through the K cyt
i þ ½Cacyt
origin. (The figure is reproduced from data in Jafri and
Keizer (1994).)
These equations are also derived from the
rapid buffer approximation (Wagner and Keizer
1994).
The release of calcium from the IP3 receptor is
@ ½Ca2þ cyt described by
¼ bcyt J diffusion þ J leak þ J release J pump :
@t

(1) J release ¼ c1 v1 Χ 3110 Ca2þ ER Ca2þ cyt ,
(5)
The buffering of calcium (bcyt) is described by where c1 is the ER/cytosol volume ratio and v1 is
the rapid buffering approximation (Wagner and the maximal calcium ion flux through the channel.
Keizer 1994): There is also a passive leak out of the ER
given by
8
>

< ½Bs cyt K cyt
s J leak ¼ c1 v2 Ca2þ Ca2þ cyt , (6)
bcyt ¼ 1 þ
2 ER
>
: K cyt þ ½Ca2þ
s cyt
91 with maximal rate v2. The pump that sequesters
>
=
½Bm cyt K cyt
m ½Be cyt K cyt
e calcium back into the ER is described by
þ
2 þ
2 ,
>
;
2
v3 ½Ca2þ cyt
m þ ½Ca cyt
K cyt s þ ½Ca cyt
K cyt
2þ 2þ
J pump ¼ , (7)
(2) ½Ca2þ 2cyt þ K 23
transmembrane fluxes have the opposite

with v3 being the maximal pump rate and K3 the
effect and the buffers and diffusion are those of
binding constant for calcium.
the ER:
The balance equation for ER calcium mirrors
that of cytosolic calcium except that the
2 3
2 2þ gER
m Dm
ER
6 Ca
D ER
þ g ER ER
D ∇ ½ Ca 2 ∇ ½ Ca ER :∇ ½ Ca ER v2 þ v Χ 3
1 110 7
m þ ½CaER
m m ER
@ ½Ca2þ ER 6 K ER 7
¼ bER 6
6
2 7,
7
1 v3 ½Ca cyt
2þ
@t 4 ½Ca2þ ½Ca2þ 5
ER cyt þ
c1 ½Ca2þ 2cyt þ k23
(8)
where with IP3, activating and inactivating calcium

bound) binding to IP3. The model for the IP3
( )1
½Bs ER K ER ½Bm ER K ER receptor was first described by DeYoung and
bER ¼ 1 þ s
2 þ
m
2 Keizer as an eight-state system and reduced to
s þ ½Ca ER
K ER m þ ½Ca ER
K ER
2þ 2þ
a two-state system used here (De Young and
(9)
Keizer 1992; Keizer and De Young 1994).
To demonstrate the model, a simulated spiral
and
wave based upon this model (Jafri and Keizer
1995) is shown here (Fig. 6).
½Bm ER K ER
m ¼
gER 2 :
s
(10)
m þ ½CaER
K ER Example of a Stochastic Model for Ca2+ Wave
The physiological events that underlie calcium
The dynamics of IP3 are described by waves are the stochastic opening and closing of
ion channels. Experiments have observed this
@ ½IP3 stochastic behavior as calcium blips, puffs, and
¼ DIP3 ∇2 ½IP3 þ I r ð½IP3 ½IP3 Þ, (8)
@t waves. Hence to understand the dynamics of
these molecular events that underlie calcium
where DIP3 is the diffusion constant for IP3 and Ir waves, stochastic models are needed. As an
is the relaxation rate of IP3 to its steady-state example, a model of a calcium wave in the
value [IP3]*. The state of the IP3 receptor is mature Xenopus laevis oocyte is presented
given by (Ullah 2011). For the simulation, each single IP3
channel was modeled with four equal and inde-
dΧ 100 pendent subunits, where the kinetics of each
¼ a2 Ca2þ cyt Χ100
dt subunit is determined by the simplified version

a5 Ca2þ cyt Χ100 þ b5 Χ110 , (9) of Sneyd and Dufour model (Sneyd and Dufour
2002). The channels were clustered into groups
dΧ 110 of 20 channels. The spatiotemporal Ca2+ dynam-
¼ a2 Ca2þ cyt Χ110 þ b2 c2 ½IP3 ics are described by the following equation:
dt

þ a5 Ca2þ cyt Χ100 b5 Χ110 , (10)
dc
!

!
¼ Dcyt ∇2 c þ J cluster d x xn
with the transition rate constants ai and bi and the dt
proportionality constant c2 for X111 (the state J serca þ J leak , (11)
dcER dp
¼ DER ∇2 cER ¼ Dp ∇2 p þ bðp pÞ þ J stim þ J PLCd ,
dt dt

!

! (13)
þ g J serca J cluster d x xn J leak ,
(12) where “c” is the cytosolic calcium concentration;
cER denotes the concentration of calcium in the
ER; Dcyt and DER are the effective diffusion coef-
C
ficients of calcium in the cytosol and in the ER,
respectively; and Dp is the diffusion coefficient of
IP3. The parameter g denotes the ratio of the
cytoplasmic volume to the ER volume and is
used to account for different volumes of two
compartments. The term Jcluster represents the
Ca2+ of an IP3 receptor containing a small cluster
of channels. The magnitude of this flux depends
on both the number of open channels within the
cluster and the Ca2+ gradient across the ER
membrane:
J cluster ¼ N open p r 2 vchannel ðcER cÞ: (14)
Here r represents the channel cross-sectional

radius. Nopen is the number of open channels, and
Calcium Waves, Models of, Fig. 6 Spiral calcium wave
vchannel is the maximum Ca2+ flux per channel.
produced by the Jafri and Keizer model for agonist-
induced calcium waves in Xenopus laevis oocytes (click Nopen is obtained from the kinetics of the IP3
for movie) receptors, which are modeled stochastically,
Calcium Waves, Models of, Fig. 7 Snapshots of Ca2+ two-dimension grid. The global Ca2+ wave is generated by
wave after injecting an IP3 stimulus. The wave propagates applying an IP3 stimulus Jstim in Eq. 13, along a line at the
to the right; the calcium released in the cytosol provides left edge of the grid. It activates the IP3 receptors and
positive feedback for the propagation of the Ca2+ wave initiates the Ca2+ wave. The IP3 then diffuses to the cyto-
across the cytosol through further activation of IP3R. IP3 sol, activates the internal IP3Rs, opens up the channels,
receptors are modeled with 20 channels per cluster at and releases more Ca2+ from the internal store ER to
a distance of 3 mm uniformly distributed throughout the generate the self-sustaining wave
i.e., the number of open channels is random, as Atri A, Amundson J et al (1993) A single-pool model for
each channel can open and close stochastically. intracellular calcium oscillations and waves in the
Xenopus laevis oocyte. Biophys J 65(4):1727–1739
The leak term Jleak represents the loss of Ca2+ Berridge MJ, Bootman MD et al (1998) Calcium–a life
through the ER membrane, into the cytosol. Sim- and death signal. Nature 395(6703):645–648
ilar to the cluster’s flux, the leak flux also depends De Young GW, Keizer J (1992) A single-pool inositol
on a concentration gradient, as well as vleak the 1,4,5-trisphosphate-receptor-based model for agonist-
stimulated oscillations in Ca2+ concentration. Proc
maximum leak conductance: Natl Acad Sci USA 89(20):9895–9899
Falcke M (2004) Reading the pattern in living cells – the
J leak ¼ vleak ðcER cÞ: (15) physics of Ca2+ signaling. Adv Phys Adv Phys
53:255–440
Fall CP, Wagner JM et al (2004) Cortically restricted
As cell function is highly dependent on a high production of IP3 leads to propagation of the fertiliza-
Ca2+ gradient across the ER membrane, the cell tion Ca2+ wave along the cell surface in a model of the
requires ATP-driven SERCA pumps (sacro-/ Xenopus egg. J Theor Biol 231(4):487–496
endoplasmic reticulum Ca2+) to be distributed Foskett JK, White C et al (2007) Inositol trisphosphate
receptor Ca2+ release channels. Physiol Rev
across the ER membrane. The Jserca term in 87(2):593–658
Eq. 11 represents this mechanism and follows Girard S, Luckhoff A et al (1992) Two-dimensional model
a similar form as Eq. 7: of calcium waves reproduces the patterns observed in
Xenopus oocytes. Biophys J 61(2):509–517
Jafri MS, Keizer J (1994) Diffusion of inositol 1,4,5-
c2
J serca ¼ vserca : (16) trisphosphate but not Ca2+ is necessary for
k2serca þ c2 a class of inositol 1,4,5-trisphosphate-induced
Ca2+ waves. Proc Natl Acad Sci
USA 91(20):9485–9489
Phospholipase C (PLCd) in the plasma mem- Jafri MS, Keizer J (1995) On the roles of Ca2+
brane stimulates production of IP3 upon increas- diffusion, Ca2+ buffers, and the endoplasmic reticu-
ing cytosolic Ca2+ levels, i.e., lum in IP3-induced Ca2+ waves. Biophys
J 69(5):2139–2153
c2 Jafri MS, Keizer J (1997) Agonist-induced calcium waves
J PLCd ¼ vPLCd : (17) in oscillatory cells: a biological example of Burgers’
ðkPLCd þ c2 Þ equation. Bull Math Biol 59(6):1125–1144
Keizer J, De Young GW (1994) Simplification of
PLCd is an important signaling effector as it
a realistic model of IP3-induced. Ca2+ oscillations.
increases sensitivity to IP3 and facilitates an intra- J Theor Biol 166:431–442
cellular calcium wave (Fig. 7). Lechleiter JD, John LM et al (1998) Ca2+ wave dispersion
and spiral wave entrainment in Xenopus laevis oocytes
overexpressing Ca2+ ATPases. Biophys Chem
Conclusions 72(1–2):123–129
Murray JD (1993) Mathematical biology. Springer,
Berlin/New York
In summary, calcium waves are ubiquitous Roth BJ, Yagodin SV et al (1995) A mathematical model
throughout biological systems. They allow the of agonist-induced propagation of calcium waves in
spatial and temporal control of calcium in differ- astrocytes. Cell Calcium 17(1):53–64
ent parts of the cell. As calcium waves are com- Sneyd J, Dufour JF (2002) A dynamic model of the type-2
inositol trisphosphate receptor. Proc Natl Acad Sci
plex nonlinear phenomena, mathematical models USA 99(4):2398–2403
have played an important role in understanding Sneyd J, Wetton BT et al (1995) Intercellular calcium
their dynamics. waves mediated by diffusion of inositol trisphosphate:
a two-dimensional model. Am J Physiol 268(6 Pt 1):
C1537–C1545
Ullah A (2011) Computational modeling of channel clus-
References tering effects on calcium signaling during Oocyte mat-
uration. PhD, Ohio University
Adkins CE, Taylor CW (1999) Lateral inhibition of ino- Wagner J, Keizer J (1994) Effects of rapid buffers on Ca2+
sitol 1,4,5-trisphosphate receptors by cytosolic diffusion and Ca2+ oscillations. Biophys
Ca(2+). Curr Biol 9(19):1115–1118 J 67(1):447–456
Calcium-Dependent Exocytosis, Biophysical Models of 535 C
control proteins such as complexins (Jahn and
Calcium-Dependent Exocytosis, Fasshauer 2012). Despite highly specialized mor-
Biophysical Models of phology, a similar exocytosis-triggering mecha-
nism is found at most high-throughput ribbon
Victor Matveev sensory synapses (Cho and von Gersdorff 2012;
Department of Mathematical Sciences, New Sterling and Matthews 2005), except for cochlear
Jersey Institute of Technology, Newark, NJ, USA hair cells which may differ in both their molecular C
exocytosis machinery and Ca2+ sensitivity
(Nouvian et al. 2011) and where exocytosis may
Synonyms require nonneuronal isoforms of synaptotagmin
(Johnson et al. 2010) or a different Ca2+ sensor,
Non-constitutive exocytosis; Regulated otoferlin (Pangrsic et al. 2012; Roux et al. 2006).
exocytosis The mechanism of exocytosis of hormone-
containing large dense-core vesicles in endocrine
cells is very similar to the mechanisms of neuro-
Definition transmitter vesicle release (Chow et al. 1992; Voets
2000), but is believed to proceed at a somewhat
Calcium-dependent exocytosis is the biochemi- slower rate due to less tight morphology of the
cally controlled fusion of the bilipid secretory release site and more lose coupling between
vesicle membrane with the bilipid cell mem- voltage-dependent calcium channels and release-
brane, triggered by the binding of several Ca2+ ready vesicles (Verhage and Toonen 2007; Wu
ions to control proteins such as synaptotagmins et al. 2009). However, this distinction may be
anchored at the interface between these two specific to only certain subclasses of endocrine
membranes. Exocytosis results in the release of cells, since a close channel-vesicle coupling char-
vesicle contents into the extracellular space, acteristic of that in neurons has been found in
namely, the release of neurotransmitter into the pancreatic beta cells (Barg et al. 2001).
synaptic cleft in the case of neuronal synapses The dependence of the neurotransmitter or
and neuromuscular junctions or the secretion of hormone release rate on Ca2+ concentration is
hormone into the bloodstream in the case of known to be steeply nonlinear, as was first
endocrine cells. Exocytosis also allows the trans- observed by Dodge and Rahamimoff (1967)
membrane proteins contained in the vesicle when studying neuromuscular junction potentials
membrane to be incorporated into the cell mem- at different extracellular Ca2+ concentrations and
brane, although such membrane protein traffick- later confirmed in studies that varied intracellular
ing is more characteristic of Ca2+-independent, Ca2+ concentration directly, using caged-Ca2+
constitutive exocytosis. release in endocrine cells, the giant calyx of
Held synapse, and other synaptic terminals
(Neher and Sakaba 2008). For low, subsaturating
Detailed Description concentrations of Ca2+, this nonlinear relation-
ship can be summarized as
In synapses, neuromuscular junctions, and endo-
Rð½CaÞ / ½Can (1)
crine cells, fast Ca2+-triggered exocytosis of
2+
a neurotransmitter or hormone-containing vesicle Here [Ca] denotes Ca concentration at the
occurs primarily through the interaction of the vesicle fusion site, R denotes instantaneous neu-
so-called SNARE (soluble N-ethylmaleimide- rotransmitter release rate, and n is the intrinsic
sensitive factor attachment protein receptor) pro- (biochemical) Ca2+ cooperativity of exocytosis,
teins spanning the synaptic and vesicle membranes which varies from three to five in most prepara-
with specific isoforms of the Ca2+-sensitive protein tions. Note however that a significantly less coop-
synaptotagmin and with involvement of other erative, near-linear Ca2+ dependence has been
C 536 Calcium-Dependent Exocytosis, Biophysical Models of
reported in mature auditory hair cells, possibly (Chapman 2002), not all of these sites are nec-
because of differences in molecular exocytosis essarily involved in fast (phasic) Ca2+-triggered
sensors mentioned above (Cho and von Gersdorff exocytosis, and therefore it is possible that the
2012; Johnson et al. 2010; Nouvian et al. 2011), relevant Ca2+-sensitive sensors could be distrib-
although it has been suggested that the linear uted among several synaptotagmin molecules,
dependence could potentially arise from the aver- which dimerize and bind Ca2+ simultaneously
aging across synaptic contacts (Heil and to trigger vesicle fusion (Mutch et al. 2011).
Neubauer 2010). However, most models (reviewed below)
The release (exocytosis) rate R is usually esti- assume for simplicity a small number of Ca2+-
mated in units of vesicles per second, and mea- binding sites, all of which have to be
sured as a membrane capacitance increase or by occupied for vesicle fusion to occur. An alter-
electrochemical detection of released molecules, native detailed model that assumes an excess of
using a carbon fiber electrode, but often assessed Ca2+-binding sites and includes Monte Carlo
only indirectly by measuring postsynaptic currents simulation of Ca2+ ion diffusion and binding
or potentials. The steep nonlinear dependence has recently been examined in Dittrich
given by Eq. 1 indicates that simultaneous binding et al. (2013).
of several Ca2+ ions to exocytosis control proteins,
most likely isoforms of synaptotagmin, is needed Sequential Ca2+-Binding Model
for release. In fact, the biochemical cooperativity Assuming for concreteness five distinct Ca2+-
can be viewed as a lower bound on the number of binding sites comprising the putative exocytosis
Ca2+-binding events required for exocytosis gate (sensor) X, the most general Ca2+-sensitive
(Dittrich et al. 2013). Although the C2A and exocytosis process can be described by the fol-
C2B domains of a given synaptotagmin molecule lowing reaction:
do possess a total of five Ca2+ ion-binding sites
½Cakþ
1
½Cakþ
2
½Cakþ
5 g
X Ð

CaX Ð

Ca2 X . . . Ca4 X Ð

Ca5 X ! Fused (2)
k1 k2 k5
where [Ca] is the Ca2+ concentration at the vesi- product g[Ca5X]. Note that the Ca2+ concentra-
cle site, k
j are the binding and unbinding rates of tion, [Ca], should be modeled independently; in
each binding site, and the final irreversible reac- the simplest case of global Ca2+ elevation pro-
tion represents the actual vesicle fusion event. In duced by caged Ca2+ release, it is approximately
a deterministic simulation, this reaction is constant, or it can be represented as a brief pulse
converted to a system of ordinary differential of certain width and amplitude when modeling
equations, using the principle of mass action: vesicle release produced by a train of action
potentials.
8
>
>
d ½X The binding and unbinding rates k j are in
>
> ¼ kþ
1 ½Ca½X þ k1 ½CaX
< dt general distinct, and if the earlier unbinding
... rates are slow, a significant accumulation of
>
>
>
: d ½Ca5 X ¼ kþ ½Ca½Ca4 X g þ k ½Ca5 X
> partially bound states of X can result during
dt 5 5 a train of stimuli; such accumulation due to
(3) slow unbinding is the basis for the so-called
bound-Ca2+ model of synaptic facilitation (see
where [CanX] represents the fraction of exocyto- entry on “▶ Facilitation, Biophysical Models”;
sis gates with n-binding sites occupied by a Ca2+ Bertram et al. 1996; Bornschein et al. 2013;
ion. The maximal fusion rate is given by the Matveev et al. 2006).
Parallel Ca2+-Binding Model exocytosis gate X are identical and can bind Ca2+
It is instructive to first consider the simplest sce- independently, leading to the following simplified
nario where all five release sites comprising version of the reaction given by Eq. 2:
5½Ca kþ 4½Ca kþ ½Ca kþ g

X Ð CaX Ð Ca2 X . . . Ca4 X Ð Ca5 X ! Fused (4)

k 2k 5k C
Here the final reaction representing vesicle where [Y] = 1[CaY] represents the fraction of
fusion is irreversible and is triggered when all unbound gates. In this case, the fusion rate is
five binding sites are occupied; exocytosis pro- given by the product g[CaY]5 (identically equal
ceeds at rate R = g [Ca5X]. Even though Eq. 4 to g[CaX] in reaction (4)), where [CaY] is the
appears to describe a series of five consecutive bound fraction (probability) of each gate Y.
binding reactions, it is equivalent to five identical Some implementations of vesicle release
reactions occurring in parallel, with fusion taking process postulate the existence of an indepen-
place when all five sites are bound, as follows: dent final release-promoting conformational
transition reaction occurring after the release
8
> ½Ca kþ sensor is fully Ca2+ bound, as, for instance,
>
<Y Ð
1
CaY,
g
5CaY ! Fused

k1 done in the model of neurotransmitter release
>
: d½CaY ¼ kþ ½Cað1 ½CaYÞ k ½CaY
> at the calyx of Held synaptic terminal by
dt Bollmann and Sakmann (2005) and Bollmann
(5) et al. (2000):
5½Ca kþ 4½Ca kþ ½Ca kþ g r

X Ð

CaX Ð Ca2 X . . . Ca4 X Ð Ca5 X Ð Ca5 X ! Fused (6)
k 2k 5k d
Here r is the maximal vesicle fusion rate. Cooperative Ca2+-Binding Model

In Bollmann et al. (2000), the parameter Some studies suggest that the Ca2+-sensitive exo-
values that we found to fit well the data from cytosis triggers exhibit strong cooperativity,
the calyx of Held synaptic terminal are whereby the target protein undergoes a
given by r = 40 ms1, k+ = 0.3 mM1ms1, conformational change with each successive
k = 3 ms1, g = 30 ms1, and d = 8 ms1. Ca2+ ion binding, which in turn leads to an
The inclusion of an additional post-binding increase in the Ca2+ affinity of the remaining
step helps to achieve more constant shape of (yet unoccupied) Ca2+-binding sites. The follow-
the release time course at different amplitudes ing widely used modification of reaction given by
of Ca2+ influx, as seen in experiment Eq. 2 has been proposed to implement this
(Bollmann et al. 2000; Yamada and Zucker possibility (Felmy et al. 2003; Heidelberger
1992). et al. 1994; Schneggenburger and Neher 2000):
5½Ca kþ 4½Ca kþ 3½Ca kþ 2½Ca kþ ½Ca kþ g

X Ð

CaX Ð Ca2 X Ð Ca3 X Ð Ca4 X Ð Ca5 X ! Fused (7)
k 2bk 2 3 4
3b k 4b k 5b k
Note that cooperative binding can be Exocytosis Rate at a Steady Ca2+

represented as either a progressive increase in Concentration
forward binding rates or decrease of backward To simulate the exocytosis rate during prolonged
rates; Eq. 7 corresponds to the latter possibility. Ca2+ elevation, for instance, to reproduce caged-
The cooperativity parameter should satisfy Ca2+ release experiments, it is sufficient to
b < 1, with a typical value used in the literature consider the equilibrium point of the reactions
of 0.25, indicating that the final Ca2+-binding summarized above. At equilibrium, the occupancy
reactions are only slowly reversible. In of the Ca2+-bound states is easily found by equat-
Schneggenburger and Neher (2000) and ing the right-hand side of the binding reaction to
Wolfel and Schneggenburger (2003), quantify- zero. For the case of independent gates, described
ing the Ca2+ dependence of vesicle release at by Eqs. 4 and 5, the equilibrium is given by
the calyx of Held synaptic terminal, the
following parameter values were obtained: kþ ½Ca ½Ca
½CaY ¼ ¼ (8)
k+ = 0.09 mM1 ms1, k = 9.5 ms1, þ
k ½Ca þ k ½Ca þ K
g = 6 ms1, and the cooperativity parameter
b = 0.25. where the ratio of the backward and forward rates
Note that the use of the term “cooperativity” K = k/k+ is referred to as the Ca2+ affinity
is ambiguous in the context of exocytosis or dissociation constant and is an important
mechanisms, since it is used to refer to two dif- Ca2+-sensitivity parameter: the lower the K, the
ferent properties of the exocytosis process: higher is the affinity (sensitivity) of the release
cooperativity can either refer to the number of sensor. It follows from Eq. 8 that K represents
Ca2+-binding sites (as inferred from the log-log the Ca2+ concentration at which half the sensors
slope of the Ca2+ dose–response curve) or indi- become bound. Since the final fusion reaction in
cate that the Ca2+-binding affinity of these sites is Eq. 5 requires the binding of all five sensors, the
not equal and increases as the first sites become release rate at equilibrium will be given by the 5th
Ca2+ bound. power of the steady-state binding occupancy of
gate Y:
Models with Independent Sets of Ca2+-
Binding Sites 5
½Ca
Several recent biophysically detailed models Rð½CaÞ ¼ g (9)
½Ca þ K
of exocytosis take into account the possibility
that Ca2+ may bind to several different At low Ca2+ concentration, this agrees with
proteins or to different domains of the same the cooperativity condition given by Eq. 1. Note
protein, such as the C2A and C2B domains
that the release rate equals 1/25 of its maximal
of synaptotagmin, each characterized by value when [Ca2+] = K.
a distinct set of rates and cooperativity values. For the case of cooperative binding, Eq. 7, the
Examples of such models are given below in steady-state fusion rate R is given by the equilib-
the context of more comprehensive models rium value of fully bound state Ca5X, which
of exocytosis – see model schemes (16) obeys a modified form of the Adair-Klotz-
and (20). Pauling equation (Weiss 1997)
g ½Ca5
Rð½CaÞ ¼ (10)
½Ca þ 5 ½Ca b K þ 10 ½Ca b K 2 þ 10 ½Ca2 b9 K 3 þ 5 ½Cab10 K 4 þ b10 K 5
5 4 4 3 7
where [Ca] is the Ca2+ concentration at the saturating part of the Hill curve and can be
release site. In the limit of very small values of greatly improved by omitting the data from sat-
b (strong cooperativity limit), the above expres- urating levels of [Ca2+]. Therefore, the most
sion approaches the well-known and widely used model-independent way of estimating the
Hill function: cooperativity factor n is to only fit the log-log
slope of the experimentally obtained Ca2+ sen-
sitivity curve below the saturation inflection, C
½Can
Rð½CaÞ g (11) which is often done in practice.
½Can þ K nD
Cumulative Release During Prolonged
where n = 5 and KD b2K (b << 1). The dis- Ca2+ Elevation
sociation constant KD quantifies the Ca2+ sensi- The Ca2+ dependence of release rate can be used
tivity of the entire 5-step process, as opposed to to quantify the total cumulative release with
the sensitivity of any individual Ca2+-binding prolonged Ca2+ elevation, as measured by the
site: when [Ca] = KD = b2K, the exocytosis total membrane capacitance increase or the total
rate reaches its half-maximal value. Note that amount of released neurotransmitter. In the case
this is the most constrained and most important of prolonged and approximately steady Ca2+ ele-
model parameter, since extensive experimental vation of duration Dt, the amount of fused vesi-
data on Ca2+ sensitivity has been collected at cles is given by Quastel et al. (1992):
various types of synaptic terminals: in general,
physiological rates of release are found when Fð½CaÞ ¼ Fmax 1 exp Req ð½CaÞ Dt (12)
[Ca2+] at the vesicle location reaches the range
of 10–50 mM (reviewed in Neher and Sakaba where Fmax represents the total amount of avail-
2008), although there are reports that highly sen- able exocytosis resources (say, vesicles) and
sitive vesicle pools exist in some cells, possibly Req([Ca]) represents the equilibrium reaction
controlled by distinct isoforms of the rate at Ca2+ concentration [Ca] attained at the
synaptotagmin sensor with Ca2+ affinity of sev- vesicle location. The exponential term can be
eral mM (reviewed in Pedersen and Sherman interpreted as the probability of release failure.
2009). For pulses raising [Ca2+] to subsaturating levels,
Note that the Hill functional form given by given Eq. 1 we have R k[Ca]n; therefore
Eq. 11 should only be viewed as a crude qualita- (Quastel et al. 1992)
tive approximation of the true Ca2+ dependence
of exocytosis rate, even at equilibrium, and tends Fð½CaÞ ¼ Fmax ½1 expðk ½Can DtÞ (13)
to be somewhat overused (Weiss 1997). This is
because the agreement between Eqs. 10 and 11 However, if [Ca2+] is high enough to saturate
becomes sufficiently accurate only for values of the release sensor during the long depolarizing
cooperativity parameter satisfying b < 0.1; in pulse, the total amount of released neurotransmit-
the absence of firm experimental evidence for ter will become limited by the duration of stimu-
such a strong cooperativity, the Hill function lation only:
should be avoided. Assuming a more moderate
but still strong cooperativity corresponding to F ¼ Fmax ½1 expðRmax DtÞ (14)
the typically used value b = 0.25, fitting the
data to a Hill function would lead to a gross As was explained above, Eqs. 12 and 13 only
underestimate of the true number of Ca2+- apply to the situation where [Ca2+] is elevated
binding sites, as illustrated in Fig. 1. This under- for a long time relative to the kinetics of the Ca2+-
estimate is mostly due to the poor match of the binding sites, as, for instance, during whole-cell
Exact: n=5, KD=K b2 =1μM Exact: n=5, KD=K b2 =1μM

Hill fit: n=3.2, KD=1.3μM Hill fit: n=4.3, KD=1.1μM
0.35
0.9
0.3
0.8
0.7 0.25
Release rate
Release rate
0.6 0.2
0.5
0.4 0.15
0.3 0.1
0.2
0.05
0.1
10−1 100 101 10−1 100

[Ca2+] (μM) [Ca2+] (μM)
Calcium-Dependent Exocytosis, Biophysical Models low quality of the fit is mostly a result of poor matching of
of, Fig. 1 Hill function fit underestimates true biochem- the saturating part of the curve by the Hill function: the
ical cooperativity and affinity of a 5th-order Ca2+-binding cooperativity is more accurately predicted if only the
reaction (Eqs. 7 and 10), even in the case of moderately non-saturating part of the data is considered (right panel)
strong cooperativity parameter (b = 0.25). Note that the
[Ca2+] elevation produced by Ca2+ uncaging and at some synapses represents a significant con-
experiments. If on the other hand [Ca2+] is under- tribution to the overall neurotransmitter release
stood as the peak concentration achieved in the under physiological conditions. A prominent fea-
vicinity of the vesicle during a brief action poten- ture of asynchronous release is that it seems to
tial, the above steady-state results would exhibit a lower, almost linear Ca2+ dependence
overestimate the true release and could only (Lou et al. 2005; Sun et al. 2007). As will be
serve as an upper bound on release rate discussed further below, there is some debate
(Shahrezaei and Delaney 2005). The “true” neu- whether the distinct components of synaptic
rotransmitter release rate would be given by transmission are all manifestations of the same
a solution to the differential equation (5), with Ca2+-dependent process or are in fact caused by
[Ca2+] representing the time-dependent Ca2+ distinct vesicle pools released via independent
concentration at the vesicle release site, which mechanisms (Chung and Raingo 2013; Smith
has to be modeled independently, using simula- et al. 2012). However, it should be noted that
tions of Ca2+ entry through Ca2+ channels, diffu- the mechanisms of asynchronous release may
sion, and binding to intracellular Ca2+ buffers. vary across distinct types of synapses (Kaeser
and Regehr 2014).
Release at Low Ca2+ Concentration: The so-called allosteric model proposed by
Asynchronous and Spontaneous Lou et al. (2005) explains the observed decrease
Transmission in apparent cooperativity of Ca2+ action at low
At very low Ca2+ concentration, depolarization- Ca2+ concentration by the presence of a slow
evoked exocytosis is more stochastic and less conformational change of the release proteins,
synchronized with the stimulus, and its rate leading to a second route of vesicle fusion,
decreases, becoming comparable to the sponta- which becomes progressively less likely at higher
neous transmission observed in the absence of Ca2+ concentration. This is implemented by the
stimulation. This delayed component of exocyto- following scheme with a reverse-cooperativity
sis is termed asynchronous or delayed exocytosis parameter f < 1:
5 ½Ca kþ ½Ca kþ
Reserve
Ð X
)*
CaX ... Ca4 X
)*
Ca5 X
pool b0 k 5 b4 k (15)
#a # af # af4 # af5
Fused Fused Fused Fused
A more detailed model proposed by Sun synaptic response at very low levels of Ca2+. This C
et al. (2007) and referred to as the dual-sensor model assumes two independent Ca2+ sensors
model provides an improved quantitative acting in parallel and each triggering a distinct
description of release at low Ca2+ concentrations, mode of neurotransmitter release:
since it reproduced more accurately the latency of
Spontaneoous
Release
r"
RRP
5 a Ca 4 a Ca a Ca g Synchronous
X0 Y0 *
) X1 Y0 * ...
)
)*
X5 Y0 !
b 2 bb 5 b4 b release
d " # 2 f Ca d "# 2 f Ca d "# 2 f Ca
X0 Y1 *
) X1 Y1 * ...
)
)*
X5 Y1 !
b 2 bb 5b b4 release
2 b d " # f Ca 2 b d "# f Ca 2 b d "# f Ca
X0 Y2 *
) X1 Y2 * ...
)
)*
X5 Y2 !
b 2 bb 5b b4 release
#g #g #g
Asynchronous Asynchronous Asynchronous
release release release
(16)
Here the synchronous release involves Multiple Vesicle Pool Models

a sensor “X” with Ca2+ cooperativity of five, Apart from the intrinsic heterogeneity in Ca2+
while a different and independent sensor “Y” is affinity and speed of exocytosis due to the exis-
responsible for asynchronous release and pos- tence of distinct release pathways guiding exocy-
sesses a cooperativity of two. Note that in this tosis of each vesicle, an additional explanation
model implementation, both sensors bind Ca2+ for the observed heterogeneous components of
cooperatively, with equal cooperativity parame- vesicle release is the existence of independent
ter b. The existence of two distinct sensors is pools of vesicles with variable degree of “pre-
supported by recent evidence that in some synap- paredness” for exocytosis. This heterogeneity is
ses asynchronous release requires a special especially pronounced in endocrine cells, which
non-synaptotagmin sensor Doc2 (Yao exhibit a fast-decaying initial phase of release
et al. 2011). However, the mechanisms of asyn- followed by a second, more slowly decaying
chronous and spontaneous release are still under component, but such behavior is also found in
debate and may vary across synaptic types other high-throughput neuronal synapses such
(Chung and Raingo 2013; Kaeser and Regehr as the calyx of Held synaptic terminal and in
2014; Smith et al. 2012). ribbon synapses (Neher 2012; Neher and Sakaba
2008; Pedersen and Sherman 2009; Verhage and Lee et al. 2013; Sorensen 2004; Verhage and
Toonen 2007). The identity of such distinct ves- Toonen 2007)). Most quantitative models of exo-
icle pools is currently under debate: they could be cytosis that include multiple releasable vesicle
distinguished either by the variations in the dis- pools are based on the two-pool model of
tance between vesicles and corresponding Heinemann et al. (1993), Voets (2000), and
voltage-gated Ca2+ channels (“positional prim- Voets et al. (1999) that was first put forward to
ing”: Wadel et al. 2007) or by differences in the quantify two temporal components of secretory
arrangement of the molecular machinery vesicle release in adrenal chromaffin cells. Here
needed for exocytosis (“molecular priming” and is a recent implementation of such a two-pool
“super-priming” (Chung and Raingo 2013; model explored by Sorensen (2004):
Docking Priming
10s
300s 30s
Reserve ko Docked k1 Slow k2 Fast
Ð Ð Ð
pool
ko pool
k1 pool ðSÞ
k2 pool ðFÞ
bs "# 3as Ca bf "# 3af Ca
CaS CaF
2bs "# 2as Ca 2bf "# 2af Ca (17)
Ca2 S Ca2 F
3bs "# as Ca 3bf "# af Ca
Ca3 S Ca3 F
# gs # gf
Slow Fast
Release Release
In this model exocytosis proceeds with insulin-secreting beta cells (Yang and Gillis
cooperativity of three from both the slow 2004). A recent multiple-pool model that
(“sustained”) and fast synchronous pools, which includes the HCSP pool along with the lower-
are replenished via two preparatory steps, affinity immediately releasable pool (IRP) and
docking and priming, from the reserve pool. implements a deterministic bi-domain model of
There are indications that in some cells, the [Ca2+] dynamics can be found in Pedersen and
slowly releasing pool may in fact have a much Sherman (2009). This model is based on an
higher sensitivity to Ca2+, in the range of several earlier model of Chen et al. (2008), and it accu-
mM rather than 10 s of mM, which allows it to rately predicts the characteristic biphasic release
participate in exocytosis despite a much greater of insulin from pancreatic beta cells, with the
separation from the Ca2+ channels compared to first phase of secretion mostly due to the docked
the immediately releasable, fast pool (reviewed vesicles that are rapidly converted to immedi-
in Pedersen and Sherman 2009). The delay in the ately releasable pool (IRP) and exocytosed in
release of the highly sensitive vesicles is response to local “microdomain” [Ca2+]
explained by the longer diffusional distance ([Ca]MD) entering the cell through L-type Ca2+
and therefore longer time required for Ca2+ con- channels, whereas the second, delayed phase of
centration to reach mM range far from the Ca2+ insulin release builds up more slowly due to the
channels. Such so-called highly Ca2+-sensitive recruitment of the reserve vesicles into the
pool (HCSP) has been described in adrenal chro- HCSP and gradual accumulation of cytosolic
maffin cells (Yang et al. 2002), rod photorecep- Ca2+ ([Ca]Cyt) entering through R-type Ca2+
tor ribbon synapses (Thoreson et al. 2004), and channels:
Reserve
pool
Almost
docked
R-type VDCC L-type VDCC C

[Ca] Cyt HCSP Docked Primed IRP [Ca] MD ð18Þ
n [Ca] Cyt n[Ca] MD

FHCSP FIRP
RHCSP RIPR
Here Ca2+-dependent transitions are indicated observed with global (whole-terminal) Ca2+ ele-
in color, according to the primary source of Ca2+ vations that activate vesicles in all pools. There-
for the corresponding transition. The R-type and fore, both sources of heterogeneity, the ones
L-type voltage-dependent Ca2+ channels intrinsic to each vesicle and the extrinsic ones
(VDCCs) contribute respectively to the cytosolic (distinct pools), seem to be relevant in many
and the microdomain pools of Ca2+, [Ca]Cyt, and synapses (Kaeser and Regehr 2014). In order to
[Ca]MD (the diffusional exchange between these take into account all factors of vesicle release
two Ca2+ pools is not shown). The Ca2+- heterogeneity and predict more accurately the
dependent exocytosis steps lead to the two exocytosis rate under low Ca2+ conditions, one
“fused” states, FHCSP and FIRP, which in turn can combine the dual-pool model with the
feed into the final released insulin states RHCSP allosteric model of Lou et al. (2005) or the
and RIRP through a final Ca2+-independent pore- dual-sensor model of Sun et al. (2007). This
expansion transition. was done, for instance, by Wolfel and
Although the existence of multiple vesicle Schneggenburger (2003), who considered allo-
pools can also be considered as a potential steric Ca2+ binding for both pools, as in reaction
explanation for asynchronous release (Chung (15) (neglecting however the exchange between
and Raingo 2013), it is likely that intrinsic the two pools due to the short timescales
sources of heterogeneity in release properties considered in that work). Finally, it should be
captured by the allosteric and the dual-sensor noted that the separation of releasable vesicles
models (see above) are also present in each into only two pools instead of a larger set of
vesicle (Kaeser and Regehr 2014; Sun pools or a continuum of states is most
et al. 2007). For instance, experiments that dis- probably a simplification, but this level of
rupt specific molecular components of release detail is sufficient to accurately quantify release
machinery indicate that at some synapses syn- under most physiological conditions. Including
chronous and asynchronous release can be inde- more than two pools would most likely lead to
pendently manipulated and that asynchronous an underdetermined model and to data over-
release is mediated by a distinct Ca2+ sensor, fitting.
Doc2 (Yao et al. 2011), and possibly involves
a distinct isoform of the vesicle SNARE protein Models with Ca2+-Dependent Vesicle Priming
VAMP4 (Raingo et al. 2012). Further, both the Vesicle docking and priming steps in reactions
synchronous and the delayed components are (17)–(18) most likely include multiple molecular
processes (involving Munc18 and other pro- a mechanism of short-term synaptic facilita-
teins) and morphological steps (e.g., positional tion (Dittman et al. 2000; Millar et al. 2005;
priming), the identity and sequence of which is Pan and Zucker 2009; Worden et al. 1997)(see
still under investigation (Lee et al. 2013; also encyclopedia entry on “▶ Facilitation,
Verhage and Toonen 2007). Experimental evi- Biophysical Models”).
dence suggests that at least some of these vesicle 2. Under conditions of high vesicle release prob-
priming steps are Ca2+ dependent in many syn- ability, Ca2+-dependent priming would mani-
apses, albeit with a lower Ca2+ cooperativity fest itself through activity-dependent
compared to exocytosis itself, with a near-linear acceleration of recovery from short-term syn-
dependence on intracellular Ca2+ concentration aptic depression (Dittman and Regehr 1998;
(Neher and Sakaba 2008). Depending on the Hosoi et al. 2007; Stevens and Wesseling
balance between the different vesicle pools at 1998; von Ruden and Neher 1993; Wang and
resting Ca2+, the Ca2+-dependent priming can Kaczmarek 1998).
manifest itself in two different properties of The simplest model of release that includes a
short-term synaptic plasticity (Zucker and Ca2+-dependent priming step can be obtained by
Regehr 2002): modifying reaction (7): for instance, Millar
1. Under conditions of low initial release-ready et al. (2005) considered the following scheme
vesicle pool size, Ca2+-dependent priming (see also Bornschein et al. (2013) and Sakaba
would cause the secretion rate to increase dur- (2008) for a slightly modified version of this
ing stimulation as a result of an increase in scheme):
vesicle “mobilization” to the release-ready,
primed pool, which can serve as
½Ca kprime kfill 5½Ca kþ 4½Ca kþ ½Ca kþ g

U Ð U Ca Ð X Ð CaX Ð . . . Ð Ca5 X ! Fused (19)
kunprime kunfill k 2bk 5b4 k
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Priming Fast Ca2þ -triggered secretion
The study of Millar et al. (2005) also consid- binding to C2A and C2B domains of
ered a more detailed version of this model that synaptotagmin, leading to the following modifica-
takes into account the independence of Ca2+ tion of the above (see also Pan and Zucker 2009):
U
kunprime "# ½Ca kprime
UCa
kunfill "# ½Ca kfill
3 kþ
1 ½Ca 2 kþ
1 ½Ca kþ
1 ½Ca
V0, 0 Ð

V0, 1 Ð V0, 2 Ð V0, 3
k1 2 b1 k
1 3 b21 k
1
k þ
2 "# 2k 2 ½Ca k þ
2 "# 2k 2 ½Ca k þ
2 "# 2k 2 ½Ca
3 kþ
1
½Ca 2 kþ
1
½Ca kþ
1
½Ca
V1, 0 Ð

V1, 1 Ð V1, 2 Ð V1, 3
k1 2 b1 k1 3 b21 k
1
2 b2 k þ
2 "# 2 k 2 ½Ca 2 b2 k þ
2 "# 2 k 2 ½Ca 2 b2 k þ
2 "# 2 k 2 ½Ca
3 kþ
1
½Ca 2 kþ
1
½Ca kþ
1
½Ca g
V2, 0 Ð

V2, 1 Ð V2, 2 Ð V2, 3 ! Fused
k1 2 b1 k1 3 b21 k
1
(20)
Here the horizontal state transitions represent junctions (Millar et al. 2005; Pan and Zucker
the binding of the three C2A sites of 2009).
synaptotagmin, while the vertical transitions cor- Vesicle priming has also been considered in
respond to the binding of the C2B sites, which is the context of more detailed models that include
assumed to be independent of the Ca2+ binding of multiple releasable vesicle pools. In such models,
the C2A domain. Both domains are assumed to the priming step is usually assumed to occur
bind Ca2+ cooperatively (with cooperativity upstream of the process of exchange between C
parameter b1 = b2 = 0.5). These models were the slow-releasing and the fast-releasing vesicle
successful in explaining the difference between pools. Adding the priming step to scheme (17)
release properties of the facilitating “tonic” and leads to the following class of models (Verhage
depressing “phasic” crustacean neuromuscular and Toonen 2007):
Docking Priming Super-priming

Reserve Docked Ca Slow Ca ð?Þ Fast
Ð Ð Ð
pool pool pool pool
"# nCa "# mCa
(21)
Can S Cam F
# #
Slow Fast
Release Release
In this scheme, “super-priming” refers to the Comprehensive Biophysical Models

conversion of the slowly releasable pool to the fast of Ca2+-Dependent Exocytosis
releasable pool, the nature and the Ca2+ depen- Several particularly detailed recent modeling
dence of which is currently unknown. Note that studies include simulations of Ca2+ dynamics
the use of the term “super-priming” is model and combine together many of the mechanisms
dependent; recently it has been used to describe reviewed above to build a more comprehensive
a newly identified additional kinetic step in the model of vesicle exocytosis:
priming of the fast pool at the calyx of Held 1. The model of Dittrich et al. (2013) is one of
synaptic terminal (Lee et al. 2013). The possible the most complete recent models of exocy-
Ca2+ dependence of the conversion of the slow tosis that includes stochastic simulations of
pool to the fast pool is unknown; it is possible Ca2+ ions diffusion, buffering, and binding to
that this conversion may even be retarded by multiple synaptotagmin molecules and takes
Ca2+ at some synapses (Neher and Sakaba 2008). into account most recent data on the copy
At the calyx of Held, the contribution of the fast number of synaptotagmin molecules and
component of release is observed to increase with SNARE complexes per vesicle (Chapman
increasing Ca2+ levels, at the expense of the slow 2002; Han et al. 2004; Mutch et al. 2011).
component (Wolfel et al. 2007). An explanation of This is the first modeling study that consid-
this observation was suggested that the slow pool ered the possibility that only a subset of
may be small at rest but increases due to activity- synaptotagmin sites have to be Ca2+
dependent conversion of vesicles from the fast bound to trigger exocytosis, and it
pool and that this conversion is less pronounced examines the effective Ca2+ cooperativity
at high Ca2+ levels, so it is not able to reduce the of exocytosis as an emergent characteristic
fast pool at more intense stimulation levels of such partial binding of a subset of Ca2+-
(Neher and Sakaba 2008). binding sites.
2. The study of Pan and Zucker (2009) builds (Bucurenciu et al. 2008; Ermolyuk et al. 2013;
a comprehensive model of release and Matveev et al. 2009, 2011; Meinrenken
short-term plasticity at tonic and phasic crus- et al. 2002, 2003; Schmidt et al. 2013; Zucker
tacean neuromuscular junctions, based on the and Fogelson 1986) and steady-state approxima-
scheme (20) reviewed above but adds positions of Ca2+ distribution near an array of open
tional priming, leading to a transition scheme Ca2+ channels (Bertram et al. 1999; Coggins and
describing vesicle exchange between the Zenisek 2009).
following distinct vesicle pools: reserve,
docked, primed channel-detached, and primed
channel-attached. One of the main aims of this Cross-References
study was to build a model that most fully
accounts for the properties of short-term syn- ▶ Facilitation, Biophysical Models
aptic facilitation observed at crustacean tonic
neuromuscular junctions, as well as the
properties of short-term depression and asyn- References
chronous release at phasic neuromuscular
junctions. Barg S, Ma X, Eliasson L, Galvanovskis J, Gopel SO,
Obermuller S, Platzer J, Renstrom E, Trus M,
The latter model of Pan and Zucker (2009) is
Atlas D, Striessnig J, Rorsman P (2001) Fast exocyto-
a more detailed example of a class of comprehen- sis with few Ca(2+) channels in insulin-secreting
sive models of Ca2+-dependent exocytosis that mouse pancreatic B cells. Biophys J 81:3308–3323
go beyond the simulation of vesicle release in Bennett MR, Farnell L, Gibson WG (2004) The facilitated
probability of quantal secretion within an array of
response to a single presynaptic depolarization, calcium channels of an active zone at the amphibian
aiming to build a comprehensive model of short- neuromuscular junction. Biophys J 86:2674–2690
term changes in synaptic transmission strength in Bertram R, Sherman A, Stanley EF (1996) Single-domain/
response to repeated stimulation. Such modeling bound calcium hypothesis of transmitter release and
facilitation. J Neurophysiol 75:1919–1931
work is reviewed in more detail in the entry
Bertram R, Smith GD, Sherman A (1999) Modeling study
“▶ Facilitation, Biophysical Models.” of the effects of overlapping Ca2+ microdomains on
neurotransmitter release. Biophys J 76:735–750
Spatial Coupling Between Vesicle and Ca2+ Bollmann JH, Sakmann B (2005) Control of synaptic
strength and timing by the release-site Ca2+ signal.
Channels
Nat Neurosci 8:426–434
Building a comprehensive model of Ca2+- Bollmann JH, Sakmann B, Borst JG (2000) Calcium sen-
dependent exocytosis at a particular class of syn- sitivity of glutamate release in a calyx-type terminal.
aptic terminal requires accurate description of Science 289:953–957
Bornschein G, Arendt O, Hallermann S, Brachtendorf S,
Ca2+ dynamics at the release site, which in turn Eilers J, Schmidt H (2013) Paired-pulse facilitation at
requires taking into account the latest data on the recurrent Purkinje neuron synapses is independent of
spatial coupling between voltage-dependent Ca2+ calbindin and parvalbumin during high-frequency
channels and synaptic vesicles (Eggermann activation. J Physiol 591:3355–3370
Bucurenciu I, Kulik A, Schwaller B, Frotscher M, Jonas
et al. 2012; Gentile and Stanley 2005; Moser
P (2008) Nanodomain coupling between Ca2+ chan-
et al. 2006; Oheim et al. 2006; Stanley 1997). nels and Ca2+ sensors promotes fast and efficient trans-
The geometric arrangement of channels and ves- mitter release at a cortical GABAergic synapse.
icles at the release site has been recently a subject Neuron 57:536–545
Chapman ER (2002) Synaptotagmin: a Ca(2+) sensor that
of intensive modeling effort, using both stochas-
triggers exocytosis? Nat Rev Mol Cell Biol 3:498–508
tic simulation of Ca2+ diffusion and buffering Chen YD, Wang S, Sherman A (2008) Identifying the
(Bennett et al. 2004; Glavinovic and Rabie targets of the amplifying pathway for insulin secretion
2001; Nadkarni et al. 2012; Scimemi and Dia- in pancreatic beta-cells by kinetic modeling of granule
exocytosis. Biophys J 95:2226–2241
mond 2012; Shahrezaei et al. 2006; Shahrezaei Cho S, von Gersdorff H (2012) Ca(2+) influx and neuro-
and Delaney 2005) as well as deterministic solu- transmitter release at ribbon synapses. Cell Calcium
tion of continuous reaction–diffusion equations 52:208–216
Chow RH, Lv R, Neher E (1992) Delay in vesicle fusion Hosoi N, Sakaba T, Neher E (2007) Quantitative analysis
revealed by electrochemical monitoring of single of calcium-dependent vesicle recruitment and its func-
secretory events in adrenal chromaffin cells. Nature tional role at the calyx of Held synapse. J Neurosci
356:60–63 27:14286–14298
Chung C, Raingo J (2013) Vesicle dynamics: how synap- Jahn R, Fasshauer D (2012) Molecular machines
tic proteins regulate different modes of neurotransmis- governing exocytosis of synaptic vesicles. Nature
sion. J Neurochem 126:146–154 490:201–207
Coggins M, Zenisek D (2009) Evidence that exocytosis is Johnson SL, Franz C, Kuhn S, Furness DN, Ruttiger L,
driven by calcium entry through multiple calcium Munkner S, Rivolta MN, Seward EP, Herschman HR,
C
channels in goldfish retinal bipolar cells. Engel J, Knipper M, Marcotti W (2010)
J Neurophysiol 101:2601–2619 Synaptotagmin IV determines the linear Ca2+ depen-
Dittman JS, Regehr WG (1998) Calcium dependence and dence of vesicle fusion at auditory ribbon synapses.
recovery kinetics of presynaptic depression at the Nat Neurosci 13:45–52
climbing fiber to Purkinje cell synapse. J Neurosci Kaeser PS, Regehr WG (2014) Molecular mechanisms for
18:6147–6162 synchronous, asynchronous, and spontaneous neuro-
Dittman JS, Kreitzer AC, Regehr WG (2000) Interplay transmitter release. Annu Rev Physiol 76:333–363
between facilitation, depression, and residual calcium Lee JS, Ho WK, Neher E, Lee SH (2013) Superpriming of
at three presynaptic terminals. J Neurosci synaptic vesicles after their recruitment to the readily
20:1374–1385 releasable pool. Proc Natl Acad Sci
Dittrich M, Pattillo JM, King JD, Cho S, Stiles JR, USA 110:15079–15084
Meriney SD (2013) An excess-calcium-binding-site Lou X, Scheuss V, Schneggenburger R (2005) Allosteric
model predicts neurotransmitter release at the neuromodulation of the presynaptic Ca2+ sensor for vesicle
muscular junction. Biophys J 104:2751–2763 fusion. Nature 435:497–501
Dodge FA, Rahamimoff R (1967) Cooperative action of Matveev V, Bertram R, Sherman A (2006) Residual bound
calcium ions in transmitter release at the neuromuscu- Ca2+ can account for the effects of Ca2+ buffers on
lar junction. J Physiol 193:419–432 synaptic facilitation. J Neurophysiol 96:3389–3397
Eggermann E, Bucurenciu I, Goswami SP, Jonas P (2012) Matveev V, Bertram R, Sherman A (2009) Ca2+ current
Nanodomain coupling between Ca(2)(+) channels and versus Ca2+ channel cooperativity of exocytosis.
sensors of exocytosis at fast mammalian synapses. Nat J Neurosci 29:12196–12209
Rev Neurosci 13:7–21 Matveev V, Bertram R, Sherman A (2011) Calcium
Ermolyuk YS, Adler FG, Surges R, Pavlov IY, Timofeeva cooperativity of exocytosis as a measure of Ca(2)+
Y, Kullman DM, Volynski KE (2013) Differential channel domain overlap. Brain Res 1398:126–138
triggering of spontaneous glutamate release by P/Q-, Meinrenken CJ, Borst JG, Sakmann B (2002) Calcium
N- and R-type Ca2+ channels. Nature Neurosci secretion coupling at calyx of held governed by
16:1754–1763 nonuniform channel-vesicle topography. J Neurosci
Felmy F, Neher E, Schneggenburger R (2003) Probing 22:1648–1667
the intracellular calcium sensitivity of transmitter Meinrenken CJ, Borst JG, Sakmann B (2003) Local routes
release during synaptic facilitation. Neuron revisited: the space and time dependence of the Ca2+
37:801–811 signal for phasic transmitter release at the rat calyx of
Gentile L, Stanley EF (2005) A unified model of presyn- Held. J Physiol 547:665–689
aptic release site gating by calcium channel domains. Millar AG, Zucker RS, Ellis-Davies GC, Charlton MP,
Eur J Neurosci 21:278–282 Atwood HL (2005) Calcium sensitivity of neurotrans-
Glavinovic MI, Rabie HR (2001) Monte Carlo evaluation mitter release differs at phasic and tonic synapses.
of quantal analysis in the light of Ca2+ dynamics and J Neurosci 25:3113–3125
the geometry of secretion. Pflugers Arch 443:132–145 Moser T, Neef A, Khimich D (2006) Mechanisms under-
Han X, Wang CT, Bai J, Chapman ER, Jackson MB lying the temporal precision of sound coding at the
(2004) Transmembrane segments of syntaxin line the inner hair cell ribbon synapse. J Physiol 576:55–62
fusion pore of Ca2+-triggered exocytosis. Science Mutch SA, Kensel-Hammes P, Gadd JC, Fujimoto BS,
304:289–292 Allen RW, Schiro PG, Lorenz RM, Kuyper CL, Kuo
Heidelberger R, Heinemann C, Neher E, Matthews JS, Bajjalieh SM, Chiu DT (2011) Protein quantifica-
G (1994) Calcium dependence of the rate of exocytosis tion at the single vesicle level reveals that a subset of
in a synaptic terminal. Nature 371:513–515 synaptic vesicle proteins are trafficked with high pre-
Heil P, Neubauer H (2010) Summing across different cision. J Neurosci 31:1461–1470
active zones can explain the quasi-linear Nadkarni S, Bartol TM, Stevens CF, Sejnowski TJ, Levine
Ca-dependencies of exocytosis by receptor cells. H (2012) Short-term plasticity constrains spatial orga-
Front Synaptic Neurosci 2:148 nization of a hippocampal presynaptic terminal. Proc
Heinemann C, von Ruden L, Chow RH, Neher E (1993) A Natl Acad Sci USA 109:14657–14662
two-step model of secretion control in neuroendocrine Neher E (2012) Introduction: regulated exocytosis. Cell
cells. Pflugers Arch 424:105–112 Calcium 52:196–198
Neher E, Sakaba T (2008) Multiple roles of calcium ions of spontaneous and asynchronous neurotransmitter
in the regulation of neurotransmitter release. Neuron release. Cell Calcium 52:226–233
59:861–872 Sorensen JB (2004) Formation, stabilisation and fusion of
Nouvian R, Neef J, Bulankina AV, Reisinger E, the readily releasable pool of secretory vesicles.
Pangrsic T, Frank T, Sikorra S, Brose N, Binz T, Pflugers Arch 448:347–362
Moser T (2011) Exocytosis at the hair cell ribbon Stanley EF (1997) The calcium channel and the organiza-
synapse apparently operates without neuronal tion of the presynaptic transmitter release face. Trends
SNARE proteins. Nat Neurosci 14:411–413 Neurosci 20:404–409
Oheim M, Kirchhoff F, Stuhmer W (2006) Calcium Sterling P, Matthews G (2005) Structure and function of
microdomains in regulated exocytosis. Cell Calcium ribbon synapses. Trends Neurosci 28:20–29
40:423–439 Stevens CF, Wesseling JF (1998) Activity-dependent
Pan B, Zucker RS (2009) A general model of synaptic modulation of the rate at which synaptic vesicles
transmission and short-term plasticity. Neuron become available to undergo exocytosis. Neuron
62:539–554 21:415–424
Pangrsic T, Reisinger E, Moser T (2012) Otoferlin: a Sun J, Pang ZP, Qin D, Fahim AT, Adachi R,
multi-C2 domain protein essential for hearing. Trends Sudhof TC (2007) A dual-Ca2+-sensor model for neu-
Neurosci 35:671–680 rotransmitter release in a central synapse. Nature
Pedersen MG, Sherman A (2009) Newcomer insulin 450:676–682
secretory granules as a highly calcium-sensitive pool. Thoreson WB, Rabl K, Townes-Anderson E, Heidelberger
Proc Natl Acad Sci USA 106:7432–7436 R (2004) A highly Ca2+-sensitive pool of vesicles
Quastel DM, Guan YY, Saint DA (1992) The relation contributes to linearity at the rod photoreceptor ribbon
between transmitter release and Ca2+ entry at the synapse. Neuron 42:595–605
mouse motor nerve terminal: role of stochastic factors Verhage M, Toonen RF (2007) Regulated exocytosis:
causing heterogeneity. Neuroscience 51:657–671 merging ideas on fusing membranes. Curr Opin Cell
Raingo J, Khvotchev M, Liu P, Darios F, Li YC, Ramirez Biol 19:402–408
DM, Adachi M, Lemieux P, Toth K, Davletov B, Voets T (2000) Dissection of three Ca2+-dependent steps
Kavalali ET (2012) VAMP4 directs synaptic vesicles leading to secretion in chromaffin cells from mouse
to a pool that selectively maintains asynchronous neu- adrenal slices. Neuron 28:537–545
rotransmission. Nat Neurosci 15:738–745 Voets T, Neher E, Moser T (1999) Mechanisms underly-
Roux I, Safieddine S, Nouvian R, Grati M, Simmler MC, ing phasic and sustained secretion in chromaffin cells
Bahloul A, Perfettini I, Le Gall M, Rostaing P, from mouse adrenal slices. Neuron 23:607–615
Hamard G, Triller A, Avan P, Moser T, Petit von Ruden L, Neher E (1993) A Ca-dependent early step
C (2006) Otoferlin, defective in a human deafness in the release of catecholamines from adrenal chro-
form, is essential for exocytosis at the auditory ribbon maffin cells. Science 262:1061–1065
synapse. Cell 127:277–289 Wadel K, Neher E, Sakaba T (2007) The coupling
Sakaba T (2008) Two Ca(2+)-dependent steps controlling between synaptic vesicles and Ca2+ channels deter-
synaptic vesicle fusion and replenishment at the cere- mines fast neurotransmitter release. Neuron
bellar basket cell terminal. Neuron 57:406–419 53:563–575
Schmidt H, Brachtendorf S, Arendt O, Hallermann S, Wang LY, Kaczmarek LK (1998) High-frequency firing
Ishiyama S, Bornschein G, Gall D, Schiffmann SN, helps replenish the readily releasable pool of synaptic
Heckmann M, Eilers J (2013) Nanodomain coupling at vesicles. Nature 394:384–388
an excitatory cortical synapse. Curr Biol 23:244–249 Weiss JN (1997) The Hill equation revisited: uses and
Schneggenburger R, Neher E (2000) Intracellular calcium misuses. FASEB J 11:835–841
dependence of transmitter release rates at a fast central Wolfel M, Schneggenburger R (2003) Presynaptic capac-
synapse. Nature 406:889–893 itance measurements and Ca2+ uncaging reveal
Scimemi A, Diamond JS (2012) The number and organi- submillisecond exocytosis kinetics and characterize
zation of Ca2+ channels in the active zone shapes the Ca2+ sensitivity of vesicle pool depletion at a fast
neurotransmitter release from Schaffer collateral syn- CNS synapse. J Neurosci 23:7059–7068
apses. J Neurosci 32:18157–18176 Wolfel M, Lou X, Schneggenburger R (2007)
Shahrezaei V, Delaney KR (2005) Brevity of the Ca2+ A mechanism intrinsic to the vesicle fusion machinery
microdomain and active zone geometry prevent Ca2 determines fast and slow transmitter release at a large
+
-sensor saturation for neurotransmitter release. CNS synapse. J Neurosci 27:3198–3210
J Neurophysiol 94:1912–1919 Worden MK, Bykhovskaia M, Hackett JT (1997) Facilita-
Shahrezaei V, Cao A, Delaney KR (2006) Ca2+ from one tion at the lobster neuromuscular junction: a stimulus-
or two channels controls fusion of a single vesicle at dependent mobilization model. J Neurophysiol
the frog neuromuscular junction. J Neurosci 78:417–428
26:13240–13249 Wu MM, Llobet A, Lagnado L (2009) Loose coupling
Smith SM, Chen W, Vyleta NP, Williams C, Lee CH, between calcium channels and sites of exocytosis in
Phillips C, Andresen MC (2012) Calcium regulation chromaffin cells. J Physiol 587:5377–5391
Calcium-Dependent Potassium Channels 549 C
Yamada MW, Zucker RS (1992) Time course of transmit- a complex function of both membrane potential
ter release calculated from stimulations of a calcium and intracellular calcium concentration (Fig. 1).
diffusion model. Biophys J 61:671–682
Yang Y, Gillis KD (2004) A highly Ca2+-sensitive pool of Both SK and BK channels are composed of four
granules is regulated by glucose and protein kinases in subunits. Intermediate-conductance (IK) channels,
insulin-secreting INS-1 cells. J Gen Physiol which are not expressed in neurons, are not consid-
124:641–651 ered further here.
Yang Y, Udayasankar S, Dunning J, Chen P, Gillis KD
(2002) A highly Ca2+-sensitive pool of vesicles is
C
regulated by protein kinase C in adrenal chromaffin
cells. Proc Natl Acad Sci USA 99:17060–17065 Detailed Description
Yao J, Gaffaney JD, Kwon SE, Chapman ER (2011) Doc2
is a Ca2+ sensor required for asynchronous neurotrans-
mitter release. Cell 147:666–677 Gating Properties
Zucker RS, Fogelson AL (1986) Relationship between SK channels activate at relatively low calcium
transmitter release and presynaptic calcium influx concentrations (100 nM to 1 mM, Fig. 1, left).
when calcium enters through discrete channels. Proc Their opening is regulated by the interaction
Natl Acad Sci USA 83:3032–3036
Zucker RS, Regehr WG (2002) Short-term synaptic plas- between calcium ions and calcium-binding pro-
ticity. Annu Rev Physiol 64:355–405 teins (calmodulins), which are constitutively
associated with each subunit. Main sources of
calcium include voltage-dependent calcium
(CaV) channels, NMDA receptors, and various
Calcium-Dependent Potassium types of intracellular stores. The time constant of
Channels SK channel activation (5–15 ms) is much faster
than the calcium oscillations involved in pace-
Guillaume Drion1,2, Fabian Philippart2, making of most neurons. These channels can
Rodolphe Sepulchre3 and Vincent Seutin2 therefore be considered as rapid “calcium fol-
1
Department of Electrical Engineering and lowers.” As a consequence, their intrinsic kinet-
Computer Science, University of Liège, Liège, ics is generally modeled as instantaneous.
Belgium BK channels are solely active during the
2
Laboratory of Pharmacology and GIGA repolarizing phase of the action potential,
Neurosciences, University of Liège, Liège, because they need both a rather strong depolari-
Belgium zation and a large calcium concentration to acti-
3
Department of Engineering, University of vate (Fig. 1, right). In addition, they deactivate
Cambridge, Cambridge, UK very quickly. Therefore, they modulate the
action potential duration and are able to generate
a fast (few milliseconds) afterhyperpolarization.
Synonyms Because of their large hyperpolarizing effect,
they are able to modulate the availability of
BK channels (KCa1.1); SK channels (KCa2.x) other voltage-dependent channels. They have no
significant role during most of the interspike
interval, contrary to SK channels.
Definition
Effect on Intrinsic Neuronal Excitability
Calcium-dependent potassium channels are In pacemaker neurons, SK channel activation
potassium-permeable channels whose gating is reg- promotes regular interspike intervals (ISIs). In
ulated by the intracellular calcium concentration. many cases, application of SK blockers has
Small-conductance (10–20 pS) calcium-activated been shown to deregularize firing and/or induce
(SK) channels are purely gated by calcium, while bursting behavior, both in vitro and in vivo (see,
the opening probability of large-conductance e.g., Feng and Jaeger 2008 and Waroux
(> 100 pS) calcium-activated (BK) channels is et al. 2005) (Fig. 2). In these neurons, SK channels
C 550 Calcium-Dependent Potassium Channels
Calcium-Dependent Potassium Channels, voltage-dependent activation of BK channels (Modified

Fig. 1 Left: Calcium-dependent activation of SK chan- from Blaz and Magleby 1986)
nels (Modified from Stocker 2004). Right: Calcium- and
Calcium-Dependent Potassium Channels, Fig. 2 Effect of the SK channel blocker apamin on intrinsic dopamine
neuron excitability (intracellular recording in a rat brain slice by J. Scuvée-Moreau)
Calcium-Dependent Potassium Channels 551 C
produce a large afterhyperpolarization (AHP), Postsynaptically, SK channels have been
whose duration varies depending on the kinetics found on the soma and dendrites of many neu-
of the calcium oscillations. Because the conduc- rons. Three roles have been proposed for them:
tance involved is rather large, the neuron is very first, they may be in close contact with NMDA
insensitive to incoming noise during this event, receptors on spines of pyramidal neurons. In this
preventing the generation of premature action case, they may produce a negative feedback in the
potentials. sense that their activation by calcium influx C
The importance of the regularity promoting through NMDA channels will provide a local
effect of SK channels has been emphasized in hyperpolarizing drive that will inhibit further
the case of cerebellar Purkinje neurons. Indeed, NMDA receptor activation. Thus, their activation
a loss of precision of pacemaking in these cells is will decrease the likelihood of NMDA-dependent
observed in a model of a disease called episodic long-term potentiation (Adelman et al. 2011).
ataxia type 2 (Alviña and Khodakhah 2010). In Second, SK channels have also been reported to
the mouse model of the disease, a loss of function contribute to dendritic integration in pyramidal-
is observed in the CaV channels which activate type neurons. More precisely, they seem to con-
SK channels of Purkinje neurons. Interestingly, tribute to the repolarization of plateau potentials
regularizing the firing rate through the use of induced by glutamate. Third, they have been
positive modulators of SK channels produces shown to underlie a slow inhibitory postsynaptic
beneficial effects. potential following activation of Gq-coupled
In pyramidal neurons, SK channels have been metabotropic receptors, synthesis of inositol tri-
shown to at least contribute to the medium dura- phosphate, and ensuing release of intracellular
tion AHP and thereby to underlie the phenome- calcium.
non known as spike frequency adaptation (Hille BK channels are present in presynaptic termi-
2001). In other types of neurons, it has been nals of several types of neurons. By modulating
shown that SK channels contribute to the termi- the action potential width, they may have a major
nation of bursts, such as in Aplysia R15 neurons effect on the amount of released
(Plant and Kim 1976). neurotransmitter.
BK channels operate during and immediately
after the action potential. They have been shown Acknowledgements This work was supported by grants
to control action potential width (Bean 2007) and n 3.4533.09 and T. 0015.13 from the F.R.S.-FNRS (VS)
and by a grant from the Belgian Science Policy (IAP P7/
therefore to promote the ability of high-frequency
10) (VS).
firing. In particular, they prevent incremental
inactivation of voltage-gated sodium channels.
This action is also performed by other potassium References
channels (e.g., Kv3) in some neurons. As pointed
out by Bean, the strong repolarizing effect of BK Adelman JP, Maylie J, Sah P (2011) Small-conductance
channels explains the counterintuitive observa- Ca2 + activated K + channels: form and function.
Annu Rev Physiol 74:245–269
tion of a broadening of action potentials during Alviña K, Khodakhah K (2010) KCa channels as thera-
blockade of some CaV channels. In addition, peutic targets in episodic ataxia type-2. J Neurosci
their progressive inactivation may contribute to 30:7249–7257
frequency-dependent broadening of action Bean BP (2007) The action potential in mammalian cen-
tral neurons. Nat Rev Neurosci 8:451–465
potentials. Blatz AL, Magleby KL (1986) Calcium-activated potas-
sium channels. Trends Neurosci 10:463–467
Effect on Synaptic Transmission Feng SS, Jaeger D (2008) The role of SK calcium-
Some studies have found evidence for the pres- dependent potassium currents in regulating the activity
of deep cerebellar nucleus neurons: a dynamic clamp
ence of some SK channel subtypes in axon ter- study. Cerebellum 7:542–546
minals, but there is scarce information on their Hille B (2001) Ion channels of excitable membranes.
physiological roles at this level. Sinauer Associates, Sunderland
C 552 Calmodulin, Models of
Plant RE, Kim M (1976) Mathematical description of

a bursting pacemaker neuron by a modification of the
Hodgkin-Huxley equations. Biophys J 16:227–244
Stocker M (2004) Ca(2+)-activated K + channels: molec-
ular determinants and function of the SK family. Nat
Waroux O, Massotte L, Alleva L, Graulich A, Thomas E,
Liégeois JF, Scuvée-Moreau J, Seutin V (2005) SK
channels control the firing pattern of midbrain dopami-
nergic neurons in vivo. Eur J Neurosci 22:3111–3121
Calmodulin, Models of
M. Neal Waxham1 and Margaret S. Cheung2

1
Department of Neurobiology and Anatomy, The
University of Texas Health Science Center at
Houston, Houston, TX, USA
2
Department of Physics, University of Houston,
Houston, TX, USA
Definition
Calmodulin (CaM) is a small (148 amino acid)

ubiquitous protein responsible for decoding
oscillatory changes in intracellular calcium and
transducing the signal to downstream targets.
Four separate calcium (Ca2+)-binding sites are Calmodulin, Models of, Fig. 1 Structures of CaM in the
absence (a and c) and presence (b and d) of Ca2+. The N-
distributed into two globular domains (the and C-terminus of CaM are colored in yellow and red,
N- and C-lobes) tethered together by a linker respectively, and Ca2+ ions are represented by green
whose structural flexibility is essential for interac- spheres. Ribbon representations of apoCaM and holoCaM
tions with its multiple target proteins. So, despite are shown in (a) and (b) and the Ca side-chain represen-
tations of the same structures are shown in (c) and (d)
its small size, there are significant complexities
in building quantitative mathematical models to
accommodate the multitude of different structural Each globular domain contains two EF hand
(liganded) states that ultimately dictate CaM’s Ca2+-binding sites, and the affinity of each site
functions. Detailed molecular modeling is for Ca2+ and their binding kinetics is fairly well
beginning to advance our understanding for how characterized (Linse et al. 1991; Bayley et al.
this single protein can interact with several unique 1996; Gaertner et al. 2004; Putkey et al. 2008;
target proteins in both Ca2+-bound (holoCaM) and Faas et al. 2011). The N-domain sites are labeled
Ca2+-free (apoCaM) states. 1 and 2 and those in the C-domain are labeled 3
and 4 (Fig. 1), and the experimentally determined
macroscopic binding constants can be found in
Detailed Description Table 1. Ca2+ binding between site 1 and site 2
and between site 3 and site 4 exhibits
CaM is typically described as a dumbbell-shaped cooperativity. Under steady-state conditions, the
protein with two globular domains connected C-lobe has a ~5- to 6-fold greater average binding
through a central helix (Crivici and Ikura 1995). affinity for Ca2+ (~1.7 mM) than the N-lobe
Calmodulin, Models of 553 C
Calmodulin, Models of, Table 1 Ca2+binding to CaM interact preferentially with apoCaM, while others
Site Kd (mM) Koff (s1) Kon (mM1s1) prefer holoCaM. In both cases, the mutual inter-
1 40 20,000 500 actions of specific amino acid residues in CaM
2 2.5 3,000 1,200 and the target molecule dictate target selectivity.
3 12.6 5,100 408 From the many CaM/target protein structures
4 0.25 9.1 36.4 available, a few generalizations about binding
can be made. The highly malleable structure of C
CaM is critical, where a helical linker connects
(~10 mM), indicating that the C-lobe is preferen- the two lobes with notable structural flexibility
tially occupied under steady-state conditions. (Crivici and Ikura 1995; Hoeflich and Ikura 2002;
However, since cellular Ca2+ levels are con- Bhattacharya et al. 2004; Slaughter et al. 2005;
stantly fluctuating and under these rapidly vary- Homouz et al. 2009; Price et al. 2010). Unwind-
ing Ca2+ conditions, the kinetics of association ing of the central helix permits the N- and C-lobes
and dissociation dictate the probability of N-lobe to adjust their orientation and relative distance to
or C-lobe saturation. The N-lobe binds Ca2+ each other to accommodate unique structural fea-
quickly and releases it quickly, while the C-lobe tures in the target molecules.
binds and releases Ca2+ more slowly. In this way, Computer simulations allow the investigation
the N-lobe has the capacity to respond in of CaM dynamics involving inter- and intra-
a switch-like manner to brief Ca2+ transients, domain movement in solution that is a
while the C-lobe kinetics lead to prerequisite to model the intricacies of target
slower activation and longer duration occupancy, interactions. Molecular dynamics simulations
loosely analogous to a capacitor. As a are typically employed to study protein dynamics
consequence, Ca2+ binding to the C-lobe of at an atomic detail but are limited to time scales
CaM is the rate-limiting step for full occupancy shorter than many biologically relevant confor-
when starting from the apoCaM state. From the mational changes and by the size (number of
intracellular perspective, both the level of resting atoms that can be simulated) of the protein of
Ca2+ and the shape, amplitude, and frequency of interest (Shaw et al. 2010). Therefore, it is com-
Ca2+ oscillations will influence the overall Ca2+ putationally intractable to explore the extensive
occupancy of the different lobes of CaM. Differ- conformational fluctuations CaM undergoes to
ent modeling strategies have been developed to interact with target molecules by using all-
address the complexities of the four separate atomistic molecular dynamics simulations. An
Ca2+-binding sites on CaM (Holmes 2000; Keller alternative strategy is to employ a physics-based
et al. 2008; Saucerman and Bers 2008; Stefan approach that utilizes low-resolution protein
et al. 2008; Kubota and Waxham 2010; Faas models to simulate the motion of CaM (and its
et al. 2011). Minimally these models incorporate targets). In this approach, each amino acid
the different Kd and on and off rates for Ca2+ (except glycine) is represented by two beads:
binding to the N- and C-lobes of CaM. Coupling a Ca bead used to capture the trace of the protein
factors have been incorporated to accommodate backbone and a second bead that represents the
the known cooperativity of Ca2+ binding into size and chemical nature of the amino acid side
some models, while others employed the strategy chain (See Fig 1.). The van der Waals interaction
to model each site within each domain indepen- between a pair of side-chain beads follows the
dently with a specific on and off rate for Ca2+ Lennard-Jones potential, and their strength of
binding. interaction depends on the amino acid sequence
In addition to Ca2+ binding, CaM can interact (chemical environment). The interaction between
with hundreds of target proteins to regulate Ca beads captures the hydrogen bonding between
numerous biochemical, metabolic, and cellular the two intervening backbones. The potential
processes (Yamniuk and Vogel 2004; Choi and for the backbone retains partial information
Husain 2006). Some of these target proteins from the experimentally determined structures.
C 554 Calmodulin, Models of
The Brownian motion of individual atoms within target molecules that favor the extended confor-
the model is described by the Langevin equations mation, a determinant in the process of ultimately
of motion where the effect of solvent is implicitly dictating target selectivity.
represented through friction and agitation: Because of the computational efficiency of
this strategy, it is also possible to address issues
d 2 X ðt Þ dXðtÞ such as how the effect of excluded volume of
m ¼ ∇U ðXÞ g þ RðtÞ: surrounding macromolecules (Minton 1981)
dt2 dt
would impact the distribution of CaM conforma-
In this formula, X is the position of the beads at tions. As one example, incorporating Ficoll 70,
time t, m denotes the mass of the bead, U(X) is the an inert, nonpolar polymer, into the simulation
potential of the system, g is the friction coeffi- environment can simulate macromolecular
cient, and R(t) is the random force caused by the crowding. In this example, a simple hard-sphere
agitation of solvent at a certain temperature. The model of Ficoll 70 (impenetrable hard spheres
trajectories can be computed by the integration of with a radius of 55 Å) is utilized to reduce the
the equations of motion. This coarse-grained computational demand while efficiently retaining
molecular simulation can be employed to effi- its excluded volume effect. Coarse-grained
ciently explore a wide range of structures in solu- molecular simulations showed that, at levels of
tion while still preserving information of protein crowding comparable to those predicted inside
conformation pertinent to a global minimum on cells, apoCaM favors a more compact conforma-
the folding energy landscape. This level of detail tion and as a consequence would favor binding to
is indispensable to capture the conformational targets where CaM must adopt a more compact
fluctuations employed by CaM as it searches for form (Homouz et al. 2009).
binding to individual targets. Based on a combined approach of coarse-
The modeling strategy is further expanded grained molecular simulations, energy landscape
with quantum chemistry calculations to capture theory, and experiments, it will be possible to
the charge distribution of residues in the Ca2+- determine the number and probability distribu-
bound state of CaM (holoCaM) (Wang et al. tion of unique conformations that will dictate
2011). Additionally, computations can be made how CaM chooses to distribute among the many
of the electrostatic interactions between charged possible intracellular targets. Ultimately, multi-
residues modulated by the ionic strength in solu- scale computational models will be needed that
tion. The electrostatic interaction Vij is can combine the oscillatory nature of the Ca2+
represented by a Debye-H€uckel (Debye and fluctuations and subsequent binding to CaM with
H€uckel 1923) term in the potential of the system: the molecular level simulations of CaM and tar-
get protein structural fluctuations.
Zi Zj r=pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
er eo kB T=2e2 I
ffi
V ij ¼ e :
4preo er
References
Zi (Zj) is the partial charge on each bead i (j), r
Bayley PM, Findlay WA et al (1996) Target recognition
is the separation between beads i and j, eo is the by calmodulin: dissecting the kinetics and affinity of
permittivity of free space, er is the relative dielec- interaction using short peptide sequences. Protein Sci
tric constant, kB is the Boltzmann constant, T is 5(7):1215–1228
Bhattacharya S, Bunick CG et al (2004) Target selectivity
the temperature, e is the charge of an electron,
in EF-hand calcium binding proteins. Biochim
and I is the ionic strength of the aqueous solution. Biophys Acta 1742:69–79
In employing such simulations, it was shown that Choi J, Husain M (2006) Calmodulin-mediated cell cycle
at higher ionic strength electrostatic repulsions regulation: new mechanisms for old observations. Cell
Cycle 5(19):2183–2186
are screened which results in stability increases
Crivici A, Ikura M (1995) Molecular and structural basis
and a more extended conformation of CaM. In of target recognition by calmodulin. Annu Rev
this situation, CaM would more likely bind to Biophys Biomol Struct 24:85–116
Capacitance, Membrane 555 C
Debye P, H€uckel E (1923) The theory of electrolytes. I.
Lowering of freezing point and related phenomena. Capacitance, Membrane
Physikalische Zeitschrift 24:185–206
Faas GC, Raghavachari S et al (2011) Calmodulin as
a direct detector of Ca2+ signals. Nat Neurosci Jorge Golowasch1 and Farzan Nadim1,2
1
14(3):301–304 Federated Department of Biological Sciences,
Gaertner TR, Putkey JA et al (2004) RC3/Neurogranin New Jersey Institute of Technology/Rutgers
and Ca2+/calmodulin-dependent protein kinase II pro-
duce opposing effects on the affinity of calmodulin for University, Newark, NJ, USA C
2
calcium. J Biol Chem 279(38):39374–39382 Department of Mathematical Sciences, New
Hoeflich KP, Ikura M (2002) Calmodulin in action: diver- Jersey Institute of Technology, Newark, NJ, USA
sity in target recognition and activation mechanisms.
Cell 108:739–742
Holmes WR (2000) Models of calmodulin trapping and
CaM kinase II activation in a dendritic spine. J Comput Definition
Homouz D, Sanabria H et al (2009) Modulation of cal- The membrane capacitance results from the fact
modulin plasticity by the effect of macromolecular
crowding. J Mol Biol 391:933–943 that the plasma membrane acts as a capacitor: the
Keller DX, Franks KM et al (2008) Calmodulin activation phospholipid bilayer is a thin insulator separating
by calcium transients in the postsynaptic density of two electrolytic media, the extracellular space
dendritic spines. PLoS One 3(4):e2045 and the cytoplasm. The membrane capacitance
Kubota Y, Waxham MN (2010) Lobe specific
Ca2+calmodulin nano-domain in neuronal spines: is proportional to the cell surface area and,
a single molecule level analysis. PLoS Comput Biol together with the membrane resistance, deter-
6(11):e1000987 mines the membrane time constant which dic-
Linse S, Helmersson A et al (1991) Calcium binding to tates how fast the cell membrane potential
calmodulin and its globular domains. J Biol Chem
266(13):8050–8054 responds to the flow of ion channel currents.
Minton A (1981) Excluded volume as a determinant of
macromolecular structure and reactivity. Biopolymers
20:2093–2120
Price ES, DeVore MS et al (2010) Detecting intramolec-
ular dynamics and multiple forster resonance energy Membrane capacitance is the electrical capaci-
transfer states by fluorescence correlation spectros- tance associated with a biological membrane,
copy. J Phys Chem B 114:5895–5902
Putkey JA, Waxham MN et al (2008) Acidic/IQ motif
expressed in units of Farads (F). The electrical
regulator of calmodulin. J Biol Chem capacitance of a biological membrane results
283(3):1401–1410 from the membrane composition of a bilayer of
Saucerman JJ, Bers DM (2008) Calmodulin mediates dif- mostly phospholipids that form an insulating
ferential sensitivity of CaMKII and calcineurin to local
Ca2+ in cardiac myocytes. Biophys
matrix to which proteins are attached or embed-
J 95(10):4597–4612 ded. The total membrane capacitance cm of a cell
Shaw DE, Maragakis P et al (2010) Atomic-level charac- is a quantity directly proportional to the mem-
terization of the structural dynamics of proteins. Sci- brane surface area and the dielectric properties of
ence 330:341–346
Slaughter BD, Bieber-Urbauer RJ et al (2005) Single-
the membrane, provided that the thickness of the
molecule tracking of sub-millisecond domain motion membrane is constant. Thus, the total membrane
in calmodulin. J Phys Chem B 109:12658–12662 capacitance is cm = CmA, where Cm is the specific
Stefan MI, Edelstein SJ et al (2008) An allosteric model membrane capacitance (typically expressed in
of calmodulin explains differential activation of
PP2B and CaMKII. Proc Natl Acad Sci U S A
units of mF/cm2) and A is the area.
105(31):10768–10773 Because of its linear relationship with cell
Wang Q, Liang KC et al (2011) The effect of macromo- surface area, capacitance is often measured
lecular crowding, ionic strength and calcium binding experimentally as a way of determining cell
on calmodulin dynamics. Plos Comput Biol 7(7)
Yamniuk AP, Vogel HJ (2004) Calmodulin’s flexibility
area and thus cell size if the cell geometry is
allows for promiscuity in its interaction with target simple, for example, spherical or cylindrical
proteins and peptides. Mol Biotechnol 27:33–57 (Streit and Lux 1987), or changes in membrane
C 556 Capacitance, Membrane
surface area (Neher and Marty 1982). These with a resistor (an RC circuit), the current
changes can be global or local, for example, response generated includes a step whose
when cells grow dendritic branches or add or amplitude is equal to the capacitive current Ic
remove membrane due to endocytosis or exocy- (Fig. 1a, red), followed by a ramp proportional
tosis at a synapse. Cell capacitance may also to the conductance (1/R) of the circuit
change due to local membrane thickness varia- (Fig. 1a). From eq. 2, c can then be calculated.
tions at lipid rafts, which are 10 % thicker than 2. Voltage step method. If a brief voltage step is
non-raft membranes (Alberts et al. 2008). applied, an initial current transient will be
Because membrane capacitance determines the generated which is primarily the current that
time constant of a neuron (tm = rmcm), it plays an charges the capacitor (Fig. 1b). By integrating
important role in the integration of the electrical the area
Ð under the curve of this transient cur-
inputs a neuron receives. It also determines the rent ( Ic dt, Fig. 1b, red), the charge Q can be
propagation velocity of action potentials which is measured and then used to determine c from
inversely proportional to cm (Matsumoto and Eq. 1.
Tasaki 1977). Finally, cell capacitance is 3. Voltage sine wave method. If a sinusoidal
a crucial parameter in detailed conductance- voltage waveform is applied to an RC circuit,
based computational models, which are widely the current response will also be sinusoidal but
used to understand the activity and output of com- shifted relative to the voltage input by
plex neuronal systems (Koch 1999). a constant time (Dt, Fig. 1c, red). The resulting
To determine the exact surface area of a cell, phase change is related to the capacitance of
a precise value of the specific membrane capaci- the circuit (Neef et al. 2007). Lock-in ampli-
tance is required. A number of different methods fiers in combination with voltage clamp
can be used to determine the exact capacitance devices generate outputs that are proportional
value, but in such calculations the fact that cell to both impedance and phase, and thus capac-
membrane surface – particularly in neurons – is itance can be determined (Neher and Marty
rarely restricted to simple shapes should always 1982; Neef et al. 2007).
be considered. In fact, the area of neurons is All voltage clamp methods give accurate
almost always distributed over surfaces that are measurements of capacitance in cells that are
intricate and complex, which affects the accuracy electrically compact (isopotential) and there-
of capacitance measurements. fore can be described as single-compartment
RC circuits. This is the case for many cell
Measurement of Membrane Capacitance types, but is almost never true for neurons
A capacitor accumulates charge proportional to whose extensive processes (dendrites and
the voltage across it: axon) render them non-isopotential. For
non-isopotential cells all these methods are
Q ¼ c DV (1)
inaccurate. Intuitively, the reason for this inac-
The rate of change of the capacitive charge is curacy is that, at locations distal from the point
the capacitive current: of current injection (i.e., the site of electrode
impalement), the membrane voltage decays
dQ dV
IC ¼ ¼c (2) along the processes relative to the command
dt dt voltage. The inability to maintain a constant
From this basic concept, several time-domain voltage throughout the cell is referred to as the
methods have been derived to measure the capac- space clamp problem.
itance of a neuron, which essentially measure IC As an example, for a simple cell that can be
under voltage clamp conditions. represented by two compartments, one near
1. Voltage ramp method. If a voltage ramp with and one far from the point of current injection,
slope dV/dt is applied to a capacitor in parallel connected by a resistor, ra, the total
Capacitance, Membrane 557 C
a Voltage Ramp c Voltage sine wave
Vm
Vm
Ic Im
Im
Ic
Δt
2Ic C
b Voltage step d Current step
Vm Iext
Vm
Im
time
time
Capacitance, Membrane, Fig. 1 Capacitance mea- with (a) and (d) typically being slower (seconds) than
surement protocols. Top traces show stimulation wave- (b) and (c) (tens to hundreds of milliseconds). Red indi-
form. (a–c), voltage clamp protocols. (d), current clamp cates the response element relevant for the measurement
protocol. Time scales for the different protocols differ, of capacitance
capacitance measured with voltage clamp several distal compartments (Golowasch

methods can be approximated by et al. 2009). The principle behind this method
is that during a current pulse currents rapidly
1 flow axially along the different cable-like
cm ¼ cn þ cf 2 (3)
1 þ r a =r f structures of a cell and “equalize” the voltage.
These currents generate voltage changes that
where cn is the capacitance of the compart- can be described by multiple exponential
ment near the point of current injection and cf terms with short time constants (Rall 1977;
and rf are the distal compartment capacitance Holmes et al. 1992). Meanwhile, the mem-
and membrane resistance, respectively brane is slowly charged homogeneously over
(Golowasch et al. 2009). Thus, the farther the the entire surface of the cell, a process that can
distal compartment, and the lower its mem- be characterized by an additional exponential
brane resistance, the bigger the error in total change of the membrane potential, which has
capacitance. For cells with more complex the slowest time constant of all the exponential
morphological structure than two connected terms. Thus, the membrane potential Vm can
compartments, the measured capacitance be described as a sum of exponential terms
deviates from the approximation given by plus the resting potential Vrest:
Eq. 3. An estimate of the capacitance
measured in voltage clamp in cells of more X
1

complex morphology is provided by V m ðtÞ ¼ V rest þ V i 1 et=ti (4)
i¼0
Taylor (2012).
4. Current clamp method. A much more accurate
method for measuring capacitance is current By dividing Eq. 4 through by the external
clamp pulses. It is appropriate even for cells current Iext, one obtains a series of resistive
that have their capacitance distributed over terms Ri = Vi/Iext. The time constant of the
C 558 Capacitance, Membrane
slowest exponential term t0 is equal to the References

product tm = rmcm, and rm can be determined
as the resistive component of the slowest Alberts B, Wilson JH, Hunt T (2008) Molecular biology of
the cell, 5th edn. Garland Science, New York
exponential term (Ro), and cm = tm/rm
Gentet LJ, Stuart GJ, Clements JD (2000) Direct measure-
(Golowasch et al. 2009). ment of specific membrane capacitance in neurons.
All these methods rely on manipulating volt- Golowasch J, Thomas G, Taylor AL, Patel A, Pineda A,
Khalil C, Nadim F (2009) Membrane capacitance
age in a range where only passive properties of
measurements revisited: dependence of capacitance
the cell are involved. Any activation of voltage- value on measurement method in nonisopotential neu-
gated currents will introduce significant errors in rons. J Neurophysiol 102:2161–2175
the capacitance estimate (Golowasch et al. 2009; Hodgkin AL, Huxley AF (1952) A quantitative descrip-
tion of membrane current and its application to
White and Hooper 2013).
conduction and excitation in nerve. J Physiol
117:500–544
Specific Membrane Capacitance Hodgkin AL, Huxley AF, Katz B (1952) Measurement of
The value most used is 1 mF/cm2, which most current-voltage relations in the membrane of the giant
axon of Loligo. J Physiol 116:424–448
likely stems from the first measurements ever
Holmes WR, Segev I, Rall W (1992) Interpretation of time
made in the famous series of experiments of constant and electrotonic length estimates in
Hodgkin, Huxley, and Katz published in 1952 multicylinder or branched neuronal structures.
(Hodgkin and Huxley 1952; Hodgkin J Neurophysiol 68:1401–1420
et al. 1952). They in fact measured values over
processing in single neurons. Oxford University Press,
a range of 0.8 and 1.5 mF/cm2, with a mean of New York
0.9 mF/cm2 (Hodgkin et al. 1952), but then used Matsumoto G, Tasaki I (1977) A study of conduction
1 mF/cm2 to simplify the complex calculation velocity in nonmyelinated nerve fibers. Biophys J
20:1–13
used for their final and seminal paper (Hodgkin
Neef A, Heinemann C, Moser T (2007) Measurements of
and Huxley 1952). Later on it was shown that membrane patch capacitance using a software-based
most cells do not have a smooth membrane whose lock-in system. Pflugers Arch 454:335–344
area can be measured optically by making Neher E, Marty A (1982) Discrete changes of cell mem-
brane capacitance observed under conditions of
a simplifying geometric assumption (e.g.,
enhanced secretion in bovine adrenal chromaffin
a perfect cylinder or a perfect sphere). Instead, cells. Proc Natl Acad Sci USA 79:6712–6716
most cell membranes have a certain degree of Niles WD, Levis RA, Cohen FS (1988) Planar
rugosity (lack of smoothness), which could lead bilayer membranes made from phospholipid mono-
layers form by a thinning process. Biophys
to an overestimate of the specific membrane
J 53:327–335
capacitance value. Using an inflation method in Rall W (1977) Core conductor theory and cable properties
the whole-cell patch clamp configuration, of neurons. In: Kandel ER (ed) Handbook of physiol-
Solsona and collaborators (Solsona et al. 1998) ogy (Section 1, The nervous system I, cellular biology
of neurons). American Physiological Society,
determined the specific membrane capacitance
Bethesda, pp 39–97
value to be closer to 0.5 mF/cm2. Furthermore, it Solsona C, Innocenti B, Fernandez JM (1998) Regulation
has been established that cell capacitance does of exocytotic fusion by cell inflation. Biophys J
not significantly vary as a function of ion channel 74:1061–1073
Streit J, Lux HD (1987) Voltage dependent calcium
density (Gentet et al. 2000), and the value of currents in PC12 growth cones and cells
0.5 mF/cm2 obtained from inflated cells is close during NGF-induced cell growth. Pflugers Arch
to that measured in artificial lipid bilayers 408:634–641
although this value also depends on lipid compo- Taylor AL (2012) What we talk about when we talk about
capacitance measured with the voltage-clamp step
sition (Niles et al. 1988).
method. J Comput Neurosci 32:167–175
White WE, Hooper SL (2013) Contamination of
Acknowledgements This work was supported in part by current-clamp measurement of neuron capacitance by
NIH grants MH064711 and MH060605 and NSF grant voltage-dependent phenomena. J Neurophysiol 110
DMS 1122291. (1):257–68
Cardiac Excitable Tissue Pathology (Ion Channels) 559 C
notch found in some ventricular cell types).
Cardiac Excitable Tissue Pathology Phase 2 is marked by a phase of high resistance
(Ion Channels) corresponding to the action potential plateau that
is maintained by inward T-type and L-type Ca2+
Jonathan D. Moreno1,2 and Colleen E. Clancy3 currents and outward delayed rectifier K+ cur-
1
Tri-Institutional MD-PhD Program, Weill rents. This critical phase permits the phenomenon
Cornell Medical College, The Rockefeller of Ca2+-induced Ca2+ release that couples elec- C
University, Sloan-Kettering Cancer Institute, trical activity to cardiac contraction. Phase 3 is
New York, NY, USA marked by inactivation of Ca2+ currents and
2
Department of Physiology and Biophysics, larger K+ currents that bring on the final repolar-
Weill Medical College of Cornell University, ization phase of the action potential, allowing the
New York, NY, USA cell to return to its resting potential in phase 4.
3
Department of Pharmacology, University of
California, Davis, Davis, CA, USA Perturbations in Ion Channels and Associated
Processes Cause Disease
A wealth of basic scientific investigations have
Definition revealed that aberrant ion channel gating can
result from naturally occurring mutations in the
The generation of the heartbeat is generated by genes encoding cardiac ion channels leading to
cellular level electrical activity in the form of cardiac arrhythmias (Noble and Rudy 2001;
action potentials that arises as a result of timed Priori et al. 1999a, b). Importantly, ion channels
opening and closing of a variety of cardiac ion also interact with, and are regulated by, accessory
channels. When this tightly coupled interplay of proteins that dynamically influence the action
ion channels is perturbed by any number of varied potential. Many such examples of indirect pertur-
mechanisms including, but not limited to, genetic bations underlying disease exist over the wide
perturbations, altered Ca2+ signaling, deranged range of cardiac ion channels. Since they cannot
subcellular signaling, disruptions in cell ultra- all be covered in this brief overview, we will
structure or coupling, ionic dysregulation, or describe some perturbations affecting cardiac
direct insult, an arrhythmogenic substrate can Na+ current mutations as paradigms for the
develop. types of perturbations that have been linked to
cardiac electrophysiologic pathology.
Since 1995 (Bennett et al. 1995), more than
Detailed Description 200 mutations have been linked to SCN5A, the
gene encoding the cardiac Na+ channel a subunit
Ionic Mechanisms of Cardiac Action Potential on chromosome 3p21 (Wang et al. 1995). Ini-
Action potentials can be divided into tially segregated into distinct disease phenotypes
noncontractile pacemaker cells and contractile such as long-QT3 syndrome (LQT3), Brugada
cells that require an external stimulus, such as syndrome (BrS), progressive cardiac conduction
atrial and ventricular cells (Demir 2004). Numer- defect (PCCD), sick sinus syndrome (SSS), atrial
ous action potential morphologies exist and vary fibrillation, and dilated cardiomyopathy (DCM),
depending on the location in the myocardium. it is now known that multiple genetic defects can
The action potential of contractile cells generally lead to overlapping syndromes with multiple
has 4 phases. Phase 0 is the rapid depolarizing clinical characteristics existing in one patient
phase resulting from fast Na+ channel activation (Ruan et al. 2009).
that causes rapid depolarization of the cell mem- LQT3 is a subset of the congenital long
brane potential. Phase 1 occurs when Na+ channel QT syndrome and is characterized by a delay
inactivate, and K+ ions pass outward through Ito in cardiac cellular repolarization, leading to
(transient outward currents, which underly the cardiac arrhythmias and sudden death, often in
C 560 Cardiac Excitable Tissue Pathology (Ion Channels)
young people. It has been proposed that delayed a loss of Na+ channel function (Tateyama et al.
ventricular repolarization promotes a substrate 2004; Vecchietti et al. 2007; Liu et al. 2002). BrS
for triggered activity via early afterdepolar- is an arrhythmic syndrome characterized by right
izations (EADs), which result from reactivation bundle branch block, ST-segment elevation on
of L-type Ca2+ channels (Thomas et al. 2007; the ECG, and ventricular tachyarrhythmias that
Viswanathan and Rudy 1999; Lankipalli et al. result in sudden death (Brugada et al. 1998). The
2005). reduction in INa has been shown to occur by
The first identified mutation in SCN5A, con- several mechanisms in BrS, including reduced
sequently shown to be a LQT3 variant, was rates of recovery from inactivation, faster inacti-
DKPQ, a 3 amino acid deletion (lysine, proline, vation subsequent to channel opening, and pro-
glutamine at positions 1505–1507) in the linker tein trafficking defects (Clancy and Kass 2002;
region between domains III and IV (Bennett et al. Grant et al. 2002; Tan et al. 2001; Bebarova et al.
1995). This motif is known to be critical for fast 2008).
inactivation of the Na+ channel, and this mutation BrS may result in heterogeneity of repolariza-
results in persistent non-inactivating current tion across the right ventricular myocardial wall.
(Bennett et al. 1995) (termed channel “bursting”) Repolarization in the epicardial cell layer is
during normal cellular repolarization. thought to be more sensitive to suppression of
While the DKPQ mutation is one example of the peak Na+ current due to a more prominent
altered gating, evidence suggests that mutation- transient outward K+ current (Ito) (Yan and
induced gain of function in cardiac INa can exist Antzelevitch 1999). Mutations in SCN5A that
in at least three distinct forms. In addition to reduce the peak INa could selectively hasten the
channel bursting, a second abnormality in Na+ repolarization in the epicardium, leading to a loss
channel gating that has been elucidated and of the epicardial action potential plateau phase
linked to phenotype results from steady-state and an increased propensity towards reentrant
channel reopening called window current (Ruan arrhythmias. This theory may explain the ST-
et al. 2009; Wang et al. 1996). Window currents segment elevation resulting from a dispersion of
arise as a result of reopenings that occur over action potential plateau potentials observed in
voltage ranges where steady-state inactivation patients with BrS.
and activation overlap. A third original mecha- Another proposed mechanism underlying the
nism was demonstrated in channels containing Brugada syndrome phenotype is that loss of Na+
the I1768V mutation, which does not result in current, observed as a depolarizing shift of the
an obvious gain of channel function. However, channel activation curve, may result in sufficiently
experiments and mathematical simulations slow conduction such that some areas of the myo-
revealed that under nonequilibrium conditions cardium are undergoing repolarization, while
during repolarization, channel reopening results others are still depolarizing (Zhang et al. 2007).
from faster recovery from inactivation at mem- Indeed, a simulation-based investigation of this
brane potentials that facilitate the activation tran- mechanism was investigated and shown to be plau-
sition (Clancy et al. 2003). sible and account for the observed ST-segment
In each of these mechanisms, pathologic late elevation in these patients (Zhang et al. 2007;
Na+ current due to channel reopening causes Petitprez et al. 2008; Miyoshi et al. 2005).
severe prolongation of the AP plateau (demon-
strated on the body surface ECG as QT interval Heterogeneity of Ion Channel Expression in
prolongation) and sets the stage for potentially the Heart: Dispersion of Action Potential
deadly arrhythmic triggers (Clancy et al. 2003). Morphology Leading to Arrhythmia
In contrast to LQTS-associated mutations that It has been shown experimentally that the myo-
cause a gain of Na+ channel function, mutations cardium displays distinct cell types that are het-
that have been identified in patients with congen- erogeneous in ion channel expression across the
ital forms of Brugada syndrome (BrS) result in ventricular wall. This heterogeneity plays a key
Cardiac Excitable Tissue Pathology (Ion Channels) 561 C
role in the normal sequence of ventricular exci- (Wingo et al. 2004), and the Ca2+-binding
tation and in arrhythmogenesis (Noble and Rudy protein calmodulin (CaM) (Mori et al. 2000)
2001; Antzelevitch et al. 1999). Potassium cur- was found on the carboxy terminus of the Na+
rents (IKr and IKs) in particular, which are essen- channel. These discoveries led to studies
tial in ventricular repolarization, show marked that revealed inactivation of INa can be
differences in expression between endocardial, modulated by Ca2+, CaM, and/or the Ca2+/CaM/
midmyocardial, and epicardial cells (Liu and CaM-kinase signaling cascade (Undrovinas and C
Antzelevitch 1995). Midmyocardial cells, for Maltsev 2008; Wagner et al. 2006; Maltsev et al.
example, have reduced IKs density and display 2008).
longer APD than other cell types, and these dif- In studies examining the interaction of the Na+
ferences in multicellular models can give rise to channel and INaL with CaMKII, it was found that
transmural heterogeneity which may contribute CaMKII coimmunoprecipitates and phosphory-
to the arrhythmogenic potential in long-QT syn- lates Na+ channels (Wagner et al. 2006).
drome (Viswanathan and Rudy 1999, 2000). Overexpression of CaMKIIdC in rabbit myocytes
(acute) and transgenic mice (chronic) led to
Mutations Can Result in Multiple Phenotypes (1) enhanced INaL, (2) slowed fast inactivation
The relationship between genetic mutations and (but enhanced intermediate inactivation),
clinical syndromes is complex, as the revelation (3) slowed recovery from inactivation, (4) shifted
of novel mutations suggests paradoxical pheno- steady-state availability in hyperpolarizing direc-
typic overlap or exclusivity. Multiple loci in the tion that was Ca2+ dependent, and (5) a rise in
cardiac Na+ channel have been identified where intracellular Na (Undrovinas and Maltsev 2008;
the same mutation can result in different disease Wagner et al. 2006); importantly, these results
phenotypes. An insertion of an aspartic acid res- were reversible with CaMKII inhibition (acute
idue (1795insD) in the carboxy terminus of only). Interestingly, Bers and Grandi (2009)
NaV1.5 can result in either BrS or LQTS (Bezzina note the striking similarity of these CaMKII-
et al. 1999). Additionally, mutation of the same induced changes with the LQT3- and Brugada-
residue to a histidine (Y1795H) or cysteine linked mutant 1795insD (Bezzina et al. 1999).
(Y1795C) results in BrS and LQTS, respectively, Maltsev and Undrovinas (Maltsev et al. 2008)
indicating the proximal region of the C-terminus found that in both normal and failing canine
as a potentially important structure in sodium ventricular myocytes, the Ca2+/CaM/CaMKII
channel function (Roden et al. 1996). Further signaling cascade increased INaL and Na+ influx
review can be found in Moreno and Clancy by slowing inactivation kinetics, but conversely
(2009). found a positive shift in the steady-state avail-
ability curve (Undrovinas and Maltsev 2008;
Calcium Dynamics and Its Regulation of the Maltsev et al. 2008). Evidence of CaMKII
Na+ Channel upregulation has been confirmed in other studies
The changing dynamics of calcium during the examining the effects of CaMKII in the heart
cardiac action potential is a complex and highly failure setting (Ai et al. 2005; Currie et al. 2004;
coupled system that includes L- and T-type Kirchhefer et al. 1999).
Ca2+ currents, ryanodine receptors, Na+/Ca2+
exchangers and SERCA pumps, Ca2+-binding Conclusions
proteins and buffering with calmodulin and tro- Understanding how perturbations to cardiac ion
ponin, and multiple sarcoplasmic reticulum (SR) channels emerge as altered behavior in higher
subspaces (Rice and Jafri 2001). Each of these dimensions of cardiac tissue and lead to patho-
processes has the potential to alter electrophysi- physiology including cardiac arrhythmia and
ological homeostasis and lead to arrhythmia. heart disease is a critical area of research. The
Ca2+ modulation of the Na+ channel has been ubiquity of ion channels throughout all excitable
demonstrated after direct binding sites for Ca2+ tissue and the myriad mutations that lead to
C 562 Cardiac Excitable Tissue Pathology (Ion Channels)
disease that have thus far been discovered have Demir SS (2004) Computational modeling of cardiac ven-
provided a tremendous substrate for a fundamen- tricular action potentials in rat and mouse: review. Jpn
J Physiol 54:523–530
tal understanding of ion channel biophysics. The Grant AO, Carboni MP, Neplioueva V, Starmer CF,
challenge continues to be integration of funda- Memmi M, Napolitano C, Priori S (2002) Long QT
mental mechanisms at the ion channel scale to the syndrome, Brugada syndrome, and conduction system
complex emergent phenomena of arrhythmia in disease are linked to a single sodium channel mutation.
J Clin Invest 110:1201–1209
higher dimensions. Kirchhefer U, Schmitz W, Scholz H, Neumann J (1999)
Activity of cAMP-dependent protein kinase and Ca2+/
calmodulin-dependent protein kinase in failing and
References nonfailing human hearts. Cardiovasc Res
42(1):254–261
Ai X, Curran JW, Shannon TR, Bers DM, Pogwizd SM Lankipalli RS, Zhu T, Guo D, Yan GX (2005) Mechanisms
(2005) Ca2+/calmodulin-dependent protein kinase underlying arrhythmogenesis in long QT syndrome.
modulates cardiac ryanodine receptor phosphorylation J Electrocardiol 38:69–73
and sarcoplasmic reticulum Ca2+ leak in heart failure. Liu DW, Antzelevitch C (1995) Characteristics of the
Circ Res 97(12):1314–1322 delayed rectifier current (IKr and IKs) in canine ven-
Antzelevitch C, Yan G, Shimizu W, Burashinikov tricular epicardial, midmyocardial, and endocardial
A (1999) Electrical heterogeneity, the ECG, and car- myocytes. A weaker IKs contributes to the longer
diac arrhythmias. In: DP Zipes, Jaife J (ed) Cardiac action potential of the M cell. Circ Res 76(3):351–365
electrophysiology: from cell to bedside. Saunders, Liu H, Tateyama M, Clancy C, Abriel H, Kass R (2002)
Philadelphia, pp 222-238 Channel openings are necessary but not sufficient for
Bebarova M, O’Hara T, Geelen JL, Jongbloed RJ, use-dependent block of cardiac Na + channels by
Timmermans C, Arens YH, Rodriguez LM, Rudy Y, flecainide evidence from the analysis of disease-linked
Volders PG (2008) Subepicardial phase 0 block and mutations. Journal of General Physiology 120:39–51
discontinuous transmural conduction underlie right Maltsev VA, Reznikov V, Undrovinas NA, Sabbah HN,
precordial ST-segment elevation by a SCN5A loss- Undrovinas A (2008) Modulation of late sodium cur-
of-function mutation. Am J Physiol Heart Circ Physiol rent by Ca2+, calmodulin, and CaMKII in normal and
295(1):H48–H58 failing dog cardiomyocytes: similarities and differ-
Bennett PB, Yazawa K, Makita N, George AL ences. Am J Physiol Heart Circ Physiol 294(4):
(1995) Molecular mechanism for an inherited cardiac H1597–H1608
arrhythmia. Nature 376:683–685 Miyoshi S, Mitamura H, Fukuda Y, Tanimoto K,
Bers DM, Grandi E (2009) Calcium/calmodulin- Hagiwara Y, Kanki H, Takatsuki S, Murata M, Miya-
dependent kinase II regulation of cardiac ion channels. zaki T, Ogawa S (2005) Link between SCN5A muta-
J Cardiovasc Pharmacol 54(3):180–187 tion and the Brugada syndrome ECG phenotype:
Bezzina C, Veldkamp MW, van Den Berg MP, simulation study. Circ J 69(5):567–575
Postma AV, Rook MB, Viersma JW, van Langen IM, Moreno JD, Clancy CE (2009) Using computational
Tan-Sindhunata G, Bink-Boelkens MT, van Der Hout modeling to predict arrhythmogenesis and antiarrhyth-
AH, Mannens MM, Wilde AA (1999) A single Na(+) mic therapy. Drug Discov Today Dis Models
channel mutation causing both long-QT and Brugada 6(3):71–84
syndromes. Circ Res 85:1206–1213 Mori M, Konno T, Ozawa T, Murata M, Imoto K,
Brugada J, Brugada R, Brugada P (1998) Right bundle- Nagayama K (2000) Novel interaction of the voltage-
branch block and ST-segment elevation in leads V-1 dependent sodium channel (VDSC) with calmodulin:
through V-3 - A marker for sudden death in patients does VDSC acquire calmodulin-mediated Ca2 + -sen-
without demonstrable structural heart disease. Circu- sitivity? Biochemistry 39(6):1316–1323
lation 97(5):457–460 Noble D, Rudy Y (2001) Models of cardiac ventricular
Clancy CE, Kass RS (2002) Defective cardiac ion chan- action potentials: interative interaction between exper-
nels: from mutations to clinical syndromes. J Clin iment and simulation. Phil Trans R Soc Lond
Invest 110(8):1075–1077 A 359:1127–1142
Clancy CE, Tateyama M, Liu H, Wehrens XH, Kass Petitprez S, Jespersen T, Pruvot E, Keller DI, Corbaz C,
R (2003) Non-equilibrium gating in cardiac Na + Schlapfer J, Abriel H, Kucera JP (2008) Analyses of
channels: an original mechanism of arrhythmia. Cir- a novel SCN5A mutation (C1850S): conduction vs.
culation 107:2233–2237 repolarization disorder hypotheses in the Brugada syn-
Currie S, Loughrey CM, Craig MA, Smith GL (2004) Cal- drome. Cardiovasc Res 78(3):494–504
cium/calmodulin-dependent protein kinase IIdelta Priori SG, Barhanin J, Hauer RN, Haverkamp W, Jongsma
associates with the ryanodine receptor complex and HJ, Kleber AG, McKenna WJ, Roden DM, Rudy Y,
regulates channel function in rabbit heart. Biochem Schwartz K, Schwartz PJ, Towbin JA, Wilde AM
J 377(Pt 2):357–366 (1999a) Genetic and molecular basis of cardiac
Cardiac Excitable Tissue Pathology (Ischemia) 563 C
arrhythmias: impact on clinical management parts Na + channel mutations in the congenital long QT
I and II. Circulation 99(4):518–528 syndrome. Proc Natl Acad Sci USA
Priori SG, Barhanin J, Hauer RN, Haverkamp W, Jongsma 93(23):13200–13205
HJ, Kleber AG, McKenna WJ, Roden DM, Rudy Y, Wingo TL, Shah VN, Anderson ME, Lybrand TP,
Schwartz K, Schwartz PJ, Towbin JA, Wilde AM Chazin WJ, Balser JR (2004) An EF-hand in
(1999b) Genetic and molecular basis of cardiac the sodium channel couples intracellular calcium to
arrhythmias: impact on clinical management part III. cardiac excitability. Nat Struct Mol Biol
Circulation 99(5):674–681 11(3):219–225
Rice J, Jafri MS (2001) Modelling calcium handling in Yan GX, Antzelevitch C (1999) Cellular basis for the
C
cardiac cells. Phil Trans R Soc Lond A 359:1143–1157 Brugada syndrome and other mechanisms of
Roden DM, Lazzara R, Rosen M, Schwartz PJ, Towbin J, arrhythmogenesis associated with ST-segment eleva-
Vincent GM (1996) Multiple mechanisms in the long- tion. Circulation 100(15):1660–1666
QT syndrome. Current knowledge, gaps, and future Zhang ZS, Tranquillo J, Neplioueva V, Bursac N, Grant
directions. The SADS Foundation Task Force on AO (2007) Sodium channel kinetic changes that pro-
LQTS. Circulation 94(8):1996–2012 duce Brugada syndrome or progressive cardiac con-
Ruan Y, Liu N, Priori SG (2009) Sodium channel muta- duction system disease. Am J Physiol Heart Circ
tions and arrhythmias. Nat Rev Cardiol 6:337–348 Physiol 292(1):H399–H407
Tan HL, Bink-Boelkens MT, Bezzina CR, Viswanathan
PC, Beaufort-Krol GC, van Tintelen PJ, van den Berg
MP, Wilde AA, Balser JR (2001) A sodium-channel
mutation causes isolated cardiac conduction disease. Cardiac Excitable Tissue Pathology
Nature 409(6823):1043–1047
Tateyama M, Liu H, Yang AS, Cormier JW, Kass RS
(Ischemia)
(2004) Structural effects of an LQT-3 mutation on
heart Na + channel gating. Biophys J 86:1843–1851 Byron N. Roberts and Colleen E. Clancy
Thomas G, Gurung IS, Killeen MJ, Hakim P, Goddard Department of Pharmacology, University of
CA, Mahaut-Smith MP, Colledge WH, Grace AA,
Huang CL (2007) Effects of L-type Ca2+ channel
California, Davis, Davis, CA, USA
antagonism on ventricular arrhythmogenesis in murine
hearts containing a modification in the Scn5a gene
modelling human long QT syndrome 3. J Physiol Definition
(Lond) 578:85–97
Undrovinas A, Maltsev VA (2008) Late sodium current is
a new therapeutic target to improve contractility and Ischemia is a condition that results when tissue is
rhythm in failing heart. Cardiovasc Hematol Agents inadequately perfused with blood, leading to
Med Chem 6(4):348–359 insufficient delivery of oxygen and necessary
Vecchietti S, Grandi E, Severi S, Rivolta I, Napolitano C,
Priori SG, Cavalcanti S (2007) In silico assessment of
components, as well as impaired removal of met-
Y1795C and Y1795H SCN5A mutations: implication abolic waste products. In the context of cardiac
for inherited arrhythmogenic syndromes. Am J Physiol tissue, ischemia can produce derangements of
Heart Circ Physiol 292(1):H56–65. Epub 2006 Sep 15 electrical activity, reduced mechanical pumping
Viswanathan PC, Rudy Y (1999) Pause induced early
afterdepolarizations in the long QT syndrome:
ability, myocardial infarction, and sudden cardiac
a simulation study. Cardiovasc Res 42:530–542 death.
Viswanathan PC, Rudy Y (2000) Cellular arrhythmogenic A wealth of experimental data has been gen-
effects of congenital and acquired long-QT syndrome erated describing the fundamental structural and
in the heterogeneous myocardium. Circulation
101:1192–1198
functional changes that occur in cardiac tissue
Wagner S, Dybkova N, Rasenack EC, Jacobshagen C, during ischemia. However, it is increasingly
Fabritz L, Kirchhof P, Maier SK, Zhang T, Hasenfuss clear that the many complex changes that occur
G, Brown JH, Bers DM, Maier LS (2006) Ca2+/cal- during the time course of ischemia and their
modulin-dependent protein kinase II regulates cardiac
Na + channels. J Clin Invest 116(12):3127–3138
resultant effects on emergent electrical and
Wang Q, Shen J, Splawski I, Atkinson D, Li Z, Robinson mechanical functioning in the heart cannot be
JL, Moss AJ, Towbin JA, Keating MT (1995) SCN5A fully observed or readily predicted via study of
mutations associated with an inherited cardiac arrhyth- constituent elements alone. For this reason,
mia, long QT syndrome. Cell 80:805–811
Wang DW, Yazawa K, George AL, Bennett PB
a number of modeling and simulation studies
(1996) Characterization of human cardiac have been undertaken to reveal fundamental
C 564 Cardiac Excitable Tissue Pathology (Ischemia)
biological principles and mechanisms of func- More recently, Crampin and Smith developed
tional changes observed in cardiac ischemia and a computational model of cardiac acidosis that
to predict how interactions between system com- includes a dynamical model of pH regulation via
ponents contribute to pathological electrical and inclusion of model representations of NHE as
mechanical changes in cardiac cells and tissue. well as the sodium-bicarbonate symporter
Modeling and simulation have also been used to (NBC), anion exchanger (AE), and chloride-
predict emergent disease processes and treatment hydroxide exchanger (CHE) – all of which
effects in ischemia. At present, there is no alter- respond to changes in pH (Crampin and Smith
native, efficient, and cost-effective experimental 2006). The model simulations predicted effects
or clinical strategy that can achieve these goals. of changes in pH on the time course of key
Enabling computationally based investigations ionic species during acidosis, including intra-
are advances in numerical techniques and com- cellular protons, sodium, and calcium. The
puting (Nivala et al. 2012; Abramson et al. 2010; Crampin and Smith model also incorporated
Rocha et al. 2011; Neic et al. 2012), the dynamic models of direct and allosteric regula-
implementation of customizable solvers such tion of subcellular components by pH, including
as Continuity (http://www.continuity.ucsd.edu/ the sodium-calcium exchanger (NCX), sodium-
Continuity), modeling platforms like CHASTE potassium pump (NaK), SERCA pump, and
and OpenCMISS (Bernabeu et al. 2009; Bradley ryanodine receptors (RyRs), which all exhibit
et al. 2011), and infrastructures aimed at facili- decreased function at subnormal pH. In addi-
tating standardization, interoperability, and tion, the model integrated experimentally
dissemination of models (e.g., CellML and observed pH-dependent changes in the conduc-
FieldML) (Bernabeu et al. 2009; Bradley et al. tance and/or gating kinetics of key cardiac ionic
2011; Christie et al. 2009; Quinn et al. 2011; currents (e.g., INa, ICa, Ito, IKr, IKs, IK1). This
Wimalaratne et al. 2009). model was then used to make predictions about
the effect of acidosis on electrophysiology and
calcium cycling, the latter of which is
Detailed Description a surrogate readout for excitation-contraction
coupling (Crampin and Smith 2006; Crampin
Effects of Derangements to pH and et al. 2006). Simulations suggested that critical
Phosphometabolites to suppressed contractility are elevated sodium,
During ischemia, one of the initial changes in the inhibition of sodium-calcium exchange, and the
cardiac cell is the acidification of both the intra- direct interaction of protons with the contractile
and extracellular compartments (Carmeliet machinery.
1999). This acidification is the result of increased A key calcium cycling component from the
proton production inside the cell as it undergoes Crampin and Smith model, the cardiac sarcoplas-
a shift from mitochondrial to glycolytic ATP mic/endoplasmic calcium (SERCA) pump, was
production, which is exacerbated as proton- subsequently modeled in more detail with partic-
absorbing phosphocreatine (PCr) is depleted. ular inclusion of multiple phosphate species
Some increased proton production is compen- (MgATP, MgADP, and Pi) to allow detailed
sated for by the sodium-hydrogen exchanger computational investigations on effects of altered
(NHE), which moves sodium into the cell and metabolic function on the SERCA pump (Tran
exports protons, a process that promotes extracel- et al. 2009). Model predictions from this study
lular acidification. Shaw and Rudy initially showed that binding of MgATP had a large inhib-
developed mathematical models that incorpo- itory effect on the maximal reverse rate of the
rated electrophysiological intracellular pH mod- SERCA pump. During ischemia, when ATP
ulation of membrane components to investigate availability is low, reverse SERCA flux would
pH effects on excitability of cardiac cells and be more prominent. This pathological effect is
tissues (Shaw and Rudy 1997a, b). expected to contribute to calcium overload and
resultant calcium-dependent arrhythmias that increase the likelihood of reentrant
observed in ischemia. arrhythmias.
Another pH derangement in ischemic tissue
is impaired proton removal that is exacerbated Accumulation of Extracellular K
by the sustained presence of carbon dioxide in One of the hallmark indicators of ischemia in
ischemic tissue (Carmeliet 1999). The recent cardiac tissue is an increase in extracellular potas-
modeling and simulation study by Roberts sium that arises from impairments in cellular C
and Christini incorporated dynamic concentra- metabolism and predisposes the heart to lethal
tion changes for extracellular carbon dioxide arrhythmias (Kleber 1983; Cascio et al. 1992).
and bicarbonate into the Crampin and Smith Hyperkalemia depolarizes the resting membrane
model (Crampin and Smith 2006) and then uti- potential, which reduces sodium channel avail-
lized the model to show how NHE inhibitors, ability and cell excitability. The consequent
which have been expected to reduce ischemia- depression in upstroke velocity, coupled with
reperfusion injury through normalization of decreased cellular coupling in response to low
Na homeostasis, made things worse by pH and ATP availability (Sugiura et al. 1990;
prolonging intracellular acidosis because of Swietach et al. 2007), leads to a significant – and
impaired proton removal. The resulting potentially proarrhythmic – reduction of tissue
prolonged acidosis was predicted to inhibit mul- conduction velocity (Kagiyama et al. 1982;
tiple subcellular components, especially NaK Carmeliet 1999).
(which is responsible for removing sodium), Shaw and Rudy developed an ionic-based the-
with the unintended and paradoxical conse- oretical model of the cardiac ventricular cell to
quence of perpetuating sodium overload (Rob- simulate conditions of acute ischemia including
erts and Christini 2011). accumulation of extracellular potassium, acido-
Roberts and Christini also applied a sis, and anoxia and to predict and examine their
comprehensive ischemia-reperfusion model that effects on ion concentrations, ionic currents, and
incorporates both pH- and phosphometabolite- the cardiac action potential (Shaw and Rudy
dependent effects to dissect out specific roles for 1997a). They used the model to predict the under-
pH- or phosphometabolite-related processes in lying ionic mechanisms responsible for reduced
generating changes in excitability and action excitability and shortening of the action potential
potential morphology that are associated with duration. Shaw and Rudy showed that extracel-
ischemia-linked development of cardiac arrhyth- lular potassium accumulation was the primary
mias (Roberts and Christini 2012). The model effector of cell excitability by depolarizing rest-
predictions suggested that phosphometabolites ing membrane potential and reducing sodium
played a larger role than pH changes in persistent channel availability. Additional depression of
shortening of APD, reduction of action potential excitability was induced by acidosis. Major
amplitude (APA), and depolarization of the rest- changes in action potential duration were
ing membrane potential. Reduced phosphome- predicted by the model to be caused by anoxia-
tabolite availability and pH recovery were dependent opening of the ATP-sensitive potas-
predicted to affect multiple ion channels and sium current (IK(ATP)), consistent with predic-
exchangers. Some of these effects were the result tions from an earlier simulation study by Ferrero
of direct modulation by phosphometabolites and/ et al. (1996).
or acidosis, while others resulted from elevated Potassium accumulation exhibits a multi-
sodium and calcium loads during reperfusion. phasic time course resulting from a presumed
The results from Roberts and Christini suggested combination of three mechanisms that includes
that therapies directed at increasing phosphome- (1) increased K(+) efflux, (2) decreased K(+)
tabolites availability during reperfusion may be influx, and (3) reduction in the volume of the
more beneficial than aiding pH recovery with extracellular space. Because the relative contri-
regard to mitigating action potential changes butions of each of these mechanisms have been
challenging to determine experimentally, model- Zhou et al. subsequently developed a reaction-

ing and simulation studies have been undertaken diffusion mathematical model of ROS-induced
to predict how ionic mechanisms contribute to ROS release in the mitochondrial network and
accumulation of extracellular potassium. demonstrated that such a model was sufficient to
Terkildsen and coauthors (2007) developed reproduce the experimentally observed emergent
a model that was used to simulate changes in macroscopic properties of the mitochondrial net-
metabolite concentrations during acute global work (Zhou et al. 2010). Model simulations were
ischemia and investigate their effects in also used to confirm experimental observations
a dynamic model of cellular electrophysiology. that suggested depolarization waves of the
The model incorporated a metabolically sensitive mitochondrial membrane potential could arise
description of the Na(+)-K(+) pump and (similar to cell membrane in cardiomyocytes)
ATP-sensitive potassium current, in addition to following stimulation with localized laser-flash
cell volume regulation to allow a quantitative or antioxidant depletion. Moreover, the model
description of ischemic hyperkalemia. Simula- predicted similar sensitivity of the wave propa-
tions predicted that reduced NaK activity plays gation rate to changes in quantities including
the largest role in extracellular potassium superoxide, diffusion, production, and free radi-
accumulation, followed by increased potassium cal scavenging to that observed experimentally.
efflux via IK(ATP) and a shrinking extracellular Finally, the model predicted novel mechanisms
space due to cell swelling. for synchronization of electrical activity, arising
from clusters of oscillating mitochondria that
Reactive Oxygen Species entrain neighboring mitochrondria.
Reactive oxygen species (ROS) play a role in Oxidative stress that occurs during ischemia
ischemic tissue pathology and result from can also result in modification of ionic compo-
increased production (e.g., increased free radical nents both directly and indirectly through activa-
generation) coupled with reduced availability tion of modulatory cell signaling pathways.
of scavenging compounds (Murphy and Because these networks of interactions are
Steenbergen 2008). ROS can damage the cell extraordinarily complicated and typically studied
membrane via lipid peroxidation and increased in isolation in experiment, computational model-
leakiness and has been implicated in the collapse ing approaches have been used to predict effects
of mitochondrial function (Akar and O’Rourke of ischemia-linked oxidation on emergent prop-
2011; Xie and Akar 2010; Akar et al. 2005; erties of cardiac cells and tissues. Christensen and
Aon et al. 2006, 2009). Modeling and simulation coauthors undertook a modeling and simulations
have been important tools to investigate studies that predicted effects of alterations in the
how metabolic dysfunction that originates in the cardiac Na current by oxidized calmodulin kinase
mitochondria can disrupt cardiac electrophysiol- II (CaMKII) (Christensen et al. 2009). The model
ogy and lead to arrhythmias (Zhou and O’Rourke predicted that phosphorylation of Na channels by
2012). Zhou et al. formulated a model of oxidation-activated CaMKII causes a reduction
the cardiac action potential in which in Na channel availability, a reduction in
a novel link between the membrane potential of cellular excitability, and a slowing of conduction
the mitochondrial and the cellular membrane velocity of excitatory waves in tissue, increasing
was established (Zhou et al. 2009). The model the likelihood of conduction block in the ische-
predictions suggested that effects of oxidative mic border zone.
stress on mitochondrial membrane potential Wagner et al. subsequently used a combined
are a primary mechanism for shortening of experimental and computational approach to
the cellular action potential duration and probe the role for ROS-activated CaMKII effects
suppression of cellular electrical excitability on Na overload that is commonly observed in
(Zhou et al. 2009). ischemia (Wagner et al. 2011). The model
showed that although ROS-activated CaMKII afterdepolarizations can give rise to spontaneous
effects on Na current may play a role in Na excitatory waves if they excite a large enough
overload in the cells, the model suggested that region. Reentrant arrhythmias require such
this mechanism alone was not sufficient to a trigger and are perpetuated by slowed conduc-
account for the large increases in Na observed tion and unidirectional conduction block, mech-
experimentally. anisms that have all been explored in quantitative
detail using modeling and simulation approaches C
Disruption of Electrical Function and (Starmer 1997; Quan and Rudy 1990; Cherry
Arrhythmia Triggers et al. 2012; Jie and Trayanova 2010; Qu et al.
A number of modeling studies have been under- 2006; Jie et al. 2008).
taken to better understand how single-cell elec-
trical changes translate to tissue-level pathologies Mechanical Dysfunction
during ischemia (Shaw and Rudy 1997b; Romero Ischemia produces profound effects on cardiac
et al. 2009; Rodriguez et al. 2004; Jie and contractile function. Ischemia can produce an
Trayanova 2010). Shaw and Rudy carried out increase in diastolic pressure due to contracture,
simulations in a multicellular one-dimensional as well a reduction or absence of developed pres-
ventricular fiber representation to determine sure (Allen and Orchard 1987). Mathematical
mechanisms of slowed conduction and conduc- models that contain explicit representations of
tion failure during acute ischemia (Shaw and cross-bridge formation and tension development
Rudy 1997b). They simulated conditions of ele- have been used to better understand how ische-
vated extracellular potassium, acidosis, and mic metabolic pathology leads to impaired con-
anoxia. The model predictions suggested that tractile force (Rice et al. 2008; Crampin and
elevation in extracellular potassium was the Smith 2006; Tran et al. 2010; Loiselle et al.
major determinant of conduction, causing super- 2008; Campbell et al. 2010; Ch’en et al. 1998).
normal or depressed conduction, and was the Tran et al. developed a thermodynamically
only factor that could singularly cause conduc- consistent and metabolite-sensitive model of car-
tion block when extracellular potassium varied diac cross-bridge cycling to explore the influence
within the range reported for acute ischemia. This of ischemic metabolites on force production
study demonstrated the potential for cardiac under normal and ischemic conditions (Tran
L-type Ca2+ current to support conduction of et al. 2010). The model recapitulated many of
excitatory waves and defined a “safety factor” the experimentally observed effects of MgATP,
for conduction that accounts not only for various MgADP, Pi, and H+ on force development and
ionic currents but also the source-sink relation- retained the force/length/Ca2+ properties of ear-
ship in coupled tissue. lier models upon which the formulation is based
The source-sink relationship and its link to (Rice et al. 2008; Crampin and Smith 2006;
reentry vulnerability were explored in another Loiselle et al. 2008; Ch’en et al. 1998). The
modeling and simulation study by Romero and model simulations suggested that the force-
colleagues (2009). These simulations suggested MgADP relationship (positive or negative)
that the source-sink relationship was the ultimate depends on the concentrations of the other metab-
cause of the unidirectional conduction block that olites and [H+].
can lead to reentrant arrhythmias. During ischemia, afterdepolarizations can
The most dangerous arrhythmias that arise give rise to a spontaneous excitatory wave, and
during ischemia are ventricular tachycardia and if the ectopic focus spreads to a large enough
fibrillation. These arrhythmias can be caused region, it can trigger a lethal arrhythmia.
either by rapid repeated activation of an abnormal A three-dimensional tissue model incorporating
(ectopic) self-exciting region, or they can mani- mechanosensitive channels and regional
fest as reentrant arrhythmias. During ischemia, ischemia has been used to study how
mechanoelectrical feedback can trigger prema- Heidlauf T, Krittian S, Ladd D, Little C, Mithraratne
ture ventricular beats and have the potential K, Nash M, Nickerson D, Nielsen P, Nordbo O,
Omholt S, Pashaei A, Paterson D, Rajagopal V,
to trigger reentry (Jie and Trayanova 2010; Jie Reeve A, Rohrle O, Safaei S, Sebastian R, Steghofer
et al. 2010). M, Wu T, Yu T, Zhang HY, Hunter P (2011)
Opencmiss: a multi-physics & multi-scale computa-
tional infrastructure for the vph/physiome project.
Prog Biophys Mol Biol 107:32–47
Conclusions Campbell SG, Lionetti FV, Campbell KS, McCulloch AD
(2010) Coupling of adjacent tropomyosins enhances
Understanding how ischemia-induced alterations cross-bridge-mediated cooperative activation in
in multiple cardiac components emerge as altered a markov model of the cardiac thin filament. Biophys
J 98:2254–2264
behavior in higher dimensions of cardiac tissue Carmeliet E (1999) Cardiac ionic currents and acute ische-
and cause pathophysiology–including cardiac mia: from channels to arrhythmias. Physiol Rev
arrhythmia and mechanical dysfunction is 79:917–1017
a critical area of research. The ubiquity of patho- Cascio WE, Yan GX, Kleber AG (1992) Early changes in
extracellular potassium in ischemic rabbit myocar-
logical ischemia as a disease process affecting dium. The role of extracellular carbon dioxide accu-
most excitable tissue and the myriad perturbations mulation and diffusion. Circ Res 70:409–422
that occur simultaneously make this a disease Ch’en FF, Vaughan-Jones RD, Clarke K, Noble D (1998)
that can benefit from modeling and simulations Modelling myocardial ischaemia and reperfusion.
Prog Biophys Mol Biol 69:515–538
approaches to improve the fundamental under- Cherry EM, Fenton FH, Gilmour RF (2012) Mechanisms
standing of the disease process and to reveal of ventricular arrhythmias: a dynamical systems-
experimentally testable targets for therapy. based perspective. Am J Physiol-Heart C 302:
H2451–H2463
Christensen MD, Dun W, Boyden PA, Anderson ME,
Mohler PJ, Hund TJ (2009) Oxidized calmodulin
References kinase ii regulates conduction following myocardial
infarction: a computational analysis. PLoS Comput
Abramson D, Bernabeu MO, Bethwaite B, Burrage K, Biol 5:e1000583
Corrias A, Enticott C, Garic S, Gavaghan D, Peachey Christie GR, Nielsen PMF, Blackett SA, Bradley CP,
T, Pitt-Francis J, Pueyo E, Rodriguez B, Sher A, Tan Hunter PJ (2009) Fieldml: concepts and implementa-
J (2010) High-throughput cardiac science on the grid. tion. Philos Trans R Soc A 367:1869–1884
Philos Trans R Soc A 368:3907–3923 Crampin EJ, Smith NP (2006) A dynamic model of exci-
Akar FG, O’Rourke B (2011) Mitochondria are sources of tation-contraction coupling during acidosis in cardiac
metabolic sink and arrhythmias. Pharmacol Ther ventricular myocytes. Biophys J 90:3074–3090
131:287–294 Crampin EJ, Smith NP, Langham AE, Clayton RH,
Akar FG, Aon MA, Tomaselli GF, O’Rourke B (2005) Orchard CH (2006) Acidosis in models of cardiac
The mitochondrial origin of postischemic arrhythmias. ventricular myocytes. Philos Transact A Math Phys
J Clin Invest 115:3527–3535 Eng Sci 364:1171–1186
Allen DG, Orchard CH (1987) Myocardial contractile Ferrero JM Jr, Saiz J, Ferrero JM, Thakor NV (1996) Sim-
function during ischemia and hypoxia. Circ Res ulation of action potentials from metabolically
60:153–168 impaired cardiac myocytes. Role of atp-sensitive k +
Aon MA, Cortassa S, Akar FG, O’Rourke B (2006) Mito- current. Circ Res 79:208–221
chondrial criticality: a new concept at the turning point Jie X, Trayanova NA (2010) Mechanisms for initiation of
of life or death. Biochim Biophys Acta 1762:232–240 reentry in acute regional ischemia phase 1b. Heart
Aon MA, Cortassa S, Akar FG, Brown DA, Zhou L, Rhythm 7:379–386
O’Rourke B (2009) From mitochondrial dynamics to Jie X, Rodriguez B, de Groot JR, Coronel R, Trayanova
arrhythmias. Int J Biochem Cell Biol 41:1940–1948 N (2008) Reentry in survived subepicardium coupled
Bernabeu MO, Bordas R, Pathmanathan P, Pitt-Francis J, to depolarized and inexcitable midmyocardium:
Cooper J, Garny A, Gavaghan DJ, Rodriguez B, South- insights into arrhythmogenesis in ischemia phase 1b.
ern JA, Whiteley JP (2009) Chaste: incorporating Heart Rhythm 5:1036–1044
a novel multi-scale spatial and temporal algorithm Jie X, Gurev V, Trayanova N (2010) Mechanisms of
into a large-scale open source library. Philos Trans mechanically induced spontaneous arrhythmias in
R Soc A 367:1907–1930 acute regional ischemia. Circ Res 106:185–192
Bradley C, Bowery A, Britten R, Budelmann V, Camara Kleber AG (1983) Resting membrane potential, extracel-
O, Christie R, Cookson A, Frangi AF, Gamage TB, lular potassium activity, and intracellular sodium
activity during acute global ischemia in isolated per- Rodriguez B, Tice BM, Eason JC, Aguel F, Ferrero JM Jr,
fused guinea pig hearts. Circ Res 52:442–450 Trayanova N (2004) Effect of acute global ischemia on
Loiselle DS, Crampin EJ, Niederer SA, Smith NP, Barclay the upper limit of vulnerability: a simulation study.
CJ (2008) Energetic consequences of mechanical Am J Physiol Heart Circ Physiol 286:H2078–2088
loads. Prog Biophys Mol Biol 97:348–366 Romero L, Trenor B, Alonso JM, Tobon C, Saiz J, Ferrero
Murphy E, Steenbergen C (2008) Mechanisms underlying JM Jr (2009) The relative role of refractoriness and
acute protection from cardiac ischemia-reperfusion source-sink relationship in reentry generation during
injury. Physiol Rev 88:581–609 simulated acute ischemia. Ann Biomed Eng
Neic A, Liebmann M, Hoetzl E, Mitchell L, Vigmond EJ, 37:1560–1571
C
Haase G, Plank G (2012) Accelerating cardiac Shaw RM, Rudy Y (1997a) Electrophysiologic effects of
bidomain simulations using graphics processing acute myocardial ischemia: a theoretical study of
units. IEEE Trans Biomed Eng 59:2281–2290 altered cell excitability and action potential duration.
Nivala M, de Lange E, Rovetti R, Qu Z (2012) Computa- Cardiovasc Res 35:256–272
tional modeling and numerical methods for spatiotem- Shaw RM, Rudy Y (1997b) Electrophysiologic effects of
poral calcium cycling in ventricular myocytes. Front acute myocardial ischemia. A mechanistic investiga-
Physiol 3:114 tion of action potential conduction and conduction
Qu ZL, Garfinkel A, Weiss JN (2006) Vulnerable window failure. Circ Res 80:124–138
for conduction block in a one-dimensional cable of Starmer CF (1997) The cardiac vulnerable period and
cardiac cells, 2: multiple extrasystoles. Biophys reentrant arrhythmias: targets of anti- and
J 91:805–815 proarrhythmic processes. Pace 20:445–454
Quan WL, Rudy Y (1990) Unidirectional block and reen- Terkildsen JR, Crampin EJ, Smith NP (2007) The balance
try of cardiac excitation - a model study. Circ Res between inactivation and activation of the na + -
66:367–382 k + pump underlies the triphasic accumulation of
Quinn TA, Granite S, Allessie MA, Antzelevitch C, extracellular k + during myocardial ischemia. Am
Bollensdorff C, Bub G, Burton RAB, Cerbai E, Chen J Physiol Heart Circ Physiol 293:H3036–3045
PS, Delmar M, DiFrancesco D, Earm YE, Efimov IR, Tran K, Smith NP, Loiselle DS, Crampin EJ
Egger M, Entcheva E, Fink M, Fischmeister R, Franz (2009) A thermodynamic model of the cardiac sarco-
MR, Garny A, Giles WR, Hannes T, Harding SE, plasmic/endoplasmic ca(2+) (serca) pump. Biophys
Hunter PJ, Iribe G, Jalife J, Johnson CR, Kass RS, J 96:2029–2042
Kodama I, Koren G, Lord P, Markhasin VS, Matsuoka Tran K, Smith NP, Loiselle DS, Crampin EJ
S, McCulloch AD, Mirams GR, Morley GE, Nattel S, (2010) A metabolite-sensitive, thermodynamically
Noble D, Olesen SP, Panfilov AV, Trayanova NA, constrained model of cardiac cross-bridge cycling:
Ravens U, Richard S, Rosenbaum DS, Rudy Y, implications for force development during ischemia.
Sachs F, Sachse FB, Saint DA, Schotten U, Solovyova Biophys J 98:267–276
O, Taggart P, Tung L, Varro A, Volders PG, Wang K, Wagner S, Ruff HM, Weber SL, Bellmann S, Sowa T,
Weiss JN, Wettwer E, White E, Wilders R, Winslow Schulte T, Anderson ME, Grandi E, Bers DM, Backs J,
RL, Kohl P (2011) Minimum information about Belardinelli L, Maier LS (2011) Reactive oxygen spe-
a cardiac electrophysiology experiment (micee): cies-activated ca/calmodulin kinase iidelta is required
standardised reporting for model reproducibility, inter- for late i(na) augmentation leading to cellular na and ca
operability, and data sharing. Prog Biophys Mol Bio overload. Circ Res 108:555–565
107:4–10 Wimalaratne SM, Halstead MDB, Lloyd CM, Crampin
Rice JJ, Wang F, Bers DM, de Tombe PP (2008) Approx- EJ, Nielsen PF (2009) Biophysical annotation and
imate model of cooperative activation and crossbridge representation of cellml models. Bioinformatics
cycling in cardiac muscle using ordinary differential 25:2263–2270
equations. Biophys J 95:2368–2390 Xie C, Akar FG (2010) Mitochondrial dysfunction: a new
Roberts BN, Christini DJ (2011) Nhe inhibition does not frontier in the search for elusive arrhythmia mecha-
improve na(+) or ca(2+) overload during reperfusion: nisms. Front Physiol 1:163
using modeling to illuminate the mechanisms under- Zhou L, O’Rourke B (2012) Cardiac mitochondrial net-
lying a therapeutic failure. PLoS Comput Biol work excitability: insights from computational analy-
7:e1002241 sis. Am J Physiol Heart Circ Physiol 302:H2178–2189
Roberts BN, Christini DJ (2012) The relative influences of Zhou L, Cortassa S, Wei AC, Aon MA, Winslow RL,
phosphometabolites and ph on action potential mor- O’Rourke B (2009) Modeling cardiac action potential
phology during myocardial reperfusion: a simulation shortening driven by oxidative stress-induced mito-
study. PLoS One 7:e47117 chondrial oscillations in guinea pig cardiomyocytes.
Rocha BM, Campos FO, Amorim RM, Plank G, dos Biophys J 97:1843–1852
Santos RW, Liebmann M, Haase G (2011) Accelerat- Zhou L, Aon MA, Almas T, Cortassa S, Winslow RL,
ing cardiac excitation spread simulations using O’Rourke B (2010) A reaction-diffusion model of ros-
graphics processing units. Concurr Comp-Pract induced ros release in a mitochondrial network. PLoS
E 23:708–720 Comput Biol 6:e1000657
C 570 Categorical Decisions
Behavioral Tasks for Studying Categorical

Categorical Decisions Decisions
Both traditions of decision-making have exten-
Paul Masset1 and Adam Kepecs2 sively used two-alternative forced choice (2AFC)
1
Watson School of Biological Sciences, Cold task designs. 2AFC is the simplest task in which
Spring Harbor Laboratory, Cold Spring Harbor, the process of a decision between alternatives can
NY, USA be studied. In a 2AFC task, the subject is forced to
2
Cold Spring Harbor Laboratory, Cold Spring make a choice between two alternatives based on
Harbor, NY, USA the stimulus she has experienced.
An example of a widely used perceptual
decision-making task is the visual random dot
Definition task (Newsome et al. 1989; Gold and Shadlen
2007). In this task an ambiguous perceptual stim-
Categorical decision-making is the process of ulus is instantiated by a cloud of moving dots.
committing to a particular option from Subjects are asked to report whether the apparent
a discrete set of alternatives. motion is to a given direction or its opposite.
A random process controls the moment-to-
moment movement of individual dots. The
Detailed Description strength of apparent motion and hence stimulus
discriminability are controlled by the fraction of
Introduction coherently moving dots. To perform correctly on
Studies of categorical decision-making attempt the task, subjects need to view the stimulus over
to understand behavior by probing how some period of time to extract the direction of
different features of complex and changing coherent motion.
environments guide the selection of choices. An example of a value-based decision-making
While the parameters underlying these features task is the matching paradigm in which subjects
often span a continuous range, the potential set have to choose between two options that differ in
of possible behavioral options is discrete. The the probability and size of rewards (Platt and
neuroscientific study of decision-making Glimcher 1999; Sugrue et al. 2004). Both reward
draws heavily on the fields of psychology, eco- probability and size are set by the experimenter
nomics, statistics, and ecology. Neuroscientific and change across trial blocks. Here a subject has
approaches to decision-making aim to to integrate across trials (and not across time
reveal computational principles that can be within a trial) to infer the value of the stimuli
mapped onto their neurobiological and make the correct choice (Fig. 1).
implementation.
There are two dominant traditions in neuro- Models of Decision-Making
science and psychology to study categorical Computational studies of decision-making in
decisions: perceptual and value-based decision- neuroscience often focus on algorithmic descrip-
making. Perceptual decision-making focuses on tions of the decision process and attempt to
how accurate decisions are reached by resolving describe both behavior and its neural correlates.
perceptual uncertainty. In value-based decision- Models of perceptual decision-making
making, the stimuli themselves are not ambigu- describe how ambiguous perceptual information
ous, rather the value or utility of different is processed and leads to the selection of a choice.
options needs to be estimated based on prior At every instant, a subject extracts some features
experience. In both cases, the goal is to system- from the stimulus that will form the evidence that
atically manipulate different features of the guides the decision-making process. One popular
environment in order to understand how they class of models that have been proposed to
guide behavior. describe how the evidence is used is based on
Categorical Decisions 571 C
a Integration
Fore-period Delay Motion Choice
C
b Integration Integration
Fore-period Fore-period Fore-period
Outcome Outcome
Delay Delay Delay
Choice Choice Choice
Categorical Decisions, Fig. 1 (a) Schematic of the targets. Choice: the subject makes a saccade to one of the
random dot task. Foreperiod: a fixation point appears at two targets. In this paradigm, the integration of evidence
the center of the screen. Delay: once the subject fixates the occurs during the stimulus presentation. (b) Schematic of
point, the two saccade targets appear on either side of the matching task. Foreperiod and delay: same as in (a).
fixation point. The trial is ready to start. Motion: the cloud Choice: the subject makes a saccade to one of the two
of randomly moving dots appears. A subset of these dots targets. In this paradigm, the integration of evidence
moves coherently in the direction of one of the saccade occurs by integration over the previous outcome history
sequential sampling (Bogacz et al. 2006). They These models have been popular, as they can
have been chiefly motivated by sequential prob- account for the dependence on the stimulus of
ability ratio test, in which the ratio of the proba- both the choice and the reaction time (tchoice
bility of possible alternatives is computed as where V(tchoice) = B). They also explain the
evidence is accumulated. These models can be widely observed trade-off between the speed
solved as a first passage time problem; a and accuracy of decisions. A lower threshold
commitment to a decision occurs once the prob- B implies that less accumulation of evidence
ability ratio reaches a set value. As shown in is necessary to commit to a decision, which
Eq. 1, the decision variable V(t) is updated by leads to faster reaction times at the expense of
integrating the momentary evidence e(t). Once reduced accuracy. Neural recordings have
the decision variable reaches a set bound B, revealed evidence for the existence of such
the corresponding alternative is chosen and the accumulators in neural circuits implicated in per-
response is initiated (Eq. 2). More complex ceptual decision-making (Gold and Shadlen
models have been developed taking into account 2007).
physiological characteristics such as adaptation In value-based decision-making, models have
and the role of time in the decision process used ideas from reinforcement learning. In
(Drugowitsch et al. 2012; Brunton et al. 2013). a matching task, even if a full model of the task
Accumulation: structure is not available to the subject, a
reinforcement-learning algorithm can estimate
dV the values of the stimulus features. The value of
¼ eðtÞ þ noise (1)
dt a stimulus is the subjective utility of its outcome,
which needs to be estimated based on experience.
Decision rule: There is a large class of reinforcement-learning
model, but the simplest and most often used
V ðt Þ B (2) one is temporal difference learning (TDRL).
C 572 Categorical Decisions
As shown in Eq. 3, the update rule for the value leads to larger reward (e.g., selecting the option
V(T) of each alternative after trial T has two with the highest value, choice = argmax(Vi))
terms. The first is an update term that depends and balancing exploitation with exploration
only on the previous outcome history. It is using a probabilistic decision rule (e.g., selecting
a temporal discount, proportional (with an option with a probability given by a softmax
a constant g) to the value of the previous trial rule where PChoose R = eVR/b/(eVR/b + eVL/b),
V(T 1). The second term is the prediction and b controls the trade-off between exploration
error d, which is proportional (with a constant and exploitation, or by a relative value rule where
a) to the difference between the outcome r(T) and Pchoice= R = (VR/(VR + VL))) (Lee et al. 2012).
the expected outcome computed from the out- TD learning:
come history V(T 1). Once the value of each
option is computed, the choice is made using V ðT Þ ¼ gVðT 1 þ adðT (3)
a decision rule that optimizes the behavior
according to the task contingencies. The decision dðT Þ ¼ rðT VðT 1 (4)
rule often represents a trade-off between
exploiting the currently available evidence that Decision rule:
e =b
VR
VR
PChoice¼R ¼ maxðV R , V L Þ or or
e
VR =
b
VL =
þe b VR þ VL
Despite its simplicity, this model is able to These neural networks implement dynamical
predict trial-to-trial variation in behavior in systems on which sensory stimuli act as perturba-
value-based decision-making tasks. For instance, tions. Depending on their strength, these perturba-
an implementation of TDRL with the relative tions can push the network toward different
value decision rule (PChoose R = (VR/(VR + attractor states leading to a categorical decision
VL))) can account for the observation that the (Fig. 2).
ratio of responses matches the ratio of inferred
stimulus values (Sugrue et al. 2004). Importantly, Open Questions and Future Directions
fitting these models to behavior has enabled Decisions are, given the available evidence, not
researchers to extract decision variables that can always perfect. Understanding the sources of
be mapped onto neural activity (Lee et al. 2012). error, the sources of noise, and uncertainty in
For example, activity in the dopaminergic system the behavior will be essential to understanding
correlates with the prediction error signal how decisions are computed in the brain. In the
(Schultz et al. 1997), and activity in the lateral two frameworks presented above, noise has dis-
intraparietal cortex correlates with stimulus value tinct origins. In perceptual decision-making, the
in a matching task (Platt and Glimcher 1999; sensory perception itself can be noisy. In value-
Sugrue et al. 2004). based decision-making, the external world envi-
These models describe the computations ronment can be stochastic, or the internal model
performed during the decision-making process of the task contingencies may be incorrect (Beck
but not their implementation at the neuronal level. et al. 2012), leading to suboptimal behavior.
Attractor networks, neural networks composed of Stochastic behavior, however, can also be advan-
spiking neurons connected through synapses, have tageous in competitive environments so that com-
been proposed as a biologically plausible imple- petitors are not able to predict choice behavior.
mentation of these computations (Wang 2012). Understanding the contribution of different
Categorical Decisions 573 C
a Stimulus Evidence Integration Decision rule
Evidence
time
V(t) > B Choice
Evidence
time
C
b Outcome history Integration Value Decision rule
Outcomes ValueR
trials
Pchoice = f(VL,VR) Choice
Outcomes ValueL
trials
Categorical Decisions, Fig. 2 (a) Stimulus: the stimu- representation of the outcome history. Integration: the
lus is being sampled. Evidence: the relevant parameters outcome history is integrated over past trials to obtain
from the stimulus are extracted to obtain the momentary a decision variable. Value computation: the value of
evidence at a given time. Integration: the evidence is each alternative is computed using the integrated outcome
integrated over time to obtain the decision variable. Race history. Decision rule: the probability of making a given
to bound: once the decision variable reaches choice is a function of the computed values that depends
a predetermined bound the corresponding action is on the decision rule used
selected. (b) Outcome history: the subject has an internal
sources of noise to decisions is an exciting direc- References

tion in the field.
Although reaction time is often used as a crit- Beck JM, Ma WJ, Pitkow X, Latham PE, Pouget A (2012)
Not noisy, just wrong: the role of suboptimal inference
ical constraint for models of decision-making,
in behavioral variability. Neuron 74:30–39
there are multiple processes contributing to it, Bogacz R, Brown E, Moehlis J, Holmes P, Cohen JD
including non-decision time, such as the time (2006) The physics of optimal decision making:
required for movement planning. Clever behav- a formal analysis of models of performance in
two-alternative forced choice tasks. Psychol Rev
ioral manipulations can be used to separate dif-
113:700–765
ferent components of reaction time and reveal Brunton BW, Botvinick MM, Brody CD (2013) Rats and
that simple discriminations can unfold in as little humans can optimally accumulate evidence for
as 30 ms (Stanford et al. 2010). decision-making. Science 340:95–98
Drugowitsch J, Moreno-Bote R, Churchland AK, Shadlen
The goal of models is to provide trial-by-trial
MN, Pouget A (2012) The cost of accumulating evi-
accounts for behavioral outputs rather than just dence in perceptual decision making. J Neurosci
fitting averages and observed distributions of 32:3612–3628
decision variables. Models, however, may be Gold JI, Shadlen MN (2007) The neural basis of decision
making. Annu Rev Neurosci 30:535–574
underconstrained when only using categorical
Kepecs A, Mainen ZF (2012) A computational framework
choices as behavioral readouts of graded deci- for the study of confidence in humans and animals.
sion variables. Additional variables, such as Philos Trans R Soc Lond B Biol Sci 367:1322–1337
reaction time (Gold and Shadlen 2007), confi- Lee D, Seo H, Jung MW (2012) Neural basis of reinforce-
ment learning and decision making. Annu Rev
dence report (Kepecs and Mainen 2012), or
changes of mind (Resulaj et al. 2009), can be Newsome WT, Britten KH, Movshon JA (1989) Neuronal
used to further constrain models. Using these correlates of a perceptual decision. Nature 341:52–54
variables can improve trial-by-trial fits, identify Platt ML, Glimcher PW (1999) Neural correlates of deci-
sion variables in parietal cortex. Nature 400:233–238
different sources of noise, and help us under-
Resulaj A, Kiani R, Wolpert DM, Shadlen MN
stand the computational processes supporting (2009) Changes of mind in decision-making. Nature
decision-making. 461:263–266
C 574 Cav1.1–4 “Long-Lasting” Type (L Type)
Schultz W, Dayan P, Montague RR (1997) A neural sub-

strate of prediction and reward. Science Cell Centered Database
275:1593–1599
Sugrue LP, Corrado GS, Newsome WT (2004) Matching
behavior and the representation of value in the parietal Maryann E. Martone
cortex. Science 304:1782–1786 Department of Neuroscience, University of
Stanford TR, Shankar S, Massoglia DP, Costello MG, California, San Diego, La Jolla, CA, USA
Salinas E (2010) Perceptual decision making in less
than 30 milliseconds. Nat Neurosci 13:379–385
Wang XJ (2012) Neural dynamics and circuit mechanisms
of decision-making. Curr. Opin Neurobiol 22:1–8 Synonyms
CCDB
Cav1.1–4 “Long-Lasting” Type

(L Type) Definition
▶ High-Voltage-Activated Calcium Channels The Cell Centered Database (CCDB; http://ccdb.

ucsd.edu) project was launched in 2002 to pro-
vide online repository of high-resolution 3D light
and electron microscopic reconstructions of cells
and subcellular structures (Martone et al. 2002,
Cav2.1 “Purkinje” Type (P/Q Type) 2003, 2008). The CCDB was designed to com-
plement the structural databases of gene, protein,
▶ High-Voltage-Activated Calcium Channels and tissue structure by hosting data in the dimen-
sional range known as the “mesoscale,” roughly
encompassing the structures that sit between
gross morphology and molecular structure, e.g.,
Cav2.2 “Non-L Type” (N Type) cellular networks, cellular and subcellular
microdomains along with their macromolecular
▶ High-Voltage-Activated Calcium Channels constituents. The study of mesoscale structures
continues to present a challenge to experimental-
ists, because to build a comprehensive under-
standing of complex tissues in this dimensional
Cav2.3 “Residual” Type (R Type) range requires the ability to aggregate data
obtained by multiple researchers across tech-
▶ High-Voltage-Activated Calcium Channels niques and spatial scales.
Cav3 The types of imaging data stored in the CCDB are
quite heterogeneous, ranging from large-scale
maps of protein distributions taken by
confocal (Orloff et al. 2012) microscopy to 3D
reconstruction of individual cells, subcellular
structures, and organelles reconstructed using
CCDB electron tomography, time series of subcellular
dynamics, and a growing collection of serial
▶ Cell Centered Database block face electron microscopic datasets
Cell Centered Database 575 C
(e.g., http://ccdb.ucsd.edu/sand/main?mpid= additional images of cells and tissues. The CIL
8244&event=displaySum). While many of the provides an easy to use graphical interface, where
datasets in the current CCDB derive from the users can browse contents by clicking on various
nervous system, reflecting our own area of organelle systems in a graphic of a eukaryotic and
expertise, the schema of the CCDB is essen- prokaryotic cell. Datasets in the CIL are anno-
tially agnostic with regard to cell type. tated with terms from 14 different ontologies in
From its inception, the CCDB envisioned that 16 different fields as well as a basic description C
it would be part of a series of distributed data- and technical details. CCDB is also registered
bases that could be cross-queried and navigated with the Neuroscience Information Framework
and has worked actively to integrate with other (http://neuinfo.org) as part of its extensive data
relevant resources. In 2012, the CCDB joined federation. Currently, the CCDB and CIL are
with the Cell Image Library (CIL: http:// maintained as separate resources, but we are work-
cellimagelibrary.org), developed by the Ameri- ing to merge the front and back ends to provide
can Society of Cell Biology. All public CCDB a unified resource for cellular imaging data.
data is now available directly through the CIL In addition to the public CCDB, the CCDB
and through CIL’s cell-centered interface also maintains a private space available via the
(Fig. 1), which additionally hosts over 10,000 CCDB portal (http://ccdb-portal.crbs.ucsd.edu/).
Cell Centered Database, Fig. 1 The CCDB may be Neuroscience Information Framework (bottom). Through
accessed through its home page (upper left), via the Cell CIL and NIF, CCDB data is linked to other relevant data at
Image Library interface (middle panel) or via the different scales and systems
C 576 Cell Centered Database
Private space is used for the data management CCDB data model is the “microscopy product”
pipeline for the National Center for Microscopy table. The microscopy product refers to a set
and Imaging Research (http://ncmir.ucsd.edu). of related 2D images taken by light
Private space is also used for registration of (epifluorescence, transmitted light, confocal, or
datasets prior to publication. Once data are multiphoton) or electron microscopy (conven-
approved, they are released to the public site. tional or high-voltage transmission electron
Although the predominant data type housed microscopy). For each type, the images compris-
within the CCDB is imaging data, the CCDB ing a single microscopy product are systemati-
was not designed to be a simple image repository, cally related to each other along a single
that is, the data model is not designed for the dimension. For example, a tilt series contains
storage of individual images. Rather, it is images that are related to each other through
designed around the process of 2D, 3D, and 4D a tilt angle; a mosaic contains images that are
reconstruction from light or electron micro- related through X, Y coordinates, and optical
graphs, capturing key steps in the process from section, through focus and serial section series
experiment to analysis. At the center of the contain images that are related through Z depth.
Cell Centered Database, Fig. 2 Organization of search microscopy product. In this case, the 2D dataset, 3D
results through the CCDB interface. Users are shown reconstruction, and segmentation are all available
thumbnails for the types of data available for each
CellML 577 C
A given set of data may be more than one prod- for use during the review process. Data will be
uct, for example, it is possible for a set of images released to the public upon publication.
to be both a mosaic and a tilt series.
Linked to each microscopy product are any
reconstructions and analysis products, e.g., seg- References
mentations and models, created from either the
raw data or the reconstruction. The CCDB makes Martone ME, Gupta A, Wong M, Qian X, Sosinsky G, C
Ludaescher B, Ellisman MH (2002) A cell centered
the 2D image datasets, reconstructions, and seg-
database for electron tomographic data. J Struct Biol
mentations available for viewing and download 138:145–155
(Fig. 2). Because each project may have different Martone ME, Zhang S, Gupta A, Qian X, He H, Price DL,
aims, the CCDB does not require that all types Wong M, Santini S, Ellisman MH (2003) The cell-
centered database: a database for multiscale structural
of products be available for each dataset.
and protein localization data from light and electron
The microscopy product identifier serves as the microscopy. Neuroinformatics 1(4):379–395
accession number for the CCDB. This value Martone ME, Tran J, Wong WW, Sargis J, Fong L, Larson
was chosen because a given dataset may only be S, Lamont SP, Gupta A, Ellisman MH (2008) The cell
centered database project: an update on building com-
taken once, and therefore, each microscopy prod-
munity resources for managing and sharing 3D imag-
uct represents a unique dataset. In contrast, mul- ing data. J Struct Biol 161(3):220–231, Epub 16 Oct
tiple reconstructions may be derived from 2007, PMID: 18054501
a single microscopy product, and multiple seg- Orloff DN1, Iwasa JH, Martone ME, Ellisman MH, Kane
CM. The cell: an image library-CCDB: a curated
mentations may be derived from a single
repository of microscopy data. Nucleic Acids Res.
reconstruction. 2013 Jan;41(Database issue):D1241-50. doi: 10.1093/
The CCDB was designed with an eye towards nar/gks1257. Epub 29 Nov 2012
encouraging reanalysis and reuse of mesoscale
data and for mining the content of these datasets.
Towards this end, the CCDB provided not only
the final data product, usually a 3D reconstruc- CellML
tion, but the raw data, specimen preparation
details, the imaging parameters, and any derived Randall D. Britten
data products created as a result of analysis of the Auckland Bioengineering Institute, University of
reconstruction. Thus, the schema of the CCDB Auckland, Auckland, New Zealand
ensures that researchers can trace the provenance
of a piece of data and understand the specimen
preparation and imaging conditions that led to it Definition
while making raw and derived data available for
reanalysis. CellML is a format for specifying mathematical
Data from the CCDB/CIL is available free of models, focusing on the needs of systems biol-
charge for browsing online and for download and ogy, but applicable to a broader range of fields
reuse through a CC-BY license. Users should cite that use mathematical modeling. It focuses on
the CCDB accession number and acknowledge models that are formulated as ordinary differen-
both the CCDB and the data contributor in any tial equation (ODE) systems and differential
subsequent use. CCDB/CIL encourages submis- algebraic equation (DAE) systems. Please see
sions from the biological imaging community Miller et al. 2013 for more details. (Note that
and accepts high-quality imaging data regardless the article mentions some of the tools that were
of whether they are part of a published study. available for working with CellML at the time it
Submitted data are curated and annotated by was written. Specifically, it mentions OpenCell
staff prior to release. If data are part of and COR. Since then, these two projects have
a published study, the CCDB/CIL can provide been merged into a single project called
an accession number and private review space OpenCOR.)
C 578 Center-Surround Modulations
Further Reading found in many different brain areas, where the

responses of neurons to a localized stimulus are
Miller AK, Britten RD, Lloyd CM, Nickerson DP, Nielsen strongly and often nonlinearly influenced by its
PM (2013) CellML. In: Dubitzky W, Wolkenhauer O,
context. Computationally, these processes are
Cho K, Yokota H (eds) Encyclopedia of Systems Biol-
ogy. Springer, New York, pp. 378–381 thought to be essential for constructing complex,
global representations in higher brain areas (i.e.,
the perception of “objects”) from simple, local
features forming the sensory input.
Center-Surround Modulations
▶ Center-Surround Processing, Computational
Role of
CSP can be introduced and studied from two
different perspectives:
• From neurophysiological and psychophysical
evidence: Experiments reveal a rich repertoire
Center-Surround Processing, of cases where the surround of a stimulus
Computational Role of influences the brain’s processing of the center
of a stimulus. Important examples are antago-
Udo Ernst nistic (▶ Retinal Neurophysiology) receptive
Department of Computational Neuroscience, fields (▶ Spectrotemporal Receptive Fields
Institute for Theoretical Physics, University of ▶ Receptive Field Modeling) of retinal gan-
Bremen, Bremen, Germany glion cells (▶ Retinal Neurophysiology)
which have different polarities in center and
surround (e.g., Geffen et al. 2007),
Synonyms nonclassical receptive fields in the visual cor-
tex (▶ Hierarchical Models of the Visual Sys-
Center-surround modulations; Contextual inte- tem) revealing strong nonlinearities in the
gration; Contextual processing; Nonclassical integration of center and surround (e.g., Free-
receptive fields man et al. 2001), and enhanced or reduced
visibilities of target stimuli induced by the
presence of flanking elements in psychophys-
Definition ical studies (e.g., Polat et al. 1998).
• From theory and first principles: Information
Center-surround processing (CSP) describes the integration becomes necessary if “global”
integration of localized “center” with contextual information is distributed over many “local”
(“surround”) information into a more global rep- channels such that the knowledge of informa-
resentation. It has been studied mostly as tion in one local channel is not sufficient to
a sensory computation, in particular in the early recover the “global” information in an unam-
visual system (Carandini et al. 2005). Although biguous way. CSP is a special form of informa-
center and surround are often defined in space, tion integration, with the additional
CSP is a more general processing in the sense that requirements that there must be some form of
it is also performed in other physical dimensions topology (in order to define center and sur-
such as in the auditory domain or in time round) and that there is a limited range over
(Schwartz et al. 2007). In neuroscientific con- which information can be integrated. This per-
texts, CSP is defined rather mechanistically as spective on CSP is explored in normative
the difference between the processing of local- approaches (▶ Bayesian Approaches in Com-
ized and extended stimuli. Signatures for CSP are putational Neuroscience: Overview ▶ Cortical
Center-Surround Processing, Computational Role of 579 C
Function, Normative Models of ▶ Visual Illu- regions with homogeneous content, as, e.g., uni-
sions, Models of), e.g., probabilistic modeling, formly painted walls or open sky. Edges or any
with strong conceptual cross-links to Gestalt other discontinuities, however, are enhanced by
psychology (▶ Visual Illusions, Models of). this filter (Fig. 1c). In discrete representation, the
The definition of center and surround strongly Mexican hat acts like a Laplacian operator
varies among studies – it can be defined in terms which computes the second spatial derivative
of the stimulus components used in an experi- (curvature) of a two-dimensional function. C
ment, in terms of the synaptic input fields of • Data compression (e.g., Zhaoping 2006): If an
a neuron (or population), in terms of a cell’s image is composed of many regions with
response properties or receptive field organiza- homogeneous content (see paragraph above),
tion, or in terms of the modulation of the neuronal as is typical for natural scenes, the output of
response by stimuli presented outside the recep- a DOG filter will yield few values which are
tive field (the “nonclassical receptive field”). CSP substantially different from zero, namely, at
is not restricted to concentric arrangements, even locations with discontinuities. The image rep-
if the name suggests this specific form for center- resentation will thus become compressed or
surround (CS) interactions. sparsified (▶ Independent Component Analy-
The following paragraphs will explain various sis of Images) (Olshausen and Field 1996).
aspects of CSP within different concepts and The term sparse denotes a representation in
theoretical approaches, in an attempt to structure which only few units (or neurons) are active,
research on CSP. Note that these aspects are while all other units are (almost) silent (e.g.,
not isolated from each other but have mutual Rozell et al. 2008).
cross-links. • Decorrelation (spatial and temporal): In most
natural image ensembles the light intensities
From Image Processing to Object of neighboring locations are highly correlated,
Recognition even over long distances. DOG filters reduce
In (applied and neural) image processing, CSP these spatial correlations as well as temporal
plays a fundamental role in both preprocessing variations in light intensities which are shared
and in higher-order computations performed on by a region of pixels in an image.
a visual scene. The simplest CS computations can • Whitening (e.g., Graham et al. 2006): Natural
be expressed as linear spatial filters applied to an images typically have power spectra
input. For example, a Mexican hat (▶ Center- exhibiting power law characteristics. DOG fil-
Surround Processing, Network Models of ters with increasing amplitudes for higher spa-
▶ Flicker-Induced Phosphenes) filter, which can tial frequencies can be used to decompose the
be realized as a Difference-of-Gaussians (DOG) image into spatial frequencies and then to
(▶ Center-Surround Processing, Network selectively enhance higher-frequency image
Models of ▶ Flicker-Induced Phosphenes) content to achieve a flat power spectrum
M (Fig. 1a) defined as whitening (▶ Independent Component Analy-
sis of Images), Fig. 1c).
More complex operations are achieved by
jxj2 jxj2
MðxÞ ¼ Aþ exp 2 A exp 2 nonlinear CSP. Robust image processing often
2sþ 2s requires computations to be invariant against
changes in illumination, spatial scale, contrast, or
can be used for edge enhancement, data com- other properties of a scene which are unimportant
pression, decorrelation (▶ Sensory Coding, for a certain (recognition) task. One example is
Efficiency), and whitening (▶ Independent divisive normalization (▶ Color Vision, Compu-
Component Analysis of Images): tational Methods for ▶ Computation with Popula-
• Edge enhancement: Mexican hat with zero inte- tion Codes) or inhibition where the output of a unit
gral, applied to images, completely suppress processing the “center” is divided by the summed
C 580 Center-Surround Processing, Computational Role of
Center-Surround Processing, Computational Role Conversely, stimulating with a bright spot in the surround
of, Fig. 1 From image processing to object recogni- leads to a decreased firing rate in the ON center and to an
tion. (a) Mexican hat function (difference of Gaussians) increased firing rate in the OFF-center cell (Kretzberg and
which is positive in the center (red) and negative in the Ernst 2013) (Textbook Neurosciences, Springer 2013,
surround (blue). (b) ON- and OFF-center receptive fields Box 18.2a). (c) Comparison between a whitened image
of retinal ganglion cells. Stimulating with a bright spot in (right) to the original image (left). The whitened image
the RF’s center leads to an increased firing rate for the ON contains far less pixel values that are different from the
center and to a decreased firing rate in the OFF-center cell. background level than the original image (Olshausen and
output of units processing the surround. Divisive “proximity” (Fig. 2a). Gestalt laws are linked to
normalization plays an important role for achiev- CSP by the fact that the surround of a feature
ing contrast-invariance of a preceding filter oper- (in the above example, the proximity of neighbor-
ation, which is, e.g., realized in neural responses to ing elements) determines how the central part of
visual input patterns. a stimulus is perceived. Also in Gestalt psychol-
Most applied image-processing algorithms ogy, surround refers not only to the dimension
make heavy use of CSP in a hierarchical manner. space but also to other dimensions such as C
Algorithms inspired by properties of the visual time: Elements having the same motion vector
system include HMAX (Riesenhuber and Poggio (i.e., a “common fate”) are also likely to be
1999), SIFT, and HOG (Dalal and Triggs 2005). grouped together. In the literature, such grouping
As an example, the Scale Invariant Feature processes often run under the terms feature inte-
Transform (SIFT) (▶ Color Vision, Computa- gration or binding (see also Herzog 2009).
tional Methods for) algorithm (Fig. 1d) builds While Gestalt psychologists defined criteria of
on a sequence, or chain, of center-surround oper- feature integration mostly in an empirical way,
ations (Lowe 2004; Lindeberg 2012). Parts of this more recent work strives to ground Gestalt prin-
procedure are (a) DOG filtering on different spa- ciples in natural scene statistics which then
tial scales to find the so-called keypoint candi- implies specific forms of CSP (Geisler
dates (i.e., characteristic points of interest in et al. 2001; Sigman et al. 2001; Simoncelli and
a scene), then (b) detection of corners by inspec- Olshausen 2001). Central idea is that the laws
tion of the keypoint candidates’ surrounding according to which the surround of a local
regions, and finally (c) the generation of “descrip- image patch influences its processing shall be
tors” by computing the differences between the determined by the statistical properties of objects
local orientation of the image patch at a keypoint in natural scenes (if object recognition is the
and the orientations of the surrounding patches. ultimate goal of visual processing). A good
These CS operations realize the computational example for the close match between Gestalt
principles of (a) scale-invariance, (b) selection processing and natural image statistics is contour
of salient features, and (c) orientation-/rotation- integration ▶ Somatosensory Cortex: Neural
invariance, respectively. Coding of Shape. Contour integration refers to
the (cognitive) process which links oriented line
Feature Integration and Gestalt Psychology segments that are aligned in colinear and
In the 1920s, a group of psychologists around cocircular way (thus obeying the Gestalt law of
Max Wertheimer (*1880, +1943) expressed “good continuation”) (Fig. 2b). Computationally,
a set of principles (Todorovic 2007) according contour integration can be performed using an
to which basic image features (e.g., dots, edges, association field that quantifies which particular
lines) are grouped by our visual system to facili- configurations of two oriented edges are to be
tate higher-order cognitive functions such as linked together (Field et al. 1993). Association
figure-ground segregation or object recognition. fields extracted from psychophysical experi-
For example, if dots are arranged on a regular ments closely match the pairwise edge statistics
grid, and if the horizontal (vertical) distance computed from human-labeled natural scenes
between these dots is slightly larger than (Fig. 2c) (Geisler et al. 2001).
the vertical (horizontal) distance, the dots are Models emerging from the feature integration
perceived as arranged in columns (rows). perspective on CSP are predominantly “bottom-
This phenomenon refers to the Gestalt law of up,” seeking to implement empirically observed
ä
Center-Surround Processing, Computational Role descriptors for a visual scene (Wikipedia, accessed 6 Oct
of, Fig. 1 (continued) Field 1996) (http://redwood.berke- 2013, http://en.wikipedia.org/wiki/Scale-invariant_fea-
ley.edu/bruno/sparsenet/). (d) Example for image spots ture_transform. The image was published under the Cre-
selected by the SIFT algorithm as suitable, characteristic ative Commons License)
principles of CSP in suitable neural architectures

(Li 1998), but see also Ernst et al. (2012) for
a normative approach.
Normative and Probabilistic Approaches

A different perspective on CSP emerges from
normative frameworks for understanding brain
function. They predict properties of CSP from
general principles for neural computation and
cognition. A prominent example for a normative
framework is the hypothesis that the brain holds
an internal model of the outside world (von
Helmholtz 1856) which is constantly matched to
the signals received by our sensory system. Such
an internal model could consist of the statistical
properties of objects and their composition in nat-
ural scenes which are used to infer the
presence (or absence) of these objects in a stream
of action potentials (▶ Hodgkin-Huxley Model
▶ Excitability: Types I, II, and III) from the retina.
In the framework of ▶ Bayesian Inference with
Spiking Neurons, CSP emerges as a competition
between neurons with input fields that overlap in
their surrounds (Fig. 3a): Both neurons would
respond equally when only the optimally matching
center stimuli were presented alone, giving the
same evidence for two potential explanations
(Fig. 3a, left). The additional appearance of a sur-
rounding stimulus would disambiguate the situation
by giving more evidence for the first of the two
explanations (Fig. 3a, right). In consequence, the
Center-Surround Processing, Computational Role
of, Fig. 2 Feature integration and Gestalt psychology. responses will become modulated accordingly – the
(a) Dot pattern visualizing the Gestalt law of proximity. In second explanation is made less probable and thus
the left pattern, the horizontal spacing of the dot grid is “explained away,” and the first explanation
smaller than the vertical spacing. In consequence, the dots becomes more likely. This effect can account for a
are perceived as being grouped in horizontal lines. The
effect reverses in the right pattern, where spacing in the range of experimentally observed CS modulations.
vertical dimension is smaller. (b) Stimulus used to study While the objective of Bayesian frameworks
contour integration (Ernst et al. 2012): in a background of is to perform statistical inference as best as pos-
randomly oriented edge elements (Gabor patches), sible, other objectives as, e.g., the goal to obtain
a horizontal contour of colinearly and cocircularly aligned
edges has been embedded. Superimposed is an association a “sparse” representation also imply specific
field, which visualizes the typical interactions needed to forms of CSP. In neurophysiological contexts,
integrate the local edge elements into the global percept of
a contour. (c) Association field computed from natural ä
image statistics. This association field was derived from
the pairwise edge statistics of contours labeled by human Center-Surround Processing, Computational Role
observers as belonging to the same object. The color of, Fig. 2 (continued) to the same contour, versus the
denotes the likelihood ratio of having an edge with the same configuration occurring in the plain edge statistics
corresponding orientation and distance to a horizontal (Geisler et al. 2006) (Geisler, Super, Perry and Gallogly,
edge in the center of the association field which belongs Vision Research 41 (2001), pp 711, Figure 3f)
a
evidence
C
p p
explanation
''TAG'' ''GUT'' ''TAG'' ''GUT''

small aperture
b CRF
0.2 0.4 0.6

0 20 40 60 80
PSTH <Oti >

Response
trial
0.0
20 40 60 80 100
4xCRF large aperture
0.0 0.2 0.4 0.6

0 20 40 60 80
Response
PSTH <Oti >

trial
20 40 60 80 100
time (s)
Center-Surround Processing, Computational Role thus decreasing its likelihood. (b) While neurons fire with
of, Fig. 3 Normative and probabilistic approaches. high rates and more regular when stimulated locally
(a) Contextual information integration and “explaining within their classical receptive field (top, left and right),
away.” In this simplified scheme, a generative model is their activity becomes more sparse and more irregular for
shown with two possible causes (explanations) for a (natural) stimulation which extends beyond their cRF
a stimulus (observation): the presence of the words “tag” (bottom, left and right), thus resembling more the typical
and “gut,” each having three letters. Given the clear evi- activation patterns observed in vivo in a natural environ-
dence of only one letter “g,” these two explanations are ment. This experimental observation (left, Vinje and Gal-
equally likely (left). The additional, contextual evidence lant 2000) (Vinje and Gallant, Science 18 (2000), 1273ff,
of the letter “a,” however, disambiguates the situation and Figure 1b/c) has recently been explained by “explaining
increases the likelihood for the presence of “tag.” At the away” in a normative model for contextual processing
same time, this increased likelihood for “tag” explains (right, Lochmann et al. 2012) (Lochmann, Ernst and
away the presence of the letter “g” as evidence for “gut,” Deneve, J. Neurosci 32 (2012), Figure 4d)
a sparse representation is beneficial to reduce the of objects efficient coding (▶ Efficient Popula-
number of action potentials transmitted and thus tion Coding) (Fig. 3b). Interestingly, requiring
energy consumption. On a conceptual level, spar- a sparse representation and performing optimal
sity is a suitable constraint for situations in which inference lead to similar dynamics in CSP
a scene that is represented by a flood of signals on (Lochmann et al. 2012; Zhu and Rozell 2013).
the sensory levels can be equally well described Taken together, one of the major factors shap-
by the presence of typically only a small number ing CSP in normative approaches is competition.
Center-Surround Processing, Computational Role idealized response pattern for an orientation-contrast sen-
of, Fig. 4 Neurophysiological and psychophysical evi- sitive cell as in Sillito et al. 1995, with the orientation of
dence for CSP. (a) Response of a V1 cell to circular the center on the horizontal axis, and the orientation of the
gratings of increasing diameter, for low stimulus contrasts surround annulus on the vertical axis. Highest firing rates
(blue) and high contrasts (red). Surround suppression thus are observed if center and surround are orthogonal (red
depends on contrast, with the preferred stimulus size being lines), lowest firing rates if center and surround have same
larger for low contrasts (Shushruth et al. 2009) (Textbook orientations (blue line). (c) Contextual influences from the
Neurosciences, Springer 2013, Box 18.7c). (b) Typical, surround are seldom homogeneous, but can be divided
Competition is realized by different mechanisms response properties which are not explained by
as, e.g., divisive and subtractive inhibition and this simple model – in particular, modulations of
unspecific and input-targeted inhibition (e.g., the neural response to stimulus components
Carandini and Heeger 1994), and leads not only inside the cRF by stimulus components outside
to suppression but also to enhancement of neural the cRF, which would give no response when
responses. Together with the statistical properties presented alone (e.g., Sillito et al. 1995; Levitt
of the elementary building blocks of natural and Lund 1997; Cavanaugh et al. 2002; Freeman C
scenes, normative frameworks currently provide et al. 2001; Series et al. 2003).
a comprehensive account for CSP as performed in cRFs of neurons often reveal a concentric orga-
the early stages of the visual system (Srinivasan nization of their spatial receptive fields, in particu-
1982; Rao and Ballard 1999; Spratling 2010; lar neurons of the early visual system. A prominent
Lochmann et al. 2012; Zhu and Rozell 2013). example is retinal ganglion cells or neurons from
the lateral geniculate nucleus (Fig. 1b). They are
Neurophysiological and Psychophysical organized either in an ON-OFF or in an OFF-ON
Evidence for CSP (antagonistic) layout, meaning that these cells
(a) Physiological Correlates respond preferentially to a bright spot on a dark
The response characteristics of neurons are often background or to a dark spot on a bright back-
described in terms of their receptive fields (RFs) ground, respectively. Mathematically, such cRFs
(▶ Spectrotemporal Receptive Fields ▶ Recep- can be described with Mexican hat functions (see
tive Field Modeling). One distinguishes above). Antagonistic cRFs are not limited to
between the “classical” and “extra-classical” vision, e.g., also neurons in the auditory system
(or “nonclassical”) part of a RF (cRF and possess RFs in frequency space which resemble a
ncRF). Heuristically, the cRF describes the neu- one-dimensional cross section of a Mexican hat
ron’s response to localized and simple stimuli, (Machens et al. 2004; Theunissen et al. 2000).
while the ncRF characterizes the response char- ncRFs come in many different varieties, and
acteristics for extended and complex stimuli. although they have extensively been studied in
Technically, the cRF is obtained by using stimu- the visual system, there is yet no coherent empir-
lation with sparse or dense noise stimuli and ical classification scheme. The following exam-
applying reverse correlation (▶ Spike Triggered ples illustrate some prominent effects that were
Average) methods to compute the RF of a cell confirmed by different experimental studies, and
under a linearity assumption (i.e., that the for which normative (see above) or bottom-up
response to arbitrary combinations of localized modeling can provide a unifying account:
stimuli is the sum of responses to these stimuli • Surround suppression (Fig. 4a): If a stimulus
presented in isolation). In essence, ncRFs are localized in feature space is made larger, the
Center-Surround Processing, Computational Role corresponds to 500 mm in the cortex. The dendrites of
of, Fig. 4 (continued) into “near surround” and “far sur- the cell are shown in black, and its axon in white (This
round.” On the left, model scheme for explaining the figure is courtesy of Dr. Zóltan Kisvarday and Ms. Kitti
origin of differences between near- and far-surround inter- Csorba) (Textbook Neurosciences, Springer 2013, Box
actions. On the right, responses of a V1 cell to a stimulus 18.8). (e) Example for crowding: the letters “H” have the
composed of a center grating and an annulus of same same size and visual eccentricity from the red fixation
(preferred) orientation with decreasing inner radius, spot. However, the right “H” is more difficult to perceive
eliciting far-surround facilitation from about 8-deg. annu- because of the presence of contextual information, i.e., the
lus width (Ichida, Schwabe, Bressloff, Angelucci, other letters. (f) Famous visual illusion, demonstrating the
J. Neurophysiol 98, 2168ff (2007), Figures 1b and 2a contextual influence of the lines in the background,
right). (d) Orientation preference map and layout of hor- suggesting depth in space, on size perception of the
izontal connections of a layer II pyramidal cell in cat A17. figures – although being of the same physical size, the
Preferred stimulus orientation is coded according to the right figure is perceived as being bigger than the left one
color palette in the lower right corner; the black bar
neural response initially grows, but then CSP is at least partly mediated by the topology
becomes more and more suppressed (e.g., of anatomical connections between neurons
Shushruth et al. 2009; Solomon et al. 2006). within and between cortical areas. Typically,
Typically, suppression is most pronounced axons or dendrites extend radially from neural
when the stimulus extends beyond the cRF populations and have different ranges for excit-
borders. If the stimulus consists of oriented atory and inhibitory connections. This property
features (i.e., gratings, bars), surround sup- naturally causes antagonistic response profiles in
pression is the strongest if these features are the afferent cells. Furthermore, connections were
similar and the weakest or even opposite if shown to be response specific: For example,
these features are different. This effect might excitatory long-ranging horizontal axons in the
also depend on stimulus strength (or contrast). primary visual cortex preferentially connect neu-
Changing the contrast can reverse the polarity rons with similar cRFs (“neurons that fire
of the modulation. together, wire together”) (Fig. 4d). Depending
• Contrast enhancement (Fig. 4b): If the stimulus on the effective sign of these interactions, they
consists of components that are different (or might mediate surround suppression (see above),
orthogonal) in the surround, neural responses but also enhancement of colinear stimulus con-
often become enhanced or disinhibited (e.g., figurations as would be necessary for integrating
Sillito et al. 1995). Computationally, contrast contours.
enhancement is similar to edge enhancement (c) Psychophysical Evidence
by DOG filtering, but extends to more complex Signatures of CSP are extensively studied with
feature combinations such as oriented gratings psychophysical methods (see paragraph above:
(corner detection). Feature Integration and Gestalt Psychology). In
• Sparsity of neural representation for natural the corresponding literature, these effects figure
scenes (Fig. 3c): Signatures of CSP are also under various terms such as masking (the percep-
revealed when stimulating visual cortical neu- tion/visibility of localized stimuli depends on
rons with natural scenes. Responses are strong their spatial and temporal context), crowding
and sustained when neurons are stimulated (detection and identification of localized stimuli
with natural scenes seen through a circular depends on the nature and density of flanking
aperture with a diameter matching their cR- elements) (Fig. 4e), popout (a much higher
F. However, when the aperture is widened, the saliency of certain feature combinations which
responses become weaker, more transient, and are embedded in surrounding elements), fill-in
sparse, thus better matching the neural dynam- (complementing missing sensory information by
ics observed in vivo (Vinje and Gallant 2000). the brain), and others. Many visual illusions yield
• Near and far surround (Fig. 4c): In the visual good examples for CSP (Bach 2013). At a first
cortex, the surround is further subdivided into glance, they might be regarded as glitches of the
a “near surround” and a “far surround.” These visual system, but indeed are signatures that the
can have differential influence on firing rates, brain does “its best” to make sense of artificially
e.g., the far surround exerts an enhancement of composed stimulus arrangements with a content
neural activity, while the near surround sup- that does not match normal experience or natural
presses neural responses (e.g., Schwabe image statistics (Fig. 4f).
et al. 2006; Ichida et al. 2007).
Note that this list is not complete and that there Cross-References
typically is a large amount of variability in ncRF
properties even for neurons of one type in one ▶ Bayesian Approaches in Computational
particular brain area. Neuroscience: Overview
(b) Anatomical Correlates ▶ Bayesian Inference with Spiking Neurons
▶ Center-Surround Processing, Network Models Geisler WS, Perry JS, Super BJ, Gallogly DP (2001) Edge
of co-occurrence in natural images predicts contour
grouping performance. Vision Res 41:711–724
▶ Color Vision, Computational Methods for Graham DJ, Chandler DM, Field DJ (2006) Can the theory
▶ Computation with Population Codes of “whitening” explain the center-surround properties
▶ Cortical Function, Normative Models of of retinal ganglion cell receptive fields? Vision Res
▶ Efficient Population Coding 46(18):2901–2913
Herzog M (2009) Binding Problem. In: Binder MD,
▶ Excitability: Types I, II, and III Hirokawa N, Windhorst U (eds) Encyclopedia of Neu-
C
▶ Flicker-Induced Phosphenes roscience. Springer, Berlin Heidelberg, pp 388–391
▶ Hierarchical Models of the Visual System Ichida JM, Schwabe L, Bressloff PC, Angelucci A (2007)
▶ Hodgkin-Huxley Model Response facilitation from the “suppressive” receptive
field surround of macaque V1 neurons. J Neurophysiol
▶ Independent Component Analysis of Images 98:2168–2181
▶ Receptive Field Modeling Kretzberg J, Ernst UA (2013) Vision. In: Galizia L
▶ Retinal Neurophysiology (ed) Neurosciences, 1st edn. SpringerSpektrum, Ber-
▶ Sensory Coding, Efficiency lin/Heidelberg, pp 363–407
Levitt JB, Lund JS (1997) Contrast dependence of contex-
▶ Somatosensory Cortex: Neural Coding of tual effects in primate visual cortex. Nature 387:73–76
Shape Li Z (1998) A neural model of contour integration in the
▶ Spectrotemporal Receptive Fields primary visual cortex. Neural Comput 10:903–940
▶ Spike Triggered Average Lindeberg T (2012) Scale-invariant feature transform.
Scholarpedia 7(5):10491. Accessed 6 Oct 2013
▶ Visual Illusions, Models of Lochmann T, Ernst UA, Deneve S (2012) Perceptual
inference predicts contextual modulations of sensory
responses. J Neurosci 32:4179–4195
References Lowe DG (2004) Distinctive image features from scale-
invariant keypoints. Int J Comput Vision 60(2):91–110
Bach M (2013) 106 Visual phenomena & optical Machens CK, Wehr MS, Zador AM (2004) Linearity of
illusions. http://www.michaelbach.de/ot/. Accessed cortical receptive fields measured with natural sounds.
6 Oct 2013 J Neurosci 24:1089–1100
Carandini M, Heeger DJ (1994) Summation and division Olshausen BA, Field DJ (1996) Emergence of simple-cell
by neurons in primate visual cortex. Science receptive field properties by learning a sparse code for
264:1333–1336 natural images. Nature 381:607–609
Carandini M, Demb JB, Mante V, Tolhurst DJ, Dan Y, Polat U, Mizobe K, Pettet MW, Kasamatsu T, Norcia AM
Olshausen BA, Gallant JL, Rust NC (2005) Do we (1998) Collinear stimuli regulate visual responses
know what the early visual system does? J Neurosci depending on cell’s contrast threshold. Nature
25:10577–10597 391:580–584
Cavanaugh JR, Bair W, Movshon JA (2002) Nature and Rao RP, Ballard DH (1999) Predictive coding in the
interaction of signals from the receptive field center visual cortex: a functional interpretation of some
and surround in macaque V1 neurons. J Neurophysiol extra-classical receptive-field effects. Nat Neurosci
88:2530–2546 2:79–87
Dalal N, Triggs B (2005) Histograms of oriented gradients Riesenhuber M, Poggio T (1999) Hierarchical models of
for human detection. In: CVPR 2005 (IEEE computer object recognition in cortex. Nat Neurosci 2:1019–1025
society conference on computer vision and pattern Rozell CJ, Johnson DH, Baraniuk RG, Olshausen BA
recognition), vol 1, pp 886–893 (2008) Sparse coding via thresholding and local com-
Ernst UA, Mandon S, Schinkel–Bielefeld N, Neitzel SD, petition in neural circuits. Neural Comput
Kreiter AK, Pawelzik KR (2012) Optimality of human 20:2526–2563
contour integration. PLoS Comput Biol 8(5): Schwabe L, Obermayer K, Angelucci A, Bressloff PC
e1002520. doi:10.1371/journal.pcbi.1002520 (2006) The role of feedback in shaping the extra-
Field DJ, Hayes A, Hess RF (1993) Contour integration by classical receptive field of cortical neurons:
the human visual system: evidence for a local “Asso- a recurrent network model. J Neurosci 26:9117–9129
ciation Field”. Vision Res 33(2):173–193 Schwartz O, Hsu A, Dayan P (2007) Space and time in
Freeman RR, Ohzawa I, Walker G (2001) Beyond the visual context. Nat Rev Neurosci 8:522–535
classical receptive field in the visual cortex. Prog Series P, Lorenceau J, Fregnac Y (2003) The “silent”
Brain Res 134:157–170 surround of V1 receptive fields: theory and experi-
Geffen MN, de Vries SE, Meister M (2007) Retinal gan- ments. J Physiol (Paris) 97:453–474
glion cells can rapidly change polarity from off to Shushruth S, Ichida JM, Levitt JB, Angelucci A (2009)
on. PLoS Biol 5:e65 Comparison of spatial summation properties of
C 588 Center-Surround Processing, Network Models of
neurons in macaque V1 and V2. J Neurophysiol Definition

102:21069–22083
Sigman M, Cecchi GA, Gilbert CD, Magnasco MO
(2001) On a common circle: natural scenes and Gestalt Surround modulations of neuronal responses by
rules. Proc Natl Acad Sci USA 98:1935–1940 stimuli outside the receptive field are omnipres-
Sillito AM, Grieve KL, Jones HE, Cudeiro J, Davis ent in sensory systems, and network models seek
J (1995) Visual cortical mechanisms detecting focal to explain them as an emergent network phenom-
orientation discontinuities. Nature 378:492–496
Simoncelli EP, Olshausen BA (2001) Natural image sta- enon. They are particularly useful because recur-
tistics and neural representation. Annu Rev Neurosci rent networks often behave counterintuitive, i.e.,
24:1193–1216 different from the neural dynamics assumed by
Solomon SG, Lee BB, Sun H (2006) Suppressive sur- phenomenological models.
rounds and contrast gain in magnocellular-pathway
retinal ganglion cells of macaque. J Neurosci
26:8715–8726
Spratling MW (2010) Predictive coding as a model of Detailed Description
response properties in cortical area V1. J Neurosci
30:3531–3543
Srinivasan MV, Laughlin SB, Dubs A (1982) Predictive Lateral inhibition is an example of center-
coding: a fresh view of inhibition in the retina. Proc surround (CS) processing. As a perceptual phe-
R Soc Lond B Biol Sci 216:427–459 nomenon, it was already known to Ernst Mach in
Theunissen FE, Sen K, Doupe AJ (2000) Spectral- the mid-nineteenth century (Yantis 2014). The
temporal receptive fields of nonlinear auditory neurons
obtained using natural sounds. J Neurosci so-called Mach bands are a visual illusion,
20:2315–2331 where the border between adjacent luminance-
Todorović, D (2007). W. Metzger: Laws of Seeing. defined areas appears to have higher contrast
Gestalt Theory, 28, 176–180 than it actually has. Phenomenological models
Vinje WE, Gallant JL (2000) Sparse coding and
decorrelation in primary visual cortex during natural of this perceptual effect share the common idea
vision. Science 287:1273–1276 of spatial filtering the input image (the visual
von Helmholtz H (1856) Handbook of physiological sensory signal) with a symmetric center-surround
optics. Leopold Voss, Leipzig kernel (Pessoa 1996). Such a filter-based descrip-
Zhaoping L (2006) Theoretical understanding of the early
visual processes by data compression and data selection is agnostic regarding the actual neural mech-
tion. Network 17:301–334 anism, but network models with short-range
Zhu M, Rozell CJ (2013) Visual nonclassical receptive excitation and long-range inhibition are plausible
field effects emerge from sparse coding in a dynamical candidates for implementing them. End-stopped
system. PLoS Comp Biol 9(8):e1003191. doi:10.1371/
journal.pcbi.1003191 neurons in primary visual cortex (V1), first
described by Hubel and Wiesel in 1965 (Hubel
and Wiesel 1965), are another example, where
the spatial surround of the receptive field modu-
Center-Surround Processing, lates the strength and selectivity of orientation-
Network Models of tuned responses. Recent modeling studies
emphasize the role of strong local and balanced
Lars Schwabe intracortical feedback in surround modulation
Department of Computer Science and Electrical (e.g., Ozeki et al. 2009) and how it may interact
Engineering, Adaptive and Regenerative with inter-areal feedback connections (Shushruth
Software Systems, Universit€at Rostock, Rostock, et al. 2012) in CS processing. Strongly recurrent
Germany neural networks can exhibit counterintuitive
emergent properties that are not anticipated by
the often implicitly assumed circuitry underlying,
Synonyms e.g., filter-based or other computational
approaches to CS. The role of CS processing in
Models of end-stopped cells; Models of lateral perception could be the amplification of relevant
inhibition; Network models of contextual effects stimulus features in noisy and cluttered
Center-Surround Processing, Network Models of 589 C
environments such as figure-ground segmenta- Horizontal cells, contacted by the photoreceptors,
tion in vision. Hence, the explicit integration of inhibit neighboring photoreceptors. This explains
different pathways in network modeling CS why a full illumination (causing lateral inhibi-
processing is a promising approach to understand tion) leads to less RGC spikes than a spatially
how sensory systems extract relevant information restricted illumination (with less or no lateral
from noisy and cluttered environments. inhibition). Figure 1a illustrates the increase
in RGC spike discharges in response to C
Models of Center-Surround Processing in a visual stimulus and shows how cone
Early Vision photoreceptors, C, are connected to horizontal
The signaling cascades in the retina triggered by cells, H, and bipolar cells, B. This basic motif,
photons impinging on photoreceptors are well namely, realizing suppression of responses to
characterized. For example, light falling onto a large stimulus via inhibition of neighboring
cones leads to their hyperpolarization and cells, has been very influential and was thus
a decrease in glutamate release onto bipolar employed to explain surround suppression in the
cells. In response to this decrease, on-center bipo- cortex as well.
lar cells depolarize, which leads to an increase of Figure 1b (top) shows two image patches of
the glutamate they release onto retinal ganglion different size. Within these patches the texture
cells (RGCs). As a consequence, these RGCs are is dominated by an oblique orientation,
on-center cells that increase their spike discharge. a stimulus feature that triggers responses of
a b c
1 2 3 2
1
To RGCs and
Optic Nerve
I I I
1
B B B I I I E E E
I I I
C C C E E E 2
E E E
1 2 3
H
Light From LGN From LGN
Center-Surround Processing, Network Models of, neurons with RF center at location 2 (in the visual field)
Fig. 1 Illustration of three center-surround effects, each in primary visual cortex (V1) are suppressed when the
with a possible underlying network model. (a) The fre- size of the oriented texture is increased. The retinal
quency of spike discharges of retinal ganglion cells circuit for lateral inhibition, Fig. 1a, is not directly
(RGCs) increases when the center of the RGC’s receptive applicable, because the horizontal connections in V1
field (RF) is illuminated (white dot) and returns to the are made only by excitatory cells. Those need to contact
pre-stimulus baseline when the illumination is turned local inhibitory neurons to implement a “sign switch”
off. When the center and the surround of the RF are that realizes the lateral inhibition. (c) Interconnected
illuminated, then the spike discharge does not increase lateral inhibition circuits across different visual areas.
(not shown). A retinal circuit model explains this selec- The feedback connections from areas downstream to V1
tivity via lateral inhibition: Light falls onto the cones, C, are known to be spatially more extensive than the
which triggers lateral inhibition of neighboring cones with horizontal connections within V1 (Fig. 1b). Via such
horizontal cells, H. The on-center bipolar cells forward pathways the lateral inhibition can affect neurons with
this signal to RGCs. (b) Responses of orientation-selective very distant RF centers
C 590 Center-Surround Processing, Network Models of
orientation-selective neurons in primary visual (see the local coupling between E and I neurons
cortex (V1). Interestingly, however, is the obser- in Fig. 1b). It was shown that an increased exci-
vation that orientation-selective V1 neurons with tation of inhibitory neurons from outside the local
their receptive field (RF) center at location recurrent circuitry, such as via the long-range
2 (Fig. 1b) decrease their spike discharges when horizontal connections, ultimately leads to
the texture is increased in size. Hubel and Wiesel a decrease of both the excitatory and inhibitory
first described such a phenomenon in V1 of cats responses, which can easily be understood using
in 1965 (Hubel and Wiesel 1965). Qualitatively, phase-plane analysis (Ozeki et al. 2009). While
this is the same phenomenon as the selectivity to this study is still compatible with horizontal con-
illumination in the RF center of on-center RGCs: nections as the substrate of surround suppression
When the stimulated area in the visual field is in V1, it shows that network models need to be set
increased, the response decreases. Experimental up explicitly and analyzed rigorously to over-
studies even support another analogy to the reti- come the potentially too simplistic intuition that
nal circuit, namely, that the observed surround worked for simple circuits (e.g., Fig. 1a) but fails
suppression in V1 may be inherited partly by for strongly recurrent systems. Recently, this
surround suppression in the LGN (Webb model has been used to investigate the orientation
et al. 2005) (as the suppression of RGCs is tuning of surround suppression (Shushruth
inherited from the lateral inhibition acting et al. 2012).
between their presynaptic cells).
Network models of surround suppression in Models of Center-Surround Processing
V1 make use of the idea of lateral inhibition. In Across Multiple Cortical Areas
V1, however, the lateral connections to neighbor- For some forms of surround suppression
ing neurons (in terms of distinct neighboring RF observed in V1, the long-range horizontal con-
centers) are made only by excitatory cells. One nections may even be too short. Feedback con-
proposal has been that these lateral connections nections from downstream cortical areas into V1
target local inhibitory neurons (Somers have been proposed as the substrate of lateral
et al. 1998), which then perform a “sign switch” inhibition in cases when the stimulus in the sur-
of this excitatory and horizontally propagating round is beyond the reach of the V1 horizontal
signal (Fig. 1b). The detailed layout of the topo- connections (Schwabe et al. 2006), because this
graphic map in V1 determines how long such feedback is spatially even more extensive. More-
horizontal connections need to be in order to over, it may play a crucial role in center-surround
link neurons with different RF centers. processing in other conditions as well, in partic-
Intracortical long-range connections (up to ular during natural vision. Figure 1c shows an
2–3 mm in macaque V1) are widely believed to image of a tiger almost hidden behind grass and
carry this horizontal signal, but much shorter the spatial extent of a RF in a visual area down-
connections (up to 0.25 mm) have been proposed stream to V1 (large red circle). The larger RF
as a potential substrate as well (Wielaard and encloses a smaller RF in V1, which is located at
Sajda 2006). These shorter connections may the barely visible neck of the tiger. Surround
account for some forms of surround suppression, suppression is known to enhance the responses
but they fail to account for physiologically to borders such as inside the RF of the V1 neuron.
observed response suppression for large stimuli Due to the barely visible border, a global-to-local
in which in the surround is way beyond the reach integration of spatially adjacent cues will be ben-
of these short-range connections. eficial to enhance responses to elongated borders
The role of short-range connections received (the tiger’s neck) in cluttered scenes.
renewed interest after a modeling study (Ozeki
et al. 2009) that emphasized an apparently coun- Modeling Techniques
terintuitive property of strongly recurrent net- For most of the effects described here, mean-field
works of excitatory and inhibitory neurons models that describe the dynamics of the average
Central Vestibular Signal Processing 591 C
neural activity in a population of neurons are Wielaard J, Sajda P (2006) Extraclassical receptive field
often sufficient. Some studies employ large- phenomena & short-range connectivity in V1. Cereb
Cortex 16(11):1531–1545
scale simulations of networks of spiking model Yantis S (2014) Sensation and perception. Worth Pub-
neurons. They are often biologically more appro- lishers, New York
priate as they can account for aspects that are hard
to integrate into mean-field models such as het- Further Reading
erogeneity of parameters. Schwartz O, Hsu A, Dayan P (2007) Space and time in C
visual context. Nat Rev Neurosci 8:522–553
Seriès P, Lorenceau J, Frégnac Y (2003) The “silent”
Discussion and Future Directions surround of V1 receptive fields: theory and experi-
Given that strongly recurrent networks can have ments. J Physiol Paris 97:453–474
counterintuitive properties (Ozeki et al. 2009;
Shushruth et al. 2012), it may be important to
employ them as building blocks for larger
models, including the neural mass models used Central Vestibular Signal Processing
in brain imaging. In particular, studying how
adaptation at various time scales and via multiple Rebecca Lim, Thomas P. Wellings and
mechanisms affects the effective network con- Alan M. Brichta
nectivity and hence the resulting center-surround School of Biomedical Sciences and Pharmacy,
computations is an important task for future Hunter Medical Research Institute,
studies. Finally, we need to understand how The University of Newcastle, Callaghan, NSW,
interconnected cortical areas collaborate in Australia
center-surround computations at each stage.
Definition
References
Hubel DH, Wiesel TN (1965) Receptive fields and func-

The vestibular nuclear complex (VNC) is a major
tional architecture in two nonstriate visual areas site of central vestibular processing and tradition-
(18 and 19) of the cat. J Neurophysiol 28(2):230–289 ally consists of four major neuroanatomical
Ozeki H, Finn IM, Schaffer ES, Miller KD, Ferster regions: medial, lateral, superior, and descending
D (2009) Inhibitory stabilization of the cortical net-
work underlies visual surround suppression. Neuron
(spinal) vestibular nuclei. However, recent cellu-
62:578–592 lar and molecular studies indicate that further sub-
Pessoa L (1996) Mach bands: how many models are divisions, based on more detailed analyses, better
possible? Recent experimental findings and modeling describe functional groups within these nuclei.
attempts. Vision Res 36(19):3205–3227
Schwabe L, Obermayer K, Angelucci A, Bressloff PC
The distribution of calcium-binding proteins and
(2006) The role of feedback in shaping the extra- marker genes is redefining the VNC into specific
classical receptive field of cortical neurons: regions with identifiable functional implications.
a recurrent network model. J Neurosci
26(36):9117–9129
Shushruth S, Mangapathy P, Ichida JM, Bressloff PC,
Schwabe L, Angelucci A (2012) Strong recurrent net- Detailed Description
works compute the orientation tuning of surround
modulation in the primate primary visual cortex. The mammalian vestibular system is responsible
J Neurosci 32(1):308–321
Somers DC, Todorov EV, Siapas AG, Toth LJ, Kim DS,
for our sense of balance and can be divided into
Sur M (1998) A local circuit approach to understand- peripheral and central components (Goldberg
ing integration of long-range inputs in primary visual et al. 2012). The peripheral vestibular system is
cortex. Cereb Cortex 8(3):204–217 located in the inner ear and composed of five
Webb BS, Dhruv NT, Solomon SG, Tailby C, Lennie PJ
(2005) Early and late mechanisms of surround sup-
sense organs (three semicircular canals and two
pression in striate cortex of macaque. Neurosci otolith organs) that collectively generate signals
25(50):11666–11675 in response to rotational and translational head
C 592 Central Vestibular Signal Processing
movements. These “balance” signals are then provides a much-needed cellular basis for under-
transmitted, via the eighth cranial nerve, to various standing this pivotal central vestibular region.
targets in the central nervous system (CNS). Col- For example, one division of the VNC, the medial
lectively, these CNS targets comprise the central vestibular nucleus (MVN), can itself be further
vestibular system and are responsible not only for subdivided simply based on the individual neu-
creating our perception of balance but also for ron’s projection target. These targets include ocu-
contributing to the remarkably precise stabiliza- lar motor neurons, cerebellar neurons, and local
tion of gaze and posture, as well as general motor neurons within the ipsi- and contralateral VNCs
control (Cullen 2012; Goldberg et al. 2012). (Highstein and Holstein 2006). Further subdivi-
sions have also been described based on specific
differences in intrinsic excitability across MVN
Subdividing the Vestibular Nuclear
neurons with different axonal projections and
Complex
neurotransmitter (Takazawa et al. 2004;
Sekirnjak and du Lac 2006; Bagnall et al. 2007;
Precisely how these complex and demanding
Kolkman et al. 2011a, b; Shin et al. 2011). Still
functions are accomplished, particularly by the
other subdivisions have been demarcated by
central vestibular system, is not understood. This
immunoreactivity (ir) to functionally important
is due in large part to our rudimentary under-
calcium-binding proteins (CBPs) such as
standing of the functional neuronal groups that
calbindin (CB) and calretinin (CR). Subdivisions
constitute the central vestibular system. This lack
based on these CBPs have proved consistent
of knowledge is best illustrated by the vestibular
across a number of species including rodents,
nuclear complex (VNC), a critically important
cat, monkeys, and human (Resibois and Rogers
brainstem region that distributes incoming
1992; Rogers and Resibois 1992; Kevetter 1996;
peripheral vestibular signals to other parts of the
Baizer and Baker 2005, 2006; Baizer and
brainstem, cortex, cerebellum, and spinal cord.
Broussard 2010) strongly suggesting a conserved
Despite it being a major site for multisensory
purpose. While the function of these distinct CBP
processing, we know very little about this region
regions has yet to be determined, it has been
beyond the classical neuroanatomical studies that
suggested that regions with dense CB-ir represent
first described four major divisions or vestibular
a zone of dense cerebellar input (Baizer et al.
nuclei (Brodal and Pompeiano 1957). This lack
2013) and thus may play a role in adaptive mod-
of detailed knowledge is because it has been
ifications of reflexive eye movements (Rambold
difficult to identify, classify, and manipulate
et al. 2002). While the CR-ir region is likely to
functionally distinct cell types that are often spa-
receive direct peripheral vestibular input and pro-
tially intermingled within the VNC (Kodama
ject to other parts of the VNC and suggests
et al. 2012). In the last decade, however, immu-
a coordinating rather than command role in eye
nohistochemical and molecular studies, together
movement generation.
with electrophysiological analyses, have begun
to identify discrete subdivisions within these
individual vestibular nuclei and support the
Functional Organization of the MVN
notion of a vastly more complex functional orga-
Revealed by Combined Studies
nization that has been long suspected (Scudder
and Fuchs 1992; Bagnall et al. 2007).
Molecular studies are beginning to uncover fur-
ther complexities in the internal organization of
Analysis of the Medial Vestibular the VNC. In addition to classifying neurons into
Nucleus distinct subdivisions, molecular studies have the
added advantage of ascribing these subdivisions
Revealing the complexity of both functional particular functions. In a recent study by Kodama
organization and connectivity within the VNC et al. (2012), using single-cell expression
Central Vestibular Signal Processing 593 C
profiling of genes related to neurotransmitters and References
ion channels, they were able to demonstrate can-
didate molecular markers that subdivided MVN Bagnall MW, Stevens RJ, du Lac S (2007) Transgenic
mouse lines subdivide medial vestibular nucleus neu-
into six distinct types. By combining this approach
rons into discrete, neurochemically distinct
with anterograde and retrograde labeling tech- populations. J Neurosci 27:2318–2330
niques, they were able to determine the functional Baizer JS, Baker JF (2005) Immunoreactivity for calcium-
role of four out of the six molecularly defined binding proteins defines subregions of the vestibular C
nuclear complex of the cat. Exp Brain Res 164:78–91
classes of MVN neurons. The six types were iden-
Baizer JS, Baker JF (2006) Immunoreactivity for
tified as three classes of excitatory neurons (E1, calretinin and calbindin in the vestibular nuclear com-
E2, E3) and three classes of inhibitory neurons plex of the monkey. Exp Brain Res 172:103–113
(I1, I2, I3), each associated with a specific set Baizer JS, Broussard DM (2010) Expression of calcium-
binding proteins and nNOS in the human vestibular
of marker genes. For example, E1 were large
and precerebellar brainstem. J Comp Neurol
glutamatergic MVN neurons that uniquely 518:872–895
expressed gene markers Spp1 (secreted phospho- Baizer JS, Paolone NA, Sherwood CC, Hof PR
protein 1) and Slc17a6 (vesicular glutamate trans- (2013) Neurochemical organization of the vestibular
brainstem in the common chimpanzee (Pan troglo-
porter 2) and projected to oculomotor or abducens
dytes). Brain Struct Funct 218(6):1463–1485
motor nucleus, while glutamatergic E2 neurons Brodal A, Pompeiano O (1957) The vestibular nuclei in
expressed Crh (corticotropin-releasing hormone) cat. J Anat 91:438–454
and Slc17a7 (vesicular glutamate transporter 1) Cullen KE (2012) The vestibular system: multimodal
integration and encoding of self-motion for motor con-
and projected to the cerebellar flocculus.
trol. Trends Neurosci 35:185–196
Goldberg JM, Wilson VJ, Cullen KE (2012) The vestibu-
lar system: a sixth sense. Oxford University Press,
Future Studies of the Central Vestibular
New York
System Highstein SM, Holstein GR (2006) The anatomy of the
vestibular nuclei. Prog Brain Res 151:157–203
As proposed by Kodama et al. (2012), this strat- Kevetter GA (1996) Pattern of selected calcium-binding
proteins in the vestibular nuclear complex of two
egy of classifying neurons into functionally dis-
rodent species. J Comp Neurol 365:575–584
tinct groups could also be applied to any Kodama T, Guerrero S, Shin M, Moghadam S, Faulstich
heterogeneous neuronal population, including M, du Lac S (2012) Neuronal classification and marker
those regions associated with central vestibular gene identification via single-cell expression profiling
of brainstem vestibular neurons subserving cerebellar
system. For example, quantitative single-cell
learning. J Neurosci 32:7819–7831
transcript analyses, in situ hybridization, immu- Kolkman KE, Moghadam SH, du Lac S (2011a) Intrinsic
nohistochemistry, and electrophysiological physiology of identified neurons in the prepositus
studies could be used to identify functional orga- hypoglossi and medial vestibular nuclei. J Vestib Res
21:33–47
nization of neurons in the cerebellar and ocular
Kolkman KE, McElvain LE, du Lac S (2011b) Diverse
motor nuclei, as well as higher cortical centers precerebellar neurons share similar intrinsic excitabil-
processing vestibular information. With the ity. J Neurosci 31:16665–16674
expanded knowledge promised by these com- Rambold H, Churchland A, Selig Y, Jasmin L,
Lisberger SG (2002) Partial ablations of the flocculus
bined studies, the chances of understanding how
and ventral paraflocculus in monkeys cause linked
the central vestibular system unobtrusively deficits in smooth pursuit eye movements and
achieves the extraordinary feat of creating our adaptive modification of the VOR. J Neurophysiol
sense of balance and its attendant actions are 87:912–924
Resibois A, Rogers JH (1992) Calretinin in rat brain: an
becoming a realistic possibility.
immunohistochemical study. Neuroscience
46:101–134
Cross-References Rogers JH, Resibois A (1992) Calretinin and calbindin-
D28k in rat brain: patterns of partial co-localization.
Neuroscience 51:843–865
▶ Peripheral Vestibular Signal Processing Scudder CA, Fuchs AF (1992) Physiological and
▶ Vestibular System: Overview behavioral identification of vestibular nucleus
C 594 Cercal System
neurons mediating the horizontal vestibuloocular Detailed Description

reflex in trained rhesus monkeys. J Neurophysiol
68:244–264
Sekirnjak C, du Lac S (2006) Physiological and anatom- Overview
ical properties of mouse medial vestibular nucleus The cercal system was shown in early studies to
neurons projecting to the oculomotor nucleus. be of critical importance for oriented escape
J Neurophysiol 95:3012–3023 responses (Hoyle 1958; Camhi 1980; Boyan
Shin M, Moghadam SH, Sekirnjak C, Bagnall MW,
Kolkman KE, Jacobs R, Faulstich M, du Lac et al. 1986). However, it is a considerable over-
S (2011) Multiple types of cerebellar target neurons simplification to classify this system as an
and their circuitry in the vestibulo-ocular reflex. “escape system.” Rather, the cercal system
J Neurosci 31:10776–10786 should be thought of as functioning as a low-
Takazawa T, Saito Y, Tsuzuki K, Ozawa S (2004) Mem-
brane and firing properties of glutamatergic and frequency, near-field extension of the animal’s
GABAergic neurons in the rat medial vestibular auditory system. Air particle displacement
nucleus. J Neurophysiol 92:3106–3120 directed at the cerci can elicit at least 14 distinct
responses, including evasion, flight, offensive
reactions, scanning, freezing, and various reac-
tions during male stridulation. Responses can
Cercal System depend on the behavioral state of the animal as
well as the context of the environment (Baba and
Hiroto Ogawa1 and John P. Miller2 Shimozawa 1997; Casas and Dangles 2010). In
1
Department of Biological Science, Faculty of addition to the filiform receptors sensitive to air
Science, Hokkaido University, Sapporo, Japan particle displacement, bristle hairs, clavate hairs,
2
Department of Cell Biology and Neuroscience, and campaniform sensilla are distributed on the
Montana State University, Bozeman, MT, USA cercus. The cercal system is, therefore, also
responsible for mediating behavior related to ori-
entation relative to gravity as well as a variety of
Definition touch and chemotactic sensations via other types
of sensors on the cerci (Sakaguchi and Murphey
The cercal system is a mechanosensory system in 1983; Murphey 1985; Heusslein and Gnatzy
orthopteran insects, which mediates the detec- 1987). The subsequent detailed description will
tion, localization, and identification of air cur- be limited to the sensory-processing system for
rents surrounding animals. The receptor organ the direction and dynamics of airflow mediated
for this modality is a pair of antenna-like append- by filiform receptors.
ages called cerci at the rear of the abdomen,
covered with mechanosensory hairs like the bris- Functional Organization of the
tles on a bottlebrush (Fig. 1). Air currents in the Mechanosensory Hair Array
animal’s immediate environment move these In Acheta domesticus, the species used for most
hairs and, thereby, activate the receptor neurons studies, each cercus is approximately 1 cm long
at the base of the hairs. The cercal system is in a normal adult cricket and is covered with
implemented around an internal representation 500–750 filiform mechanosensory hairs (Palka
of stimulus direction and dynamics that demon- et al. 1977; Miller et al. 2011). These hairs
strates the essential features of neural maps found range in length from 50 mm to 1.5 mm. Each
in more complex systems, including mammalian hair is supported at its base by a viscoelastic
visual and auditory systems. Over the last several socket that enables the hair to pivot within its
decades, this system of the cricket has been used socket, rather than bending along its shaft (Keil
as a simple “model system” for investigations of and Steinbrecht 1984). Each hair’s directional
development, mechanoreceptor biomechanics, movement axis is determined by the orientation
neural maps, and neural coding (Jacobs et al. of a hinge-like structure in the socket. The 750
2008). hairs on each cercus are arrayed with their
Cercal System 595 C
Cercal System, Fig. 1 The cricket cercal system. (a) A reconstruction of a single identified giant interneuron
Gryllus bimaculatus (female). The cerci are the two (GI right 10-3) is shown in blue, and several filiform
antenna-like structures, covered with fine hairs, extending sensory afferent arbors are shown in other colors. (c)
from the rear of the abdomen. (b) Computer reconstruc- Micrograph of the GI left 10-2 and filiform afferent
tion of the outline (blue) of a terminal abdominal ganglion stained with fluorescent dye
(TAG) with several reconstructed nerve cells.
movement axes in diverse orientations within the and dynamical sensitivities that appear to be
horizontal plane. On a global scale, the direction- derived largely from the mechanical properties
ality of hairs is organized in bands along the long of the hairs themselves (Shimozawa and Kanou
axis of the cerci. There is a systematic rotation in 1984). The amplitude of the response of each
the movement directions of the hairs around the sensory receptor cell to any air current stimulus
circumference of each cercus, and there is depends upon the velocity and direction of that
a nonuniform distribution of hair densities along stimulus, and the directional tuning curves of the
the length of the cercus, with greater density receptor afferents are well described by cosine
nearer the base than near the tip (Palka et al. functions (Landolfa and Jacobs 1995; Landolfa
1977; Miller et al. 2011). This global organiza- and Miller 1995).
tion of the array insures that air currents of suffi-
cient velocity will deflect all of the filiform hairs Representation of Stimulus Direction
to some extent: each hair will be deflected from The sensory afferents of filiform sensilla project
its rest position by an amount that is proportional in an orderly array into the terminal abdominal
to the cosine of the angle between the air current ganglion to form a continuous representation of
direction and the hair’s movement axis. the direction of air currents in the horizontal
Each mechanosensory hair is innervated by plane (Bacon and Murphey 1984; Jacobs and
a single spike-generating mechanosensory recep- Theunissen 1996; Paydar et al. 1999; Jacobs and
tor neuron. These receptors display directional Theunissen 2000). The synaptic terminals from
C 596 Cercal System
Cercal System, Fig. 2 Anatomical prediction of synap- peak activation (inset). (b) Reconstructed dendrites of
tic connectivity between filiform sensory afferents and right GI 10-2 within the afferent neural map of directions.
interneurons. (a) A reconstruction of right GI 10-2 The neural map is predicted by the “statistical cloud”
shown in yellow and afferent arbors from 12 different corresponding to the density of synaptic terminals for
filiform hair receptors shown in different colors. The that stimulus direction (inset)
color of each afferent corresponds to its direction of
afferents having similar peak directional sensitiv- approximately 10 pairs of identified ascending
ities arborize in adjacent areas, and the spatial interneurons whose axons project to motor cen-
segregation between afferent arbors increases as ters in the thorax and integrative centers in the
the difference in their directional tuning brain. These ascending interneurons have been
increases. A representation of the stimulus direc- called “giant interneurons (GIs)” due to their
tion in a three-dimensional neural map of the large-diameter axons which enable very rapid
entire filiform afferents has been developed, conduction of action potentials to the higher
based on anatomical and physiological measure- ganglia. Like the afferents, these interneurons
ments taken from a large sample of individual are also sensitive to the direction and dynamics
afferents (Fig. 2; Troyer et al. 1994; Jacobs and of air current stimuli (Jacobs et al. 1986; Miller
Theunissen 1996; Paydar et al. 1999). By com- et al. 1991; Theunissen and Miller 1991;
bining the predicted responses of each class of Theunissen et al. 1996; Ogawa et al. 2004). The
afferents with this information about the spatial directional sensitivity of the ascending interneu-
location of their terminal arborizations within the rons is primarily based on the relative positions of
neural map, predictions have been made about their dendrites within the afferent map of the
the spatial patterns of synaptic activation that mechanosensory afferents (Fig. 2; Jacobs and
would result from sustained, unidirectional air Theunissen 2000). However, the overall direc-
currents (Troyer et al. 1994; Paydar et al. 1999; tional tuning and dynamical sensitivity of each
Jacobs and Theunissen 2000). The predictions individual interneuron also depends on the spatial
were consistent with spatial patterns of ensemble distribution of synapses onto the complex geom-
activity of filiform afferents visualized by cal- etry of the dendritic arbors (Fig. 3; Ogawa et al.
cium imaging experiments (Ogawa et al. 2006). 2008). The frequency sensitivity of each inter-
neuron depends on two independent factors:
Dendritic Integration in the Giant (a) its inherent frequency filtering properties
Interneurons that is a function of intrinsic membrane and syn-
The 1,500 sensory afferents synapse with a group aptic transmission characteristics and (b) the
of approximately 30 local interneurons and position of its dendrites within the afferent map.
Cercal System 597 C
Cercal System, Fig. 3 Relationship between electro- spikes (black line) and the mean amplitude of Ca2+
tonic geometry of the dendritic arbor and directional increases at three different dendrites X, Y, and Z (red,
tuning of dendritic Ca2+ responses in GI 10-3. (a) Image green, and blue lines, respectively). Each value of the
showing the compartmental model of GI 10-3 with the Ca2+- and action-potential responses was scaled to the
spike-initiation zone (SIZ) indicated. (b) Plots of inward maximal responses of each measurement. The curve
electrotonic distance (ED) versus path distance relative to shown by the orange line was predicted by calculating
the SIZ in the model of GI 10-3. The X axis is the path the weighted sum of the tuning curves of the individual
distance from the soma to dendritic compartments along dendritic responses based on electrotonic distances. Ant
each process. (c) Polar plots indicate the mean number of anterior, post posterior
Neural Encoding and Transmission approximately 8 (Theunissen and Miller 1991;

Some or perhaps all of the 20 projecting sensory Theunissen et al. 1996). More recent research has
interneurons in this system represent the general demonstrated that these same four GIs (and per-
parameters of air currents via an ensemble haps all of the other projecting INs) also respond
coarse-coding scheme. For example, four GIs in with high sensitivity to complex dynamic
a well-studied group (the left and right GIs 10-2 multidirectional features of air currents on
and 10-3) have been shown to be tuned broadly to a smaller spatial scale than the physical dimen-
the direction and velocity of bulk airflow move- sions of the cerci (Mulder-Rosi et al. 2010), using
ments and project a coarse-coded image of these a coding scheme that is more complex than
parameters to higher centers that is accurate to simple linear encoding (Aldworth et al. 2011).
C 598 Cercal System
The velocity threshold for these features is lower Jacobs GA, Miller JP, Murphey RK (1986) Cellular mech-
than the threshold for simple unidirectional bulk anisms underlying directional sensitivity of an identi-
fied sensory interneuron. J Neurosci 6:2298–2311
flow stimuli used in previous studies. Thus, in Jacobs GA, Miller JP, Aldworth Z (2008) Computational
addition to participating in the ensemble mechanisms of mechanosensory processing in the
encoding of bulk air flow stimulus characteris- cricket. J Exp Biol 211:1819–1828
tics, the projecting GIs are also tuned to detect Keil TA, Steinbrecht RA (1984) Mechanosensitive and
olfactory sensilla of insects. In: King RC, Akai
complex dynamical signal features that may cor- H (eds) Insect ultrastructure, vol 2. Plenum, New
respond to important naturalistic stimuli such as York, pp 477–516
approaching predators. Four main factors under- Landolfa M, Jacobs GA (1995) Direction sensitivity of the
lie the specialized feature sensitivity of these filiform hair population of the cricket cercal system.
J Comp Physiol A 177:759–766
interneurons: (1) the specific characteristics of Landolfa M, Miller JP (1995) Stimulus-response proper-
the distribution of mechanosensory hairs along ties of cricket cercal filiform receptors. J Comp Physiol
the two cerci, (2) the operation of the cerci as A 177:749–757
delay lines for the sensory afferent inputs located Miller JP, Theunissen FE, Jacobs GA (1991) Representa-
tion of sensory information in the cricket cercal sen-
at different distances out along the cerci (Mulder- sory system. I. Response properties of the primary
Rosi et al. 2010), (3) the highly organized nature interneurons. J Neurophysiol 66:1680–1689
of the afferent terminal arborizations into a map Miller JP, Krueger S, Heys JJ, Gedeon T (2011) Quanti-
of directional and dynamical stimulus parameters tative characterization of the filiform mechanosensory
hair array on the cricket cercus. PLoS One 6(11):
within the terminal abdominal ganglion, and e27873
(4) the architecture of the interneurons’ dendritic Mulder-Rosi J, Cummins GI, Miller JP (2010) The cricket
branches within this afferent map. cercal system implements delay-line processing.
Murphey RK (1985) A second cricket cercal sensory sys-
References tem: bristle hairs and the interneurons they activate.
Aldworth ZN, Dimitrov AG, Cummins GI, Gedeon T, Ogawa H, Baba Y, Oka K (2004) Directional sensitivity of
Miller JP (2011) Temporal encoding in a nervous sys- dendritic calcium responses to wind stimuli in cricket
tem. PLoS Comput Biol 7(5):e1002041 giant interneurons. Neurosci Lett 358:185–188
Baba Y, Shimozawa T (1997) Diversity of motor Ogawa H, Cummins GI, Jacobs GA, Miller JP
responses initiated by a wind stimulus in the freely (2006) Visualization of ensemble activity patterns of
moving cricket, Gryllus bimaculatus. Zool Sci mechanosensory afferents in the cricket cercal sensory
14:587–594 system with calcium imaging. J Neurobiol 66:293–307
Bacon JP, Murphey RK (1984) Receptive fields of cricket Ogawa H, Cummins GI, Jacobs GA, Oka K (2008) Den-
are related by their dendritic structure. J Physiol dritic design implements algorithms for extraction of
(Lond) 352:601–623 sensory information. J Neurosci 28:4592–4603
Boyan GS, Ashman S, Ball EE (1986) Initiation and Palka J, Levine R, Schubiger M (1977) The cercus-to-
modulation of flight by a single giant interneuron in giant interneuron system of crickets. I. Some aspects
the cercal system of the locust. Naturwissenschaften of the sensory cells. J Comp Physiol A 119:267–283
73:272–274 Paydar S, Doan CA, Jacobs GA (1999) Neural mapping of
Camhi JM (1980) The escape system of the cockroach. Sci direction and frequency in the cricket cercal sensory
Am 243:144–157 system. J Neurosci 19:1771–1781
Casas J, Dangles O (2010) Physical ecology of fluid flow Sakaguchi DS, Murphey RK (1983) The equilibrium
sensing in arthropods. Annu Rev Entomol 55:505–520 detecting system of the cricket: physiology and mor-
Heusslein R, Gnatzy W (1987) Central projections of phology of an identified interneuron. J Comp Physiol
campaniform sensilla on the cerci of crickets and A 150:141–152
cockroaches. Cell Tissue Res 247:591–598 Shimozawa T, Kanou M (1984) The aerodynamics and
Hoyle G (1958) The leap of the grasshopper. Sci Am sensory physiology of range fractionation in the cercal
198:30–35 filiform sensilla of the cricket Gryllus bimaculatus.
Jacobs GA, Theunissen F (1996) Functional organization J Comp Physiol A 155:495–505
of a neural map in the cricket cercal sensory system. Theunissen FE, Miller JP (1991) Representation of sen-
J Neurosci 16:769–784 sory information in the cricket cercal sensory system.
Jacobs GA, Theunissen F (2000) Extraction of sensory II. Information theoretic calculation of system accu-
parameters from a neural map by primary sensory racy and optimal tuning curve widths of four primary
interneurons. J Neurosci 20:2934–2943 interneurons. J Neurophysiol 66:1690–1703
Chaos, Neural Population Models and 599 C
Theunissen FE, Roddey JC, Stufflebeam S, Clague H, exhibit unpredictable behavior, in that arbitrarily
Miller JP (1996) Information theoretical analysis of small deviations in the initial conditions, after
dynamical encoding by four identified primary sensory
interneurons in the cricket cercal system. sufficient time has passed, lead to completely
J Neurophysiol 75:1345–1364 different behavior. This is referred to as deter-
Troyer TW, Levin JE, Jacobs GA (1994) Construction and ministic chaos and has been shown to arise in
analysis of a data base representing a neural map. neural population models (NPM).
Microsc Res Tech 29:329–343
C
Cerebellar Long Term Depression Chaotic Systems
Deterministic chaos occurs in systems
▶ Long Term Depression in the Granule Cell-
completely described by deterministic differen-
Purkinje Cell Synapse
tial equations. These systems may be quite sim-
ple, but they have to be nonlinear and live in
phase spaces of at least three dimensions
Cerebellar LTD (Poincaré-Bendixson theorem). Chaos means
that arbitrarily small differences in initial condi-
▶ Long Term Depression in the Granule Cell- tions nevertheless lead to strongly diverging tra-
Purkinje Cell Synapse jectories of system behavior after some time.
Consequently, under any realistic conditions the
system’s long-term behavior is unpredictable. In
spite of the impossibility to predict the system’s
Cerebrovascular Disease behavior in any single case, the chaotic dynamics
are often not entirely arbitrary, but can be
▶ Brain Ischemia and Stroke described by so-called strange attractors. Just
like fixed points or limit cycles, strange attractors
eventually attract all trajectories that start within
their basin of attraction. However, unlike those
Channelopathy
simple forms of attractors, strange attractors are
not simply points or closed orbits, but form com-
▶ Epilepsy: Abnormal Ion Channels
plex manifolds of fractal dimension in phase
▶ Neuropharmacological Modeling, Pharmaco-
space. In spite of their spatial extent, they fill
genomics and Ion Channel Modulation
only an infinitesimally small fraction of the
phase space. Examples are the Lorenz attractor
(Lorenz 1963), which describes a simple hydro-
Chaos, Neural Population Models dynamic system, and the Rössler attractor
and (Rössler 1976), which was inspired by a taffy
puller machine.
Thomas R. Knösche The rate with which nearby trajectories con-
Max Planck Institute for Human Cognitive and verge or diverge is described by the Lyapunov
Brain Sciences, Leipzig, Germany spectrum, comprising one Lyapunov exponent
per phase space dimension (Politi 2013). If the
largest Lyapunov exponent (LLE) is positive, the
Definition system is chaotic (in case of more than one
positive Lyapunov exponent, one speaks of
Nonlinear dynamic systems, even if completely hyperchaos), while zero or negative LLEs indi-
described by deterministic equations, may cate phase space contraction, that is, trajectories
C 600 Chaos, Neural Population Models and
Chaos, Neural
Population Models and,
Fig. 1 Largest Lyapunov
exponent of a periodically
driven Jansen-Rit model, as
a function of stimulus
amplitude and frequency.
Positive exponents
(magenta to black) reflect
diverging trajectories and
thus chaos. Zero exponents
(white) indicate neutral
stability, and negative
exponents (cyan to yellow)
reflect frequency locking.
Several parameter regions
with scattered, presumably
fractal, patterns of chaotic
regime are shown at a finer
resolution (From Spiegler
et al. (2011))
from a certain region in phase space tend to population. They also showed that chaotic and
converge. non-chaotic regions in parameter space form
Whether or not a system is chaotic may a complicated, apparently fractal, pattern. The
depend on system parameters. The transition time series and frequency spectra corresponding
from deterministic to chaotic asymptotic behav- to the chaotic dynamics closely resembled real
ior upon change of one or several system param- EEG in the gamma range (30–100 Hz).
eters is referred to as route to chaos. Very often Spiegler and colleagues (2011) investigated
this happens through an infinite succession of a periodically forced Jansen-Rit model and
period-doubling bifurcations of decreasing spac- observed, when varying amplitude and frequency
ing (Feigenbaum 1978). The chaotic region is of the input, a complicated, presumably fractal,
usually interrupted by islands of stability in pattern of chaotic and non-chaotic regimes
a fractal fashion. (Fig. 1). Interestingly, they showed that this
pattern matches the estimated chaoticity
Chaos in NPMs (in terms of the LLE) of the time series of EEG
In several types of NPMs, chaotic dynamic under stimulation with light flicker of variable
behaviors have been demonstrated. Freeman frequency.
(1987) constructed an NPM of the olfactory sys-
tem and demonstrated different chaotic regimes. Brain Dynamics: Stochastic Processes or
More recently, Dafilis and colleagues (2001, Deterministic Chaos?
2009) investigated the parameter space for Time series of functional brain measurements,
a cortical model of two neural populations, one for example, electroencephalography or magne-
excitatory and one inhibitory (Liley model). They toencephalography, feature a high degree of
showed that chaos can occur, when the excitatory irregularity and unpredictability. There has been
input to the inhibitory population does not greatly a discussion whether stochastic processes or
exceed the excitatory input to the excitatory deterministic chaos is the main reason for this
Chloride Channels 601 C
behavior (Stam 2005). Both can clearly explain Korn H, Faure P (2003) Is there chaos in the brain?
the observations in terms of time series and spec- Experimental evidence and related models. C R Biol
326:787–840
tra. Stochastic elements in functional activity Lorenz EN (1963) Deterministic nonperiodic flow.
could originate from stochastic input from the J Atmos Sci 20(2):130–141
outer and inner environments and from the very Politi A (2013) Lyapunov exponent. Scholarpedia
high number of relatively independent processes 8(3):2722
Rössler OE (1976) An equation for continuous chaos.
that are present in the brain at any given time, Phys Lett 57A(5):397–398
C
while deterministic chaos can originate from the Skarda CA, Freeman WJ (1987) How brains make chaos
intrinsic dynamics of the neural circuits, as in order to make sense of the world. Behav Brain Sci
shown above. While stochasticity would be sim- 10:161–173
Spiegler A, Knösche TR, Schwab K, Haueisen J, Atay FM
ply an epiphenomenon of brain function, chaos (2011) Modeling brain resonance phenomena
could play an important role in information using a neural mass model. PLoS Comput Biol 7:
processing (Skarda and Freeman 1987). Experi- e1002298
mentally, chaos has been shown to exist at the Stam CJ (2005) Nonlinear dynamical analysis of EEG and
MEG: review of an emerging field. Clin Neurophysiol
single-cell level. There is, however, growing evi- 116(2005):2266–2301
dence for the existence of chaos at all levels of
brain function and for its role as part of the neural
code. For an extensive discussion, refer to Korn
and Faure (2003). Characterization of Motor Learning
in Continuous Tracking Tasks
Cross-References ▶ Human Balancing Tasks: Power Laws, Inter-
mittency, and Lévy Flights
▶ Bifurcation Analysis
▶ Bifurcations Dynamics of Single Neurons and
Small Networks
▶ Bifurcations, Neural Population Models and
▶ Jansen-Rit Model Chloride Channels
▶ Neural Population Model
▶ Pattern Formation in Neural Population Steven A. Prescott
Models Neurosciences and Mental Health, The Hospital
▶ Phase Transitions, Neural Population Models for Sick Children, Toronto, ON, Canada
and Department of Physiology, University of
Toronto, Toronto, ON, Canada
References
Definition
Dafilis MP, Liley DTJ, Cadusch PJ (2001) Robust chaos in
a model of the electroencephalogram: implications for Chloride channels play an important role in reg-
brain dynamics. Chaos 11(3):474–478 ulating neuronal excitability, especially in the
Dafilis MP, Frascoli F, Cadusch PJ, Liley DTJ context of fast synaptic inhibition mediated by
(2009) Chaos and generalised multistability in
a mesoscopic model of the electroencephalogram. GABAA and glycine receptors. But in order for
Physica D 238:1056–1060 chloride channels to reduce excitability, chloride
Feigenbaum MJ (1978) Quantitative universality for driving force must be maintained via transporters
a class of nonlinear transformations. J Stat Phys that extrude chloride from the cell. Other trans-
19(1):25–52
Freeman WJ (1987) Simulation of chaotic EEG patterns porters load chloride into the cell. The compli-
with a dynamic model of the olfactory system. Biol cated and dynamic balancing of chloride influx
Cybern 56:139–150 and efflux, which also involves a variety of other
C 602 Chloride Channels
ion species, directly affects how chloride chan- chloride flux is sometimes necessary (Doyon
nels regulate excitability. et al. 2011).
Chloride Flux and the Electrochemical

Detailed Description Gradient
The magnitude and direction of chloride flux
Function across the cell membrane depends on the electro-
Chloride ions participate in many cellular func- chemical chloride gradient. That gradient is
tions, from helping regulate cellular excitability described by the driving force, which is the dif-
to helping regulate cell volume. Those func- ference between membrane potential and chlo-
tions are mediated through a variety of chloride ride reversal potential ECl. ECl is dictated by the
channels and transporters (Duran et al. 2010; ratio of intracellular and extracellular concentra-
Fig. 1). The regulation of neuronal excitability tions, [Cl]i and [Cl]o, as explained by the
garners the most attention among neuroscien- Nernst equation. Under normal conditions (i.e.,
tists. As the dominant charge carrier through low [Cl]i and membrane potential between rest-
GABAA and glycine receptors, chloride is ing level and spike threshold), opening a chloride
directly implicated in the efficacy of fast synap- channel will allow chloride to flow down its elec-
tic inhibition. Therefore, if chloride regulation trochemical gradient and into the neuron. This
is compromised, which occurs under a variety influx represents an outward current since the
of pathological conditions including epilepsy direction of current is defined by the direction in
and neuropathic pain (Kahle et al. 2008; which positive charge flows.
Price et al. 2009), the efficacy of fast synaptic Notably, physiological conditions are often
inhibition can be compromised with disastrous deliberately violated for the purpose of experi-
effects. mental testing. For example, in whole cell patch
Understanding (and modeling) synaptic inhi- clamp recordings, high levels of chloride are
bition requires that we understand (and correctly sometimes included in the pipette solution
account for) chloride flux. Synaptic inhibition is (which will dialyze the cell) in order to increase
most often modeled as a simple current. But the chloride driving force so as to improve the
keeping in mind that chloride current reflects detection and measurement of chloride currents.
chloride flux and that chloride flux can sometimes Likewise, voltage can be clamped at values that,
deplete the chloride gradient, explicitly modeling under normal conditions, would occur only
chloride leak GABAA Potassium-chloride Sodium-potassium Sodium-potassium-chloride GABAA chloride leak

channels (CIC-2) receptor co-transporter (KCC2) ATPase co-transporter (NKCC1) receptor channels (CIC-2)
Cl− Cl− 2K+ Na+ K+ 2Cl−
outside
inside
ATP
HCO3− Cl− K+ 3Na+ Cl− HCO3− Cl−
low intracelluar chloride conditions high intracelluar chloride conditions
Chloride Channels, Fig. 1 Cartoon depicting mem- chloride gradient is maintained by cotransporters (red)
brane proteins that contribute to chloride flux. Channels and exchangers (not shown) that harness the gradient of
(blue) let chloride move down its electrochemical gradi- another ion species to move chloride against its gradient.
ent. Importantly, chloride can move in either direction Maintenance of the other ion gradients ultimately relies on
through those channels depending on the direction of the the sodium-potassium pump (mauve). Numerous other
gradient, notwithstanding rectification of the channel. The proteins involved in chloride flux are not shown
Chloride Channels 603 C
briefly during a spike. Under these artificial con- cotransporters that move both ion species in the
ditions, chloride may flow in the direction oppo- same direction, exchangers move each ion spe-
site to what it would normally. To be clear, there cies in opposite directions.
is nothing wrong with such experiments, but one It has also been suggested that certain rectify-
must consider the relevant factors when ing leak channels, like ClC-2, serve as one-way
interpreting the results. valves that selectively let chloride out of the cell.
However, a passive leak process can only ever C
Chloride Transporters and Chloride let chloride flow down its gradient, unlike
Homeostasis cotransport, which moves chloride against its
The ability of chloride to reduce excitability gradient by harnessing a different ion gradient.
requires that chloride leak into the neuron. In short, the role of ClC-2 channels is context
Passive movement of chloride down its gradient dependent, but they probably work in mature,
and into the cell would eventually deplete the healthy, central neurons to regulate excitability
chloride gradient were it not for other mecha- by leaking chloride into the cell. But to make this
nisms that serve to replenish the gradient by matter more confusing, certain members of the
extruding chloride from the cell. (Notably, pas- ClC family actually function as exchangers rather
sive chloride flux can deplete the gradient, but it than as channels (Duran et al. 2010).
cannot reverse it, which is a topic we will return One should appreciate that there are other
to later). The most important chloride extruder chloride channels not shown in Fig. 1. For
in central neurons is the potassium-chloride instance, many neurons express chloride chan-
cotransporter KCC2 (Payne et al. 2003). KCC2 nels that are activated by increased intracellular
harnesses the potassium gradient such that chlo- calcium (Duran et al. 2010; Hartzell et al. 2005).
ride ions can piggyback potassium ions out of Of particular clinical importance, the cystic fibro-
the neuron through an electroneutral process sis transmembrane conductance regulator
(Fig. 1). Cotransport is not equivalent to (CFTR) is a chloride channel involved in regu-
pumping: Pumping is active (i.e., it depends lating water flow across epithelia; mutations in
directly on ATP hydrolysis), whereas this channel lead to, among other problems,
cotransport is only secondarily active (i.e., it dehydrated mucus that is excessively viscous
depends on ion gradients established and and which therefore blocks ducts, leading to cys-
maintained by pumps). tic fibrosis (Gadsby et al. 2006). Beyond their
The sodium-potassium-chloride cotransporter involvement in phasic synaptic transmission,
NKCC1 is also expressed in neurons and GABA receptors can (depending on their subunit
serves to load chloride into cells via the sodium composition) be exquisitely sensitive to GABA,
gradient. In central neurons, KCC2 expression remaining open even in response to ambient
increases during development, whereas NKCC1 levels of extracellular GABA and thus contribut-
expression decreases (Ben-Ari et al. 2012). The ing a tonic chloride conductance to the cell (Lee
same is not true for peripheral neurons, which is and Maguire 2014; Sigel and Steinmann 2012).
why GABA is depolarizing (although nonethe- In all cases, channels rely on the passive move-
less inhibitory because of processes like sodium ment of chloride down its electrochemical
channel inactivation) in mature peripheral gradient.
neurons.
Other proteins not depicted in Fig. 1 can also Synaptic Inhibition and Chloride Flux
contribute to chloride homeostasis. These include As already explained, synaptic inhibition by
sodium-dependent and sodium-independent GABAA receptors relies on chloride influx to
anion exchangers. Exchangers, or antiporters, produce an outward current. However, GABAA
function like cotransporters in that they move receptors are often described as mediating
one ion against its gradient by harnessing the shunting inhibition because they prevent the
energy of a different gradient, but unlike depolarization caused by concurrent excitatory
C 604 Chloride Channels
input rather than causing gross hyperpolariza- Summary

tion. This occurs because the GABA reversal Chloride plays a crucial role in helping regulate
potential is close to the resting membrane neuronal excitability, especially in the context of
potential, which means that chloride driving fast synaptic inhibition. Inhibition normally
force is small, especially in the absence of relies on chloride moving into the neuron, and
a depolarizing event. Shunting inhibition works inhibition may thus fail if chloride is allowed to
by GABAA receptors mediating a chloride- accumulate intracellularly. Indeed, the direction
based current that is proportional yet opposite and magnitude of chloride flux depend on the
in direction to the inward cation-based current electrochemical chloride gradient, which in turn
mediated by AMPA/NMDA receptors. To be depends on the regulation of intracellular chlo-
clear, cations entering via AMPA/NMDA recep- ride levels. A variety of transporters work to
tors do not escape via the open GABAA chan- regulate intracellular chloride levels, implicating
nels, contrary to what one might assume based other ion species in the process. This dynamic
on the term “shunting.” Consequently, even balancing of chloride influx and efflux is trans-
shunting inhibition involves chloride influx parent under normal conditions, but the complex
and, therefore, is not immune to the effects of interplay between processes must be accounted
chloride dysregulation. for when any one of those processes goes awry.
Notably, GABAA channels do not always
mediate chloride influx. For example, in imma-
ture neurons and mature peripheral neurons that References
maintain a high intracellular chloride concentra-
Ben-Ari Y, Khalilov I, Kahle KT, Cherubini E (2012)
tion through the actions of NKCC1, GABAA The GABA excitatory/inhibitory shift in brain matu-
receptor activation leads to depolarization caused ration and neurological disorders. Neuroscientist
by chloride efflux. This is in contrast to the 18:467–486
depolarizing GABA responses that occur in cen- Doyon N, Prescott SA, Castonguay A, Godin AG,
Kroger H, De Koninck Y (2011) Efficacy of synaptic
tral neurons after pathological reduction of KCC2 inhibition depends on multiple dynamically
and/or during sustained inhibitory input. Under interacting mechanisms implicated in chloride homeo-
conditions in which chloride extrusion capacity is stasis. PLoS Comput Biol 7:e1002149
reduced (but chloride is not actively loaded into Duran C, Thompson CH, Xiao Q, Hartzell HC
(2010) Chloride channels: often enigmatic, rarely pre-
the neuron, e.g., by NKCC1), chloride influx can dictable. Annu Rev Physiol 72:95–121
rapidly deplete the chloride gradient, especially Gadsby DC, Vergani P, Csanady L (2006) The ABC
in small compartments like dendrites that have protein turned chloride channel whose failure causes
a high surface area-to-volume ratio (Doyon cystic fibrosis. Nature 440:477–483
Grover LM, Lambert NA, Schwartzkroin PA, Teyler TJ
et al. 2011), but chloride flux does not reverse (1993) Role of HCO3 ions in depolarizing GABAA
direction; instead, GABA-induced depolarization receptor-mediated responses in pyramidal cells of rat
occurs because the associated bicarbonate efflux hippocampus. J Neurophysiol 69:1541–1555
becomes dominant relative to chloride influx Hartzell C, Putzier I, Arreola J (2005) Calcium-
activated chloride channels. Annu Rev Physiol
(Grover et al. 1993). The relative permeability 67:719–758
of GABAA channels to chloride versus bicarbon- Kahle KT, Staley KJ, Nahed BV, Gamba G, Hebert SC,
ate is about 4:1 and should be accounted for by Lifton RP, Mount DB (2008) Roles of the cation-
calculating the “GABA” reversal potential by chloride cotransporters in neurological disease. Nat
Clin Pract Neurol 4:490–503
using the Goldman-Hodgkin-Katz equation Lee V, Maguire J (2014) The impact of tonic GABA
(which can accommodate multiple ion species) receptor-mediated inhibition on neuronal excitability
rather than the Nernst equation (which accommo- varies across brain region and cell type. Front Neural
dates only one ion species). Because of the bicar- Circuits 8:3
Payne JA, Rivera C, Voipio J, Kaila K (2003) Cation-
bonate component, the functional reversal chloride co-transporters in neuronal communication,
potential for GABAA channels is usually several development and trauma. Trends Neurosci
millivolts less negative than ECl. 26:199–206
Choice Behavior 605 C
Price TJ, Cervero F, Gold MS, Hammond DL, Prescott SA approximations of specific, isolated aspects of
(2009) Chloride regulation in the pain pathway. Brain human behavior. Biologists are interested in
Res Rev 60:149–170
Sigel E, Steinmann ME (2012) Structure, function, and choice behavior, too. Optimal foraging theory,
modulation of GABAA receptors. J Biol Chem for instance, prescribes and explains how animals
287:40224–40231 should make optimal foraging decisions in vari-
able environments if they were to maximize
Darwinian fitness. Neuroscientists have started C
to reveal the neural mechanisms underlying
Choice Behavior choice behaviors in animal and human models.
Thus, despite a superficially obvious dominance
Marijn van Wingerden and Tobias Kalenscher of economic theory in the study of human deci-
Department of Comparative Psychology, sion making, the science of choice behavior com-
Institute of Experimental Psychology, Heinrich- prises many different approaches from different
Heine University D€usseldorf, D€usseldorf, schools of thinking and applies to human as well
Germany as animal subjects. This entry is meant to give an
overview of the different approaches to animal
and human choice behavior.
Synonyms
Choice Behavior in Microeconomics:
Consumer behavior; Decision-making; Prefer- Revealed Preferences
ence mapping In economic theory, choices, measured as the
response distribution among a range of alterna-
tives, are the observable behavioral output of
Definition a process that draws upon preferences. Prefer-
ences are unobservable cognitive structures spe-
The pattern of response allocation when making cific to an individual in a particular context. It is
a selection among several choice alternatives. usually assumed that a decision maker behaves as
if he or she was maximizing a utility function.
A utility function is a mathematical construct
Detailed Description linking objective quantities of a (single) good
A to the subjective value (utility) it renders
Choice behavior – the description of response (Fig. 1a). This utility function is usually assumed
allocation patterns when making a selection to be concave and decelerating, meaning that the
among several choice alternatives, as well as the additional utility gained per unit of A decreases
identification of the choice-underlying psycho- as a function of the amount of A already owned.
logical and neural mechanisms and their evolu- For example, increasing the wealth of a very poor
tionary origins – has been the topic of study in beggar by $100 yields a very large additional
many scientific disciplines. Economics, mainly utility to the beggar, whereas adding $100 to the
micro- and macroeconomics, has generated for- fortune of a very rich person yields little addi-
mal models of choice behavior that are designed tional utility. The increase in utility as a function
to make accurate predictions of (aggregate) of goods already owned is called marginal utility.
human behavior while making as few assump- The marginal utility function (Fig. 1b), obtained
tions as possible and while being agnostic about by plotting the increase in utility due to receiving
the choice-underlying mental mechanisms. Psy- an additional unit of A, follows an exponential
chologists are traditionally more concerned with decay (Friedman and Savage 1948; Bernoulli
describing and explaining the observable reality 1954; Kahneman and Tversky 1979).
of human decision, but psychological theory is Preferences and utility functions are not visi-
often also tested with animal models as ble, but can only be inferred from observable
C 606 Choice Behavior
Choice Behavior, Fig. 1 Utility and marginal utility captures the notion that a small increase given a poor
functions. (a) A utility function links objective quantities reference is more valuable than a similar small increase
of a good with the subjective pleasure (“utility”) derived against a wealthy background. (b) Each additional unit of
from these goods. The concave, decelerating function a good thus confers less additional or marginal utility
choices. In the analysis of consumer choice

behavior, preferences are often inferred by pitting
against each other two or more bundles,
consisting typically of different quantities of
good A and good B and observing the pattern of
responses of the consumer. These choices or
revealed preferences are thought to reflect the
outcome of the consumer’s internal decision-
making processes and are an inherently relative
property. The consumer is assumed to maximize
her/his utility (Von Neumann and Morgenstern
1944), or subjective pleasure, derived from
obtaining each of the alternatives within her/his
budget constraints. This is illustrated by plotting Choice Behavior, Fig. 2 Revealed preferences, the bud-
revealed preferences on a budget line diagram get line, and indifference curve. If a consumer is asked to
(Fig. 2). (Although the concept “utility” is often spend a budget of 10 units between good A and good B,
equated with subjective pleasure, this is an impre- both priced at 1 unit per good, her/his chosen bundle
(purple circle on the blue budget line) reveals her/his
cise paraphrase of its historical meaning, since preference. The indifference curve (red) links hypotheti-
economic theory is agnostic about the actual cal bundles that would render equal utility to the chosen
mental mechanisms producing choice. Thus, bundle. The marginal rate of substitution moves away
strictly speaking, the term utility is inherently from 1:1 as either good A or good B becomes scarcer
and the utility attached to each unit of the scarce good
descriptive and does not refer to any actual increases
hedonic or emotional states.) When asked to
spend her/his budget between goods A and
B within the allotted budget, the consumer available to the consumer within the budget.
reveals her/his preferences between A and B by The weak axiom of revealed preferences
selecting a certain bundle on the budget line. This (WARP; Samuelson 1938) states that the selected
bundle represents the best possible alternative bundle is at least as good or better than other
bundles on the budget line and strictly preferred substitution and can be read out from the slope
over all other bundles that contain less goods than of the indifference curve.
the chosen bundle (every bundle inside the trian-
gle). This prediction is derived from the assump- Choice Behavior in Animals: Optimal
tion that, if the consumer prefers A over B, he or Foraging
she cannot simultaneously prefer B over A and It is very likely that evolution has shaped the way
that more of a good thing is better than less of in which people make decisions. The evolution of C
a good thing (cf. Glimcher 2010). WARP is an decision making can be investigated by studying
appealing theory with only minimal assumptions, animal choice behavior. Biologists usually
but has little predictive power outside the assume animals forage for food, choose mating
revealed preference set. WARP has been partners, or make other decisions affecting their
extended to GARP, the generalized axiom of fitness. According to this view, natural selection
revealed preferences, refined, and shown to be has favored the development of specific choice
a necessary and sufficient condition for choice behaviors that maximize the decision maker’s
data to be consistent with the maximization of Darwinian fitness, e.g., that optimize their energy
a continuous, concave, and monotonic utility intake and reproduction success (Kalenscher and
function (Houthakker 1950; Afriat 1967; van Wingerden 2011). Thus, like humans, ani-
Harbaugh et al. 2001). Satisfying GARP implies mals encounter a plethora of choices in relation to
internally consistent decision making and transi- food intake and energy expenditure. In the field of
tivity, i.e., the preservation of a preference order, behavioral ecology, optimal foraging theory is
presumably corresponding to an at least ordinal a normative theory that prescribes that animals
utility ranking of the choice alternatives. should make foraging decisions that maximize
The chosen bundle is also informative about their fitness, operationalized as the ratio of food
the consumer’s relative preference between intake over expenditure to ensure survival
A and B. More precisely, one could ask which (Charnov 1976) or reproductive success
bundles other than the chosen one would be (Hamilton 1964a, b). When taking into account
equally valuable (i.e., render equal utility) to the that a given food patch becomes depleted over
consumer. Let us assume a consumer wanted to time and obtaining additional energy intake
buy fruit and chose a bundle containing 4 apples therefore is associated with progressively larger
and 3 oranges. And let us now assume that one amounts of energy spent foraging, the marginal
apple is taken away from this bundle. How many rate of energy intake mirrors the shape of the
oranges would we need to add to this new bundle marginal utility function. Indeed, animals are
to make it equally valuable to the old bundle? sensitive to the marginal rate of energy intake
Removing one apple may be compensated by when making foraging decisions.
adding, say, one orange. However, the less
apples the consumer has in his bundle, the Choice Distribution: Herrnstein’s Matching
more valuable each apple becomes (evident Law
from the concave shape of utility functions), so When making repeated decisions between
if yet another apple is removed, one may have to options A and B, a decision maker should consis-
add 2 additional oranges to make the next bundle tently choose the option that represents the
equally attractive. Such equally valued bundles highest utility. Typically, however, this is not
can be linked with an indifference curve (Fig. 2), how animals behave in such a choice situation.
a function that is convex to the origin, tangential Since the reward rate associated with a given
to the originally chosen bundle, and represents option might change over time, it pays to switch
combinations of good A and good B that render between exploitation and exploration of choice
equal utility. The rate at which a consumer is options. The matching law (Herrnstein 1961)
willing to give up one unit of A in favor of one states that animals distribute their responses
unit of B is called the marginal rate of across choice options to match the relative rates
of reward obtained from those options. So, for hunt. A predator could opt for initiating a hunt
instance, if option A provides three units of subject to high risk with regard to the actual
reward and option B provides one unit, then outcome or rather pursuing a low-risk foraging
option A is typically chosen three times as often strategy with a tighter outcome distribution. In
as B. Indeed, when given a choice, animals do humans, risk and probabilities can be simulta-
explore the alternatives as predicted by the neously or independently instructed, whereas ani-
matching law (Kalenscher et al. 2003; Sugrue mals typically have to sample the outcome
et al. 2004). distribution to arrive at estimates of outcome
risk and probabilities (Schultz et al. 2008, 2011;
Choice Behavior Under Risk: Expected Utility Tobler et al. 2009; O’Neill and Schultz 2010).
Often, the outcome associated with a choice The most prominent normatively flavored
option is not deterministic, but prob- framework for understanding decision making
abilistic. Probabilistic outcomes can be drawn under risk in humans is the expected utility theory
from a distribution with known variance (such (EUT; Von Neumann and Morgenstern 1944).
as a dice toss), in which case economists say EUT states that a decision maker should integrate
that the outcomes are subject to risk. Thus, risk the utilities of all possible outcomes of a decision
is often defined as outcome variance. Risk con- weighted by their probabilities into an expected
trasts with ambiguity, which occurs when the utility estimate, which can then be used to com-
variance in the outcome distribution is not (yet) pare option alternatives. EUT proposes a set of
known (Burke and Tobler 2011). In addition, mathematical equations to describe choice
uncertainty, defined as the inverse of predictabil- behavior based on four axioms that define
ity of the outcome, has been shown to affect a rational decision maker (rational in this case
decision making too. Risk, probability, and means that the decision maker exhibits stable
uncertainty are related concepts, as illustrated in preferences and behaves consistently with
the following example. Assume a decision maker respect to them). These axioms are completeness,
chooses between the following gambles: continuity, transitivity, and independence. The
(A) Win $90 or $110 with equal probability first three prescribe that a decision maker can
(p = 0.5) and will rank option alternatives consistently,
or while the fourth ensures that alternatives are eval-
(B) Win $50 or $150 with equal probability uated independently, such that, for instance, the
(p = 0.5) rank of an alternative should not be affected by
Both alternatives have equal expected value the availability of an inferior alternative. If the
(the sum of the products of the probabilities times axioms of EUT are met, option A should
reward magnitudes is $100 in both cases). How- be preferred over option B if and only if
ever, alternative (B) has the higher outcome var- U(A) > U(B), that is, if the utility derived from
iance and is therefore the riskier alternative. For option A exceeds that obtained from option B. If
each alternative, the uncertainty of each outcome a consumer’s choices imply a concave utility
is maximal because p = 0.5. Changing the function, EUT predicts risk aversion if the choice
reward probabilities would skew the outcome is between a risky outcome (i.e., equiprobable
distribution and would affect the alternatives’ low and high payouts) versus a safe outcome
expected values. Uncertainty decreases towards (certain, medium-sized outcome, Fig. 3). Due to
the probability extremes and is zero at p = 0 or the decreasing rate of marginal utility implied
p = 1. in concave utility functions, high outcomes
Risk also seems to play a role in animal for- are discounted more relative to small outcomes,
aging decisions. For example, the actual caloric so that
yield associated with hunting a certain type of
prey varies depending on the specifics of the
prey and the amount of energy spent during the Uhigh þ Ulow < 2 Umed
Choice Behavior, Fig. 3 Risk aversion and risk prone- with increasing level of stimulus magnitude. A concave
ness. Utility functions can explain risk attitude according utility/fitness function predicts risk aversion when choos-
to expected utility theory and risk sensitivity theory. (a) ing between a medium-sized, certain reward (RM) and
Utility (for humans) or Darwinian fitness (for animals) as a risky option offering large or small rewards (RS and
a function of the magnitude of a commodity. The utility/ RL) with equal probabilities. (b) A convex function pre-
fitness curve is concave and is a decelerating function of dicts risk proneness (Taken from Kalenscher and van
the current level of stimulus magnitude (wealth/amount) Wingerden (2011), with permission)
because the marginal utility/fitness increment decreases
resulting in net risk aversion. In the study of theory (DUT; Samuelson 1937; Koopmans
foraging decisions, the concept of outcome vari- 1960; Prelec and Loewenstein 1991; Kalenscher
ance has been captured in risk-sensitive foraging and Pennartz 2008). DUT states that the value of
theory by assuming that energy intake is a delayed reward decreases with the duration of
nonlinearly coupled to Darwinian fitness, the delay according to a discount function. There-
resulting in a decreasing rate of marginal fitness fore, in maximizing the utility of her/his choices,
that in turn predicts risk aversion (Kacelnik and a decision maker should take into account the
Bateson 1996; Kacelnik 1997; Stephens 2008). discounted utility, derived by reducing the utility
Indeed, humans and animals are mostly risk of each of the alternatives according to this time-
averse in many domains and situations, although dependent discount factor.
they occasionally exhibit risk-seeking behavior The shape of the discount function has been
(Friedman and Savage 1948; McCoy and Platt the subject of debate. Originally, a decreasing
2005; Tobler et al. 2007; Glimcher 2008). exponential with a constant discount factor was
proposed (Samuelson 1937). An exponential
Choice Behavior with Delayed Outcomes: discount function predicts that preferences are
Discounted Utility stationary, i.e., that the addition of a common
Besides outcome variance, delay to reward is delay to both alternatives in an intertemporal
another important factor affecting choice behav- choice task should not interfere with the prefer-
ior. Humans and other animals, when given ence order (Fig. 4a; Koopmans 1960). Formally
a choice between two equally sized but differen- put, if a decision maker is indifferent between
tially delayed outcomes, will prefer the sooner option A delivered at time t1 and option
reward to the later reward (Samuelson 1937; B delivered at time t1 + Dt, that is, choosing
McDiarmid and Rilling 1965; Rachlin and both options about 50 % of the time, then pref-
Green 1972; Ainslie 1974; Kalenscher and van erence stationarity states that he or she should
Wingerden 2011). This process, called delay also be indifferent between option A delivered at
discounting, has been formalized in the norma- a later time t2 and option B delivered at t2 + Dt
tive economic framework of discounted utility (Eq. 1):
a CONSTANT DISCOUNTING b HYPERBOLIC DISCOUNTING

Discounted Utility
Small Reward Value

Large Reward Value
UL
Us<UL US
US
Us>UL UL
UL UL
Us<UL US Us<UL US
Small, Early Large, Small, Early Large, Small, Early Large, Small, Early Large,
Reward Delayed Reward Delayed Reward Delayed Reward Delayed
Reward Reward Reward Reward
PROXIMAL REWARDS DISTANT REWARDS PROXIMAL REWARDS DISTANT REWARDS
Choice Behavior, Fig. 4 Constant vs. hyperbolic place a premium on short-term availability of rewards,
discounting. Constant vs. hyperbolic discounting of future deflecting the discount into an upward direction for tem-
events. The figure describes a choice between a small, porally close rewards. The resulting hyperbolic discount
short-term outcome or a large, long-term outcome function can explain preference reversals over time. Due
(proximal) and another situation in which both outcomes to the steeper utility decay for short delays, the utility of
are deferred into the future by the same time interval the small, short-term commodity is higher than that of the
(distant). (a) Constant (here: exponential) utility function large, delayed reward for temporally proximal outcomes,
of a large, delayed (gray line) and small, short-term com- but the utility order reverses when both outcomes are
modity (black line). With exponential discounting, pref- deferred into the future (Taken from Kalenscher and
erence stationarity holds when the rewards are deferred by Pennartz (2008), with permission)
the same time interval into the future. (b) People seem to
If Aðt1 Þ Bðt1 þ DtÞ then Aðt2 Þ Ecological Rationality: Sequential

Bðt2 þ DtÞ (1) Background-Foreground Choices
It has been argued that the (binary) choice
However, empirical evidence shows that in between two differentially delayed alternatives
many situations, decision makers exhibit time- is an artificial testing situation that does not cor-
inconsistent preferences (see “▶ Decision- respond to the contexts in which animals make,
Making, Bias). This means that a decision maker and arguably preindustrial age humans made,
prefers A(t1) over B(t1 + Dt) but B(t2 + Dt) over their decisions. Stephens (2008) and coworkers
A(t2) after the addition of some common delay have emphasized the use of sequential choice
t2–t1 (Ainslie 1975; Green et al. 1994; McClure paradigms to approach foraging behavior, using
et al. 2004). Given this evidence, a hyperbolic patches as canonical units of foraging. A patch
discount function, which can explain preference constitutes a finite-sized resource that can be
reversals after adding common delays, has been exploited more than once. Importantly, the mar-
proposed (Fig. 4b; Ainslie 1975; Mazur 1984, ginal rate of return decreases with each subse-
1988; Rachlin et al. 1991; Green et al. 2010). quent extraction until the patch is depleted. The
The tendency of humans and other animals to put foraging animal thus faces a sequential stream of
a premium on the short-term availability of choices to either stay and further exploit the patch
rewards has been called impulsivity or failure of or go and explore another source, which likely
self-control (see “▶ Decision-Making, Bias”). depends on parameters such as the current rate of
return given the patch depletion state and the
travel time to the next patch. This type of sequen-
tial choice has a background-foreground struc-
ture in which the forager default or background
strategy is to explore the environment for food,
which can be temporarily put on hold to follow
the foreground strategy of exploiting the current C
foraging options, such as gathering additional
nectar or attacking a prey. Stephens et al. model
intertemporal choice behavior using a patch par-
adigm by splitting the large, delayed reward in
two smaller rewards. After receiving an initial
small reward, the animal can choose to stay for
an additional delay, resulting in the delivery of
the second reward, or go and start a new trial
(Fig. 5). By adjusting the time delays in the
patch preparation to the delays typically used in
the self-control or intertemporal choice para-
digm, the mean outcome rates over time associ-
ated with the choice options can be matched
precisely (Stephens and Anderson 2001).
Whereas animals typically make many “impul- Choice Behavior, Fig. 5 Self-control and patch-
foraging paradigms. Two “equivalent” decision contexts.
sive” decisions in the binary intertemporal choice In the self-control paradigm, the subject makes a binary
paradigm, resulting in a suboptimal rate of energy choice between a small immediate outcome (G1 after t1)
intake, they attain much higher energy intake and a large delayed outcome (G2 after t2). In the patch
rates in the economically equivalent patch para- paradigm, the forager’s choice has equivalent conse-
quences, but the forager decides whether to stay in the
digms. Modeling choice behavior through deci- patch and collect G2–G1 additional units of food after t2–t1
sion rules, Stephens and Anderson show that the or to leave immediately after obtaining the small amount
identical “shortsighted” decision rules produce G1. In the latter, but not the former case, the inter-trial
optimal choices, in the sense of maximizing interval (t) is part of a key delay to the next reward
opportunity (Taken from Stephens and Anderson (2001))
energy intake rates over time, in one paradigm
(the patch situation), but very clearly not in the
other paradigm (the binary intertemporal choice
situation). Specifically, in the patch paradigm, the maximizing a utility function, but economic the-
inter-trial interval becomes a key delay, included ory does not assume that there actually is such
in the stay/leave decision, so that the short-term a thing as a mental representation of a utility
and long-term optimality of the choice for the function. Recently, this agnostic view has been
larger, delayed reward becomes aligned (see challenged by advances in the neurosciences
“▶ Decision-Making, Bias”). aimed at investigating whether there actually is
a neural representation of (cardinal) utility in the
Neural Correlates of Choice Behavior brain. This new “neuroeconomics” paradigm
Economic theory is traditionally agnostic regard- (Glimcher and Rustichini 2004) has proven
ing the mental mechanisms underlying choice quite fruitful in finding correlates of inferred
behavior, but is only concerned with accurately utilities and, conversely, in predicting choice
describing and predicting choices. As a conse- from neural data, supporting the validity of an
quence, a common notion in economics is that underlying utility function guiding choice
a decision maker behaves “as if” he or she was behavior.
The neural correlates of value-based decision influence choice behavior. Assuming that
making (as distinct from, e.g., perceptual deci- revealed preferences reflect internal decision-
sion making or categorization) have been making processes using the ordinal utility rank-
investigated in both humans and animals. A ing of choice alternatives and that these processes
well-established finding in the research on value rely on cardinal representations of value in the
coding in the brain is that frontal brain regions brain, it follows that neural signals in some brain
such as the orbitofrontal cortex (OFC), lateral areas should be sensitive to these parameters
prefrontal cortex (lPFC), ventromedial prefrontal (Burke and Tobler 2011).
cortex (vmPFC), and anterior cingulate cortex
(ACC) show differential levels of activity with Risk
regard to the (expected) value of a stimulus (see In humans, signals of risk, defined as outcome
recent reviews by Rangel and Hare 2010; Louie variance or reward volatility, can be found as
and Glimcher 2012). Though the values associ- BOLD activity correlates in the anterior cingulate
ated with stimuli and actions influence neural cortex (Behrens et al. 2007; Christopoulos
processing in many other cortical and subcortical et al. 2009), OFC (Tobler et al. 2007), and VStr
brain regions, electrophysiological recordings (Kuhnen and Knutson 2005). In nonhuman pri-
made from primate OFC (area 11/13, which has mates, a population of neurons recorded from
homologues in rodents and humans; Wallis 2012) macaque OFC and posterior cingulate cortex
indicate that neural representations of value asso- (CGp) show firing rates that correlate with risk
ciated with outcomes found in that brain region levels. When given a choice between safe and
are independent of the action required to obtain risky outcomes, under certain conditions, these
the said outcome, its physical location, and, to animals prefer the risky outcome, with their pref-
some degree, its sensory features (Padoa- erence increasing as the risk increases (McCoy
Schioppa and Assad 2006, 2008; Padoa-Schioppa and Platt 2005; O’Neill and Schultz 2010); how-
2007). When monkeys were faced with ever, this risk-seeking behavior is linked to the
a binary choice, some neurons showed variation objective levels of reward, so it reverses to risk-
in firing rate that correlated with the derived neutral or risk-averse behavior when higher out-
subjective utility estimate of the chosen alterna- comes are at stake. In macaques, these neuronal
tive, called chosen value correlates. Such an risk signals occur separately from expected value
abstract representation of subjective value in signals present in other OFC neurons, but in rats,
a goods-based model (Padoa-Schioppa 2011) of risk does modulate reward-related neuronal
choice alternatives approximates economic responses (Roitman and Roitman 2010). Dopa-
value as defined in normative frameworks such mine neurons also code a tonic risk signal in the
as EUT or DUT (see above). In humans, similar delay interval between stimulus and outcome
representations of expected and obtained (Fiorillo et al. 2003, 2005; Tobler et al. 2005). It
value, which could inform decision-making pro- thus appears that the brain maintains an isolated
cesses and decision evaluation, are present in the measure of outcome variance, tracking the vola-
ventromedial prefrontal cortex (Kable and tility of the reward environment. From an effi-
Glimcher 2007, 2009; Kahnt et al. 2010) and cient coding perspective, this measure of risk
ventral striatum (Knutson et al. 2001; Tobler should influence the neuronal representation of
et al. 2007; Rolls et al. 2008; Christopoulos outcomes by scaling these representations to their
et al. 2009). mean (expected outcome) and variance (risk) to
facilitate reward discrimination (Schultz
Neural Coding of Risk, Probability, and Delays et al. 2011; Louie and Glimcher 2012). Record-
Associated with Outcomes ings of neural activity in primate and human OFC
As outlined above, risk/outcome variance, prob- and primate lateral intraparietal area (LIP) show
ability/uncertainty, and delay to reward all that this is indeed the case (Elliott et al. 2008;
Padoa-Schioppa 2009; Louie et al. 2011; Louie discriminates between different probabilities
and Glimcher 2012). (Van Duuren et al. 2009). In monkeys, firing
The neural basis of risk attitudes has recently rate modulation by outcome probability was
been investigated, indicating that value signals found in ACC, OFC, and lateral prefrontal cortex
related to risky outcomes in lateral prefrontal by Kennerly et al. (2009) and Wallis and
cortex are modulated by individual risk attitudes Kennerley (2011). Interestingly, the OFC
(Tobler et al. 2009). The value signal of risky contained a larger proportion of neurons that C
options was attenuated by increasing risk in seemed to code an isolated probability signal,
risk-averse participants, but increased in risk- whereas this parameter was more often integrated
seeking participants. In addition, anticipation of with other decision parameters in ACC and lPFC.
risk differentially modulated neuronal activity in
ventral striatum (VStr) and anterior insula Delay
(AI) for risk seekers and risk avoiders (Rudorf Humans and other animals prefer immediate over
et al. 2012). delayed rewards (see above). One view on the
neural basis of these time preferences holds that
Probability immediate and delayed rewards are actually
While a neural correlate of risk associated with processed by different neural systems that com-
a certain stimulus is informative for tracking the pete for behavioral control. The so-called b sys-
volatility of the expected outcome stream, tem is placed in the mesolimbic reward areas
a stimulus that predicts reward with p = 0.75 such as VStr and OFC and is thought to process
should clearly be valued higher than a stimulus the immediate value of reward, while the d sys-
that predicts reward with p = 0.25, in spite of tem, located in prefrontal and parietal regions,
equal risk associated with both stimuli. In con- performs basic exponential discounting of
trast to humans, who can be instructed on proba- delayed rewards (McClure et al. 2004, 2007).
bilities, animals have to experience the outcome Together, the b/d system combines into a
distribution. As probabilities associated with hyperbolic-like discounting function consistent
stimuli become known through repeated sam- with theoretical and experimental data (Ainslie
pling, neural signals of outcome probabilities 1975; Mazur 1984, 1988; Rachlin et al. 1991)
could emerge. Platt and Glimcher (1999) with more steep discounting in the b than the d
presented monkeys with pairs of stimuli that dif- network. Although the view that discounting is
fered in the probability of associated reward, the consequence of two separate neural choice
while recording the activity from neurons in systems has been criticized (Kable and Glimcher
LIP. They found that LIP neurons coded the 2007), recent models posit that the discounted
action (a saccade) required to obtain the reward future reward value is neutrally represented as
and that these neuron’s firing rates were modified a singular signal in ventromedial prefrontal cor-
by both the probability of receiving a reward and tex which, however, can be modified by activity
its magnitude. When the probability of reward in brain structures that were associated with the d
varied, animals exhibited matching behavior system, mainly dorsolateral prefrontal cortex
(see above), mirrored by neural activity in LIP (Hare et al. 2009; Figner et al. 2010). In primates,
that tracked the expected reward associated with single neurons in LIP (Louie and Glimcher
each saccade target across trials (Sugrue 2010), OFC (Roesch and Olson 2005), and stria-
et al. 2004). Dopamine neurons in the ventral tum (Cai et al. 2011) encode subjective value
tegmental area (VTA) of macaques also seem to modulated by delay to reward. Likewise, single-
incorporate probability in coding an expected neuron activity in the PFC homologue of pigeons
outcome signal at the time of stimulus onset, in (Kalenscher et al. 2005) and in rat OFC (Roesch
line with EUT (Fiorillo et al. 2003). In rats, neural et al. 2006) and other limbic structures (reviewed
activity in OFC single cells and ensembles in Roesch and Bryden 2011) is sensitive to delay.
However, in pigeons, the reward signal integrates Ainslie G (1974) Impulse control in pigeons. J Exp Anal
delay to reward with magnitude, whereas this is Behav 21:485–489
Ainslie G (1975) Specious reward: a behavioral theory of
not the case in rat OFC, amygdala, and ventral impulsiveness and impulse control. Psychol Bull
striatum. 82:463–496
The importance of the OFC for the represen- Behrens TEJ, Woolrich MW, Walton ME, Rushworth
tation to delayed rewards is underlined by the MFS (2007) Learning the value of information in an
uncertain world. Nat Neurosci 10:1214–1221
numerous lesion studies that have implicated Bernoulli D (1954) Exposition of a new theory on the
this region, sometimes with opposing effects measurement of risk. Econometrica 22:23
(Mobini et al. 2002; Winstanley et al. 2004; Burke CJ, Tobler PN (2011) Coding of reward probability
Rudebeck et al. 2006; Mar et al. 2011). In rats, and risk by single neurons in animals. Front Neurosci
5:121
Mar et al. found that the exact anatomical locus Cai X, Kim S, Lee D (2011) Heterogeneous coding of
of the lesion differentially affected delay temporally discounted values in the dorsal and ventral
discounting, with lateral OFC lesions increasing striatum during intertemporal choice. Neuron
and more medial lesion decreasing discounting, 69:170–182
Charnov EL (1976) Optimal foraging, the marginal value
respectively. theorem. Theor Popul Biol 9:129–136
Christopoulos GI, Tobler PN, Bossaerts P, Dolan RJ,
Neural Representation of Value in Foraging Schultz W (2009) Neural correlates of value, risk,
As reviewed earlier, the binary choice context and risk aversion contributing to decision making
under risk. J Neurosci 29:12574–12583
presented in most of the studies discussed above Elliott R, Agnew Z, Deakin JFW (2008) Medial
might not reflect the choices for which neural orbitofrontal cortex codes relative rather than absolute
processes have presumably been adapted in the value of financial rewards in humans. Eur J Neurosci
course of evolution (Stephens 2008; Kalenscher 27:2213–2218
Figner B, Knoch D, Johnson EJ, Krosch AR, Lisanby SH,
and van Wingerden 2011). If choice behavior Fehr E, Weber EU (2010) Lateral prefrontal cortex and
relies partly on evolutionarily old adaptations to self-control in intertemporal choice. Nat Neurosci
the stay/leave decisions made during foraging, it 13:538–539
follows that the relative value of these alterna- Fiorillo CD, Tobler PN, Schultz W (2003) Discrete coding
of reward probability and uncertainty by dopamine
tives should be represented, perhaps separately, neurons. Science 299:1898–1902
in phylogenetically older brain structures. The Fiorillo CD, Tobler PN, Schultz W (2005) Evidence that
challenge in foraging can be conceptualized as the delay-period activity of dopamine neurons corre-
calculating the relative outcome rate of the cur- sponds to reward uncertainty rather than
backpropagating TD errors. Behav Brain Funct BBF
rent foreground environment versus the mean 1:7
outcome rate of the background. Recent work in Friedman M, Savage LJ (1948) The utility analysis of
humans and primates indicates that dorsal ACC choices involving risk. J Polit Econ 56:279
seems to code signals that relate to stay/go deci- Glimcher PW (2008) Understanding risk: a guide for the
perplexed. Cogn Affect Behav Neurosci 8:348–354
sions in patch-foraging paradigms (Hayden Glimcher PW (2010) Foundations of neuroeconomic anal-
et al. 2011; Kolling et al. 2012). While the ysis. Oxford University Press, New York
OFC/vmPFC seems mainly concerned with cur- Glimcher PW, Rustichini A (2004) Neuroeconomics: the
rent or chosen value, dACC activity scales with consilience of brain and decision. Science
306:447–452
the relative level of reward in the background in Green L, Fristoe N, Myerson J (1994) Temporal
both humans and macaques, possibly serving as discounting and preference reversals in choice
a reference estimate for stay/leave foraging between delayed outcomes. Psychon Bull Rev
decisions. 1:383–389
Green L, Myerson J, Calvert AL (2010) Pigeons’
discounting of probabilistic and delayed reinforcers.
J Exp Anal Behav 94:113–123
References Hamilton WD (1964a) The genetical evolution of social
behaviour. I. J Theor Biol 7:1–16
Afriat S (1967) The construction of utility functions from Hamilton WD (1964b) The genetical evolution of social
expenditure data. Int Econ Rev 8:67–77 behaviour. II. J Theor Biol 7:17–52
Harbaugh W, Krause K, Berry T (2001) GARP for kids: on Koopmans TC (1960) Stationary ordinal utility and impa-
the development of rational choice behavior. Am Econ tience. Econometrica 28:287–309
Rev 91(5):1539–1545 Kuhnen CM, Knutson B (2005) The neural basis of finan-
Hare TA, Camerer CF, Rangel A (2009) Self-control in cial risk taking. Neuron 47:763–770
decision-making involves modulation of the vmPFC Louie K, Glimcher PW (2010) Separating value from
valuation system. Science 324:646–648 choice: delay discounting activity in the lateral
Hayden BY, Pearson JM, Platt ML (2011) Neuronal basis intraparietal area. J Neurosci 30:5498–5507
of sequential foraging decisions in a patchy environ- Louie K, Glimcher PW (2012) Efficient coding and the
ment. Nat Neurosci 14:933–939 neural representation of value. Ann N Y Acad Sci
C
Herrnstein RJ (1961) Relative and absolute strength of 1251:13–32
response as a function of frequency of reinforcement. Louie K, Grattan LE, Glimcher PW (2011) Reward value-
J Exp Anal Behav 4:267–272 based gain control: divisive normalization in parietal
Houthakker H (1950) Revealed preference and the utility cortex. J Neurosci 31:10627–10639
function. Economica 17:159–174 Mar AC, Walker ALJ, Theobald DE, Eagle DM, Robbins
Kable JW, Glimcher PW (2007) The neural correlates of TW (2011) Dissociable effects of lesions to
subjective value during intertemporal choice. Nat orbitofrontal cortex subregions on impulsive choice
Neurosci 10:1625–1633 in the rat. J Neurosci 31:6398–6404
Kable JW, Glimcher PW (2009) The neurobiology of Mazur JE (1984) Tests of an equivalence rule for fixed and
decision: consensus and controversy. Neuron variable reinforcer delays. J Exp Psychol Anim Behav
63:733–745 Process 10:426–436
Kacelnik A (1997) Normative and descriptive models of Mazur JE (1988) Estimation of indifference points with an
decision making: time discounting and risk sensitivity. adjusting-delay procedure. J Exp Anal Behav
Ciba Found Symp 208:51–70 49:37–47
Kacelnik A, Bateson M (1996) Risky theories – the effects McClure SM, Laibson DI, Loewenstein G, Cohen JD
of variance on foraging decisions. Integr Comp Biol (2004) Separate neural systems value immediate and
36:402–434 delayed monetary rewards. Science 306:503–507
Kahneman D, Tversky A (1979) Prospect theory: an anal- McClure SM, Ericson KM, Laibson DI, Loewenstein G,
ysis of decision under risk. Econometrica 47:263 Cohen JD (2007) Time discounting for primary
Kahnt T, Heinzle J, Park SQ, Haynes J-D (2010) The rewards. J Neurosci 27:5796–5804
neural code of reward anticipation in human McCoy AN, Platt ML (2005) Risk-sensitive neurons in
orbitofrontal cortex. Proc Natl Acad Sci U S macaque posterior cingulate cortex. Nat Neurosci
A 107:6010–6015 8:1220–1227
Kalenscher T, Pennartz CM (2008) Is a bird in the hand McDiarmid CG, Rilling ME (1965) Reinforcement delay
worth two in the future? The neuroeconomics of and reinforcement rate as determinants of schedule
intertemporal decision-making. Prog Neurobiol preference. Psychon Sci 2:195–196
84:284–315 Mobini S, Body S, Ho MY, Bradshaw CM, Szabadi E,
Kalenscher T, Van Wingerden M (2011) Why we should Deakin JF, Anderson IM (2002) Effects of lesions of
use animals to study economic decision the orbitofrontal cortex on sensitivity to delayed and
making – a perspective. Front Neurosci 5:82 probabilistic reinforcement. Psychopharmacology
Kalenscher T, Diekamp B, G€ unt€
urk€
un O (2003) (Berl) 160:290–298
Neural architecture of choice behaviour in O’Neill M, Schultz W (2010) Coding of reward risk by
a concurrent interval schedule. Eur J Neurosci orbitofrontal neurons is mostly distinct from coding of
18:2627–2637 reward value. Neuron 68:789–800
Kalenscher T, Windmann S, Diekamp B, Rose J, Padoa-Schioppa C (2007) Orbitofrontal cortex and the
G€unt€urk€un O, Colombo M (2005) Single units in the computation of economic value. Ann N Y Acad Sci
pigeon brain integrate reward amount and time-to- 1121:232–253
reward in an impulsive choice task. Curr Biol Padoa-Schioppa C (2009) Range-adapting representation
15:594–602 of economic value in the orbitofrontal cortex.
Kennerley SW, Dahmubed AF, Lara AH, Wallis JD J Neurosci 29:14004–14014
(2009) Neurons in the frontal lobe encode the value Padoa-Schioppa C (2011) Neurobiology of economic
of multiple decision variables. J Cogn Neurosci choice: a good-based model. Annu Rev Neurosci
21:1162–1178 34:333–359
Knutson B, Adams CM, Fong GW, Hommer D (2001) Padoa-Schioppa C, Assad JA (2006) Neurons in the
Anticipation of increasing monetary reward selec- orbitofrontal cortex encode economic value. Nature
tively recruits nucleus accumbens. J Neurosci 21: 441:223–226
RC159 Padoa-Schioppa C, Assad JA (2008) The representation
Kolling N, Behrens TEJ, Mars RB, Rushworth MFS of economic value in the orbitofrontal cortex is
(2012) Neural mechanisms of foraging. Science invariant for changes of menu. Nat Neurosci
336:95–98 11:95–102
C 616 Chronic Electrical Stimulation
Platt ML, Glimcher PW (1999) Neural correlates of deci- Tobler PN, O’Doherty JP, Dolan RJ, Schultz W (2007)
sion variables in parietal cortex. Nature 400:233–238 Reward value coding distinct from risk attitude-related
Prelec D, Loewenstein G (1991) Decision making over uncertainty coding in human reward systems.
time and under uncertainty: a common approach. J Neurophysiol 97:1621–1632
Manag Sci 37:770–786 Tobler PN, Christopoulos GI, O’Doherty JP, Dolan RJ,
Rachlin H, Green L (1972) Commitment, choice and self Schultz W (2009) Risk-dependent reward value signal
control. J Exp Anal Behav 17:15–22 in human prefrontal cortex. Proc Natl Acad Sci U S
Rachlin H, Raineri A, Cross D (1991) Subjective proba- A 106:7185–7190
bility and delay. J Exp Anal Behav 55:233–244 Van Duuren E, Van Der Plasse G, Lankelma J, Joosten
Rangel A, Hare T (2010) Neural computations associated RNJMA, Feenstra MGP, Pennartz CMA, Van
with goal-directed choice. Curr Opin Neurobiol Duuren E, Van der Plasse G (2009) Single-cell and
20:262–270 population coding of expected reward probability in
Roesch MR, Bryden DW (2011) Impact of size and delay the orbitofrontal cortex of the rat. J Neurosci
on neural activity in the rat limbic corticostriatal sys- 29:8965–8976
tem. Front Neurosci 5:130 Von Neumann J, Morgenstern O (1944) Theory of games
Roesch MR, Olson CR (2005) Neuronal activity in pri- and economic behavior. Princeton University Press,
mate orbitofrontal cortex reflects the value of time. Princeton
J Neurophysiol 94:2457–2471 Wallis JD (2012) Cross-species studies of orbitofrontal
Roesch MR, Taylor AR, Schoenbaum G (2006) Encoding cortex and value-based decision-making. Nat Neurosci
of time-discounted rewards in orbitofrontal cortex is 15:13–19
independent of value representation. Neuron Wallis JD, Kennerley SW (2011) Contrasting reward sig-
51:509–520 nals in the orbitofrontal cortex and anterior cingulate
Roitman JD, Roitman MF (2010) Risk-preference differ- cortex. Ann N Y Acad Sci 1239:33–42
entiates orbitofrontal cortex responses to freely chosen Winstanley CA, Theobald DEH, Cardinal RN, Robbins
reward outcomes. Eur J Neurosci 31:1492–1500 TW (2004) Contrasting roles of basolateral amygdala
Rolls ET, McCabe C, Redoute J (2008) Expected value, and orbitofrontal cortex in impulsive choice.
reward outcome, and temporal difference error repre- J Neurosci 24:4718–4722
sentations in a probabilistic decision task. Cereb
Cortex 18:652–663
Rudebeck PH, Walton ME, Smyth AN, Bannerman DM,
Rushworth MF (2006) Separate neural pathways pro-
cess different decision costs. Nat Neurosci 9:1161–1168 Chronic Electrical Stimulation
Rudorf S, Preuschoff K, Weber B (2012) Neural correlates
of anticipation risk reflect risk preferences. J Neurosci
▶ Parkinson’s Disease: Deep Brain Stimulation
32:16683–16692
Samuelson PA (1937) A note on measurement of utility.
Rev Econ Stud 4:155–161
Samuelson PA (1938) A note on the pure theory of con-
sumer’s behaviour. Economica 5:61–71
Schultz W, Preuschoff K, Camerer C, Hsu M, Fiorillo CD, Clone Nomenclature
Tobler PN, Bossaerts P (2008) Explicit neural signals
reflecting reward uncertainty. Philos Trans R Soc Lond ▶ Nomenclature of Ion Channels (IUPHAR
B Biol Sci 363:3801–3811
Scheme)
Schultz W, O’Neill M, Tobler PN, Kobayashi S (2011)
Neuronal signals for reward risk in frontal cortex. Ann
N Y Acad Sci 1239:109–117
Stephens DW (2008) Decision ecology: foraging and the
ecology of animal decision making. Cogn Affect
Behav Neurosci 8:475–484 Closed-Loop Neuroprosthesis
Stephens DW, Anderson D (2001) The adaptive value of
preference for immediacy: when shortsighted rules ▶ Recurrent Brain-Computer Interfaces
have farsighted consequences. Behav Ecol
12:330–339
Sugrue LP, Corrado GS, Newsome WT (2004) Matching
behavior and the representation of value in the parietal
cortex. Science 304:1782–1787
Tobler PN, Fiorillo CD, Schultz W (2005) Adaptive cod-
Cochlear Implant
ing of reward value by dopamine neurons. Science
307:1642–1645 ▶ Auditory Prosthesis
Cochlear Inner Hair Cell, Model 617 C
Cochlear Inner Hair Cell, Model
Enrique A. Lopez-Poveda
Instituto de Neurociencias de Castilla y Leon,
Instituto de Investigación Biomédica de
Salamanca, Departamento de Cirugı́a, Facultad C
de Medicina, University of Salamanca,
Salamanca, Spain
Synonyms
Auditory sensory receptor cell, model; Auditory

transducer, model
Definition
An algorithmic description of how deflections of

the stereocilia in an auditory sensory receptor
cell, or cochlear inner hair cell, cause fluctuations
in the cell’s intracellular voltage or receptor
potential. Cochlear Inner Hair Cell, Model, Fig. 1 An inner hair
cell and its equivalent electrical circuit model. Stereocilia
deflection toward the tallest cilium increases the inward
flow of K+, which increases the cell’s intracellular poten-
Detailed Description tial (V). In the model, this is described by a conductance
(gA) in the apical portion of the cell’s membrane, which is
Cochlear inner hair cells (IHCs) are auditory typically assumed to be the sum of a mechanically sensi-
sensory receptor cells located in the organ of tive K+ conductance and leakage conductance (not
shown). The “excess” of intracellular K+ is reduced by
Corti. They have stereocilia that deflect following an outward K+ current through voltage-gated basolateral
mechanical motion of the organ of Corti (Fig. 1). channels. In the model, this is described by fast-activating
The fluid surrounding the stereocilia bundle, the (gK,f) and slow-activating (gK,s) basolateral K+ conduc-
endolymph, has a positive electrical potential tances. The cell’s intracellular voltage (V) controls the
inward flow of calcium (Ca2+), which controls the release
(80–120 mV), while the electrical potential of neurotransmitter vesicles from within the cell. The
within the IHC is negative at rest (around intracellular voltage also depends on other parameters:
60 mV). Stereocilia deflection toward the Et endocochlear potential; Rt, RP epithelium resistances;
tallest cilium in the bundle increases the inward CA, CB capacitance of the apical and basal portions of the
cell’s membrane, respectively; EK,f and EK,s reversal
flow of positive ions, mostly potassium (K+), and potentials of the fast and slow basolateral K+ currents,
thus increases the cell’s intracellular potential. respectively (Model adapted from Lopez-Poveda and
This mechanism effectively transduces Eustaquio-Martin 2006)
a mechanical signal (stereocilia motion) into an
electrical signal (the IHC intracellular or receptor
potential). This mechanism is important because synaptic cleft, which then triggers action poten-
the IHC receptor potential controls the influx of tials in nearby auditory nerve terminals.
calcium ions (Ca2+) through voltage-gated chan- Increasing the amount of stereocilia deflection
nels, and intracellular calcium triggers the release increases the proportion of open transducer chan-
of neurotransmitter from within IHC to the nels and the magnitude of the receptor potential.
C 618 Cochlear Inner Hair Cell, Model
The rate of increase is nonlinear due to two biophysical analogs and signal-processing ana-
mechanisms: the saturation of transducer current logs, as reviewed elsewhere (Meddis and
(Lukashkin and Russell 1998) and the outward Lopez-Poveda 2010).
flow of K+ through the IHC basolateral mem- Signal-processing models consider the IHC as
brane (Kros and Crawford 1990). While a cascade of an asymmetric, saturating, nonlinear
stereocilia deflection in one direction increases gain, which accounts for the growth of the trans-
the receptor potential, deflection in the opposite ducer current with increasing stereocilia motion,
direction closes transducer ion channels and pre- followed by a low-pass filter, which accounts for
vents the inward flow of K+ ions to the cell. This the resistor-capacitor filtering of the IHC mem-
asymmetric and saturating gating of transducer brane. The order and cutoff frequency of this
channels explains why the receptor potential filter are chosen so as to mimic as closely as
shows an AC and a DC component. It also possible the physiological low-pass characteris-
explains why the IHC is said to operate as tics of the IHC. These models are easy to imple-
a saturating, half-wave rectifier. ment and fast to evaluate and require very few
The IHC receptor potential is also determined parameters. They, however, neglect important
by the resistor-capacitor effect of the IHC mem- aspects of IHC processing and are limited in
brane and by the homeostasis of the organ of scope. For instance, they disregard the voltage
Corti. The amplitude ratio between the AC and and time dependence of activation of basolateral
DC receptor potential components decreases with K+ currents that could be significant for hearing
increasing stimulus frequency (Russell and brief and intense sounds (Kros 1996). Another
Sellick 1978), which explains the common notion shortcoming of these models is that their param-
that the IHC operates as a low-pass filter. Less eters do not represent physiological variables;
acknowledged, however, is that the AC receptor hence, these models may be inappropriate to
potential alone shows a band-pass response tuned model hearing impairment associated with IHC
at a frequency around 500 Hz (Sellick and Russell dysfunction.
1980) or 1 kHz (Dallos 1985) for low stimulus Biophysical IHC models are electrical circuit
intensities. The rising slope of the AC frequency analogs of the full organ of Corti (an early review
response probably results from receptor potential is given by Mountain and Hubbard 1996).
responding to the velocity of vibration of the Figure 1 illustrates an example model of this
organ of Corti for frequencies below approxi- type by Lopez-Poveda and Eustaquio-Martı́n
mately 200–500 Hz and to its displacement (2006). It consists of several elements that
above that frequency (Sellick and Russell 1980). describe the electrical properties of the apical
The IHC responds nonlinearly also in time. and basal portions of the IHC and its surrounding
The time-dependent activation of basolateral K+ fluids. It is assumed that the IHC intracellular
channels induces nonlinear, time-dependent space is equipotential and thus can be represented
adaptation of the receptor potential which could by a single node (V). It is also assumed that the
contribute partly to auditory nerve adaptation IHC intracellular potential is primarily controlled
(Kros and Crawford 1990; Kros 1996). by the interplay of a transducer, variable (inward)
K+ current that results from stereocilia deflec-
tions, and a basolateral (outward) K+ current
Approaches to Model the IHC Transfer that eliminates the excess of intracellular K+
Function from within the IHC. The magnitude of the trans-
ducer current is calculated from stereocilia dis-
IHC models aim at simulating the cell’s receptor placement using a Boltzmann function that
potential in response to a time-varying signal describes the gating of transducer channels. The
describing stereocilia motion. Traditionally, the excess of intracellular K+ is eliminated through
IHC transfer function has been modeled using two basolateral voltage- and time-dependent
Cochlear Inner Hair Cell, Model 619 C
nonlinear activating conductances: one (gK,f) Cross-References
with fast and one (gK,s) with slow-activation
kinetics. The activation of these two conduc- ▶ Anatomy and Physiology of the Mammalian
tances is modeled using a Hodgkin-Huxley Auditory System
approach. The reversal potential of each of the ▶ Auditory Nerve Response, Afferent Signals
currents involved is accounted for by a shunt ▶ Auditory Sensing Systems: Overview
battery (EK,s and EK,f). The capacitive effects of C
the IHC membrane are modeled with a single
capacitor (CA + CB). The receptor potential also References
depends on the endocochlear potential (Et),
which is simulated with a battery. The values of Dallos P (1985) Response characteristics of mammalian
the model parameters are typically set to physio- cochlear hair cells. J Neurosci 5:1591–1608
Guinan JJ (2012) How are inner hair cells stimulated?
logical values or are optimized to reproduce
Evidence for multiple mechanical drives. Hear Res
physiological data. This relatively simple 292:35–50
electrical circuit accounts for a wide range of Kros CJ (1996) Physiology of mammalian cochlear hair
well-reported in vitro and in vivo IHC response cells. In: Dallos P, Popper AN, Fay RR (eds) The
cochlea. Springer, New York, pp 318–385
characteristics (Lopez-Poveda and Eustaquio-
Kros CJ, Crawford AC (1990) Potassium currents in inner
Martin 2006). hair cells isolated from the guinea-pig cochlea.
There exist simpler biophysical models J Physiol 421:263–291
that consider voltage- and time-independent Lopez-Poveda EA, Eustaquio-Martin A (2006)
A biophysical model of the inner hair cell: the contri-
basolateral K+ currents (Shamma et al. 1986).
bution of potassium currents to peripheral auditory
There also exist more sophisticated models that compression. J Assoc Res Otolaryngol 7:218–235
incorporate the role of transmembrane chloride Lopez-Poveda EA, Barrios LF, Alves-Pinto A (2007) Psy-
and sodium currents and pumps in shaping the chophysical estimates of level-dependent best-
frequency shifts in the apical region of the
IHC intracellular potential (Zeddies and Siegel
human basilar membrane. J Acoust Soc Am
2004). 121:3646–3654
Lukashkin AN, Russell IJ (1998) A descriptive model of
the receptor potential nonlinearities generated by the
hair cell mechanoelectrical transducer. J Acoust Soc
Applications Am 103:973–980
Meddis R, Lopez-Poveda EA (2010) Auditory periphery:
Both signal-processing and biophysical IHC from pinna to auditory nerve. In: Meddis R, Lopez-
models have been used successfully in composite Poveda EA, Popper AN, Fay RR (eds) Computational
models of the auditory system. Springer, New York,
models of the peripheral auditory system and
pp 7–38
related applications (reviewed by Meddis Meddis R, Lopez-Poveda EA, Popper AN, Fay RR
et al. 2010). In this case, it is typically assumed (2010) Computational models of the auditory system.
that the input to the IHC model is basilar mem- Springer, New York
Mountain DC, Hubbard AE (1996) Computational anal-
brane motion rather than stereocilia motion, and a
ysis of hair cell and auditory nerve processes. In:
high-pass filter is used to couple between the two Hawkins HL, McMullen TA, Popper AN, Fay RR
(Shamma et al. 1986). This high-pass filter is (eds) Auditory computation. Springer, New York,
important to mimic the abovementioned band- pp 121–156
Russell IJ, Sellick PM (1978) Intracellular studies of hair
pass behavior of the AC receptor potential,
cells in the mammalian cochlea. J Physiol
which could be significant for low-frequency 284:261–290
hearing (Lopez-Poveda et al. 2007). Recent evi- Sellick PM, Russell IJ (1980) The responses of inner hair
dence suggests that the motion of the reticular cells to basilar membrane velocity during low fre-
quency auditory stimulation in the guinea pig cochlea.
lamina would be a more accurate mechanical
Hear Res 2:439–445
input to the IHC than basilar membrane motion Shamma SA, Chadwick RS, Wilbur WJ, Morrish KA,
(Guinan 2012). Rinzel J (1986) A biophysical model of cochlear
C 620 Cocktail Party Problem
processing: intensity dependence of pure tone Such solution regimes manifest themselves as
responses. J Acoust Soc Am 80:133–145 a large number (typically three or more)
Zeddies DG, Siegel JH (2004) A biophysical model of an
inner hair cell. J Acoust Soc Am 116:426–441 coexisting stable periodic solutions. These solu-
tions coexist at a given fixed parameter set, per-
Further Reading sist indefinitely, and can be switched by between
Geisler DC (1996) From sound to synapse: physiology of external perturbations.
the mammalian ear. Oxford University Press, New
York
Cocktail Party Problem Overview

Many biological dynamical systems exhibit
▶ Auditory Perceptual Organization multistability. Such systems have two or more
stable attractors for a fixed set of parameters.
Multistable periodic systems exhibit two or
more oscillatory attractors for a fixed set of
Coding Accuracy of Neural parameters. Examples of multistability can
Populations be found in vertebrate motoneurons (e.g.,
Hounsgaard and Kiehn 1989, where two distinct
▶ Efficient Population Coding spike firing rates of action potentials coexist),
invertebrate neurons (e.g., Lechner et al. 1996,
coexisting bursting and spiking oscillations), and
small networks of coupled neurons (e.g.,
Coding Efficiency of Neural Kleinfeld et al. 1990). Outside of neuroscience,
Populations such multistable oscillations can be found in
models of intracellular calcium oscillations
▶ Efficient Population Coding (Haberichter et al. 2001) and coupled genetic
oscillators (Koseska et al. 2007).
Bursting pacemaker neurons are single neuron
oscillators with a unique oscillatory state – they
Coexistence alternate between an active episode comprised of
a series of two or more action potentials and
▶ Multistability in Neurodynamics: Overview a silent phase characterized by a lack of the firing
of action potentials. Such oscillations are auton-
omous and exhibited by neurons in a variety of
neural systems. These include invertebrate
Coexistence of Bursting Regimes neurons (Kandel 1979), cortical neurons
(Chagnac-Amitai and Connors 1989), and
Robert Butera brainstem neurons (Smith et al. 1991). Such prop-
School of Electrical and Computer Engineering, erties are considered to be common to neurons
Laboratory for Neuroengineering, Georgia involved in either central pattern generation or
Institute of Technology, Atlanta, GA, USA neurosecretory functions.
Bursting neurons can have multiple stable
solutions. The most common and simplest case
Definition is coexistence of silent, bursting, and/or periodic
spiking solutions. The coexistence of two such
Coexisting bursting solutions are exhibited in states is often referred to as bistability. Coexis-
some models of bursting (link) neurons. tence of periodic spiking and bursting solutions
Coexistence of Bursting Regimes 621 C
has been suggested in computational (Butera 1. The solution trajectory, during the silent phase
et al. 1995) and experimental (Lechner of the burst, converges to a low-dimensional
et al. 1996) studies of Aplysia neuron R15. Burst- manifold (two dimensional in the examples
ing neurons are often described as having their studied).
dynamics on two distinct time scales, one fast, 2. A first-return map constructed on this mani-
underlying the firing of action potentials, and one fold possesses a series of discontinuous solu-
slow, underlying the slow dynamics. The slower tions, with each continuous set corresponding C
dynamics of a bursting system are often studied in to all solution trajectories with the same num-
the lower-dimensional space of the slower state ber of action potentials.
variables (e.g., Butera et al. 1997). 3. The first-return map can be decomposed into
Canavier et al. (1993) have shown for the first two distinct maps – a contractile map
time that a modeled bursting neuron can exhibit corresponding to the silent phase of the burst
a large number of stable coexisting bursting solu- on the low-dimensional manifold and an
tions. These solutions all exist at the same fixed expansive map which is discontinuous, with
parameter set. The occurrence of a given solution each distinct set of points corresponding to
is dependent on the choice of initial conditions, bursts with the same number of action
and an appropriate well-timed single perturbation potentials.
can switch the oscillation from one stable solu-
tion to another. Simulated neuromodulators can Computational Example
alter the number of stable attractors (Canavier Figure 1 illustrates an example of a multistable
et al. 1994). Two unique properties are exhibited bursting model. The model consists of two slow
by these solutions. First, the distinct bursting variables u1 and u2 and a quadratic integrate-and-
oscillations are concentric when viewed in the fire-type model for the membrane potential vari-
state space of the slow variables. Second, each able v. At this parameter set (identical to those
distinct bursting oscillation has a distinct number used in Fig. 2 of Newman and Butera 2010), three
of action potentials per cycle. Some of the stable coexisting bursting solutions exist. Panel
bursting solutions were limit cycles, while others A illustrates the (u1,u2) phase space of the three
were chaotic attractors. Appropriately timed tran- solutions, all of which are stable limit cycles. The
sient inputs are capable of switching the trajec- trajectories wind counterclockwise. When u1 is at
tory from one attractor to the other, but otherwise a sufficiently high value, the model starts firing
the attractors are stable, as well as robust to action potentials, and with each firing of an action
a modest amount of noise. potential, the slow variables u1 and u2 are dis-
cretely perturbed. Panel B illustrates the time
Necessary Components course of the three limit cycle solutions plotting
This phenomena was further investigated in v versus time. The solutions in panel B are color
a reduced 4-variable ion-channel-based model coded to correspond to the three phase-plane
(Butera 1998), and necessary components were solution projections in panel A.
identified that give rise to multistability. These The dynamics of such a model can be studied
assumptions were validated in a simplified 3- by choosing a Poincare section (LINK) that spans
variable model. Further studies (Newman and all possible trajectories during the silent phase of
Butera 2010) generalized the 3-variable model the burst cycle. One such section is chosen and
to one that is scalable to an arbitrarily large num- indicated by the line labeled S in Fig. 1. Initial
ber of stable solutions and demonstrated experi- conditions on this section y1 are systematically
mental responses of bursting neuron R15 to chosen, and subsequent crossings of the section in
transient input that are consistent with these positive (y2) and negative (y3) direction are
mechanisms. Several key components underlie recorded. Using such an approach, a return map
these solutions in all computational models can be constructed. Figure 2a plots a return map
studied: of initial conditions along Section S in Fig. 1,
C 622 Coexistence of Bursting Regimes
Coexistence of Bursting Regimes, Fig. 1 Multi- stable limit cycle solutions that coexist. (b) Plot of
stability in a minimal bursting model. Model is from voltage versus time for the three stable limit cycle
Newman and Butera (2010), with parameters identical to solutions, color coded to link to the identical solution
those used in Fig. 2 of that paper. (a) Phase-plane projec- trajectories in panel (a)
tion of the two slow variables of the model for the three
Coexistence of Bursting Regimes, Fig. 2 Mappings y1 to the corresponding crossing of S indicated by y2 on

for understanding multistability in bursting models. (a) Fig. 1. The dots represent the mapping of all points y2 to
First-return map calculated by choosing initial conditions the second cross of S indicated by y3 on Fig. 1. This map is
corresponding to positive y values on the Section S in inverted so that both maps share y2 on the y-axis. The
Fig. 1. This is the mapping y1 ! y3, indicated by the color coding associates these maps with the solution tra-
example trajectory in Fig. 1. (b) Decomposed first-return jectories in Fig. 1 that have the same number of action
map. The dotted line represents the mapping of all points potentials as all the trajectories in these maps
plotting y3 versus y1. Note the discontinuous number of action potentials as the limit cycle
nature of the maps – each continuous family of solutions in Fig. 1 are colored accordingly.
solutions corresponds to all trajectories with the To understand how the fast and slow processes
same number of action potentials during a burst. interplay, the map can be decomposed into two
The data points corresponding to the same separate maps. The map of y1 ! y2 represents
Coexistence of Bursting Regimes 623 C
the silent phase of the burst. This map is contin- Cross-References
uous, since action potentials do not fire, and is
plotted by the dotted line in Fig. 2b. The map of ▶ Multistability in Neurodynamics: Overview
y2 ! y3 highlights the effect of the action poten-
tials in breaking up the map into discontinuous
regions based on the number of action potentials
fired. This is also plotted in Fig. 2b, only inverted, References C
so that both maps share y2 on the y-axis of the
map. The intersection of these two maps is iden- Aida T, Davis P (1991) Storage of optical pulse data
sequences in loop memory using multistable oscilla-
tical to the fixed points in the first-return map of tions. Electron Lett 27:1544–1546
panel A. By studying the effects of parameters on Aida T, Otani N, Davis P (1994) Digital implementation
these maps, one can deduce how various compo- of a nonlinear delayed-feedback system. IEEE Trans
nents of the model contribute to modulating the Circuits Syst I 41:238–242
Butera RJ (1998) Multirhythmic bursting. Chaos
map into having fewer or additional intersections, 8:274–284
controlling the amount of multistability in the Butera RJ, Clark JW, Canavier CC, Baxter DA, Byrne JH
system. (1995) Analysis of the effects of modulatory agents on
a modeled bursting neuron: dynamic interactions
between voltage and calcium dependent systems.
Implications J Comput Neurosci 2:19–44
The experimental validation of such models is Butera RJ, Clark JW, Byrne JH (1997) Transient
challenging, as neurons exist not in isolation, but responses of a modeled bursting neuron: analysis
in complex networks. Newman and Butera with equilibrium and averaged nullclines. Biol Cybern
77:307–322
(2010) illustrated experimental responses from Butera RJ, McSpadden N, Mason J (2002) Theory and
bursting neuron R15 whose dynamics in design of a bio-inspired multistable oscillator. In: Pro-
response to transient inputs are consistent with ceedings of the 2002 IEEE International Symposium
the dynamics described by this class of models. on Circuits and Systems, pp 301–304. IEEE Press,
Piscataway, NJ.
They further postulate that this class of Canavier CC, Baxter DA, Clark JW, Byrne JH
multistability can exist in a large number of (1993) Nonlinear dynamics in a model neuron provide
bursting neuron models with circle/circle a novel mechanism for transient synaptic inputs to
dynamics (Izhikevich 2007) and that even if produce long-term alterations of postsynaptic activity.
only one attractor exists, such topology can Canavier CC, Baxter DA, Clark JW, Byrne JH (1994) Mul-
give rise to complex transient responses to tran- tiple modes of activity in a model neuron suggest
sient input. a novel mechanism for the effects of neuromodulators.
While speculative, this type of bursting could J Neurophysiol 72:872–882
Chagnac-Amitai Y, Connors BW (1989) Synchronized
underlie a form of memory not dependent on excitation and inhibition driven by intrinsically burst-
a slow time-scale process, but rather inherent in ing neurons in neocortex. J Neurophysiol
the stability of the attractors. Canavier 62:1149–1162
et al. (1994) showed how neuromodulators Foss J, Longtin A, Mensour B, Milton J (1996)
Multistability and delayed recurrent loops. Phys Rev
could alter the number and stability of the Lett 76:708–711
attractors. Haberichter T, Marhl M, Heinrich R (2001)
The feasibility of this idea has also been dem- Birhythmicity, trirhythmicity and chaos in bursting
onstrated in analog circuit implementations of calcium oscillations. Biophys Chem 90:17–30
Hounsgaard J, Kiehn O (1989) Serotonin-induced
these dynamics (Butera et al. 2002). Other bistability of turtle motoneurones caused by a
dynamical systems that exhibit such large-scale nifedipine-sensitive calcium plateau potential.
multistability typically rely upon the role of J Physiol Lond 414:265–282
delayed feedback and have been demonstrated Izhikevich EM (2007) Dynamical systems in neurosci-
ence: the geometry of excitability and bursting. MIT
in neural systems (Foss et al. 1996) and Press, Cambridge
engineered optical (Aida and Davis 1991) and Kandel ER (1979) Behavioral biology of aplysia. Free-
digital (Aida et al. 1994) systems. man, San Francisco
C 624 Cognition and Control
Kleinfeld D, Raccuia-Behling F, Chiel HJ (1990) Circuits Detailed Description

constructed from identified aplysia neurons exhibit
multiple patterns of persistent activity. Biophys
J 57:697–715 Cognitive tasks are full of uncertainty. From
Koseska A, Volkov E, Zaikin A, Kurths J (2007) Inherent deciding what to call a new object, to determining
multistability in arrays of autoinducer coupled genetic what causes what in the environment, to learning
oscillators. Phys Rev E 73:031916 a language, there is generally not enough infor-
Lechner HA, Baxter DA, Clark JW, Byrne JH
(1996) Bistability and its regulation by serotonin in mation to uniquely determine the answer. Bayes-
the endogenously bursting neuron R15 in Aplysia. ian models of cognition provide a principled way
J Neurophysiol 75:957–962 to deal with this uncertainty, using probability
Newman J, Butera RJ (2010) Mechanism, dynamics, and theory to evaluate hypotheses and produce
biological existence of multistability in a large class of
bursting neurons. Chaos 20:023118 responses. Below, I briefly review three key
Smith JC, Ellenberger HH, Ballanyi K, Richter DW, areas in which Bayesian models of cognition
Feldman JL (1991) Pre-Botzinger complex: have been applied: categorization, learning and
a brainstem region that may generate respiratory causal inference, and language.
rhythm in mammals. Science 254:726–729
Categorization
Cognition and Control Studies of categorization investigate how people

group stimuli and generalize from previous expe-
▶ Embodied Cognition, Dynamic Field Theory of rience to new situations. Classically, categories
were defined by a set of necessary and sufficient
conditions that determined whether an object was
a member of the category. However, all-or-none
Cognition, Bayesian Models of category membership clashes with empirical
investigations demonstrating that in human
Adam N. Sanborn categorization, objects are more or less members
Department of Psychology, University of of a category (Rosch 1973; Rosch and
Warwick, Coventry, UK Mervis 1975).
Bayesian models of cognition provide a link
between the classical definition and empirical
Synonyms observations. People might assume that
a necessary and sufficient set of conditions
Probabilistic models of cognition; Rational describe the category, but do not know the exact
models of cognition set of conditions. For example, it might be
unclear exactly what wavelengths of light define
the category of red. Uncertainty about the bound-
Definition ary can be dealt with in a Bayesian model, which
produces a graded category as a result of combin-
Bayesian models provide a principled way to deal ing sets of more or less likely conditions (Shepard
with uncertainty. In cognitive tasks the 1987; Tenenbaum and Griffiths 2001).
uncertainty is mainly in how to represent the Another classic representational question in
observed data, as there is usually not enough categorization is whether categories are
data to uniquely determine the representation. represented by prototypes or exemplars. Proto-
Bayesian models of cognition implement type representations compare new stimuli to
a general method for dealing with representa- a single abstract example, while exemplar repre-
tional uncertainty which can be applied to explain sentations compare new stimuli to all previously
human performance in various cognitive tasks. observed examples (Medin and Schaffer 1978;
Cognition, Bayesian Models of 625 C
Nosofsky 1986; Reed 1972). Each representation While the absence of observing “He carried
is appropriate for different problems, and me the ball” is not proof that this construction
Bayesian models have been used to combine is ungrammatical (it could be just rarely
these representations together, weighting repre- employed), it is evidence for the hypothesis that
sentations by how well they suit the problem carried is an exception. By adjusting the weights
(Anderson 1991). on the hypotheses given the observed data,
Bayesian models of language are able to make C
sensible inferences despite uncertainty about how
Learning and Causal Inference language should be represented (Perfors
et al. 2010).
Learning research has a long history in cognition,
with initial investigations focusing how animals
were able to associate stimuli such as tones and Cross-References
lights with particular rewards. Although early
work investigated associations, the demonstrated ▶ Perception, Bayesian Models of
power and flexibility of animal learning indicated
that more complex representations were used
(Rescorla 1988). References
This type of learning has been modeled in
a Bayesian framework as causal inference, in Anderson JR (1991) The adaptive nature of human cate-
which the animal attempts to determine the gorization. Psychol Rev 98(3):409–429
Baker CL (1979) Syntactic theory and the projection
causal structure of the environment in order to problem. Linguist Inq 10(4):533–581
make predictions (Courville et al. 2006). Learn- Courville AC, Daw ND, Touretzky DS (2006) Bayesian
ing a causal structure is extremely difficult, and theories of conditioning in a changing world. Trends
for some learning problems even infinite obser- Cogn Sci 10:294–300
Gold EM (1967) Language identification in the limit. Info
vations of outcomes are not enough to uniquely Control 10(5):447–474
determine the structure. Bayesian models deal Kemp C, Tenenbaum JB (2008) The discovery of struc-
with this uncertainty and can even take into tural form. Proc Natl Acad Sci 105(31):10687–10692
account vast spaces of possible structural forms Medin DL, Schaffer MM (1978) Context theory of classi-
fication learning. Psychol Rev 85:207–238
(Kemp and Tenenbaum 2008). Nosofsky RM (1986) Attention, similarity, and the
identification-categorization relationship. J Exp
Psychol Gen 115:39–57
Language Perfors A, Tenenbaum JB, Wonnacott E (2010) Variabil-
ity, negative evidence, and the acquisition of verb
argument constructions. J Child Lang 37(3):607–642
In language learning a difficult problem is to Reed SK (1972) Pattern recognition and categorization.
work out what is and is not grammatical, and Cogn Psychol 3:393–407
this is another problem for which too few exam- Rescorla RA (1988) Pavlovian conditioning: It’s not what
you think it is. Am Psychol 43(3):151
ples exist to uniquely determine the answer Rosch E (1973) On the internal structure of perceptual and
(Gold 1967). Learning grammaticality is diffi- semantic categories. In: Moore TE (ed) Cognitive
cult: while people almost always observe development and acquisition of language. Academic,
only correct examples of language use, they New York
Rosch E, Mervis CB (1975) Family resemblances: studies
learn both grammatical rules and also the in the internal structure of categories. Cogn Psychol
exceptions to the rule. An example of this is 7:573–605
the dative alternation: for most verbs it is gram- Shepard RN (1987) Toward a universal law of generali-
matical to alternate between the constructions zation for psychological science. Science
237:1317–1323
“He threw the ball to me” and “He threw me the Tenenbaum JB, Griffiths TL (2001) Generalization, sim-
ball,” but an exception is the verb carried ilarity, and Bayesian inference. Behav Brain Sci
(Baker 1979). 24:629–641
C 626 Coherence and Granger Causality Spectral Analysis
The injected tracer substances diffuse from the

Coherence and Granger Causality extracellular space into the neurons and are then
Spectral Analysis driven by active transport mechanisms toward the
axon terminals (anterograde tracer) or toward the
▶ Spectral Interdependency Methods soma (retrograde tracer). For each tracer injec-
tion, the core information to record is (i) the
location of the injection zone, (ii) whether the
tracer is anterograde or retrograde, (iii) which
Coherence of Local Field Potentials locations in the brain are labeled by the tracer
after the incubation period, and if available (iv)
▶ Local Field Potential, Synchrony of the laminar distribution of labeled sites and (v)
quantitative measures.
This information is collated from the literature
by manually parsing figures and textual state-
Collations of Connectivity Data on ments, in which locations are expressed as brain
the Macaque Brain (CoCoMac) region names. What makes the task challenging is
that the relevant literature spans many decades and
Rembrandt Bakker1,2 and Markus Diesmann3,4 that nomenclatures and brain parcellations
1
Nijmegen and Institute of Neuroscience and (atlases) vary widely both in time and across
Medicine (INM–6) Donders Institute, Radboud research groups. CoCoMac uses transformations
University, Nijmegen, The Netherlands based on spatial relations (Stephan et al. 2000) to
2
J€
ulich Research Centre, J€ulich, Germany map data onto a user-specified set of brain regions.
3
Institute of Neuroscience and Medicine (INM-6)
and Institute for Advanced Simulation (IAS-6),
J€ulich Research Centre and JARA, J€ulich, References
Germany
4
Medical Faculty, RWTH Aachen University, Stephan KE, Zilles K, Kötter R (2000) Coordinate-
Aachen, Germany independent mapping of structural and functional
data by objective relational transformation (ORT).
Philos Trans R Soc Lond B Biol Sci 355:37–54
Synonyms
Compilation of published tracer injection studies Collision Avoidance Models, Visually

in the Macaque brain; Macaque axonal projection Guided
matrix; Macaque brain wiring diagram
Damián Oliva
Departamento de Ciencia y Tecnologı́a,
Definition Universidad Nacional de Quilmes, CONICET,
Buenos Aires, Argentina
Collations of Connectivity Data on the Macaque
Brain (CoCoMac) is a database that contains the
results of hundreds of publications in which Definition
tracer substances are injected into macaque
brains to reveal the afferent and efferent projec- Visually guided collision avoidance models
tions of neurons in the injection zones. The goal is (VGCAMs) describe the operation of sensory
to get complete coverage of the brain, so that a and motor neuronal circuits detecting potential
full axonal projection matrix can be extracted collisions and producing motor avoidance
from the database. responses. To describe these sensory–motor
Collision Avoidance Models, Visually Guided 627 C
transformations, VGCAMs usually begin as phe- simulated using their two-dimensional projection
nomenological descriptions of responses and on a computer screen (Gabbiani et al. 1999). Such
then focus on the study of associated neural cir- looming stimuli usually represent dark circles or
cuits and biophysical mechanisms. At the sensory squares of various sizes approaching at constant
level, several biophysical processes operate in speeds on a direct collision course with the ani-
parallel to shape and tune the visual response of mal (Fig. 1a). Let l denote the object half-size and
collision detecting neurons. VGCAMs also gen- x the object position with respect to the animal C
erate descriptions about the operation of interme- eye, i.e., x > 0 before collision, and x = 0 at
diate circuits linking sensory input and motor collision. Defining t = 0 as the time of expected
output, when experimental data about these inter- collision and t < 0 as the time before collision,
mediate circuits is incomplete. the position of an object approaching at constant
velocity v is described by the equation:
Detailed Description xðtÞ ¼ v t: (1)
Visually Guided Collision Avoidance According to these conventions, the velocity

Behaviors v is negative, reflecting the fact that the object is
Most visual animals are highly effective at approaching. The angular size of the object at the
detecting and avoiding collisions, which may retina, y(t), is given by
occur either by encounters with obstacles during
navigation (Tammero and Dickinson 2002b) or l
yðtÞ ¼ 2 tan 1 : (2)
by moving objects that directly approach them vt
(e.g., predators; Rind and Simmons 1999;
Gabbiani et al. 2004; Simmons et al. 2010; The square drawn on the monitor screen
Fotowat and Gabbiani 2011; Card 2012). In (Fig. 1a) has a half-size, lscreen(t), that depends
both cases, an important visual cue to predict on the distance from the monitor to the eye of the
a potential collision is the expansion of the animal, xeye-screen, and the angular size, y(t), as
approaching surface or object. The detection of
this visual cue triggers motor programs con- lscreen ¼ xscreeneye tan ðyðtÞ=2Þ: (3)
trolled by neural circuits to generate quick and
reliable avoidance responses. To be effective, the The angular edge velocity of the object, c(t),
associated maneuvers need to be executed in is defined as
a timely and accurate manner, which implies
that the approaching object must be precisely
1 dy yðtÞ0 l=v
monitored in real time. For this reason, most cðtÞ ¼ ¼ ¼ : (4)
animals possess movement detector neurons spe-
2 dt 2 ðl=vÞ2 þ t2
cially tuned to detect objects approaching on
a collision course. Thus, visually guided collision As follows from Eqs. 2 and 4, the optical vari-
avoidance behaviors (VGCABs) provide ables y(t) and c(t) are nonlinear functions of time
a favorable model for studies of sensory coding and depend only on the ratio l/|v| between the
and sensory–motor transformations. object’s half-size l and its approach speed v. For
a given object size, a faster approach speed
Looming Stimuli implies a faster expansion rate and a smaller l/|v|
To describe the visuomotor transformation that ratio. Varying the l/|v| ratio thus allows investiga-
begins in sensory neurons and ends in the ani- tors to manipulate the temporal dynamics of the
mal’s muscles, one needs well-controlled stimu- looming stimulus.
lation conditions. If we stimulate the animal For the stimulus to be uniquely determined by
monocularly, approaching objects can be the parameter l/|v|, simulations must begin from
C 628 Collision Avoidance Models, Visually Guided
b 60 120
Screen l/|v| = 479 ms
a Virtual Object
50 θ0 =3.9⬚ 100
Animal
Edge Velocity ψ(⬚/s)

Eye
Angular Size θ (⬚)

lscreen
l 40 Angular Size 80
q
Edge Velocity
v 30 60
20 40
xeye-screen
x(t) 10 20
0 0
Initial Distance −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0
Time to collision (sec)
Collision Avoidance Models, Visually Guided, the position of the object in a reference system centered at
Fig. 1 Looming stimuli. (a) Simulation of an object’s the animal’s eye, y(t) is the angle subtended by the object
approach at constant speed. The eye of the animal is at the animal’s eye, and lscreen is the square half-size on the
stimulated from the left side by simulating squares of monitor screen. (b) Time course of y(t) and of the angular
half-size l approaching at a constant speed v towards the edge velocity c(t) as a function of time before collision for
center of the eye. The figure shows the virtual object at an approach with l/|v| = 479 ms and y0 = 3.9º. Both
two different times during the simulated approach. x(t) is functions are nonlinear in time; see Eqs. 2–4
an angular size smaller than the minimum one collision (Gibson 1958; Lee and Reddish 1981;
detected by the animal. If the object starts from Wagner 1982; Wang and Frost 1992).
a higher angular size, the initial angular size, y0, Invertebrate VGCABs vary considerably
is another independent parameter characterizing depending on the animal’s motor system, loco-
the stimulus (Oliva and Tomsic 2012). motion mode, and living environment. Some
VGCABs can be described as all-or-none thresh-
VGCABs in Invertebrates old responses, triggered when an optical variable
Invertebrates are important models for the neuro- exceeds a certain value, after which the animal
biology of visually guided behaviors due to their displays a pre-programmed motor response. This
accessible nervous systems and easily quantifi- type of VGCAB was described in the crayfish
able behavioral output (Reichardt et al. 1983; defensive reflex (Glantz 1974a) and in flies (leg
Egelhaaf et al. 1989; Borst and Haag 2002; extension reflex for landing, Borst and Bahde
Srinivasan and Zhang 2004). Phenomenological 1988; Tammero and Dickinson 2002a; saccades
descriptions usually rely on input–output func- in free flight, Tammero and Dickinson 2002b).
tions assuming that the motor output is regulated Other avoidance responses are continuously
by an input optical variable. The input optical adjusted to account for modifications of the tra-
variable may, for example, be the angular size of jectory and/or velocity of obstacles and predatory
the object, y(t), or its time derivatives like the attacks. Such behaviors have been studied in the
angular velocity, y’(t),or the angular accelera- context of animal navigation (e.g., Srinivasan and
tion, y”(t) (Fig. 1b). A simple but useful such Zhang 2004) and less frequently in the context of
variable is time-to-contact, t(t), which can be predator avoidance (Land and Layne 1995; Oliva
computed when y is small and when the approach and Tomsic 2012). Finally, there are behaviors
velocity is constant: t ~ y(t)/y’(t). Time-to- where the avoidance response is composed of
contact gives a running estimate of the time distinctive preparatory stages, each one being
before collision that could trigger a motor reac- triggered when an optical variable reaches
tion at a constant delay prior to the anticipated a particular threshold. Examples of such
multistage responses are observed in locusts (e.g., opposite direction, while frontal image expansion
Santer et al. 2005b; Fotowat and Gabbiani 2007; causes a landing response. Note that lateral
Fotowat et al. 2011), flies (Card and Dickinson expansion on one side inhibits the opposite
2008; Fotowat et al. 2009), and crabs (Nalbach expansion-detecting pathway, preventing a turn
1990; Hemmi and Pfeil 2010). from being immediately followed by another one
Figure 2 illustrates behavioral responses to in the opposite direction (Fig. 2Bv).
looming stimuli in invertebrates. They exemplify Other insects, such as locusts, also present C
ballistic-like, regulated, and multistage responses collision avoidance responses during flight.
in systems where neural and behavioral analyses When looming stimuli are presented, flying
have been carried out. locusts may, for example, produce a gliding
Ballistic-Like Escape Behaviors: Fig. 2A behavior to evade aerial predators (Santer
shows a device designed to study the collision et al. 2005a).
avoidance and landing responses to looming Continuously Regulated Escape Behaviors:
stimuli in Drosophila (Tammero and Dickinson Fig. 2C shows a system designed to study colli-
2002a). Tethered flies were placed inside an sion avoidance responses to looming stimuli of
arena in which stimuli of varying trajectories initial size y0 in the crab Neohelice granulata
and angular rates of expansion were applied (Oliva et al. 2007). The escape response consists
(c constant for each stimulus). Flies have two of a threshold-type decision for initiating the
types of responses to these stimuli: leg exten- escape run, followed by a visually regulated
sions, which are typical of the “landing mechanism for continual control of the escape
response,” and a change in their wing-stroke speed (Oliva and Tomsic 2012). The escape deci-
amplitude, characteristic of collision avoidance sion and the regulated speed mechanism can be
responses. An expanding stimulus elicits both described by a phenomenological input–output
landing and collision avoidance responses, but function fIO that describes the crab speed, vc(t),
two different pathways appear to mediate the as a function of an optical variable, u2(t d),
reactions. The azimuthal position of the stimulus evaluated d milliseconds earlier:
affects the probability of the two behaviors: the

probability of the landing response is greater vc ðtÞ ¼ f IO u2 ðt d if u2 ðt d > 0
(5)
when stimuli are applied frontally, while the vc ðtÞ ¼ 0 if u2 ðt d 0:
probability of a collision avoidance reaction is
strongest laterally. Both the latencies and sensi- The optical variable u2(t) is the product of the
tivities of the two responses differ as a function of angular increment since the onset of expansion,
the angular velocity of expansion. Dy(t) = y(t) y0, and of the stimulus angular
These experiments and knowledge about velocity, y0 (t):
motion-detection circuits in flies have inspired
a computational model to describe collision
avoidance and landing responses (Fig. 2B). u2 ðtÞ ¼ ðDyðtÞ Dyesc Þ y0 ðtÞ: (6)
According to this model, the fly estimates optic
flow during flight using a two-dimensional array The parameter Dyesc is a threshold indepen-
of Reichardt motion detectors (Fig. 2Bi) (Borst dent of looming stimulus dynamics. In addition to
and Egelhaaf 1989). The motion information is regulating the speed of escape (Eqs. 5 and 6), the
pooled such that image expansion in the two crab makes continuous adjustments to the escape
lateral fields and in the frontal fields of view is direction as a function of the path of the object
calculated (Fig. 2Bii). The outputs of these three approach (Land and Layne 1995). Thus, the pred-
expansion calculations are then temporally inte- ator is maintained laterally and the crab runs in
grated and passed through threshold detectors the opposite direction.
(Fig. 2Biii–iv). Expansion detected laterally trig- Multistage Escape Behaviors: Fig. 2D shows
gers a collision avoidance response in the a looming-evoked multistage escape behavior in
Collision Avoidance Models, Visually Guided, eliciting a behavioral response (Adapted from Tammero
Fig. 2 VGCAB in invertebrates. (A) Schematized and Dickinson (2002a)). (B) Computational model to
experimental setup for measuring a fly’s response to describe collision avoidance and landing responses in
image expansion. During tethered flight, the fly’s wing- flies (see text for details; Adapted from Tammero and
stroke amplitude and frequency are measured by optically Dickinson (2002a)). (C) Measurement of the escape
tracking the shadows cast from an infrared (IR) diode by response in the crab Neohelice granulata. Locomotor
each of the wings on an optical wing-beat analyzer. The activity was studied in a walking simulator consisting of
difference between the amplitudes of each wing stroke a Styrofoam ball that could be freely rotated by the animal.
controls the visual display, allowing the fly to orient The crab was held in position by a weightless rod attached
actively towards the position of the 15 15 square. At to its carapace that could slide up and down within a guide
periodic intervals, the square symmetrically expands, located above the animal. Locomotion was assessed by
locusts (Santer et al. 2005b; Fotowat and a linear relationship between the mean takeoff
Gabbiani 2007). The locust jump consists of sev- time and l/|v| (Fotowat et al. 2009).
eral distinct phases, during which the animal ori- Another important aspect of visually guided
ents itself away from the approaching object jumping escape responses is related to the control
using its middle legs and stores the energy of the escape’s direction. Using high-speed vide-
required for takeoff in the elastic elements of its ography, Santer et al. (2005b) showed that, after
hind legs (Heitler 1974; Burrows 1996). By mon- the start of co-contraction, locusts produce tilting C
itoring the position of the hind leg femur–tibia postural movements mediated by the pro- and
joint, Fotowat and Gabbiani (2007) showed that mesothoracic legs. These movements rotate the
after an initial flexion of the tibia, the joint moves long axis of the locust towards the direction of
to align the leg parallel to the body. This stage is jump up to 50 to either side of their long axis. In
called the initial joint movement (IJM). Subse- addition, escape circuitry allows locusts to con-
quently, the flexor and extensor muscles of the trol the timing, distance, and elevation of these
hind legs contract simultaneously to store the jumps (Santer et al. 2005b; Simmons et al. 2010).
mechanical energy required for the jump, In another series of studies on escape’s direc-
a stage known as co-contraction. This leads to tionality, Card and Dickinson (2008) described
a final femur–tibia joint movement (FJM), which the escape behavior of the fruit fly, Drosophila,
is followed by cessation of activity in the flexors, showing that the fly’s escape response is com-
also known as triggering. After this, the energy posed of different stages occurring while the
accumulated in elastic hind leg elements is object is approaching: (1) freeze, (2) a body
released and takeoff occurs. lean or leg postural adjustment, (3) wing eleva-
Figure 2E shows a linear relationship between tion, and (4) jump. They found that flies can use
the mean takeoff time and the value of the visual information to plan a jump directly away
looming stimulus parameter l/|v|. This linear rela- from a looming threat by a series of postural
tionship is equivalent to takeoff occurring with adjustments that determine the direction of their
a fixed offset relative to the time the stimulus has escape. This is accomplished by regulating the
reached a fixed angular threshold size on the position of their center of mass in relation to the
retina (Gabbiani et al. 1999). Fotowat and middle leg points of contact. Thanks to this strat-
Gabbiani (2007) studied the execution times for egy, the extension of the middle legs pushes them
the IJM and FJM, also showing a linear relation- away from the expanding visual stimulus.
ship for tIJM and tFJM as a function of l/|v|. Thus, Collectively, the results from these studies
the initial and final preparatory phases and take- suggest that the invertebrate central nervous sys-
off are triggered when the approaching object tem is capable of mapping particular azimuthal
crosses successive threshold angular sizes on the positions in visual space to a set of trajectories
animal’s retina. A similar study conducted for the in motor space. Furthermore, it was proposed
jumping escape response of Drosophila that independent descending pathways with dif-
melanogaster to looming stimuli found again ferent thresholds of activation regulate these
Collision Avoidance Models, Visually Guided, Fig. 2 behavior (IJM and FJM; see main text). The last frame
(continued) recording the rotations of the ball with two corresponds to the average timing of takeoff [time to
computer mice. The looming stimuli were applied from collision (ttc) is negative before takeoff] (Adapted from
the right (Adapted from Oliva and Tomsic (2012)). (D) Fotowat and Gabbiani (2007)). (E) Mean timing of takeoff
Looming-evoked multistage escape behavior in locust. and SD relative to expected collision (time remaining to
Drawing of four frames of a looming stimulus with collision positive before expected collision). Locusts were
l/|v| = 60 ms. Frame 1 corresponds to the first stimulus presented looming stimuli at l/|v| = 40, 60, 80, 100, and
frame. Frames 2 and 3 correspond to the average timing of 120 ms (n indicates number of jumps recorded across all
the first two preparatory stages of the jump escape animals (Adapted from Fotowat and Gabbiani (2007))
Collision Avoidance Models, Visually Guided, histograms show the mean spike rate recorded from
Fig. 3 Response of crabs’ MLG1 and MLG2 neurons 11 MLG1 and 13 MLG2 neurons from different animals
to looming and receding stimuli. (a) Responses to (one trial per stimulus per neuron). Data is subdivided into
a black looming and receding stimulus for MLG1 neurons. 100 ms bins and is plotted as mean s.e.m. The angular
(b) Same as (a) for MLG2 neurons. Sample recordings size of the looming and receding object is given in the
from a single MLG1 and a single MLG2 neuron illustrate bottom traces of each panel. Arrowheads signal the begin-
the type of responses of these neurons. Peristimulus time ning of the stimulus (Adapted from Oliva et al. (2007))
visuomotor transformations (Santer et al. 2008; The most important property of LSNs is their
Card and Dickinson 2008; Card 2012). preferential response to looming rather than to
These behavioral data raise the question: What receding or translating stimuli. For example,
information do visual neurons encode to regulate Fig. 3a shows the response to looming and reced-
these VGCABs? In the next section, we describe ing stimuli of the motion detector neurons MLG1
a set of neurons specially tuned to detect stimuli and MLG2 in the crab Neohelice granulata. Both
on a collision course. These neurons are impor- stimuli have the same value of l/|v| and therefore
tant elements of the descending pathways that act on the same number of photoreceptors but with
regulate VGCABs. a temporarily reversed photoreceptors’ activation
dynamics. As seen in the figure, the response to
Looming-Sensitive Neurons in Invertebrates expansion is much larger than that to contraction.
Neurons tightly tuned to detect objects In contrast to neurons involved in optomotor
approaching on a direct collision course, called response behaviors, such as those broadly studied
looming-sensitive neurons (LSNs), have been in the fly (Borst and Haag 2002), LSNs have
characterized in locusts (Schlotterer 1977; Rind a weak response or are inhibited by whole field
and Simmons 1992; Gabbiani et al. 1999; Gray motion (Rind and Simmons 1992; Gabbiani
et al. 2010), flies (Borst 1991; Fotowat et al. 2002; Medan et al. 2007). Furthermore,
et al. 2009), praying mantis (Yamawaki and Toh LSNs do not have a strong directional sensitivity
2009), crayfish (Glantz 1974a), and crabs (Medan across their visual field (Krapp and Gabbiani
et al. 2007; Oliva et al. 2007). 2005; Medan et al. 2007), respond to small object
movement, and have large receptive fields
(Rowell et al. 1977).
Some LSNs exhibit adaptation processes
inhibiting their response to stimuli that are con-
stant over time. This phenomenon generates
selectivity to looming stimuli (which produce
a temporal increase in excitation), as opposed to C
translating movements that produce constant
excitation (Peron and Gabbiani 2009b, c).
Another phenomenon observed in many LSNs
is the reduction of their responses to repetitive
stimulation (known as habituation; Glantz 1974b;
O’Shea and Rowell 1975; Tomsic et al. 2003;
Gray 2005). This property is expected, since if
these neurons are an important part of the warn-
ing and escape system, the animal must have
mechanisms to protect against repetitive stimula-
tion that is harmless. Furthermore, LSNs must
possess mechanisms to avoid habituation to
wide-field motion stimuli experienced during
locomotion; lateral presynaptic inhibition is Collision Avoidance Models, Visually Guided,
thought to represent one such mechanism Fig. 4 Schematics of LGMD circuitry. The lobula
(O’Shea and Rowell 1975). giant movement detector (LGMD) receives excitatory syn-
aptic inputs (circles shaded in dark gray) on a large den-
How do sensory circuits encoding looming
dritic field (A). An important characteristic of this pathway
stimuli work and what are the biophysical pro- is the presence of a lateral inhibitory network between
cesses tuning LSNs to looming stimuli? To date, excitatory afferent fibers presynaptic to the LGMD
the most complete answers to these questions (inhibitory synapses are illustrated by circles shaded in
light gray). ON and OFF feedforward inhibition impinges
have been obtained in the lobula giant motion
on two separate dendritic fields (B and C, respectively).
detector (LGMD) neuron of the locust. In the The LGMD makes a strong synapse with the descending
following section, we will see that various bio- contralateral movement detector (DCMD) that contacts
physical processes work in parallel to shape its motor and interneurons involved in generating jumps
and flight steering (Adapted from Gabbiani et al. (2005))
visual responses.
LGMD/DCMD Neurons in Locust

The lobula giant motion detector (LGMD) is an identified neuron, the descending contralateral
identified neuron in the third visual neuropil of movement detector (DCMD; Rowell 1971;
the locust optic lobe (lobula) that responds vigor- O’Shea and Williams 1974). The synaptic con-
ously to objects approaching on a collision course tact between the LGMD and the DCMD is so
(Fig. 4). The locust LGMD is an extensively powerful that LGMD’s spikes are transmitted in
studied looming-sensitive neuron (O’Shea and a 1–1 manner to the DCMD. Thus, under visual
Rowell 1976; Rowell et al. 1977; Schlotterer stimulation, each spike in the DCMD is caused by
1977; Rind and Simmons 1992; Hatsopoulos a spike in the LGMD. The DCMD large axon
et al. 1995; Judge and Rind 1997; Gabbiani travels down the animal’s contralateral nerve
et al. 1999, 2002; Krapp and Gabbiani 2005; cord and contacts motor and interneurons
Guest and Gray 2006; Santer et al. 2008; Peron involved in generating jumps and flight steering
et al. 2009; Peron and Gabbiani 2009a, b; (Burrows and Rowell 1973; O’Shea et al. 1974;
Fotowat and Gabbiani 2011; Jones and Gabbiani Simmons 1980). Therefore, the LGMD/DCMD
2012b). It makes a strong synapse with another neurons provide an excellent system for studying
the sensory–motor transformations that occur to collision, tpeak, as a function of l/|v| reveals
during looming-evoked collision avoidance a relation that can be fitted by a linear function
behaviors (Fotowat and Gabbiani 2011). with slope, a, and intercept, d (Fig. 5b). Thus, the
As in many LSNs, the LGMD responds to the LGMD’s peak firing time acts as an angular
movement of small objects irrespective of the threshold detector, signaling, with a delay of d
object’s location in the neuron’s receptive field. ms, the time at which an object reaches a fixed
This response is small and weakly directionally angular size ythres = 2 tan1(1/a) on the retina
selective. The neuron is also inhibited by whole (Fig. 5a; Gabbiani et al. 1999).
field motion and, as previously mentioned, it is Using data from electrophysiological and ana-
vigorously excited by objects approaching on tomical characterizations of the LGMD (Palka
a collision course with the animal (Schlotterer 1967; Rowell et al. 1977) and the linear relation
1977; Rind and Simmons 1992; Hatsopoulos from Fig. 5b, Gabbiani et al. (1999) proposed an
et al. 1995; Gabbiani et al. 1999). input–output function, f, describing the firing rate
The LGMD neuron has three dendritic sub- of the LGMD and DCMD neurons that should
fields that integrate different processes (O’Shea satisfy the following three assumptions:
and Williams 1974). The main subfield (1) f should depend only on the angular size y(t)
(A) receives ~15,000 cholinergic excitatory syn- and angular edge velocity c(t) of the approaching
apses arranged retinotopically. The remaining object. These two variables are thought to be
two dendritic subfields (B and C) receive ~1,000 represented in the form of inhibitory (size-
feedforward GABAergic inhibitory synapses dependent) and excitatory (motion-dependent)
(Rowell et al. 1977; Gabbiani et al. 2005). inputs to the various dendritic subfields of the
Current evidence is consistent with subfields LGMD. (2) The firing rate at time t depends
B and C providing phasic ON and OFF inhibition, only on the value of y and c at time t d.
respectively (Gabbiani et al. 2004). (3) The peak of f(t) should satisfy the linear
Figure 4 shows a proposed model for the relation between tpeak and l/|v| shown in Fig. 5b.
LGMD/DCMD circuit (Rowell et al. 1977; Rind These assumptions imply that f must be described
and Bramwell 1996; Gabbiani et al. 1999). In this by the following phenomenological model:
model, excitatory presynaptic inputs arriving at
dendritic subfield A experience lateral inhibition

f ðtÞ ¼ g cðt dÞ eayðtdÞ , (7)
(LI) that reduces LGMD responses to wide-field
stimuli. On the other hand, there are also
feedforward inhibitory (FFI) processes acting on where g() is a static nonlinearity that character-
subfields B and C, which increase with the size of izes the transformation between the kinematic
the moving object, generating the selectivity of variable = c(t d)eay(td) equation and the
this neuron to small object movement and con- firing rate.
trolling the rapid saturation to looming stimuli. Gabbiani et al. (2001) investigated the depen-
dence of this angular threshold computation on
LGMD/DCMD Responses to Looming Stimuli several stimulus parameters that alter the spatial
Figure 5a shows the typical response of the and temporal activation patterns of inputs of the
LGMD neuron to looming stimuli as the looming LGMD. It was shown that the same angular
parameter l/|v| is varied over one order of magni- threshold computation was implemented by the
tude (5–50 ms; Gabbiani et al. 1999). The overall LGMD in three different locust species and that
time course of the LGMD’s firing rate consists of the computation was invariant to changes in tar-
an increase followed by a peak and a subsequent get shape (from solid squares to solid discs), to
decrease. A feature of the LGMD response to changes in target texture (checkerboard and con-
looming stimuli is that the peak firing time concentric patterns), and to changes in contrast
sistently shifts towards collision as the parameter (Gabbiani et al. 1999). Finally, the angular
l/|v| is reduced. Plotting peak firing time relative threshold computation did not depend on object
a
80
angle θ(t) (deg)
60
40 l/|v|=50
l/|v|=10
l/|v|=30
θ thres
20 C
0 l
=10
150 |v| b y
300
Peak time relative to collision (msec)

100
δ
250
50
α
firing frequency (Hz)
0 l =30 200 δ x
1
|v|
100 150
δ
50
100
0 l =50 50
|v|
60
40 δ 0
20 −50
0 0 10 20 30 40 50 60
−300 −250 −200 −150 −100 −50 0 l (msec)
time to collision (msec) |v|
Collision Avoidance Models, Visually Guided, (with each panel corresponding to a different value of
Fig. 5 LGMD response to looming stimuli. (a) Peak l/|v|). The position of the peak firing rate is indicated by
firing rate signals that an angular threshold was exceeded the dotted lines while the dashed lines indicate the
d ms earlier. The time course of angular size is depicted in moment in time preceding the peak by d ms. Note that
the top panel for three approaches at different values of the value of ythres is independent of l/|v|, as read from the
l/|v|. The mean angular size subtended by the object d ms ordinate ythres on the top panel. (b) Relation between the
prior to the peak for each approach is indicated by the time of peak firing rate and l/|v| is linear. The parameters d
crosses. The corresponding time course of the instanta- and ythres are obtained from the linear fit (see inset)
neous firing rate is illustrated in the bottom three panels (Adapted from Gabbiani et al. (1999))
approach angle, over at least 135 in the horizon- inhibitory inputs to the LGMD, as summarized
tal plane. in Eq. 8:
These findings raise the question: How does
the LGMD implement the multiplicative opera- f ðtÞ ¼ g½exp½logðcðt dÞÞ a yðt dÞ: (8)
tion between angular edge velocity c and angular
size y in Eq. 7? The proposed scheme is based on The model of Eqs. 7 and 8 accurately
excitation being logarithmically related to angu- describes the LGMD/DCMD firing rate around
lar velocity and added to inhibition signaling the time of maximum activity (tpeak). However,
angular size, followed by an approximate expo- for the prediction to be accurate at all times, it is
nentiation at the spike initiation zone (SIZ) that necessary to relax the second assumption and
converts membrane voltage to firing rate add a dependency to the static nonlinearity g,
(Gabbiani et al. 2002). Together, these steps with respect to the parameter l/|v| (Gabbiani
result in a multiplication of the excitatory and et al. 1999):
h i
f ðtÞ ¼ gl=jvj cðt dÞ eayðtdÞ : (9) Experimental Characterization and
Computational Modeling of Presynaptic
Circuitry and Feedforward Excitation to the
This dependence of the g on l/|v| is attributable LGMD Subfield A
both to the properties of the conductances shap- The LGMD receives retinotopic excitatory input
ing the response of the LGMD and to its presyn- from an entire visual hemifield (see Figs. 4 and
aptic elements that were not taken into account 6A). Using electrophysiology, in vivo imaging,
in the derivation of Eqs. 7 and 8 (Gabbiani and pharmacology, it has been shown that
et al. 1999). Although the proposed computation this excitatory projection activates calcium-
(Eqs. 7–9) has a high predictive power and pro- permeable nicotinic acetylcholine receptors and
vides insight into the LGMD computation, we preserves retinotopy down to the level of a single
will see next that many biophysical processes ommatidium (Peron et al. 2009). Figure 6Bi
act in parallel to shape and tune the LGMD shows the proposed mechanism that results in
response to looming stimuli. speed sensitivity of the LGMD’s excitatory
a facet b (i)
a
tin
size dependent inhibition re
Ph Vm
(i)
speed dependent
Stimulus Photoreceptor RF
excitation
Vm
LMC (ii)
(ii) (iii) (iv)

Med single
Strength
Edge
Resp
facet gexc
Speed (σ x/σ t ) activation
Strength
LGMD Im
(iii) Lum Change Edge Speed
Cum Gauss Fit (t, σt)
Vm
(iv)
Resp Peak
Time
Lum Change
Duration (4σ t ) activation
time
SIZ
Lum Change
Duration
Collision Avoidance Models, Visually Guided, terminating onto the LGMD’s large excitatory dendritic
Fig. 6 Circuitry, physiology, and modeling of the field (green rake). (iii–iv) Voltage and current clamping
excitatory presynaptic circuit to the LGMD. (A) the LGMD to resolve postsynaptic currents. In addition,
Recordings from photoreceptors, LMCs, and the LGMD the LGMD receives inhibition related to stimulus size
in both voltage (Im) and current (Vm) clamp to single-facet onto distinct dendrites (red line). (B) Model of velocity
stimuli (right, cyan traces) enable analysis of how edge encoding by single photoreceptors and its relation with the
speed is coded along the pathway. Five facets on the eye LGMD excitatory synaptic conductance’s strength and
(top, dashed square) mapped onto five distinct locations latency. (Bi) As a dark object crosses the Gaussian-shaped
within the excitatory dendritic field (matching square and receptive field, it produces a luminance (Lum) change
hues) symbolize its retinotopic organization. whose duration depends on stimulus speed (Bii). (Biii)
(Ai) Photoreceptors (Ph) hyperpolarize (right, cyan The luminance change duration (LD = 4 st) and the
trace) as dark edges cross their receptive fields over single edge velocity c determine the LGMD excitatory synaptic
facets on the eye (top, yellow circles). (Aii) Subsequently, conductance’s strength and latency (Biv) (Adapted from
that information propagates along an excitatory pathway Jones and Gabbiani (2010, 2012b))
comprising LMCs and medullary interneurons (Med),
input (Jones and Gabbiani 2010). When information, the luminance input to each facet
a translating edge stimulus is applied, the figure’s can be calculated as a function of time, and,
edges move with angular velocity c through the using the model for the excitatory conductance
array of facets on the animal’s eye (Fig. 6Ai). The (Fig. 6Biii–iv), the excitatory postsynaptic cur-
faster the movement of a dark edge across a facet rent in the LGMD compartment associated with
(Fig. 6Bi), the more rapidly the luminance facet number n, iexc,n, is described by
encountered by the underlying photoreceptors C
will decrease (Fig. 6Bii). More specifically, an iexc, n ðtÞ ¼ gexc, n ðtÞ ðV n ðtÞ Eexc Þ, (10)
edge traveling at a constant speed c, and sweep-
ing through a Gaussian receptive field of width where Vdend,n is the dendritic compartment poten-
sx, will produce a local, time-dependent lumi- tial in the region associated to the facet, Eexc is the
nance change that has the functional form of excitatory synaptic reversal potential, and gexc,n is
a cumulative Gaussian with a duration deter- the excitatory conductance (current and conduc-
mined by its standard deviation (SD), st = sx/c. tance are expressed per unit area).
Below photoreceptors, the most likely next Finally, another effect related to the latency
stage in the visual pathway is the large monopolar reduction for faster luminance changes, defined
cells (LMCs) in the lamina (Fig. 6Aii; James and as the synchronization effect, was characterized
Osorio 1996; Rind and Bramwell 1996). Record- in the LGMD (Jones and Gabbiani 2010). By
ings from LMCs revealed that these neurons selectively stimulating an array of neighboring
responded to these luminance changes with photoreceptors, latency reduction (Fig. 6Biii)
increasing depolarization proportional to 1/st favors summation of inputs activated sequen-
yielding a local response proportional to the stim- tially throughout the looming sequence, making
ulus speed c = sx/st and consistent with the high- the neuron maximally sensitive to looming stim-
pass filtering properties reported in other insect uli. Therefore, the LGMD’s selectivity arises par-
species (Laughlin and Hardie 1978). Another tially from presynaptic synchronization of a large
effect was that the latency of the peak response population of inputs during a looming stimulus
in LMCs cells decreased as the stimulus duration and subsequent postsynaptic detection within the
became shorter (or as the edge velocity was LGMD neuron.
increased).
Because recording from the transmedullary Experimental Characterization of Lateral
neurons that synapse onto the LGMD is not yet Inhibition (LI) and Feedforward Inhibition (FFI) in
technically feasible, Jones and Gabbiani LGMD to Looming Stimuli
(2010) measured the effect of that stage of the To analyze separately the action of LI and FFI,
excitatory pathway by voltage clamping the Rowell et al. (1977) conducted intracellular
LGMD to resolve excitatory postsynaptic cur- recordings between dendritic field A and the
rents elicited by single-facet stimulation LGMD’s SIZ. When measuring membrane volt-
(Fig. 6Aiii–iv). Based on these experiments, age in this region, IPSPs associated with the FFI
Jones and Gabbiani (2012b) described the excit- could be observed to evaluate their effect on the
atory synaptic conductance’s strength and LGMD firing rate.
latency as a function of the edge velocity c and Through this approach, it was shown that LI or
of the luminance change duration (LD = 4 st), FFI can be selectively activated with a translating
respectively (Fig. 6Biii). square-wave grating stimulus. If the grating has
To construct a computational model of the wavelength l and angular velocity o, the contrast
LGMD excitatory current to looming stimuli, it frequency is defined as fc = o/l. For contrast
is necessary to incorporate information about the frequencies below 12 Hz, there was a reduction
spatial distribution of ommatidia on the locust in the LGMD firing rate, but no IPSPs were
eye (Krapp and Gabbiani 2005). Based on this observed. In this range of frequencies, the grating
stimulus preferentially activates presynaptic computational model of LI was developed in

LI. For frequencies greater than 12 Hz, IPSPs Rind and Bramwell (1996).
were observed, which indicated the activation
of FFI. Spike-Frequency Adaptation Process and Its
A subsequent experiment was designed to ana- Influence in the LGMD Tuning to Looming Stimuli
lyze the influence of LI on looming stimuli Another biophysical process that affects the
(Gabbiani et al. 2002). In this experiment, tuning of the LGMD to looming stimuli is
looming and a low-frequency wide-field drifting spike-frequency adaptation (SFA). SFA is
grating were simultaneously applied. As indi- a reduction of a neuron’s firing rate to a
cated above, LI was thus preferentially activated. stimulus of constant intensity. The LGMD
By comparing the difference between the exhibits rapid adaptation to both constant current
response to this stimulus and the response to injection (Gabbiani and Krapp 2006) and to trans-
a looming stimulus alone, it was shown that the lating objects moving at constant angular speed
effect of LI was mainly produced for angular (Peron and Gabbiani 2009a, b). Using pharma-
sizes y <23 . cology and Ca2+ imaging techniques, Peron and
As we saw earlier, the phenomenological Gabbiani (2009a) showed that SFA is mediated
model (Eqs. 7–9) assumes that inhibition depends by a Ca2+-dependent K + current IAHP governed
on the angular size, exp(a y), and is mainly by a large influx of calcium in the region between
produced by FFI on LGMD’s dendritic fields the spike initiation zone (SIZ) and the excitatory
B and C. To test this hypothesis experimentally, dendritic field (see Fig. 6A). Intracellular block of
FFI was inactivated through the local application this current minimally affected the LGMD’s
of a GABA antagonist (picrotoxin) in these den- response to looming stimuli but substantially
dritic subfields (Gabbiani et al. 2002, 2005). increased its response to translating ones. Thus,
After blocking feedforward inhibition by means SFA contributes to the neuron’s tuning by selec-
of picrotoxin injection, the duration of LGMD tively decreasing its responses to non-preferred
responses to looming stimuli was greatly stimuli, such as translating and receding stimuli
prolonged. This manipulation also increased the (Peron and Gabbiani 2009b). Peron and Gabbiani
number of spikes and peak firing rates relative to (2009a, b) constructed a computational model
control trials. These results suggested that with three compartments that simulate the den-
feedforward inhibition plays an important role drites, the region where IAHP was localized, and
in controlling feedforward excitation. Based on the axon. This computational model could
these experimental results, it was concluded explain several aspects of the LGMD responses
that the multiplicative operation is most likely to visual stimuli, current pulses, and pharmaco-
caused by the interaction between postsynaptic logical treatments (Peron and Gabbiani 2009a).
feedforward excitation and inhibition within the Furthermore, it was shown that SFA acts as
LGMD itself, rather than through presynaptic a selectivity filter for specific temporal input pro-
inhibition. files. In the context of time-varying stimuli, adap-
Although LI does not have a significant influ- tation is least effective at suppressing spikes in
ence on the LGMD response to looming stimuli response to stimuli where not only the first but
for angles greater than 23 , this inhibition is also the second derivative of the input’s intensity
important to protect the LGMD against habitua- is positive. Looming stimuli exhibit precisely this
tion to visual stimuli that the animal experiences property, and this may explain the insensitivity of
while walking or flying (O’Shea and Rowell the LGMD’s looming response to SFA. In con-
1975). Furthermore, it can strongly influence the clusion, SFA provides an example of how
response of this neuron to the movement of small a membrane conductance can tune a neuron for
stimuli (Rind and Simmons 1999; Peron and specific time-varying stimuli (Peron and
Gabbiani 2009b). A biologically inspired Gabbiani 2009b).
Computational/Biophysical Model of LGMD different compartments of the model were inter-
Response to Looming Stimuli related and how they were related to the angular
The biophysical processes involved in the LGMD speed and size of the looming stimulus. To eval-
computation described phenomenologically uate the signal transformations in the context of
(Eqs. 7–9) could be understood in a detailed the proposed computation, Jones and Gabbiani
computational/biophysical model (Jones and (2012b) first quantified the relationship between
Gabbiani 2012b). The model took into account angular velocity, the total excitatory synaptic C
the morphology of the neuron (Peron et al. 2007) conductance to the LGMD model, and the den-
and the action of different intrinsic and synaptic dritic and SIZ membrane voltage.
currents acting on the LGMD (Fig. 6A). The Figure 7c shows the normalized excitatory
model neuron had a rake-like dendritic tree, synaptic conductance and both the dendritic and
where it received excitatory synaptic inputs (see SIZ Vm during looming. Because the dendritic
Eq. 10). It also possessed active conductances and SIZ Vm are similar to one another, it was
implementing realistic SFA within the SIZ. In concluded that the velocity tuning transformation
addition to the time-varying excitatory input, occurs locally within the dendrites of the model
the model included inhibitory inputs stimulated and that the SIZ Vm was a close reflection of the
with a fixed magnitude and delay after the stim- local dendritic Vm. Using the data in Fig. 7c, the
ulus had reached each new facet. Therefore, the excitatory conductance and resulting Vm at the
size-dependent inhibitory synaptic current acting SIZ were plotted against each other (Fig. 7d),
on the LGMD (Fig. 6A), iinh,was described by and from this figure, we observe that the transfor-
mation between the two signals can be described
iinh ðtÞ ¼ ginh ðtÞ ðV m ðtÞ Einh Þ: (11) as a logarithmic compression. The light dashed
lines show logarithmic fits to the curves and the
Here, Vm is the dendritic compartment poten- inset shows that the relationship is approximately
tial, Einh is the inhibitory synaptic reversal poten- linear when displayed on a logarithmic abscissa.
tial, and ginh is the inhibitory conductance. Due to To understand how the logarithmic compression
these assumptions, the synaptic inhibition was occurs, Jones and Gabbiani analyzed the den-
approximately proportional to the area covered dritic membrane potential, concluding that the
by the looming stimulus, consistent with previous cause of the compression was due to a reduction
descriptions of feedforward inhibition onto the in the driving force of synaptic excitatory cur-
LGMD (Hatsopoulos et al. 1995; Gabbiani rents as an increasing number of synapses are
et al. 2005). Finally, the SIZ compartment activated by the looming stimulus (see Eq. 10).
contained Hodgkin–Huxley active conductances On the other hand, simulations showed that,
for spike production, and synaptic noise in the unlike the excitatory process, the size-dependent
model was adjusted to fit the variability of exper- inhibitory conductance was more subtly
imental data to single-facet stimulation (Jones transformed. Since the reversal potential of inhib-
and Gabbiani 2012a, b). itory inputs were only 10 mV below the resting
As seen in Fig. 7a and b, the model displayed potential, the best way to measure the effect of
looming responses similar to those observed inhibition on the output of the model was to
in vivo and reproduces the linear relationship consider the difference in Vm of simulations
between tpeak and l/|v| described above. But, with and without inhibition; they called this dif-
how does the LGMD implement the logarithmic ference: Vm inhibitory influence. Figure 7e shows
transformation of angular edge velocity c and the its time evolution with the angular size of the
multiplication with angular size y in Eqs. 7–9? stimulus. This time course was consistent with
The biophysical/computational model the feedforward inhibition observed experimen-
addressed these questions by examining how sig- tally in the firing rate of LGMD (Gabbiani
nals at different levels of the circuit and in et al. 2005).
Collision Avoidance Models, Visually Guided, the brighter solid traces are the Vm near the SIZ. For
Fig. 7 Computational/biophysical model of LGMD comparison, the time course of gexc(t) is plotted. (d) The
response to looming stimuli. (a) Spiking responses of relationship between gexc(t) and the resulting membrane
the LGMD model to looming stimuli. Top traces show the potential near the SIZ. Colors show responses to differ-
stimulus angular size for l/|v| = 10, 40, and 80 ms. The ent looming stimuli as in (a). Solid lines show the
rasters show spiking in 50 simulated presentations for observed relationship, while dashed lighter lines show
each l/|v| value, and bottom traces show the mean instan- fitted functions of the form Vm = A log(B gexc). The
taneous firing rates of the model computed by convolution inset shows the data on a logarithmic x-axis, demon-
of individual spike trains with a Gaussian filter (SD, strating that the relationship is approximately linear
20 ms). Envelopes show SEM. (b) The timing of the with a slight dependence on the l/|v| parameter.
model LGMD peak firing rate (circles, error bars show (e) Time course of the Vm inhibitory influence is
SEM) depends linearly on l/|v| (black dashed line). This shown, plotted with the angular size of the stimulus.
corresponds with the peak response occurring 14.7 ms Signals are all normalized to their maximum across l/|v|
after the stimulus has reached a constant angular size of values. (f) Instantaneous firing rate function for the
23.7 . The gray line shows the best fit linear relationship model. The solid black line depicts the averaged rela-
found in Jones and Gabbiani (2010). (c) Membrane poten- tionship between median filtered SIZ Vm and IFR, while
tial over time in multiple locations. Lighter traces show the dotted lines are one SD away from it. The red
the membrane potential over time in a dendritic compart- dashed line shows the best fitting power-law function
ment corresponding to the center of the visual field, while (Adapted from Jones and Gabbiani (2012b))
The observation that the excitatory synaptic Collision Avoidance Models, Visually Guided,
conductance activated by a looming stimulus is Table 1 Phenomenological models in LSNs
compressed nearly logarithmically in the mem- Optical
brane potential of the model, while the inhibitory Animal Neuron variable References
influence remains much closer to signaling angu- Crayfish Jittery y0 Glantz (1974a)
movement
lar size, matches the proposed phenomenological fibers
model that excitatory and inhibitory inputs are Locust LGMD/DCMD = y0 Gabbiani
C
effectively multiplied by an addition in logarith- eay et al. (1999)
mic coordinates. The subsequent approximate Pigeon Neurons , r, t r = y0 Sun and Frost
exponentiation is then implemented via a t = y/y0 (1998)
nonlinear transformation at the SIZ of the = y0
LGMD (Eq. 8; Gabbiani et al. 2002). eay
Figure 7f shows the transformation between Goldfish Mauthner cell k=y Preuss
eay et al. (2006)
membrane potential at the SIZ and instantaneous
firing rate. This relation was fit with a power law,
suggesting that this supralinear transformation
plays a role similar to the exponentiation pro- are the angular velocity y0 (t) and the angular size
posed in Eq. 8 (Gabbiani et al. 2002). y(t). Using this information, an animal could
In conclusion, this computational model compute in parallel several variables of interest,
(Jones and Gabbiani 2012b) suggests a plausible such as the time to collision t and the function.
biophysical implementation of the computation This seems to be the case of pigeons, in which
performed by the LGMD: the total excitatory extracellular recordings in the optic tectum
input to the LGMD is expansively dependent on showed three types of neurons (called r, t,
angular velocity but is logarithmically com- and ) that encode information on collisions
pressed by local reduction of the driving force (Sun and Frost 1998; Laurent and Gabbiani
in the dendrites, resulting in membrane potential 1998). These results indicate that the pigeon’s
changes that are a saturating function of velocity. brain reconstructs objects approaching using sev-
In the main dendritic trunk, this excitatory signal eral computations. Each variable provides
is summed with an inhibitory signal related to the a different piece of information about the state
stimulus’ angular size; in the biophysical model, of the environment, and the animal presumably
this inhibitory influence tracks angular size makes an informed decision on the basis of these
sublinearly. The resulting membrane potential is different inputs. Moreover, given the results of
transformed by a final nonlinear, expansive Table 1, it will be interesting to investigate
spike-thresholding step at the LGMD’s SIZ. The whether the biophysical processes described in
output firing rate thus reflects a multiplication the LGMD neuron (see previous section) can be
between the stimulus’ angular velocity and generalized to other species.
a negative exponential of its angular size.
Linking Sensory Coding and Behavioral
Phenomenological Models in Other LSNs Responses
Studies to characterize other LSNs have been In recent years, significant progress has been
carried out in different animal species. Table 1 made in understanding the function of sensory
summarizes some phenomenological models circuits and in the phenomenological description
describing LSN responses to collision stimuli. of behavioral and neuronal responses to looming
These models use input–output functions of the stimuli. However, important aspects are still
type R(t) = g(z(t d)), where z is an input optical unknown about the mechanisms that nervous sys-
variable, R(t) is the neuronal firing rate, and d is tems use to decide whether, when, in which direc-
the delay time. The examples in Table 1 show tion, and how intensely to perform an escape
that the optical variables most commonly used response (Santer et al. 2008; Fotowat and
Gabbiani 2011; Hemmi and Tomsic 2012; Card DCMD response affects the final motor output on
2012; Herberholz and Marquart 2012). a trial-by-trial basis (Fotowat et al. 2011).
One approach to study these issues (generally These experiments found little evidence for an
used when experimental data about intermediate involvement of the DCMD in the initial prepara-
circuits is incomplete) is to use VGCAMs to tory movements leading to the jump. The DCMD
propose feasible descriptions about the operation firing rate threshold predicted only 36 % of the
of intermediate circuits linking sensory input and variance of co-contraction onset, indicating that
motor output. For example, Glantz (1974a) pro- other neurons still play an important role at this
posed a model for the crayfish’s defensive reflex. stage. This result is not surprising since, as we
Using the fact that crayfish LSNs encode the saw in section 2, escape responses can require
angular velocity of expansion (see Table 1), the complex calculations prior to initiation.
existence of a firing rate integrator was proposed, Nevertheless, the DCMD activity played an
such that the behavioral response was displayed increasingly important role as collision became
when an accumulated number of spikes exceeded imminent. For instance, after co-contraction
a threshold value. This model could explain that onset, the number of DCMD spikes precisely
the behavioral response was produced when the predicted the hind leg’s motoneuron activity.
angular size increased by a fixed value of 8 . The number of DCMD spikes from
Another example of this approach is the com- co-contraction onset was highly predictive of
putational model for eliciting collision avoidance jump occurrence. Additionally, the time
and landing responses in flies (Tammero and of DCMD peak firing rate could predict 75 % of
Dickinson 2002a, b) (see Fig. 2b). It is notewor- the trial-to-trial variability of the jump time from
thy that these VGCAMs can be used to frame the DCMD peak firing time. In conclusion, the
hypotheses about circuitry, which can be directly transformation of sensory activity into the
tested by current or future experiments. motor program leading to visually guided jumps
To study experimentally how sensory informa- appears to rely on distinct attributes of a neuron’s
tion control decisions and generate different motor time-varying discharge from a specific time point
escape patterns, it is extremely useful to have an (co-contraction onset).
experimental preparation which allows the mea- Is the DCMD the sole source of looming infor-
surement of neural activity in sensory neurons and mation to the jump motor circuitry, or are there
interneurons, electromyographic activity in impor- other parallel visual pathways that could generate
tant muscles, and behavioral response in a freely escapes? Ablating the DCMD altered the flexion
behaving animal. In this direction, Santer timing but did not eliminate it, a fact that
et al. (2008) studied looming behavior in tethered suggested that pathways beyond the DCMD
locusts walking on a ball, where locusts perform must necessarily be involved (Santer
pre-takeoff leg adjustments, but do not complete et al. 2008). Locusts possess, besides the
the final hind leg extension for the takeoff jump. DCMD, a second pathway called the descending
Nevertheless, using simultaneous high-speed vid- ipsilateral motion detector (DIMD) found in the
eography and extracellular recording, it was pos- ipsilateral nerve cord, which has near identical
sible to correlate DCMD responses with the timing looming responses to those of the DCMD
of hind leg flexion. (Fotowat et al. 2011). In the absence of these
Recently, a study which simultaneously mea- two neurons, locusts still flex their hind legs in
sured the DCMD neuron activity, muscle activity preparation for co-contraction; however, they
in the hind legs, and the animal’s acceleration rarely take off, and when they do, it is after
was conducted in freely behaving locusts projected collision. These results suggest that
(Fotowat et al. 2011). It used a miniature telem- the activity of the DCMD or DIMD is likely not
etry system to transmit data output to a remote necessary for the initial flexion stage but plays
receiver wirelessly (Harrison et al. 2011). This a critical role in generating correctly timed jump
system permitted study of how variability in the takeoffs.
Conclusion Cross-References
Understanding how sensory stimuli are processed
by the nervous system to generate complex ▶ Decision-Making, Models
behaviors is a central goal of systems and com- ▶ Optic Flow Processing
putational neuroscience. In view of their biolog- ▶ Reduced Morphology Models
ical importance, visually guided collision ▶ Visual Motion Detection in Drosophila
avoidance behaviors (VGCABs) provide ▶ Visual Processing in Free Flight C
a favorable model for studies of sensory coding,
sensory–motor transformations, and decision-
making processes, as many animals have devel- References
oped specialized neural circuitry to sense and
execute them. Furthermore, these behaviors Borst A (1991) Fly visual interneurons responsive to
must be controlled by rather simple neural cir- image expansion. Zool Jb Physiol 95:305–313
Borst A, Bahde S (1988) Spatio-temporal integration of
cuits to generate quick and reliable avoidance motion: a simple strategy for safe landing in flies.
responses. At the sensory level, VGCAMs usu- Naturwissenschaften 75:265–267
ally begin as phenomenological descriptions Borst A, Egelhaaf M (1989) Principles of visual motion
and then focus on the study of associated neural detection. Trends Neurosci 12:297–306
Borst A, Haag J (2002) Neural networks in the cockpit of
circuits and biophysical mechanisms for the fly. J Comp Physiol A Neuroethol Sens Neural
implementing these computations. In this sense, Behav Physiol 188:419–437
an outstanding example is the case of the locust’s Burrows M (1996) The neurobiology of an insect brain.
LGMD/DCMD neurons, whose properties have Oxford University Press, Oxford
Burrows M, Rowell CHF (1973) Connections between
been studied for over 40 years (Burrows 1996). descending visual interneurons and metathoracic
Using experimental techniques such as electro- motoneurons in the locust. J Comp Physiol
physiology, pharmacology, calcium imaging, and 85:221–234
computational modeling techniques, this work Card GM (2012) Escape behaviors in insects. Curr Opin
Neurobiol 22:180–186
has shown that different biophysical processes Card GM, Dickinson MH (2008) Visually mediated motor
work in parallel to tune the LGMD response to planning in the escape response of Drosophila. Curr
looming stimuli. Biol 18:1300–1307
At the behavioral level, studies in inverte- Egelhaaf M, Borst A, Reichardt W (1989) Computational
structure of a biological motion-detection system as
brates showed that VGCABs vary considerably revealed by local detector analysis in the fly’s nervous
depending on the animals’ motor structure and system. J Opt Soc Am A 6:1070–1087
living environment, ranging from ballistic-like or Fotowat H, Gabbiani F (2007) Relationship between the
multistage to continually adjusted behaviors. For phases of sensory and motor activity during a looming
evoked multistage escape behavior. J Neurosci
this reason, we need comparative studies across 27:10047–10059
different species to determine the generality of Fotowat H, Gabbiani F (2011) Collision detection as
the discovered mechanisms. VGCABs require a model for sensory-motor integration. Annu Rev
computations integrating information from dif- Neurosci 2011(34):1–19
Fotowat H, Fayyazuddin A, Bellen HJ, Gabbiani F (2009)
ferent sensory channels to decide whether, A novel neuronal pathway for visually guided escape
when, in which direction, and how intensely to in Drosophila melanogaster. J Neurophysiol
perform an escape response. Therefore, it will be 102:875–885
interesting to study how downstream circuits Fotowat H, Harrison RR, Gabbiani F (2011) Multiplexing
of motor information in the discharge of a collision
interpret these sensory messages in order to detecting neuron during escape behaviors. Neuron
make an appropriate motor decision. Finally, 69:147–158
VGCAMs can be used to frame hypotheses Gabbiani F, Krapp HG (2006) Spike-frequency adaptation
about circuitry, which can be directly tested by and intrinsic properties of an identified, looming sen-
sitive neuron. J Neurophysiol 96:2951–2962
techniques such as calcium imaging and electro- Gabbiani F, Krapp HG, Laurent G (1999) Computation of
physiological recordings in freely behaving object approach by a wide-field, motion-sensitive neu-
animals. ron. J Neurosci 19:1122–1141
Gabbiani F, Mo C, Laurent G (2001) Invariance of angular Jones PW, Gabbiani F (2012a) Impact of neural noise on a
threshold computation in a wide-field looming- sensory-motor pathway signaling impending collision.
sensitive neuron. J Neurosci 21:314–329 J Neurophysiol 107:1067–1079
Gabbiani F, Krapp HG, Koch C, Laurent G (2002) Multi- Jones PW, Gabbiani F (2012b) Logarithmic compression
plicative computation in a visual neuron sensitive to of sensory signals within the dendritic tree of a
looming. Nature 21:320–324 collision-sensitive neuron. J Neurosci 23: 8539–8546.
Gabbiani F, Krapp HG, Hatsopoulos N, Mo CH, Koch C, 32(14):4923–4934
Laurent G (2004) Multiplication and stimulus invari- Judge S, Rind F (1997) The locust DCMD, a movement-
ance in a looming-sensitive neuron. J Physiol Paris detecting neurone tightly tuned to collision trajecto-
98:19–34 ries. J Exp Biol 200:2209–2216
Gabbiani F, Cohen I, Laurent G (2005) Time-dependent Krapp HG, Gabbiani F (2005) Spatial distribution
activation of feed-forward inhibition in a looming sen- of inputs and local receptive field properties of a
sitive neuron. J Neurophysiol 94:2150–2161 wide-field, looming sensitive neuron. J Neurophysiol
Gibson JJ (1958) Visually controlled locomotion 93:2240–2253
and visual orientation in animals. Br J Psychol Land M, Layne JE (1995) The visual control of behaviour
49:182–194 in fiddler crabs. II. Tracking control systems in court-
Glantz RM (1974a) Defense reflex and motion detector ship and defence. J Comp Physiol A 177:91
responsiveness to approaching targets: the motion Laughlin SB, Hardie RC (1978) Common strategies for
detector trigger to the defense reflex pathway. light adaptation in the peripheral visual systems of fly
J Comp Physiol 95:297–314 and dragonfly. J Comp Physiol 128:319–340
Glantz RM (1974b) Habituation of the motion detectors of Laurent G, Gabbiani F (1998) Collision-avoidance:
the crayfish optic nerve. J Neurobiol 5:489–510 nature’s many solutions. Nat Neurosci 1:261–263
Gray JR (2005) Habituated visual neurons in locusts Lee DN, Reddish PE (1981) Plummeting gannets:
remain sensitive to novel looming objects. J Exp Biol a paradigm of ecological optics. Nature 293:293–294
208:2515–2532 Medan V, Oliva D, Tomsic D (2007) Characterization of
Gray JR, Blincow E, Robertson RM (2010) A pair of lobula giant neurons responsive to visual stimuli that
motion-sensitive neurons in the locust encode elicit escape behaviors in the crab Chasmagnathus.
approaches of a looming object. J Comp Physiol J Neurophysiol 98:2414–2428
A 196:927–938 Nalbach HO (1990) Visually elicited escape in crabs. In:
Guest B, Gray J (2006) Responses of a looming-sensitive Wiese K, Krent WD, Tautz J, Reichert H, Mulloney B
neuron to compound and paired object approaches. (eds) Frontiers in crustacean neurobiology. Birkhauser
J Neurophysiol 95:1428–1441 Verlag, Basel, pp 165–172
Harrison RR, Fotowat H, Chan R, Kier RJ, Olberg R, O’Shea M, Rowell CHF (1975) Protection from habitua-
Leonardo A, Gabbiani F (2011) Wireless neural/ tion by lateral inhibition. Nature 254:53–55
EMG telemetry systems for small freely moving O’Shea M, Rowell CHF (1976) The neuronal basis
animals. IEEE Trans Biomed Circuits Syst of a sensory analyser, the acridid movement
5:103–111 detector system. II. Response decrement, conver-
Hatsopoulos N, Gabbiani F, Laurent G (1995) Elementary gence, and the nature of the excitatory afferents to
computation of object approach by wide-field visual the fan-like dendrites of the LGMD. J Exp Biol
neuron. Science 270:1000–1003 65:289–308
Heitler WJ (1974) The locust jump. J Comp Physiol O’Shea M, Williams JLD (1974) The anatomy and output
A 89:93–104 connection of a locust visual interneurone; the lobula
Hemmi JM, Pfeil A (2010) A multi-stage anti-predator giant movement detector (LGMD) neurone. J Comp
response increases information on predation risk. Physiol 91:257–266
J Exp Biol 213:1484–1489 O’Shea M, Rowell CHF, Williams JLD (1974) The
Hemmi JM, Tomsic D (2012) The neuroethology of anatomy of a locust visual interneurone; the
escape in crabs: from sensory ecology to neurons and descending contralateral movement detector. J Exp
back. Curr Opin Neurobiol 22:194–200 Biol 60:1–12
Herberholz J, Marquart G (2012) Decision making and Oliva D, Tomsic D (2012) Visuo-motor transformations
behavioral choice during predator avoidance. Front involved in the escape response to looming stimuli in
Neurosci 6:125 the crab Neohelice (=Chasmagnathus) granulata.
James AC, Osorio D (1996) Characterisation of columnar J Exp Biol 215:3488–3500
neurons and visual signal processing in the medulla of Oliva D, Medan V, Tomsic D (2007) Escape behavior and
the locust optic lobe by system identification tech- neuronal responses to looming stimuli in the crab
niques. J Comp Physiol A Neuroethol Sens Neural Chasmagnathus granulatus (Decapoda: Grapsidae).
Behav Physiol 178:183–199 J Exp Biol 210:865–880
Jones PW, Gabbiani F (2010) Synchronized neural input Palka J (1967) An inhibitory process influencing visual
shapes stimulus selectivity in a collision-detecting responses in a fibre of the ventral nerve cord of locusts.
neuron. Curr Biol 20:2052–2057 J Insect Physiol 13:235–248
Color Vision, Computational Methods for 645 C
Peron SP, Gabbiani F (2009a) Spike frequency adaptation Srinivasan MV, Zhang S (2004) Visual motor computa-
mediates looming stimulus selectivity in a collision tions in insects. Annu Rev Neurosci 27:679–696
detecting neuron. Nat Neurosci 12:318–326 Sun H, Frost BJ (1998) Computation of different optical
Peron SP, Gabbiani F (2009b) Role of spike-frequency variables of looming objects in pigeon nucleus
adaptation in shaping neuronal response to dynamic rotundus neurons. Nat Neurosci 1:296–303
stimuli. Biol Cybern 100:505–520 Tammero LF, Dickinson MH (2002a) Collision-
Peron SP, Krapp HG, Gabbiani F (2007) Influence of avoidance and landing responses are mediated by sep-
electrotonic structure and synaptic mapping on the arate pathways in the fruit fly Drosophila
receptive field properties of a collision-detecting neu- melanogaster. J Exp Biol 205:2785–2798
C
ron. J Neurophysiol 97:159–177 Tammero LF, Dickinson MH (2002b) The influence of
Peron SP, Jones PW, Gabbiani F (2009) Precise subcellu- visual landscape on the free flight of the fruit fly
lar input retinotopy and its computational conse- Drosophila melanogaster. J Exp Biol 205:327–343
quences in an identified visual interneuron. Neuron Tomsic D, Berón de Astrada M, Sztarker J (2003) Identi-
63:830–842 fication of individual neurons reflecting short- and
Preuss T, Osei-Bonsu PE, Weiss SA, Wang C, Faber DS long-term visual memory in an arthropod. J Neurosci
(2006) Neural representation of object approach in a 23:8539–8546
decision-making motor circuit. J Neurosci Wagner H (1982) Flow-field variables trigger landing in
26:3454–3464 flies. Nature 297:147–148
Reichardt W, Poggio T, Hausen K (1983) Figure-ground Wang Y, Frost BJ (1992) Time to collision is signalled by
discrimination by relative movement in the visual sys- neurons in the nucleus rotundus of pigeons. Nature
tem of the fly. Part II: towards the neural circuitry. Biol 356:236–238
Cybern 46(Suppl):1–l30 Yamawaki Y, Toh Y (2009) Responses of descending
Rind FC, Bramwell DI (1996) Neural network based on neurons to looming stimuli in the praying mantis
the input organization of an identified neuron signaling Tenodera aridifolia. J Comp Physiol A 195:253–264
impending collision. J Neurophysiol 75:967–985
Rind FC, Simmons PJ (1992) Orthopteran DCMD neuron:
a reevaluation of responses to moving objects.
I. -Selective responses to approaching objects.
J Neurophysiol 68:1654–1666 Color Computational Vision
Rind FC, Simmons PJ (1999) Seeing what is coming:
building collision sensitive neurones. Trends Neurosci
▶ Color Vision, Computational Methods for
22:215–220
Rowell CHF (1971) The orthopteran descending move-
ment detector (DMD) neurones: a characterization and
review. J Comp Physiol A 73:167–194
Rowell CHF, O’Shea M, Williams JL (1977) The neuronal
basis of a sensory analyser, the acridid movement
Color Vision, Computational
detector system. IV. The preference for small field Methods for
stimuli. J Exp Biol 68:157–185
Santer RD, Simmons PJ, Rind FC (2005a) Gliding behav- C. Alejandro Parraga
iour elicited by lateral looming stimuli in flying
Computer Vision Centre/Computer Science
locusts. J Comp Physiol A 191:61–73
Santer RD, Yamawaki Y, Rind FC, Simmons PJ (2005b) Department, Universitat Autònoma de
Motor activity and trajectory control during escape Barcelona, Barcelona, Spain
jumping in the locust Locusta migratoria. J Comp
Physiol A 191:965–975
Santer RD, Yamawaki Y, Rind FC, Simmons PJ
(2008) Preparing for escape: an examination of the Synonyms
role of the DCMD neuron in locust escape jumps.
J Comp Physiol A 194:69–77 Color computational vision; Computational neu-
Schlotterer GR (1977) Response of the locust descending
roscience of color
movement detector neuron to rapidly approaching and
withdrawing visual stimuli. Can J Zool 55:1372–1376
Simmons PJ (1980) Connexions between a movement-
detecting visual interneurone and flight motoneurones Definition
of a locust. J Exp Biol 86:87–97
Simmons PJ, Rind FC, Santer RD (2010) Escapes with and
without preparation: the neuroethology of visual star- The study of color vision has been aided by a
tle in locusts. J Insect Physiol 2010(56):876–883 whole battery of computational methods that
C 646 Color Vision, Computational Methods for
attempt to describe the mechanisms that lead these mechanisms are deeply rooted in the life-
to our perception of colors in terms of the styles of our primate ancestors and their need to
information-processing properties of the visual avoid predators, find suitable mates, and forage
system. Their scope is highly interdisciplinary, for nutritious fruits and leaves in the African
linking apparently dissimilar disciplines such as forest. These tasks and a changing visual envi-
mathematics, physics, computer science, neuro- ronment may have determined the properties of
science, cognitive science, and psychology. our color vision and its drive to collect informa-
Since the sensation of color is a feature of our tion about the part of the environment that
brains, computational approaches usually include changes the least: the surfaces of objects. This
biological features of neural systems in their knowledge may help categorize the objects
descriptions, from retinal light-receptor interac- (edible? dangerous?) and therefore improve our
tion to subcortical color opponency, cortical sig- chances of survival. In view of this, it is easy to
nal decoding, and color categorization. They understand why our brain has evolved its non-
produce hypotheses that are usually tested by linearities to favor a representation of the world
behavioral or psychophysical experiments. that enhances and improves on the mere physics
of the visual environment.
The very properties that aid survival also
Detailed Description make the problem of predicting color sensation
from the wavelength patterns of light stimulating
Although the sensation of hue is an invention of our retinas very complex and mathematically
our brains, it nevertheless allows us to identify difficult to address. Color vision scientists
objects by one of their key physical features: their have made a great deal of advance, making
color. Our cortical neurons receive information sense of what is in essence a complex nonlinear
about the world “out there” and process it in problem. Each subdiscipline studies the phenom-
highly nonlinear ways to make sense of the envi- enon from its own unique perspective, with
ronment, highlighting and grouping important computational approaches contributing to our
items over the rest. In the case of color, this understanding at every level, from neurophysiol-
information is carried by the distribution of ogy to behavior.
energy in the photons (wavelength distribution)
of the light reflected by objects. Our visual sys- The Retina
tem captures these photons through a very sophis- The retina is perhaps the best known part of our
ticated mechanism in our retinas, one that visual system. It contains three types of photore-
converts wavelength information into neural ceptors or cones that are selectively stimulated by
spikes and ultimately into the vivid sensation photons according to their energy (wavelength).
we experience. However, the result of this pro- These receptors (named L, M, and S because of
cess bares only a small resemblance to the phys- their preferences for long, medium, and short
ical characteristics of the light originally falling wavelengths, respectively) are light-sensitive
on our retinas. For instance, two objects whose neurons which fire electrical impulses when
reflected light may differ by only a few more exposed to light of wavelengths near their pre-
energetic photons may be assigned completely ferred value. Their sensitivity has been quantified
different color categories by our brain. The oppo- (Stockman and Sharpe 2000) in the form of three
site may also occur: two objects reflecting spectral sensitivity functions (see Fig. 1). Notice
remarkably different photon/energy distributions that since each of the cones could be stimulated
may look as having the same color to us. Our by photons with energies within a wide range,
perception of color at a point in space may long-wavelength-preferring cones could be stim-
depend not only on the physical characteristics ulated by medium-wavelength light. Given that
of the light reflected from that point but also on cones also respond to light intensity, intense light
that of the points surrounding it. The origins of of suboptimal wavelength can produce the same
Color Vision, Computational Methods for, output of at least two different types of photosensitive
Fig. 1 Schematics of the chromatic signal processing in neurons. Two of the post-receptoral mechanisms are chro-
the retina. The figure on the left shows the three human matically opponent and one (L + M) is achromatic. Panels
photoreceptors and how they connect to form the three on the right show the spectral sensitivity of each chro-
post-receptoral mechanisms by addition and subtraction. matic mechanism. The terms L + M, L M, etc. refer to
Chromatic information is obtained by sampling the spec- the sign of the operation and do not represent the weight of
trum around different wavelengths and comparing the each mechanism in the sum
firing rate as low-intensity light of optimal wave- workings of three opposed mechanisms: red
length. This is the reason why many (in fact infi- green, blue yellow, and light dark (see bottom
nite) wavelength distributions can produce of Fig. 1). Each theory was an attempt to explain
similar outputs from the cones. Two lights of different phenomena such as how mixtures of
different wavelength distributions that produce colored light create a different color or the obser-
the same retinal output are called metamers. vation that some opposing pairs of sensations
Receptoral and post-receptoral mechanisms cannot be simultaneously experienced (e.g.,
were anticipated in the nineteenth century (well there is no reddish-green color) and afterimages
before any physiological observation) in the form (illusory colors that appear after our retinas have
of a trichromatic theory by Young (1802) and been overexposed to colored stimuli) (Gregory
Helmholtz (1867) and a color opponent theory 1998). Our current knowledge of the interaction
by Hering (1875). The first postulates that color between cone photoreceptors and light can be
vision is based on three fundamental mechanisms expressed mathematically as the integral of the
as illustrated at the top of Fig. 1, and the second product between the electromagnetic power dis-
postulates that color vision arises from the tribution of the light and each of the cones’
spectral sensitivity functions (L, M, and S curves images obtained by commercial cameras. These
shown in Fig. 1): are also modeled by Eq. 1, but their sensor sensi-
tivities are usually unknown. Some devices are
ð
capable of producing images in standard, device-
ei ¼ Si ðlÞPðlÞ dl (1)
l independent color spaces such as the CIE1931
XYZ system or the sRGB color standard, which
In Eq. 1, ei represents the excitations of the three are easily convertible following a simple set of
classes of cones (i = L, M, S), Si (l) is their formulae (Poynton 2003; Green and MacDonald
corresponding spectral sensitivity, and P(l) is 2002). Once we know the XYZ coordinates, we
the spectral power distribution of the light falling can obtain the chromaticity coordinates x, y using
on them. Since in real life light measurements are the following:
only available at a series of finite wavelengths,
the integral in Eq. 1 is generally approximated to X
a finite sum over the visible spectrum x ¼
XþYþZ (2)
(400–700 nm). Y
y ¼
XþYþZ
Computational Methods for the Retina
In practice, it is unlikely that we know the full From there, it is possible to obtain the cone exci-
spectrum of the light falling on the retina. It is tations using Y (luminance) and the following
more common to operate retinal models based on transformation (Wyszecki and Stiles 1982a):
eL ¼ Y ½0:15514x=y þ 0:54312 0:03286 ð1 x yÞ=y

eM ¼ Y ½0:15514x=y þ 0:45684 0:03286 ð1 x yÞ=y (3)
eS ¼ Y ½0:00801 ð1 x yÞ=y
In the case of Eq. 3, ei values correspond to the strengths of the two post-receptoral neural
Smith and Pokorny cone fundamentals (Smith opponent mechanisms (see below). Equa-
and Pokorny 1975) and the chromaticities ought tion 3 provides a good approximation for
to be derived from the Judd-modified CIE1931 transforming the output of cameras and monitors
standard color matching functions (Judd 1951). once we can express it in the CIE1931 XYZ
The plot at the top of Fig. 1 shows an example of system. However, in most cases this is not
cone fundamentals, with the sensitivity of each possible, and devices such as cameras and mon-
cone plotted as a function of wavelength. Color itors need to be characterized to convert their
matching functions are obtained from psycho- outputs to cone excitations. There are several
physical experiments where subjects have to methods for characterizing cameras which
visually match the sensation produced by spec- include gamut mapping between a known target
trally narrowband stimuli to a mixture of three and the camera output (Cheung et al. 2004;
spectrally broadband lights. It is possible to Parraga et al. 2010), estimation of the camera’s
mathematically obtain cone fundamentals from sensor sensitivities (Westland and Ripamonti
such experiments, given certain assumptions 2004; Barnard and Funt 2002), and mixtures
(Wyszecki and Stiles 1982b). Although Smith (Parraga et al. 2010). They transform the output
and Pokorny did not specify the coefficient nec- of a commercial digital camera into a device-
essary to obtain es, the value of 0.00801 was independent color space such as the
chosen to take into account the relative CIE1931 XYZ.
Receptoral Gain Control Mechanisms variations in intensity between shaded and illu-
In the retina, there is a nonlinear relationship minated objects can reach 9 orders of magnitude),
between the input and the output of the cone which cannot be encoded directly by the output of
photoreceptors, which is often characterized as a single neurons (e.g., retinal neurons can encode
gain control mechanism (output signals are mod- ranges of approx. 50–1) (De Valois 2004).
ified by signals from the same or different cones).
This mechanism converts cone excitations into a Post-Receptoral Color Processing C
contrast representation such as The signals transmitted by cones are combined
and encoded in a series of steps by neurons in the
Dei retina (the last processing layer is formed by
Ci ¼ ; emi ¼ L, M, S (4)
ebi e0i neurons called ganglion cells) and transmitted to
an area inside the thalamus called the lateral
where ebi represents the cone excitation produced geniculate nucleus (LGN). The majority of these
by the background (the adaptation state), e0i is a signals (80 % of the fibers reaching the LGN)
constant, and Dei represents the excitation differ- passes through a layer of ganglion cells called
ence between the cone considered and the back- parvocellular (or parvo) cells (De Valois 2004).
ground (Stockman and Brainard 2010). The At this stage, the information is organized both
i subindex discriminates between the three types spatially and chromatically in opponent signals.
of photoreceptors. For example, a parvo cell in the central retina
One of the reasons for this normalization is to may receive input from a single L or M cone
reduce the massive intensity range present in the and weight it against input from a small group
visual environment (in a typical sunny day, of surrounding cones (see Fig. 2).
Color Vision, Computational Methods for, intensity and color. If the light falling on the cell’s recep-
Fig. 2 Schematics of the spatiochromatic opponent tive field changes its intensity, both center and surround
processing in the retina. Parvocellular ganglion cells col- are stimulated, therefore increasing their antagonism. If
lect input from a single cone and compare it with a pool of the same light changes its color toward red, the center is
surrounding cones. In the case of LM opponency, this stimulated and the surround is depressed, contributing less
leads to four possible combinations, as illustrated in the to reduce the ganglion cell output signal (therefore the +
figure. The right part of the figure shows an example of sign)
how the same cell can carry information about both
Each ganglion cell is stimulated by light fall- L-cone excitation (the center). A similar algo-
ing on a physical part of the retina that contains rithm can be applied to obtain S(L + M)
the cones to which it is connected (its receptive opponency. Similarly to parvo, there exist in the
field). As a result of the wiring shown in Fig. 2, retina two other types of ganglion cells. The
parvo cells output a signal that is spatially and koniocellular (or konio) cells that carry informa-
chromatically opponent and conveys both intention about S-cone excitation relative to their sur-
sity (i.e., brightness) and wavelength information rounds (S(L + M) and S + (L + M)) and the
(i.e., color) from the light falling on its receptive magnocellular (or magno) cells that receive
field. Thus parvo cells perform a “double duty,” excitatory output from all three types of cones,
carrying information about both intensity and therefore, are insensitive to color (De Valois
color. The right part of Fig. 2 shows how intensity 2004). Figure 1 illustrates this opponency from
changes in light increase the cell’s antagonism, a spectral point of view.
conveying information about the smallest possi- The three types of ganglion cells described
ble unit (a cone), while chromatic changes pro- above connect to specialized layers of neurons
duce a synergistic response in the center and in the LGN. The presence of these LGN cells
surround, thus increasing the number of cones representing different aspects of the light cap-
contributing to the signal. For this reason, parvo tured by cones has inspired the creation of color
cells in Fig. 2 convey finer detail information spaces based on the properties of these neurons.
(higher spatial frequencies) about intensity than The best known is the MacLeod–Boynton cone-
for color. In other words, when light varies spa- excitation space (MacLeod and Boynton 1979)
tially in color, the same cells can only carry which is based on the coefficients of Eq. 3 plus an
information about its coarse spatial distribution arbitrary condition that the relationships among
(lower spatial frequencies). This imbalance has the sensitivities of the three receptoral mecha-
been confirmed psychophysically by our greater nisms at 400 nm (see Fig. 1) will be given by
ability to detect fine patterns of achromatic light Eq. 5:
and coarse patterns of colored light (De Valois
and De Valois 1988) and matches the informa- SL ð400Þ þ SM ð400Þ:36em ¼ :36emSS ð400Þ (5)
tion content of natural scenes (Parraga
et al. 2002). This combination of spatial and where Si (400) is the sensitivity of each cone
chromatic opponency in parvo cells appears to photoreceptor at 400 nm. To obtain the
be a sophisticated mechanism to optimally coordinates of the MacLeod–Boynton receptor
extract information from the visual environ- space, it is necessary to convert the L-, M-,
ment, carrying both chromatic and achromatic and S-cone excitations to coordinates l, m, and
signals over different spatial frequency bands s, respectively, using the following
(multiplexing). transformation:
Computationally, it is common to model the
center–surround receptive field structure by con- L M S
l¼ ; emm ¼ ; ems ¼ (6)
volving the color opponent representations with LþM LþM LþM
Gaussian kernels of opposite sign and different
standard deviations (difference of Gaussians). The normalization provided by Eq. 5 constrains
For example, one kernel is applied to the L-cone the value of s in Eq. 6 between 0 and 1.
representation and another (of greater standard An extension of cone-excitation spaces are
deviation) to the M-cone representation and the cone-contrast spaces, which address the issues
results are subtracted from each other. This proposed by the receptoral gain control mechanisms
cess is equivalent to subtracting a blurred version described in Eq. 4. Two popular cone-contrast
of the M-cone-excitation image (the surround) spaces were created by Derrington et al. (1984)
from another slightly blurred version of the and Boynton (1986) following different
Color Vision,
Computational Methods
for, Fig. 3 The figure
shows the influence of
surrounding patterns on the
perceived color of a central
area. In both figures, the
inverted “U” bar has the C
same RGB values, but it is
perceived differently
according to its
surroundings. This
phenomenon is called
“chromatic induction”
assumptions about the neutral points (crossings to many other properties of the visual stimulus
of the LM and S(L + M) functions of Fig. 1). such as contrast, object orientation, shape,
Cone-contrast representations take into account etc. This is thought to be the reason behind many
the cone excitation produced by the background illusory phenomena where the color of the area we
and rescale the signal accordingly. The new coor- gaze is influenced by patterns of the surrounding
dinates become area (chromatic induction, see Fig. 3) and color
constancy (Brainard 2004).
Dei Originally, color processing in the visual cor-
Ci ¼ ; emi ¼ L, M, S (7)
ebi tex was thought to be modular and separated from
other forms of perception. This view was
where Dei is the cone-excitation difference supported by studies in cerebral achromatopsia
between the signal and a given reference level (a lesion in the brain that causes loss of the ability
(the background). eib is the cone excitation pro- to recognize colors without loss of form and
duced by the background. In this new represen- motion perception) and others (Zeki 1993).
tation, the background excitation becomes the Today’s prevalent view is that V1 is strongly
origin of the new system. Cone-contrast spaces stimulated by color, with about 50 % of its neu-
provide a representation of the input signal arriv- rons responsive to some kind of chromatic stim-
ing at later stages in the visual pathway. ulus, with LM being the preferred direction
(Shapley and Hawken 2011). Areas V2 and V4
Cortical Color Processing also respond to color; however, many of their
LGN neurons project to the primary visual cortex neurons respond equally to achromatic patterns.
in the occipital part of the brain (also known as Two intriguing classes of color-processing neu-
striate cortex or area V1). From there, visual rons were found in the primary visual cortex:
information is sent to the so-called extrastriate single- and double-opponent cells (Shapley and
visual cortical areas (named V2, V3, V4, and Hawken 2011; Hurlbert 2003; see Fig. 4). The
V5). Each part of the retina is mapped into V1 former respond best to large areas of color, while
with a very precise correspondence between the the latter have preference for color boundaries,
visual field and its related location in the primary patterns, and textures. Both types of neurons have
visual cortex. A large portion of V1 is dedicated been linked to the perceptual phenomena shown
to the center of the retina, where the abundance of in Fig. 3.
cones is highest and the receptive field smallest. In There have been reports in the literature about
contrast to the well-defined, opponent-color a spatial segregation of color-selective neurons in
information-carrying neurons found in the retina V1 and V2. A number of optical imaging electro-
and LGN, color-sensitive cortical neurons respond physiological and microelectrode studies have
signals, and other post-receptoral stages which

may include normalizations by neighboring units
and weighting parameters. The center–surround
receptive field interaction is usually modeled via
a pyramid of 2-dimensional Gaussian or Gabor
operators that convolve the image to obtain chro-
matic and achromatic information at different spa-
tial scales and orientations. A final layer can also
be implemented to account for single- and double-
opponent cell processing.
Chromatic induction demonstrations like
those shown in Fig. 3 have inspired some com-
putational attempts to model color processing in
the visual cortex, trying to predict the effects of
surround stimuli on the perceived color of a cen-
tral patch. Examples of these are the models of
Singer and D’Zmura (1995) and Spitzer and
Barkan (2005) and Otazu et al.’s CIWaM
(2010). The latter consists of a cone-receptor
stage that separates the input into two chromatic
and one achromatic channels: a wavelet pyramid
to model spatiochromatic processing in the cor-
tex, including normalization effects by neighbor-
Color Vision, Computational Methods for, ing neurons (divisive normalization) (Heeger
Fig. 4 Cartoon examples of single-opponent and 1992), and a weighting function to mimic the
double-opponent color-processing cells. Single-opponent spatial filtering of the human visual system.
cells are insensitive to edges and do not prefer any oriented
A later version of CIWaM was capable of
stimuli. They respond strongly to colored patches. On the
other hand, double-opponent cells are strongly selective predicting unconstrained eye movements in
for chromatic edges and have a preferred orientation human subjects (Murray et al. 2011).
A family of computational approaches to color
found the presence of clusters of color-selective processing takes advantage of the tools currently
neurons, although there are still negative results available in computer science for tasks such as
and the issue has not been definitively resolved object recognition. For example, one approach
(Shapley and Hawken 2011). consists of applying shape-based descriptors
such as scale-invariant feature transform (SIFT)
Computational Models of Color Processing in on the post-receptoral color representations. SIFT
the Visual Cortex will learn “interesting features” of an object in a
Computational models of cortical processing can set of training images which can be later used to
be categorized as either primarily phenomenolog- identify the object in the dataset. Other
ical, i.e., trying to capture the observed phenom- approaches involve a set of shape-based descrip-
ena in a convenient mathematical form, or tors computed from gray-scale pixel intensities,
mechanistic, i.e., trying to show how observed which are in turn concatenated with hue histo-
phenomena arise from biologically grounded neu- grams. An improvement has been proposed by
ral mechanisms. A typical color vision model con- Van de Weijer and Schmid who concatenated a
tains a preprocessing stage where the input image hue histogram with gray-scale SIFT image
is converted to an L-, M-, or S-cone-excitation descriptors (Gevers 2012), while Zhang
representation, a cone opponency stage that et al. (2012) described a hierarchical model of
weights the output of cones to produce opponent color processing inspired in the primate visual
cortex. They produced chromatically opponent users expect the image to represent the scene
channels using operators that resemble single- “as they saw it,” not the actual physical values.
and double-opponent cells and outperformed the Many popular computational approaches try to
object recognition of previous approaches. solve this problem by taking into account the
information intrinsic to the image, rather than
Color Constancy the physiological or anatomical mechanisms of
The human visual system has the ability to per- the visual system. For practical reasons, methods C
ceive objects as having a well-defined color were developed to return an image that approxi-
despite the fact that they reflect light according mately represents the intrinsic color of objects,
to their own pigmentation, geometry, and the discounting the effects of the illumination
wavelength content of the illumination. During (canonical image). This image can be then
a typical day, natural daylight changes its color “reilluminated” to create conditions close to the
quite dramatically, but we are mostly unaware of observer’s perception. Illuminant removal
its full effect on the surfaces of objects. As we methods contain different statistical hypothesis,
have seen before, the distribution of power as a the most popular being the “gray world” assump-
function of wavelength in the light reaching our tion, which states that the average reflectance of a
retinas determines how we perceive color. In the scene is achromatic. Another popular hypothesis
case of light reflected from objects, this power (called “MAX-RGB”) takes advantage of some
distribution depends on several factors. Equa- properties of white surfaces and specular reflec-
tion 8 provides a simplified description of the tions and is based on the assumption that the
problem (Brainard 2004): maximum reflectance of a scene is achromatic.
Once a hypothesis is introduced, the computa-
CðlÞ ¼ EðlÞRðlÞ (8) tional algorithm returns an image that meets the
hypothesis. Many of these solutions are strongly
Here C(l) is the spectral power distribution of the dependent on whether the input image complies
light reaching our retina, E(l) is the spectral with the hypothesis and are prone to fail in par-
power distribution of the illumination, and R(l) ticular cases. For example, a “gray world” algo-
is the spectral reflectance of the surface. Equa- rithm may fail when the average color of the
tion 8 assumes that the illumination is spatially scene is not gray. The main drawback of this
uniform and surfaces are Lambertian (diffusely family of computational approaches is that they
reflective, with no specularities). Of these factors, do not provide much insight about the mecha-
only R(l) is intrinsic to the objects, while E(l) nisms that are behind color constancy. See
changes according to the type of illumination, Gevers (2012) for a comparison of illuminant
physical location, time of the day, etc. In sum- removal methods.
mary, to produce a stable scene in terms of color,
the human visual system disentangles the two Color Appearance
factors at the right side of Eq. 8. This is a math- The models and the physiologically based color
ematically ill-posed problem, given that there are spaces described above have been created to han-
infinite possible combinations of illumination dle the perception of colors viewed in isolation
and reflectance that produce the same reflected and in controlled, fully adapted conditions. How-
light. To solve it, the human visual system may ever, the perceptual appearance of colors in nat-
resort to information from previous experience, uralistic viewing conditions depends on so many
scene features such as specular reflections and external factors (surround, background, illumi-
interreflections, memory of colors, etc. nant, etc.) that a series of sophisticated models
The problem of illumination removal is ubiq- have been created to account for them. These
uitous. For example, in computer vision, there are so-called color appearance models are not
applications that require illumination stability in constrained by attempts to incorporate elements
order for algorithms to work. In photography, of the physiology (Fairchild 1998).
Cross-References MacLeod DIA, Boynton RM (1979) Chromaticity dia-

gram showing cone excitation by stimuli of equal
luminance. J Opt Soc Am 69:1183–1187
▶ Center-Surround Processing, Computational Murray N, Vanrell M, Otazu X, Parraga CA (2011)
Role of Saliency estimation using a non-parametric low-level
▶ Center-Surround Processing, Network Models vision model. In: Computer vision and pattern recog-
of nition (CVPR), 2011 I.E. conference on, pp 433–440
Otazu X, Parraga CA, Vanrell M (2010) Towards a unified
▶ Computational Neuroanatomy: Overview model for chromatic induction. J Vis 10(5):1–24
▶ Cortex: Overview Parraga CA, Troscianko T, Tolhurst DJ
▶ Lateral Geniculate Nucleus (LGN) Models (2002) Spatiochromatic properties of natural images
▶ Visual Illusions, Models of and human vision. Curr Biol 12:483–487
Parraga CA, Baldrich R, Vanrell M (2010) Accurate map-
ping of natural scenes radiance to cone activation
space: a new image dataset. In: CGIV 2010/
References MCS’10 – 5th European conference on colour in
graphics, imaging, and vision – 12th international
Barnard K, Funt B (2002) Camera characterization for symposium on multispectral colour science. Society
color research. Color Res Appl 27:153–164 for Imaging Science and Technology, Joensuu,
Boynton RM (1986) A system of photometry and color- pp 50–57
imetry based on cone excitations. Color Res Appl Poynton CA (2003) Digital video and HDTV: algorithms
11:244–252 and interfaces. Morgan Kaufmann, Amsterdam/ Boston
Brainard DH (2004) Color constancy. In: Chalupa LM, Shapley R, Hawken MJ (2011) Color in the cortex: single-
Werner JS (eds) The visual neurosciences. MIT Press, and double-opponent cells. Vision Res 51:701–717
Cambridge, MA, pp 948–961 Singer B, D’Zmura M (1995) Contrast gain control: a
Cheung V, Westland S, Connah D, Ripamonti C (2004) bilinear model for chromatic selectivity. J Opt Soc
A comparative study of the characterisation of colour Am A Opt Image Sci Vis 12:667–685
cameras by means of neural networks and polynomial Smith VC, Pokorny J (1975) Spectral sensitivity of the
transforms. Color Technol 120:19–25 foveal cone photopigments between 400 and 500 nm.
De Valois R (2004) Neural coding of color. In: Werner JS, Vision Res 15:161–171
Chalupa LM (eds) The visual neurosciences. MIT Spitzer H, Barkan Y (2005) Computational adaptation
Press, Cambridge, MA, p 1001 model and its predictions for color induction of first
De Valois RL, De Valois KK (1988) Spatial vision. and second orders. Vision Res 45:3323–3342
Oxford University Press, New York Stockman A, Brainard DH (2010) Color vision mecha-
Derrington AM, Krauskopf J, Lennie P (1984) Chromatic nisms. In: Bass M, Mahajan VN (eds) OSA handbook
mechanisms in lateral geniculate-nucleus of macaque. of optics. McGraw-Hill, New York, pp 11.11–11.104
J Physiol 357:241–265 Stockman A, Sharpe LT (2000) The spectral sensitivities
Fairchild MD (1998) Color appearance models. Addison- of the middle- and long-wavelength-sensitive cones
Wesley, Reading/Harlow derived from measurements in observers of known
Gevers T (2012) Color in computer vision: fundamentals genotype. Vision Res 40:1711–1737
and applications. Wiley, Hoboken Westland S, Ripamonti C (2004) Characterization of cam-
Green P, MacDonald L (2002) Colour engineering: eras. In: Computational colour science: using
achieving device independent colour. Wiley, MATLAB. Wiley, Chichester, pp 127–128
Chichester Wyszecki G, Stiles WS (1982a) Theories and models of
Gregory RL (1998) Seeing colours. In: Eye and brain: the color vision. In: Color science: concepts and methods,
psychology of seeing. Oxford University Press, quantitative data and formulae. Wiley, New York/
Oxford, pp 121–134 Chichester, p 615
Heeger DJ (1992) Normalization of cell responses in cat Wyszecki G, Stiles WS (1982b) Colorimetry. In: Color
striate cortex. Vis Neurosci 9:181–197 science: concepts and methods, quantitative data and
Helmholtz HV (1867) Handbuch f physiologischen Optik. formulae. Wiley, New York/Chichester, pp 117–145
Voss, Leipzig Young T (1802) On the theory of light and colours. Philos
Hering E (1875) Zur Lehre vom Lichtsinne. Sechs Trans R Soc Lond 92:12–48
Mittheilungen an die Kaiserl. Akademie der Zeki S (1993) A vision of the brain. Blackwell, Oxford/
Wissenschaften in Wien, 2nd edn. Gerold, Wien Boston
Hurlbert A (2003) Colour vision: primary visual cortex Zhang J, Barhomi Y, Serre T (2012) A new biologically
shows its influence. Curr Biol 13:R270–R272 inspired color image descriptor. In: Fitzgibbon AW,
Judd DB (1951) Report of U.S. secretariat committee on Lazebnik S, Perona P, Sato Y, Schmid C (eds) ECCV
colorimetry and artificial daylight. In: Twelfth session 2012 – 12th European conference on computer vision,
of the CIE. Bureau Central de la CIE, Stockholm, p 11 7–13 Oct 2012. Springer, Florence, pp 312–324
Comparative Analysis of Half-Center Central Pattern Generators (CPGs) 655 C
flow is determined by the neuron’s own voltage
Comparative Analysis of Half-Center and the ion concentrations in the extracellular
Central Pattern Generators (CPGs) medium surrounding the neuron, independent of
the activity of other neurons nearby. Other cur-
Jonathan Rubin rents are synaptic and are induced by the activity
Department of Mathematics and Center for the of other neurons. When active, neurons release
Neural Basis of Cognition, University of neurotransmitters that activate synaptic currents C
Pittsburgh, Pittsburgh, PA, USA in other neurons, their postsynaptic targets. In
half-center CPGs, these synaptic currents are
inhibitory. If the voltage of a postsynaptic neuron
Definition is at a typical physiological level, then such a
current lowers its voltage, making it harder for
CPGs are networks of neurons that can produce the postsynaptic neuron to activate in the short
rhythmic activity, featuring a stereotyped term. Thus, the primary features of a half-center
sequence of phases that repeats in a periodic- CPG consist of the sets of ion currents intrinsic to
like manner, without receiving rhythmic inputs its components; the time scales, voltage depen-
from an outside source. Half-center CPGs, also dences, strengths, and other characteristics of
known as half-center oscillators, are CPGs these currents; the pattern of synaptic connections
consisting of two components connected by between its components; and the characteristics of
reciprocal synaptic inhibition. While the coupled the synaptic currents that these connections induce.
network exhibits a two-phase pattern, with one To compare properties of different half-center
component exhibiting a high activity level and CPGs efficiently, however, it is useful to group
the other at a low activity level in each phase, the together various half-center CPGs with common
individual components are not able to switch emergent features that lead to similar properties.
between sustained high and low activity phases
in isolation. Comparative analysis of half-center Classifying Half-Center CPGs: Low and High
CPGs is the study of how specific features of half- Activity States and Transitions Between
center CPGs translate into particular properties. Them
For simplicity, it is convenient to refer to compo-
nents of CPGs as individual neurons, although the
Detailed Description following discussion can be generalized to cases
where the components are neuronal populations.
A half-center CPG consists of a pair of compo- The activity level of a neuron can be quantified
nents and the coupling between them. Each using its membrane potential or its rate of gener-
component may be an individual neuron or ating spikes, which are brief pulses of elevated
a population of neurons. In a half-center oscilla- potential also known as action potentials. Usually
tion, each component switches back and forth high potentials are associated with high rates of
between sustained low activity and high activity spike generation, although there are circum-
states. The level of activity in a neuron is related stances in which this link does not hold. An
to its membrane potential. Transmembrane cur- isolated neuron can be in a low activity state, in
rents allow ions to pass through channels into or a high activity state, or in a state where it spends
out of a neuron, altering its membrane potential a significant time in each state with rapid transi-
or voltage. These channels may be blocked by tions between states. By definition, the neuronal
two types of gating processes, and changes in the components of a half-center CPG cannot be in the
extent of blockage are called (de)activation or latter state in isolation, as can be checked exper-
(de)inactivation, depending on the gating type. imentally by preventing the induction of
Many of these ion currents are intrinsic to synaptic currents, although they must enter this
a neuron, meaning that the extent to which they rhythmic state once they are coupled synaptically
C 656 Comparative Analysis of Half-Center Central Pattern Generators (CPGs)
(Hooper 2001). For an isolated neuron to main- mechanisms, taking into account both intrinsic
tain a low activity state, there must be a low and synaptic properties of a model, rather than
voltage at which its intrinsic ion currents are in terms of the intrinsic attractors of model units
balanced. A high activity state may correspond alone.
to the balancing of a neuron’s intrinsic currents at A transition-based classification of half-center
a high voltage or to sustained spike generation, CPG models uses the categories of intrinsic
without extended pauses between spikes. If we escape, intrinsic release, synaptic escape, and
represent a neuron by a system of ordinary dif- synaptic release (see Wang and Rinzel 1992;
ferential equations encoding the relations Skinner et al. 1994), defined in terms of which
between membrane potential and ion current neuron, and which processes associated with that
flows, ignoring spatial effects and stochasticity, neuron, control the transitions between activity
then states of balanced current are equilibrium states. To set up and understand this classifica-
points of the system while a state of sustained tion, it is useful to note that certain processes
spiking is a periodic solution. If a biological neu- associated with voltage changes are relatively
ron or mathematical model neuron tends to one of fast, while others are relatively slow. The former
these states over time from an appropriate set of group includes changes in activation or inactiva-
starting conditions, then the attained state is an tion of certain currents in response to changes in
attractor for the system. voltage as well as changes of voltage itself in
These ideas lead to a classification of half- response to changes in availability of currents,
center CPGs in terms of which attractors are while the latter includes changes in other (in)
present in the absence of inputs, assuming these activation levels as well as changes in concentra-
are unique. Theoretically, the possibilities are: tions of certain ions within a neuron.
• L-L: both neurons have attracting equilibrium • In intrinsic escape, the slow processes of the
points in low activity states in the absence of neuron in a low activity state allow it to make
inputs. a transition to a high activity state. In doing so,
• L-H: one neuron has an attracting equilibrium its inhibition to the other neuron turns on,
point at a low activity state and one has an causing the other neuron to make a high-to-
attracting high activity equilibrium point or low transition.
oscillation in the absence of inputs. • In intrinsic release, the slow processes of the
• H-H: both neurons have attracting high activ- neuron in a high activity state cause it to make
ity states, either equilibrium points or oscilla- a transition to a low activity state. In doing so,
tions, in the absence of inputs. its inhibition to the other neuron turns off,
In each case, when the neurons are coupled causing the other neuron to make a low-to-
with inhibition, both must transition repetitively high transition.
between low and high states, taking turns being • In synaptic escape, the threshold for activation
active, for a CPG rhythm to result. Thus, it is of synaptic inputs is low. The neuron in a low
important to check that this alternation is possible activity state undergoes small increases in
in each of the cases in this classification. To keep voltage due to its slow processes. Due to
things relatively simple, it is helpful to assume the low synaptic activation threshold, this
that the synaptic inputs from a neuron turn on voltage increase is enough to introduce some
quickly when that neuron’s activity level inhibition of the high activity neuron. This
becomes sufficiently high and turn off quickly inhibition suffices to induce a high-to-low
when its activity level is sufficiently low, transition in the high activity neuron, and
although theoretical work has considered slower the resultant loss of inhibition of the low
input changes as well. It turns out that there are activity neuron causes it to switch to a high
certain types of transitions between states that activity state.
can occur in each case and that the most useful • In synaptic release, the threshold for activa-
classification is in terms of these transition tion of synaptic inputs is high. The neuron in
a high activity state undergoes small decreases inhibited neuron to activate, inhibiting its partner,
in voltage due to its slow processes. Due to the in a process called post-inhibitory rebound. The
high synaptic activation threshold, this voltage whole cascade can repeat periodically and sustain
decrease is enough to relieve some inhibition a half-center oscillation.
of the low activity neuron. This loss of inhibi-
tion suffices to induce a low-to-high transition Example 2 – H-H/Synaptic Release with
in the low activity neuron, and the resultant IK–Ca Suppose that both neurons have attracting C
inhibition of the high activity neuron causes it high activity states without inputs and have
to drop to a low activity state. a calcium-dependent outward potassium
This classification is simplest if it is assumed current, IK–Ca, which is activated at these
that the two neurons in the half-center CPG attracting states. If one neuron is put into a low
model are identical. More generally, it makes activity state in the absence of inputs, it will
sense under the assumption that both transitions transition to high activity on its own, but suppose
in the network are of the same type. that the other neuron has just been allowed to
A generalization of the classification would become active and supplies sufficient inhibition
allow for two different transition types within to block this transition. Over time, the active
one network oscillation; one neuron might acti- neuron’s IK–Ca activates, and this outward current
vate through intrinsic escape, for example, reduces its voltage. If this voltage reduction is
while the other neuron might activate through sufficient, it may reduce the inhibition level to the
intrinsic release. Another observation is that other neuron. As a result, the other neuron may
none of these transition types explicitly become active and inhibit its partner, causing
specifies a particular timing between the transi- a high-to-low transition in the formerly active
tion of one neuron and the corresponding transi- neuron.
tion of the other neuron. For example, one
neuron might undergo a high-to-low transition Example 3 – H-H/Intrinsic Escape with IK–Ca In
intrinsically, releasing the other neuron from the above scenario, it is also possible that the
inhibition. This change might be enough to neuron in the low activity state can escape with-
induce an immediate low-to-high transition in out any reduction of inhibition, because its IK–Ca
the released neuron. On the other hand, additional deactivates over time. Depending on model
time might be needed before the released neuron parameters, this effect may be enough to allow
could activate. it to transition to the high state and suppress the
Finally, note that this classification is related formerly active neuron.
to the classification based on single neuron
attractors. This relation can be illustrated using These examples demonstrate that different
examples involving particular currents. transition mechanisms can arise even from the
same intrinsic neuron features, depending on the
Example 1 – L-L/Intrinsic Release with quantitative details. Similarly, the same transi-
IT Suppose that both neurons have attractors at tion mechanism can result despite the involve-
low activity states without inputs and have T-type ment of different neuron features. In particular,
calcium currents, IT, which are inactivated at distinct current types can be associated with the
these attracting states. If one neuron is placed in same transition mechanism; for example, the
a high activity state and the other in a low activity depolarization-activated outward current IK–Ca
state, the low activity neuron is inhibited and its or the depolarization-inactivated inward current
voltage is reduced. IT is a positive inward current IT can be replaced by the depolarization-
that slowly deinactivates at low voltages. Thus, if inactivated inward current INaP in the above
the high state neuron transitions to a low state, the examples, as long as model parameters are
rapid release from inhibition combined with the tuned appropriately (Fig. 1a; cf. Izhikevich
enhanced availability of IT can allow the formerly 2007).
C 658 Comparative Analysis of Half-Center Central Pattern Generators (CPGs)
Comparative Analysis of Half-Center Central Pattern oscillations in a 1 s simulation as a function of applied

Generators (CPGs), Fig. 1 Half-center oscillations with current Iapp when transitions are by escape (top row) or
transitions by escape and release. (a) A two-neuron model release (bottom row). For example, five oscillations mean
with reciprocal inhibition and persistent sodium current that one cell oscillated three times and the other two times
generates half-center oscillations using escape (top row) during the simulation
or release (bottom row). (b) Number of completed
Transition-Dependent Properties of Half- thus accelerate the escape of the suppressed neu-
Center CPGs ron. This effect shortens the period of the CPG
A classification of half-center CPGs in terms of oscillation (Fig. 1b). In a half-center CPG with
transition mechanisms is useful because various a moderate synaptic threshold such that transi-
properties are determined, or predicted to be tions remain intrinsic, a switch from intrinsic
determined, by which transition mechanism is release to intrinsic escape may be achieved by
present. Consider a theoretical half-center CPG applying progressively stronger excitatory cur-
composed of two identical neurons with low rent, shifting the neurons’ attractors from low
activity attractors and with transitions by intrinsic activity to high activity states; this switch may
release. Suppose that an equal excitatory current, be abrupt or may be interrupted by a current range
which raises voltage, is applied to both neurons. lacking oscillations, in which release is blocked
Whichever neuron is active controls the next but escape is not yet possible. In either case, the
phase transition in the coupled network, and dependence of CPG period on applied current
therefore the effect of this current will be deter- level is non-monotonic.
mined by its impact on the neurons when they are A more complicated situation is the applica-
active, as long as the current is not so large that it tion of unequal excitatory current levels to the
changes the transition mechanism (Skinner et al. two neurons in a half-center CPG (Daun et al.
1993; Curtu et al. 2008). Since the current is 2009). This scenario is likely functionally rele-
excitatory, it will delay the high-to-low transition vant, as modulatory inputs to CPGs may target
of each active neuron, prolonging the period of distinct components differently. In a CPG with
the CPG oscillation (Fig. 1b). Alternatively, sup- transitions by intrinsic escape, the neuron receiv-
pose that the neurons in the CPG have high activ- ing a stronger current will escape faster, but under
ity attractors and that the CPG rhythm involves fairly general conditions, the network equili-
transitions by intrinsic escape. In this case, brates such that the duration of the other neuron’s
whichever neuron is suppressed controls the suppressed phase remains approximately con-
next phase transition in the coupled network. If stant. In contrast, if transitions occur through
an excitatory current is applied equally to both intrinsic release, then current availabilities tend
neurons, this current will raise their voltages and to equilibrate such that little change in either
Comparative Analysis of Half-Center Central Pattern a function of the phase at which stimulation is applied to
Generators (CPGs), Fig. 2 PRCs for half-center oscil- the neuron with a voltage trace of the corresponding type
lations depend on whether transitions occur by release shown in the bottom panels (Figure reproduced from
(left) or escape (right). Top curves (solid blue or dashed Zhang and Lewis (2013) with permission of Springer-
green) show Z1/T, a measure of phase advance or delay, as Verlag)
oscillation phase duration results. If transitions same event next occurs is time T. For each
occur through synaptic release, then the low-to- t [0, T], define a phase f = t/T [0, 1].
high transition of the neuron with greater stimu- A phase response curve (PRC) can be defined
lation is accelerated but that of the other neuron is by plotting the advance or delay in phase intro-
delayed, such that both phase durations are duced by a transient current injection as
altered. Thus, of these possibilities, only intrinsic a function of the oscillation phase at which the
escape allows for independent modulation of the injection occurs (Ermentrout and Terman 2010).
durations of the two activity phases in a half- The PRCs that result when the transitions within
center oscillation (Daun et al. 2009), as would an oscillation are due to escape are quite different
appear to be desirable to maximize the flexibility from the PRCs for release-based transitions
of behaviors driven by CPG rhythms and is (Fig. 2; Zhang and Lewis 2013). For example,
observed in data on fictive locomotion and simu- label the voltages of neurons in a half-center CPG
lations that reproduce it (Yakovenko et al. 2005; as v1, v2. Suppose that f = 0 occurs when the
Rybak et al. 2006). voltage v1 reaches its maximum. If the two neu-
A third scenario that can yield different rons are identical, then v2 reaches its maximum at
response properties of half-center CPGs is the f = 1/2. If transitions occur by escape, then
application of transient stimulation. Suppose stimulation of neuron 1 at f (0, 1/2) will
that a CPG is oscillating with a period T. Assign have little impact, since the next phase transition
time 0 to correspond to the time when the voltage relies on the escape of unstimulated neuron 2.
of one cell increases through some fixed level. Stimulation at f (1/2,1), when neuron 1 is
Under this time labeling, the time at which this suppressed, may accelerate its escape and hence
C 660 Compartment Model for Nuclear Medicine
advance the phase of the oscillation, although Wang X-J, Rinzel J (1992) Alternating and synchronous
details relating to the timing, amplitude, and rhythms in reciprocally inhibitory model neurons.
Neural Comput 4:84–97
duration of stimulation matter. If transitions Yakovenko S, McCrea DA, Stecina K, Prochazka
occur instead by release, then stimulation of neu- A (2005) Control of locomotor cycle durations.
ron 1 at f in some range within (0, 1/2) will help J Neurophysiol 94:1057–1065
it continue to suppress neuron 2 and hence cause Zhang C, Lewis TJ (2013) Phase response properties
of half-center oscillators. J Comput Neurosci
a delay, especially if v1 is dropping toward syn- 35:55–74
aptic threshold when the stimulation arrives.
Stimulation at f (1/2,1) in this case arrives
when neuron 1 is suppressed and has little effect
on when neuron 2 releases neuron 1 and hence Compartment Model for Nuclear
relatively little impact on phase. In summary, the Medicine
advance or delay in oscillations induced by
a transient stimulation at a known oscillation ▶ Kinetic Models for PET/SPECT Imaging
phase depends on whether the transitions in
a network are by escape or release. A major
phase delay can only arise in the release case,
and stimulation in the later part of the active
phase of the neuron being stimulated is particu- Compartmental Models of Spinal
larly well suited to distinguish escape (little Motoneurons
change in phase) from release (significant phase
delay) (Zhang and Lewis 2013). Randy Powers
Department of Physiology and Biophysics,
University of Washington, Seattle, WA, USA
References
Curtu R, Shpiro A, Rubin N, Rinzel J (2008) Mechanisms Definition

for frequency control in neuronal competition models.
SIAM J Appl Dyn Syst 7:609–649 Compartmental neuron models represent the
Daun S, Rubin JE, Rybak IA (2009) Control of oscillation
complex geometry of a neuron as a set of electri-
periods and phase durations in half-center central pat-
tern generators: a comparative mechanistic analysis. cally connected compartments, each of which is
J Comput Neurosci 27:3–36 considered to be isopotential (i.e., the membrane
Ermentrout GB, Terman D (2010) Mathematical founda- voltage does not vary along the length of the
compartment). Compartmental models of spinal
Hooper S (2001) Central pattern generators. In: Encyclo-
pedia of Life Sciences. Wiley. http://onlinelibrary. motoneurons can be divided into three groups
wiley.com/book/10.1002/047001590X. Accessed 17 based on the extent to which they rely on
Aug 2013 a simplified version of dendritic morphology:
Izhikevich E (2007) Dynamical systems in neuroscience:
(1) detailed compartmental models, which
the geometry of excitability and bursting. MIT Press,
Cambridge attempt to accurately represent the dendritic mor-
Rybak IA, Shevtsova NA, Lafreniere-Roula M, McCrea phology based on anatomical reconstructions of
DA (2006) Modeling spinal circuitry involved in loco- stained motoneurons; (2) cable models, in which
motor pattern generation: insights from deletions dur-
the branching structure of the dendritic tree is
ing fictive locomotion. J Physiol 577:617–639
Skinner FK, Turrigiano GG, Marder E (1993) Frequency collapsed into one or more equivalent cables;
and burst duration in oscillating neurons and two-cell and (3) two-compartment models, in which the
networks. Biol Cybern 69:375–383 dendritic tree is represented by a single lumped
Skinner FK, Kopell N, Marder E (1994) Mechanisms for
compartment that is connected to a compartment
oscillation and frequency control in reciprocally inhib-
itory model neural networks. J Comput Neurosci representing the soma and initial portion of
1:69–87 the axon.
Compartmental Models of Spinal Motoneurons 661 C
Detailed Description Formulation of Compartmental Models
Although the equivalent cylinder approximation
Motoneurons have the largest dendritic trees of any and linear cable theory were extremely useful in
neuron in the mammalian central nervous system providing information about the basic electrical
(Kernell 2006). Unlike pyramidal and Purkinje properties of motoneurons as well as the flow
neurons, whose dendritic trees primarily project of synaptic current in the absence of voltage-
in a single plane, motoneuron dendrites typically dependent conductances, the compartmental C
project radially from the soma in all directions and representation is needed to model current and
branch extensively. As a result, it is extremely voltage changes along the cable in the presence
difficult to record directly from motoneuron den- of voltage-dependent conductances. In the com-
drites, and our understanding of the properties of partmental representation, the cable (or branched
motoneuron dendrites is based almost exclusively dendritic tree) is divided into discrete compart-
on indirect evidence. Compartmental models of ments, each of which is assumed to be
motoneurons are therefore useful tools to infer isopotential along its length. This representation
dendritic properties based on intracellular and replaces the cable equation, which is partial dif-
whole-cell recordings from the soma. ferential equation that is first order in time and
second order in space, with a set of coupled,
Equivalent Cylinder Representation of the ordinary first-order differential equations
Dendritic Tree describing current flow in each compartment
Early compartmental models of motoneurons and between adjacent compartments.
represented the dendritic tree as an equivalent The basic equation describing current flow in
cylinder. Wilfrid Rall showed that if dendritic a compartment is based on Kirchhoff’s current
branching followed certain rules, the entire den- law, which states that for each compartment the
dritic tree could be collapsed into a single equiv- net current through the membrane (im) is equal to
alent constant diameter cylinder. This allowed the net longitudinal current that enters and leaves
the use of linear cable theory to derive analytical the compartment through the interior of the com-
solutions for the effects of injecting current at one partment. The membrane current is in turn com-
point on the cable on the current and voltage posed of capacitative current and ionic current.
measured at other points along the cable The membrane equation for compartment j,
(Jack et al. 1975; Koch 1999; Rall 1977; Rall connected to compartments j 1 and j + 1,
et al. 1992). This theory provided estimates of is thus
the electrical length of the dendrites based on

measurements of the decay of somatic membrane cm_j dVj =dt þ Iion_j ¼ gj1, j Vj1 Vj
voltage following a brief pulse of injected cur-
rent. The distal ends of motoneuron dendrites gj, jþ1 Vj Vjþ1 ,
were estimated to be an average of 1–2 length
constants (l) away from the soma. (A length con- where Vj1, Vj, and Vj+1 are the membrane
stant is distance over which the voltage produced potentials across compartments j1, j, and
by a steady-state injection of current applied at j + 1; cm_j is the capacitance of compartment j;
one point on the cable falls to 1/e of its value at and gj1,j and gj,j+1 are the coupling conduc-
the point of current injection.) These estimates tances (reciprocal of the axial resistances)
implied that distal synapses could exert an appre- between adjacent compartments (Segev
ciable effect on the somatic membrane potential, et al. 1989). If we divide the ionic current into
contrary to earlier suggestions (Eccles 1964). The that following across a voltage-independent leak
theory also was used to estimate the relative conductance (gL) and that flowing across
location of synapses based on the size and time a number of different voltage-dependent con-
course of synaptic potentials measured at ductances (Sgi(V)) and rearrange terms, the
the soma. basic differential equation is
C 662 Compartmental Models of Spinal Motoneurons

dVj =dt ¼ 1=cm_j gL Vj VL Sgk Vj Vj Vi þ Vj1 Vj þ gj1, j Vj1 Vj gj, jþ1 Vj Vjþ1 ,
where VL is the resting (leak) conductance and nine- and six-fold times those on the soma. The
the Vis are the reversal potentials for ions flowing increased density and lower threshold of channels
through the voltage-dependent conductances. on the initial segment are common features of
The voltage-dependent conductances are the compartmental models of motoneurons (and
product of a fixed maximal conductance and other neurons as well) and fit well with the avail-
one or more state variables that describe how able experimental evidence (Bender and Trussell
the proportion of open channels changes with 2012).
voltage and time. A separate set of differential Until recently, it was generally assumed that
equations describes the evolutions of the state normal motoneuron dendrites are passive, i.e.,
variables, generally of the following form: they have no voltage- or calcium-dependent
channels. As described later, it is now clear that
dm=dt ¼ ðm1 ðVÞ mÞ=tm ðVÞ, motoneuron dendrites have a variety of active
conductances. However, even early recordings
where m1(V) is the steady-state voltage depen- of motoneuron properties indicated that follow-
dence of channel open probability and tm(V) is ing axotomy motoneurons develop a number of
the voltage-dependent time constant. Channels features, such as all-or-none partial spikes
can also be gated by internal calcium superimposed upon synaptic potentials (Kuno
(or sodium), and in these cases additional differ- and Llinas 1970), which are consistent with the
ential equations are needed to represent the buff- presence of active dendritic conductances. The
ering, extrusion, and diffusion of calcium input–output properties of motoneurons are also
(or sodium). altered following axotomy. In normal motoneu-
rons the relation between injected current and
Early Models of Action Potential Generation firing rate is composed of up to three linear seg-
and Repetitive Firing ments: a shallow primary range at lower levels of
Dodge and Cooley (1973) adapted the Hodgkin injected current and a steeper, secondary range at
and Huxley (1952) description of the gating of the higher levels, followed in some cases by
sodium (Na) and potassium (K) conductances a shallow tertiary range at the highest levels of
underlying the action potential in squid axons to injected current (Kernell 2006). Following
model the generation of action potentials in spi- axotomy, the frequency-current (F-I) relation is
nal motoneurons. The motoneuron was characterized by a single linear range whose
represented as a series of cylindrical compart- slope is similar to that of the secondary range in
ments representing the proximal axon and soma, normal motoneurons (Heyer and Llinas 1977).
and the dendritic tree was represented as an Traub and Llinas (1977) explored the potential
equivalent cylinder divided into 30 compart- mechanisms underlying the differences in repet-
ments. The model provided a good match to itive firing in normal and axotomized motoneu-
experimentally recorded action potentials and rons using compartmental models of varying
somatic voltage-clamp recordings when Na and complexity, ranging from a single cable repre-
K channels were confined to the soma and axon sentation of motoneuron dendrites to up to four
initial segment. In addition, in order to match the branching dendrites. An essential feature of all of
experimental data, the voltage dependence of the models was the inclusion of a slow potassium
sodium channels on the initial segment was conductance underlying the medium-duration
shifted to 10 mV more hyperpolarized than afterhyperpolarization (mAHP). This conduc-
those on the soma, and the densities of Na and tance is activated by the influx of calcium
K channels on the initial segment were roughly through calcium channels that are opened during
the spike. Both the calcium (Ca) and calcium- evoked by electrical stimulation of multiple Ia
activated potassium channels (KCa) were placed afferents (composite Ia EPSPs) or from the
in compartments corresponding to the proximal stretch-evoked activation of single Ia afferents
dendrites, and Na channels were extended to the (unitary or single-fiber Ia EPSPs) provided
proximal dendrites to provide sufficient depolar- a number of lines of evidence that the synaptic
ization to activate the calcium channels. The terminals of Ia afferents are distributed widely
authors examined the effects of several different across the dendritic trees of motoneurons (see C
spatial distributions of Na and KCa channels on Burke 1967, and references therein). Single-
model behavior and found a range of distributions fiber Ia EPSPs exhibit a wide variety of shapes
that could reproduce the F-I behavior of normal that can be quantified by plotting the EPSP width
and axotomized motoneurons. Secondary range at half amplitude (half-width) versus the time to
firing resulted from a decrease in action potential- peak. EPSPs with short times to peak and half-
induced dendritic depolarization at higher levels widths were presumed to arrive from proximal
of injected current, resulting in a decrease in KCa synapses, whereas distal synapses were thought
activation. The behavior of axotomized motoneu- to underlie those with long time to peak and half-
rons was reproduced by lowering the Na density widths. Simulations with a 10-compartment
on the proximal dendrites (leading to secondary cable model (Rall et al. 1967) could reproduce
range firing at low levels of injected current) and the experimentally recorded range of EPSP
putting hot spots of Na channels at 0.6–0.7 l from shapes if it was assumed that single Ia afferents
the soma. The distal Na hot spots mediated local gave rise to a number of synaptic terminals at
spikelike depolarizations in response to synaptic different distances from the soma, an assumption
inputs that propagated decrementally to the soma, that was later validated by experimental evidence
leading to the spikelike partial responses (Burke and Glenn 1996). This model was also
superimposed on the synaptic potentials recorded able to account for a number of other experimen-
at the soma. tal findings related to the summation of Ia com-
Although the mechanism for partial responses posite EPSPs, their sensitivity to changes in
proposed by Traub and Llinas (1977) is consis- somatic membrane potential and the ability to
tent with the available experimental evidence detect changes in membrane impedance during
(Sernagor et al. 1986), secondary range firing the EPSP.
does not seem to depend upon a saturation or The acquisition of detailed anatomical data on
reduction of the KCa-mediated mAHP the structure of motoneuron dendrites and the
(Schwindt and Calvin 1973). More recent com- projection of single Ia afferent fibers have guided
partmental models based on an alternative the development of models that provide a much
mechanism of secondary range firing are consid- more precise representation of the structure and
ered later. specific membrane properties of motoneuron
dendrites as well as the spatial distribution of Ia
Synaptic Integration in Passive Dendritic afferent synaptic terminals. Detailed models of
Trees reconstructed spinal motoneurons have shown
Both cable models and models with a detailed that the experimentally recorded voltage
representation of the dendritic tree have been responses to pulses and steps of current injected
used to investigate the relation between the struc- into the soma can only be reproduced if the spe-
ture and specific membrane properties of moto- cific membrane resistivity is lower in the soma
neuron dendrites and the delivery of current from than in the dendrites (Clements and Redman
dendritic synapses to the soma. The best-studied 1989; Fleshman et al. 1988). Although part of
synaptic input to spinal motoneurons is the mono- this decrease in somatic resistivity may arise
synaptic excitatory input from primary muscle from the leak induced by penetration of the
spindle (Ia) afferents. Somatic recordings of soma with a sharp electrode, this is unlikely to
excitatory postsynaptic potentials (EPSPs) be the only factor contributing to variations in
membrane resistivity (see Sect. 3.3.2 in Powers Korogod et al. (2000) calculated synaptic
and Binder 2001). Unfortunately, there are charge transfer for synapses throughout the den-
a number of different spatial distributions of spe- dritic tree in detailed compartmental models of
cific membrane resistivity that can reproduce the reconstructed motoneurons at different levels of
experimental data; for example, a step distribu- tonic synaptic input (simulated by changing the
tion (low in the soma and high and uniform in effective membrane resistivity). In models of cat
the dendrites) and a sigmoidal distribution spinal motoneurons, all of the synapses exhibited
(resistivity rises as a sigmoidal function of dis- high transfer effectiveness (i.e., they transmitted
tance from the soma) give equally good fits to the at least 50 % of their charge) at rest. However, at
experimental data. Somewhat surprisingly, when high levels of tonic synaptic activity, only the
Ia afferent terminals are distributed onto the most proximal synapses were effective. Cushing
models in accordance with the available ana- et al. (2005) used compartmental models of
tomical data, the step and sigmoidal models reconstructed motoneurons to explicitly estimate
give equally good fits to the experimentally the total amount of synaptic current that could be
recorded EPSPs (Segev et al. 1990). Nonethe- delivered to the soma during increasing activa-
less, the spatial distribution of voltage changes tion of excitatory dendritic synapses. The relation
produced by activating Ia afferents will depend between the effective synaptic current (i.e., the
upon the exact spatial distribution of specific portion of current that reached the soma) and the
membrane resistivity; this may be particularly level of excitatory drive was highly nonlinear,
important when considering how synaptic inputs due to increases in membrane conductance and
are modified by voltage-dependent dendritic decreases in synaptic driving force as increasing
conductances. numbers of synapses were activated. The effec-
tive synaptic current produced by high-frequency
Limited Charge Transfer in Passive Dendritic (100 Hz) activation of all of the excitatory syn-
Trees apses was well below that needed to maximally
Although analysis based on both cable theory and drive motoneurons. This result fits with the
compartmental models indicates that a significant experimental finding that the effective synaptic
fraction of the charge entering through even the currents produced by high-frequency activation
most distal synapses on motoneurons can reach of various peripheral and descending inputs to
the soma, that situation changes as more and motoneurons were rather modest (Powers and
more synapses are activated. The total conduc- Binder 2001).
tance of a patch of membrane (in the absence of
voltage-dependent conductances) is the sum of Repetitive Firing and Synaptic Integration in
the resting conductance and the conductance con- Motoneuron Models with Active Dendritic
tributed by activated synapses, and even at rela- Trees
tively low levels of synaptic activation, the total Many of the early experimental studies of synap-
conductance begins to be dominated by the syn- tic integration and repetitive firing in motoneu-
aptic conductance term (Barrett 1975). This rons were based on recordings obtained in
increased conductance together with a decrease anesthetized animals. An alternative preparation
in driving force as the membrane is depolarized is the unanesthetized, decerebrate animal, which
toward the synaptic reversal potential reduces the has been used for over a century to study spinal
amount of synaptic charge reaching the soma. and postural reflexes. In this preparation, the
Barrett (1975) concluded that physiological cerebrum is removed under anesthesia, rendering
levels of activation of excitatory input to moto- the animal insensate, obviating the need for fur-
neurons could deliver enough current to produce ther anesthetic. In the absence of anesthetic and
high rates of repetitive discharge, but recent sim- inhibitory inputs from the cortex, monoaminergic
ulations using detailed compartmental models neurons in the brainstem that project to the
suggest that this may not be the case. spinal cord exhibit tonic activity. This tonic
monoaminergic drive facilitates the activation of currents at more depolarized voltages (Heckman
a number of voltage-sensitive channels, includ- et al. 2003). This N-shaped relation has two stable
ing those present on motoneuron dendrites steady-state voltages: one near the resting poten-
(Heckman et al. 2003). Under these conditions, tial and one corresponding to a depolarized pla-
motoneurons exhibit a number of behaviors that teau potential. Bistable behavior occurs when
are not observed in anesthetized preparations, injected or synaptic current produces a dendritic
and these same behaviors occur in in vitro prep- plateau that persists when the excitatory input C
arations in the presence of monoamines stops, since the plateau depolarization represents
(Heckman et al. 2003). These behaviors include a stable state. In response to a triangular injected
the following: (1) bistable discharge, the moto- current waveform, the onset of the dendritic pla-
neuron continues to discharge after the cessation teau leads to an increase in the slope of the F-I
of a brief excitatory stimulus; (2) hysteresis, in relation (secondary range firing), followed by
response to a triangular injected current wave- a decrease in slope (tertiary range firing) once
form, the motoneuron stops discharging at a stable plateau is reached (Elbasiouny
a lower level of injected current than is required et al. 2006; Powers et al. 2012). Hysteresis in
to recruit the motoneuron on the ascending limb the F-I relation occurs because the site of the
of the command; and (3) voltage-dependent syn- dendritic plateau is at a significant electrical dis-
aptic amplification, excitatory synaptic currents tance from the soma, so that the plateau is turned
are increased with depolarization. All of these off at a lower somatic voltage that was required to
behaviors are thought to be mediated by persis- initiate the plateau.
tent inward currents (PICs) generated primarily Simulations based on detailed compartmental
by low-threshold calcium channels on the den- models suggest that the best matches to the exper-
drites (Heckman et al. 2003). imental data are obtained when the Ca channels
PIC-mediated behaviors can be replicated by responsible for the plateaus are restricted to
a variety of compartmental models that place the a relatively narrow range of distances centered
spike-generating (Na and K) conductances on or at roughly 0.5 l away from the soma, although
near the soma compartment and PIC-generating the exact length and location of these dendritic
conductances on the dendrites. These include “hot spots” depend somewhat on which types of
two-compartment models (Booth et al. 1997; experimental data (both electrophysiological and
Kim and Jones 2010; Venugopal et al. 2011), anatomical) are used to constrain the models (Bui
cable models (Carlin et al. 2009; Powers et al. 2006; Elbasiouny et al. 2005; Grande
et al. 2012; Shapiro and Lee 2007), and detailed et al. 2007b). The presence of multiple hot spots
compartmental models (e.g., Bui et al. 2006; of Ca channels may allow motoneurons to
Elbasiouny et al. 2005, 2006). The consensus amplify synaptic inputs in a graded manner as
view based on these simulations is that the synaptic input is increased, since the effective
bistable behavior and hysteresis observed in thresholds for activating Ca channels may vary
somatic recordings arise from a plateau depolar- along the dendritic tree as a result of variations in
ization in the dendrites following voltage- the local geometry (Bui et al. 2006; Elbasiouny
dependent activation of the PIC. The PIC acts in et al. 2006; Carlin et al. 2009). A nonuniform
combination with the leak conductance (and distribution of Ca channels also can allow
probably voltage-dependent K conductances) to differential amplification of different synaptic
give rise to a dendritic current–voltage (I-V) rela- inputs, with increased amplification occurring as
tion that is N-shaped, reflecting the contribution the spatial distributions of synaptic terminals and
of the leak current at hyperpolarized voltages Ca channels show increasing overlap (Bui
(leading to a positive slope in the I-V relation) et al. 2008; Grande et al. 2007a).
and the predominance of the PIC at intermediate There is also evidence that dendritic
voltages (leading to a negative I-V slope), (as well as somatic and axonal) Na channels
followed by a predominance of outward contribute to PICs and synaptic amplification
(Heckman et al. 2003; Powers et al. 2012). One of different channel types. As a result, the actual
intriguing possibility, based on a recent simula- spatial distribution of a particular channel may be
tion study with a two-cable compartmental significantly different from the one that provides
model, is that bistable firing and synaptic ampli- the best fit to experimental data for a given
fication arise in different parts of the dendritic model, since the optimal distribution is likely to
tree with different PIC-generating channels vary with the channel properties and the presence
(Shapiro and Lee 2007). In this model, all-or- of other channels. Nonetheless, the spatial sepa-
none plateau generation by Ca channels gives ration of spike-generating and plateau-generating
rise to bistable behavior, whereas graded conductances is a common feature of many dif-
activation of Na channels produces synaptic ferent types of compartmental models of spinal
amplification. motoneurons and at this point appears to provide
Although there is not enough experimental a fairly accurate representation of a range of
and simulation data to specify the spatial distri- motoneuron behaviors.
bution of Na and Ca channels mediating PIC, the
data do support the idea that most of the PIC is
derived from dendritic channels (Heckman Cross-References
et al. 2003), and this has important implications
for the synaptic control of motoneuron firing rate. ▶ GENESIS, the GEneral NEural SImulation
Since most of the synaptic input is on the den- System
drites, the dendritic plateaus have a lower thresh- ▶ NEURON Simulation Environment
old for activation (as measured at the soma) when
driven by excitatory synaptic input than by cur-
rent injected into the soma (Bui et al. (2006) in References
fact used experimental data of this type to set
their channel distributions). As a result, in both Barrett JN (1975) Motoneuron dendrites: role in synaptic
integration. Fed Proc 34:1398–1407
cable (Powers et al. 2012) and detailed compart-
Bender KJ, Trussell LO (2012) The physiology of
mental models (Elbasiouny et al.), an increase in the axon initial segment. Annu Rev Neurosci
excitatory drive can lead to a rapid increase in 35:249–265
firing rate as soon as the motoneuron starts to fire Booth V, Rinzel J, Kiehn O (1997) Compartmental model
of vertebrate motoneurons for Ca2+-dependent spiking
(secondary range firing), followed by a much
and plateau potentials under pharmacological treat-
shallower increase in rate with further increases ment. J Neurophysiol 78:3371–3385
in excitation (tertiary range firing). Similar pat- Bui TV, Ter-Mikaelian M, Bedrossian D, Rose PK
terns of firing are seen in human motor units (2006) Computational estimation of the distribution
of L-type Ca(2+) channels in motoneurons based on
during slowly increasing isometric contractions,
variable threshold of activation of persistent inward
supporting the hypothesis that PICs provide an currents. J Neurophysiol 95:225–241
important contribution to motoneuron discharge Bui TV, Grande G, Rose PK (2008) Relative location
during a variety of motor behaviors (Heckman of inhibitory synapses and persistent inward
currents determines the magnitude and mode of syn-
et al. 2008).
aptic amplification in motoneurons. J Neurophysiol
99:583–594
Limitations Burke RE (1967) Composite nature of the monosynaptic
As is true of almost all compartmental models, excitatory postsynaptic potential. J Neurophysiol
30:1114–1137
the experimental data is not sufficient to constrain Burke RE, Glenn LL (1996) Horseradish peroxidase study
the parameters (channel properties and channel of the spatial and electrotonic distribution of group Ia
distribution) of compartmental models of spinal synapses on type-identified ankle extensor motoneu-
motoneurons. This limitation is particularly rons in the cat. J Comp Neurol 372:465–485
Carlin KP, Bui TV, Dai Y, Brownstone RM (2009) Stair-
severe for motoneuron models due to the lack of
case currents in motoneurons: insight into the spatial
direct dendritic recordings and the limited immu- arrangement of calcium channels in the dendritic tree.
nocytochemical data on the dendritic distribution J Neurosci 29:5343–5353
Clements JD, Redman SJ (1989) Cable properties of cat reconfiguration of the effective dendritic field of moto-
spinal motoneurones measured by combining voltage neurons. J Comp Neurol 422:18–34
clamp, current clamp and intracellular staining. Kuno M, Llinas R (1970) Enhancement of synaptic trans-
J Physiol Lond 409:63–87 mission by dendritic potentials in chromatolysed
Cushing S, Bui T, Rose PK (2005) Effect of nonlinear motoneurones of the cat. J Physiol 210:807–821
summation of synaptic currents on the input–output Powers RK, Binder MD (2001) Input–output functions of
properties of spinal motoneurons. J Neurophysiol mammalian motoneurons. Rev Physiol Biochem
94:3465–3478 Pharmacol 143:137–263
Dodge FA, Cooley JW (1973) Action potential of the Powers RK, Elbasiouny SM, Rymer WZ, Heckman CJ
C
motoneuron. IBM J Res Dev 17:219–229 (2012a) Contribution of intrinsic properties and syn-
Eccles JC (1964) The physiology of synapses. Springer, aptic inputs to motoneuron discharge patterns:
Berlin a simulation study. J Neurophysiol 107:808–823
Elbasiouny SM, Bennett DJ, Mushahwar VK (2005) Sim- Powers RK, Nardelli P, Cope TC (2012b) Frequency-
ulation of dendritic CaV1.3 channels in cat lumbar dependent amplification of stretch-evoked excitatory
motoneurons: spatial distribution. J Neurophysiol input in spinal motoneurons. J Neurophysiol
94:3961–3974 108:753–759
Elbasiouny SM, Bennett DJ, Mushahwar VK (2006) Sim- Rall W (1977) Core conductor theory and cable properties
ulation of Ca2+ persistent inward currents in spinal of neurons. In: Handbook of physiology, The nervous
motoneurones: mode of activation and integration of system, cellular biology of neurons. American Physi-
synaptic inputs. J Physiol 570:355–374 ological Society, Bethesda, pp 39–97
Fleshman JW, Segev I, Burke RB (1988) Electrotonic Rall W, Burke RE, Smith TG, Nelson PG, Frank K (1967)
architecture of type-identified alpha-motoneurons in Dendritic location of synapses and possible mecha-
the cat spinal cord. J Neurophysiol 60:60–85 nisms for the monosynaptic EPSP in motoneurons.
Grande G, Bui TV, Rose PK (2007a) Effect of localized J Neurophysiol 30:1169–1193
innervation of the dendritic trees of feline motoneu- Rall W, Burke RE, Holmes WR, Jack JJ, Redman SJ,
rons on the amplification of synaptic input: Segev I (1992) Matching dendritic neuron models to
a computational study. J Physiol 583:611–630 experimental data. Physiol Rev 86:S159–S185
Grande G, Bui TV, Rose PK (2007b) Estimates of the Schwindt PC, Calvin WH (1973) Nature of conductances
location of L-type Ca2+ channels in motoneurons of underlying rhythmic firing in cat spinal motoneurons.
different sizes: a computational study. J Neurophysiol J Neurophysiol 36:955–973
97:4023–4035 Segev I, Fleshman JW, Burke RB (1989) Compartmental
Heckman CJ, Lee RH, Brownstone RM models of complex neurons. In: Koch C, Segev I (eds)
(2003) Hyperexcitable dendrites in motoneurons and Methods in neuronal modeling: from synapses to net-
their neuromodulatory control during motor behavior. works. MIT Press, Cambridge, MA
Trends Neurosci 26:688–695 Segev I, Fleshman JWJ, Burke RE (1990) Computer sim-
Heckman CJ, Johnson M, Mottram C, Schuster J (2008) ulation of group Ia EPSPs using morphologically real-
Persistent inward currents in spinal motoneurons and istic models of cat alpha-motoneurons. J Neurophysiol
their influence on human motoneuron firing patterns. 64:648–660
Neuroscientist 14:264–275 Sernagor E, Yarom Y, Werman R (1986) Sodium-
Heyer CB, Llinas R (1977) Control of rhythmic firing in dependent regenerative responses in dendrites of
normal and axotomized cat spinal motoneurons. axotomized motoneurons in the cat. Proc Natl Acad
J Neurophysiol 40:480–488 Sci U S A 83:7966–7970
Hodgkin AL, Huxley AF (1952) A quantitative descrip- Shapiro NP, Lee R (2007) Synaptic amplification versus
tion of membrane current and its application to con- bistablility in motoneuron dendritic processing: a
duction and excitation in nerve. J Physiol 116:500–544 top-down modeling approach. J Neurophysiol
(London) 97:3948–3960
Jack JJ, Noble D, Tsien RW (1975) Electric current flow in Traub RD, Llinas R (1977) The spatial distribution of
excitable cells. Oxford University Press, Oxford ionic conductances in normal and axotomized moto-
Kernell D (2006) The motoneurone and its muscle fibers. neurons. Neuroscience 2:829–849
Oxford University Press, Oxford Venugopal S, Hamm TM, Crook SM, Jung R (2011) Mod-
Kim H, Jones KE (2010) Asymmetric electrotonic cou- ulation of inhibitory strength and kinetics facilitates
pling between the soma and dendrites alters the regulation of persistent inward currents and motoneu-
bistable firing behaviour of reduced models. ron excitability following spinal cord injury.
J Comput Neurosci J Neurophysiol 106:2167–2179
processing in single neurons. Oxford University Press,
New York Further Reading
Korogod SM, Kulagina IB, Horcholle-Bossavit G, Genesis http://www.genesis-sim.org
Gogan P, Tyc-Dumont S (2000) Activity-dependent Neuron http://www.neuron.yale.edu/neuron
C 668 Compilation of Published Tracer Injection Studies in the Macaque Brain
subcortical areas. Prefrontal cortex, motor cortex,

Compilation of Published Tracer and the striatum are among the most densely
Injection Studies in the Macaque innervated areas. DA release tends to be fairly
Brain homogeneous across DA neurons, by virtue of
electrical coupling between the axons of adjacent
▶ Collations of Connectivity Data on the neurons, which induces highly synchronous fir-
Macaque Brain (CoCoMac) ing. As pointed out by Paul Glimcher, these prop-
erties suggest that DA neurons “cannot say much
to the rest of the brain but what they say must be
widely heard” (Glimcher 2011).
Complex Network Theory in DA affects its targets primarily by two classes
Neuroscience of receptor – D1 and D2 – and the distinct prop-
erties of these receptors have functional conse-
▶ Network Theory in Neuroscience quences which are discussed below. D2 receptors
tend to be activated earlier, and at lower concen-
trations, than D1 receptors (Lapish et al. 2007;
Schultz 2007). As a consequence, D2 receptors
Computation with Dopaminergic are more sensitive to phasic (transient) DA
Modulation release than D1 receptors, with the opposite pat-
tern for tonic (background) DA release (Grace
Sam Gershman 1991). D1 and D2 receptors also have different
Department of Psychology and Neuroscience postsynaptic effects: D1 receptor activation
Institute, Princeton University, Princeton, increases both NMDA (excitatory) and GABA
NJ, USA (inhibitory) currents, whereas D2 receptor acti-
vation decreases them (Trantham-Davidson
et al. 2004).
Definition
Reinforcement Learning Models
Dopamine (DA) is a neuromodulator released One of the most influential hypotheses about DA
by midbrain neurons with widespread projec- is that it plays an integral role in a system that
tions throughout the brain. Dopaminergic mod- learns to predict future reward. This hypothesis
ulation has diverse effects on cellular, motor, can be formalized in terms of the theory of rein-
and cognitive functions, including reinforce- forcement learning (RL) first developed in com-
ment learning, working memory, and attention. puter science (Sutton and Barto 1998). At each
Dysregulation of dopamine also plays a central time t the agent occupies a state st (e.g., the
role in the breakdown of these functions in agent’s location or the surrounding stimuli) and
disorders such as Parkinson’s disease and receives a reward rt. The goal of an RL system is
schizophrenia. to estimate the expected discounted future return,
or value, of visiting a sequence of states starting
in state st:
" #
X
Dopamine Basics V ðst Þ ¼ E k
g r tþk , (1)
DA is a neuromodulator released by neurons in k¼0
two nuclei of the midbrain: the ventral tegmental

area (VTA) and the substantia nigra pars where g is a discount factor that accounts for the
compacta (SNc). The axons of these neurons fact that distal rewards are less valuable than
project to a large number of cortical and proximal rewards. The expectation E represents
Computation with Dopaminergic Modulation 669 C
an average of the quantity inside the brackets increase reward. For example, electrophysiolog-
over stochastic sequences of states. ical studies have shown that DA responses
If we make certain assumptions about the to reward are sensitive to actions (Morris
environment (specifically, that it is a Markov et al. 2006; Roesch et al. 2007), and DA receptor
decision process; see Sutton and Barto 1998), antagonism in the striatum suppresses the initia-
then the value can be estimated by a simple learn- tion of reward-seeking actions (Berridge and
ing rule known as temporal difference learning: Robinson 1998). According to RL theory, there C
are a number of ways to use prediction errors
V^ðst Þ V^ðst Þ þ adt , (2) to estimate the value of actions, some of which
have been integrated into models of DA
where a is a learning rate and dt is a prediction (e.g., Frank 2005; Houk et al. 1995; Montague
error defined as: et al. 1996; Suri and Schultz 1998).
Since the seminal work of Schultz
et al. (1997), the prediction error theory of DA
dt ¼ r t þ gV^ðstþ1 Þ V^ðst Þ: (3) has been corroborated by converging evidence
from humans and other animals (see Glimcher
Intuitively, if the prediction error is greater 2011 for a review). At the same time, it is clear
than zero, then more reward was received than that such a simple theory is unable to account for
was predicted, and hence V^ðst Þ is increased. If the many of the complexities of DA function. These
prediction error is less than zero, then V^ðst Þ is lacunae have stimulated a number of significant
decreased. Another implication of this learning extensions of the theory, some of which are
rule is that prediction errors will gradually prop- described below. Interestingly, these extensions
agate backwards in time as values at earlier time mirror the technical development of RL theory
points come to reflect reward received at later in computer science. Biological and artificial
time points. agents face similar challenges in optimizing
Schultz et al. (1997) observed that the phasic reward – and may have even arrived at similar
firing of DA neurons in an appetitive condition- solutions.
ing experiment was conspicuously consistent One development of the prediction error the-
with a hypothetical prediction error. In these ory concerns the phasic DA firing in response to
experiments, a conditioned stimulus (CS, e.g., novel stimuli (e.g., Horvitz et al. 1997). The
a tone) is repeatedly followed by an uncondi- original prediction error theory fails to anticipate
tioned stimulus (US, e.g., juice reward). Schultz this finding, since novelty does not by itself pre-
et al. noticed that DA firing increased in response dict reward. However, as pointed out by Kakade
to an unexpected reward (e.g., in early training and Dayan (2002), novelty responses can be
trials). After multiple training trials, if the reward rationalized by considering the problem of opti-
is delivered following the CS, then DA firing is at mizing reward. All agents face a dilemma
baseline following reward delivery but increases between exploiting the currently best option and
following the CS onset. In this case, the reward is exploring other options that might be potentially
expected and hence the prediction error at the better (Cohen et al. 2007). One way to encourage
time of reward is zero, whereas the prediction exploration is by initializing the values of all
error at the time of the CS is greater than zero states to be greater than zero. As shown by
due to the temporal propagation described above. Kakade and Dayan, this has the effect of gener-
Finally, if the reward is omitted following the CS, ating positive prediction errors when these states
DA firing pauses (decreases below baseline), are visited for the first time, consistent with the
consistent with the occurrence of a negative pre- DA novelty response.
diction error. A second development has been to incorporate
In addition to its role in predicting reward, DA timing variability and partial observability into
plays an important role in selecting actions to the TD model (Daw et al. 2006). The original
C 670 Computation with Dopaminergic Modulation
prediction error theory (Montague et al. 1996; express D2 receptors. The net effect of DA in the
Schultz et al. 1997) used a delay line to represent striatum is to increase activity in the direct path-
different temporal epochs as distinct states. Daw way and suppress activity in the indirect pathway,
et al. (2006) replaced the delay line with two resulting in disinhibition of action selection in the
assumptions: (1) States are occupied for some frontal cortex via modulation of basal ganglia
random amount of time before a transition occurs output pathways. One consequence of this differ-
(timing variability); and (2) states are not directly ential effect is that neurons in the direct pathway
observed, but must be inferred from noisy undergo long-term potentiation following a DA
sensory data (partial observability). Daw burst, whereas neurons in the indirect pathway
et al. showed that these extensions could account undergo long-term depression. Another contribu-
for a number of timing properties of phasic DA tion of this model is a division of error-driven
firing. For example, if a reward is delivered early learning into positive and negative components:
or late, a phasic DA burst is observed (Hollerman D1-expressing neurons learn more from positive
and Schultz 1998), contrary to the delay line prediction errors, whereas D2-expressing neu-
model, which predicts a negative prediction rons learn more from negative prediction errors.
error (i.e., a pause in firing). The Daw Thus, the model of Frank (2005) provides
et al. model correctly predicts a positive predic- a neurally plausible mechanism by which DA
tion error, because it infers that a transition to the can influence action selection in corticostriatal
“reward state” has occurred early or late. circuits.
A third development of the theory concerns
the role of tonic DA. It has long been recognized Gain Modulation and Working Memory
that tonic levels of DA appear to be regulated The sensitivity of a neuron to stimulation can be
independently of phasic levels (Grace 1991). In enhanced by the application of DA, without
particular, increased tonic DA has been associ- changing the spontaneous firing rate (Clarke
ated with the invigoration of behavior (Schultz et al. 1987; Foote and Morrison 1987). In other
2007), and the depletion of tonic DA results in words, DA increases the signal-to-noise ratio of
a loss of vigor (Salamone et al. 2001; Sokolowski a neuron. Servan-Schreiber et al. (1990) intro-
and Salamone 1998). To account for these obser- duced a connectionist implementation of this
vations, Niv et al. (2007) proposed that tonic DA idea by positing that DA increases the gain b of
signals an estimate of an environment’s average a sigmoidal activation function:
reward. When the average reward is high, it is
rational for an agent to respond more vigorously 1
a¼ , (4)
(since more reward can be collected), whereas 1 þ ebxb
when the average reward is low, the energetic
costs may outweigh the benefits of responding. where x is the synaptic input of the neuron, b is
This model predicts not only a lower rate of the baseline membrane potential, and a is the
responding with depleted tonic DA but also firing rate. This basic schema for dopaminergic
a longer latency to respond, consistent with gain modulation has been reproduced in much
experimental evidence (Denk et al. 2005). greater biological detail by a number of authors
A fourth development has been to integrate the (e.g., Durstewitz et al. 1999; Moyer et al. 2007).
prediction error theory into more biologically An important function of gain modulation is to
detailed models of RL in the basal ganglia. For support active maintenance in working memory:
example, Frank (2005) proposed a model in If a group neurons are recurrently connected to
which DA differentially affects the “direct” and form an attractor network, increasing the gain
“indirect” pathways of the basal ganglia. The will increase the stability of high activity states
direct pathway striatal neurons in the direct path- (i.e., actively maintained memories) while
way primarily express D1 receptors, whereas suppressing low activity states corresponding to
striatal neurons in the indirect pathway primarily spontaneous background activity. Consistent
Computation with Dopaminergic Modulation 671 C
with this idea, DA appears to play an important A number of authors have proposed that many
role in supporting active maintenance of working cognitive deficits in schizophrenia can be
memory representations in the prefrontal cortex explained by a breakdown in the regulation of
(Cohen et al. 2002). gating (Braver and Cohen 1999; Durstewitz and
DA has been hypothesized to play Seamans 2008; Frank 2008; Waltz et al. 2007).
a complementary role in updating working mem- The gain modulation view leads Braver and
ory (Braver and Cohen 1999; Durstewitz Cohen (1999) to propose that reduced prefrontal C
et al. 1999; O’Reilly and Frank 2006). According DA in schizophrenics results in a lower signal-to-
to these models, a phasic burst of DA acts as noise ratio; combined with noisier DA signaling,
a “gating” signal, transiently amplifying afferent this produces gating of irrelevant information
inputs while suppressing local inhibitory signals. into working memory and can cause psychotic
The effect of this gating signal is the encoding of episodes in schizophrenia. Another view, pro-
afferent input into the current attractor state. posed by Frank and colleagues (Frank 2008;
O’Reilly and Frank (2006) have proposed Waltz et al. 2007), is that striatal D1 receptor
a biologically detailed refinement of this idea, function is compromised in schizophrenia,
whereby “stripes” in the prefrontal cortex resulting in impaired learning from positive pre-
(groups of neurons connected to distinct areas of diction errors but spared learning from negative
the basal ganglia). Levitt et al. (1993) encode prediction errors.
representations in working memory that are The model of Frank and colleagues also sheds
updated by phasic DA bursts. This model uses light on the cognitive deficits in Parkinson’s dis-
an RL mechanism to learn when to gate, provid- ease (Frank 2005; Moustafa et al. 2008b; Wiecki
ing a link to the models of corticostriatal plastic- and Frank 2010). Unmedicated patients with
ity described above. Parkinson’s disease appear to be better at learning
How can the active maintenance and from negative outcomes compared to positive
updating roles of DA be reconciled? One outcomes, and this pattern is reversed by DA
suggestion is that these roles are mediated medication (Cools et al. 2006; Frank et al. 2004;
by different receptors. Specifically, Cohen Moustafa et al. 2008a). Frank et al. explained
et al. (2002) have suggested that phasic DA these findings in terms of the effects of DA levels
activity primarily affects updating via D2 recep- on D1 and D2 receptors in the striatum: Depleted
tors, whereas tonic DA activity primarily affects DA in unmedicated patients results in D2 domi-
active maintenance via D1 receptors. These two nance (as in schizophrenia), whereas elevated
modes interact, with tonic DA inhibiting phasic DA in medicated patients results in D1 domi-
DA by regulating presynaptic autoreceptors nance. This account may explain the nature of
(Grace 1991). This hypothesis is consistent cognitive deficits following DA medication, such
with the observation that prolonged pharmaco- as failures to avoid non-rewarding or punishing
logical elevation of tonic DA results in persev- stimuli (Cools et al. 2006; Frank et al. 2004;
erative and stereotyped behavior (Arnsten 1998; Moustafa et al. 2008a) and susceptibility to path-
Sawaguchi and Goldman-Rakic 1991), which ological gambling and addiction (Dagher and
could arise from a failure to update due to Robbins 2009).
reduced phasic DA. In addition to these cognitive deficits,
Parkinson’s disease is associated with
Computational Modeling of Dopamine- bradykinesia, an overall slowing of movements,
Associated Disorders and this deficit is relieved by DA medication.
The models reviewed above furnish a set of However, Parkinson’s patients can learn to
hypotheses about the etiology of psychiatric dis- make movements at a task-specified velocity
orders arising from DA dysfunction. This section and with the same level of accuracy as controls;
focuses on the two most well-studied examples: the key difference is that Parkinson’s patients
schizophrenia and Parkinson’s disease. take longer to reach the performance criterion
C 672 Computation with Dopaminergic Modulation
(Mazzoni et al. 2007). The tonic DA model of decisions about delay oreffort. Psychopharmacology
Niv et al. (2007) offers a perspective on these (Berl) 179:587–596
Durstewitz D, Seamans JK (2008) The dual-state theory of
findings: since tonic DA is reduced in prefrontal cortex dopamine function with relevance to
Parkinson’s disease, the average reward will be COMT genotypes and schizophrenia. Biol Psychiatry
perceived as lower, thus reducing the opportunity 64:739–749
cost of not responding. Thus, bradykinesia may Durstewitz D, Kelc M, G€ unt€urk€
un O (1999)
A neurocomputational theory of the dopaminergic
actually reflect an optimal decision by the motor modulation of working memory functions. J Neurosci
system based on incorrect information about 19:2807–2822
average reward provided by tonic DA (Niv and Foote L, Morrison JH (1987) Extrathalamic modulation of
Rivlin-Etzion 2007). cortical function. Annu Rev Neurosci 10:67–95
Frank MJ (2005) Dynamic dopamine modulation in the
basal ganglia: a neurocomputational account of cogni-
tive deficits in medicated and nonmedicated Parkin-
Cross-References sonism. J Cogn Neurosci 17:51–72
Frank MJ (2008) Schizophrenia: a computational rein-
forcement learning perspective. Schizophr Bull
▶ Computational Models of Neuromodulation 34:1008–1011
▶ Reward-Based Learning, Model-Based and Frank MJ, Seeberger L, O’Reilly RC (2004) By carrot or
Model-Free by stick: cognitive reinforcement learning in Parkin-
sonism. Science 306:1940–1943
Glimcher PW (2011) Understanding dopamine and rein-
forcement learning: the dopamine reward prediction
References error hypothesis. Proc Natl Acad Sci
108:15647–15654
Arnsten AFT (1998) Catecholamine modulation of pre- Grace AA (1991) Phasic versus tonic dopamine release
frontal cortical cognitive function. Trends Cogn Sci and the modulation of dopamine system responsivity:
2:436–447 a hypothesis of the etiology of schizophrenia. Neuro-
Berridge KC, Robinson TE (1998) What is the role of science 41:1–24
dopamine in reward: hedonic impact, reward Hollerman JR, Schultz W (1998) Dopamine neurons
learning, or incentive salience? Brain Res Rev report an error in the temporal prediction of reward
28:309–369 during learning. Nat Neurosci 1:304–309
Braver TS, Cohen JD (1999) Dopamine, cognitive control, Horvitz JC, Stewart T, Jacobs B (1997) Burst activity of
and schizophrenia: the gating model. Prog Brain Res ventral tegmental dopamine neurons is elicited by
121:327–349 sensory stimuli in the awake cat. Brain Res
Clarke CR, Geffen GM, Geffen LB (1987) Catechol- 759:251–258
amines and attention I: animal and clinical studies. Houk JC, Adams JL, Barto AG (1995) A model of how the
Neurosci Biobehav Rev 11:341–352 basal ganglia generate and use neural signals that pre-
Cohen JD, Braver TS, Brown JW (2002) Computational dict reinforcement. In: Houk JC, Davis JL, Beiser DG
perspectives on dopamine function in prefrontal cor- (eds) Models of information processing in the basal
tex. Curr Opin Neurobiol 12:223–229 ganglia. MIT Press, Cambridge, MA, pp 249–270
Cohen JD, McClure SM, Yu AJ (2007) Should I stay or Kakade S, Dayan P (2002) Dopamine: generalization and
should I go? How the human brain manages the trade- bonuses. Neural Netw 15:549–559
off between exploitation and exploration. Philos Trans Lapish CC, Kroener S, Durstewitz D, Lavin A, Seamans
R Soc Lond B Biol Sci 362:933–942 JK (2007) The ability of the mesocortical dopamine
Cools R, Altamirano L, D’Esposito M (2006) Reversal system to operate in distinct temporal modes. Psycho-
learning in Parkinson’s disease depends on medication pharmacology (Berl) 191:609–625
status and outcome valence. Neuropsychologia Levitt JB, Lewis DA, Yoshioka T, Lund JS (1993) Topog-
44:1663–1673 raphy of pyramidal neuron intrinsic connections in
Dagher A, Robbins T (2009) Personality addiction dopa- macaque monkey prefrontal cortex (areas 9 and 46).
mine: Insights from Parkinson’s disease. Neuron J Comp Neurol 338:360–376
61:502–510 Mazzoni P, Hristova A, Krakauer JW (2007) Why don’t
Daw ND, Courville AC, Touretzky DS (2006) Represen- we move faster? Parkinson’s disease, movement vigor,
tation and timing in theories of the dopamine system. and implicit motivation. J Neurosci 27:7105–7116
Neural Comput 18:1637–1677 Montague PR, Dayan P, Sejnowski TJ
Denk F, Walton ME, Jennings KA, Sharp T, Rushworth (1996) A framework for mesencephalic dopamine sys-
MF, Bannerman DM (2005) Differential involvement tems based on predictive Hebbian learning. J Neurosci
of serotonin and dopamine systems in cost–benefit 16(5):1936–1947
Computation with Population Codes 673 C
Morris G, Nevet A, Arkadir D, Vaadia E, Bergman H Waltz JA, Frank MJ, Robinson BM, Gold JM
(2006) Midbrain dopamine neurons encode decisions (2007) Selective reinforcement learning deficits in
for future action. Nature Neuroscience 9:1057–1063 schizophrenia support predictions from computational
Moustafa AA, Cohen MX, Sherman SJ, Frank MJ (2008a) models of striatal-corticaldysfunction. Biol Psychiatry
A role for dopamine in temporal decision making and 62:756–764
reward maximization in Parkinsonism. J Neurosci Wiecki TV, Frank MJ (2010) Neurocomputational models
28:12294–12304 of motor and cognitive deficits in Parkinson’s disease.
Moustafa AA, Sherman SJ, Frank MJ (2008b) Prog Brain Res 183:275–297
A dopaminergic basis for working memory, learning,
C
and attentional shifting in Parkinson’s disease.
Neuropsychologia 46:3144–3156
Moyer JT, Wolf JA, Finkel LH (2007) Effects of dopami-
nergic modulation on the integrative properties of the Computation with Local Field
ventral striatal medium spiny neuron. J Neurophysiol Potentials in Olfaction
98:3731–3748
Niv Y, Rivlin-Etzion M (2007) Parkinson’s disease: fight-
▶ Local Field Potential in Olfaction
ing the will? J Neurosci 27:11777–11779
Niv Y, Daw ND, Joel D, Dayan P (2007) Tonic dopamine:
opportunity costs and the controlof response vigor.
Psychopharmacology (Berl) 191:507–520
O’Reilly RC, Frank MJ (2006) Making working memory
work: a computational model of learning in the pre-
Computation with Population Codes
frontal cortex and basal ganglia. Neural Comput
18:283–328 E. Paxon Frady1 and William B. Kristan2
Roesch MR, Calu DJ, Schoenbaum G (2007) Dopamine 1
Department of Neurosciences, UC San Diego,
neurons encode the better options in rats deciding
La Jolla, CA, USA
between differently delayed or sized rewards. Nature 2
Neuroscience 10:1615:1624 Division of Biology, UC San Diego, La Jolla,
Salamone JD, Wisniecki A, Carlson BB, Correa M (2001) CA, USA
Nucleus accumbens dopamine depletions make ani-
mals highly sensitive to high fixed ratio requirements
but do not impair primary food reinforcement. Neuro-
science 5:863–870 Definition
Sawaguchi T, Goldman-Rakic PS (1991) D1 dopamine
receptors in prefrontal cortex: involvement in working Population coding is a computational theory pos-
memory. Science 251:947–950
tulating that information is represented and
Schultz W (2007) Multiple dopamine functions at differ-
ent time courses. Annu Rev Neurosci 30:259–288 processed by a large number of neurons. In such
Schultz W, Dayan P, Montague PR (1997) A neural sub- a coding scheme, each neuron on its own encodes
strate of prediction and reward. Science only a small amount of the information that is
275:1593–1599
distributed across the population. Population cod-
Servan-Schreiber D, Printz H, Cohen JD
(1990) A network model of catecholamine effects: ing provides robustness, because even if individ-
gain, signal-to-noise ratio, and behavior. Science ual neurons are noisy or fail altogether, the
249:892–895 population code still processes information
Sokolowski JD, Salamone JD (1998) The role of
effectively.
accumbens dopamine in lever pressing and response
allocation: effects of 6-OHDA injected into core and
dorsomedial shell. Pharmacol Biochem Behav
Suri RE, Schultz W (1998) Learning of sequential move-
ments by neural network model with dopamine-like
reinforcement signal. Exp Brain Res 121:350–354 Fundamentals of Computation
Sutton RS, Barto AG (1998) Reinforcement learning: an To understand how a population code computes,
introduction. MIT Press, Cambridge, MA we will first start by describing the basics of
Trantham-Davidson H, Neely LC, Lavin A, Seamans JK
computation by analyzing silicon-based com-
(2004) Mechanisms underlying differential D1 versus
D2 dopamine receptor regulation of inhibition in pre- puters. There are two essential ingredients to
frontal cortex. J Neurosci 24:10652–10659 computation: representation and transformation.
C 674 Computation with Population Codes
Information in computers is represented using have a great deal of noise of many sorts – both
discrete binary digits, called bits. A bit is often from the intrinsic properties of the component
thought of as a “yes” or “no” response or a “1” or neurons and from ongoing neural activity from
“0”. A bit represents two states and is the simplest sources other than the signal of interest. The
form of information. A computer is built by uti- neurons must perform their computations in the
lizing a large number of bits to represent infor- face of many potential failures caused by this
mation. The number of elements that can be noise. The brain appears to solve this noise prob-
represented grows exponentially with the number lem by distributing information across a large
of bits that are used. Eight bits (a “byte”) can population of neurons. Having more neurons
represent up to 256 (28) different values. Large entails robustness to the encoded information,
numbers are represented by more and more bits: because a large number of independent neurons
an integer is typically represented by 32 bits, and represent the information. By spreading the infor-
a “double” is a decimal number that is mation about different variables across many
represented by 64 bits. A computer can represent neurons, very little information is lost if a single
anything by using more and more bits of infor- neuron misfires or even dies.
mation. A 2-h movie, for example, is currently Theoretical work in systems neuroscience has
represented with about 8 billion bits. developed the “population coding” framework of
Computation is in essence the transformation neural computation, based on this premise that
of information from one representation to information is processed across a large popula-
another. For instance, a computer represents – tion of neurons. Population coding computes dif-
or “encodes” – each letter in this text by a byte ferently than a computer, but it still must perform
(8 bits). For you to read this text, the computer the same fundamental operations of information
must transform each letter into an image, which is processing: representation and transformation.
made of pixels that are also represented by bits. The brain receives information from the outside
The computer transforms the byte that represents world and must represent this information in the
each letter into a different sequence of bits that activity of neurons. The brain then transforms
represents the pixel colors. The specific rules of this information in different brain regions to
the transformation are called the “algorithm.” other representations that are useful for percep-
The monitor then “decodes” the bits that repre- tion or behavior. For instance, to grab a baseball,
sent the pixel colors and creates an image. your brain transforms visual information that is
A computer can perform “universal” computa- stored in an eye-centered coordinate system into
tion, which means that it has the ability to trans- a body-centered coordinate system. The same
form any string of bits into any other arbitrary information is stored in both coordinate systems,
string of bits. This ability is what makes but with different representations. The motor cor-
a computer so powerful, because it can transform tex then uses the body-centered coordinate sys-
any type of information in one representation into tem to generate a motor sequence that contracts
any other representation. and relaxes the correct muscles to reach out while
The brain, too, must represent a large number opening the hand and then stop the reaching and
of variables produced by the sensory environ- close the hand as it envelopes the baseball. For
ment or intrinsic states. To perform their func- the brain to perform such a behavior, it must have
tions, neurons in the brain must be able to a representation of all the relevant information, as
encode, transform, and decode this information. well as the ability to transform one representation
However, unlike the bits on a computer, neurons into another.
are intrinsically noisy. The failure of a single bit
in a computer could have catastrophic effects on Representation of Single Variable in
the entire system, but by design there is virtually Population Codes
no noise in a computer, so these types of failures The first step for a population code is to represent
are rare. Biological systems, on the other hand, information through the activity of neurons.
Computation with Population Codes, Fig. 1 The presented, the neurons respond based on their receptive
standard model of population coding. (a) A population fields shown in a. (c) The population vector estimator is
of neurons have receptive fields that are sensitive to dif- a voting method that decodes the orientation from the
ferent orientations. Each bell-shaped curve is from neural activities by fitting a cosine function to the
a different neuron, and the neuron’s preferred orientations responses. (d) The maximum likelihood estimator fits
are spread out across the stimulus space. Y-axis: spikes a template derived from the receptive field shapes (From
per second. (b) When an orientation of y = 90 is Deneve et al. 1999)
A good way to understand how a population code A large number of neurons make up the pop-
can represent information is to decode a single ulation, where each neuron has its own receptive
variable from neural responses. A classic exam- field. The preferred orientations of each neuron
ple of population coding is provided by determin- are spread out across the full stimulus space.
ing how a simple cell in visual cortex responds to Figure 1a shows several example receptive fields
bars oriented at different angles. A given simple from eight different neurons; each bell-shaped
cell responds maximally to a bar that is oriented curve is the receptive field of a different neuron.
at the neuron’s “preferred” angle, but this neuron These receptive fields partially overlap and are
also responds to orientations that are close to this centered at different preferred angles. When
preferred angle. The relationship between the a stimulus is presented, all neurons respond
stimulus and the firing rate of the neuron creates based on their receptive fields. When an oriented
the neuron’s “receptive field,” which peaks at the bar is presented that has an angle of y = 90 , then
preferred angle and decays as the stimulus angle several neurons with preferred orientations close
moves away from the preferred angle. The recep- to 90 will fire (Fig. 1b). The closer the neuron’s
tive field is typically a bell-shaped curve, often preferred orientation is to 90 , the higher the
taken to be Gaussian for nonperiodic variables neuron’s firing rate, but noise can make the firing
and a cosine function for periodic variables rate variable for a given stimulus presentation.
(Deneve et al. 1999). The activity of the population is described by

r j ¼ f j yjYj þ ϵ (1) nonperiodic variables. A standard algorithm for
decoding a nonperiodic variable from a popula-
where rj is the firing rate of neuron j, Yj is the tion code is the OLE:
preferred orientation of neuron j, fj is a function X
that describes the neuron’s receptive field, and e yÔLE ¼ Yj r j (4)
is noise. This is the “standard model” of popula- j
tion coding.
There are various approaches used to come up
with an estimate, y^ , of the orientation of the Center of Mass Estimator (COM)
stimulus encoded by the firing rates of the popu- A simple extension of the OLE, the COM can
lation of neurons. All approaches obtain an esti- account for the background firing (one form of
mate by computing a function from the activity noise) by comparing the neural responses when
across the population R, where R = (r1, r2, . . .). the stimulus is presented to the background rate
Because of noise, the response across the popu- and when the stimulus is absent:
lation varies during different presentations of the X
same identical stimulus. This variability means Yj r j b
y^COM ¼ X
j
^ is a random variable. Differ-
that the estimate, y, (5)
r b
j j
ent techniques for decoding the estimate from the
population activity have been developed, each
with different tradeoffs: in bias (whether the y^ is where b is the spontaneous activity of the neu-
right on average), variance (how close y^ is to y on rons. All three of these estimators are “voting
average), computational efficiency, and biologi- methods,” because each spike from a neuron is
cal plausibility (Pouget et al. 1998). like a vote for that neuron’s preferred stimulus
One interpretation of the population code rep- feature, and the estimate is a simple combination
resentation is that each neuron “votes” for its pre- of all of the votes. Unlike these voting methods,
ferred stimulus. Such schemes are typically other estimators can decode the population activ-
computationally efficient and biologically plausi- ity optimally, which means the variance of the
ble, but have higher variances than other decoding estimate is minimized and there is no bias. These
methods (they are only optimal under specific estimators are less biologically plausible, how-
conditions as discussed below). Three commonly ever, and less computationally efficient than the
used voting schemes are the following. voting methods.
Population Vector Estimator (PVE) Maximum Likelihood Estimator (ML)

Bar orientation is a spatially periodic variable, This estimator finds the value of y that maximizes
and the PVE decodes such periodic variables by the conditional probability of the population
fitting a cosine function (Fig. 1c) to the activity response, P(R|y), and is optimal when y is uni-
patterns, R: formly distributed:
y^PVE ¼ phaseðzÞ (2) y^ML ¼ argmaxy PðRjyÞ (6)
The ML estimator is similar to the PVE in that

where
X it is effectively fitting a template to the population
z¼ r j eiYj (3) response (Paradiso 1988) (Fig. 1d). However, the
j PVE assumes that the shape of the template is
a cosine, whereas the ML derives the template
Optimal Linear Estimator (OLE) directly from the receptive fields of the neurons,
Because the PVE works only for periodic vari- which means that the decoder must have knowl-
ables, other measures must be used for edge of the receptive field shapes of the encoding
neurons. This may be a biological implausibility, activation, which would correspond to a different
but it is also possible that the receptive field population vector in the 10-D space. Together,
shapes are learned over time (Sanger 2003). the responses to all possible orientations would
form a 1-D “manifold” (a curvy line) in the 10-D
Maximum a Posteriori Estimator (MAP) response space.
This estimator is an extension of the ML estima- In addition to representing information about
tor that can incorporate the prior distributions of one stimulus variable, a population of neurons C
R and y. The MAP estimate maximizes the full can encode multiple variables simultaneously.
posterior distribution of the encoded variable, A signal could have several feature dimensions,
P(y|R), which is derived from the conditional all of which are encoded by the brain. For
probability through Bayes’ theorem: instance, a population of neurons could simulta-
neously encode both the orientation and the
y^MAP ¼ argmaxy PðyjRÞ (7) thickness of a bar stimulus. In this population,
each neuron would have a two-parameter recep-
PðRjyÞPðyÞ tive field that has both a preferred orientation and
PðyjRÞ ¼ (8) a preferred thickness:
Pð RÞ

This is a useful modification if certain values r j ¼ f j y, fjYj , Fj þ ϵ (9)
of y are more likely than others, e.g., if horizontal
orientations are seen more frequently than verti- where F is the thickness of the bar and Fj is the
cal orientations. If the prior distribution is thickness preferred by neuron j. In this represen-
flat – meaning that all values of y are equally tation, each combination of orientation and thick-
likely – then the MAP estimate is identical to ness would correspond to a single point in the
the ML estimate. high-dimensional response space (with each
dimension again corresponding to the activity of
one neuron). All possible combinations of bar
High-Dimensional Population Codes
orientation and bar thickness would form a
The activity of a population of neurons can be
2-dimensional manifold (a curvy plane). In one
considered to form a high-dimensional response
direction along this manifold, the information
space: two neurons would form a two-
about orientation is represented, and in another
dimensional space (a plane), three neurons
direction information about thickness is
would form a three-dimensional space, etc. If
represented. In general, the population can
the activity of three neurons is plotted as the x,
encode multiple dimensions of the stimulus,
y, and z coordinates in a three-dimensional
where each neuron can be sensitive to any com-
Cartesian system, then a particular stimulus acti-
bination of stimulus dimensions:
vates these neurons in a particular firing pattern,
which would correspond to a point, called the
“population vector,” in three-dimensional space. r j ¼ f j y, f, c, . . . jYj , Fj , Cj , . . . þ ϵ (10)
Each additional neuron would add a dimension to
this response space. As the firing rate of neuron An individual neuron can have a preference
j increases, the population vector would move up for an arbitrary combination of each of the stim-
in the jth dimension. If we had ten neurons ulus dimensions. The preferred orientation and
representing the 1-dimensional orientation, then thickness of a given neuron can be chosen inde-
each presentation of a particular orientation pendently. For instance, a neuron with a preferred
would activate the ten neurons in a specific way orientation of Yj = 90 could be just as likely to
that would correspond to a population vector in respond to a thick bar as a thin bar.
the 10-dimensional response space. Each orien- The information about a particular feature of
tation would cause a different pattern of the stimulus does not necessarily correspond to
the firing of a particular neuron. In the population the response space must have as many dimen-
code, the neurons represent stimuli through their sions as the number of features. Because we
joint activities. When representing multiple stim- have made the receptive fields perfectly corre-
ulus dimensions, the neurons are activated by all lated, every combination of thickness and orien-
features of the stimulus with different strengths. tation no longer forms a 2-dimensional manifold,
The orientation of a stimulus would thus be only but instead forms a 1-dimensional manifold.
one of many components of the response for Every point in the 1-dimensional manifold repre-
a given neuron. The information about orienta- sents different combinations of the two dimen-
tion is encoded in the combination of activities sions, which means that the information about
across the population. This can be thought of as both features is not present in the population
a particular direction in the high-dimensional activity.
space, but not as the firing of a particular neuron. To form a complete basis of a stimulus space
Just as adding more neurons adds more dimen- with d dimensions, there must be at least
sions to the population code, adding more fea- d neurons with receptive fields that are “linearly
tures to the stimuli would increase the number of independent.” This means that no receptive field
dimensions in the manifold representing those could be made out of a linear combination of the
features. In general, the larger the number of other receptive fields. The number of stimulus
dimensions (i.e., neurons), the larger the number dimensions that a population code could possibly
of features that a particular neuronal population represent is less than or equal to the number of
can represent. This relationship between the neurons. Typically a signal can be fully charac-
number of neurons and the number of features terized by only a few features, and the number of
that can be encoded requires a consideration of neurons responding to a stimulus is usually much
“basis sets.” greater than the number of stimulus dimensions.
If there are d stimulus dimensions, and N neurons
Basis Sets in the population, and N > d, then the population
For the same neurons to encode more than one code is a d-dimensional manifold in an
feature, the activity of the neuronal population N-dimensional space. To form a complete basis
must create a “complete basis set” for these fea- of a d-dimensional stimulus, there must be at least
tures. To understand what such a requirement d neurons with receptive fields that are linearly
means, consider a particular set of neurons with independent. Not all of the neurons need to be
receptive fields that respond to both orientation linearly independent, however; with N neurons,
and thickness, but these receptive fields are actu- some of the receptive fields can be linear combi-
ally perfectly correlated: neurons that prefer 90 nations of others, as long as there is a subset of at
orientations also prefer thick bars, and neurons least d linearly independent receptive fields.
that prefer 0 also prefer thin bars. If the receptive If the stimulus dimensions (d) are low com-
fields are correlated in such a way, then there are pared to the number of neural dimensions (N), the
different combinations of orientations and thick- neurons can form an “overcomplete basis set” of
nesses that activate the population in exactly the the stimulus space. There are at least d neurons
same way. If the population response is the same with linearly independent receptive fields in an
to two different stimuli, then there is less infor- overcomplete basis set, but the other neurons can
mation about the stimulus available in the activ- be linearly dependent. The overcomplete basis
ity. Technically, this inability of the population to gives robustness to the population code because
fully encode the two features arises because we redundant information about the stimulus is
have intentionally made the receptive fields “lin- duplicated across many neurons in the
early dependent.” For the neurons to represent all N-dimensional space. Even if one of the
the dimensions of the stimulus, they must form N neurons completely failed, the N–1 neurons
a complete basis of the stimulus space: the man- would still form a complete basis of the
ifold determined by the neurons’ firing rates in d-dimensional stimulus space, and no
information about the stimulus would be lost. An and motor-commands as well. The nervous sys-
interesting feature of this robustness is that none tem of the leech has provided insights into how
of the N–1 receptive fields needs to be exactly the the same population of neurons can encode dif-
same as the neurons that failed. The information ferent modalities of information. The leech has
carried by the failed neuron is also represented by a very simple nervous system, where the activity
some linear combination of the N–1 neural of a significant fraction of cells (>100) can be
responses. This property means that no two neu- recorded simultaneously with voltage-sensitive C
rons are necessarily “redundant” – every neuron dyes (VSDs). These recordings show that almost
could have a unique response to a stimulus – but every cell in the leech nervous system is
there is redundant information in the population multifunctional: each neuron responds to many
vector. This redundancy helps the neuronal pop- different stimuli and during several different
ulation maintain all of the information even in the behaviors. In effect, each neuron has a high-
presence of large amounts of noise. dimensional receptive field that is sensitive to
The receptive field functions (fj) play a critical many of the sensory, motor, and intrinsic vari-
role in the formation of a basis set. The receptive ables. Neurons in the leech encode the sensory
field shapes considered so far are Gaussian (aka, environment using a population code. The local
bell-shaped curves), because these types of functions bend response is a reflex in which the leech bends
match many biological receptive fields and because its body away from a stimulus. Local bending
these types of functions are capable of forming requires information about the stimulus location
a basis set. There are other types of receptive field to be propagated through the nervous system.
functions (such as sigmoids or threshold-linear) that The neurons involved in this behavior have
can also form basis sets, and it is possible to have overlapping receptive fields, each with preferred
different types of basis functions together in a single touch locations, and encode the information
population code. The key property is that the basis about touch location just like the orientation
set must be able to represent (or closely approxi- information in the standard model (Lewis and
mate) any type of nonlinear function by a linear Kristan 1998; Fig. 1a). Neurons are also encoding
combination of the basis functions: information about the motor patterns. For
instance, virtually all the neurons that respond
X
gð yÞ ¼ wj f j yjYj (11) while the leech swims also respond while the
j leech crawls (Briggman and Kristan 2006;
Fig. 2a). Swimming and crawling are both oscil-
where wj are the coefficients and g(y) is some latory behaviors, and many of the neurons’ volt-
nonlinear function. This requirement of a basis ages oscillate with the muscle contractions at
set is similar to the decomposition of a signal by different phases (Fig. 2b, c). There is no correla-
a Fourier transform, where a linear combination tion between the phase at which a neuron oscil-
of sines and cosines can closely approximate any lates during swimming and the phase at which it
nonlinear function. Some receptive field func- oscillates during crawling. This may be
tions do not have this property. For instance, if a consequence of the population forming
all the receptive fields were linear, then any linear a complete basis of all the motor patterns,
combination of them would also be linear. This because uncorrelated responses can arise from
means that nonlinear functions cannot be fully linearly independent receptive fields.
represented, which means that a population code The swim and crawl motor patterns are
based entirely on linear receptive fields could not represented as different population vectors in
fully represent any signal. the high-dimensional response space of neural
Nervous systems can utilize the population activity (Fig. 2d). The swim motor pattern is off
coding framework beyond representing extrinsic in one direction, and the crawl motor pattern
features of the stimulus. The brain must is off in another. To visualize the population
represent intrinsic information about decisions code recorded through VSD imaging, the
Computation with Population Codes, dimensional neural space to be reduced to 3 dimensions,

Fig. 2 Populations encode multiple modalities of infor- which can be visualized. (f) Trajectories of the neural
mation. (a) Neurons that oscillate with the swimming and responses in PCA space during swimming and crawling
crawling behaviors were observed with voltage-sensitive behaviors (left). The black dots indicate equally spaced
dyes. Virtually all neurons are multifunctional: those intervals in time. These behaviors can be predicted from
involved in swimming were also involved in crawling. the population response using linear discriminant analy-
(b, c) Image of the dorsal side of the leech ganglion sis, which chooses the direction that maximally separates
where neuron somas are shown as circles. Neurons the population vectors during the two behaviors (arrow).
involved in crawling (b) and swimming (c) are colored The decision to swim or crawl can be determined before
based on the relative phase of their oscillations. (d) The the motor patterns start (right): the traces show the firing
swimming and crawling central pattern generators (CPG) of several motor neurons during swimming (blue) or
are encoded as different population vectors in the high- crawling (red). In one outlier, the population vector
dimensional neural space. The population begins at an appeared to head in the swim direction, but changed
initial rest state and then makes an excursion to one of course and went to the crawl pattern (green) (From
the CPGs after a stimulus. (e) The first three principal Briggman et al. 2005; Briggman and Kristan 2006)
components of neural responses. This allows the high-
high-dimensional neural activity was reduced to with equal probability. This decision can be
three dimensions using principal components decoded from the population activity before the
analysis. Each principal component corresponds motor pattern can be seen. The information about
to different combinations of activity patterns the decision is encoded by the same population of
across all the neurons (Fig. 2e). The combinations neurons and can be decoded as a particular direc-
of activity that are most correlated are chosen as tion in the high-dimensional space. By finding the
the first three principal components, and the pop- point in time where the population vectors during
ulation vector can be visualized as a point in swim and crawl separate, the decision direction
three-dimensional space. By plotting the popula- can be decoded as the linear discriminant that
tion vector over time, the two behaviors can be maximally separates the two classes of behavior.
distinguished in the neural activity (Fig. 2f). In The decision direction is actually a different
fact, the neural activity predicts the future behav- direction than the information about either
ioral state before the swim or crawl behavior motor pattern (Briggman et al. 2005). This
actually begins. If the leech is stimulated at shows that the leech nervous system encodes all
a particular site, it “chooses” to swim or crawl of the pertinent information as different
directions along a manifold in high-dimensional population encodes the position of the eyes, ex
space, regardless of whether the information is (Fig. 3b). The eye position neurons were given
sensory, motor, or intrinsic. sigmoidal receptive fields, which were chosen
because of their biological relevance; each of
Transformations with Population Codes these neurons has a different inflection point
Each area of the brain represents information (similar to a preferred stimulus) in their receptive
using its own population code by combining fields, and they too form a population code. C
information from the population codes of its To combine the information from the retina
inputs. Decoding a population code with an esti- (rx) and the eye position (ex), an intermediate
mator is a strategy for an experimentalist to get an (“hidden”) layer was added (Fig. 3c) that forms
understanding of how information is represented a population code to represent the combined
by neuronal activity. In the brain, however, the information about retinal activity and eye posi-
stimulus dimensions are never directly extracted tion. The key to the transformation was to make
from a population code of neural activity. the population code of the hidden layer have
Instead, one population code is mapped directly a complete basis set of the information from
onto another population code. This mapping is both input populations. This projected the two
still effectively an estimation problem, because dimensions of the stimulus (retinal position and
the direct mapping is equivalent to decoding the eye position Fig. 3a, b) onto a 2-dimensional
relevant information from the inputs and then manifold in high-dimensional space (Fig. 3c).
encoding them into the local code. This transformation allowed the relevant dimen-
Mapping one population code onto another sion of head-centered coordinates (Fig. 3d) to be
requires nonlinear transformations. An excellent a simple linear combination of the population
example of this problem is the transformation of code formed in the hidden layer, which could
a retinotopic map onto a head-centered map easily be read out by the next neuronal popula-
(Pouget and Sejnowski 1997; Fig. 3). As the tion. The advantage of using a population code
eyes move around, neurons that receive informa- with a complete basis set is that any reference
tion from the retina have receptive fields that frame is just a linear combination away.
move with the eyes. These neurons encode the In fact, information about eye position, head
visual scene in retinotopic coordinates. Another position, body position, joint position, etc. can all
population of neurons encodes the position of the be combined into a single population code. With
eyes. With the information from these two such a code, different motor systems can read out
populations, the brain can encode the visual the information about visual space with different
input in reference to the head position, which reference frames all from the same population
would be useful if an object was moving towards (Pouget and Sejnowski 1997) with just a simple
your head and you wanted to move your head out linear transformation. Further, the information
of the way. The goal is to transform the popula- about each reference frame can be decoded by
tion code in the retinotopic coordinates into these motor systems in parallel, because the pop-
a population code in head-centered coordinates ulation code is representing all of the necessary
by combining the retinotopic information with information simultaneously.
the eye position information. For simplification, The key to forming this basis set was to find
we will consider a model of only the horizontal how the activity in the two population codes
spatial dimension of the stimulus, x. In this should be combined. It is standard procedure
model, each neuron in the retinotopic population that synaptic connections can compute X linear
has a receptive field that is based on where the transformations of the inputs ( r j ¼ wij r i ),
stimulus falls on the retina. For these receptive i.e., the synaptic strengths multiply the iinput fir-
fields, the relevant stimulus dimension is the ing rates to produce the downstream firing rates.
position on the retina, rx, and each neuron has But this modification alone was not sufficient to
a preferred retinal position (Fig. 3a). A second form a complete basis set of the data; gain control
Computation with Population Codes, information. (d) The head-centered reference frame can
Fig. 3 Transformation of population codes. (a) be decoded from the hidden layer with a linear transfor-
A population of neurons form a retinotopic map of the mation. (e, f) The receptive field of a neuron in the hidden
visual world, with individual neurons having receptive layer (same as panel C) has a receptive field that is gain
fields based on where the stimulus falls on the retina (rx). modulated by the eye position units. This allows the hid-
(b) Another population of neurons encode the position of den layer to form a complete basis set of the inputs. (e)
the eyes (ex). (c) The hidden layer combines information Full 2-D receptive field. (f) Slices of the receptive field
from the retinotopic map and eye position units and forms shown in black from (e) (From Pouget and Sejnowski
a population code that is a complete basis set of the input 1997)
was also needed. Gain control is a scaling of the shunting inhibition could be the mechanism of
postsynaptic firing rate and is seen in many dif- such multiplicative transformations (Blomfield
ferent biological neural circuits. Gain control is 1974; Vu and Krasne 1992). Shunting inhibition
a nonlinear operation, which meets the necessary alters the resistance of the neuron by opening
conditions for forming a basis set (other nonline- inhibitory channels with a reversal potential that
arities are also possible, but gain control is the is the same as the resting potential. Changing the
most biologically plausible (Pouget and resistance of the neuron scales the voltage (and
Sejnowski 1997)). Figure 3e, f shows how neu- thus the firing rate) through Ohm’s law (V = IR).
rons in the hidden layer are gain modulated by the More recent studies have suggested that this mech-
eye position, where the receptive field shape is anism is not sufficient when spiking dynamics are
maintained, but the amplitude is altered. These considered and that shunting inhibition acts addi-
neurons are inspired by the activity of neurons in tively (Holt and Koch 1997). Many alternatives use
parietal cortex. neural circuitry to implement gain control
The mechanism of gain control in neurons is (Baca et al. 2008; Salinas and Abbott 1996). How-
still controversial. Early studies postulated that ever, shunting inhibition is receiving renewed
interest as the mechanism for gain control, as weights for each component are spread out at
models show that adding noise (Doiron et al. every angle, mainly along the edge of the unit
2000; Chance et al. 2002) or using multi- circle. Figure 4c shows the shapes of all of the
compartmental neurons (Capaday and van filters sorted based on the angle, which shows
Vreeswijk 2006) can implement multiplicative how the filters for all the neurons vary smoothly
operations in spiking neuron models. as combinations of the two components. These
neurons represent temporal variations of the odor C
Population Coding in Time signal as a two-dimensional manifold in the
Many signals that the brain processes change in high-dimensional neural space. The neurons’
time. This temporal information can also be receptive fields are spread out almost uniformly,
encoded by population codes. To encode temporal which creates an overcomplete basis set that is
information, the receptive fields must also vary in minimally redundant and maximally robust.
time. An illustration of how temporal population These types of receptive fields allow the neurons
coding is implemented by the nervous system to encode a representation of some temporal his-
comes from studies of olfactory processing in the tory of the signal in the population code.
insect antennal lobe. While randomly puffing This encoding of temporal history in the pop-
pulses of odor, several neurons were recorded as ulation response can be seen in a simple model of
they responded to the fluctuating odor stimulus. these neurons (Fig. 4d). In the model, a population
Neurons respond with different activity patterns of neurons are randomly given one of the eight
during such an odor stimulus; some neurons receptive fields made out of different combina-
respond immediately when the odor turns on and tions of the two main components (Fig. 4e).
then go silent, some neurons stay on while the odor When the population is given a step odor stimu-
is present, and still other neurons respond only lus, the neurons respond differently based on
when the odor turns off. The firing rates of these their receptive fields. Just like the responses
neurons were approximated by a “linear-nonlinear seen in the antennal lobe, the model neurons go
model” and closely matched the recorded rates. through different fluctuations when the stimulus
This model found a “temporal receptive field” that comes on and when it goes off (Fig. 4f). The
describes a preferred temporal fluctuation in the stimulus information is encoded as a point in the
odors where the neuron fires maximally. Just like high-dimensional activity space, which in this
receptive fields considered before, temporal fluc- case moves in time (Fig. 4g). Before the onset
tuations in the stimulus that are close to the pre- of the odor step, the population is at some equi-
ferred temporal fluctuation also cause the neuron librium (denoted as B). After the step, the popu-
to fire. The temporal receptive field integrates the lation vector makes a rapid excursion from
fluctuation pattern of the odor stimulus over time baseline and eventually reaches a fixed point
(this is the linear part of the model), and this sum is (FP), which indicates that the firing rates have
then passed through a nonlinear function (which is reached a steady state. When the odor is turned
a threshold-linear filter). off, the population vector leaves the fixed point,
The temporal receptive fields found for all of makes another excursion, and settles back to the
the neurons can be well described by only two original baseline. If a different odor is given, then
main components (Fig. 4a). These components the neurons progress through a similar pattern,
are determined by normalizing the responses and but along a completely different trajectory.
calculating the principal components across time Each point along these trajectories encodes the
for all of the temporal responses of the neurons. recent history of the odor fluctuations. Since the
Each neuron’s receptive field is effectively odor filters are responsive only to a window of
explained as a linear combination (i.e., a a few seconds, the information in the population
weighted sum) of just two components. Figure 4b code contains only a few seconds of history. At the
shows the weights of the two components that best fixed point after the odor turns on, the population
match the derived filters from the data, where the vector represents the presence of a particular odor
Computation with Population Codes, filter. (e) The receptive fields were chosen based on eight
Fig. 4 Population coding in time. (a) The first two prin- different combinations of the principal components
cipal components (PC1, PC2) of the temporal receptive derived in (a). (f) Responses of four antennal lobe projec-
fields of many neurons in the insect antennal lobe. (b) The tion neurons (PN1-4) to a step of odorant (top, gray-
coefficients of the first two components for each receptive shaded area), compared with the responses of the model
field. The receptive fields are spread out uniformly around (bottom) that shows similar patterns of activation. (g) The
the unit circle as different combinations of the two main population vector trajectories in PCA space of the
components. The colors are based on the angle of the recorded neurons from f (top) when two different odors
coefficients around the circle (F). (c) All normalized were presented. The population vector encodes the odor-
receptive fields plotted and sorted vertically based on the ant presence and recent fluctuation history. The model
angle from (b). (d, e) A simple model of the insect anten- neurons show a similar pattern (bottom). B indicates the
nal lobe population code. (d) The signal (s(t)) is encoded baseline when no odor is present; FP indicates fixed points
by the responses of a population of neurons (r(n)(t)) with of the population vector when different odors are present
different temporal receptive fields and a threshold-linear (From Geffen et al. 2009)
and the fact that the odor has not changed in the last system and a necessary foundation for generating
few seconds. During the excursion, the population the wide variety of perceptual, cognitive, and
vector continuously encodes the odor presence and behavioral states seen in the brain.
the sudden change in odor that occurred when the
odor step turned on. Different odor fluctuations
would move the population vector along different References
trajectories in the high-dimensional space, and the C
recent history of the stimulus fluctuations can be Baca SM, Marin-Burgin A, Wagenaar DA, Kristan WB Jr
(2008) Widespread inhibition proportional to excita-
decoded by the immediate position of the popula-
tion controls the gain of a leech behavioral circuit.
tion vector. Neuron 57:276–289
Alterations in the neural responses over longer Blomfield S (1974) Arithmetical operations performed by
timescales would be required to encode a longer nerve cells. Brain Res 69:115–124
Briggman KL, Kristan WB Jr (2006) Imaging dedicated
history of temporal information. Other types of
and multifunctional neural circuits generating distinct
temporal fluctuations seen in nervous systems behaviors. J Neurosci 26(42):10925–10933
may be serving this purpose. A full range of dif- Briggman KL, Abarbanel HDI, Kristan WB Jr
ferent time scales could be implemented by vari- (2005) Optical imaging of neuronal populations during
decision-making. Science 307:896–901
ous neural mechanisms. For instance, adaptation
Capaday C, Van Vreeswijk C (2006) Direct control of
or neuromodulators can encode information on the firing rate gain by dendritic shunting inhibition. J Int
order of seconds, short-term plasticity could be Neurosci 5(2):199–222
involved in representing information on the order Chance FS, Abbott LF, Reyes AD (2002) Gain modulation
from background synaptic input. Neuron 35:773–782
of tens of seconds, and long-term plasticity could
Deneve S, Latham PE, Pouget A (1999) Reading popula-
encode even larger temporal scales. tion codes: a neural implementation of ideal observers.
Nat Neurosci 2(8):740–745
Conclusion Doiron B, Longtin A, Berman N, Maler L (2000) Subtrac-
tive and divisive inhibition: effect of voltage-
Population coding is a powerful framework for
dependent inhibitory conductances and noise. Neural
neural computation. By spreading information Comput 13:227–248
across a large population of neurons, the brain Geffen MN, Broome BM, Laurent G, Meister M (2009)
solves many of the problems posed by biology. Neural encoding of rapidly fluctuating odors. Neuron
61:570–586
Noise is mitigated by coding information in the
Holt GR, Koch C (1997) Shunting inhibition does not have
joint activities of many neurons, instead of indi- a divisive effect on firing rates. Neural Comput
vidual neurons that can easily fail. Further robust- 9:1001–1013
ness is achieved by coding with an overcomplete Lewis JE, Kristan WB Jr (1998) A neuronal network for
computing population vectors in the leech. Nature
basis set, which contains redundant information
391:76–79
that can compensate for information lost due to Paradiso MA (1988) A theory of the use of visual orien-
noise. An overcomplete basis set enables tation information which exploits the columnar struc-
a population of neurons to represent all of the ture of striate cortex. Biol. Cybern. 58:35–49
Pouget A, Sejnowski TJ (1997) Spatial transformations in
features of the stimulus, the intrinsic states,
the parietal cortex using basis functions. J Cogn
and/or motor activity as well as the recent history Neurosci 9(2):222–237
of these states simultaneously. This facilitates Pouget A, Zhang K, Deneve S, Latham PE (1998) Statisti-
parallel processing from downstream systems, cally efficient estimation using population coding.
Neural Comput 10:373–401
since they can decode different reference frames
Salinas E, Abbott LF (1996) A model of multiplicative
from the same population. Population codes can neural responses in parietal cortex. Proc Natl Acad Sci
be arbitrarily transformed into other representa- 93:11956–11961
tions and combined with information from other Sanger TD (2003) Neural population codes. Curr Opin
population codes in a biologically plausible man-
Vu ET, Krasne FB (1992) Evidence for a computational
ner. Population coding is a promising theoretical distinction between proximal and distal neuronal inhi-
basis for universal computation in the nervous bition. Science 255(5052):1710–1712
C 686 Computation with Serotonergic Modulation
exert opposing and/or complementary effects on

Computation with Serotonergic neuronal physiology and act on the order of mil-
Modulation liseconds to minutes (Bockaert et al. 2006).
Lastly, in recent years, there has been a growing
Matthew Lewis understanding of the functional heterogeneity of
Cornell University, Ithaca, NY, USA serotonergic neurons in central nervous system,
suggesting the existence of specific serotonergic
subsystems, which modulate specific neural
Synonyms structures and behavior in a context- and brain
state-dependent manner (Hale and Lowry 2011).
5-hydroxytryptamine (5-HT) Given the biochemical and behavioral complex-
ity of the serotonergic system, it is difficult to
discern an underlying principle that governs
Definition its role, as has been attempted for other
neuromodulators (Doya 2002; Montague
Serotonin is an aminergic neuromodulator found et al. 2004; Ashton-Jones and Cohen 2005;
throughout nervous systems of invertebrate and Bouret and Sara 2005; Hasselmo and Sarter
vertebrate animals. Serotonergic modulation of 2011).
neural circuits plays an important role in Early neurophysiology experiments
a multitude of neural functions ranging from performed in invertebrate preparations have
motor control, memory systems, decision- shown that 5-HT modulates the intrinsic mem-
making, and sensory processing. brane properties of neurons, altering action
potential waveforms and membrane conduc-
tances, through the activity of 5-HTR-dependent
Detailed Description signaling pathways (Kleinhaus and Angstadt
1995; Katz 1998). Through these same pathways,
Introduction 5-HT has been shown to play a role in the mod-
Serotonin (5-HT) is an ancient neuromodulator ulation of synapses, where it is capable to aug-
fundamentally tied to the evolution, development or attenuate the strength of synapses within
ment, and optimal function of the nervous system a neural circuit. While 5-HT alters the properties
of animals. 5-HT and serotonergic receptors of neurons and synapses within particular neural
(5-HTRs) are found throughout the body and circuits, it also plays a role in the interactions
influence almost every aspect of animal physiol- between distinct neural structures, modulating
ogy, ranging from autonomic functions such as the relative strengths of opposing circuits, lead-
cardiovascular/smooth muscle tone and thermo- ing to changes in brain state and production of
regulation to higher-level processes such as sen- adaptive behaviors (Brezina 2010; Dayan and
sory processing, cognitive control, memory, and Huys 2009). Most recently, computational neu-
emotion. The role of 5-HT is complex and multi- roscientists, influenced by computational theories
faceted, in part due to a variety of properties of of learning and decision-making, have argued
the serotonergic system. In both invertebrate and that given the widespread presence of 5-HT in
vertebrate species, serotonergic fibers project the vertebrate brain, 5-HT might act as a global
throughout the nervous system releasing 5-HT signal, reporting behaviorally relevant state
at synapses and non-synaptically through a pro- information (i.e., reward, hunger, arousal, pun-
cess known as volume transmission, a common ishment) to the rest of the brain, facilitating the
feature of neuromodulatory transmission (Zoli formation of both aversive and appetitive cogni-
et al. 1999; Dayan and Huys 2009). There are tive states. Finally while many scientists have
many different types of 5-HTRs (at least 17 dif- proposed many theories for a single function of
ferent receptors within humans), many of which 5-HT (Soubrie 1986; Deakin and Graeff 1991;
Computation with Serotonergic Modulation 687 C
Computation with Serotonergic Modulation, a major axon tract of the mammalian brain. Serotonergic
Fig. 1 Schematic of the mammalian 5-HT system. neurons are found throughout the forebrain where they
Serotonergic neurons (shown in red) are restricted to the modulate sensory processing, motor cortex, and cognitive
midbrain and hindbrain, where they are divided into two function. A caudal-projecting group projects to hindbrain
distinct populations. A rostral-projecting group projects to and spinal cord structures where it modulates transmission
the forebrain and midbrain structures (caudal linear of sensory signaling, locomotor central pattern generators
nucleus, CLi; dorsal raphe nucleus, DRN; median raphe (CPGs), and autonomic functions (raphe magnus, RM;
nucleus, MRN). The majority of serotonergic axonal fibers raphe pallidus, RP; raphe obscurus, RO)
travel to the forebrain through the forebrain bundle,
Jacobs and Fornal 1999; Tecott 2007; Cools altering both intrinsic membrane properties and
et al. 2007; Dayan and Huys 2009), it is important synaptic interactions among neurons within the
to note that 5-HT does not act in a CPG. Modulatory inputs play an essential role in
“neurochemical vacuum” and there are many normal network function. In many experimental
important interactions between the serotonergic systems, the removal of neuromodulatory inputs
and other neuromodulatory systems both at the through a variety of methods can significantly
cellular and network level (Katz and Harris- impair or completely abolish rhythmic motor
Warrick 1990; Daw et al. 2002; Marder and activity (Harris-Warrick 2011). In vertebrates,
Thirumalai 2002). I will here briefly review the 5-HT plays an important role in the generation
main functions ascribed to 5-HT modulation in of rhythmic locomotion behaviors (Fig. 1). In the
(1) rhythmic motor networks and (2) modulation neonatal rodent spinal, bath application of 5-HT
of sensory systems. Note that this is by no means is reliable and commonly used method to induce
a complete or exhaustive review of 5-HT function in vitro fictive locomotion. Several 5-HTRs have
in the central nervous system, but rather an been identified in the spinal cord. Pharmacologi-
overview of the best described roles of 5-HT cal blockade 5-HT2 or 5-HT7 receptors prevent
modulation. the initiation of fictive locomotion following
brainstem stimulation. Conversely pharmacolog-
5-HT and Rhythmic Motor Networks ical activation of 5-HT2 and 5-HT7 receptors
Serotonin plays a prominent role in the initiation induces rhythmic outputs within the spinal cord
and coordination of rhythmic motor behaviors in that resembles electrically driven locomotion.
both invertebrate and vertebrate nervous systems Within spinal CPG neurons, 5-HT alters intrinsic
(Jacobs and Fornal 1993; Katz 1998; Jordan and membrane properties through the modulation
Slawinska 2011). Rhythmic motor behaviors of particular membrane currents. 5-HT lowers
(i.e., swimming, breathing, walking) are driven the voltage threshold of activation of the
by the activity of biological neural networks hyperpolarization-activated cation current (Ih)
known as central pattern generators (CPGs). and persistent sodium current (INaP), inducing
These networks generate stable, repetitive motor a depolarization from rest (Jordan and Slawinska
outputs even in the absence of sensory feedback 2011). Computational modeling of the spinal
or extrinsic cues. Neuromodulators alter CPGs by CPG suggests that the spinal locomotion is
C 688 Computation with Serotonergic Modulation
organized by the activity of two interconnected (Jacobs and Fornal 1999). DRN neurons also
levels of a half-center oscillator, a rhythm- respond transiently to particular types of sensory
generating layer and a pattern-forming layer stimuli (Heym et al. 1982; Ranade and Mainen
(McCrea and Rybak 2008). 5-HT is thought to 2009). Lastly, recent work has suggested that
modulate excitatory and inhibitory neurons within DRN neurons may be transmitting signals related
these two functional layers, providing tonic excit- to features of behavioral paradigms such as reward,
atory drive and important flexibility to perform aversion, and executive control (Miyazaki
coordinated rhythmic movements (Harris-Warrick et al. 2012; Nakamura 2013).
2011; Jordan and Slawinska 2011). 5-HT has been shown to modulate sensory
responses across a number of neural structures
Serotonergic Modulation of Sensory Systems across species and sensory modalities (Andersen
Just as in the motor networks mentioned above, and Dafny 1982; Kloppenburg and Erber 1995;
neuromodulators also continuously modify sen- Mooney et al. 1996; Hurley and Pollak 1999;
sory circuits to appropriately guide behaviors Xiang and Prince 2003; Kloppenburg and Mercer
(Birmingham and Tuauck 2003; Hurley 2008; Dacks et al. 2009). In many regions of the
et al. 2004; Devore and Linster 2012). brain, patterns of neuromodulatory input are
Neuromodulatory neurons effecting sensory highly heterogeneous and differentially target
processing may either be intrinsic to an affected particular neuron types within specific sensory
sensory circuit or located in distal brain regions circuits (McLean and Shipley 1987; Klepper
providing centrifugal inputs to many brain and Herbert 1991; Mooney et al. 1996; Kang
regions (sensory or non-sensory) concurrently. et al. 2001; Paspalas and Papadopoulos 2001).
Rather than transmitting information about ongo- Through complex innervation patterns and the
ing stimuli, neuromodulatory inputs instead are multitude of 5-HTRs expressed within sensory
thought to dynamically filter neural representa- circuits, 5-HT induces a variety of effects on
tions in a manner related to environmental events sensory representations from gross attenuation
and/or the brain’s internal state (Hurley (Petzold et al. 2009) or enhancement (Hurley
et al. 2004; Devore and Linster 2012). 5-HT is and Pollak 2001; Deemyad et al. 2013) to alter-
an important mediator of these functional ation of receptive field structure (Hurley and
changes. Pollak 1999; Dacks et al 2009) and changes in
In vertebrates, the majority of sensory circuits neural sensitivity to environmental cues
receive centrifugal serotonergic inputs from the (Kloppenburg and Mercer 2008; Yokogawa
median raphé (MRN) and dorsal raphé nuclei et al. 2012).
(DRN) located along the midline of the brainstem.
Interestingly, DRN neurons receive inputs from
many sensory processing regions, including the References
retina and vestibular nuclei (Kawano et al. 1996;
Andersen E, Dafny N (1982) Dorsal raphe nucleus mod-
Ye and Kim 2001). It should be noted that the
ulates sensory evoked responses in caudate and sep-
majority of DRN inputs are not from sensory tum. Int J Neurosci 17:151–155
structures but instead from other raphé nuclei, Ashton-Jones G, Cohen JD (2005) An integrative theory
midbrain, and cortical structures (Hornung 2010). of locus coeruleus-norepinephrine function: adaptive
gain and optimal performance. Annu Rev Neurosci
Tonic DRN activity is tied to the phase of the
28:403–450
sleep-wake cycle, with activity highest during Birmingham JT, Tauck DL (2003) Neuromodulation in
the waking state. Compared to alert waking invertebrate sensory systems: from biophysics to
states, DRN neurons show decreased tonic activity behavior. J Exp Biol 206:3541–3546
Bockaert J, Claeysen S, Becamel C, Dumuis A, Marin P
during quiet waking states, even less activity (2006) Neuronal 5-HT metabotropic receptors: fine-
during slow-wave sleep, and are completely tuning their structure, signaling, and roles in synaptic
silent during rapid-eye movement (REM) sleep modulation. Cell Tissue Res 326:553–572
Computation with Serotonergic Modulation 689 C
Bouret S, Sara SJ (2005) Network reset: a simplified over- Jacobs BL, Fornal CA (1993) 5-HT and motor control:
arching theory of locus coeruleus noradrenaline func- a hypothesis. Trends Neurosci 16:346–352
tion. Trends Neurosci 28:574–582 Jacobs BL, Fornal CA (1999) Activity of serotonergic
Brezina V (2010) Beyond the wiring diagram: signalling neurons in behaving animals. Neuropsychophar-
through complex neuromodulator networks. Philos macology 21(2 Suppl):9S–15S
Trans R Soc Lond B 365:2363–2374 Jordan LM, Slawinska U (2011) Modulation of rhythmic
Cools R, Roberts AC, Robbins TW (2007) Serotoninergic movement: control of coordination. Prog Brain Res
regulation of emotional and behavioural control pro- 188:181–195
cesses. Trends Cogn Sci 12:31–40 Kang T, Lee J, Choi K, Park S, Jeong Y, Jo S, Won
C
Dacks AM, Green DS, Root CM, Nighorn AJ, Wang JW M (2001) Distribution of serotonin immunoreactivity
(2009) Serotonin modulates olfactory processing in in the main olfactory bulb of the Mongolian gerbil.
the antennal lobe of Drosophila. J Neurogenet Anat Histol Embryol 30:117–120
23:366–377 Katz PS (1998) Neuromodulation intrinsic to the central
Daw ND, Kakade S, Dayan P (2002) Opponent interac- pattern generator for escape swimming in Tritonia.
tions between serotonin and dopamine. Neural Netw Ann NY Acad Sci 860:181–188
15:603–616 Katz PS, Harris-Warrick RM (1990) Actions of identified
Dayan P, Huys Q (2009) Serotonin in affective control. neuromodulatory neurons in simple motor system.
Annu Rev Neurosci 32:95–126 Trends Neurosci 13:367–373
Deakin JF, Graeff FG (1991) 5-HT and mechanisms of Kawano H, Decker K, Reuss S (1996) Is there a direct
defence. J Psychopharmacol 5:305–315 retina-raphe-suprachiasmatic nucleus pathway in the
Deemyad T, Metzen MG, Pan Y, Chacron MJ (2013) rat? Neurosci Lett 212:143–146
Serotonin selectively enhances perception and Kleinhaus AL, Angstadt JD (1995) Diversity and modu-
sensory neural responses to stimuli generated by lation of ionic conductances in leech neurons.
same-sex conspecifics. Proc Natl Acad Sci J Neurobiol 27:419–433
USA 110:19609–19614 Klepper A, Herbert H (1991) Distribution and origin of
Devore S, Linster C (2012) Noradrenergic and cholinergic noradrenergic and serotonergic fibers in the cochlear
modulation of olfactory bulb sensory processing. Front nucleus and inferior colliculus of the rat. Brain Res
Behav Neurosci 6:52 557:190–201
Doya K (2002) Metalearning and neuromodulation. Neu- Kloppenburg P, Erber J (1995) The modulatory effects of
ral Netw 15:495–506 serotonin and octopamine in the visual system of the
Hale MW, Lowry CA (2011) Functional topography of honey bee (Apis mellifera L.) II. Electrophysiological
midbrain and pontine serotonergic systems: implica- analysis of motion-sensitive neurons in the lobula.
tions for synaptic regulation of serotonergic circuits. J Comp Physiol A 176:119–129
Psychopharmacology (Berl) 213:243–264 Kloppenburg P, Mercer AR (2008) Serotonin modulation
Harris-Warrick RM (2011) Neuromodulation and flexibil- of moth central olfactory neurons. Annu Rev Entomol
ity in Central Pattern Generator networks. Curr Opin 53:179–190
Neurbiol 21:685–692 Marder E, Thirumalai V (2002) Cellular, synaptic and
Hasselmo ME, Sarter M (2011) Modes and models of network effects of neuromodulation. Neural Netw
forebrain cholinergic neuromodulation of cognition. 15:479–493
Neuropsychopharmacology 36:52–73 McCrea DA, Rybak IA (2008) Organization of mamma-
Heym J, Steinfels GF, Jacobs BL (1982) Activity of lian locomotor rhythms and pattern generation. Brain
serotonin-containing neurons in the nucleus raphe Res Rev 57:134–146
pallidus of freely moving cats. Brain Res 251:259–276 McLean JH, Shipley MT (1987) Serotonergic afferents to
Hornung J (2010) The neuroanatomy of the serotonergic the rat olfactory bulb: I. Origins and laminar specificity
system. In: M€uller CP, Jacobs BL (eds) Handbook of of serotonergic inputs in the adult rat. J Neurosci
the behavioral neurobiology of serotonin, 1st edn. 7:3016–3028
Elsevier, London, pp 51–64 Miyazaki K, Miyazaki KW, Doya K (2012) The role of
Hurley LM, Pollak CD (1999) Serotonin differentially serotonin in the regulation of patience and impulsivity.
modulates responses to tones and frequency- Mol Neurobiol 45:213–224
modulated sweeps in the inferior colliculus. Montague PR, Hyman SE, Cohen JD (2004) Computa-
J Neurosci 19:8071–8082 tional roles for dopamine in behavioural control.
Hurley LM, Pollak GD (2001) Serotonin effects on fre- Nature 431:760–767
quency tuning of inferior colliculus neurons. Mooney RD, Huang X, Shi MY, Bennett-Clarke CA,
J Neurophysiol 85:828–842 Rhoades RW (1996) Serotonin modulates retinotectal
Hurley LM, Devilbiss DM, Waterhouse BD and corticotectal convergence in the superior
(2004) A matter of focus: monoaminergic modulation colliculus. Prog Brain Res 112:57–59
of stimulus coding in mammalian sensory networks. Nakamura K (2013) The role of the dorsal raphe nucleus in
Curr Opin Neurobiol 14:488–495 reward-seeking behavior. Front Integr Neurosci 7:60
C 690 Computational Analysis of Rodent Spinal CPG
Paspalas C, Papadopoulos G (2001) Serotonergic afferents and the full synaptic architecture is far from being
preferentially innervate distinct subclasses of definitively mapped out. A few mathematical and
peptidergic interneurons in the rat visual cortex.
Brain Res 891:158–167 computational models have been proposed which
Petzold GC, Hagiwara A, Murthy VN (2009) Serotonergic (1) integrate experimental results into a unified,
modulation of odor input to the mammalian olfactory interrogable conceptual framework and (2) enable
bulb. Nat Neurosci 12:784–791 the testing of various hypotheses of CPG organi-
Ranade SP, Mainen ZF (2009) Transient firing of dorsal
raphe neurons encodes diverse and specific sensory, zation in silico. The ultimate goal of computa-
motor, and reward events. J Neurophysiol tional modeling and analysis of the rodent spinal
102:3026–3037 CPG is to link specific network structures and
Soubrié P (1986) Reconciling the role of central serotonin intrinsic neuronal properties to experimentally
neurons in human and animal behavior. Behav Brain
Sci 9:319–364 observed features of locomotor output.
Tecott LH (2007) Serotonin and the orchestration of
energy balance. Cell Metab 6:352–361
Xiang Z, Prince D (2003) Heterogeneous actions of sero- Detailed Description
tonin on interneurons in rat visual cortex.
Ye Y, Kim DO (2001) Connections between the dorsal Background
raphe nucleus and a hindbrain region consisting of the In order for it to walk effectively, a rodent’s fore-
cochlear nucleus and neighboring structures. Acta limb and hind limb muscles must repeatedly flex
Otolaryngol 121:284–288
Yokogawa T, Hannan MC, Burgess HA (2012) The dorsal and contract in the correct sequence and with the
raphe modulates sensory responsiveness during correct timing (onset and duration) of flexion and
arousal in zebrafish. J Neurosci 32:15205–15215 contraction. In rodents and other mammals, the
Zoli M, Jansson A, Sykova E, Agnati LF, Fuxe K (1999) basic patterning of such locomotor activity is
Volume transmission in the CNS and its relevance for
neuropsychopharmacology. Trends Pharmacol Sci generated and maintained by the neuronal net-
20:142–150 works in the spinal cord, which are most likely
organized into one or more central pattern gener-
ators (CPGs).
Early experiments with decerebrate cats
established that higher order input from the
Computational Analysis of Rodent brain is not needed for locomotion: Cats whose
Spinal CPG brainstems have been transected at the level of
the mesencephalic locomotor region (MLR) will
William Erik Sherwood walk with the proper gait and speed when placed
University of Utah, Salt Lake City, UT, USA on a moving treadmill. (If the brainstem is
transected above the MLR, the cats will walk
spontaneously; if the transection is made below
Definition the MLR, the electrical or pharmacological stim-
ulation of the brainstem is required to produce
The production of patterned locomotor output by walking.) Thus, the circuitry which patterns neu-
the rodent neuromuscular system is regulated by romuscular activity for walking, as well as other
the central pattern generator(s) of the spinal cord. basic gates such as hopping and jumping, must be
The rodent spinal CPG is capable of producing located below the brainstem in the spinal cord.
a wide range of gaits, and in particular the default The closely related phenomenon of “fictive
“fundamental locomotor rhythm” of walking. locomotion” has been studied in experiments on
Fictive locomotor experiments and transgenic rodents. In these preparations, the whole spinal
rodent lines have produced a wealth of informa- cord is removed from the animal and then stimu-
tion about the neuronal constituency and synaptic lated either electrically or using neuromodulatory
organization of the rodent spinal CPG, but many agents, e.g., serotonin, NMDA, dopamine, or ace-
neuronal subpopulations have yet to be identified tylcholine. The stimulation evokes neural activity
Computational Analysis of Rodent Spinal CPG 691 C
among the interneurons and motoneurons of the motoneurons (MNs), ipsilateral coordinating
cord, and the observed firing patterns correspond interneurons (ICNs), commissural interneurons
closely to the activity patterns observed in intact, (CINs), and rhythmogenic interneurons (RGNs).
locomoting animals, as explained below. Moto- Motoneurons directly innervate the flexor or
neuron activity is measured at the dorsal roots, extensor muscles on a single side of the animal,
the proximal ends of bundles of motoneuron and their axons form the ventral roots of each
axons that innervate the muscles of the intact spinal segment. Though they act primarily to C
animal, and typically occurs in rhythmic bursts. excite muscular contraction, they may also limit
The motoneurons (and interneurons) active dur- their own firing by exciting Renshaw cells which
ing locomotion are located in the lower thoracic in turn inhibit the motoneurons. The most signif-
and lumbar spinal segments; those active during icant flexor-related motoneuron clusters are
hind limb movement are found in lumbar seg- found in L1 and L2, and the main extensor-
ments L1–L6, with the strongest coordinated related motoneuron clusters are located in L5
activity often measured from L2 and L5. Firing and L6.
of L2 motoneurons corresponds to hind limb The rhythmogenic interneurons appear to be
flexor muscle activity, while firing of L5 moto- a heterogeneous group of interneurons which act
neurons corresponds to hind limb extensor mus- in different ways to support the rhythmic
cle activity. (oscillatory) drive that underlies the locomotor
The phasing of motoneuron bursts – between pattern. Neither the composition nor the location
the L2 and L5 segments and between the left and of the RGN population is known with much
right sides of each segment – takes one of four certainty. Each lumbar segment has some
forms: total (contralateral) intra- and (ipsilateral) rhythmogenic capacity, as does each side of
intersegmental synchrony, (contralateral) intra- a given segment, but some segments exhibit
segmental alternation with (ipsilateral) stronger rhythmogenic activity than others in fic-
intersegmental synchrony, (contralateral) intra- tive locomotion experiments. Some RGNs, i.e.,
segmental synchrony with (ipsilateral) those in the Hb9 and V2a subclasses, may be able
intersegmental alternation, and simultaneous to burst endogenously and may thereby effect
(contralateral) intrasegmental alternation with direct rhythmic excitation of MNs. But quite
(ipsilateral) intersegmental alternation. These likely there are other, as yet unidentified, sub-
modes correspond to tonic-clonic seizing, populations of RGNs, perhaps organized in a
rhythmic left-right alternation, hopping, and fic- half-center oscillator configuration, which con-
tive locomotion (in experimental preparation) tribute to the generation of the rhythmic motor
or walking (in the intact animal). This last pattern.
mode, simultaneous ipsilateral flexor-extensor Experiments using transgenic rodent (mainly
antisynchrony and contralateral flexor-flexor/ mouse) lines have provided the most detailed
extensor-extensor antisynchrony, is also called picture of the intra-segmental connectivity of
the fundamental locomotor rhythm of the spinal the spinal cord. A small number of ICN subpop-
locomotor CPG. Since the rhythmic neural firing ulations have been identified to date, including
patterns evoked in fictive locomotor experiments the Renshaw cells already mentioned, as well as
occur in the absence of sensory feedback, it may Ia, V1, and V2b interneurons. These interneurons
be inferred that the neural programs for basic act via inhibition to enforce alternation between
locomotor gates, and the mechanisms for extensor-associated and flexor-associated neuron
switching between them, are encoded in the cir- groups on a single side of the midline. Some
cuitry of the spinal locomotor CPG. subtypes, in particular V1, help regulate locomo-
tor speed.
Neurobiology of the Rodent Spinal CPG The axons of commissural interneurons cross
The neurons of the rodent spinal CPG may be the midline of the spinal cord and may synapse
divided into four major functional classes: onto neurons in the same spinal segment or
several segments away. CINs mediate all cross- are defined solely in terms of generic oscillator
cord communication and play a central role in cells and their connections (Buono 2001; Buono
coordinating the rhythmic alternation of flexion and Golubitsky 2001; Golubitsky et al. 1999;
and extension between the left and right sides. As Stewart et al. 2003). In the locomotor CPG con-
with the ICNs, our most detailed knowledge of text, they aim to answer the question, given
spinal cross-cord connectivity comes from trans- a particular arrangement of (generic) neurons
genic rodent studies. The main CIN pathways are and synaptic connections, which locomotor
inhibitory: some run from L1, L2, and L3 to L4, gaits are even theoretically possible? Few, if
L5, and L6, and vice versa, while others connect any, assumptions are made about the cells’ intrin-
opposite sides of the same spinal segments. Pre- sic dynamics or the properties of the connections.
sumably all of these connections work to enforce The cell networks’ dynamics are described purely
the contralateral flexor-flexor and extensor- in terms of their constituent cells’ phases, and
extensor antisynchrony needed for walking. model analysis relies on the symmetry proper-
Some experimental work has also indicated that ties of the network. This approach is essentially
there are excitatory CIN connections within and mathematical, rather than computational, and
between segments (e.g., mediated by V0 Dbx1 eschews almost all biological details. It is quite
and V3 interneuron subpopulations), but less is powerful, however, in that it can establish defin-
known about them, and the evidence that they itively the complete set of possible phase rela-
play a significant role in cross-cord coordination tionships which can exist in a network of
is indeterminate. It is known that contralateral oscillators with a given symmetric coupling
MNs are the postsynaptic targets of some CINs, structure. It can also be used to determine
but for many CIN pathways the targets are whether the set of possible outputs of two dif-
not known precisely, and the overall organization ferent networks is functionally identical.
of CIN connections is only incompletely Although real CPGs have many cell types with
understood. complicated cellular dynamics, as well as
a variety of synaptic and electrical couplings,
Computational Models and may not be fully symmetric, often their
Computational models of the rodent spinal CPG structures do appear to conform (at least
have two main goals: (1) to integrate the myriad roughly) to the results of the mathematical the-
experimentally determined neurobiological ory. Especially relevant to the modeling of the
details into a unified, comprehensible framework rodent spinal locomotor CPG are results which
and (2) to help researchers investigate hypotheses establish the minimal network structures
regarding CPG organization under simulated (minimal in the sense of least number of oscil-
experimental conditions. There are also two com- latory cells and connections) which are capable
peting imperatives which inform the construction of producing all known bipedal and quadrupedal
of any computational neuroscience model: com- gaits. In particular, eight coupled oscillatory
prehensive inclusion of all known experimental cells arranged in two rings of four cells (a
data in the pursuit of biological realism and the two-layer network) are needed to produce all
simplification of messy details, imperfect data, primary and secondary gaits in quadrupeds.
and biological complexity in order to improve
model comprehensibility and to make model Rybak et al. Model
analysis tractable. Extant models of the rodent Over the past decade, Rybak and collaborators
spinal CPG resolve the tension between these have developed a large-scale model of the
competing aims at different balance points. mammalian CPG which incorporates a great
deal of experimental data from both cat and
Coupled Cell Networks rodent preparations, achieving remarkable bio-
Most simplified are coupled cell network models, logical realism at the expense of some analytic
originally developed by Golubitsky et al., which tractability (Ivashko et al. 2003; Dietz et al. 2013;
McCrea and Rybak 2007; Rybak et al. 2006a,b). excitatory CINs, and it omits the tonic drive from
The model is discussed in detail in the article the MLR and afferent feedback.
(Two-Level Model of Mammalian Locomotor When properly tuned, the Rybak et al. model
CPG); here we present only a brief summary. successfully reproduces the basic fictive locomo-
The Rybak et al. model comprises a few tor rhythms found in the cat and rodent spinal
100 Hodgkin-Huxley-style model neurons, preparations, with good matches to the oscillation
divided into classes aligned with the rubric of periods and duty cycles recorded in experiments. C
MNs, ICNs, CINs, and RGNs described above. Perturbation experiments with the model show
Where possible, the intrinsic properties of the that it is capable of recovering from phase-
model neurons are tuned to match those of iden- resetting perturbations rapidly and with minimal
tified subpopulations of the mammalian spinal interruption of the locomotor rhythm. Further-
CPG. Experimental data on synaptic connectivity more, the model reproduces the phenomenon of
also inform the model’s synaptic architecture. deletions, spontaneous omissions of a motoneu-
However, some features of the model’s neuronal ron pool’s activity during a phase of the locomo-
and synaptic components are, of necessity, tor rhythm when it is normally active. In most
hypothetical. cases, deletions do not change the overall phase
The chief distinguishing feature of the Rybak of the locomotor rhythm once the deleted moto-
et al. model is its two-layer architecture: there is neuron pool rejoins the rhythm, and they are thus
a clear partitioning of interneuronal subnetworks termed “non-resetting deletions.” A small minor-
into a rhythm-generating layer and a pattern ity of deletions do lead to a shift in the phasing of
formation layer. The rhythm-generating layer the fundamental locomotor rhythm and are called
has a synaptic architecture that supports “resetting deletions.” The dual layer architecture
rhythmogenesis via a standard half-center oscil- of the Rybak et al. model can explain the origin of
lator dynamic, but it also includes model neurons the two types of deletions: failures at the rhythm-
capable of intrinsic bursting. The output of the generating layer produce resetting deletions,
rhythm-generating layer feeds forward (primarily while failures at the pattern-generating layer pro-
excitation, but also with a few inhibitory path- duce non-resetting deletions.
ways) to the pattern formation layer. The pattern
formation layer includes reciprocal inhibitory cir- Sherwood et al. Model
cuits which enhance and modulate the flexor- Intermediate in biological realism and analytic
extensor alternation communicated from the tractability is the model of Sherwood et al. for
rhythm-generating layer. The output of the pat- the rodent spinal hind limb CPG (Sherwood et al.
tern formation layer feeds forward to the moto- 2011). It uses Hodgkin-Huxley-style neurons and
neurons, acting via direct excitation on dynamic synapses (with parameters and proper-
motoneurons as well as via polysynaptic inhibi- ties parameters similar to those of corresponding
tion. There is no direct pathway from the rhythm- cells in the Rybak et al. model), and it is informed
generating layer to the motoneurons. Both the by experimental data on the neurobiology of the
rhythm-generating and pattern-generating layers rodent spinal CPG circa 2008. It was developed,
may receive direct tonic excitation, simulating however, with the mathematical results from
input from the MLR. In its original incarnation, coupled cell network theory in mind, and to this
the model was used to simulate the CPG network end it represents various populations of MNs,
of a single hemicord of the decerebrate cat prep- CINs, and RGNs by single representative model
aration, and it admitted afferent feedback at each neurons (as per the methods of coupled cell net-
neuronal layer. The model was later modified to work theory, the ICNs were functionally sub-
model the rodent fictive locomotor preparations sumed into the synaptic connections between
described above. The modified model includes the other three categories of neuron). Further-
two hemicord CPG units connected at the more, it is organized as a single layer with mul-
rhythm-generating level via both inhibitory and tiple symmetries; all synaptic connections known
Computational Analysis of Rodent Spinal CPG, Fig. 1 Diagram of the full CPG model of Sherwood et al. Flex
flexor, Ext extensor (From Sherwood et al. 2011)
or hypothesized at the time were included. A key antisynchrony (the walking rhythm). To this end,
departure from conventional CPG architectures they explored the space of symmetric synaptic
(i.e., half-center oscillators) is that each RGN weightings in the CPG model extensively and
bursts endogenous, so that rhythmogenesis is an also studied the phasing properties of function-
intrinsic neuronal property, rather than a network ally significant subnetworks (related to quotient
property. In many respects, the Sherwood networks in coupled oscillator theory) of the full
et al. model is a “minimal” biologically realistic CPG model.
model of the rodent spinal CPG (see Fig. 1). Among the findings was that a multitude of
While it was known from theoretical consid- synaptic weightings produced functionally
erations that such an architecture is capable of equivalent CPG output, echoing results from
producing each of the fictive locomotor gaits seen studies of invertebrate CPGs (cf. Prinz et al.
experimentally, Sherwood et al. sought to inves- 2004). Strong inhibitory connections were
tigate how effectively the fundamental locomotor required to produce the fundamental locomotor
rhythm was achieved by the CPG when starting rhythm, while small amounts of interfering cross-
from an unorganized state, how the CPG cord excitation could destabilize it. In all config-
responded to perturbations of the locomotor urations leading to the fundamental locomotor
rhythm, and how the distribution of synaptic rhythm, the model CPG recovered very slowly
weights affected these two aspects of CPG out- from perturbations, much more slowly than could
put. Emphasis was placed on understanding be biologically plausible. Analysis of CPG sub-
the role of CINs in organizing simultaneous networks revealed that the endogenous bursting
ipsilateral flexor-extensor antisynchrony properties of the coupled RGNs dominated the
and contralateral flexor-flexor/extensor-extensor phasing and synchronization properties of the
CPG network. This was a major cause of the slow Stewart I, Golubitsky M, Pivato M (2003) Symmetry
network phase resetting in response to perturba- groupoids and patterns of synchrony in coupled cell
networks. SIAM J Appl Dyn Syst 2(4):609–646
tion. It was surmised that rhythmogenesis by
single endogenously bursting neurons was not
Further Reading
functionally equivalent to – and was in some Dougherty KJ, Kiehn O (2010a) Functional organization
way more fragile than – rhythmogenesis by con- of V2a-related locomotor circuits in the rodent spinal
ventional half-center oscillator networks. Subse- cord. Ann N Y Acad Sci 1198(1):85–93 C
quent investigation of the phase response Dougherty KJ, Kiehn O (2010b) Firing and cellular prop-
erties of V2a interneurons in the rodent spinal cord.
properties of endogenously bursting neurons J Neurosci 30(1):24–37
revealed that such neurons have a much richer Harris-Warrick RM (2011) Neuromodulation and flexibil-
phase response structure than tonically firing neu- ity in central pattern generator networks. Curr Opin
rons and thus more complicated synchronization Neurobiol 21(5):685–692
Kiehn O (2006) Locomotor circuits in the mammalian
behavior in some network architectures, such as spinal cord. Annu Rev Neurosci 29:279–306
that of the Sherwood et al. CPG model (Fig. 1). Kiehn O (2011) Development and functional organization
of spinal locomotor circuits. Curr Opin Neurobiol
21(1):100–109
Kiehn O, Butt SJB (2003) Physiological, anatomical and
References genetic identification of CPG neurons in the develop-
ing mammalian spinal cord. Prog Neurobiol
Buono P-L (2001) Models of central pattern generators for 70:347–361
quadruped locomotion II. Secondary gaits. J Math Biol Kiehn O, Kjaerulff O (1998) Distribution of central pat-
42:327–346 tern generators for rhythmic motor outputs in the spi-
Buono P-L, Golubitsky M (2001) Models of central pat- nal cord of limbed vertebrates. Ann N Y Acad Sci
tern generators for quadruped locomotion I. Primary 860:110–129
gaits. J Math Biol 42:291–326 Kiehn O, Kullander K (2004) Central pattern generators
Dietz SB, Shevtsova NA, Rybak IA, Harris-Warrick RM deciphered by molecular genetics. Neuron 41:317–321
(2013) An asymmetric model of the spinal locomotor Kiehn O, Tresch MC (2002) Gap junctions and motor
central pattern generator: insights from afferent stim- behavior. Trends Neurosci 25(2):108–115
ulations. BMC Neurosci 14(Suppl 1):P161 Kiehn O, Sillar KT, Kjaerulff O, McDearmid JR
Golubitsky M, Stewart I, Buono P-L, Collins JJ (1999) Effects of noradrenaline on locomotor
(1999) Symmetry in locomotor central pattern gener- rhythm-generating networks in the isolated neonatal
ators and animal gaits. Nature 401:693–695 rat spinal cord. J Neurophysiol 82:741–746
Ivashko DG, Prilutsky BI, Markin SN, Chapin JK, Rybak Kiehn O, Kjaerulff O, Tresch MC, Harris-Warrick RM
IA (2003) Modeling the spinal cord neural circuitry (2000) Contributions of intrinsic motor neuron prop-
controlling cat hind limb movement during locomo- erties to the production of rhythmic motor output in the
tion. Neurocomputing 52–54:621–629 mammalian spinal cord. Brain Res Bull 53(5):649–659
McCrea DA, Rybak IA (2007) Modeling the mammalian Kiehn O, Quinlan KA, Restrepo CE, Lundfald L,
locomotor CPG: insights from mistakes and perturba- Borgius L, Talpalar AE, Endo T (2008) Excitatory
tions. Prog Brain Res 165:235–253 components of the mammalian locomotor CPG. Brain
Prinz AA, Bucher D, Marder E (2004) Similar network Res Brain Res Rev 57:56–63
activity from disparate circuit parameters. Nat Lundfald L, Restrepo CE, Butt SJB, Peng C-Y, Droho S,
Neurosci 7(12):1345–1352 Endo T, Zeilhofer HU, Sharma K, Kiehn O (2007)
Rybak IA, Shevtsova NA, Lafreniere-Roula M, McCrea Phenotype of V2-derived interneurons and their rela-
DA (2006a) Modelling spinal circuitry involved tionship to the axon guidance molecule EphA4 in the
in locomotor pattern generation: insights from the developing mouse spinal cord. Eur J Neurosci
effects of afferent stimulation. J Physiol 26:2989–3002
577(2):641–658 Meehan CF, Grondahl L, Nielsen JB, Hultborn H (2012)
Rybak IA, Stecina K, Shevtsova NA, McCrea DA (2006b) Fictive locomotion in the adult decerebrate and spinal
Modelling spinal circuitry involved in locomotor pat- mouse in vivo. J Physiol 590(2):289–300
tern generation: insights from deletions during fictive Nishimaru H, Restrepo CE, Kiehn O (2006) Activity of
locomotion. J Physiol 577(2):617–639 renshaw cells during locomotor-like rhythmic activity
Sherwood WE, Harris-Warrick R, Guckenheimer J (2011) in the isolated spinal cord of neonatal mice. J Neurosci
Synaptic patterning of left-right alternation in 26(20):5320–5328
a computational model of the rodent hindlimb Quinlan KA, Kiehn O (2007) Segmental, synaptic actions
central pattern generator. J Comput Neurosci of commissural interneurons in the mouse spinal cord.
30(2):323–360 J Neurosci 27(24):6521–6530
C 696 Computational Maps
Sherwood WE, Guckenheimer J (2010) Dissecting the (intracortical or epicortical) if the electrical con-
phase response of a model bursting neuron. SIAM tacts of the stimulator are positioned within
J Appl Dyn Syst 9(3):659–768
Zhong G, Shevtsova NA, Rybak IA, Harris-Warrick RM the cortex or on its surface, or noninvasive
(2012) Neuronal activity in the isolated mouse spinal transcranial stimulation if the neurons are stimu-
cord during spontaneous deletions in fictive locomo- lated by electrical currents induced across scalp
tion: insights into locomotor central pattern generator by either external magnetic field (transcranial
organization. J Physiol 590(19):4735–4759
magnetic stimulation, TMS) or electrical current
administered via scalp electrodes (transcranial
direct-current stimulation, tDCS). Apart from
direct brain stimulation, other stimulation forms
Computational Maps and targets may also be used, such as spinal cord
stimulation (e.g., for the treatment of pain), vagus
▶ Cortical Maps, Activity-Dependent nerve stimulation (for the treatment of epilepsy),
Development or acoustic stimulation (for the treatment of tin-
nitus). Novel, model-based approaches, which
use methods from synergetics, nonlinear dynam-
ics, and statistical physics to specifically restore
Computational Model-Based brain function and connectivity, demonstrate
Development of Novel Stimulation how insights into the dynamics of complex sys-
Algorithms tems contribute to the development of novel
therapies.
Oleksandr V. Popovych1 and
Peter Alexander Tass1,2,3
1
Institute of Neuroscience and Medicine – Detailed Description
Neuromodulation (INM-7), Research Center
J€ulich, J€
ulich, Germany In the nervous system synchronization processes
2
Department of Neurosurgery, Stanford play an important role, e.g., in the context of
University, Stanford, CA, USA information processing (Singer 1989) and motor
3
Department of Neuromodulation, University of control (Andres and Gerloff 1999). However,
Cologne, Cologne, Germany pathological, excessive synchronization may
strongly impair brain function and is a hallmark
of several neurological disorders like Parkinson’s
Definition disease (PD), essential tremor, epilepsy, and tin-
nitus (Alberts et al. 1969; Elble and Koller 1990;
In the context of this entry, the model-based Nini et al. 1995; Llinas et al. 1999; Milton and
development of stimulation algorithms desig- Jung 2003; Smirnov et al. 2008; Roberts
nates a process of creation and computational et al. 2010). For example, parkinsonian resting
testing of new techniques for control and modu- tremor appears to be caused by a pacemaker-like
lation of undesirable (pathological) neuronal population of neurons which fires in
dynamics. Abnormal brain activity has been a synchronized and periodic manner (Alberts
observed in several neurological disorders et al. 1969; Smirnov et al. 2008). In contrast,
including Parkinson’s disease, essential tremor, under healthy conditions, these neurons fire in
epilepsy, tinnitus, and others. Brain stimulation is an uncorrelated, i.e., desynchronized manner
used for the therapy of patients suffering, for (Nini et al. 1995). Permanent deep brain stimula-
example, from Parkinson’s disease, epilepsy, or tion (DBS) at high frequencies (>100 Hz) is the
mental disorders. Brain stimulation is called deep standard therapy for medically refractory PD and
brain stimulation (DBS) if structures deeply essential tremor (Benabid et al. 1991, 2002;
inside the brain are targeted, cortical stimulation Blond et al. 1992). High-frequency (HF) DBS
Computational Model-Based Development of Novel Stimulation Algorithms 697 C
has been developed empirically, mainly based on presented in detail: coordinated reset (CR) stim-
experimental results, clinical observations, and ulation (Tass 2003a, b), multisite linear delayed
animal experiments (Bergman et al. 1990). DBS feedback (MLDF) stimulation (Hauptmann et al.
was approved by the Food and Drug Administra- 2005a, b, c, 2007a, b), nonlinear delayed
tion (FDA) as a treatment for essential tremor in feedback (NDF) stimulation (Popovych et al.
1997, for Parkinson’s disease in 2002, and for 2005, 2006a, b, 2008; Popovych and Tass
dystonia in 2003. The treatment of severe neuro- 2010), and proportional-integro-differential feed- C
logical and psychiatric diseases with DBS is back (PIDF) stimulation (Pyragas et al. 2007).
a rapidly growing and promising field. However, These techniques have the common objective of
in spite of many beneficial effects, in some reducing the extent of synchronization of the
patients, DBS may not help or may cause side target population by reestablishing a normal
effects, or the therapeutic effects may disappear desynchronized physiological activity in
over time (Tasker 1998; Volkmann 2004; a formerly highly synchronized population of
Rodriguez-Oroz et al. 2005; Deuschl et al. 2006). neurons. For other stimulation methods, we
With the objective of finding more effective refer to Rosenblum and Pikovsky (2004a),
stimulation techniques with less side effects, a Tukhlina et al. (2007), Kiss et al. (2007), Luo
model-based development of novel stimulation et al. (2009), Danzl et al. (2009), and Nabi and
methods has been initiated (Tass 1999, 2002a, Moehlis (2011).
b, 2003b; Rosenblum and Pikovsky 2004a; CR stimulation, in its original realization, uses
Hauptmann et al. 2005b; Popovych et al. 2005, short electrical pulse trains to subsequently reset
2006a; Tass et al. 2006; Hauptmann et al. 2007c; subpopulations of the neuronal network, which
Tukhlina et al. 2007; Popovych and Tass 2010). induces a desynchronized state (Tass 2003a, b).
In these studies, relevant neuronal target The stimulation is applied through a small num-
populations were modeled mathematically, and ber of stimulation sites which are equally spaced
stimulation techniques have been developed uti- within the neuronal population. Desynchro-
lizing principles from nonlinear dynamics and nization is achieved by utilizing self-organization
statistical physics (Tass 1999). One goal of this principles, in particular, the slaving principle
approach is to control the pathological neural induced by the pathological neuronal interactions
dynamics appropriately in order to achieve (i.e., interactions which have the potential to
a mild and efficient relief of symptoms (Tass induce a pathological synchronization) (Tass
1999). The second, more ambitious goal is to 2003a, b). Several studies addressed the optimal
stimulate in a way that the formerly affected parameter calibration of CR stimulation
neuronal populations unlearn their pathological (Lysyansky et al. 2011a, 2013; L€ucken
connectivity and, hence, their tendency to pro- et al. 2013) and also suggested CR stimulation
duce pathological synchronization (Tass and for activation of inactive or hypoactive neuronal
Majtanik 2006). Put otherwise, the very goal of population by multifrequency and phase-shifted
this approach is to induce long-lasting therapeutic activation of the stimulated neuronal networks
effects which outlast the cessation of stimulation. (Lysyansky et al. 2011b).
To this end, stimulation algorithms have been MLDF (Hauptmann et al. 2005a, b, c, 2007a,
developed and optimized to exploit dynamic b) and NDF (Popovych et al. 2005, 2006a, b,
self-organization principles and plasticity rules 2008; Popovych and Tass 2010) stimulation
(Tass and Majtanik 2006; Tass and Hauptmann use delayed feedback for stabilizing
2006, 2007, 2009; Hauptmann and Tass 2007, a desynchronized state which is intended to be
2009, 2010; Tass and Popovych 2012; Popovych as close to the physiological mode of action as
and Tass 2012). possible. Here, the local field potential (LFP) of
Several novel stimulation techniques have the target population is measured, amplified,
computationally been developed in the past. In delayed, and fed back into the ensemble. The
this entry, four of these control methods will be PIDF (Pyragas et al. 2007) utilizes an
C 698 Computational Model-Based Development of Novel Stimulation Algorithms
instantaneous LFP and is designed for a a synaptic-independent blockade of the intrinsic

particularly difficult situation characterized by voltage-gated currents after its cessation
a separate registration and stimulation setup. (Beurrier et al. 2001, 2002). Another hypothesis
It has been shown experimentally that synap- is that HF DBS appears to block neuronal activity
tic plasticity enhances neuronal synchronization in relevant target areas during stimulation
(Nowotny et al. 2003). From the kindling phe- (Benabid et al. 2002). In single-compartment
nomenon in the context of epilepsy, it is well conductance-based biophysical models of iso-
known that neural networks may learn patholog- lated STN neurons, the HF stimulation may
ical strong interactions (Speckmann and Elger cause a suppression of neuronal activity on an
1991; Morimoto et al. 2004). The novel elementary membrane level, where a neuron’s
desynchronizing stimulation protocols are resting state or low-amplitude subthreshold oscil-
designed to invert this pathological process, so lations can get stabilized (Pyragas et al. 2013).
that the affected neuronal populations unlearn The obtained theoretical results resemble the
their pathological connectivity, and physiologi- clinically observed relations between stimulation
cal neuronal activity is reestablished on a long- amplitude and stimulation frequency required
term basis (Tass and Majtanik 2006; Hauptmann to suppress parkinsonian tremor (Benabid
and Tass 2007, 2009, 2010; Tass and Hauptmann et al. 1991). Several contributing mechanisms
2007; Tass and Popovych 2012; Popovych and resulting in the observed effects of HF DBS
Tass 2012). In a nutshell, the novel stimulation might be membrane inhibition, jamming, excita-
techniques aim at a well-directed employment of tion of excitatory and inhibitory afferents, exci-
fundamental principles of dynamic brain action tation of efferents, and plasticity (Benabid
to induce long-lasting therapeutic effects. et al. 2005). In particular, HF stimulation of
afferent axons projecting to STN can account
Standard High-Frequency Stimulation for a therapeutic effect of HF DBS within STN
For high-frequency (HF) deep brain stimulation (Gradinaru et al. 2009).
(DBS), the golden standard for the therapy To precisely evaluate the contribution of these
of medically refractory movement disorders, different mechanisms, spatially extended multi-
depth electrodes are chronically implanted in compartment neuron models were used to dem-
the thalamic ventralis intermedius nucleus or the onstrate the effects of extracellular stimulation on
subthalamic nucleus (STN) (Benabid et al. 1991, the different structures of the stimulated neuronal
2002), and permanent HF (>100 Hz) periodic population (Grill and McIntyre 2001). Depending
electrical pulse-train stimulation is applied. HF on the stimulation amplitude and the shape of the
DBS has been developed empirically, mainly stimulation pulses, either the cells were activated
based on intraoperative observations, and, as directly or fibers mediating excitatory or strong
yet, the mechanism of HF DBS is not sufficiently inhibitory action were activated (Grill and
understood (McIntyre et al. 2004b). HF DBS McIntyre 2001). Modeling studies indicate that
strongly alters the neuronal firing and mimics already at the level of single neurons, the activa-
the effect of tissue lesioning, e.g., by suppressing tion of a larger number of structures can take
neuronal firing, which, in turn, suppresses the place with different and possibly conflicting
symptoms (Benabid et al. 2002; Filali impacts on the single neuron dynamics (Grill
et al. 2004; McIntyre et al. 2004b; Volkmann and McIntyre 2001).
2004). However, HF DBS is a reversible tech- Experimental and modeling studies also
nique and has a much lower rate of side effects reveal that the globus pallidum interior
than lesioning with thermocoagulation (GPi) – one structure of the basal
(Schuurman et al. 2000). ganglia – might strongly be involved in the mech-
HF DBS seems to induce a regular bursting in anisms of DBS (Hashimoto et al. 2003; Rubin
STN neurons during stimulation and a transient and Terman 2004; Miocinovic et al. 2006). HF
suppression of the neuronal activity mediated by stimulation of the STN regularizes GPi firing
(Hashimoto et al. 2003), and this restores the strong coupling, the population automatically
responsiveness of the thalamus (Rubin and relaxes into the desired desynchronized state.
Terman 2004). Accordingly, one may distinguish The scheme of the stimulation setup is presented
between local and nonlocal effects of HF stimu- in Fig. 1a. Several stimulation sites are placed
lation. Locally, in the vicinity of the stimulation within the target network, and weak resetting
electrode, axons rather than cell bodies (somas) stimulation signals are administered via these
get activated (McIntyre et al. 2004a), while the stimulation sites. In this way the oscillatory pop- C
latter can even be effectively inhibited by ulation is divided into several subpopulations,
HF stimulation (Benabid et al. 2002; Welter where each of them is assigned to the
et al. 2004; Meissner et al. 2005). The corresponding stimulation site and receives the
stimulation-induced axonal activity propagates stimulation signal mostly from that stimulation
antidromically and orthodromically (Hammond site. CR stimulation means that a synchronized
et al. 2008) and can change the firing in the output population of neurons is stimulated with
structures downstream to the neuronal target pop- a sequence of brief resetting stimuli (typically
ulation. The pathological discharge patterns in brief HF stimulus trains; see Fig. 1b) via the
the target population can be replaced by HF spik- different sites. The delay between the subsequent
ing or suppressed depending on whether the resetting stimuli can be chosen as t = T/n with
efferent fibers of the stimulated nucleus are excit- respect to that at the preceding site (Fig. 1a),
atory or inhibitory, respectively (Hashimoto where n is the number of stimulation sites, and
et al. 2003; Anderson et al. 2003; McIntyre T approximates the mean period of the collective
et al. 2004b). In other words, local and nonlocal dynamics of synchronized oscillators (Tass
effects of HF DBS may differ considerably. 2003a, b).
HF DBS, although being clinically beneficial The subsequent reset of the different subpop-
and widely used, still has a number of limitations. ulations induces a so-called cluster state, i.e., the
On the one hand, HF DBS may cause adverse whole population divided into n subpopulations
effects like dysarthria, dysesthesia, cerebellar which differ with respect to their mean phase.
ataxia, and memory decline (Volkmann 2004; This effect is illustrated in Fig. 1d where a snap-
Rodriguez-Oroz et al. 2005; Freund 2005). On shot of the distribution of the phases cj of the
the other hand, HF DBS may be ineffective, or stimulated oscillators is shown. The phases of the
its therapeutic effect may wear off over time population of oscillators stimulated via, e.g., four
(Kumar et al. 2003; Rodriguez-Oroz et al. 2005). sites form four phase clusters equidistantly
For instance, 11–15 % of PD patients have unsat- (or close to that) distributed over the oscillation
isfactory outcomes although their depth electrodes period. To estimate the extent and type of
are properly placed (Limousin et al. 1999). This synchronization of the whole population of
initiated the development of new stimulation algo- N oscillators, the cluster variables
rithms with an extensive use of mathematical
1X
models and methods from nonlinear dynamics N
and statistical physics (Tass 1999). Zm ðtÞ ¼ Rm ðtÞeiCm ðtÞ ¼ eimcj ðtÞ , (1)
N j¼1
Coordinated Reset Stimulation
Coordinated reset (CR) stimulation was designed can be used. Rm(t) and Cm(t) are the
to specifically counteract pathological synchrony corresponding real amplitude and real mean
by desynchronization (Tass 2003a, b). The idea phase, where 0 Rm(t) 1 for all time t (Daido
behind CR stimulation (Tass 2003a, b) is to 1992; Tass 1999). Cluster variables are conve-
robustly shift the stimulated population into nient for characterizing synchronized states of
a dynamical state which is not the desired different types: Perfect in-phase synchronization
desynchronized state, but sufficiently close to it. corresponds to R1 = 1, whereas an incoherent
Close in the sense that due to the pathologically state, with uniformly distributed phases, is
Computational Model-
Based Development of
Novel Stimulation
Algorithms,
Fig. 1 Stimulation setup
of CR stimulation method.
(a) Brief and mild resetting
stimuli are administered at
different sites at subsequent
times and effectively divide
the stimulated population
into several, say four,
subpopulations. (b)
Stimulation signals of CR
stimulation in the form of
short charge-balanced
pulse trains. (c–f)
Snapshots of the phase
distribution density
histograms before, during,
and after (poststimulation
desynchronizing transient)
CR stimulation and during
resynchronization as
indicated in the plots. In
(d), the phases cj of the
stimulation neurons form
four phase clusters
equidistantly (or close to
that) distributed over the
period of oscillations. In
(e), a nearly uniform
distribution of the oscillator
phases is formed during the
poststimulation transient.
Plots (c–f) are adapted from
Lysyansky et al. (2013)
associated with Rm = 0, m = 1,2,3, . . . Small four-cluster state; (e) 0.17 and 0.08,
values of R1 combined with large values of Rm poststimulation desynchronization transient; and
are indicative of an m-cluster state consisting of (f) 0.75 and 0.2, resynchronization.
m distinct and equally spaced clusters, where all From the cluster state, the neurons typically
oscillators within the same cluster have similar relax to a uniformly desynchronized state
phases. In Fig. 1c–d, for instance, the variables R1 (Fig. 1e) before they revert back to the in-phase
and R4 approximately attain the following synchronized state (Fig. 1f), if left unperturbed.
values, respectively: (c) 0.98 and 0.72, synchro- To understand how a stimulus-induced clustering
nization; (d) 0.07 and 0.55, stimulation-induced leads to an effective desynchronization, the
dynamics of the leading modes Z1, Z2, . . ., can be at time t if nfire(t) = 1. Varying the threshold
considered. When the coupling among oscillators parameter 0.99 in a reasonable range does not
becomes sufficiently large, e.g., it exceeds change the results.
a certain critical value, Z1 from Eq. 1 becomes As shown in Fig. 2, stimulation starts when the
an order parameter (Kuramoto 1984), which neurons are synchronized, and the collective fir-
according to the slaving principle (Haken 1983) ing demonstrates high-amplitude rhythmic oscil-
governs the dynamics of the other, stable modes lations (upper-right insert in Figs. 2 and 1c). After C
Zm (m = 2,3, . . .) on the center manifold (Pliss a few periods of stimulation (Fig. 1b), the oscil-
1964): The order parameter Z1 acts on a slow latory population is shifted into a cluster state
timescale, whereas the stable modes Zm act on (bottom-right insert in Figs. 2 and 1d). Then the
a fast timescale and relax to values given by the stimulation is switched off (Fig. 1b) and the
order parameter Z1 (Wunderlin and Haken 1975; ensemble returns to a synchronized state on this
Haken 1983). In a system with large number of way passing through a uniformly desynchronized
oscillators, this relationship reads (Tass 1999): state (bottom-left insert in Figs. 2 and 1e, f). And
the procedure is repeated (Fig. 1b) such that the
Rm / Rn1 with n 2, m ¼ 2, 3, 4, . . . (2) ensemble is kept in a transient desynchronized
state. The relaxation to a clustered state is due to
Small values of R1 in the stimulation-induced the system being attracted by the center manifold
cluster state lead to small values of other as characterized by Eq. 2. By imposing a cluster
order parameters and to onset of desynchro- state, the stimulation does only half of the
nization during the poststimulation transient. desynchronizing work. The rest, namely,
Stimulation-free neurons will eventually approaching a uniformly desynchronized state,
resynchronize if left unperturbed for sufficiently is done by the system itself. In this way, the
long time (Fig. 1f). Hence, to maintain coupling, which causes the synchronization, is
a desynchronized neuronal firing, CR stimuli used for improving the desynchronizing effect.
have to be administered repetitively. In summary, by shifting the system into an
CR stimulation exploits transient responses unstable cluster state, the system reacts by auto-
which are due to the oscillators’ (pathologically matically running through a transient
strong) interactions. The general stimulation pro- desynchronized state.
tocol of the intermittent CR stimulation (Tass The effectively desynchronizing intermittent
et al. 2009; Lysyansky et al. 2011a) is illustrated CR stimulation (Fig. 1b) can be used to prevent
in Fig. 2. Here, the collective dynamics is visual- from resynchronization. For this, the repetitive
ized by considering the collective firing of the stimulus administration can be organized either
neurons. A single firing/bursting model neuron regardless of the state of the stimulated ensemble
fires/bursts whenever its phase is close to zero (open-loop control) or in a demand-controlled
(modulo 2p) (Kuramoto 1984; Ermentrout and way (closed-loop control), where the following
Kopell 1991; Grannan et al. 1993; Hansel three different control strategies can be utilized:
et al. 1993; Tass 1999). The collective firing can (i) Periodic administration of CR stimuli: The
be illustrated with the relative number of neurons most simple, open-loop type of stimulation
producing an action potential or burst at time t is a periodic administration of CR stimuli.
given by Here, the time intervals of fixed length of
CR stimulation (ON cycles) are typically
number of neurons with coscj > 0:99 recurrently followed by time intervals of
nfire ðtÞ ¼ :
N fixed length where the stimulation is switched
(3) off (OFF cycles) (Fig. 1b).
(ii) Demand-controlled timing of the administra-
0 nfire(t) 1 for all t. nfire(t) = 0 means that tion of identical stimuli: Whenever the pop-
no neuron fires/bursts, while all neurons fire/burst ulation tends to resynchronize, the same
Computational Model-Based Development of Novel neurons is maintained by repetitive administration of CR

Stimulation Algorithms, Fig. 2 A general scheme of stimuli intermingled with epochs of no stimulation
the intermittent CR stimulation. Desynchronized firing of
stimulus is administered (Fig. 3). The stron- times tk = kvT, where k = 0,1,2,3,. . . is the
ger the synchronization among the neurons index labeling the different stimuli,
is, the more often a stimulus has to be admin- T ¼ T þ e is a time interval in the range of
istered to maintain an uncorrelated firing. In the period T of the population without
an experimental application, ideally, one has stimulation, and n is a small integer such as
to observe the synchronized oscillation dur- 2 or 3. This means that a 1: n entrainment of
ing a sufficiently long period of time in order the four subpopulations is performed, where
to perform a frequency analysis which yields the spontaneous frequency of the neurons is
the period T of the population in the absence approximately v times larger compared to
of stimulation and, thus, the critical stimula- the frequency of stimulus administration.
tion parameter t (the time delay between The smaller the | e |, the smaller is the stim-
successive HF pulse trains administered via ulation strength necessary to achieve an
different stimulation sites; see Fig. 1). Alter- entrainment.
natively, one can also choose T according to
the known frequency range of the patholog- The closed-loop variants (ii) and (iii) require
ical oscillation. that the ensembles activity can be measured
(iii) Periodically administered HF pulse trains appropriately. Here, either the start times of iden-
of demand-controlled length: The stimuli tical CR stimuli or the length of periodically
are periodically administered with offset administered stimuli is calculated from the values
Computational Model-Based Development of Novel soon as the amplitude R1 of the recovering order parame-
Stimulation Algorithms, Fig. 3 Desynchronizing ter reaches the value of 0.5, the stimulus is administered
effect of the demand-controlled intermittent CR stimula- again. Periodic stimulation with demand-controlled
tion. Time course of R1 from Eq. 1 (a, c) and of nfire from length of HF pulse train (c, d): The stimulus is adminis-
Eq. 3 (b, d) during different types of stimulation. Demand- tered periodically, where the length of the HF pulse trains
controlled timing of stimulus administration (a, b): As is adapted to R1 (First published in Tass (2003b))
of R1. For example, in the case (ii), the stimula- suppress the firing in the case of the standard
tion is started if R1 becomes larger than a certain permanent HF pulse-train stimulation is about
threshold (Fig. 3, upper two plots), whereas in the five and eight times larger than that used for
case (iii), the stimulation period is longer for blocking the resynchronization in Fig. 3a, b and
larger values of R1 measured at the onset of the 3c, d, respectively. This illustrates the effective-
stimulation (Fig. 3, bottom two plots); see refer- ness of the demand-controlled CR stimulation.
ences (Tass 2003a, b) for details. The latter can effectively desynchronize stimu-
Applying the standard, permanent HF stimulated oscillators with a significantly smaller
lation (Benabid et al. 1991; Blond et al. 1992) (in amount of stimulation current compared to the
a first approximation) corresponds to stimulating standard permanent HF pulse-train stimulation.
each neuron with the same HF pulse train. During The efficacy of CR stimulation can further be
a permanent HF stimulation, a high-frequency improved by an optimal choice of the stimulation
entrainment of the order parameter Z1 captures parameters. Several computational studies on
Z1 in a small portion of the complex plane (Tass neuronal models of different complexity have
2001), so that the individual neurons’ firing is addressed this problem and showed that intermit-
stopped, but no desynchronization occurs (see tent m:n ON-OFF CR stimulation, where
also Pyragas et al. (2013)). In contrast, during m cycles with stimulation ON are recurrently
stimulation, R1 can be even larger compared to followed by n cycles with stimulation OFF, is
its pre-stimulus level, and after stimulation, the most effective for a weak stimulation intensity
synchronous firing continues immediately. To and short ON intervals (Lysyansky et al. 2011a).
suppress the firing with such a simple pulse The stimulation-induced cluster state leads to the
train persistently, it has to be administered per- longest desynchronized poststimulation transient
manently. The number of single pulses used to which can further be prolonged for nonuniform
timing of the stimuli onsets (L€ucken et al. 2013).

The number of stimulation sites is another impor-
tant stimulation parameter, and its optimal choice
essentially depends on the properties of the neu-
ronal tissue (Lysyansky et al. 2013). For a weak
(strong) spatial decay rate of the stimulation cur-
rent with distance to the stimulation site, CR
stimulation can optimally be delivered via small
(large) number of stimulation sites.
Spike Timing-Dependent Plasticity

It is well known that synapses in the brain are
involved in the adaptation process, where their
efficacy is governed by spike timing-dependent
plasticity (STDP) (Gerstner et al. 1996; Markram
et al. 1997; Feldman 2000; Wittenberg and Wang Computational Model-Based Development of Novel
2006; Caporale and Dan 2008). Plasticity is Stimulation Algorithms, Fig. 4 (a) Normalized plas-
ticity function versus the difference of the spike/burst
a fundamental property of the nervous system: timing of post- and presynaptic neuron Dt = tpost tpre. (b)
In order to learn and to adapt to sensory inputs, The average update rate of synaptic weights
neurons continuously regulate the strength of Dke ¼ WðxÞrðxÞdðxÞ for the uniform distribution
their synaptic connections in relation to the density r= 1/2 e of Dt [e, e]
mutual timing properties of their firing or burst-
ing (Hebb 1949; Gerstner et al. 1996; Markram collective dynamics may coexist. If, for instance,
et al. 1997; Debanne et al. 1998; Kilgard and the initial coupling is weak, the neurons get
Merzenich 1998; Abbott and Nelson 2000; desynchronized, and the spiking time differences
Feldman 2000; Song et al. 2000; van Hemmen Dt get broadly (e.g., uniformly) distributed. Then
2001; Zhou et al. 2003). However, plasticity may STDP (Fig. 4a) leads to further suppression of the
not only lead to desired learning and optimization coupling (see Fig. 4b for large e where Dk < 0),
processes. Rather neuronal populations can learn the neurons remain desynchronized, and
pathologically strong interactions which may a weakly coupled desynchronized state stabilizes.
lead, e.g., to the emergence of epilepsy On the other hand, if the initial coupling is strong,
(Morimoto et al. 2004; Speckmann and Elger the neurons fire synchronously, i.e., the spiking
1991). This is well known from the so-called time differences are narrowly distributed around
kindling phenomenon (Goddar 1967), where pre- zero (see Fig. 4b for small e where Dk > 0), and
paratory stimulation induces the spontaneous synaptic weights get further potentiated. Such
production of epileptic seizures without gross initial conditions result in a strongly coupled
morphological changes (Morimoto et al. 2004). synchronized state. Along with the two regimes
An example of the STDP function is illus- mentioned above, some other states with inter-
trated in Fig. 4a (see also references Bi and Poo mediate coupling strength and synchronization
(1998) and Bi (2002)), and the synaptic weights may coexist as well.
are updated in a point-like process by the incre- The impact of plasticity on synaptic weights
ment W(Dt) as soon as the pre- and postsynaptic and collective neural dynamics has been
neurons fire. For such a plasticity rule, the synap- accounted for by several theoretical studies on
tic weight is potentiated or depressed depending desynchronizing stimulation methods (Tass and
on whether the postsynaptic firing follows or Majtanik 2006; Hauptmann and Tass 2007; Tass
advances the presynaptic firing, respectively. and Hauptmann 2007). They have initiated an
STDP can lead to multistability where states approach which is targeted on unlearning patho-
with different connectivity and different logically strong synaptic interactions by
desynchronizing brain stimulation and which has Tass and Majtanik 2006) and for the treatment of
further been developed in subsequent papers tinnitus by noninvasive acoustic stimulation
(Hauptmann and Tass 2009, 2010; Tass and (Tass and Popovych 2012), respectively. For
Popovych 2012; Popovych and Tass 2012). This both stimulation modalities, there exist a range
approach exploits plasticity in two different ways: of stimulation intensities, where CR stimulation
On the one hand, due to plasticity, desynchronizing can effectively change the collective dynamics of
stimulation may decrease the strength of the neu- the stimulated neural network to a weakly C
rons’ synapses by decreasing the rate of coinci- coupled and desynchronized regime which stably
dences. On the other hand, neuronal networks persists after the stimulation is completely
with synaptic plasticity may exhibit bi- or switched off (Fig. 5, green curves; Popovych
multistability (Seliger et al. 2002; Tass and and Tass 2012). Desynchronizing CR stimulation
Majtanik 2006; Hauptmann and Tass 2007; Tass results in an anti-kindling of the pathological
and Hauptmann 2007; Maistrenko et al. 2007). connectivity, and, finally, the physiological
Accordingly, by decreasing the mean synaptic weakly coupled and desynchronized state is
weight, desynchronizing stimulation may shift reestablished. This may lead to long-lasting ther-
a neuronal population from a stable synchronized apeutic effects of CR stimulation. On the other
(pathological) state to a stable desynchronized hand, too weak or too strong CR stimulation does
(healthy) state, where the neuronal population not cause long-lasting effects, and the synaptic
remains thereafter, if left unperturbed. connectivity returns to the pre-stimulation level
In PD neuronal populations of the basal after cessation of stimulation (Fig. 5, blue and red
ganglia are strongly synchronized (Beurrier curves, respectively). In the latter case, however,
et al. 2002; Schnitzler et al. 2006; Timmermann CR stimulation may induce an acute (during
et al. 2007). By the same token, abnormal neural stimulation) normalization of abnormal
synchrony is a hallmark of tinnitus (Llinas connectivity.
et al. 1999; Weisz et al. 2005; Roberts The theoretical findings on the properties of
et al. 2010; Adamchic et al. 2013). In the syn- CR stimulation have been verified experimen-
chronized regime, synaptic plasticity results in tally. The resetting impact and the induced
a further amplification of the synchronized activ- transient desynchronization of an electrical
ity by a strengthening of the synaptic connections short-pulse stimulation, on which the CR tech-
(Nowotny et al. 2003). Properly designed stimu- nique is based, have been reported in vivo for
lation may be used to break this vicious circle and coupled neuronal bursters in paddlefish (Neiman
to induce an anti-kindling (Tass and Majtanik et al. 2007). Taking into account STDP, the long-
2006; Hauptmann and Tass 2007, 2009, 2010; lasting desynchronizing effects of CR stimulation
Tass and Hauptmann 2007; Tass and Popovych have been investigated in detail in theoretical
2012; Popovych and Tass 2012), which finally studies (Tass and Majtanik 2006; Tass and
may reestablish the normal level of connectivity, Hauptmann 2007; Hauptmann and Tass 2007),
associated with a mostly uncorrelated neuronal and the results have been confirmed experimen-
activity. In this way a sustained long-lasting tally in vitro in rat hippocampal slice (Tass
desynchronization can be achieved, and thera- et al. 2009). The beneficial therapeutic long-
peutic aftereffects can be established after the lasting aftereffects of the weak electrical CR
cessation of desynchronizing stimulation as stimulation have been observed in the MPTP-
predicted computationally (Tass et al. 2012a, b). treated macaque monkeys, in contrast to
The anti-kindling process in neural networks a stronger CR stimulation and to standard HF
with STDP is illustrated in Fig. 5 for CR stimu- DBS (Tass et al. 2012b). It was shown that uni-
lation for two stimulation modalities: somatic lateral CR stimulation delivered to the STN of
electrical and synaptically mediated (e.g., sen- parkinsonian MPTP monkeys for only 2 h per day
sory) stimulation, which were suggested for during 5 days leads to significant and sustained
desynchronizing electrical DBS (Tass 2003a, b; therapeutic aftereffects for at least 30 days, while
Computational Model-Based Development of Novel indicated by the red bar and by the vertical dashed lines for
Stimulation Algorithms, Fig. 5 Stimulation-induced direct electrical CR stimulation, indirect excitatory CR
rewiring and desynchronization of the neural ensemble stimulation, and indirect inhibitory CR stimulation
with STDP by desynchronizing coordinated (CR) stimu- (as indicated in the plots) delivered to a neuronal popula-
lation. The time courses of the mean synaptic weights tion in a strongly coupled and synchronized regime; see
CðtÞ are plotted for different stimulation intensities I as (Popovych and Tass 2012) for details (Adapted from
indicated in the legends. The stimulation time interval is Popovych and Tass (2012))
standard 130 Hz DBS and strong CR stimulation signals arrive at the neural target population as
have no aftereffects (Tass et al. 2012b). postsynaptic potentials. Sensory CR stimulation
Computational studies revealed that CR stim- has been suggested for desynchronizing the neu-
ulation is effective for a number of stimulation ral synchrony underlying tinnitus (Tass and
setups and demonstrates a great applicability. CR Popovych 2012) and successively verified in a
stimulation has been suggested for counteraction clinical proof of concept study in tinnitus patients
of a cerebral hypoactivity found in a number of treated with noninvasive acoustic CR stimulation
diseases (Candy et al. 1983; Wolkin et al. 1992; (Tass et al. 2012a; Silchenko et al. 2013;
Salehi et al. 1994; Kishimoto et al. 1998; Bearden Adamchic et al. 2013). It turned out that acoustic
et al. 2001), without promoting pathological syn- CR stimulation can significantly counteract both
chronization, by a multifrequency and phase- tinnitus symptoms and the underlying pathologi-
shifted activation of the stimulated neuronal cal neuronal synchronization by normalizing the
networks (Lysyansky et al. 2011b). Other com- effective connectivity and restoring the func-
putational studies showed that CR stimulation tional patterns of activity (Tass et al. 2012a;
can be effective in inducing desynchronization Silchenko et al. 2013; Adamchic et al. 2013).
for direct somatic stimulation and as well as for
excitatory or inhibitory synaptically meditated Multisite Linear Delayed Feedback
stimulation (Popovych and Tass 2012); see Similarly as in the case of CR stimulation,
Fig. 5. The latter stimulation setup might corre- multisite linear delayed feedback (MLDF)
spond to stimulation of afferent or efferent fibers (Hauptmann et al. 2005a, b, c, 2007a, b) is admin-
or sensory stimulation where the stimulation istered via several stimulation sites, e.g., via four
sites as illustrated in Fig. 1a. The individual stim- cj(t) of the stimulated neuronal subpopulation
ulation signals Sm(t) of each of the stimulation assigned to the stimulation site m are attracted
sites are however derived from the delayed mean to the phase C(ttm) of the corresponding stim-
field Z(t) of the stimulated ensemble using differ- ulation signal. Hence, the phases of all oscillators
ent time delays for different stimulation signals. stimulated with MLDF become symmetrically
The mean field characterizes the collective mac- redistributed on the circle (0,2p) in a cluster
roscopic dynamics of the oscillators and can be state. The order parameter R1(t) is thus mini- C
viewed as the ensemble average of the signals mized. Depending on the value of the delay t,
zj(t), j = 1,. . .,N, of individual oscillators, the stimulation can induce different cluster
XN
ZðtÞ ¼ N 1 j¼1 zj ðtÞ. states in the stimulated ensemble, where the
corresponding order parameter Rm attains large
For n stimulation sites, the stimulation signals values.
are calculated as Sm(t) = KZ(t tm), m = 1,. . .,n, As shown in Fig. 6b, c, the in-phase synchro-
where K is the amplification parameter, and the nization in the stimulated ensemble is effectively
values of delay tm, for example, for n = 4 are suppressed (for time t > 200, where both cou-
calculated from the following relation pling and stimulation are switched on), where the
order parameter R1(t) = |Z1(t)| from Eq. 1 attains
11 2ðm 1Þ small values (Fig. 6b, c, red curve). This indicates
tm ¼ t, m ¼ 1, 2, 3, 4: (4)
8 a symmetrical redistribution of the oscillator
phases cj(t) over the unit circle. For the parame-
The delays tm are symmetrically distributed ter t close to the mean period T of the stimulation-
with respect to the main delay t, where the free ensemble, a four-cluster state is induced by
smallest time delay between neighboring stimu- the stimulation, where the order parameters R1
lation sites is chosen as t/4. In the case and R2 are small, whereas R4 is relatively large
t = T (mean period of the ensemble), the delays (Fig. 6b). In the inset showing four trajectories
tm are uniformly distributed over the mean period from each of the stimulated subpopulations, the
T. In another realization, instead of four delays emerging four-cluster state induced by MLDF is
tm, m = 1, . . .,4, one can use only two of them, illustrated. For t closer to, for example, 2T, the
e.g., t1 and t2. One can put t3 = t1 and t4 = t2, stimulation induces a two-cluster state, where R1
where the polarity of the stimulation signals S3(t) is small, whereas R2 and R4 are large (Fig. 6c).
and S4(t) is reversed: S3(t) = S1(t) and The oscillators thus split into two clusters, which
S4(t) = S2(t). Assuming that the mean field of is also illustrated in the inset in Fig. 6c.
the ensemble uniformly oscillates with period MLDF robustly suppresses in-phase synchro-
T = t, the alternating polarity of the signal cor- nization as shown in Fig. 7a, where the
responds to a shift in time by half a period. There- time-averaged order parameter R1 attains small
fore, under this condition, the stimulation signal values for a broad range of parameters t and K.
S3(t) = S1(t) = KZ(t t1) approximates the On the other hand, depending on system and
stimulation signal S1(t + t/2) which is shifted in stimulation parameters, MLDF can induce either
time by half of the period, which, in turn, is equal a two-cluster state, where the second order
to KZ(t t3), where t3 is calculated according to parameter R2 attains relatively large values
Eq. 4. Analogous arguments are applicable to the (e.g., for t 2-T; see Fig. 7b), or a four-cluster
stimulation signal S4(t) = S2(t) = KZ(t t2). state, where R2 becomes small and the fourth
If the phase C(t) of the mean field Z(t) order parameter R4 increases (e.g., for t T;
(see also Z1 from Eq. 1) uniformly rotates with see Fig. 7c). Therefore, the whole stimulated
a constant frequency O = 2p/t, the phases population splits into two or four distinct sub-
Fm(t) = C(ttm) of the stimulation signals populations. Within the phase clusters, the indi-
Sm(t) are equidistantly distributed over the unit vidual oscillators have phases close to each
circle as illustrated in Fig. 6a. Then the phases other, while the different phase clusters are
Novel Stimulation
Algorithms,
Fig. 6 Control of
synchronization by
multisite linear delayed
feedback (MLDF)
stimulation. (a)
Distribution of the phases
Fm(t) of the stimulation
signals Sm(t) administered
via four stimulation sites
(as in Fig. 1a), with the
delayed mean phase,
Fm(t) = C(t\tm), with
delays tm from Eq. 4 for
t = T. (b, c) Time courses
of the amplitudes of the
cluster variables (Eq. 1) the
order parameters R1, R2,
and R4. In the subplots, four
trajectories from each of the
four stimulated
subpopulations assigned to
each of the four different
stimulation sites are shown
for t (320, 340).
Parameter t = T in (b) and
t = 2T in (c)
equidistantly distributed on the circle. Hence, the same stimulation signal S(t). In this case the
depending on the values of the parameters t and stimulation signal S(t) attains the form
K, MLDF with four stimulation sites may cause S(t) = KZ(t t). For stimulation with SLDF, in
either a two-cluster state, where R1 is close to the corresponding two-parameter diagram
zero and R2 is large, or a four-cluster state, (Fig. 7d), islands of perfect desynchronization
where both R1 and R2 are small, but R4 is large. are complemented by areas of stimulation-
The cluster states become less pronounced, and enhanced synchronization. In the limit N ! 1,
the phases redistribute on the circle even more the order parameter R1 = 0 in the desynchro-
uniformly if a local coupling as well as a spatially nization regions, where the phases are uniformly
decaying profile of the current spread is taken distributed on the circle (0, 2p) (Rosenblum and
into account (Hauptmann et al. 2005a). Pikovsky 2004a, b). This is the state of complete
In Fig. 7d a similar two-parameter diagram for desynchronization, where the stimulated oscilla-
the averaged order parameter R1(t) is presented tors rotate with different frequencies indicating
for a single-site linear delayed feedback (SLDF) an absence of any cluster state whatsoever. The
suggested for synchronization control in refer- island-like structure of desynchronization
ences (Rosenblum and Pikovsky 2004a, b). The regions in parameter space of SLDF (Fig. 7d)
stimulation is performed via one stimulation was also experimentally confirmed for arrays of
electrode in such a way that all oscillators of the coupled electrochemical oscillators (Zhai
ensemble (in a first approximation) receive et al. 2008).
Computational Model-Based Development of Novel impact of the single-site linear delayed feedback (SLDF)
Stimulation Algorithms, Fig. 7 Impact of the MLDF on the oscillatory population is illustrated, where the
stimulation versus parameters t and stimulus amplifica- values of the order parameter hR1i are depicted in color
tion K. The time-averaged order parameters hR1i, hR2i and versus parameters t and K (First published in Popovych
hR4i are depicted in plots (a–c), respectively, and encoded et al. (2006))
in color ranging from 0 (blue) to 1 (red). In plot (d), the
The important property of the stimulation with and one stimulating site is required; see Fig. 8a.
the multi- and single-site linear delayed feedback All stimulated oscillators receive the same stim-
is the inherit demand-controlled character of the ulation signal S(t) which is constructed from the
methods. As soon as the desired desynchronized measured mean field of the ensemble. It is
state is achieved, the values of the order parame- assumed that the measured mean field Z(t) of
ter R1(t), i.e., the amplitude of the mean field, the ensemble has the form of a complex analytic
become small. Along with the order parameter, signal Z(t) = X(t) + iY(t), where X(t) and Y(t) are
in the desynchronized state, the amplitude of the the real and imaginary parts of Z(t), respectively.
stimulation signal S(t) vanishes as well. Stimula- If only the real part X(t) of the mean filed is
tion with multi- and single-site linear delayed measured, the imaginary part can be calculated,
feedback thus represents a noninvasive control e.g., with the Hilbert transform (Pikovsky
method for desynchronization of coupled oscilla- et al. 2001). The stimulation signal is then
tors. The stimulated ensemble is then subjected to constructed by a nonlinear combination of
a highly effective control at a minimal amount of a delayed complex conjugate mean field with
stimulation force. the instantaneous mean field (Popovych
et al. 2005, 2006a, b, 2008; Popovych and Tass
Nonlinear Delayed Feedback 2010)
As in the case of the single-site linear delayed
feedback, for the stimulation with nonlinear
delayed feedback (NDF), only one registering SðtÞ ¼ KZ2 ðtÞZ ðt ntÞ (5)
Novel Stimulation
Algorithms,
Fig. 8 Control of
synchronization by
nonlinear delayed feedback
(NDF) stimulation. (a) The
macroscopic activity (mean
field) of the controlled
population is measured,
delayed, nonlinearly
combined with the
instantaneous mean field,
amplified, and fed back via
a single stimulation site. (b,
c) Desynchronization of
strongly synchronized
oscillators by NDF. Time
courses of the order
parameter R1(t) (red
curves) and the amplitude
of the stimulation signal |
S(t)| (blue curves) are
plotted for delays (b) t = T/
2 and (c) t = T, where T is
the mean period of the
stimulation-free ensemble.
In the insets, trajectories of
two selected oscillators are
depicted in the stimulated
regime (First published in
Popovych et al. (2008))
where K is a stimulus amplification parameter, t with different individual frequencies just as in

is a time delay, and the asterisk denotes complex the coupling- and stimulation-free regime.
conjugacy. As soon as a desynchronized state is achieved,
The desynchronizing effect of the stimulation the stimulation force decreases, and the stimu-
with NDF is illustrated in Fig. 8b, c. NDF (with lated system is subjected to a highly effective
onset at t = 550) desynchronizes the stimulated control with a minimal amount of stimulation
oscillators, and the order parameter R1 reaches force. Also, as soon as the oscillators
values of approximately the same order of mag- resynchronize, the mean field starts to exhibit
nitude as in the uncoupled regime (t < 400). This large-amplitude oscillations, and the stimulation
indicates a high level of desynchronization. The signal increases in amplitude and brings the
stimulation does not destroy the normal oscilla- ensemble back to a desynchronized state. This
tory activity of the individual oscillators. In the demand-controlled character of the nonlinear
insets in Fig. 8b, c, individual trajectories of two delayed feedback is illustrated in Fig. 8c where
selected oscillators of the stimulated ensemble the onsets of resynchronization (increase of R1(t),
are plotted. The stimulated oscillators rotate red curve) at times around t 850, 1050, and
1200 lead to an increase of the amplitude of frequency control of the oscillatory population
the stimulation signal |S(t)| [blue curve], stimulated with NDF.
which in turn results in a suppression of the
resynchronization. Mixed Nonlinear Delayed Feedback
The impact of the nonlinear delayed feedback The NDF method can also be applied for
on the stimulated oscillators is twofold. On the desynchronization and decoupling of two
one hand, the stimulation can effectively (or more) interacting oscillator populations. For C
desynchronize even strongly interacting oscilla- this, mixed NDF can be used (Popovych and Tass
tors within a large range of values of the stimulus 2010), see Fig. 10a. For a drive-response cou-
amplification K; see Fig. 9a. This effect is very pling scheme, the coupling within population
robust with respect to a variation of the delay t 2 is assumed to be weak, so that, being isolated
and, as a result, with respect to a variation of the from population 1, no synchronization emerges
mean frequency O of the stimulated ensemble. in population 2. In contrast, the coupling in pop-
On the other hand, in a weakly coupled ensemble, ulation 1 is strong enough to cause synchroniza-
the stimulation can induce synchronization in tion within population 1. It then drives the second
island-like regions of small values of the stimulus population, which synchronizes because of the
amplification K complemented by domains of driving and sends a response signal back to pop-
desynchronization; see Fig. 9b. ulation 1. The second ensemble is stimulated with
An increase of the stimulus amplification signal S(t), which is constructed from the mixed
parameter K results in a gradual decay of the mean field Ze according to the rule of NDF from
order parameter R1 for both strongly and weakly Eq. 5. The mixed mean field Ze = eW1 + (1 e)
coupled oscillators, which indicates the onset of W2 is a linear combination of the mean fields W1
desynchronization in the stimulated ensemble. and W2 of populations 1 and 2, respectively.
Simultaneously, the amplitude of the stimulation The level of mixing of the mean fields W1 and
signal |S(t)| decays as well, indicating the W2 within the stimulation signal is given by the
demand-controlled character of the nonlinear parameter e. Depending on e the mixed NDF can
delayed feedback stimulation. For a fixed delay have different desynchronizing effects on
t > 0, the order parameter and the amplitude of populations 1 and 2.
the stimulation signal decay as |K| increases • Small e: Mostly population 2 contributes to
according to the following power law (Fig. 9c): the stimulation signal. The mixed NDF
desynchronizes the driven and stimulated
R1 jK j1=2 , jSj jK j1=2 : (6) population 2 (Fig. 10c), but the driving ensem-
ble 1 remains unaffected and exhibits strongly
For large values of K, all stimulated oscillators synchronized dynamics (Fig. 10b). The
rotate with different frequencies close to the nat- populations get effectively decoupled from
ural frequencies (Popovych et al. 2005, 2006a, each other.
2008). The oscillators thus exhibit a uniform • Intermediate e: Both populations equally con-
desynchronous dynamics without any kind of tribute to the stimulation signal. Both ensem-
cluster states. In addition, depending on the bles remain synchronized (Fig. 10b, c).
values of the delay t, the nonlinear delayed feed- • Large e: Mostly population 1 contributes to
back can significantly change the mean frequency the stimulation signal. Both ensembles are
O, i.e., the frequency of the mean field Z(t) of the effectively desynchronized by the mixed
stimulated ensemble (Popovych et al. 2005, NDF (Fig. 10b, c).
2006a, 2008). The macroscopic dynamics can In the latter case, the desynchronization
thus be either accelerated or slowed down, induced by the mixed NDF in the driven and
whereas the individual dynamics remains close stimulated population 2 propagates to the drive
to the original one. This enables an approach for population 1 which is not directly stimulated.
Computational Model-Based Development of Novel encoded in color ranging from red (synchronization) to
Stimulation Algorithms, Fig. 9 Robustness of the NDF blue (desynchronization) versus delay t and stimulus
effects. (a) Stimulation-induced desynchronization and amplification K. (c) Log-log plot of the time-averaged
(b) stimulation-induced synchronization in ensembles of order parameter hR1i and amplitude of the stimulation
(a) strongly coupled and (b) weakly coupled oscillators. signal h|S(t)|i versus K. The dashed line has the slope
The time-averaged values of the order parameter hR1i are 0.5 and is given for comparison
This is indicative of an indirect control of syn- signal is considered to be the mean field W1 of the
chronization by the mixed NDF. measured subpopulation. Below, the main atten-
tion will be paid to the proportional-differential
Proportional-Integro-Differential Feedback (PD) feedback only (for more details, see refer-
For a particularly difficult situation, where mea- ence Pyragas et al. 2007). Then, the stimulation
surement and stimulation are not possible at the signal S(t) administered to the second, stimulated
same time and at the same place, there is another subpopulation is constructed as
control method which is based on a proportional-
integro-differential feedback (PIDF). The SðtÞ ¼ PW 1 ðtÞ þ DW_ 1 ðtÞ, (7)
scheme of this stimulation protocol is sketched
in Fig. 11a, see also Fig. 10a for e = 1 except for where the parameters P and D define the strength
the measured signal being processed by a PIDF of the proportional and differential feedback,
algorithm. The controlled ensemble of N coupled respectively.
oscillators is divided into two separate subpopu- The effect of the stimulation with PD feedback
lations of N1 and N2 = N N1 oscillators, one is illustrated in Fig. 11b. As the strength of the
being exclusively measured and the other being feedback (parameters P and D) increases, the
exclusively stimulated. In this way a separate stimulation results in a complete desynchro-
stimulation-registration setup is realized, where nization of the whole ensemble. The threshold
the recording and stimulating sites are spatially of the onset of desynchronization depends on
separated and the measured signal is not the relative splitting N1:N2 of the oscillators
corrupted by stimulation artifacts. The observed between subpopulations and on the mean
Computational Model-Based Development of Novel order parameter hR1i of (b) intrinsically synchronized,
Stimulation Algorithms, Fig. 10 Desynchronization drive population 1 and (c) stimulated population 2 driven
and decoupling of interacting populations by the mixed into a synchronized state by population 1 versus time
NDF. (a) Stimulation setup: The measured mean fields of delay t and mixing parameter e. The color coding as in
populations 1 and 2 are linearly combined into a mixed Fig. 9 (First published in Popovych and Tass (2010),
mean field, processed by the NDF algorithm (Eq. 5) and copyright (2010) by the American Physical Society)
fed back to the target population 2. (b, c) Time-averaged
frequency O: The threshold is larger for the subpopulation (Fig. 11c). In this sense, the PIDF
smaller number of oscillators N2 in the stimulated stimulation method appears to be very effective
populations or for larger frequency O. The latter even for a complicated stimulation protocol with
dependence can be eliminated if an integral com- a separate stimulation-registration setup.
ponent is included in the stimulation signal; see
reference (Pyragas et al. 2007). Moreover, if the Closed-Loop DBS
coupling in the ensemble is rather weak, the The standard setup of HF DBS utilizes an open-
desynchronization can be achieved by applying loop stimulation protocol where a permanent HF
the proportional feedback only. In contrast, in the electrical pulse train is administered to the target
case of strong coupling, the stimulation signal nucleus (Benabid et al. 1991; Volkmann 2004).
additionally requires the differential feedback Systematic investigations of the influence of the
for robust desynchronization. As illustrated in stimulation parameters were performed with
the two-parameter diagrams in Fig. 11c, d, there a focus on the optimization of standard HF DBS
exists a certain threshold in parameters P and via an appropriate parameter calibration
D values, where the stimulation with PIDF (Rizzone et al. 2001; Moro et al. 2002; Rubin
desynchronizes both subpopulations in the and Terman 2004) including computational
target ensemble: stimulated subpopulation approaches toward a closed-loop optimization
(Fig. 11d) and also measured, non-stimulated setup (Feng et al. 2007a, b).
Computational Model-Based Development of Novel different splitting N1:N2 and different mean frequencies O.
Stimulation Algorithms, Fig. 11 PIDF control: (a) The (c, d) The time-averaged order parameters hR1i (encoded
mean field is measured in one part of the controlled in color) of (c) the measured subpopulation and (d) stim-
ensemble and, after processing according to proportional- ulated subpopulation versus stimulation parameters P and
integro-differential feedback (PIDF) algorithm, adminis- D. The white curve is the parameter threshold for the onset
tered to the other part of the ensemble. (b) The time- of desynchronization in the subpopulations (First
averaged order parameter hR1i of the whole ensemble published in Pyragas et al. (2007), used with permission
versus the strength of the PD feedback (with P = D) for from EPL)
In monkeys rendered parkinsonian with (comprising seven pulses at 130 Hz) was
the neurotoxin 1-methyl-4-phenyl-1,2,3,6- delivered through a pair of electrodes located
tetrahydropyridine (MPTP), a closed-loop DBS in the GPi at a predetermined, fixed latency
was tested under acute conditions (Rosin (80 ms) following each action potential recorded
et al. 2011). To this end, a short train through an electrode placed in the primary motor
cortex (M1). This type of stimulation caused refractory movement disorders (Benabid
a strong decrease of the firing rate of pallidal et al. 1991; Volkmann 2004). To overcome lim-
neurons together with a pronounced decrease of itations of this type of therapy, a model-based
the oscillatory neuronal activity at the tremor development of novel deep brain stimulation
frequency (4–7 Hz) and at the double tremor techniques has been initiated (Tass 1999). The
frequency (9–15 Hz) along with an amelioration ingredients of this approach are dynamic neuro-
of the MPTP-induced akinesia. After cessation of nal self-organization principles combined with C
this type of closed-loop DBS, the initial firing fundamental plasticity rules of the nervous sys-
pattern reverted back, i.e., pallidal firing rate tem (Tass 2003a, b; Tass and Majtanik 2006;
and pallidal oscillatory activity attained Tass and Hauptmann 2006, 2007; Hauptmann
pre-stimulus levels (Rosin et al. 2011). In con- and Tass 2007, 2009, 2010; Tass and Popovych
trast, standard continuous 130 Hz DBS caused 2012; Popovych and Tass 2012).
a less pronounced decrease of the pallidal firing The control methods discussed in this entry
rate, the oscillatory neuronal activity, and the differ with respect to the stimulation setup and
amelioration of the akinesia (Rosin et al. 2011). stimulation effects as well as other properties
Another study (Little et al. 2013) confirmed such as robustness and applicability. For exam-
the efficacy of the closed-loop adaptive DBS ple, CR stimulation, in an open-loop protocol, is
(aDBS) in PD patients, where the onsets and off- simple to apply, does not require sophisticated
sets of HF stimulation were triggered by calibration, and effectively causes desynchro-
a threshold crossing by LFP in beta band mea- nization of the stimulated oscillators. As men-
sured via the same stimulation electrode tioned above, a number of computational and
implanted in STN. The stimulation trigger thresh- experimental studies confirmed the applicability
old for the LFP amplitude was heuristically deter- and efficacy of CR stimulation under different
mined in such a way that a reduction of the stimulation modalities.
stimulation time of approximately 50 % was Other smart feedback methods are more diffi-
achieved while maintaining clinical effect. The cult to realize conceptually and technically and
onset of HF stimulation was delayed by 30–40 ms await experimental proof of concept. In modeling
after the crossing of the threshold by LFP, and the studies, they effectively result in a sustained
stimulation was sustained until beta amplitude desynchronized regime. The feedback methods
fell below the threshold again (Little can be applied under a variety of conditions and
et al. 2013). For the same stimulation intensity posses an intrinsic demand-controlled character,
and stimulation frequency (130 Hz), the aDBS where the stimulation signal is significantly
can be about 30 % more effective than standard reduced or even vanishes as soon as desynchro-
continuous HF DBS, while less than 50 % of the nization is achieved. The experimental and clin-
total electrical energy is delivered in the aDBS ical realization of these methods is a challenging
mode as compared to continuous HF DB- task, first of all, from the technical side, since
S. Moreover, despite of the used fixed beta stimulation signals have to fulfill all safety
threshold, the triggered stimulation duration aspects like charge density limits. These limits
(per 10-s block) progressively drops over time strongly affect the applicability of the slow feed-
during stimulation in the aDBS mode, which back signals, i.e., slow compared to the time-
suggests that aDBS may lead to positive adaptive scales of HF and CR pulses. In this way, the
effects in pathological parkinsonian networks application conditions of the methods have to be
(Little et al. 2013). handled with care, and stimulation setups
and effects have to clearly be distinguished for
Summary different feedback methods. Otherwise, one can
High-frequency deep brain stimulation is the come up with misleading conclusions and
standard therapy for the treatment of medically erroneous interpretations of the efficacy of
feedback methods; see the computational study delivered via depth electrodes. More sophisti-
(Dovzhenok et al. 2013) where nonlinear and cated deep brain stimulation techniques are in
linear techniques were not properly distin- the process of being established for clinical use.
guished. To this end, only an experimental proof Delayed feedback Delayed feedback is a
can finally assess the applicability and perfor- method for the creation of a closed-loop forc-
mance of the control methods. ing, where a portion of the measured output
In forthcoming studies, the mathematical signal of a system is time delayed, linearly or
modeling needs to be refined to incorporate fur- nonlinearly processed, and fed back into the
ther anatomical and physiological details, for system. This approach is often used to control
instance, contributions of glial cells (Silchenko the dynamic behavior of complex systems. In
and Tass 2008). By the same token, control tech- this entry, delayed feedback is used to control
niques have to be optimized and further devel- synchronization in ensembles of coupled
oped. As currently done for CR stimulation, oscillators, e.g., neurons.
clinical studies are necessary to evaluate the Order parameter The order parameter is
therapeutic effects of the novel stimulation tech- a quantity characterizing a phase transition or
niques under real conditions. This interdisciplin- phase change in the transformation of
ary endeavor might finally provide superior a complex system from one phase (state) to
therapies for patients with neurological or psy- another. The order parameter is convenient for
chiatric diseases. characterizing the onset and extent of synchro-
nization in larger ensembles: Perfect phase
synchronization corresponds to a large value
Glossary of the order parameter, whereas an incoherent
(desynchronized) state is associated with
Coordinated reset stimulation Coordinated a small value of the order parameter. In syn-
reset (CR) stimulation is an effectively ergetics it has been shown that the dynamics of
desynchronizing control technique, where complex systems may be governed by only
a population of synchronized oscillators is a few order parameters.
stimulated via several stimulation sites in Synchronization Synchronization (from Greek
such a way that spatially and timely coordi- syn = the same, common and chronos = time)
nated phase reset is achieved in subpopula- means the adjustment of rhythms of self-
tions assigned to each of the stimulation sustained oscillators due to their weak interac-
sites. This method is suggested for counteraction. The interacting oscillators can be regular
tion of abnormal neuronal synchronization (periodic) or chaotic. There are several differ-
characteristic for several neurological diseases ent forms of synchronization including phase,
and amelioration of their symptoms. It has complete, generalized, and lag synchroniza-
successively been verified in a number of tion, etc. In this entry, we focus on phase
experimental and clinical studies. synchronization. In the simplest form, the
Deep brain stimulation Electrical deep brain oscillators, rotating with the same frequency,
stimulation (DBS) is the standard therapy for become phase synchronized (phase locked) to
medically refractory movement disorders, each other, if the difference of their phases
e.g., Parkinson’s disease and essential tremor. remains bounded, e.g., constant. In the pres-
It requires a surgical treatment, where depth ence of noise, phase synchronization is char-
electrodes are chronically implanted in target acterized by the presence of one or more
areas like the thalamic ventralis intermedius prominent peaks of the distribution of the
nucleus or the subthalamic nucleus. For phase difference. Put otherwise, the oscillators
standard DBS, electrical high-frequency adjust their rhythms, while their amplitude
(>100 Hz) stimulation is permanently dynamics need not be correlated.
References Caporale N, Dan Y (2008) Spike timing dependent plas-
ticity: a Hebbian learning rule. Annu Rev Neurosci
Abbott L, Nelson S (2000) Synaptic plasticity: taming the 31:25–46
beast. Nat Neurosci 3:1178–1183 Daido H (1992) Order function and macroscopic mutual
Adamchic I, Toth T, Hauptmann C, Tass PA entrainment in uniformly coupled limit-cycle oscilla-
(2013) Reversing pathologically increased EEG tors. Prog Theor Phys 88:1213–1218
power by acoustic coordinated reset neuromodulation. Danzl P, Hespanha J, Moehlis J (2009) Event-based
minimum-time control of oscillatory neuron models.
Hum Brain Mapp. doi:10.1002/hbm.22314
Biol Cybern 101(5–6):387–399
C
Alberts WW, Wright EJ, Feinstein B (1969) Cortical
potentials and parkinsonian tremor. Nature Debanne D, Gahweiler B, Thompson S (1998) Long-term
221:670–672 synaptic plasticity between pairs of individual CA3
Anderson ME, Postupna N, Ruffo M (2003) Effects of pyramidal cells in rat hippocampus slice cultures.
high-frequency stimulation in the internal globus J Physiol 507:237–247
pallidus on the activity of thalamic neurons in the Deuschl G, Schade-Brittinger C, Krack P, Volkmann J,
awake monkey. J Neurophysiol 89(2):1150–1160 Sch€afer H, Bötzel K, Daniels C, Deutschl€ander A,
Andres F, Gerloff C (1999) Coherence of sequential Dillmann U, Eisner W, Gruber D, Hamel W,
movements and motor learning. J Clin Neurophysiol Herzog J, Hilker R, Klebe S, Kloß M, Koy J,
16:520–527 Krause M, Kupsch A, Lorenz D, Lorenzl S,
Bearden CE, Hoffman KM, Cannon TD (2001) The neuro- Mehdorn H, Moringlane J, Oertel W, Pinsker M,
psychology and neuroanatomy of bipolar affective dis- Reichmann H, Reuß A, Schneider GH, Schnitzler A,
order: a critical review. Bipolar Disord 3(3):106–150 Steude U, Sturm V, Timmermann L, Tronnier V,
Benabid AL, Pollak P, Gervason C, Hoffmann D, Gao Trottenberg T, Wojtecki L, Wolf E, Poewe W,
DM, Hommel M, Perret JE, de Rougemount J (1991) Voges J (2006) A randomized trial of deep-brain
Long-term suppression of tremor by chronic stimula- stimulation for Parkinson’s disease. N Engl J Med
tion of ventral intermediate thalamic nucleus. Lancet 355:896–908
337:403–406 Dovzhenok A, Park C, Worth RM, Rubchinsky LL
Benabid AL, Benazzous A, Pollak P (2002) Mechanisms (2013) Failure of delayed feedback deep brain stimu-
of deep brain stimulation. Mov Disord 17:S73–S74 lation for intermittent pathological synchronization in
Benabid AL, Wallace B, Mitrofanis J, Xia R, Piallat B, Parkinson’s disease. PLoS One 8(3):e58264
Chabardes S, Berger F (2005) A putative generalized Elble RJ, Koller WC (1990) Tremor. John Hopkins Uni-
model of the effects and mechanism of action of high versity Press, Baltimore
frequency electrical stimulation of the central nervous Ermentrout B, Kopell N (1991) Multiple pulse interac-
system. Acta Neurol Belg 105(3):149–157 tions and averaging in systems of coupled neural
Bergman H, Wichmann T, Delong MR (1990) Reversal of assemblies. J Math Biol 29:195–217
experimental Parkinsonism by lesions of the Feldman DE (2000) Timing-based LTP and LTD at verti-
subthalamic nucleus. Science 249(4975):1436–1438 cal inputs to layer II/III pyramidal cells in rat barrel
Beurrier C, Bioulac B, Audin J, Hammond C (2001) High- cortex. Neuron 27(1):45–56
frequency stimulation produces a transient blockade of Feng X, Greenwald B, Rabitz H, Shea-Brown E, Kosut
voltage-gated currents in subthalamic neurons. R (2007a) Toward closed-loop optimization of deep
J Neurophysiol 85:1351–1356 brain stimulation for Parkinson’s disease: concepts
Beurrier C, Garcia L, Bioulac B, Hammond C (2002) and lessons from a computational model. J Neural
Subthalamic nucleus: a clock inside basal ganglia? Eng 4(2):L14–L21
Thal Relat Syst 2:1–8 Feng XJ, Shea-Brown E, Greenwald B, Kosut R, Rabitz
Bi G-Q (2002) Spatiotemporal specificity of synaptic plas- H (2007b) Optimal deep brain stimulation of the
ticity: cellular rules and mechanisms. Biol Cybern subthalamic nucleus – a computational study.
87:319–332 J Comput Neurosci 23(3):265–282
Bi G-Q, Poo M-M (1998) Synaptic modifications in cul- Filali M, Hutchison WD, Palter VN, Lozano AM,
tured hippocampal neurons: dependence on spike Dostrovsky JO (2004) Stimulation-induced inhibition
timing, synaptic strength, and postsynaptic cell type. of neuronal firing in human subthalamic nucleus. Exp
J Neurosci 18:10464–10472 Brain Res 156(3):274–281
Blond S, Caparros-Lefebvre D, Parker F, Assaker R, Freund H-J (2005) Long-term effects of deep brain
Petit H, Guieu J-D, Christiaens J-L (1992) Control of stimulation in Parkinson’s disease. Brain
tremor and involuntary movement disorders by 128:2222–2223
chronic stereotactic stimulation of the ventral interme- Gerstner W, Kempter R, van Hemmen JL, Wagner H
diate thalamic nucleus. J Neurosurg 77:62–68 (1996) A neuronal learning rule for sub-millisecond
Candy JM, Perry RH, Perry EK, Irving D, Blessed G, temporal coding. Nature 383:76–78
Fairbairn AF, Tomlinson BE (1983) Pathological- Goddar G (1967) Development of epileptic seizures
changes in the nucleus of Meynert in Alzheimer’s through brain stimulation at low intensity. Nature
and Parkinson’s diseases. J Neurol Sci 59(2):277–289 214:1020–1021
Gradinaru V, Mogri M, Thompson KR, Henderson JM, Kishimoto H, Yamada K, Iseki E, Kosaka K, Okoshi
Deisseroth K (2009) Optical deconstruction of Parkin- T (1998) Brain imaging of affective disorders and
sonian neural circuitry. Science 324(5925):354–359 schizophrenia. Psychiatry Clin Neurosci 52(Suppl):
Grannan ER, Kleinfeld D, Sompolinsky H (1993) S212–S214
Stimulus-dependent synchronization of neuronal Kiss IZ, Rusin CG, Kori H, Hudson JL (2007) Engineering
assemblies. Neural Comput 5:550–569 complex dynamical structures: sequential patterns and
Grill WM, McIntyre CC (2001) Extracellular excitation of desynchronization. Science 316(5833):1886–1889
central neurons: implications for the mechanisms of Kumar R, Lozano AM, Sime E, Lang AE (2003) Long-term
deep brain stimulation. Thal Relat Syst 1:269–277 follow-up of thalamic deep brain stimulation for essential
Haken H (1983) Advanced synergetics. Springer, Berlin and Parkinsonian tremor. Neurology 61(11):1601–1604
Hammond C, Ammari R, Bioulac B, Garcia L (2008) Kuramoto Y (1984) Chemical oscillations, waves, and
Latest view on the mechanism of action of deep brain turbulence. Springer, Berlin
stimulation. Mov Disord 23(15):2111–2121 Limousin P, Speelman JD, Gielen F, Janssens M (1999)
Hansel D, Mato G, Meunier C (1993) Phase dynamics of Multicentre European study of thalamic stimulation in
weakly coupled Hodgkin-Huxley neurons. Europhys Parkinsonian and essential tremor. J Neurol Neurosurg
Lett 23:367–372 Psychiatry 66:289–296
Hashimoto T, Elder CM, Okun MS, Patrick SK, Vitek JL Little S, Pogosyan A, Neal S, Zavala B, Zrinzo L, Hariz M,
(2003) Stimulation of the subthalamic nucleus changes Foltynie T, Limousin P, Ashkan K, FitzGerald J,
the firing pattern of pallidal neurons. J Neurosci Green AL, Aziz TZ, Brown P (2013) Adaptive deep
23(5):1916–1923 brain stimulation in advanced Parkinson disease. Ann
Hauptmann C, Tass PA (2007) Therapeutic rewiring by Neurol 74(3):449–457
means of desynchronizing brain stimulation. Llinas RR, Ribary U, Jeanmonod D, Kronberg E, Mitra PP
Biosystems 89:173–181 (1999) Thalamocortical dysrhythmia: a neurological
Hauptmann C, Tass PA (2009) Cumulative and after- and neuropsychiatric syndrome characterized by mag-
effects of short and weak coordinated reset stimula- netoencephalography. Proc Natl Acad Sci U S
tion: a modeling study. J Neural Eng 6(1):016004 A 96(26):15222–15227
Hauptmann C, Tass PA (2010) Restoration of segregated, L€
ucken L, Yanchuk S, Popovych OV, Tass PA
physiological neuronal connectivity by (2013) Desynchronization boost by non-uniform coor-
desynchronizing stimulation. J Neural Eng 7:056008 dinated reset stimulation in ensembles of pulse-
Hauptmann C, Popovych O, Tass PA (2005a) Delayed coupled neurons. Front Comput Neurosci 7(63):63
feedback control of synchronization in locally coupled Luo M, Wu YJ, Peng JH (2009) Washout filter aided mean
neuronal networks. Neurocomputing 65–66:759–767 field feedback desynchronization in an ensemble of
Hauptmann C, Popovych O, Tass PA (2005b) Effectively globally coupled neural oscillators. Biol Cybern
desynchronizing deep brain stimulation based on 101(3):241–246
a coordinated delayed feedback stimulation via several Lysyansky B, Popovych OV, Tass PA (2011a)
sites: a computational study. Biol Cybern 93:463–470 Desynchronizing anti-resonance effect of m:n on-off
Hauptmann C, Popovych O, Tass PA (2005c) Multisite coordinated reset stimulation. J Neural Eng
coordinated delayed feedback for an effective 8(3):036019
desynchronization of neuronal networks. Stoch Dyn Lysyansky B, Popovych OV, Tass PA (2011b) Multi-
5(2):307–319 frequency activation of neuronal networks by coordi-
Hauptmann C, Omelchenko O, Popovych OV, nated reset stimulation. Interface Focus 1(1):75–85
Maistrenko Y, Tass PA (2007a) Control of spatially Lysyansky B, Popovych OV, Tass PA (2013) Optimal
patterned synchrony with multisite delayed feedback. number of stimulation contacts for coordinated reset
Phys Rev E 76:066209 neuromodulation. Front Neuroeng 6(5):5
Hauptmann C, Popovych O, Tass PA (2007b) Demand- Maistrenko YL, Lysyansky B, Hauptmann C, Burylko O,
controlled desynchronization of oscillatory networks Tass PA (2007) Multistability in the Kuramoto model
by means of a multisite delayed feedback stimulation. with synaptic plasticity. Phys Rev E 75(6):066207
Comput Visual Sci 10:71–78 Markram H, L€ ubke J, Frotscher M, Sakmann B (1997)
Hauptmann C, Popovych O, Tass PA (2007c) Regulation of synaptic efficacy by coincidence of post-
Desynchronizing the abnormally synchronized neural synaptic APs and EPSPs. Science 275:213–215
activity in the subthalamic nucleus: a modeling study. McIntyre CC, Grill WM, Sherman DL, Thakor NV
Expert Rev Med Devices 4(5):633–650 (2004a) Cellular effects of deep brain stimulation:
Hebb D (1949) The organization of behavior: model-based analysis of activation and inhibition.
a neuropsychological theory, Wiley book in clinical J Neurophysiol 91(4):1457–1469
psychology. Wiley, New York McIntyre CC, Savasta M, Goff LK-L, Vitek JL (2004b)
Kilgard MP, Merzenich MM (1998) Cortical map reorga- Uncovering the mechanism(s) of action of deep brain
nization enabled by nucleus basalis activity. Science stimulation: activation, inhibition, or both. Clin
279(5357):1714–1718 Neurophysiol 115(6):1239–1248
Meissner W, Leblois A, Hansel D, Bioulac B, Gross CE, Popovych OV, Hauptmann C, Tass PA (2008) Impact of
Benazzouz A, Boraud T (2005) Subthalamic high fre- nonlinear delayed feedback on synchronized oscilla-
quency stimulation resets subthalamic firing tors. J Biol Phys 34:367–379
and reduces abnormal oscillations. Brain Pyragas K, Popovych OV, Tass PA (2007) Controlling
128:2372–2382 synchrony in oscillatory networks with a separate
Milton J, Jung P (eds) (2003) Epilepsy as a dynamic stimulation-registration setup. Europhys Lett
disease. Biological and medical physics, biomedical 80:40002
engineering. Springer, Berlin Pyragas K, Novicenko V, Tass PA (2013) Mechanism of
Miocinovic S, Parent M, Butson CR, Hahn PJ, Russo GS, suppression of sustained neuronal spiking under high-
C
Vitek JL, McIntyre CC (2006) Computational analysis frequency stimulation. Biol Cybern 107:669–684
of subthalamic nucleus and lenticular fasciculus acti- Rizzone M, Lanotte M, Bergamasco B, Tavella A,
vation during therapeutic deep brain stimulation. Torre E, Faccani G, Melcarne A, Lopiano L (2001)
J Neurophysiol 96(3):1569–1580 Deep brain stimulation of the subthalamic nucleus in
Morimoto K, Fahnestock M, Racine RJ (2004) Kindling Parkinson’s disease: effects of variation in stimulation
and status epilepticus models of epilepsy: rewiring the parameters. J Neurol Neurosurg Psychiatry
brain. Prog Neurobiol 73(1):1–60 71(2):215–219
Moro E, Esselink RJA, Xie J, Hommel M, Benabid AL, Roberts LE, Eggermont JJ, Caspary DM, Shore SE,
Pollak P (2002) The impact on Parkinson’s disease of Melcher JR, Kaltenbach JA (2010) Ringing ears: the
electrical parameter settings in stn stimulation. Neu- neuroscience of tinnitus. J Neurosci
rology 59(5):706–713 30(45):14972–14979
Nabi A, Moehlis J (2011) Single input optimal control for Rodriguez-Oroz MC, Obeso JA, Lang AE, Houeto JL,
globally coupled neuron networks. J Neural Eng Pollak P, Rehncrona S, Kulisevsky J, Albanese A,
8(6):065008 Volkmann J, Hariz MI, Quinn NP, Speelman JD,
Neiman A, Russell D, Yakusheva T, DiLullo A, Tass PA Guridi J, Zamarbide I, Gironell A, Molet J, Pascual-
(2007) Response clustering in transient stochastic syn- Sedano B, Pidoux B, Bonnet AM, Agid Y, Xie J,
chronization and desynchronization of coupled neuro- Benabid AL, Lozano AM, Saint-Cyr J, Romito L,
nal bursters. Phys Rev E 76:021908 Contarino MF, Scerrati M, Fraix V, Van Blercom
Nini A, Feingold A, Slovin H, Bergmann H (1995) Neu- N (2005) Bilateral deep brain stimulation in
rons in the globus pallidus do not show correlated Parkinson’s disease: a multicentre study with 4 years
activity in the normal monkey, but phase-locked oscil- follow-up. Brain 128:2240–2249
lations appear in the MPTP model of Parkinsonism. Rosenblum MG, Pikovsky AS (2004a) Controlling syn-
J Neurophysiol 74:1800–1805 chronization in an ensemble of globally coupled oscil-
Nowotny T, Zhigulin VP, Selverston AI, Abarbanel HDI, lators. Phys Rev Lett 92:114102
Rabinovich MI (2003) Enhancement of synchroniza- Rosenblum MG, Pikovsky AS (2004b) Delayed feedback
tion in a hybrid neural circuit by spike-timing depen- control of collective synchrony: an approach to sup-
dent plasticity. J Neurosci 23(30):9776–9785 pression of pathological brain rhythms. Phys Rev
Pikovsky A, Rosenblum M, Kurths J (2001) Synchroniza- E 70:041904
tion, a universal concept in nonlinear sciences. Cam- Rosin B, Slovik M, Mitelman R, Rivlin-Etzion M, Haber
bridge University Press, Cambridge SN, Israel Z, Vaadia E, Bergman H (2011) Closed-
Pliss V (1964) Principal reduction in the theory of stability loop deep brain stimulation is superior in ameliorating
of motion. Izv Akad Nauk SSSR Math Ser Parkinsonism. Neuron 72(2):370–384
28:1297–1324 Rubin JE, Terman D (2004) High frequency stimulation of
Popovych OV, Tass PA (2010) Synchronization control of the subthalamic nucleus eliminates pathological tha-
interacting oscillatory ensembles by mixed nonlinear lamic rhythmicity in a computational model. J Comput
delayed feedback. Phys Rev E 82(2):026204 Neurosci 16(3):211–235
Popovych OV, Tass PA (2012) Desynchronizing electrical Salehi A, Lucassen PJ, Pool CW, Gonatas NK, Ravid R,
and sensory coordinated reset neuromodulation. Front Swaab DF (1994) Decreased neuronal-activity in the
Hum Neurosci 6:58 nucleus basalis of Meynert in Alzheimer’s-disease as
Popovych OV, Hauptmann C, Tass PA (2005) Effective suggested by the size of the Golgi-apparatus. Neuro-
desynchronization by nonlinear delayed feedback. science 59(4):871–880
Phys Rev Lett 94:164102 Schnitzler A, Timmermann L, Gross J (2006) Physiolog-
Popovych OV, Hauptmann C, Tass PA (2006a) Control of ical and pathological oscillatory networks in the
neuronal synchrony by nonlinear delayed feedback. human motor system. J Physiol Paris 99:3–7
Biol Cybern 95:69–85 Schuurman PR, Bosch DA, Bossuyt PM, Bonsel GJ, van
Popovych OV, Hauptmann C, Tass PA (2006b) Someren EJ, de Bie RM, Merkus MP, Speelman JD
Desynchronization and decoupling of interacting (2000) A comparison of continuous thalamic stimula-
oscillators by nonlinear delayed feedback. Int tion and thalamotomy for suppression of severe
J Bifurc Chaos 16(7):1977–1987 tremor. N Engl J Med 342:461–468
Seliger P, Young SC, Tsimring LS (2002) Plasticity and Tass PA, Popovych OV (2012) Unlearning tinnitus-
learning in a network of coupled phase oscillators. related cerebral synchrony with acoustic coordinated
Phys Rev E 65(4):041906 reset stimulation: theoretical concept and modelling.
Silchenko A, Tass P (2008) Computational modeling of Biol Cybern 106:27–36
paroxysmal depolarization shifts in neurons induced Tass PA, Hauptmann C, Popovych OV (2006) Develop-
by the glutamate release from astrocytes. Biol Cybern ment of therapeutic brain stimulation techniques with
98:61–74 methods from nonlinear dynamics and statistical phys-
Silchenko AN, Adamchic I, Hauptmann C, Tass PA ics. Int J Bifurc Chaos 16(7):1889–1911
(2013) Impact of acoustic coordinated reset Tass PA, Silchenko AN, Hauptmann C, Barnikol UB,
neuromodulation on effective connectivity in a neural Speckmann EJ (2009) Long-lasting desynchronization
network of phantom sound. Neuroimage 77:133–147 in rat hippocampal slice induced by coordinated reset
Singer W (1989) Search for coherence: a basic principle of stimulation. Phys Rev E 80(1):011902
cortical self-organization. Concepts Neurosci 1:1–26 Tass P, Adamchic I, Freund H-J, von Stackelberg T,
Smirnov DA, Barnikol UB, Barnikol TT, Bezruchko BP, Hauptmann C (2012a) Counteracting tinnitus by
Hauptmann C, Buhrle C, Maarouf M, Sturm V, Freund acoustic coordinated reset neuromodulation. Restor
H-J, Tass PA (2008) The generation of Parkinsonian Neurol Neurosci 30:137–159
tremor as revealed by directional coupling analysis. Tass PA, Qin L, Hauptmann C, Doveros S, Bezard E,
Europhys Lett 83(2):20003 Boraud T, Meissner WG (2012b) Coordinated reset
Song S, Miller K, Abbott L (2000) Competitive Hebbian has sustained aftereffects in Parkinsonian monkeys.
learning through spike-timing-dependent synaptic Ann Neurol 72(5):816–820
plasticity. Nat Neurosci 3:919–926 Timmermann L, Florin E, Reck C (2007) Pathological
Speckmann E, Elger C (1991) The neurophysiological cerebral oscillatory activity in Parkinson’s disease:
basis of epileptic activity: a condensed overview. Epi- a critical review on methods, data and hypotheses.
lepsy Res Suppl 2:1–7 Expert Rev Med Devices 4(5):651–661
Tasker RR (1998) Deep brain stimulation is preferable to Tukhlina N, Rosenblum M, Pikovsky A, Kurths J (2007)
thalamotomy for tremor suppression. Surg Neurol Feedback suppression of neural synchrony by
49:145–154 vanishing stimulation. Phys Rev E 75:011918
Tass PA (1999) Phase resetting in medicine and biology: van Hemmen J (2001) Theory of synaptic plasticity. In:
stochastic modelling and data analysis. Springer, Moss F, Gielen S (eds) Handbook of biological phys-
Berlin ics, vol 4. Elsevier, Amsterdam, pp 771–823
Tass PA (2001) Effective desynchronization with Volkmann J (2004) Deep brain stimulation for the treat-
a resetting pulse train followed by a single pulse. ment of Parkinson’s disease. J Clin Neurophysiol
Europhys Lett 55:171–177 21(1):6–17
Tass PA (2002a) Desynchronization of brain rhythms with Weisz N, Moratti S, Meinzer M, Dohrmann K, Elbert
soft phase-resetting techniques. Biol Cybern T (2005) Tinnitus perception and distress is related to
87:102–115 abnormal spontaneous brain activity as measured by
Tass PA (2002b) Effective desynchronization with bipolar magnetoencephalography. PLoS Med 2(6):e153
double-pulse stimulation. Phys Rev E 66:036226 Welter ML, Houeto JL, Bonnet AM, Bejjani PB,
Tass PA (2003a) Desynchronization by means of Mesnage V, Dormont D, Navarro S, Cornu P,
a coordinated reset of neural subpopulations – a Agid Y, Pidoux B (2004) Effects of high-frequency
novel technique for demand controlled deep brain stimulation on subthalamic neuronal activity in Par-
stimulation. Prog Theory Phys Suppl 150:281–296 kinsonian patients. Arch Neurol 61(1):89–96
Tass PA (2003b) A model of desynchronizing deep Wittenberg GM, Wang SSH (2006) Malleability of spike-
brain stimulation with a demand-controlled coordi- timing-dependent plasticity at the CA3-CA1 synapse.
nated reset of neural subpopulations. Biol Cybern J Neurosci 26(24):6610–6617
89:81–88 Wolkin A, Sanfilipo M, Wolf AP, Angrist B, Brodie JD,
Tass P, Hauptmann C (2006) Long-term anti-kindling Rotrosen J (1992) Negative symptoms and
effects induced by short-term, weak desynchronizing hypofrontality in chronic-schizophrenia. Arch Gen
stimulation. Nonlinear Phenom Complex Syst Psychiatry 49(12):959–965
9:298–312 Wunderlin A, Haken H (1975) Scaling theory for
Tass PA, Hauptmann C (2007) Therapeutic modulation of nonequilibrium systems. Z Phys B 21:393–401
synaptic connectivity with desynchronizing brain Zhai Y, Kiss IZ, Hudson JL (2008) Control of
stimulation. Int J Psychophysiol 64(1):53–61 complex dynamics with time-delayed feedback
Tass PA, Hauptmann C (2009) Anti-kindling achieved by in populations of chemical oscillators: desynchro-
stimulation targeting slow synaptic dynamics. Restor nization and clustering. Ind Eng Chem Res
Neurol Neurosci 27(6):591–611 47(10):3502–3514
Tass PA, Majtanik M (2006) Long-term anti-kindling Zhou Q, Tao HW, Poo MM (2003) Reversal and stabili-
effects of desynchronizing brain stimulation: a theo- zation of synaptic modifications in a developing visual
retical study. Biol Cybern 94(1):58–66 system. Science 300(5627):1953–1957
Computational Modeling of Olfactory Behavior 721 C
odor perception. We here briefly review this
Computational Modeling approach and highlight some examples from the
last 10 years.
▶ Metabotropic Receptors (G Protein Coupled
Receptors)
▶ Polypeptide and Protein Modeling for Drug Detailed Description
Design C
Our standard approach to modeling olfactory
behavior is as follows: (1) determine via behav-
ioral pharmacology or similar experiments which
Computational Modeling of olfactory brain areas are likely to underlie
Olfactory Behavior a certain aspect of olfactory perception; (2) create
a computational simulation of those brain areas,
Christiane Linster1 and Thomas A. Cleland2 incorporating as much constraining experimental
1
Computational Physiology Lab, Department of detail as possible given the level of modeling
Neurobiology and Behavior, Cornell University, chosen; (3) use the model to simulate specific
Ithaca, NY, USA relationships between neural activity patterns
2
Computational Physiology Lab, Department and perception, generating testable hypotheses;
of Psychology, Cornell University, Ithaca, (4) based on these hypotheses, perturb both
NY, USA experimental and computational variables via
lesions, training parameters, pharmacology, and
other experimental tools. The outcome of these
Definition studies then serves to inform the computational
models and help determine the direction of
Computational modeling is an essential tool for future study.
developing an understanding of how nervous sys- One critical aspect of this process is the selec-
tems compute. This is particularly so for question of the level of detail at which to build the
tions that span levels of analysis, attempting to model. Our modeling is nearly all based on spe-
integrate cellular, neuromodulatory, and electro- cific neural systems – as opposed to generic neu-
physiological data with behavioral performance. ral networks – but the level of detail implemented
In neuroscience, computational techniques must be appropriate given the available experi-
are used to study the mechanisms underlying mental data in order for the model to contribute
neuronal or network responses to simple and meaningfully to this research approach. More
complex inputs, analyze interactions among the abstract models of particular networks, made
parameters governing the properties of a neuron with relatively reduced, single-compartment neu-
or network, and determine the coordinated mech- rons, are appropriate when the data include gen-
anisms that underlie experimentally observed eral information about neurons being excitatory
rich phenomena such as coherent oscillations or or inhibitory, reasonable estimates of synaptic
synaptic plasticity. In particular, computational connectivity patterns, and neuromodulator
modeling has been successful in associating neu- effects that are thought to excite, inhibit, and/or
ral activity with behavioral function, proposing affect the excitability of particular cell types.
neurophysiological mechanisms for observed Building simpler models with fewer free param-
behavioral capabilities, and generating novel, eters enables a relatively efficient exploration of
testable hypotheses. In our lab, computational what such a network can do and how the different
models of behavioral phenomena have enabled effects of neuromodulators, for example, could
us to elucidate relationships among odorant phys- combine to produce a functional outcome that
ical properties, top-down neuromodulatory sig- corresponds to behavioral data. Following up on
nals, olfactory neural network operations, and hypotheses generated by such models can
C 722 Computational Modeling of Olfactory Behavior
increase the efficiency of experimental design, new data often have served their scientific
ultimately producing data sufficient to build the purpose well.
richer and more complex biophysical models. Over the last two decades, we have developed
Biophysical models require considerably computational models of many behavior-related
more experimental data in order to be useful. processes within olfaction, including the cholin-
They define physical membrane properties such ergic, noradrenergic, and dopaminergic modula-
as surface area, capacitance, and cell morphol- tion of bulbar circuitry (Linster and Gervais
ogy; model many individual mechanisms such as 1996; Hasselmo et al. 1997; Linster and
sodium channels, calcium diffusion, and NMDA Hasselmo 1997; Linster and Cleland 2002;
receptors; deploy these mechanisms into or with Linster et al. 2003, 2011; Mandairon
respect to the cell membrane; and often are mor- et al. 2006a; Escanilla et al. 2009; Devore and
phologically multicompartmental. Oligocom- Linster 2012; Devore et al. 2012; de Almeida
partmental model neurons, in particular, are et al. 2013; Li and Cleland 2013), perceptual
built using perhaps 2–20 compartments that learning (Mandairon et al. 2006b), mixture
serve to accurately model spike propagation processing in the olfactory bulb (Linster and
times, segregate electrotonically distant parts of Cleland 2004), the bulbar mechanisms regulating
the neuron, construct cells with heterogeneous differentiation among similar odors (Cleland and
channel distributions, and/or implement other Sethupathy 2006; Cleland et al. 2007; Cleland
phenomena in which spatial heterogeneity mat- and Linster 2012), the role of spike synchroniza-
ters. (Neuron models with hundreds of compart- tion in antennal lobe odor processing (Linster
ments, in contrast, usually are constructed to et al. 1994; Linster and Cleland 2001) and syn-
replicate the arbor of a specific neuron measured aptic plasticity in olfactory bulb and cortex
from imaging data.) If there are constraining data (Linster et al. 2007, 2009; Linster and Cleland
for enough of these parameters, then such models 2010), and the multiple concerted processes
are singularly powerful, as they are able to imple- required to enable concentration invariance in
ment underappreciated computational elements odor perception (Cleland and Linster 1999;
like shunt inhibition, intrinsic resonance proper- Cleland et al. 2007, 2011). In each case,
ties, and the partial isolation of spines, and at their a critical factor affecting model validity has
best they can predict the network-level conse- been the comparison between model output and
quences of, for example, localized effects on ion behavioral results. How does one determine
channel properties. In the absence of sufficient whether and how the values of model output vari-
data, however, such models lose most of their ables correspond to measured indices of behav-
predictive value, as they simply contain too ior? While this is always ultimately an
many free parameters which, if unconstrained, experimental question, we have determined sev-
can enable meaningless fits to many datasets; eral behavioral indices that can be reliably
simpler models should be used in such situations. predicted by specific output metrics from our
Irrespective of the level of complexity, however, computational models, including perceptual sim-
useful models usually do not simply replicate ilarity, discrimination performance, learning rate,
established data but employ their capacity to memory capacity, and memory persistence. For
quantify and juxtapose diverse datasets in order example, we first demonstrated that the pairwise
to explain what is known and explore what that overlap in neuronal activation patterns across the
might imply. Models can range from being quite olfactory bulb input layer could predict pairwise
tightly based on cellular data (e.g., Li and Cleland perceptual similarity (Cleland et al. 2002). To the
2013) to being quite speculative (e.g., Linster and extent that this result is generalizable as a basic
Cleland 2010), but in either case, the task is to principle of olfactory coding, one then can, in
understand how best to ask the next question. subsequent studies, compare changes in pairwise
While models that stand the test of time are laud- overlap within the computational model to
able, models that are subsequently superseded by behavioral changes in perceptual discrimination.
This in turn enables, for example, the effects of combined electrophysiological and behavioral
neuromodulators on olfactory discrimination to experiments, when fine-scale spike synchrony
be modeled and tested against behavioral data was disrupted, bees’ odor discrimination perfor-
(e.g., Mandairon et al. 2006a). Importantly, mance was impaired, even though the overall
while this principle of pattern overlap among profile of evoked activity in antennal lobe neu-
glomeruli has been largely supported experimen- rons remained intact (Stopfer et al. 1997, 2003).
tally, its basis is inductive and it cannot be reli- Using the same behavioral paradigm, it subse- C
ably extrapolated; e.g., it may not extend reliably quently was demonstrated that honeybees dis-
to spatial patterns among mitral cells or within criminate odorants better when they are
piriform cortex. Like theoretical models in gen- presented at high concentrations than when they
eral, these models serve to organize our under- are presented at lower concentrations (Bhagavan
standing of complex datasets until a superior and Smith 1997). Neurophysiologically, this
theory can be constructed. result was counterintuitive because calcium
We here describe three specific examples of imaging studies have shown that higher concen-
the modeling of olfactory behavior that provided tration odor stimuli activate correspondingly
useful insights into the underlying neural mecha- larger and more overlapping areas of the antennal
nisms. First, an abstract model of spike timing lobe (Strauch et al. 2012), which, given that over-
and oscillations in the honeybee antennal lobe lap predicts perceptual similarity, should result in
explained how neuronal synchronization patterns poorer discrimination. Computational modeling
could underlie seemingly paradoxical behavioral resolved this behavioral conundrum by combin-
generalization effects observed across different ing the findings of these two lines of research
odor concentrations (Linster and Cleland 2001, (Linster and Cleland 2001; Cleland and Linster
2002). Second, a large-scale, reduced model of 2002). The model exhibited oscillations and
olfactory bulb outlined a feedback normalization spike synchronization patterns similar to those
network enabling concentration invariance in recorded in the antennal lobe and demonstrated
bulbar output, consistent with behavioral perfor- that while higher-intensity odor inputs would
mance, and showed that network properties in the evoke broader neural activity overall, an increas-
bulb appear to be optimized for this function ingly narrow subpopulation of neurons within
(Cleland et al. 2007). Third, a moderately this broad ensemble would be increasingly
detailed computational model of cholinergic strongly synchronized. Synchrony-sensitive
modulation in the olfactory bulb glomerular follower neurons and plasticity processes
layer predicted coordinated network-level and responding selectively to these highly synchro-
behavioral outcomes from the properties and nous inputs would therefore interpret higher con-
localization of nicotinic cholinergic receptors in centration inputs as more discriminable from one
the olfactory bulb (Mandairon et al. 2006a). another, matching behavioral observations, while
blocking these synchronization processes would
Example 1: Odor Concentration and Spike impair fine odor discrimination. In contrast,
Timing in the Insect Antennal Lobe Honeybee lower concentration inputs would evoke weaker
olfaction is an important model system for olfac- synchrony among antennal lobe projection neu-
tory learning due to its well-developed and effi- rons and generate less activity or plasticity in
cient behavioral conditioning paradigms and the follower cells (Fig. 1). This work illustrated
relatively well-understood neural circuits how computational modeling can fuse separate
involved in odor learning. The behavioral rele- datasets derived from a common system to
vance of spike timing regulation in olfaction was explain seemingly unrelated results, constructing
first demonstrated in the honeybee antennal lobe, a common theoretical basis upon which subse-
in which it was demonstrated that stimulus- quent experiments can be based. Since this time,
evoked synchronous spiking was necessary for the study of spike synchronization processes
fine odor discrimination. Specifically, in within olfactory coding research has grown
A Projection Neuron (PN)

Local Interneuron (LN)
Inhibitory synapse
Associative Neuron (AN)
PN
PN Bi
Bii
O1 O1
O2 O2
Ci Cii
low high D 1.2
1.0
Generalization
0.8
0.6 Low odor concentration
0.4
0.2
High odor concentration
0
Cond Test
500 ms Odor Odor
Computational Modeling of Olfactory Behavior, a glomerulus receiving input from sensory neurons
Fig. 1 Spike synchronization and odor perception in the expressing a shared olfactory receptor; darker colors sym-
honeybee antennal lobe. A. Network diagram of the hon- bolize stronger activation, and striped glomeruli indicate
eybee antennal lobe model. The network was created glomeruli substantially activated by both inputs. Glomer-
based on known neuron types and was capable of gener- uli are mapped onto a hypothetical one-dimensional cir-
ating odor-evoked, GABAergic neuron-dependent field cular axis of receptive field similarity (Cleland and Linster
potential oscillations that shaped spike timing as 2002). O1 and O2 denote the centers of the glomerular
described experimentally. Briefly, input from olfactory patterns activated by two different odors. C. Increased
sensory neurons (top) directly activates projection neu- odor concentration also generates stronger oscillations in
rons (PNs) as well as inhibitory local interneurons (LNs). activated glomeruli, resulting in greater synchrony among
Local interneurons are reciprocally connected with each a narrower population of the most strongly activated pro-
other in a feedback network which generates synchronous jection neurons. D. Increased synchrony leads to reduced
oscillations when activated. Local interneurons also generalization between two odors presented at higher
inhibit projection neurons and are capable of synchroniz- concentrations than when presented at lower concentra-
ing projection neuron spikes with respect to these network tions, despite their larger overlap at the input to the anten-
oscillations. Because local interneurons in the model have nal lobe. Odor conditioning was simulated by presenting
higher activation thresholds than projection neurons, a conditioning odor stimulus to the model at a low (25 %
highly activated glomeruli oscillate more strongly, and of maximum activation) or high (100 % of maximum
the corresponding projection neurons are most strongly activation) intensity and training the network on the
synchronized. The associative neuron (AN), odor. Subsequently, the network was presented with
corresponding to the reward-activated neuron VUMmx1 a novel odor, and the degree of overlap between the
in honeybees, is sensitive to synchronized spike inputs. B. representations of the conditioned and novel odors was
In the model, increasing odor concentration at the input of calculated (generalization). The graph shows the degree of
the antennal lobe generates more widespread activation generalization to the novel test odor after conditioning at
patterns and hence increased overlap among odor repre- low and high stimulus intensities (Figures adapted from
sentations (Bi designates a lower concentration odorant, Cleland and Linster (2002))
Bii a higher concentration odorant). Each circle represents
substantially, particularly with the elaboration of global feedback normalization at minimal meta-
spike timing-dependent synaptic plasticity algo- bolic cost (as predicted by the simulations)
rithms in the brain (Song et al. 2000; Gao and corresponded closely to that determined experi-
Strowbridge 2009; Linster and Cleland 2010). mentally (Cleland et al. 2007). These simulations
showed that a partially localized network of micro-
Example 2: The Problem of Concentration circuits could underlie a nonlocalized, uniform
Invariance in the Olfactory Bulb Odorants elicit effect, normalizing odor responses in order to ren- C
widely distributed patterns of activity across the der them more concentration invariant. In this
olfactory bulb input layer, as evidenced by 2- example, behavioral experiments were necessary
deoxyglucose activity mapping (Johnson and to validate the basis for the computational model
Leon 2007), intrinsic imaging (Meister and by demonstrating that global normalization cor-
Bonhoeffer 2001), and other imaging techniques. rectly predicted perceptual similarities. Interest-
These activation patterns are substantially altered ingly, in subsequent experimental studies, the role
when odor concentration is changed (Johnson of these laterally projecting interneurons in broadly
et al. 1999; Johnson and Leon 2000; Meister inhibiting mitral cells across the olfactory bulb was
and Bonhoeffer 2001), to the extent that the pat- confirmed (Marbach and Albeanu 2011), even
tern evoked by a given odor may more closely though the interneurons themselves since have
resemble that of a different odor than that of itself been shown to be GABAergic and dopaminergic
when presented at a different concentration rather than glutamatergic as was initially believed
(Cleland et al. 2007). Despite this fact, animals (Kiyokage et al. 2010). This provides an excellent
can identify a given odor over a reasonable con- example for how a model can serve its purpose,
centration range, and – critically – neural being correct at a certain level and enabling the
response patterns at the output of the olfactory coordinated interpretation of diverse datasets,
bulb are less affected by odor concentration than while also becoming outdated by experimental
are those measured at the input. Interestingly, progress and requiring update or replacement.
simple normalization (z-scoring) of these odor-
evoked patterns rendered them reasonably con- Example 3: Effects of Nicotinic Receptor Activa-
sistent across concentrations (Johnson et al. 1999; tion on Odor Discrimination: Behavioral Exper-
Johnson and Leon 2000), and behavioral tests iments and Biophysical Modeling The cellular
measuring the pairwise perceptual similarities of effects of nicotinic cholinergic receptor activa-
odorants showed that perception was better tion in the olfactory bulb are relatively well
predicted by normalized activity patterns than understood (Castillo et al. 1999; Pressler
by non-normalized activity (Cleland et al. 2007) et al. 2007). Nicotinic cholinergic activation
(Fig. 2). These results suggested that these odor- opens cation currents both in mitral cells, exciting
specific patterns are actively normalized within them and increasing their odor-evoked activation
olfactory bulb circuitry. Computational modeling levels, and in GABAergic periglomerular cells,
based on these data demonstrated how increasing their inhibition of mitral cell primary
established neural circuits in the olfactory bulb dendrites. Hence, nicotinic receptor activation
glomerular layer could perform such simultaneously excites and (indirectly) inhibits
a normalization function. Specifically, a laterally mitral cells in the bulb – two effects that initially
interconnected network of excitatory interneu- appear to cancel one another out. However, com-
rons, synapsing onto local inhibitory interneurons putational modeling of these two neuron types
and with a broad distribution of lateral projection embedded within their glomerular microcircuit
distances estimated from anatomical data showed that rather than canceling each other
(Aungst et al. 2003), was shown capable of out, these two effects complement one another
performing feedback normalization uniformly (Linster and Hasselmo 1997; Linster and
across the bulb; moreover, the optimal connec- Cleland 2002). Specifically, while the increased
tivity and spread of this network to achieve inhibition on mitral cell primary dendrites
a b 90
Neural activity
%correct trials
Increasing odor 80
concentration
70
60
50
Odorant similarity 0.001 0.1 10
Odor concentration
Behavior
c
0.8
Dissimilarity index
raw data
0.6
0.4
normalized data
0.2
0
Odor A low − Odor A −
Odor A high Odor B
Computational Modeling of Olfactory Behavior, (c) In certain cases, the raw and normalized activation
Fig. 2 Normalization in the olfactory bulb glomerular patterns evoked by a given odorant at different concentra-
layer. (a) Schematic depiction of odor representations in tions are more different from one another than they are to
the olfactory bulb glomerular layer. Odorants are ordered the representation of an entirely different odorant. The
along a hypothetical axis of similarity onto which individ- graph depicts (black lines) indices of dissimilarity
ual glomeruli can be mapped. Note that this is purely for between glomerular activation patterns, calculated from
representational purposes and that such orderly mapping raw or normalized 2-deoxyglucose uptake data, computed
does not exist in the system. Glomerular activation between a single odor (A) presented at two concentrations
becomes stronger and spreads by activating additional and between that odor and a different odor (B), both
glomeruli as odor intensity increases; consequently, the presented at the same concentration (Cleland
overlap between two odor representations increases. (b) et al. 2007). For comparison, the graph also depicts (red
Despite the increase in overlap between odorants at higher line) the corresponding behavioral results (i.e., degree of
concentrations, rats and mice can differentiate odorants perceptual dissimilarity) in response to the same odor
better as concentration increases. The graph shows the pairs. The normalized activation patterns correctly predict
percent correct trials in a forced choice go/no-go task in behavioral perception whereas the raw data do not.
rats as a function of odor concentration (Wei et al. 2006).
sharpens mitral cell responses to odorants by nicotinic receptors are activated, separating the
suppressing weak responses, the concomitant representations of highly similar odors and
excitation near the soma maintains or enhances predicting that they would become easier to dis-
the strength of response within this narrower criminate. We tested this prediction behaviorally
receptive field. As a consequence, receptive fields and found that, indeed, the blockade of nicotinic
are sharper but stronger (Mandairon et al. 2006a). cholinergic receptors in the olfactory bulb
When comparing ensemble responses to reduced rats’ capacity to differentiate between
multiple odorants in this model, the overlap in perceptually similar odorants, whereas it did
mitral cell activation patterns is reduced when not affect the perception of chemically and
b
Activation
OSN
a
OSN PG
C
Mi
PG
Activation
OSN
PG
Mi
Mi
c
3
Discrimination index
0
Nicotinic Control Enhanced
receptors ACh release
blocked
Computational Modeling of Olfactory Behavior, panel, Activity profiles in the unmodulated state. PG inhi-
Fig. 3 Nicotinic cholinergic neuromodulation of odor bition inhibits the weakly excited edges of the OSN odor
discrimination in the olfactory bulb glomerular layer. (a) representation, sharpening the resulting mitral cell repre-
Schematic of a simplified glomerular microcircuit anno- sentation (Cleland and Sethupathy 2006; Cleland and
tated with relative spike rates to indicate levels of activa- Linster 2012). Bottom panel, Effects of nicotinic cholin-
tion. Olfactory bulb output neurons (mitral cells; Mi) and ergic neuromodulation. Nicotinic modulation increases
GABAergic interneurons (PG) receive direct afferent PG cell activity, inhibiting mitral cells more strongly and
input from sensory neurons (OSNs). PG cells directly thereby sharpening their tuning curves (solid red line).
inhibit mitral cells and suppress their responses to weak Concurrently, the nicotinic activation of mitral cells
inputs (left panel), whereas strongly activated mitral cells amplifies their responses to odor inputs (dotted red line).
overcome this inhibition (right panel). Both PG and mitral (c) Behavioral experiments demonstrate that animals’ dis-
cells express nicotinic cholinergic receptors and are fur- crimination between similar odors is reduced when nico-
ther activated by cholinergic neuromodulation. (b) Effects tinic receptors are blocked and increased when
of nicotinic receptor activation. Graphs depict the tuning cholinergic inputs are potentiated. The ordinate depicts
curves of neuron types to a range of similar odorants; the an index of discrimination averaged among pairs of sim-
abscissa depicts a hypothetical axis of odor similarity. Top ilar odorants (Chaudhury et al. 2009)
perceptually dissimilar odor pairs. Moreover, in the olfactory bulb glomerular layer. Front Integr
enhancing cholinergic modulation during behav- Neurosci 6:5
Cleland TA, Sethupathy P (2006) Non-topographical con-
ior improved rats’ ability to differentiate between trast enhancement in the olfactory bulb. BMC
perceptually similar odorants (Mandairon Neurosci 7:7
et al. 2006a; Chaudhury et al. 2009) (Fig. 3). In Cleland TA, Morse A, Yue EL, Linster C (2002) Behav-
this study, the known cellular effects of nicotinic ioral models of odor similarity. Behav Neurosci
116:222–231
receptor activation were sufficient to predict Cleland TA, Johnson BA, Leon M, Linster C (2007) Rela-
perceptual effects once constructed into an appro- tional representation in the olfactory system. Proc Natl
priate network model, suggesting that the com- Acad Sci USA 104:1953–1958
putational model captured the essence of this Cleland TA, Chen SY, Hozer KW, Ukatu HN, Wong KJ,
Zheng F (2011) Sequential mechanisms underlying
neuromodulatory process. concentration invariance in biological olfaction.
Front Neuroeng 4:21
de Almeida L, Idiart M, Linster C (2013) A model of
Conclusion cholinergic modulation in olfactory bulb and piriform
cortex. J Neurophysiol 109:1360–1377
Devore S, Linster C (2012) Noradrenergic and cholinergic
Computational modeling of the neurons and net- modulation of olfactory bulb sensory processing. Front
works of the brain provides essential tools Behav Neurosci 6:52
and frameworks for understanding the neural Devore S, Manella LC, Linster C (2012) Blocking mus-
carinic receptors in the olfactory bulb impairs perfor-
mechanisms underlying animal behavior. By mance on an olfactory short-term memory task. Front
juxtaposing diverse experimental results within Behav Neurosci 6:59
appropriate theoretical frameworks, constructing Escanilla O, Yuhas C, Marzan D, Linster C (2009) Dopa-
functional scenarios, and assessing the likelihood minergic modulation of olfactory bulb processing
affects odor discrimination learning in rats. Behav
and consequences of each, modeling facilitates Neurosci 123:828–833
the theoretical analysis of complex datasets and Gao Y, Strowbridge BW (2009) Long-term plasticity of
the construction of testable hypotheses, providing excitatory inputs to granule cells in the rat olfactory
an essential link between physiology and bulb. Nat Neurosci 12:731–733
Hasselmo ME, Linster C, Patil M, Ma D, Cekic M (1997)
behavior. Noradrenergic suppression of synaptic transmission
may influence cortical signal-to-noise ratio.
References Johnson BA, Leon M (2000) Modular representations of
odorants in the glomerular layer of the rat olfactory
Aungst JL, Heyward PM, Puche AC, Karnup SV, bulb and the effects of stimulus concentration. J Comp
Hayar A, Szabo G, Shipley MT (2003) Centre- Neurol 422:496–509
surround inhibition among olfactory bulb glomeruli. Johnson BA, Leon M (2007) Chemotopic odorant coding in
Nature 426:623–629 a mammalian olfactory system. J Comp Neurol 503:1–34
Bhagavan S, Smith BH (1997) Olfactory conditioning in Johnson BA, Woo CC, Hingco EE, Pham KL, Leon
the honey bee, Apis mellifera: effects of odor intensity. M (1999) Multidimensional chemotopic responses to
Physiol Behav 61:107–117 n-aliphatic acid odorants in the rat olfactory bulb.
Castillo PE, Carleton A, Vincent JD, Lledo PM (1999) J Comp Neurol 409:529–548
Multiple and opposing roles of cholinergic transmission Kiyokage E, Pan YZ, Shao Z, Kobayashi K, Szabo G,
in the main olfactory bulb. J Neurosci 19:9180–9191 Yanagawa Y, Obata K, Okano H, Toida K, Puche AC,
Chaudhury D, Escanilla O, Linster C (2009) Bulbar ace- Shipley MT (2010) Molecular identity of periglomerular
tylcholine enhances neural and perceptual odor dis- and short axon cells. J Neurosci 30:1185–1196
crimination. J Neurosci 29:52–60 Li G, Cleland TA (2013) A two-layer biophysical model
Cleland TA, Linster C (1999) Concentration tuning medi- of cholinergic neuromodulation in olfactory bulb.
ated by spare receptor capacity in olfactory sensory neu- J Neurosci 33:3037–3058
rons: a theoretical study. Neural Comput 11:1673–1690 Linster C, Cleland TA (2001) How spike synchronization
Cleland TA, Linster C (2002) How synchronization prop- among olfactory neurons can contribute to sensory
erties among second-order sensory neurons can medi- discrimination. J Comput Neurosci 10:187–193
ate stimulus salience. Behav Neurosci 116:212–221 Linster C, Cleland TA (2002) Cholinergic modulation of
Cleland TA, Linster C (2012) On-center/inhibitory- sensory representations in the olfactory bulb. Neural
surround decorrelation via intraglomerular inhibition Netw 15:709–717
Computational Models of Deep Brain Stimulation (DBS) 729 C
Linster C, Cleland TA (2004) Configurational and ele- Stopfer M, Jayaraman V, Laurent G (2003) Intensity ver-
mental odor mixture perception can arise from local sus identity coding in an olfactory system. Neuron
inhibition. J Comput Neurosci 16:39–47 39:991–1004
Linster C, Cleland TA (2010) Decorrelation of odor rep- Strauch M, Ditzen M, Galizia CG (2012) Keeping their
resentations via spike timing-dependent plasticity. distance? Odor response patterns along the concentra-
Front Comput Neurosci 4:157 tion range. Front Syst Neurosci 6:71
Linster C, Gervais R (1996) Investigation of the role of Wei CJ, Linster C, Cleland TA (2006) Dopamine D
interneurons and their modulation by centrifugal fibers (2) receptor activation modulates perceived odor
in a neural model of the olfactory bulb. J Comput intensity. Behav Neurosci 120:393–400
C
Linster C, Hasselmo M (1997) Modulation of inhibition in
a model of olfactory bulb reduces overlap in the neural
representation of olfactory stimuli. Behav Brain Res Computational Models of Deep Brain
84:117–127 Stimulation (DBS)
Linster C, Kerszberg M, Masson C (1994) How
neurons may compute: the case of insect sexual
pheromone discrimination. J Comput Neurosci Jonathan Rubin1 and Cameron McIntyre2
1
1:231–238 Department of Mathematics and Center for the
Linster C, Maloney M, Patil M, Hasselmo ME Neural Basis of Cognition, University of
(2003) Enhanced cholinergic suppression of previ-
ously strengthened synapses enables the formation of Pittsburgh, Pittsburgh, PA, USA
2
self-organized representations in olfactory cortex. Departments of Biomedical Engineering,
Neurobiol Learn Mem 80:302–314 Neurology, and Neurosurgery, Case Western
Linster C, Henry L, Kadohisa M, Wilson DA (2007) Syn- Reserve University School of Medicine,
aptic adaptation and odor-background segmentation.
Neurobiol Learn Mem 87:352–360 Cleveland, OH, USA
Linster C, Menon AV, Singh CY, Wilson DA (2009) Odor-
specific habituation arises from interaction of afferent
synaptic adaptation and intrinsic synaptic potentiation Definition
in olfactory cortex. Learn Mem 16:452–459
Linster C, Nai Q, Ennis M (2011) Nonlinear effects of
noradrenergic modulation of olfactory bulb function in DBS is a procedure in which a device is
adult rodents. J Neurophysiol 105:1432–1443 implanted within the brain and used to deliver
Mandairon N, Ferretti CJ, Stack CM, Rubin DB, pulses of electrical activity to the surrounding
Cleland TA, Linster C (2006a) Cholinergic modula-
tion in the olfactory bulb influences spontaneous olfac- tissue. While parts of the device may extend
tory discrimination in adult rats. Eur J Neurosci elsewhere, the electrical stimulation is delivered
24:3234–3244 in the interior of the brain, rather than outside of it
Mandairon N, Stack C, Kiselycznyk C, Linster C (2006b) or on its surface. Computational models of DBS
Broad activation of the olfactory bulb produces long-
lasting changes in odor perception. Proc Natl Acad Sci are systems of equations representing some
USA 103:13543–13548 aspects of the DBS process, such as the electric
Marbach F, Albeanu DF (2011) Photostimulation of short field generated by a certain form of DBS or the
axon cells reveals widespread inhibition in the mouse predicted voltage responses of neurons impacted
olfactory bulb. Soc Neurosci Abstr 475:19
Meister M, Bonhoeffer T (2001) Tuning and topography by this field, for which a computer can be used to
in an odor map on the rat olfactory bulb. J Neurosci obtain high-accuracy approximate solutions.
21:1351–1360 These predicted neural responses to DBS can
Pressler RT, Inoue T, Strowbridge BW (2007) Muscarinic then be introduced into network models of brain
receptor activation modulates granule cell excitability
and potentiates inhibition onto mitral cells in the rat circuits, enabling theoretical investigation of the
olfactory bulb. J Neurosci 27:10969–10981 effects of DBS on brain activity.
Song S, Miller KD, Abbott LF (2000) Competitive
Hebbian learning through spike-timing-dependent
synaptic plasticity. Nat Neurosci 3:919–926 Detailed Description
Stopfer M, Bhagavan S, Smith BH, Laurent G (1997)
Impaired odour discrimination on desynchronization
of odour-encoding neural assemblies. Nature While DBS is being used or tested for use in
390:70–74 treating a variety of conditions, computational
C 730 Computational Models of Deep Brain Stimulation (DBS)
models of DBS to date have primarily focused on from patient-specific DBS models and the
DBS for Parkinson’s disease (PD) and related corresponding electric field generated by human
states (collectively known as parkinsonism). DBS electrodes. The last decade has seen these
Therefore, this entry will focus on DBS models models evolve from point source electrodes in an
related to parkinsonism. infinite homogeneous medium, to clinical elec-
trodes in the human brain, to detailed representa-
DBS for Parkinsonism tions of the electrode-tissue interface (Chaturvedi
In DBS, a set of electrodes is surgically implanted et al. 2010; Yousif and Liu 2009). The electric
into a brain area selected based on patient symp- field predictions from modern DBS models have
toms. The electrodes are linked to a subcutaneous been validated against experimental measure-
pulse generator outside of the brain, which causes ments in the brains of nonhuman primates
electrical stimulation to be delivered through the (Miocinovic et al. 2009) and electrophysiological
electrodes with a prescribed frequency, ampli- measurements in human patients (Chaturvedi
tude, and timing. These stimulation parameters et al. 2010). Such models have been used to
can be tuned externally to try to optimize the design novel DBS electrodes (Martens
therapeutic results of DBS. et al. 2011; Vasques et al. 2010) and evolve the
The mechanisms underlying the therapeutic concept of current steering (Butson and McIntyre
efficacy of chronic high-frequency (~100 Hz) 2008).
brain stimulation remain mysterious and contro- Patient-specific DBS models have had their
versial. While earlier theories about DBS focused greatest clinical impact in the analysis of the
on local effects on neuronal firing or ion channel target of the stimulation. In the case of
activation, recent research efforts have reasoned subthalamic nucleus (STN) DBS for PD,
that since disorders treated with DBS are funda- patient-specific models have demonstrated that
mentally disorders of a specific brain network, activation of the white matter dorsal to the STN
the benefits of DBS are likely to derive from is most associated with therapeutic benefit (Maks
circuit-level effects. The basic working hypothe- et al. 2009; Butson et al. 2011), supporting early
sis is that DBS interacts with the diseased net- clinical hypotheses (Voges et al. 2002). Further,
work to eliminate or subdue the underlying when this region is explicitly targeted via surgical
pathological neural activity (Montgomery and electrode placement (Plaha et al. 2006) or stimu-
Baker 2000; Rubin and Terman 2004; Garcia lation parameter selection (Frankemolle
et al. 2005; Wichmann and DeLong 2006; et al. 2010), outcomes improve relative to stimu-
McIntyre and Hahn 2010; Rubin et al. 2012). lation concentrated on the STN itself.
Computational models have played an important
role in the evolution of this hypothesis, Computational Models Focusing on the
suggesting that changes in the underlying dynam- Network Response to DBS for Parkinsonism
ics of the stimulated brain networks may repre- There are several different characteristics that
sent the core mechanisms of DBS and that those can be used to classify those existing models
basic dynamical changes can be achieved via that focus on the network response to DBS for
activation, inhibition, or lesion. parkinsonism. Computational models are gener-
ally motivated by an underlying hypothesis for
Computational Models of the Electric Field the central mechanism of DBS efficacy, although
Associated with DBS for Parkinsonism such hypotheses need not be mutually exclusive.
Computational modeling of DBS has become Depending on which hypothesis is under study,
established as a useful tool for investigating ther- models may vary in their mathematical formula-
apeutic mechanisms and designing techniques to tions, level of abstraction of model neurons,
optimize clinical application of DBS technology selection of brain structures modeled, and repre-
(McIntyre et al. 2007). A prime example comes sentation of DBS inputs.
Computational Models of Deep Brain Stimulation (DBS) 731 C
For example, a common general hypothesis is models. These currents may be delivered at mul-
that DBS works by eliminating pathological syn- tiple sites, with various timings, or may be
chrony in basal ganglia structures. This idea can be derived from computationally generated feed-
explored using model equations representing back signals (Hauptmann et al. 2005). Alterna-
many individual neurons or using models in tively, a more involved field model of DBS may
which variables correspond to entire populations be coupled to a detailed neuronal model for con-
of neurons (e.g., one equation for each basal sideration of spatial effects (Yousif et al. 2012). C
ganglia area). In the former approach, which Some computational modeling studies have
allows the timing of individual neuron spikes to circumvented the need for an explicit represen-
be considered, neurons have been represented tation of DBS by incorporating experimentally
either as abstract oscillators or as more biophysical recorded data from basal ganglia structures, col-
units. In the latter, model equations can track phase lected with and without DBS, as inputs into
density, with synchronization corresponding to a computational model (Guo et al. 2008; Dorval
a peaked phase distribution (Tass 2002), or mean et al. 2008; Meijer et al. 2011; Cleary
voltage, which only allows the consideration of et al. 2013). Even without experimental record-
synchrony between populations, not within ings or modeling of DBS currents, DBS effects
a population. Similarly, both of these frameworks may also be included in a model through a
can be utilized to consider alternative hypotheses phase-resetting curve, which is a mathematical
not emphasizing synchrony. They have been used, representation of how neurons respond to gen-
for example, to consider DBS-induced changes in eral inputs (Wilson et al. 2011), or through
burstiness (Rubin and Terman 2004; Modolo altered probability of spike generation in sto-
et al. 2008; Hahn and McIntyre 2010) or firing chastic process models of neuronal circuits
rates (Moroney et al. 2008; Humphries and Gurney (Rosenbaum et al. 2012).
2012) within various basal ganglia nuclei. In par-
ticular, models have been used to consider the References
specific hypothesis that DBS can replace patho-
logically bursty and irregular firing with more Butson CR, McIntyre CC (2008) Current steering to con-
regularized outputs. Some of these models trol the volume of tissue activated during deep brain
include additional brain structures such as tha- stimulation. Brain Stimul 1(1):7–15
Butson CR, Cooper SE, Henderson JM, Wolgamuth B,
lamic nuclei downstream from the basal ganglia, McIntyre CC (2011) Probabilistic analysis of activa-
so that predictions can be made about the impact tion volumes generated during deep brain stimulation.
that altered basal ganglia firing patterns have on Neuroimage 54(3):2096–2104
the activity outside of the basal ganglia (Rubin Chaturvedi A, Butson CR, Lempka SF, Cooper SE,
McIntyre CC (2010) Patient-specific models of deep
and Terman 2004), have considered effects of brain stimulation: influence of field model complexity
DBS applied directly to thalamic targets (Grill on neural activation predictions. Brain Stimul
et al. 2004), or have included some form of 3(2):65–67
cortical inputs (Hahn and McIntyre 2010). Cleary DR, Raslan AM, Rubin JE, Bahgat D,
Viswanathan A, Heinricher MM, Burchiel KJ
Finally, neurons need not be modeled explicitly (2013) Deep brain stimulation entrains local neuronal
at all; instead, they can be represented as sto- firing in human globus pallidus internus.
chastic processes yielding discrete spiking J Neurophysiol 109(4):978–987
events, which may be useful for considering Dorval AD, Russo GS, Hashimoto T, Xu W, Grill WM,
Vitek JL (2008) Deep brain stimulation reduces neu-
hypotheses about differences in correlations, ronal entropy in the MPTP-primate model of
information transfer, or signal propagation with Parkinson’s disease. J Neurophysiol
and without DBS (Reitsma et al. 2011; 100(5):2807–2818
Rosenbaum et al. 2014). Frankemolle AM, Wu J, Noecker AM, Voelcker-Rehage-
C, Ho JC, Vitek JL, McIntyre CC, Alberts JL
DBS inputs themselves are represented simply (2010) Reversing cognitive-motor impairments in
as applied currents within many computational Parkinson’s disease patients using a computational
C 732 Computational Models of Deep Brain Stimulation (DBS)
modelling approach to deep brain stimulation pro- Moroney R, Heida C, Geelen J (2008) Increased
gramming. Brain 133(Pt 3):746–761 bradykinesia in Parkinson’s disease with increased
Garcia L, D’Alessandro G, Bioulac B, Hammond movement complexity: elbow flexion-extension
C (2005) High-frequency stimulation in Parkinson’s movements. J Comput Neurosci 25:501–519
disease: more or less? Trends Neurosci 28:209–216 Plaha P, Ben-Shlomo Y, Patel NK, Gill SS (2006) Stimula-
Grill WM, Snyder AN, Miocinovic S (2004) Deep tion of the caudal zona incerta is superior to stimulation
brain stimulation creates an informational lesion of the subthalamic nucleus in improving contralateral
of the stimulated nucleus. Neuroreport parkinsonism. Brain 129(Pt 7):1732–1747
15(7):1137–1140 Reitsma P, Doiron B, Rubin J (2011) Correlation transfer
Guo Y, Rubin JE, McIntyre CC, Vitek JL, Terman from basal ganglia to thalamus in Parkinson’s disease.
D (2008) Thalamocortical relay fidelity varies across Front Comput Neurosci 5. doi: 10.3389/
subthalamic nucleus deep brain stimulation protocols fncom.2011.00058
in a data-driven computational model. J Neurophysiol Rosenbaum R, Rubin J, Doiron B (2012) Short term syn-
99(3):1477–1492 aptic depression imposes a frequency dependent filter
Hahn PJ, McIntyre CC (2010) Modeling shifts in the rate on synaptic information transfer. PLoS Comp Biol 8:
and pattern of subthalamopallidal network activity e1002557
during deep brain stimulation. J Comput Neurosci Rosenbaum R, Zimnik A, Zheng F, Turner RS, Alzheimer
28(3):425–441 C, Doiron B, Rubin JE (2014) Axonal and synaptic
Hauptmann C, Popovych O, Tass PA (2005) Effectively failure suppress the transmission of firing rate oscilla-
desynchronizing deep brain stimulation based on tions, synchrony, and information during high
a coordinated delayed feedback stimulation via several frequency deep brain stimulation. Neurobiol Dis
sites: a computational study. Biol Cybern 62:86–99
93(6):463–470 Rubin JE, Terman D (2004) High frequency stimulation of
Humphries MD, Gurney K (2012) Network effects of the subthalamic nucleus eliminates pathological tha-
subthalamic deep brain stimulation drive a unique lamic rhythmicity in a computational model. J Comput
mixture of responses in basal ganglia output. Eur Neurosci 16(3):211–235
J Neurosci 36(2):2240–2251 Rubin JE, McIntyre CC, Turner RS, Wichmann T (2012)
Maks CB, Butson CR, Walter BL, Vitek JL, McIntyre CC Basal ganglia activity patterns in parkinsonism and
(2009) Deep brain stimulation activation volumes and computational modeling of their downstream effects.
their association with neurophysiological mapping and Eur J Neurosci 36(2):2213–2228
therapeutic outcomes. J Neurol Neurosurg Psychiatry Tass PA (2002) Desynchronization of brain rhythms with
80(6):659–666 soft phase-resetting techniques. Biol Cybern
Martens HC, Toader E, Decre MM, Anderson DJ, 87:102–115
Vetter R, Kipke DR, Baker KB, Johnson MD, Vitek Vasques X, Cif L, Mennessier G, Coubes P (2010) A
JL (2011) Spatial steering of deep brain stimulation target-specific electrode and lead design for internal
volumes using a novel lead design. Clin Neurophysiol globus pallidus deep brain stimulation. Stereotact
122(3):558–566 Funct Neurosurg 88(3):129–137
McIntyre CC, Hahn PJ (2010) Network perspectives on Voges J, Volkmann J, Allert N, Lehrke R, Koulousakis A,
the mechanisms of deep brain stimulation. Neurobiol Freund HJ, Strum V (2002) Bilateral high-frequency
Dis 38(3):329–337 stimulation in the subthalamic nucleus for the treat-
McIntyre CC, Miocinovic S, Butson CR (2007) Computa- ment of Parkinson’s disease: correlation of therapeutic
tional analysis of deep brain stimulation. Expert Rev effect with anatomical electrode position. J Neurosurg
Med Devices 4(5):615–622 96(2):269–279
Meijer HGE, Krupa M, Cagnan H, Lourens MAJ, Heida T, Wichmann T, DeLong MR (2006) Deep brain stimulation
Martens HCF, Van Gils SA (2011) From Parkinsonian for neurologic and neuropsychiatric disorders. Neuron
thalamic activity to restoring thalamic relay using deep 52:197–204
brain stimulation: new insights from computational Wilson CJ, Bryce Beverlin II, Netoff T (2011) Chaotic
modeling. J Neural Eng 8(6):066005 desynchronization as the therapeutic mechanism of
Miocinovic S, Lempka SF, Russo GS, Maks CB, Butson deep brain stimulation. Front Syst Neurosci
CR, Sakaie KE, Vitek JL, McIntyre CC (2009) Exper- 5. doi:10.3389/fnsys.2011.00050
imental and theoretical characterization of the voltage Yousif N, Liu X (2009) Investigating the depth
distribution generated by deep brain stimulation. Exp electrode-brain interface in deep brain stimulation
Neurol 216(1):166–176 using finite element models with graded complexity
Modolo J, Henry J, Beuter A (2008) Dynamics of the in structure and solution. J Neurosci Methods 184
subthalamo-pallidal complex during deep brain stim- (1):142–51
ulation in Parkinson’s disease. J Biol Phys 34:351–366 Yousif N, Borisyuk R, Pavese N, Nandi D, Bain P (2012)
Montgomery E Jr, Baker K (2000) Mechanism of deep Spatiotemporal visualization of deep brain stimula-
brain stimulation and future technical developments. tion‐induced effects in the subthalamic nucleus. Eur
Neurol Res 22:259–266 J Neurosci 36(2):2252–2259
Computational Models of Mammalian Respiratory CPG 733 C
provide mechanistic explanations for the activity
Computational Models of patterns and operation of identified or postulated
Mammalian Respiratory CPG key components of the brainstem respiratory CPG
and associated motor output patterns during the
Jeffrey C. Smith normal respiratory cycle. The basic activity pat-
Cellular and Systems Neurobiology Section, terns of brainstem–spinal neurons and efferent
NINDS/NIH, Bethesda, MD, USA motor behaviors of the CPG to be explained have C
been established experimentally from electrophys-
iological recording studies (see Lindsey et al. 2012
Definition for a comprehensive review). Essential neurophys-
iological features as they are presently understood
Rhythmic breathing movements in mammals are that need to be taken into account in computational
produced by a central pattern generator (CPG), modeling analyses are summarized below.
consisting of specialized neuronal networks During normal breathing in mammals, the
located in the brainstem that are capable of rhythmic motor pattern consists of three neural
endogenously producing patterned rhythmic activity epochs or “phases”: inspiration (I, phase
activity. This activity emerges from the intrinsic 1), post-inspiration (post-I, phase 2), and the later
biophysical properties and synaptic interconnec- second stage (called E-2) of expiration constitut-
tions of spatially distributed neuron populations ing phase 3 of the respiratory cycle (Richter and
within the pontine–medullary circuits comprising Smith 2014) (Fig. 1). This three-phase motor
the CPG. These circuits are embedded in a larger pattern is phylogenetically conserved in all mam-
respiratory neural control system engaging vari- malian species from rodents to humans and
ous central nervous system (CNS) and peripheral reflects the coordinated activity of numerous
afferent inputs that regulate the neural activity populations of brainstem and spinal motoneurons
patterns including the oscillation period and that are driven and controlled by the output of the
amplitude of the output rhythmic motor activity brainstem CPG circuits. At the brainstem level,
to adjust it to the internal and/or external the cranial motoneurons control musculature
environment for sensorimotor integration and of the upper airways (e.g., hypoglossal,
physiological homeostasis of O2 and CO2 in the glossopharyngeal, laryngeal muscles), and moto-
body. Computational models of the brainstem neurons at the spinal cord level innervate muscles
respiratory CPG that incorporate various levels of the respiratory pump (diaphragm, thoracic,
of biological and mathematical detail have been abdominal musculature). This neural pattern is
in development for several decades with the also reflected in the activity profiles of the ante-
objective of explaining the dynamical operation cedent interneuron populations (Fig. 1) in spe-
of the GPG neurons and circuits, particularly to cific anatomically defined regions of the
delineate mechanisms of respiratory rhythm and medulla oblongata and pons generating the
pattern generation. Recent data-driven models three-phasic rhythmic drive. That is, there are
that have explored functional architecture of the distinct populations of neurons that are predom-
CPG circuits and advanced understanding of inantly active with some temporal dispersion dur-
these mechanisms are reviewed. ing the inspiratory phase or during the post-I or
later E-2 phase of expiration. This has been deter-
mined experimentally by extracellular recording
Detailed Description and mapping of single neuron or neuron popula-
tion activity, including with large arrays of
Respiratory Motor Patterns and CPG Activity: microelectrodes enabling simultaneous record-
Neurophysiology ings at multiple brainstem sites for temporal cor-
A major objective of the models described in this relation of activity, allowing inferences about
entry is to reproduce experimental data and network connectivity to guide computational
C 734 Computational Models of Mammalian Respiratory CPG
Computational Models of Mammalian Respiratory active inspiratory (I) and expiratory (E) neurons are typi-
CPG, Fig. 1 Patterns of neuronal population activity cally classified with nomenclature based on profiles of
from extracellular recordings in the medulla and motor their firing patterns (e.g., decrementing, augmenting spik-
output patterns recorded from spinal (PN) and cranial ing frequency) and their predominant activity phase. As
nerves (central vagus, cVN, and hypoglossal, HN) during depicted, pre-I/I neurons in the pre-BötC start firing before
the three-phase respiratory cycle. Neuronal population and continue into inspiration; early-I neurons have peak
recordings have been obtained in key structural–- spiking in early-I followed by a decrementing pattern;
functional compartments (BötC, pre-BötC, and rVRG, rVRG aug-I neurons have an augmenting (also called
refer to Fig. 2). Recordings were obtained from juvenile ramping) firing pattern; post-I neurons, a decrementing
rat brainstem–spinal cord preparations and are integrated firing pattern; and aug-E neurons, an augmenting firing
and averaged over multiple cycles to show relative timing pattern during E-2 (After Smith et al. 2009)
and activity profiles during the three-phase cycle. The
analyses of circuit interactions (Lindsey that control contraction of the diaphragm via the
et al. 2012). There have also been numerous phrenic nerve (PN) to inflate the lungs. Medullary
intracellular recording studies that reveal the inspiratory neurons also drive cranial hypoglos-
spiking activity and underlying synaptic inputs sal motoneurons (Fig. 2) to activate via hypoglos-
or biophysical properties of the different neuron sal nerves (HN) tongue protrussor muscles to
types, which are critical for conductance-based prevent obstruction of the oropharyngeal airway
modeling analyses of neuronal circuit activity during inspiration. Other cranial motoneurons
and neuronal interactions (Richter and Smith with projecting axons in branches of the vagus
2014). nerve innervate laryngeal muscles that control
Each phase of the respiratory cycle is func- abduction of the glottis and vocal cords to main-
tionally and behaviorally distinct. During inspi- tain upper airway patency. The post-I phase of
ration, medullary inspiratory premotoneurons motor output is evident in recordings from the
(i.e., bulbospinal neurons with direct monosyn- recurrent laryngeal or central vagus nerves (cVN)
aptic connections to spinal motoneurons, see (e.g., Richter and Smith 2014) and is essential for
Fig. 2) drive the activity of phrenic motoneurons controlling musculature regulating the caliber of
Computational Models of Mammalian Respiratory transmission circuits in the rVRG. BötC excitatory post-I
CPG, Fig. 2 Structural–functional compartments of the neurons generate the post-I phase activity transmitted to
brainstem respiratory CPG arranged from the caudal cranial vagal motoneurons (I/E Mns, brown), which also
medulla to the rostral pons and their excitatory/inhibitory receive inspiratory inputs from excitatory pre-BötC and
circuit components, viewed in a parasagittal plane through rVRG neurons. The pre-I/I excitatory neurons of the
the lateral brainstem. Each side of the brainstem contains pre-BötC are the excitatory “kernel” circuit neurons that
these structures with extensive bilateral interconnections project via premotor circuits to cranial motoneurons
(not shown). The major compartments are the caudal and including hypoglossal and vagal motoneurons (brown)
rostral ventral respiratory groups (cVRG, rVRG), and to rVRG excitatory aug-I bulbospinal premotor neu-
pre-BötC, BötC, RTN/pFRG, and pontine respiratory rons with projections to spinal cord inspiratory phrenic
group (PRG) in the dorsolateral pons. Locations of key motoneurons (IMn, brown). Excitatory bulbospinal aug-E
populations of medullary inspiratory (I), expiratory (E), expiratory neurons of cVRG and driven by BötC aug-E
and tonically active (tonic) interneurons involved in respi- excitatory neurons project to thoracic intercostal (IC) and
ratory rhythm and pattern generation are indicated. Com- abdominal (Abd) spinal expiratory motoneurons (EMns).
partmentalized circuit components of the pre-BötC and Circuit neurons within these compartments receive tonic
BötC are core CPG microcircuits incorporating the excit- and in some cases phasic/rhythmic excitatory drive from
atory (red) pre-I/I population in pre-BötC and three inhib- the pontine, RTN/pFRG, and raphé compartments, where
itory (blue) populations (early-I, post-I, and aug-E) the latter two drives are in part chemosensory signals
neurons within the pre-BötC–BötC that mutually inhibit dependent on the blood/brain CO2 level. Additional
each other. The BötC expiratory neurons are a primary descriptions of the roles of the different populations of
source of synaptic inhibition that controls alternation excitatory and inhibitory respiratory neurons are in the
between I and E phases, including in downstream main text (After Smith et al. 2009)
the upper airway during expiration as well as activity (Fig. 1), this phase is also critical for
integrating breathing with several other orofacial dynamically off-switching inspiration in the
motor behaviors including vocalization and core microcircuits. During the E-2 phase of
swallowing. During this phase the larynx is ini- expiration, the active medullary expiratory
tially partially adducted (closed), which stalls bulbospinal neurons drive spinal thoracic and
airflow to provide more time for gas exchange, lumbar motoneurons (Fig. 1) to produce low
and then slowly abducted (opened) to allow exha- levels of intercostal and abdominal muscle activ-
lation. In the premotor network producing this ity during normal (quiet) breathing. At higher
levels of respiratory motor activity, such as segregation of neuron types, although not exclu-
when levels of carbon dioxide (CO2) are elevated sive, suggests a specific compartmental organiza-
or levels of oxygen (O2) are reduced, forced or tion of mechanisms producing inspiratory and
active exhalation occurs during this phase, which expiratory activity and that these populations
engages new circuit elements (Abdala et al. 2009; form microcircuits that interact within and
Rubin et al. 2011). between these compartments. These microcir-
cuits, depicted in Figs. 2 and 3, are postulated to
CPG Network Architecture and Circuit represent key neural building blocks of the CPG
Components network (Smith et al. 2007, 2013).
Respiratory CPG circuits like all other GPG (and Experimental results also suggest that (1) the
CNS) circuits are constructed from core sets of BötC contains neuronal machinery essential for
interacting excitatory (glutamatergic) and inhibi- generating expiratory activity, including
tory (glycinergic and/or GABAergic) populations populations of inhibitory neurons active during
of interneurons whose activity is transmitted via the post-I and E-2 expiratory phases; (2) the
mono- and polysynaptic circuits to motoneurons. pre-BötC, a critical CPG component, has excit-
The critical populations and their circuit archi- atory circuits that constitute the kernel inspira-
tectures have not been completely defined from tory oscillator, these neurons have local intrinsic
experimental studies for any mammalian CPG, mechanisms for inspiratory rhythm generation,
and CPG models attempt to define hypothetical and there are also pre-BötC inhibitory interneu-
canonical circuit architectures or circuit “motifs” rons driven locally by the excitatory neurons and
(Stein et al. 1997; Shepard and Grillner 2010; are thus also active during inspiration; (3) the
Grillner and Jessell 2009) that can explain rVRG contains the main population of excitatory
the dynamical interactions and functional contri- premotor bulbospinal inspiratory neurons that are
butions of these excitatory and inhibitory driven by pre-BötC rhythmogenic excitatory neu-
populations to produce appropriately coordinated rons and project to phrenic motoneurons as well
circuit activity and motor output patterns. as cranial motoneurons so that activity of this
A central hypothesis in the respiratory neuro- population produces phrenic nerve (PN) inspira-
biology field is that these key populations of tory motor output and the inspiratory activity of
interneurons, which constitute the neural machin- cranial nerves (e.g., the laryngeal nerve innervat-
ery for generating the three phases of the respira- ing the upper airway); (4) generation of the
tory cycle, are bilaterally distributed in three normal three-phase respiratory pattern requires
anatomically defined and functionally distinct pontine circuits in the rostral brainstem, specifi-
compartments in the ventrolateral medulla cally excitatory inputs to more caudal compart-
(Fig. 2): the so-called Bötzinger complex ments especially the BötC to control generation
(BötC), pre-Bötzinger complex (pre-BötC), and of the post-I activity phase; and (5) other excit-
rostral ventral respiratory group (rVRG), which atory inputs such as from the retrotrapezoid
collectively are components of a more extensive nucleus or parafacial respiratory group
rostro-caudally distributed “ventral respiratory (RTN/pFRG) in the more rostral medulla, which
column” (VRC) on each side of the brainstem is a major source of chemosensory (CO2/pH and
(Smith et al. 2013; Fig. 2). Electrophysiological O2-related) afferent input for homeostatic respi-
recording studies have shown that major ratory regulation, adaptively adjust the core CPG
populations of inspiratory neurons are concen- circuit activity in response to various metabolic
trated in the pre-BötC (pre-I/I and early-I neu- demands. Other brainstem sources of afferent
rons, see Fig. 1 for nomenclature) and rVRG input are from the serotonergic raphé nuclei
(aug-I neurons), while populations of expiratory (also with CO2 chemosensory properties, Fig. 2)
neurons are in the BötC (post-I neurons, aug-E and the nuclei of the tractus solitarius (NTS),
neurons active during the E-2 phase) and more which also project to the RTN and pons convey-
caudally in the cVRG (aug-E). This spatial ing mechanosensory signals and input from
Computational Models of Mammalian Respiratory received tonic excitatory synaptic drive from the sources
CPG, Fig. 3 Circuit configuration of the conductance- depicted. Representative simulated integrated activities of
based neuronal population model, incorporating the the main populations of BötC, pre-BötC, and rVRG during
core mutual inhibitory ring-like structure within the three-phase respiratory cycle are illustrated at the
pre-BötC–BötC microcircuits and the inspiratory bottom. Integrated population activities are total number
pre-BötC excitatory kernel circuits composed of pre-I/I of spikes per sec/number of neurons (n = 50 neurons per
spiking neurons with INaP functioning as an important population). Detailed descriptions of the model compo-
conductance mechanism. Each neuronal component is nents and their interactions are in the main text (From
represented in the model by a population of neurons. In Smith et al. 2009)
the model simulations, all of these circuit components
peripheral chemoreceptors (sensing O2) (Smith inhibitory neuron populations in the pre-BötC
et al. 2013). active during inspiration, and (3) the excitatory
Based on this spatial and functional organiza- populations in the pre-BötC and rVRG that gen-
tion, the current neurophysiological concept is erate inspiration. The operation of this circuitry is
that the respiratory rhythm and normal three- controlled in part by other compartmentalized
phase inspiratory–expiratory pattern emerge network components located in the more rostral
from circuit-based dynamic interactions between medulla and pons that provide neural input
(1) the inhibitory neuronal populations in the “drives,” regulating dynamic behavior of the
BötC that are active during expiration, (2) the core CPG circuitry. The compartmental network
organization is thought to be a basic design prin- of CPG operation. This includes rhythmogenic
ciple allowing flexible multilevel control of dif- mechanisms embedded within the interacting
ferent circuit components for regulation of pontine–medullary circuits that generate differ-
inspiratory and expiratory phase activity (Smith ent experimentally observed motor patterns
et al. 2013), giving rise to a repertoire of underlying distinct (patho)physiological respira-
respiratory motor behaviors. The recent models tory behaviors (see Lindsey et al. 2012; Smith
described below attempt to computationally repet al. 2013 for recent reviews).
resent these processes.
Conductance-Based Computational Models
Recent Computational Models: Overview The conductance-based models developed by
A basic problem in understanding the operation Rybak et al. (2007), Smith et al. (2007), and
of the postulated core CPG circuits described Rybak and Smith (2009) are extensions of their
above is to define how the intrinsic biophysical/ earlier models that incorporate synaptically
electrophysiological properties of the constituent coupled populations of inspiratory and expiratory
neuron populations, and population connectivity/ spiking single-compartment neurons with bio-
synaptic interactions within specific circuit archi- physical properties and ion channel kinetics of
tectures, generate the spatiotemporal patterns of membrane currents/conductances modeled in the
neuronal activity required for the three-phase Hodgkin–Huxley style. These model formula-
organization of the respiratory motor pattern. tions incorporate experimentally characterized
Computational models of the brainstem respira- or postulated biophysical properties for respira-
tory CPG that incorporate various levels of bio- tory neurons. The model neurons are assembled
logical and mathematical detail have been in into individual populations for each functional
development for decades, exploring various type of inspiratory or expiratory neuron with het-
facets of the dynamical operation of the GPG erogeneity of neurons built into individual
neurons and circuits. For a comprehensive review populations by random distributions of some
including historical perspectives, see Lindsey neuronal biophysical parameters and simulation
et al. (2012). Two types of recent models are conditions. Excitatory or inhibitory synaptic
described below – conductance-based and interactions among populations are modeled by
activity-based models of CPG circuits. These incorporating postsynaptic conductances in tar-
models incorporate the most recent experimental get neurons that are activated by spiking of neu-
information on the structural–functional rons providing the activity-dependent synaptic
organization of pontine–medullary network inputs. The model circuitry, shown schematically
components (Fig. 2), as identified from neuroan- in Fig. 3, also includes the tonically spiking neu-
atomical reconstruction or electrophysiological rons representing those identified electrophysio-
connectivity studies, and key biophysical/electro- logically in the pons, RTN, and raphé nuclei that
physiological properties inferred from neuronal provide excitatory synaptic drive to the postu-
recording studies. The models represent at differ- lated microcircuits in the three major medullary
ent levels of complexity how the inhibitory and compartments (BötC, pre-BötC, and rVRG) as
excitatory network interactions and intrinsic bio- described below.
physical mechanisms within the core circuitry, The BötC compartment model microcircuit
operating under control by external excitatory contains two populations of rhythmically active
inputs such as from RTN/pFRG and the pons, inhibitory expiratory neurons, the post-
contribute to rhythmic pattern generation in dif- inspiratory (post-I) and the augmenting expira-
ferent network states. In addition to explaining tory (aug-E) neurons active during the E-2 phase.
circuit dynamics generating the normal three- These populations inhibit neuronal populations
phase respiratory pattern, these models have within the pre-BötC and rVRG and reciprocally
explored a number of other dynamical features inhibit each other (Fig. 2). The BötC
compartment contains an excitatory post-I popu- drive increases the inspiratory burst frequency
lation (post-I(e)) that contributes to the simulated over a large dynamic range (Purvis et al. 2007)
central vagus (cVN) cranial motor output. All and ultimately switches population activity to a
BötC neurons have intrinsic spike frequency mode of tonic asynchronous spiking. In the CPG
adaptation properties defined by a high-voltage- model, where the pre-BötC circuits are embedded
activated calcium (ICaL) with calcium flux in the larger network providing inhibitory and
activating calcium-dependent potassium currents excitatory inputs, under normal conditions, most C
(IK,Ca, see Eq. 2 below). neurons of the pre-BötC excitatory population are
The pre-BötC compartment in the model assumed to operate in a tonic-spiking mode due
includes two rhythmically active inspiratory neu- to high tonic excitatory input and are phasically
ronal populations: pre-I/I and early-I(1) (Figs. 1 inhibited by the expiratory neurons (post-I,
and 2), where the pre-I/I population is the key aug-E) (discussed in Smith et al. 2007).
excitatory population of pre-BötC serving as a The early-I(1) population within the pre-BötC
major source of network inspiratory activity. compartment is a population of inhibitory inter-
This population projects to the aug-I population neurons with adapting properties (also governed
of premotor inspiratory neurons of the rVRG and by ICaL and IK,Ca) and receives excitation from
via an excitatory hypoglossal premotor neuronal the pre-I/I population. In the modeled network,
population to the hypoglossal motor output this population serves as a major source of inhi-
(HN) (see Fig. 2; not shown in Fig. 3). Motoneu- bition of all expiratory neurons during inspiration
ron populations are not explicitly incorporated in (see Fig. 2).
the model, and the activity of premotoneuron Thus an important circuit architectural feature
populations is considered to be representative of of this model is that the minimal core structure
motor outputs since the activity patterns of of the CPG is represented by BötC inhibitory
motoneurons have been shown electrophysiolog- expiratory (post-I and aug-E) neurons and the
ically to generally follow the activity of inhibitory inspiratory neurons in the pre-BötC,
premotoneurons providing excitatory monosyn- coupled in a ring-like network with mutual
aptic inputs. The pre-I/I populations of excitatory inhibitory interactions (Fig. 3). This inhibitory
neurons, which are synaptically coupled by network interacts with the key excitatory kernel
mutual excitatory connections, play a critical network of pre-BötC inspiratory neurons to coor-
role in generating the inspiratory rhythm. In the dinate the generation of inspiratory and expira-
model, as found experimentally, some of these tory activity patterns. In this model synaptic
cells have intrinsic oscillatory bursting properties inhibition is essential for orchestrating phase
dependent on a persistent sodium current transitions to generate the three-phase respiratory
(INaP) – an experimentally established conduc- pattern.
tance mechanism underlying membrane The rVRG compartment contains the aug-I
voltage-dependent intrinsic rhythmogenic behav- excitatory, and early-I(2) inhibitory, populations.
ior of pre-BötC circuits in which endogenous Aug-I represents the rhythmically active popula-
oscillatory bursting is a function of the level of tion of premotor inspiratory neurons projecting to
neuronal excitability or tonic excitatory drive and exciting phrenic motoneurons. Activity of
(Smith et al. 2007; Richter and Smith 2014). All this population in the model represents PN output
pre-BötC excitatory neurons in this model have and contributes to the inspiratory component of
INaP. Earlier conductance-based models of INaP- cVN discharge. The role of the inhibitory early-
dependent bursting in a network model of iso- I(2) population of adapting neurons in the model
lated pre-BötC excitatory circuits (Butera is in shaping the ramping patterns of aug-I neu-
et al. 1999a, b; see also Butera et al. 2005 for rons. The tonic excitatory drives to these neurons
review) have shown that an increase in the aver- contribute importantly to the augmenting activity
age neuronal excitability or in external excitatory pattern.
Single Model Neuron Conductance

where ENa, EK, ECa, EL, ESynE, and ESynI are the
Descriptions and Population Modeling:
reversal potentials and gNa , gNaP , gK , gCaL , and
Mathematical Formulations
gK, Ca are maximum conductances of the
Neurons in this model are represented as single-
corresponding channels; gSynE and gSynI are
compartment cells with ionic currents for mem-
excitatory and inhibitory synaptic conductances,
brane channels described following the classic
respectively, and gL is the constant leak
Hodgkin–Huxley formulation as summarized
conductance.
below, following closely the description given
The kinetics of the activation and inactivation
in the original papers (Rybak et al. 2007; Smith
gating variables mi and hi (indexes indicate
et al. 2007). The simplified ionic current (I) equa-
individual ionic currents), respectively, are
tion for the pre-BötC excitatory neurons is
described as
dV
C ¼ I Na I NaP I K I L I SynE I SynI; d
dt tmi ðV Þ mi ¼ m1i ðV Þ mi ;
(1) dt (4)
d
thi ðV Þ hi ¼ h1i ðV Þ hi :
and the early-I (1, 2), aug-I, post-I, and aug-E dt
neurons are modeled by
The expressions for steady-state activation
and inactivation variables, time constants, value
dV of maximal conductances for individual neuron
C ¼ I Na I K I CaL I K, Ca I L
dt types, and the sources of experimental data from
I SynE I SynI; (2) which model parameters were obtained are
presented in the original papers (Rybak
et al. 2007; Smith et al. 2007).
where V is the membrane potential, C is the
For currents gated by the internal calcium
membrane capacitance, and t is time. The ionic
concentration in the model (IK, Ca), the kinetics
currents represented on the right side of these
of intracellular calcium concentration Ca is
current balance equations are INa [fast
described as follows:
(transient) sodium current], INaP [persistent
(slow inactivating) sodium current], IK (delayed
rectifier potassium current), ICaL (high-voltage- d
Ca ¼ kCa I CaL ð1 PB Þ
activated calcium-L current), IK,Ca (calcium- dt
dependent potassium current), IL (leakage þ ðCa0 CaÞ=tCa ; (5)
current), and ISynE and ISynI (excitatory and
inhibitory synaptic currents, respectively).
where the first term constitutes influx (with the
Expressions for individual currents are as
coefficient kCa) and buffering (with the probabil-
follows:
ity PB) such that PB = B/(Ca + B + K), where B is
the total buffer concentration and K is the rate
I Na ¼ gNa m3Na hNa ðV ENa Þ;
parameter. The second term describes kinetics of
I NaP ¼ gNaP mNaP hNaP ðV ENa Þ;
a calcium pump with resting level of calcium
I K ¼ g m4K ðV Ek Þ;
(3) concentration Ca0 and time constant tCa. The
I CaL ¼ gCaL mCaL hCaL ðV ECa Þ;
I K, Ca ¼ gK, Ca m2K, Ca ðV EK Þ; calcium reversal potential in the model is a func-
I L ¼ gL ðV EL Þ; tion of Ca:

I SynE ¼ gSynE V ESynE ; ECa ¼ 13:27 lnð4=CaÞ ðat rest Ca ¼ Cao
(6)
I SynI ¼ gSynI V ESynI ; ¼ 5 105 mM and ECa ¼ 150 mV :
For modeling multiple neuronal populations each neuron of population A received the
and their synaptic interactions, each functional corresponding excitatory or inhibitory synaptic
type of neuron (pre-I/I, early-I, post-I, aug-E) is input from each neuron of population B or from
represented by a population of 50 neurons. Con- d, respectively. The excitatory (gSynE) and inhib-
nections between populations in this model are itory postsynaptic (gSynI) conductances are
specified so that if a population A was assigned to activated by the network activity-dependent
receive an excitatory or inhibitory input from a excitatory or inhibitory inputs, respectively, for- C
population B or external excitatory drive d, then mulated as
X X X
gSynEi ðtÞ ¼ gE S wji exp t tkj =tSynE þ gEd Sfwdmi g dmi
tkj <t
X
j
X X
m
(7)
gSynIi ðtÞ ¼ gI S wji exp t tkj =tSynI þ gId Sfwdmi g dmi
j tkj <t m
where the function S{x} = x, if x 0, and 0 if conditions for values of membrane potential, cal-
x < 0. The first terms in the above two Eq. 7 cium concentrations, and channel conductances.
describe the dynamics of integrated synaptic
inputs from other network neurons. The second Dynamical Operation of Model Core Circuitry
term describes the integrated inputs from external Examples of the simulated output of this model
drives dmi. The activity dependence of these are shown in Figs. 3 and 4, which illustrate neu-
inputs is handled by considering that each action ron population activity patterns (Fig. 3) and the
potential arriving to neuron i from neuron j at membrane potential trajectories (Fig. 4) includ-
time tkj increases the excitatory synaptic conducing spiking patterns and the temporal profiles of
tance by gE wji if the synaptic weight wji > 0 or postsynaptic inhibitory conductances in single
increases the inhibitory postsynaptic conduc- representative neurons from each of the four
tance by gI wji if the synaptic weight wji < 0. The main neuron populations in the pre-BötC and
parameters gE and gI define the increase in the BötC circuitry. The integrated population activi-
excitatory or inhibitory synaptic conductance, ties are comparable to electrophysiological
respectively, produced by one action potential recordings (Fig. 1) obtained from an experimen-
arriving for |wji| = 1, and tSynE and tSynE represent tal preparation (in situ perfused brainstem–spinal
the decay time constants for the excitatory and cord preparation from juvenile rat) (Smith
inhibitory conductances, respectively. In the sec- et al. 2007, 2009), which spontaneously generates
ond term of Eq. 7, the parameters gEd and gId the three-phase respiratory pattern that is
define the increase in excitatory or inhibitory recorded at the neuron population level in med-
postsynaptic conductance, respectively, pro- ullary compartments and from cranial and spinal
duced by external input drive dmi = 1 with a nerves monitoring motor output. The model
synaptic weight of |wdmi| = 1. In the original parameters chosen and model simulations were
model, gId ¼ 0. The relative weights of synaptic designed to reproduce experimental data
connections (wji and wdmi) during the three-phase obtained from this preparation (see Smith
operation and other states of the network are et al. 2007 for details). The simulations have
provided in the original publications. Heteroge- reproduced some essential temporal features of
neity of neurons to account for biological vari- the three-phase organization of the neuron popu-
ability within each model population was lation activities and network outputs. This corre-
incorporated by a random distribution of EL spondence has been interpreted to imply that the
(mean values SD) as well as the initial minimal architecture for the core pre-BötC and
Computational Models of Mammalian Respiratory description given in the text, which provides an explana-
CPG, Fig. 4 Simulated membrane potential trajectories tion of the dynamical operation of these microcircuits.
and time courses of inhibitory postsynaptic conductances This explanation emphasizes the roles of synaptic inhibi-
in single neurons from the four major populations of the tion and intrinsic cellular biophysical properties in orches-
core pre-BötC and BötC circuitry in the conductance- trating phase transitions and neuronal rhythmic activity
based model with connections as depicted in Fig. 3. patterns underlying generation of the three-phase respira-
Sequential events beginning with the inspiratory tory pattern (The figure is from Smith et al. 2013 with
(I) period are numbered and correspond to the stepwise permissions)
BötC circuits represented in the model defines patterns, and underlying rhythmogenic mecha-
essential excitatory and inhibitory population nisms as the network was physically reduced
interactions for generating the three-phase rhyth- and functionally reconfigured. Regarding the
mic pattern. In additional tests of this model, latter, this model has delineated mechanistically
simulations could mimic a number of experimen- how the core CPG components could be built
tal observations including results from experi- hierarchically by assembling the postulated
ments examining (1) pharmacological local pre-BötC and BötC microcircuits with
perturbations of the cycle frequency and ampli- their particular electrophysiological properties
tude of population activity by blocking INaP, and their regulation by external excitatory drive
(2) disruptions of network activity by reducing inputs. Thus the model has shown that this type
inhibitory synaptic conductances, and (3) trans- of circuit assembly has the capability for func-
formations of CPG network activity, motor tional reorganization under different (patho)
physiological conditions, which is considered an three-phase respiratory pattern are generally in
important property of the CPG. accord with classical descriptions of Richter and
The membrane trajectories and postsynaptic colleagues, but in the current model, involvement
inhibitory conductances from representative sin- of specific compartmental microcircuits and their
gle model neurons (Fig. 4) serve to illustrate in more recently elaborated electrophysiological
principle the dynamical operation of the core properties are considered (see Richter and Smith
rhythmogenic microcircuits, particularly how 2014 for recent review). C
the mutual inhibitory interactions within the
ring-like architecture of circuit connections can Some Model Limitations
orchestrate phase transitions to produce the three- This model, as noted above, is a highly simplified
phase respiratory pattern. Sequential events representation of microcircuits within the CPG
(numbered in the figure) beginning with the inspi- that attempts to define key features of the archi-
ratory (I) period are as follows. (1) During the tecture and dynamics of postulated core circuitry
inspiratory phase, activity of the pre-BötC early-I incorporating simplified sets of neuronal bio-
inhibitory population declines due in part to physical properties. A key feature of the model
intrinsic spike frequency adaptation. This enables is the asymmetric strengths of mutual synaptic
slow disinhibition of the BötC post-I neurons, interactions within the ring-like inhibitory
evident in the trajectory of the post-I neuron connectome. In combination with strong intrinsic
inhibitory conductance (2). Spiking of post-I spike frequency adaptation properties of two of
inhibitory neurons (3) initiates inhibition of inspi- the inhibitory populations (post-I and early-I),
ratory activity by the abrupt rise of inhibitory this connectome enables the orderly progression
conductance in pre-I/I (4) and early-I neurons, of the cycle activity phases. Such asymmetry has
terminating the inspiratory phase. Simultaneous not yet been examined and verified experimen-
inhibition of BötC aug-E neurons (5) with the tally. Furthermore, all neurons in this model are
rapid onset of inhibitory postsynaptic conduc- assumed to receive tonic excitation that facili-
tance also silences these neurons. Activity of the tates the phase transitions. The mapping of excit-
aug-E neuron population (6) is disinhibited and atory input drives onto the circuitry and the
develops later in expiration, forming the second relative strengths of drives have also not been
expiratory (E-2) phase, as the post-I neuron spik- analyzed and confirmed experimentally. Some
ing frequency declines (adaptation). Aug-E neu- of the inputs from the sources represented in the
ron activity causes inhibitory conductance in model may in fact be rhythmic, and some models
post-I neurons to further reduce their spike have attempted to explicitly incorporate state-
frequency (7), while the simultaneous inhibition dependent phasic inputs (Rubin et al. 2011;
of early-I neurons (8) transiently prevents Lindsey et al. 2012). The interacting processes
reactivation of these neurons. The pre-I/I neurons leading to stable transitions between each activity
in pre-BötC are only weakly inhibited in the phase are dynamically complex (see section
model during the E-2 phase and receive high “Activity-Based Computational Models”
tonic excitatory drive, which in combination below), and the processes operating in the real
with the recovery from INaP inactivation in these network are undoubtedly not fully represented in
cells enables escape from inhibition and onset of the model. As shown from electrophysiological
pre-inspiratory phase spiking activity (9) initiat- recordings, the three phases of the cycle are not
ing a wave of synaptic excitation in the network. strictly sequential with some temporal overlap,
The resultant strong excitation to the early-I reflected by the spiking activity profiles of popu-
inhibitory neurons and their rapid onset of spik- lation neurons, particularly subsets of post-I neu-
ing (10) initiates strong inhibition of post-I and rons that spike into the E-2 phase, and pre-I/I
aug-E neurons, completing the E-to-I transition neurons that begin spiking in the E-2 phase.
and initiation of the next cycle. These model- These features are present in the model simula-
based descriptions for the generation of the tions. However, there is also temporal dispersion
in the spiking patterns of different neurons within intrinsic burst generation mechanisms of
a population active during a given phase. For pre-BötC excitatory neurons in a more simplified
example, it has been established experimentally format have been developed by Rubin
that there are also expiratory neurons with spiking et al. (2009, 2011). These models have been
confined primarily to the post-I phase. Currently it used for computational studies of circuit dynam-
is not clear whether this reflects heterogeneity of ics in various states, including during the three-
synaptic inputs and/or electrophysiological prop- phase rhythmic pattern generation state as well as
erties within a given neuronal population, which is structurally reduced states of the CPG network
a feature of the current model, or if this reflects that are considered to be functionally important.
functionally distinct subpopulations. Activity-based neural models (Ermentrout and
This model assumes that the essential circuit Terman 2010) are an established general form
interactions generating the respiratory rhythm are of models and have particular utility for (1) test-
distributed between the BötC and pre-BötC ing if the dynamical processes operating as
regions, which is consistent with current experi- described by more complex conductance-based
mental information but not yet definitively models, such as for the core circuitry of the respi-
established. The spatial segregation of neurons ratory CPG as described above, can be ade-
with particular spiking patterns is not exclusive. quately represented by a more minimal
For example, post-I spiking neurons are also construct within a dynamical system’s analytical
located in the pre-BötC, which may imply that and computational framework – this could imply
some of the inhibitory circuit interactions leading that the basic circuit architecture and the pro-
to inspiratory phase termination occur locally posed key dynamical properties of circuit ele-
within this compartment. Regardless of the par- ments can be a robust representation of the
ticular spatial arrangements, the circuit motifs essential dynamical processes that are not a func-
and dynamic processes inherent in the present tion of parameter tuning in the very large, com-
model are a plausible solution to the general plex parameter sets of the conductance-based
problem of identifying canonical circuit architec- model – and (2) applying numerical methods
ture and neuronal interactions for generating the including bifurcation analysis, phase plane anal-
three-phase activity pattern. The mutual inhibi- ysis, and fast–slow variable decomposition
tory connectome appears to be an important methods to analyze dynamic mechanisms under-
architecture for orchestrating the three-phase lying transitions between respiratory phases.
rhythmic pattern. More complex architectures These approaches have also enabled analyses of
with distinct subpopulations of glycinergic and how gradual changes of excitatory drives to par-
GABAergic inhibitory neurons and their postu- ticular cells in the network can control oscillation
lated network interactions, but retaining key fea- period and phase durations, which is an important
tures of the inhibitory circuit interactions physiological problem. In addition, transitions to
represented by the model described here, have other oscillatory regimes, such as experimentally
been explored (Shevtsova et al. 2011). For more established two-phase and one-(active) phase
biophysically complex conductance-based rhythmic activity states underlying important
models of pre-BötC excitatory circuit neurons, respiratory motor behaviors might be explained
incorporating additional sodium- and calcium- more rigorously, as well as the emergence of
based mechanisms for inspiratory rhythm gener- intermediate activity patterns in transitions
ation, see for example Jasinski et al. 2013. between rhythmic states that have been observed
experimentally.
Activity-Based Computational Models This model was constructed with the same
Activity-based (non-spiking) models of the core spatially organized network architecture as the
pre-BötC–BötC circuitry that incorporate the conductance-based model and correspondingly
essential spike frequency adaptation properties contains four neurons (i {1, 2, 3, 4}), represen-
of inhibitory neurons and INaP-dependent tative of the same neuronal populations, and three
sources of excitatory drive (dk; k {1, 2, 3}). Following closely the description given in
The four neurons include one excitatory pre-I/I Rubin et al. (2009), the membrane potential of
(i = 1) and one inhibitory early-I (i = 2) neuron, the INaP-containing excitatory pre-I/I neuron, V1,
for these neuronal types in the pre-BötC as in the is described by the differential equation:
conductance-based model, and two inhibitory
neurons, post-I (i = 3) and aug-E (i = 4), dV 1
C ¼ I NaP I K I L1 I SynE1 I SynI1 :
representing these populations in the BötC. dt C
Each type of neuron in this model is described (8)
by the dependent variable Vi, which represents an
average voltage for that (i-th) population, and The other three neuron types (i {2, 3, 4}) in
each output f(Vi) (0 f(Vi) 1) represents an the circuit explicitly have dynamic adaptation
averaged or integrated population activity at the of the depolarized voltage trajectory defined
corresponding average voltage. It is assumed that by the outward potassium current IADi. Each of
the dynamics of the average voltages can be these neurons has a membrane potential Vi that
represented in the model by a conductance-like evolves as
analytical framework since regimes in which
neurons within each population are synchronized dV i
C ¼ I ADi I Li I SynEi I SynIi : (9)
are considered (Rubin et al. 2009). Thus, the dt
model was formulated to also explicitly represent
dynamics of ionic currents so that the role of INaP In the above equations, C is the neuronal capac-
causing oscillatory activity and current causing itance, IK represents the potassium delayed recti-
neuronal adaptation properties could be incorpo- fier current, and ILi (i {1, 2, 3, 4}) is the
rated. In essence, this is an interesting novel, leakage current. ISynEi and ISynIi (i {1, 2, 3,
hybrid form of activity-based models that is 4}) are the excitatory and inhibitory synaptic
meant to capture some of the essential dynamics currents, respectively, described by
of conductance-based models.
I NaP ¼ gNaP mNaP hNaP ðV 1 ENa Þ,

I K ¼ gK m4K ðV 1 EK Þ,
I ADi ¼ gAD mADi ðV i EK Þ,
I Li ¼ gL ðV i EL Þ,
X3
I SynEi ¼ gSynE V i ESynE cki d k , for i 6¼ 2, and
k¼1 (10)
!
X
3
I SynE2 ¼ gSynE V 2 ESynE a12 f 1 ðV 1 Þ þ cki d k ,
k¼1
X4
I SynIi ¼ gSynI V i ESynI bji f j V j :
j¼2
j6¼i
In these equations gNaP , gK , gAD , gL , gSynE ; and potentials; a12 defines the weight of the excitatory
gSynI are the maximum conductances of the synaptic input from the pre-I/I to the early-I
corresponding currents and ENa, EK, EL, ESynE, neuron (see Fig. 5); bji defines the weight of
and ESynI are the corresponding reversal the inhibitory input from neuron j to neuron
i (i {1, 2, 3, 4}, j {2, 3, 4}); and cki defines Rubin et al. noted that at most points in the
the weight of the excitatory synaptic input to eight-dimensional phase space representing the
neuron i from drive k (dk, k {1, 2, 3}). The eight variables of this model, the voltages
asymmetries in synaptic weights for the mutual V1,. . .,V4 evolve much faster than the inactivation
inhibitory interactions between early-I, post-I, variable hNaP and the adaptation variables mADi
and aug-E neurons (e.g., b23 > b32, b34 > b43, (i = 2,3,4). Accordingly, the voltages represent
but b24 = b42) specify the order in which these the fast variables, and the system of differential
cell types generate action potentials to produce equations for V1,. . .,V4 represent the fast
the three-phase rhythmic pattern. For the analysis subsystem. Equilibrium points of this subsystem
of effects of total independent excitatory drives occur where the right-hand sides of all equations
(Di, i {1, 2, 3, 4}) to different target neurons, in the fast subsystem are zero and typically form
Di were varied without decomposition into smooth equilibrium surfaces. These investigators
weighted sums of drives dk. In these cases, in accordingly used bifurcation analysis to identify
the equations for ISynEi and ISynIi of the system fast subsystem equilibrium surfaces and phase
X 3 transitional features by numerically tracking the
(Eq. 10), the term cki dk was replaced in the fast subsystem equilibrium surfaces along which
k¼1
the fast system variables are relatively constant
model with Di (see Fig. 5).
and detecting bifurcation curves, where solutions
The output activity of each neuron type
jump away (“fast jumps”) from each surface.
(indirectly representing the rate of spiking activ-
Thus, they demonstrated how bifurcation dia-
ity) is defined by the nonlinear function fi(Vi)
grams (see Rubin et al. 2009) can be created by
treating the variables hNaP and mADi as fixed
f i ðV i Þ ¼ 1= 1 þ exp V i V 1=2 =kVi ,if1,2, 3, 4g;
parameters and following curves of equilibrium
(11) points of the differential Eqs. 8 and 9 for the
voltages V1,. . .,V4. They also demonstrated that
where V1/2 is the half-activity voltage and kVi in certain cases, bifurcation curves could be gen-
defines the slope of the output function for each erated by simultaneously varying two activation
neuron. variables while holding the other two fixed,
The two slow variables in the model (hNaP), whereas in other cases, all four activation vari-
representing the slow inactivation of INaP in neu- ables could be varied simultaneously along a
ron 1 (pre-I/I), and mADi (i {2, 3, 4}), defining one-dimensional curve in four-dimensional
the time course of adaptation in the other three space. The dynamics of the model phase transi-
neurons (each with fixed time constant tADi and tions, analyzed with this approach, yielded new
maximal adaptation), are defined by kADi: insights into the operation of the circuit in differ-
d ent states.
thNaP ðV 1 Þ hNaP ¼ h1NaP ðV 1 Þ hNaP ,
dt (12)
d
tADi mADi ¼ kADi f i ðV i Þ mADi :
dt Dynamical Operation During Three-Phase
Voltage-dependent activation and inactivation Rhythmic Pattern Generation: New Insights
variables and time constants for INaP (based on The three-phasic operation of this activity model,
the INaP kinetic model of Butera et al. 1999a) and as illustrated in Fig. 5, generates stable oscilla-
activation of IK channels in the pre-I/I neuron are tions with period T, duration of inspiration TI,
described as follows: and duration of expiration TE, which, as the
model was parameterized, are within the physio-
mNaP ¼ 1=ð1 þ expððV 1 þ 40Þ=6ÞÞ, logical ranges observed experimentally in the
h1NaP ¼ 1=ð1 þ expððV 1 þ 48Þ=6ÞÞ, brainstem–spinal cord preparations of juvenile
(13)
thNaP ¼ thNapmax =coshððV 1 þ 48Þ=12Þ, rats whose activity the model was intended to
mK ¼ 1=ð1 þ expððV 1 þ 29Þ=4ÞÞ: mimic (as with the conductance-based model).
Computational Models of Mammalian Respiratory inspiration (TI) and expiration (TE) with T = 2.5 s,
CPG, Fig. 5 Activity profiles of the four main types of TI = 0.9 s, and TE = 1.6 s that are within the observed
neurons in the pre-BötC–BötC microcircuitry and control experimental range. The voltage trajectories show adap-
of respiratory cycle period and phase durations by tonic tation properties of post-I and early-I neurons, slow
excitatory drives from simulations with the activity-based depolarizing drift in voltage of pre-I/I neurons, and
model of Rubin et al. (a) The distinct populations of augmenting activity profile of aug-E neurons.
inspiratory and expiratory neurons of the core circuitry Compare the profiles with integrated population activities
within the BötC and pre-BötC compartments provide of the conductance-based model shown in Fig. 3. (c)
multiple nodes that may be targeted by external drives Dynamic ranges of T, TI, and TE with changes in total
for control of cycle period and phase durations. (b) Sim- drive selectively to each neuron type D1. . .D4, where the
ulations show activity profiles of each neuron type of the excitatory postsynaptic input to each neuron has been
core circuitry producing a three-phase activity pattern weighted in proportion to the drive (Modified from
with respiratory cycle period (T) and durations of Rubin et al. 2009)
The voltage trajectories show characteristic slow by release (as distinguished from escape), with
adaptation properties of post-I and early-I neu- mAD2 growing during inspiration.
rons, depolarizing drift in voltage of pre-I/I neu- As discussed above for the conductance-based
rons during the expiratory period due to slow model, it is important to emphasize that although
recovery from inactivation of INaP, and the the expiratory period can be conditionally
augmenting activity profile typical of aug-E neu- subdivided into post-inspiratory (with a dominant
rons, all of which are comparable to the temporal post-I neuronal activity) and late E-2 expiratory
profiles of integrated population activities in the (with reduced post-I and increasing aug-E activ-
conductance-based model (see Fig. 3). During the ity) phases, as the model is constructed, there is
expiratory phases, the adapting (decrementing) no distinct switching between post-I and aug-E
activity of the post-I neuron, which is defined by neuron activity phases in the so-called three-
the dynamic increase in its adaptation variable phase rhythmic pattern. Mathematically, there
mAD3 (see Eqs. 9 and 12), causes a decline in are only two fast jumps observed by bifurcation
inhibition from the post-I neuron that shapes the analysis, and accordingly the rhythmic pattern is
augmenting profile of aug-E neuron activity and not strictly three-phase, although functionally
produces slow depolarization of the pre-I/I and this is the case, leading these investigators to
early-I neurons. Simultaneously, the pre-I/I neu- term the normal respiratory pattern as a “bio”
rons are depolarizing due to the slow recovery three-phase rhythm. Nonetheless, in the model
from inactivation of INaP (slow increase of hNaP, V1 and V2 are able to escape from this complex
Eq. 12) so that at some moment the pre-I neurons temporal pattern of expiratory phase inhibition
rapidly activate, providing excitation of the for the transition to the inspiratory phase, as
early-I neurons. This inhibits both post-I and occurs naturally. A novel interesting result of
aug-E neurons, completing the switch from expi- the analysis is that with a higher excitatory drive
ration to the inspiratory phase in accord with the to the network, the early-I neurons can release
conductance-based descriptions. The analysis from inhibition first and initiate the transition to
indicates that the moment of pre-I/I activation the inspiratory phase. Thus this model in general
occurs when the system reaches a bifurcation has delineated possible contributions of escape
curve along the post-I/aug-E equilibrium surface, and release mechanisms to various phase transi-
allowing the fast voltage variables to evolve tions during the cycle.
quickly until they again reach an equilibrium Results of these computational analyses have
surface, yielding a fast jump to the inspiratory also generally suggested that the basic architecture
phase. and key properties of the core circuit neurons, as
During the inspiratory phase, according to this represented in both the simplified activity-based
analysis, the system drifts near a new equilibrium model with reduced parameter sets and the more
surface, with its dynamics governed by Eq. 12. complex conductance-based model, constitute a
The early-I neuron, with its adapting activity, is robust dynamical structure. The mutual inhibitory
defined by the dynamic increase in its adaptation ring-like architecture interacting with the embed-
variable mAD2 (Eqs. 9 and 12). The associated ded excitatory kernel intrinsically gives rise to
decline in inhibition from this neuron produces stable phase transitions. This bolsters the conclu-
the slow depolarization of the post-I and aug-E sion that this represents a plausible canonical cir-
neurons, and eventually, the system reaches cuit arrangement within the CPG for generation of
another bifurcation curve, and a second fast the three-phase rhythmic pattern.
jump occurs. Within this fast jump, the post-I
neuron rapidly activates and inhibits both inspi- Control of Oscillation Period and Phase
ratory neurons (pre-I/I and early-I) and the aug-E Durations in the Three-Phase Rhythm-
neuron initially, producing the inspiratory to Generating State
expiratory phase transition. The model analysis One important problem in respiratory neural con-
indicates that this transition is always facilitated trol addressed with this model is to delineate
mechanistically how the excitatory drives from from inhibition and corresponding earlier switch
various sources (signaling changes in physiolog- to the inspiratory phase. Interestingly, as noted
ical conditions) control the frequency and dura- above, with a sufficient increase in drive, the
tions of inspiratory–expiratory phase activity. early-I cell can replace the pre-I/I cell as the initi-
Rubin et al. (2009) noted that each of the ator of the transition to inspiration. Experimental
inhibitory and excitatory populations of the manipulation of neuron excitability within the
pre-BötC–BötC circuitry potentially represent pre-BötC modulates respiratory cycle frequency C
nodes for control of period and phase durations over a substantial dynamic range and reinforces
in the intact respiratory network. To theoretically the basic concept that the pre-BötC is a key
investigate this control with the activity-based rhythmogenic structure. According to the model-
model, the effects of variations in the total exciting results, both the excitatory and inhibitory neu-
atory drive Di to each neuron on the oscillation ron populations in the pre-BötC can play a role and
T and on the durations of TI and TE have been represent different mechanisms for control of
analyzed. Some of the results are summarized in respiratory frequency and phase durations. This
Fig. 5. new insight points to potentially distinct mecha-
In general this model showed that variations in nisms at this level of the circuitry.
drive to the different network cells impact the These modeling results illustrate in principle
inspiratory and expiratory phase durations, and how the inspiratory and expiratory neuron
the overall oscillatory period, differently with the populations can be distinct nodes for differential
model operating in the regime generating the control of period and phase durations. As pointed
three-phase pattern (Fig. 5). For example, out by Rubin et al. (2009), the significant changes
increasing the drive to aug-E neurons lengthened in the oscillation period following changes in
T (reduced respiratory frequency) by 2.5 times drives to particular neurons revealed by the
by prolonging TE and by a small increase in TI. model are not typical for classical mutual inhibi-
The prolongation of TE results from a stronger tory half-center models. Mathematically, the dif-
aug-E inhibition of inspiratory neurons that ferent effects on period observed relate in part
necessitates a greater recovery from adaptation to the asymmetry between decrementing and
and inactivation before the inspiratory phase can augmenting activity patterns in different cells.
initiate, causing a delay in the onset of this phase. Phase transitions in the model occur via combina-
In contrast elevated drive to post-I neurons tions of adaptation, release from inhibition, and
slightly decreased T (increase frequency) as TE recovery from adaptation (or from INaP inactiva-
becomes shorter because the additional drive to tion in the pre-I/I neurons) that provides the
post-I leads to stronger inhibition on aug-E. The escape mechanism. Importantly, the activity-
resulting weaker activation of aug-E leads to a net based modeling approach allowed the simulated
weaker inhibition of pre-I/I and early-I neurons, changes in phase durations and oscillation periods
allowing a slightly earlier escape to initiate the to be linked mathematically to the different bal-
inspiratory phase. ances of escape and release from inhibition and the
With the drive targeting neurons in the contributions of neuronal spike frequency adapta-
pre-BötC, the cycle period can be reduced and tion, all of which collectively underlie the phase
frequency increased by a factor of 4 from its initial transitions. Thus, the modeling results generally
value with increasing drive to the pre-I/I excitatory demonstrate how activity patterns of different cell
neurons. This occurs primarily by shortening TE types in the CPG (1) dictate the forms of the
because the increase in drive to the pre-I/I neuron transition curves between phases and underlie
allows this neuron to escape from inhibition ear- the distinctive role each cell population plays in
lier. Increasing drive to the pre-BötC early-I inhib- the phase transitions and (2) govern the effects that
itory neurons can also reduce T by a factor of 2 changes in excitatory drive have on the phase
but by reducing TI and shortening of TE, which transition curves and overall respiratory activity
results from an earlier escape of the early-I neuron pattern.
In other results from this theoretical analysis, circuit architectures and the dynamical operation
it has been demonstrated that the different nodes of postulated core microcircuits of the CPG. With
in the circuitry can not only differentially provide the recent advent of new connectomics and
control of oscillation period and phase durations optogenetics approaches for reconstructing circuit
but also can allow for control of the mode of structure–function, increasingly more detailed
rhythmic pattern generation by the external experimental information will become available.
drives. The detailed treatment of this aspect is As with all CNS circuits, CPG circuit dynamics
presented in Rubin et al. (2009). This analysis cannot be easily intuited even when circuit orga-
has generally reinforced the conclusion from the nization has been definitively established. Many
conductance-based model that the core CPG cir- opportunities therefore exist for future concerted
cuits appear to contain multiple functionally experimental and computational studies investi-
embedded oscillatory mechanisms, ranging gating the operation of this vital CPG.
from mutual inhibitory network-based interac-
tions between pre-BötC and BötC circuits resem-
bling classical inhibitory half-center type References
structures to conductance-based endogenous
rhythmogenic mechanisms within the pre-BötC, Abdala APL, Rybak IA, Smith JC, Paton JFR
(2009) Abdominal expiratory activity in the rat
with important physiological implications.
brainstem-spinal cord in situ: patterns, origins, and
Another important extension of this activity- implications for respiratory rhythm generation.
based model has been the inclusion of an addi- J Physiol 587:3539–3559
tional state-dependent oscillatory mechanism for Butera RJ, Rinzel J, Smith JC (1999a) Models of respira-
tory rhythm generation in the pre-Bötzinger complex.
homeostatic regulation, in the RTN/pFRG, that
I. Bursting pacemaker neurons. J Neurophysiol
becomes active under conditions of high respira- 82:382–397
tory drive (i.e., by elevated CO2 in the system) Butera RJ, Rinzel J, Smith JC (1999b) Models of respira-
and dynamically interacts with pre-BötC–BötC tory rhythm generation in the pre-Bötzinger complex.
II. Populations of coupled pacemaker neurons.
oscillator to produce forced expiration – a
respiratory motor behavior distinct from the nor- Butera RJ, Rubin J, Terman D, Smith JC (2005) Oscilla-
mal three-phase motor output that occurs at tory bursting mechanisms in respiratory pacemaker
normal resting levels of CO2. Again, the applica- neurons and networks. In: Coombes S, Bressloff PC
(eds) Bursting. The genesis of rhythm in the
tion of analytical methods from dynamical sys-
nervous system. World Scientific Press, London,
tems theory has elucidated possible mechanisms pp 303–347
and dynamic synchronization of these oscillators. Ermentrout GB, Terman D (2010) Mathematical founda-
These computational studies are reviewed by tions of neuroscience. Springer, New York
Grillner S, Jessell TM (2009) Measured motion: searching
Molkov in this volume.
for simplicity in spinal locomotor networks. Curr Opin
Summary Perspectives Jasinski PE, Molkov YI, Shevtsova NA, Smith JC, Rybak
The brainstem respiratory CPG has provided IA (2013) Sodium and calcium mechanisms of rhyth-
mic bursting in excitatory neural networks of the
unique opportunities to experimentally unravel
pre-Bötzinger complex: a computational modeling
how cellular and circuit properties are integrated study. Eur J Neurosci 37:212–230
in the mammalian CNS to produce rhythmic Lindsey B, Rybak IA, Smith JC (2012) Computational
motor behavior. Some of the properties of key models and emergent properties of respiratory neural
networks. Compr Physiol 2:1619–1670
components of this CPG have been experimen-
Purvis L, Smith JC, Koizumi H, Butera RJ (2007) Intrinsic
tally investigated in extraordinary detail, yet bursters increase the robustness of rhythm generation
many aspects of neuronal and circuit structural– in an excitatory network. J Neurophysiol
functional properties remain to be established 97:1515–1526
Richter DW, Smith JC (2014) Respiratory rhythm gener-
(Smith et al. 2013). The data-driven computa-
ation in vivo. Physiology 29:558–571
tional modeling approaches described have been Rubin JE, Shevtsova NA, Ermentrout GB, Smith JC,
instrumental in exploring plausible canonical Rybak IA (2009) Multiple rhythmic states in a
Computational Models of Modulation of Oscillatory Dynamics 751 C
model of the respiratory CPG. J Neurophysiol can be of different nature: (i) quantitative where
101:2146–2165 the key parameters of oscillations, e.g., amplitude
Rubin J, Bacaak BJ, Molkov YI, Shevtsova NA, Smith JC,
Rybak IA (2011) Interacting oscillations in neural or dominant frequency, vary over the course of
control of breathing: modeling and qualitative analy- a simulation and/or (ii) qualitative where the
sis. J Comput Neurosci 30:607–632 oscillatory regime undergoes state transitions.
Rybak IA, Smith JC (2009) Computational modeling of Modulatory mechanisms that underlie changes
the respiratory network. In: Binder MD, Hirokawa N,
Windhorst U (eds) Encyclopedia of neuroscience, part in the oscillatory behavior of a model could C
3. Springer, Berlin, pp 824–832 also be of various origins ranging from an
Rybak IA, Abdula APL, Markin SN, Paton JFR, Smith JC external stimulus, entraining input, and specific
(2007) Spatial organization and state-dependent mech- excitatory/inhibitory drive to simulated
anisms for respiratory rhythm and pattern generation.
Prog Brain Res 165:201–220 neuromodulation or attentional effects.
Shepard GM, Grillner S (2010) Handbook of brain micro-
circuits. Oxford University Press, New York
Shevtsova NA et al (2011) Computational modelling of Detailed Description
5-HT receptor-mediated reorganization of the
brainstem respiratory network. Eur J Neurosci
34:1276–1279 Background and Scope
Smith JC, Abdala APL, Koizumi H, Rybak IA, Paton JFR Oscillations constitute an inextricable aspect of
(2007) Spatial and functional architecture of the mam- neural dynamics. Their origins and underlying
malian brainstem respiratory network: a hierarchy of
three oscillatory mechanisms. J Neurophysiol mechanisms at a cellular as well as population
98:3370–3387 level have been under investigation using both
Smith JC, Abdala APL, Rybak IA, Paton JFR (2009) Struc- experimental and theoretical approaches. The
tural and functional architecture of respiratory net- importance of oscillatory effects in the brain can
works in the mammalian brainstem. Philos Trans
R Soc Lond B Biol Sci 364:2577–2587 largely be explained by abundant biological evi-
Smith JC, Abdala APL, Borgmann A, Rybak IA, Paton JFR dence for their correlations with various aspects
(2013) Brainstem respiratory networks: building blocks of cognition and behavior. In this light, task- or
and microcircuits. Trends Neurosci 36:152–162 function-dependent modulations of oscillatory
Stein PSG, Grillner S, Selverston AI, Stuart DG
(1997) Neurons, networks, and motor behavior. MIT dynamics bear special relevance to understanding
Press, Cambridge, MA the mechanisms, role, and nature of this prevalent
phenomenon manifested at different scales of
neural system organization. Computational
models that capture such modulatory effects and
demonstrate their impact on the ongoing neural
Computational Models of dynamics and/or elucidate their functional impli-
Modulation of Oscillatory Dynamics cations on a circuit or system level serve as
a valuable tool in unraveling the mystery of oscil-
Pawel Andrzej Herman lations. The emphasis is here on spiking neural
Computational Biology, KTH Royal Institute of network models that simulate biologically rele-
Technology, Stockholm, Sweden vant processes involved in dynamical modulation
of ongoing neural activity in the brain. In other
words, experimental scenarios where the effect of
Definition static manipulations of model parameters on the
emergent oscillatory dynamics is tested in inde-
The term refers to a wide range of spiking neural pendent simulations lie outside the definition of
network models that not only exhibit oscillatory the term of computational models of modulations
behavior but also simulate neural effects and/or adopted here.
processes underlying ongoing changes in the The most common manifestations of modula-
oscillatory dynamics. These changes in the oscil- tory effects are quantitative changes in parame-
latory dynamics, often referred to as modulations, ters of neural rhythms, such as amplitude or peak
C 752 Computational Models of Modulation of Oscillatory Dynamics
frequency, and qualitative shifts of the oscillatory Stimulus- or State-Dependent Modulations

regime including synchronization effects. There of Oscillatory Dynamics
is a wide range of mechanisms exploited in com- Neural computation theories assume that in the
putational models to simulate modulations of macroscale brain processes information
oscillations. For example, the transitions between depending on incoming stimuli and its current
different oscillatory states in a multi-stable model dynamical as well as functional state. In the
can be induced by an external stimulus (e.g., same spirit, modulation of the dynamics
Battaglia and Hansel 2006; Börgers et al. 2008; exhibited by network models often results from
Herman et al. 2013; Lundqvist 2013a, b) or an external input/stimulus they receive and/or
endogenously controlled by intrinsic cellular or some intrinsic endogenous state transitions they
network dynamics (Herman et al. 2013). Alterna- undergo. Some computational studies attribute
tively, a periodic modulation of oscillations can functional interpretation to different dynamical
be imposed by varying noise conditions or other states of models. In this regard, special credit
changing sources in order to study implications of should be given to earlier work on spiking attrac-
changes in oscillatory dynamics (Lundqvist tor networks (Amit and Brunel 1997; Compte
et al. 2013a). From a functional perspective, et al. 2003; Lundqvist et al. 2010) used to simu-
modulation of oscillatory phenomena in compu- late a so-called memory delay period when
tational models is commonly studied in the con- humans and primates held one or several items
text of neuromodulatory (Destexhe et al. 1994; in working memory. The network dynamics dur-
Liljenström and Hasselmo 1995; Li and Cleland ing memory-coding interval is qualitatively dif-
2013) and attentional effects (Börgers et al. 2005, ferent from that exhibited beforehand. More
2008; Buia and Tiesinga 2006; Gu and recently, the abundant work on cortical attractor
Liljenström 2007; Buehlmann and Deco 2008) models by Lundqvist and Herman (Lundqvist
as well as perceptual and working memory pro- et al. 2011, 2013a, b, c; Herman et al. 2013)
cesses (Lundqvist et al. 2012, 2013a, b; Herman deserves here special attention since their
et al. 2013). Then, depending on the modeling biophysically detailed networks have been used
context, simulated modulation is most often to simulate a wide range of memory processes
related to cortical or hippocampal dynamics. and perceptual phenomena with strong oscilla-
Since gamma rhythm with the peak frequency tory correlates in agreement with biological
above 20–30 Hz constitutes the most prevalent data. In Lundqvist et al. (2011) and Herman
spectral component of activity of oscillatory spik- et al. (2013), endogenous periodic switching
ing network models, it is worth recalling its origin between different network states with distinct
(Buzsaki and Wang 2012). The key neural mech- spectral manifestations and functional meaning
anism involved is slow and fast feedback inhibi- was reported. In particular, transitions from a
tion on a network level (Brunel and Hakim 1999; memory-coding state accompanied by elevated
Whittington et al. 2000) or feedback inhibition on spectral power in theta and gamma bands to
a cellular level (adaptation and other slow a noncoding idling regime with increased alpha-
hyperpolarizing activity-dependent currents, band activity and suppressed theta-gamma corre-
e.g., Rotstein et al. 2005; Lundqvist et al. 2006; lates, and vice versa, serve as a prime example of
Herman et al. 2013). The network-level feedback intrinsically driven modulation of the model’s
inhibition can be exhibited either between excit- oscillatory dynamics. It should be added here
atory and inhibitory cells leading to gamma-like that such switching dynamics mediated by neural
oscillations at around 30 Hz (Brunel and Hakim adaptation as well as short-term synaptic plastic-
1999; Whittington et al. 2000) or between inhib- ity (depression and augmentation) can be linked
itory neurons resulting in faster oscillations at to a so-called periodic memory replay phenome-
frequencies higher than 60 Hz. In vivo both non implicated in working memory maintenance
types of interactions are likely to contribute to (Sandberg et al. 2003; Fuentemilla et al. 2010;
the broadband gamma (Ainsworth et al. 2011). Lundqvist et al. 2011; Herman et al. 2013).
Another modulatory effect in attractor network model (Lundqvist et al. 2013a) serves as one of
models simulating working memory is associated the most representative examples of imposing
with varying oscillatory power in theta, alpha, rhythmic modulation in a network by an external
and gamma frequency bands depending on the source. The authors used an attractor network to
memory load, i.e., the number of memory items simulate a threshold-stimulus detection task and
to be replayed by the network (Herman imposed a slow sinusoidal modulation (0.1 Hz)
et al. 2013). This phenomenon has been widely of excitatory noise onto each principal cell with C
reported in experimental studies (e.g., the aim of generating a slowly varying oscillatory
Raghavachari et al. 2001; Jacobs and Kahana component in the synthesized field potentials.
2009). All these effects can be explained by the This allowed for investigating its effect on
so-called bistability of the network dynamics processing weak stimuli. Since the original net-
(it can exhibit two qualitatively different regimes work without a sinusoidal noise source produces
of operation), controlled by the interaction of oscillatory activity in three distinct frequency
excitatory and inhibitory cells, and by synaptic bands roughly corresponding to electrophysio-
(Wang et al. 2006; Mongillo et al. 2008) and logical theta, alpha, and gamma depending on
neural mechanisms regulating lifetime of the current functional and dynamical state, as
memory-coding states. a result of introducing the aforementioned fluc-
Changes or transitions of the dynamical tuating component, a strong phase-amplitude
regime of neural network models can also be coupling emerged. In other words, a modulatory
induced by external stimuli. As a result, effect of the phase of the slow fluctuations on the
a stimulus drives the modulation of the model’s power of faster rhythms (>1 Hz) was obtained in
ongoing oscillatory dynamics. This was the orig- agreement with experimental data (Monto
inal approach to simulating a memory delay et al. 2008) and in line with the concept of nested
period (Amit and Brunel 1997; Compte oscillations (cf. section “Models of Nested
et al. 2003; Lundqvist et al. 2006), where the Oscillations”).
stimulus forced cells in the targeted cell assembly Another prominent example of inducing oscil-
to fire at an elevated rate, simultaneously latory behavior in a model and studying the emer-
suppressing the activity in other cells not belong- gent modulatory dynamics is a combined
ing to the given assembly. This suppression is simulation and experimental study by Jones
usually mediated by disynaptic inhibition via et al. (2009). In their laminar model of somato-
a population of interneurons. As mentioned ear- sensory cortex, both feedforward and feedback
lier, in a biologically detailed version of such pathways were stimulated with exogenous 10 Hz
attractor networks (Lundqvist et al. 2006, 2010, spiking input mimicking Rolandic mu rhythm.
2013a, b, c; Herman 2013) interactions between For the feedfoward pathway layer 4 served as
pyramidal and basket cells in layers 2/3 mediate the target, whereas for the feedback pathway,
the competition between different memory pat- the input was delivered to distal dendrites of
terns and produce at the same time gamma-like layer 2/3 cells. By manipulating the relative
oscillations following a stimulus presentation. timing of these inputs, the authors could study
There have also been other detailed models properties of the total modulated rhythmic activ-
devised to study the impact of stimulation on ity and its effect on evoked sensory responses in
oscillatory power changes (Jones et al. 2009). a simulated somatosensory task.
However, they did not have a clear functional
interpretation. Models of Nested Oscillations
Nested oscillation is a phenomenon where the
Imposed Rhythmic Modulation of Neural phase of underlying slow rhythm modulates the
Dynamics in Computational Models power of faster oscillations. It has been
A study on functional and dynamical implica- a prominent finding for theta and gamma
tions of infra-slow fluctuations in a cortical rhythms, particularly in memory tasks
(e.g., Chrobak and Buzsaki 1998; Canolty changing the ratio between theta and gamma, as
et al. 2006; Jensen and Colgin 2007; Kendrick a potential effect of learning. Their simulation
et al. 2011). This classical example of cross- results suggested that shallow nesting of
frequency modulatory dynamics has received gamma on theta, reported experimentally in
some attention in the computational neuroscience inferotemporal cortex in a learning scenario, opti-
community. It has been modeled in both hippo- mized coupling between the two rhythms and
campal networks (White et al., 2000; Rotstein maximized responses in the downstream neurons.
et al. 2005; Neymotin et al. 2011) and cortical A different approach to phase-amplitude mod-
models (Kendrick et al. 2011; Herman ulation was adopted in a cortical laminar model
et al. 2013). The origin of oscillatory nesting is proposed by Kramer et al. (2008). The authors
still a matter of debate. In modeling work the suggested that layer interactions lead to the emer-
emphasis has so far been on the role of intrinsic gence of lower-beta rhythm as a result of period
properties of interneurons (White et al. 2000; concatenation of simultaneous faster rhythms. In
Tiesinga et al. 2001; Rotstein et al. 2005; another cortical model the occurrence of theta-
Neymotin et al. 2011). For generating a slower gamma coupling in layer 2/3 was obtained as
theta rhythm, Neymotin et al. (2011), Rotstein a result of the attractor dynamics mediated by
et al. (2005), and White et al. (2000) relied on recurrent connectivity (Herman et al. 2013).
a specific subtype of interneurons with slow syn- While the neural mechanism underlying
aptic dynamics (oriens lacunosum moleculare) gamma-band activity was analogous to the other
due to their prevalence in hippocampus. In models, i.e., local inhibitory-excitatory loop,
Rotstein et al.’s (2005), their coupling with fast- the slower rhythm reflected the activation of
spiking interneurons was needed to sustain syn- a distributed memory pattern in the network
chronized rhythmic activity at a population level. stored by means of the aforementioned long-
Then, as a result of special tuning of the coupling range recurrent connections. Most importantly,
parameters to avoid interneurons firing without due to this functional aspect of the dynamics of
phasic excitation and with introduction of theta-gamma modulation, Herman et al.’s (2013)
AMPA-mediated excitation, the authors man- model lends itself to an interpretation of
aged to obtain the effect of nested oscillations. a computational role of nested oscillations. In
A similar circuit of interneurons with fast and the context of memory and learning, one could
slow synaptic dynamics was proposed by White hypothesize that slow and spatially coherent
et al. (2000) as main contributors to hippocampal rhythm would provide a window for bursting
rhythms. Neymotin et al. (2011), on the other and synaptic wiring of cell assemblies whereas
hand, introduced an additional external driving the local populations constituting an assembly
input to oriens lacunosum moleculare interneu- would exhibit activity on faster timescales
rons from medial septum. Then, to generate corresponding to gamma rhythm. In that frame-
gamma-band activity, classical basket-pyramidal work, gamma oscillations would mediate local
cell loop was exploited, also enhanced by medial computations, possibly of competitive nature.
septum. Modulatory effect of theta rhythm with
respect to the amplitude of gamma oscillations Attentional and Neuromodulatory Effects on
was obtained due to the coupling between two Oscillatory Dynamics in Computational
different types of interneurons in certain analogy Models
to Rotstein et al.’s (2005) and White et al.’s Neuromodulation provides another dimension to
(2000) minimalistic models. Similarly, the neural the dynamics of neural networks by modifying
network model developed by Kendrick et al. intrinsic properties of neurons and synapses
(2011) exhibited theta-nested gamma as a result (Marder and Thirumalai 2002; Marder 2012).
of combining fast and slow inhibitory interneu- Although there have been a number of computa-
rons. Interestingly, the model allowed the authors tional studies that attempt to enhance our
to investigate functional consequences of understanding of the computational role of
neuromodulation (Fellous and Linster 1998), less enhancement can be obtained by decreasing neu-
attention in modeling work has been given to its ral adaptation, which could reflect cholinergic
implications for the oscillatory dynamics in modulation (Börgers et al. 2005). Alternatively,
neural circuits. Historically, it is important to cholinergic-mediated reduction of inhibitory
mention theoretical studies on extrinsic drive on excitatory cells can be used as
neuromodulation of thalamic reticular cells a biasing mechanism (Buia and Tiesinga 2006;
(Destexhe et al. 1994) and the effects of cholin- Börgers et al. 2008). Gu and Liljenström (2007) C
ergic modulation on the olfactory cortex simulated different effects of cholinergic modu-
(Liljenström and Hasselmo 1995). Destexhe lation on extracortical and intracortical projec-
et al. (1994) illustrated that a neuromodulator tions in a multilayer cortex and obtained the
starts a cascade of cellular processes that leads effect of attentional increase of gamma-band syn-
to burst activity of single neurons, which gener- chronization in the superficial layers.
ates population oscillations in the frequency
range of spindle waves. Liljenström and
Hasselmo (1995) contributed to the understand- Cross-References
ing of the implications of acetylcholine effects at
the cellular level for cortical oscillations. In ▶ Attentional Top-Down Modulation, Models of
a rather simple model of EEG and field potentials ▶ Computational Models of Neuromodulation
recordings from the piriform cortex, they man- ▶ Gamma Rhythm, Neural Population Models of
aged to explain the mechanism of altered evoked the
gamma oscillations and demonstrated how sup- ▶ Hippocampal Theta, Gamma, and Theta/
pression of both neural adaptation and intrinsic Gamma Network Models
excitatory synaptic transmission coupled with the ▶ Neuromodulation: Overview
activation of inhibitory interneurons mediated
dynamical shifts between non-oscillatory and
theta rhythmic regimes. More recently, the mech- References
anism underlying cholinergic neuromodulation
in the olfactory system was scrutinized in Ainsworth M, Lee S, Roopun AK, Cunningham MO,
a biophysical model of the olfactory bulb Traub RD, Kopell NJ, Whittington MA (2011) Dual
gamma rhythm generators control interlaminar syn-
(Li and Cleland 2013). Muscarinic chrony in auditory cortex. J Neurosci 31:17040–17051
neuromodulation was shown to strengthen the Amit DJ, Brunel N (1997) Model of global spontaneous
gamma-band oscillatory component in the syn- activity and local structured activity during delay
thesized bulbar local field potential. periods in the cerebral cortex. Cerebral Cortex
7:237–252
A neuromodulation phenomenon has also Battaglia D, Hansel D (2006) Synchronous chaos and
been linked to attentional modulation of the neu- broad band gamma rhythm in a minimal multi-layer
ral dynamics. In a few computational studies, the model of primary visual cortex. PLoS Comput Biol
effect of attention on basic stimulus processing 7:e1002176
Börgers C, Epstein S, Kopell NJ (2005) Background
has been demonstrated using a biased competi- gamma rhythmicity and attention in cortical local cir-
tion approach. The basic idea is that a slight cuits: a computational study. Proc Natl Acad Sci
additional external input to one of the 102:702–707
competing neural populations (attended) is Börgers C, Epstein S, Kopell NJ (2008) Gamma oscilla-
tions mediate stimulus competition and attentional
amplified and the activity in the other population selection in a cortical network model. Proc Natl Acad
(non-attended) is suppressed (Deco and Thiele Sci 105:18023–18028
2009). A desirable feature of these attention Brunel N, Hakim V (1999) Fast global oscillations in
models is the emergence of synchronization of networks of integrate-and-fire neurons with low firing
rates. Neural Comput 11:1621–1671
neural activity in the gamma band consistently Buehlmann A, Deco G (2008) The neuronal basis of
with experimental data (Fries at al. 2001). attention: rate versus synchronization modulation.
Such attention-induced effects of synchrony J Neurosci 28:7679–7686
Buia C, Tiesinga P (2006) Attentional modulation of firing cortex. PLoS Comput Biol 4(9):e1000169.
rate and synchrony in a model cortical network. doi:10.1371/journal.pcbi.1000169
J Comput Neurosci 20:247–264 Li G, Cleland TA (2013) A two-layer biophysical model
Buzsaki G, Wang XJ (2012) Mechanisms of gamma oscil- of cholinergic neuromodulation in olfactory bulb.
lations. Annu Rev Neurosci 35:203–225 J Neurosci 33:3037–3058
Canolty RT, Edwards E, Dalai SS, Soltani M, Nagarajan Liljenström H, Hasselmo ME (1995) Cholinergic modu-
SS, Kirsch HE, Berger MS, Barbaro NM, Knight RT lation of cortical oscillatory dynamics. J Neurophysiol
(2006) High gamma power is phase-locked to theta 74:288–297
oscillations in human neocortex. Science Lundqvist M, Compte A, Lansner A (2010) Bistable, irreg-
313:1626–1628 ular firing and population oscillations in a modular
Chrobak JJ, Buzsaki G (1998) Gamma oscillations in the attractor memory network. PLoS Comp Biol 6:
entorhinal cortex of the freely behaving rat. J Neurosci e1000803
18:388–398 Lundqvist M, Compte A, Lansner A (2010) Bistable,
Compte A, Sanchez-Vives MV, McCormick DA, Wang irregular firing and population oscillations in a modu-
X-J (2003) Cellular and network mechanisms of slow lar attractor memory network. PLoS Comput Biol 6(6):
oscillatory activity (<1 Hz) and wave propagations in e1000803. doi:10.1371/journal.pcbi.1000803
a cortical network model. J Neurophysiol Lundqvist M, Herman P, Lansner A (2011) Theta and
90:3441–3454 gamma power increases and alpha/beta power
Deco G, Thiele A (2009) Attention – oscillations and decreases with memory load in an attractor network
neuropharmacology. Eur J Neurosci 30:347–354 model. J Cogn Neurosci 23:3008–3020
Destexhe A, Contreras D, Sejnowski TJ, Steriade M Lundqvist M, Herman PA, Palva M, Palva S, Silverstein D,
(1994) Modeling the control of reticular thalamic Lansner A (2013a) Stimulus detection rate and latency,
oscillations by neuromodulators. Neuroreport 5 firing rates and 1-40Hz oscillatory power are modulated
(17):2217–2220 by infra-slow fluctuations in a bistable attractor network
Fellous JM, Linster C (1998) Computational models of model. Neuroimage 83:458–471
neuromodulation. Neural Comput 10(4):771–805 Lundqvist M, Herman PA, Lansner A (2013b) Effect of
Fries P, Reynolds JH, Rorie AE, Desimone R (2001) prestimulus alpha power, phase, and synchronization
Modulation of oscillatory neural synchronization by on stimulus detection rates in a biophysical attractor
selective visual attention. Science 291:1560–1563 network model. J Neurosci 33:11817–11824
Fuentemilla L, Penny WD, Cashdollar N, Bunzeck N, Lundqvist M, Herman PA, Lansner A (2013c) Attractor
Duzel E (2010) Theta-coupled periodic replay in hypothesis of associative cortex: insights from a
working memory. Curr Biol 20:1–7 biophysically detailed network model. In: Signorelli
Gu Y, Liljenström H (2007) A neural network model of F (ed) Functional brain mapping and the endeavor to
attention modulated neurodynamics. Cogn Neurodyn understand the working brain. ISBN: 978-953-51-
1:275–285 1160-3, InTech, Rijeka, Croatia, DOI: 10.5772/56229
Herman PA, Lundqvist M, Lansner A (2013) Nested theta Marder E (2012) Neuromodulation of neuronal circuits:
to gamma oscillations and precise spatiotemporal fir- back to the future. Neuron 76:1–11
ing during memory retrieval in a simulated attractor Marder E, Thirumalai V (2002) Cellular, synaptic and net-
network. Brain Res 1536:68–87 work effects of neuromodulation. Neural Netw
Jacobs J, Kahana M (2009) Neural representations of 15:479–493
individual stimuli in humans revealed by gamma- Mongillo G, Barak O, Tsodyks M (2008) Synaptic theory
band electrocortigraphic activity. J Neurosci of working memory. Science 319:1543–1546
19:10203–10214 Monto S, Palva S, Voipio J, Palva JM (2008) Very slow
Jensen O, Colgin LL (2007) Cross-frequency coupling EEG fluctuations predict the dynamics of stimulus
between neuronal oscillations. Trends Cogn Sci detection and oscillation amplitudes in humans.
11:267–269 J Neurosci 28:8268–8272
Jones SR, Pritchett DL, Sikora MA, Stufflebeam SM, Neymotin SA, Lazarewicz MT, Sherif M, Contreras D,
H€am€al€ainen M, Moore CI (2009) Quantitative analysis Finkel LH, Lytton WW (2011) Ketamine disrupts theta
and biophysically realistic neural modeling of the modulation of gamma in a computer model. J Neurosci
MEG mu rhythm: rhythmogenesis and modulation of 31:11733–11743
sensory-evoked responses. J Neurophysiol Raghavachari S, Kahana MJ, Rizzuto DS, Caplan JB,
102(6):3554–3572 Kirschen MP, Bourgeois B (2001) Gating of human
Kendrick KM, Zhan Y, Nicol AU, Zhang X, Feng J (2011) theta oscillations by a working memory task.
Learning alters theta amplitude, theta-gamma cou- J Neurosci 21:3175–3183
pling and neuronal synchronization in inferotemporal Rotstein HG, Pervouchine DD, Acker CD, Gillies MJ,
cortex. BMC Neurosci 12:55 White JA, Buhl EH, Whittington MA, Kopell N
Kramer MA, Roopun AK, Carracedo LM, Traub RD, (2005) Slow and fast inhibition and an H-current inter-
Whittington MA, Kopell N (2008) Rhythm generation act to create a theta rhythm in a model of CA1 inter-
through period concatenation in rat somatosensory neuron network. J Neurophysiol 94:1509–1518
Computational Models of Motor Pools 757 C
Sandberg A, Tegnér J, Lansner A (2003) A working mem- network. Efforts at modeling this fundamentally
ory model based on fast Hebbian learning. Network important system date back more than 50 years,
14(4):789–802
Tiesinga PHE, Fellous J-M, José JV, Sejnowski TJ (2001) to Wilfred Rall’s early interest in the
Computational model of carbachol-induced delta, input–output relations of the monosynaptic reflex
theta and gamma oscillations in hippocampus. Hippo- arc (Rall 1955). But the system is inherently
campus 11:251–274 a neuromechanical one and the first models of
Wang Y, Markram H, Goodman PH, Berger TK, Ma J,
Goldman-Rakic PS (2006) Heterogeneity in the pyra- motor unit pools in the 1960s and 1970s, C
midal network of the medial prefrontal cortex. Nat Henneman and Burke, were based solely on
Neurosci 9:534–542 recruitment of motor units at their maximum
White JA, Banks MI, Pearce RA, Kopell NJ (2000) Net- forces to clearly illustrate the neuromechanical
works of interneurons with fast and slow
y-aminobutyric acid type A (GABAA) kinetics pro- nature of the system (Kernell 2006). The key
vide substrate for mixed gamma-theta rhythm. Proc point is the great diversity of motor unit proper-
Natl Acad Sci 14:8128–8133 ties and their organization via Henneman’s size
Whittington MA, Traub RD, Kopell N, Ermentrout B, principle (Henneman and Mendell 1981). Burke
Buhl EH (2000) Inhibition-based rhythms: experimen-
tal and mathematical observations on network dynam- defined motor unit types to express this diversity
ics. Int J Psychophys 38:315–336 (S (slow), FR (fast, fatigue resistant), and FF
(fast, fatigable)) (Burke 1981), but all electrical
and mechanical properties exhibit a continuous
variation (Heckman and Enoka 2012). S motor
units have motoneurons with low thresholds to
Computational Models of Motor synaptic activation, innervate relatively few mus-
Pools cle fibers, and thus produce small force. These
fibers are all slow but have tremendous fatigue
C. J. Heckman resistance. FRs and FFs have progressively
Feinberg School of Medicine, Northwestern higher thresholds for their motoneurons, greater
University, Chicago, IL, USA forces and speeds, but lower fatigue resistances
for their muscle units. These differences are not
small: threshold varies tenfold, force more than
Definition tenfold, speed fivefold, and fatigue more than
tenfold (Kernell 2006). The recruitment models
The motor unit comprises the motoneuron, its by Henneman established that this sequence is
axon, and all the muscle fibers innervated by the optimal for fatigue resistance, energy efficiency,
axon’s terminal branches (Heckman and Enoka and precision of motor output. Burke and col-
2012). The muscle fibers for one motor unit are leagues used this recruitment model approach to
sometimes referred to as the muscle unit. Each show how usage of motor units compared to the
muscle is innervated by a pool of motoneurons in forces and speeds required in a wide range of
the ventral portion of the spinal cord or, for the movements. Posture can be accomplished by
face and neck, in the brainstem. Thus, the motor recruitment solely of S units, while FF units
unit pool comprises all the motoneurons in one only become important for high speeds and
motor nucleus and the muscle innervated by those forces.
motoneurons. In contrast, a motor pool only More detailed models of motor unit pools by
includes the motoneurons. Heckman and Binder (1991) and by Fuglevand
and colleagues (1993) in the early 1990s included
rate modulation and revealed that the overall
Detailed Description form of the motor pool input–output function
was sigmoidal, with progressive recruitment of
The motor unit pool has long been an appealing larger motor units at low levels giving way to
target for computational modeling of a neural saturation of forces at high levels. The Heckman
C 758 Computational Models of Neural Retina
and Binder work was designed to understand how Henneman E, Mendell LM (1981) Functional organiza-
the organization of synaptic input to the system tion of motoneuron pool and its inputs. In: Brooks VB
(ed) Handbook of physiology, The nervous system,
affected motor output; this approach continues motor control. American Physiological Society,
and has lead to improved understanding of how Bethesda, pp 423–507
neuromodulatory inputs that alter the electrical Kernell D (2006) The motoneuron and its muscle fibres.
properties of motoneurons are fundamental in Oxford University Press, Oxford
Rall W (1955) A statistical theory of monosynaptic
shaping normal motor output (Heckman and input–output relations. J Cell Comp Physiol
Enoka 2012). The work by Fuglevand and col- 46:373–412
leagues has inspired many studies of the structure Venugopal S, Hamm TM, Crook SM, Jung R (2011) Mod-
of the output in human subjects, such as relations ulation of inhibitory strength and kinetics facilitates
regulation of persistent inward currents and motoneu-
between force and EMG (Heckman and Enoka ron excitability following spinal cord injury.
2012). Models have now progressed to where the J Neurophysiol 106:2167–2179
distortions in motor units that occur in injury and
neurodegenerative diseases are beginning to be
explored (e.g., Elbasiouny et al. 2010; Venugopal
et al. 2011). In all these efforts, a major advantage Computational Models of Neural
of this system is the close correspondence Retina
between theory and experiment that it affords.
Many of the key parameters of motoneurons can Socrates Dokos
be measured by intracellular recordings in animal Graduate School of Biomedical Engineering,
preparations, while all of the output variables, University of New South Wales, Sydney, NSW,
including not only muscle force and EMG but Australia
also firing patterns of single motor units, can be
measured in humans. Thus, motor pool computa-
tional models continue to provide unique insights Synonyms
into the structure of motor commands in both
normal and pathological states. Computational models of retinal function; Reti-
nal network models
References
Definition
Burke RE (1981) Motor units: anatomy, physiology, and
functional organization. In: Brooks VB (ed) Handbook
of physiology, The nervous system, motor control.
Computational models of the neural retina
American Physiological Society, Bethesda, simulate the response of the retina to input
pp 345–422 light. In their most detailed form, the models
Elbasiouny SM, Amendola J, Durand J, Heckman CJ yield neural output as a spatially varying
(2010) Evidence from computer simulations for alter-
ations in the membrane biophysical properties and
pattern of spike trains which fully encode the
dendritic processing of synaptic inputs in mutant incident dynamic image. In vivo, this retinal
superoxide dismutase-1 motoneurons. J Neurosci output is relayed to higher vision processing
30:5544–5558 centers in the brain such as the lateral
Fuglevand AJ, Winter DA, Patla AE (1993) Models of
recruitment and rate coding organization in motor-unit
geniculate nucleus and the primary visual cor-
pools. J Neurophysiol 70:2470–2488 tex in humans and primates. Neural retinal
Heckman CJ, Binder MD (1991) Computer simulation of models assist not only in the understanding of
the steady-state input–output function of the cat medial normal and abnormal retinal function, but are
gastrocnemius motoneuron pool. J Neurophysiol
65:952–967
useful for investigating retinal response to arti-
Heckman CJ, Enoka RM (2012) Motor unit. Compr Phys ficial stimuli such as electrical stimulation by
2:2629–2682 a vision prostheses.
Computational Models of Neural Retina 759 C
Detailed Description light stimulus s(r,t) (r denotes the 2D spatial
position and t time), an RGC stimulus current
The vertebrate retina consists of a thin layer of Is(r,t) is generated according to a linear convolu-
neural tissue which lines the back of the eye. It tion operation:
acts to transduce incoming light signals sensed by
ð ð1
rod and cone photoreceptor cells into neural spike
I s ðr, tÞ ¼ hðj, tÞsðr j, t tÞdjdt C
patterns output by retinal ganglion cells (RGCs), ℝ2 0
whose axons together form the optic nerve. The
ratio of photoreceptor cells to RGCs is approxi- where h(j, t) denotes the spatiotemporal convo-
mately 100:1 (dependent on eccentricity from the lution kernel which may feature a DoG spatial
fovea), indicating that considerable neural characteristic (McLaughlin et al. 2000). The spa-
processing occurs in the retina itself prior to the tial integral in the variable j is carried out over
transfer of visual information via the optic nerve ℝ2, and the lower bound of the time integral in
to the brain. variable t is zero, since the kernel is assumed to
Each RGC is centered on an approximately be causal: that is, future values of the light stim-
circular receptive field of photoreceptors within ulus do not affect the output at time t. The stim-
an associated annular surround. Incident light fall- ulus current Is(r,t) represents total RGC synaptic
ing within these regions will have opposite effects input at position r and time t. This input is inte-
on RGC activation, eliciting two main types of grated with respect to time and when the inte-
response depending on the cell type. On-center grated signal exceeds a given threshold, an
RGCs are activated by light falling within their RGC spike is generated and the integrator is
receptive field and inhibited by light falling in their reset to zero. The resulting RGC output neural
surround, while the reverse holds for off-center response function consists of a train of unit
RGCs. Both these spatial RGC responses can be impulses given by
mathematically modeled using a simple difference
of Gaussians (DoG) (Rodieck 1965). X
1
rðr, tÞ ¼ d t t r, i
Computational models of the retinal response i¼1
to light may be broadly classified into integrate
and fire, block-structured, and network models. where d(t) is the Dirac delta function and tr. i is
The models vary in complexity from simulating the time of the ith spike at position r. Some level
the response of a single retinal output neuron to of stochasticity in spike timing can be obtained
incident light, through to simulating a spatiotem- by updating the integrator threshold after each
poral neural response to a 2D time-varying spike from a random distribution of values. Fur-
image. thermore, to account for a finite RGC memory,
Is(r,t) can be pre-multiplied by an exponentially
Integrate and Fire Models decaying time window function prior to integra-
The simplest neural retina models are of the inte- tion: this is the basis of leaky integrate and fire
grate and fire type (Gestri et al. 1980; Reich models. In such models, Is values further back in
et al. 1997), shown in Fig. 1 fully generalized to time contribute less to the generation of
the spatiotemporal domain. For a given input a subsequent output spike.
Computational Models
of Neural Retina,
Fig. 1 Generalized
spatiotemporal integrate
and fire retinal model
C 760 Computational Models of Neural Retina
ð1
Block-Structured Models
CeD qðr, t t dt
t
vðr, tÞ ¼
Block-structured models of the retina simulate 80
the response of output retinal neurons to input < 1
, if vðr, tÞ > 0
gðr, tÞ ¼ 4
: 1 þ ½vðr, tÞ
light using a combination of linear and nonlinear
input/output functional block elements. In con- 1, otherwise
trast to integrate and fire models, no block has its
output reset when a neural spike is triggered. In where C and D are parameters governing the
one such model, the input light signal s(t) is low-pass temporal filter characteristics. Modula-
passed through a linear filter to produce tion of neural activation by g(r,t) corresponds to
a generator potential g(t), which corresponds to the phenomenon of contrast gain control,
the RGC membrane voltage (Keat et al. 2001). whereby RGCs fire with increased frequency
An output spike is triggered whenever a positive- when the input light stimulus changes rapidly
going g(t) value crosses a fixed threshold. (Wilke et al. 2001).
A feedback loop in the form of negative after The output of this retinal model is the RGC
potentials can be used to ensure that the model scalar firing rate f(r,t), expressed as the number of
produces clustered spikes, as observed in the real spikes per unit time. This rate can be converted
retina. into an output spike train by assigning the
A more complex block-structured architecture interspike interval t to a random variable drawn
of the retina is shown in Fig. 2 (Wilke et al. 2001). from an inhomogeneous Poisson distribution
In this model, a spatiotemporal light stimulus with probability density function:
s(r, t) is transformed through linear convolution
into a neural activation signal u(r,t) according to
ð ð1 pðr, tÞjt ¼ f ðr, tÞef ðr, tÞt
uðr, tÞ ¼ hðj, tÞsðr j, t tÞdjdt . This
ℝ2 0 A remaining class of block-structured
signal is then multiplied by a local modulation models of the retina is based on a cascaded
factor g(r,t) between 0 and 1, scaled, and rectified architecture consisting of a dynamic linear con-
to produce the output local RGC firing rate f(r,t): volution input/output block followed by a static
nonlinearity, also known as a Wiener system
qðr, tÞ ¼ gðr, t uðr, t (Juusola et al. 1995; Chichilnisky 2001). An
Aqðr, tÞ, if Aqðr, tÞ > B advantage of these models is that several itera-
f ðr, tÞ ¼
0, otherwise tive techniques are available to identify constit-
uent block characteristics from input/output
where A is the output rate scaling factor and B the experimental data (Westwick and Kearney
basal firing rate. The g(r,t) modulation factor 2003).
itself is determined by a linear convolution of
q(r,t) with a low-pass temporal filter to determine Network Models
an intermediate signal v(r,t), which is then fed Network models of the retina consist of coupled
through a static nonlinearity to yield g(r,t): systems of discrete neural elements whose
of Neural Retina,
Fig. 2 Block-structured
spatiotemporal model of
Wilke et al. (2001)
Computational Models of Neuromodulation 761 C
outputs modulate other elements through excita- Keat J, Reinagel P, Reid RC, Meister M (2001) Predicting
tion or inhibition. The simplest such retinal net- every spike: a model for the responses of visual neu-
rons. Neuron 30:803–817
work model is the Hartline-Ratliff formulation McLaughlin D, Shapley R, Shelley M, Wielaard DJ
(Hadeler and Kuhn 1987), in which the firing (2000) A network neuronal model of macaque primary
rate of neuron j, fj, is given by visual cortex (V1): orientation selectivity and dynam-
ics in the input layer 4Ca. Proc Natl Acad Sci U S
! A 97:8087–8092
X
n
Reich DS, Victor JD, Knight BW, Ozaki T, Kaplan
C
f j ¼ bj max 0, ej bjk f k E (1997) Response variability and timing precision of
k¼0 neuronal spike trains in vivo. J Neurophysiol
77:2836–2841
where ej is the excitatory input to neuron j due to Rekeczky C, Roska B, Nemeth E, Werblin FS (2001) The
incident light, bj is a scaling factor, and coeffi- network behind spatio-temporal patterns: building
low-complexity retinal models in CNN based on mor-
cient bjk characterizes the inhibitory action of phology, pharmacology and physiology. Int J Circuit
neuron k on neuron j. The max() function acts Theory Appl 29:197–239
as a rectifier, preventing negative values of Rodieck RW (1965) Quantitative analysis of cat retinal
firing rate. ganglion cell response to visual stimuli. Vision Res
5:583–601
More complex retinal network models are Westwick DT, Kearney RE (2003) Identification of
based on ionic membrane conductance formula- nonlinear physiological systems. IEEE Press, New
tions of discrete neurons and their synaptic Jersey
interconnections (Rekeczky et al. 2001; Cottaris Wilke SD, Thiel A, Eurich CW, Greschner M, Bongard M,
Ammerm€ uller J, Schwegler H (2001) Population cod-
and Elfar 2005; Wohrer and Kornprobst 2009). ing of motion patterns in the early visual system.
Despite the fact that biophysically detailed J Comp Physiol A 187:549–558
ionic RGC models are available (Fohlmeister Wohrer A, Kornprobst P (2009) Virtual retina:
and Miller 1997), network models of the whole a biological retina model and simulator, with contrast
gain control. J Comput Neurosci 26:219–249
retina adopt simplified ionic formulations for
computational efficiency. Such complex network
Further Reading
models can incorporate multiple neural layers of Martins JC, Sousa LA (2009) Bioelectronic vision: retina
the retina with up to 100,000 or more discrete models, evaluation metrics, and system design. World
neurons. Scientific, New Jersey
References
Computational Models of
Chichilnisky EJ (2001) A simple white noise analysis of
neuronal light responses. Network: Comput Neural
Neuromodulation
Syst 12:199–213
Cottaris NP, Elfar SD (2005) How the retinal network Angela J. Yu
reacts to epiretinal stimulation to form the prosthetic Department of Cognitive Science, University of
visual input to the cortex. J Neural Eng 2:S74–S90
Fohlmeister JF, Miller RF (1997) Impulse encoding mech-
California, San Diego, La Jolla, CA, USA
anisms of ganglion cells in the tiger salamander retina.
Gestri G, Mastebroek HAK, Zaagman WH (1980) Sto- Definition
chastic constancy, variability and adaptation of spike
generation: performance of a giant neuron in the visual
system of the fly. Biol Cybern 38:31–40 Neuromodulatory systems serve a special meta-
Hadeler KP, Kuhn D (1987) Stationary states of the processing role in the brain. Due to their anatom-
Hartline-Ratliff model. Biol Cybern 56:411–417 ical privileges of having massively extensive
Juusola M, Weckstrom M, Uusitalo RO, Korenberg MJ,
French AS (1995) Nonlinear models of the first syn-
arborization patterns throughout the central and
apse in the light-adapted fly retina. J Neurophysiol peripheral nervous systems and to their physio-
74:2538–2547 logical capacity of finely controlling how other
C 762 Computational Models of Neuromodulation
neurons communicate with each other and plas- (Sutton 1988), which can learn not only which
ticize, they are ideally positioned to regulate the stimuli predict future rewards but also the tem-
way information is acquired, processed, utilized, poral delay until such a reward appears. Compu-
and stored in the brain. As such, neuromodulation tational models have suggested reward prediction
has been a fertile ground for computational error to be important for at least three different
models for neural information processing, which functions. First, the most obvious, is the learning
have strived to explain not only how the major of environmental cues that predict future rewards
neuromodulators coordinate to enable normal as well as actions that lead to greater or lesser
sensory, motor, and cognitive functions but also amount of reward. There is evidence that D1
how dysfunctions arise in psychiatric and neuro- dopamine receptors in the direct or “go” pathway
logical conditions when these neuromodulatory in the striatum mediate the facilitation of actions
systems are impaired. with surprisingly good outcomes or positive pre-
diction errors, whereas D2 dopamine receptors in
the indirect or “no-go” pathway in the striatum
Detailed Description underlie the inhibition of actions that lead to
negative prediction errors (Gerfen et al. 1990;
Although much still remains unknown or Smith et al. 1998; Frank et al. 2004; Frank
opaque about neuromodulatory functions, a 2005; Cohen and Frank 2009; Kravitz
computationally sophisticated and coherent pic- et al. 2012). A second role suggested for dopa-
ture is beginning to emerge on the back of recent mine, related to the theory of incentive salience
modeling works (Dayan 2012). The four major (Berridge and Robinson 1998; McClure
neuromodulators that have received most theo- et al. 2003) and potentially via D1 receptors in
retical treatment are dopamine, serotonin, acetyl- the nucleus accumbens (Murschall and Hauber
choline, and norepinephrine. They all play 2006; Frank 2005; Surmeier et al. 2007; Surmeier
important roles in regulating cortical functions et al. 2010), is that it controls the expression of
such as inference, learning, and decision-making. Pavlovian approach responses associated with
Broadly speaking, dopamine and serotonin the prospect of reward (Ikemoto and
appear to be critical for the processing of stimuli Panksepp 1999), in turn facilitating the initiation
and events with salient affective value, such as and learning of instrumental responses that boost
rewards and punishments, and the learning and reward (Satoh et al. 2003; Nakamura and
execution of actions that lead to greater subjec- Hikosaka 2006; Talmi et al. 2008). A third role
tive utility (more rewards or fewer punishments); suggested for the dopamine-mediated prediction
in contrast, acetylcholine and norepinephrine error is that over the long term, the average pre-
appear to be most critical for signaling various diction error reflects the average reward rate,
forms of uncertainty that plague cortical compu- which represents an opportunity cost that can
tations, arising from intrinsic neuronal processing drive the level of response vigor (Niv
noise, external environmental changeability, as et al. 2007) or the trade-off between speed and
well as imperfect sensory observations. accuracy (Dayanik and Yu 2012).
While animals need to seek out rewarding
circumstances and actions, they also need to
Reward and Punishment avoid punishing or threatening ones. The learning
of avoidance behavior, or behavioral inhibition,
One of the best understood computational func- appears to depend on both dopamine and seroto-
tions of neuromodulation is the role of dopamine nin (5-HT). Unexpected punishments (Ungless
(DA) in signaling reward prediction errors et al. 2004; Cohen et al. 2012) and the
(Montague et al. 1996; Schultz et al. 1997), nondelivery of expected reward (Schultz
associated with a reinforcement learning algo- et al. 1997) have both been shown to activate
rithm known as temporal difference learning the dopaminergic system, which enable the
learning of avoidance behavior via D2 receptors a stimulus or event should upregulate the learn-
in the indirect pathway of the striatum (Frank ing accorded to that stimulus/event, relative to
et al. 2004; Kravitz et al. 2012). Complementing other better understood aspect of the environment
the dopaminergic system, serotonin has been (Yu and Dayan 2005b; Behrens et al. 2007;
suggested to play an opponent role, by signaling Nassar et al. 2012). Strictly speaking, exact
prediction errors associated with future Bayesian statistical computations need not distin-
“version” rather than reward (Deakin 1983; guish among the various forms of uncertainty, as C
Daw et al. 2002; Cools 2011; Boureau and they are all reflected as probability distributions,
Dayan 2011). The suggestion is that initial hint or properties of distributions, of the relevant ran-
of future aversion would activate the serotonergic dom variables under consideration. However, the
prediction error (Schweimer and Ungless 2010) brain can hardly be expected to implement arbi-
but that once the threat of that aversive outcome trarily precise and complex Bayesian computa-
is avoided due to actions or circumstances, then tions, given its limited processing hardware, and
the dopaminergic neurons would be activated by thus needs to represent the most common and
the associated reward prediction error, signaling critical aspects of uncertainty so as to achieve
safety, and thus reinforce the avoidance action near-optimal but efficient computations (Yu and
(Johnson et al. 2001; Moutoussis et al. 2008; Dayan 2003, 2005a, b).
Maia 2010). Aside from learning avoidance Based on a wealth of physiological, pharma-
actions, serotonin has also been thought to be cological, and behavioral data implicating acetyl-
important for mitigating impulsivity and enabling choline (ACh) and norepinephrine (NE) in
general behavioral inhibition, such as delaying specific aspects of a range of cognitive processes,
actions in order to receive a larger reward it has been proposed that they respectively signal
(Miyazaki et al. 2011; Doya 2002). expected and unexpected uncertainty, with the
former associated with known stochasticity or
ignorance in the current environment and the
Uncertainty latter arising from unexpected gross changes in
the environment that require drastic revision of
Uncertainty of various forms plagues the brain’s internal representation of the behavioral context
interactions with the environment. In a Bayesian (Yu and Dayan 2005b). The two forms of uncer-
statistical framework, optimal inference and pre- tainty are expected to be partly synergistic and
diction, based on unreliable observations in partly antagonistic: a rise in unexpected uncer-
changing contexts, require the representation tainty should lead to an increase of expected
and utilization of different forms of uncertainty. uncertainty (due to known ignorance), while
For example, in the context of multimodal cue expected uncertainty in turn has a suppressive
integration, uncertainty about the state of the influence on unexpected uncertainty, since the
world induced by unreliability of a cue should level of evidence for a fundamental change
lead to the relative downgrading of that source of induced by a certain magnitude of prediction
information relative to other more reliable cues error should normalized by the degree of
(Ernst and Banks 2002; Battaglia et al. 2003). expected variability (prediction error) (Yu and
Analogously, in the context of integrating Dayan 2005b). Consistent with this, Sara and
top-down and bottom-up information, uncer- colleagues have found similarly antagonistic
tainty about the validity of an internal model of interactions between ACh and NE in a series of
the environment should lead to relative suppres- learning and memory studies (Sara 1989;
sion of top-down expectations in relation to Ammassari-Teule et al. 1991; Sara et al. 1992;
bottom-up, novel sensory inputs (Yu and Dayan Dyon-Laurent et al. 1993, 1994). They demon-
2003; Körding and Wolpert 2004; Yu and Dayan strated that learning and memory deficits
2005b). On the other hand, uncertainty about the caused by cholinergic lesions can be
predictive or other statistical properties of alleviated by the administration of clonidine
C 764 Computational Models of Neuromodulation
(Sara 1989; Ammassari-Teule et al. 1991; Sara References

et al. 1992; Dyon-Laurent et al. 1993, 1994),
a noradrenergic a-2 agonist that decreases the Ammassari-Teule M, Maho C, Sara SJ (1991) Clonidine
reverses spatial learning deficits and reinstates y fre-
level of NE (Coull et al. 1997).
quencies in rats with partial fornix section. Behav
In recent years, a number of experimental Brain Res 45:1–8
studies have confirmed major aspects of the Aston-Jones G, Cohen JD (2005) An integrative theory of
uncertainty-based account of acetylcholine and locus coeruleus-norepinephrine function: adaptive
gain and optimal performance. Annu Rev Neurosci
norepinephrine functions. Pupil diameter,
28:403–450
known to be correlated with noradrenergic neu- Battaglia PW, Jacobs RA, Aslin RN (2003) Bayesian inte-
ronal activation level (Aston-Jones and Cohen gration of visual and auditory signals for spatial local-
2005), has been shown to be, on the one hand, ization. J Opt Soc Am A Opt Image Sci Vis
20(7):1391–1397
correlated with the discarding of previous beliefs
Behrens TEJ, Woolrich MW, Walton ME, Rushworth
and strategies and the exploration of novel ones MFS (2007) Learning the value of information in an
(Gilzenrat et al. 2010; Jepma and Nieuwenhuis uncertain world. Nature Neurosci 10(9):1214–1221
2011) and, on the other hand, reflective of sub- Berridge KC, Robinson TE (1998) What is the role of
dopamine in reward: hedonic impact, reward learning,
jects unexpected uncertainty beyond predicted
or incentive salience? Brain Res Rev 28:309–369
variability Preuschoff11. Moreover, there is evi- Boureau Y-L, Dayan P (2011) Opponency revisited: com-
dence that pupil diameter is reflective of unex- petition and cooperation between dopamine and sero-
pected uncertainty (operationalized here as tonin. Neuropsychopharmacology 36:74–97
Cohen JD, Frank MJ (2009) Neurocomputational models
change point probability) and accompanies the
of basal ganglia function in learning, memory and
discarding of previous beliefs and the incorpora- choice. Behav Brain Res 199:141–156
tion of new data into a novel belief state (Nassar Cohen JY, Haesler S, Vong L, Lowell BB, Uchida
et al. 2012). Interestingly, pupil dilation is con- N (2012) Neuron-type-specific signals for reward and
punishment in the ventral tegmental area. Nature
trolled by two muscles, the sphincter for contrac-
482:85–88
tion and the dilator for dilation, which are Cools R (2011) Dopaminergic control of the striatum for
respectively innervated by acetylcholine and high-level cognition. Curr Opin Neurobiol
norepinephrine (Hutchins and Corbett 1997). 21:402–407
Córdova, Yu, Chiba (2004) Soc Neurosci Abstr
The administration of cholinergic antagonists,
Coull JT, Frith CD, Dolan RJ, Frackowiak RS, Grasby PM
scopolamine (muscarinic) and mecamylamine (1997) The neural correlates of the noradrenergic mod-
(nicotinic), increases pupil size in normal elderly ulation of human attention, arousal and learning. Eur
volunteers (Little et al. 1998). Moreover, J Neurosci 9(3):589–598
Daw ND, Kakade S, Dayan P (2002) Opponent interac-
Alzheimer’s disease (AD) patients, who have
tions between serotonin and dopamine. Neural Netw
a severe and specific cholinergic deficit, have 15:603–616
larger pupil size than controls both tonically Dayan P (2012) Twenty-five lessons from computational
and in response to light, while AD patients neuromodulation. Neuron 76:240–256
Dayanik S, Yu AJ (2012) Reward-rate maximization in
medicated with acetylcholinesterase inhibitor,
sequential identification under a stochastic deadline.
which increases the endogenous level of acetyl- SIAM J Control Optim 51(4):2922–2948
choline, show a pupil response much closer to Deakin JFW (1983) Roles of brain serotonergic neurons
controls (Fotiou et al. 2009). This raises the in escape, avoidance and other behaviors.
J Psychopharmacol 43:563–577
intriguing possibility that the pupil dilation
Doya K (2002) Metalearning and neuromodulation. Neu-
results might reflect the joint influence of both ral Netw 15(4–6):495–506
norepinephrine and acetylcholine. Separately, Dyon-Laurent C, Romand S, Biegon A, Sara SJ
pharmacological and lesion studies directly (1993) Functional reorganization of the noradrenergic
system after partial fornix section: a behavioral and
manipulating acetylcholine levels have also
autoradiographic study. Exp Brain Res 96:203–211
produced behavioral results consistent with the Dyon-Laurent C, Hervé A, Sara SJ (1994) Noradrenergic
cholinergic expected uncertainty hypothesis hyperactivity in hippocampus after partial denerva-
(Thiel and Fink 2008; Córdova and Chiba tion: pharmacological, behavioral, and electrophysio-
logical studies. Exp Brain Res 99:259–266
2004).
Ernst MO, Banks MS (2002) Humans integrate visual and Moutoussis M, Bentall RP, Williams J, Dayan P (2008)
haptic information in a statistically optimal fashion. A temporal difference account of avoidance learning.
Nature 415(6870):429–433 Network 19:137–160
Fotiou DF, Stergiou V, Tsiptsios D, Lithari C, Nakou M, Murschall A, Hauber W (2006) Inactivation of the ventral
Karlovasitou A (2009) Cholinergic deficiency in tegmental area abolished the general excitatory influ-
Alzheimer’s and Parkinson’s disease: evaluation with ence of Pavlovian cues on instrumental performance.
pupillometry. Int J Psychophysiol 73(2):143–149 Learn Mem 13:123–126
Frank MJ (2005) Dynamic dopamine modulation in the Nakamura K, Hikosaka O (2006) Role of dopamine in the
basal ganglia: a neurocomputational account of cogni- primate caudate nucleus in reward modulation of sac-
C
tive deficits in medicated and nonmedicated Parkin- cades. J Neurosci 26:5360–5369
sonism. J Cogn Neurosci 17:51–72 Nassar MR, Rumsey KM, Wilson RC, Parikh K, Heasly B,
Frank MJ, Seeberger LC, OReilly RC (2004) By carrot or Gold JI (2012) Rational regulation of learning dynam-
by stick: cognitive reinforcement learning in Parkin- ics by pupil-linked arousal systems. Nat Neurosci
sonism. Science 306:1940–1943 15:1040–1046
Gerfen CR, Engber TM, Mahan LC, Susel Z, Chase TN, Niv Y, Daw ND, Joel D, Dayan P (2007) Tonic dopamine:
Monsma FJJ, Sibley DR (1990) D1 and D2 dopamine opportunity cost and the control of response vigor.
receptor-regulated gene expression of striatonigral and Psychopharmacology (Berl) 191:507–520
striatopallidal neurons. Science 250:1429–1432 Sara SJ (1989) Noradrenergic-cholinergic interaction: its
Gilzenrat MS, Nieuwenhuis S, Jepma M, Cohen JD possible role in memory dysfunction associated with
(2010) Pupil diameter tracks changes in control state senile dementia. Arch Gerontol Geriatr Suppl
predicted by the adaptive gain theory of locus 1:99–108
coeruleus function. Cogn Affect Behav Neurosci Sara SJ, Dyon-Laurent C, Guibert B, Leviel V (1992)
10:252–269 Noradrenergic hyperactivity after fornix section: role
Hutchins JB, Corbett JJ (1997) The visual system. In: in cholinergic dependent memory performance. Exp
Haines DE (ed) Fundamental neuroscience. Churchill Brain Res 89:125–132
Livingstone, New York, pp 265–284 Satoh T, Nakai S, Sat T, Kimura M (2003) Correlated
Ikemoto S, Panksepp J (1999) The role of nucleus coding of motivation and outcome of decision by
accumbens dopamine in motivated behavior: dopamine neurons. J Neurosci 23:9913–9923
a unifying interpretation with special reference to Schultz W, Dayan P, Montague PR (1997) A neural sub-
reward-seeking. Brain Res Rev 31:6–41 strate of prediction and reward. Science
Jepma M, Nieuwenhuis S (2011) Pupil diameter predicts 275:1593–1599
changes in the exploration-exploitation trade-off: evi- Schweimer JV, Ungless MA (2010) Phasic responses in
dence for the adaptive gain theory. J Cogn Neurosci dorsal raphe serotonin neurons to noxious stimuli.
23:1587–1596 Neuroscience 171:1209–1215
Johnson J, Li W, Li J, Klopf A (2001) A computational Smith Y, Bevan MD, Shink E, Bolam JP (1998) Microcir-
model of learned avoidance behavior in a one-way cuitry of the direct and indirect pathways of the basal
avoidance experiment. Adapt Behav 9:91 ganglia. Neuroscience 86:353–387
Körding KP, Wolpert DM (2004) Bayesian integration in Surmeier DJ, Ding J, Day M, Wang Z, Shen W (2007) D1
sensorimotor learning. Nature 427:244–247 and D2 dopamine receptor modulation of striatal
Kravitz AV, Tye LD, Kreitzer AC (2012) Distinct roles glutamatergic signaling in striatal medium spiny neu-
for direct and indirect pathway striatal neurons in rons. Trends Neurosci 30:228–235
reinforcement. Nat Neurosci 15:816–818 Surmeier DJ, Shen W, Day M, Gertler T, Chan S, Tian X,
Little JT, Johnson DN, Minichiello M, Weingartner H, Plotkin JL (2010) The role of dopamine in modulating
Sunderland T (1998) Combined nicotinic and musca- the structure and function of striatal circuits. Prog
rinic blockade in elderly normal volunteers: cognitive, Brain Res 183:149–167
behavioral, and physiological responses. Neuropsy- Sutton RS (1988) Learning to predict by the methods of
chopharmacology 19(1):60–69 temporal differences. Mach Learn 3(1):9–44
Maia TV (2010) Two-factor theory, the actor-critic model, Talmi D, Seymour B, Dayan P, Dolan RJ (2008) Human
and conditioned avoidance. Learn Behav 38:50–67 Pavlovian-instrumental transfer. J Neurosci
McClure SM, Daw ND, Montague PR 28:360–368
(2003) A computational substrate for incentive Thiel CM, Fink GR (2008) Effects of the cholinergic
salience. Trends Neurosci 26:423–428 agonist nicotine on reorienting of visual spatial atten-
Miyazaki K, Miyazaki KW, Doya K (2011) Activation of tion and top-down attentional control. Neuroscience
dorsal Raphe serotonin neurons underlies waiting for 152(2):381–390
delayed rewards. J Neurosci 31:469–479 Ungless MA, Magill PJ, Bolam JP (2004) Uniform inhi-
Montague PR, Dayan P, Sejnowski TJ bition of dopamine neurons in the ventral tegmental
(1996) A framework for mesencephalic dopamine sys- area by aversive stimuli. Science 303:2040–2042
tems based on predictive hebbian learning. J Neurosci Yu AJ, Dayan P (2003) Expected and unexpected uncer-
16:1936–1947 tainty: ACh and NE in the neocortex. In: Becker STS,
C 766 Computational Models of Painful Loading
Obermayer K (eds) Advances in neural information clinician in order to suppress the pathological
processing systems 15. MIT Press, Cambridge, MA, symptoms, while not inducing any unwanted
pp 157–164
Yu AJ, Dayan P (2005a) Inference, attention, and decision side effects.
in a Bayesian neural architecture. In: Saul LK, Weiss Y, Computational models can be used to estimate
Bottou L (eds) Advances in neural information the optimal parameter settings for DBS to aid this
processing systems 17. MIT Press, Cambridge, MA process.
Yu AJ, Dayan P (2005b) Uncertainty, neuromodulation,
and attention. Neuron 46:681–692
Computational Models of Painful Deep Brain Stimulation

Loading Over 20 years ago (Benabid et al. 1987), deep
brain stimulation (DBS) was introduced as
▶ Biomechanical Model of Low Back Pain a clinical therapy for a number of neurological
and more recently psychological disorders. The
treatment involves chronic electrical stimulation
via quadripolar electrodes implanted into various
regions of the human brain in a disorder-specific
Computational Models of Retinal
manner. Most commonly, DBS is used to treat
Function
a range of movement disorders such as essential
tremor, Parkinson’s disease, and dystonia
▶ Computational Models of Neural Retina
(Vidailhet et al. 2005; Deuschl et al. 2006;
Kupsch et al. 2006), with electrodes typically
implanted into regions of the thalamus or the
basal ganglia. The treatment works well, and
Computational Models of Visual 70–80 % of patients experience benefit from the
Illusions treatment (Parkinson’s UK).
▶ Visual Illusions, Models of Parameter Settings

Despite the success of the treatment, however,
there remains a lack of understanding of how
this electrical current changes neuronal activity
Computational Models Supporting to lead to the observed therapeutic benefits
Parameter Finding for Deep Brain (Benabid 2007; Kringelbach et al. 2007). This
Stimulation lack of understanding in turn restricts the clinical
efficacy of DBS, further development of the tech-
Nada Yousif nology, and identification of novel clinical appli-
Division of Brain Sciences, Imperial College cations. Of particular importance to the patient
London, London, UK and clinician, however, is the identification of the
optimal set of parameters for stimulation. These
parameters include stimulation frequency, pulse
Definition width, amplitude, and contact configuration (see
Fig. 1). This process of parameter selection for an
Deep brain stimulation is a surgical therapy individual patient is time consuming and difficult
involving electrical stimulation of the brain via (Rizzone et al. 2001; Moro et al. 2002; Kuncel
chronically implanted electrodes. and Grill 2004), relying mostly on trial and error
Once implanted, there are a number of stimu- and the intuition of an experienced clinician. If
lation parameters which must be set by the we knew how DBS was achieving the clinical
Computational Models Supporting Parameter Finding for Deep Brain Stimulation 767 C
Computational Models Supporting Parameter Find- can be configured in different ways. If monopolar stimu-
ing for Deep Brain Stimulation, Fig. 1 Once the lation is selected, a single contact is used as the cathode,
quadripolar DBS electrodes are implanted, an optimal and the case of the stimulator, which is usually implanted
set of parameters must be chosen by the clinician. The in the chest, is used as the return electrode. If bipolar
three parameters are stimulation frequency, pulse width, stimulation is selected, two contacts are used, and
and amplitude of the waveform. In addition, the electrode a near-field dipole is created
improvement at the neuronal level, patients’ shown that using neuron models compared to
symptoms could be better targeted. axon models predicts that a much smaller region
of surrounding tissue will be activated, which
FEM Models will clearly impact on the use of such models
In recent years, computational models have been for predicting parameter settings.
used to investigate this issue (Wei and Grill 2005;
Yousif and Liu 2009; Butson et al. 2011) via a Making Predictions
two-step approach: first, the spread of the electric Previously, there have been attempts to link the
field is modelled in a 3-dimensional model of the predictions made by FEM models to clinical stud-
electrode and the surrounding tissue in a finite ies or observations (Butson and McIntyre 2007;
element method (FEM) model (Fig. 2a). The Frankemolle et al. 2010; Chaturvedi et al. 2010).
potential distribution induced by stimulation can Similarly a limited amount of work has been done
then be calculated by solving the Laplace equa- in order to link model results to measurements of
tion within the geometry. the electric field in the macaque brain
(Miocinovic et al. 2009). This not only allows
Neuronal Models the correlation of results with the observed clin-
For the second step, the electric field estimation is ical improvement but can also constrain the
applied to a model of neuronal structures model parameters as the DBS-induced electric
(Fig. 2b) to predict the effect that stimulation field can be directly measured. This process of
has on the firing properties in the surrounding validation is crucial for the success of these
tissue. At present, the vast majority of such models in aiding the understanding of DBS and
models focus on arrays of unconnected axons, for establishing their credibility to clinicians who
but recent work has shown that by modelling will ideally use such models in clinical practice.
single neurons and networks of neurons, we can In recent work, the use of such models to
observe subtle effects of DBS on neural activity visualize the impact of specific parameter set-
which would be overlooked by axon models tings for an individual patient has emerged.
(Yousif et al. 2010, 2012). It has also been Yousif et al. reported a clinical case where
C 768 Computational Models Supporting Parameter Finding for Deep Brain Stimulation
Supporting Parameter
Finding for Deep Brain
Stimulation, Fig. 2 To
model the impact of DBS
on the human brain,
a two-step approach is
typically taken. (a) The
geometry of interest is
modelled in 3-dimensions
and the equation of interest
solved over this geometry
to give a potential
distribution (schematized
here in color). (b) This
potential distribution can
then be applied to
a dynamic model of neural
activity, typically axons or
neurons
a bipolar setting was not tolerated by a patient parameters can be found which will guide the
with essential tremor during the first DBS pro- clinician’s search for the optimal set of parame-
gramming session (Yousif et al. 2012). However, ters. Such systems have now started to emerge
when the polarity of the same two contacts from commercial companies. In conclusion, such
was reversed, the patient was able to tolerate a method which theoretically predicts a set of
a 70 % increase in amplitude. By utilizing a com- parameters would streamline the process, reduce
putational modelling approach to visualize the the stress and discomfort to patients and reduce
effect of the two settings on the firing of neural the time required for tuning, checking, and
fibers, it was possible to confirm that reversing adjusting parameters.
the polarity, stimulated fibers of the nearby
internal capsule. Stimulation of this fiber bundle
has been reported to induce side effects from References
a thalamic electrode (Limousin et al. 1999).
Benabid AL (2007) What the future holds for deep
Hence, such an approach was used to help the brain stimulation. Expert Rev Med Devices
clinician to understand why one configuration 4(6):895–903
was superior to another, when simple consider- Benabid AL, Pollak P et al (1987) Combined
ation of the electric field induced by each would (thalamotomy and stimulation) stereotactic surgery of
the VIM thalamic nucleus for bilateral Parkinson dis-
indicate them to be equivalent. ease. Appl Neurophysiol 50(1–6):344–346
Future work should take a prospective Butson CR, McIntyre CC (2007) Differences among
approach to parameter settings. For each patient implanted pulse generator waveforms cause variations
that is treated with DBS, an anatomically accu- in the neural response to deep brain stimulation. Clin
Neurophysiol 118(8):1889–1894
rate model of the current spread can be produced Butson CR, Cooper SE et al (2011) Probabilistic analysis
and coupled to axon and/or neuron models. Prior of activation volumes generated during deep brain
to the process of clinically finding parameters, stimulation. Neuroimage 54(3):2096–2104
such a model can then be simulated at different Chaturvedi A, Butson CR, Lempka SF, Cooper SE,
McIntyre CC (2010) Patient-specific models of deep
parameter settings, to maximally activate axons brain stimulation: influence of field model complexity
or change the firing rate of neurons within the on neural activation predictions. Brain Stimul
target brain structure. In this way, a set of 3(2):65–67
Computational Models to Optimize the Electrodes and Waveforms for Deep Brain Stimulation 769 C
Deuschl G, Schade-Brittinger C et al (2006) A randomized
trial of deep-brain stimulation for Parkinson’s disease. Computational Models to Optimize
N Engl J Med 355(9):896–908
Frankemolle A, Wu J, Noecker AM, Voelcker-Rehage C, the Electrodes and Waveforms for
Ho JC, Vitek JL, McIntyre CC, Alberts JL Deep Brain Stimulation
(2010) Reversing cognitive-motor impairments in
Parkinson’s disease patients using a computational Bryan Howell1 and Warren M. Grill1,2,3
modelling approach to deep brain stimulation pro-
gramming. Brain 133(3):746–761
1
Department of Biomedical Engineering, C
Kringelbach ML, Jenkinson N et al (2007) Translational Duke University, Durham, NC, USA
2
principles of deep brain stimulation. Nat Rev Neurosci Department of Neurobiology, Duke University
8(8):623–635 Medical Center, Durham, NC, USA
Kuncel AM, Grill WM (2004) Selection of stimulus 3
parameters for deep brain stimulation. Clin Department of Surgery, Duke University
Neurophysiol 115(11):2431–2441 Medical Center, Durham, NC, USA
Kupsch A, Benecke R et al (2006) Pallidal deep-brain
stimulation in primary generalized or segmental dys-
tonia. N Engl J Med 355(19):1978–1990
Limousin P, Speelman JD et al (1999) Multicentre Euro- Definition
pean study of thalamic stimulation in Parkinsonian and
essential tremor. J Neurol Neurosurg Psychiatry A computational model used to optimize elec-
66(3):289–296 trodes and waveforms for deep brain stimulation
Miocinovic S, Lempka SF et al (2009) Experimental and
theoretical characterization of the voltage distribution (DBS) is typically composed of two separate
generated by deep brain stimulation. Exp Neurol models: a volume conductor model of an elec-
216(1):166–176 trode in brain tissue used to calculate electrical
Moro E, Esselink RJ et al (2002) The impact on potential distributions and cable models of neural
Parkinson’s disease of electrical parameter settings in
STN stimulation. Neurology 59(5):706–713 elements (e.g., axons, local cells) used to calcu-
Rizzone M, Lanotte M et al (2001) Deep brain stimulation late the response of neurons to the imposed stim-
of the subthalamic nucleus in Parkinson’s disease: ulation. The integrated model is interfaced with
effects of variation in stimulation parameters. an optimization algorithm to search for solutions
J Neurol Neurosurg Psychiatry 71(2):215–219
Vidailhet M, Vercueil L et al (2005) Bilateral deep- that increase the performance of DBS, where
brain stimulation of the globus pallidus in performance could be related to the efficiency,
primary generalized dystonia. N Engl J Med selectivity, and/or safety of stimulation. The
352(5):459–467 results from the two models are combined to
Wei XF, Grill WM (2005) Current density distributions,
field distributions and impedance analysis of seg- evaluate a specified cost function, and the opti-
mented deep brain stimulation electrodes. J Neural mization algorithm searches for a set of parame-
Eng 2(4):139–147 ters that minimize the cost.
Yousif N, Liu X (2009) Investigating the depth
electrode–brain interface in deep brain stimulation
using finite element models with graded complexity Detailed Description
in structure and solution. J Neurosci Methods
184(1):142–151
Yousif N, Purswani N et al (2010) Evaluating the impact DBS is an established therapy for treating neuro-
of the deep brain stimulation induced electric field on logical disorders, in which an implanted array of
subthalamic neurons: a computational modelling cylindrical electrodes is used to deliver rectangular
study. J Neurosci Methods 188(1):105–112
waveforms of electrical pulses to specific regions
Yousif N, Borisyuk R et al (2012a) Spatiotemporal
visualization of deep brain stimulation-induced of the brain. Trial and error experimentation is not
effects in the subthalamic nucleus. Eur J Neurosci an efficient approach to design improved elec-
36(2):2252–2259 trodes or waveforms and is unlikely to lead to an
Yousif N, Pavese N et al (2012b) Reversing the polarity of
optimal result. Therefore, model-based design is
bipolar stimulation in deep brain stimulation for essen-
tial tremor: a theoretical explanation for a useful clin- being used to develop improved DBS electrodes
ical intervention. Neurocase 20(1):10–17 (Butson and McIntyre 2006; Keane et al. 2012)
C 770 Computational Models to Optimize the Electrodes and Waveforms for Deep Brain Stimulation
!
and waveforms (Sahin and Tie 2007; Yousif et al. resistive, and E is derived from the gradient (∇)
2008; Foutz and McIntyre 2010; Wongsarnpigoon of the scalar potentials (F):
and Grill 2010). !
E ¼ ∇F: (1)
Parameterization of Electrodes and It is also valid to assume that the resistivity (r)
Waveforms of brain tissue is linear or independent of! F
There are an infinite number of possible electrode (Nicholson and Freeman 1975). Therefore, E is
designs and waveform shapes. In real DBS sys- directly
proportional
!
to the current density distri-
tems, however, practical limitations exist, such as bution J :
limitations on the size and number of electrodes
! !
and the maximum rate of change (slew rate) of E ¼r J : (2)
the output voltage in the implantable stimulator.
Therefore, the choice of parameterization should Substitution of Eq. 2 – Ohm’s law – into Eq. 1
define as large a number of electrodes or wave- and taking the divergence of the result yields
forms as possible using the least number of Poisson’s equation:

!
parameters while keeping practical limitations
∇2 F ¼ ∇ r J (3)
in mind. Figure 1 gives an example.
which is the partial differential equation solved in
Volume Conductor Models of Electrodes in the forward field problem for DBS.
Brain Tissue The electrical properties of brain tissue are

! process is
The first step in the optimization highly inhomogeneous (position dependent) and
calculating the electric field E generated by anisotropic (direction dependent). Therefore,
the passage of current through brain tissue. Poisson’s equation cannot be solved analytically,
Although the electrical properties of brain tissue and numerical methods are required to approxi-
vary with frequency, it is valid to assume that the mate a solution. The finite element method
capacitive, inductive, and wave propagation (FEM) is the preferred choice for modeling
effects can be ignored for typical DBS parameters DBS because it can be readily applied to the
(Bossetti et al. 2008). Under quasi-static complex geometries of the brain. For example,
conditions, brain tissue can be treated as purely patient-specific FEM models are the current state
Computational Models to Optimize the Electrodes parameterized by the positions of the individual electrodes
and Waveforms for Deep Brain Stimulation, along the shaft (z1 z3). Waveforms (of voltage or cur-
Fig. 1 Parameterization of an electrode (array) and rent) applied to each contact – or between contacts – can
waveform. An individual cylindrical electrode (left) can be parameterized either by using the coefficients of
be parameterized by its radius (r), height (h), and the a number of splines to define the waveform (right) or by
fraction of its surface that is conductive (f). In addition, discretizing the waveform into a finite time series
the configuration of the array (middle) can be
Computational Models to Optimize the Electrodes and Waveforms for Deep Brain Stimulation 771 C
of the art: The geometries of brain structures
are defined by individual brain images, and the
tensor conductivities of each brain region are
determined using diffusion-tensor magnetic res-
onance imaging (Butson et al. 2007; Chaturvedi
et al. 2010). Patient-specific models show good
agreement with clinical results, but this perfor- C
mance comes at the expense of complexity and
associated computational time.
In cases where an iterative optimization pro-
cess requires solving Poisson’s equation hun-
dreds to thousands of times, simplified analytic
and semi-analytic solutions offer a compromise
between the predictive capabilities of the model Computational Models to Optimize the Electrodes
and computational time. For example, if brain and Waveforms for Deep Brain Stimulation,
Fig. 2 Modeling neural elements with the cable equation.
tissue is approximated as an infinite homoge- An isolated patch of membrane on a neuron (top) is
neous medium with a single bulk r, then cylin- modeled using a one-dimensional cable (middle). In that
drical harmonics (i.e., Bessel functions) can be cable, the flow of charge is quantified using the cable
used to calculate semi-analytically the potentials equation (bottom), where Vm is the membrane voltage, Vr
is the resting voltage, Rm is the membrane resistance, Cm is
generated by the DBS electrode array. Although the membrane capacitance, Ra is the axoplasmic resis-
this method requires solving a system of tance, Fe is the external potential, and n denotes the
unknowns, it can still be used to evaluate the index of the compartment
effect of the electrode geometry on F. However,
if electrode geometry is not important and only of current (Zhang and Grill 2010), which yields
the temporal dynamics matter, then the DBS the following analytic solution (Li and Uren
electrode can be approximated as a point source 1998):
1=2
I ri, j
Fðx, y, zÞ ¼ 1=2 (4)
4p r1, 1 x2 þ r2, 2 y2 þ r3, 3 z2 þ r1, 2 xy þ r1, 3 xz þ r2, 3 yz
where I is the stimulus current, | | is the determi- cylindrical membrane compartments, each with
nate operator, ri, j is the resistivity tensor, and x, y, circuit elements modeling the membrane and axo-
and z are the spatial coordinates relative to the point plasm. If the soma is also approximated as a series
source (4) minimizes computational time and can of cylindrical compartments, then the flow of
be combined with the method of images to account charge through the entire neuron can be described
for simple inhomogeneities in the brain tissue. using the cable equation (Fig. 2). The cable equa-
tion can be solved numerically using the finite
difference method (Carnevale and Hines 1997).
Cable Models of Neural Elements
The second step in the optimization process is
simulating the response of target neural elements Optimization Algorithms
to the electric potentials calculated with the vol- The final step in the optimization process is using
ume conductor model. Axons and dendrites are the results from the computational model of
much longer than they are wide and azimuthally DBS to evaluate and minimize a cost function
symmetric about their longitudinal axis, so they via a numerical optimization algorithm. The
are modeled as a one-dimensional cable of choice of cost function depends on the goal of
C 772 Computational Models to Optimize the Electrodes and Waveforms for Deep Brain Stimulation
the optimization. If the goal is to extend the bat- (Sahin and Tie 2007). For all other PWs,
tery life of the device, efficiency is increased by Gaussian waveforms and triangular waveforms
minimizing the transferred electrical energy (E): are the most energy efficient; and the PW that
minimizes E depends on a number of factors,
ð
including the spatiotemporal distribution of the
E ¼ IV dt, (5)
potentials and the geometry, position, orienta-
tion, and electrophysiology of the target neural
where I and V are the applied voltage and applied element (Foutz and McIntyre 2010;
current, respectively. If the goal is to reduce Wongsarnpigoon and Grill 2010).
the risk of tissue damage, then the injected Although little work has been done on opti-
charge (Q), mizing electrodes for DBS, emerging work
ð shows that the same factors that impact waveform
efficiency (see above) are also important in
Q ¼ I dt, (6)
designing more efficient and selective electrode
geometries (Howell and Grill 2013).
is minimized because Q is correlated with tissue
injury (McCreery et al. 1990; Shannon 1992). Or,
if the goal is to mitigate the sensitivity of clinical References
outcomes to the malposition of the electrode,
selectivity is increased by constructing a curve, Bossetti CA, Birdno MJ et al (2008) Analysis of the quasi-
p(x), of the proportion of a nontarget population static approximation for calculating potentials gener-
activated versus selected proportions of ated by neural stimulation. J Neural Eng 5(1):44–53
Butson CR, Cooper SE et al (2007) Patient-specific anal-
a target population and minimizing the area ysis of the volume of tissue activated during deep brain
(A) under the curve: stimulation. Neuroimage 34(2):661–670
Butson CR, McIntyre CC (2006) Role of electrode design
ð on the volume of tissue activated during deep brain
A ¼ pðxÞ dx: (7) stimulation. J Neural Eng 3(1):1–8
Carnevale NT, Hines ML (1997) The NEURON book.
Cambridge University Press
The cost function must be minimized numer- Chaturvedi A, Butson CR et al (2010) Patient-specific
models of deep brain stimulation: influence of field
ically, as there is no guarantee that it will be
model complexity on neural activation predictions.
continuously differentiable over the parameter Brain Stimul 3(2):65–67
space defining the electrode and/or waveform. Foutz TJ, McIntyre CC (2010) Evaluation of novel stim-
Using iterative methods (e.g., Newton’s method ulus waveforms for deep brain stimulation. J Neural
Eng 7(6):066008
and conjugate gradient) to minimize cost is
Howell B, Grill WM (2013) Model-based optimization of
impractical because they require approximating electrode designs for deep brain stimulation. In: Neu-
a Jacobian and/or Hessian, which may or may not ral engineering (NER), 2013 6th international IEEE/
exist for all space. Therefore, stochastic search EMBS conference on San Diego, California
Keane M, Deyo S et al (2012) Improved spatial targeting
algorithms are used to identify solutions that
with directionally segmented deep brain stimulation
minimize cost, including simulated annealing, leads for treating essential tremor. J Neural Eng
genetic algorithm, and particle swarm – inspired 9(4):046005
by the natural processes of crystallization, evolu- Li P, Uren NF (1998) Analytical solution for the electric
potential due to a point source in an arbitrarily aniso-
tion, and the collective behavior of self-organized
tropic half-space. J Eng Math 33(2):129–140
systems, respectively. McCreery DB, Agnew WF et al (1990) Charge density and
charge per phase as cofactors in neural injury induced
Optimal Electrodes and Waveforms by electrical stimulation. Biomed Eng IEEE Trans
37(10):996–1001
Computational studies of DBS show that Nicholson C, Freeman JA (1975) Theory of current
rectangular waveforms are the most energy source-density analysis and determination of conduc-
efficient for pulse widths (PWs) 40 ms tivity tensor for anuran cerebellum. 38(2):356–368
Computational Olfaction 773 C
Sahin M, Tie Y (2007) Non-rectangular waveforms for Detailed Description
neural stimulation with practical electrodes. J Neural
Eng 4(3):227
Shannon RV (1992) A model of safe levels for Neural Networks Involved in Olfactory
electrical stimulation. Biomed Eng IEEE Trans Computation
39(4):424–426 Computational models of the olfactory system have
Wongsarnpigoon A, Grill WM (2010) Energy-efficient contributed immensely to the framing of experi-
waveform shapes for neural stimulation revealed with
a genetic algorithm. J Neural Eng 7(4):046009 mental problems and the construction of complex C
Yousif N, Bayford R et al (2008) The influence of reac- hypotheses regarding its function. Computational
tivity of the electrode–brain interface on the crossing models from detailed biophysical to large-scale
electric current in therapeutic deep brain stimulation. simple neuron models have proposed how olfactory
Neuroscience 156(3):597–606
Zhang TC, Grill WM (2010) Modeling deep brain networks compute the functions necessary to
stimulation: point source approximation versus realis- extract and recognize important chemical signals.
tic representation of the electrode. J Neural Eng Signal-to-noise ratio in the olfactory system is
7(6):066009 inherently low at the level of olfactory sensory
neurons (OSN), which are broadly tuned to odor-
ants. OSNs sharing a common response profile to
odorants project their axons to specific glomeruli
Computational Neuroscience of within the olfactory bulb (Mombaerts
Color et al. 1996), with a very high convergence. This
high convergence ratio has been proposed to
▶ Color Vision, Computational Methods for improve the signal-to-noise ratio during odorant
detection (van Drongelen et al. 1978).
Mitral cells, along with middle and deep tufted
cells, are the principal output neurons of the
olfactory bulb. They are directly postsynaptic to
Computational Olfaction OSNs and as such the pattern of mitral/tufted cell
activation across the bulb constitutes the output
Christiane Linster representations. Besides input from a single class
Computational Physiology Lab, Department of of OSNs, mitral/tufted activation pattern also
Neurobiology and Behavior, Cornell University, depends on the activity of several classes of bul-
Ithaca, NY, USA bar interneurons. For example, some form of
normalization of stimulus concentrations is
clearly evident in the concentration-response pro-
Definition files of mitral cells; this function has been pro-
posed to be the role of an extensive network of
In natural environments, airborne chemical stim- excitatory and inhibitory neurons in the glomer-
uli are distributed unpredictably in time and ular layer (Cleland et al. 2007). Contrast
space, and odorants from innumerable sources enhancement, the effects of which have been
intermix freely. The olfactory system must per- directly observed in the olfactory bulb (Yokoi
form computations to detect potential signals of et al. 1995), can be functionally defined as
interest within these noisy signals, extract these a process of competition between neurons
signals, form representations, compare these to expressing similar response profiles. Simplified
those of previously experienced, differentiate rel- models of the olfactory system have been able to
evant from irrelevant stimuli, and cue an appro- implement contrast enhancement using lateral
priate response. Computational modeling inhibition between glomeruli (Cleland and
suggests that subnetworks of the olfactory system Linster 2002; Linster and Cleland 2001; Linster
may be dedicated to perform specific computa- and Gervais 1996; Linster and Hasselmo 1997,
tions underlying these functions. 1999; Linster and Smith 1997). Alternatively
C 774 Computational Olfaction
to these highly simplified representations, synaptic depression can be employed to filter

non-topographical models of contrast enhance- out stable background odorants, whereas long-
ment are also capable of distributing inhibition term synaptic plasticity can store associations
in proportion to receptive field similarity but use between neurons responding to the same odor
local computations and broad feedback inhibition stimuli. Indeed, the extensive intrinsic feedback
to create contrast (Cleland and Sethupathy 2004). network in this cortex and its integration with
The net activation of mitral cells is then trans- afferent inputs closely resembles the structure of
lated into trains of action potentials, the precise traditional theoretical associative memory net-
timing of which appears to be regulated by coor- works as first described by Marr (Marr 1971).
dinated oscillations measurable in field record-
ings across the OB. These bulbar oscillations are
thought to depend on an extensive excitatory- Conclusions
inhibitory network of mitral cell secondary den-
drites and granule cell interneurons, as well as Computational models have an established and
reciprocal cortical connections that modulate growing role within systems neuroscience. As
bulbar activity according to behavioral state our understanding of neural processing and inter-
(Kay 2003; Lagier et al. 2004; Martin actions becomes more sophisticated, computer
et al. 2004; Ravel et al. 2003). Mitral cell spike models of these systems are increasingly neces-
synchronization patterns can mediate a second sary in order to understand and interpret experi-
level of feature extraction to the odor representa- mental results. In the olfactory system in
tion (David et al. 2009), as follower neurons in particular, computational modeling will no
diverse central olfactory structures (Linster and doubt be essential to understanding the integra-
Cleland 2003) process incoming mitral cell tion of the many factors influencing the construc-
spikes via synaptic learning rules, the best tion and transformation of odor representations.
known of which rely upon precise spike timing
(Cleland and Linster 2002; Song et al. 2000). The
mechanisms of oscillations have been the focus References
of a number of computational modeling efforts,
some proposing oscillators governed by an Ambros-Ingerson J, Granger R, Lynch G (1990) Simula-
tion of paleocortex performs hierarchical clustering.
excitatory-inhibitory feedback loop, others pro- Science 247:1344–1348
posing intrinsic oscillations in mitral cells syn- Bathellier B, Lagier S, Faure P, Lledo PM (2006) Circuit
chronized by slow common inhibitory inputs properties generating gamma oscillations in a network
(Bathellier et al. 2008, 2006; David et al. 2009; model of the olfactory bulb. J Neurophysiol
95:2678–2691
Erdi et al. 1993; Freeman 1987; Li and Hopfield Bathellier B, Carleton A, Gerstner W (2008) Gamma
1989). oscillations in a nonlinear regime: a minimal model
Even though some models propose associative approach using heterogeneous integrate-and-fire net-
memory functions for olfactory bulb (Erdi works. Neural Comput 20:2973–3002
Cleland TA, Linster C (2002) How synchronization
et al. 1993; Hendin et al. 1998), olfactory asso- properties among second-order sensory neurons can
ciative memory functions have been more com- mediate stimulus salience. Behav Neurosci
monly attributed to the piriform cortex, one of the 116:212–221
targets of mitral cell axons projecting from the Cleland TA, Sethupathy P (2004) Non-topographical con-
trast enhancement disambiguates high-dimensional
olfactory bulb. Specifically, the piriform cortex odor representations. Soc Neurosci Abstr 531.2
has been proposed to mediate the associative Cleland TA, Johnson BA, Leon M, Linster C (2007) Rela-
memory functions necessary for odor-context tional representation in the olfactory system. Proc Natl
learning (de Almeida et al 2013; Haberly 2001; Acad Sci U S A 104:1953–1958
David FO, Hugues E, Cenier T, Fourcaud-Trocme N,
Haberly and Bower 1989; Hasselmo et al. 1990) Buonviso N (2009) Specific entrainment of mitral
and hierarchical clustering (Ambros-Ingerson cells during gamma oscillation in the rat olfactory
et al. 1990). For example, cortical short-term bulb. PLoS Comput Biol 5:e1000551
Computational Psychiatry 775 C
de Almeida L, Idiart M, Linster C (2013) A model of oscillatory responses in the rat olfactory bulb:
cholinergic modulation in olfactory bulb and piriform a correlate of odor recognition? J Neurosci
cortex. J Neurophysiol 109:1360–1377 24:389–397
Erdi P, Grobler T, Barna G, Kaski K (1993) Dynamics of Mombaerts P, Wang F, Dulac C, Chao SK, Nemes A,
the olfactory bulb: bifurcations, learning, and memory. Mendelsohn M, Edmondson J, Axel R (1996) Visual-
Biol Cybern 69:57–66 izing an olfactory sensory map. Cell 87:675–686
Freeman WJ (1987) Simulation of chaotic EEG patterns Ravel N, Chabaud P, Martin C, Gaveau V, Hugues E,
with a dynamic model of the olfactory system. Biol Tallon-Baudry C, Bertrand O, Gervais R (2003) Olfac-
Cybern 56:139–150 tory learning modifies the expression of odour-induced
C
Haberly LB (2001) Parallel-distributed processing in oscillatory responses in the gamma (60–90 Hz) and
olfactory cortex: new insights from morphological beta (15–40 Hz) bands in the rat olfactory bulb. Eur
and physiological analysis of neuronal circuitry. J Neurosci 17:350–358
Chem Senses 26:551–576 Song S, Miller KD, Abbott LF (2000) Competitive
Haberly LB, Bower JM (1989) Olfactory cortex: model Hebbian learning through spike-timing-dependent
circuit for study of associative memory? Trends synaptic plasticity. Nat Neurosci 3:919–926
Neurosci 12:258–264 van Drongelen W, Holley A, Doving KB (1978) Conver-
Hasselmo ME, Wilson MA, Anderson BP, Bower JM gence in the olfactory system: quantitative aspects of
(1990) Associative memory function in piriform odour sensitivity. J Theor Biol 71:39–48
(olfactory) cortex: computational modeling and neu- Yokoi M, Mori K, Nakanishi S (1995) Refinement of odor
ropharmacology. Cold Spring Harb Symp Quant Biol molecule tuning by dendrodendritic synaptic inhibi-
55:599–610 tion in the olfactory bulb. Proc Natl Acad Sci U S A
Hendin O, Horn D, Tsodyks MV (1998) Associative mem- 92:3371–3375
ory and segmentation in an oscillatory neural model of
the olfactory bulb. J Comput Neurosci 5:157–169
Further Reading
Kay LM (2003) Two species of gamma oscillations in the
Escanilla O, Arrellanos A, Karnow A, Ennis M, Linster C
olfactory bulb: dependence on behavioral state and
(2010) Noradrenergic modulation of behavioral odor
synaptic interactions. J Integr Neurosci 2:31–44
detection and discrimination thresholds in the olfac-
Lagier S, Carleton A, Lledo PM (2004) Interplay between
tory bulb. Eur J Neurosci 32:458–468
local GABAergic interneurons and relay neurons gen-
erates gamma oscillations in the rat olfactory bulb.
J Neurosci 24:4382–4392
Li Z, Hopfield JJ (1989) Modeling the olfactory bulb and its
neural oscillatory processings. Biol Cybern 61:379–392 Computational Psychiatry
Linster C, Cleland TA (2001) How spike synchronization
among olfactory neurons can contribute to sensory
discrimination. J Comput Neurosci 10:187–193 Quentin J. M. Huys
Linster C, Cleland TA (2003) Decorrelation of odor rep- Translational Neuromodeling Unit, Institute of
resentations via spike timing-dependent plasticity. Biomedical Engineering, ETH Z€urich and
Front Comput Neurosci 4:157 University of Z€urich, Z€urich, Switzerland
Linster C, Gervais R (1996) Investigation of the role of
interneurons and their modulation by centrifugal fibers Department of Psychiatry, Psychotherapy and
in a neural model of the olfactory bulb. J Comput Psychosomatics, Hospital of Psychiatry,
Neurosci 3:225–246 University of Z€urich, Z€urich, Switzerland
Linster C, Hasselmo M (1997) Modulation of inhibition in
a model of olfactory bulb reduces overlap in the neural
representation of olfactory stimuli. Behav Brain Res
84:117–127
Definition
Linster C, Hasselmo ME (1999) Behavioral responses to
aliphatic aldehydes can be predicted from known elec- Computational psychiatry is a heterogeneous
trophysiological responses of mitral cells in the olfac- field at the intersection of computational neuro-
tory bulb. Physiol Behav 66:497–502
science and psychiatry. Incorporating methods
Linster C, Smith BH (1997) A computational model of the
response of honey bee antennal lobe circuitry to odor from psychiatry, psychology, neuroscience,
mixtures: overshadowing, blocking and unblocking can behavioral economics, and machine learning,
arise from lateral inhibition. Behav Brain Res 87:1–14 computational psychiatry focuses on building
Marr D (1971) Simple memory: a theory for archicortex.
mathematical models of neural or cognitive phe-
Philos Trans R Soc Lond B Biol Sci 262:23–81
Martin C, Gervais R, Hugues E, Messaoudi B, Ravel nomena relevant to psychiatric diseases. The
N (2004) Learning modulation of odor-induced models span a wide range – from biologically
C 776 Computational Psychiatry
detailed models of neurons or networks to connectionist (McClelland et al. 1986) or dynam-

abstract models describing high-level cognitive ical (King et al. 1984) and emphasized properties
abilities of an organism. Psychiatric diseases are of the brain as a neural network. The second, and
conceptualized either as an extreme of normal now most common, class is models of reinforce-
function or as a consequence of alterations in ment learning (Montague et al. 1996; Schultz
parts of the model. et al. 1997; Sutton and Barto 1998). The third
As in computational neuroscience more gen- and fourth classes emphasize explicit model-
erally, the building of models forces key concepts based planning or social cognitive processes,
to be made concrete and hidden assumptions to respectively.
be made explicit. One critical function of these
models in the setting of psychiatry is their ability Connectionist and Neural Network Models
to bridge between low-level biological and high- Connectionist approaches model psychological
level cognitive features. While many neurobio- functions using a neural network in which partic-
logical alterations are known, the exclusively ular neurons take on specific computational roles,
atheoretical focus of standard psychiatric nosol- for instance, representing a sensory input. Key
ogy on high-level symptoms has as yet prevented aspects of the neural networks are based on par-
an integration of these bodies of knowledge. ticular features derived from biology (or known
David Marr pointed out that models at different to be involved in psychopathology). This allows
levels of description may be independent (Marr the consequences of these features for complex
1982). Nevertheless, algorithmic details may computation to be probed and hence is one direct
constrain functions at the computational level. approach to examining the link between Marr’s
The models used in computational psychiatry levels. In psychiatry, connectionist models have
make these constraints explicit and thereby aim been prominently applied to schizophrenia.
to provide normative conduits between the dif- Patients with schizophrenia or mania can char-
ferent levels at which neural systems are ana- acteristically display rapidly changing, loose
lyzed (Stephan et al. 2006; Huys et al. 2011; associations in their speech. Early work exam-
Hasler 2012; Montague et al. 2012). This in turn ined how parameters governing the dynamics of
allows for a principled approach to study dys- associative networks might reproduce this. An
functions and indeed may allow the dysfunctions increase in noise, corresponding to a decrease in
observed in psychiatry to inform neuroscience the dynamic gain, led to less specific memories,
in general. mirroring a broadening of associations in schizo-
Practically, it underpins hopes that computa- phrenia, and less stable, constantly altering mem-
tional techniques may facilitate the development ories, possibly mirroring the pressure of speech
of a psychiatric nomenclature based on an under- observed in mania. Spurious memories reminis-
standing of the underlying neuroscience. Compu- cent of hallucinations arose when overloading the
tational models enhance experimental designs by network with memories beyond its capacity
allowing more intricate neural and/or cognitive (Grossberg and Pepe 1970; Hoffman 1987).
processes to be inferred from complex features of Patients with schizophrenia also show impair-
the data, often via Bayesian inference. These ments in cognitive flexibility and control tasks
aspects motivate hopes that it may facilitate the that require the inhibition of a prepotent response.
development of clinical treatment and decision Cohen et al. (1996) used a connectionist network
tools informed by advances in neuroscience. partitioned into four modules to model this. One
module represented stimuli, another other
responses, and a third the prefrontal cortex.
Detailed Description A fourth module integrated the inputs from all
modules. The prefrontal cortex had a critical
This entry describes four types of models applied function in maintaining representations of task-
to psychiatric diseases. The earliest models were relevant variables. When this memory function
was impaired by a reduction in its gain, the net- with new experience by computing a prediction
work could accurately capture the impairments in error (PE) d which compares the obtained with
performance seen in schizophrenia. As the gain the expected future reward Qðs, aÞ. This expec-
was thought to depend on dopaminergic input, tation Qðs, aÞ can then be updated by letting
this work correctly predicted a prefrontal reduc- Qðs, aÞ Qðs, aÞ þ ϵd , where 0 ϵ 1 is
tion of dopamine in schizophrenia. Later work a learning rate that determines how rapidly the
further refined the role of dopamine in gating of qualities are allowed to change over time. That is, C
information into the prefrontal cortex (Braver rather than inferring decisions from an under-
et al. 1999; Frank 2005). standing of the world, model-free choices are
Many psychiatric disorders are relapsing- the result of repeated trial-and-error learning.
remitting, with periods of well-being punctuated This motivates the characterization of habits as
by times of illness. This is prominent in bipolar model-free (Daw et al. 2005). Seminal work has
disorder, which can cycle rapidly between phases shown that phasic signals by dopaminergic neu-
of depression and mania. Conversely, loss of rons are proportional to this PE d (Montague
diurnal sleep-wake rhythms is also common. et al. 1996; Schultz et al. 1997).
These phenomena can be described by dynamical The application of reinforcement learning
systems models of neural networks with feedback models to psychiatric diseases is motivated by
loops and delays interacting to produce oscilla- a trio of facts. First, decisions are central to psy-
tory phenomena at various timescales (Mackey chiatric disorders, and reinforcement learning
and Milton 1987; Milton 2010). The complexities techniques allow for a principled approach to
of the local circuit, with synthesis, breakdown, decision-making. Decisions are the final common
and reuptake, can add substantial further com- pathway preceding many aberrant behaviors, and
plexity and facilitate the emergence of highly poor decisions can have profound consequences
variable, chaotic solutions potentially relevant for affected individuals. For instance, around
to multiple disorders (King et al. 1984). a third of the life stress associated with major
depressive disorder is due to poor decisions by
Reinforcement Learning Models individuals (Kendler et al. 1999). Second, many
Reinforcement learning (RL; Sutton and Barto psychoactive substances and clinically useful
1998) describes a set of techniques aimed at medications influence neuromodulators: dopa-
choosing that action which maximizes the long- mine (antipsychotics such as haloperidol or
term expected reward. In terms of applications to clozapine and stimulants such as methylpheni-
psychiatric diseases, it is useful to differentiate date), serotonin (selective serotonin reuptake
between two approaches to solving this problem inhibitors, tricyclic antidepressants), noradrena-
(Daw et al. 2005): a model-based and a model- line (tricyclic antidepressants, serotonin-
free one. In the model-based approaches, the noradrenaline reuptake inhibitors), or
agent has a model M of the world that describes acetylcholine (certain antipsychotics, cholines-
the consequences of actions and the desirability terase inhibitors). Third, and unifying the first
of the consequences. For instance, a player may two, these same neuromodulators are central to
know the rules of chess. The best move can then computational models of decision-making. This
be inferred by considering all moves, their con- is particularly clear for phasic dopamine, which
sequences, and the subsequent moves iteratively. appears to report a signal akin to the PE d used for
For most problems of interest, this model- learning in model-free RL (Montague et al. 1996;
based decision procedure is computationally Schultz et al. 1997; Sutton and Barto 1998).
unachievable. Model-free approaches replace
computation by experience, maintaining Predominantly Model-Free RL Accounts of
a lookup table Qðs, aÞ of the expected reward Psychiatric Diseases
(and hence goodness) of each behavior a in situ- The simplest application of reinforcement learn-
ation s. This can be iteratively updated online ing models is in examining anhedonia, a central
component of depression. Anhedonia describes a pronounced slowing (bradykinesia) and exper-

a reduced ability to experience pleasure and is imentally in facilitated learning of action omis-
usually measured by verbal reports. People who sion compared to action production (no-go rather
report such a lack of pleasure are less likely to than go learning) that is normalized by dopami-
choose the more rewarding stimulus in a variety nergic medication. The effects of dopamine on go
of standard learning (Costello 1972; Henriques and no-go learning are accounted for by biolog-
et al. 1994; Pizzagalli et al. 2005) and stimulus ically detailed computational models of the basal
reactivity (Bylsma et al. 2008) tasks. Decisions in ganglia (Frank 2005). Two critical structural
some of these tasks can be captured by simple aspects of the basal ganglia inform these models:
model-free reinforcement learning. Depression (a) the presence of parallel corticobasal ganglia-
interferes with this learning by reducing the PE thalamocortical loops for go and no-go learning
signal d (Steele et al. 2007; Kumar et al. 2008; and (b) the prevalence of activating D1-type
Chase et al. 2010a, b; Gradin et al. 2011), leading dopamine receptors on the go and inhibiting
to reduced expectations of rewards over time. D2-type receptors on the no-go loop (Maia and
The PE d is composed of the difference between Frank 2011). Positive prediction errors are thus
expected reward and obtained reward, and anhe- linked with go and negative prediction errors with
donia appears to reduce the former, while alter- no-go learning (Frank et al. 2007; Guitart-Masip
ations of tonic dopamine affect the latter et al. 2012). Such considerations are also relevant
(Chowdhury et al. 2013; Huys et al. 2013). for schizophrenia, Tourette’s syndrome, and
Addictive substances release dopamine either attention-deficit hyperactivity disorder (Maia
directly or indirectly and modify dopamine and Frank 2011). The work on Parkinson’s dis-
receptors chronically (Volkow et al. 2009). This ease is remarkable in that it has informed under-
might simply suggest that addictive substances standing of the function of the basal ganglia and
thereby result in overly large values of the drug- dopamine in the healthy brain.
taking action by adding an irreducible floor to the Finally, dopamine also plays a particular role
PE signal d. This would result in a persistent in schizophrenia research because (a) the positive
signal for learning, increasing the value of drug- symptoms of psychosis respond to dopaminergic
taking actions without bound (Redish 2004; D2 antagonists such as haloperidol (Seeman
Dayan 2009). Other reinforcement learning et al. 1976; Kapur et al. 2000; Laruelle
models address the transformation from hedonic et al. 2003) and because of (b) findings of
to compulsive drug taking (Everitt and Robbins increased dopamine synthesis, release, and syn-
2005) by emphasizing the difference between aptic levels in psychotic states (Heinz 2002;
model-free and model-based reinforcement Kapur 2003) from the first episode onward
learning (Redish et al. 2008). Drugs of abuse are (Egerton et al. 2013). Model-free reinforcement
known to lead to a shift toward habitual behavior learning models that link phasic dopamine to
(Dickinson et al. 2000; Nelson and Killcross prediction error signals have been instrumental
2006), which can be captured by a shift from in relating the dopaminergic dysfunctions to cog-
model-based to model-free decision-making nitive phenomena. They have provided detailed
(Daw et al. 2005). Neurobiologically, this shift accounts of the effects of dopamine manipula-
might be accompanied by a progressive tions on both learning and expression of learned
engraining of drug-taking actions through pro- contingencies in preclinical animal models such
gressively more dorsal corticostriatal loops as the conditioned avoidance response and latent
(Belin and Everitt 2008) and impairments of inhibition (Smith et al. 2004, 2005, 2007;
orbitofrontal function (Lucantonio et al. 2012). Moutoussis et al. 2008). In humans, reinforce-
Parkinson’s disease is characterized by ment learning models of decision-making have
a progressive reduction of dopamine-producing been combined with fMRI to examine the corre-
cells in the midbrain and treated with dopamine lates of putatively dopaminergic PEs. Schizo-
precursors or agonists. Clinically, it results in phrenia patients have a reduced response to
prediction errors in striatal and midbrain regions to the definition of control in control systems
(Juckel et al. 2006; Corlett et al. 2007; Jensen theory. The psychological construct whereby
et al. 2008; Murray et al. 2008) that is normalized controllability is related to particular desirable
by treatment with second-generation antipsy- outcomes can be captured by defining controlla-
chotic medications (Juckel et al. 2006). bility as the fraction of rewards reliably achiev-
Passivity phenomena are another important able through appropriate action selection (Dayan
characteristic feature of schizophrenia, where and Huys 2009). These definitions of controlla- C
movements or sensations are experienced as not bility can be used as prior beliefs and thereby
originating from oneself. Incorrect expectations account qualitatively for learned helplessness
about the consequences of one’s own actions may and chronic mild stress (Willner 1997) effects
lead to the assignment of the sensory conse- (Dayan and Huys 2009). Further refinements
quences of one’s own actions to external sources accounting specifically for generalization issues
(Gray et al. 1991; Fletcher and Frith 2009). This and for the computational costs of goal-directed
may arise from an impaired connectivity between decisions allow for more quantitatively detailed
areas generating the predictions and those accounts (Lieder et al. 2013).
assessing the input (Ford et al. 2007), speaking Serotonin plays an important role in research
to a generalized notion of PEs that is less directly on depression because selective serotonin inhib-
related to phasic dopaminergic prediction errors itors (SSRIs) are first-line pharmacotherapeutic
(Rao and Ballard 1999) but rather to the conse- agents (Gelder et al. 2006) and because dietary
quences of dopamine in modulating synaptic manipulations that are thought to acutely reduce
plasticity (Friston and Frith 1995; Stephan serotonin (acute tryptophan depletion) can lead to
et al. 2006; Friston 2008; Stephan et al. 2009). a rapid recurrence of symptoms in formerly
depressed patients (Smith et al. 1997). However,
Predominantly Model-Based Accounts of animal work has related increased serotonin to
Psychiatric Diseases behavioral inhibition, aversive expectations, and
Higher-level cognitive aspects of psychiatric dis- expression of helplessness (Soubrié 1986; Maier
eases are captured by reinforcement learning and Watkins 2005), suggesting that a reduction in
models in which decisions are the product of serotonin should improve depressive symptoms.
searching an explicit model of the task at hand. Reinforcement learning models have linked the
These have been applied to the concept of help- inefficient avoidance of negative events due to
lessness in depression, to the role of serotonin, serotonin in a model-free system with an
and to delusional belief formation in increased exposure to more negative events.
schizophrenia. These experiences might in turn underlie more
In depression research, the concept of negative moods due to serotonin reductions
helplessness – a perception of lack of (Dayan and Huys 2008; Boureau and Dayan
control – arose from animal behavior and 2011; Dayan 2012). Further work has character-
describes animals who fail to escape from shocks ized the interaction of this model-free behavioral
after experiencing other stressors that were not inhibition with internal cognitive planning
under their control (Seligman and Maier 1967). models and has suggested that serotonin might
Similar findings exist in humans (e.g., Miller and also inhibit internal thought processes. This
Seligman 1975). The simplest mathematical for- would be instrumental in facilitating planning
mulation of controllability is as a low action- by pruning overly large decision trees to
outcome contingency p(o|a) (Maier and a computationally manageable size (Huys
Seligman 1976). More detailed accounts view et al. 2012). While serotonin is also central to
controllability as relating not to individual avail- helplessness, this computational relationship has
able actions but to the average achievability of not yet been examined.
different outcomes by different choices of actions “Jumping to conclusions” is a phenomenon
(Dayan and Huys 2009). This definition is close that aims to capture delusional belief formation.
It describes a tendency for patients with paranoid Autistic spectrum disorder (ASD) on the other
ideation to declare a strongly held belief based on hand does not affect the maintenance of cooper-
data that should not be sufficient to warrant such ation in the trust task, but the cingulate cortex
strong beliefs (Garety et al. 2005). However, a does not tag “self”-responses accurately (Tomlin
model-based reinforcement learning model of the et al. 2006; Chiu et al. 2008). In related cooper-
standard task used to measure jumping to conclu- ative games (e.g., the “stag hunt” game), ASD
sions (the “beads-in-a-jar” task) suggested that patients have a less rich representation of the
the reason patients with schizophrenia jumped other’s strategic abilities (Yoshida et al. 2008,
to conclusions was not due to aberrantly strong 2010a, b). Responses of healthy volunteers to
beliefs but due to taking into account their future patients with a variety of psychiatric disorders
inability to exploit further data (Moutoussis in the trust task fall into distinguishable classes
et al. 2011). (Koshelev et al. 2010). This might allow the use
of healthy volunteers as “biosensors” and formal-
Game-Theoretical Approaches to Social izes one important aspect and use of empathy or
Dysfunction even countertransference in psychiatric clinical
Psychiatric disorders have a profoundly detri- practice. Formally, these tasks are partially
mental effect on social function. Social interac- observable Markov decision problems and can
tions are extremely complex and therefore be modeled as such (Yoshida et al. 2008; Ray
difficult to investigate directly. Game-theoretical et al. 2009).
approaches allow interpersonal cognition to be
probed parametrically because the social commu- Limitations
nication channel is highly restricted yet func- Computational approaches to psychiatric disor-
tional. Economic games benefit from having ders focus on decision-making, valuation, and
been extensively examined both theoretically cognition. While this is critical to many psychi-
and experimentally in economics (Camerer atric disorders, immunological, endocrine, or
2003). vegetative dysfunctions are so far largely beyond
Borderline personality disorder (BPD), which these accounts.
is characterized by unstable personal relation-
ships and inappropriate emotional responses,
Acknowledgments I thank Dominik Bach, Kay
leads to a breakdown of cooperation in trust H. Brodersen, Anthony Cruickshank, Peter Dayan, Marc
tasks. In this task, the first participant (the inves- Guitart-Masip, Helene Haker, Gregor Hasler, Falk Lieder,
tor) chooses which fraction of an endowment to Tiago Maia, John Milton, Michael Moutoussis, Peggy
Seriès, and Klaas Enno Stephan for informative comments
invest. The investment is multiplied, and the
and discussions on an earlier version of this contribution.
trustee chooses what fraction of the multiplied
amount to return to the investor. When played
over multiple rounds, both players are best off References
when they cooperate (i.e., when the investor
invests and the trustee faithfully returns Belin D, Everitt BJ (2008) Cocaine seeking habits depend
a substantial fraction). However, this cooperation upon dopamine-dependent serial connectivity linking
is easily broken. Healthy participants maintain the ventral with the dorsal striatum. Neuron
57(3):432–441
cooperation by repairing lost trust through signals
Boureau YL, Dayan P (2011) Opponency revisited:
(e.g., by investing a large fraction to signal will- competition and cooperation between dopamine
ingness to cooperate; King-Casas et al. 2005). and serotonin. Neuropsychopharmacology
Patients with BPD fail to recognize social 36(1):74–97
Braver TS, Barch DM, Cohen JD (1999) Cognition and
signals used to repair social bonds after tempo- control in schizophrenia: a computational model of
rary ruptures, potentially due to a failure of the dopamine and prefrontal function. Biol Psychiatry
insula to respond (King-Casas et al. 2008). 46(3):312–328
Bylsma LM, Morris BH, Rottenberg J (2008) A meta- Everitt BJ, Robbins TW (2005) Neural systems of rein-
analysis of emotional reactivity in major depressive forcement for drug addiction: from actions to habits to
disorder. Clin Psychol Rev 28(4):676–691 compulsion. Nat Neurosci 8(11):1481–1489
Camerer CF (2003) Behavioral game theory: experiments Fletcher PC, Frith CD (2009) Perceiving is believing:
in strategic interaction. Princeton University Press, a Bayesian approach to explaining the positive
Princeton symptoms of schizophrenia. Nat Rev Neurosci
Chase HW, Frank MJ, Michael A, Bullmore ET, Sahakian 10(1):48–58
BJ, Robbins TW (2010a) Approach and avoidance Ford JM, Roach BJ, Faustman WO, Mathalon DH
learning in patients with major depression and healthy (2007) Synch before you speak: auditory hallucina-
C
controls: relation to anhedonia. Psychol Med tions in schizophrenia. Am J Psychiatry
40(3):433–440 164(3):458–466
Chase HW, Michael A, Bullmore ET, Sahakian BJ, Rob- Frank MJ (2005) Dynamic dopamine modulation in the
bins TW (2010b) Paradoxical enhancement of choice basal ganglia: a neurocomputational account of cogni-
reaction time performance in patients with major tive deficits in medicated and nonmedicated Parkin-
depression. J Psychopharmacol 24(4):471–479 sonism. J Cogn Neurosci 17(1):51–72
Chiu PH, Kayali MA, Kishida KT, Tomlin D, Klinger LG, Frank MJ, Moustafa AA, Haughey HM, Curran T,
Klinger MR, Montague PR (2008) Self responses Hutchison KE (2007) Genetic triple dissociation
along cingulate cortex reveal quantitative neural phe- reveals multiple roles for dopamine in reinforcement
notype for high-functioning autism. Neuron learning. Proc Natl Acad Sci U S A 104(41):
57(3):463–473 16311–16316
Chowdhury R, Guitart-Masip M, Lambert C, Dayan P, Friston K (2008) Hierarchical models in the brain. PLoS
Huys QJM, D€uzel E, Dolan RJ (2013) Dopamine Comput Biol 4(11):e1000211
restores reward prediction errors in older age. Nat Friston KJ, Frith CD (1995) Schizophrenia: a disconnec-
Neurosci 16:648–653 tion syndrome? Clin Neurosci 3(2):89–97
Cohen JD, Braver TS, O’Reilly RC Garety PA, Freeman D, Jolley S, Dunn G, Bebbington PE,
(1996) A computational approach to prefrontal cortex, Fowler DG, Kuipers E, Dudley R (2005) Reasoning,
cognitive control and schizophrenia: recent develop- emotions, and delusional conviction in psychosis.
ments and current challenges. Philos Trans R Soc Lond J Abnorm Psychol 114(3):373–384
B Biol Sci 351(1346):1515–1527 Gelder M, Harrison P, Cowen P (2006) Shorter Oxford
Corlett PR, Murray GK, Honey GD, Aitken MRF, Shanks textbook of psychiatry. Oxford University Press,
DR, Robbins TW, Bullmore ET, Dickinson A, Oxford
Fletcher PC (2007) Disrupted prediction-error signal Gradin VB, Kumar P, Waiter G, Ahearn T, Stickle C,
in psychosis: evidence for an associative account of Milders M, Reid I, Hall J, Steele JD (2011) Expected
delusions. Brain 130(Pt 9):2387–2400 value and prediction error abnormalities in depression
Costello CG (1972) Depression: loss of reinforcers or loss and schizophrenia. Brain 134(Pt 6):1751–1764
of reinforcer effectiveness? Behav Ther 3:240–247 Gray J, Feldon J, Rawlins J, Hemsley D, Smith A (1991)
Daw ND, Niv Y, Dayan P (2005) Uncertainty-based com- The neuropsychology of schizophrenia. Behav Brain
petition between prefrontal and dorsolateral striatal Sci 14(01):1–20
systems for behavioral control. Nat Neurosci Grossberg S, Pepe J (1970) Schizophrenia: possible
8(12):1704–1711 dependence of associational span, bowing, and pri-
Dayan P (2009) Dopamine, reinforcement learning, and macy vs. recency on spiking threshold. Behav Sci
addiction. Pharmacopsychiatry 42(Suppl 1):S56–S65 15(4):359–362
Dayan P (2012) Instrumental vigour in punishment and Guitart-Masip M, Huys QJM, Fuentemilla L, Dayan P,
reward. Eur J Neurosci 35(7):1152–1168 Duzel E, Dolan RJ (2012) Go and no-go learning in
Dayan P, Huys QJM (2008) Serotonin, inhibition, and reward and punishment: interactions between affect
negative mood. PLoS Comput Biol 4(2):e4 and effect. Neuroimage 62(1):154–166
Dayan P, Huys QJM (2009) Serotonin in affective control. Hasler G (2012) Can the neuroeconomics revolution rev-
Annu Rev Neurosci 32:95–126 olutionize psychiatry? Neurosci Biobehav Rev
Dickinson A, Smith J, Mirenowicz J (2000) Dissociation 36(1):64–78
of Pavlovian and instrumental incentive learning under Heinz A (2002) Dopaminergic dysfunction in alcoholism
dopamine antagonists. Behav Neurosci and schizophrenia-psychopathological and behavioral
114(3):468–483 correlates. Eur Psychiatry 17(1):9–16
Egerton A, Chaddock CA, Winton-Brown TT, Bloomfield Henriques JB, Glowacki JM, Davidson RJ (1994) Reward
MAP, Bhattacharyya S, Allen P, McGuire PK, fails to alter response bias in depression. J Abnorm
Howes OD (2013) Presynaptic striatal dopamine dys- Psychol 103(3):460–466
function in people at ultra-high risk for psychosis: Hoffman RE (1987) Computer simulations of neural infor-
findings in a second cohort. Biol Psychiatry mation processing and the schizophrenia-mania
74(2):106–112 dichotomy. Arch Gen Psychiatry 44(2):178–188
Huys QJM, Moutoussis M, Williams J (2011) Are com- Lieder F, Goodman N, Huys QJM (2013) Learned help-
putational models of any use to psychiatry? Neural lessness and generalization. Proc Annu Conf Cognit
Netw 24(6):544–551 Sci Soc
Huys QJM, Eshel N, O’Nions E, Sheridan L, Dayan P, Lucantonio F, Stalnaker TA, Shaham Y, Niv Y,
Roiser JP (2012) Bonsai trees in your head: how the Schoenbaum G (2012) The impact of orbitofrontal
Pavlovian system sculpts goal-directed choices by dysfunction on cocaine addiction. Nat Neurosci
pruning decision trees. PLoS Comput Biol 8(3): 15(3):358–366
e1002410 Mackey MC, Milton JG (1987) Dynamical diseases. Ann
Huys QJM, Pizzagalli DA, Bogdan R, Dayan P (2013) NY Acad Sci 504:16–32
Mapping anhedonia onto reinforcement learning: Maia TV, Frank MJ (2011) From reinforcement learning
a behavioural meta-analysis. Biol Mood Anxiety models to psychiatric and neurological disorders. Nat
Disord 3(1):12 Neurosci 14(2):154–162
Jensen J, Willeit M, Zipursky RB, Savina I, Smith AJ, Maier S, Seligman M (1976) Learned helplessness: theory
Menon M, Crawley AP, Kapur S (2008) The formation and evidence. J Exp Psychol Gen 105(1):3–46
of abnormal associations in schizophrenia: neural and Maier SF, Watkins LR (2005) Stressor controllability and
behavioral evidence. Neuropsychopharmacology learned helplessness: the roles of the dorsal raphe
33(3):473–479 nucleus, serotonin, and corticotropin-releasing factor.
Juckel G, Schlagenhauf F, Koslowski M, Filonov D, Neurosci Biobehav Rev 29(4–5):829–841
Wstenberg T, Villringer A, Knutson B, Kienast T, Marr D (1982) Vision. Freeman, New York
Gallinat J, Wrase J, Heinz A (2006) Dysfunction McClelland J, Rumelhart D, Hinton GE (1986) Parallel
of ventral striatal reward prediction in schizophrenic distributed processing. MIT Press, Cambridge, MA
patients treated with typical, not atypical, Miller WR, Seligman ME (1975) Depression and learned
neuroleptics. Psychopharmacology (Berl) helplessness in man. J Abnorm Psychol 84(3):228–238
187(2):222–228 Milton JG (2010) Epilepsy as a dynamic disease: a tutorial
Kapur S (2003) Psychosis as a state of aberrant salience: of the past with an eye to the future. Epilepsy Behav
a framework linking biology, phenomenology, and 18(1–2):33–44
pharmacology in schizophrenia. Am J Psychiatry Montague PR, Dayan P, Sejnowski TJ (1996)
160(1):13–23 A framework for mesencephalic dopamine systems
Kapur S, Zipursky R, Jones C, Remington G, based on predictive Hebbian learning. J Neurosci
Houle S (2000) Relationship between dopamine d 16(5):1936–1947
(2) occupancy, clinical response, and side effects: a Montague PR, Dolan RJ, Friston KJ, Dayan P (2012) Com-
double-blind pet study of first-episode schizophrenia. putational psychiatry. Trends Cogn Sci 16(1):72–80
Am J Psychiatry 157(4):514–520 Moutoussis M, Bentall RP, Williams J, Dayan P (2008)
Kendler KS, Karkowski LM, Prescott CA (1999) Causal A temporal difference account of avoidance learning.
relationship between stressful life events and the Network 19(2):137–160
onset of major depression. Am J Psychiatry Moutoussis M, Bentall RP, El-Deredy W, Dayan P
156:837–841 (2011) Bayesian modelling of jumping-to-conclusions
King R, Barchas JD, Huberman BA (1984) Chaotic behav- bias in delusional patients. Cognit Neuropsychiatr
ior in dopamine neurodynamics. Proc Natl Acad Sci 16(5):422–447
U S A 81(4):1244–1247 Murray GK, Corlett PR, Clark L, Pessiglione M, Black-
King-Casas B, Tomlin D, Anen C, Camerer CF, Quartz well AD, Honey G, Jones PB, Bullmore ET, Robbins
SR, Montague PR (2005) Getting to know you: repu- TW, Fletcher PC (2008) Substantia nigra/ventral teg-
tation and trust in a two-person economic exchange. mental reward prediction error disruption in psychosis.
Science 308(5718):78–83 Mol Psychiatry 13(3):239, 267–276
King-Casas B, Sharp C, Lomax-Bream L, Lohrenz T, Nelson A, Killcross S (2006) Amphetamine exposure
Fonagy P, Montague PR (2008) The rupture and repair enhances habit formation. J Neurosci
of cooperation in borderline personality disorder. Sci- 26(14):3805–3812
ence 321(5890):806–810 Pizzagalli DA, Jahn AL, O’Shea JP (2005) Toward an
Koshelev M, Lohrenz T, Vannucci M, Montague PR objective characterization of an anhedonic phenotype:
(2010) Biosensor approach to psychopathology classi- a signal-detection approach. Biol Psychiatry
fication. PLoS Comput Biol 6(10):e1000966 57(4):319–327
Kumar P, Waiter G, Ahearn T, Milders M, Reid I, Steele Rao RP, Ballard DH (1999) Predictive coding in the visual
JD (2008) Abnormal temporal difference reward- cortex: a functional interpretation of some extra-
learning signals in major depression. Brain classical receptive-field effects. Nat Neurosci
131(Pt 8):2084–2093 2(1):79–87
Laruelle M, Kegeles LS, Abi-Darghama A (2003) Gluta- Ray D, King-Casas B, Montague PR, Dayan P (2009)
mate, dopamine, and schizophrenia: from pathophysi- Bayesian model of behaviour in economic games. In:
ology to treatment. Ann N Y Acad Sci Koller D, Schuurmans D, Bengio Y, Bottou L (eds)
1003(1):138–158 Advances in neural information processing systems 21.
Conductance Scaling 783 C
Proceedings of the twenty-second annual conference Yoshida W, Dziobek I, Kliemann D, Heekeren HR,
on neural information processing systems, Vancouver. Friston KJ, Dolan RJ (2010a) Cooperation and
Curran Associates, New York, pp 1345–1352 heterogeneity of the autistic mind. J Neurosci
Redish AD (2004) Addiction as a computational process 30(26):8815–8818
gone awry. Science 306(5703):1944–1947 Yoshida W, Seymour B, Friston KJ, Dolan RJ (2010b)
Redish AD, Jensen S, Johnson A (2008) A unified frame- Neural mechanisms of belief inference during cooper-
work for addiction: vulnerabilities in the decision pro- ative games. J Neurosci 30(32):10744–10751
cess. Behav Brain Sci 31(4):415–437, discussion
437–487
C
Schultz W, Dayan P, Montague PR (1997) A neural sub-
strate of prediction and reward. Science
275(5306):1593–1599 Computational Structural Biology
Seeman P, Lee T, Chau-Wong M, Wong K (1976) Anti-
psychotic drug doses and neuroleptic/dopamine recep-
▶ Polypeptide and Protein Modeling for Drug
tors. Nature 261(5562):717–719
Seligman ME, Maier SF (1967) Failure to escape trau- Design
matic shock. J Exp Psychol 74(1):1–9
Smith KA, Fairburn CG, Cowen PJ (1997) Relapse of
depression after rapid depletion of tryptophan. Lancet
249:915–919
Smith AJ, Li M, Becker S, Kapur S (2004) A model of Computational Synthesis of
antipsychotic action in conditioned avoidance: a com- Motoneuron Morphology
putational approach. Neuropsychopharmacology
29(6):1040–1049
Smith AJ, Becker S, Kapur S (2005) A computational
model of the functional role of the ventral-striatal d2 Morphology
receptor in the expression of previously acquired
behaviors. Neural Comput 17(2):361–395
Smith AJ, Li M, Becker S, Kapur S (2007) Linking
animal models of psychosis to computational models
of dopamine function. Neuropsychopharmacology Computer Generation of
32(1):54–66 Motoneuron Morphology
Soubrié P (1986) Reconciling the role of central serotonin
neurons in human and animal behaviour. Behav Brain
Sci 9:319–364
Steele JD, Kumar P, Ebmeier KP (2007) Blunted response Morphology
to feedback information in depressive illness. Brain
130(Pt 9):2367–2374
Stephan KE, Baldeweg T, Friston KJ (2006) Synaptic
plasticity and dysconnection in schizophrenia. Biol
Psychiatry 59(10):929–939 Conditional Rate Function
Stephan KE, Penny WD, Daunizeau J, Moran RJ, Friston
KJ (2009) Bayesian model selection for group studies.
▶ Correlation Analysis of Parallel Spike Trains
Neuroimage 46(4):1004–1017
Sutton RS, Barto AG (1998) Reinforcement learning: an
introduction. MIT Press, Cambridge, MA
Tomlin D, Kayali MA, King-Casas B, Anen C, Camerer
CF, Quartz SR, Montague PR (2006) Agent-specific
responses in the cingulate cortex during economic Conductance Co-regulation
exchanges. Science 312(5776):1047–1050
Volkow ND, Fowler JS, Wang GJ, Baler R, Telang ▶ Neuronal Parameter Co-regulation
F (2009) Imaging dopamine’s role in drug abuse and
addiction. Neuropharmacology 56(Suppl 1):3–8
Willner P (1997) Validity, reliability and utility of the
chronic mild stress model of depression: a 10-year
review and evaluation. Psychopharmacology (Berl) Conductance Scaling
134:319–329
Yoshida W, Dolan RJ, Friston KJ (2008) Game theory of
mind. PLoS Comput Biol 4(12):e1000254 ▶ Neuronal Parameter Co-regulation
C 784 Conductance-Based Models of Nonlinear Dynamics in Vertebrate Motoneurons
plateau potentials that develop and persist on

Conductance-Based Models of a slower timescale (order of seconds to minutes)
Nonlinear Dynamics in Vertebrate relative to the millisecond timescale action
Motoneurons potential events. In most adult vertebrates, pla-
teau potentials are generated by noninactivating
Sharmila Venugopal “persistent inward currents” (Lee and Heckman
Department of Physiology, David Geffen School 1998b) predominantly composed of low-voltage-
of Medicine, University of California Los Angeles, activated L-type calcium and to a lesser extent by
Los Angeles, CA, USA persistent sodium currents (Lee and Heckman
1996; Li and Bennett 2003). It is generally
accepted that such currents are of dendritic
Definition origin (Carlin et al. 2000). When activated,
they produce a persistent plateau in the somatic
The term “conductance-based models of membrane voltage that can be sufficient to sus-
nonlinear dynamics in vertebrate motoneurons” tain action potentials even after the removal of
refers to conductance-based motoneuron models plateau producing depolarizing input – also
that capture the complex nonlinear membrane known as “self-sustained firing” (see Fig. 1).
properties (e.g., bistability) of vertebrate moto- Additionally, activation of plateau potentials
neurons observed both in vivo and in vitro under can induce higher-frequency firing for the same
certain conditions. Numerous conductance- depolarizing current injection that produced a
based motoneuron models with widely varying low-frequency firing before plateau activation
morphological complexities and different ionic (see Fig. 1).
current composition exist that reproduce the
experimentally observed nonlinear membrane Models of Motoneuron Bistability
properties. This article focuses on models that Models of motoneuron bistability have faithfully
assume minimally required morphological captured the nonlinear dynamics arising at the
complexity (e.g., two compartments) and ionic soma in experimental settings.
conductances. The resulting reduced set of There are two basic features of these models
equations and parameters describing such that are critical for reproducing the experimen-
models have enabled detailed analyses of model tally observed bistable membrane properties in
behavior using geometric dynamical systems motoneurons. They include (1) the presence of at
methods to provide insight into the basis of least two weak moderately coupled compart-
nonlinear membrane properties in vertebrate ments and (2) sufficiently disparate membrane
motoneurons. properties between the compartments (e.g., den-
dritic localization of persistent inward currents).
A classic two-compartment conductance-based
Detailed Description model that first explored the phenomenology of
vertebrate motoneuron bistability is the Booth-
Dynamics of vertebrate motoneurons are phylo- Rinzel model (Booth and Rinzel 1995) that
genetically conserved sharing common phenom- included dimensionless equations. Similar
enology from turtles to humans (Hornby two-compartment models incorporated more
et al. 2002). Under certain experimental condi- realistic biophysically based ionic conductances
tions, these neurons display nonlinear membrane to generate physiologically realistic membrane
properties such as bistability (Bennett et al. 2001; voltage responses and firing behavior in the spe-
Hounsgaard and Kiehn 1985; Schwindt and Crill cific motoneurons being modeled (e.g., Booth
1977; Sigvardt et al. 1985; Lee and Heckman et al. 1997; Kurian et al. 2011; Venugopal
1998a). Physiologically, bistability arises due to et al. 2011, 2012).
Conductance-Based Models of Nonlinear Dynamics in Vertebrate Motoneurons 785 C
Conductance-Based Models of Nonlinear Dynamics self-sustained firing (light green). Black circles indicate
in Vertebrate Motoneurons, Fig. 1 Bistable mem- zero Isoma. Right: Figure shows the two levels of Isoma, low
brane behavior in vertebrate motoneurons. Left: current (black triangle) that initially induces a -
Somatic current injection (Isoma) marked by red asterisk low-frequency firing in the soma (orange) but does not
induced somatic spikes (dark green) that in turn also activate the dendrite followed by a brief high Isoma (red
activates dendritic plateau (dendrite switches from OFF asterisk) that turns on the plateau. Subsequent return of
to ON state). This dendritic plateau persists even after the Isoma to its initial value (black triangle) results in a higher-
somatic current returns to its initial level (demarcated by frequency firing (yellow) due to the additional persistent
vertical dotted line) at which the soma was silent. This depolarization by the dendritic plateau
sustained plateau now drives somatic spikes resulting in
Conductance-Based Models of Nonlinear Dynamics compartments, namely, soma and dendrite. Arrows show
in Vertebrate Motoneurons, Fig. 2 Minimal moto- the directions of inward and outward currents (see text for
neuron model of bistability. Left: A schematic showing description of currents). Right: Equations guiding the
the morphological and electrophysiological components membrane voltage dynamics in somatic and dendritic
of the model consisting of two electrotonically coupled compartments
A Minimal Motoneuron Model for Shown here are minimal set of ionic currents
Understanding Bistable Membrane Behavior that include fast sodium (INa) and delayed recti-
Figure 2 shows the model schematic of a minimal fier potassium (IKdr) in the soma for action poten-
two-compartment vertebrate motoneuron. The tial generation, a voltage-dependent persistent
equations guiding this minimal model are given inward current in the dendrite (IPIC) that mediates
below: dendritic plateau, a leak current (Ileak) in both the
C 786 Conductance-Based Models of Nonlinear Dynamics in Vertebrate Motoneurons
compartments to maintain the resting potential, one can easily incorporate additional terms for
a coupling current (Icoup) that arises due to the ionic currents as well as increase the morpholog-
electrotonic coupling of the compartments whose ical complexity as relevant. Alternatively, such
magnitude depends on the coupling conductance minimal models could guide the development of
(gc), and the proportion of area (p) suggested to artificial motoneurons that may act as bistable
be occupied by the soma. The term Isoma refers to switches in neuromorphic circuits.
somatic current injection. The somatic and den-
dritic voltages are given by VS and VD, respec- Hysteresis in the Frequency-Current
tively. For a detailed description of the model, see Relationship and Mechanism Underlying
(Booth and Rinzel 1995). Note that these form Bistability
a general set of equations and conductances guid- Motoneuron bistability described above can give
ing the somatic and dendritic compartments and rise to a counterclockwise hysteresis in the
Conductance-Based Models of Nonlinear Dynamics Blue dotted lines demarcate the amplitude of dendritic
in Vertebrate Motoneurons, Fig. 3 Bifurcation ana- oscillations due to soma firing in the primary range (only
lyses of emergence of f-I hysteresis and bistability. Top: stable oscillations are shown here). Red dotted lines
Schematic of f-I relationship in response to a somatic demarcate the amplitude of dendritic oscillations due to
triangular ramp current injection (Isoma trace not shown, soma firing in the tertiary range. Note that there are no
but see curved arrow showing direction of Isoma changes). stable steady-state oscillations for the secondary range in
Spiking begins in the soma in the primary frequency range this diagram suggesting that these frequencies are transi-
(blue) and continues into a steeper secondary range (red) tory. Bistability-1 (region between left orange and
as Isoma ramps up. On the descending Isoma ramp, tertiary magenta vertical lines) marks the region where tertiary
frequencies (green) give rise to a counterclockwise hys- firing and (stable) non-spiking state can coexist underly-
teresis in the f-I relationship. Bottom: Bifurcation curve ing self-sustained firing described earlier. Bistability-2
showing the various steady-state regimes. The bold lines (region between magenta and right orange vertical lines)
in the steady-state S-shaped curve indicate stable fixed marks the region where primary and tertiary firing can
points and broken lines indicate unstable fixed points. coexist as described earlier in Fig. 1 (right)
Conductance-Based Models of Nonlinear Dynamics in Vertebrate Motoneurons 787 C
injected current firing frequency relationship (f-I) inputs in turn regulating motoneuron firing fre-
(see Fig. 3, top) although existence of f-I hyster- quencies and muscle output.
esis itself is not sufficient for bistability. In summary, conductance-based mathemati-
Of importance in the f-I curve shown are the cal models of vertebrate motoneurons have been
three ranges of firing frequencies, (1) primary, useful platforms to examine nonlinear dynamics
(2) secondary, and (3) tertiary ranges (see Kernell in these neurons and have provided important
1965; Schwindt 1973 for the first descriptions), insights into the basis of experimentally observed C
that together lead to a counterclockwise hystere- membrane properties in this important class of
sis to the f-I relationship. Use of geometric cells in the nervous system.
dynamical systems methods has provided insight
into the mechanism of bistability and the emer-
gence of the three frequency ranges. Figure 3
Cross-References
(bottom) shows a typical bifurcation structure of
the dendritic voltage VD with Isoma as the param-
▶ Compartmental Models of Spinal
eter. In the class of two-compartment models
Motoneurons
described above, this corresponds to an
▶ Morphologically Detailed Motoneuron Models
S-shaped curve for a range of coupling conduc-
tances characterized as weak-moderate (e.g., see
Booth and Rinzel 1995; Kurian et al. 2011). The
lower branch (dendrite OFF) consists of stable References
fixed points (non-spiking steady states/resting)
Bennett DJ, Li Y, Siu M (2001) Plateau potentials in
and a stable oscillatory regime. The upper
sacrocaudal motoneurons of chronic spinal rats,
branch (dendrite ON) also consists of stable recorded in vitro. J Neurophysiol 86:1955–1971
fixed points (non-spiking plateau state) and Booth V, Rinzel J (1995) A minimal, compartmental
a stable oscillatory regime. The height and model for a dendritic origin of bistability of motoneu-
ron firing patterns. J Comput Neurosci 2:299–312
width of the S shape as well as that of the
Booth V, Rinzel J, Kiehn O (1997) Compartmental
regimes described are dependent on the proper- model of vertebrate motoneurons for Ca2+
ties of the ionic conductances considered (e.g., dependent spiking and plateau potentials under
note differences among the various models pharmacological treatment. J Neurophysiol
78:3371–3385
Booth et al. 1997; Kurian et al. 2011; Venugopal
Carlin KP, Jones KE, Jiang Z, Jordan LM, Brownstone
et al. 2011). Superimposing the dendritic volt- RM (2000) Dendritic L-type calcium currents in
age changes during a somatic triangular ramp mouse spinal motoneurons: implications for
current injection in the model clarifies that bistability. Eur J Neurosci 12:1635–1646
Hornby TG, McDonagh JC, Reinking RM, Stuart DG
(1) the primary firing frequencies correspond to
(2002) Motoneurons: a preferred firing range across
the oscillatory frequencies emerging at the vertebrate species? Muscle Nerve 25:632–648
lower branch, (2) tertiary frequencies corre- Hounsgaard J, Kiehn O (1985) Ca2+ dependent bistability
spond to the oscillatory frequencies emerging induced by serotonin in spinal motoneurons. Exp Brain
Res 57:422–425
at the upper branch, and (3) secondary frequen-
Kernell D (1965) High-frequency repetitive firing of cat
cies would arise during plateau development and lumbosacral motoneurons stimulated by long-lasting
are transitory see (Venugopal et al. 2011). The injected currents. Acta Physiol Scand 65:74–86
same is illustrated using the steady-state bifur- Kurian M, Crook SM, Jung R (2011) Motoneuron model
of self-sustained firing after spinal cord injury.
cation diagram (see Fig. 3) of dendritic voltage J Comput Neurosci 31:625–645
in response to Isoma clarifying that only 1 and 3 Lee RH, Heckman CJ (1996) Influence of voltage-
frequencies are stable oscillatory states, while sensitive dendritic conductances on bistable firing
the secondary range arising during plateau and effective synaptic current in cat spinal motoneu-
rons in vivo. J Neurophysiol 76:2107–2110
development is not evident as a stable steady
Lee RH, Heckman CJ (1998a) Bistability in spinal moto-
state. This critical insight suggests the possible neurons in vivo: systematic variations in rhythmic
window of modulation of the plateau by synaptic firing patterns. J Neurophysiol 80:572–582
C 788 Connectionist Models of CPG Networks
Lee RH, Heckman CJ (1998b) Bistability of spinal moto- Network Topology: At the connectional
neurons in vivo: systematic variations in persistent level, a CPG network consists of “unit CPGs”
inward currents. J Neurophysiol 80:583–593
Li Y, Bennett DJ (2003) Persistent sodium and calcium representing the nodes of the network, and the
currents cause plateau potentials in motoneurons of reciprocal connections between unit CPGs corre-
chronic spinal rats. J Neurophysiol 90:857–869 spond to the branches of the network. Physiolog-
Schwindt PC (1973) Membrane-potential trajectories ically, a unit CPG represents a group of neurons
underlying motoneuron rhythmic firing at high rates.
J Neurophysiol 36:434–439 that can generate a recurrent burst of activity
Schwindt PC, Crill WE (1977) A persistent negative resis- (Grillner 2006).
tance in cat lumbar motoneurons. Brain Res Simple inputs such as tonic excitation can
120:173–178 produce rhythmic outputs in these networks.
Sigvardt KA, Grillner S, Wallén P, Donger PAM
(1985) Activation of NMDA receptors elicits fictive Increasing the stimulus strength typically
locomotion and bistable membrane properties in the increases the frequency of the network rhythm.
lamprey spinal cord. Brain Res 336:390–395 Two unit CPGs are reciprocally coupled by
Venugopal S, Hamm TM, Crook SM, Jung R (2011) Mod- mutual inhibition to generate alternating activity
ulation of inhibitory strength and kinetics facilitates
regulation of persistent inward currents and motoneu- patterns (Fig. 1a) or by mutual excitation to gen-
ron excitability following spinal cord injury. erate synchronized bursts (Fig. 1b). Many unit
J Neurophysiol 106:2167–2179 CPGs connected in a chain with mutual excita-
Venugopal S, Hamm TM, Jung R (2012) Differential tion can generate a progressive phase lag in the
contributions of somatic and dendritic calcium-
dependent potassium currents to the control of moto- bursts of activity along the CPG chain (Fig. 1c). It
neuron excitability following spinal cord injury. is generally accepted that rhythmogenesis is
Cognit Neurodyn 6(3):283–293 intrinsic to the neurons forming the unit CPG.
However, connectionist models of CPG networks
have unequivocally demonstrated that rhythmic
patterns can be a resultant of a topology of con-
nections among unit CPGs (Buchanan 1992).
Connectionist Models of CPG Hence, based on the nature of connections
Networks between unit CPGs, a CPG network can generate
and shape patterns of activity.
Sharmila Venugopal Reconfiguration of CPGs: Central pattern
Department of Physiology, David Geffen generators can reconfigure to switch from one
School of Medicine, University of California activity pattern to another to produce multiple
Los Angeles, Los Angeles, CA, USA forms of the same behavior (e.g., walking, run-
ning, and galloping as different forms of locomo-
tion) or functionally distinct behaviors (e.g.,
Detailed Description normal breathing versus gasping). Mathematical
and simulation models suggest that the mecha-
Central Pattern Generators (CPGs) are neural nisms underlying the reconfiguration of CPGs
networks that can produce coordinated patterns may include (1) altered strength and duration of
of rhythmic activity to orchestrate repetitive the excitatory drive to the CPG network (e.g.,
behaviors such as feeding, locomotion, and res- Jung et al. 1996), (2) altered strength and duration
piration. CPGs are known to be composed of of the mutual connections between unit CPGs
phylogenetically conserved connectional units (e.g., Venugopal et al. 2007), and (3) recruitment
that are typically organized into chains to enable and de-recruitment of unit CPGs (e.g., Nasse
coordinated activation of the effector organs et al. 2008). Such reconfiguration of a given
(muscles). Through integration of experimental CPG network architecture can confer
and computational approaches at the cellular and multifunctionality.
network levels, essential connectional elements Unit CPG Models: Biophysical models
of CPGs have been revealed (Grillner 2006). of unit CPGs are typically based on
Connectionist Models of CPG Networks 789 C
Connectionist Models of CPG Networks, the duration of CPG drive signal, and blue and red wave-
Fig. 1 Schematic of typical connectional elements of forms represent out-of-phase (a) and in-phase (b) activity.
a CPG network. Unit CPGs with (a) reciprocal inhibitory In (c), the thin black lines indicate reduced drive to the
(blue) and (b) reciprocal excitatory (red) connections are corresponding unit CPGs such that the top unit leads and
shown. (c) A chain of unit CPGs with reciprocal excitatory subsequent phase lags are produced in the onset of activity
connections. The thick black lines in a, b, and c represent as we go down the chain
conductance-based Hodgkin-Huxley formalism physiological ranges (Prinz et al. 2004). Alterna-

to model the constituent neurons with one or tively, artificial CPG networks for engineering
more compartments (e.g., see review Grillner applications have utilized artificial neural net-
et al. 2005). In models that exclusively focused work learning algorithms to estimate network
on network properties, nonlinear oscillators have parameters (e.g., Ijspeert 2008).
been widely used to represent unit CPGs (e.g., see
review Ijspeert 2008). It appears that the network
activity pattern of interest can be produced Cross-References
irrespective of the type of the oscillator model
(e.g., Collins and Richmond 1994). This provides ▶ Pulse-Coupled Oscillators
various options for choosing a suitable oscillator
to represent the unit CPGs/nodes, particularly
when mathematical analyses are adopted for References
gaining mechanistic insight into network behav-
ior. Moreover, simple oscillator models are suit- Buchanan JT (1992) Neural network simulations of
able choices for prosthetic and/or robotic coupled locomotor oscillators in the lamprey spinal
applications. cord. Biol Cybern 66:367–374
Collins JJ, Richmond SA (1994) Hard-wired central pat-
Parameter Tuning in CPG Networks: Iden- tern generators for quadrupedal locomotion. Biol
tification of intrinsic and synaptic properties/ Cybern 71:375–385
parameters of CPG networks that produce the Grillner S, Kozlov A, Kotaleski JH (2005) Integrative
shape and the nuances of activity patterns is at neuroscience: linking levels of analyses. Curr Opin
the core of neurobiological investigation of Grillner S (2006) Biological pattern generation: the cellu-
CPGs. Biophysical models of CPGs often adopt lar and computational logic of networks in motion.
brute force methods for parameter tuning within Neuron 52:751–766
C 790 Connectivity Analysis in Normal and Pathological Brains
Ijspeert AJ (2008) Central pattern generators for locomo- combining segregation and integration of com-
tion control in animals and robots: a review. Neural ponents. Brain connectivity can be intuitively
Netw 21:642–653
Jung R, Kiemel T, Cohen AH (1996) Dynamic behavior of represented as graphs, where nodes represent
a neural network model of locomotor control in the neural elements, ranging in scale from individual
lamprey. J Neurophysiol 75(3):1074–1086 cells to large-scale neural populations (e.g., cor-
Nasse J, Terman D, Venugopal S, Hermann G, Rogers R, tical areas) and links, representing structural or
Travers JB (2008) Local circuit input to the medullary
reticular formation from the rostral nucleus of the functional associations between the nodes. This
solitary tract. Am J Physiol Regul Integr Comp Physiol simplifying approach is based on the assumption
295(5):R1391–R1408 that neural elements are intrinsically homoge-
Prinz AA, Bucher D, Marder E (2004) Similar network neous and that their interactions are determined
activity from disparate circuit parameters. Nat
Neurosci 7(12):1345–1352 by anatomical connections.
Venugopal S, Travers JB, Terman DH (2007) A compu- Several types of connectivity can be distin-
tational model for motor pattern switching between guished (Friston 2004):
taste-induced ingestion and rejection oromotor behav- • Structural: the physical coupling of neural
iors. J Comput Neurosci 22(2):223–238
elements by neuronal connections
• Functional: statistical dependencies in the
dynamics of neural elements
• Effective: the dynamic or functional impact
Connectivity Analysis in Normal and exerted by one neural element on another
Pathological Brains Node elements can be defined by morpholog-
ical, anatomical, or functional criteria, for
Claus C. Hilgetag instance, anatomical landmarks, cytoarchi-
Department of Computational Neuroscience, tectonic characterization, or functional activation
University Medical Center Hamburg–Eppendorf, patterns, respectively. Moreover, nodes can also
Hamburg University, Hamburg, Germany be parcellated by their characteristic connec-
Department of Health Sciences, Boston tional fingerprint. Importantly, the way in which
University, Boston, USA node parcels are defined also affects the density
and topological features of the studied networks
(de Reus and van den Heuvel 2013).
Synonyms Structural connectivity is mostly derived
from invasive approaches in animal models
Connectomics; Graph analysis; Network science (histochemical or viral pathway tracing, polar-
ized light imaging (PLI) of fibers, or optogenetic
manipulation of projection neurons) and nonin-
Definition vasive investigations of the human brain (indirect
characterization of tracts by diffusion-weighted
Connectivity analysis focuses on the neural net- magnetic resonance imaging, dwMRI).
work basis of the brain and characterizes topo- Depending on the studied type of connectivity
logical and other aspects of brain network and empirical constraints, one can consider
organization that appear relevant for understand- directed versus undirected connectivity. For
ing normal and pathological brain function. instance, human brain structural connectivity
derived from diffusion-based MR imaging is typ-
ically undirected, whereas histochemical tracing
Detailed Description in animal models allows identifying the origins
and terminations of projections. Another distinc-
Brains as Networks tion can be made of binary versus weighted net-
The network perspective of the brain integrates works. The former contain information on the
aspects of local versus distributed brain function, existence or absence of links, while the latter
Connectivity Analysis in Normal and Pathological Brains 791 C
also include information on the strength of edges inverse, local efficiency) is frequently taken as
(Rubinov and Sporns 2011). an indicator for the potential length of mini-
Network organization matters functionally, in mal interaction distances in a network.
the ultimate sense that there are systematic con- • Local clustering (cliquishness): the proportion
nectivity differences between the brains of of existing connections among the neighbors
healthy subjects and patient groups (Bullmore of a node.
and Sporns 2012). Central aspects of connectivity C
organization are network geometry (i.e., the spa- Global Network Features
tial embedding of networks) versus topology • Overall density of wiring: the total number of
(nonspatial features of connectional organiza- existing links in a network, out of all
tion) (Bullmore and Sporns 2009; Sporns possible ones.
et al. 2004). • The average shortest path (or its inverse,
global efficiency).
• Network diameter: the longest shortest path.
Topological Features of Brain • Further features can are considered at the scale
Connectivity of the whole network including assortivity: the
similarity of the degree of linked nodes, as
Topology describes the connectional organiza- well as averages of node or circuit-based indi-
tion of a brain network, independent of its spatial ces, such as clustering.
arrangement. Topological features can be loosely • Existence, number, and size of modules,
ordered from more local to more global aspects. which are formed by nodes that are more fre-
Some features are summarized in the following quently or densely linked with each other than
list. They are described in more detail in the with the rest of the network.
recent research literature (Kaiser 2011; Rubinov • Hubs, which can be identified by a variety of
and Sporns 2010; Sporns 2011; Stam and measures, such as the node degree, node cen-
Reijneveld 2007). trality (in terms of representation in the net-
work paths), or vulnerability, and which may
Node Features be global or provincial hubs depending on
• Degree: the total number of incoming their placement in the network.
(afferent, in-degree) and outgoing (efferent, • Hierarchy: the encapsulation of features
out-degree) links of a node; the degree distri- within features (e.g., modules of modules in
bution of nodes is a characteristic indicator of hierarchical modular networks); however,
different classes of networks. hierarchy can also be defined in various other
• Strength: the graded degree of weighted ways, e.g., by the local versus global access of
connections. nodes in a network (M€uller-Linow et al. 2008).
• Further node indices can be defined, such as It is important to note that network features at the
the proportion between incoming and outgo- same or different scales may be interrelated. For
ing projections (Kötter and Stephan 2003). instance, dense networks may result in densely
wired motifs.
Circuit Features
• Motifs or circuits are subnetworks consisting Benchmark Networks
of relatively few nodes or links; motif spectra Benchmark networks can be based on the degree
appear to be characteristic for different classes distribution of nodes. In so-called Erdös-Rènyi or
of biological networks (Milo et al. 2004). random networks, the probability, p(k), of nodes
• Paths are ordered sequences of unique edges, possessing an edge degree k follows a Gaussian
while walks are sequences of nonunique distribution, whereas the degree of nodes is con-
edges; cycles are paths leading back to the stant in regular or lattice networks. Small-world
same node; the length of shortest paths (or its networks can be seen as regular networks with
C 792 Connectivity Analysis in Normal and Pathological Brains
a fraction of random links; they possess cluster- interactions between remote regions. The latter
ing comparable to regular networks and average pathologies can be considered as disconnexion
shortest paths similar to random graphs. In scale- (disrupted connectivity) (Geschwind 1965) or
free networks, p(k) follows a power law; thus, dysconnection (maladjusted connectivity) syn-
these networks lack a characteristic scale. Bench- dromes (Stephan et al. 2009). Systematic differ-
mark networks can also be constructed for ences in structural and functional connectivity
a variety of other topological measures, for have been found between healthy controls and
instance, yielding random modular graphs or dif- different patient samples, for example, with
ferent types of hierarchical networks (Kaiser autism or schizophrenia spectrum disorders
2011). (Bullmore and Sporns 2012).
There are also several established datasets for A well-studied example of the network per-
biological connectomes, in particular for cortico- spective is Alzheimer’s disease (AD), in which
cortical connections in the cat (Scannell highly connected hub nodes of human functional
et al. 1999), the rhesus macaque (Felleman and brain connectivity show high metabolic activity
Van Essen 1991; Markov et al. 2014) as well as as well as an affinity for the accumulation of beta-
mesoscopic connections of the human brain amyloid protein, thus, becoming the origin for the
based on diffusion MRI (Hagmann et al. 2008). propagation of AD (Buckner et al. 2009).
Moreover, connectomes or partial connectomes Generally, however, it is currently unclear
at the cellular level exist for Caenorhabditis which network features contribute causally to dif-
elegans (Varshney et al. 2011), drosophila ferent aspects of network dynamics or function.
(Chiang et al. 2011), and zebra fish (Stobb
et al. 2012). From these datasets, further bench-
marks can be derived, such as randomized ver- Cross-References
sions preserving features, for example, the degree
distribution, or pairwise degree distribution. ▶ Connectome, General
▶ Human Connectome Project
Geometry of Brain Connectivity ▶ Neuropathologies and Networks
▶ Wiring Principles, Optimization
Brain networks are spatially embedded. The dis-
tance among neural elements may influence
physiological and functional parameters through
References
delays and metabolic costs and place constraints
on the existence and strength of links. Conse- Buckner RL, Sepulcre J, Talukdar T, Krienen FM, Liu H,
quently, it is a popular idea that brain networks Hedden T, Andrews-Hanna JR, Sperling RA, Johnson
should have compact wiring. However, other KA (2009) Cortical hubs revealed by intrinsic func-
structural, functional, metabolic, or evolutionary tional connectivity: mapping, assessment of stability,
and relation to Alzheimer’s disease. J Neurosci
constraints might also be important and there 29(6):1860–1873
may be trade-offs between various, partly antag- Bullmore E, Sporns O (2009) Complex brain networks:
onistic constraints on brain connectivity (Chen graph theoretical analysis of structural and functional
et al. 2013; Kaiser and Hilgetag 2006). systems. Nat Rev Neurosci 10(3):186–198
Bullmore E, Sporns O (2012) The economy of brain net-
work organization. Nat Rev Neurosci 13(5):336–349
Chen Y, Wang S, Hilgetag CC, Zhou C (2013) Trade-off
Pathological Brain Connectivity between multiple constraints enables simultaneous
formation of modules and hubs in neural systems.
PLoS Comput Biol 9(3):e1002937
Aberrations in brain function arise from pertur- Chiang A-S, Lin C-Y, Chuang C-C, Chang H-M, Hsieh
bations of local neural structures and mechanisms C-H, Yeh C-W, Shih C-T, Wu J-J, Wang G-T, Chen
(i.e., node perturbations) as well as perturbed Y-C, Wu C-C, Chen G-Y, Ching Y-T, Lee P-C,
Connectome, Drosophila 793 C
Lin C-Y, Lin H-H, Wu C-C, Hsu H-W, Huang Y-A, Sporns O, Chialvo DR, Kaiser M, Hilgetag CC
Chen J-Y, Chiang H-J, Lu C-F, Ni R-F, Yeh C-Y, (2004) Organization, development and function of
Hwang J-K (2011) Three-dimensional reconstruction complex brain networks. Trends Cogn Sci
of brain-wide wiring networks in Drosophila at single- 8(9):418–425
cell resolution. Curr Biol 21(1):1–11 Stam CJ, Reijneveld JC (2007) Graph theoretical analysis
de Reus MA, van den Heuvel MP (2013) The parcellation- of complex networks in the brain. Nonlinear Biomed
based connectome: limitations and extensions. Phys 1(1):3
Neuroimage 80:397–404 Stephan KE, Friston KJ, Frith CD (2009) Dysconnection
Felleman DJ, Van Essen DC (1991) Distributed hierarchi- in schizophrenia: from abnormal synaptic plasticity to
C
cal processing in the primate cerebral cortex. Cereb failures of self-monitoring. Schizophr Bull
Cortex 1(1):1–47 (New York, NY: 1991) 35(3):509–527
Friston KJ (2004) Functional and effective connectivity in Stobb M, Peterson JM, Mazzag B, Gahtan E (2012) Graph
neuroimaging: a synthesis. Hum Brain theoretical model of a sensorimotor connectome in
Mapp 2(1–2):56–78 zebrafish. PLoS ONE 7(5):e37292
Geschwind N (1965) Disconnexion syndromes in animals Varshney LR, Chen BL, Paniagua E, Hall DH, Chklovskii
and man. I. Brain J Neurol 88(2):237–294 DB (2011) Structural properties of the Caenorhabditis
Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey elegans neuronal network. PLoS Comput Biol 7(2):
CJ, Van Wedeen J, Sporns O (2008) Mapping the e1001066
structural core of human cerebral cortex. PLoS Biol
6(7):1479–1493
Further Reading
Kaiser M (2011) A tutorial in connectome analysis: topo-
Special issue: connectivity (2012) NeuroImage 62(4). http://
logical and spatial features of brain networks.
www.sciencedirect.com/science/journal/10538119/62
Neuroimage 57(3):892–907
Special issue: the connectome (2013) Trends Cognit Sci
Kaiser M, Hilgetag CC (2006) Nonoptimal component
17(12). http://www.sciencedirect.com/science/jour-
placement, but short processing paths, due to long-
nal/13646613/17
distance projections in neural systems. PLoS Comput
Sporns O (2011) Networks of the brain. MIT Press,
Biol 2(7):e95
Cambridge
Kötter R, Stephan KE (2003) Network participation indi-
ces: characterizing component roles for information
processing in neural networks. Neural Netw
16(9):1261–1275
Markov NT, Ercsey-Ravasz MM, Ribeiro Gomes AR,
Lamy C, Magrou L, Vezoli J, Misery P, Falchier A, Connectivity of the Visual Areas
Quilodran R, Gariel MA, Sallet J, Gamanut R,
Huissoud C, Clavagnier S, Giroud P, Sappey-Marinier-
D, Barone P, Dehay C, Toroczkai Z, Knoblauch K, Van ▶ Multiscale Brain Connectivity
Essen DC, Kennedy H (2014) A weighted and directed
interareal connectivity matrix for macaque cerebral cor-
tex. Cereb Cortex 24(1):17–36
Milo R, Itzkovitz S, Kashtan N, Levitt R, Shen-Orr S, Connectome, Drosophila
Ayzenshtat I, Sheffer M, Alon U (2004) Superfamilies
of evolved and designed networks. Science
(NY) 303(5663):1538–1542 Arjun Bharioke1, Louis K. Scheffer1,
M€uller-Linow M, Hilgetag CC, Hutt M-T (2008) Organi- Dmitri B. Chklovskii1 and Ian A. Meinertzhagen2
zation of excitable dynamics in hierarchical biological 1
Janelia Farm Research Campus, Howard
networks. PLoS Comput Biol 4(9):e1000190
Hughes Medical Institute, Ashburn, VA, USA
Rubinov M, Sporns O (2010) Complex network measures 2
of brain connectivity: uses and interpretations. Department of Psychology and Neuroscience,
Neuroimage 52(3):1059–1069 Dalhousie University, Halifax, NS, Canada
Rubinov M, Sporns O (2011) Weight-conserving charac-
terization of complex functional brain networks.
Neuroimage 56(4):2068–2079
Scannell JW, Burns GA, Hilgetag CC, O’Neil MA, Young Definition
MP (1999) The connectional organization of the
cortico-thalamic system of the cat. Cereb Cortex The Drosophila connectome is a comprehensive
9(3):277–299 (New York, NY: 1991)
description of all the connections between sub-
Sporns O (2011) The non-random brain: efficiency,
economy, and complex dynamics. Front Comput units comprising the central nervous system
Neurosci 5:5 (CNS) of the fruit fly, Drosophila melanogaster.
C 794 Connectome, Drosophila
connectomes have been reported. In particular,

cellular connectomes have been reported for
a cartridge of the lamina (Meinertzhagen and
O’Neil 1991; Rivera-Alba et al. 2011) and
a column of the medulla (Takemura et al. 2013),
the first and second neuropil compartments,
respectively, within the fly optic lobe (transmit-
ting visual information). Despite their incomplete
nature, these partial connectomes have already
provided valuable insights into the computational
Connectome, Drosophila, Fig. 1 Drosophila brain functions of neuronal circuits.
with anatomically identified neuropil compartments indi-
vidually colored (www.flybrain.org (AB00122)). Scale
bar: 100 mm
Existing Neuropil-Level Connectomes
The CNS of Drosophila consists of the brain Large-scale, compartmental maps of the Dro-
and ventral ganglion. It has a volume of ~0.07 sophila brain have been obtained through the
mm3 (see Rein et al. 2002, Table 1). To compare, use of genetic reporters imaged by light micros-
this is similar to the volume of a single cortical copy (Chiang et al. 2011). Each reporter line
column in mouse barrel cortex (~0.09 mm3) fluorescently labels a sparse subset of neurons
(Lefort et al. 2009; Hooks et al. 2011). The Dro- (Jenett et al. 2012; Ito et al. 2013). Aligning
sophila brain can be subdivided into smaller images from multiple lines to a standard Dro-
neuropil compartments (Fig. 1), identified sophila brain, the overlapping arbors of locally
through anatomical boundaries (Ito et al. 2014), arborizing neurons are used to define individual
as well as the overlapping arborizations of the compartments: local processing units (LPUs)
neurons constituting each neuropil compartment (Chiang et al. 2011). The connections between
(Chiang et al. 2011; Jenett et al. 2012). In total, these LPUs are found by identifying neurons
across these compartments, the Drosophila projecting between them. In total, the resulting
brain is thought to contain ~90,000 cells, of connectome contains 58 tracts between the 41
which ~90 % are neurons. These neurons can be identified LPUs, a sparse subset (<10 %) of the
classified into discrete cell types, with stereotyp- 820 numerically possible LPU-LPU connections.
ical shapes and synaptic connections, using both
human observation and genetic reporter lines that
label individual types. Cellular Connectomes and EM
A Drosophila connectome, within the brain,
can be constructed at different levels of resolu- A neuropil-level connectome, generated from
tion. At low resolution, a connectome can be said light microscopy, has insufficient resolution to
to comprise neuropil compartments, with con- identify synapses, and hence cannot detail the
nections between them defined by neural path- true connections between neurons. The large-
ways, as identified by light microscopy. In scale, dense identification of all synapses within
contrast, at high resolution, a connectome com- a region of neuropil requires higher resolution
prises individual neurons and all the chemical EM-based methods, typically serial-section
synapses between them. The components of transmission EM (ssTEM) (Fig. 2). However,
such a cellular connectome are identified through these techniques are labor intensive and time
electron microscopy (EM). consuming. Therefore, only a few partial
As of 2013, neuropil-level connectomes have cellular connectomes have been reconstructed,
been constructed for the entire Drosophila brain specifically for two Drosophila optic lobe
(Chiang et al. 2011), but only partial cellular neuropils.
Connectome, Drosophila, Fig. 2 Reconstructing isolate neuronal cross sections (neurite profiles in single
a cellular connectome using ssTEM. (a) TEM image of colors). (c) Profiles in consecutive sections (left) are
part of a 40 nm thick section, one of 2,769 sections imaged linked to construct 3D neurites (right). (d) A subset of
to reconstruct the medulla connectome. (b) Applying reconstructed neurons, embedded within the imaged
computational segmentation algorithms, followed by neuropil. (e) Rendering of neurites belonging to 379
human proofreading on the TEM images, is necessary to reconstructed neurons. Scale bar: 500 nm (a, b)
Optic Lobe Cellular Connectomes non-optic-lobe neuropils, though the number of

postsynaptic targets may vary across neuropils.
The optic lobe of flies has been the subject of
experimental interest for over 50 years, driven by Lamina
the advanced visual behaviors of flies (Heisen- The lamina receives most of its input directly
berg and wolf 1984). Despite the large size of the from photoreceptors in the retina. It is thought
optic lobe, encompassing almost 50 % of the total to be mostly responsible for encoding contrast
CNS volume (Rein et al. 2002), the modular and for light adaptation. A cellular connectome
structure of its underlying neuropil compart- of the repeating unit of the lamina, the lamina
ments, corresponding to the ommatidia array of cartridge, has been reconstructed twice, in two
the compound eye, is highly amenable to different wild-type strains (Meinertzhagen and
connectomics. O’Neil 1991; Rivera-Alba et al. 2011). Both
Within the optic lobes, synapses are divergent reconstructions are similar, with most differences
polyads (Fig. 3), i.e., each presynaptic speciali- apparently introduced by improvements in recon-
zation (or “T-bar”) is typically associated struction quality and by differences in the inter-
with four to six postsynaptic dendrites. This pretation of some postsynaptic elements. The
divergence is expected to generalize to consensus lamina connectome contains the
C 796 Connectome, Drosophila
Connectome, Drosophila, Fig. 3 Divergent polyadic

synapse. The presynaptic process, identified with a T-bar
(red arrow), is associated with four postsynaptic dendrites
(blue arrowheads). These postsynaptic processes can be
distinguished from adjacent, unconnected processes (e.g.,
green circle) by their postsynaptic densities. Scale bar:
250 nm
terminals of six photoreceptors, along with

12 types of neurons (Rivera-Alba et al. 2011;
Tuthill et al. 2013). In total, each cartridge con-
tains ~480 presynaptic sites, contacting ~1,250 Connectome, Drosophila, Fig. 4 3D graph of
postsynaptic sites. a connectome module: the connections between neurons
found in every medulla column. The dominant direction of
signal flow is oriented into the page. Cell types with
Medulla stronger connections are positioned closer to each other.
Directly downstream of the lamina, the medulla The colors identify three spatially segregated groups of
is the single largest neuropil of the fly’s brain and neurons and, hence, three pathways of information trans-
receives visual input both through the lamina and mission within the medulla
directly from a subset of photoreceptors. It is
thought to be responsible for (1) the computation are far more cross-column connections in the
of local visual motion, by the comparison of medulla than in the lamina (Riviera-Alba et al.
inputs from different points in space, and 2011).
(2) color vision, utilizing the direct inputs from
spectrally tuned photoreceptors. Characteristics of the Drosophila Connectome
A connectome of the repeating unit within the Presynaptic terminals within the existing
medulla, the column, was reconstructed recently connectome reconstructions are typically of uni-
(Takemura et al. 2013) (Fig. 4). It reports the form size, and the number of synaptic contacts
connections between 56 identified cell types, of between different neuron pairs ranges over
which 27 are modular (contained in every more than two orders of magnitude (1–150).
column). As in the lamina, each presynaptic Hence the synaptic weight for any pair is thought
site contacts multiple postsynaptic elements to be well approximated by the number of synap-
(~3.8 on average). Within a column, the modular ses in parallel. However, it is not known if it will
cell types are connected by ~2,500 synaptic con- be possible to generalize this quantitative assump-
tacts. However, this is only a subset of the ~5,800 tion to the rest of the Drosophila connectome.
synaptic contacts that these same modular cell Synapse numbers for identified neurons are
types make in total, showing clearly that there consistent across neurons of the same type in
the lamina and medulla. However, as has been the exploration of color vision circuits
seen in both the antennal lobe (Chou et al. 2010) (Franceschini et al. 1981).
and the mushroom body (Murthy et al. 2008;
Caron et al. 2013), this may not generalize to
other parts of the brain. In general, learning asso- The Future
ciated or other forms of plasticity may decrease
the stereotypy in the numbers of synaptic con- We are obviously at the early stages of assem- C
tacts, thereby increasing the difficulty of bling the components of the Drosophila cellular
obtaining a single, defined connectome. connectome. However, from the current rapid
pace of improvements in sample preparation,
imaging, and software tools for reconstruction,
The Utility of Connectome Information: it seems possible that the cellular connectome
Functional Studies will become available for the entire fly’s CNS.
This will require improvements in staining and
One rationale for compiling a cellular sample preparation, increased reliability and
connectome is to gain insight into the neuronal speed of imaging, and improved automated
implementations of behavior. The existing optic image analysis algorithms. Additionally, infor-
lobe connectomes have already demonstrated mation on the polarity of synaptic transmission,
such utility. They have been used to direct the and the presence of electrical synapses, neither
attention of genetic experiments towards visible from EM, must be incorporated from
a specific neuron(s) (Gao et al. 2008) and to other techniques (Meinertzhagen and Lee,
provide the neuronal circuit substrates for com- 2011). However, even prior to the completion of
putational analyses (Takemura et al. 2013). For the entire connectome, the demonstrated utility
example, the comprehensive nature of the cellu- of the existing pieces of the Drosophila
lar connectome reconstruction in the medulla connectome suggests that each additional recon-
allowed the identification of candidate neurons struction should provide valuable biological
comprising a circuit computing local motion, insight.
a computation that had been intensely studied
for more than 50 years (Borst et al. 2010), yet
with limited prior insight. References
Borst A, Haag J, Reiff DF (2010) Fly motion vision. Annu

In Progress Rev Neurosci 33:49–70
Caron SJ, Ruta V, Abbott L, Axel R (2013) Random
convergence of olfactory inputs in the Drosophila
As of 2013, ongoing Drosophila reconstruction mushroom body. Nature 497:113–117
projects include the larval CNS connectome Chiang A-S et al (2011) Three-dimensional reconstruction
(Cardona et al.), and the connectome for seven of brain-wide wiring networks in Drosophila at single-
cell resolution. Curr Biol 21:1–11
medulla columns, utilizing a new imaging tech- Chou Y-H et al (2010) Diversity and wiring variability of
nique: focused ion beam milling scanning EM olfactory local interneurons in the Drosophila antennal
(FIB-SEM). FIB allows improved z-axis resolu- lobe. Nat Neurosci 13:439–449
tion, enabling more complete tracing of neurites Franceschini N, Kirschfeld K, Minke B (1981) Fluores-
cence of photoreceptor cells observed in vivo. Science
thinner than the section thickness. This results in 213:1264–1267
a more complete connectome, given that only Gao S et al (2008) The neural substrate of spectral prefer-
50 % of the postsynaptic sites could be traced to ence in Drosophila. Neuron 60:328–342
an identified neuron in the ssTEM dataset. Heisenberg M, Wolf R (1984) Vision in Drosophila.
Genetics of microbehaviour. Springer
Further, by including multiple columns, the Hooks B et al (2011) Laminar analysis of excitatory local
resulting connectome should also allow an esti- circuits in vibrissal motor and sensory cortical areas.
mate of the variation between columns, as well as PLoS Biol 9:e1000572
C 798 Connectome, General
Ito K, Shinomiya K, Ito M et al (2014) A systematic (macro [or regional] vs. micro [or cellular]
nomenclature for the insect brain. Neuron (in press) connectome) and the other based on research
Ito M, Masuda N, Shinomiya K, Endo K, Ito K (2013)
Systematic analysis of neural projections reveals focus (data acquisition technology vs. theoretical
clonal composition of the Drosophila brain. Curr Biol frameworks and analysis). The end goal of
23:644–655 connectomics is to understand how the brain
Jenett A et al (2012) A GAL4-driver line resource for works based on its intricate circuitry, i.e., to go
Drosophila neurobiology. Cell Rep 2:991–1001
Large VRML model with brain strutures labeled. Flybrain from complete structure to detailed function. See
on-line: http://www.flybrain.org. Accession number: Seung (2012) and Sporns (2011, 2012) for
AB00122 a general overview of connectomics.
Lefort S, Tomm C, Floyd Sarria J-C, Petersen CC (2009)
The excitatory neuronal network of the C2 barrel col-
umn in mouse primary somatosensory cortex. Neuron
Meinertzhagen IA, Lee C-H (2011) The genetic analysis
of functional connectomics in Drosophila. Adv Genet The main underlying assumption in
80:99–151
Meinertzhagen I, O’neil S (1991) Synaptic organization of connectomics is that the brain’s connective archi-
columnar elements in the lamina of the wild type in tecture determines in large part its function
Drosophila melanogaster. J Comp Neurol 305:232– (Sporns 2012; Seung 2012). This assumption
263 can be put in an evolutionary context, as noted
Murthy M, Fiete I, Laurent G (2008) Testing odor
response stereotypy in the Drosophila mushroom by Swanson (2003): Evolution from simple neu-
body. Neuron 59:1009–1023 ronal networks in primitive animals to complex
Rein K, Zöckler M, Mader MT, Grübel C, Heisenberg M brains of animals like humans happened mostly
(2002) The Drosophila standard brain. Curr Biol in its architecture, not in its components.
12:227–231
Rivera-Alba M et al (2011) Wiring economy and volume
exclusion determine neuronal placement in the Dro- Current Status of Connectomics
sophila brain. Curr Biol 21:2000–2005 Connectomics aims to obtain the connectome
Takemura S-Y et al (2013) A visual motion detection from the whole brain and analyze it in full detail
circuit suggested by Drosophila connectomics. Nature
500:175–181 at subcellular resolution. However, because of
Tuthill JC, Nern A, Holtz SL, Rubin GM, Reiser MB technological limitations, current efforts are
(2013) Contributions of the 12 neuron classes in the focused on either small volumes (on the order of
fly lamina to motion vision. Neuron 79:128–140 100 um) of nm resolution samples or large vol-
umes (whole brain) at hundreds of um resolution.
Connectomes at these two different scales are
called micro (or cellular) connectome and
Connectome, General macro (or regional) connectome, respectively
(see, e.g., DeFelipe 2010).
Yoonsuck Choe Currently, the only full connectome that is
Department of Computer Science and available is that of the nematode Caenorhabiditis
Engineering, Texas A&M University, College elegans. All 302 neurons, their connections, and
Station, TX, USA other somatic cells have been mapped out fully
(White et al. 1986). Various connectivity analysis
(e.g., Watts and Strogatz 1998; Kaiser and
Definition Hilgetag 2006; Sohn et al. 2011) and open-source
simulation efforts (OpenWorm, ▶ http: //
Connectome is the full connection matrix (or the openworm. org) are based on the C. elegans
full wiring diagram) of all neurons in the brain connectome data.
(Sporns et al. 2005). Current efforts in The Drosophila connectome has also been
connectomics (study of the connectome) can be extensively mapped (379 neurons and 8,637
grouped along two axes, one based on scale chemical synapses) in the medulla of the optic
Connectome, General 799 C
lobe (Takemura et al. 2013). However, due to the rabbit retina (Anderson et al. 2011). At a similar
complexity of the Drosophila brain compared to scale, Bock et al. (2011) combined functional
that of the C. elegans (~ 90,000 cells in its brain, imaging (two-photon calcium imaging) and serial
of which ~90 % are neurons), the full connectome transmission EM to image 450 um wide and
has not been mapped. See “▶ Connectome, Dro- 350 um deep volume from the mouse visual cortex
sophila” for details. (this data set is available through the Open
Cellular-level connectome (nm resolution) of Connectome Project ▶ http://openconnecto- C
larger animals seems to be beyond the horizon at meproject.org). In a similar vein, Mishchenko
the moment, although technologies for obtaining et al. (2010) studied the ultrastructure of rat hippo-
such data are actively being developed. For an campal neuropil from serial sectioning transmis-
overview of such methods, see “▶ Physical Sec- sion EM images (45–50 nm thick sections, 2.2 nm
tioning Microscopy,” especially those employing in-plane resolution).
electron microscopy, e.g., serial block-face EM At the um scale, light microscopy techniques
(Denk and Horstmann 2004) and automated tape- are being developed for connectomics, with the
collecting ultramicrotome (Hayworth et al. 2006). aim of obtaining local circuits based on histolog-
These techniques have been used to map small ical stains (e.g., knife-edge scanning microscopy
volumes of neural circuitry (~1003um3) at nm and related techniques; Mayerich et al. 2008;
resolution (on the order of 10 nm), for example, Chung et al. 2011; Li et al. 2010a; Fig. 1) and
the mouse retina (Helmstaedter et al. 2013) or the long-range projections based on tracer injections
Connectome, General, Fig. 1 Circuits in the mouse et al. 2008). The Golgi data set shown is an overlay of
hippocampus. Screenshot from the Knife-Edge Scanning thirty 1 um-thick sections at an interval of 2 um,
Microscope. Brain Atlas http://kesm.org, inverted and representing a tissue thickness of 60 um. The voxel reso-
contrast enhanced (Chung et al. 2011; Mayerich lution is 0.7 um 0.6 um 1.0 um
(Hintiryan et al. 2012; Mitra 2012). The typical spatial scales. It is also notable that connectivity
volume that can be imaged with these methods is matrices derived from tracer injection methods
~1 cm3 at the resolution of ~1 um. See “▶ Phys- are available at this scale, such as that of the
ical Sectioning Microscopy” and Kleinfeld macaque (“▶ Collations of Connectivity Data
et al. (2011) for related approaches such as on the Macaque Brain (CoCoMac)”; Felleman
array tomography, serial two-photon tomogra- and Van Essen 1991) and the cat (de Reus and
phy, and all-optical histology. Optical sectioning van den Heuvel 2013). Simulation of an entire
techniques such as scanned light-sheet micros- thalamocortical network of the human brain
copy (Keller et al. 2008) are also promising, based on dMRI connectivity data (Izhikevich
especially when combined with tissue-clearing and Edelman 2008) and extensive connectivity
methods like CLARITY (Chung and Diesseroth analysis exists at this scale (see Chapter 5, The
2013). Fluorescence probes are routinely used to Connectome at the Macroscale, in Sporns 2012,
label distinct neuronal types or molecular signa- and “▶ Human Connectome Project”).
tures such as presynaptic vs. postsynaptic densi- Across the multiple scales, connectomics
ties for the detection of synaptic connections. research is already producing deep insights into
Most of the above techniques have been applied the organization and function of the brain, e.g.,
to mouse brain imaging (see “▶ Connectome, (1) discovery of new cell types based on mor-
Mouse” for publicly available data). However, phology and connectivity (Helmstaedter
the available data are only partial and in many et al. 2013), (2) organizing principles such as
cases geometric reconstructions are not available; the “rich club” (nodes with high connectivity
thus analysis based on the data is scarce. Some are more heavily connected among themselves;
reconstructions are available for small volumes van den Heuvel and Sporns 2011; de Reus and
of the mouse neuromuscular junction van den Heuvel 2013), (3) emergence of activa-
connectome (~1003 um3 at ~0.2 um resolution), tion patterns and sensitivity of the pattern to small
where the brainbow technique was used local changes (Izhikevich and Edelman 2008),
(Srinivasan et al. 2010). (4) wiring optimization principles (see “▶ Wiring
At the highest spatial scale, neuroimaging Principles, Optimization”), (5) changes in the
methods are used to map the macro (or regional) connectome due to neurodegeneration (Zhou
connectome. These techniques are mostly based et al. 2012), and (6) identification of functionally
on diffusion magnetic resonance imaging (dMRI; distinct neuronal populations based on connec-
Park et al. 2008; Hagmann et al. 2007, 2008; tivity (Sohn et al. 2011).
Johansen-Berg and Rushworth 2009), with the
exception of 3D-polarized light imaging From Raw Data to Geometry, Connectivity,
(3D-PLI; Axer et al. 2011). Major white matter and Databases
tracts can be imaged using these approaches. The The imaging data obtained using the techniques
volume covered by these techniques is often as outlined above are prone to noise and distortions;
large as the human brain (on the order of thus, noise and artifact removal using image
103 cm3), but the imaging resolution is low, processing algorithms is a necessary first step.
about 2 mm for dMRI and about 30 um for Next, denoised images need to be registered so
3D-PLI. Serial sectioning has been employed on that an aligned 3D volume can be constructed
whole human brain at 20 um sectioning thickness (note that some methods do not require such
with in-plane resolution of 10 um 10 um, but a registration step). The last step of assembling
the histological stain used only labeled cell bod- the connectome is to reconstruct the geometry of
ies; thus, connectivity could not be established all neurons and their connections from the image
(Amunts et al. 2013). Although the resolution is volume (see “▶ Reconstruction, Techniques and
low, data from this scale has been analyzed most Validation,” and “▶ Reconstruction, Electron
intensely compared to connectomes at different Microscopy”). This is no trivial task, and
researchers are exploring (1) manual (e.g., Fiala Analysis and Utilization of Connectome Data
2005), (2) automated (Jain et al. 2010; Bebis et al. Once the connectome is reconstructed and con-
2010; Edwards et al. 2013), (3) crowd-sourcing- nectivity established, we need to analyze the data
based (e.g., the EyeWire Project http://eyewire. to gain new insights in brain structure and
org; Helmstaedter et al. 2013), and (4) a hybrid of function.
automated and manual methods (Chklovskii Graph theoretical analysis: With a full con-
et al. 2010; Mishchenko et al. 2010). nectivity matrix, we can use standard graph the- C
Once the full geometric description of all neu- oretic measures such as in-degree, out-degree,
rons is known, the connectivity needs to be clustering index, etc. (Barabasi 2002; Bullmore
established. In electron microscopy (EM) data, and Sporns 2009; Rubinov and Sporns 2010) and
synapses can be inferred based on synaptic densi- look for motifs (Watts and Strogatz 1998). See
ties and synaptic vesicles and by locating gap Kaiser (2011) for a tutorial.
junctions (Xu et al. 2013). Several manual and Basic circuit analysis: Stereotypical patterns
automated methods have been developed for this, of local circuits are a hallmark of brain architec-
e.g., Chklovskii et al. (2010), Xu et al. (2013), and ture. These patterns are called basic circuits, and
Kreshuk et al. (2011). Aside from EM data, using these as a building block, large-scale circuit
exact connectivity cannot be established, unless analysis can be conducted (Shepherd 2004).
molecular markers such as synaptophysin-GFP Connectome data can also enable systematic dis-
are used to explicitly mark the synapses covery of new basic circuit patterns.
(Li et al. 2010b). For example, data based on Estimating dynamic parameters: Certain
light microscopy typically show only a fraction dynamic parameters such as conduction delay
(~1 % for Golgi) of the entire population of can be estimated based on axon length and diam-
neurons. In this case, we need to estimate con- eter. Simply calculating the delay distribution can
nectivity. Methods like those proposed by already provide great insights into brain function.
Kalisman et al. (2003) can be used. One way to For example, Thiel et al. (2003) showed that the
test the validity of such an approximation is as complexity of network dynamics critically
follows. First,conduct a systematic simulation depends on the delay distribution. (Also see
study with a full synthetic circuit to establish a Sporns et al. (2000) on the relationship between
baseline. Next, drop a certain proportion of neuroanatomy and brain dynamics.) Compared to
connections and observe the resulting change connectivity data, collections of delay data are
in behavior. The degree of redundancy in the relatively scarce, with few exceptions being the
connections (both for real and synthetic circuits) works by Lamme and Roelfsema (2000) and
will play an important role here. Nowak and Bullier (1997).
Informatics is also an important component in Bridging the gap between structure and func-
connectomics: how to represent and organize the tion: The connectome is fundamentally a static
image data and reconstructed geometry and con- structure. How can physiological properties be
nectivity. Standards for describing neuronal mor- inferred from just the structure? Toledo-
phology such as “NeuroML” and classification Rodriguez et al. (2004) show a possibly powerful
and taxonomy schemes such as the Foundational solution to this problem: use gene expression data
Model of Connectivity (FMC) are being devel- (also see Markram 2006). They found that gene
oped (Swanson and Bota 2010). Due to the mas- expression and electrophysiological properties
sive amounts of data regardless of the imaging are closely correlated. For approaches like this,
modality or scale, connectome data are generally existing gene expression data (e.g., the Allen
made available through web-based multi- Brain Atlas (Lein et al. 2007) and techniques
scale visualization environments (see Chung that combine molecular and EM imaging serve
et al. 2011; Hintiryan et al. 2012; Mitra 2012; as great resources (e.g., array tomography
Mikula et al. 2007). (Micheva and Smith 2007)) are great resources
for this kind of approach. These approaches can Koene et al. 2009; Koene 2007; Ascoli
also help determine the sign and strength of the et al. 2001; Eberhard et al. 2006). Connectomics
synaptic connections (together with dynamic data can help narrow down on the range of vari-
parameters discussed right above) which are ous parameters for these simulations (for param-
important properties that influence function. For eter constraining procedures, see Druckmann
example, the function of a circuit can instantly et al. 2008).
become nontrivial to analyze when inhibition is Investigate the effect of link fidelity: A great
involved, e.g., in various circuits that incorporate matter of debate in connectomics is whether indi-
disinhibition. vidual connections matter (detailed EM info
Linking with dynamic data: To obtain a more needed) or whether they can be averaged
complete picture of the brain, eventually we will (diffusion MRI is enough). Some results suggest
need to combine structural and functional imag- that dropping even a single spike in the initial
ing of the brain. Near real-time calcium imaging condition can have a global effect on the entire
of whole animals such as the zebra fish is already cortex within 0.5 s (see Izhikevich and Edelman’s
possible (Ahrens et al. 2013), and at the higher (2008) large-scale simulation study of the
spatial scale, neuroimaging methods like magne- thalamocortical system based on diffusion MRI
toencephalography (MEG) can be used (Larson- data). However, considering that the brain in
Prior et al. 2013). Functional and effective a normal operating environment is always
connectivity estimated using functional MRI anchored to the present input stimulus thus con-
(Friston 2011) could also be compared and stantly resetting the initial condition, this may not
contrasted with connectomics data. Some of the be a serious issue. Issues like these can be studied
limitations of functional and effective connectiv- based on circuit data estimated from the
ity estimation can be addressed through connectomics data sets.
optogenetics (Lee et al. 2010). Investigate changes in the connectome in
Inter- and intra-specimen variability estima- brain disorders: An important practical applica-
tion: Simply measuring the morphological vari- tion of connectomics is to understand how brain
ability among the same class of neurons can disorders change or deteriorate the connectome.
provide valuable insights into how redundant or Preliminary work in this direction has already
specialized the functions are (Gerald Edelman, appeared: Zhou et al. (2012) showed how con-
personal communication, 2009). Even when connectivity in health predicts regional vulnerability
nectivity is not known, just examining the den- to disease (note: in this work, fMRI-based func-
dritic trees can give deep insights into neural tional connectivity was studied, not structural
computation (Mel 1994; Polsky et al. 2004). connectivity). See “▶ Neuropathologies and Net-
Brute-force parameter search and simulation: works” for details.
Of course a straightforward yet potentially valu- Novel visualization methods for discovery: The
able approach is to start with computational sim- amount of connectome data that already exists
ulations based on detailed neuronal morphology surpasses our capability to analyze and understand
(cf. the Blue Brain Project, Markram 2006). The them. In most cases, the connectome data are
reconstructed geometry can be used to construct stored as numerical values, and humans are not
multi-compartment models (see, e.g., Dayan and good at dealing with such data directly. To enable
Abbott 2001). Appropriate parameters such as discovery, novel real-time visualization methods
channel conductance, capacitance, etc. need to may be necessary. Visualization should not be
be figured out. Tools like NEURON, GENESIS, confined to rendering the geometry – various
neuroConstruct, NeuGEN, Netmorph, etc. can be forms of filtering and statistical summary visuali-
used for multi-compartment simulation and zation based on local and global geometric prop-
parameterized synthetic circuit generation/simu- erties (such as orientation, anisotropy, etc.) need to
lation/analysis (Hines and Carnevale 1997; be developed. See Margulies et al. (2013) on visu-
Bower and Beeman 1998; Gleeson et al. 2007; alizing dMRI human connectome data and
Mayerich and Hart (2007) on visualizing serial EM (note that whether brain simulations can cause
data. Visual analytics for abstract information mind is a whole separate issue; see, e.g., Searle
extracted from the connectome will also be neces- 1980).
sary (see Meyer et al. 2010 for a discussion on Some may argue that simulating the brain is not
visual analytics in general). enough. That is, the entire body must be simulated.
This kind of argument can be based on philosoph-
Challenges and Open Questions ical grounds, e.g., the criticism of brain-body dual- C
A hotly debated open question in connectomics is ism by Bennett and Hacker (2003) (also see the
regarding the scale: What level of detail is counterpoint raised by Churchland 2005). How-
enough to completely replicate the architecture ever, the argument can also be based on practical
and function of the brain (Koene 2012)? This grounds. All sensorimotor organs and the exten-
question arises partly due to the incompleteness sive peripheral nervous system form a major
of various data acquisition technology chunk of the animal nervous system, and these
operating at scales ranging across several orders nervous networks define the environment-to-
of magnitude, from EM to dMRI. There is no body interface. Without understanding this inter-
definite answer to this problem. In the next face, it would be difficult to understand the func-
decade or so, for practical reasons, we will have tion of the brain.
to live with a patchwork of connectomics data Finally, an important question is about the role
from various scales, so we need to invent novel of theory in the analysis of connectomics data.
approaches to register and integrate connectome Connectomics research is data-driven by its nature,
data from different imaging modalities at differ- so one of the hopes is that data can lead us to
ent scales. principles of brain organization and function.
One of the ultimate goals of connectomics is to However, due to the massiveness of the data, with-
construct a high-fidelity computer simulation of out theoretical guidance, it would be difficult to
the brain (Whole Brain Emulation; Koene 2012). make progress. This is a classical chicken-or-egg
We can consider an interesting open question: Is problem: We want to study the connectome to
it possible that we can have such a high-fidelity get to the theory, but without theory we get lost
computer simulation of the brain but cannot in the data. There is some skepticism about
understand what is going on in the simulation? connectomics research due to problems like this,
This is a real possibility. Based on this, some but as argued by Lichtman and Sanes (2008), the
might even argue that modeling and simulation human genome project had similar concerns yet
do not give you any additional understanding the data ultimately led to spectacular scientific
about the phenomena that you are studying. How- discoveries, and similar results are expected from
ever, we need to consider that simulations give us connectomics. Since very specific theories can
two very important tools: (1) full access to often be misleading unless grounded on detailed
the system state (readout) and (2) full control data, it would be better to start with broader per-
over the system state (intervention). With spectives, such as the importance of action and the
a completely replicated simulation, we can con- sensorimotor loop (Llinás et al. 1994; Fuster 1997;
duct a full battery of experiments, including Choe et al. 2007), the role of time and prediction
a unique opportunity to selectively damage or (Choe et al. 2012), the evolutionary origin of brain
turn off parts of the system, which is an important (Humphrey 1992; Kwon and Choe 2009; Chung
requirement for inferring causality in a complex and Choe 2011), and the developmental processes
system (see Pearl 2001 for the importance of that shape the brain (Weng et al. 2001;
intervention for such an inference). Furthermore, Miikkulainen et al. 2005).
high-fidelity simulations can lead to a rather
exotic possibility of allowing the simulation
Acknowledgments Part of the contents of this chapter
itself to experiment on itself to objectively study (section “Analysis and Utilization of Connectome Data”)
subjective phenomena such as consciousness first appeared in Choe et al. (2011).
Cross-References Chklovskii DB, Vitaladevuni S, Scheffer LK (2010) Semi-

automated reconstruction of neural circuits using elec-
tron microscopy. Curr Opin Neurobiol 20(5):667–675.
▶ Connectome, Drosophila Neuronal and glial cell biology – new technologies
▶ Connectome, Mouse Choe Y, Yang H-F, Eng DC-Y (2007) Autonomous
▶ Determining Network Structure from Data: learning of the semantics of internal sensory states
Nonlinear Modeling Methods based on motor exploration. Int J Hum Robot
4:211–243
▶ Human Connectome Project Choe Y, Mayerich D, Kwon J, Miller DE, Chung JR,
▶ Multiscale Brain Connectivity Sung C, Keyser J, Abbott LC (2011) Knife-edge
▶ Network Theory in Neuroscience scanning microscopy for connectomics research. In:
▶ Neuropathologies and Networks Proceedings of the international joint conference on
neural networks. IEEE Press, Piscataway,
▶ Wiring Principles, Optimization pp 2258–2265
Choe Y, Kwon J, Chung JR (2012) Time, consciousness,
References and mind uploading. Int J Mach Conscious 4:257–274
Chung JR, Choe Y (2011) Emergence of memory in reac-
Ahrens MB, Orger MB, Robson DN, Li JM, Keller PJ tive agents equipped with environmental markers.
(2013) Whole-brain functional imaging at cellular res- IEEE Trans Auton Ment Develop 3:257–271
olution using light-sheet microscopy. Nat Methods Chung K, Diesseroth K (2013) CLARITY for mapping the
10:413–420 nervous system. Nat Methods 10:508–513
Amunts K, Lepage C, Borgeat L, Mohlberg H, Chung JR, Sung C, Mayerich D, Kwon J, Miller DE,
Dickscheid T, Rousseau M-E, Bludau S, Bazin P-L, Huffman T, Abbott LC, Keyser J, Choe Y (2011)
Lewis LB, Oros-Peusquens A-M, Shah NJ, Lippert T, Multiscale exploration of mouse brain microstructures
Zilles K, Evans AC (2013) Bigbrain: an ultrahigh- using the knife-edge scanning microscope brain atlas.
resolution 3d human brain model. Science Front Neuroinform 5:29
340:1472–1475 Churchland PM (2005) Cleansing science. Inquiry
Anderson JR, Jones BW, Watt CB, Shaw MV, Yang J-H, 5:464–477
DeMill D, Lauritzen JS, Lin Y, Rapp KD, Dayan P, Abbott LF (2001) Theoretical neuroscience.
Mastronarde D, Koshevoy P, Grimm B, Tasdizen T, MIT Press, Cambridge, MA
Whitaker R, Marc RE (2011) Exploring the retinal de Reus MA, van den Heuvel MP (2013) Rich club orga-
connectome. Mol Vis 17:355–379 nization and intermodule communication in the cat
Ascoli G, Krichmar J, Scorcioni R, Nasuto S, Senft connectome. J Neurosci 33:12929–12939
S (2001) Computer generation and quantitative mor- DeFelipe J (2010) From the connectome to the synaptome:
phometric analysis of virtual neurons. Anat Embryol an epic love story. Science 330:1198–1201
204:283–301 Denk W, Horstmann H (2004) Serial block-face scanning
Axer M, Amunts K, Grassel D, Palm C, Dammers J, Axer H, electron microscopy to reconstruct three-dimensional
Pietrzyk U, Zilles K (2011) A novel approach to the tissue nanostructure. PLoS Biol 19:e329
human connectome: ultra-high resolution mapping of Druckmann S, Berger TK, Hill S, Schurmann F,
fiber tracts in the brain. Neuroimage 54:1091–1101 Markram H, Segev I (2008) Evaluating automated
Barabasi A-L (2002) Linked: the new science of networks. parameter constraining procedures of neuron models
Perseus Publishing, Cambridge, MA by experimental and surrogate data. Biol Cybern
Bebis G, Boyle R, Parvin B, Koracin D, Chung R (Eds.) 99:371–379
(2010) Advances in Visual Computing: 6th Interna- Eberhard JP, Wanner A, Wittum G (2006) NeuGen: a tool
tional Symposium, ISVC 2010, Las Vegas, NV, USA, for the generation of realistic morphology of cortical
Nov 29-Dec 1, 2010, Proceedings (Vol. 1). Springer neurons and neural networks. Neurocomputing
Bennett MR, Hacker PMS (2003) Philosophical founda- 70:327–342
tions of neuroscience. Blackwell, Malden, MA Edwards J, Daniel E, Kinney J, Bartol T, Sejnowski T,
Bock DD, Kerlin AM, Andermann ML, Hood G, Johnston D, Harris K, Bajaj C (2013) Volrovern:
Wetzel AW, Yurgenson S, Soucy ER, Kim HS, enhancing surface and volumetric reconstruction for
Reid RC (2011) Network anatomy and in vivo realistic dynamical simulation of cellular and subcel-
physiology of visual cortical neurons. Nature lular function. Neuroinformatics 11:1–13
471:177–182 Felleman DJ, Van Essen DC (1991) Distributed hierarchi-
Bower JM, Beeman D (1998) The book of GENESIS: cal processing in primate cerebral cortex. Cereb Cor-
exploring realistic neural models with the GEneral tex 1:1–47
NEural SImulation System. Telos, Santa Clara Fiala JC (2005) Reconstruct: a free editor for serial section
Bullmore E, Sporns O (2009) Complex brain networks: microscopy. J Microsc 218:52–61
graph theoretical analysis of structural and functional Friston K (2011) Functional and effective connectivity:
systems. Nat Rev Neurosci 10:186–198 a review. Brain Connect 1:13–36
Fuster JM (1997) The prefrontal cortex: anatomy, physi- automated histology in the pursuit of connectomes.
ology, and the frontal lobe, 3rd edn. Lippencott-Raven, J Neurosci 31:16125–16138
Philadelphia Koene RA (2007) Large scale high resolution network
Gleeson P, Steuber V, Silver R (2007) NeuroConstruct: generation: producing known validation sets for serial
a tool for modeling networks of neurons in 3D space. reconstruction methods that use histological images of
Neuron 54:219–235 neural tissue. In: International conference on complex
Hagmann P, Kurant M, Gigandet X, Thiran P, Wedeen VJ, systems. [Presentation]
Meuli R, Thiran J-P (2007) Mapping human whole- Koene R (2012) Fundamentals of whole brain emulation:
brain structural networks with diffusion MRI. PLoS state, transition, and update representations. Int J Mach
C
One 2:e597 Conscious 4:5–21
Hagmann P, Cammoun L, Gigandet X, Meuli R, Koene RA, Tijms B, van Hees P, Postma F, de Ridder A,
Honey CJ, Wedeen VJ, Sporns O (2008) Mapping Ramakers GJA, van Pelt J, van Ooyen A (2009)
the structural core of human cerebral cortex. PLoS NETMORPH: a framework for the stochastic
Biol 6:e159 generation of large scale neuronal networks with real-
Hayworth KJ, Kasthuri N, Schalek R, Lichtman JW istic neuron morphologies. Neuroinformatics
(2006) Automating the collection of ultrathin sections 7:1539–2791
for large volume TEM reconstructions. Microsc Kreshuk A, Straehle C, Sommer C, Koethe U, Cantoni M,
Microanal 12(Suppl S02):86–87 Knott G, Hamprecht FA (2011) Automated detection
Helmstaedter M, Briggman KL, Turaga SC, Jain V, Seung and segmentation of synaptic contacts in nearly isotro-
HS, Denk W (2013) Connectomic reconstruction of pic serial electron microscopy images. PLoS One 6:
the inner plexiform layer in the mouse retina. Nature e24899
500:168–174 Kwon J, Choe Y (2009) Facilitating neural dynamics for
Hines ML, Carnevale NT (1997) The NEURON simula- delay compensation: a road to predictive neural
tion environment. Neural Comput 9:1179–1209 dynamics? Neural Netw 22:267–276
Hintiryan H, Gou L, Zingg B, Yamashita S, Lyden HM, Lamme VAF, Roelfsema PR (2000) The distinct modes of
Song MY, Grewal AK, Zhang X, Toga AW, Dong vision offered by feedforward and recurrent
H-W (2012) Comprehensive connectivity of the processing. Trends Neurosci 23:571–579
mouse main olfactory bulb: analysis and online digital Larson-Prior L, Oostenveld R, Penna SD, Michalareas G,
atlas. Front Neuroanat 6:30 Prior F, Babajani-Feremi A, Schoffelen J-M,
Humphrey N (1992) A history of the mind. HarperCollins, Marzetti L, de Pasquale F, Pompeo FD, Stout J,
New York Woolrich M, Luo Q, Bucholz R, Fries P, Pizzella V,
Izhikevich EM, Edelman GM (2008) Large-scale model Romani G, Corbetta M, Snyder A (2013) Adding
of mammalian thalamocortical systems. Proc Natl dynamics to the human connectome project with ME-
Acad Sci U S A 105:3593–3598 G. Neuroimage 80:190–201
Jain V, Seung HS, Turaga SC (2010) Machines that learn Lee JH, Durand R, Gradinaru V, Zhang F, Goshen I, Kim
to segment images: a crucial technology for D-S, Fenno LE, Ramakrishnan C, Deisseroth K (2010)
connectomics. Curr Opin Neurobiol 20:653–666 Global and local fMRI signals driven by neurons
Johansen-Berg H, Rushworth MF (2009) Using diffusion defined optogenetically by type and wiring. Nature
imaging to study human connectional anatomy. Annu 465:788–792
Rev Neurosci 32:75–94 Lein ES et al (2007) Genome-wide atlas of gene expres-
Kaiser M (2011) A tutorial in connectome analysis: topo- sion in the adult mouse brain. Nature 445:168–176
logical and spatial features of brain networks. Li A, Gong H, Zhang B, Wang Q, Wan C, Wu J, Liu Q,
Neuroimage 57:892–907 Zeng S, Luo Q (2010a) Microoptical sectioning tomog-
Kaiser M, Hilgetag CC (2006) Nonoptimal component raphy to obtain a high-resolution atlas of the mouse
placement, but short processing paths, due to long- brain. Science 330:1404–1408
distance projections in neural systems. PLoS Comput Li L, Tasic B, Micheva KD, Ivanov VM, Spletter ML,
Biol 2:805–815 Smith SJ, Luo L (2010b) Visualizing the distribution
Kalisman N, Silberberg G, Markram H (2003) Deriving of synapses from individual neurons in the mouse
physical connectivity from neuronal morphology. Biol brain. PLoS One 5:e11503
Cybern 88:210–218 Lichtman JW, Sanes JR (2008) Ome sweet ome: what can
Keller PJ, Schmidt AD, Wittbrodt J, Stelzer EHK the genome tell us about the connectome? Curr Opin
(2008) Reconstruction of zebrafish early embryonic Neurobiol 18(3):346–353
development by scanned light sheet microscopy. Sci- Llinás R, Ribary E, Joliot M, Wang G (1994) Content and
ence 322:1065–1069 context in temporal thalamocortical binding. In:
Kleinfeld D, Bharioke A, Blinder P, Bock DD, Briggman Buzsaki G (ed) Temporal coding in the brain. Springer,
KL, Chklovskii DB, Denk W, Helm-Staedter M, Berlin
Kaufhold JP, Lee W-CA, Meyer HS, Micheva KD, Margulies DS, Bttger J, Watanabe A, Gorgolewski KJ
Oberlaender M, Prohaska S, Reid RC, Smith SJ, (2013) Visualizing the human connectome.
Takemura S, Tsai PS, Sakmann B (2011) Large-scale Neuroimage 80:445–461
Markram H (2006) The blue brain project. Nat Rev organization of the caenorhabditis elegans
Neurosci 7:153–160 connectome. PLoS Comput Biol 7:e1001139
Mayerich D, Hart J (2007) Volume visualization in serial Sporns O (2011) Networks of the brain. MIT Press, Cam-
electron microscopy using local variance. In: IEEE bridge, MA
symposium on biological data visualization (BioVis), Sporns O (2012) Discovering the human connectome.
IEEE, Piscataway, NJ pp 1–6 MIT Press, Cambridge, MA
Mayerich D, Abbott LC, McCormick BH (2008) Knife-edge Sporns O, Tononi G, Edelman GM (2000) Connectivity
scanning microscopy for imaging and reconstruction of and complexity: the relationship between neuroanat-
three-dimensional anatomical structures of the mouse omy and brain dynamics. Neural Netw 13:909–922
brain. J Microsc 231:134–143 Sporns O, Tononi G, Kotter R (2005) The human
Mel BW (1994) Information processing in dendritic trees. connectome: a structural description of the human
Neural Comput 6:1031–1085 brain. PLoS Comput Biol 1:e42
Meyer J, Thomas J, Diehl S, Fisher BD, Keim DA, Srinivasan R, Li Q, Zhou X, Lu J, Lichtman J, Wong ST
Laidlaw D, Miksch S, Mueller K, Ribarsky W, (2010) Reconstruction of the neuromuscular junction
Preim B, Ynnerman A (2010) From visualization to connectome. Bioinformatics 26:i64–i70
visually enabled reasoning. Sci Vis: Adv Concepts Swanson LW (2003) Brain architecture: understanding the
1:227–245 basic plan. Oxford University Press, Oxford
Micheva K, Smith SJ (2007) Array tomography: a new Swanson LW, Bota M (2010) Foundational model of
tool for imaging the molecular architecture and ultrastructural connectivity in the nervous system with
structure of neural circuits. Neuron 55:25–36 a schema for wiring diagrams, connectome, and basic
Miikkulainen R, Bednar JA, Choe Y, Sirosh J (2005) Com- plan architecture. Proc Natl Acad Sci
putational maps in the visual cortex. Springer, Berlin, 107:20610–20617
p 538. URL: http://www.computationalmaps.org Takemura S-Y, Bharioke A, Lu Z, Nern A,
Mikula S, Trotts I, Stone JM, Jones EG (2007) Internet- Vitaladevuni S, Rivlin PK, Katz WT, Olbris DJ,
enabled high-resolution brain mapping and virtual Plaza SM, Winston P, Zhao T, Horne JA, Fetter RD,
microscopy. Neuroimage 35:9–15 Takemura S, Blazek K, Chang L, Ogundeyi O,
Mishchenko Y, Hu T, Spacek J, Mendenhall J, Harris KM, Saunders MA, Shapiro V, Sigmund C, Rubin GM,
Chklovskii DB (2010) Ultrastructural analysis of hip- Scheffer LK, Meinertzhagen IA, Chklovskii DB
pocampal neuropil from the connectomics perspective. (2013) A visual motion detection circuit suggested by
Neuron 67:1009–1020 drosophila connectomics. Nature 500:175–181
Mitra PP (2012) Technical white paper: mouse brain Thiel A, Schwegler H, Eurich CW (2003) Complex dynam-
architecture project. Technical report, Cold Spring ics is abolished in delayed recurrent systems with dis-
Harbor Laboratory. http://brainarchitecture.org tributed feedback times. Complexity 8:102–108
Nowak LG, Bullier J (1997) The timing of information Toledo-Rodriguez M, Blumenfeld B, Wu C, Luo J,
transfer in the visual system. In: Rockland KS, Kass Attali B, Goodman P, Markram H (2004) Correlation
JH, Peters A (eds) Cerebral cortex, vol 12. Plenum maps allow neuronal electrical properties to be
Press, New York, pp 205–241 predicted from single-cell gene expression profiles in
Park H-J, Kim JJ, Lee S-K, Seok JH, Chun J, Kim DI, Lee rat neocortex. Cereb Cortex 14:1310–1327
JD (2008) Corpus callosal connection mapping using van den Heuvel MP, Sporns O (2011) Rich-club organi-
cortical gray matter parcellation and dt-mri. Hum zation of the human connectome. J Neurosci
Brain Mapp 29:503–516 31:15775–15786
Pearl J (2001) Reasoning with cause and effect. AI Mag Watts DJ, Strogatz SH (1998) Collective dynamics of
23:95–111 small-world networks. Nature 393:440–442
Polsky A, Mel BW, Schiller J (2004) Computational sub- Weng J, McClelland JL, Pentland A, Sporns O,
units in thin dendrites of pyramidal cells. Nat Neurosci Stockman I, Sur M, Thelen E (2001) Autonomous
7:821–827 mental development by robots and animals. Science
Rubinov M, Sporns O (2010) Complex network measures 291(5504):599–600
of brain connectivity: uses and interpretations. White JG, Southgate E, Thomson JN, Brenner S (1986)
Neuroimage 52:1059–1069 The structure of the nervous system of the nematode
Searle JR (1980) Minds, brains and programs. Behav caenorhabditis elegans. Philos Trans R Soc Lond
Brain Sci 3:417–424 B 314:1–340
Seung HS (2012) Connectome: how the brain’s wiring Xu SWEM, Jarrell TA, Wang Y, Cook SJ, Hall DH
makes us who we are. Houghton Mifflin Harcourt, (2013) Computer assisted assembly of connectomes
Boston, MA from electron micrographs: application to
Shepherd GM (2004) The synaptic organization of the caenorhabditis elegans. PLoS One 8:e54050
brain, 5th edn. Oxford University Press, Oxford, Zhou J, Gennatas ED, Kramer JH, Miller BL, Seeley WW
UK/New York (2012) Predicting regional neurodegeneration from the
Sohn Y, Choi M-K, Ahn Y-Y, Lee J, Jeong J (2011) healthy brain functional connectome. Neuron,
Topological cluster analysis reveals the systemic 73:1216–1227
Connectome, Mouse 807 C
array of transgenic, knockin, and knockout vari-
Connectome, Human ants and molecular probes to infer function from
the connectome structure (Chinwalla et al. 2002).
▶ Human Connectome Project
Acquisition of the Mouse Connectome
C
Rapid development in imaging technology in the
Connectome, Mouse past decade is taking us ever closer to the com-
pletion of the mouse connectome. See “▶ Physi-
Yoonsuck Choe1, Jaerock Kwon2, cal Sectioning Microscopy” for a detailed review
David Mayerich3 and Louise C. Abbott4 of such imaging technologies.
1
Engineering, Texas A&M University, College
Station, TX, USA Public Data Sources
2
Department of Electrical and Computer
Engineering, Kettering University, Flint, Here, we will review existing public data sources
MI, USA relevant to mouse connectome research. As men-
3
Beckman Institute for Advanced Science and tioned above, to date, a complete mouse
Technology, University of Illinois at Urbana- connectome does not exist. The existing data
Champaign, Urbana, IL, USA collections focus on different aspects of the
4
Department of Veterinary Integrative mouse connectome and are in many ways com-
Biosciences, Texas A&M University, College plementary to each other, especially due to the
Station, TX, USA different scales and the size of the tissue volume
represented in each data: cellular (or micro)
connectome (sections “Knife-Edge Scanning
Definition Microscope (KESM) Brain Atlas” through
“Mouse Brain Architecture Project” below) ver-
The mouse connectome is a collection of all sus regional (or macro) connectome (section
neuronal connections in the mouse brain, i.e., “Electron Microscopy Data” below).
a complete wiring diagram of the mouse brain.
Knife-Edge Scanning Microscope (KESM)
Brain Atlas
Detailed Description The Knife-Edge Scanning Microscope (KESM)
is a physical sectioning microscope that enables
The concept of connectome (Sporns et al. 2005; submicrometer-resolution imaging of whole
Seung 2012; Sporns 2012) has rapidly grown into small animal brains (Mayerich et al. 2008). The
a central theme in neuroscience and specifically KESM Brain Atlas is an online atlas (http://kesm.
in computational neuroscience and computa- org) at Texas A&M University that serves whole
tional neuroanatomy. So far, as of 2013, the mouse (C57BL/6) brain data acquired using the
only existing connectome is that of the nematode KESM (Chung et al. 2011). Currently, three data
C. elegans (White et al. 1986). Once obtained, the sets are available, two from mouse brains stained
connectome of the mouse is expected to have with Golgi (neuronal morphology) and one with
significant scientific and clinical impact, due to India ink (vasculature). The two Golgi data sets
(1) strong architectural and genetic similarities provide rich neuronal morphology data spanning
between the mouse and humans, (2) the extensive the entire mouse brain with a voxel resolution of
body of experimental studies involving the 0.6 0.7 1.0 mm. Due to the sparse nature of
organism, and (3) the availability of a large the Golgi stain, exact synaptic connections and
C 808 Connectome, Mouse
long-range projections are not visible in these registration. Injection sites were systematically
data sets. However, the overall architectural chosen from a regular grid rather than anatomi-
plan of the mouse brain can be studied, based on cally identified regions. The prepared brains were
the diverse patterns of dendritic arbors across sectioned coronally to 50um thick sections and
different regions of the mouse brain and major imaged using a virtual slide scanner. All the data
projection pathways such as the various commis- can be explored online using an interactive visu-
sures and the radiating projections within the alization tool called iConnectome. Detailed spec-
striatum. imen preparation and data acquisition methods
and case studies focusing on the pathways in the
Allen Mouse Brain Connectivity Atlas mouse olfactory bulb are available in Hintiryan
The Allen Mouse Brain Connectivity Atlas et al. (2012). Latest work from the same project
(http://connectivity.brain-map.org) provides axo- include a mouse neocortical connectivity atlas
nal projection data from ~300 regions in the based on manual reconstruction (Zingg et al.
mouse brain, associated with C57BL/6 J mice 2014).
and ~100 Cre driver lines backcrossed to the
C57BL/6J mice (Allen Institute for Brain Science Mouse Brain Architecture Project
2012). For this atlas, conventional and viral The Mouse Brain Architecture Project (MPA,
tracers were used to label the projections, and http://brainarchitecture.org) at the Cold Spring
serial two-photon tomography (Ragan Harbor Laboratory also hosts long-range projec-
et al. 2012) was used to section and image the tion data similar to the Allen Mouse Connectivity
prepared whole mouse brains. The resolution of Atlas and the Mouse Connectome Project. The
the data sets is 0.3 0.3 mm on the imaging MPA includes anterograde and retrograde tracer
plane, with a z-stack interval of 100 um. The data from 235 regions in the mouse brain,
resulting image stacks are registered to the refer- with the majority of the injection sites in the
ence atlas of the Allen Mouse Brain Atlas and the cerebrum (218). The basic methodology is simi-
Mouse Brain in Stereotaxic Coordinates (Paxinos lar to that of the other long-range projection
and Franklin 2001). Quantitative data from the atlases: inject tracer in C57BL/6 mouse brain,
projections along with detailed annotations of the slice (cryosection), whole-slide digital imaging
injection site are entered into an online (alternating slices processed with conventional
neuroinformatics platform for public use. The histology and tracer detection, respectively), and
platform also includes search and visualization make the processed data available online (Mitra
tools for an extensive analysis of the axonal pro- 2012). However, one main feature distinguishes
jection patterns in the mouse brain. the MBA project from other atlases: the choice of
injection coordinates. In MBA, the injection sites
The Mouse Connectome Project are determined by an algorithm to sample the
The Mouse Connectome Project (http:// brain in “the most homogeneous way compatible
mouseconnectome.org) at the University of with separation of brain regions” (Grange and
California, Los Angeles, hosts long-range projec- Mitra 2011). Injection sites in the cerebral cortex
tion data similar to the Allen Mouse Brain Con- are selected based on a separate algorithm
nectivity Atlas. The collection hosts 245 injection adapted to the geometry of the cortex.
cases (125 viewable at the moment, with
a projected total case of 400 [as of July 2013]), Electron Microscopy Data
with each case containing two injection sites Electron microscopy (EM) data from selected
using anterograde and retrograde tracers. All the regions of the mouse brain and nervous system
brains used for data acquisition were that of the are also available to the public.
C57BL/6J mice and were counterstained with The Open Connectome Project (http://www.
a fluorescent Nissl stain for cytoarchitectural openconnectomeproject.org/) hosts several EM
Connectome, Mouse 809 C
image stacks, which include data from the References
mouse visual cortex (Bock et al. 2011). The
data set was acquired using a custom Transmis- Allen Institute for Brain Science (2012) ALLEN mouse
brain connectivity atlas technical white paper: over-
sion Electron Microscope (TEM), spanning
view. Technical report October 2012 v.4. Allen Institute
a volume ~450 350 52 mm, at a resolution for Brain Science. http://connectivity.brain-map.org
of <50 nm in depth and <5 nm in the Bock DD, Kerlin AM, Andermann ML, Hood G, Wetzel
imaging plane. AW, Yurgenson S, Soucy ER, Kim HS, Reid RC C
(2011) Network anatomy and in vivo physiology of
The Whole-Brain Ultrastructural Mapping
visual cortical neurons. Nature 471:177–182
Project (http://connectomes.org) hosts EM data Chinwalla AT, Cook LL, Delehaunty KD, Fewell GA,
that spans an entire mouse coronal section at Fulton LA, Fulton RS, Mauceli E (2002) Initial
80 nm voxel resolution, acquired using block- sequencing and comparative analysis of the mouse
genome. Nature, 420(6915), 520–562
face electron microscopy. The data was made
Chung JR, Sung C, Mayerich D, Kwon J, Miller DE,
possible by a whole-brain EM staining method Huffman T, Abbott LC, Keyser J, Choe Y (2011)
reported in Mikula et al. (2012). Multiscale exploration of mouse brain microstructures
using the knife-edge scanning microscope brain atlas.
Front Neuroinform 5:29
Grange P, Mitra PP (2011) Algorithmic choice of coordi-
Summary and Future Prospects nates for injections into the brain: encoding
a neuroanatomical atlas on a grid. arXiv:1104.2616v1
In this article, we reviewed mouse connectome [q-Bio.QM]
Hintiryan H, Gou L, Zingg B, Yamashita S, Lyden HM,
data that are publicly available online. As we
Song MY, Grewal AK, Zhang X, Toga AW, Dong
have seen, a complete connectome of the mouse H-W (2012) Comprehensive connectivity of the
is not yet available. However, data from multiple mouse main olfactory bulb: analysis and online digital
scales and different imaging modalities allow us atlas. Front Neuroanat 6:30. doi:10.3389/
fnana.2012.00030
to piece together an initial draft of the mouse
Mayerich D, Abbott LC, McCormick BH (2008) Knife-
connectome. Such a draft, although rough, can edge scanning microscopy for imaging and reconstruc-
provide a valuable roadmap for the acquisition of tion of three-dimensional anatomical structures of the
the full mouse connectome. mouse brain. J Microsc 231:134–143
Mikula S, Binding J, Denk W (2012) Staining and embed-
Looking into the future, we can consider the
ding the whole mouse brain for electron microscopy.
acquisition of the raw connectome data as only Nat Methods 9:1198–1201
a first step toward the understanding of mouse Mitra PP (2012) Technical white paper: mouse brain
brain function. The immediate issue would be the architecture project. Technical report, Cold Spring
Harbor Laboratory. http://brainarchitecture.org/
geometric reconstruction of the neuronal mor-
mouse/documentation/project-white-paper
phologies and projections. Next, synaptic con- Paxinos G, Franklin KBJ (2001) The mouse brain in
nections need to be established. Based on this, stereotaxic coordinates, 2nd edn. Academic, San
we can have a full wiring diagram. Inferring Diego
Ragan T, Kadiri LR, Venkataraju KU, Bahlmann K,
function from this wiring diagram would require
Sutin J, Taranda J, Arganda-Carreras I, Kim Y,
a coordinated effort to integrate a vast amount Seung HS, Osten P (2012) Serial two-photon tomog-
of experimental research on mouse brain raphy for automated ex vivo mouse brain imaging. Nat
function – physiological, molecular, genetic, Methods 9:255–258
Seung HS (2012) Connectome: how the brain’s wiring
behavioral, and ecological – into the mouse
makes us who we are. Houghton Mifflin Harcourt,
connectome. Boston
Sporns O (2012) Discovering the human connectome.
MIT Press, Cambridge, MA
Sporns O, Tononi G, Kötter R (2005) The human
Cross-References connectome: a structural description of the human
brain. PLoS Comput Biol 1:e42
▶ Connectome, General White JG, Southgate E, Thomson JN, Brenner S (1986)
▶ Physical Sectioning Microscopy The structure of the nervous system of the nematode
C 810 Connectomics
Caenorhabditis elegans. Philos Trans R Soc Lond the order of a few ms up to a few tens of ms
B 314:1–340 (Eggermont 1985). A major example of context
Zingg B, Hintiryan H, Gou L, Song MY, Bay M,
Bienkowski MS, Dong HW (2014) Neural Networks dependence consists of the suppression of the
of the Mouse Neocortex. Cell, 156(5), 1096–1111 responses to a stimulus that is preceded shortly
beforehand by another similar stimulus (often
called “forward masking”). Forward masking at
the level of the mechanical vibrations of the
Connectomics cochlea may account for most properties of
psychoacoustical forward masking (Plack and
▶ Connectivity Analysis in Normal and Patho- Oxenham 1998). At the level of the inferior
logical Brains colliculus, neural forward masking has been
used to account for the precedence effect, in
which the first arriving waveform dominates spa-
tial perception (Litovsky et al. 1999). In auditory
cortex, forward masking caused by a short stim-
Consumer Behavior ulus is strong and may last 100 s of ms after the
offset of the sound, far beyond psychoacoustical
▶ Choice Behavior forward masking (Wehr and Zador 2005;
Nakamoto et al. 2006).
A major type of context dependence studied
in the vertebrate auditory system is called
Context-Dependent Processing in “stimulus-specific adaptation” (SSA). SSA is
Auditory Cortex the reduction in the response to a repeated
sound which does not generalize, or generalizes
Israel Nelken only partially, to other, even very similar sounds
Edmond and Lily Safra Center for Brain (Ulanovsky et al. 2003; Nelken and Ulanovsky
Sciences, The Hebrew University of Jerusalem, 2007). SSA is usually studied using oddball
Jerusalem, Israel sequences composed of tone pips of two frequen-
cies, one of which is common (usually appearing
in 80–95 % of all tone presentations) and the
Definition other rare. By presenting two oddball sequences,
with the role of the two frequencies switched, it is
Context-dependent processing refers to the possible to compare the responses to the same
experimental finding that a neuron may respond tone frequency when common and when rare.
to the same stimulus in different ways, depending As a rule, the response to the same stimulus
on the recent history of the stimulation. Context when rare is larger than when it is common
dependence differs from plasticity in that the (Nelken and Ulanovsky 2007). SSA is often stud-
changes in response properties are fast and ied with tone pips of <100 ms duration, presented
reversible. at rates of about 2–3 Hz, and with frequency
separation ((f2-f1)/f1) of ~30 %. However, SSA
in oddball sequences can be recorded for short
Detailed Description tone pips even when interstimulus interval is as
long as 2 s or with frequency separation of below
Responses in the auditory system are context 10 %, far below the tuning width of the recorded
dependent at multiple time scales. Already at neurons (Malmierca et al. 2009; Antunes
the level of the auditory nerve, firing rate adapta- et al. 2010; Taaseh et al. 2011). SSA with
tion constitutes a simple form of context depen- these properties has been shown in all major
dence, although the times involved are mostly on mammalian models of the auditory system, from
Context-Dependent Processing in Auditory Cortex 811 C
rodents through carnivores to primates, and in auditory nerve up to many tens of seconds in the
nonmammalian vertebrates (barn owls) as well auditory cortex (Yaron et al. 2012). At the long
(Gutfreund 2012). end, it may be indistinguishable from fast plas-
The role of SSA in sound processing is not ticity, such as that documented in ferret auditory
fully understood. SSA may confer on cortical cortex (Fritz et al. 2003). In these experiments,
neurons some amount of deviance sensitivity the responses to a tone frequency have been rap-
(Taaseh et al. 2011), a property that is found in idly and selectively enhanced once that fre- C
human event-related potentials that are associ- quency became behaviorally relevant. These
ated with A1 activation (Grimm and Escera changes could occur within a minute or so,
2012). In humans, deviance selectivity has been about the same time scale studied by Yaron
studied most intensively in relationship to et al. (2012).
mismatch negativity (MMN), a brain potential In summary, context sensitivity in the audi-
that peaks >150 ms after sound onset, much tory system is related to a large number of sep-
later than the earliest cortical responses that arate processes that may influence sound coding
show SSA. Together with other evidences quite dramatically as early as in the auditory
(Farley et al. 2010), these findings suggest that midbrain. While some of them are well studied,
SSA is not the direct neuronal correlate of MMN, their overall effects on sound coding and the
although it may lie upstream of the generation of interactions between them are not, and this con-
MMN and confer on MMN some of its special tinuum of time scales, which span the range
properties. It is tempting to speculate that SSA in from sound processing time scales of
auditory cortex is one of several inputs to the 1–100 ms to the time scales that would be
neural activation that is eventually measured as already considered as representing plastic
MMN on the scalp. changes, remains an important frontier of audi-
While SSA and forward masking are the best- tory neuroscience.
studied context-dependent processes in the audi-
tory system, they are not the only ones. As early
as the IC, there are processes that adapt neuronal References
response properties to the statistics of the incom-
ing sound streams (Dean et al. 2005). Thus, for Antunes FM, Nelken I, Covey E, Malmierca MS
(2010) Stimulus-specific adaptation in the auditory
example, in the IC (and to some extent even in the thalamus of the anesthetized rat. PLoS One 5(11):
auditory nerve), sound level coding adapts to the e14071
statistics of the sound levels to which the animal Dean I, Harper NS, McAlpine D (2005) Neural population
is exposed. Roughly speaking, neurons shift their coding of sound level adapts to stimulus statistics. Nat
Neurosci 8(12):1684–1689
level responses so that the steepest part of their Eggermont JJ (1985) Peripheral auditory adaptation and
dynamic range overlays the sound level range fatigue: a model oriented review. Hear Res
with the highest occurrence probability. Simi- 18(1):57–71
larly, temporal response properties of auditory Farley BJ, Quirk MC, Doherty JJ, Christian EP (2010) -
Stimulus-specific adaptation in auditory cortex is an
neurons may change with changes in the noise NMDA-independent process distinct from the sensory
level. For example, in the IC, at good signal-to- novelty encoded by the mismatch negativity.
noise levels, neurons are mostly temporally band- J Neurosci 30(49):16475–16484
pass, but in the presence of noise, they become Fritz J, Shamma S, Elhilali M, Klein D (2003) Rapid task-
related plasticity of spectrotemporal receptive fields in
low-pass (Lesica and Grothe 2008). Such primary auditory cortex. Nat Neurosci
changes may occur over a few hundreds of 6(11):1216–1223
ms. Changes with similar flavor have been also Grimm S, Escera C (2012) Auditory deviance detection
documented in bird songs (Nagel and Doupe revisited: evidence for a hierarchical novelty system.
Int J Psychophysiol 85(1):88–92
2006). Gutfreund Y (2012) Stimulus-specific adaptation, habitu-
Context sensitivity may be present over many ation and change detection in the gaze control system.
different time scales, from a few ms in the Biol Cybern 106(11–12):657–68
C 812 Contextual Integration
Lesica NA, Grothe B (2008) Dynamic spectrotemporal

feature selectivity in the auditory midbrain. Continuum Spine Theory
J Neurosci 28(21):5412–5421
Litovsky RY, Colburn HS, Yost WA, Guzman SJ
(1999) The precedence effect. J Acoust Soc Am ▶ Dendritic Spines: Continuum Theory
106(4 Pt 1):1633–1654
Malmierca MS, Cristaudo S, Perez-Gonzalez D, Covey
E (2009) Stimulus-specific adaptation in the inferior
colliculus of the anesthetized rat. J Neurosci
29(17):5483–5493 Continuum Theory for Active Spines
Nagel KI, Doupe AJ (2006) Temporal processing and
adaptation in the songbird auditory forebrain. Neuron
51(6):845–859 ▶ Dendritic Spines: Continuum Theory
Nakamoto KT, Zhang J, Kitzes LM (2006) Temporal
nonlinearity during recovery from sequential inhibi-
tion by neurons in the cat primary auditory cortex.
Nelken I, Ulanovsky N (2007) Change detection, Contour Completion and Pop-Out,
mismatch negativity and stimulus-specific adaptation Models of
in animal models. J Psychophysiol 21:214–223
Plack CJ, Oxenham AJ (1998) Basilar-membrane
nonlinearity and the growth of forward masking. ▶ Center-Surround Processing, Network Models of
J Acoust Soc Am 103(3):1598–1608
Taaseh N, Yaron A, Nelken I (2011) Stimulus-specific
adaptation and deviance detection in the rat auditory
cortex. PLoS One 6(8):e23369
Ulanovsky N, Las L, Nelken I (2003) Processing of Control of Aquatic and Terrestrial
low-probability sounds by cortical neurons. Nat
Neurosci 6(4):391–398
Gaits in Salamander
Wehr M, Zador AM (2005) Synaptic mechanisms of for-
ward suppression in rat auditory cortex. Neuron Auke Jan Ijspeert1 and Jean-Marie Cabelguen2
47(3):437–445 1
EPFL École Polytechnique Fédérale de
Yaron A, Hershenhoren I, Nelken I (2012) Sensitivity to
complex statistical regularities in rat auditory cortex.
2
Neuron 76(3):603–615 Neurocentre Magendie, INSERM U 862 –
Bordeaux University, Bordeaux, France
Contextual Integration Synonyms
▶ Center-Surround Processing, Computational Swimming and walking modes of the salamander

Role of
Definition
Contextual Processing Salamanders are amphibian animals that are

capable of several aquatic and terrestrial gaits.
▶ Center-Surround Processing, Computational These gaits are in large part controlled by central
Role of pattern generator (CPG) networks in the spinal
cord. These networks can be modeled at several
levels of abstraction from detailed models based
on Hodgkin-Huxley type of neurons to abstract
Continuum Spine Model systems of coupled oscillators. The models have
been instrumental in testing some hypotheses
▶ Dendritic Spines: Continuum Theory concerning gait generation and gait transition in
Control of Aquatic and Terrestrial Gaits in Salamander 813 C
the salamander. One of the main hypotheses is gaits are only two out of several others (the most
that the salamander CPG is constructed out of two frequent ones), and the animal exhibits a large
main subnetworks: a lamprey-like swimming movement repertoire.
CPG for the axial musculature and slower CPGs Like many other vertebrate animals, the peri-
for the limbs. Some of the models have been odic and coordinated activations of axial and
tested on a salamander-like robot and demon- limb muscles are produced by central pattern
strated their ability to make transitions between generators (CPGs) located in the spinal cord C
swimming and walking gaits by varying the level (Grillner 1981). One can distinguish an axial
of tonic input applied to the CPGs. CPG that activates the axial musculature in the
trunk and tail and limb CPGs that activate limbs
muscles. Both types of CPGs are composed of
Detailed Description multiple-coupled oscillators, i.e., neural pools
that can oscillate independently (Cheng
Locomotion in Salamanders et al. 1998, 2002) and that are coupled through
Salamanders are key animals from an evolution- projections between interneurons for producing
ary point of view. Their body plan is close to coordinated activity patterns. For instance, it has
early terrestrial tetrapods (Cohen 1988); Edwards been shown that surgically isolated segments and
1977), and they offer an interesting opportunity hemisegments of the axial CPG can produce
to study the transition from aquatic to terrestrial rhythmic activity when pharmacologically acti-
locomotion. Also they provide a useful link vated (Ryczko et al. 2010a). Like the lamprey, the
between popular animal models for studying ver- axial CPG is therefore a double chain of oscilla-
tebrate locomotion such as on one hand swim- tors with reciprocal inhibition to produce left-
ming animals like the lamprey and zebra fish and right alternation. The limb CPG appears to be
the other hand mammal tetrapods like the rat composed of a set of independent oscillators for
and cat. flexor and extensor muscles, which are coupled
In water, salamanders typically use an with inhibition in parallel with weaker excitation
anguilliform swimming gait, i.e., a gait in which (Cheng et al. 1998, 2002).
a traveling wave of lateral undulation is propa- One can therefore hypothesize that the
gated all along the body from head to tail salamander axial CPG has been inherited from
(see Fig. 3, bottom). The wavelength of the older, swimming vertebrates like the lamprey
undulation is kept more or less constant and is (Edwards 1977; Cohen 1988; Delvolvé
close to one body length. The animal can et al. 1997; Ijspeert 2001) and that the circuit
change speed by modulating the frequency of has been extended during evolution by phyloge-
the undulation, and the swimming frequency netically more recent networks controlling the
range is relatively high, e.g., 2.22–4.49 Hz limbs. This hypothesis is consistent with the
(mean: 3.09 Hz) in Pleurodeles waltl (Delvolvé observation that salamanders and lampreys
et al. 1997). share the same anguilliform swimming mode.
On ground, salamanders switch to either walk- For homologies between lamprey and salaman-
ing trot gaits (i.e., a gait in which diagonal pairs der, see Ryczko et al. (2010b).
of limbs are moved in synchrony) or diagonal A very interesting property of the salamander
sequence walking gaits (see Fig. 3, top, for locomotor network is that electrical stimulation
a walking trot). The frequencies are significantly of the mesencephalic locomotor region (MLR,
lower, e.g., 0.58–2.17 Hz (mean: 1.23 Hz) in a part of the brain stem that controls the initiation
Pleurodeles waltl (Delvolvé et al. 1997), and in and maintenance of locomotion; Dubuc 2009) in
a given animal do not overlap with those a decerebrated salamander will induce gait tran-
of swimming. A review of the kinematics of sitions depending on the level of stimulation
salamander locomotion can be found in (Cabelguen et al. 2003). The level of activation
Karakasiliotis et al. (2013). Note that these two of that region determines the frequency and the
C 814 Control of Aquatic and Terrestrial Gaits in Salamander
mode of locomotion, with (slow) stepping move- Using Numerical Models and Robots to
ments at low intensity and (faster) swimming Decode the Gait Generation Mechanisms
movements at high intensity (Cabelguen Different types of models have been developed to
et al. 2003). Interestingly, a gradual increase of address the questions listed above. As will be
stimulation will lead to an abrupt switch of gaits briefly described later, the models vary in their
when a certain stimulation threshold is reached. level of abstraction, from detailed models com-
From a dynamical systems point of view, the posed of Hodgkin-Huxley-like neurons (Bicanski
MLR and spinal cord can therefore make et al 2013b) to integrate-and-fire neurons (Kn€usel
a bifurcation between two very different regimes et al. 2013), to leaky-integrator neurons (Ijspeert
under the control of a simple tonic (i.e., 2001; Ijspeert et al. 2005), and to abstract models
non-oscillating) signal. Similar gait transitions made of coupled phase oscillators (Ijspeert
under MLR stimulation have been observed in et al. 2007; Kn€usel et al. 2013). Some models
other vertebrates, e.g., transitions between walk- have been tested together with a model of the
ing, trotting, and galloping in decerebrate cats body either in simulation or on board of a real
(Shik et al. 1966). robot (Ijspeert 2001; Ijspeert et al. 2005, 2007).
More detailed reviews of salamander locomo- See Bicanski et al. (2013a) for a review. The next
tion and neurophysiology can be found in sections will present two models, one based on
Chevallier et al. (2008), Karakasiliotis coupled oscillators and one based on Hodgkin-
et al. (2013), and Bicanski et al. (2013a). While Huxley-like neurons.
we now have a good grasp of the general organi-
zation and functioning of the salamander loco- Oscillator Models of the Salamander
motor networks, many aspects still require more Locomotor Circuit
investigation and can strongly benefit from com- Models made of abstract oscillators are useful to
putational studies. In particular the following investigate and test hypotheses concerning the
questions need to be addressed: general topology of the couplings between neural
1. What are the mechanisms of rhythm genera- oscillators and the mechanisms of gait transition,
tion and how are the local oscillators i.e., questions 2 and 4 as listed above.
implemented, i.e., which types of interneu- These models are based on the assumptions
rons are involved and how are they that the rhythm generation ability of a local net-
interconnected? work of neurons can be represented by a simple
2. How are the oscillators of the whole network, oscillator model, without the need to model the
i.e., axial and limb CPGs, coupled together? many individual neurons that compose the local
3. How much does sensory feedback affect pat- network (in the order of several hundreds, the
tern generation? exact number is not known), and that the global
4. What are the mechanisms of gait transition, in dynamics of the coupled oscillator network is
particular the switch between walking and mainly dependent on high-level properties such
swimming under MLR stimulation? as intrinsic frequencies of the oscillators, topol-
5. What is the interplay between multiple ogy of inter-oscillator couplings, and strengths of
descending pathways and the spinal circuits couplings rather than on the local rhythm gener-
that allows the animal to produce its large ation mechanisms. Such an approach is based on
movement repertoire? dynamical systems theory and has been useful to
As we will see next, computational models can help understand the dynamics of the lamprey
play a significant role in helping answering these swimming network (Kopell 1995).
questions and decoding the mechanisms of gait A model based on coupled oscillators was
generation and gait transition. Note that many of used to investigate the spinal mechanisms under-
these questions are relevant for vertebrate ani- lying the gait transition between stepping and
mals in general. swimming (question 4 above). To compare
a b
d MLR d Brainstem
CPG model
(spinal cord)
17 18
1 9 C
7 8
2 10 1
PD controller
3 11
3
4 12
19 20
5 13 10
9
4
6 14
5
7 15
ϕi Vi
8 16 6
∼
ϕ i
Control of Aquatic and Terrestrial Gaits in Salaman- is composed of a body CPG – a double chain of 16 oscil-
der, Fig. 1 Configuration of a CPG model based on lators with nearest neighbor coupling for driving the spine
coupled oscillators (a) and salamander robot (b) (From motors – and a limb CPG, four oscillators for driving the
Ijspeert et al. (2007)). The robot is driven by 10 DC limb motors. The CPG model receives left and right drive
motors, which actuate six hinge joints for the spine signals d from the MLR region in the brain stem. The
(black disks in the schematic view of the robot) and four velocity, direction, and type of gait exhibited by the robot
rotational joints for the limbs (black cylinders). The CPG can be adjusted by modifying these two signals
characteristics of the resulting gaits with those of

where yi and ri represent the phase and amplitude
the salamander, the mathematical model was
of the oscillator i, ni and Ri determine the intrinsic
coupled with a mechanical model of the body,
frequency and amplitude (i.e., the frequency and
i.e., an amphibious salamander-like robot
amplitude toward which the oscillator converges
Salamandra robotica (Fig. 1 right) (Ijspeert
when isolated), and ai is a positive constant.
et al. 2007).
Oscillators are coupled together with diffusive
The model represents the rhythmic activity of
coupling with a coupling weight wij and a phase
a local neural oscillator using the so-called
bias fij. The phase bias determines the phase lag
amplitude-controlled phase oscillator. It is an
that the coupling tends to induce between oscilla-
oscillator with two state variables: the phase and
tors. The positive oscillatory signal xi represents
the amplitude of oscillations. Its time evolution of
the periodic bursts produced by the oscillator.
an oscillator i is governed by the following
Such an oscillator is essentially a Kuramoto oscil-
equations:
lator (Strogatz 2000) for the phase, extended
X
with a critically damped second-order linear dif-
y_ i ¼ 2p vi þ r j wij sin yj yi fij (1) ferential equation for the amplitude. An isolated
j
oscillator will converge to a steady-state regime

a
i
with sinusoidal output xi1 with frequency ni and
r€i ¼ ai ðRi r i Þ r_i (2) amplitude Ri: xi1 (t) = Ri (1 + cos(ni t + y0)).
4
The CPG model is used to test four hypothe-
xi ¼ r i ð1 þ cos ðyi ÞÞ (3) ses: (1) the axial CPG of the salamander is
organized like that of the lamprey, i.e., as hypotheses 1 and 2. The so-called saturation
a double chain of oscillators that produces trav- function is implemented to adjust the intrinsic
eling waves when activated. (2) A strong cou- frequency and amplitude of limb and axial oscil-
pling from limb oscillators to axial oscillators lators depending on the drive (representing the
allows the limb oscillators to impose their descending stimulation coming from the MLR)
(slower) frequency on the axial network when according to hypotheses 3 and 4. The equations of
limbs are activated. (3) The limb oscillators can- the coupled differential equations are numeri-
not oscillate at high frequencies, i.e., they satu- cally integrated on board of a microcontroller of
rate and stop oscillating at high levels of the robot.
stimulation. (4) Limb oscillators have a lower When a low drive is applied to the model, it
intrinsic frequency than the axial oscillators for leads to salamander-like stepping movements,
the same level of drive. with periodic motions of the limbs that are in
Following these hypotheses, the model is antiphase between fore- and hind limbs and with
made of 20 coupled oscillators, 16 for the axial a standing wave of trunk and tail undulation
CPG and 4 for the limb CPGs (Fig. 1, left). (Fig. 2). When the drive is increased, both fre-
The topology and strengths of couplings satisfy quency and amplitude of the lateral undulation
Control of Aquatic and Terrestrial Gaits in Salaman- releases a traveling wave in the axial CPG as needed for
der, Fig. 2 Transition from stepping to swimming. anguilliform swimming. The gait transition is accompa-
The transition occurs when the drive applied to the CPG nied by a jump in instantaneous frequency, with low
model is linearly increased (From Ijspeert et al. (2007)). frequencies for stepping and high frequencies for swim-
Below the first threshold, the circuit is silent. Once passing ming (third panel from top). Note that the large variations
this first threshold (at t = 4s), the CPG produces stepping- in instantaneous frequencies at t = 4s and t = 20s are due
like patterns with a standing wave in the axial CPG (top) to short transient periods during which the phases of
and periodic activation of the limb CPGs (second from oscillators rapidly change before reaching steady-state
top panel). When reaching the limb saturation threshold regimes
(t = 20s), the limb oscillators stop oscillating, and this
are increased leading to faster stepping. When the
drive passes the saturation threshold of the limb
oscillators, these stop oscillating, and the axial
oscillators rapidly switch to a traveling wave
regime. Also because of the saturation, the limb
oscillators do not impose anymore slower fre-
quencies, and the frequency of oscillations then C
rapidly increases, similarly to what is observed in
the real animal. The model therefore correctly
replicates the abrupt gait transition induced by
MLR stimulation (Cabelguen et al. 2003). It
also provides an explanation of why swimming
frequencies are systematically higher than
stepping frequencies in kinematic recordings of
freely moving animals (Frolich and Biewener
1992; Delvolvé et al. 1997).
Interestingly, despite the robot being only
an approximation of the salamander body
(e.g., in terms of number of joints and the
simplified limb design), the movements pro-
duced by the robot are strikingly similar to
those of the animal both in water and on ground
(Fig. 3). By adjusting the drive level, the speed
of locomotion can be adjusted both during
stepping and swimming, and rapid switches
between the two gaits can be performed. By
providing more drive to the axial oscillators on
one side compared to the other, turning move-
Control of Aquatic and Terrestrial Gaits in Salaman-
ments can be induced. The higher drive on one
der, Fig. 3 Stepping and swimming gaits in the sala-
side leads to higher amplitude of bending mander and the robot. Top two panels: walking trot gait
toward that side. where the fore- and hind limbs are in antiphase and with
the trunk and tail performing a standing wave. Bottom two
panels: anguilliform swimming gait with the body making
Hodgkin-Huxley Models of the Salamander
a lateral undulation traveling from head to tail (Adapted
Oscillatory Circuits from the supplementary material of Ijspeert et al. (2007))
Models based on Hodgkin-Huxley (HH) neuron
equations are useful to investigate the mecha-
nisms of rhythm generation in the salamander
(Question 1). Unlike the lamprey, the types of on one side of the axial CPG) containing sparsely
interneurons involved in the local oscillator net- connected excitatory and inhibitory neuron
works of the salamander and the topology of their populations (Bicanski et al. 2013b).
coupling are yet unknown. However, various Hemisegments in the same segment are
experiments have shown that salamander and connected through both excitatory and inhibitory
lamprey segmental circuits share many similari- contralateral projections (Fig. 4). The model
ties (Ryczko et al. 2010b). matches a large range of experimental findings,
Models of the salamander segmental circuitry in particular the pharmacological stimulation by
can therefore be developed with lamprey-like NMDA (N-methyl-D-aspartate) of isolated axial
spinal hemisegments (i.e., oscillator networks hemisegments and segments (Fig. 4, bottom).
Control of Aquatic and Terrestrial Gaits in Salaman- hemisegment, i.e., a total of 160 neurons for the whole
der, Fig. 4 Salamander neural oscillator model made segmental network. Bottom: a representative average
of HH neurons. Top: topology of the circuit made of membrane potential trace in mV of the left and right
populations of excitatory (EIN) and inhibitory (CCIN) hemisegmental EIN populations (Adapted from Bicanski
neurons. Each drawn neural unit represents a population et al. (2013b))
of neuron models: 50 EINs and 30 CCINs per
The model reproduces most of the effects of potentially serve as a building block for the
various pharmacological experiments (Ryczko axial and limb oscillators in salamanders.
et al. 2010a) that have been performed on isolated
salamander spinal cords (e.g., the blockade of Sensory Feedback, Rich Motor Skills, and
AMPA synapses, glycinergic synapses, calcium- Descending Pathways
activated potassium current, persistent sodium The models above do not include sensory feed-
current, and h-current). back, and it is likely that (multimodal) sensory
The model can not only replicate NMDA- feedback plays an important role in shaping the
induced oscillations, which are typically slower motor patterns and in adjusting their timing dur-
than in vivo, but also the range of in vivo fre- ing locomotion. For instance, the difference of
quency when receiving excitatory input from interaction forces between water and ground
a population of simulated brain stem neurons could possibly explain difference between trav-
(i.e., a more realistic way of stimulating spinal eling and standing waves in axial CPG during
cord circuits). By slightly adjusting the conduc- swimming and stepping. This was shown in
tances involved in burst termination, neural oscil- a neuromechanical simulation that combined
lators with different frequency ranges can be a simplified musculoskeletal model with a CPG
implemented, suggesting that the same circuit model based on leaky-integrator neurons
can be used both to represent slower limb oscil- (Ijspeert et al. 2005). In the model, a simulated
lators and faster axial oscillators. When a slower swimming salamander placed on dry ground
limb oscillator is coupled to a faster axial oscil- tends to have lateral undulations that are close
lator, it can impose its slower frequency as to a standing wave even when the CPG produces
needed during stepping and similarly to the oscil- a traveling wave. In other words, the isotropic
lator model above. The model therefore suggests friction forces on dry ground imposed the stand-
that a lamprey-like segmental network can ing wave of body movements. When sensory
feedback from simulated stretch-sensitive cells above. Preliminary studies show that multiple
was added to the circuit, the CPG activity was different motor behaviors can be generated by
transformed into a standing wave that resembles simply being able to activate subparts of the
the standing waves of EMG patterns recording circuits differentially through different indepen-
during stepping on ground (Delvolvé et al. 1997). dent pathways (e.g., by providing different levels
This illustrates how physical interaction with the of drives to the neck, trunk, and tail oscillators in
environment together with sensory feedback can the axial CPG). But clearly the models will C
modulate CPG patterns. require modeling more complex circuits and
Note that the coupled oscillator model more sensory streams to account for the richness
described above (Ijspeert et al. 2007) relied on of salamander motor skills. In particular the limb
long-range couplings from limb to axial oscilla- circuits are so far oversimplified and require more
tors, with the forelimb oscillators sending pro- detailed studies. Continuing to combine experi-
jections to all trunk oscillators and the hind limb ments in neurophysiology together with model-
oscillators sending projections to all tail oscilla- ing and robotic studies will be instrumental in
tors. Such a global coupling from limb to axial making progress in this exciting field.
oscillators meant that a standing wave was
imposed on the axial CPG as soon as the limb
CPGs were active. Since then, experiments on References
isolated spinal cords have shown that the axial
Bicanski A, Ryczko D, Knuesel J, Harischandra N,
CPG can produce traveling waves even when the
Charrier V, Ekeberg Ö, Cabelguen JM, Ijspeert
limb CPGs are rhythmically active (Ryczko A (2013a) Decoding the mechanisms of gait genera-
et al. 2009). The frequencies of limb and axial tion in salamanders by combining neurobiology,
CPGs are then identical and slow, supporting the modeling and robotics. Biol Cybern 107(5):545–564
Bicanski A, Ryczko D, Cabelguen JM, Ijspeert A (2013b)
idea of strong couplings from limb to axial CPGs
From lamprey to salamander: an exploratory modeling
(hypothesis 2 in the coupled oscillator model). study on the architecture of the spinal locomotor net-
But the fact that traveling waves are then works in the salamander. Biol Cybern 107(5):565–587
observed instead of standing waves suggests Cabelguen J-M, Bourcier-Lucas C, Dubuc R (2003)
Bimodal locomotion elicited by electrical stimulation
that the coupling from limb to axial CPGs is
of the midbrain in the salamander Notophthalmus
only local, close to the girdles, and not global as viridescens. J Neurosci 23:2434–2439
in Ijspeert et al. (2007). It is therefore likely that Cheng J, Stein RB, Jovanovic K, Yoshida K, Bennett DJ,
the standing waves of EMGs during stepping are Han Y (1998) Identification, localization, and modu-
lation of neural networks for walking in the mudpuppy
due either to sensory feedback or to modulation
(Necturus maculatus) spinal cord. J Neurosci
from descending pathways (rather than a central 18:4295–4304
effect based on global couplings). Cheng J, Jovanovic K, Aoyagi Y, Bennett DJ, Han Y,
An important topic that remains to be Stein RB (2002) Differential distribution of interneu-
rons in the neural networks that control walking in the
addressed is how the salamander produces its
mudpuppy (Necturus maculatus) spinal cord. Exp
large repertoire of movements (Question Brain Res 145:190–198
5 above). In addition to anguilliform swimming Chevallier S, Ijspeert AJ, Ryczko D, Nagy F, Cabelguen
and walking trots described above, the salaman- JM (2008) Organisation of the spinal central pattern
generators for locomotion in the salamander: biology
der can perform a large variety of movements
and modelling. Brain Res Rev 57:147–161
such as backward walking, paddling limbs in Cohen AH (1988) Evolution of the vertebrate central
phase, aquatic stepping, etc. The mechanisms pattern generator for locomotion. In: Cohen AH,
underlying this large repertoire remain to be Rossignol S, Grillner S (eds) Neural control of rhyth-
mic movements in vertebrates. Wiley-Interscience,
decoded, but it is most likely that they involve
New York
adding more descending pathways and more Delvolvé I, Bem T, Cabelguen JM (1997) Epaxial and
complex spinal circuits to the models presented limb muscle activity during swimming and terrestrial
C 820 Control of Breathing, Integration of Adaptive Reflexes
stepping in the adult newt, Pleurodeles waltl.

J Neurophysiol 78:638–650 Control of Breathing, Integration of
Dubuc R (2009) Locomotor regions in the midbrain
(MLR) and diencephalon (DLR). In: Binder MD, Adaptive Reflexes
Hirokawa N, Windhorst U (eds) Encyclopedia of neu-
roscience. Springer, Berlin Chung Tin1 and Chi-Sang Poon2
Edwards JL (1977) The evolution of terrestrial locomo- 1
Department of Mechanical and Biomedical
tion. In: Hecht MK, Goody PC, Hecht BM (eds) Major
patterns in vertebrate evolution. Plenum, New York, Engineering, City University of Hong Kong,
pp 553–576 Kowloon Tong, Hong Kong
2
Frolich L, Biewener A (1992) Kinematic and electromyo- Institute for Medical Engineering and Science,
graphic analysis of the functional role of the body axis Harvard-MIT Division of Health Sciences and
during terrestrial and aquatic locomotion in the sala-
mander Ambystoma tigrinum. J Exp Biol 162:107–130 Technology, Massachusetts Institute of
Grillner S (1981) Control of locomotion in bipeds, tetra- Technology, Cambridge, MA, USA
pods, and fish. In: Brooks VB (ed) Handbook of phys-
iology, the nervous system, vol 2, Motor control,
section 1. American Physiology Society, Bethesda,
pp 1179–1236 Definition
Ijspeert AJ (2001) A connectionist central pattern genera-
tor for the aquatic and terrestrial gaits of a simulated
Models of respiratory regulation – as with many
salamander. Biol Cybern 84:331–348
Ijspeert AJ, Crespi A, Cabelguen J-M (2005) Simulation physiological control systems – have tradition-
and robotics studies of salamander locomotion: apply- ally invoked the Sherringtonian reflex paradigm
ing neurobiological principles to the control of loco- (Sherrington 1906) in which the response to any
motion in robots. Neuroinformatics 3:171–195
afferent input is assumed to follow a static
Ijspeert AJ, Crespi A, Ryczko D, Cabelguen J-M
(2007) From swimming to walking with stimulus–response relationship that adds to
a salamander robot driven by a spinal cord model. other stimulus–response relationships to produce
Science 315:1416–1420 the overall response. Such a reductionist analysis
Karakasiliotis K, Schilling N, Cabelguen J-M, Ijspeert AJ
has set the foundation for our understanding of
(2013) Where are we in understanding salamander
locomotion: biological and robotic perspectives on various respiratory control mechanisms, such as
kinematics. Biol Cybern 107(5):529–544 central and peripheral chemoreflex, lung inflation
Kn€usel J, Bicanski A, Ryczko D, Cabelguen JM, Ijspeert and deflation reflexes, and exercise reflex via
AJ (2013) A salamander’s flexible spinal network for
various chemical and mechanical afferent feed-
locomotion, modeled at two levels of abstraction.
Integr Comp Biol 53(2):269–282 backs and feedforward commands. Under this
Kopell N (1995) Chains of coupled oscillators. In: Sherringtonian framework, such feedback and
Arbib MA (ed) The handbook of brain theory and feedforward controls are taken as a combination
neural networks. MIT Press, Cambridge, MA,
of simple reflex drives acting in parallel that are
pp 178–183
Ryczko D, Lamarque S, Didier H, Cabelguen JM integrated algebraically by the brainstem respira-
(2009) Dynamics of the axial locomotor network in tory controller in determining the resultant respi-
the isolated spinal cord of the salamander. Soc ratory rhythm (inspiratory/expiratory duration)
Neurosci. Program 565.8, Abstract EE6
and motor pattern (tidal volume) in response to
Ryczko D, Charrier V, Ijspeert A, Cabelguen JM (2010a)
Segmental oscillators in axial motor circuits of the various exogenous and endogenous stimuli.
salamander: distribution and bursting mechanisms. However, this simplified reductionist framework
J Neurophysiol 104:2677–2692 of respiratory control has been challenged
Ryczko D, Dubuc R, Cabelguen J-M (2010b)
by recent evidence indicating that such
Rhythmogenesis in axial locomotor networks: an
interspecies comparison. Prog Brain Res 187:189–211 stimulus–response relationships are not merely
Shik ML, Severin FV, Orlovskii GN (1966) Control of mediated by static reflexes in the form of classical
walking and running by means of electric stimulation proportional feedback and feedforward controls
of the midbrain. Biofizika 11:659–666
but may involve adaptive neural mechanisms that
Strogatz S (2000) From Kuramoto to Crawford: exploring
the onset of synchronization in populations of coupled conform to adaptive control systems in modern
oscillators. Phys D 143(1–4):1–20 engineering control theory.
Control of Breathing, Integration of Adaptive Reflexes 821 C
According to Sherrington (1906), reflexes are conditioning differs from nonassociative learning
animal instincts evolved through natural selec- in that the plasticity is not only use dependent but
tion to serve specific survival purposes in adap- also pairing specific, requiring the interaction of
tation to challenging environments, hence the two coincident afferent inputs called the condi-
term “adaptive reflex.” On the other hand, it is tioned and unconditioned stimuli. A canonical
now well recognized that many reflexes in lower example of classical conditioning is the mamma-
animals (e.g., Aplysia gill and siphon withdrawal lian eye-blink reflex (Thompson and Steinmetz C
reflex) and mammalian higher brain mechanisms 2009). Another form of associative learning is
(e.g., eye-blink reflex) may also adapt in real time operant conditioning (also called instrumental
to modify the stimulus–response relationships conditioning or reinforcement learning), in
through nonassociative and associative learning, which the response to a stimulus may be modified
two forms of use-dependent (activity-dependent) by feedback of the resultant response in a closed
neural plasticity that contribute importantly to loop in order to optimize certain performance
brain intelligence. This entry will focus on the measure (Hawkins et al. 2006). Experimental
roles of nonassociative and associative learning observations suggest that associative learning
in adaptive modulation of chemoreflex, may play an important role in modulating the
mechanoreflex, and exercise reflex control of respiratory controller gain through integration of
breathing. a variety of respiratory afferent information. This
This entry is organized as follows. The next suggests that the respiratory controller may not
section presents the notion of nonassociative simply add multiple stimuli linearly but may
learning in respiratory control. Nonassociative integrate them in a nonlinear, time-dependent
learning is the progressive decrease (short-term manner to bring out an optimal response in the
depression, STD) or increase (short-term poten- face of changing environmental challenges. Such
tiation, STP) in the evoked response upon con- optimal respiratory control suggests a close cor-
tinued or repetitive afferent stimulation, relation between homeostatic regulation in phys-
evidencing use-dependent plasticity (Poon and iology and the internal model principle in
Schmid 2012). Two traditional forms of engineering (Poon et al. 2007) – which states
nonassociative learning (habituation and sensiti- that a feedback regulator under external distur-
zation evidencing STD and STP, respectively) bances may maintain regulation and stability pro-
have been extensively studied in the classic vided a suitably reduplicated (internal) model of
experiment of Aplysia gill and the disturbance signal is adapted in the feedback
siphon withdrawal reflex (Hawkins et al. 1998). path (Francis and Wonham 1975). The notion of
A third form of nonassociative learning called internal model is now well recognized in senso-
“desensitization” – also evidencing STD but is rimotor integration (Davidson and Wolpert 2005;
expressed in a secondary pathway – was first Tin and Poon 2005). This section concludes with
discovered in the pontine modulation of the a discussion of the relevance of internal model
Hering–Breuer reflex in rats (Siniaia learning in respiratory control.
et al. 2000). The three forms of nonassociative
learning – habituation, sensitization and
desensitization – have been shown to have Detailed Description
significant computational relevance and can
be formulated as a neural integrator/ Nonassociative Learning in
differentiator that modulates the stimulus– Vagal–Mechanical Reflex and Chemoreflex
response relationship in a time-dependent Adaptation
manner. Nonassociative learning with activity-dependent
The third section discusses the role of asso- STP and STD of neurotransmission is prevalent
ciative learning in respiratory control. Associa- in respiratory control. For instance, during hyp-
tive learning such as classical (Pavlovian) oxia or sustained carotid sinus nerve stimulation,
STP is evidenced by a progressive increase in the x_ ¼ ax þ bu

(1)
amplitude of respiratory motor response, y ¼ cx þ du
followed by a gradual decay (afterdischarge) of
the heightened response upon removal of the where x and x_ are the state variable and its time
stimulus (Poon et al. 1999; Song and Poon derivative, respectively; y and u are the output
2009b). Such neural plasticity phenomena have and input of the integrator or differentiator;
important computational implications. a, b, c, and d are the system parameters with
a > 0 to guarantee stability.
Nonassociative Learning as Neural Integrator and With a constant-step input, u = u0 for 0 < t
Differentiator T, where T is the duration of stimulation, Eq. 1
Functionally, the temporal patterns of STP and can be solved analytically for x(t) (and hence y(t))
STD in response to a stimulus can be modeled assuming x(0) = 0:
respectively as a neural integrator and neural
differentiator, or equivalently a low-pass and cbu
yðtÞ ¼ du þ ð1 eat Þ for 0 < t T
high-pass filter (Poon and Siniaia 2000). This a
analogy provides a strong mathematical founda- (2)
tion that underpins nonassociative learning in the
perspective of adaptive control. Numeric integra- The response y(t) corresponds to a neural inte-
tion and differentiation are elemental calculus grator if both terms on the right-hand side of Eq. 2
operations. They also underlie all temporal have the same sign, or a neural differentiator if
dynamics and kinematics phenomena in Nature. they have opposite signs. Similarly, their
An analog integrator or differentiator is any phys- afterdischarge during off-transient can be
ical process with observable input and output modeled by the following equation:
signals that demonstrate integral or differential
transformation in real time, respectively. Analog yðtÞ ¼ ðyðT Þ yð0ÞÞeat for y > T (3)
integrators and differentiators are limited by their
leakages, i.e., physical bypasses that offset the Hence, neural integrator and differentiator can
integration and differentiation processes. Leaky be characterized by the above exponential func-
integrators or low-pass filters are commonly used tions in the time domain. Higher-order integrator
in electrophysiology experiments to obtain and differentiator can be achieved by combining
a moving time average of neuronal firing, which multiple exponentials with different gains and
gives an analog “neurogram” of spike frequency time constants.
as an output. From linear time-invariant systems theory,
In the time domain, a leaky integrator’s neural integrator and differentiator can also be
response to a constant-step input exhibits expo- expressed in frequency domain using Laplace
nential saturation during on-transient and expo- transform. In this regard, they perform low-pass
nential decay during off-transient, whereas the and high-pass filtering to the input signal. The
corresponding response of a leaky differentiator time constant of a leaky integrator or
demonstrates exponential decay from an initial differentiator is inversely proportional to the cut-
overshoot during on-transient and exponential off frequency of the equivalent low-pass or high-
recovery from rebound undershoot during pass filter, respectively. In addition to frequency
off-transient (Fig. 1a). As such, STP and STD filtering, a leaky integrator/differentiator also
are closely related to neural integrator/ introduces phase shift (phase lag/phase lead) in
differentiator as they display characteristic inte- the input–output relationship (Fig. 1b, c).
gral/differential responses. A first-order neural An integrator or differentiator that sustains an
integrator or differentiator can be readily off-transient response with short-term memory
described by the following state-space model decay upon removal of the stimulus is said to be
(Poon and Young 2006): biphasic (or else, monophasic); it is termed
a Integrator Differentiator
Norm. Response
3 3
1 1
C
0 60 120 0 60 120
Time (sec) Time (sec)
b
Relative Gain
1 1
0.1 0.1
0.1 1 10 100 0.1 1 10 100
Frequency (cpm) Frequency (cpm)
c R2
C R2
V0
V0
Is R1 R3 Is R1 C V0
R3
Control of Breathing, Integration of Adaptive frequency domain, an integrator/differentiator behaves

Reflexes, Fig. 1 Time and frequency response charac- like a low-pass/high-pass filter. The passband in both
teristics of integrator (left panels) and differentiator (right cases is the frequency range where the transmission gain
panels). (a) The temporal response of a leaky integrator to (normalized to unity) is highest and relatively constant.
a constant-step stimulus (horizontal bar) consists of abrupt The high and low cutoff frequencies (vertical dotted lines)
reflex increase/decrease of the response at stimulus onset/ of these filters are inversely proportional to the time con-
cessation followed by exponentially increasing/decaying stants of the corresponding integrator and differentiator
(potentiation/afterdischarge) on-/off-transients. A leaky shown in a. (c) Examples of RC integrator (C = 3.77,
differentiator has similar reflex components but with R1 = 1.60, R2 = 24.0, R3 = 0.73) and differentiator
exponentially decaying/rebound-increasing (C = 1.0, R1 = 1.75, R2 = 24.0, R3 = 10.2) circuits
(accommodation/aftercharge) on-/off-transients, which with a current source (IS) input and voltage (V0) output.
are opposite to those of an integrator (overlaying dotted Units are arbitrary and RC values correspond to parameter
lines). In both cases the off-transients may be rectified, values as defined in Eq. 1 for integrator (a = 0.125,
with the response becoming monophasic (not shown) b = 0.125, c = 1, d = 0.5) and differentiator
instead of biphasic. The time scales chosen are typical of (a = 0.125, b = 0.125, c = 1, d = 1.5) (Adapted
oculomotor integrator and respiratory integrator. (b) In the from Poon and Young (2006))
inverted if the gain is negative. Under these broad (Poon and Young 2006). The STD or STP
definitions, nonassociative learning may be induced in a primary stimulus–response pathway
classified into four basic forms depending on (i.e., one that is directly activated by the primary
whether the corresponding neural integrator or stimulus) is necessarily monophasic, in that all
differentiator response is monophasic or biphasic reflex and adaptation effects mediated by that
Differentiator and Integrator Double Integrator

a b
Tonic Bias Tonic Bias
Secondary Secondary
Primary
Primary
c Double Differentiator d Integrator and Differentiator
Tonic Bias Tonic Bias

Secondary Secondary
Primary Primary
Control of Breathing, Integration of Adaptive secondary integrator is biphasic. Resultant response (last
Reflexes, Fig. 2 Compound neural integrator and box) at output neuron shows a compound differentiator–-
differentiator models. Boxes show model simulations in integrator characteristic. (b) A double integrator may arise
arbitrary units and parameter values. (a) Primary differen- from STP in both primary and secondary pathways. (c)
tiator–secondary integrator is realized by activity- A double differentiator is similar to a double integrator but
dependent habituation–sensitization in corresponding with STD instead of STP in both pathways. (d) A primary
primary–secondary pathways. Step application of primary integrator–secondary differentiator can be realized in
input at a constant firing rate (inset) induces synaptic STD a similar fashion with a STP-STD combination. Structur-
and STP (filled and open inner triangle) in primary and ally, primary differentiator (a, c) and secondary
secondary pathway, respectively, with corresponding differentiator (b, d) are comprised of the primary reflex
temporal differentiation and integration of the transmitted in conjunction with an inverted primary and secondary
signals (upper and lower boxes). Primary differentiator is integrator, respectively, demonstrating the complementar-
monophasic (with abrupt termination of response due to ities of integrator and differentiator (Adapted from Poon
auto-chopping at end of stimulus train), whereas and Young (2006))
pathway must vanish abruptly and any associated primary stimulus ends. Such a post-stimulus
memory effects become latent once the primary response provides a simple marker for discrimi-
stimulus ceases (Fig. 2). Habituation and primary nating the memory effects in the primary and
sensitization are monophasic STD and STP secondary pathways. The nonassociative learning
expressed in the primary stimulus–response path- effects corresponding to STD and STP mediated
way. The monophasic effects reflect the fact that by a secondary pathway have been termed desen-
the primary pathway is disfacilitated (gated off) sitization and secondary sensitization, respec-
once the primary stimulus ceases. Hence, habitu- tively, in contradistinction to their counterparts
ation and primary sensitization act not only as a (habituation and primary sensitization) that are
high-pass or low-pass filter in the frequency mediated by the primary pathway.
domain but also as a temporal filter in the time A primary stimulus may induce STP and/or
domain that gates off any residual memory effect STD in the primary and secondary pathways
of the STD and STP in the primary pathway once simultaneously to form compound integrator/
the primary stimulus is removed. differentiator. The dynamical order of
Alternatively, the primary stimulus may a compound integrator/differentiator is the num-
induce STP or STD of neurotransmission indi- ber of integrators/differentiators it is composed
rectly in a collateral (secondary) pathway which of, and its memory order is the number of com-
is driven by another (secondary) input (Fig. 2). In ponent integrators/differentiators that are
this event the resulting integrator/differentiator biphasic with short-term memory. The four pos-
response is biphasic with a post-stimulus short- sible second-order compound integrator/
term memory response, as the secondary input differentiator configurations are summarized in
keeps the secondary pathway active after the Fig. 2. Under this framework, the classical
dual-process theory of nonassociative learning afferent-induced phasic STD that is NMDA recep-
with concurrent habituation and secondary sensi- tor independent as well as long-term depression
tization (Groves and Thompson 1970) is a special that is NMDAR receptor dependent (Zhou
case of the second-order compound integrator/ et al. 1997). The biphasic STD component with
differentiator structure. a post-stimulus rebound response that is
dependent on NMDA receptor activity effectively
Short-Term Depression of Vagally Mediated discriminates the adaptive modulation of the C
Hering–Breuer Reflex Hering–Breuer reflex due to desensitization in
The classic Hering–Breuer inflation reflex the secondary (pontine) pathway from that due to
describes the inspiration inhibition and expiration habituation in the primary pathway. Pontine
prolongation consequent to lung inflation, which desensitization of the Hering–Breuer reflex in
is sensed by slowly adapting pulmonary stretch vagal–intact animals plays an important role in
receptors in the airways and conveyed by the restoring entrainment of the respiratory rhythm to
vagus nerve to the respiratory controller. This cyclical mechanical ventilation upon changing
“reflex” is presumed to protect the lungs from ventilator frequency and/or end-expiratory airway
potential damages due to overexpansion. How- pressure (Dutschmann et al. 2009; MacDonald
ever, new data accumulated over the past decade et al. 2007).
suggest that such a reflex is not static but is
dynamic and adaptive, conforming to Short-Term Potentiation and Depression of
nonassociative learning (MacDonald et al. 2009; Peripheral Chemoreflex
Poon et al. 2000; Siniaia et al. 2000). In addition to pulmonary stretch receptor affer-
During sustained lung inflation or electrical ents mediated by the vagus nerve, the respiratory
vagal stimulation in vagotomized rats, the imme- control system also senses arterial partial pres-
diate Hering–Breuer reflex response is countered sures of CO2 (PaCO2 ) and O2 (PaO2 ) and arterial pH
by an ensuing STD that gradually lessens the through central and peripheral chemoreceptors in
initial lengthening of expiratory duration (TE) order to regulate blood gas concentrations and
and decrease of respiratory frequency ( fR). This acid–base balance to maintain homeostasis.
is followed by a rebound shortening of TE and Abrupt exposure to hypoxic gas or brief electrical
increase of fR upon termination of lung inflation stimulation of the carotid sinus nerve (CSN) in
or vagal stimulation before these post-stimulus animal models elicits complex peripheral
STD responses gradually die out. The on- and chemoreflex response of inspiratory and expira-
off-transient STD responses in TE and fR resemble tory variables that vary in a time-dependent man-
a biphasic differentiator. In both cases, the ner (Poon et al. 1999, 2000; Song and Poon
on-transients are best fitted by second-order expo- 2009b). These include a biphasic STP of
nential functions, while the off-transients are best inspiratory amplitude (with post-stimulus
fitted with a first-order exponential. The curve- afterdischarge) and STD-like shortening of inspi-
fitting results suggest a second-order differentiator ratory duration (TI) resulting in a STP-like pro-
with monophasic STD (habituation) in the primary gressive increase of fR, as well as rapid shortening
pathway and biphasic STD (desensitization) in of TE followed by a biphasic STP-like lengthen-
a secondary pathway (Fig. 2). Studies show that ing of TE, resulting in post-hypoxic decline of
the biphasic STD component is abolished by respiratory frequency. The hypoxia-/CSN-
NMDA receptor blockade or lesioning of the induced STP and STD of inspiratory variables
Kölliker-Fuse nucleus in the pontine pneumotaxic resemble a biphasic neural integrator, whereas
region in vagotomized rat, but similar interven- the corresponding STP of the TE response in
tions have no effect on the monophasic STD com- combination with an initial rapid shortening of
ponent (Poon et al. 2000; Siniaia et al. 2000). The TE at the onset of hypoxia/CSN stimulation
latter is thought to be mediated by relay neurons in resembles a biphasic neural differentiator. These
the nucleus tractus solitarius, which demonstrate nonassociative learning effects provide low-pass
and high-pass filtering of the peripheral chemo- induce long-term facilitation of respiratory
receptor afferent input as discussed in the previ- (phrenic) motor output as well as hypoglossal
ous section. (particularly genioglossal) and laryngeal adduc-
Detailed analysis of the dynamics of the tor motor outputs that serve to increase upper
peripheral chemoreflex response to CSN stimu- airway patency (Fuller et al. 2000; McGuire
lation delivered repetitively breath-by-breath et al. 2007; Xing et al. 2013). Similarly, episodic
during either the inspiratory or expiratory phase modulation of vagally mediated mechanical
of the respiratory cycle reveals that peripheral feedback induces long-term facilitation of hypo-
chemoafferent signaling is relayed by parallel glossal motoneuron activity (Tadjalli et al. 2010).
integrator/differentiator central pathways that It is believed that such long-term adaptive mod-
independently modulate inspiratory amplitude, ulations of the vagal respiratory reflex and
TI and TE with varying time constants (Young peripheral chemoreflex are protective responses
et al. 2003). These include biphasic fast–slow to obstructive sleep apnea, which is characterized
compound integrator pathways that mediate the by repeated episodes of hypoxemia and disrup-
successive augmentation of inspiratory ampli- tion of vagal feedback.
tude and shortening of TI, and biphasic fast–slow
compound integrator/differentiator pathways that Associative Learning in Vagal–Mechanical
mediate the initial shortening and subsequent Reflex and Chemoreflex Adaptation
prolongation of TE. These biphasic integrator/ In associative learning, the learning process is
differentiator pathways also account for the dependent on the pairing of two coincident
off-transients of all these response components. inputs. Its significance in respiratory control has
Blockade of NMDA receptor slows down the traditionally been underappreciated in favor of
dynamics of the inspiratory-related neural inte- the Sherringtonian framework. However, recent
grators (Poon et al. 1999) and abolishes the slow evidence increasingly points to the existence of
expiratory-related neural differentiator response associative learning in respiratory control similar
(Poon et al. 2000). Lesioning of lateral to classical studies in Aplysia (Glanzman 1995)
parabrachial nucleus selectively attenuates the and cerebellum (Thompson 1988). Specifically,
rapid shortening of TE during hypoxia, but pre- vagal–mechanical input and central and periph-
sents no observable effects on the hypoxic eral chemoafferent inputs may indeed interact
response of other respiratory variables (Song associatively (rather than linearly) in modulating
and Poon 2009b). Similarly, lesioning of lateral the resultant ventilatory output.
parabrachial nucleus selectively attenuates the
shortening of TE and increase of inspiratory Associative Learning in Central–Peripheral
amplitude during hypercapnia, but presents no Chemoreceptor Interaction
observable effects on the hypercapnic response A generally accepted simplification in current
of TI (Song and Poon 2009a). mathematical models of respiratory control is
that feedback signals from central and peripheral
Long-Term Facilitation of Vagal–Mechanical chemoreceptors add linearly to determine the
Reflex and Peripheral Chemoreflex total “reflex” response (Cunningham et al. 1986;
In addition to short-term adaptive modulation of Gray 1946; Grodins et al. 1954), an assumption
vagal–mechanical reflex and peripheral that is supported by indirect evidence (e.g.,
chemoreflex with STP and STD in primary and (Mardimae et al. 2012)). However, a significant
secondary pathways following sustained afferent number of reports in the literature present
stimulations, long-term adaptive modulation of conflicting observations indicating that the inter-
these reflexes may ensue when the afferent acti- action of central and peripheral chemoreceptor
vations become repetitive. Specifically, repeated inputs may not be purely additive but could
exposures to alternating episodes of hypoxia and range from hypoadditive on one end of the spec-
hyperoxia in humans and animal models may trum to hyperadditive on the other. A recent study
in awake dogs (Blain et al. 2010) provides one of (Grodins et al. 1954). However, despite extensive
the strongest evidence for hyperadditive search over the past century, none of the hypothe-
peripheral–central chemoreceptor interaction, sized feedforward or feedback mechanisms has
by abruptly stimulating the central chemorecep- been demonstrated conclusively as the true “exer-
tor while holding the peripheral chemoreceptor cise stimulus” that is obligatory to exercise hyper-
activity constant at varying background levels. pnea (Poon et al. 2007). Much of the confusion lies
On the other hand, other studies (Day and Wilson in the fact that feedback mechanisms such as C
2007, 2009; Tin et al. 2012) demonstrate chemoreflex and skeletal muscle afferent feedback
hypoadditive interaction when the peripheral exert only transient effects on the exercise ventila-
chemoreceptor is stimulated while the central tory response but do not materially contribute to the
chemoreceptor activity is held constant at vary- steady-state exercise hyperpnea response (Poon
ing background levels. In particular, (Tin and Tin 2013). This persistent lack of conclusive
et al. 2012) showed that the responses in inspira- evidence of a feedforward exercise stimulus as
tory amplitude and TE (and hence fR and neural hypothesized under the Sherringtonian reflex
ventilation) to hypoxia varied inversely with paradigm raises serious questions regarding
the background CO2 level in spontaneously the validity of such oversimplified models for exer-
breathing (vagal–intact or vagotomized) and cise hyperpnea.
mechanically ventilated rats. The results show
a hypoadditive hypercapnic–hypoxic interaction Optimization Model of Exercise Hyperpnea and
in vivo that resembles the reported hypoadditive CO2 Breathing
central–peripheral chemoreceptor interaction in In contrast to Sherringtonian reflex, it has long
situ for these respiratory variables in the rat been conjectured that many respiratory control
(Day and Wilson 2007, 2009), regardless of dif- phenomena may boil down to the general idea
ferences in vagal feedback, body temperature and of optimality implied in natural selection (Priban
ventilation method. and Fincham 1965). Poon (1987) first introduced
Collectively, these reports point to a complex the following mathematical cost function to inte-
nonlinear interaction of peripheral and central grate the chemical and mechanical components
chemoreceptor inputs that contradicts the tradi- of breathing:
tional Sherringtonian paradigm. The mechanism
and site of such nonlinear interaction remain to be J ¼ J c þ J m ¼ ½aðPaCO2 bÞ2 þ lnW_ o
2
(4a)
investigated.
2 2
where W_ o ¼ V_ E = 1 V_ E =V_ max
Associative Learning in Mechanical–Chemical (4b)
Interaction
Another limitation of the classical chemoreflex/ The terms Jc, Jm in Eq. 4a represent the com-
mechanoreflex model under the Sherringtonian peting chemical and mechanical costs of breath-
paradigm is that it cannot explain the control of ing (a, b are parameters) in conformance to
ventilation during muscular exercise. In this case, Steven’s power law and Weber–Fechner law of
ventilation increases directly with increasing met- psychophysics, respectively. The term W_ o that
abolic demands (in terms of metabolic CO2 output, defines Jm is a measure of the work rate of breath-
V_ CO2 ) without any appreciable increases in PaCO2 ing as perceived by the respiratory controller
and arterial pH, i.e., without apparent activation of subject to the mechanical limitation of the respi-
the chemoreflex mechanism. This discrepancy has ratory system. This is given by the mechanical
led to the postulation of some exercise-induced plant equation (4b), where V_ max is the maximal
feedforward signal that drives respiratory ventila- ventilation that could be sustained
by the respira-
tory pump and the factor 1 V_ E =V_ max repre-
2
tion in proportion to metabolic demands with
near constancy of arterial CO2 and pH levels sents the pump’s neuromechanical efficiency.
a b
Exercise Load Exercise Load
Chemoreflex Chemoreflex
hyperpnea compensation hyperpnea compensation
Behavioral, Behavioral,
physiological & Feedforward Feedback Feedback Optimal sensorimotor integration physiological &
defense inputs defense inputs
Exercise Chemical Mechanical

Chemical Mechanical
stimulus stimulus stimulus
input input
Control of Breathing, Integration of Adaptive exercise stimuli. (b) The optimization model integrates
Reflexes, Fig. 3 Two views for respiratory control. (a) various afferent–efferent signals in a single model to char-
Classical reflex model assumes additive, reducible, and acterize the complex interactions among these signals
superposable characteristics of chemical, mechanical, and (Adapted from Poon et al. (2007))
The chemical cost Jc is a function of total chem- Equation 6 simulates the linear V_ E PaCO2
ical feedback. At rest and during moderate exer- relationship during CO2 inhalation and propor-
cise it is determined primarily by PaCO2 , which is tional V_ E V_ CO2 relationship during exercise at
constrained by the pulmonary gas exchange constant PaCO2 level. Thus, the optimization
(chemical plant) equation: model successfully reproduces normocapnic
exercise hyperpnea and hypercapnic CO2
863V_ CO2 response by simply adopting an optimal control
PaCO2 ¼ PICO2 þ (5)
VD law (Eq. 4a) in place of the Sherringtonian reflex
V_ E 1 control law, without the need of an explicit
VT
feedforward exercise stimulus. In more general
In Eq. 5, VD/VT is the physiological dead terms, the model addresses the important ques-
space-to-tidal volume ratio and is a measure of tion of how the respiratory controller balances
the inefficiency of pulmonary gas exchange, and multiple competing objectives in order to opti-
PICO2 is the partial pressure of CO2 in the inspired mize the work of breathing while striving to
air during CO2 inhalation ( PICO2 = 0 when maintain homeostasis of arterial CO2 and pH
breathing room air). levels in a unified framework (Fig. 3). The
The optimal total ventilation V_ E is one that form of the optimal solution (Eq. 6a & 6b)
minimizes J (Eq. 4a) subject to the pulmonary gas implies that the respiratory controller is respon-
exchange process (Eq. 5) and mechanical con- sive to an apparent metabolic CO2 load
straint (Eq. 4b), thus weighing the chemoafferent V_ CO2 =ð1 V D =V T Þ which tracks not only the
feedback against respiratory mechanical feed- actual metabolic CO2 load V_ CO2 but also the
back, giving: inefficiency of pulmonary gas exchange (VD/VT)
that adds to the overall challenge of CO2 elimi-
V_ E0 nation. This is in close agreement with the obser-
V_ E ¼ (6a) vation of heightened exercise ventilation at
1 þ V E0 =V_ max
_
normal PaCO2 in patients with increased physio-
logical VD/VT, such as in congestive heart failure.
V_ CO2
where V_ E0 ¼ 863a2 ðPaCO2 bÞ The controller’s responsiveness to an apparent
VD
1 metabolic load that is influenced by inefficiency
VT of pulmonary CO2 elimination caused by physi-
(6b) ological VD/VT is indicative of its ability of
cognition and perception (Poon and Tin 2013). interaction at increasing V_ E levels (Poon 1987).
This is analogous to temperature sensing, where In this event, the resultant V_ E V_ CO2 curve would
the perceived (or “real-feel”) temperature is demonstrate an apparent additive component
determined not only by the ambient temperature with a positive y-intercept. Hence, mechanical
but also by environmental factors (humidity, limitation due to a finite V_ max provides a
wind chill, etc.) that affect body heat dissipation. plausible explanation of the apparent additive
On the other hand, for any sizable external CO2–exercise interaction component found in C
dead space load added in series with the airway, experiments that include moderate to high exer-
CO2 elimination is also impaired by an elevated cise V_ E levels (Poon and Greene 1985) or in
series VD/VT with resultant increase in the appar- patients with pronounced ventilatory limitations
ent metabolic CO2 load like in congestive heart (Paoletti et al. 2011).
failure patients. However, because the increase in
VD/VT now comes with rebreathing of dead space Optimization Model of Mechanical Load
CO2, the controller may (at the beginning of each Compensation
inspiration) mistake the dead space load for an A remarkable property of the respiratory control-
airway CO2 load (at PICO2 PaCO2 ) that remains ler is its capability to compensate for ventilatory
until the dead space is cleared toward late loading and unloading at rest and during moder-
inspiration. This ambiguity may render the con- ate exercise, a phenomenon known as ventilatory
troller with an illusion of a virtual PICO2 (PÎCO2 ), load compensation. In normal subjects, this is
which amounts to a virtual airway CO2 load achieved by adopting an optimal respiratory pat-
(Poon and Tin 2013). Hence, the effect of exter- tern that minimizes the increases in mechanical
nal dead space loading is equivalent to work rate of breathing caused by the ventilatory
a combination of airway CO2 loading and parallel load, provided the latter is not mechanically lim-
dead space loading. Such illusive perception iting at the ventilatory levels of interest (i.e., at
caused by CO2 rebreathing provides a plausible V_ E << V_ max ) (Poon 1987, 1989a, b). This capa-
explanation of the hypercapnic effect of dead bility is also evidenced in patients with early
space loading as opposed to the eucapnic state stages of certain cardio respiratory diseases such
of increased physiological VD/VT in congestive as pulmonary fibrosis and emphysema. The opti-
heart failure. mization model provides a conceptual framework
In addition, Eq. 6b predicts a synergetic for understanding this behavior. To model the
CO2–exercise interaction such that the slope of integrative control of V_ E and respiratory pattern,
the V_ E V_ CO2 relationship is increased at Eq. 4a has been extended by expressing W_ o
increasing PaCO2 levels while inhaling CO2, as explicitly in terms of the isometric respiratory
demonstrated experimentally (Poon et al. 2007; driving pressure P(t) (Poon et al. 1992). The
Poon and Tin 2013). In contrast, the traditional mechanical plant in this case is defined by the
Sherringtonian reflex model would predict an following equation of motion:
additive CO2-exercise interaction with a parallel
shift of the V_ E V_ CO2 curve by CO2 inhalation PðtÞ ¼ V_ ðtÞRrs þ V ðtÞErs (7)
instead (Poon 1987). On the other hand, Eq. 6a
shows that the magnitude of CO2–exercise inter- whereby all ventilatory variables can be derived
action is also influenced by mechanical con- successively from the P(t) waveform as follows:
straints. The term V_ max in Eqs. 4 and 6a
represents a hypothetical maximum ventilation PðtÞ ¼ V_ ðtÞ, V ðtÞ ! V T , T I , T E ! V_ E (8)
that can be sustained by the respiratory muscles,
such that the effect of mechanical limitation where Rrs, Ers are respectively the total (extrinsic
intensifies as V_ E approaches V_ max resulting in and intrinsic) respiratory resistance and
a less than multiplicative CO2–exercise elastance; V_ ðtÞ, V(t) are instantaneous respiratory
airflow and volume; and TI and TE are inspiratory “internal model.” It is now well recognized that
and expiratory durations. The resultant V_ E is forward and inverse internal models (models of the
given by the variables VT, TI and TE. This inte- external environment and its inverse) are wide-
grated model captures both the optimal ventila- spread in many types of sensorimotor integration
tory response characteristics of Eq. 6 and the in the mammalian brain (Davidson and Wolpert
corresponding optimal respiratory pattern. The 2005; Tin and Poon 2005).
strength of ventilatory load compensation is The optimal solution (Eq. 6) suggests that the
dependent on the type and magnitude of the ven- respiratory controller may be tracking an appar-
tilatory load relative to those of the background ent metabolic CO2 load V_ CO2 =ð1 V D =V T Þ
ventilatory stimulus. The parameters Rrs and Ers instead of the actual metabolic CO2 load V_ CO2 ,
can explicitly model the effects of resistive, elas- perhaps through continuous algorithmic identifi-
tic, and threshold loading, which have been stud- cation of the causal relationship of the controller
ied singly and in combination, as well as under output and resultant chemical–mechanical affer-
the influence of chemical stimulus and exercise ent feedbacks. The projected estimate of this
(Poon 1989a, b; Sidney and Poon 1995). Com- apparent metabolic CO2 load may take the form
puter simulations with the optimization model of some controller drive signals (indirect adaptive
have proved to be consistent with the experimen- control) or modifiable gain parameters (direct
tal observations (Poon et al. 1992). In all cases, adaptive control), like in engineering adaptive
ventilatory loading provokes compensatory control theory (Fig. 4). Either way, the “exercise
inspiratory, postinspiratory, and expiratory stimulus” may amount to some “mental percept”
motor responses that fully or partially restore signal emergent from optimal sensorimotor inte-
V_ E to the control level in normal or disease states gration rather than an explicit feedforward signal
or when ventilation is stimulated by chemical or to the controller. Similar internal model sensori-
exercise inputs. motor integration architecture may also underlie
the optimal adaptation of the respiratory pattern
to changes in ventilatory mechanical loads.
Internal Model Paradigm in Respiratory Control
Traditional Sherringtonian reflex models of Hebbian Feedback Covariance Learning and
homeostatic regulation rely mainly on feedback, Direct Internal Model Control
feedforward, and set-point operations. These Hebbian synaptic plasticity was first postulated
classical control engineering schemes have lim- by D.O. Hebb over 50 years ago as a mechanism
ited predictive power for complex behaviors, of learning and memory (Hebb 1949). In its orig-
such as integrative physiological control at inal form, Hebb suggested that a synaptic con-
multiple levels of organization, and are nonrobust nection can be strengthened if pre- and
to disturbances (Poon 1996). In contrast, the postsynaptic activities coincided during a short
optimization model suggests that the respiratory time interval. To circumvent limitations to the
controller may be self-adapting, continuously classical Hebbian model such as instability and
maintaining optimality in the presence of saturation and the lack of reinforcement feedback
unknown disturbances. dynamics in the learning mechanism, Young and
In this context, the homeostatic regulation Poon (1998) have introduced a Hebbian feedback
problem in physiology is closely related to the covariance adaptive control algorithm as a novel
internal model principle in modern control systems paradigm for direct internal model adaptation and
engineering, which states that a feedback regulator reinforcement learning (operant or instrumental
under external disturbances may maintain regula- conditioning). Unlike classical adaptive control
tion and stability provided a suitably reduplicated laws which operate deterministically, the feed-
model of the disturbance signal is adapted in the back covariance adaptive control law is driven
feedback path (Francis and Wonham 1975). Such by the intrinsic fluctuations of the control signal
a reduplicated model of the disturbance is called (such fluctuations are ubiquitous in physiological
a Indirect Adaptive Control b Direct Adaptive Control
Ventilatory Controller Ventilatory Controller

P(t)
•
PaCO2 VE
Internal model W
• C
VCO2
estimate •
•
VCO2 estimate P(t)
PaCO2 VE
x W
c
Efference
Controller copy
• V(t)
Ventilatory VE P(t) Mechanical Chemical
E I
Controller plant plant
Vagal
feedback
PaCO2, PaO2, pH
Chemoreceptors
Control of Breathing, Integration of Adaptive ventilatory drive. (b) In direct adaptive control, the esti-
Reflexes, Fig. 4 Hypothetical internal model structures mate is directly incorporated in the controller as a variable
for self-tuning adaptive control of respiration. (a) In indi- feedback gain. (c) Two-tier respiratory control system
rect adaptive control, the metabolic level is continuously structure with (central and peripheral) chemoreceptor
estimated by the controller based on an internal model of afferent feedback driving an adaptive ventilatory control-
the causal relationship between the respiratory motor out- ler and vagal volume-related feedback driving an adaptive
put and chemical feedback. The resultant neuronal esti- central pattern generator signals (Adapted from Poon
mate provides an indirect “feedforward” signal to the et al. (2007))
feedback controller (with fixed gain W) to generate the
signals) and resultant covariation of the feedback and efferent signals, respectively, about their
signal that is routed back to the controller through corresponding mean values. The adaptation law
the plant dynamics. In its simplest form, the con- states that the controller gain (or synaptic weight),
troller gain, w, that determines the optimal w, will be potentiated if du and dy are positively
input–output relationship is computed by corre- correlated and will be weakened otherwise. The
lating the corresponding intrinsic fluctuations of parameter w thus encodes an inverse internal
the control (efferent) signal u and feedback model that may track any disturbances in the
(afferent) signal y: plant dynamics, such as changes in the metabolic
V_ CO2 in the chemical plant of the respiratory sys-
dw tem. Young and Poon have shown that, with suit-
¼ kðdy duÞ (9)
dt able modifications of the adaptation rule (Eq. 9) to
account for the dynamics and delays in the feed-
where k is the adaptation constant and du and dy back loop, this feedback covariance-based internal
are the temporal variations of the afferent model paradigm allows robust direct adaptive
control of a general class of linear and nonlinear Dutschmann M, Mörschel M, Rybak IA, Dick TE
dynamical systems with stability and convergence (2009) Learning to breathe: control of the
inspiratory-expiratory phase transition shifts from
guaranteed by Lyapunov theory. sensory- to central-dominated during postnatal devel-
In summary, multiple lines of evidence have opment in rats. J Physiol 587(Pt 20):4931–4948.
strongly suggested that mammalian respiratory doi:10.1113/jphysiol.2009.174599
control possesses rich adaptive signal processing Francis B, Wonham W (1975) The internal model princi-
ple for linear multivariable regulators. Appl Math
capabilities in the form of both nonassociative Optim 2(2):170–194. Retrieved from http://link.
learning and associative learning, which have springer.com/article/10.1007/BF01447855
been ignored in the traditional Sherringtonian Fuller D, Bach K, Baker T, Kinkead R, Mitchell G (2000)
framework. New paradigms of adaptive respiratory Long term facilitation of phrenic motor output. Respir
Physiol Neurobiol 121:135–146. Retrieved from
control will potentially open up new opportunities http://www.sciencedirect.com/science/article/pii/S0034
for more in-depth understanding of homeostatic 568700001249
regulation and sensorimotor control in general. Glanzman D (1995) The cellular basis of classical condi-
tioning in Aplysia californica—it’s less simple than
you think. Trends Neurosci 18(1):30–36. Retrieved
Cross-References from http://www.sciencedirect.com/science/article/
pii/016622369593947V
Gray J (1946) The multiple factor theory of the control of
▶ Computational Models of Mammalian respiratory ventilation. Science 103:739–744.
Respiratory CPG Retrieved from http://www.sciencemag.org/content/
▶ Hebbian Learning 103/2687/739.short
Grodins F, Gray J, Schroeder K, Norins A, Jones R (1954)
▶ Learning Rules: Overview Respiratory responses to CO2 inhalation. A theoretical
▶ Pre-Botzinger Complex Rhythm Generation study of a nonlinear biological regulator. J Appl
▶ Reward-Based Learning, Model-Based and Physiol 7:283–308. Retrieved from http://jap.physiol-
Model-Free ogy.org/content/7/3/283.short
Groves P, Thompson R (1970) Habituation: a dual-
▶ Short-Term Plasticity, Biophysical Models process theory. Psychol Rev 77:419–450. Retrieved
from http://doi.apa.org/psycinfo/1971-02046-001
Hawkins R, Cohen T, Greene W, Kandel E (1998) Rela-
References tionships between dishabituation, sensitization, and
inhibition of the gill- and siphon-withdrawal reflex in
Blain GM, Smith CA, Henderson KS, Dempsey JA Aplysia californica: effects of response measure, test
(2010) Peripheral chemoreceptors determine the respi- time, and training stimulus. Behav Neurosci
ratory sensitivity of central chemoreceptors to CO2. 112(1):24–38
J Physiol 588(Pt 13):2455–2471. doi:10.1113/ Hawkins R, Clark G, Kandel E (2006) Operant condition-
jphysiol.2010.187211 ing of gill withdrawal in Aplysia. J Neurosci
Cunningham D, Robbins P, Wolff C (1986) Integration of 26:2443–2448. Retrieved from http://www.jneurosci.
respiratory responses to changes in alveolar partial pres- org/content/26/9/2443.short
sures of CO2 and O2 and in arterial pH. In: Fishman A, Hebb DO (1949) The organization of behavior:
Cherniack NS, Widdicombe J (eds) Handbook of phys- a neuropsychological theory, vol 39, Theology today.
iology, Sect. 3: the respiratory system. American Phys- Wiley, New York. doi:10.1177/004057368303900411
iological Society, Bethesda, pp 475–528 MacDonald SM, Song G, Poon C-S (2007) Nonassociative
Davidson P, Wolpert D (2005) Widespread access to learning promotes respiratory entrainment to mechan-
predictive models in the motor system: a ical ventilation. PLoS One 2(9):e865. doi:10.1371/
short review. J Neural Eng 2(3):S313–S319. Retrieved journal.pone.0000865
from http://iopscience.iop.org/1741-2552/2/3/S11 MacDonald SM, Tin C, Song G, Poon CS (2009)
Day TA, Wilson RJA (2007) Brainstem PCO2 modulates Use-dependent learning and memory of the Hering-
phrenic responses to specific carotid body hypoxia in Breuer inflation reflex in rats. Exp Physiol
an in situ dual perfused rat preparation. J Physiol 94(2):269–278. doi:expphysiol.2008.045344 [pii]
578(3):843–857. doi:10.1113/jphysiol.2006.119594 10.1113/expphysiol.2008.045344
Day TA, Wilson RJA (2009) A negative interaction Mardimae A, Balaban D, Machina M, Battisti-
between brainstem and peripheral respiratory chemo- Charbonney A, Han J, Katznelson R, Minkovich LL,
receptors modulates peripheral chemoreflex magni- Fedorko L, Murphy PM, Wasowicz M, Naughton F,
tude. J Physiol 587(Pt 4):883–896. doi:10.1113/ Meineri M, Fisher JA, Duffin J (2012) The interaction
jphysiol.2008.160689 of carbon dioxide and hypoxia in the control of
cerebral blood flow. Pflugers Arch Eur J Physiol Poon CS, Siniaia MS, Young DL, Eldridge FL
464:345–351. Retrieved from http://link.springer. (1999) Short-term potentiation of carotid chemoreflex:
com/article/10.1007/s00424-012-1148-1 an NMDAR-dependent neural integrator. Neuroreport
McGuire M, MacDonald S, Song G, Poon CS 10(11):2261–2265. Retrieved from http://www.ncbi.
(2007) Phrenic long-term facilitation is robust to nlm.nih.gov/pubmed/10439445
hypercapnia and hypocapnia but not hyperventilatory Poon CS, Young DL, Siniaia MS (2000) High-pass filter-
hypotension under PEEP. Respir Physiol Neurobiol ing of carotid-vagal influences on expiration in rat: role
158:107–111. Retrieved from http://www. of N-methyl-D-aspartate receptors. Neurosci Lett
sciencedirect.com/science/article/pii/S156990487000493 284(1–2):5–8. Retrieved from http://www.ncbi.nlm.
C
Paoletti P, De Filippis F, Fraioli F, Cinquanta A, Valli G, nih.gov/pubmed/10771148
Laveneziana P, Vaccaro F, Martolini D, Palange Poon C, Tin C, Yu Y (2007) Homeostasis of exercise
P (2011) Cardiopulmonary exercise testing (CPET) hyperpnea and optimal sensorimotor integration: the
in pulmonary emphysema. Respir Physiol Neurobiol internal model paradigm. Respir Physiol Neurobiol
2. doi:10.1016/j.resp.2011.07.013 159(1):1–20. Retrieved from http://www.sciencedirect.
Poon CS (1987) Ventilatory control in hypercapnia and com/science/article/pii/S1569904807000857
exercise: optimization hypothesis. J Appl Physiol Priban I, Fincham W (1965) Self-adaptive control and the
62(6):2447. Retrieved from http://scholar.google.com/ respiratory system. Nature 208(5008):339–343.
scholar?hl=en&btnG= Search&q=intitle:Ventilatory Retrieved from http://adsabs.harvard.edu/abs/
+control+in+hypercapnia+and+exercise:+optimization 1965Natur.208.339P
+hypothesis#0 Sherrington C (1906) The integrative action of the nervous
Poon CS (1989a) Effects of inspiratory elastic load on system. Yale University Press, New Haven. Retrieved
respiratory control in hypercapnia and exercise. from http://www.itiroyal.com/Irradiation and rein-
J Appl Physiol (Bethesda, 1985) 66(5):2400–2406. forcement, Sir Charles S. Sherrington 1906.pdf
Retrieved from http://www.ncbi.nlm.nih.gov/ Sidney D, Poon C (1995) Ventilatory responses to dead
pubmed/2501282 space and CO2 breathing under inspiratory resistive
Poon CS (1989b) Effects of inspiratory resistive load on load. J Appl Physiol 78(2):555–561. Retrieved from
respiratory control in hypercapnia and exercise. J Appl http://jap.physiology.org/content/78/2/555.short
Physiol (Bethesda, 1985) 66(5):2391–2399. Retrieved Siniaia MS, Young DL, Poon CS (2000) Habituation and
from http://www.ncbi.nlm.nih.gov/pubmed/2501281 desensitization of the Hering-Breuer reflex in rat.
Poon CS (1996) Self-tuning optimal regulation of respira- J Physiol 523(Pt 2):479–491. Retrieved from http://
tory motor output by Hebbian covariance learning. www.pubmedcentral.nih.gov/articlerender.fcgi?artid=
Neural Netw Off J Int Neural Netw Soc 2269814&tool=pmcentrez&rendertype=abstract
9(8):1367–1383. Retrieved from http://www.ncbi. Song G, Poon C (2009a) Lateral parabrachial nucleus
nlm.nih.gov/pubmed/12662540 mediates shortening of expiration and increase of
Poon CS, Greene JG (1985) Control of exercise hyperpnea inspiratory drive during hypercapnia. Respir Physiol
during hypercapnia in humans. J Appl Physiol Neurobiol 165(1):9–12. doi:10.1016/j.
(Bethesda, 1985) 59(3):792–797. Retrieved from resp.2008.10.009
http://www.ncbi.nlm.nih.gov/pubmed/4055568 Song G, Poon CS (2009b) Lateral parabrachial nucleus
Poon C, Schmid S (2012) Nonassociative learning. In: mediates shortening of expiration during hypoxia.
Seel N (ed) Encyclopedia of the sciences of learning. Respir Physiol Neurobiol 165(1):1–8. doi:10.1016/j.
Springer, New York, pp 2475–2477 resp.2008.10.007.Lateral
Poon C, Siniaia M (2000) Plasticity of cardiorespiratory Tadjalli A, Duffin J, Peever J (2010) Identification of
neural processing: classification and computational a novel form of noradrenergic-dependent respiratory
functions. Respir Physiol 122(2–3):83–109. Retrieved motor plasticity triggered by vagal feedback.
from http://www.sciencedirect.com/science/article/ J Neurosci Off J Soc Neurosci 30(50):16886–16895.
pii/S0034568700001523 doi:10.1523/JNEUROSCI.3394-10.2010
Poon C, Tin C (2013) Mechanism of augmented exercise Thompson R (1988) The neural basis of basic associative
hyperpnea in chronic heart failure and dead space learning of discrete behavioral responses. Trends
loading. Respir Physiol Neurobiol 186(1):114–130. Neurosci 11(4):152–155. Retrieved from http://www.
Retrieved from http://www.sciencedirect.com/sci- sciencedirect.com/science/article/pii/0166223688901415
ence/article/pii/S1569904812003795 Thompson R, Steinmetz J (2009) The role of the cerebel-
Poon C-S, Young DL (2006) Nonassociative learning as lum in classical conditioning of discrete behavioral
gated neural integrator and differentiator in stimulus- responses. Neuroscience 162:732–755. Retrieved
response pathways. Behav Brain Funct BBF 2:29. from http://www.sciencedirect.com/science/article/
doi:10.1186/1744-9081-2-29 pii/S0306452209000967
Poon CS, Lin SL, Knudson OB (1992) Optimization char- Tin C, Poon C-S (2005) Internal models in sensorimotor
acter of inspiratory neural drive. J Appl Physiol integration: perspectives from adaptive control theory.
(Bethesda, 1985) 72(5):2005–2017. Retrieved from J Neural Eng 2(3):S147–S163. doi:10.1088/1741-
http://www.ncbi.nlm.nih.gov/pubmed/1601812 2560/2/3/S01
C 834 Control of Locomotion and Scratching in Turtles
Tin C, Song G, Poon C-S (2012) Hypercapnia attenuates swimming and for three forms of scratching,
inspiratory amplitude and expiratory time responsive- a rhythmic behavior in which the same-side
ness to hypoxia in vagotomized and vagal-intact rats.
Respir Physiol Neurobiol 181(1):79–87. doi:10.1016/ (ipsilateral) hind limb repeatedly rubs against
j.resp.2012.01.008 a specific location on the body surface that has
Xing T, Fong A, Bautista T, Pilowsky P (2013) Acute been irritated or tickled. Stimulation in
intermittent hypoxia induced neural plasticity in respi- a transition zone, at the border of the receptive
ratory motor control. Clin Exp Pharmacol Physiol .
Retrieved from http://onlinelibrary.wiley.com/doi/ fields for two forms of scratching, or simulta-
10.1111/1440-1681.12129/abstract neous two-site stimulation for two forms of
Young D, Poon C (1998) Hebbian covariance learning. scratching, or for one form of scratching plus
A nexus for respiratory variability, memory, and forward swimming, can evoke motor pattern
optimization? Adv Exp Med Biol 450:73–83.
Retrieved from http://www.ncbi.nlm.nih.gov/ blends. Blends may be switches, in which con-
pubmed/10026966 secutive cycles alternate between two different
Young DL, Eldridge FL, Poon C-S (2003) Integration- motor patterns, or hybrids, in which each cycle
differentiation and gating of carotid afferent traffic exhibits characteristics of both motor patterns.
that shapes the respiratory pattern. J Appl Physiol
(Bethesda, 1985) 94(3):1213–1229. doi:10.1152/ The ability of the turtle spinal cord to produce
japplphysiol.00639.2002 blends suggests that the corresponding central
Zhou Z, Champagnat J, Poon CS (1997) Phasic and long- pattern generators are partly shared or that there
term depression in brainstem nucleus tractus solitarius are strong interactions between components
neurons: differing roles of AMPA receptor desensiti-
zation. J Neurosci Off J Soc Neurosci of each. The spinal cord on the opposite
17(14):5349–5356. Retrieved from http://www.ncbi. (contralateral) side to the scratching limb contrib-
nlm.nih.gov/pubmed/9204919 utes to generating scratching motor patterns, spe-
cifically to the hip-extensor phase of scratching;
this circuitry may be part of a bilateral network
that contributes to flexor-extensor and left-right
Control of Locomotion and coordination for both swimming and scratching.
Scratching in Turtles Many spinal interneurons, which are neurons
contained entirely within the spinal cord, are
Ari Berkowitz multifunctional, meaning that they generate
Department of Biology, Cellular & Behavioral action potentials rhythmically during forward
Neurobiology Graduate Program, University of swimming and all three forms of scratching,
Oklahoma, Norman, OK, USA often bilaterally, and are often activated during
limb withdrawal (flexion reflex) as well.
Multifunctional interneurons may be components
Definition of multiple central pattern generators. Other spi-
nal interneurons, however, are behaviorally spe-
The spinal cord contains central pattern genera- cialized: scratching-specialized neurons and
tors, which are networks of neurons that can flexion reflex-selective neurons are not active,
generate coordinated patterns of activity in and often are inhibited, during other kinds of
motor neurons (motor patterns) in the absence motor patterns, and may play a special role in
of sensory feedback. The turtle spinal cord con- selecting or generating one type of motor pattern.
tains central pattern generators for several kinds
of limb movements and is convenient to study
physiologically because turtles, being diving ani- Detailed Description
mals, have evolved mechanisms to keep tissue
healthy under low-oxygen (hypoxic) conditions. Why Study Locomotion and Scratching in
Turtle locomotion includes swimming and Turtles?
stepping. Spinal cord central pattern generators Turtles have been used as a model system to
have been demonstrated in turtles for forward study the control of locomotion and scratching
Control of Locomotion and Scratching in Turtles 835 C
Control of Locomotion and Scratching in Turtles, limb withdrawal (flexion reflex) motor patterns it gener-
Fig. 1 Schematic illustration of the turtle spinal cord ates with different kinds of mechanical or electrical
preparation and the forward swimming, scratching, and stimulation
in vertebrates since the 1970s. One major reason spinal cord mechanisms have been studied for
for their use is that turtles, being diving animals, forward swimming.
have evolved cellular and molecular mechanisms
to maintain tissue health under conditions of There Are Three Forms of Turtle Scratching
reduced oxygen (hypoxia) (Lutz and Milton In addition, there are three forms of scratching
2004). Such conditions often occur during phys- performed by each hind limb in response to irri-
iological experiments on reduced or in vitro ani- tation or tickling of the skin or shell (which has
mal preparations, which are often required to sensory innervation). In each form of scratching,
study neuronal mechanisms. Thus, researchers the hind limb rubs against a distinct region of the
are often able to acquire useful physiological body surface (Fig. 1). A different part of the hind
data from healthy turtle tissue for about three limb and a different knee-hip synergy are used in
times as long as they could from comparable each form, to target the different regions of the
mammalian tissue (Hounsgaard and Nicholson body surface (Mortin et al. 1985). The different
1990). synergies required to reach each region are anal-
It is also relatively easy to evoke both loco- ogous to the different ways humans use to scratch
motion and scratching in turtles, even in reduced the upper and lower back. In both cases, biome-
preparations and without using exogenous drugs chanical constraints necessitate the use of differ-
(Stein 2005). The forms of locomotion that have ent limb muscle synergies. Also in both cases,
been studied behaviorally include forward swim- there is a narrow transition zone at the borders of
ming, backward swimming (or backpaddling), the receptive fields for different forms of
and forward stepping. Of these, brainstem and scratching. An irritating stimulus within
a transition zone can evoke either of the two simultaneous stimulation of sites in two scratch
forms of scratching or sometimes a blend of the receptive fields (Mortin et al. 1985; Stein
two forms. Blends can be switches, in which the et al. 1986). Blends between swimming and
limb changes forms on a cycle-by-cycle basis, or scratching have also been seen (Earhart and
hybrids, in which each cycle includes elements of Stein 2000). Thus, coordinated hind limb move-
the two forms. Both switches and hybrids can ment blends can be produced without the brain.
result in successful rubbing. The receptive fields The several forms of swimming and
for rostral and pocket scratching are innervated scratching each involve rhythmic contractions
by contiguous spinal cord segments (i.e., their of the same hind limb muscles (Lennard and
dermatomes are adjacent and partly overlapping) Stein 1977; Robertson et al. 1985). Nonetheless,
(Mortin and Stein 1990). The spinal cord seg- they can be distinguished from one another based
ments that provide sensory innervation of the on a set of criteria that involve the amplitudes and
caudal scratch receptive field, however, are timing (i.e., phasing) of muscle contractions.
more caudal and are not contiguous with the During scratching, knee extension is the major
segment innervating the pocket scratch receptive determinant of force generated by the limb
field, being separated by segments that provide against the body surface. A burst of knee-
sensory innervation of the hind limb instead. extensor muscle activity occurs during the latter
portion of each hip-flexor burst in rostral
The Turtle Spinal Cord Contains Central scratching, during each hip-extensor burst in
Pattern Generators for Swimming and pocket scratching, and toward the end of the
Scratching hip-extensor burst and the start of the next flexor
The turtle spinal cord can produce the different burst in caudal scratching. Forward swimming
forms of locomotion and scratching even when shows a similar knee-hip synergy to rostral
all input from the brain is removed. Forward scratching, but the knee-extensor and hip-
swimming movements of the opposite hind limb flexor bursts are weaker and briefer, and the
can be evoked by electrical stimulation in hip-extensor bursts are stronger and longer, in
a region of the spinal cord white matter (the forward swimming as compared to rostral
lateral funiculus) in post-cervical spinal cord seg- scratching. When forward swim-evoking and ros-
ments (Lennard and Stein 1977). This stimulation tral scratch-evoking stimuli are delivered simul-
is thought to generate action potentials in the taneously, the animal produces an adaptive blend
descending axons of reticulospinal neurons of the two (a hybrid), in which the amplitude of
(which have been cut off from their cell bodies). hip extension is similar to forward swimming and
Selective lesion of the white matter in this region the amplitudes of knee extension and hip flexion
reduces or eliminates voluntary swimming in are similar to rostral scratching. The single crite-
otherwise intact turtles. rion of the timing of the knee-extensor burst
It has been found in limbed vertebrates gener- within the hip cycle reliably distinguishes rostral
ally that locomotion and scratching can be pro- scratching from pocket scratching or caudal
duced without the brain. Moreover, vertebrates scratching, but does not reliably distinguish
without limbs can produce axial locomotion pocket scratching from caudal scratching or ros-
without the brain. Insects can also produce loco- tral scratching from forward swimming.
motion and scratching without the brain (via the The turtle spinal cord can generate the basic
thoracic ganglia). Thus, there is a remarkable patterns of motor neuron activity (i.e., motor pat-
conservation of the capacity to produce these terns) that underlie forward swimming and the
rhythmic movements using the central nervous three forms of scratching even without
system below the level of the brain. In turtles, at movement-related sensory feedback from the
least, the spinal cord is also able to generate limbs (Robertson et al. 1985; Juranek and Currie
blends of two forms of scratching appropriately 2000; Fig. 1). This can be demonstrated by cut-
during stimulation of a transition zone or during ting all of the dorsal roots that carry sensory input
from the limbs to the spinal cord, but it is usually flexor and extensor units to generate a rhythm
demonstrated by chemically blocking acetylcho- (Stein 2008). A normally occurring variation in
line receptors at neuromuscular junctions system- rostral scratching (and sometimes other forms of
ically (i.e., the animal is immobilized). The same scratching) is missing the hip-extensor phase
kind of electrical stimulation in the lateral funic- (i.e., missing both the hip-extensor bursts and
ulus that evokes swimming movements in the the quiescent periods between successive
moving animal evokes forward swimming hip-flexor bursts). This is called hip-extensor C
motor patterns in the immobilized animal. The deletion scratching. Hip-flexor motor neurons
same kind of mechanical stimulation of the shell continue to burst rhythmically during
or skin that evokes scratching movements in the hip-extensor deletion scratching. This finding
moving animal evokes each form of scratching suggests that a hip-flexor rhythm can occur in
motor pattern in the immobilized animal. The the absence of a hip-extensor rhythm, in contrast
patterns of electrical activity in motor nerves to the half-center hypothesis. Knee-extensor
that innervate particular hind limb muscles are bursts can occur normally and can alternate with
essentially identical in immobilized animals to knee-flexor bursts during hip-extensor deletion
the patterns of electrical activity in those muscles rostral scratching, which suggests that the turtle
during actual swimming and scratching move- spinal cord contains sufficient neural circuitry to
ments. Thus, the spinal cord contains central pat- generate rhythmic activity in knee-extensor
tern generators for forward swimming and each motor neurons without rhythmic activity in
form of scratching. hip-extensor motor neurons (Stein and Daniels-
The spinal cord contains sufficient neural cir- McQueen 2004). This finding is inconsistent with
cuitry not only to generate the motor patterns for a simple half-center hypothesis in which all limb
forward swimming and each form of scratching extensor motor neurons (i.e., both hip and knee)
but also to generate adaptive blends of two such are controlled together. The existence of different
motor patterns (Robertson et al. 1985; Stein knee-hip synergies for the different forms of
et al. 1986; Juranek and Currie 2000). Stimula- scratching is also inconsistent with such
tion in a scratch transition zone (rostral pocket or a simple half-center hypothesis.
pocket caudal) or two-site stimulation (e.g., ros- The capacity to produce rhythmic, hind limb
tral + caudal) can evoke appropriate blends of motor patterns is distributed across several seg-
two forms of scratching motor patterns in the ments of the spinal cord, including all five seg-
immobilized preparation, including both ments of the hind limb enlargement and the spinal
switches and hybrids (Robertson et al. 1985; cord segments just rostral to the enlargement
Stein et al. 1986). Blends between forward swim- (Mortin and Stein 1989). Nonetheless, the more
ming and a form of scratching have also been rostral segments of the enlargement and the
described in the immobilized preparation immediately pre-enlargement segments are
(Juranek and Currie 2000). Thus, the capacity to more important for generating these rhythms
produce pure-form locomotion and scratching, as than the caudal enlargement segments are. Spinal
well as coordinated and adaptive blends, is cord segments still further caudal are not neces-
programmed into the spinal cord, in turtles at sary to generate these rhythmic motor patterns.
least. Similar conclusions have been reached for chick
Several features of scratching motor patterns embryo rhythmic activity (Ho and O’Donovan
are consistent with the unit-burst generator 1993), rodent locomotor-like activity (Cazalets
hypothesis of Sten Grillner (Grillner 1981) that et al. 1995; Kjaerulff and Kiehn 1996; Cowley
each joint is controlled by a pair of rhythmogenic and Schmidt 1997), and cat scratching motor
modules and inconsistent with the “half-center” patterns (Berkinblit et al. 1978; Deliagina
hypothesis originally proposed by T. Graham et al. 1983). Thus, there appears to be a
Brown (Brown 1911), at least in its simplest conserved rostral weighting of spinal cord central
form, which requires mutual inhibition between pattern generation across limbed vertebrates.
The importance of the rostral part of the hind limb slows rostral scratching motor patterns and
enlargement for motor pattern generation is not blocking glycine receptors speeds them up
directly related to the locations of motor neurons; (Currie and Lee 1997).
hip-flexor and knee-extensor motor neurons are
located in the rostral segments of the enlarge- Motor Neurons During Locomotion and
ment, but hip-extensor motor neurons are located Scratching
in caudal segments of the enlargement, in turtles Hind limb motor neurons may contribute actively
(Ruigrok and Crowe 1984) as in mammals to the amplitude and perhaps the timing of
(Romanes 1964). turtle scratching and swimming rhythms. Motor
It is possible to evoke some of these hind limb neurons receive alternating but temporally
motor patterns in vitro. The turtle spinal cord with overlapping excitatory, depolarizing inputs and
a region of the rostral shell still attached to it via inhibitory, chloride-dependent, hyperpolarizing
peripheral nerves can be used to evoke rostral inputs during scratching motor patterns in vivo
scratching motor patterns using natural stimula- (Robertson and Stein 1988; Stein 2010). Intrinsic
tion of the shell in vitro (Keifer and Stein 1983; properties of motor neurons can be studied in
Alaburda and Hounsgaard 2003). Also, the turtle more detail in slice preparations, which can be
spinal cord with a particular cutaneous nerve (the made especially thick for turtle spinal cord,
ventral posterior pocket) attached can be used to because of its unusual resistance to hypoxia
evoke pocket scratching motor patterns via nerve (Hounsgaard and Nicholson 1990). Some motor
electrical stimulation in vitro (Currie and Lee neuron dendrites can produce calcium-dependent
1996). action potentials (Hounsgaard and Mintz 1988).
Motor neurons also show bistability or plateau
Neurotransmitters and Ion Channels potentials, primarily in the dendrites, meaning
Glutamate and glycine have been shown to play that their membrane potentials are stable at two
important roles in scratching motor patterns. Bath different levels, one of which (the resting poten-
application of glutamate agonists (NMDA and tial) is below action potential threshold and the
AMPA) to the turtle spinal cord in vivo and other of which is above it (Hounsgaard and Mintz
in vitro can evoke rhythmic activity of motor 1988). This bistability might contribute to the
nerves that is similar to scratching motor patterns strength and duration of each motor neuron
(Currie 1999). Scratch sensory stimulation can burst during scratching.
reset these rhythms, indicating interaction, at The bistability depends on L-type (persistent)
least, between the networks that produce voltage-dependent calcium channels, especially
glutamatergic rhythms and those that produce CaV1.3 channels (Hounsgaard and Mintz 1988;
scratching. Blocking NMDA receptors reduces Alaburda and Hounsgaard 2003; Simon
the rate of the scratching rhythm and the temporal et al. 2003), as well as calcium release from the
summation of sensory inputs that evoke endoplasmic reticulum via ryanodine receptors
scratching motor patterns for rostral scratching (Mejia-Gervacio et al. 2004). There are also
in vivo (Currie and Stein 1992) and for pocket calcium-dependent nonspecific cation channels
scratching in vitro (Currie and Lee 1996). Some in turtle motor neurons, but they are not required
individual spinal cord interneurons (i.e., neurons for plateau potentials (Perrier and Hounsgaard
contained entirely within the central nervous sys- 1999). The plateau potentials can be affected by
tem) are similarly active for many seconds fol- a number of neuromodulators (Delgado-Lezama
lowing a single scratch stimulus and may et al. 1997; Svirskis and Hounsgaard 1998). The
contribute to temporal summation during plateau potentials are augmented by serotonin
scratching (Currie and Stein 1990). Such inter- (especially via 5-HT2-type receptors)
neurons can also have their multisecond activity (Hounsgaard and Kiehn 1989; Perrier and
reduced by blocking NMDA receptors (Currie Hounsgaard 2003), group I metabotropic gluta-
and Stein 1992). Glycine, on the other hand, mate receptors, and muscarinic acetylcholine
receptors. Plateau potentials are generally have axon projections that terminate in the ven-
increased by neuromodulators that close potas- tral horn (Berkowitz 2005, 2008; Berkowitz
sium channels (Hounsgaard and Mintz 1988). et al. 2006), consistent with their contributing to
A compartmental model of turtle motor neurons motor output during all of these motor patterns,
also suggests that neuromodulators close potas- although their synaptic targets are not yet known.
sium channels and that serotonin has multiple Thus, there is a strong likelihood that the net-
effects via different types of ion channels that works generating these distinct forms of rhythmic C
are differentially distributed in the cell (Booth hind limb movement partly overlap in turtles.
et al. 1997). In slice preparations, motor neurons
can produce oscillations in the presence of the Multifunctional and Specialized Spinal
glutamate agonist NMDA even when all action Interneurons
potentials are blocked (by tetrodotoxin) (Guertin Most scratch-activated interneurons are activated
and Hounsgaard 1998). These NMDA-dependent during stimulation of the body surface for all
motor neuron oscillations disappear if L-type cal- three forms of ipsilateral scratching and often
cium channels are blocked. contralateral scratching as well (Berkowitz
Although intrinsic properties of motor neurons 2001). Nonetheless, most such interneurons are
may contribute to shaping their bursts during most strongly activated during stimulation of one
scratching (Alaburda and Hounsgaard 2003), the particular region of the body surface and are
synaptic input to motor neurons is so strong dur- progressively less strongly activated during stim-
ing rostral scratching motor patterns in vitro that ulation of sites progressively further away from
it likely overwhelms and suppresses motor neu- this preferred region. Thus, the activity of these
ron intrinsic properties like plateau potentials, neurons does not precisely indicate which site has
which may consequently play a relatively minor been stimulated. Instead, each is broadly tuned to
role during scratching (Alaburda et al. 2005). a region of the body surface. Different scratch-
activated interneurons within one animal’s spinal
Spinal Interneurons Involved in Locomotion cord are broadly tuned to different regions of the
and Scratching body surface. The combination of several such
Much is still unknown about the spinal cord interneurons with different tuning (i.e., popula-
interneurons and pathways that mediate locomotion coding) can in principle specify precisely
tion and scratching in turtles and in limbed ver- which site has been stimulated and thus which
tebrates generally. However, some things are form of scratching should be generated to initiate
known (Berkowitz 2010). Turtles have been par- contact with that site (Berkowitz 2009). If,
ticularly useful for studying the extent to which indeed, the turtle spinal cord uses a large popula-
networks of interneurons that generate forward tion of broadly tuned interneurons to code stimu-
swimming and the three forms of scratching in lus locations and corresponding movement
each hind limb are shared versus separate. It is directions, this would be similar to population
known that many spinal cord interneurons are coding in many animals and for many behaviors,
rhythmically activated during forward swimming from leech local bending to primate arm reaching
and all three forms of scratching (Berkowitz and (Briggman and Kristan 2008).
Stein 1994a, b; Berkowitz 2001a, b, 2002). These Nonetheless, some turtle spinal cord interneu-
multifunctional interneurons are found through- rons are behaviorally specialized (Fig. 2;
out the hind limb enlargement spinal cord, as well Berkowitz 2010). Some interneurons are rhyth-
as in spinal cord segments rostral to the hind limb mically activated during scratching motor pat-
enlargement (Berkowitz 2001a, b). In the trans- terns, but are not activated at all or are even
verse plane, these interneurons are scattered inhibited during forward swimming motor pat-
through the ventral part of the dorsal horn, the terns (Berkowitz 2002, 2008). In addition, there
intermediate zone, and the dorsal part of the ven- are interneurons that are strongly activated during
tral horn. At least some of these interneurons limb withdrawal (flexion reflex) motor patterns
while interneurons that are rhythmically active

during the hip-flexor phase are generally active
during hip-extensor deletions, suggesting that
there is an entire module of interneurons and
motor neurons that is inactive during
hip-extensor deletions (Stein and Daniels-
McQueen 2002; Stein 2008).
In addition to rhythmic, hip-related interneu-
rons, some interneurons do shift their peak firing
within the hip activity cycle between scratching
and swimming or between one form of scratching
and another (Berkowitz 2001, 2002). Such inter-
Control of Locomotion and Scratching in Turtles, neurons could be involved directly in control of
Fig. 2 Schematic Venn diagram illustrating other joints, such as the knee or ankle, consistent
multifunctional and specialized types of spinal with the unit-burst generator hypothesis. Rhyth-
interneurons
mic interneurons have been studied that generate
bursts of action potentials correlated with knee-
extensor or knee-flexor motor neuron activities
that are inhibited during forward swimming and (Stein and Daniels-McQueen 2003). Some spinal
each form of scratching motor pattern (Berkowitz cord ventral horn and dorsal horn interneurons
2007). The relative roles and effects of shared also show the same kind of bistability as seen in
versus behaviorally specialized interneurons are motor neurons (Hounsgaard and Kjaerulff 1992;
not yet understood in turtles. There is also evi- Russo and Hounsgaard 1994, 1996).
dence from hatchling tadpoles and larval Among the spinal cord interneurons that are
zebrafish that a combination of shared and behav- rhythmically activated during forward swimming
iorally specialized interneurons contributes to and all three forms of scratching is a class of
different forms of axial locomotion (Berkowitz interneurons called transverse interneurons, or
et al. 2010). Less is known about the extent to T neurons, that can be delineated by their
which circuitry is shared in the control of differ- somato-dendritic morphologies (Berkowitz
ent forms of locomotion and scratching in et al. 2006). These interneurons have dendrites
mammals. that extend far dorsoventrally and mediolaterally
(i.e., in the transverse plane) but are short
Rhythmic Activity of Spinal Interneurons rostrocaudally; they also tend to have a soma
Spinal cord interneurons that are activated during that is longer mediolaterally than rostrocaudally.
swimming and scratching motor patterns tend to T neurons tend to be very rhythmically active
be rhythmically active or at least have firing rates during both swimming and scratching
that are rhythmically modulated. Most such inter- (Berkowitz 2008). They also tend to have higher
neurons tend to fire action potentials (i.e., have peak firing rates during scratching, larger mem-
a preferred phase) during a particular phase of the brane potential oscillations during scratching,
hip flexion-extension cycle and tend to maintain and briefer afterhyperpolarizations following
this preferred phase across all three forms of action potentials than other scratch-activated
scratching and often during forward swimming interneurons (Berkowitz et al. 2006). T neurons
as well (Berkowitz 2001, 2002). Thus, control are also typically activated during ipsilateral limb
of hip movement timing appears to dominate withdrawal. At least some T neurons have axon
spinal cord interneuron rhythmic activity. Inter- terminals in the ventral horn in the hind limb
neurons that are rhythmically active during the enlargement. Thus, T neurons are good candi-
hip-extensor phase of scratching motor patterns dates to play an important role in motor control
are generally quiet during hip-extensor deletions, for a wide range of limb movements.
central pattern generators for swimming and
scratching have common members or at least
strong interactions between them. Another way
of addressing this question is to deliver two stim-
uli simultaneously that each individually evokes
one of the two motor patterns. If the sets of
interneurons that generate the two motor patterns C
Control of Locomotion and Scratching in Turtles, are entirely separate (i.e., the central pattern gen-
Fig. 3 Hypothetical architecture to account for how erators are separate), then simultaneous stimula-
broadly tuned spinal interneurons could generate the cor-
rect knee-hip synergies for rostral and pocket scratching. tion should generate a superposition of the two
Each interneuron would project to two motor neuron pools rhythms within motor neurons. If many of the
(directly or indirectly), with differently tuned interneurons central pattern generator interneurons are shared
having different dual projections. Knee-extensor motor (or have strong interactions between them, which
neurons would integrate overlapping excitatory and inhib-
itory inputs from differently tuned interneurons, thus might be equivalent), however, then one stimulus
determining the timing of knee-extensor bursts within might modify the motor pattern produced by the
the hip cycle. Whichever group of interneurons was other stimulus in a coordinated manner
more active would determine the knee-hip synergy (Berkowitz and Hao 2011).
As noted above, simultaneous stimulation of
A hypothetical architecture has been proposed the body surface in regions that evoke two differ-
that could in principle generate the correct knee- ent forms of scratching can evoke coordinated
hip synergies for rostral scratching and pocket and adaptive blends of the two (Robertson
scratching using only broadly tuned interneurons et al. 1985; Stein et al. 1986). In addition, brief
that preferentially fire action potentials in forward swim stimulation can interrupt and reset
a particular phase of the hip cycle (Fig. 3; the rostral scratching rhythm and vice versa
Berkowitz 2001). Each such interneuron would (Juranek and Currie 2000). (Also, brief stimula-
project to both a knee motor neuron pool and tion to evoke a limb withdrawal or flexion reflex
a hip motor neuron pool. According to this motor pattern can reset the scratching rhythm
hypothesis, interneurons that are broadly tuned (Currie and Stein 1989).) Finally, prolonged
to a rostral scratch region would project to both simultaneous forward swim stimulation and
knee-extensor and hip-flexor interneurons, while scratch stimulation (for all three forms of
interneurons that are broadly tuned to a pocket scratching) can generate swim-scratch blends,
scratch region would project to both knee- generate a swim rhythm when the swim stimula-
extensor and hip-extensor motor neurons. The tion alone is insufficient, generate a pure-form
knee-hip synergy would then be determined in swim rhythm with a rhythm frequency greater
motor neurons, especially knee-extensor motor than that evoked by either stimulation alone,
neurons, that integrate excitatory and inhibitory and sometimes disrupt one phase of the rhythm
inputs with different but overlapping time or disrupt rhythm generation entirely (Hao
courses. There is evidence from intracellular et al. 2011). These physiological findings suggest
recordings of knee-extensor motor neurons that that there are either substantial common elements
they do receive excitatory and inhibitory postsyn- to the swimming and scratching central pattern
aptic potentials that overlap in time during rostral generators or strong interactions between the
and pocket scratching motor patterns (Robertson generators.
and Stein 1988). Computer simulations were also generated of
simplified models of caudal scratch and forward
Dual Stimulation swim networks that were (1) entirely separate,
The large set of spinal interneurons that are rhyth- (2) separate but interacting, and (3) shared central
mically activated during both swimming and pattern generators (Hao et al. 2011). The results
scratching motor patterns suggests that the of these simulations suggested that shared or
strongly interacting central pattern generators right hind limbs (Field and Stein 1997), which
can account for the physiological findings but might rely on the same spinal cord circuitry that
completely separate central pattern generators underlies left-right coordination during bilateral
cannot. The strongly interacting simulated net- swimming. Bilaterally evoked scratching usually
works and especially the shared simulated net- involves alternation between the two hind limbs,
work replicated the physiological findings of but can involve synchronized (in-phase) move-
increased rhythm frequency under dual stimula- ments, especially for caudal scratch stimulation
tion and summation of a subthreshold swim stim- near the tail. On occasion, the left and right hind
ulus with a suprathreshold scratch stimulus to limbs display 2:1 coordination, in which one limb
generate a suprathreshold swim. The success of moves through two cycles in exactly the time that
these models depended on neurons switching the other limb moves through one cycle (Stein
their phase in the swim rhythm via escape from and McCullough 1998).
inhibition. Many spinal cord interneurons that are
rhythmically activated during ipsilateral
Left-Right Hind Limb Coordination scratching motor patterns are also robustly and
Coordination between the left and right hind rhythmically activated during contralateral
limbs also can be generated by the spinal cord. scratching motor patterns (Berkowitz and Stein
In intact turtles performing forward swimming 1994a, b; Berkowitz 2001a, b). Such interneurons
spontaneously, forward swimming involves might be components of networks that produce
alternating (i.e., out-of-phase) coordination the coordinated activity in contralateral motor
between left and right hind limb movements neurons during scratching and potentially during
(Field and Stein 1997). Surprisingly, such coor- swimming as well. Some such interneurons are
dination continues after a cut along the midline of rhythmically active in the same phase of the
the spinal cord (separating the left and right sides) ipsilateral hip-flexor rhythm no matter which
throughout the hind limb enlargement and the side is stimulated, while other such interneurons
immediately pre-enlargement segments (Samara switch to the opposite phase when the opposite
and Currie 2007). This shows that commissural side of the body is stimulated. Collectively, such
pathways in more rostral regions (including the interneurons could mediate the in-phase and
forelimb enlargement) are sufficient to produce out-of-phase left-right limb coordination that
left-right hind limb alternation. has been observed during scratching and
Scratching, however, is essentially a unilateral locomotion.
behavior; the ipsilateral hind limb rhythmically But does the contralateral spinal cord actually
rubs against the stimulated site, while the contra- play a role in generating unilateral scratching
lateral hind limb remains relatively still (Mortin motor patterns? Surprisingly, the answer is yes,
et al. 1985). A close look at the activity of con- as shown by experiments in which the spinal cord
tralateral motor neurons during unilateral scratch on one side was removed, throughout the hind
stimulation, however, has revealed that they typ- limb enlargement (Stein et al. 1995). Stimulation
ically are weakly but rhythmically activated in of the body surface on the intact side evoked
alternation with their ipsilateral counterparts dur- a strong rostral scratching rhythm on that side
ing unilateral scratching (Currie and Stein 1989). that was consistently missing the hip-extensor
It is not clear if this contralateral motor neuron phase (i.e., hip-extensor deletion rostral
activation is adaptive during scratching; it might scratching). Stimulation of the body surface on
be an epiphenomenon due to activation of net- the lesioned side (but rostral to the lesion) evoked
works that are shared between swimming and a hip-extensor rhythm alone on the intact side.
scratching. In addition, bilateral scratch stimula- Both of these findings suggest that the contralat-
tion and unilateral scratch stimulation near the eral spinal cord contributes a specific component
midline (e.g., near the tail) can evoke bilaterally to the scratching motor pattern, namely, the
coordinated scratching movements of the left and hip-extensor phase.
In other experiments, a hemisection (i.e., when glycine receptors are blocked (Currie and
a spinal cord transection limited to one side, Lee 1997). Glycinergic inhibition cannot be nec-
e.g., the left) was used to eliminate stimulation essary for coordination during scratching, how-
of the left hind limb enlargement via ipsilateral ever, as left-right coordination can continue
descending axons on the left and contralateral under glycine receptor blockade.
descending axons from the right (Stein
et al. 1998). In this case, stimulation of the body C
surface on the intact right side (rostral to the
References
hemisection) evoked hip-extensor deletion ros-
tral scratching. Stimulation of the body surface Alaburda A, Hounsgaard J (2003) Metabotropic modula-
on the left side (rostral to the hemisection) in tion of motoneurons by scratch-like spinal network
combination with the stimulation on the right activity. J Neurosci 23:8625–8629
side evoked a normal rostral scratching rhythm Alaburda A, Russo R, MacAulay N, Hounsgaard J (2005)
Periodic high-conductance states in spinal neurons
(including the hip-extensor phase), i.e., the rostral during scratch-like network activity in adult turtles.
scratch motor pattern was “reconstructed” by this J Neurosci 25:6316–6321
two-site stimulation. This finding suggests that Berkinblit MB, Deliagina TG, Feldman AG, Gelfand IM,
descending, crossed axons activate interneurons Orlovsky GN (1978) Generation of scratching
I. Activity of spinal interneurons during scratching.
that produce the hip-extensor phase of scratching J Neurophysiol 41:1040–1057
and that these interneurons are coordinated with Berkowitz A (2001a) Broadly tuned spinal neurons for
interneurons that are activated through unilateral each form of fictive scratching in spinal turtles.
pathways. In addition, crossed pathways in J Neurophysiol 86:1017–1025
Berkowitz A (2001b) Rhythmicity of spinal neurons acti-
pre-enlargement segments are sufficient to create vated during each form of fictive scratching in spinal
the left-right alternation during bilateral rostral turtles. J Neurophysiol 86:1026–1036
scratch stimulation. Berkowitz A (2002) Both shared and specialized spinal
All of these findings, in combination with the circuitry for scratching and swimming in turtles.
single-neuron recordings of interneurons that are Berkowitz A (2005) Physiology and morphology indicate
bilaterally activated during scratching, suggest that individual spinal interneurons contribute to diverse
that there is a bilateral set of interneurons that limb movements. J Neurophysiol 94:4455–4470
each contribute to both left and right hind limb Berkowitz A (2007) Spinal interneurons that are selec-
tively activated during fictive flexion reflex.
scratching (Stein et al. 1995). This set of shared J Neurosci 27:4634–4641
neurons might contribute to bilateral coordina- Berkowitz A (2008) Physiology and morphology of
tion for both scratching and swimming. Surpris- shared and specialized spinal interneurons for locomo-
ingly, though, one side (left or right) of the hind tion and scratching. J Neurophysiol 99:2887–2901
Berkowitz A (2009) Population coding. In: Encyclopedia
limb enlargement is sufficient to produce rhyth- of neuroscience, vol 7. Elsevier, Oxford, pp 757–764
mic hip-flexor, hip-extensor, and knee-extensor Berkowitz A (2010) Multifunctional and specialized spi-
muscle activities with appropriate phasing during nal interneurons for turtle limb movements. Ann N Y
forward swimming movements evoked by elec- Acad Sci 1198:119–132
Berkowitz A, Hao ZZ (2011) Partly shared spinal cord
trical stimulation of the contralateral lateral networks for locomotion and scratching. Integr Comp
funiculus (in the rostral spinal cord), though Biol 51:890–902
cycle frequency and muscle burst amplitude are Berkowitz A, Stein PS (1994a) Activity of descending
lower than otherwise. Thus, the contralateral hind propriospinal axons in the turtle hindlimb
enlargement during two forms of fictive scratching:
limb enlargement spinal cord may actually play broad tuning to regions of the body surface.
a more important role in generating unilateral J Neurosci 14:5089–5104
scratching than in generating unilateral Berkowitz A, Stein PS (1994b) Activity of descending
swimming. propriospinal axons in the turtle hindlimb enlargement
during two forms of fictive scratching: phase analyses.
Crossed glycinergic inhibition appears to con- J Neurosci 14:5105–5119
tribute to bilateral coordination during Berkowitz A, Yosten GL, Ballard RM (2006) Somato-
scratching, as coordination is more variable dendritic morphology predicts physiology for neurons
that contribute to several kinds of limb movements. Grillner S (1981) Control of locomotion in bipeds, tetra-
J Neurophysiol 95:2821–2831 pods, and fish. In: Brooks V (ed) Handbook of physi-
Berkowitz A, Roberts A, Soffe SR (2010) Roles for ology, Sect. 1. The nervous system. Motor control,
multifunctional and specialized spinal interneurons vol 2. American Physiological Society, Bethesda,
during motor pattern generation in tadpoles, zebrafish pp 1179–1236
larvae, and turtles. Front Behav Neurosci 4:36. Guertin PA, Hounsgaard J (1998) NMDA-Induced intrin-
doi:10.3389/fnbeh.2010.00036 sic voltage oscillations depend on L-type calcium
Booth V, Rinzel J, Kiehn O (1997) Compartmental model channels in spinal motoneurons of adult turtles.
of vertebrate motoneurons for Ca2+-dependent spiking J Neurophysiol 80:3380–3382
and plateau potentials under pharmacological treat- Hao ZZ, Spardy LE, Nguyen EB, Rubin JE, Berkowitz
ment. J Neurophysiol 78:3371–3385 A (2011) Strong interactions between spinal cord net-
Briggman KL, Kristan WB (2008) Multifunctional works for locomotion and scratching. J Neurophysiol
pattern-generating circuits. Annu Rev Neurosci 106:1766–1781
31:271–294 Ho S, O’Donovan MJ (1993) Regionalization and
Brown TG (1911) The intrinsic factors in the act of pro- intersegmental coordination of rhythm-generating net-
gression in the mammal. Proc R Soc Lond 84:308–319 works in the spinal cord of the chick embryo.
Cazalets JR, Borde M, Clarac F (1995) Localization and J Neurosci 13:1354–1371
organization of the central pattern generator for Hounsgaard J, Kiehn O (1989) Serotonin-induced
hindlimb locomotion in newborn rat. J Neurosci bistability of turtle motoneurones caused by a
15:4943–4951 nifedipine-sensitive calcium plateau potential.
Cowley KC, Schmidt BJ (1997) Regional distribution of J Physiol 414:265–282
the locomotor pattern-generating network in the neo- Hounsgaard J, Kjaerulff O (1992) Ca2+-mediated plateau
natal rat spinal cord. J Neurophysiol 77:247–259 potentials in a subpopulation of interneurons in the
Currie SN (1999) Fictive hindlimb motor patterns evoked ventral horn of the turtle spinal cord. Eur J Neurosci
by AMPA and NMDA in turtle spinal cord-hindlimb 4:183–188
nerve preparations. J Physiol Paris 93:199–211 Hounsgaard J, Mintz I (1988) Calcium conductance and
Currie SN, Lee S (1996) Sensory-evoked pocket scratch firing properties of spinal motoneurones in the turtle.
motor patterns in the in vitro turtle spinal cord: reduc- J Physiol 398:591–603
tion of excitability by an N-methyl-D-aspartate antag- Hounsgaard J, Nicholson C (1990) The isolated turtle
onist. J Neurophysiol 76:81–92 brain and the physiology of neuronal circuits. In:
Currie SN, Lee S (1997) Glycinergic inhibition contrib- Jahnsen H (ed) Preparations of vertebrate central ner-
utes to the generation of rostral scratch motor vous system in vitro. Wiley, New York, pp 155–181
patterns in the turtle spinal cord. J Neurosci Juranek J, Currie SN (2000) Electrically evoked fictive
17:3322–3333 swimming in the low-spinal immobilized turtle.
Currie SN, Stein PS (1989) Interruptions of fictive scratch J Neurophysiol 83:146–155
motor rhythms by activation of cutaneous flexion Keifer J, Stein PS (1983) In vitro motor program for the
reflex afferents in the turtle. J Neurosci 9:488–496 rostral scratch reflex generated by the turtle spinal
Currie SN, Stein PS (1990) Cutaneous stimulation evokes cord. Brain Res 266:148–151
long-lasting excitation of spinal interneurons in the Kjaerulff O, Kiehn O (1996) Distribution of networks
turtle. J Neurophysiol 64:1134–1148 generating and coordinating locomotor activity in the
Currie SN, Stein PS (1992) Glutamate antagonists applied neonatal rat spinal cord in vitro: a lesion study.
to midbody spinal cord segments reduce the excitabil- J Neurosci 16:5777–5794
ity of the fictive rostral scratch reflex in the turtle. Lennard PR, Stein PS (1977) Swimming movements
Brain Res 581:91–100 elicited by electrical stimulation of turtle spinal cord.
Delgado-Lezama R, Perrier JF, Nedergaard S, Svirskis G, I Low-spinal and intact preparations. J Neurophysiol
Hounsgaard J (1997) Metabotropic synaptic regulation 40:768–778
of intrinsic response properties of turtle spinal Lutz PL, Milton SL (2004) Negotiating brain anoxia sur-
motoneurones. J Physiol 504(Pt 1):97–102 vival in the turtle. J Exp Biol 207:3141–3147
Deliagina TG, Orlovsky GN, Pavlova GA (1983) The Mejia-Gervacio S, Hounsgaard J, Diaz-Munoz M (2004)
capacity for generation of rhythmic oscillations is dis- Roles of ryanodine and inositol triphosphate receptors
tributed in the lumbosacral spinal cord of the cat. Exp in regulation of plateau potentials in turtle spinal moto-
Brain Res 53:81–90 neurons. Neuroscience 123:123–130
Earhart GM, Stein PS (2000) Scratch-swim hybrids in the Mortin LI, Stein PS (1989) Spinal cord segments
spinal turtle: blending of rostral scratch and forward containing key elements of the central pattern genera-
swim. J Neurophysiol 83:156–165 tors for three forms of scratch reflex in the turtle.
Field EC, Stein PS (1997) Spinal cord coordination of J Neurosci 9:2285–2296
hindlimb movements in the turtle: interlimb temporal Mortin LI, Stein PS (1990) Cutaneous dermatomes for
relationships during bilateral scratching and swim- initiation of three forms of the scratch reflex in the
ming. J Neurophysiol 78:1404–1413 spinal turtle. J Comp Neurol 295:515–529
Coordinate Transformations, Role of Spinal Circuitry in 845 C
Mortin LI, Keifer J, Stein PS (1985) Three forms of the Stein PS, Daniels-McQueen S (2004) Variations in motor
scratch reflex in the spinal turtle: movement analyses. patterns during fictive rostral scratching in the turtle:
J Neurophysiol 53:1501–1516 knee-related deletions. J Neurophysiol 91:2380–2384
Perrier JF, Hounsgaard J (1999) Ca(2+)-activated Stein PS, McCullough ML (1998) Example of 2:1
nonselective cationic current (I(CAN)) in turtle moto- interlimb coordination during fictive rostral scratching
neurons. J Neurophysiol 82:730–735 in a spinal turtle. J Neurophysiol 79:1132–1134
Perrier JF, Hounsgaard J (2003) 5-HT2 receptors promote Stein PS, Camp AW, Robertson GA, Mortin LI
plateau potentials in turtle spinal motoneurons by (1986) Blends of rostral and caudal scratch reflex
facilitating an L-type calcium current. J Neurophysiol motor patterns elicited by simultaneous stimulation of
C
89:954–959 two sites in the spinal turtle. J Neurosci 6:2259–2266
Robertson GA, Stein PS (1988) Synaptic control of Stein PS, Victor JC, Field EC, Currie SN (1995) Bilateral
hindlimb motoneurones during three forms of the fic- control of hindlimb scratching in the spinal turtle:
tive scratch reflex in the turtle. J Physiol 404:101–128 contralateral spinal circuitry contributes to the normal
Robertson GA, Mortin LI, Keifer J, Stein PS (1985) Three ipsilateral motor pattern of fictive rostral scratching.
forms of the scratch reflex in the spinal turtle: central J Neurosci 15:4343–4355
generation of motor patterns. J Neurophysiol Stein PSG, McCullough ML, Currie SN (1998) Reconstruc-
53:1517–1534 tion of flexor/extensor alternation during fictive rostral
Romanes GJ (1964) The motor pools of the spinal cord. scratching by two-site stimulation in the spinal turtle with
Prog Brain Res 11:93–119 a transverse spinal hemisection. J Neurosci 18:467–479
Ruigrok TJH, Crowe A (1984) The organization of moto- Svirskis G, Hounsgaard J (1998) Transmitter regulation of
neurons in the turtle lumbar spinal cord. J Comp plateau properties in turtle motoneurons.
Neurol 228:24–37 J Neurophysiol 79:45–50
Russo RE, Hounsgaard J (1994) Short-term plasticity in
turtle dorsal horn neurons mediated by L-type Ca2+
channels. Neuroscience 61:191–197
Russo RE, Hounsgaard J (1996) Plateau-generating Convection-Diffusion Equation
neurones in the dorsal horn in an in vitro preparation
of the turtle spinal cord. J Physiol 493(Pt 1):39–54
Samara RF, Currie SN (2007) Crossed commissural path- ▶ Fokker-Planck Equation
ways in the spinal hindlimb enlargement are not nec-
essary for right left hindlimb alternation during turtle
swimming. J Neurophysiol 98:2223–2231
Samara RF, Currie SN (2008a) Electrically evoked
locomotor activity in the turtle spinal cord Coordinate Transformations, Role of
hemi-enlargement preparation. Neurosci Lett Spinal Circuitry in
441:105–109
Samara RF, Currie SN (2008b) Location of spinal cord
pathways that control hindlimb movement amplitude Yifat Prut
and interlimb coordination during voluntary swim- Department of Medical Neurobiology, Hadassah
ming in turtles. J Neurophysiol 99:1953–1968 Medical School, The Hebrew University,
Simon M, Perrier JF, Hounsgaard J (2003) Subcellular Jerusalem, Israel
distribution of L-type Ca2+ channels responsible for
plateau potentials in motoneurons from the
lumbar spinal cord of the turtle. Eur J Neurosci
18:258–266 Synonyms
Stein PS (2005) Neuronal control of turtle hindlimb motor
rhythms. J Comp Physiol A 191:213–229
Stein PS (2008) Motor pattern deletions and modular orga- Descending control of spinal circuitry in volun-
nization of turtle spinal cord. Brain Res Rev 57:118–124 tary movements; Role of descending pathways in
Stein PS (2010) Alternation of agonists and antagonists extrinsic to intrinsic coordinate transformation;
during turtle hindlimb motor rhythms. Ann N Y Acad Spinal control of voluntary movements
Sci 1198:105–118
Stein PS, Daniels-McQueen S (2002) Modular organiza-
tion of turtle spinal interneurons during normal and
deletion fictive rostral scratching. J Neurosci Definition
22:6800–6809
Stein PS, Daniels-McQueen S (2003) Timing of knee-
related spinal neurons during fictive rostral scratching The process through which a motor target in
in the turtle. J Neurophysiol 90:3585–3593 the external environment is translated into a
C 846 Coordinate Transformations, Role of Spinal Circuitry in
body-centered coordinate system is called coor- which in itself could provide segmental neurons
dinate transformation. In the motor system, this with sensory-related information (Kuypers 1987;
translation is often referred to as the transforma- Toyoshima and Sakai 1982). Other sources of
tion of an externally defined (e.g., in a retinotopic descending input such as the vestibulospinal and
coordinate frame) target into a muscle-based rubrospinal pathways may provide further state
(or intrinsically defined) motor command. Multi- variables related to posture, balance, etc. that
ple cortical areas are known to be involved in this could be relevant to the spinal processing of
process. Many of these areas, including pre- and coordinate transformation (Baldissera
post-central sites, project downstream to activate et al. 1981). In addition to these sources of
directly and indirectly spinal neurons. The inter- processed information, spinal neurons integrate
action of this extensive system of descending proprioceptive inputs which are relayed via a rich
pathways with spinal circuitry may further adjust afferent system. Finally, neighboring INs have
motor commands before they reach target mus- been shown to engage with local interactions
cles in a final processing step of coordinate (Jankowska 1992) which can modify converging
transformation. inputs before finally affecting MNs.
This entry reviews the mounting evidence that
spinal neurons provide the final processing step
Detailed Description for completing the coordinate transformation in
the motor system. It further develops the concept
When making a movement directed toward that in addition to the spatial coordinate system,
a target in the world around us, the nervous sys- spinal processing also modifies the temporal pat-
tem is faced with a computationally difficult tern of descending commands to match it with the
task. The target which is initially defined in required pattern of muscle activation. Finally, the
a retinotopic coordinate system must be trans- possible implications of this organizational
lated into a detailed body-centered motor com- scheme in which motor commands are finalized
mand that specifies the required spatiotemporal downstream to M1 are discussed in terms of
activation profile of all muscles (Shadmehr and current research in the field of motor control.
Wise 2005). This computation is completed on
the fly and in an unconscious manner. It is gen-
erally assumed that this process is the outcome of Activity of Spinal Interneurons During
cortical processing in which multimodal sensory Voluntary Movement
information is integrated and converted into
appropriate motor commands (Sabes 2011). There is abundant work on the organization of
However, there are several reasons why segmen- segmental circuitry in reduced preparations (i.e.,
tal interneurons (INs) could also contribute to the cats, rodents, and in some cases monkeys). These
process of coordinate transformation. First, the studies have mostly concentrated on the input to
segmental network of interneurons is situated these neurons from local, peripheral, and/or
between the motor cortical network and descending sources. Nonetheless, understanding
corresponding motoneurons (MNs) such that the functional contribution of spinal neurons to
most of the cortical information sent to MNs is any type of motor behavior requires recording
relayed via INs (Porter and Lemon 1993). Sec- single cell properties and cell-to-cell interactions
ond, multiple cortical areas transmit information during task performance, since these characteris-
to spinal neurons, including the motor and tics considerably affect the input-to-output trans-
premotor areas, which could provide INs with formation of any network. The ability to induce
diverse information that contains both high- and fictive locomotion in reduced preparations (e.g.,
low-level features of the motor command (Dum isolated spinal cord) has led to extensive study of
and Strick 2002). Some descending output interneuron (IN) activity in these conditions.
emerges from sensory areas of the parietal lobe, However, these studies concentrate on spinal
network activity rather than its control by firing rates than cortical neurons. However, this
descending corticospinal pathways, which how- enhanced firing is more regular (in terms of the
ever is the fundamental component of voluntary inter-spike interval) and less correlated among
motor control. Early studies of cervical spinal neighboring interneurons than that of cortical
neurons during normal motor behavior of pri- neurons. Specifically, the spinal noise correlation
mates (Fetz et al. 1999) showed that there is was found to be independent of the signal corre-
both directional and torque tuning of these neu- lation recorded for these cells (Averbeck and Lee C
rons as well as connectivity with muscles. The 2004). Hence neighboring neurons with similar
connections of INs with muscles were biased response properties do not express any tendency
toward flexor muscles unlike the generally even for rate co-modulation on a trial-to-trial basis.
corticomuscular connectivity found for motor Thus spinal neurons may “amplify” a cortical
cortical neurons that have a functional link with command and provide a more rigorous drive to
flexors and extensors muscles (Fetz 1992). Later working muscles while at the same time spinal
studies revealed that spinal neurons are not only circuitry can eliminate any redundancy in the
active during motor action but modify their rate command which may exist at the motor cortical
when preparing for movement during an level (Lee et al. 1998).
instructed delay period (Prut and Fetz 1999).
This finding demonstrated that spinal INs are
not merely related to motor execution but also Integration of Sensorimotor
receive information (which can be excitatory or Information in the Primary Motor Cortex
inhibitory) during motor preparation. Via ascend-
ing pathways, this spinal activation can back- The next step in understanding the contribution of
modify cortical activity during the pre-movement spinal neurons to coordinate transformation is to
delay period when coordinate transformation characterize the properties of the motor com-
apparently takes place. Further investigation of mand emitted by the motor cortex. The motor
spinal activity during a two-dimensional wrist cortex integrates multiple sources of information
task (Harel et al. 2008) revealed that the tuning in various modalities as required for performing
properties (such as the width and height of tuning some form of sensory-to-motor transformation.
curves) were very similar for both motor cortical Information about higher aspects of motor
and spinal neurons but that the distribution of actions and external targets is relayed to the
preferred directions (PDs) tended to be more motor cortex via the premotor areas. Neurons in
biased across spinal neurons as compared to the the primary motor cortex may thus express
uniform distribution of cortical PDs responses to a large set of environmental cues,
(Georgopoulos et al. 1986). This result provided as long as they are informative about the action to
a first indication that spinal circuitry process be made. For example, a general visual cue may
arriving input in order to match it to working produce a very weak response in M1, whereas
muscles. Nonetheless, the response properties of a similar cue that contains directional informa-
spinal neurons have not been found to be dramat- tion may trigger a robust response in the same cell
ically different from those found for motor corti- (Asher et al. 2010). Furthermore, even variables
cal cells. Thus a general conclusion that can be which normally are not expected to be coded for
drawn from studies of the task-related activity of in the motor cortex (such as the color of a cue)
spinal neurons is that these neurons express a rich may induce a distinguishable response as long as
spectrum of firing properties that are related to they are relevant to the ensuing movement (Zach
different aspects of the task including both motor et al. 2008). In this respect the motor cortex
and non-motor properties. apparently codes the motor-relevant information
Some differences have been reported in the regardless of its specific modality.
firing properties of spinal and cortical neurons. Along with information about the environ-
First, spinal neurons generally exhibit higher ment and targets, the motor cortex receives
somatosensory information from the somatosen- command is specified is a topic of ongoing debate.
sory cortex which provides it with details about The classical view suggested a muscle-based
the state of the plant. It appears that while visual coordinate system (Evarts et al. 1984), whereas
information can be suppressed at the level of later views posited that cortical neurons operate
motor cortex based on its relevance to the ensuing in an extrinsic-based coordinate frame
action, motor cortical response to somatosensory (Georgopoulos et al. 1986). However, the general
cues (e.g., unexpected perturbation of the arm) is consensus today is that motor cortical neurons
extensive and robust. This is reasonable since any operate in a system that combines both coordinate
change in the state of the periphery is important systems and thus mostly contains a mixed repre-
for the motor machinery as regards updating and sentation of both target-related and muscle-
correcting the planned and/or executed move- related information, in a fairly heterogeneous
ment. Nonetheless, there is evidence that as manner across the motor cortex (Kalaska 2009).
soon as a movement has been initiated, some Early studies used a behavioral paradigm in
sensory information (i.e., cutaneous information) which monkeys were trained to perform a
is attenuated in a way which could potentially center-out wrist task with three different hand
serve to enhance the feedforward component of postures: pronation, mid-position, and supina-
motor control. This implies that at least this com- tion (Kakei et al. 2003). The use of a single-
ponent of sensory information is important for joint task that relies on the near-isotropic wrist
planning actions, but not for its online guidance. joint allowed for a systematic exploration of the
In addition to the corticocortical information impact of hand posture on spatial tuning proper-
that is integrated by motor cortical neurons, ties of motor cortical neurons (compared to
thalamocortical input relays information from a similar approach that was used in a reaching
the cerebellum and the basal ganglia. This infor- task; Scott and Kalaska 1997). The authors were
mation appears to be more crucial to action selec- able to distinguish between three different coor-
tion and action timing (Hikosaka 1998) and thus dinate systems: external, muscle-based, and
may not play a dominant role in coordinate trans- joint-based. They found that the majority of
formation. Hence motor cortical neurons appear neurons exhibited a small shift in their preferred
to integrate different sources of corticocortical direction (30 shifts on average when the hand
information in a weighted manner which is rotated from pronation to supination and the
related to the importance of these sources to muscle rotated 75 ). Moreover, recordings in
motor performance. One operative framework the premotor ventral cortex (PMv) revealed that
that captures this property of motor cortical activ- neurons in this high-level motor cortical site do
ity is the Optimal Feedback Control model which not change their PD during the task; i.e., their
posits that motor cortical activity manipulates tuning is not affected by hand posture. A study
and optimizes feedback gains in order to mini- that used the same paradigm in primates during
mize some task-specific cost function (Scott spinal recordings revealed that although the
2004; Todorov and Jordan 2002). However, this majority of spinal neurons express PD shifts that
theory provides no explicit account of the trans- were similar to muscles, cortical neurons recorded
formation of coordinates by the system. at the same time expressed, on average, a much
smaller PD shift. The authors concluded that the
transformation from external to internal coordi-
The Coordinate Frame of Motor Cortical nate frame is completed downstream of M1.
Information More generally, the transformation of the coordi-
nate frame does not take place within a single
The end result of motor cortical integration is site; rather, there is a gradual and successive tran-
a motor command to the working muscles. The sition of the coordinate system via inter-site
exact coordinate system within which this motor interactions.
Theoretical studies have argued that it is preparation to execution from an external to an
impossible to deduce the coordinate frame from internal mode (Harel et al. 2008). This result is
PD shifts (Shah et al. 2004) because when similar in spirit but very different in detail from
shifting postures, neurons modify both their PDs the mental rotation of the population vector
and firing rates (often quantified as a dynamic reported by Georgopoulos and colleagues
index which is the difference in firing range in (1989) where the PV shifted from the cued to
the two postures over the sum of these ranges). the actual movement but remained in an exter- C
This suggests that in practice it is possible to nally based frame of coordinates. Accordingly,
obtain any type of coordinate system irrespective integration of different sources of information
of the extent of PD shift, since there can be (e.g., a postural variable and a directional cue)
a tradeoff between changes in PD and dynamic should be modeled using time-dependent weight
indices. Nevertheless, spinal neurons and cortical variables.
neurons were shown to exhibit similarly symmet-
ric distribution of dynamic indices but different
PD shifts. Thus spinal neurons participate in the Spinal Targets of Descending Pathways
process of coordinate transformation by provid-
ing the final computational step which includes The possibility that the final step in coordinate
removing any external-based information (which transformation takes place at the spinal level does
is non-separable at the cortical level) and thus not necessarily imply that spinal interneurons are
rotating the PD to fit the exclusively muscle- actively or exclusively involved in this processing
based frame. stage. The termination pattern of corticospinal
pathways is rich and extensive in both human
and nonhuman primates. These pathways tend to
Dynamic Properties of Coordinate affect interneurons that are located in the interme-
Transformation diate zone, and thus most of the descending corti-
cal command is first processed at this level, though
The concept of coordination transformation is it is possible that some of the command (e.g., a
often viewed in a “timeless” manner. In other finger-related command) is relayed directly to
words, the coordinate system of each neuron is MNs. The detailed scheme of termination of this
considered to be a fixed property that does not pathway could, in itself, provide a processing link
change over time. However, this is not necessar- for cortical commands. Any organization scheme
ily the case. During a behavioral trial the nervous that respects one descending parameter but ignores
system collects increasingly more information another will preserve the information related to the
about task-relevant parameters. In addition, the former but not the latter attribute. For example, if
relevance of external cues and the internal state all cortical neurons that contain similar PD termi-
can change along the trial. For example, at cue nate onto the same spinal target, the PD informa-
onset, the spatial information about the cue tion will be preserved at the postsynaptic level.
(which is often given for a short period of time) However, if the PD distribution of cortical cells
is critical for correct task performance, whereas that converge upon the same spinal target is uni-
the posture of the hand or other muscle-based formly distributed, the directional information will
parameters is less relevant at this stage. By con- be averaged out. This by no means implies that the
trast, at the “go” signal, any plan for movement spinal target will not be directionally tuned; sim-
relies on information about posture and muscle ply this tuning will not be driven by cortical
states. Indeed it was shown that the coordinate commands.
frame of cortical neurons (which unlike spinal Studies stimulating in the motor cortex while
neurons have access to information in different recording in the spinal cord uncovered the CS
modalities) gradually shifts during motor connectivity pattern in awake behaving primates.
Coordinate Transformations, Role of Spinal Circuitry cortical inputs and produce a motor command which is
in, Fig. 1 A Model of the functional organization of integrated and rotated into a muscle-based command. At
the CS system. (a) Motor cortical neurons integrate infor- this level the response properties are more uniform than at
mation about the motor targets (x1,. . .,xn) and about the the cortical level. To the left of each layer, a diagram
peripheral state (y1,..,ym) with time-varying weight con- illustrates the mean response profile of cells. (b) Frequen-
stants (Wx(t) and Wy(t)), respectively. This integration cies of different response patterns for cortical neurons
occurs at random and the resultant response properties of (orange), spinal interneurons (blue), and motoneurons
cortical neurons are heterogeneous (depicted as differ- (black). Response pattern was quantified using a phasic-
ences in the shading of cortical neurons). Lateral interac- tonic index which approached 1 for phasic cells and 0 for
tions can further modify the response properties and the tonic cells. Cortical cells had a broader distribution with
coordinate frame of the cortical layer. Cortical neurons are a significantly higher mean value (a tendency toward
characterized by a time-dependent response pattern that phasic firing), whereas spinal interneurons had
appears as a transient signal. Spinal interneurons integrate a narrower distribution with a lower mean value
These studies revealed a very broad CS connec- Role of Spinal Interneurons as a “Neural
tivity pattern with many cortical sites that affect Integrator”
spinal targets. Furthermore, interacting cortical
and spinal neurons express aligned directional- In the process of translating the cortical com-
torque tuning at movement onset, whereas mand into a signal to drive muscles, the spatial
unconnected sites have inverse tuning directions. coordinate frame as well as the temporal profile
This finding further suggests that although infor- of the task-related activity needs to be adjusted.
mation on directional torque is preserved in the Cortical neurons have often been shown to code
CS system, it is organized in a very broad man- time derivatives of motor parameters (e.g., veloc-
ner; namely, cortical neurons with vaguely simi- ity; Schwartz et al. 1988), yet at the pre-muscle
lar tuning will tend to affect a single spinal site. In level, this information needs to be integrated into
practice this result is consistent with a nonspecific the required shape of force profile. The alterna-
connectivity pattern so that a heterogeneous pop- tive that the intrinsic properties of MNs could by
ulation of cortical neurons converges upon spinal themselves integrate a transient motor command
circuitry (Fig. 1). This differs from the assump- for voluntary action seems unlikely given the
tion of a unique population of output motor cor- short time scales available for this operation
tical cells that deliver a muscle-based motor (Heckman et al. 2009). We found that when
command and are specifically connected with sustained torque is required, motor cortical firing
premotor spinal targets. is different from spinal firing in two main ways.
First, cortical firing is transient, whereas spinal the characteristics of motor synergies. This result
firing is tonic. Second, cortical neuron firing is is consistent with earlier studies that showed that
less informative than spinal neurons about static prolonged activation of the motor cortex in pri-
variables such as hand posture. Similar properties mates produces a finite repertoire of movements
have been reported for the eye movement system which mimic natural actions (i.e., defensive ges-
where cerebellar receiving neurons fire in relation tures, self-feeding, etc. ; Graziano and Aflalo
to eye acceleration, vestibular neurons fire in 2007). Could these results be interpreted as an C
diverse relations to eye movement parameters, indication that the motor primitives in fact reside
but the abducens MNs fire in relation to eye in the motor cortex? In that case, the translation
position (Joshua et al. 2013). In this system it from external to internal coordinate frame
was suggested that a “soft” feedforward hierar- appears less intuitive. For example, we would
chical network can act as an integrator (Joshua need to assume that motor cortical neurons that
et al. 2013; Miri et al. 2011). Similar events may are part of a single module randomly sample the
also take place in the corticospinal system where more externally based variable space, since oth-
cortical neurons exhibit fairly diverse firing prop- erwise the resultant motor activation might be
erties and broadly converge upon spinal neurons; biased. More data are needed to understand in
here, the needed integration is already present at what way the modularly organized system deals
the spinal interneuronal level. with coordinate transformation.
Modular Organization Versus Implications for BMI

Coordinate Transformation
In recent years many studies have demonstrated
A conceptual, data-based model of the functional that useful signals can be extracted from motor
organization of the motor system suggests that cortical activity to drive prosthetic devices
the motor system is characterized by a modular (Hatsopoulos and Donoghue 2009; Schwartz
organization principle (Bizzi et al. 2000, 2008). 2007). These devices range from an on-screen
In this scheme, motor actions are constructed by cursor to a humanoid robotic limb with multiple
the spatiotemporal superpositioning of basic degrees of freedom. The most common approach
modules that can be viewed as elementary build- employed in these studies is to assume a fixed
ing blocks of movements (often referred to as coordinate frame (either extrinsic or intrinsic) of
motor primitives). Each module provides an acti- these signals. However, findings have shown that
vation of a set of muscles (i.e., a muscle synergy). the activity of cortical neurons is affected by the
These modules were first extrapolated from work operative context. For example, hand posture
on spinal stimulations of spinalized frogs and induces a consistent bias of PDs across motor
appeared to offer an elegant way to mediate coor- cortical neurons, which thus cannot be eliminated
dinate transformations. In this model, a cortical by population averaging. Furthermore, trial time
activation signal (presumably expressed in an systematically affects the PDs of cortical neu-
external frame of coordinates) activates the nec- rons. Finally, it is conceivable that the behavioral
essary set of modules which in turn incorporate context may entail other parameters such as men-
the descending command into the ongoing motor tal states which are difficult to control for but may
state. Although considerable data on motor nevertheless affect the PD of cortical neurons in
behavior supported the existence of a robust set a consistent manner that cannot be eliminated by
of motor synergies, their neural substrate averaging. In this respect using motor cortical
remained unclear. More recently, it was shown activity as part of a long-term control signal
that electrical activation of the motor cortex in should be consistently updated with an adaptive
primates produces motor responses which have algorithm that reweighs the activity of different
cells to prevent a gradual deviation between the Heckman CJ, Mottram C, Quinlan K, Theiss R, Schuster
intended and the reconstructed actions. J (2009) Motoneuron excitability: the importance of
neuromodulatory inputs. Clin Neurophysiol
120:2040–2054
Cross-References Hikosaka O (1998) Neural systems for control of volun-
tary action–a hypothesis. Adv Biophys 35:81–102
Jankowska E (1992) Interneuronal relay in spinal path-
▶ Brain-Machine Interface: Overview ways from proprioceptors. Prog Neurobiol
▶ Cortical Motor Prosthesis 38:335–378
▶ Decision-Making, Motor Planning Joshua M, Medina JF, Lisberger SG (2013) Diversity of
neural responses in the brainstem during smooth pur-
▶ General Overview of Spinal Anatomy and suit eye movements constrains the circuit mechanisms
Physiology Organization of neural integration. J Neurosci 33:6633–6647
Kakei S, Hoffman DS, Strick PL (2003) Sensorimotor
transformations in cortical motor areas. Neurosci Res
References 46:1–10
Kalaska JF (2009) From intention to action: motor cortex
Asher I, Zinger N, Yanai Y, Israel Z, Prut Y (2010) and the control of reaching movements. Adv Exp Med
Population-based corticospinal interactions in Biol 629:139–178
macaques are correlated with visuomotor processing. Kuypers HG (1987) Some aspects of the organization of
Cereb Cortex 20:241–252 the output of the motor cortex. Ciba Found Symp
Averbeck BB, Lee D (2004) Coding and transmission of 132:63–82
information by neural ensembles. Trends Neurosci Lee D, Port NL, Kruse W, Georgopoulos AP (1998) Var-
27:225–230 iability and correlated noise in the discharge of neu-
Baldissera F, Hultborn H, Illert M (1981) Integration in rons in motor and parietal areas of the primate cortex.
spinal neuronal systems. In: Mountcastle JM, J Neurosci 18:1161–1170
Brookhart VB (eds) The nervous system. American Miri A, Daie K, Arrenberg AB, Baier H, Aksay E, Tank
Physiological Society, Bethesda, pp 509–95 DW (2011) Spatial gradients and multidimensional
Bizzi E, Tresch MC, Saltiel P, d’Avella A (2000) New dynamics in a neural integrator circuit. Nat Neurosci
perspectives on spinal motor systems. Nat Rev 14:1150–1159
Neurosci 1:101–108 Porter R, Lemon RN (1993) Corticospinal function and
Bizzi E, Cheung VC, d’Avella A, Saltiel P, Tresch voluntary movement. Calrendon Press, Oxford
M (2008) Combining modules for movement. Brain Prut Y, Fetz EE (1999) Primate spinal interneurons show
Res Rev 57:125–133 pre-movement instructed delay activity. Nature
Dum RP, Strick PL (2002) Motor areas in the frontal lobe 401:590–594
of the primate. Physiol Behav 77:677–682 Sabes PN (2011) Sensory integration for reaching:
Evarts EV, Shinoda Y, Wise SP (1984) Neurophysiologi- models of optimality in the context of behavior and
cal approaches to higher brain functions. Wiley, New the underlying neural circuits. Prog Brain Res
York 191:195–209
Fetz EE (1992) Are movement parameters recognizably Schwartz AB (2007) Useful signals from motor cortex.
coded in the activity of single neurons. Behav Brain J Physiol 579:581–601
Sci 15:679–690 Schwartz AB, Kettner RE, Georgopoulos AP (1988) Pri-
Fetz EE, Perlmutter SI, Prut Y, Maier MA (1999) Primate mate motor cortex and free arm movements to visual
spinal interneurons: muscle fields and response prop- targets in three- dimensional space. I. Relations
erties during voluntary movement [in process citation]. between single cell discharge and direction of move-
Prog Brain Res 123:323–330 ment. J Neurosci 8:2913–2927
Georgopoulos AP, Schwartz AB, Kettner RE (1986) Neu- Scott SH (2004) Optimal feedback control and the neural
ronal population coding of movement direction. Sci- basis of volitional motor control. Nat Rev Neurosci
ence 233:1416–1419 5:532–546
Georgopoulos AP, Lurito JT, Petrides M, Schwartz AB, Scott SH, Kalaska JF (1997) Reaching movements with
Massey JT (1989) Mental rotation of the neuronal similar hand paths but different arm orientations.
population vector. Science 243:234–236 I. Activity of individual cells in motor cortex.
Graziano MS, Aflalo TN (2007) Mapping behavioral rep- J Neurophysiol 77:826–852
ertoire onto the cortex. Neuron 56:239–251 Shadmehr R, Wise SP (2005) The computational neurobi-
Harel R, Asher I, Cohen O, Israel Z, Shalit U et al (2008) ology of reaching and pointing: a foundation for motor
Computation in spinal circuitry: lessons from behav- learning. MIT Press, Cambridge, MA. xvii, p 575
ing primates. Behav Brain Res 194:119–128 Shah A, Fagg AH, Barto AG (2004) Cortical involvement
Hatsopoulos NG, Donoghue JP (2009) The science of neu- in the recruitment of wrist muscles. J Neurophysiol
ral interface systems. Annu Rev Neurosci 32:249–266 91:2445–2456
Coordination Dynamics 853 C
Todorov E, Jordan MI (2002) Optimal feedback control as Detailed Description
a theory of motor coordination. Nat Neurosci
5:1226–1235
Toyoshima K, Sakai H (1982) Exact cortical extent of the Coordination
origin of the corticospinal tract (CST) and the quanti- That the different limbs are coordinated during
tative contribution to the CST in different cytoarchi- locomotion seems obvious, although the phenom-
tectonic areas. A study with horseradish peroxidase in enon of relative coordination, in which one limb
the monkey. J Hirnforsch 23:257–269
Zach N, Inbar D, Grinvald Y, Bergman H, Vaadia makes an extra cycle every now and then, shows C
E (2008) Emergence of novel representations in pri- that coordination requires a quantitative approach
mary motor cortex and premotor neurons during asso- (Holst 1973). Rhythmic movements like dancing
ciative learning. J Neurosci 28:9545–9556 and making music, but also movements like
chewing and speaking that are approximately
rhythmic, are coordinated in the sense that the
relative timing of movement components remains
Coordination Dynamics invariant in time and, within limits, across
changes of movement parameters such as speed
Gregor Schöner and Eva Nowak and amplitude (Schöner 2002). Direct evidence
Institut f€
ur Neuroinformatik, Ruhr-Universit€at for stable relative timing is obtained when a coor-
Bochum, Bochum, Germany dination pattern recovers from a perturbation, the
lagging component catching up and the leading
component falling back to reinstate the
Synonyms pre-perturbation pattern (Schöner and Kelso
1988). Discrete motor acts are also coordinated
Haken-Kelso-Bunz model in the sense that their durations are aligned when
the movements are performed concurrently, but
not when they are performed separately (Kelso
Definition et al. 1979). The coordination of movements with
external events, for example, intercepting
Voluntary movements are typically coordinated a moving object may show the same characteris-
in the sense that their components form ordered tics of relative timing (Warren 2006).
patterns. This includes the coordination of differ- To assess the convergence of coordination
ent spatial components of a movement, the coor- patterns that defines their stability, a distance
dination of different limbs, and also the measure among patterns of coordination is
coordination of movement with perceived events. needed. The relative phase is such a measure. It
More formally, coordination is characterized by is defined as the temporal offset of matching
stable relative timing of the movement compo- events in two timed movements, measured in
nents. Coordination dynamics is a theoretical percent of the movement time (Schöner 2002).
approach to understanding how coordination
arises which postulates that neural oscillators Neural Oscillators
control the timing of each component. The cou- A key theoretical idea for understanding coordi-
pling among such neural oscillators leads to sta- nation is that the timing of voluntary motor acts is
ble relative timing. Mathematical models of generated from neural oscillators, whose stable
coordination dynamics have been formulated at limit cycles drive the effector systems (Schöner
the level of the neural oscillators and their cou- 2002). This idea is not limited to rhythmic move-
pling, but also at the level of relative phase itself ment. For discrete motor acts, the limit cycle may
as a macroscopic, phenomenological variable. be instantiated via a bifurcation at the beginning
The observation of instabilities in relative timing of the movement and terminated by another
has provided support to the notion of coordina- bifurcation at the end of a single cycle (Schöner
tion dynamics. 1990). Perceived events may drive these
C 854 Coordination Dynamics
dφ/dt = f(φ)
relative phase
neutrally stable
φ
relative phase
stabilized
by coupling
Coordination Dynamics, Fig. 1 The rate of change of how the negative slope of the relative phase dynamics
relative phase, df/dt, plotted against relative phase, f, has leads to convergence to the fixed point which accounts
a set of marginally stable fixed points when the underlying for recovery from perturbations and resistance to noise.
oscillators are weakly coupled (blue solid line) that coa- The loss of stability for weak coupling leads to predictions
lesces into a single-stable fixed point for sufficiently of increased variance and relaxation time (Schöner
strong coupling (red solid line). The red arrows illustrate et al. 1986)
bifurcations, accounting for the timing of discrete patterns of coordination (homologous limbs
motor acts relative to the environment (Schöner alternate) when movements become faster
1994). (Schöner and Kelso 1988).
Coordination is brought about by coupling The coupled neural oscillators themselves
multiple such oscillators. Coupling generically have also been used as a level of description of
induces phase locking (see entry on Multistability coordination patterns (Haken et al. 1985;
of Coupled Neural Oscillators). This can be visu- Grossberg et al. 1997) to account for the role of
alized by looking at the rate of change of relative movement amplitude. Behavioral data suggest
phase as a function of relative phase (Fig. 1). that what determines the stability of the coordi-
Limit cycle oscillators always have one nation pattern may be the spatial arrangement of
Lyapunov exponent that is zero, which reflects movement components, not necessarily the ana-
that there is no resistance to perturbations that tomical nature of the components (Mechsner
shift the oscillator along the limit cycle. In et al. 2001). If this were true, then perhaps corti-
uncoupled oscillators, the dynamics of the rela- cal oscillators that control movement in space,
tive phase has a line of fixed points that are rather than spinal oscillators directly linked to the
marginally stable. Coupling most strongly affects effectors, would be the basis for coordination.
the direction in phase space, in which the vector
field is zero, leading generically to a stable fixed-
point attractor for relative phase that represents Cross-References
a stable pattern of coordination.
▶ Bifurcations, Neural Population Models and
Link to Experiment ▶ Embodied Cognition, Dynamic Field Theory of
The dynamics of relative phase has been used as ▶ Multistability of Coupled Neuronal Oscillators
a direct phenomenological model of the dynam-
ics of coordination (Schöner et al. 1986). In par- References
ticular, experimental evidence for loss of
stability, observed through enhanced fluctuations Grossberg S, Pribe C, Cohen MA (1997) Neural control of
and slowed relaxation after perturbations, has interlimb oscillations. I. Human bimanual coordina-
tion. Biol Cybern 77:131–140
highlighted that stability is a necessary concept Haken H, Kelso JAS, Bunz H (1985) A theoretical model
for an understanding of coordination. Loss of of phase transitions in human hand movements. Biol
stability is typically observed for antiphase Cybern 51:347–356
Correlation Analysis of Parallel Spike Trains 855 C
Holst E (1973) Relative coordination as a phenomenon applied to one of the signals. In case the two
and as a method of analysis of central nervous func- signals are simultaneously recorded spike trains,
tion. In: The behavioral physiology of animals and
man: the collected papers of Erich von holst, the cross-correlation becomes a count of the num-
vol 1. University of Miami Press, Coral Gables, ber of coincidences of firing for the two spike
pp 33–135 trains as a function of the time delay between
Kelso JAS, Southard DL, Goodman D (1979) On the them. If one considers one spike train as the
nature of human interlimb coordination. Science
203:1029–1031 input to a system and the other spike train as the C
Mechsner F, Kerzel D, Knoblich G, Prinz W (2001) Percep- output, then the cross-correlation function
tual basis of bimanual coordination. Nature 414:69–73 between the input and output spike trains normal-
Schöner G (1990) A dynamic theory of coordination of ized on the input autocorrelation function is equal
discrete movement. Biol Cybern 63:257–270
Schöner G (1994) Dynamic theory of action-perception to the impulse response of the linear part of the
patterns: the time-before-contact paradigm. Hum Mov system. Cross-correlation can also, and in the
Sci 3:415–439 nervous system more generally, result from com-
Schöner G (2002) Timing, clocks, and dynamical systems. mon input to the two spike trains, i.e., from pro-
Brain Cogn 48:31–51
Schöner G, Kelso JAS (1988) Dynamic pattern generation in viding a sensory stimulus or from rhythmic or
behavioral and neural systems. Science 239:1513–1520 other spontaneous activity in the brain.
Schöner G, Haken H, Kelso JAS (1986) A stochastic the-
ory of phase transitions in human hand movement.
Warren WH (2006) The dynamics of perception and Detailed Description
action. Psychol Rev 113(2):358–389
I will use the definition of cross-correlation as
Further Reading standard in the field of linear systems analysis
Grimme B, Fuchs S, Perrier P, Schöner G (2011) Limb as applied to the interaction between two neurons
versus speech motor control: a conceptual review. (Marmarelis and Marmarelis 1978). I will (except
Motor Control 15:5–33
when noted) assume stationarity of the spike
trains. The cross-correlation function reflects the
similarity of two spike trains x(t) and y(t), or in
sampled form x(n) and y(n), as a function
Correlation Analysis of Parallel of the delay between them, expressed as a time
Spike Trains delay t or as the number of bins k. T is the
duration of the spike trains, and N is the number
Jos J. Eggermont of bins in the spike train. Note that the time delay
Department of Physiology and Pharmacology, can be positive and negative.
University of Calgary, Calgary, AB, Canada
ðT
Department of Psychology, University of 1
Rxy ðtÞ ¼ ðxðtÞyðt þ tÞdt or in sampled form :
Calgary, Calgary, AB, Canada T
0
1X N
Rxy ðkÞ ¼ xðnÞyðn þ kÞ
Synonyms N n¼1
(1)
Conditional rate function; Cross-correlogram;
Linear system kernel The cross-covariance function is obtained by
subtracting the means mx and my from the two
signals (I omit the sampled form here):
Definition
ðT
1
Cross-correlation is a measure of the similarity of Cxy ðtÞ ¼ ððxðtÞ mx Þ yðt þ tÞ my dt (2)
T
two signals as a function of the time lag or lead 0
C 856 Correlation Analysis of Parallel Spike Trains
The cross-correlation coefficient function is i.e., the number of standard deviations above the
obtained by dividing the cross-covariance func- mean-corrected R, which is equal to zero:
tion by the time-average variances, s2x and s2y , of
the two spike trains (again the sampled form is rffiffiffiffiffiffiffiffiffiffiffi
NxNy NxNy
omitted): zxy ðtÞ ¼ Rxy ðtÞ = (6)
N N
ðT
1 qffiffiffiffiffiffiffiffiffiffi
rxy ðtÞ ¼ ððxðtÞ mx Þ yðt þ tÞ my dt= s2x s2y Statistical significance should be set conserva-
T
0 tively, e.g., at z 4.
(3) The standard normal distribution that is the
basis for the interpretation of z, as the number of
The expressions in the above equations are standard deviations between a value x and the
defined empirical estimates of the underlying sta- mean m of the distribution, is based on the
tistics. Any deviation from a value of 1 for the z-score z = (xm)/s. For neural cross-
correlation coefficient is due to (1) nonlinearity correlograms we generally do not have normal
of the system, (2) the presence of noise in the distributions, but we assume that the bin filling
spike generation of the system, and (3) the pres- with coincidences follows a Poisson process
ence of other spike train inputs to the system that (the bin size should be so small that the probabil-
contribute to the output spike train y(t) but are ity of >1 coincidence is vanishingly small). For
not accounted for by the observed input spike a Poisson process the standard deviation is the
train x(t). square root of the mean, leading to the standard
deviation of the number of coincidences
qffiffiffiffiffiffiffiffi
Representation of the Cross-Correlogram in Nx Ny
sxy ¼ N .
Terms of Coincidences
In case of Poisson spike trains, the cross- If one wants to assess significance based on
correlation coefficient becomes (Eggermont the cross-correlation coefficient, the standard
deviation thereof is equal to

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1992)

N N
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi SD rxy ¼ N1 1 Nx 2 y . For large numbers
u !ffi
NxNy u t N 2x N 2y
of bins in the record and relatively low firing
rxy ðtÞ ¼ Rxy ðtÞ = Nx Ny
N N N rates, this can be approximated as

qffiffiffi
(4) SD rxy ¼ N1 . So for a 900 s record length
and 2 ms bin width, D, the SD = 0.0015. So
Here, Nx and Ny are the number of spikes in
taking 4 SD would make rxy values above 0.006
spike trains x(t) and y(t), respectively, and N is
significant, but one can wonder whether values
the number of bins in the total spike train dura-
N N close to 0.006 are relevant. For larger numbers of
tion. xN y is the expected number of coincidences
spikes, the SD decreases compared to this
in case of independence. For small numbers of
estimate.
spikes compared to the number of bins (N) in the
record, i.e., low firing rates, this can be approxi-
mated by Representations of the Cross-Correlogram in
Terms of Firing Rate
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi The cross-correlogram is sometimes defined as a
NxNy
rxy ðtÞ ¼ Rxy ðtÞ = NxNy (5) first-order conditional rate function Ry|x(t) =
N
Rxy(t)/mx (where mx is the mean firing rate for
spike train x) that measures, for one cell y and
For statistical purposes this can also be close to any particular time t, the average instanta-
converted (holds for all firing rates) to a z-score, neous rate or the likelihood of generating a spike,
conditional on an x spike t time units away recurrence times. One observes that with the
(Brillinger et al. 1976). If the trains are indepen- exception of the first 10 s or so, the occurrence
dent, then Ry|x(t) = my, where my is the mean firing of the recurrence times is stationary. The most
rate for spike train y, for all t. If y spikes are important events that determine the various order
independent of “later” x spikes (causality), then cross-correlograms here (shown in part c, d) are
Ry|x(t) = my for all t < 0. This “causality” may the first-order recurrences, i.e., there are hardly
apparently be violated if both x and y spikes are any bursts (indicated by ■ and symbols) in the C
caused with some jitter by a third, unobserved, y-train (a) but quite a few in the x-train (b), which
input. Deviation of Ry|x(t) from my is suggestive had the higher firing rate. For calculation of the
of dependency of the y-train on what happened in correlograms, we assume periodic stationarity:
the x-train t time units earlier. For all natural spike this means that for each stimulus presentation at
trains, Ry|x(t) will be flat and essentially equal to my intervals of P sec, the response of the neurons is
at large |t| values because inevitably (with the statistically the same, albeit it is nonstationary in
exception for oscillatory correlograms) any influ- firing rate. By using ensemble averaging, i.e., in
ence of x on y will have vanished. p Assuming
ffiffiffiffiffiffiffiffiffiffiffiffiffi Eqs.1, 2, and 3, the time variable is now t modulo
a Poisson distribution, one
pffiffiffiffiffiffiffiffiffiffiffiffiffi can graph Ryjx ðtÞ . P and stimulus-dependent cross-correlations can
The variance of Ryjx ðtÞ is approximately con- be obtained.
stant for all t at a level of (4DTmx)1, so SD lines The top correlogram is that based on only first-
can be drawn at (4DTmx)2 and become smaller order recurrence times, the second one on
in proportion to √T. This provides a second-order recurrence times, and the third one
good approximation if D is not too large and the on third-order recurrence times; the fourth shows
correlation width of the process is not too large the all-order cross-correlogram as commonly
(Voigt and Young 1990). used. All correlograms are on the same scale.
The shaded correlograms are the shift predictors
(see next section), calculated from the time-
Time-Dependent Cross-Correlation Functions dependent shift correlograms shown in the bot-
Visual inspection of the time-dependent cross- tom row that show only modest stimulus locking
correlation function, Rxy(t,t), with t as the run- to the stimulus. The shape of the cross-
ning time in the spike trains and t as the lag-lead correlogram suggests common input as the main
time between spikes in trains x and y, shown here source of correlated spike times (see later in
in an example (Fig. 1), is often the best first step this entry).
in judging stationarity. Rxy(t,t) without any aver-
aging (panels a and b) represents the recurrence
times (Van Stokkum et al. 1986) between the two
spike trains x(t) and y(t) as a function of t. Panels Pair Correlation Under Stimulus
c and d represent the time-averaged Rxy(t). The Conditions
example reflects simultaneous recordings from
two neurons in the auditory midbrain of the Euro- A peak in the cross-correlogram can be the result
pean grass frog in response to species-specific of a neural interaction between units (“noise cor-
vocalizations. The left-hand and right-hand col- relation”) but also the result of a coupling of
umns refer to the same data, but on the left-hand the firings of the neurons to a stimulus (“signal
side, the correlation shown is Rxy(t,t), whereas correlation”) (Gawne and Richmond 1993).
the right-hand side shows Ryx(t,t), i.e., the refer- As we will see, the interpretation of the
ence unit has changed. cross-correlogram will require both the
Vertically, the time over the first 150 s since autocorrelograms of the two units and an estimate
stimulus onset, t, is shown, and horizontally the of the correlation due to stimulus coupling. Fol-
lag-lead times, t, between the x and y spikes. lowing our earlier approach (Eggermont
Different symbols indicate different order et al. 1983), we consider two spike trains x(t)
Correlation Analysis of Parallel Spike Trains, nonsimultaneous (shaded) recurrence-time histograms of

Fig. 1 Time-dependent cross-correlation diagrams of order 1, 2, and 3, and cross-correlation histogram derived
spike trains. The different symbols used indicate the from (a) and (e), respectively; (d) as (c) but now derived
order of the spikes following the reference spike: “|” from (b) and (f); (e) nonsimultaneous time-dependent
indicates the first-order recurrence time (defined in Van cross-correlation diagram (defined as the correlation
Stokkum et al. 1986), i.e., the first spike in the y-train between the 1 s duration spike train x to stimulus presen-
following or preceding a spike in the x-train (for the left- tation at t and spike train y to the next stimulus presenta-
hand column); “” indicates the second-order recurrence tion at time t + 1 s), spikes from unit x are used as triggers;
time, i.e., the second spike in the y-train following or (f) as (e), but now with spikes from unit y as triggers (With
preceding a spike in the x-train; and finally “ ” indicates kind permission from Springer Science + Business
the third-order recurrence time. (a) Simultaneous time- Media: Biological Cybernetics, Representation of Time-
dependent cross-correlation diagram (no averaging Dependent Correlation and Recurrence Time Functions,
involved), spikes from unit x (2,188 events) are used as volume 55, 1986, page 17–24, I. H. M. van Stokkum,
triggers; (b) as (a) but now with spikes from unit P. I. M. Johannesma, and J. J. Eggermont, Fig. 6)
y (528 events) as triggers; (c) simultaneous and
and y(t) of duration T and represented as 1X N X M

a sequence of d-functions: Rxy ðtÞ ¼ d t ti tj (8)
T i¼1 j¼1
X
N X
M
x ðt Þ ¼ dðt ti Þ and yðtÞ ¼ d t tj
i¼1 j¼1
For a bin width D and expectation E = mxmy of
(7) Rxy(t) under independence of firing of the two
where ti and tj are the spike times. The cross- spike trains, a cross-covariance histogram is
correlation function is given by defined:
Correlation Analysis of Parallel Spike Trains, (e.g., Gawne and Richmond 1993). Cross-
Table 1 Overview of possible covariance histograms correlations obtained under spontaneous firing
for a double-unit recording with two presentations of the
stimulus ensemble conditions or after appropriate correction for cor-
relation resulting from stimulus locking will be
x1(t) y1(t) x2(t) y2(t)
called neural correlation, they exclusively are
x1(t) Cx1x1 Cx1y1 Cx1x2 Cx1y2
noise correlations. During stimulation the “total”
y1(t) Cy1y1 Cx2y1 Cy1y2
neural synchrony consists of a component due to C
x2(t) Cx2x2 Cx2y2
stimulus synchrony and one that is a result of
y2(t) Cy2y2
neural connectivity, the neural correlation.
The nonsimultaneous cross-covariance histo-
ð
tþD=2
1 grams between spike trains x1(t) and y2(t), respec-
Cxy ðt, DÞ ¼ ðRxyðsÞ E ds (9) tively, and x2(t) and y1(t) represent correlation
D
tD=2 between spike trains x(t) and y(t) as a result of
stimulus coupling. These stimulus-induced cor-
If one presents the stimulus twice, then the relations are known as the shift predictors.
resulting spike trains are x1(t) and y1(t) for the Another nonsimultaneous cross-covariance his-
first stimulus presentation, respectively, and x2(t) togram is obtained by taking both spike trains
and y2(t) for the second one. Presenting the stimu- for the same unit to obtain Cx1x2 and Cy1y2.
lus twice allows us to estimate the correlation These histograms represent the amount of cou-
between the firing of the two spike trains based pling of the units to the stimulus individually and
on their time locking to the stimulus. In the text have been called “existence functions” (Aertsen
I assume a long stimulus consisting of various et al. 1979) because any significant peak around
elements. The predictor for that is to correlate t = 0 indicates a stimulus coupling effect and
spike train x for stimulus presentation 1 with hence the existence of a stimulus–response rela-
spike train y for stimulus presentation 2, under tion. This shuffled autocorrelogram technique
the assumption that memory effects have long was rediscovered by Joris et al. (2006).
subsided between the two stimulus presentations. There is no model-free approach for the sep-
This result is called the shift predictor. For those aration of correlations produced by direct neural
four spike trains, one can compute ten covariance interaction and those caused by stimulus cou-
histograms (or correlograms) as shown in Table 1. pling. Under the assumption that the effects of
On the main diagonal two estimates, each of an external stimulus and the effects of neural
the units’ autocovariance histograms is found. connectivity to the neural synchrony are
Under long-term stationary firing conditions, additive, Perkel et al. (1967) proposed a
Cx1x1 = Cx2x2 and one commonly uses correction for the effects of the stimulus. Two
Cxx = (Cx1x1 + Cx2x2)/2 as an estimator of the formally identical stimulus predictors were
autocovariance function for spike train x(t). For suggested, one resulting from a cross-
the off-diagonal elements, it is noted that, e.g., correlation of the two single-unit poststimulus-
Cx1y1(t) = Cy1x1(t), therefore the table is time histograms (PSTHs), which can be used
symmetric provided that the sign of t is when the stimulus is periodic. One constructs
adjusted. We define Cx1y1(t) and Cx2y2(t) as the PSTHs, i.e., the cross-correlograms between
simultaneous cross-covariance histograms, and stimulus onsets and spikes, RSx(s) and RSy(s),
again we may use their average as an estimate and calculates their correlation integral modulo
for the cross-covariance function. Following our stimulus period P (here s is the poststimulus
previously introduced terminology (Eggermont time):
et al. 1983; Epping and Eggermont 1987),
cross-correlations obtained under stimulus con- ðP
ditions will be called neural synchrony, they RScorr ðtÞ ¼ RSx ðsÞRSy ðs þ tÞds (10)
will be a mix of signal and noise correlations 0
One can also calculate the cross-correlogram obtained, there are a number of possibilities: the
between one spike train and a shifted (by 1 stim- interneuronal coupling itself changes due to
ulus period) version of the second spike train. application of a stimulus (facilitation, depression,
This latter procedure, resulting in the shift pre- short-term learning), the finding is an artifact due
dictor, has been the most popular and has the to the selection of the wrong model, or the finding
advantage that it can be applied to very long can be explained on the basis of a pool of neurons
stimuli (such as a set of animal vocalizations) affecting the firing behavior of the pair under
that allow only one or a few repetitions; in such study in a stimulus-dependent way.
a case calculating the PSTH would be useless. It must be remarked that the additive
The shift predictor was assumed to represent the model (Perkel et al. 1967) is the only one rou-
amount of correlation that could be attributed to tinely used in stimulus-correction procedures
the locking of the neuron’s firings to the stimulus. (e.g., Voigt and Young 1990; Eggermont 1994).
Subtracting the shift predictor from the neural However, Srinivasan and Bernard (1976) have
synchrony then would result in the neural shown that coincidence-detecting neurons can
correlation. fire at a rate proportional to the product of the
The assumption of additivity of stimulus- rates of two input neurons, provided that their
induced correlation and neural correlation may spike trains are statistically independent. The
well be the weakest link in the chain of reasoning criteria determining whether the action is addi-
that leads to an interpretation of stimulus- tive or multiplicative are the size of the excit-
dependent neural correlations. A strong warning atory postsynaptic potential (EPSP) relative to
against the use of the shift predictor in estimates threshold and the time constant of decay for the
of the neural interaction strength came from stud- EPSP. Specifically, EPSP values smaller than
ies using realistic model neurons (Melssen and the threshold of firing and with a duration less
Epping 1987). Because stimulus effects and neu- than the average interspike interval of the indi-
ral connectivity effects on the neural synchrony vidual spike trains cause the action to be multi-
are generally nonadditive, it was concluded that plicative. Larger EPSP values or those with
the use of the shift predictor to separate stimulus longer duration make the action additive. An
and neural effects was of limited value. Yang and alternative formulation from which this effect
Shamma (1990) underlined these objections to can be understood is that we consider a spike
the use of correlation methods in identifying neu- generator with probability of firing, g, which
ral connectivity and arrived at conclusions simi- depends sigmoidally on the EPSP value
lar to those of Melssen and Epping (1987) using u (Eggermont et al. 1983).
simulations with similar model neurons. In case
of strong stimulus locking, even all synchrony gðzjuÞ ¼ expbu=ð1 þ expbuÞ (11)
can be due to stimulus correlation (Eggermont
1994). Under “adequate” stimulus conditions Here u is simply considered to be the linear
that evoke strongly stimulus-driven and repro- summation of contributions ui caused by incom-
ducible spike trains, especially for auditory stim- ing spikes zi possibly originating from different
uli, the neural synchrony becomes identical to the projecting neurons. Actually g(z|u) describes the
expected correlation based on the shift predictor, well-known S-shaped curve. For relatively
and consequently no inference can be made about small u values, i.e., in the purely exponential
possible underlying connectivity (changes). One region of the curve and temporal overlap of u1
may avoid this overcorrection effect due to stim- and u2, we have:
ulus locking by applying suboptimal stimulation.
Thus, a sufficient amount of “neural noise” in the
gðzju1 , u2 Þ ¼ expbðu1 þ u2 Þ=ð1 þ expbðu1 þ u2 ÞÞ
record is required to allow conclusions about
changes in functional neural connectivity. Thus, expbu1 • expbu2 gðzju1 Þ • gðzju2 Þ
when stimulus-dependent neural correlations are (12)
If the u values are large enough to be in the Practical Estimation of Stimulus Correlation
linear, intermediate, part of the S-curve, then During spontaneous activity one estimates the
after expansion of the exponential and long expected value of the cross-correlogram under
division: the assumption of independence and subtracts
this in order to obtain the neural correlation. Dur-
1 ing stimulation, the effect of a common input to
gðzjuÞ þ 1=4bu (13) both spike trains as a result of a correlation of the C
2
neuronal firing to the stimulus has to be taken into
and in case u1 and u2 are nonoverlapping (low account if one wants an estimate of the true neural
spike rates), averaging over time gives correlation. If the stimulus is periodic, one can,
for instance, use the shift predictor obtained by
calculating the cross-correlation between spike
gðzju1 , u2 Þ gðzju1 Þ þ gðzju2 Þ (14)
trains x(t) and y(t + P), i.e., between the original
x-train and the y-train shifted over one stimulus
Thus, simply on basis of the size u and dura- period. This method is the only one that can be
tion of the EPSP, the firing probability of the used if a long stimulus sequence, e.g., a sequence
output neuron can be multiplicative Eq. 13 or of vocalizations or other natural sounds, is only
additive Eq. 14 with respect to the input spike repeated once, as discussed above.
trains. In fact u = u(t) and additional conditions As an example of this latter stimulus-
regarding the decay time constant may enter the correction procedure, we show recordings from
description. the auditory midbrain to 500 ms long 30 Hz
Let us now consider that a multiplicative amplitude-modulated noise-burst stimulation
effect of the stimulus on spike train activity is repeated every 3 s (Fig. 2). The shift predictor is
present, e.g., by two input spike trains both shaded and shows the 30 Hz modulation in
depending on the stimulus and fulfilling the con- the firing rate that obviously is stimulus locked.
ditions for the EPSP but only one of these input The triangular shape of the predictor is the
trains is observed together with the output train. result of the approximately rectangular enve-
In such a case the degree of multiplicative action lope for the 500 ms long noise burst. The neural
is strongly dependent on the type of stimulus. correlation consists of a narrow zero-lag cen-
Stimuli such as continuous noise that cause the tered peak approximately 5 ms in width (W) at
input spike trains to have a relatively large time half peak amplitude.
jitter, i.e., have a broad existence function,
Cii’(z), (Aertsen et al. 1979), produce results Rate Correlation and Event Correlation
that better fit a multiplicative model than stimuli Cross-correlograms frequently show a narrow
that tend to lock the spikes to a considerable peak on a broader pedestal (e.g., Fig. 2a) espe-
degree (e.g., tone pips) as suggested by the cially when units are recorded on the same elec-
Srinivasan and Bernard (1976) model results. trode (Eggermont 1992; Nowak et al. 1995). The
Thus, also multiplicative interaction effects broader pedestal is typically the result of covari-
may be stimulus dependent. In case they are ance in firing rate between the two spike trains,
not, the application of a correction procedure e.g., caused by the stimulus, whereas the sharp
based on an additive model, however, errone- peak indicates event synchrony, i.e., firing coin-
ously will indicate a stimulus dependence of cidences (Neven and Aertsen 1992). The rate
the neural correlation (Eggermont 1994). covariance under spontaneous conditions can
Eggermont (2006) introduced a method to cor- easily be estimated by calculating a cross-
rect for these multiplicative effects, which are correlogram for 50 ms bins that slide in 10 ms
substantial when cortical network oscillations steps (Eggermont and Smith 1995). Subtracting
are present and will be discussed later on in the rate covariance from the overall correlogram
this entry. will result in the event correlation.
correlation function, which is likely to be

wrongly interpreted as spike coordination.”
Stimulus-Dependent Neural Correlation

An example, exhibiting stimulus-dependent
correlation, is provided in Fig. 3. As discussed
above, this may potentially result from using
the invalid assumption of “neural synchrony
equals the sum of stimulus correlation and
neural correlation.” The two spontaneously
active units shown were recorded in the auditory
midbrain of the grass frog on electrodes with an
estimated tip separation of 130 mm. The sponta-
neous rates were 0.34 (A) and 2.5 (B) spikes/s,
respectively. Their spectro-temporal sensitivi-
ties, represented as dot displays of neural activ-
ity, as determined with tone pips presented once
per second, are shown in Fig. 3a, b. The spon-
Correlation Analysis of Parallel Spike Trains,
taneous activity of unit a was suppressed by
Fig. 2 (a) Simultaneous and nonsimultaneous (i.e., the
shift predictor; shaded) cross-correlation histograms for frequencies below 1,000 Hz (Fig. 3a), and the
two units recorded from the auditory midbrain in the grass very low spontaneous activity did not result in
frog under amplitude-modulated noise-burst stimulation many spikes in the 256 ms time window,
of 500 ms duration. The triangular shift predictor results
whereas unit b was activated in this frequency
from the approximately rectangular stimulus envelope.
A 30 Hz stimulus-locked oscillation is visible as well. range with a latency of 67 ms (Fig. 3b). Neural
(b) and (c) difference histogram shown on different time synchrony histograms and shift predictors
scales; W is the width of the peak at half amplitude. Bin (shaded) of spontaneous and stimulus-evoked
width was 4 ms in (a) and (b), and 1 ms in (c). The number
activity are shown in Fig. 3c–f. For spontaneous
of spikes for each unit is indicated in the top right of panel
A (From Epping and Eggermont (1987) firing the neural synchrony and shift predictor
are flat and hardly different (Fig. 3c). For tone-
pip stimulation the shift predictor (Fig. 3d)
This generally assumed model, even if cor- shows a decrement due to the antagonistic
rect, may still result in an inaccurate estimation (suppression vs. activation) stimulus influence
of the true event correlation. Staude et al. (2008) on the two units. Just visible on top of the shift
decomposed the raw cross-correlation as the predictor is a small, rather symmetric peak
sum of the modulation rate covariance and around the origin, likely indicating shared neu-
spike coordination based on the mean condi- ral input. The shift predictor to sinusoidally
tional covariance of the processes given the amplitude-modulated (AM) 200 ms tone bursts
rates. Both, rate covariation and spike coordina- is broad (Fig. 3e) with a sharp (W = 6 ms) peak
tion alike, contribute to the raw correlation func- of the neural synchrony visible on top. Figure 3f
tion by triangular peaks. This separation does shows results for a low-pass noise modulated
not always have to result in a broad triangular tone. One observes an asymmetrical peak with
peak for rate correlation and a narrow one for flanking valleys due to the autocorrelation struc-
spike correlation; at least in their procedure the ture of the individual spike trains. The two units
reverse can be possible as well. Staude responded antagonistically to the tone-pip stim-
et al. (2008) stated “without prior knowledge of ulus, but more alike to the other stimulus ensem-
the underlying model class, rate estimation will bles; e.g., when stimulated with sinusoidally
most likely be suboptimal. The resulting predic- AM tone bursts (Fig. 3e) with a carrier fre-
tor leaves a residual central peak in the corrected quency of 500 Hz, the units both were excited
by a band-limited range of AM rates around
100 Hz. This suggests that the unit pair exhibits
stimulus dependencies of both stimulus-induced
cross-correlations as revealed by the shift pre-
dictor as well as neurally mediated cross-
correlations because of changes in form and
width of peaks in the neural synchrony as com- C
pared with the shift predictor.
Correcting for Common Input Firing

Periodicities and Firing Rates
According to de la Rocha et al. (2007), the
stimulus-corrected cross-correlation coefficient
for common input still depends on the geometric
mean of the firing rates of the two neurons, both
in cortical slice recordings and in their simulated
data. This effect was also visible for in vivo
recordings from visual cortex (Kr€uger 1991),
and it was used to argue that nearly all of
the cortical cell correlations resulted from
common retinal input. This suggests that
additivity-corrected procedures are inadequate
and that multiplicative interactions of stimulus
and spontaneous firing activity occur. This can
be corrected for by a deconvolution procedure as
described below.
The example shown (Fig. 4) was based on
spontaneous multielectrode recordings from pri-
mary auditory cortex in an anesthetized cat where
the EEG showed spindle waves in the frequency
range of 8–10 Hz. In anesthetized animals (sleep)
spindle oscillations are frequently observed. The
oscillations likely reflect common input as the
correlograms typically appear symmetric around
zero lag, so their effect should disappear in the
corrected correlogram using a deconvolution pro-
cedure. These periodic oscillations are reflected
Correlation Analysis of Parallel Spike Trains, in the firing times of the spike trains and in the
Fig. 3 Unit pair with a stimulus-dependent neural corre- oscillatory character of the cross-correlogram
lation. (a) and (b) dot displays of single-unit activity to (red curves). This effect was corrected by
tone pips with carrier frequencies in the range
100–3,200 Hz. The stimulus inhibits the unit shown in
deconvolving the cross-correlation function by
(a), whereas it activates the unit shown in (a). (c–f) simul- the geometric mean of the autocorrelation func-
taneous and nonsimultaneous (shift predictor; shaded) tions of the two spike trains (Eggermont and
cross-correlation histograms for a variety of stimuli. Bin Smith 1996). The corrected result is shown as
width is 4 ms (From Epping and Eggermont (1987))
the blue curves. One notices also the strong
contribution (>50 %) of these spindle-related
correlations on the peak value of the cross-
correlogram.
Correlation Analysis of Parallel Spike Trains, correlograms for spontaneous activity in cat primary audi-
Fig. 4 The deconvolution-based correction procedure tory cortex. The blue curves represent the corrected cross-
eliminates the oscillatory aspects resulting from correlograms. All units showed an oscillatory autocorre-
anesthesia-induced spindle oscillations in auditory cortex. lation at the same (spindle) frequency, and the effect
Here we show all pair correlations between the spike thereof is eliminated using the deconvolution procedure.
activity between two 8-electrode arrays; one array with All correlograms are scaled to the peak of the raw cross-
electrodes numbered 1–8 and the other with electrodes correlation coefficient value (the red curve), which is
numbered 9–16. The red curves show the raw cross- shown in each panel (From Eggermont (unpublished))
The deconvolution procedure is carried out in Correlation and Connectivity

the frequency domain by taking the Fourier trans-
form of Cxy(t), resulting in the cross spectrum In general there is no unique relationship
Sxy(f), and dividing by the square root of the between known connectivity and observed
product of the two Fourier-transformed autocor- cross-correlation (Meyer and van Vreeswijk
relation functions, i.e., the power spectra. This 2002). Correlations are in principle determined
gives the complex coherence: uniquely by connectivity, including the strengths
and delay dependence of the connections. From
Sxy ð f Þ the neural correlation one can, under certain
gxy ðf Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (15) assumptions about the integration of neural
Sxx ð f ÞSyy ð f Þ
input and the shape on the neuron’s response
curve, estimate the strength of the neural interac-
An inverse Fourier Transform results in the tion. Because the estimate of the peak neural
corrected “correlation coefficient” (Eggermont correlation is very sensitive about the shape of
2006). and working point on the neuron’s response curve
Correlation Analysis of Parallel Spike Trains, excitatory input. (f) Neural shared input. Stimulus-driven
Fig. 5 Potential functional connectivity underlying activity in (a–d); spontaneous activity in (e) and (f). Note
cross-correlation functions. Simultaneous and different time scales. Simple connectivity schemes that
nonsimultaneous (shift predictor; shaded) cross- can account for the observed phenomena are shown as
correlation histograms are superimposed. (a) Neural syn- insets. S, stimulus; 1, neuron 1; 2, neuron 2; u, unobserved
chrony, incremental correlogram. (b) Neural synchrony, neuron. Open and closed triangles indicate excitatory and
decremental correlogram. (c) Unidirectional excitatory inhibitory connections, respectively (From Epping and
input. (d) Neural shared input. (e) Unidirectional Eggermont (1987))
(Melssen and Epping 1987), it is generally not shifted to positive lag times, and unilateral inhi-
permitted to equate the strength of the neural bition shows a dip at positive lag times.
correlation estimated from extracellular record- In auditory cortex evidence for direct excita-
ings with the strength of the neural interaction. tion or inhibition has been scant, unless the neu-
When an actual stimulus is applied, one of its ron pairs were recorded on the same electrode,
effects, especially for cortical neurons, will be and nearly all correlograms are of the common
to shift the neuron’s working point, and apparent input type (Eggermont 1992, 2000). In visual
changes in neural correlation may be found with- cortex several interaction types have been
out a concomitant change in the strength of the shown for recordings with dual electrodes from
neural interaction (Aertsen and Gerstein 1985; cells at different depth in the same cortical col-
Melssen and Epping 1987). umn (Toyama et al. 1981). However, more vari-
The simplest interactions shown by a pair of ety in correlogram shapes was found in the
neurons are that one neuron makes a direct auditory midbrain (Epping and Eggermont
excitatory or inhibitory connection with the 1987). In the examples shown in Fig. 5, we
other or that both neurons share a common excit- show (A) stimulus-induced common excitatory
atory or inhibitory input. It has been shown in effects without neural correlation for two units
simulation studies that common excitatory input recorded on the same electrode and (B) stimulus-
and common inhibitory input result in a peak in induced common suppressive effects for a pair
the cross-correlogram around zero-lag time. recorded on electrodes separated by 130 mm. In
Reciprocal excitatory and inhibitory input to the both cases the shift predictor is equal to the neural
neurons provided by the same source result in synchrony. Weak stimulus-induced activation
a central dip in the cross-correlogram (Moore combined with a direct excitatory effect from
et al. 1970). Unilateral excitation shows a peak one neuron upon the other recorded on the same
electrode (C) is inferred because the shift predic- multiple single-unit recording, the changes in
tor is hardly visible and the peak of the the MU values will be positively, but nonlinearly,
correlogram is displaced by 6 ms from the origin related to those for the corresponding SU ones.
and highly asymmetrical. This is interpreted as So if SU correlations go up, so will the MU
a unidirectional excitatory influence of unit 1 on correlation. The only difference is that the
unit 2. In these three examples the stimulus R-values are larger for MU than for the
consisted of random frequency tone pips corresponding SU ones but smaller than the sum
presented once per second. Part D shows of all the relevant SU pair R-values. Trying to
stimulus-induced (Poisson-distributed clicks) estimate functional connectivity from sorted
common input for a pair recorded with electrodes multiunit activity is not encouraged.
separated by 290 mm putatively via an
unobserved neuron (D), because it has
a pronounced shift predictor. For two spontane- Correlations and the Brain
ous recordings (E,F), reason for the flat shift pre-
dictors, unidirectional excitatory input was Calculating a shift predictor or any other type of
observed for a single electrode pair (E), and com- stimulus predictor is a way to extract the effects
mon input for a dual electrode pair (separation of stimulation on the effective neural connectiv-
160 mm; F). The highly asymmetrical peak of ity despite the drawbacks that we discussed.
the neural synchrony displaced from the origin However, as we have previously argued (Tomita
in part E indicates that unit 1 exerts and Eggermont 2005), it is unlikely that the ner-
a unidirectional excitatory influence on unit vous system performs a correction for stimulus-
2. However, there may be a complication due to induced correlation as estimated by the various
the dead time of the spike sorting procedure, and predictors. It is the actual spike coincidences, i.e.,
it may well be a common input neural correlation the neural synchrony, that are affecting the poten-
as well. This is a known drawback of using sorted tial for firing in a target neuron and not the
units recorded on the same electrode. stimulus-corrected ones. Thus, raw correlations
may effectively estimate those coincident firings
between neurons that could play a role in neural
Effects of Spike Sorting on Pair population coding of sound. The effect of remov-
Correlations ing common periodicities in spike firing and com-
mon bursting activity from the cross-correlation
Multiunit (MU) recording is common in struc- can be justified because these contributions typi-
tures with high cell densities and especially using cally do not occur, or occur much less, in awake
multielectrode arrays. Spike sorting can result in animals and are often the result of a using anes-
well-sorted units, but by no means can one be thesia. Nevertheless, in our data for 15-min
sure that one is dealing with single units (SU), steady-state stimuli, the effects of a stimulus cor-
that is, only the case when doing intracellular or rection based on the shift predictor are minimal
patch-clamp recordings. Evidence from neural (Fig. 6), so using the stimulus-correction proce-
modeling suggests that interpretation of neural dure has minimal effect on the interpretation of
correlations from multiunit recordings may be the data. However, for transient and optimal stim-
ambiguous as it is not a linear combination of uli, the stimulus correction can be extensive and
correlations for the various single-unit pairs even remove all correlation (Eggermont 1994),
(Bedenbauch and Gerstein 1997; Gerstein so in this case one has to consider what the cor-
2000). However, changes in single-unit correla- relation estimate needs to support. If one is
tion strengths will be accompanied by compara- concerned with estimating connectivity or decid-
ble changes in the correlation between multiunit ing which wiring scheme is most likely, then both
activity as shown in Eggermont (2000). Thus, if stimulus corrections and corrections for common,
a number of single units contribute to each network-induced, activity need to be performed.
Correlation Analysis of Parallel Spike Trains, function is shown in solid line; the peak cross-correlation
Fig. 6 This set of 3 cross-correlograms illustrates the is approximately 0.082. The frequency tuning of the two
stimulus-correction procedure applied to spike trains recording sites was nearly identical, yet the stimulus cor-
recorded from primary auditory cortex in the cat during relation obtained as a shift predictor (dashed line) shows
a 900 s long random frequency stimulus where the second only modest stimulus locking. The stimulus-corrected
half was equal to the first. Stimuli were tone pips presented cross-correlation coefficient function is shown in a
randomly in frequency and time with an overall rate of connected-dot line (With kind permission from Springer
28/s across 7 octaves (112 frequencies, so an average rate Science + Business Media: Analysis of parallel spike
of 0.25 frequencies/s). The recording comprised two trains, Chapter 5, Pair-correlation in the time and fre-
arrays of 16 electrodes each. Here we show the pair quency domain, 2010, page 77–102, J. J. Eggermont,
correlation between two neighboring electrodes separated Fig. 5.3)
by 0.5 mm. The uncorrected cross-correlation coefficient
If one is concerned with how the brain performs Bedenbaugh P, Gerstein GL (1997) Multiunit normalized
its task in vivo, then these corrections may cross correlation differs from the average single-unit
normalized correlation. Neural Comput 9:1265–1275
obscure the very aspects one wants to understand. Brillinger DR, Bryant HL Jr, Segundo JP (1976) Identifi-
cation of synaptic interactions. Biol Cybern 22:
213–228
Cross-References de la Rocha J, Doiron B, Shea-Brown E, Josić K, Reyes
A (2007) Correlation between neural spike trains
increases with firing rate. Nature 448:802–806
▶ Spike Train Eggermont JJ (1992) Neural interaction in cat primary
auditory cortex. Dependence on recording depth,
electrode separation, and age. J Neurophysiol
68:1216–1228
References Eggermont JJ (1994) Neural interaction in cat primary
auditory cortex II. Effects of sound stimulation.
Aertsen AM, Gerstein GL (1985) Evaluation of neuronal J Neurophysiol 71:246–270
connectivity: sensitivity of cross-correlation. Brain Eggermont JJ (2000) Sound-induced synchronization of
Res 340:341–354 neural activity between and within three auditory cor-
Aertsen AM, Smolders JW, Johannesma PI (1979) Neural tical areas. J Neurophysiol 83:2708–2722
representation of the acoustic biotope: on the existence Eggermont JJ (2006) Properties of correlated neural activ-
of stimulus-event relations for sensory neurons. Biol ity clusters in cat auditory cortex resemble those of
Cybern 32:175–185 neural assemblies. J Neurophysiol 96:746–764
C 868 Cortical Columns and Neural Population Models
Eggermont JJ (2010) Pair-correlation in the time and Toyama K, Kimura M, Tanaka K (1981) Organization of
frequency domain. In: Gr€ un S, Rotter S (eds) Analysis cat visual cortex as investigated by cross-correlation
of parallel spike train. Springer, New York, pp 77–102 technique. J Neurophysiol 46:202–214
Eggermont JJ, Smith GM (1995) Rate covariance domi- van Stokkum IH, Johannesma PI, Eggermont JJ
nates spontaneous cortical unit-pair correlograms. (1986) Representation of time-dependent correlation
Neuroreport 6:2125–2128 and recurrence time functions. A new method to ana-
Eggermont JJ, Smith GM (1996) Neural connectivity only lyse non-stationary point processes. Biol Cybern
accounts for a small part of neural correlation in audi- 55:17–24
tory cortex. Exp Brain Res 110:379–391 Voigt HF, Young ED (1990) Cross-correlation analysis of
Eggermont JJ, Epping WJ, Aertsen AM (1983) Stimulus inhibitory interactions in dorsal cochlear nucleus.
dependent neural correlations in the auditory midbrain J Neurophysiol 64:1590–1610
of the grassfrog (Rana temporaria L.). Biol Cybern Yang XW, Shamma SA (1990) Identification of connec-
47:103–117 tivity in neural networks. Biophys J 57:987–999
Epping WJ, Eggermont JJ (1987) Coherent neural activity
in the auditory midbrain of the grassfrog.
Gawne TJ, Richmond BJ (1993) How independent are the
messages carried by adjacent inferior temporal cortical Cortical Columns and Neural
neurons? J Neurosci 13:2758–2771 Population Models
Gerstein GL (2000) Cross-correlation measures of
unresolved multi-neuron recordings. J Neurosci
Methods 100:41–51 ▶ Mesoscopic Anatomy and Neural Population
Joris PX, Louage DH, Cardoen L, van der Heijden Models
M (2006) Correlation index: a new metric to quantify
temporal coding. Hear Res 216–217:19–30
Kr€uger J (1991) Spike train correlations on slow time scales
in monkey visual cortex. In: Kr€ uger J (ed) Neuronal
cooperativity. Springer, Berlin/Heidelberg, pp 105–132 Cortical Columns, Models of
Marmarelis PZ, Marmarelis VZ (1978) Analysis of phys-
iological systems: the white noise approach. Plenum,
Sean Hill
New York
Melssen WJ, Epping WJ (1987) Detection and estimation Brain Mind Institute, École Polytechnique
of neural connectivity based on crosscorrelation anal- Fédérale de Lausanne, Lausanne, Switzerland
ysis. Biol Cybern 57:403–414
Meyer C, van Vreeswijk C (2002) Temporal correlations
in stochastic networks of spiking neurons. Neural
Comput 14:369–404 Definition
Moore GP, Segundo JP, Perkel DH, Levitan H (1970)
Statistical signs of synaptic interaction in neurons. A cortical column is a module in the neocortex,
approximately 300–600 mm in diameter, com-
Neven H, Aertsen A (1992) Rate coherence and event
coherence in the visual cortex: a neuronal model of posed of many minicolumns coupled by short-
object recognition. Biol Cybern 67:309–322 range horizontal connections. Neurons within
Nowak LG, Munk MH, Nelson JI, James AC, Bullier a column often share dynamic response proper-
J (1995) Structural basis of cortical synchronization.
ties. In general, a column is part of a continuous
I. Three types of interhemispheric coupling.
J Neurophysiol 74:2379–2400 sheet of neocortex, where the boundaries of indi-
Perkel DH, Gerstein GL, Moore GP (1967) Neuronal vidual columns cannot be determined by anatomy
spike trains and stochastic point processes. alone, although there are exceptions such as the
II. Simultaneous spike trains. Biophys J 7:419–440
barrel field columns in rodent neocortex. Col-
Srinivasan MV, Bernhard GD (1976) A proposed mecha-
nism for multiplication of neural signals. Biol Cybern umns with similar response properties are often
21:227–236 connected by long-range horizontal connections.
Staude B, Rotter S, Gr€ un S (2008) Can spike coordination Numerous efforts have been undertaken to
be differentiated from rate covariation? Neural
Comput 20:1973–1999
develop computational models of the cortical
Tomita M, Eggermont JJ (2005) Cross-correlation and column, which is often equated with the quest to
joint spectro-temporal receptive field properties in identify the fundamental computational unit of
auditory cortex. J Neurophysiol 93:378–392 the mammalian brain. These efforts span from
Cortical Columns, Models of 869 C
highly simplified “mean-field” models that model neocortex is thought to have occurred as a result
interactions between homogeneous cell adding additional cortical columns rather than
populations to detailed biophysical models with increasing the size of the cortical column (Rakic
heterogeneous neuron morphologies and diverse 1995). Working in cat visual cortex, Hubel and
electrophysiological behavior. Orientation selec- Wiesel demonstrated functional cortical
tivity in the primary visual cortex is an example columns with a diameter of around 0.5 mm
of a computation performed by a cortical column, (Hubel and Wiesel 1977). Later studies showed C
which has been observed experimentally and that different patterns of stimuli produced differ-
modeled computationally. However, a common ent functional architectures with different spatial
goal of modeling the cortical column is to iden- fidelity (Mountcastle 1998). In Neuroscience,
tify a more general putative canonical computa- there has been difficulty adhering to a single def-
tion occurring throughout the neocortex. inition of a cortical column leading to debate and
confusion about both its existence and function
(Horton and Adams 2005). However, a common
Detailed Description goal of modeling the cortical column is to iden-
tify a putative but general canonical computation
Background and the specific neuronal mechanisms that under-
The neocortex originated in a mammalian ances- lie the emergence of diverse functional architec-
tor approximately 250 million years ago and its tures throughout the neocortex.
evolution has been critical for the emergence of
human cognition and behavior. The types of neu- Modeling Approaches
rons and their local synaptic connectivity in dif- An early qualitative model of cortical circuitry
ferent parts of the neocortex appear to be was proposed by Hubel and Wiesel (1962) to
variations on a basic stereotypical circuit design. describe the mechanisms underlying simple and
Lorente de Nó (1938) was the first to suggest that complex cell responses in nonhuman primate
“vertical chains” of neurons traversing the six cortex based on strongly tuned convergent
layers of the neocortex formed elementary func- thalamocortical input.
tional units: “all the elements of the cortex are Douglas and Martin (1991) provided the first
represented in it, and therefore it may be called an canonical microcircuit model from their studies
elementary unit, in which theoretically, the whole on visual cortex in the cat. This model was based
process of the transmission of impulses from the on microanatomical studies and in vivo intracel-
afferent fibre to the efferent axon may be accom- lular recordings. It provided an account of the
plished.” Vernon Mountcastle and colleagues intracellular responses to pulsed stimulations in
subsequently showed that neurons separated by terms of an inseparable combination of excitatory
a horizontal distance of less than a millimeter and inhibitory responses of three basic
share similar response properties (Mountcastle populations of neurons. Their model proposed
1957; Mountcastle et al. 1957; Powell and that thalamic input is not the major source of
Mountcastle 1959). The term “column” was pro- excitatory input to cortical neurons, and instead
posed by Mountcastle who suggested “there is an intracortical input was the dominant source of
elementary unit of organization in the somatic excitation.
cortex made up of a vertical group of cells Somers et al. (1995) implemented a large-
extending through all the cellular layers” scale computational model of orientation selec-
(Mountcastle 1957). The hypothesis is that the tivity in cortical simple cells that demonstrated
column is a functional cortical module, com- that sharp orientation tuning could emerge
posed of minicolumns, and approximately from local recurrent intracortical excitation at
300–600 mm in diameter, which is replicated the columnar level and be balanced by
throughout the neocortex as the primary compu- nonspecific inhibition with only weakly tuned
tational unit. The evolutionary expansion of the thalamocortical excitation.
C 870 Cortical Columns, Models of
Much work has been focused on explaining numbers and distribution of the somata, den-
the function of the cortical column during wake- drites, and thalamocortical synapses of nine
fulness. Hill and Tononi (2005) focused on the cortical cell types, soma location, dendritic
observation that the cortical column is morphology, synapse innervation, and predic-
multifunctional and can exhibit strikingly differ- tions of synaptic connectivity for
ent functions in wakefulness and sleep. Their a thalamocortical circuit in the rat vibrissal cor-
large-scale model of thalamocortical visual cir- tex (Oberlaender et al. 2012).
cuitry was based on point neurons organized in
repeating columnar structures with patchy long- Summary
range connectivity and illustrated how columns A broad diversity of approaches has been taken to
transition between spontaneous activity and modeling the cortical column. In the end, they
orientation-selective responses during wakeful- share a common goal of understanding whether
ness and slow oscillations during sleep. or not there is a canonical computation performed
Traub et al. (2005) has developed a model of by cortical microcircuitry and what are the mech-
a single thalamocortical column (in this case anisms that govern and generate it.
a cortical minicolumn coupled with thalamic
and reticular thalamic neurons) based on bio-
physical models of neurons with schematic mor- References
phologies employed to more accurately capture
the positioning of synapses on different compart- Douglas RJ, Martin KA (1991) A functional microcircuit
ments of neurons. This model exhibited a broad for cat visual cortex. J Physiol 440:735–769
Helmstaedter M, de Kock CPJ, Feldmeyer D, Bruno RM,
range of oscillatory, epileptic, and sleeplike Sakmann B (2007) Reconstruction of an average cor-
phenomena. tical column in silico. Brain Res Rev 55(2):193–203
Launched in 2005, the Blue Brain Project Hill S, Tononi G (2005) Modeling sleep and wakefulness
(bluebrain.epfl.ch) proposed a data-driven pro- in the thalamocortical system. J Neurophysiol
93(3):1671–1698
cess to construct highly detailed biophysical Hill SL, Wang Y, Riachi I, Schurmann F, Markram
models of the cortical microcircuitry that under- H (2012) PNAS plus: statistical connectivity provides
lies a column (Markram 2006) using extensive a sufficient foundation for specific functional connec-
measurements of the young rat (P14) somatosen- tivity in neocortical neural microcircuits. Proc Natl
Acad Sci 109(42):E2885–E2894. doi:10.1073/
sory cortex, including single-cell gene expres- pnas.1202128109
sion, whole-cell patch clamp electrophysiology, Horton JC, Adams DL (2005) The cortical column:
morphological reconstructions, synaptic connec- a structure without a function. Philos Trans R Soc
tivity and plasticity, and more. The models cap- Lond B Biol Sci 360(1456):837–862
Hubel DH, Wiesel TN (1962) Receptive fields, binocular
ture the experimentally observed diversity and interaction and functional architecture in the cat’s
composition of 55 morphological and 12 electro- visual cortex. J Physiol 160:106–154
physiological classes. The group has since shown Hubel DH, Wiesel TN (1977) Functional architecture of
that the resulting models can be used to predict macaque monkey cortex. Proc R Soc Lond B 198:1–59
Lorente de Nó R (1938) Architectonics and structure of
synaptic connectivity patterns (Hill et al. 2012) the cerebral cortex. In: Fulton JF (ed) Physiology of
and cellular contributions to the local field poten- the nervous system. Oxford University Press, New
tial (Reimann et al. 2013). York, pp 291–330
A detailed anatomical model of an average Markram H (2006) The Blue Brain Project. Nat Rev
barrel column within the somatosensory cortex Mountcastle VB (1957) Modality and topographic prop-
of the rat has been proposed by Helmstaedter erties of single neurons of cat’s somatic sensory cortex.
et al. (2007). The barrel cortex provides an J Neurophysiol 20:408–434
example where both the anatomical and func- Mountcastle VB (1998) Perceptual neuroscience. The
cerebral cortex. Harvard University Press, Cambridge
tional cortical column is well defined. Work Mountcastle VB, Davies PW, Berman AL (1957) Response
from this group has led to detailed anatomical properties of neurons of cat’s somatic sensory cortex to
reconstructions that take into account the peripheral stimuli. J Neurophysiol 20:374–407
Cortical Function, Normative Models of 871 C
Oberlaender M, de Kock CP, Bruno RM, Ramirez A, the chosen objective function, if such a solution
Meyer HS, Dercksen VJ, Helmstaedter M, Sakmann can be found.
B (2012) Cell type-specific three-dimensional struc-
ture of thalamocortical circuits in a column of rat
vibrissal cortex. Cereb Cortex 22(10):2375–2391
Powell TPS, Mountcastle VB (1959) Some aspects of the Detailed Description
functional organization of the cortex of the postcentral
gyrus of the monkey: a correlation of findings obtained
in a single unit analysis with cytoarchitecture. Bull Normative models of cortical function are much C
Johns Hopkins Hosp 105:133–162 like other computational models of the cortex.
Rakic P (1995) A small step for the cell, a giant leap for They try to describe and explain experimental
mankind: a hypothesis of neocortical expansion during data and make predictions that can be used to
evolution. Trends Neurosci 18(9):383–388
Reimann MW, Anastassiou CA, Perin R, Hill SL, falsify them (Dayan and Abbott 2001). But rather
Markram H, Koch C (2013) A biophysically detailed than directly asking how the brain solves
model of neocortical local field potentials predicts the a particular computational problem, one first
critical role of active membrane currents. Neuron asks what the optimal solution to this problem
79(2):375–390
Somers DC, Nelson SB, Sur M (1995) An emergent model is. Then the actual solution embodied in the brain
of orientation selectivity in cat visual cortical simple may be understood as a specific implementation
cells. J Neurosci 15(8):5448–5465 of this optimal solution or a reasonable approxi-
Traub RD, Contreras D, Cunningham MO, Murray H, mation to it. Thus, by using normative models,
LeBeau FE, Roopun A, Bibbig A, Wilent WB, Higley
MJ, Whittington MA (2005) Single-column one approaches the question of how the brain
thalamocortical network model exhibiting gamma works by first asking how the brain ought to
oscillations, sleep spindles, and epileptogenic bursts. work. The great strength of this approach is that
J Neurophysiol 93(4):2194–2232 it can lead to an understanding of brain function
at a deeper level. Rather than only describing how
something works in the brain, normative models
reveal why the brain may be operating in this
Cortical Field Theory way. Normative modeling approaches are not
restricted to cortical function, but may be used
▶ Neural Field Model, Continuum at essentially all levels of nervous system model-
ing and beyond.
There is a wide range of existing and conceiv-
able normative models of cortical function due to
the vast range of functions in which the cortex is
Cortical Function, Normative involved. These include perception, cognition,
Models of memory, and motor control. In each of these
areas, normative models can be used to try to
Jochen Triesch better understand aspects of cortical function.
Frankfurt Institute for Advanced Studies, While the cortex is clearly involved in all of the
Frankfurt am Main, Germany mentioned functions, it is worth noting, however,
that all of these functions are also present in
animal species without any cortex. In such spe-
Definition cies, similar computational functions may be
implemented in a very different way. In much
Normative models of cortical function are the same way as different computer algorithms
a class of models that are related to or derived or different implementations of the same algo-
from a well-defined mathematical objective rithm can solve the same problem, different ner-
function. They allow to understand the brain’s vous system structures in different species might
solution to a certain computational problem in have evolved to solve the same problem in
relation to an optimal solution that maximizes different ways.
C 872 Cortical Maps, Intrinsic Processes
A generic task for a perceptual system that operates, they reveal why this way of operation
may be implemented in sensory cortical areas is is appropriate or even optimal.
to represent the sensory signals in an efficient
way. Sensory stimuli are typically represented
by large populations of neurons in sensory corti- Cross-References
cal areas. Normative models of sensory coding
ask how a population of neurons can encode ▶ Bayesian Approaches in Computational
a given class of sensory signals optimally. To Neuroscience: Overview
this end, certain assumptions about noise in the ▶ Cortex: Overview
sensory signals or the processing neurons or ▶ Predictive Coding
a limited energy budget may be considered. ▶ Reinforcement Learning in Cortical Networks
Sparse coding models and predictive coding
models are popular examples of normative
models of sensory coding. References
Another generic computational problem in
which the cortex is involved is to make inferences Dayan P, Abbott LF (2001) Theoretical neuroscience:
computational and mathematical modeling of neural
from incomplete and noisy sensory inputs. Exam-
systems. MIT Press, Cambridge, MA
ples are recognizing the face of a friend in Doya K, Ishii S, Pouget A, Rao RPN (eds) (2007) Bayesian
a poorly lit room or following a conversation in brain: probabilistic approaches to neural coding. MIT
a noisy environment. A popular normative frame- Press, Cambridge, MA
work for describing how such inferences ought to
be made is that of Bayesian inference. Corre-
spondingly, much current research aims to under-
stand learning and information processing in the Cortical Maps, Intrinsic Processes
cortex from the perspective of Bayesian inference
and attempts to elucidate whether the cortex may Geoffrey J. Goodhill
implement a specific kind of Bayesian inference Queensland Brain Institute and School of
and learning algorithm or some approximation Mathematics and Physics, University of
to it. Correspondingly, Bayesian inference is Queensland, St Lucia, QLD, Australia
a major topic in contemporary computational
neuroscience (Doya et al. 2007).
Another generic computational problem in Definition
which the cortex is also involved is to make
decisions and control our actions. A popular nor- A cortical map is a neural representation of an
mative framework for describing how decisions input space (e.g., the two dimensions of visual
ought to be made in an optimal way and how an space and/or features such as orientation prefer-
agent ought to select actions are those of decision ence and ocular dominance) such that similar
theory and reinforcement learning. Conse- features in the input space are represented by
quently, much research tries to address how the neurons which are nearby in the cortex.
cortex and other brain structures such as the basal
ganglia may be implementing approximations to
optimal decision making or specific reinforce- Detailed Description
ment learning algorithms.
Overall, normative modeling approaches A common way in which axonal projections con-
have already proved to be very fruitful in pro- nect areas in the brain is via topographic maps.
viding a deeper understanding of brain function Although there is no precise mathematical defini-
in general and cortical function in particular. tion of this term, roughly speaking it means that
Beyond merely describing how the cortex neurons that are physically nearby in the input
Cortical Maps, Intrinsic Processes 873 C
region, or alternatively neurons that have similar Numerous experimental manipulations to the
functional properties, map to neurons that are retinotectal system have provided a rich dataset
nearby in the target region. The formation of of mapping outcomes following either molecular
these maps during neural development depends disruptions of the gradients (Lemke and Reber
on both intrinsic (molecularly guided) and extrin- 2005; Huberman et al. 2008; Cang and Feldheim
sic (activity-dependent) cues and has served as 2013) or surgical disruptions of the input or target
a paradigm model for how such cues guide brain structures (such as removing or rotating parts of C
development more generally. Activity-dependent the eye and/or the tectum) (Udin and Fawcett
effects on map development are the subject of 1988; Goodhill and Richards 1999). Together
a separate chapter; here we consider the role of these experiments provide many constraints for
intrinsic cues. theories of map formation.
The formation of topographic maps has been Early models in the 1970s and 1980s addressed
particularly well studied in the visual system. basic principles for how gradient matching could
A paradigmatic example is the projection from establish maps in general and how versions of
the retina to the tectum (frogs, fish, birds)/supe- these principles could reproduce many of the sur-
rior colliculus (mammals). To explain how such gical manipulation experiments in the retinotectal
maps develop, Sperry (1963) defined the key system (Willshaw and Price 2003; Goodhill and
theoretical idea of chemospecificity: a process Xu 2005). These principles include competition
of matching of chemically defined labels between for target space and other types of fiber-fiber inter-
the input and target structures. One particularly actions in addition to gradient matching. In the
elegant form of this idea is that the input structure 1990s, models were proposed based closely on
contains a spatial concentration gradient of the properties of the newly discovered Eph and
a receptor, which is matched to a spatial concen- ephrin gradients, though often without also con-
tration gradient of a ligand in the target structure. sidering the surgical manipulation data (Goodhill
In the 1990s, this prediction was experimentally and Richards 1999). More recent models have
confirmed via the discovery of gradients of Eph been particularly concerned with data from recent
receptors in the retina, and their ligands the gene disruption experiments, where, for instance,
ephrins in the tectum/superior colliculus (Fig. 1; Eph levels are changed in only a subset of
McLaughlin and O’Leary 2005). retinal cells or where a much smaller than normal
number of retinal cells send axons to the tectum
(Simpson et al. 2009). The mathematical princi-
ples underlying all these models are quite diverse,
though many are based on the idea of optimizing
an energy function for the map. Although no one
modeling approach has yet become canonical,
together these models have helped to clarify the
relative roles and importance of the different bio-
logical mechanisms (most notably gradients, com-
petition, and fiber-fiber interactions) involved in
map formation.
Intrinsic processes of map development in the
cortex have been far less studied than for retino-
Cortical Maps, Intrinsic Processes, Fig. 1 Gradient
matching in the retinotectal system. Eph receptors are tectal maps; however, there is recent evidence that
expressed in an increasing nasal (N) to temporal (T) gra- similar principles of matching between Eph and
dient in the retina, while ephrin ligands are expressed in an ephrin gradients also guide their initial formation
increasing rostral (R) to caudal (C) gradient in the tectum.
(Huberman et al. 2008). An intriguing question is
The Eph-ephrin interaction is repulsive, so that regions of
retina with high levels of Eph receptor map to regions of whether the formation of functional maps in the
tectum with low levels of ephrin ligand, and vice versa cortex, such as ocular dominance and orientation
C 874 Cortical Maps, Activity-Dependent Development
maps in primary visual cortex, could also be partly Simpson HD, Mortimer D, Goodhill GJ (2009) Theoretical
controlled by intrinsic processes. Recent data has models of neural circuit development. Curr Top Dev
Biol 87:1–51
found molecular differences between regions of Sperry RW (1963) Chemoaffinity in the orderly growth of
very young cat visual cortex which are reminiscent nerve fiber patterns and connections. Proc Natl Acad
of the patterns of ocular dominance columns that Sci U S A 50:703–710
subsequently develop, though a causative link has Udin SB, Fawcett JW (1988) Formation of topographic
maps. Annu Rev Neurosci 11:289–327
yet to be established. Similarly, clonally related Willshaw DJ, Price DJ (2003) Models of topographic map
neurons (those arising from a single progenitor formation. In: Van Ooyen A (ed) Modeling neural
cell) in rat visual cortex have been found to have development. MIT Press, Cambridge, MA,
orientation preferences that are more similar than pp 213–244
non-clonally related neurons, challenging purely
activity-dependent accounts for the development
of orientation preference.
In summary, models up to now have played Cortical Maps, Activity-Dependent
a critical role in demonstrating the effectiveness Development
(or otherwise) of principles for how map devel-
opment could occur and dissecting how the pre- Nicholas Swindale
cise details of particular biological processes University of British Columbia, Vancouver,
affect map structure. However, the constant BC, Canada
flow of new data continues to provide new ques-
tions and challenges, which will provide inspira-
tion for modeling work in this area for some time Synonyms
to come.
Computational maps; Self-organizing maps;
Topographic maps
Cross-References
▶ Cortical Maps, Activity-Dependent Definition

Development
The term “cortical map” refers to the existence of
a nonrandom relationship between the position
of a neuron in the cerebral cortex and the value of
References
some property that can be assigned to it on the
Cang J, Feldheim DA (2013) Developmental mechanisms basis of physiological or anatomical tests. Typi-
of topographic map formation and alignment. Annu cally the property in question is a receptive field
Rev Neurosci 36:51–77 parameter of a sensory neuron, e.g., the position
Goodhill GJ, Richards LJ (1999) Retinotectal maps: mol- in space of the receptive field of a visual sensory
ecules, models, and misplaced data. Trends Neurosci
22:529–534 neuron, or the color of a stimulus that activates it,
Goodhill GJ, Xu J (2005) The development of retinotectal but it might also be a projective field, e.g., the
maps: a review of models based on molecular gradi- position in the body of a muscle or group of
ents. Network 16:5–34 muscles activated by neurons in motor cortex,
Huberman AD, Feller MB, Chapman B (2008) Mecha-
nisms underlying development of visual maps and or, more speculatively, some aspect of knowl-
receptive fields. Annu Rev Neurosci 31:479–509 edge or behavior coded for, or produced, by
Lemke G, Reber M (2005) Retinotectal mapping: new activity in single neurons or groups of neighbor-
Insights from molecular genetics. Annu Rev Cell Dev ing neurons (perhaps as measured by fMRI
Biol 21:551–580
McLaughlin T, O’Leary DDM (2005) Molecular gradients experiments). A further qualification is that map
and development of retinotopic maps. Annu Rev properties often remain constant in value with
Neurosci 28:327–355 depth in the cortex and vary only with lateral
Cortical Maps, Activity-Dependent Development 875 C
(equivalently tangential) position. This property, to be modeled mathematically (Goodhill and Xu
termed “columnar organization,” (Hubel and 2005; Goodhill 2007; Eglen and Gjorgjieva
Wiesel 1977; Mountcastle 1998) means that the 2009). Much of this work can be applied to the
cortex is normally treated as a two-dimensional development of maps in LGN and cortex as well.
array for modeling purposes. “Activity-
dependent development” in the context of Ocular Dominance Maps
“cortical maps” refers to a class of computational When the visual fields of the two eyes overlap C
neuroscience model that seeks to explain the substantially, as they do in carnivores and most
development of cortical maps as the outcome of primates, the retinotopic maps also overlap, i.e.,
mathematically specified rules governing axonal cells in the target structures will have receptive
growth and synapse formation in which neural fields that are in approximately the same location
activity plays a primary role. All of these models in visual space. In such cases, the relative
use Hebbian learning rules to achieve a mapping strength of the inputs from the two eyes,
between cells in one or more input layers a property termed ocular dominance, is often
representing either cell classes in the retina(s) or found to vary systematically with position across
the lateral geniculate nucleus (LGN) layers and the map, allowing it to be subdivided into inter-
connections with cells in a single output layer that digitating regions referred to as ocular dominance
represents the cortex. The interest of the models columns (Hubel and Wiesel 1977). The layout of
lies not so much in what can be computed with the columns is generally periodic with a spacing
them, but in the fact that they are often able to in the range 0.8–2.0 mm depending on species. In
replicate, sometimes in a very detailed way, the old-world primates (such as macaque monkeys),
geometrical properties of different maps and their the morphology is branching stripes (Fig. 1a),
spatial relationships (Erwin et al. 1995; Swindale whereas in cats and ferrets, there is a less ordered
1996, 2003; Goodhill 2007). pattern of elongated blobs and patches (Fig. 1b).
Ocular dominance columns are not found in all
species. They may be variably present in the
Detailed Description visual cortex of new-world monkeys (Adams
and Horton 2003) and are largely absent in the
Varieties of Cortical Map visual cortices of mice and rats which have only
In retinotopic maps, there is an orderly mapping a very small binocular visual field (H€ubener et al.
of receptive field position in retinotopic coordi- 2008).
nates (elevation and azimuth) in a brain region, as
a result of topographically ordered connections Orientation Maps
between retinal ganglion cells and target struc- The majority of neurons in the primary visual
tures. Such maps possess local order (neighbor- cortex of all species that have been examined
ing retinal neurons tend to project to neighboring show orientation selectivity, i.e., the neuron will
regions in the target structure) and global order, respond selectively to the orientation of a bar or
i.e., iso-elevation and iso-azimuthal contours edge flashed in, or moved across, the receptive
have a predictable and consistent layout in the field. The response is typically a bell-shaped
target structure. Retinotopic maps occur in the function of stimulus orientation with a width in
optic tectum in lower vertebrates such as frogs the range of 20–60 at half height and a peak that
and fish and in the superior colliculus, LGN, and defines the cell’s preferred orientation. In carni-
visual cortical areas of mammals. The rules vores and primates, preferred orientation has
governing the formation of maps in the optic a columnar organization and varies smoothly
tectum of frogs and fish, and their responses to and systematically with tangential position
experimental manipulations such as removing (Fig. 1c). As with ocular dominance columns,
parts of the retina and/or the tectum, were the periodicity is generally in the range
among the first to be studied experimentally and 0.8–2.0 mm. However, when ocular dominance
Cortical Maps, Activity-Dependent Development, of ocular dominance (outlined by upper layer of black
Fig. 1 Ocular dominance column arrangements observed lines), spatial frequency (lowest layer), and orientation
in macaque monkey visual cortex (a) and area 17 of cat (colored lines, middle layer). Singularities in this panel
visual cortex (b). White areas are regions of cortex labeled are identifiable as points where colored lines meet (Panel
by injection of radioactive tracer into one eye. (c) Layout a is adapted from Hubel and Wiesel (1977); b is adapted
of orientation preference in area 17 of cat visual cortex. from LeVay et al. (1978). c is from H€ubener et al. (1997).
Singularities (pinwheels) are identifiable as points where Figure d was kindly supplied to the author by M. H€ubener)
many color boundaries intersect. (d) Layout of domains
and orientation maps are simultaneously present Singularities can be classified by sign, where pos-
in a visual cortical area, which is often the case, itive sign means that orientations rotate in
the periodicities are usually different. Because a clockwise direction if a circular path enclosing
orientation is, by definition, a property that is a singularity is traversed in a clockwise direction,
cyclic in the range 0–p, it is often convenient to while negative sign means that orientations rotate
represent it as an angle-doubled complex in an anticlockwise direction for clockwise motion
number z = aei2y, where y is the orientation around the circuit. Positive and negative singular-
and the length a of the vector represents ities are found in approximately equal numbers in
response strength or tuning sharpness, or some visual cortex maps (Obermayer and Blasdel 1997).
combination of the two. Singularities (more com-
monly referred to as “pinwheels”) occur fre- Direction Preference
quently in orientation maps (Fig. 1c, d). In addition to selectivity for orientation, many
Mathematically these correspond to positions visual cortical neurons give different responses
where z = 0 which will happen at locations to the two possible directions of motion of a long
where zero crossings in the real and imaginary edge. This can range from the equal responses to
components of z intersect (Swindale 1982). For both directions to no response at all in the so-
most types of periodic layout of z, this will be called null direction, with the responses of most
a frequent occurrence (Wolf and Geisel 1998). neurons falling between these two extremes.
Direction preference is represented continuously Fourier transform of a 2D Gaussian, which is
in cat and ferret primary visual cortex with the another Gaussian function. The resulting function
topological consequence that lines, across which turns out to be an excellent description of the
direction preference reverses, extend between spatial receptive field profile of many visual cor-
and connect pairs of orientation singularities of tex neurons (Ringach 2002). Bandwidths for
opposite sign (Swindale et al. 1987). This most visual cortical neurons are about 1–2
arrangement has been confirmed experimentally octaves, while within a single cortical area (e.g., C
(Swindale et al. 1987; Shmuel and Grinvald area 17 of the cat), preferred spatial frequency
1996; Weliky et al. 1996). can vary from cell to cell over a range of about
2–3 (Movshon et al. 1978; Born and Tootell
Spatial Frequency Maps 1991). Experiments in cats, monkeys, and ferrets
As an alternative to bars and edges, many exper- have shown that preferred spatial frequency
imenters have used moving sine wave gratings to varies periodically with position, with an overall
study the responses of visual cortex neurons. layout that is independent of, but geometrically
Since the Fourier transform of an extended sine similar to, that of ocular dominance (H€ubener
wave grating is a point in a 2D spatial frequency et al. 1997; Issa et al. 2000; Yu et al. 2005;
space (and a moving grating is a point in a 3D Nauhaus et al. 2012).
spatiotemporal frequency space), such stimuli
allow the use of linear systems theory to charac- Additional Maps
terize neuronal responses in space and time. Maps of retinal position, ocular dominance, ori-
Experimentally, it is found that the majority of entation, direction of motion, and spatial fre-
primary visual cortex neurons will respond quency are all simultaneously present in area 17
with an elevated, or modulated, firing rate to of the cat (Fig. 1d) and ferret. Area V1 in
a moving grating stimulus, with the magnitude macaque monkeys also contains a retinotopic
of the response varying with the orientation, map plus maps of ocular dominance, orientation,
direction of motion, and spatial frequency of the and spatial frequency. Additional maps are either
grating. Cells are also tuned for spatial frequency, known or very likely to be present. For example,
with the position of the peak defining the pre- there is an ordered sensitivity to stereoscopic
ferred spatial frequency of the cell and the range disparity in cat area 18 (Kara and Boyd 2009),
defining a bandwidth (De Valois and De Valois and several investigators have reported maps for
1990). color preference in macaque V1 (e.g., Lu and Roe
In a 2D space whose coordinates are horizon- 2008). How many different features may simul-
tal frequency on the x-axis and vertical frequency taneously be represented in maps in a single cor-
on the y-axis (or equivalently in coordinates, spa- tical area is currently unknown.
tial frequency = r and orientation = y), tuning is
typically an approximately Gaussian-shaped Rats, Mice, and Squirrels
region centered on a preferred orientation and It is worth noting specifically that the visual cor-
spatial frequency (De Valois and De Valois tices of these animals have retinotopic maps but
1990). Transformed back into visual space, such that the maps for preferred orientation and ocular
a tuning function corresponds to a Gabor func- dominance appear to be random, in both tangen-
tion. This can be explained in the following way: tial and vertical (columnar) directions (Van
a Gaussian blob in spatial frequency space can be Hooser et al. 2005; Ohki et al. 2005; Bonin
represented as a delta function positioned at the et al. 2011). This is a violation of the principle
preferred (r, y) value convolved with a 2D Gauss- of columnar organization and shows that it does
ian. In spatial coordinates, this corresponds to the not have always to be present. This might suggest
Fourier transform of a delta function, which that orientation and ocular dominance columns
equals a sine wave grating, multiplied by the are adaptive features that arose in more recently
evolved mammals with larger (and perhaps bet- of molecularly guided (innate, genetic, nature) or
ter) cortices. However, both rats and mice have neural activity-dependent (environmental, learn-
well-developed visual abilities, and squirrels in ing based, nurture) cues. Although it is frequently
particular have good visual abilities so it is hard pointed out that there can be complex interactions
to avoid the conclusion that columns are not between genes and the environment, the nature-
essential features of a working cortex. nurture dimension continues to be a useful one
along which to characterize models and research
Maps in Other Cortical Areas findings. Initial research, begun in the 1960s,
The bulk of our knowledge of possible forms of showed that cortical maps develop early in life
cortical map organization comes from studies of and that their structure can be altered by expo-
visual cortical areas. Maps are present in other sure, or lack of exposure, to specific stimuli dur-
sensory areas (Mountcastle 1998; Inan and Crair ing and after the period of map development.
2007), but it is not clear if multiple features are These findings inspired the development of vari-
overlaid in them with comparable complexity to ous classes of computational neuroscience model
visual cortex. Auditory cortex contains an that sought to explain cortical map development
ordered one-dimensional map of frequency, i.e., as the outcome of a self-organizing, activity-
preferred auditory frequency is constant dependent process in which neural activity plays
along one axis but varies, approximately as an a critical role. However, it is now known that in
exponential function of distance, along the the superior colliculus and optic tectum of lower
orthogonal axis. There is evidence that other audi- vertebrates, retinotopic maps develop under the
tory response properties, such as envelope period- influence of molecular guidance cues
icity, may be mapped along the orthogonal (McLaughlin and O’Leary 2005) and it is highly
dimension (Langner and Ochse 2006). It has long likely that retinotopic and somatotopic maps in
been known that somatosensory cortex has an the cerebral cortex do so as well (Rakic et al.
ordered map of the body surface: early studies of 2009; Cang et al. 2005; Inan and Crair 2007).
this area showed that in addition to the map of the Models based on correlated neural activity and
body surface (analogous to the retinotopic map), Hebbian learning however remain the best can-
there were local subdivisions into regions which didates for explaining the refinement of
received input from different classes of sensory retinotopic maps in the cortex (Huberman et al.
receptor, e.g., superficial and deep. These repre- 2008) as well as for the development of orienta-
sentations were found to be columnar, and studies tion and direction preference maps (White and
showing it, done in the late 1950s, were among the Fitzpatrick 2007).
first to demonstrate columnar organization
(Mountcastle 1998). Taste has recently been Sequence of Visual Cortex Map Development
shown to have an ordered representation in While the timing of developmental events mea-
mouse gustatory cortex where there are small sured in days from conception varies consider-
(~ 200–400 mm diameter) patches of cells, about ably in different species, the general order is
1–2 mm apart, that respond selectively to either similar (Fig. 2; Workman et al. 2013). Retinal
bitter, salty, sweet, or umami tastants, each of ganglion cells are generated very early in the
which has a correspondingly specialized receptor sequence, but the retina is unlikely to be respon-
in the taste buds (Chen et al. 2011). sive to patterned light stimuli until shortly before
the time of eye opening, at which point cells in
Development of Cortical Maps both the LGN and cortex have finished dividing
There is a long-standing debate about the extent and migrating and LGN axons have already made
to which human knowledge is innate or acquired synaptic connections in layer IV of the cortex.
through learning. In the narrower context of cor- Visual responsiveness is likely to depend not just
tical maps, the debate likewise centers on the on the presence of rods and cones but also on the
extent to which map development is the outcome maturation of the bipolar neurons which link rods
Cortical Maps, Activity-Dependent Development, timing of ocular dominance column development; the
Fig. 2 Timeline of development of the retina, LGN, and mouse does not possess maps of ocular dominance, orien-
visual cortex maps in mouse, cat, ferret, and macaque tation, or direction preference, nor does the macaque have
monkey. Note that while the actual timing of events dif- a direction preference map; the diagram should not be
fers substantially in the different species, the overall order relied upon for exact timing details given that it is
is similar in all of them. Several qualifications apply to the a synthesis across several species (Layout taken from
figure (discussed in the text): there is still debate over the Huberman et al. 2008; data from Workman et al. (2013))
and cones to ganglion cells, and this happens species. Mice and ferrets are born relatively
relatively late in the sequence. Prior to this, and immature, before the retina is functional and
beginning from the time that retinal ganglion before layers in the LGN and cortex have fully
cells have reached the LGN, retinal ganglion formed. Cats are born shortly before the retina is
cells fire coordinated bursts of action potentials mature and about a week before the eyes open.
that move in a wavelike fashion across the retina, Macaque monkeys are born very much later in
outwards from an originating center (Wong 1999; the sequence with the eyes opening in utero about
Warland et al. 2006). (Stages I, II, and III differ in 6 weeks before birth. As light may well penetrate
terms of cell type and transmitters and will not be into the uterus, the possibility of intrauterine
discussed here.) The centers and directions of visual stimulation exists, and this might also
wavefront motion are seemingly random but explain why the eyes would open before birth.
with a refractory behavior so that cells that have However, the extent to which vision can occur in
fired in a burst are unlikely to participate in utero remains unknown.
another one for some time. Such refractory Following birth, the trajectory of visual devel-
areas often cause propagating waves to die out. opment is species dependent. In macaque mon-
The retinal waves continue throughout the period keys, born much later relative to developmental
when the LGN layers are segregating and during milestones than other species, ocular dominance
the period when the LGN axons are innervating columns are present at birth (Horton and Hocking
layer IV of the cortex. Interference with the 1996), while orientation maps have been reported
waves has been shown to prevent segregation of very soon afterwards (Wiesel and Hubel 1974).
LGN layers and to reduce the refinement of the Ocular dominance and orientation maps begin to
retinotopic maps in the cortex (Torborg and develop in cats about a week after eye opening
Feller 2005; Firth et al. 2005; Huberman et al. and are adultlike 2–3 weeks later. Orientation
2008). As Fig. 2 shows, the timing of birth rela- maps start to develop in ferrets at the time of
tive to development varies greatly in different eye opening, followed about a week later by
direction preference maps. Ocular dominance performed outside of a critical period: in mice,
columns have been reported to be present prena- this is within 5 weeks after birth, cats about 3
tally in the ferret as well as in ferrets in which months and monkeys about 6 months. Manipu-
both eyes were removed (Katz and Crowley lation of the distribution of oriented stimuli has
2002). This observation, plus the recent discov- been done by either placing young animals in
ery of a molecular marker for ocular dominance a cylinder on which vertical lines are the only
(Tomita et al. 2013), suggests that neural activity visible contour or by fitting cylindrical lenses to
might not play a significant role in the formation the eyes: these cause orientations not matching
of ocular dominance columns. However, pat- the axis of the cylinder to be blurred. Such
terned spontaneous activity may still be present manipulations have been found to bias the dis-
in the LGN of blinded ferrets, and its role cannot tribution of orientation preferences and, corre-
be ruled out. Likewise, a role for retinal waves, or spondingly, the structure of orientation maps in
even for intrauterine visual stimulation, cannot be cat visual cortex (Sengpiel and Kind 2002). The
ruled out in the development of ocular dominance development of direction preference has also
columns which are present at birth in monkeys. been shown to be prevented by rearing animals
Although orientation maps normally develop in in stroboscopic light which abolishes motion
the presence of visually driven activity in cats and cues (White and Fitzpatrick 2007).
ferrets, the presence of the maps in animals that
have been dark-reared, as well as in newborn Vector Representation of Map Parameters
monkeys, suggests that either neural activity is As outlined above, the receptive fields of neurons
not involved or that retinal waves, together with in V1 can be parameterized along a variety
patterned spontaneous LGN and cortical activity, of dimensions represented in maps (Fig. 1d).
are able to cause maps to form in the absence of For example, in cat visual cortex, we might rep-
visually driven input. The observation that resent the list as follows: receptive field position
blocking activity in ON-center retinal ganglion (x, y), ocular dominance (d), preferred orientation
cells prevents the development of orientation (itself a vector, z = (a, b)), and preferred spatial
maps (Chapman and Gödecke 2000) suggests frequency (f). We might represent the list of
that neural activity is involved however. properties measured for cells at a point, i, on the
As noted above, mice do not have maps of cortical surface, by the vector vi = {xi, yi, di, ai,
ocular dominance or orientation, though individ- bi, fi}. The complete mapping can be conceptual-
ual cells have these properties which hence may ized as the projection of the 2D cortical sheet into
arise through activity-dependent interactions. a higher-dimensional feature space S. This con-
Although the role played by patterned neural struction replies on a number of assumptions.
activity, whether spontaneous or visually Firstly, the representation of the cortex as a 2D
driven, in the development of orientation and sheet is based on the hypothesis of columnar
ocular dominance maps remains uncertain, organization. This appears generally to be true
there is no doubt that alterations in the patterns although exceptions to it have been claimed. Sec-
of visual input can cause modification of map ond, it should be noted that the values of xi, yi,
structure. Monocular deprivation (closing one etc. are the best values of response functions that
eye), if done shortly after birth, causes a loss are often Gaussian, or approximately Gaussian,
of inputs from the deprived eye and a gain in the in shape; hence, associating a point in the cortex
number and strength of inputs from the other with a single position in S is an oversimplifica-
eye (Hubel and Wiesel 1977). In cats and mon- tion. Instead it would be more correct to think of
keys, there is a corresponding change in the positions in the cortex each populating S with
widths of the ocular dominance columns with spikes falling off in density from a central point.
active geniculate inputs taking over cortical ter- Thirdly, the representation of the receptive field
ritory formerly innervated by the deprived eye. as a point assumes that the best values x, y, d,
Monocular deprivation has little effect if etc. are independent of each other, e.g., that the
preferred orientation is independent of the spatial
frequency of the stimulus and its precise position.
More formally this means (or assumes) that the
response to a stimulus, u, R(u), can be expressed
as the separable product of receptive field func-
tions for each of the dimensions, i.e., that
Ri(u) = F1(x xi).F2(y yi). . . where F() is C
typically a Gaussian. The assumption of separa-
bility (see Yu et al. 2005 for discussion) and the
consequent ability to describe cortical receptive
fields as positions, or regions, in a feature space
underlie many of the models that will be
described in the following sections.
The Models
Cortical Maps, Activity-Dependent Development,
Linear Correlation-Based Models
Fig. 3 The setting for the ocular dominance column
All activity-based models of cortical map model described in the text
development assume that maps result from
a combination of Hebbian synaptic plasticity,
spatially structured activity in the input layer(s), imposed separately (i.e., following updates of
and, necessarily, some form of structured local the weight values at each computational step),
connectivity in the cortex. (If no local cortical while the main description of the model will
connectivity was present, maps could not arise take the form of equations describing the rate of
since the position of a cell in the cortical sheet weight change as a function of the weights, dis-
would have no influence on the type of receptive tance-dependent correlation functions in the
field that emerged.) The class of correlation- input layers, and the cortical interaction function.
based models discussed in this section further However, normalization, and the way in which it
assume (a) that cortical neurons sum their inputs is done, is arguably as important as the rest of the
linearly, i.e., that the activity of a cortical unit is model as this step is often the only one that allows
given by the dot product of a weight vector w and weight values to decrease. This is obviously an
an input vector u; (b) that weight changes take essential element in developing response selec-
place slowly on the timescale of input activity tivity. Models that use the BCM learning rule,
variation, allowing the rate of weight change to which explicitly includes a mechanism for
be calculated as an average of activity; (c) that decreasing weights as part of the learning rule
local connectivity in the cortex is described by (Dayan and Abbott 2001), are an exception
a Mexican Hat function, typically a difference of to this.
Gaussians; and finally (d) that the models enforce The use of these assumptions is illustrated in
competition between synapses in the form of a simplified model of ocular dominance column
normalization rules. These rules can enforce con- development presented by Dayan and Abbott
stancy of the sum of the inputs to a neuron; con- (2001). This model is itself based on earlier
stancy of the sum of the outputs of an input models. In this setting (Fig. 3), two units with
neuron, or sometimes both. Normalization can activities uL and uR connect, via sets of modifi-
be done by either multiplicatively scaling all the able feedforward weights wL and wR, to a single
weights or subtracting from them. If subtractive layer of cortical units with activities vi, where i
normalization is done, additional rules have to be indexes position on the cortical surface. The cor-
implemented to keep weights within realistic bio- tical units are connected by a set of fixed recur-
logical bounds, i.e., zero and some maximum rent (i.e., lateral) weights Mi,i* with the property
value. Normalization constraints are usually that the strength of the connection varies as
a Mexican hat-shaped function of the distance where Q = huuTi is the input correlation matrix.
d = |i i*| between two positions i and i*. The For two inputs, this has the form
vector u = (uL, uR) represents the input activi-

ties, the matrix W = (wL, wR) represents the qS qD
feedforward weights, v represents the activities qD qS
of the cortical units, and M represents the con-
nections between cortical units. A simple inter- where qS is the within-eye correlation (for one
pretation of Hebbian learning is that weights unit, this is just the r.m.s firing rate) and qD is the
change in proportion to the time-averaged prod- between-eye correlation. In general, we expect
uct of the activities of the input (pre-) and output the firing patterns to be different in the two
(post-) units, i.e., that eyes, which guarantees that qS > qD.
Equation 4 can be solved for sum and differ-
dW T ence terms of wR and wL, i.e., by defining w+ =
tw ¼ vu (1)
dt wR + wL and w = wR wL. We then have
where tw is a learning rate constant. To evaluate dw

tw ¼ ðqS qD ÞKw (5a)
this, we first have to calculate the output activi- dt
ties, v, which will depend on recurrent connec-
tions as well as the input activities and the and
connection weights. This can be simplified by
finding the steady-state solution of the following dwþ
tw ¼ ðqS þ qD ÞKwþ : (5b)
recurrent model: dt
dv Solutions to Eq. 5a for qS > qD take the form

tv ¼ v þ Wu þ Mv (2) of a sum of exponentially growing modes deter-
dt
mined by the eigenvalues and eigenvectors of K.
which is given by The eigenvectors are the discrete Fourier compo-
nents of K and hence are periodic functions of
v ¼ Wu þ Mv: cortical distance |i i*|. The dominant, most
rapidly growing mode will be the one with the
This has the solution largest amplitude (Fig. 4). Solutions to Eq. 5b
represent even more rapidly growing modes;
v ¼ KWu (3) however, these represent an increase in the values
of the summed weights from both eyes to each
where K = (I M) 1 and I is the identity cortical cell. Application of normalization rules
matrix. Note that K can be regarded as will ensure that dw+/dt = 0. Periodic fluctuations
a redefined cortical connection matrix which in the value of w = wR wL can be identified
allows a one-step calculation of the cortical with the reciprocal periodic fluctuations in ocular
response to a given input vector as dominance observed in the real cortex.
A more elaborate version of the above model
X
Nv
was presented by Miller et al. (1989). In this model,
vi ¼ K i, i wR, i uR þ wL, i uL : two input layers (rather than just two units),
i ¼1 labeled R and L, project to a sheet of cortical
units. Position in the input layers is given by a or
Substituting Eq. 3 into Eq. 1 gives b and position in the cortex is given by x or y (note
that for 2D arrays these are all 2D position vectors,
dW and not indices). Units in each input layer
tw ¼ KW uuT ¼ KWQ (4)
dt project to a small area of the output layer as
Cortical Maps, Activity-Dependent Development, components (dashed line). (b) The Fourier transform of
Fig. 4 (a) A Mexican Hat (difference of Gaussians) lat- the function, K~ , showing a maximum at a positive wave-
eral interaction function, K, showing the largest eigenvec- length, k (From Dayan and Abbott 2001)
tor (e) or (equivalently) the largest of its Fourier
defined by an arbor function A(x a). A = 1 if a cortical point x having a receptive field
the distance |x a| is less than a fixed arbor centered at position a = x in the input layer.
radius and is zero elsewhere. This formulation Connection strengths from the R input layer
enforces a rigid retinotopic mapping with change according to
dwR ðx, a, tÞ X
¼ eAðx aÞ K ðx yÞ½QRR ða bÞwR ðy, b, tÞ þ QRL ða bÞwL ðy, b, tÞ (6)
dt y, b
where e is a constant determining the overall models although the reasons for it, and possible
growth rate, K is the cortical interaction function, quantification of the morphology so that model
and the functions QRR and QRL describe the outputs and biological data can be more closely
dependence of the auto- and cross-correlation compared, have received relatively little theoret-
functions of LGN activity with distance. Miller ical attention.
et al. assumed that the within-eye correlation Equation 6 can be applied to the development of
functions were positive Gaussians, while the orientation preference. In this model, the two input
between-eye correlations were either zero or neg- layers represent ON- and OFF-center retinal inputs.
ative Gaussians. The equation for dwL/dt is Model parameters are adjusted so that segregation
obtained by interchanging L and R in Eq. 6. of inputs occurs on a small scale, within receptive
The behavior of this system is similar to that of fields, leading to alternating bands of ON and OFF
Eq. 4 with reciprocal periodic fluctuations in the inputs to each cortical unit and hence resulting in
ocular dominance of the cortical receptive fields orientation selectivity. Although neighboring cor-
developing across the surface of the cortex. As tical cells have similar orientation preferences (the
with Eq. 4, the periodicity is predictable from the result of the overlap in their receptive fields), the
frequency for which the Fourier power spectrum overall layout of orientation in this model is
of K has a maximum amplitude. Less predictably, nonperiodic, which is at odds with real orientation
the model (and others like it) generates spatial maps. A combined model, with ON and OFF layers
patterns of branching stripes and elongated blobs for each of the two eyes, giving four input layers,
that closely resemble the biological patterns. The has also been presented as a model for the joint
production of stripes from circularly symmetric development of orientation and ocular dominance
interactions can be regarded as a success of the maps (Erwin and Miller 1998).
Competitive Hebbian Models point, i*, for which vi is a maximum. Weights are
Equation 3 shows that the cortical response to an then changed according to
input vector is the product of the vector and the
weight vector for the same point, filtered by the wi ðt þ 1Þ ¼ aðtÞ½wi ðtÞ þ ehði , iÞuðtÞ (7)
lateral cortical interaction function. The
response is proportional to the match (or projec- where a is a normalization constant chosen to
tion) of the input vector in the space of weights keep the sum of the weight values (or the sum
(which has as many dimensions as there are of the squares) constant for each cortical point, e
inputs to each cell) and the weights of that cor- is a learning rate constant and h(i*,i) defines the
tical unit. This satisfies the intuitive notion that cortical neighborhood function in terms of
the stimulus that matches the receptive field the distance between points i and i*, and u(t) is
profile best will be the one that gives the largest the input presented at timestep t. The neighbor-
response. If the stimulus does not match the hood function h is normally a Gaussian and for
weights exactly, Hebbian modification will cortical simulations typically is a few hundred
have the effect of changing the weights so as to microns in width. In simulations employing this
make that cortical point more responsive to the rule, static input patterns are presented one at
stimulus in future presentations. In weight a time and weights are updated at each step.
space, this means that the weight vector moves Adaptive weight updating means that cortical
towards the position of the stimulus in the same units become selectively responsive to those
space. As shown in Eq. 4, the assumption of inputs presented during training and lose their
linear cortical responses allows the calculation responsiveness to stimuli not presented. The cor-
of weight changes to be based on time-averaged tical neighborhood function, which causes points
correlation functions rather than on a sequential near the winning one to become more responsive
presentation of many different input patterns. to the same stimulus, irrespective of their weight
This speeds up simulations but does not allow values, enforces continuity in the representation
the possible effects of nonlinearities in the cor- of the those points in the stimulus space. Hence,
tical response to be taken into account. Kohonen the Kohonen algorithm performs a dimension-
(1997) took an alternative approach to simplify- reducing, topology-preserving mapping of points
ing the calculation of cortical activities by find- in a high-dimensional input space onto a two-
ing the cortical point, i*, whose weight vector dimensional cortex.
best matched the input vector. Weight values of In the low-dimensional version, stimuli are
this point, and, crucially, values of surrounding represented not as activity distributions in sheets
cortical points in a region defined by of cells but as points in a reduced feature space
a neighborhood function, are then adjusted to whose dimensions are the receptive field param-
make these points more responsive to the stimu- eters represented in the map. Weights in this
lus. The fact that only a single cortical point is formulation represent stimulus selectivities of
chosen, no matter what the stimulus, makes the the cortical units rather than single synapses.
learning competitive and highly nonlinear. The The learning rule for this version of the model is
model has been applied in two forms, a high-
dimensional version and a low-dimensional
wi ðt þ 1Þ ¼ wi ðtÞ þ e uðtÞ wi ðtÞ hði , iÞ (8)
version. The setting for the high-dimensional
version is similar to the one above where one
or more sheets of cells project to a cortical sheet. Here, point i* is defined as the one for which
In this case however, lateral cortical connections |u w| is a minimum, i.e., it is the cortical point
are assumed rather than specified explicitly. The that is closest (most responsive) to u. At each
winning cortical location is found by calculating step, learning has the effect of moving wi*, and
vi = wiu for all cortical locations and finding the its cortical neighbors, closer to u, the currently
presented stimulus. For visual cortex maps, the In the limit of b ! 1, the output is a delta
dimensions of the feature space inhabited by function positioned at the largest value of vi
w and u might be chosen to include retinotopic leading to the winner-take-all behavior of
position (2 dimensions), ocular dominance (1), the Kohonen model. In the limit b ! 0, vi !
and orientation (2) to give a five-dimensional wiu giving the linear noncompetitive learning
space, though other parameters such as spatial model.
frequency or direction of motion might also be C
added. This rule has a simple geometrical inter-
pretation. Points i index positions in the cortex The Elastic Net Model
which may typically be represented on The low-dimensional version of the Kohonen
a rectilinear or hexagonal lattice. The function mapping algorithm (Eq. 8) causes a 2D cortical
wi maps these points into the feature space; sheet to project into a higher-dimensional fea-
hence, one can imagine the cortex as a 2D lattice ture space in such a way that the projection is (a)
of points folded into the space. Points u in the locally continuous (resulting from the neighbor-
space are chosen one at a time, according to hood function) and (b) the sheet approximates
a probability distribution p(u), and for each, the the distribution of stimuli within the space. The
cortical point that is closest is found. This point elastic net algorithm, devised by Durbin and
and its neighbors are moved towards the stimu- Willshaw (1987), achieves a similar end, but
lus. This process is repeated (possibly for many by different means. The term “elastic net”
thousands of stimuli) until a map has formed. applies to a lattice of cortical points which can
Both high- and low-dimensional forms of the be conceived as an elastic net with adjacent
Kohonen model have been used to model visual nodes on a square lattice connected by
cortex map development (Erwin et al. 1995; elastic. Under the influence of the elastic forces,
Swindale 1996, 2003). The low-dimensional ver- the net would collapse to a point. However,
sion has the advantage of being simple to imple- stimuli – fixed points in the feature
ment and it is relatively fast to compute map space – exert attractive forces on the net, which
layouts for reasonably sized areas of cortex get stronger the shorter the distance between the
(e.g., a simulation involving five dimensions stimulus and the cortical point. The algorithm
and 200 200 sized cortex corresponding to an calculates a stable configuration of the net (i.e.,
area of about 4 4 mm might take a few hours or cortical receptive field values) for a given, fixed
less to compute). Unlike the linear correlation- set of stimuli distributed in the space in some
based model discussed above, a fixed retinotopy manner. The learning rule is
is not assumed, allowing investigations of possi-
ble relations between local magnification of the X
wi ðt þ 1Þ ¼ wi ðtÞ þ a Fi, j : vj wi
retinotopic map and the rates of change of other
j
map parameters. X
Piepenbrock and Obermayer (2000) intro- þ bs ðwk wi Þ (10)
duced a model that bridges the extremes of linear kN i
and competitive cortical activation models. In

this framework, cortical activation is produced where a and b are rate constants, and the summa-
by a softmax function: tion in the second term is over the nearest neigh-
bors k of point i. The “force” Fi,j exerted by
0 1
b stimulus j on cortical point i is a Gaussian func-
exp bwi u B wu C tion of the distance between wi and vj (i.e., the
vi ¼ X Bor vi ¼ X i C
expðbwi uÞ @ bA
ðwi uÞ response, assuming Gaussian receptive fields)
i i normalized by the sum of the responses from all
(9) other cortical units, i.e.,

2
exp vj wi =2s2
Fi , j ¼ X
2 : (11)
exp vj wp =2s2
p
It should be noted that this formulation results

in the elastic net algorithm being nonlocal, i.e., it
is necessary to sum over every point in the cortex
to determine the learning rate of any particular
cortical point. This might be criticized as
unbiological. However, normalization ensures
that stimuli that are far away from any cortical
point do not get ignored and exert forces that are
as large as those exerted by stimuli that are closer
to cortical units. Parameter s scales the distance
over which stimuli exert attractive forces. Ini-
tially this distance may be large, and it is typically
reduced in size (a process referred to as “simu- Cortical Maps, Activity-Dependent Development,
lated annealing”) in order to make individual Fig. 5 Cortical map model produced by the elastic
cortical points approach specific stimuli. net algorithm. Black lines (lowest layer) represent ocular
dominance boundaries, green lines (top layer) represent
The expression for weight change can be inte-
spatial frequency domain boundaries, thin blue- and red-
grated to yield an energy function: colored lines (middle layer) show iso-orientation domain
X X
2 boundaries, and thick red lines mark lines across which
E ¼ as log exp vi wj =2s2 direction preference reverses. These lines begin and end in
i j orientation singularities. Note the frequent orthogonal
intersections of all types of boundary. This figure can be
1 X 2
þ bs wj wjþ1 compared with the experimental data shown in Fig. 1d
2 j
(From Carreira-Perpinãń et al. 2005)
(12)
By definition, this has the property that Dwj =
@E/@wj. At each iteration, @E(= Dwj@wj) is this part of E measures what has elsewhere been
guaranteed to be negative provided that Dwj is termed “coverage uniformity” – the idea being
small enough that its sign is always the same as that the cortical maps ought to be organized so
@wj. Hence, the learning rule will always lead that the net response to a stimulus should be
to a decrease in the value of E. The expression invariant with respect to parameters such as reti-
for E has an interesting interpretation. The term nal location, eye of stimulation, and orientation
exp(|vi wj|2/2s2) can be interpreted as the (Swindale 1996). The second term in Eq. 12 mea-
response of cortical point j to a stimulus i, sures the sum of all the distances between the net
a Gaussian function of the distance between the points as projected into the feature space. Hence,
stimulus vi, and the receptive field center, wj. The it is a measure of how smooth the map is. It can be
summation over cortical points indexed by j thus minimized by contracting the net to a point, but
measures the total response of the cortex to this will clearly lead to extremely unequal
a given stimulus (assuming Gaussian tuning responses to different stimuli. Depending on the
functions). The log of this term is then summed ratio between parameters a and b, maps that
over i, that is, over all the stimuli. As in an minimize E will represent a trade-off between
entropy measure, the magnitude of this sum will coverage uniformity and local smoothness. This
be largest, and hence the contribution to E most fact, plus the finding that maps generated with the
negative, when the distribution of responses over elastic net algorithm (Fig. 5) resemble real visual
the stimuli (for a fixed total sum) is flat. Hence, cortex maps (for a range of values of a and b),
suggests that cortical maps also represent strengthening of inhibitory synapses in propor-
a compromise between coverage uniformity and tion to the product of pre- and postsynaptic activ-
local smoothness. Local smoothness is not in ities. This rule was initially proposed by Barlow
itself a functional constraint, but it may be related and Foldiak (1989) and was termed “anti-
to axonal connection lengths, if it is assumed that Hebbian” learning by them. It has the effect of
the total number of connections between cell decorrelating cortical responses to correlated
pairs increases with the similarity of their recep- input patterns and leads to a more sparse repre- C
tive fields. Since axons take up a significant pro- sentation of visual images. It is likely that the
portion of the volume of cortex, arranging a map LISSOM learning rule will lead to the develop-
so that cells with similar receptive fields are as ment of receptive fields that are the independent
close to each other as possible (subject to the components of the input images which, as shown
conflicting constraint of coverage uniformity) elsewhere for natural images, are oriented Gabor-
will reduce the volume of cortex. like filters closely resembling the receptive fields
of simple cells (Olshausen and Field 1996;
Models Based on Plasticity of Lateral Hyv€arinen and Hoyer 2001).
Intracortical Connections The assumptions of LISSOM have mixed
The LISSOM (Laterally Interconnected experimental support: there is evidence for
Synergetically Self-Organizing Map) model was Hebbian plasticity of inhibitory cortical synap-
devised initially by Miikkulainen and has since ses (Gaiarsa and Ben-Ari 2006); however,
been extended and explored by him and his col- experimental results show that longer-range
leagues (Miikkulainen et al. 2005). LISSOM dif- (1–5 mm) lateral cortical connections (which
fers from most other cortical map models by are the ones that would seem most likely to be
having modifiable short range excitatory and lon- involved in decorrelating images) are excitatory
ger-range inhibitory lateral connections in the while inhibitory connections travel shorter dis-
cortex. The model exists in various forms but tances (Gilbert 1992). However, these shorter
a standard implementation has a single layer range connections might still play a role in
of retinal units, representing a visual input, decorrelating cortical responses. The fact that
connecting to layers of ON- and OFF-center much cortical map development takes place in
LGN units via fixed, topographically ordered the absence of visually driven activity (and
weights. These units project to a single layer of hence interesting higher-order structure) sug-
laterally interconnected cortical units. LGN and gests that natural images do not play a role in
cortical units compute their responses as the dot the formation of simple cell receptive fields so
product of the input and weight vectors with that the types of developmental plasticity that
a sigmoidal activation function that constrains LISSOM would seem best at modeling are per-
activities to lie between 0 and 1. Cortical activi- haps those that occur at relatively late stages of
ties are computed by iterating an activation equa- map development.
tion which includes the lateral inputs. Responses
are calculated to individual stimulus patterns and Overview of “Simple” Models of Map
then the LGN to cortex weights and the lateral Development
weights are modified according the product of The models outlined above are variably, some-
pre- and postsynaptic activities. Following this, times highly, simplified abstractions of cortical
weights are normalized divisively. map development. The advantages of such sim-
Much of the interest of the LISSOM model plification are many. They make simulations
comes from studying its behavior when trained computationally feasible and they allow model
with structured visual inputs (e.g., natural behavior to be understood in terms of functional
images) which have higher-order correlations properties, such as wirelength minimization, that
that are ignored or absent in standard correlation are likely to have been important constraints on
models. LISSOM incorporates Hebbian the evolution of developmental mechanisms.
Because the number of parameters is often necessitate highly detailed, though also highly
small, empirical studies of the effects of chang- constrained, computer modeling of biochemical,
ing them is much easier. Mathematical simplic- cellular, and multicellular interactions and envi-
ity also allows for the possibility of making ronmental stimulation. A major value of such
predictions based on analysis rather than com- models may lie in their ability to predict
puter simulation and for discovering relation- a phenotype given one or more genetic mutations
ships between different kinds of models. This known to play a role in mental illness. Perhaps in
form of understanding has allowed many of the anticipation of such a scenario (which may be
models to be usefully applied to problems other decades away), several recent models of visual
than simulating cortical map development. For system development have simulated the behaviors
example, the elastic net algorithm can produce of growing axons, incorporating branching and
solutions to the traveling salesman problem, retraction rules, the effect of molecular markers
while the Kohonen algorithm, initially inspired (Ephrin and eph receptor gradients), neuronal
by the problem of cortical map development, has growth factors, synaptic plasticity rules, and
arguably had more applications in areas outside detailed simulations of retinal input activity pat-
developmental neuroscience than within it terns (Yates et al. 2004; Godfrey et al. 2009;
(Kohonen 1997). Many forms of connection Grimbert and Cang 2012). While they are capable
between Kohonen, elastic net, and correlation- of correctly replicating known macroscopic devel-
based models have been drawn, the upshot being opmental features as well as a variety of experi-
that all of the models can be shown to perform mental results, these models currently have the
gradient descent, at least approximately, on disadvantage that computations are slow to com-
a function that represents a trade-off between plete and this, plus their complexity, makes it very
the conflicting constraints of coverage unifor- hard to assess and predict the effects of changes in
mity and local smoothness (Goodhill and rules and parameter values on model output. This
Sejnowski 1997). situation is likely to change as computers get faster
and as the models get more tightly constrained by
“Realistic” Models of Cortical Map biological findings.
Development
These strengths notwithstanding, abstract Experimental Tests of Model Predictions
models can be criticized because of the many For the class of “simple” model discussed in this
details they leave out. These details are likely to article, the interplay between experiment and
play roles in development that will be best theory has been vigorous and interesting. Despite
understood in the context of a computational their success however, the assumptions on which
model. Real cortical development is the result most, if not all, of the models are based remain
of the guided outgrowth, branching, and retrac- unproved. The most common of these is the pos-
tion of axons, together with the reciprocal exten- tulate of radially symmetric Mexican Hat con-
sion and retraction of dendrites and the nectivity in the cortex. Although this exists in
coordinated formation and loss of synaptic con- the retina, there is no evidence for a greater
nections between the two. The field of research spread of inhibitory connections relative to the
studying these processes is large and rapidly spread of excitatory connections in the cortex.
progressing and a time should come when the The connections that go furthest in the visual
vast majority of the genetic and molecular mech- cortex are excitatory and have a patchy, asym-
anisms underlying these processes will have been metric spread which is at odds with radial sym-
characterized. Arguably, such a reductionist metry (Gilbert 1992). It remains possible that
description will not be complete until it can be there is a second class of lateral excitatory con-
shown that larger-scale phenomena can be cor- nections which are extremely local (<250 mm),
rectly predicted as the outcome of lower-level and these, combined with the known inhibitory
molecular interactions. This will almost certainly ones, might provide the desired connectivity.
However, this remains a speculation. Neither the afferents, as well as in ferrets whose eyes had
Kohonen nor elastic net models explicitly require been removed (Katz and Crowley 2002). The
Mexican Hat connectivity; however, the elastic implication that molecular cues, rather than neu-
net model has a normalization term that requires ral activity, may be involved in segregation is
summing over the entire cortex, and since no also supported by the recent finding of
known intracortical connections span such dis- a molecular marker for ocular dominance col-
tances, it is not clear how this might be umns (Tomita et al. 2013). Whether the molecu- C
implemented biologically. The winner-take-all lar markers play a primary or a passive role in
mechanism of Kohonen likewise requires segregation, or interact with activity-dependent
a biologically unfeasible comparison across the mechanisms, remains to be determined however.
entire cortex (as does the related model of The role of correlated activity in the develop-
Piepenbrock and Obermayer). In these cases, ment of orientation selectivity is also unclear,
more plausible models might be constructed that mostly because fewer experiments have been
avoided global comparisons. done to study the question. Blocking activity in
By definition, all the models discussed here retinal ON-center ganglion cells in ferret retinas
rely (implicitly if not explicitly) on the presence prevents the development of orientation selectivity
of local correlations in the patterns of neural (Chapman and Godecke 2000). This finding, plus
activity in the inputs to the cortex which come evidence that differences in the spontaneous firing
primarily from the LGN. Here, the experimental patterns of ON- and OFF-center ganglion cells
evidence is more supportive. Retinal wave activ- emerge about halfway during the period of retinal
ity has now been fairly thoroughly characterized. wave activity (Kerschensteiner and Wong 2008),
The waves introduce correlations between gan- supports the ON-OFF competition-based model of
glion cell activity that fall off approximately as Miller (1994). However, the model has the weak-
Gaussian functions with distance (Wong et al. ness that it does not produce realistic periodic
1993; Stafford et al. 2009). The wave activity is patterns of orientation preference while the param-
transmitted through the LGN to the cortex eter regime required for simultaneous emergence
(Ackman et al. 2012) and hence is capable of of both ocular dominance and orientation prefer-
influencing cortical development in the manner ence is narrow (Erwin and Miller 1998).
proposed by correlation-based models. The In spite of their relative lack of empirical sup-
cross-eye correlations for spontaneous activity port, many of the models, especially those based
would be expected to be zero although Ackman on Kohonen and elastic net algorithms, have been
et al.’s data suggest the possibility of cross-eye spectacularly successful in reproducing and in
coordination which has so far been assumed not several cases correctly predicting experimental
to exist. There is a considerable body of evidence observations. Among these successes can be
showing that blocking or altering patterns of counted the visual resemblance of the layouts of
spontaneous activity in the retina interferes with ocular dominance and orientation domains pro-
cortical map development. Examples of this duced by the majority of the models and those
include the finding that blocking spontaneous observed experimentally. The presence of half-
retinal activity prevents refinement of retinotopic rotation singularities in the orientation maps,
maps in mouse visual cortex and in ferrets dis- their overall density, their tendency to lie in the
rupts ocular dominance column formation, centers of ocular dominance stripes, and the sta-
though it does not completely abolish it tistics of the distributions of positive and negative
(Huberman et al. 2008). This evidence that the singularities are all correctly reproduced (Erwin
initial stages of ocular dominance column forma- et al. 1995; Swindale 1996, 2003; Goodhill
tion are activity dependent has to be set against 2007). The Kohonen and elastic net models all
the finding that ocular dominance columns could produce maps in which gradients of change in
be observed in ferrets almost simultaneously with specific maps (e.g., ocular dominance or spatial
the initial innervation of layer IV by LGN frequency) tend to be orthogonal to each other,
most likely because of the algorithms’ optimiza- Cross-References

tion of coverage (Durbin and Mitchison 1990;
Swindale 2000; Carreira-Perpinãń et al. 2005). ▶ Emergence of Orientation Selectivity in the
Orthogonal gradient relationships have been Cerebral Cortex, Modeling
observed experimentally between orientation ▶ Hierarchical Models of the Visual System
and ocular dominance, ocular dominance and ▶ Network Theory in Neuroscience
spatial frequency, ocular dominance and dispar- ▶ Somatosensory Cortex: Organization
ity, and spatial frequency and orientation ▶ Topographica
(H€ubener et al. 1997; Yu et al. 2005; Nauhaus ▶ Visual Hallucinations and Migraine Auras
et al. 2012). A prediction that reducing the num- ▶ Wiring Principles, Optimization
ber of feature dimensions present in a map should
lead to an increase in the spacing of the remaining
features (Swindale 2004; Carreira-Perpinãń References
et al. 2005) was also tested and confirmed (Farley
Ackman JB, Burbridge TJ, Crair MC (2012) Retinal
et al. 2007). waves coordinate patterned activity throughout the
developing visual system. Nature 490:219–225
Final Words Adams DL, Horton JC (2003) Capricious expression of
cortical columns in the primate brain. Nat Neurosci
“Nothing in biology makes sense except in the
6:113–114
light of evolution” and the models presented Barlow HB, Földiák P (1989) Adaptation and
here arguably only make sense because they decorrelation in the cortex. In: Durbin RM, Mial C,
show how activity-based Hebbian learning Mitchison GJ (eds) The computing neuron. Addison-
Wesley, Wokingham, pp 54–72 (Chap 4)
rules can optimize a variety of properties likely
Bonin V, Histed MH, Yurgenson S, Reid RC (2011) Local
to have been subject to selection pressure. These diversity and fine-scale organization of receptive fields
include efficiency of information representation in mouse visual cortex. J Neurosci 31:18506–18521
as reflected by coverage uniformity and sparse- Born RT, Tootell RB (1991) Spatial frequency tuning of
single units in macaque supragranular striate cortex.
ness of coding: good coverage is needed to rep-
Proc Natl Acad Sci U S A 88:7066–7070
resent the information in images faithfully, Cang JH, Kaneko M, Yamada J, Woods G, Stryker MP,
while sparse coding minimizes energy use Feldheim DA (2005) Ephrin-As guide the formation of
(Vincent et al. 2005). Minimizing wirelength functional maps in the visual cortex. Neuron
48:577–589
leads to a reduction in brain volume and faster
Carreira-Perpinãń MÁ, Lister R, Goodhill GJ
processing (Chklovskii and Koulakov 2004). It (2005) A computational model for the development
may be pointed out however that these perfor- of multiple maps in primary visual cortex. Cereb Cor-
mance optimizations offer little insight into the tex 15:1222–1233
Chapman B, Gödecke I (2000) Cortical cell orientation
kinds of cognitive learning that take place in the
selectivity fails to develop in the absence of ON-center
cortex after birth. At the other end of the nature- retinal ganglion cell activity. J Neurosci 20:1922–1930
nurture debate, the forms of topographic connec- Chen X, Gabitto M, Peng Y, Ryba NJP, Zuker CS
tivity that molecular gradients are known to gen- (2011) A gustotopic map of taste qualities in the mam-
malian brain. Science 333:1262–1266
erate seem a long way removed from the kinds of
Chklovskii DB, Koulakov AA (2004) Maps in the brain:
connectivity that would be needed to establish what can we learn from them? Annu Rev Neurosci
such complex things as universal rules of gram- 27:369–392
mar or neonatal abilities such as speech and Dayan P, Abbott LF (2001) Theoretical neuroscience.
MIT Press, Cambridge, Massachusetts
face recognition (Marcus 2004). A considerable
De Valois RL, De Valois KK (1990) Spatial vision.
amount of work remains to be done to show how Oxford University Press, New York Oxford
innate, genetically based mechanisms can lead Durbin R, Willshaw DJ (1987) An analogue approach to
to the establishment of neural structures on the travelling salesman problem using an elastic net
method. Nature 326:689–691
which Hebbian (or other) correlation-based
Durbin R and Mitchison G (1990) A dimension reduction
learning rules can operate to produce the behav- framework for understanding cortical maps. Nature
iors that are critical to the functions of the cortex. 343:644–647
Eglen SJ, Gjorgjieva J (2009) Self-organization in the Huberman AD, Feller MB, Chapman B (2008) Mecha-
developing nervous system: theoretical models. nisms underlying development of visual maps and
HFSP J 3:176–185 receptive fields. Annu Rev Neurosci 31:479–509
Erwin E, Miller KD (1998) Correlation-based develop- Hyv€arinen A, Hoyer PO (2001) A two-layer sparse coding
ment of ocularly matched orientation and ocular dom- model learns simple and complex cell receptive fields
inance maps: determination of required input and topography from natural images. Vision Res
activities. J Neurosci 18:9870–9895 41:2413–2423
Erwin E, Obermayer K, Schulten K (1995) Models of Inan M, Crair MC (2007) Development of cortical maps:
orientation and ocular dominance columns in the perspectives from the barrel cortex. Neuroscientist
C
visual cortex: a critical comparison. Neural Comput 13:49–61
7:425–468 Issa NP, Trepel C, Stryker MP (2000) Spatial frequency
Farley BJ, Yu H, Jin DZ, Sur M (2007) Alteration of visual maps in cat visual cortex. J Neurosci 20:8504–8514
input results in a coordinated reorganization of Kara P, Boyd JD (2009) A micro-architecture for binocu-
multiple visual cortex maps. J Neurosci 27: lar disparity and ocular dominance in visual cortex.
10299–10310 Nature 458:627–631
Firth SI, Wang CT, Feller MB (2005) Retinal waves: Katz LC, Crowley JC (2002) Development of cortical
mechanisms and function in visual system develop- circuits: lessons from ocular dominance columns. Nat
ment. Cell Calcium 37:425-432 http://dx.doi.org/ Rev Neurosci 3:34–42
10.1016/j.ceca.2005.01.010 Kerschensteiner D, Wong ROL (2008) A precisely timed
Gaiarsa J-L, Ben-Ari Y (2006) Long-term plasticity at asynchronous pattern of ON and OFF retinal ganglion
inhibitory synapses: a phenomenon that has been cell activity during propagation of retinal waves. Neu-
overlooked. In: Kittler JT, Moss SJ (eds) The dynamic ron 58:851–858
synapse: molecular methods in ionotropic receptor Kohonen T (1997) Self-organizing maps. Springer, Berlin
biology. CRC Press, Boca Raton (Chap 2) Langner G, Ochse M (2006) The neural basis of pitch and
Gilbert CD (1992) Horizontal integration and cortical harmony in the auditory system. Music Sci 10(1
dynamics. Neuron 9:1–13 suppl):185–208
Godfrey KB, Eglen SJ, Swindale NV (2009) A multi- LeVay S, Stryker MP, Shatz CJ (1978) Ocular dominance
component model of the developing retinocollicular columns and their development in layer IV of the cat’s
pathway incorporating axonal and synaptic growth. visual cortex: a quantitative study. J Comp Neurol
PLoS Comput Biol 5:1–22 (e1000600) 179:223–244
Goodhill GJ and Sejnowski TJ (1997) A unifying objec- Lu HD, Roe AW (2008) Functional organization of color
tive function for topographic mappings. Neural domains in V1 and V2 of macaque monkey revealed
Comput 9:1291-1303 by optical imaging. Cereb Cortex 18:516–533
Goodhill GJ (2007) Contributions of theoretical modeling Marcus G (2004) The birth of the mind. Basic Books,
to the understanding of neural map development. Neu- New York
ron 56:301–311 McLaughlin T, O’Leary DD (2005) Molecular gradients
Goodhill GJ, Xu J (2005) The development of retinotectal and development of retinotopic maps. Annu Rev
maps: a review of models based on molecular gradi- Neurosci 28:327–355
ents. Network 16:5–34 Miikkulainen R, Bednar J, Choe Y, Sirosh J (2005) Com-
Grimbert F, Cang J (2012) New model of retinocollicular putational maps in the visual cortex. Springer, New
mapping predicts the mechanisms of axonal York, NY
competition and explains the role of reverse molecular Miller KD (1994) A model for the development of simple
signaling during development. J Neurosci cell receptive fields and the ordered arrangement of
32:9755–9768 orientation columns through activity-dependent com-
Horton JC, Hocking DR (1996) An adult-like pattern of petition between ON- and OFF-center inputs.
ocular dominance columns in striate cortex of newborn J Neurosci 14:409–441
monkeys prior to visual experience. J Neurosci Miller KD, Keller JB, Stryker MP (1989) Ocular domi-
16:1791–1807 nance column development: analysis and simulation.
Hubel DH, Wiesel TN (1977) Functional architecture of Science 245:605–615
macaque monkey visual cortex. Proc R Soc Mountcastle VB (1998) Perceptual neuroscience: the cere-
B 198:1–59 bral cortex. Harvard University Press, Cambridge,
H€ubener M, Shoham D, Grinvald A, Bonhoeffer T (1997) Massachusetts
Spatial relationships among three columnar systems in Movshon JA, Thompson ID, Tolhurst DJ (1978) Spatial
cat area 17. J Neurosci 17:9270–9284 summation in the receptive fields of simple cells in the
H€ubener M, Hofer SB, Mrsic-Flogel TD (2008) Ocular cat’s striate cortex. J Physiol 283:53–77
dominance plasticity. In: Chalupa LM, Williams RW Nauhaus I, Nielsen KJ, Disney AA, Callaway EM
(eds) Eye, retina, and visual system of the mouse. MIT (2012) Orthogonal micro-organization of orientation
Press, Cambridge, Massachusetts London, England and spatial frequency in primate primary visual cortex.
pp 439–448 Nat Neurosci 15:1683–1690
C 892 Cortical Motor Prosthesis
Obermayer K, Blasdel G (1997) Singularities in primate Vincent BT, Baddeley RJ, Troscianko T, Gilchrist ID
orientation maps. Neural Comput 9:555–575 (2005) Is the early visual system optimised to be
Ohki K, Chung S, Ch’ng YH, Kara P, Reid RC energy efficient? Network 16:175–190
(2005) Functional imaging with cellular resolution Warland DK, Huberman AD, Chalupa LM (2006) Dynam-
reveals precise micro-architecture in visual cortex. ics of spontaneous activity in the fetal macaque retina
Nature 433:597–603 during development of retinogeniculate pathways.
Olshausen B, Field D (1996) Emergence of simple-cell J Neurosci 26:5190–5197
receptive fields properties by learning a sparse code for Weliky M, Bosking WH, Fitzpatrick DA (1996) System-
natural images. Nature 381:607–609 atic map of direction preference in primary visual
Piepenbrock C, Obermayer K (2000) The effect of cortex. Nature 379:725–728
intracortical competition on the formation of topo- White L, Fitzpatrick D (2007) Vision and cortical map
graphic maps in models of Hebbian learning. Biol development. Neuron 56:327–338
Cybern 82:345–353 Wiesel TN, Hubel DH (1974) Ordered arrangement of
Rakic P, Ayoub AE, Breunig JJ, Dominguez MH orientation columns in monkeys lacking visual expe-
(2009) Decision by division: making cortical maps. rience. J Comp Neurol 158:307–318
Trends Neurosci 32:291–301 Wolf F, Geisel T (1998) Spontaneous pinwheel
Ringach DL (2002) Spatial structure and symmetry of annihilation during visual development. Nature
simple-cell receptive fields in macaque primary visual 395:73–78
cortex. J Neurophysiol 88:455–463 Wong RO (1999) Retinal waves and visual system devel-
Sengpiel F, Kind PC (2002) The role of activity in development. Annu Rev Neurosci 22:29–47
opment of the visual system. Curr Biol 12: Wong ROL, Meister M, Shatz CJ (1993) Transient period
R818–R826 of correlated bursting activity during development of
Shmuel A, Grinvald A (1996) Functional organization the mammalian retina. Neuron 11:923–938
for direction of motion and its relationship to Workman AD, Charvet CJ, Clancy B, Darlington RB,
orientation maps in cat area 18. J Neurosci 16: Finlay BL (2013) Modeling transformations of
6945–6964 neurodevelopmental sequences across mammalian
Stafford BK, Sher A, Litke AM, Feldheim DA (2009) Spa- species. J Neurosci 33:7368–7383
tial-temporal patterns of retinal waves underlying Yates PA, Holub AD, McLaughlin T, Sejnowski TJ,
activity-dependent refinement of retinofugal projec- O’Leary DD (2004) Computational modeling of
tions. Neuron 64:200–212 retinotopic map development to define contributions
Swindale NV (1982) A model for the formation of orien- of EphA-ephrinA gradients, axon-axon interactions,
tation columns. Proc R Soc B 215:211–230 and patterned activity. J Neurobiol 59:95–113
Swindale NV (1996) The development of topography in Yu H, Farley BJ, Jin DZ, Sur M (2005) The coordinated
the visual cortex: a review of models. Netw Comput mapping of visual space and response features in
Neural Syst 7:161–247 visual cortex. Neuron 47:267–280
Swindale NV (2000) How many maps are there in visual
cortex? Cereb Cortex 10:633–643
Swindale NV (2003) Development of ocular dominance
stripes, orientation selectivity, and orientation
columns. In: van Ooyen A (ed) Modeling neural Cortical Motor Prosthesis
development. MIT Press, Cambridge, MA (Chap 12)
Swindale NV (2004) How different feature spaces
may be represented in cortical maps. Network Karthikeyan Balasubramanian and
15:217–242 Nicholas G. Hatsopoulos
Swindale NV, Matsubara JA, Cynader MS (1987) Surface Department of Organismal Biology and
organization of orientation and direction selectivity in
cat area 18. J Neurosci 7:1414–1427
Anatomy, University of Chicago, Chicago,
Tomita K, Sperling M, Cambridge SB, Bonhoeffer T, IL, USA
H€ubener M (2013) A molecular correlate of ocular
dominance columns in the developing mammalian
visual cortex. Cereb Cortex 23(11):2531–2541 doi:
10.1093/cercor/bhs232
Abbreviations
Torborg CL, Feller MB (2005) Spontaneous patterned
retinal activity and the refinement of retinal projec- ALS Amyotrophic lateral sclerosis
tions Progress in Neurobiol 76:213–235 BCI Brain-computer interface
Van Hooser SD, Heimel JA, Chung S, Nelson SB, Toth LJ
(2005) Orientation selectivity without orientation
BMI Brain-machine interface
maps in visual cortex of a highly visual mammal. ECG Electrocardiogram
J Neurosci 25:19–28 ECoG Electrocorticogram
Cortical Motor Prosthesis 893 C
EEG Electroencephalogram spikes, local field potentials, ECoGs, and EEGs,
EMG Electromyogram or a combination of these signals, can be used to
EOG Electrooculogram control a motor prosthesis. Temporal and spectral
ERP Event-related potential modulations of these signals are typically
SCI Spinal cord injury mapped (or decoded) to generate a resultant con-
trol signal, which is then fed to an actuator. The
user is normally provided with a feedback signal C
Definition on the current state of the device. Visual feedback
has been the most common type, although, more
Neuromotor prostheses or, more commonly recently, augmented feedback systems utilize
referred to as brain-machine interfaces (BMIs) tactile, proprioception, and other sensory substi-
or brain-computer interfaces (BCIs), refer to sys- tution modalities.
tems controlling prosthetic devices via an inter- BMIs have been shown to have great potential
face with ensembles of the neurons often from in restoring motor function in patients with severe
the cortex. Electrical potentials emanating motor dysfunction, to extend the motor capabili-
from neurons in the vicinity of an electrode ties of amputees, and to assist in neural rehabili-
interface are decoded to extract useful control tation. Several experimental trials on human
signals for external devices, typically an artificial subjects and nonhuman primates have shown
limb or a robot. Nonelectric potentials such as the the viability of BMIs. Nevertheless, their clinical
metabolic signals are also being used in applicability is still limited and is a subject of
some BMIs. further research. A typical BMI is comprised of
an electrode interface, a signal processing setup,
a decoding algorithm, an actuator, and a sensory
Detailed Description feedback system. The following sections discuss
the various signals acquired using different elec-
Cortically controlled BMIs utilize voluntary trode systems, some of the potential decoders,
modulations of cortical neurons in controlling actuator control strategies, and feedback modali-
an external prosthetic device. The system-level ties. The translational nature and the possible
architecture of a BMI setup is shown in Fig. 1. pathways of future BMIs are also discussed
Individual neural signals, i.e., action potential briefly.
Cortical Motor
Prosthesis,
Fig. 1 Schematic diagram
of a closed-loop
BMI. Neural signals
acquired through various
electrode technologies are
processed to extract
information such as spike
rate, spectral features,
etc. The decoder then maps
neural information to
a control variable of the
actuator. The state of the
actuator is provided as
feedback to the user
Cortical Motor EEG

Prosthesis, Electrode
Fig. 2 Schematic
representation of various ECoG
Electrode Scalp
recording modalities. EEG
Layer
electrodes are placed over
the scalp. ECoG arrays
collect signals from the
Cranial
surface of the cortex, while Layer
microelectrode arrays Microelectrode
penetrate the brain Array
parenchyma
Dura mater
Cortex
Neural Signals Cortical Motor Prosthesis, Table 1 Types of neural

signals
Recorded neural signals can be classified Spatial
according to their temporal and spatial resolution Temporal resolution
Signal type resolution (ms) (mm)
and also by the degree of invasiveness needed to
Action potential 1 10–1
acquire them (see Fig. 2). Table 1 shows the
spikes
various classes of neural signals acquired from Local field potentials ~10 10–1–1
different electrode interfaces. These signals can Electrocorticogram ~10 1–10
be decoded to estimate the neural state and/or Electroencephalogram ~25 >10
motor intentions and subsequently used to control fMRI ~1,000 1
a prosthetic device. fNIRS ~1,000 >10
Action Potential Spikes

Neurons exhibit brief 1 ms de- and repolariza-
tions, called action potentials or spikes (Fig. 3a), that M1 neurons encode signals representing limb
which are carried forward by other synaptically movement in space. Earlier experiments demon-
linked neurons. They can be spontaneously gen- strated encoding of kinetic variables in M1, while
erated or evoked by a stimulus. Volitional inten- a majority of the present BMIs attempt to decode
tions or mental imagery of motor actions also kinematic variables such as direction, position,
elicits neural spikes. These frequency-modulated and velocity.
discrete signal events can be recorded and
mapped to an actuator or end effector. Local Field Potentials
Multielectrode arrays, tetrodes, and microwire Local field potentials (LFPs) are another class of
arrays are some of the widely used penetrating neurophysiological signaling recorded from pen-
electrodes used in recording spike activity. Pri- etrating microelectrodes as continuous
mary motor cortex (M1) has been the prima facie low-frequency signals (unlike spikes which are
location to record neural signals for BMIs, discrete events) (see Fig. 3b). The “spatial reach”
although multiple frontal and parietal cortical of these signal sources ranges from <1 mm
areas have also been used by researchers. Over (Katzner et al. 2009) to over 1 cm at times
the past 50 years, strong evidence has emerged depending on the electrode size. Mathematical
a b
200 250 254
μVolts
0
μVolts
0
C
−249
−250
−250
0.0 ms 1000
0 msec 1.5 00 msec

msec 1500
Two Single Spike Units LFP Trace
Cortical Motor Prosthesis, Fig. 3 Spike waveforms and local field potential trace
models and electrophysiological experiments signals. Computational modeling studies

suggest that LFPs represent the aggregate synap- (Slutzky et al. 2010) reported optimal electrode
tic activity from a group of neurons near the spacing (for maximized information and mini-
electrode tip. Some experimenters, however, mized data redundancy) to be 0.6 mm (subdural)
argue that the LFPs are also comprised of and 0.6–0.9 mm (epidural) for rat cortices and
“volume-conducted” electrical activity from 1.7–1.8 mm (subdural) and 9–13 mm (epidural)
nonlocal sources. Despite the controversy as to for human recordings. ECoGs were originally
the origin of the LFP, changes in LFP signals employed in the domain of clinical studies where
have been shown to be linked to motor functions human subjects with epileptic seizures or tumors
as well as to other important cortical processes were examined. Recently, ECoGs are shown to
such as memory. Its increased spatial resolution have potentials for BMI applications. Higher spa-
compared to noninvasive recording, spectral tial resolution, bandwidth, signal-to-noise ratio
fidelity, and long-term stability for chronic (SNR), immunity to artifacts such as EMG or
recording has made this signal modality a viable EOG, and signal stability are some of the advan-
option for BMI applications. Information from tages of ECoGs over other semi-/noninvasive
LFPs are extracted from different frequency recording techniques. Similar to LFPs, ECoG
bands (0–300 Hz) (Andersen et al. 2004; Flint activity can be observed at various frequency
et al. 2012, 2013; Hwang and Andersen 2013). bands, i.e., theta (4–8 Hz), mu (8–12 Hz), beta
LFPs in certain frequency bands also exhibit (18–25 Hz), and gamma (>40 Hz). The
widespread synchronous couplings with neurons low-frequency bands (mu/beta) are spatially less
in motor cortex (Donoghue et al. 1998). specific and exhibit amplitude attenuation on
movement onset, while the gamma oscillations
Electrocorticogram increase in their amplitude and are spatially spe-
ECoG signals are continuous neural recordings cific (Brunner et al. 2009). Schalk et al. (2007)
from the surface of the cortex either beneath have also observed time-domain features, called
(subdural) or above (epidural) the dura mater. the local motor potentials (LMP) that could be
A strip or grid of electrodes with interelectrode used for decoding. Discrete decoding in an online
distances of ~10 mm (in human implantations) paradigm (Milekovic et al. 2012) has also been
and electrode diameter of ~4 mm (with 2.3 mm reported to establish the viability of ECoG signals
exposed) is placed on the cortex to record ECoG for BMI applications.
Electroencephalogram imaging technique that uses BOLD responses

EEG signals are the least invasive in the hierar- to infer brain activity is functional magnetic
chy of neuroelectrophysiological recordings. Dif- resonance imaging (fMRI) (Wolpaw and
ferential electrodes placed on the surface of the Wolpaw 2012). The oxy-Hb is nonmagnetic,
scalp record the potential difference generated by while the deoxy-Hb is slightly magnetic causing
neuronal sources in and around ~10 cm. Position- a perturbation in the incident magnetic field,
ing the electrodes is often referred to as the elec- thereby changing the strength of the magnetic
trode montage and varies between 21, 74, and field. The proportional changes in the magnetic
142 electrodes located at different scalp land- field due to the presence of deoxy-Hb can then be
marks. Events that elicit neural activity would used in estimating the BOLD response. Since
generally be detected as event-related potentials hemodynamic changes are slow, these imaging
(ERPs) in EEG signals. The averaged ERP com- techniques have temporal resolutions on the
ponent could be exogenous, meaning it is order of seconds.
recorded within 150 ms of the event, or endoge-
nous, a more sustained component. P300 signals
are ERPs occurring at an average latency of Decoding Systems
300 ms from stimulus event but can vary in
latency between 250 and 750 ms and have been Neurons of the motor cortex can encode kinetic
a subject of interest in BCI applications (Farwell variables such as force and torque (Fagg
and Donchin 1988; Li et al. 2010; McFarland et al. 2009; Cabel et al. 2001; Cheney and Fetz
et al. 1997, 2010; Nijboer et al. 2008b). EEGs 1980; Sergio et al. 2005; Smith et al. 1975; Taira
are easily affected by non-brain physiological et al. 1996) or kinematic variables such as posi-
signals such as EMG, EOG, and ECG, which tion, velocity, and acceleration (Ashe and
needs to be detected and eliminated for reliable Georgopoulos 1994; Fu et al. (1993, 1995)).
BMI control. Some neurons encode combinations of these
variables (Paninski et al. 2004; Wang
Metabolic Signals et al. 2007; Lawhern et al. 2010). Systems that
Apart from the neuroelectrophysiological sig- are used to decipher the encoded values or
nals, researchers are also investigating meta- parameters are termed decoders. However,
bolic signals for their suitability in BMI/BCI decoders can map neural data to an output with-
applications. Sitaram et al. (2009) have used out the need to learn any underlying encoding
functional near-infrared spectroscopy (fNIRS) model. Decoders can be static, adaptive, or
signals for a BCI discrete classification system. stochastic. Linear static decoders, such as the
The fNIRS technique utilizes blood-oxygen- Wiener filters, are widely used due to their
level-dependent (BOLD) responses to detect simplicity.
brain activity. As the neural activity increases,
the total hemoglobin and the oxyhemoglobin Static Decoders
(oxy-Hb) concentrations in the blood increase, One of the earliest decoding algorithm in
while the deoxyhemoglobin (deoxy-Hb) con- systems neuroscience is the so-called “popula-
centration decreases, in the vicinity of the active tion vector” algorithm that has been used to
brain region. Since oxy-Hb and deoxy-Hb relate ensemble neural firings to movement
absorb light wavelengths differently, changes parameters, typically direction and velocity.
in their concentration can be measured using Georgopoulos et al. (1986) observed that neu-
optical diodes serving as light transmitters and rons tend to fire the most when the actual move-
receivers. The fNIRS has spatial and temporal ment vector is aligned with a direction termed
resolutions in the order of centimeters and sec- the “preferred direction” and follow a cosine
onds, respectively. The most widely studied tuning curve in their rate of discharge. The firing
rate (z(D)) is related to the movement direction Cortical Motor Prosthesis, Table 2 Kalman filter
as given in Eq. 1: algorithm
xk ¼ Axk1 State prediction
zðDÞ ¼ a þ g cos yPD (1) Sk ¼ ASk1 AT þ R Covariance prediction
1 Kalman gain
K k ¼ Sk CT CSk CT þ Qk
where yPD is the angle between the preferred xk ¼ xk þ K k ðzk Cxk Þ State update
direction P of the neuron and the actual move- Sk ¼ ðI K k CÞSk Covariance update C
ment direction D. The movement direction can
then be decoded by the weighted sum of the
neural firing rate from an ensemble of
N neurons, as given by Eq. 2:
Adaptive Decoders
Decoders capable of adapting to the changing
X
N
V ðDÞ ¼ fi ðDÞ (2) input signal statistics are referred to as adaptive
i¼1 decoders. The neural activity matrix is often
noisy and the intended motor output can be esti-
where fi(D) = (zi(D) ai)Pi. mated using optimal linear filters such as the
Note that each neuron’s firing history over Kalman filter (which adapts the filter gain
multiple time points is not explicitly included in according to the changing covariance estimate
the population decoding. of the neural signal). BMIs use a discretized,
Wiener filters take into account these time- time-invariant (constant parameters) version of
delayed neural activities and map them to the filter. Kalman filters are linear Gaussian-
a control variable, either kinetic or kinematic structured filters and can be easily understood in
(see Eq. 3): a state-space framework. It is intended to esti-
mate the state vector, xk, at each time step taking
y^ ¼ Zb (3) into account the observed firing rate zk. In BMIs,
the state vector could be the kinematics of
a prosthetic limb, xk = [position, velocity,
where the neural activity matrix Z (columns
acceleration]T. The state transition equation for
corresponding to different time delays of multiple
xk is a linear function given by Eq. 5:
neurons and rows corresponding to different time
points) is mapped to the control variable y^ via the
system coefficients b (a column vector where xk ¼ Axk1 þ ϵk (5)
each component corresponds to a different time
delay). Note that the Weiner filter does not begin where A is the state transition matrix and
with an explicit encoding model like the popula- ϵk N ð0, RÞ models the transition uncertainty.
tion vector algorithm. The Weiner filter has the Similarly, the neural firing rates are again linearly
advantage that a closed-form solution for the beta related to the state vector:
coefficients exists that is optimal by minimizing
the error between the predicted and training zk ¼ Cxk þ dk (6)
control variable. With limited training data, reg-
ularization terms can be added to the error equa- The firing rate noise is described by
tion such that large coefficients are penalized dk N ð0, QÞ. The algorithm for Kalman filter-
resulting in the ridge-regression solution for b. ing can be described as a two-step process where
Solving for b in a ridge-regression framework is the state transition and its covariance are
shown in Eq. 4, with a ridge parameter l: predicted in the first step and subsequently
updated using the observed neural firing rates.
1
b ¼ Z T Z þ lI ZT y (4) Table 2 shows the pseudo-code version of the
Cortical Motor Prosthesis, Table 3 Kalman filter Here, the prior and the likelihood can be arbi-
matrices
trary distributions, although Kalman filter
! !1
XN X
N assumes them to be Gaussian. In order to decode
A¼ xk xTk1 xk1 xTk1 posterior distributions that are nonlinear and
k¼2 k¼2
! non-Gaussian, nonparametric methods such as
X
N X
N
1
R ¼ N1 xk xTk A xk1 xTk particle filters can be used. Particle filters
k¼2 k¼2
! !1 represent the posterior distribution using a set
X
N X
N
of random samples, or particles, (Np), and their
C¼ zk xTk xk xTk
k¼1 k¼1 associated state estimate and weights, {xi,t, Wi,t},
!
X
N X
N where i = 1, . . . ,Np. There are several flavors of
1
Q¼M zk zTk C xk zTk particle filtering available, and one interesting
k¼1 k¼1
variant, the auxiliary particle filter (APF) (Chen
2003), is discussed here.
Kalman filter. The matrices A, R, C, Q are chosen The objective is to determine an approxima-
such that the joint probability between the kine- tion of the posterior distribution:
matic variable XN and the firing rates ZN from
X
Np

N training sets, p(XN, ZN), is maximized (Black p^ðxt jZt Þ W i,t d xt xi,t (8)
and Donoghue 2007). The closed-form solution i¼1
yields the matrices as described in Table 3
(Wu et al. 2006). As the state transition pro- In order to estimate the weights W , the APF
gresses with time, it becomes nearly independent approach uses a two-step process which samples
of the initial state, and hence, a reasonably chosen weights from a proposal distribution (qa) and
random initial condition should suffice for updates the weight according to some likelihood
a satisfactory filtering. estimate. Since the actual posterior distribution is
Recently, a modified version of the Kalman unknown and cannot be sampled from, the pro-
filter, referred to as recalibrated feedback posal distribution is used. This could be a simple
intention-trained Kalman filter (ReFIT-KF), was distribution with Gaussian parameters or a kernel
reported that allows users to maintain zero velocity that reflects neural tuning curves from the ensem-
when the target is reached (Gilja et al. 2012). ble. An algorithmic representation of the APF is
According to this algorithm, in a 2D cursor reaching given in Table 4. These steps are repeated for every
experiment that decodes both position and velocity, time step. Particle filters have been used in
the uncertainty in the estimate of the position pk was decoding the hippocampal place cells (Brown
“causally intervened” with a more certain position et al. 2001; Mountney et al. 2009) and retinal
p^k such that the intended movement is always neurons (Brown et al. 2003), and a real-time imple-
directed towards the minimal path to the target. mentation is discussed in Mountney et al. (2011).
Nonparametric Decoders
Under the Bayesian framework, the decoding Decoder Initialization
paradigm can be viewed as estimating the maxi-
mum a posteriori value of p(xt|Zt), i.e., the move- The process of mapping a priori neural signals to
ment variable conditioned on the firing rate any control variables, thereby determining the
(Wu et al. 2006). The posterior probability is decoder filter coefficients, is termed decoder ini-
related to the prior and the likelihood (under tialization. The decoder coefficient matrix to
Markov assumptions) as transform neural firing rates into control
ð signals can be estimated by concurrently record-
pðxt jZt Þ / pðzt jxt Þ pðxt jxt1 Þpðxt1 jZt1 Þdxt1 ing the neural data and the actual kinematic/
kinetic variable, as the subject executes the task.
(7) This popular mapping approach is often referred
Cortical Motor Prosthesis, Table 4 Particle filter where the subject can imagine a particular
algorithm movement or observe a relevant task and the

xî,t qa xt jxi, t1 , zt Sample from proposal neural signals are recorded and mapped to the
distribution observed movement. In a BMI experiment
Wi,0 = 1/Np, i = 1, . . . , Np Initialize the weights reported in Balasubramanian et al. (2013), neu-
pðzt j^
xi,t Þpðxî,t jxî,t1 Þ Calculate the 1st
W ^ i,t ¼ W i, t1 ral activity collected from the primary motor
qa ðxî,t jxî,t1 , zt Þ stage weights
^
W i, t
cortex, as a chronically amputated macaque C
W i, t ¼ N p Normalize weights
X was observing a robot moving in a desired tra-
W ^ n,t jectory, was used to initialize a decoder.
n¼1
Clanton et al. (2013) have used observation-
x^t qb ðx^t j^
xt1 , zt Þ Resample from the
new proposal based decoder initialization in controlling a
distribution multiple DOF robot involving translation,
pðzt j^
xi, t Þpðxî,t jxî, t1 Þ Calculate the 2nd rotation, and grasp aperture.
W i,t ¼ W i,t1 q x^ x^ , z
b ð i, t j i,t1 t Þ stage weights
W i,t Normalize weights
W i,t ¼ Np
aux
Unsupervised Decoder Initialization
X
W n,t The unsupervised initialization approach ignores
n¼1
the unobservable kinematic variable and shifts
X
Np
Posterior distribution
the emphasis on the neural signal dynamics to
pðxt jZ t Þ W aux î,t
i,t d xt x
i¼1 determine the decoder filter coefficients. In one
unsupervised approach, decoders are assumed to
function as linear correlators and the coefficients
are sampled from a multivariate distribution with
to as biomimetic decoding. When the subjects covariances matching the neural data. Such
have intact nervous systems and limbs, a decoder is expected to have a matching fre-
performing the kinematic task is trivial and the quency response to the neural dynamics. A
biomimetic decoding is preferred. However, the minimum-norm Mahalanobis distance solution
mapping becomes nontrivial when the kinematic was implemented in Badreldin et al. (2013) for
variables are not observable, such as in paralyzed initializing a velocity decoder. The coefficients
or amputee patients. This could be resolved by were chosen such that the Gaussian properties of
resorting to either (i) an observation-based ini- the output are preserved.
tialization or (ii) an unsupervised decoder that At a system level, the decoders and the
does not need the kinematic variable. Subsequent ensemble of neurons driving them can be
sections below explain briefly the different viewed as two coupled systems, and either
decoder initialization approaches. Once the or both can be adaptive. Adaptive decoders
decoder is initialized, it could remain static or (discussed in the previous section) can mini-
adaptive. mize control errors by adapting the filter coeffi-
cients. On the other hand, when a static decoder
Observation-Based Initialization is initialized using the unsupervised approach,
Several regions of the cortex have neurons adaptation occurs at the neuronal side of the
exhibiting congruent activity during active coupled system.
movement as well as passive observation of
the same task (Dushanova and Donoghue Operant Conditioning
2010; Tkach et al. 2008). Primary motor corti- A proven technique in learning to adapt to
cal activity profiles were first reported to be a decoder is operant conditioning, which relies
similar when recorded from macaques engaged on providing a reinforcement stimulus whenever
in a particular task and also while merely a desired behavior occurs (Seel 2012). As learn-
observing the same task (Tkach et al. 2008). ing occurs, the occurrence frequency of the
This idea could be used to initialize a decoder desired behavior increases. In BMI experiments
with nonhuman primates, desired behavior can be solution, mimicking regularities in biological
reinforced using juice reward or fruit bits. Posi- motion. For nonredundant robots (number of
tive reinforcements have been used by Fetz and DOFs 6), a unique inverse kinematic solution
colleagues in training monkeys to volitionally is achievable for a given end-effector definition.
modulate single-neuron firing activity (Fetz Redundant robots can have infinite solutions to
1969; Fetz and Finocchio 1971). Fetz and a given spatial coordination and orientation and
Finocchio (1975) also showed that volitional neu- hence need numerical controllers or constrained
ral modulations can be achieved with complete controllers such as null-space controllers (Nemec
dissociation from any peripheral muscle contrac- and Zlajpah 2000; Corke 2011). Finer motor con-
tions. This degree of flexibility of motor cortex trols such as grasping can be controlled in
(Wander et al. 2013) has been harnessed in devel- a reduced dimension synergistic space rather
oping BMIs further. Several experiments have than the joint space. A large proportion of the
successfully used operant conditioning as variance observed in natural grasp kinematics
a mode in BMI training. Recent experiments can be accounted for with few principal compo-
with a primate model of amputation have used nents (Ingram et al. 2008; Mason et al. 2001;
operant conditioning in training macaques to con- Santello 2002). For example, the hand has more
trol a multiple DOF robot (Balasubramanian than 20 DOFs (say, n) and any hand posture
et al. 2013). belongs to an Rn dimension representation. How-
ever, natural grasping behavior resides in a much
smaller representational space so that dimension-
Actuator Modalities and Controlled ality reduction techniques such as principal com-
Devices ponent analysis can be used. The principal
components (or eigengrasps (Ciocarlie
Experimental BMI studies have reported exten- et al. 2007)) of this high-dimension space form
sively on cortical control of virtual objects such a synergistic basis for the complex postures of the
as computer cursors in 2D or 3D space (Black hand. A single basis can then be used to represent
et al. 2003; Hatsopoulos et al. 2004; Santhanam grasp aperture and is independent of the number
et al. 2006; Serruya et al. 2002; Shpigelman of fingers. Similarly, other bases can represent
et al. 2008; Taylor et al. 2002). Tasks include pronation/supination, thumb abduction/adduction,
reaching predefined target locations either in etc. These basis sets offer a direct mapping
sequence or at random, tracking a randomly between a decoded control signal and a control
moving target, free-choice targets with varied synergy.
reward probability, etc. These experiments were
useful in devising a translational pathway to
neural motor prosthesis and rehabilitation. Feedback in a BMI
Robots, whole-arm manipulators, and virtual
avatars allow BMIs to operate many degrees of In BMIs, feedback about the state of the pros-
freedom in a 3D space (Carmena et al. 2003; thesis is essential to compensate for any move-
Lebedev et al. 2005). When controlling ment errors. Visual feedback is the predominant
a multiple DOF robot using BMIs, decoded sig- feedback modality that uses the natural sensory
nals can (a) operate directly on the joint space pathway. Additionally, proprioceptive, tactile
controlling the individual joints or (b) define the (Nicolelis 2003), and, to some extent, auditory
end-effector position and orientation, while feedback (Nijboer et al. 2008a; Schreuder
the controllers determine the instantaneous et al. 2010) are being investigated as promising
joint configuration. Smooth movement controls modalities for BMIs. Artificial feedback sys-
can be achieved with a minimum-jerk tems provide tactile and/or proprioceptive
(Flash and Hogan 1985; Shadmehr 2005) feedback in the form of mechanical stimulation
of peripheral nerves or electrical stimulation situations. Haptic feedback can be substituted by
of residual nerves (Hatsopoulos and Donoghue alternative feedback modalities such as
2009). Extended physiological proprioception vibrotactile or pneumatic pressure and delivered
(EPP), an approach which uses residual to a different region of the body (Kaczmarek
tactile and proprioceptive sensations as feed- et al. 1991). Vibrotactile feedback to the tongue
back pathways, is another important natural has been a successful substitution approach used
form of feedback to augment the visual system. by visually impaired persons in detecting obsta- C
Subjects with spinal cord injury (SCI) and with cles (Chebat et al. 2011). Several studies have
subsequent motor dysfunction may retain some used vibrotactile stimulation as substituted feed-
residual somatosensation (Eide et al. 1996; back delivered to the biceps of an intact arm, or
Finnerup et al. 2004; Grill et al. 2009), which residual arm in cases of amputees, to indicate
could be taken advantage of for providing EPP a 1D kinetic variable, such as the output of
feedback. a pressure sensor (Aniruddha et al. 2007). The
Pacinian mechanoreceptors (Mountcastle
Multisensory Feedback et al. 1972) are, in general, the target receptors
Current BMIs are exploring different ways to and determine the applied vibration frequency
effectively integrate information from multiple (~200 Hz). Parameters such as the frequency,
sensory modalities. It is well established that intensity, timbre, duration, and body location
sensory inputs alter motor decisions, when pro- can be modulated at the receptor level to deliver
vided as feedback. Experimental evidence has various forms of feedback information (Cincotti
suggested that augmenting feedback with more et al. 2007).
than one sensory modality provides usable Another direct substitution for force feedback
mutual information to the motor cortex is the use of a pneumatic cuff delivering radial
(Suminski et al. 2009). Alterations in movement pressure to the biceps in proportion to the grasp
trajectories were observed as early as 70 ms pressure exerted by the prosthesis (Patterson and
following a proprioceptive cue (Crago Katz 1992). In the BCI context, users with
et al. 1976). Visual feedback is relatively substituted feedback were able to generate rela-
slow and influences the motor commands tively more accurate control signals.
around ~150 ms. Macaques performing a target
tracking task in a 2D space were able to reach Bidirectional Brain-Machine Interfaces
the target faster when the visual feedback was Rich somatosensory feedback can be provided
augmented with proprioception. An exoskeleton directly to the brain via electrical stimulation of
was used by Suminski et al. (2010) to provide the somatosensory cortex. Intracortical microsti-
kinesthetic feedback, in addition to vision, mulation (ICMS) is a technique used to deliver
about cursor positions in the 2D plane. The small amounts of currents (~25–60 mA) to neu-
exoskeleton is one way of providing EPP to rons via electrodes or microwires (O’Doherty
BMI users. Under congruent feedback condi- et al. 2011). The delivery rate of stimuli to the
tions, BMI performance improved by 40 % in neurons is a parameter controlled in delivering
terms of the time to target and path length (the discriminable tactile feedback. O’Doherty
ratio between the actual path and the minimal et al. (2009) has reported using a stimulation
path to target). frequency of 30 and 60 Hz with pulse widths of
150–200 ms. Berg et al. (2013) have used ICMS to
Sensory Substitution deliver graded percepts of force stimuli exerted
In neuromotor prosthesis control, providing bio- on a prosthetic finger. Amplitude of four graded
logically plausible sensory feedback (primarily signals representing the different mechanical
haptic feedback) to the user is a challenging stimuli was modulated in proportion to time-
issue and substituted feedback is viable in certain varied force output of the prosthetic finger, and
equivalent currents were delivered. ICMS to S1 three-dimensional cursor control in a human

can be used for directional instruction, active tetraplegic (Wang et al. 2013). In contrast to
tactile sense, or even to represent somatotopic these invasive methods, EEG has been exten-
mapping of a body in 3D space (Lebedev and sively used in human subjects relying primarily
Nicolelis 2011). on voluntary modulation of slower frequency
oscillations below 40 Hz over somatosensory
and motor cortices. Despite the weaker informa-
BMI Applications in Human Patients tion content in EEG-based recordings, it has been
possible to control up to three dimensions
Although currently there do not exist commer- of cursor movement using voluntary modulation
cially available brain-machine interfaces for of slow sensorimotor rhythms (McFarland
human patient use, there have been attempts in et al. 2010).
research and preclinical settings to examine the
capabilities of such systems in tetraplegic
patients. Phil Kennedy and colleagues were the
first to demonstrate that severely disabled
Cross-References
“locked-in” patients (due to brainstem stroke
▶ Brain-Machine Interface: Overview
and ALS) could control switches and cursor
▶ Neural Decoding
movements using spikes and local field potentials
from implanted microwires encased in a glass
cone (Kennedy and Bakay 1998; Kennedy
et al. 2000, 2004). Building on this early work, References
an FDA clinical trial was initiated on four
patients with tetraplegia due to spinal cord injury, Andersen RA, Musallam S, Pesaran B (2004) Selecting
ALS, and brainstem stroke using the BrainGate the signals for a brain-machine interface. Curr Opin
Neurobiol 14(6):720–726
system. This system consisted of a silicon-based
Aniruddha C, Vikram A, Ander R, Soumyadipta A,
100-microelectrode array implanted in the pri- Nitish T (2007) A brain-computer interface with
mary motor cortex. It was shown in two of these vibrotactile biofeedback for haptic information.
patients that effective two-dimensional cursor J NeuroEng Rehab 4
Ashe J, Georgopoulos AP (1994) Movement parameters
control was possible using spikes from approxi-
and neural activity in motor cortex and area 5. Cereb
mately 27 neurons (Hochberg et al. 2006). More Cortex 4(6):590–600
recently, the second-generation BrainGate sys- Badreldin I, Southerland J, Mukta V, Eleryan A,
tem (BrainGate 2) has demonstrated that chronic Balasubramanian K, Fagg A, Hatsopoulos N, Oweiss
K (2013) Unsupervised decoder initialization for
tetraplegics could control three-dimensional
brain-machine interfaces using neural state space
reaching and one-dimensional grasping of a dynamics. In: Neural engineering (NER), 2013 inter-
multi-articulate robot using motor cortical spik- national IEEE/EMBS conference on, San Diego, Cal-
ing activity from the implanted silicon-based ifornia, USA, pp 997–1000
Balasubramanian K, Southerland J, Mukta V, Qian K,
electrode array (Hochberg et al. 2012). In addi-
Eleryan A, Fagg AH, Sluzky M, Oweiss K,
tion, another group has shown that an individual Hatsopoulos N (2013) Operant conditioning of
with tetraplegia could control up to seven a multiple degree-of-freedom brain-machine interface
independent degrees of freedom of a prosthetic in a primate model of amputation. In: Engineering in
medicine and biology society (EMBC), 2013 annual
limb including three-dimensional translation, international conference of the IEEE, Osaka, Japan,
three-dimensional orientation, and grasp pp 303–306
using the same implanted array technology Berg J, Dammann J, Tenore F, Tabot G, Boback J,
(Collinger et al. 2013). A somewhat less invasive Manfredi L, Peterson M, Katyal K, Johannes M,
Makhlin A, Wilcox R, Franklin R, Vogelstein R,
approach using field potential oscillations up to
Hatsopoulos N, Bensmaia S (2013) Behavioral dem-
200 Hz recorded from the surface of the cortex onstration of a somatosensory neuroprosthesis. Neural
via ECoG electrodes has demonstrated effective Syst Rehab Eng IEEE Trans 21(3):500–507
Black MJ, Bienenstock E, Donoghue JP, Serruya M, degree of freedom cortical control of a robotic arm.
Wu W, Gao Y (2003) Connecting brains with In: Brain-computer interface research. Springer, Ber-
machines: the neural control of 2D cursor movement. lin, pp 73–81
In: Neural engineering (NER), 2003 international Collinger JL, Wodlinger B, Downey JE, Wang W, Tyler-
IEEE/EMBS conference on, Capri Island, Italy, Kabara EC, Weber DJ, McMorland AJ, Velliste M,
pp 580–583 Boninger ML, Schwartz AB (2013) High-performance
Black MJ, Donoghue JP (2007) Probabilistically neuroprosthetic control by an individual with
modeling and decoding neural population activity in tetraplegia. The Lancet 381(9866):557–564
motor cortex. In: Guido Dornhege THDJM, del Millán Corke P (2011) Robotics, vision and control: fundamental
C
JR, M€uller K-R (eds) Toward Brain-Computer Interfac- algorithms in MATLAB, vol 73. Springer, Berlin
ing, MIT Press, Cambridge, MA, pp. 147–159 Crago PE, Houk JC, Hasan Z (1976) Regulatory
Brown EN, Nguyen DP, Frank LM, Wilson MA, Solo actions of human stretch reflex. J Neurophysiol
V (2001) An analysis of neural receptive field plastic- 39(5):925–935
ity by point process adaptive filtering. Proc Natl Acad Donoghue JP, Sanes JN, Hatsopoulos NG, Gaál G (1998)
Sci 98(21):12261–12266 Neural discharge and local field potential oscillations
Brown EN, Barbieri R, Eden UT, Frank LM (2003) Like- in primate motor cortex during voluntary movements.
lihood methods for neural spike train data analysis. In: J Neurophysiol 79(1):159–173
Feng J (ed.), Computational neuroscience: A compre- Dushanova J, Donoghue J (2010) Neurons in primary
hensive approach. Chapman & Hall/CRC London, PP motor cortex engaged during action observation. Eur
253–286 J Neurosci 31(2):386–398
Brunner P, Ritaccio AL, Lynch TM, Emrich JF, Eide PK, Jørum E, Stenehjem AE (1996) Somatosensory
Wilson JA, Williams JC, Aarnoutse EJ, Ramsey NF, findings in patients with spinal cord injury and central
Leuthardt EC, Bischof H, Schalk G (2009) dysaesthesia pain. J Neurol Neurosurg Psychiatry
A practical procedure for real-time functional 60(4):411–415
mapping of eloquent cortex using electrocorti- Fagg A, Ojakangas G, Miller L, Hatsopoulos N (2009)
cographic signals in humans. Epilepsy Behav Kinetic trajectory decoding using motor cortical
15(3):278–286 ensembles. Neural Syst Rehab Eng IEEE Trans
Cabel DW, Cisek P, Scott SH (2001) Neural activity in 17(5):487–496
primary motor cortex related to mechanical loads Farwell LA, Donchin E (1988) Talking off the top of your
applied to the shoulder and elbow during a postural head: toward a mental prosthesis utilizing event-
task. J Neurophysiol 86(4):2102–2108 related brain potentials. Electroencephalogr Clin
Carmena JM, Lebedev MA, Crist RE, O’Doherty JE, Neurophysiol 70(6):510–523
Santucci DM, Dimitrov DF, Patil PG, Henriquez CS, Fetz EE (1969) Operant conditioning of cortical unit
Nicolelis MA (2003) Learning to control activity. Science 163(3870):955–958
a brain–machine interface for reaching and grasping Fetz EE, Finocchio DV (1971) Operant conditioning of
by primates. PLoS Biol 1(2):e42 specific patterns of neural and muscular activity. Sci-
Chebat D-R, Schneider FC, Kupers R, Ptito M (2011) ence 174(4007):431–435
Navigation with a sensory substitution device in Fetz E, Finocchio D (1975) Correlations between activity
congenitally blind individuals. Neuroreport of motor cortex cells and arm muscles during operantly
22(7):342–347 conditioned response patterns. Exp Brain Res
Chen Z (2003) Bayesian filtering: from kalman filters to 23(3):217–240
particle filters, and beyond. Statistics 182(1):1–69 Finnerup NB, Gyldensted C, Fuglsang-Frederiksen A,
Cheney PD, Fetz EE (1980) Functional classes of primate Bach FW, Jensen TS (2004) Sensory perception in
corticomotoneuronal cells and their relation to active complete spinal cord injury. Acta Neurol Scand
force. J Neurophysiol 44(4):773–791 109(3):194–199
Cincotti F, Kauhanen L, Aloise F, Palom€aki T, Flash T, Hogan N (1985) The coordination of arm move-
Caporusso N, Jyl€anki P, Mattia D, Babiloni F, ments: an experimentally confirmed mathematical
Vanacker G, Nuttin M, Marciani MG, del Millán JR model. J Neurosci 5(7):1688–1703
(2007) Vibrotactile feedback for brain-computer inter- Flint R, Wright Z, Slutzky M (2012) Control of a biomi-
face operation. Intell Neurosci 12 metic brain machine interface with local field poten-
Ciocarlie M, Goldfeder C, Allen P (2007) Dexterous tials: Performance and stability of a static decoder over
grasping via eigengrasps: a low-dimensional approach 200 days. In: Engineering in medicine and biology
to a high-complexity problem. In: Proceedings of the society (EMBC), 2012 annual international conference
robotics: science and systems 2007 manipulation of the IEEE, San Diego, California, USA,
workshop - sensing and adapting to the real world. pp 6719–6722
Robotics: science and systems conference, Atlanta, Flint RD, Wright ZA, Scheid MR, Slutzky MW
Georgia USA (2013) Long term, stable brain machine interface per-
Clanton ST, McMorland AJ, Zohny Z, Jeffries SM, Ras- formance using local field potentials and multiunit
mussen RG, Flesher SN, Velliste M (2013) Seven spikes. J Neural Eng 10(5):056005
Fu Q, Suarez J, Ebner T (1993) Neuronal specification of Kennedy P, Andreasen D, Ehirim P, King B, Kirby T,
direction and distance during reaching movements in Mao H, Moore M (2004) Using human extra-cortical
the superior precentral premotor area and primary local field potentials to control a switch. J Neural Eng
motor cortex of monkeys. J Neurophysiol 1(2):72
70(5):2097–2116 Lawhern V, Wu W, Hatsopoulos N, Paninski L (2010)
Fu Q, Flament D, Coltz J, Ebner T (1995) Temporal Population decoding of motor cortical activity using
encoding of movement kinematics in the discharge of a generalized linear model with hidden states.
primate primary motor and premotor neurons. J Neurosci Methods 189(2):267–280
J Neurophysiol 73(2):836–854 Lebedev MA, Nicolelis MA (2011) Toward a whole-body
Georgopoulos AP, Schwartz AB, Kettner RE (1986) Neu- neuroprosthetic, chapter-3. In: Jens Schouenborg MG,
ronal population coding of movement direction. Sci- Danielsen N (eds) Brain machine interfaces: implica-
ence 233(4771):1416–1419 tions for science, clinical practice and society,
Gilja V, Nuyujukian P, Chestek CA, Cunningham JP, vol 194, Progress in brain research. Elsevier, Amster-
Byron MY, Fan JM, Church-land MM, dam, pp 47–60
Kaufman MT, Kao JC, Ryu SI, Shenoy KV (2012) Lebedev MA, Carmena JM, O’Doherty JE,
A high-performance neural prosthesis enabled by con- Zacksenhouse M, Henriquez CS, Principe JC, Nicolelis
trol algorithm design. Nature Neurosci 15:1752–1757 MA (2005) Cortical ensemble adaptation to represent
Grill WM, Norman SE, Bellamkonda RV (2009) velocity of an artificial actuator controlled by a brain-
Implanted neural interfaces: biochallenges and machine interface. J Neurosci 25(19):4681–4693
engineered solutions. Annu Rev Biomed Eng 11:1–24 Li Y, Long J, Yu T, Yu Z, Wang C, Zhang H, Guan C
Hatsopoulos NG, Donoghue JP (2009) The (2010) An EEG-based BCI system for 2-d cursor con-
science of neural interface systems. Ann Rev Neurosci trol by combining mu/beta rhythm and p300 potential.
32:249 Biomed Eng IEEE Trans 57(10):2495–2505
Hatsopoulos N, Joshi J, O’Leary JG (2004) Decoding Mason C, Gomez J, Ebner T (2001) Hand synergies during
continuous and discrete motor behaviors using motor reach-to-grasp. J Neurophysiol 86(6):2896–2910
and premotor cortical ensembles. J Neurophysiol McFarland DJ, Lefkowicz AT, Wolpaw JR
92(2):1165–1174 (1997) Design and operation of an EEG-based brain-
Hochberg LR, Serruya MD, Friehs GM, Mukand JA, computer interface with digital signal processing tech-
Saleh M, Caplan AH, Branner A, Chen D, Penn RD, nology. Behav Res Methods Instrum Comput
Donoghue JP (2006) Neuronal ensemble control of 29(3):337–345
prosthetic devices by a human with tetraplegia. Nature McFarland DJ, Sarnacki WA, Wolpaw JR (2010) Electro-
442(7099):164–171 encephalographic (EEG) control of three-dimensional
Hochberg LR, Bacher D, Jarosiewicz B, Masse NY, movement. J Neural Eng 7(3):036007
Simeral JD, Vogel J, Haddadin S, Liu J, Cash SS, Milekovic T, Fischer J, Pistohl T, Ruescher J, Schulze-
van der Smagt P et al (2012) Reach and grasp by Bonhage A, Aertsen A, Rickert J, Ball T, Mehring C
people with tetraplegia using a neurally controlled (2012) An online brain–machine interface using
robotic arm. Nature 485(7398):372–375 decoding of movement direction from the human elec-
Hwang EJ, Andersen RA (2013) The utility of trocorticogram. J Neural Eng 9(4):046003
multichannel local field potentials for brain-machine Mountcastle VB, LaMotte RH, Carli G (1972) Detection
interfaces. J Neural Eng 10(4):046005 thresholds for stimuli in humans and monkeys: com-
Ingram JN, Körding KP, Howard IS, Wolpert DM parison with threshold events in mechanoreceptive
(2008) The statistics of natural hand movements. Exp afferent nerve fibers innervating the monkey hand.
Brain Res 188(2):223–236 J Neurophysiol 35:122–136
Kaczmarek K, Webster J, Bach-y Rita P, Tompkins WJ Mountney J, Obeid I, Silage D (2011). Modular particle
(1991) Electrotactile and vibrotactile displays for filtering FPGA hardware architecture for brain
sensory substitution systems. Biom Eng IEEE Trans machine interfaces. In: Engineering in medicine and
38(1):1–16 biology society, EMBC, 2011 annual international
Katzner S, Nauhaus I, Benucci A, Bonin V, Ringach DL, conference of the IEEE, pp 4617–4620
Carandini M (2009) Local origin of field potentials in Mountney J, Sobel M, Obeid I (2009). Bayesian
visual cortex. Neuron 61(1):35–41 auxiliary particle filters for estimating neural tuning
Kennedy PR, Bakay RA (1998) Restoration of neural parameters. In: Engineering in medicine and biology
output from a paralyzed patient by a direct brain con- society. EMBC 2009. Annual international conference
nection. Neuroreport 9(8):1707–1711 of the IEEE, Minneapolis, Minnesota, USA,
Kennedy PR, Bakay RA, Moore MM, Adams K, pp 5705–5708
Goldwaithe J (2000) Direct control of a computer Nemec B, Zlajpah L (2000) Null space velocity control
from the human central nervous system. Rehab Eng with dynamically consistent pseudo-inverse. Robotica
IEEE Trans 8(2):198–202 18(5):513–518
Nicolelis MA (2003) Brain-machine interfaces to restore Shpigelman L, Lalazar H, Vaadia E (2008) Kernel-arma
motor function and probe neural circuits. Nat Rev for hand tracking and brainmachine interfacing during
Neurosci 4(5):417–422 3d motor control. In: Koller D, Schuurmans D, Bengio
Nijboer F, Furdea A, Gunst I, Mellinger J, McFarland DJ, Y, Bottou L (eds) Advances in neural information
Birbaumer N, K€ubler A (2008a) An auditory brain- processing systems. Curran Associates, Inc. Red
computer interface (bci). J Neurosci Methods Hook, NY, 21:1489–1496
167(1):43–50 Sitaram R, Caria A, Birbaumer N (2009)
Nijboer F, Sellers E, Mellinger J, Jordan M, Matuz T, Hemodynamic brain-computer interfaces for commu-
Furdea A, Halder S, Mochty U, Krusienski D, nication and rehabilitation. Neural Netw 22(9):
C
Vaughan T, Wolpaw J, Birbaumer N, K€ ubler 1320–1328
A (2008b) A p300-based brain-computer interface for Slutzky MW, Jordan LR, Krieg T, Chen M, Mogul DJ,
people with amyotrophic lateral sclerosis. Clin Miller LE (2010) Optimal spacing of surface electrode
Neurophysiol 119(8):1909–1916 arrays for brain–machine interface applications.
O’Doherty JE, Lebedev M, Hanson TL, Fitzsimmons N, J Neural Eng 7(2):026004
Nicolelis MA (2009) A brain-machine interface Smith A, Hepp-Reymond MC, Wyss U (1975) Relation
instructed by direct intracortical microstimulation. of activity in precentral cortical neurons to force and
Front Integ Neurosci 3(20) rate of force change during isometric contractions of
O’Doherty JE, Lebedev MA, Ifft PJ, Zhuang KZ, finger muscles. Exp Brain Res 23(3):315–332
Shokur S, Bleuler H, Nicolelis MA (2011) Active tac- Suminski AJ, Tkach DC, Hatsopoulos NG (2009)
tile exploration using a brain-machine-brain interface. Exploiting multiple sensory modalities in brain-
Nature 479(7372):228–231 machine interfaces. Neural Netw 22(9):1224–1234
Paninski L, Fellows MR, Hatsopoulos NG, Donoghue JP Suminski AJ, Tkach DC, Fagg AH, Hatsopoulos NG
(2004) Spatiotemporal tuning of motor cortical neu- (2010) Incorporating feedback from multiple sensory
rons for hand position and velocity. J Neurophysiol modalities enhances brain-machine interface control.
91(1):515–532 J Neurosci 30(50):16777–16787
Patterson PE, Katz JA (1992) Design and evaluation of Taira M, Boline J, Smyrnis N, Georgopoulos AP,
a sensory feedback system that provides grasping pres- Ashe J (1996) On the relations between single cell
sure in a myoelectric hand. J Rehabil Res Dev 29(1):1–8 activity in the motor cortex and the direction and
Santello M (2002) Kinematic synergies for the control of magnitude of three-dimensional static isometric
hand shape. Arch Italiennes de Biologie force. Exp Brain Res 109(3):367–376
140(3):221–228 Taylor DM, Tillery SIH, Schwartz AB (2002) Direct cor-
Santhanam G, Ryu SI, Byron MY, Afshar A, Shenoy KV tical control of 3d neuro-prosthetic devices. Science
(2006) A high-performance brain–computer interface. 296(5574):1829–1832
Nature 442(7099):195–198 Tkach D, Reimer J, Hatsopoulos NG (2008) Observation-
Schalk G, Kubánek J, Miller K, Anderson N, Leuthardt E, based learning for brain–machine interfaces. Curr
Ojemann J, Limbrick D, Moran D, Gerhardt L, Opin Neurobiol 18(6):589–594
Wolpaw J (2007) Decoding two-dimensional move- Wander JD, Blakely T, Miller KJ, Weaver KE, Johnson
ment trajectories using electrocorticographic signals in LA, Olson JD, Fetz EE, Rao RPN, Ojemann JG
humans. J Neural Eng 4(3):264 (2013) Distributed cortical adaptation during learning
Schreuder M, Blankertz B, Tangermann M (2010) A new of a brain-computer interface task. Proc Natl Acad Sci
auditory multi-class brain-computer interface para- 110(26):10818–10823
digm: spatial hearing as an informative cue. PLoS Wang W, Chan SS, Heldman DA, Moran DW (2007)
ONE 5(4):e9813 Motor cortical representation of position and
Seel N (2012) Operant conditioning. In: Seel N velocity during reaching. J Neurophysiol 97(6):
(ed) Encyclopedia of the sciences of learning. 4258–4270
Springer, New York, p 2526 Wang W, Collinger JL, Degenhart AD, Tyler-Kabara EC,
Sergio LE, Hamel-Pâquet C, Kalaska JF (2005) Motor Schwartz AB, Moran DW, Weber DJ, Wodlinger B,
cortex neural correlates of output kinematics and Vinjamuri RK, Ashmore RC et al (2013) An
kinetics during isometric-force and arm-reaching electrocorticographic brain interface in an individual
tasks. J Neurophysiol 94(4):2353–2378 with tetraplegia. PLoS One 8(2):e55344
Serruya MD, Hatsopoulos NG, Paninski L, Fellows MR, Wolpaw J, Wolpaw EW (2012) Brain-computer inter-
Donoghue JP (2002) Brain-machine interface: instant faces: principles and practice. Oxford University
neural control of a movement signal. Nature Press, New York
416(6877):141–142 Wu W, Gao Y, Bienenstock E, Donoghue JP, Black MJ
Shadmehr R (2005) The computational neurobiology of (2006) Bayesian population decoding of motor cortical
reaching and pointing: a foundation for motor learn- activity using a kalman filter. Neural Comput
ing. MIT Press, Cambridge, MA 18(1):80–118
C 906 Cortical Representation of Speech Sounds
waves, which occur at a frequency of 7–14 Hz. As

Cortical Representation of Speech sleep deepens, waves with slower frequencies
Sounds (0.3–4 Hz), including delta waves and slow oscil-
lations, appear and progressively dominate the
▶ Neural Coding of Speech Sounds EEG. During paradoxical sleep, also called
rapid eye movement (REM) sleep, EEG activities
resemble those of wakefulness.
The cellular mechanisms of slow-wave sleep
Corticothalamic Feedback: oscillations have been investigated since the first
Large-Scale Synchrony extracellular and intracellular recordings in
mammals. The major brain regions which have
Alain Destexhe been identified are the thalamus and cerebral
Unit of Neuroscience Information and cortex, which are densely linked by means of
Complexity (UNIC), Centre National de la reciprocal projections. The activities of thalamic
Recherche Scientifique (CNRS), Gif-sur-Yvette, and cortical neurons during sleep have been
France largely documented by electrophysiological stud-
ies. The cellular mechanisms underlying these
oscillations depend on many factors, such as the
Definition connectivity and intrinsic properties of the differ-
ent types of thalamic and cortical neurons. To
The thalamus is a pacemaker for several types of determine these cellular mechanisms, it is neces-
brain oscillations, such as sleep spindles. Although sary to use computational models, which are
these oscillations are generated within thalamic based on experimental data, and suggest mecha-
circuits, their triggering and large-scale synchrony nisms and, if possible, predictions to test them.
depends on cerebral cortex. In particular, it This type of interaction has been quite successful
depends on the descending excitatory action of in the (still ongoing) exploration of the mecha-
the cortex on the thalamus. Computational models nisms of sleep oscillations.
show that the inhibitory-dominant character of this The cellular mechanisms leading to the genesis
“corticothalamic feedback” can explain the large- of spindle oscillations by thalamic circuits were
scale synchrony of oscillations in cerebral cortex. reviewed in another entry (“▶ Spindle Oscilla-
tions: Models”). Spindle oscillations are generated
by interacting thalamic neurons, as illustrated in
Detailed Description Fig. 1. Thalamocortical (TC) and thalamic reticu-
lar (RE) cells, endowed with their intrinsic burst-
Introduction ing properties and connected via glutamate
The fact that the brain produces oscillations was (AMPA) and GABAA receptors, can generate
discovered as soon as electroencephalogram spindle oscillations reproducing the typical fea-
(EEG) recordings of brain activity were avail- tures observed in intracellular recordings
able. The first human EEG recordings reported (reviewed in Destexhe and Sejnowski 2001, 2003).
oscillations, which type, frequency, and ampli-
tude highly depend on behavioral state (reviewed Propagation or Large-Scale Synchrony?
in Steriade 2003). In an alert, awake subject, the The initiation of spindle oscillations and their
EEG is dominated by low-amplitude fast activity distribution in large-scale circuits were investi-
(“desynchronized EEG”) with high-frequency gated using multiple recordings in vivo
oscillations (beta, gamma), whereas during (Contreras et al. 1996) and in vitro (Kim
slow-wave sleep, the EEG shifts to large- et al. 1995). Spindle oscillations in vitro showed
amplitude, slow oscillations. The early stage of traveling wave patterns, with the oscillation typ-
slow-wave sleep is associated with EEG spindle ically starting on one side of the slice and
Corticothalamic Feedback: Large-Scale Synchrony 907 C
a b
GABAA
40 mV
RE1 RE2
AMPA RE
GABAA C
+
TC1 TC2 GABAB TC
0.5 sec
Corticothalamic Feedback: Large-Scale Synchrony, with glutamatergic (AMPA) and GABAergic receptors
Fig. 1 Computational model of spindle waves generated (GABAA and GABAB). (b) Activity of two model neurons
by TC-RE interactions. (a) Scheme of a simple circuit during simulated spindle waves (Modified from Destexhe
consisting of two TC and two RE cells, interconnected et al. 1996)
In vivo experiments
Intact Decorticated
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Th8
200 µV
1s
Corticothalamic Feedback: Large-Scale Synchrony, (intact), the oscillations were synchronized over large
Fig. 2 The large-scale synchrony of spindle oscillations. distances (7 mm between Th1 and Th8). Right: after
Multiple extracellular field potential recordings removal of the cortex (decorticated), the large-scale syn-
(Th1–Th8) in the thalamus of cats in vivo are shown in chrony was largely abolished (Modified from Contreras
two cases. Left: in the intact thalamocortical system et al. 1996)
propagating to the other side, at a constant prop- and RE layers, consistent with anatomical data.
agation velocity (Kim et al. 1995). Traveling Under these conditions, the models generated
spindle waves were simulated by computational traveling waves consistent with in vitro data.
models of networks of interconnected TC and RE However, in contrast to thalamic slices, the
cells (one-dimensional extensions of the circuit intact thalamocortical system in vivo showed
shown in Fig. 1), in two independent modeling that spindle oscillations are remarkably synchro-
studies (Destexhe et al. 1996; Golomb nized over extended thalamic and cortical
et al. 1996). These models were similar in spirit regions, with little evidence for propagation
to the circuit shown in Fig. 1a but assumed that (Fig. 2, left). This large-scale synchrony was lost
there was a topographic connectivity between TC when the cortex was removed (Fig. 2, right).
C 908 Corticothalamic Feedback: Large-Scale Synchrony
Computational models
Thalamus with cortex Isolated thalamic network
Th1
Th2
Th3
Th4
Th5
Th6
Th7
Th8 50 mV
10 s
Corticothalamic Feedback: Large-Scale Synchrony, was present (left) and local synchrony with traveling wave
Fig. 3 Modeling the large-scale synchrony of spindle activity in the isolated thalamus (right) (Modified from
oscillations. Computational models of thalamocortical Destexhe et al. 1998)
networks displaying large-scale synchrony when cortex
In this case, propagating oscillations were seen, et al. 1998). The principal prediction of this
suggesting that although the oscillation is gener- model was that, in order to generate large-scale
ated by the thalamus, its large-scale synchrony coherent oscillations, one has to take into account
depends on the cortex (Contreras et al. 1996). the powerful action of cortex on thalamus (via
corticothalamic fibers) and, most importantly, the
Thalamocortical Networks cortex had to recruit the thalamus primarily
To investigate possible mechanisms underlying through the RE nucleus. Because of the powerful
large-scale synchrony, a thalamocortical network inhibitory action of RE cells, the action of
model was developed by combining a previous corticothalamic input is “inhibitory dominant”
model of thalamic slices with a model of deep on TC cells, which property is essential to main-
cortical layers (see details in Destexhe tain large-scale synchrony by synchronizing the
Corticothalamic Feedback: Large-Scale Synchrony 909 C
rebound bursts of TC cells (Destexhe et al. 1998). In conclusion, in this entry, we have reviewed
This property of inhibitory dominance is gener- evidence that the mutual interactions between
ally observed intracellularly in thalamic neurons cortex and thalamus are a very powerful mecha-
when stimulating the cortex. In these conditions, nism to synchronize oscillations over multiple
the same model was capable of generating large- brain areas. This mechanism was revealed by
scale synchrony in the presence of the cortex and performing recordings in thalamus and cortex,
traveling wave activity in the isolated thalamus interconnected or isolated from each other, as C
(Fig. 3). Consistent with these models, propagat- well as by using computational models. The
ing activity has indeed been observed in the thal- models revealed that a critical ingredient is that
amus of decorticated cats in vivo (Contreras the cortex recruits the thalamus essentially
et al. 1996). through inhibition. This mechanism can explain
The “inhibitory-dominant” character of all experimental measurements, and it can also
corticothalamic feedback on thalamic relay cells explain the genesis of a form of hypersyn-
has multiple important consequences. Not only it chronized pathological oscillations by the
can explain the large-scale synchrony of spindle thalamocortical system.
oscillation, but it can also explain the genesis of
some pathological situations such as absence epi-
leptic seizures (Destexhe et al. 1998), as detailed Cross-References
in another entry (“▶ Slow Oscillations and Epi-
lepsy: Network Models”). As a result of inhibi- ▶ Slow Oscillations and Epilepsy: Network
tory dominance, a too strong corticothalamic Models
feedback can over-activate thalamic GABAB ▶ Spindle Oscillations: Models
receptors and entrain the physiologically intact
thalamus into hypersynchronous rhythms at
~3 Hz. This scheme may explain the genesis of References
absence seizures, which are hypersynchronous
~3 Hz rhythms that appear suddenly in the Bal T, Debay D, Destexhe A (2000) Cortical feedback
thalamocortical system. These seizures can be controls the frequency and synchrony of oscillations in
the visual thalamus. J Neurosci 20:7478–7488
provoked experimentally by increasing cortical Blumenfeld H, McCormick DA (2000) Corticothalamic
excitability while keeping a physiologically inputs control the pattern of activity generated in
intact thalamus (reviewed in Gloor and Fariello thalamocortical networks. J Neurosci 20:5153–5162
1988). The thalamocortical model accounts for Contreras D, Destexhe A, Sejnowski TJ, Steriade
M (1996) Control of spatiotemporal coherence of
those experiments and can simulate seizures a thalamic oscillation by corticothalamic feedback.
based on inhibitory-dominant corticothalamic Science 274:771–774
feedback (Destexhe et al. 1998). This model Destexhe A, Sejnowski TJ (2001) Thalamocortical assem-
directly predicted that manipulating blies. Oxford University Press, Oxford
Destexhe A, Sejnowski TJ (2003) Interactions between
corticothalamic feedback should entrain intact membrane conductances underlying thalamocortical
thalamic circuits to generate hypersynchronous slow-wave oscillations. Physiol Rev 83:1401–1453
rhythms at ~3 Hz, a prediction which has been Destexhe A, Bal T, McCormick DA, Sejnowski TJ
verified by two independent studies (Blumenfeld (1996) Ionic mechanisms underlying synchronized
oscillations and propagating waves in a model of ferret
and McCormick 2000; Bal et al. 2000). A similar thalamic slices. J Neurophysiol 76:2049–2070
mechanism, with a different balance between Destexhe A, Contreras D, Steriade M (1998) Mechanisms
GABAA and GABAB receptors, can also generate underlying the synchronizing action of
faster hypersynchronous rhythms (around corticothalamic feedback through inhibition of tha-
lamic relay cells. J Neurophysiol 79:999–1016
5–10 Hz), as observed in rat or mouse experimen- Gloor P, Fariello RG (1988) Generalized epilepsy: some
tal models of absence seizures. of its cellular mechanisms differ from those of focal
epilepsy. Trends Neurosci 11:63–68
C 910 Coupled Oscillations in Neural Control of Breathing
Golomb D, Wang XJ, Rinzel J (1996) Propagation of

spindle waves in a thalamic slice model.
Kim U, Bal T, McCormick DA (1995) Spindle waves are
propagating synchronized oscillations in the ferret
LGNd in vitro. J Neurophysiol 74:1301–1323
Steriade M (2003) Neuronal substrates of sleep and epi-
lepsy. Cambridge University Press, Cambridge, UK
Further Reading
Steriade M, Jones EG, McCormick DA (eds) (1997)
Thalamus, vol 2. Elsevier, Amsterdam/Holland
Coupled Oscillations in Neural Control of Breathing,

Fig. 1 Parasagittal view of brainstem at the level of
nucleus ambiguus. Abbreviations: 5n trigeminal nerve, 7
Coupled Oscillations in Neural facial nucleus, 7n facial nerve, AmbC compact part of
Control of Breathing nucleus ambiguus, BötC bötzinger complex, cVRG caudal
division of ventral respiratory group, KF kölliker-fuse
nucleus, Mo5 motor nucleus of the trigeminal nerve, PB
Yaroslav Molkov parabrachial nucleus, pFRG parafacial respiratory group,
Department of Mathematical Sciences, Indiana Pn basilar pontine nuclei, pre-BötC pre-bötzinger com-
University – Purdue University Indianapolis, plex, RTN retrotrapezoid nucleus, rVRG rostral division of
Indianapolis, IN, USA ventral respiratory group, scp superior cerebellar pedun-
cle, SO superior olive, sol solitary tract, sp5 spinal trigem-
inal tract, VRC ventral respiratory column of the medulla,
VRG ventral respiratory group (Adapted from Alheid and
Synonyms McCrimmon (2008) with permission)
Dual-generator concept A distinct site of neural oscillations involved

with the respiratory function, located more rostral
to the pre-BötC, was identified in vitro, in the
Definition isolated brainstem-spinal cord preparation. The
source of these oscillations, which was termed
The respiratory rhythm in mammals is generated parafacial respiratory group (pFRG), was found
by a respiratory central pattern generator (CPG) to reside within the retrotrapezoid nucleus (RTN)
in the brainstem and encompasses both pons and or partially overlaps with it (Onimaru and
medulla (see Fig. 1; Alheid and McCrimmon Homma 2003; Janczewski and Feldman 2006).
2008; Smith et al. 2013). The pre-Bötzinger com- The pFRG oscillations emerge with increased met-
plex (pre-BötC), located in the ventrolateral abolic demands (e.g., hypercapnia) and via abdom-
region of the medulla, is the putative kernel of inal nerve (AbN) drive abdominal muscles
rhythmic inspiratory activity. The pre-BötC enabling forced lung deflation (active expiration).
interacts with the expiratory neurons of Bötzinger This rhythmic activity is coupled to the pre-BötC
complex (BötC) to generate primary respiratory oscillations so that generated bursts represent
oscillations. These oscillations are defined by late-expiratory (late-E) or sometimes biphasic
both the intrinsic biophysical properties of neu- (containing pre-inspiratory, pre-I, and post-inspira-
rons involved and the network interactions within tory, post-I, components) discharges (Janczewski
and between the pre-BötC and BötC. Through the and Feldman 2006; Abdala et al. 2009).
premotor neurons of the rostral ventral respira- Several competing concepts concerning the
tory group (rVRG), these oscillations produce appearance of RTN/pFRG oscillations and their
rhythmic drive to phrenic motoneurons that physiological function have been suggested
through the phrenic nerve (PN) control move- (Janczewski et al. 2002; Onimaru and Homma
ments of the diaphragm and lung inflation. 2003; Abdala et al. 2009; Molkov et al. 2010).
Coupled Oscillations in Neural Control of Breathing 911 C
Coupled Oscillations in Neural Control of Breathing, inspiratory phases of respiratory cycles (some highlighted
Fig. 2 Quantal acceleration of late-E discharges in the in by vertical pink bars), and interburst intervals correspond
situ brainstem-spinal cord preparation. The phrenic nerve to expiratory phases. Late-E discharges in AbN occur at
(PN, red) and abdominal nerve (AbN, black) activities are the end of expiration, just prior to inspiration (highlighted
shown by pairs of traces: raw recording (lower trace) and by gray bars) (Adapted from Molkov et al. (2010) with
integrated activity (upper trace). PN bursts define permission)
However, there are still heated debates on the been that opioids reduce the excitability of pre-
exact physiological role of RTN/pFRG oscilla- BötC neurons either directly or indirectly (via the
tions, the specific conditions for their emergence, suppression of excitatory synaptic transmission
and the nature and mechanisms of the interac- within the pre-BötC).
tions between the BötC/pre-BötC and RTN/ Similarly, quantal acceleration of late-E
pFRG oscillators. abdominal activity was observed with develop-
ment of hypercapnia (elevation of CO2) in the in
situ arterially perfused rat brainstem-spinal cord
Detailed Description preparation (Fig. 2, see Abdala et al. 2009;
Molkov et al. 2010) and in vivo (Iizuka and
From mathematical viewpoint, the most striking Fregosi 2007). Experimental studies in the per-
experimental observations that evidence the exis- fused in situ preparation have shown that under
tence of two distinct oscillators in neural control baseline normocapnic metabolic conditions (5 %
of breathing are the so-called quantal slowing of CO2), the abdominal nerve motor output (AbN)
phrenic rhythm (Janczewski and Feldman 2006; typically exhibits a low-amplitude expiratory
Mellen et al. 2003) and quantal acceleration of (post-I) activity. Switching to hypercapnic
abdominal discharges (Abdala et al. 2009; (7–10 % CO2) conditions evoked large
Molkov et al. 2010). amplitude late-E (also called pre-I) bursts in the
The quantal slowing of the pre-BötC and/or AbN. These late-E discharges emerge in AbN at
PN oscillations has been demonstrated with 7 % CO2 followed by a progressive increase in
administration of opioid agonists in the in vitro their frequency as the CO2 concentration in the
isolated brainstem-spinal cord preparation and perfusate is incremented to 10 %. Importantly,
in vivo (Mellen et al. 2003; Janczewski and although the frequency of late-E bursts increases
Feldman 2006). This phenomenon consists of with CO2, these bursts remained coupled (phase-
a stepwise reduction in the frequency of pre- locked) to the inspiratory bursts in the PN
BötC (and/or PN) oscillations resulting from (Fig. 2). With the development of hypercapnia,
deletion (missing) of a single or a series of inspi- the ratio of late-E burst frequency to the PN burst
ratory bursts in the output activity recorded. frequency showed a stepwise (quantal) increase
Although the exact pharmacological effect of from 1:5 and 1:4 (not shown) to 1:3, 1:2, and,
opioids is unknown, the general assumption has finally, 1:1 (at 10 % CO2). On returning CO2 to
C 912 Coupled Oscillations in Neural Control of Breathing
Coupled Oscillations in Neural Control of Breathing, neurons. Excitatory late-E neurons send their axons to
Fig. 3 The schematic of the model. Each population bulbospinal premotor expiratory (E) neurons in cVRG,
(shown as a sphere) consists of 50 single-compartment representing the source of AbN activity. The model
neurons described in the Hodgkin-Huxley style. The includes three sources of tonic excitatory drive: pons,
populations are denoted according to their firing pattern: RTN, and raphé (shown as green triangles) which project
post-I post-inspiratory, aug-E augmenting expiratory, pre- to multiple neural populations in the model. The late-E
I/I pre-inspiratory/inspiratory, early-I early inspiratory, population receives an additional external drive simulat-
ramp-I ramping inspiratory, late-E late expiratory. BötC/ ing the effect of hypercapnia. All model equations and
pre-BötC oscillator (Smith et al. 2007) controls RTN/ parameters can be found in Molkov et al. (2010) (Adapted
pFRG oscillator (the intrinsically rhythmic late-E popula- from Molkov et al. (2010) with permission)
tion) by inhibitory inputs from post-I and early-I(1)
the control levels, the ratio showed a stepwise (see Fig. 3). At some level of CO2, this excitation
reversal. overcomes the inhibition of late-E population in
Several computational models of interacting RTN/pFRG by BötC/pre-BötC circuits and initi-
neural oscillators related to generation of respi- ates late-E rhythmic activity which translates to
ratory activity with various levels of complexity AbN. The late-E bursts initially emerge at a very
have been proposed. The most detailed descrip- low frequency, but as CO2 level progresses
tion of possible interactions between the BötC/ (Fig. 4), the frequency of late-E discharges
pre-BötC and RTN/pFRG circuits (Fig. 3; quantally increases, until it reaches 1:1 ratio
Molkov et al. 2010) was developed as an exten- with the BötC/pre-BötC oscillations that define
sion of the respiratory CPG model proposed ear- PN output. In contrast, strong hypoxia or anoxia
lier by Smith et al. (2007) (see ▶ Computational can evoke the RTN/pFRG oscillations through
Models of Mammalian Respiratory CPG). a reduction of inhibition, specifically the post-I
According to that description, the BötC/pre- inhibition, and produce an effect similar to hyper-
BötC oscillator controls the emergence of late-E capnia by shifting the balance between inhibition
oscillations in RTN/pFRG and their frequency and excitation at the level of RTN/pFRG late-E
and phase by inhibiting RTN/pFRG oscillator neurons.
during both expiration (via post-I inhibition) Rubin et al. (2011) performed thorough qual-
and inspiration (via early-I inhibition). This inhi- itative analysis of the reduced model and
bition prevents late-E activity in RTN/pFRG and explained the regimes of quantal acceleration
AbN to occur under normal metabolic conditions. and quantal slowing in terms of synchronization
The excitability of the late-E population in of the BötC/pre-BötC and RTN/pFRG oscilla-
RTN/pFRG in the model is highly sensitive to tors. The dynamics of each oscillator can be
hypercapnia which causes their depolarization represented by a stable limit cycle in some
Coupled Oscillations in Neural Control of Breathing 913 C
Coupled Oscillations in Neural Control of Breathing, mimicking progressive hypercapnia. The lower panel
Fig. 4 Quantal acceleration of AbN activity in the model. shows the dependence of oscillation periods in AbN
The late-E bursts in the AbN are always phase-locked with (black circles) and PN (red squares) on the hypercapnic
PN bursts, and the ratio between AbN and PN frequencies drive (horizontal axis). The period of AbN late-E dis-
quantally increased through 1:3 (at drive to late-E = 0.32) charges sequentially jumps from 4 PN periods to 3, then
to 1:2 (0.34) and to 1:1 (0.36) as “hypercapnic” drive to to 2, and finally to 1 (Adapted from Molkov et al. (2010)
the late-E population of RTN/pFRG gradually increases with permission)
phase space. The phase space of a system of two BötC oscillators, the system proceeds through
coupled oscillators is a Cartesian product of the a series of phase-locked resonances with 1:N
phase spaces of each oscillator. The ratios between the RTN/pFRG and BötC/pre-
corresponding limit set is a 2D invariant torus, BötC frequencies, with N decreasing from an
and the behavior of this system is represented by initial higher value to 1 (N = 3 on Fig. 5). The
a trajectory on this torus. If the ratio of oscillation specifics of this coupling define a phase (late-E)
frequencies of the two oscillators is rational (i.e., of RTN/pFRG oscillations in all synchronized
equal to N/M, for some integers N and M), then states.
this trajectory is closed, indicating N:M synchro- In contrast, progressive suppression of pre-
nization between oscillators, where N and BötC excitability (e.g., by opioid agonists) results
M represent numbers of rotations around two in a decrease in frequency of BötC/pre-BötC
orthogonal circles that together span the torus. rhythm generator and ultimately in cessation of
With changing conditions, the system of its oscillations. In this regime, the pre-BötC neu-
coupled oscillators can proceed through regimes rons depend on recruitment by the late-E neuron
characterized by different relations between to activate. Because of this coupling, the
oscillation frequencies. A progressive increase pre-BötC frequency is quantally reduced through
of RTN/pFRG excitability as defined by the a series of resonances with M:1 ratios between
CO2 level results in an emergence and subsequent the RTN/pFRG and BötC/pre-BötC frequencies
increase in the frequency of its oscillations. Due with M increasing from 1 to higher values (see
to coupling between RTN/pFRG and BötC/pre- M = 4 on Fig. 5).
C 914 Coupling of Local Field Potentials
Coupled Oscillations in Neural Control of Breathing, variables of these oscillators. Subthreshold activities of
Fig. 5 The phase trajectories in the reduced model illus- late-E neuron (left) and early-I neuron (right) are shown
trating two synchronized late-E: pre-BötC states during by dashed lines. Suprathreshold parts of trajectories (solid
hypercapnia (1:3 regime, left panel) and after suppression lines) correspond to bursts of the late-E (left) and pre-
of pre-BötC (4:1 regime, right panel). The reduced model BotC (right) neurons. Cyclical subthreshold movement
by Rubin et al. (2011) describes neural populations as two- on the left panel indicates two periods of early-I activity
dimensional relaxation oscillators with one fast and one during which the late-E neuron fails to burst. Similar
slow variable. Fast variables V2 and V5 represent mem- movement on the right shows four late-E oscillations
brane potentials of early-I neuron in pre-BötC and late-E during which pre-BotC neurons fail to activate (Adapted
neuron in RTN/pFRG, respectively; m2 and h5 are slow from Rubin et al. (2011) with permission)
Cross-References networks for respiratory rhythm generation. Neuron

37:821–826
Molkov YI, Abdala AP, Bacak BJ, Smith JC, Paton JF,
▶ Computational Models of Mammalian Rybak IA (2010) Late-expiratory activity: emergence
Respiratory CPG and interactions with the respiratory CPG.
Onimaru H, Homma I (2003) A novel functional neuron
group for respiratory rhythm generation in the ventral
References medulla. J Neurosci 23(4):1478–1486
Rubin JE, Bacak BJ, Molkov YI, Shevtsova NA, Smith
Abdala AP, Rybak IA, Smith JC, Paton JF (2009) Abdom- JC, Rybak IA (2011) Interacting oscillations in neural
inal expiratory activity in the rat brainstem-spinal cord control of breathing: modeling and qualitative analy-
in situ: patterns, origins and implications for respira- sis. J Comput Neurosci 30(3):607–632
tory rhythm generation. J Physiol 587(Pt Smith JC, Abdala AP, Koizumi H, Rybak IA, Paton JF
14):3539–3559 (2007) Spatial and functional architecture of the mam-
Alheid GF, McCrimmon DR (2008) The chemical neuro- malian brain stem respiratory network: a hierarchy of
anatomy of breathing. Respir Physiol Neurobiol three oscillatory mechanisms. J Neurophysiol
164(1–2):3–11 98:3370–3387
Iizuka M, Fregosi RF (2007) Influence of hypercapnic aci- Smith JC, Abdala AP, Borgmann A, Rybak IA, Paton JF
dosis and hypoxia on abdominal expiratory nerve activity (2013) Brainstem respiratory networks: building
in the rat. Respir Physiol Neurobiol 157:196–205 blocks and microcircuits. Trends Neurosci
Janczewski WA, Feldman JL (2006) Distinct rhythm gen- 36(3):152–162
erators for inspiration and expiration in the juvenile
rat. J Physiol 570:407–420
Janczewski WA, Onimaru H, Homma I, Feldman JL
(2002) Opioid-resistant respiratory pathway from the
preinspiratory neurones to abdominal muscles: in vivo
and in vitro study in the newborn rat. J Physiol
545:1017–1026
Coupling of Local Field Potentials
Mellen NM, Janczewski WA, Bocchiaro CM, Feldman JL
(2003) Opioid-induced quantal slowing reveals dual ▶ Local Field Potential, Synchrony of
Current Source Density (CSD) Analysis 915 C
Definition
CPG
Current source density analysis (CSD) is a class
▶ Locomotor Pattern Generation in the Rodent of methods of analysis of extracellular electric
Spinal Cord potentials recorded at multiple sites leading to
▶ Vertebrate Pattern Generation: Overview estimates of current sources generating the mea-
sured potentials. It is usually applied to C
low-frequency part of the potential (called the
local field potential, LFP) and to simultaneous
Cross-Correlation of Local Field recordings or to recordings taken with fixed
Potentials time reference to the onset of specific stimulus
(evoked potentials, EP).
▶ Local Field Potential, Synchrony of
Cross-Correlogram Among the different mechanisms contributing to

extracellular electric potential in the tissue
▶ Correlation Analysis of Parallel Spike Trains (Buzsáki et al. 2012; Einevoll et al. 2012), trans-
membrane currents in neurons are believed to
dominate. These are ionic currents passing
through all the different membrane channels
(passive, voltage dependent, calcium dependent,
CSD
synaptic, etc.) as well as the capacitive currents
which, while charging the membrane, also con-
▶ Current Source Density (CSD) Analysis
tribute to the motion of ions in the extracellular
space, influencing the extracellular potential and
seen by the extracellular electrode. The places
where net current is entering or leaving the cell
CSD Method are called current sources or sinks. While these
sources are localized along the membrane of the
▶ Current Source Density (CSD) Analysis neuron, in practice, with a finite resolution
afforded by available electrode setups and lim-
ited by typical densities, we may only recover
coarse-grained density. This is what we have in
Current Source Density (CSD) mind when discussing estimation of current
Analysis source density. Figure 1 shows the relation
between the “microscopic” currents (a), coarse-
Daniel K. Wójcik grained field we usually have in mind (b), and a
Department of Neurophysiology, Nencki reconstruction of the CSD from measured poten-
Institute of Experimental Biology, Warsaw, tials (c).
Poland CSD analysis can be performed for signals in
full spectrum or in any selected band, although it
is usually applied to the low-frequency part
Synonyms (<500 Hz) of the extracellular potential (LFP).
We propose here such an expansion of LFP as the
CSD; CSD method; Reconstruction of current commonly used term local field potential is actu-
sources ally a misnomer, since due to the long-range
C 916 Current Source Density (CSD) Analysis
Current Source Density (CSD) Analysis, cells in the model of thalamocortical loop based on Traub
Fig. 1 Comparison of the current sources obtained in et al. (2005), 10 ms after simulated thalamocortical stim-
a simulation (a) with coarse-grained CSD (smoothed ulus. Dominating contributions to CSD come from infra-
with a Gaussian kernel of s = 75 mm) (b) and a recon- and supragranular pyramidal cells. Vertical distance given
struction with kernel CSD method (c) from LFP computed in mm from cortical surface, horizontal from the center of
at a grid of 8 14 electrodes from the data in (a). Data the simulated column (H. Gła˛bska)
were taken from the cortical part of a simulation of 3,500
nature of the electric field, LFP can be observed prefer to do the opposite, according to the
millimeters away from sources (e.g., Kajikawa sign of CSD (red for positive, blue for nega-
and Schroeder 2011; Lindén et al. 2011; Hunt tive). This situation has led to two opposite
et al. 2011; Łe˛ski et al. 2013). Despite that filter- conventions being in wide use. A reader is
ing in frequencies, it is known that fast processes, advised to always check carefully what is the
such as spikes, may still contribute to the LFP convention used in a given work.
(Buzsáki et al. 2012; Einevoll et al. 2012;
Reimann et al. 2013).
When net positive current enters the cell, Physics Behind the Relation Between the CSD
we speak of current sink, and it corresponds and LFP
to negative CSD. When net negative current To get intuitive understanding of the basic rela-
enters the cell, we speak of current source, tion between current sources and extracellular
and it corresponds to positive CSD. Since potential, consider infinite, homogeneous, and
negative CSD is observed for excitatory syn- isotropic conductive medium of conductivity s
aptic stimulation (positive current entering the (Tranquillo 2008). If we place a stimulating elec-
cell/negative CSD), many researchers prefer to trode and inject current I, it will induce current
denote current sinks by red (“hot spot”) and flow in the tissue with current density
!
current sources by blue. There are J ¼ I^
r =ð4pr 2 Þ radially at a distance r away
a comparable number of researchers who from the stimulation point (Fig. 2).
definition we get the relation between the CSD
and the potential as
→ I
J = r
4πr 2
!
! Cð r0
! 1
V r ¼ d3 r0 (1)
4ps ! !0 C
→ rr
r
or inverting this relation, we obtain the Poisson

equation
I

!
C r ¼ sDV (2)
Equations (1) and (2) are valid only in our

Current Source Density (CSD) Analysis, restricted setting. Equation (2) can be generalized
Fig. 2 Current I injected at the origin of the system for arbitrary conductivity tensor fields s:
spreads uniformly in all directions in infinite, homoge-
neous, and isotropic medium

!
C r ¼ ∇ ½s∇V (3)
In a !
purely!conductive medium, Ohm’s law
holds J ¼ sE ¼ s∇V , which gives us the
potential in space V(r) = I/(4psr). A multitude Solving Eq. (3) usually requires numerical
!
of currents Ij located at r j induce potential
X methods. Careful derivations of Eq. (3) can be
! ! !
V r ¼ I j = 4psj r r j j . It is natural found in (Nicholson 1973; Stevens 1966; Nunez
j and Srinivasan 2005).
to introduce current source density (CSD),
X
! ! !
C r ¼ I j d r r j , a scalar density
j Methods of CSD Estimation
field which is usually coarse grained to The simplest numerical approximation to the
a smooth quantity (see Fig. 1). With this Laplacian is the three-point formula

C ðxk , yl , zm Þ s½V ðxkþ1 , yl , zm Þ 2V ðxk , yl , zm þ Vðxk1 , yl , zm

þV xk , ylþ1 , zm 2Vðxk , yl , zm þ ðxk , yl1 , zm
þV ðxk , yl , zmþ1 Þ 2V ðxk , yl , zm Þ þ V ðxk , yl , zm1 Þ=d2
which was introduced by Pitts (1952). While in laminar structure of the cortex keeping a single
the original application two-dimensional case term in the above formula. This approach has
was considered, by far the most common use of been made popular in the early 1970s by work
this approach has been in the study of of Haberly and Shepherd (1973) and especially
one-dimensional cortical recordings with linear Nicholson and Freeman (1975). Much of the
multielectrodes. One then considers lateral work in this area till 1985 has been summarized
invariance of the potential as a consequence of by Ulla Mitzdorf in her still very useful review
(1985). We refer to this approach as the tradi- interpolated in between. From such a model, one
tional CSD (tCSD) method. can compute potential at the measurement points
Methodologically, there have been few using forward modeling formula, Eq. 1. This leads
advances in the estimation of CSD from the mea- to a matrix relation between the potential and the
X
surements since Pitts work till 2006. The main CSD given by V j ¼ Mjk Ck , where
deficiencies of the traditional approach are lack k
of control over the assumptions made, such as !
ð 0
extent of the sources in the dimensions which
1 ! fk x
3 0
are not probed and between the contacts, spatial Mjk ¼ d x
4ps ! !0
and measurement noise, and the necessity of x j x
exclusion of contacts at the border. To reduce
the noise, Rappelsberger et al. (1981) proposed In lower dimensionality (1D, 2D), where one
to smooth the measurements with a Hamming does not probe all directions, it is necessary to
window obtaining a five-point formula in 1D make assumptions about source behavior there.
For example, in 1D one may assume invariance of

CðzÞ ¼ s½0:23V ðz 2dÞ þ 0:08V ðz d 0:62Vðz CSD on disks of some radius R (e.g., of the size of
þ0:08V ðz þ dÞ þ 0:23ðz þ 2dÞ=d2 cortical column) orthogonal to the shaft, and in 2D
one may assume constancy or Gaussian decay of the
which gained some popularity. In 1988, Vaknin source on an interval orthogonal to the MEA plane.
et al. proposed to extend the grid of electrodes These assumptions lead to other forms of matrix
beyond the actual set of contacts copying the elements which incorporate the specific form of
outmost recordings to the added contacts. While the model (Pettersen et al. 2006; Łe˛ski et al. 2011).
in general physically questionable, this numerical For typical recordings the matrix relation
trick allowed computation of CSD values for all between the LFP and CSD can be inverted lead-
contact positions and was used in cortical studies. ing to a formula
!
for C!k
as a function of measured
Rapid development of new multielectrodes at potential, C ¼ M1 V , and in consequence to an
the beginning of the twenty-first century (Egert estimate of current sources in the whole probed
et al. 1998; Csicsvari et al 2003; Buzsáki 2004; region. A convenient feature of iCSD method is
Berdondini et al 2005; Kipke et al. 2008; Frey that the matrix M1 is estimated once for a given
et al 2009) stimulated renewed interest in LFP setup and model of sources. Also, one can easily
recordings and analysis, including CSD analysis. incorporate different boundary conditions over-
In 2006, Pettersen et al. proposed a new, model- coming naturally this limitation of the traditional
based approach to CSD estimation, which they approach (Łe˛ski et al. 2007).
termed inverse CSD (iCSD) method. The method Inverse CSD is a framework which allows one to
was later generalized to three- and incorporate different assumptions about the struc-
two-dimensional regular recording setups (Łe˛ski ture of the sources or the properties of the tissue,
et al. 2007, 2011). e.g., its conductivity (Goto et al. 2010). An interest-
The idea behind iCSD is to assume a parametric ing variant was developed for localization of single-
model of the sources of as many parameters as the cell current sources during action potential genera-
number of measurements. For example, for mea- tion (spike CSD, sCSD, Somogyvári et al. 2005,
surements of potential V1, . . . ,VN at points 2012). The flexibility of the framework is in the
! ! construction of the function space used for estima-
x 1 , . . . , x N , we may take the values of CSD at
measurement points C1,. . .,CN as parameters and tion. However, iCSD was developed for regular
spline interpolate in between (spline iCSD). Then recording grids and under the assumption of negli-
X N gible recording and position noise, which could not
! ! !
C x ¼ Ck f k x , where f k x is a func- be easily overcome within this framework.
k¼1
! !
A general solution to these problems was pro-
tion taking 1 at x k , 0 at x l6¼k , and spline vided with kernel current source density (kCSD)

method (Potworowski et al. 2012). While in iCSD e ¼ V ðxÞ ¼ a1 b1 ðxÞ þ .. . þ aM bM ðxÞ : bi : ℝd ! ℝ ,

F
the dimension of the function space in which one
does estimation is equal to the number of mea-
surements, in kCSD one constructs spaces of with the dimension of the spaces, M, much
much larger dimensionality. The flexibility in greater than the number of measurements, N.
tackling arbitrary electrode setups comes from The basis functions are related by a linear opera-
separation of the construction of estimation tor A : F e 7!F so that bi ¼ Ab~i , and C
space from the definition of the setup. The use XM
of kernel methods allows us to use standard tech- V ðxÞ ¼ ACðxÞ ¼ ai bi ðxÞ . In three dimen-
niques for dealing with noise (e.g., ridge regres- i¼1
sion, etc.) as described below. sions (isotropic, homogeneous) A ¼ sD , in

Assume potentials V1, . . . VN measured at lower dimensionality, we must assume properties
points x1, . . . , xN. To construct the framework of the sources in the directions not probed for the
of kCSD, we start with two linear spaces: the same reasons as in iCSD. For instance, if in 2D
space of sources we assume that the sources are invariant in
the direction z orthogonal to the electrode plane

e ¼ CðxÞ ¼ a1 b~1 ðxÞ þ . .. þ aM b~M ðxÞ : b~i : ℝd ! ℝ ,

F (x,y) within a layer of 2h, then the physical
CSD(x,y,z) is C(x,y) for | z | < h and 0 otherwise,
and the space of potentials and
0 1
ð ð
1 B h C
bi ðx, yÞ ¼ A b~i :¼ dx0 dy0 arsinh@qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiAb~i ðx0 , y0 Þ
2ps 2
ðx x0 Þ þ ðy y0 Þ 2
We require that b~i (bi) are linearly independent theory (Schoelkopf and Smola 2002), one can
and so they constitute bases of the linear spaces show that minimum estimation error
e and F . To efficiently estimate in such a large
F
space, we construct a kernel function X
N
err ðV ðÞÞ ¼ ð V ð x i Þ V i Þ 2 þ l k bk 2
i¼1
X
M
K ðx, x0Þ ¼ bi ðxÞbi ðx0Þ, is obtained for potential function of the form
i¼1
X
N
V ð xÞ ¼ bk K ðxk , xÞF ,
which introduces the structure of reproducible k¼1
kernel Hilbert space (RKHS) on F (Aronszajn
1950). Using representer theorem from RKHS where
2 3 2 31 2 3
b1 K ðx1 , x1 Þ þ l K ðx1 , xN Þ V1
4⋮5 ¼ 4 ⋮ ⋱ ⋮ 5 4 ⋮ 5,
bN K ð xN , x1 Þ K ð xN , xN Þ þ l VN
which can be written in more compact notation as Limitations of CSD Analysis

The main difficulty in the interpretation of CSD
b ¼ ðK þ lIÞ1 V profiles is that it is not possible to tell without
extra knowledge what is the nature of a given
with an obvious definition of terms. Having esti- observed feature, for example, if a spot of nega-
mated the potential, we use the relation between the tive CSD is due to excitatory synaptic stimulation
basis functions to construct a cross-kernel function: or a passive return current matching inhibitory
current elsewhere. Thus, one usually has to build
X
M on extra a priori knowledge of system’s anatomy
K~ ðx, xÞ ¼ bj ðxÞb~j ðxÞ and physiology (Gratiy et al. 2011) or use com-
j¼1
putational models of the studied systems
(Makarov et al. 2010; Potworowski et al. 2011).
with which the current sources estimated in the
If possible, do both.
measurement space are given by
CSD analysis allows one to better localize
~ T 1
C ðxÞ ¼ K ðxÞ ðK þ lIÞ V with K ðxÞ :¼ ~ T
neural activity by deconvolution of the inverse

K~ ðx1 , xÞ, . . . , K~ ðxn , xÞ . distance kernel. However, since usually multiple
Parameter l can be selected from data using, cell populations overlap, observed profile of the
for instance, cross-validation. For further details current sources will reflect summary activity of
on kCSD, see Potworowski et al. (2012). Figure 3 all of them. One way to overcome this problem is
shows an example of CSD estimation with kCSD to use a method of source decomposition (see
method. Einevoll et al. 2013, for a discussion). It seems
Current Source Density (CSD) Analysis, the potential measured at the electrode locations. (c) CSD
Fig. 3 Example of source reconstruction with kCSD reconstructed from the random 16 measurements. (d–f)
method. (a) Model sources shown were used to generate Sources reconstructed from 2, 4, and 8 measurements,
potentials at randomly selected electrode locations respectively
marked with black dots. (b) Potential interpolated from
that independent component analysis (ICA) fol- on substrate-integrated multielectrode arrays. Brain
lowing CSD gives functionally meaningful Res Brain Res Protoc 2(4):229–242
Einevoll GT, Lindén H, Tetzlaff T, Łe˛ski S, Pettersen KH
results (Łe˛ski et al. 2010; Makarov et al. 2010; (2012) Local field potentials. Biophysical origin and
Potworowski et al. 2011). analysis. In: Quiroga RQ and Panzer S (Ed.) Principles
To tackle both the problem of overlapping of neural coding. CRC Press, Boca Raton pp 37–61
populations and unknown origin of the given Einevoll GT, Kayser C, Logothetis NK, Panzeri S (2013)
Modelling and analysis of local field potentials for
source, it is particularly useful to generate ground studying the function of cortical circuits. Nat Rev
C
truth data as close to the system studied as possi- Neurosci 14(11):770–785
ble, for instance, from large-scale models Frey U, Egert U, Heer F, Hafizovic S, Hierlemann
(Potworowski et al. 2011; Reimann et al. 2013). A (2009) Microelectronic system for high-resolution
mapping of extracellular electric fields applied to brain
Finally, let us observe that all the CSD slices. Biosens Bioelectron 24(7):2191–2198
methods mentioned so far rely on Eq. (3) which Goto T, Hatanaka R, Ogawa T, Sumiyoshi A, Riera J,
has recently been challenged (see chapter Local Kawashima R (2010) An evaluation of the conductiv-
Field Potentials: Interaction with the Extracellu- ity profile in the somatosensory barrel cortex of Wistar
rats. J Neurophysiol 104(6):3388–3412
lar Medium for a review). In that case the phys- Gratiy SL, Devor A, Einevoll GT, Dale AM (2011) On the
ical part of the CSD model must be modified estimation of population-specific synaptic currents
(Bedard and Destexhe, Physical Review E 2011). from laminar multielectrode recordings. Front
Neuroinform 5:32
Haberly LB, Shepherd GM (1973) Current-density analy-
sis of summed evoked potentials in opossum
Cross-References prepyriform cortex. J Neurophysiol 36(4):789–802
Hunt MJ, Falinska M, Łeski S, Wójcik DK, Kasicki
▶ Independent Component Analysis of Images S (2011) Differential effects produced by ketamine
on oscillatory activity recorded in the rat hippocam-
▶ Local Field Potential Interaction with the pus, dorsal striatum and nucleus accumbens.
Extracellular Medium J Psychopharmacol 25(6):808–821
▶ Local Field Potentials (LFP) Kajikawa Y, Schroeder CE (2011) How local is the local
field potential? Neuron 72(5):847–858
Kipke DR, Shain W, Buzsáki G, Fetz E, Henderson JM,
Hetke JF, Schalk G (2008) Advanced
References neurotechnologies for chronic neural interfaces: new
horizons and clinical opportunities. J Neurosci
Aronszajn N (1950) Theory of reproducing kernels. Trans 28(46):11830–11838
Am Math Soc 68(3):337–404 Łe˛ski S, Wójcik DK, Tereszczuk J, Świejkowski DA,
Bedard C, Destexhe A (2011) A generalized theory for Kublik E, Wróbel A (2007) Inverse current-source
current-source density analysis in brain tissue. Phys density method in 3D: reconstruction fidelity, bound-
Rev E 84:041909 ary effects, and influence of distant sources.
Berdondini L, van der Wal PD, Guenat O, de Rooij NF, Neuroinformatics 5(4):207–222
Koudelka-Hep M, Seitz P, Kaufmann R, Metzler P, Łe˛ski S, Kublik E, Swiejkowski DA, Wróbel A, Wójcik
Blanc N, Rohr S (2005) High-density electrode array DK (2010) Extracting functional components of neural
for imaging in vitro electrophysiological activity. dynamics with independent component analysis and
Biosens Bioelectron 21(1):167–174 inverse current source density. J Comput Neurosci
Buzsáki G (2004) Large-scale recording of neuronal 29(3):459–473
ensembles. Nat Neurosci 7(5):446–451 Łe˛ski S, Pettersen KH, Tunstall B, Einevoll GT, Gigg J,
Buzsáki G, Anastassiou CA, Koch C (2012) The origin of Wójcik DK (2011) Inverse current source density
extracellular fields and currents–EEG, ECoG, LFP and method in two dimensions: inferring neural activation
spikes. Nat Rev Neurosci 13(6):407–420 from multielectrode recordings. Neuroinformatics
Csicsvari J, Henze DA, Jamieson B, Harris KD, Sirota A, 9(4):401–425
Barthó P, Wise KD, Buzsáki G (2003) Massively par- Łe˛ski S, Lindén H, Tetzlaff T, Pettersen KH, Einevoll GT
allel recording of unit and local field potentials with (2013) Frequency dependence of signal power and
silicon-based electrodes. J Neurophysiol 90(2): spatial reach of the local field potential. PLoS Comput
1314–1323 Biol 9(7):e1003137
Egert U, Schlosshauer B, Fennrich S, Nisch W, Fejtl M, Lindén H, Tetzlaff T, Potjans TC, Pettersen KH, Gr€ un S,
Knott T, M€uller T, H€ammerle H (1998) A novel Diesmann M, Einevoll GT (2011) Modeling the spatial
organotypic long-term culture of the rat hippocampus reach of the LFP. Neuron 72(5):859–872
C 922 Cutaneous Mechanoreceptive Afferents: Neural Coding of Texture
Makarov VA, Makarova J, Herreras O (2010) Disentan- Higley MJ, Whittington MA (2005) Single-column
glement of local field potential sources by independent thalamocortical network model exhibiting gamma
component analysis. J Comput Neurosci 29(3): oscillations, sleep spindles, and epileptogenic bursts.
445–457 J Neurophysiol 93(4):2194–2232
Mitzdorf U (1985) Current source-density method and Vaknin G, DiScenna PG, Teyler TJ (1988) A method for
application in cat cerebral cortex: investigation of calculating current source density (CSD) analysis
evoked potentials and EEG phenomena. Physiol Rev without resorting to recording sites outside the sam-
65(1):37–100 pling volume. J Neurosci Methods 24(2):131–135
Nicholson C (1973) Theoretical analysis of field potentials
in anisotropic ensembles of neuronal elements. IEEE
Further Reading
Transactions on Biomedical Engineering 20(4):
Scholarpedia
278–288
Local field potential
Nicholson C, Freeman JA (1975) Theory of current
source-density analysis and determination of conduc-
tivity tensor for anuran cerebellum. J Neurophysiol
38(2):356–368
Nunez PL, Srinivasan R (2005) Electric fields of the brain:
the neurophysics of EEG. MIT Press, Cambridge, MA Cutaneous Mechanoreceptive
Pettersen KH, Devor A, Ulbert I, Dale AM, Einevoll GT Afferents: Neural Coding of Texture
(2006) Current-source density estimation based on
inversion of electrostatic forward solution: effects of
Justin Lieber
finite extent of neuronal activity and conductivity dis-
continuities. J Neurosci Methods 154(1–2):116–133 Committee on Computational Neuroscience,
Pitts WH (1952) Investigations on synaptic transmission. University of Chicago, Chicago, IL, USA
In: von Foerster H (Ed.) Cybernetics, Transactions of
the 9th conference, Josiah Macy Foundation, New
York pp 159–166
Potworowski J, Glabska H, Leski S, Wojcik D (2011) Definition
Extracting activity of individual cell populations
from multielectrode recordings. BMC Neurosci We can identify objects by the way they feel. As
12(Suppl 1):374
we move our fingers over objects in our pocket or
Potworowski J, Jakuczun W, Łe˛ski S, Wójcik D (2012)
Kernel current source density method. Neural Comput purse, we can easily distinguish between our keys
24(2):541–575 and our phone or between two different fabrics in
Rappelsberger P, Pockberger H, Petsche H (1981) Current our dresser drawer. The tactile perception of the
source density analysis: methods and application to
surface texture of objects, i.e., of their micro-
simultaneously recorded field potentials of the rabbit’s
visual cortex. Pflugers Arch 389(2):159–170 structure and material properties, contributes to
Reimann MW, Anastassiou CA, Perin R, Hill SL, our ability to identify them by touch. Signals in
Markram H, Koch C (2013) A biophysically detailed the nerve convey information that allows us to
model of neocortical local field potentials predicts the
identify different textures, as well as attribute
critical role of active membrane currents. Neuron
79(2):375–390 different perceptual properties (e.g., roughness
Schoelkopf B, Smola A (2002) Learning with Kernels. or hardness) to them.
Massachusetts Institute of Technology, Cambridge,
MA
Somogyvári Z, Zalányi L, Ulbert I, Erdi P (2005) Model-
based source localization of extracellular action poten- Detailed Description
tials. J Neurosci Methods 147(2):126–137
Somogyvári Z, Cserpán D, Ulbert I, Erdi P (2012) Local- Spatial and Temporal Mechanisms
ization of single-cell current sources based on extra-
When we run our fingers across a textured sur-
cellular potential patterns: the spike CSD method. Eur
J Neurosci 36(10):3299–3313 face, a characteristic pattern of deformations is
Stevens CF (1966) Neurophysiology: a primer. Wiley, produced in the skin (Sripati et al. 2006), and the
New York resulting stresses and strains are transduced by
Tranquillo J (2008) Quantitative neurophysiology. Mor-
multiple populations of mechanoreceptors
gan and Claypool Publishers, San Rafael
Traub RD, Contreras D, Cunningham MO, Murray H, embedded in the skin. The spatial layout and
LeBeau FEN, Roopun A, Bibbig A, Wilent WB, timing of the consequent spatiotemporal patterns
Cutaneous Mechanoreceptive Afferents: Neural Coding of Texture 923 C
Cutaneous Mechanoreceptive Afferents: Neural Examples of the SA1, RA, and PC response to letters
Coding of Texture, Fig. 1 SA1 spatial image. SA1 serially scanned over the finger (Phillips et al. 1988).
receptors carry a spatial pattern of information. (a) Note that the SA1 spatial image is characterized by well-
When dot patterns (top) are serially scanned over the resolved edges and corners (when compared to the RA and
finger, they elicit a spatial pattern of activation in example PC spatial images)
SA1, RA, and PC receptors (Connor et al. 1990). (b)
of afferent activation convey information about et al. 1988), curvature (Yau et al. 2013), and
surface texture (Weber et al. 2013). motion (Pei et al. 2010) in ways that draw strong
As coarse textural features indent the skin, analogies to their visual cortical counterparts.
their spatial layout is reflected in the spatial pat- Although the SA1 spatial image effectively
tern of activation evoked across the population of relays information about coarse textural features,
Merkel disks, which are relatively dense in the many tangible surface features are too fine to be
fingertip skin and are innervated by slowly resolved spatially, given that the skin filters out
adapting type 1 (SA1) afferents (Johansson and very fine features (Sripati et al. 2006; Weber
Vallbo 1979) (Fig. 1). SA1 afferents thus encode et al. 2013). In fact, information about fine tex-
spatial patterns in a manner analogous to photo- tural features is not perceptually accessible until
receptors in the retina. SA1 responses are the skin is actively scanned over a surface
uniquely suited to transduce spatial features as (Hollins and Risner 2000). Movement between
they have small receptive fields; that is, they only a textured surface and the skin results in vibra-
respond to small and sharply defined patches of tions (Manfredi et al. 2014; Bensmaia and
skin. SA1 responses are strongly modulated by Hollins 2003, 2005) that reflect both spatially
spatial features like edges, corners, and indenta- resolvable features as well as fine surface
tion depth (Johnson 2001). The spatial image microgeometry. These vibrations elicit highly
carried by SA1 afferents is further elaborated in precise and repeatable temporal spiking patterns
the somatosensory cortex, where neural in rapidly adapting Meissner corpuscles
responses exhibit a sensitivity to spatial features (innervated by rapidly adapting or RA afferents)
like edges (Bensmaia et al. 2008; Phillips and Pacinian corpuscles (innervated by PC
Cutaneous Mechanoreceptive Afferents: Neural trial power spectra of PC responses. Textures may elicit
Coding of Texture, Fig. 2 The timing of PC responses sharply periodic responses from PC receptors. (d) The
encodes information about skin vibrations. (a) High- periodicity in PC responses (average power spectra,
resolution surface profiles of natural texture stimuli. (b) orange) matches the frequency of skin vibrations (black)
Texture-specific responses by an example PC receptor. PC (Weber et al. 2013)
responses are precisely timed and repeatable. (c) Trial-by-
afferents). These patterns carry information Roughness

about the skin vibrations (Talbot et al. 1968; Perceived roughness is determined by how vari-
Mackevicius et al. 2012) and thus about the tex- able the neural responses to textures are over time
ture itself (Weber et al. 2013; Fig. 2). and space. Specifically, the spatial variability
In summary, then, coarse textural features are across the SA1 population response (Fig. 3a;
encoded in the spatial pattern of activity evoked Connor et al. 1990; Connor and Johnson 1992;
in one population of afferents – SA1 Blake et al. 1997) is integrated with the temporal
fibers – while fine textural features are encoded variability in the response of RA and PC afferents
in the temporal patterns of activity evoked in the (Fig. 3b, c; Weber et al. 2013) to culminate in
other two, RA and PC fibers. a percept of roughness (Fig. 3d). In other words,
a texture is perceived as rough to the extent that
Multidimensionality of Texture Perception different SA1 afferents produce different
The sensation of exploring a surface can be responses and individual RA and PC afferents
described along a number of perceptual dimen- produce responses that change over time.
sions: things may feel rough or smooth, hard or Sensitivity to peripheral variability is
soft, sticky or slippery, and warm or cool (Hollins a common feature of sensory processing. Indeed,
et al. 2000). The dominant dimension, and the when stimuli do not change in time, central sen-
most studied one, is roughness. sory neurons across all modalities adapt toward
Cutaneous Mechanoreceptive Afferents: Neural (roughness) and neurons (variation). The gray line denotes
Coding of Texture, Fig. 3 Signals from multiple recep- the line of best fit. SA1 spatial variation barely changes
tors types are necessary to explain roughness. (a) SA1 across a fourfold scale in roughness for fine textures.
spatial variation is plotted against judgments of roughness (b and c) Temporal variation for RA and PC afferents,
for natural textures, including both coarse (plastic dots and respectively, plotted against perceived roughness. (d)
gratings, green) and fine (fabrics and sandpapers, orange) A combined model of all three predictors almost perfectly
textures. Error bars represent SEMs across subjects predicts texture roughness (Weber et al. 2013).
their baseline level of activity (Wark et al. 2007). kinesthesia, and even audition may be combined
In this way, adaptation serves to increase neural to culminate in a percept of hardness (Srinivasan
sensitivity to variations in the sensory input. Fur- and Lamotte 1995; LaMotte 2000). Thus, hard-
thermore, sensory neurons in both visual and ness perception involves a complex reconstruc-
somatosensory cortex exhibit both excitatory tion of a measureable surface property but does
and inhibitory components and thus show sensi- not seem to be based on a simple readout of
tivity to spatial differences in activation across a singular neural quantity.
the sensory sheet. Information from both spatial and temporal
mechanisms is combined to reconstruct judg-
Hardness ments of hardness. Surfaces with different com-
Judgments of surface hardness lie on a subjective pliance lead to different spatial patterns of
continuum that closely covaries with measure- deformation of the skin: soft surfaces deform
ments of surface compliance (Fig. 4a; Harper around the skin, while hard surfaces change the
and Stevens 1964). Information from spatial pat- shape of the skin. These spatial patterns of skin
terns of afferent activation, response timing, deformations are reflected in the spatial patterns
Cutaneous Mechanoreceptive Afferents: Neural (numbered from the softest (1) to hardest (9) surface) for
Coding of Texture, Fig. 4 Hardness perception. (a) three different application forces. Both SA1 and RA affer-
Judgments of hardness and softness covary closely with ents fire, while surfaces are actively loaded onto the skin
a measure of surface compliance (Harper and Stevens (left), but only SA1 afferents fire when the surfaces are
1964). (b) Afferent activity in response to compliant stim- static (right). For a given force, firing rates increase with
uli indented into the skin. SA1 (top) and RA (bottom) surface hardness
responses are plotted against surface compliance
of activation elicited in populations of SA1 affer- (SAII) afferents) (Knibestol 1975) as well as
ents (Franzén et al. 1996). SA1 and RA afferents (Birznieks et al. 2001,
Precisely timed information can also be used 2010).
to determine hardness. When judged through
a probe, hardness is most effectively perceived Thermal Conductivity
using rapid tapping motions (LaMotte 2000), Perception of a surface’s temperature is clearly
which elicit dynamic force profiles that most used for object identification (Katz and Krueger
effectively recruit RA and PC afferents. Indeed, 1989) and is incorporated into the
while SA1 responses to statically indented stim- multidimensional percept of texture (Bensmaia
uli are informative about surface compliance, RA and Hollins 2005; Hollins et al. 2000). Because
responses only convey hardness information dur- everyday objects tend to be colder than skin tem-
ing the dynamic loading phase (Fig. 4b; Condon perature, surface temperature is determined by
et al. 2013). the rate of heat flow out of the skin (Ho and
Jones 2006). This heat flow is almost certainly
Stickiness transduced by thermoreceptive afferents in the
Perceptual judgments of stickiness are often skin (Darian-Smith et al. 1973, 1979; Johnson
closely correlated with perceptual judgments of et al. 1973, 1979).
roughness (Hollins et al. 2000). When exploring
the stickiness of a surface, subjects tend to apply Conclusion
a set amount of force normal to the surface. The Spatiotemporal patterns of activation across the
resulting range of force tangential to the surface four main populations of afferents that innervate
covaries closely with measurements of perceived the glabrous skin of the hand – SA1, RA, PC, and
stickiness (Smith and Scott 1996). Tangential SAII – as well as thermoreceptors, convey exqui-
forces stretch the skin, which excites Ruffini cyl- site and multidimensional information about sur-
inders (innervated by slowly adapting type II face texture. Different information about surface
texture is conveyed in different aspects of affer- Ho HN, Jones LA (2006) Contribution of thermal cues to
ent responses, with spatial patterns mediating the material discrimination and localization. Percept
Psychophys 68(1):118–128
perception of coarse features and compliance and Hollins M, Risner SR (2000) Evidence for the duplex
temporal patterns mediating the perception of theory of tactile texture perception. Percept
fine features. Importantly, the perception of Psychophys 62(4):695–705
texture reflects the integration of different types Hollins M et al (2000) Individual differences in perceptual
space for tactile textures: evidence from
of signals from multiple different tactile multidimensional scaling. Percept Psychophys
C
submodalities. 62(8):1534–1544
Johansson RS, Vallbo AB (1979) Tactile sensibility in the
human hand: relative and absolute densities of four
types of mechanoreceptive units in glabrous skin.
References J Physiol 286:283–300
Johnson KO (2001) The roles and functions of cutaneous
Bensmaia SJ, Hollins M (2003) The vibrations of texture. mechanoreceptors. Curr Opin Neurobiol 11(4):
Somatosens Mot Res 20(1):33–43 455–461
Bensmaia S, Hollins M (2005) Pacinian representations of Johnson KO, Darian-Smith I, LaMotte C (1973) Periph-
fine surface texture. Percept Psychophys 67(5): eral neural determinants of temperature discrimination
842–854 in man: a correlative study of responses to cooling
Bensmaia SJ et al (2008) The representation of stimulus skin. J Neurophysiol 36(2):347–370
orientation in the early stages of somatosensory Johnson KO et al (1979) Coding of incremental changes in
processing. J Neurosci 28(3):776–786 skin temperature by a population of warm fibers in the
Birznieks I et al (2001) Encoding of direction of fingertip monkey: correlation with intensity discrimination in
forces by human tactile afferents. J Neurosci 21(20): man. J Neurophysiol 42(5):1332–1353
8222–8237 Katz D, Krueger LE (1989) The world of touch. Erlbaum,
Birznieks I et al (2010) Encoding of tangential torque in Hillsdale. xii 260 pp
responses of tactile afferent fibres innervating the Knibestol M (1975) Stimulus-response functions of
fingerpad of the monkey. J Physiol 588(Pt 7): slowly adapting mechanoreceptors in the human gla-
1057–1072 brous skin area. J Physiol 245(1):63–80
Blake DT, Hsiao SS, Johnson KO (1997) Neural coding LaMotte RH (2000) Softness discrimination with a tool.
mechanisms in tactile pattern recognition: the relative J Neurophysiol 83(4):1777–1786
contributions of slowly and rapidly adapting mecha- Mackevicius EL et al (2012) Millisecond precision spike
noreceptors to perceived roughness. J Neurosci timing shapes tactile perception. J Neurosci 32(44):
17(19):7480–7489 15309–15317
Condon M et al (2013) Differential sensitivity to surface Manfredi LR et al (2014) Natural scenes in
compliance by tactile afferents in the human finger tactile texture. J Neurophysiol. (doi:10.1152/jn.
pad. J Neurophysiol 111:1308–1317 00680.2013)
Connor CE, Johnson KO (1992) Neural coding of tactile Pei YC et al (2010) Shape invariant coding of motion
texture: comparison of spatial and temporal mecha- direction in somatosensory cortex. PLoS Biol 8(2):
nisms for roughness perception. J Neurosci 12(9): e1000305
3414–3426 Phillips JR, Johnson KO, Hsiao SS (1988) Spatial pattern
Connor CE et al (1990) Tactile roughness: neural codes representation and transformation in monkey somato-
that account for psychophysical magnitude estimates. sensory cortex. Proc Natl Acad Sci U S A 85(4):
J Neurosci 10(12):3823–3836 1317–1321
Darian-Smith I, Johnson KO, Dykes R (1973) “Cold” fiber Smith AM, Scott SH (1996) Subjective scaling of smooth
population innervating palmar and digital skin of the surface friction. J Neurophysiol 75(5):1957–1962
monkey: responses to cooling pulses. J Neurophysiol Srinivasan MA, Lamotte RH (1995) Tactual discrimina-
36(2):325–346 tion of softness. J Neurophysiol 73(1):88–101
Darian-Smith I et al (1979) Warm fibers innervating Sripati AP, Bensmaia SJ, Johnson KO (2006)
palmar and digital skin of the monkey: A continuum mechanical model of mechanoreceptive
responses to thermal stimuli. J Neurophysiol 42(5): afferent responses to indented spatial patterns.
1297–1315 J Neurophysiol 95(6):3852–3864
Franzén O, Johansson R, Terenius LY (1996) Somesthesis Talbot WH et al (1968) The sense of flutter-vibration:
and the neurobiology of the somatosensory cortex. In: comparison of the human capacity with response pat-
Advances in life sciences. Birkh€auser Verlag, Basel/ terns of mechanoreceptive afferents from the monkey
Boston, xx, 421 pp hand. J Neurophysiol 31(2):301–334
Harper R, Stevens SS (1964) Subjective hardness of com- Wark B, Lundstrom BN, Fairhall A (2007) Sensory adap-
pliant materials. Q J Exp Psychol 16(3):204–215 tation. Curr Opin Neurobiol 17(4):423–429
C 928 Cx3D: Cortex Simulation in 3D
Weber AI et al (2013) Spatial and temporal codes mediate doubles, and thus, exponentially more cells are
the tactile perception of natural textures. Proc Natl created over time. These cells form all the body
Acad Sci U S A 110(42):17107–17112
Yau JM, Connor CE, Hsiao SS (2013) Representation of parts, but how do they do this? There is no global
tactile curvature in macaque somatosensory area 2. controlling unit that tells each cell what to do.
J Neurophysiol 109(12):2999–3012 Each cell individually has to figure out what it has
to do in this vast sea of potentially billions of
other cells.
Cx3D is a simulation program that allows
Cx3D: Cortex Simulation in 3D modelers to try out their theories of how the
developmental process happens. Even though
Andreas Hauri Cx3D was written with the intention of studying
Institute of Neuroinformatics, ETHZ/UZH, the development of neural tissue, it is not limited
Z€urich, Switzerland to that. Cx3D allows cells to divide, grow out
axons and dendrites, secrete external and internal
chemicals, read internal and external chemicals,
Definition and move around in the environment (the extra-
cellular matrix which forms the biological envi-
Cx3D (pronounced Cortex 3D) is a simulation ronment of cells). The modeler has full control
tool for simulating the growth primarily of neu- over these actions and can write small modules
ral tissue but is also capable of simulating other called BiologicalModules that are essentially
tissue growth. Cx3D allows a modeler to simu- programs running in the cells to define their
late the behavior of cells and their axonal and behavior. BiologicalModules can be seen as an
dendritic extensions in an artificial environ- epiphenomenological model of proteins working
ment. Each cell can interact locally with its together that define one behavior of a cell, e.g., to
environment that consists of other cells and move the cell in a direction defined by an extra-
a chemical environment. The emphasis in cellular chemical gradient.
Cx3D is on the local behavior of cells and the Cx3D provides a basic physics engine that
emergence of a global structure through the prevents cells and cell parts from overlapping.
local behavior of the cells. It is not concerned The user of the simulation tool does not have to
with the simulation of an individual cell’s inter- worry about the detailed implementation of the
nal environment but how multiple cells can physical processes if he does not want to do so
organize themselves. and can focus on the biological aspects of the
model. In order to support the modeler even fur-
ther, a language has been developed that supports
Detailed Description the modeler in only allowing biologically plausi-
ble commands for the cells. This language,
In order to understand how tissue forms, and named GCode, is not directly part of Cx3D and
therefore how parts of the body of therefore does not have to be used, but it proves to
a multicellular organism form, we must study be very helpful in developing correctly behaving
how cells organize themselves into these com- models.
posites. Each cell is an individual; it has an inter- In Cx3D, the cells are compartmentalized.
nal environment of chemicals that is separated by The soma is approximated with a sphere, and
a membrane from the external environment. The the axons, dendrites, and other processes are
cell is the basic operational unit of an organism. approximated with cylinders. Each of these
In the development of an animal, each body starts compartments can potentially run a different
with the fertilized egg. The fertilized egg, a single BiologicalModule and therefore have a biologi-
cell, starts to divide and create more cells. With cally active behavior. They are all passively sub-
each cell cycle the number of cells potentially ject to the force scheme of the simulation tool.
Cytoplasmic Resistivity 929 C
Cx3D operates on three levels of simulation: into Cx3Dp (p for parallel) that runs in a multi-
a biological, a physical, and a spatial level. The core/multi-computer environment to allow for
spatial level provides the neighborhood relations these big simulations.
between the different compartments of the cells. More information on Cx3D, along with links
The physical level uses the spatial information to to publications and the latest version of the source
calculate forces between the cellular compartments code, can be found at http://www.ini.uzh.ch/
and takes care of intracellular and extracellular ~amw/seco/cx3d. C
chemical diffusion processes. The biological level
makes use of the physical level and organizes the
compartments into cells that can run
BiologicalModules. The simulation is round based
and approximates development in small time steps. Cytoplasmic Resistivity
Since tissue is built of thousands and poten-
tially millions of cells, Cx3D has been extended ▶ Resistivity, Axial
D
Decision-Making that sits at the interfaces of many areas, among

them psychology, neuroscience, economics,
▶ Choice Behavior finance, political science, engineering, and math-
ematics. The range of decision-making tasks seen
in the literature is extremely diverse, reflecting
the interdisciplinary nature of the area and the
Decision-Making Tasks diverse background of the researchers engaged in
it. This entry attempts to summarize and catego-
Angela J. Yu rize the major classes of decision-making tasks
Department of Cognitive Science, University of with respect to the formal, mathematically quan-
California, San Diego, La Jolla, CA, USA tifiable properties of the tasks and the cognitive
processes that may consequently be entailed.
Broadly, decision-making tasks can be grouped
Definition into the following categories and subcategories:
• Decision-making under uncertainty
A diverse repertoire of behavioral tasks has been – Sensory uncertainty
employed to examine the cognitive processes and – Outcome uncertainty
neural basis underlying decision-making in • Multidimensional decision-making
humans and animals. Some of these have their – Target-distractor differentiation
origins in psychology, others in cognitive neuro- – Multisensory integration
science, yet others in economics. There is also • Preference-based decision-making
a continual invention of novel or hybrid para-
digms, sometimes motivated by deep conceptual
questions stimulated by computational modeling Decision-Making Under Uncertainty
of decision-making and sometimes motivated by
specific hypotheses related to functions of neuro- Many decision-making tasks require the subjects
nal systems and brain regions that are suspected to make choices among alternatives under condi-
of playing an important role in decision-making. tions of uncertainty. Uncertainty can arise from
noise related to sensory inputs, ignorance about
stimulus or action outcomes, stochasticity or
Detailed Description imprecise representation related to (relative) tem-
poral onset of events, or some hybrid combina-
The area of decision-making is a dynamically tions of two or more of these factors. The
evolving, multifaceted area of active research difficulty of the tasks can often be controlled

D 932 Decision-Making Tasks
parametrically via the level of uncertainty present until the subject makes a perceptual
induced experimentally, and the tasks can there- response (Roitman and Shadlen 2002). The latter,
fore be used to probe fundamental limitations and known as the reaction time (RT) variant, is the
properties of human and animal cognition, as behaviorally more interesting case, as it allows
well as characterize individual and group the subject to decide not only which stimulus was
differences. seen but how much sensory data to accumulate
about the stimulus before responding. This has
made this task a simple yet effective means for
Sensory Uncertainty probing the speed-accuracy trade-off in percep-
tion (Gold and Shadlen 2002). Another property
Decision-making under sensory uncertainty, or of the task that makes it ideal for examining the
perceptual decision-making, involves tasks in speed-accuracy trade-off is that it corrupts the
which the subject views a noisy stimulus and sensory stimulus with a significant amount of
chooses a response consistent with the perceived noise that is identical and independent over
response. One such widely used task is the time, making the informational value of the sen-
random-dot coherent motion task (Newsome sory stimulus constant over time (Bogacz
and Paré 1988), in which a field of flickering et al. 2006). This property also gives the experi-
dots is displayed to the subject, a fraction menter an accessible experimental parameter for
(usually a minority) of which has simulated controlling the information rate (signal-to-noise
motion in a coherent direction, and the remainder ratio) of the stimulus, through the coherence of
either move in random directions (Newsome and the stimulus, or the percentage of coherently
Paré 1988) or have no discernible direction of moving dots.
motion (Shadlen and Newsome 2001). Based on The mathematically optimal decision policy
the stimulus, the subject then chooses the appro- for choosing among two hypotheses based on
priate response to indicate the perceived direction independent, identically distributed data, where
of coherent motion in the stimulus. This is usually the observer can choose how much data to
done by eye movements in monkeys and either observe, is known as the sequential hypothesis
eye movements or key presses in humans. Typi- ratio test (SPRT). It is the optimal policy for
cally, there are two possible directions of motion, minimizing any cost function that is monotoni-
out of which the subject must choose, making this cally related to average reaction time and proba-
a 2-alternative forced choice (2AFC) task; occa- bility of error (Wald 1947; Wald and Wolfowitz
sionally, it has been done with more than two 1948). It states that the observer should accumu-
alternatives or in a multiple-alternative forced late evidence, in the form of Bayesian posterior
choice (mAFC) format (Churchland et al. 2008). probability using Bayes rule or, equivalently, by
When the 2AFC version is done in conjunction summing up log odds ratios, until the total evi-
with recordings of neurons that are sensitive to dence exceeds a confidence bound that favors one
the direction of motion of the stimulus, the two or the other stimulus; then, at that point, the
possible directions of motion are chosen to be observer should terminate the observation pro-
aligned with the preferred and anti-preferred cess and choose the more probable alternative.
directions of the neuron being recorded The height of the decision boundary is deter-
(Shadlen and Newsome 2001); in the mAFC ver- mined by the relative importance of speed and
sion, one of the possible motion directions is accuracy in the objective function, with greater
aligned with the preferred direction of the neuron importance of speed relative to accuracy leading
(Churchland et al. 2008). to lower decision boundaries, and vice versa for
There are two major variants of the random- decreasing importance of speed relative to accu-
dot coherent motion task, one in which the stim- racy. In the limit when many data samples are
ulus is present for a fixed duration (Newsome and observed before the decision is made, SPRT con-
Paré 1988) and one in which the stimulus is verges to the drift-diffusion process first studied
Decision-Making Tasks 933 D
in connection with statistical mechanics; it is Specifically, the NoGo response requires waiting
essentially a bounded linear dynamical system until the response deadline, while a Go response
with additive Wiener noise and absorbing bound- immediately terminates the current trial (Shenoy
aries (Laming 1968; Ratcliff and Rouder 1998; and Yu 2012). Using a Bayes-risk minimizing
Bogacz et al. 2006). Many mathematical proper- decision policy that minimizes not only error
ties of the drift-diffusion process are well known, rate but also average decision delay naturally
such as the average sample size before termina- exhibits the experimentally observed Go bias.
tion, and the probability of exceeding the correct The optimal decision policy is formally equiva-
boundary, for different problem parameters lent to a DDM with a time-varying threshold that D
and task conditions. It therefore provides initially rises after stimulus onset and collapses
a convenient mathematical framework for ana- again just before the response deadline. The ini-
lyzing SPRT and thus experimental behavior in tial rise in the threshold is due to the diminishing
different task settings. temporal advantage of choosing the fast Go
A popular alternative to the 2AFC paradigm is response compared to the fixed-delay NoGo
the Go/NoGo (GNG) task, in which one stimulus response.
requires a response (Go) and the other requires
the response to be withheld (NoGo). One appar-
ent advantage of the GNG paradigm is that it Outcome Uncertainty
obviates the need for response selection
(Donders 1969), thus helping to minimize any This class of decision-making tasks is designed
confounding influences of motor planning and such that subjects have no uncertainty about the
execution when the experimental focus is on per- stimulus identity, but rather about the conse-
ceptual and cognitive processing. However, quences of choosing one alternative over another.
within-subject comparison of reaction time To induce such uncertainty, the subjects are not
(RT) and error rates (ER) in 2AFC and Go/NoGo explicitly told about the reinforcement conse-
variants of the same perceptual decision-making quences of the different alternatives, but instead
task has found systematic biases between the two they have to learn them over time. The reinforce-
(Gomez et al. 2007). Specifically, the go stimulus ments are typically in the form of a reward, such
in the Go/NoGo task elicits shorter RT and more as money for human subjects, juice for monkeys,
false-alarm responses than in the 2AFC task, or seeds for birds; but occasionally they can also
when paired with the same opposing stimulus in be in the form of a penalty, such as money taken
both paradigms, resulting in a Go bias. This raises away for human subjects, a foot shock for rats, or
the question of whether the two paradigms really air puff for rabbits. This class of task originated in
probe the same underlying processes. the study of associative learning, which showed
Existing mechanistic models of these choice that subjects’ asymptotic choice performance
tasks, mostly variants of the drift-diffusion model after substantial learning exhibited interesting
(DDM; Ratcliff and Smith 2004; Gomez features. For example, Herrnstein (Herrnstein
et al. 2007) and the related leaky competing 1961) found that pigeons, when faced with
accumulator models (Usher and McClelland a choice between pecking two different buttons
2001; Bogacz et al. 2006), capture various in a Skinner box, did not always choose the one
aspects of behavioral performance, but do not with the higher reward rate, which is mathemat-
clarify the provenance of the Go bias in GNG. ically optimal, but rather alternated among the
Recently, it has been shown that the Go bias may choices stochastically so as to “match” their
arise as a strategic adjustment in response to the underlying reward rates. Based on these results,
implicit asymmetry in the cost structure of the Herrnstein formulated the “matching law” and
2AFC and GNG tasks and need not imply any the related “melioration theory” (Herrnstein
fundamental differences in the sensory and cog- 1970), which gave a procedural account of how
nitive processes engaged in the two tasks. matching-like behavior can arise from limited
working memory. A more recent account, using Multidimensional Decision-Making

Bayesian learning theory, shows that matching-
like behavior is, at a finer timescale, the conse- There are two main types of multidimensional
quence of a maximizing choice strategy (always decision-making: (1) target-distractor differenti-
choosing the best options), coupled with a learning ation, in which subjects must exclude the inter-
procedure that assumes the world to be fering influence of a distractor stimulus or
nonstationary and thus results in internal beliefs attribute, in order to focus on the target stimulus
that fluctuate with empirical experiences, even or attribute, and (2) multisensory integration, in
those driven by noise in a truly stationary environ- which subjects must integrate multiple stimuli,
ment (Yu and Cohen 2009; Yu and Huang 2014). often differing in sensory modalities, to reach
In order to induce continual outcome uncer- a combined choice response.
tainty, many implementations of this task incor-
porate unpredictable, unannounced changes in
the stimulus-outcome contingencies during the Target-Distractor Differentiation
experimental session. A classical example of
such a manipulation is reversal learning, in One classical task of this type is the Stroop task
which the choice with the better outcome sud- (Stroop 1935), in which subjects must ignore the
denly becomes the worse choice, while the pre- meaning of a color word and report the physical
viously bad choice now becomes the better color in which the text is written. Subjects con-
choice. This type of tasks has been used to sistently respond faster and more accurately
study, among other things, how perseverative when the meaning and physical attribute of the
tendencies might change as a function of experi- color word are congruent than when they are not.
mental condition or manipulations of the neural A related task is the Simon task, in which
circuitry in the brain. Such tasks are rich for a stimulus present on one side (e.g., visual stim-
theoretical modeling of the neural representation, ulus on the left side of the screen or auditory
computation, and utilization of different forms of stimulus to the left ear) may instruct the subjects
uncertainty that arise during learning and to respond with the other, less direct response
decision-making (Yu and Dayan 2005; Nassar (e.g., right button press), and again, performance
et al. 2010). For example, there is thought to be is better when the two dimensions are congruent
a differentiation of at least two kinds of uncer- than when they are not (Simon 1967). Yet another
tainties, “expected uncertainty,” dealing with similar task is the Eriksen task (Eriksen and
known uncertainties and variabilities in the envi- Eriksen 1974), which requires the subject to iden-
ronment, and “unexpected uncertainty,” arising tify a central stimulus (e.g., the letter “H” or “S”)
from dramatic, unexpected changes in the statis- when flanked on both sides by letters that are
tical contingencies in the environment. Experi- either congruent or incongruent with the central
mentally, human subjects have been shown to be stimulus; again, subjects are significantly better
more ready to alter choice strategy under condi- in the congruent condition. These tasks have been
tions of more frequent choice-outcome contin- used to discern how the brain identifies and filters
gencies compared to when these changes are out (or fails to do so) irrelevant distractors. Care-
less frequent (Behrens et al. 2007); the ful behavioral analysis of choice accuracy as
upregulation of the neuromodulator norepineph- a function of reaction time (Cho et al. 2002) has
rine has been shown to increase the rates for given clues as to the role of cognitive processes
learning such changes (Devauges and Sara such as attentional control and stimulated theo-
1990); pupil diameter in human subjects, under retical modeling of the neural computations giv-
the control of noradrenergic and cholinergic ing rise to specific features of the behavioral
neuromodulatory systems, has also been shown dynamics in such tasks (Yu et al. 2009).
to correspond to specific model-derived uncer- A related but distinct class of tasks involves
tainty signals (Nassar et al. 2012). a prepotent “go” signal present on each trial and
Decision-Making Tasks 935 D
an occasional “stop” signal on a small fraction of less noise (and greater reliability) in the com-
trials, whereby the subject must execute the “go” bined percept (Jacobs 1999; Ernst and Banks
response only on “go” trials and not on “stop” 2002; Battaglia et al. 2003; Dayan et al. 2000;
trials. In the stop-signal task (Logan and Cowan Shams et al. 2005). This explains, for instance,
1984), the “stop” signal appears, if at all, at an why vision typically dominates over auditory
unpredictable time after the “go” stimulus, and modality in spatial localization in
the greater the onset asynchrony, the more a phenomenon known as “visual capture” – this
unlikely that the subject is able to withhold the follows directly from the Bayesian formulation,
“go” response. A related task is what has been since vision has greater spatial acuity and reli- D
called a “compelled-response task,” in which the ability than audition (Battaglia et al. 2003). Con-
subject is instructed to go before being shown the versely, when the localization task is in the
cue that indicates which of the two responses is temporal domain instead of spatial, where audi-
correct (Salinas et al. 2010). These tasks have tion has greater temporal acuity and reliability,
given rise to competing theoretical models that auditory stimuli can induce an illusory visual
postulate either the existence (Logan and Cowan percept. For example, when a single visual flash
1984) or absence (Shenoy and Yu 2011; Salinas is accompanied by several auditory beeps, the
and Stanford 2013) of a specific stopping process visual percept is that of several flashes (Shams
or pathway and, in the latter class, the precise et al. 2000). Again this phenomenon can be
capacity to stop depending on either explained by a statistically optimal ideal observer
a normative, sequential consideration of the pros model (Shams et al. 2005; Kording et al. 2007).
and cons of responding (Shenoy and Yu 2011) or Recent neurophysiological studies have also
the efficiency of the perceptual process itself begun to elucidate the neural basis of multisen-
(Salinas and Stanford 2013). However, recent sory integration [see, e.g., Driver and Noesselt
experimental results indicate systematic and (2008) for a review].
rational changes in subjects’ stopping capacity
as a function of the reward structure of the task
(Leotti and Wager 2009), suggesting not only that Preference-Based Decision-Making
the “go” response is context sensitive and strate-
gically malleable but also that it takes into One area of decision-making is devoted to the
account perceptual uncertainty and decision- study of how humans make their choices based
theoretical factors such as reward contingencies on their own internal preferences, such as in
(Shenoy and Yu 2011). consumer decision-making, instead of based on
sensory features or behavioral outcomes associ-
ated with the options. When choosing among
Multisensory Integration options that differ along multiple attribute dimen-
sions, humans consistently exhibit certain puz-
Instead of investigating how the brain selectively zling preference shifts, or even reversals,
processes certain aspects of the environment and depending on the context. Three broad categories
excludes others, simultaneously present stimuli of contextual effects are studied in the psychol-
have also been employed in experimental tasks to ogy literature. In the attraction effect, given two
examine how the brain combines multiple similarly preferred options, A and B, the intro-
sources of sensory information, in particular duction of a third option Z that is similar to B, but
across sensory modalities. For instance, several also clearly less attractive than B, results in an
studies in recent years have shown that human increase in preference for B over A (Huber and
subjects combine differentially reliable sensory Payne 1982; Heath and Chatterjee 1995). In the
inputs from different modalities in a computa- compromise effect, when B > A in one attribute
tionally optimal (Bayesian) way, such that and B < A in another attribute and Z has the same
greater weight is assigned to sensory inputs with trade-off but is even more extreme than B, then
B becomes the “compromise” option and two-alternative forced choice tasks. Psychol Rev
becomes preferred relative to A (Simonson 113(4):700–765
Busemeyer JR, Townsend JT (1993) Decision field theory.
1989). In the similarity effect, the introduction Psychol Rev 100:432–459
of a third option Z that is very similar to B in Cho RY, Nystrom LE, Brown ET, Jone AD, Braver TS,
both attribute dimensions shifts the relative pref- Holmes PJ, Cohen JD (2002) Mechanisms underlying
erence away from B to A (Tversky n.d.). dependencies of performance on stimulus history in
a two-alternative forced choice task. Cogn Affect
Traditionally, there have been two lines of Behav Neurosci 2:283–299
explanations for such effects, the first attributing Churchland AK, Kiani R, Shadlen MN (2008) Decision-
them to biases or suboptimalities in human making with multiple alternatives. Nat Neurosci
decision-making (Kahneman and Tversky 1979; 11(6):693–702
Dayan P, Kakade S, Montague PR (2000) Learning and
Kahneman et al. 1982) and the other suggesting selective attention. Nat Rev Neurosci 3:1218–1223
that they are by-products of specific architectural Devauges V, Sara SJ (1990) Activation of the noradren-
or dynamical constraints on neural processing ergic system facilitates an attentional shift in the rat.
(Busemeyer and Townsend 1993; Usher and Behav Brain Res 39(1):19–28
Donders FC (1969) On the speed of mental processes.
McClelland 2004; Trueblood 2012). A more Acta Psychol (Amst) 30:412
recent, normative account (Shenoy and Yu Driver J, Noesselt T (2008) Multisensory interplay reveals
2013) uses a Bayesian model to demonstrate crossmodal influences on “sensory-specific” brain
that these contextual effects can arise as rational regions, neural responses, and judgments. Neuron
Rev 57:11–23
consequences of three basic assumptions: Eriksen BA, Eriksen CW (1974) Effects of noise letters
(1) humans make preferential choices based on upon the identification of a target letter in a nonsearch
relative values anchored with respect to “fair task. Percept Psychophys 16:143–149
market value,” which is inferred from both prior Ernst MO, Banks MS (2002) Humans integrate visual and
haptic information in a statistically optimal fashion.
experience and the current set of available Nature 415(6870):429–433
options; (2) different attributes are imperfect sub- Gold JI, Shadlen MN (2002) Banburismus and the brain:
stitutes for one another, so that one unit of decoding the relationship between sensory stimuli,
a scarce attribute is more valuable than one unit decisions, and reward. Neuron 36:299–308
Gomez P, Ratcliff R, Perea M (2007) A model of the
of an abundant one; and (3) uncertainty in beliefs go-nogo task. J Exp Psychol 136(3):389–413.
about “market conditions” induces stochasticity doi:10.1037/0096-3445.136.3.389.A
in relative preference on repeated encounters Heath T, Chatterjee S (1995) Asymmetric decoy effects on
with the same set of options. This model not lower-quality versus higher-quality brands: meta-
analytic and experimental evidence. J Consum Res
only provides a principled explanation for why 22:268–284
specific types of contextual modulation of pref- Herrnstein RJ (1961) Relative and absolute strength of
erence choice exist, but a means to infer individ- responses as a function of frequency of reinforcement.
ual and group preferences in novel contexts given J Exp Anal Behav 4:267–272
Herrnstein RJ (1970) On the law of effect. J Exp Anal
observed choices. Behav 13:243–266
Huber J, Payne J (1982) Adding asymmetrically domi-
nated alternatives: violations of regularity and the sim-
ilarity hypothesis. J Consum Res 9(1):90–98
References Jacobs RA (1999) Optimal integration of texture and
motion cues in depth. Vis Res 39:3621–3629
Battaglia PW, Jacobs RA, Aslin RN (2003) Bayesian inte- Kahneman D, Tversky A (1979) Prospect theory: an anal-
gration of visual and auditory signals for spatial local- ysis of decisions under risk. Econometrica 47:313–327
ization. J Opt Soc Am A Opt Image Sci Vis Kahneman D, Slovic P, Tversky A (eds) (1982) Judgement
20(7):1391–1397 under uncertainty: heuristics and biases. Cambridge
Behrens TEJ, Woolrich MW, Walton ME, Rushworth University Press, Cambridge, UK
MFS (2007) Learning the value of information in an Kording KP, Beierholm U, Ma W, Quartz S,
uncertain world. Nature Neurosci 10(9):1214–1221 Tenenbaum J, Shams L (2007) Causal inference in
Bogacz R, Brown E, Moehlis J, Hu P, Holmes P, Cohen JD cue combination. PLoS One 2(9):e943
(2006) The physics of optimal decision making: Laming DRJ (1968) Information theory of choice-reaction
a formal analysis of models of performance in times. Academic, London
Decision-Making, Bias 937 D
Leotti LA, Wager TD (2009) Motivational influences on Stroop JR (1935) Studies of interference in serial verbal
response inhibition measures. J Exp Psychol Hum reactions. J Exp Psychol Gen 18:643–662
Percept Perform 36(2):430–447 Trueblood JS (2012) Multialternative context effects
Logan G, Cowan W (1984) On the ability to inhibit obtained using an inference task. Psychon Bull Rev
thought and action: a theory of an act of control. 19(5):962–968. doi:10.3758/s13423-012-0288-9
Psych Rev 91:295–327 Tversky A (n.d.) Elimination by aspects: a theory of
Nassar MR, Wilson RC, Heasly B, Gold JI (2010) An choice. Psychol Rev 79:288–299
approximately Bayesian delta-rule model explains Usher M, McClelland JL (2001) The time course of per-
the dynamics of belief updating in a changing environ- ceptual choice: the leaky, competing accumulator
ment. J Neurosci 30(37):12366–12378 model. Psychol Rev 108(3):550–592
Nassar MR, Rumsey KM, Wilson RC, Parikh K, Heasly B, Usher M, McClelland J (2004) Loss aversion and inhibi- D
Gold JI (2012) Rational regulation of learning dynam- tion in dynamical models of multialternative choice.
ics by pupil-linked arousal systems. Nature Neurosci Psychol Rev 111(3):757–769
15(7):1040–1046 Wald A (1947) Sequential analysis. Wiley, New York
Newsome WT, Paré EB (1988) A selective impairment of Wald A, Wolfowitz J (1948) Optimal character of the
motion perception following lesions of the sequential probability ratio test. Ann Math Statist
middle temporal visual area (MT). J Neurosci 19:326–339
8(6):2201–2211 Yu AJ, Cohen JD (2009) Sequential effects: superstition
Ratcliff R, Rouder JN (1998) Modeling response times for or rational behavior? Adv Neural Inf Process Syst
two-choice decisions. Psychol Sci 9:347–356 21:1873–1880
Ratcliff R, Smith PL (2004) A comparison of sequential Yu AJ, Dayan P (2005) Uncertainty, neuromodulation,
sampling models for two-choice reaction time. and attention. Neuron 46:681–692
Psychol Rev 111:333–346 Yu AJ, Huang H (2014) Maximizing masquerading as
Roitman JD, Shadlen MN (2002) Response of neurons in matching in human visual search choice behavior.
the lateral intraparietal area during a combined visual Decision (To appear)
discrimination reaction time task. J Neurosci Yu AJ, Dayan P, Cohen JD (2009) Dynamics of atten-
22(21):9475–9489 tional selection under conflict: toward a rational
Salinas E, Stanford TR (2013) Waiting is the hardest part: Bayesian account. J Exp Psychol Hum Percept Per-
comparison of two computational strategies for form 35:700–717
performing a compelled-response task. Front Comput
Neurosci 33(13):5668–5685
Salinas E, Shankar S, Gabriela Costello M, Zhu D,
Stanford TR (2010) Waiting is the hardest part: com-
parison of two computational strategies for performing Decision-Making, Bias
a compelled-response task. Front Comput Neurosci.
doi:10.3389/fn-com.2010.00153
Marijn van Wingerden and Tobias Kalenscher
Shadlen MN, Newsome WT (2001) Neural basis of
a perceptual decision in the parietal cortex (area LIP) Department of Comparative Psychology,
of the rhesus monkey. J Neurophysiol 86:1916–1935 Institute of Experimental Psychology, Heinrich-
Shams L, Kamitani Y, Shimojo S (2000) What you see is Heine University D€usseldorf, D€usseldorf,
what you hear. Nature 408:788
Germany
Shams L, Ma WJ, Beierholm U (2005) Sound-induced
flash illusion as an optimal percept. Neuroreport
16(17):1923–1927
Shenoy P, Yu AJ (2011) Rational decision-making in Definition
inhibitory control. Front Hum Neurosci. doi:10.3389/
fnhum.2011.00048
Shenoy P, Yu AJ (2012) Strategic impatience in Go/NoGo Biases in decision making are identified as pat-
versus forced-choice decision-making. Adv Neural Inf terns of choice behavior that violate normative
Process Syst 25 theory of choice allocation.
Shenoy P, Yu AJ (2013) A rational account of contextual
effects in preference choice: what makes for a bargain?
In: Proceedings of the thirty-fifth annual conference of
the cognitive science society. Berlin, Germany Detailed Description
Simon JR (1967) Ear preference in a simple reaction-time
task. J Exp Psychol 75(1):49–55
According to the axioms frequently used in eco-
Simonson I (1989) Choice based on reasons: the case of
attraction and compromise effects. J Consum Res nomic theory, a rational decision maker is an
16:158–157 agent who makes consistent choices over time,
D 938 Decision-Making, Bias
and therefore exhibits stable preferences a constant rate. This means that, if a decision
(Samuelson 1938; Friedman and Savage 1948, maker reveals a preference for, say, an SS out-
1952). Importantly, the rationality assumption come delivered after delay t1 over an LL outcome
does not prescribe that the decision maker should delivered after delay t2, with t1 < t2, she should
always optimize her objective outcome, e.g., her still prefer the SS outcome if both outcomes are
monetary payoff, but it does prescribe that an deferred into the future by the same time interval
individual should behave consistently, given the Dt, because both outcomes would be discounted
assumption of stable preferences. This rationality at the same rate. However, in contrast to the
assumption has been applied to both human and assumption of stable intertemporal preferences,
animal behavior (Kacelnik 2006). However, both in practice, most human as well as animal deci-
in the human and animal literature, numerous sion makers exhibit preference reversals, typi-
examples of violations of economic rationality, cally changing their choice behavior from
i.e., biases in decision making, can be found preferring the SS to preferring the LL outcome
(Kalenscher and Van Wingerden 2011). when a common front-end delay (Dt) is added
(Rachlin and Green 1972; Ainslie 1974; Bateson
Present Bias and Time-Inconsistent and Kacelnik 1996; Isles et al. 2003; Kalenscher
Preferences et al. 2005; Kalenscher and Pennartz 2008; Louie
A ubiquitous finding in decision making studies and Glimcher 2010; Kalenscher and van
is that decision makers prefer immediate or Wingerden 2011). Assuming that the overt
briefly delayed outcomes over longer delayed choice behavior reflects an internal ranking of
outcomes (e.g., Ainslie 1974; Green et al. 1994; discounted utility, this implies that the time-
see Kalenscher and Pennartz 2008 for a review), discounted utility attached to the SS option has
even if the delayed outcomes clearly constitute sunk below that of the LL, invalidating constant
the better long-term choice (e.g., saving for discounting.
a pension plan vs. drinking away your money DUT is agnostic about whether a given SS
today). This myopic choice behavior is usually should be preferred over an LL alternative. In
termed ‘impulsivity’. Impulsive, shortsighted the context of animal foraging, however, optimal
intertemporal decision making violates several foraging theory (Charnov 1976; Stephens and
assumptions made in rational choice theories, Krebs 1986) prescribes long-term rate maximiza-
such as Discounted Utility Theory (DUT). DUT tion to increase Darwinian fitness (see “▶ Choice
is a normative account of decision-making over Behavior”). In self-control paradigms, paradoxi-
time, which predicts a rational decision maker’s cally, animals typically fail to optimize their
preference for immediate over delayed reward long-term rates. Imagine an SS option yields
(see “▶ Choice Behavior”). DUT states that reward G1 after some time t1 while an LL option
agents should maximize discounted utility when yields G2 after t2. The choice is presented to the
making intertemporal decision, where discounted animals after some inter-trial interval t. A long-
utility reflects the reduced current value incurred term rate maximizing decision rule (Stephens and
by options that are deferred into the future. For Krebs 1986)
example, when choosing between a smaller,
sooner (SS) or a larger, later (LL) outcome, an G2 G1
> (1)
agent’s discount function combined with the t þ t2 t þ t1 þ p
duration of the delays could lead a decision
maker to prefer SS over LL or vice versa, would predict that animals should always choose
depending on the delays and magnitudes the LL alternative G2 if its long-term rate exceeds
involved. Importantly, for a rational decision that of the SS option G1. Animals typically
maker, such preferences are assumed to be stable. exhibit present-bias, a failure to maximize long-
Specifically, one of the predictions of DUT is that term rates, by choosing the SS option even when
decision makers discount future rewards at a post-feeding delay (p) is added to it, making the
trial durations equivalent (Mazur and Logue a small gain G2 with high probability P2, and
1978; Green et al. 1981). the subject prefers the first, riskier option.
How can such time-inconsistent behavior be According to the independence axiom of EUT,
explained? Two-system models of decision mak- the preference should be the same when the same
ing posit that the utility premium put on immedi- probabilistic outcome is added to both options.
ate over delayed outcomes can be traced to So, if a gain G2 with probability (1 P2) is now
different (competing) neural systems that process added to both choices, both expected outcomes
immediate vs. delayed outcomes. McClure increase by the same amount, but the second one
et al. (2004, 2007) and others have shown that becomes certain, i.e., G2 with a probability of D
‘limbic’ regions such as the ventromedial pre- 1, and in this case decision makers typically pre-
frontal cortex and ventral striatum (Hariri fer the second option, reversing their original
et al. 2006) are preferentially activated by imme- preference. In other words, decision makers
diate outcomes, while other prefrontal structures have a strong preference for outcomes that are
such as lateral prefrontal cortex are engaged by certain over outcomes that are probabilistic, even
both immediate and delayed outcomes. The if the expected utility ranking is maintained.
inverse of impulsivity, self-control, can be A recent brain imaging study (Hsu et al. 2009)
invoked to combat present-bias in decision mak- building on prospect theory (Kahneman and
ing. Hare and colleagues (2009, 2011) have Tversky 1979; Tversky and Kahneman 1992)
shown that activity in dorsolateral prefrontal cor- indicates that neural activity in the striatum
tex is associated with self control, and that this when evaluating probabilistic outcomes is better
activity can modulate value signals in ventrome- explained by a non-linear (reverse-S shaped)
dial prefrontal cortex, downregulating the value function than a strictly linear function. In animals
of liked, but unwanted outcomes when self con- studies, results that resemble Allais-like certainty
trol succeeds. effects are observed in rats (MacDonald
et al. 1991) and honeybees (Shafir et al. 2008).
Certainty Effect in Decision Making
In decision making under risk, the normatively Sign/Reflection Effects in Decision Making
flavored Expected Utility Theory framework Economic theory usually assumes a monotoni-
(EUT, see Choice Behavior) posits that monetary cally increasing utility function linking objective
outcomes are nonlinearly transformed to subjec- quantities of a good to a subjective valuation as
tive value or utility. This transformation can the basis for preference judgments (Samuelson
explain the commonly found risk-aversion when 1937, 1938; von Neumann and Morgenstern
choosing between a certain and a probabilistic 1944). When outcomes are not certain, EUT pre-
outcome. EUT assumes that, while gains are scribes that the expected utility of a choice can be
non-linearly transformed, probabilities are not, calculated simply by weighing the utilities of the
so that each outcome’s expected utility is the possible outcomes of a decision by their proba-
linearly weighted sum of the probability- bilities and summing across these outcomes (see
weighted utilities of the possible outcomes. How- “▶ Choice Behavior”). According to EUT,
ever, behavioral data show that decision makers a decision maker should base her choice on the
overvalue certain outcomes. The Allais paradox final state of wealth, taking into account all
supports this notion by showing that adding (discounted) utilities of all outcomes, and it
a probabilistic outcome to two gambles in should not matter whether changes in good quan-
a choice set, so that one of the options now tities are accounted as gains or losses. However,
becomes certain instead of probabilistic, induces humans seem to treat gains differently to losses.
decision makers to reverse their preferences Because of the concave, decelerating nature of
(Allais 1953; Kahneman and Tversky 1979; Hsu the utility function, humans tend to be risk-averse
et al. 2009). So, suppose that the choices are when choosing between gambles with gains.
a large gain G1 with a low probability P1 or However, when the gambles are presented as
D 940 Decision-Making, Bias
potential losses rather than gains, risk attitudes high (loss frame), the birds are more prone towards
often change from risk-averse to risk-seeking risk-averse behavior (selecting the fixed outcome)
(Kahneman and Tversky 1979, 1984). This is or risk-seeking behavior (selecting the probabilis-
called the reflection effect, the finding that tic outcome), respectively (Marsh and Kacelnik
risk attitudes change when the prospects flip 2002).
around 0.
Further indications that losses are treated dif- Non-stationary Preferences: Intransitivity
ferently from gains come from research on loss A rational decision maker should exhibit stable
aversion (Kahneman and Tversky 1984; Tversky preferences. Formally, if a decision maker prefers
and Kahneman 1992). The idea that losses loom A over B and B over C, she should also prefer
larger than gains is illustrated by the fact that A over C (transitivity of preferences). However,
most participants reject a 50/50 gamble on human subjects often show context-dependent or
a coin toss or fair die. In fact, the gains need to intransitive choices (Tversky 1969; Kalenscher
be about twice as big to offset the potential loss. and Pennartz 2010; Kalenscher et al. 2010).
In the brain, several areas including the VStr and Intransitive preferences seem to occur readily
ventromedial prefrontal cortex (vmPFC) show when choice options differ on several value-
a pattern of ‘neural’ loss aversion – that is, the related attributes, such as probability and reward
slope of BOLD deactivations to losses is steeper magnitude, and could stem from an attribute-wise
than activations to gains, when mixed gambles of comparison instead of an alternative-wise inte-
potential losses and gains are evaluated (Tom gration of choice attributes. For instance, consis-
et al. 2007). In animals, loss aversion is also tent with the additive difference model put
observed (Chen et al. 2006; but see Silberberg forward by Tversky (1969), Kalenscher
et al. 2008). et al. (2010) showed that participants choosing
between two gambles tend to weigh differences
Framing Effect in Decision Making in probabilities and reward magnitudes differen-
The differential treatment of losses and gains is tially. So, a large difference in probabilities (with
even more strikingly evident in the framing effect. a moderate difference in reward magnitude)
Kahneman and Tversky (1984) showed that when resulted in a strong weighting of this attribute,
a decision has to be made between a certain and and correspondingly risk-averse choices,
a probabilistic outcome on human lives saved whereas a large difference in reward magnitude
(gain frame) versus lost (loss frame), with the (with a moderate difference in probability)
exact same numerical outcomes, the gain frame resulted in a strong weighting of magnitude,
induces risk aversion whereas the loss frame together with risk-seeking choices. This
induces risk seeking choice behavior in context-dependent shift from risk-aversion to
participants. This shows that the wording risk-seeking, depending on the difference in the
used to present a decision problem is sufficient to gambles’ attributes, rendered the potential for
bias a subject between treating identical outcomes circular, intransitive preference structures. The
as gains or losses, and affecting risk attitude neural data by Kalenscher et al. support separate
accordingly. There is some evidence that tracking of the context-dependent weight of out-
decisions in the direction of the framing effect, as come probabilities (posterior cingulate cortex)
opposed to those decisions in opposition to the and magnitudes (posterior insula).
choice biases, elicit more activity in the
amygdala (De Martino et al. 2006; but see Talmi Sunk Cost Effect in Decision Making
et al. 2010), and that resilience against the The sunk cost effect is the tendency to persist in
framing bias is associated with activity in an endeavor once an investment of effort, time, or
OFC. In animals, starlings show behavior in line money has been made (Thaler 1980). A famous
with a framing effect. When the baseline expecta- example entails the continuation of the Concorde
tion of a food outcome is either low (gain frame) or supersonic airplane project, even when
profitability analyses indicated that the project References
would never make money (Arkes and Ayton
1999). Besides policy makers, participants in Ainslie G (1974) Impulse control in pigeons. J Exp Anal
Behav 21:485–489
the lab also exhibit the sunk cost effect. Navarro
Allais M (1953) Le Comportement de l’Homme Rationnel
and Fantino (2005) and others have shown that devant le Risque: Critique des Postulats et Axiomes de
human participants persist in a choice option l’Ecole Americaine. Econometrica 21:503
searching for payout beyond the moment when Arkes HR, Ayton P (1999) The sunk cost and concorde
effects: are humans less rational than lower animals?
switching, abandoning or resetting would be opti-
Psychol Bull 125:591–600 D
mal (Kogut 1990). In animals, sunk costs have Bateson M, Kacelnik A (1996) Preferences for fixed and
been studied in pigeons. Like humans, pigeons variable food sources: variability in amount and delay.
generally persisted longer on a probabilistic J Exp Anal Behav 63:313–329
Charnov EL (1976) Optimal foraging: the marginal value
choice than prescribed by optimal choice, and
theorem. Theor Popul Biol 9:129–136
this persistence was modulated by the size of Chen MK, Lakshminarayanan V, Santos LR (2006) How
the sunk costs (Navarro and Fantino 2005; basic are behavioral biases? Evidence from capuchin
Pattison et al. 2012). monkey trading behavior. J Polit Econ 114:517–537
De Martino B, Kumaran D, Seymour B, Dolan RJ
(2006) Frames, biases, and rational decision-making
Prosocial Biases in Decision Making in the human brain. Science 313:684–687
One of the tenets of microeconomic theory is that Engel C (2011) Dictator games: a meta study. Exp Econ
decisions should be made from a purely self- 14:583–610
Fehr E, Fischbacher U (2003) The nature of human altru-
interested perspective, and that other people’s
ism. Nature 425:785–791
outcomes should be disregarded. Obviously, this Friedman M, Savage LJ (1948) The utility analysis of
notion does not conform to reality. Human deci- choices involving risk. J Polit Econ 56:279
sion makers experience consistent prosocial Friedman M, Savage LJ (1952) The expected-utility
hypothesis and the measurability of utility. J Polit
biases that become manifest in costly helping
Econ 60:463–474
behavior, e.g., when an agent reduces her own Green L, Fisher EB, Perlow S, Sherman L (1981) Prefer-
outcome in favor of another’s. The most straight- ence reversal and self-control: choice as a function of
forward example of this is the Dictator Game reward amount and delay. Behav Anal Lett 1:43–51
Green L, Fristoe N, Myerson J (1994) Temporal
(DG; Fehr and Fischbacher 2003; Engel 2011),
discounting and preference reversals in choice
where a participant can split a sum of money between delayed outcomes. Psychon Bull Rev
between herself and a powerless partner. The 1:383–389
partner has to accept, and cannot retaliate. From Hare TA, Camerer CF, Rangel A (2009) Self-control in
decision-making involves modulation of the vmPFC
a game theoretical perspective, the dictator
valuation system. Science 324:646–648
should keep all the money for herself. This is Hare TA, Malmaud J, Rangel A (2011) Focusing attention
not the case, people across all cultures and on the health aspects of foods changes value signals in
socio-economic groups consistently give sums vmPFC and improves dietary choice. J Neurosci
31:11077–11087
larger than 0, even in anonymous, one-shot
Hariri AR, Brown SM, Williamson DE, Flory JD, De
games where reputation building or other strate- Wit H, Manuck SB (2006) Preference for immediate
gic motives can be ruled out (Fehr and over delayed rewards is associated with magnitude of
Fischbacher 2003; Engel 2011). This suggests ventral striatal activity. J Neurosci 26:13213–13217
Hsu M, Krajbich I, Zhao C, Camerer CF (2009) Neural
that decision makers are presumably not the self-
response to reward anticipation under risk is nonlinear
ish, own-payoff maximizers postulated by eco- in probabilities. J Neurosci 29:2231–2237
nomic theory, but do take the well-being of others Isles AR, Humby T, Wilkinson LS (2003) Measuring
into consideration. impulsivity in mice using a novel operant delayed
reinforcement task: effects of behavioural manipula-
tions and d-amphetamine. Psychopharmacology (Berl)
170:376–382
Cross-References Kacelnik A (2006) Meanings of rationality. In: Nudds M,
Hurley S (eds) Rational animals? Oxford University
Press, Oxford, pp 87–106
▶ Choice Behavior
D 942 Decision-Making, Models
Kahneman D, Tversky A (1979) Prospect theory: an anal- Silberberg A, Roma PG, Huntsberry ME, Warren-Boulton
ysis of decision under risk. Econometrica 47:263 FR, Sakagami T, Ruggiero AM, Suomi SJ (2008) On
Kahneman D, Tversky A (1984) Choices, values, and loss aversion in capuchin monkeys. J Exp Anal Behav
frames. Am Psychol 39:341–350 89:145–155
Kalenscher T, Pennartz CM (2008) Is a bird in the hand Stephens DW, Krebs JR (1986) Foraging theory, Monographs
worth two in the future? The neuroeconomics of in behavior and ecology. University Press, Princeton
intertemporal decision-making. Prog Neurobiol Talmi D, Hurlemann R, Patin A, Dolan RJ (2010) Framing
84:284–315 effect following bilateral amygdala lesion. Neuropsy-
Kalenscher T, Pennartz C (2010) Do intransitive choices chologia 48:1823–1827
reflect genuinely context-dependent preferences. In: Thaler R (1980) Toward a positive theory of consumer
Delgado MR, Phelps E, Robbins T (eds) Attention choice. J Econ Behav Organ 1:39–60
and performance XIII: decision making. Oxford Uni- Tom SM, Fox CR, Trepel C, Poldrack RA (2007) The
versity Press, Vermont neural basis of loss aversion in decision-making under
Kalenscher T, Van Wingerden M (2011) Why we should risk. Science 315:515–518
use animals to study economic decision Tversky A (1969) Intransitivity of preferences. Psychol
making – a perspective. Front Neurosci 5:82 Rev 76:31–48
Kalenscher T, G€unt€urk€ un O, Calabrese P, Gehlen W, Tversky A, Kahneman D (1992) Advances in prospect
Kalt T, Diekamp B (2005) Neural correlates of theory: cumulative representation of uncertainty.
a default response in a delayed go/no-go task. J Exp J Risk Uncertain 5:297–323
Anal Behav 84:521–535 Von Neumann J, Morgenstern O (1944) Theory of games
Kalenscher T, Tobler PN, Huijbers W, Daselaar SM, and economic behavior. University Press, Princeton
Pennartz CMA (2010) Neural signatures of intransitive
preferences. Front Hum Neurosci 4:49. doi:10.3389/
fnhum.2010.00049
Kogut CA (1990) Consumer search behavior and sunk
costs. J Econ Behav Organ 14:381–392
Decision-Making, Models
Louie K, Glimcher PW (2010) Separating value from
choice: delay discounting activity in the lateral Paul Miller
intraparietal area. J Neurosci 30:5498–5507 Department of Biology and Brandeis University,
MacDonald DNH, Kagel JH, Battalio RC (1991) Animals’
Waltham, MA, USA
choices over uncertain outcomes, further experimental
results. Econ J 101:1065–1084 Volen National Center for Complex Systems,
Marsh B, Kacelnik A (2002) Framing effects and risky Waltham, MA, USA
decisions in starlings. Proc Natl Acad Sci U S A
99:3352–3355
Mazur JE, Logue AW (1978) Choice in a “self-control”
paradigm: effects of a fading procedure. J Exp Anal Definition
Behav 30:11–17
McClure SM, Laibson DI, Loewenstein G, Cohen JD Models of decision making attempt to describe,
(2004) Separate neural systems value immediate and
using stochastic differential equations which rep-
delayed monetary rewards. Science 306:503–507
McClure SM, Ericson KM, Laibson DI, Loewenstein G, resent either neural activity or more abstract psy-
Cohen JD (2007) Time discounting for primary chological variables, the dynamical process that
rewards. J Neurosci 27:5796–5804 produces a commitment to a single action/out-
Navarro AD, Fantino E (2005) The sunk cost effect in
come as a result of incoming evidence that can
pigeons and humans. J Exp Anal Behav 83:1–13
Pattison KF, Zentall TR, Watanabe S (2012) Sunk cost: be ambiguous as to the action it supports.
pigeons (Columba livia), too, show bias to complete
a task rather than shift to another. J Comp Psychol
126:1–9
Rachlin H, Green L (1972) Commitment, choice and self
control. J Exp Anal Behav 17:15–22
Samuelson PA (1937) A note on measurement of utility. Background
Rev Econ Stud 4:155–161 Decision making can be separated into four pro-
Samuelson PA (1938) A note on the pure theory of con-
sumer’s behaviour. Economica 5:61–71
cesses (Doya 2008):
Shafir S, Reich T, Tsur E, Erev I, Lotem A (2008) Percep- 1. Acquisition of sensory information to deter-
tual accuracy and conflicting effects of certainty on mine the state of the environment and the
risk-taking behaviour. Nature 453:917–920 organism within it
Decision-Making, Models 943 D
2. Evaluation of potential actions (options) in
terms of the cost and benefit to the organism
given its belief about the current state
3. Selection of an action based on, ideally, an
optimal trade-off between the costs and
benefits
4. Use of the outcome of the action to update the
costs and benefits associated with it
Models of the dynamics of decision making D
have focused on perceptual decisions with only
two possible responses available. The term Decision-Making, Models, Fig. 1 Scheme of the
two-alternative forced-choice (TAFC) task. Two streams
two-alternative forced choice (TAFC) applies to of sensory input, each containing stimulus information, or
such tasks when two stimuli are provided, but the a signal (S1 and S2) combined with noise (s1Z1(t) and
term is now generally used for any binary choice s2Z2(t)), are compared in a decision-making circuit. The
discrimination task. circuit must produce one of two responses (A or B) indi-
cating which of the two signals is stronger. The optimal
In a perceptual decision, the response, or method for achieving this discrimination is via the sequen-
action, is directly determined by the current per- tial probability ratio test (SPRT) which requires the
cept. Thus, the decision in these tasks is essen- decision-making circuit to integrate inputs over time
tially one of perceptual categorization, namely,
process (1) above, though the same models can be have provided the dominant paradigm for analy-
used for action selection given ambiguous infor- sis of choice behavior, the restriction to only two
mation of the current state (process 3). choices is lifted in many of the more recent
Evaluation of the possible responses in terms models of decision making based on multiple
of their value or the resulting state’s utility variables, allowing for the fitting of a wider
(process 2) (Sugrue et al. 2005) given both uncer- range of data sets.
tainty in the current state, and uncertainty in the The tasks can be based on either a free-
outcomes of an action given the state, is the response paradigm, in which a subject responds
subject of expected utility theory and prospect after as much or little time as she wants, or an
theory. interrogation (forced response) paradigm, in
The necessary learning and updating of the which the stimulus duration is limited and the
values of different actions given the actual out- subject must make a response within a given
comes they produce (process 4) are the subject of time interval. The free-response paradigm is per-
instrumental conditioning and reinforcement haps more powerful, since each trial produces
learning, for example, via temporal difference two types of information: accuracy (correct or
learning (Seymour et al. 2004) and actor-critic incorrect) and response time. However, by varia-
models (Joel et al. 2002). tion of the time allowed when responses are
This entry is primarily concerned with the forced, both paradigms are valuable for
dynamics of the production of either a single constraining models, since they can provide
percept given unreliable sensory evidence a distribution of response times for both correct
(1) or a single action given uncertainty in the and incorrect trials, as well as the proportion of
outcomes (3). trials that are correct or incorrect with a given
stimulus. These behavioral data can be modified
General Features of Discrimination Tasks by task difficulty, task instructions (such as
or TAFC Tasks “respond rapidly” versus “respond accurately”),
In a TAFC task, a single decision variable can be or reward schedules and intertrial intervals.
defined representing the likelihood ratio – the Most models of the dynamics of decision mak-
probability that evidence to date favors one altering focus on tasks where the time from stimulus
native over the other. While TAFC tasks (Fig. 1) onset to response is no more than one to two
seconds, a timescale over which neural spiking these tasks. All of the models to be discussed
can be maintained. Choices requiring much more below can produce such a speed-accuracy trade-
time than this are likely to depend upon multiple off by parameter adjustment. If parameters are
memory stores, neural circuits, and strategies, adjusted so as to increase the mean response time,
which become difficult to identify, extract, and then accuracy increases. Such a trade-off is
model in a dynamical systems framework observed in behavioral tasks under two regimes:
(a state-based framework is more appropriate). one, when either instructions or the schedule of
In the standard setup of the models, two par- reward and punishment encourages participants
allel streams of noisy sensory input are available, to respond as quickly as possible while being less
with each stream supplying evidence in support concerned about making errors or two, when sub-
of one of the two allowed actions (see Fig. 1). The jects respond as accurately as possible while
sensory inputs can be of either discrete or contin- being less concerned about the time it takes to
uous quantities and can arrive discretely or con- decide. The simplest way to effect such a trade-
tinuously in time. The majority of models focus off is to adjust the intertrial interval, which if long
on continuous update in continuous time so they compared to the decision time means that accu-
can be formulated as stochastic differential equa- racy of responses impacts reward rate much more
tions (Gillespie 1992; Lawler 2006). The sensory so than the time for the decision itself. Models
evidence, which is momentary, produces can replicate such behavior when optimal perfor-
a decision variable, which indicates the likeli- mance is based on the maximal reward rate. Typ-
hood of choosing one of the two alternatives ical parameter adjustments to increase accuracy
given current evidence and all prior evidence. while slowing responses would be a multiplica-
The primary difference between models is in tive scaling down of inputs (and the concurrent
how sensory evidence determines the decision input noise) or a scaling up of the range across
variable. While most models incorporate a form which the decision variable can vary by raising
of temporal integration of evidence (Cain and a decision threshold (Figs. 2 and 3) (Ratcliff
Shea-Brown 2012) and include a negative inter- 2002; Simen et al. 2009; Balci et al. 2011).
action between the two sources of evidence, dif- A similar effect can be achieved in alternative,
ferences arise in the stability of initial states attractor-based models through the level of a
which determines whether integration is perfect global applied current, which affects the stability
and in the nature of the interaction: feedforward of the initial “undecided” state (Figs. 6–9) (Miller
between the inputs, feedforward between out- and Katz 2013).
puts, or feedback from outputs to decision vari- From a neuroscience perspective, the decision
ables (Bogacz et al. 2006). Models can also differ variable is typically interpreted as either the mean
in their choice of decision threshold – the value of firing rate of a group of neurons or a linear com-
the decision variable at which a response is bination of rates of many neurons (Beck
produced – in the free-response paradigm et al. 2008), the difference between two groups
(Simen et al. 2009; Deneve 2012; Drugowitsch being the simplest such combination. There has
et al. 2012) and in particular whether this param- been remarkable progress in matching the
eter or other model parameters, such as input observed firing patterns of neurons (Newsome
gain, which also affect the response time distri- et al. 1989; Shadlen and Newsome 2001; Huk
bution, are static or dynamical across a trial and Shadlen 2005) with the dynamics of
(Shea-Brown et al. 2008; Thura et al. 2012). a decision variable in more mathematical models
As the time available for acquisition of sen- of decision making (Glimcher 2001, 2003; Gold
sory information increases, so does the accuracy and Shadlen 2001, 2007; Smith and Ratcliff
of responses in a perceptual discrimination task. 2004; Ratcliff et al. 2007). This has led to the
Accuracy is measured as probability of choosing introduction of biophysically based models of
the response leading to more reward, which is neural circuits (Wang 2008), which have
equivalent to obtaining a veridical percept in accounted for much of the concordance between
simple mathematical models, neural activity, and accuracy. In these tasks, the sequential probabil-
behavior. ity ratio test (SPRT), introduced by Wald and
Wolfowitz (Wald 1947; Wald and Wolfowitz
Optimal Decision Making 1948), and, in its continuous form, the drift-
An optimal decision-making strategy either max- diffusion model (DDM) (Ratcliff and Smith
imizes expected reward over a given time or 2004; Ratcliff and McKoon; 2008) lead to opti-
minimizes risk. In TAFC perceptual tasks, mal choice behavior by any of these measures of
a response is either correct or an error. In the optimality (see (Bogacz et al. 2006) for
interrogation paradigm, with fixed time per deci- a thorough review). D
sion, the optimal strategy is the one leading to Using SPRT in the interrogation paradigm,
greatest accuracy, that is, the lowest expected one simply accumulates over time the
error rate. In the free-response paradigm, the log-likelihood ratio of the probabilities of each
optimal strategy either delivers the greatest accu- alternative given the stream of evidence, where
racy for a given mean response time or produces the observed sensory input per unit time has
the fastest mean response time for a given a certain probability given alternative A and
another probability given alternative B. Integrat-
ing the log-likelihood over time, after setting the
initial condition as the log-likelihood ratio of the
prior probabilities, log[P(A)/P(B)], leads to
a quantity log[P(A|S)/P(B|S)] which is greater
than zero if A is more likely than B given the
stimulus and less than zero otherwise. Thus, from
standard Bayesian theory, the optimal procedure
is to choose A or B depending on the sign of the
summed, or in the continuous limit, integrated,
log-likelihood ratio.
Decision-Making, Models, Fig. 2 The drift-diffusion
model (DDM). The DDM is a one-dimensional model,
In the free-response paradigm, a stopping cri-
so the two competing inputs and their noise terms are terion must be included. This is achieved by
first combined: in this case S = S1 S2 and setting two thresholds for the integrated
s2 = s12 + s22 log-likelihood ratio, a positive one (+a) for
Decision-Making,
Models, Fig. 3 The drift-
diffusion model is a Wiener
process (one-dimensional
Brownian motion) with
absorbing boundaries. In
the absence of the
boundaries, the probability
distribution is Gaussian,
centered at a distance St
from its starting point with
variance increasing as s2t
choice A and a negative one (b) for choice responses showed the models to be inaccurate in
B. The further the thresholds are from the origin, this regard – observed response time distributions
the lower the chance of error, but the longer the are skewed with a long tail, whereas the response
integration time before reaching a decision. Thus, times of accumulator models were much more
the thresholds reflect the fraction of errors that symmetric about the mean. These discrepancies
can be tolerated, with a ¼ log 1a
b
and b ¼ log 1b
a
led to the ascendance of Ratcliff’s drift-diffusion
where a is the probability of choosing A when model (Ratcliff 1978).
B is correct and b is the probability of choosing
B when A is correct. The Drift-Diffusion Model
The drift-diffusion model (DDM) is an integrator
The Models with thresholds (Fig. 2), or more precisely, the
Accumulator Models decision variable, x, follows a Wiener process
The first models of decision making in humans or with two absorbing boundaries (Fig. 3). It
animals were accumulator models, sometimes includes a deterministic (drift) term, S, propor-
called counter models or race models. In these tional to the rate of incoming evidence and
models, evidence accumulates separately for a diffusive noise term of variance s2, which pro-
each possible outcome. This has the advantage duces variability in response times and can lead
that if many outcomes are possible, the models to errors:
are simply extended by addition of one more
variable for each additional alternative, with evi- dx
¼ S þ sðtÞ
dence for each alternative accumulating within dt
its allotted variable. In the interrogation para-
digm, one simply reads out the highest variable, where (t) is a white noise term defined by
so the choice depends on the sign of the differ- h(t)(t0 )i = d(t t0 ).
ence of the two variables in the TAFC paradigm. If the model is scaled to a given level of noise,
Thus, if the difference in accumulated quantities then its three independent parameters are drift
matched the difference in integrated log proba- rate (S) and positions of each of the two thresh-
bilities of the two types of evidence, such readout olds (a, b) with respect to the starting point.
from an accumulator model would be equivalent When the model was introduced, these parame-
to an SPRT, so would be optimal. ters were assumed fixed for a given subject in
In the free-response paradigm, accumulator a specific task. The threshold spacing determines
models produce a choice when any one of the where one operates in the speed-accuracy trade-
accumulated variables reaches a threshold, so off, so it can be optimized as a function of the
these models can be called “race to threshold relative cost for making an incorrect response and
models” or simply “race models.” The original the time between trials. Any starting point away
accumulator models included neither interaction from the midpoint represents bias or prior infor-
between accumulators nor ability for variables to mation. The drift rate is proportional to stimulus
decrease. However, for decisions in nature or in strength.
laboratory protocols, evidence in favor of one With fixed parameters, which could be fitted to
alternative is typically evidence against the any subject’s responses, the DDM reproduces
other alternative. This is particularly problematic key features of the behavioral data: notably the
in the free-response paradigm, because the time skewed shape of response time distributions and
at which one variable reaches threshold and pro- the covariation of mean response times and
duces the corresponding choice is independent of response accuracy with task difficulty. Skewed
evidence accumulated for other choices. Thus, response time distributions arise because the var-
the behavior of simple accumulator models is iance of a Wiener process increases linearly with
not optimal. Comparisons of response time dis- time – responses much earlier than the mean
tributions of these models with behavioral response time, when the variance in the decision
variable is low, are less likely than responses two-stage models, where the first stage sets the
much later, when the variance in the decision starting point from the values of the two choices,
variable is high. A more difficult perceptual before a second stage of integration toward
choice is represented by a drift rate closer to threshold commences. Such a model is supported
zero, which increases response times and by electrophysiological data (Rorie et al. 2010).
increases the probability of error. Such covaria-
tion of response times with accuracy matches The Leaky Competing Accumulator Model
behavioral data well, so long as an additional The leaky competing accumulator (LCA), intro-
processing time is added to the model – the addi- D
duced by Usher and McClelland (2001), was
tional time representing a variable sensory trans- suggested to be in better accord with neural data
duction delay on the input side and a motor delay and to better fit some behavioral data than the
on the output side, both of which contribute to DDM. The LCA is a two-variable model, with
response times in addition to the processing each variable integrating evidence in support of
within any decision circuit. one of the two alternatives in a TAFC task. The
While the original DDM included trial-to-trial model includes a “leak” term, as decay back to
variability through the diffusive noise term, addi- baseline for each individual variable in the
tional trial-to-trial variability in the parameters absence of incoming evidence. The “competi-
was needed to account for differences between tion” in LCA is a cross-inhibition between the
the response time distributions of correct trials two variables, thus improving upon original
from those of incorrect trials. With fixed param- accumulator models in allowing evidence for
eters and no initial bias, the DDM produces iden- one variable contributing to a reduction in the
tical distributions (though with different other variable (Fig. 4).
magnitudes) for the timing of correct responses
and errors. However, response times of human dX1 pffiffiffi
subjects are typically slower when they produce t ¼ S1 kX1 bX2 þ s1 t1 ðtÞ
dt
errors, unless they are instructed to respond as dX2 pffiffiffi
quickly as possible, in which case the reverse is t ¼ S2 kX2 bX1 þ s2 t2 ðtÞ
dt
true. These behaviors are accounted for in the
DDM by including trial-to-trial variability in the
drift rate and/or the starting point. Trial-to-trial The difference in the two LCA variables, Xd =
variability in drift rate leads to slower errors, as X1 X2, follows an Ornstein-Uhlenbeck
errors become more likely on those trials when process:
the drift rate is altered from its mean toward zero
and mean responses are longer. Trial-to-trial var- dXd pffiffiffi
t ¼ Sd kXd þ bXd þ sd tðtÞ
iability in the starting point leads to faster errors, dt
as errors become more likely on those trials in pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
which the starting point is closer to the error where Sd = S1 S2 and sd ¼ s21 þ s22 . It
boundary, in which case the response is faster. should be noted that the DDM is retrieved as
Error response times are reduced more by such a special case of the LCA, if the coefficient of
variability than are correct response times since the term linear in Xd, that is, b k, is set to zero.
the correct responses include more of the trials If the only criterion for choosing the best
started at the midpoint where mean times are model were its match to behavioral data, one
longer. could use Akaike information criterion or Bayes-
Prior information or bias can be incorporated ian information criterion to assess whether inclu-
in the drift-diffusion model, either through a shift sion of a nonzero term in the LCA is justified by
in the starting point of integration or an additional the better fit to data so produced. As well as
bias current to the inputs to be integrated. A shift ability to fit human response times, the ability
in starting point is more optimal and has led to for a model to match neural activity and to be
response times and thus produced an inherently

dynamical model; memory retrieval was treated
as a set of parallel DDMs, each representing an
individual item with a “match” and “non-match”
threshold to indicate its recognition or not. The
rate of evidence accumulation in each individual
DDM was set in terms of the similarity between
a current item and one in memory.
Models such as those of Hopfield respond to
Decision-Making, Models, Fig. 4 The leaky competing
the match between a current item and memories
accumulator (LCA) model. Two separate integration proof previously encoded items, which are contained
cesses for the two separate stimuli produce two decision within the network as attractor states. The net-
variables, X1 and X2. A “leak” term proportional to work’s activity reaches a stable attractor state
k causes a decay of the decision variable through self-
inhibition, while a cross-inhibition term proportional to b
more rapidly if the match is close. The temporal
produces competition. In the interrogation paradigm, dynamics of memory retrieval or pattern comple-
a decision is made according to the greater of X1 or X2 at tion, which comprises the decision process, is not
the response time, while in the free-response paradigm, addressed carefully in neural network models, in
a choice is made when one decision variable first reaches
its threshold (given as +a and +b, respectively)
which neural units can be binary and time can be
discretized, since this is not their goal. However,
the use of attractor states to represent the outcome
generated robustly in a neural circuit should be of a decision or a perceptual categorization has
considered when assessing its value and rele- achieved success in biophysical models based on
vance. One achievement of the LCA is to repro- spiking neurons (Wang 2008), albeit with far
duce an initial increase in both variables, before simpler attractor states than those of neural
the competition between variables causes one networks.
variable to be suppressed, while the other vari-
able accelerates its rise. Such behavior matches Biophysical Models
electrophysiological data recorded in monkeys The LCA model (Usher and McClelland 2001),
during perceptual choices – in particular, firing being motivated by neurophysiology, is similar in
rates of neurons representing the response not spirit to the more detailed biophysical models
chosen increase upon stimulus onset before they that followed it. In particular, the first model of
are suppressed (cf. trajectories in Fig. 6). decision making based on spiking neurons
assumes two competing integrators, where
Neural Network Models and Attractor States integration is produced through tuned recurrent
Some of the first models of perceptual categori- excitation and competition is the result of cross-
zation, which have had significant impact on inhibition between the pools of neurons (see
neuroscience, were Hopfield networks, Cohen- Fig. 5). However, when even the most elementary
Grossberg neural networks, and Willshaw properties of spiking neurons are taken into
networks. While these models are primarily account, a few additional complications arise
aimed at formation and storage of memories, the (Wong and Wang 2006).
retrieval of a memory via corrupted information First, neurons emit spikes as a point process in
is identical to a perceptual decision. time, so that even at a constant firing rate they
A correspondence between memory retrieval supply a variable current to other cells. The var-
and decision making should not be surprising, iability in the current can be reduced with an
since Ratcliff’s introduction of the DDM – the increase in the number of cells in a group, so
archetype of decision-making models – was long as the spike times are uncorrelated – that
within a paper entitled “A theory of memory is, cells are firing asynchronously. Asynchrony is
retrieval.” Ratcliff was focused on fitting most easily achieved when neurons spike
irregularly, a feature produced by adding addi- excitation to inhibit each other, at a minimum,
tional noise to each neuron in the decision- a third group of inhibitory cells must be added.
making circuit. So, in accord with most likely The need for an extra cell group to mediate the
realizations of a decision-making circuit in vivo, inhibition adds a small delay to the cross-
biophysical models introduce additional noise inhibition compared to self-excitation, though
into the decision-making model itself, which this effect can be counteracted with fast responses
inevitably adds to any already-present stimulus in the synaptic connections to and from inhibitory
noise (Wang 2002). cells versus slower synaptic responses in excit-
Second, neurons are either excitatory or inhib- atory connections. The second effect of an inter- D
itory, so in order for two groups of cells with self- mediate cell group is to add the nonlinearity of
the inhibitory interneuron’s response function
into the cross-inhibition. Thus, the inhibitory
input to one excitatory cell group is not a linear
function of the firing rate of the other cell group.
The consequences of this and other nonlinearities
are discussed further below.
Third, biophysical models of neurons, just like
neurons in vivo, respond nonlinearly to increased
inputs. Similarly, synaptic inputs to a cell satu-
rate, so are a nonlinear function of presynaptic
firing rates. In the LCA model (Usher and
Decision-Making, Models, Fig. 5 A neural circuit
model of decision making. Integration of stimuli (S1, S2) McClelland 2001), both the neural response and
by groups of neurons with rates r1 and r2 can be achieved the feedback are linear functions passing through
through strong excitatory recurrent connections within the origin, so via the tuning of one variable, the
groups (looped connections with arrows). The mean firing
two curves can match each other and produce an
rate of cells in the inhibitory pool (“INHIB”) increases
with both r1 and r2, producing competition through inhib- integrator. Integrators require such a matching of
itory input to both cell groups (solid circles) (Wang 2002) synaptic feedback to neural response so they can
Decision-Making, Models, Fig. 6 Nonlinearity of neu- unstable, so that by I1 = I2 = 10, the only stable fixed
ral response functions produces stable fixed points, toward points correspond to one choice or the other and a decision
which neural activity evolves. Nullclines are represented is forced. With high enough applied input current,
by the solid thick red/green lines, the green line where I1 = I2 = 15, a new symmetric stable state appears at
dr2/dt = 0 at fixed r1 and the red line where dr1/dt = 0 at high firing rates, but this plays no role in the decision-
fixed r2. Crossings of the lines are fixed points of the making process. Deterministic trajectories from a range of
system, with stable fixed points denoted by solid black starting points are indicated by the thin colored lines that
circles. Decision states are stable fixed points with terminate at a fixed point. See Table 1 for an xppaut code
r2 >> r1 or r1 >> r2. As a symmetric input current is that produces the nullclines and Table 2 for the Matlab
added to the network, the spontaneous “undecided” state, code that produces the trajectories
that is, the fixed point of low rates with r1 = r2, becomes
Decision-Making, Models, Table 1 A code base on xppaut to produce nullclines and fixed points for two interacting
groups of neurons, connected in a decision-making circuit, as used to produce Figs. 6, 7, 8, and 9
# Rate-based model of two coupled neuron populations.
# I1 and I2 are the inputs to each population, which should be varied.

par I1=5
par I2=5.5
# Strength of self-excitation
par wself=1
# Cross connections are net inhibitory
par wcross=-0.2
# tau is the time constant -- the slowest synaptic time constant, of NMDA is
used.
# tau has no effect on steady state solutions.
par tau=0.075
# Sigmoidal f-I curves require three parameters, these are:

# rmax, the maximum rate
par rmax=100
# Ithresh, current to produce half-maximum rate
par Ithresh=50
# delta, determines the width of the sigmoid (sensitivity near Ithresh).
par delta=20
# Initial firing rates can be varied.

init r1=0
init r2=0
dr1/dt=-r1/tau+rmax/(1+exp(-(I1+wself*r1+wcross*r2-Ithresh)/delta))/tau
dr2/dt=-r2/tau+rmax/(1+exp(-(I2+wself*r2+wcross*r1-Ithresh)/delta))/tau
# The following are default values for the start-up of xpp.
@method=rk,total=2,bound=500,dt=.005,dtmin=1e-12,atoler=1e-7
@toler=1e-5,xhi=2,yhi=100,ylo=0 njmp=5
@ bell=off,nout=50
done
retain a stable firing rate in the absence of intersect at no more than three points (Fig. 6),
input – the firing rate produced by a given synap- leading to the possibility of two discrete stable
tic input must be exactly that needed to generate attractor states for a group of cells (with an unsta-
the same synaptic input – and this must be true for ble fixed point in between). When two such
a wide range of firing rates. Such matching of groups are coupled, such as by cross-inhibition,
synaptic feedback to neural response produces the network can have at most four stable states,
a state of marginal stability, typically called given by the combinations of low and high firing
a line attractor or continuous attractor. rates for the two groups. In the winner-takes-all
However, in the absence of symmetry or model of decision making, three of these states
a remarkable similarity in the shape of neural can be stable in the absence of input: the state
response curves and synaptic feedback curves, when both cell groups have low or spontaneous
the nonlinear curves of biophysical neurons activity and the two states with just one of the cell
Decision-Making, Models, Fig. 7 Stochastic noise in Center left: with moderate applied current, noise can
the simulation produces trial-to-trial variability in the induce transitions to one of the two “decision states.”
neural responses. With the same neural circuit of Fig. 6, Center and far right: with increased applied current, tra-
addition of noise can cause neural activity to end up in jectories always end up at a decision state – even with
different states, even with the same starting points. Small I1 = I2 = 15, the two decision states are more stable than
colored dots indicate trajectories, with black solid circles the symmetric high-rate state (far right). See Table 1 for
denoting end points after 5 s. All simulations begin with an xppaut code that produces the nullclines and Table 2 for
r1 = r2 = 0.1 Hz. Far left: with no applied current, neural the Matlab code that produces the trajectories
activity is maintained near the low-rate spontaneous state.
groups possessing high activity. The fixed point of each individual choice alternative before stim-
with both groups possessing high activity is ulus onset, if prior probability impacts the
unstable. In the presence of input, the symmetric starting point of integration. Models in which
low-activity state becomes unstable, and only the total amount of inhibition depends on the
two stable states remain: the decision states with total number of alternative choices (see Fig. 10)
one group active (“the winner”) and the other reproduce such behavior. One consequence of the
group inactive (“the loser”). Importantly, with increased inhibition is a slowing of decision times
a combination of slow synaptic time constants as the number of alternatives increases.
(NMDA receptors with a time constant of
50–100 ms are an essential ingredient of the Sequential Discrimination, Context-Dependent
model) and sufficient fine-tuning of parameters Decisions, and Prior Information
of the network, the time course for the network’s The most developed dynamical models of deci-
activity to shift from the unstable initial state to sion making pertain to the identification of an
one of the two remaining stable attractor states is incoming sensory stimulus or the comparison of
slow enough to match neural and behavioral two or more concurrent stimuli. However, many
response times. decisions require a comparison of successive
stimuli, and even when two stimuli are concur-
Extension of Models to Multiple Choices rent, our attention typically switches from one to
Many decision-making models for TAFC are the other when making a perceptual choice. Thus,
simply extended to the case of multiple alterna- models of decision making have been developed
tives (Bogacz et al. 2007b; Furman and Wang in which the stimuli are separated in time, so
2008; Niwa and Ditterich 2008; Ditterich 2010). a form of short-term memory is required, with
Electrophysiological data suggests that neural fir- the contents of short-term memory and a later
ing rates reach a threshold independent of number sensory input both affecting the choice of action
of alternatives, but neurons receive greater inhi- (Romo and Salinas 2003; Machens et al. 2005;
bition, as revealed by reduced firing rates in their Miller and Wang 2006). The process of making
initial, spontaneous activity state (Churchland a decision by combining short-term memory,
et al. 2008; Churchland and Ditterich 2012). which can represent the current “context”
This is akin to a reduction in the prior probability (Salinas 2004), with sensory input provides the
Decision-Making, Models, Fig. 8 A bias in the inputs with r1 > r2 includes the line r1 = r2. Right: with added
causes neural activity to favor one decision state over the noise, some trials with a symmetric starting point termi-
other, though noise means that “errors” can arise. Left: in nate in the state with high r2 – these correspond to “errors”
the deterministic system, more trajectories terminate at in the standard terminology of decision making. See
the attractor point of high r1, because group 1 receives Table 1 for an xppaut code that produces the nullclines
higher input current. In particular, a symmetric initial and Table 2 for the Matlab code that produces the
condition (r1 = r2) results in termination with high r1 trajectories
and low r2. That is, the basin of attraction for the state
essential ingredient for working memory tasks to an optimal, time-varying cost function
and for model-based strategies of action selection (Drugowitsch et al. 2012).
(Deco and Rolls 2005).
Prior information can make one response Testing Decision-Making Models
either a more likely alternative or a more reward- Decision-making models can be tested more
ing alternative given ambiguous sensory infor- stringently using tasks in which the stimulus is
mation and can lead to across-trial dependences not held constant across the time allotted for
in decision-making behavior. Consideration of producing a response (Zhou et al. 2009; Stanford
how much to weigh prior information compared et al. 2010; Shankar et al. 2011; R€uter et al. 2012).
to the current stimulus requires a separate choice For example, in models such as the DDM based
(Hanks et al. 2011), which establishes how long on perfect integration, if a stimulus is altered or
one should continue to acquire sensory input. even reversed for a short amount of time, so long
Such a choice is akin to setting a decision-making as the stimulus alteration is in the period of its
threshold. Factors affecting the optimal period integration, it has the same effect on response
for obtaining sensory input include the relative time and choice probability whether it is early
clarity of incoming information compared to the or late. However, in models where the initial state
strength of the prior (Deneve 2012), as well as the is unstable, a late altered stimulus has weaker
intrinsic cost of a reduced reward rate when tak- impact on the decision-making dynamics than
ing more time to decide (Drugowitsch an early one. Conversely, in models where the
et al. 2012). At some point in time, awaiting initial state is stable, such as the LCA with
further sensory evidence does not improve one’s a positive leak term, only stimuli presented
probability of a correct choice sufficiently to shortly before the final response contribute to it;
warrant any extra delay of a potential reward. the time constant of the drift term that draws the
Solution by dynamical programming (Bellman decision variables back to the initial state corre-
1957) of a model that takes into account these sponds to a time constant for the forgetting of
factors suggests that monkeys respond according earlier evidence.
Decision-Making, Models, Fig. 9 In a noisy system, “undecided” responses switch to the decision state favored
the initial state can remain deterministically stable, but by the input bias (corresponding to “correct” responses),
responses terminate in one of the decision states, with the while a few terminate in the other decision state (“incor-
bias favoring one final state over the other. Left: in the rect” responses) and one remains in the symmetric
absence of noise and a bias in the inputs, more trajectories low-rate state (an “undecided” response). See Table 1
evolve to the fixed point favored by the input, but many for an xppaut code that produces the nullclines and Table 2
terminate at the “undecided” state. Right: with added for the Matlab code that produces the trajectories
noise and symmetric initial conditions, most otherwise
Alternatively, if one sets up a task in which the relevant tasks, the difficulty is in unraveling
noise in the stimulus, controlled by the experi- the cause of a sensory input, which has been
menter, is the dominant noise source in the degraded either at source, or through sensory
decision-making process, one can analyze sepa- processing, or as a result of imperfections of
rately correct trials and error trials and align them memory encoding and recall. The requisite action
by either time of stimulus onset or time of given a sensory percept is either via
response to assess the impact of noise fluctuations a straightforward instruction for human subjects
on choice probability or response times. Care or produced by weeks to months of training in
must be taken when aligning by response times, nonhuman animal subjects. Thus, in the post-
since threshold crossings are inevitably produced training stage used to acquire data, the step from
by noise fluctuations in the direction of the percept to action can be considered very fast and
threshold crossed. However, in all cases, model independent of the parameters varied by the
predictions can be tested with experimental mea- experimenter to modify task difficulty.
surements. Current results appear to be task However, most decisions require us to select
dependent, as some data sets suggest all evidence a course of action given a percept or given
is equally impactful (supporting a perfect integra- a combination of percepts. Two general strate-
tor) (Huk and Shadlen 2005; Brunton et al. 2013), gies, termed model based or model-free (Dayan
while others suggest the weighing of sensory and Daw 2008; Dayan and Niv 2008), are possi-
evidence is higher early (Ludwig and Davies ble for action selection. Model-based strategies
2011) or higher late (Cisek et al. 2009; Thura require an evaluation of all possible conse-
et al. 2012) or oscillatory (Wyart et al. 2012) quences, with their likelihood, similar in a chess
across stimulus duration. game to calculating all the combinations of
moves in response to one move, or, conversely,
Beyond Discrimination: Action Selection all the possible causes of an observation.
In the models considered heretofore, the decision A model-free system simply learns the value of
of what action to take has been equivalent to the a given state and selects an action based on the
question of what is perceived. This is because in immediately reachable state with highest value.
Decision-Making, Models, Table 2 The Matlab function used to simulate multiple decision-making trials as plotted
in each panel of Figs. 6, 7, 8, and 9
function[ r1 r2 ] = WTAtrials( Iapp1,Iapp2,sigma );

% WTAtrials produces multiple trials of a winner-takes-all network produced
% by coupling two sigmoidal firing rate models, each representing a neural
% population.
% Iapp1 and Iapp2 are the respective current inputs to the 2 populations.
% sigma is the level of noise (added independently to each population).
% The function returns the firing rate as a function of time for each
% trial for each population.
% Number of trials is defined within the function as Ntrials = 10.
dt = 0.001; % timestep for simulation

tmax = 5.0; % max time for integration
tvec = 0:dt:tmax; % time vector
Ntrials = 10;
r1 = zeros(Ntrials, length(tvec)); % records rate of one cell group

r2 = zeros(Ntrials, length(tvec)); % records rate of other cell group
tau = 0.05; % time constant (cf NMDA receptors)
rmax = 100; % max rate of cells

Ithresh = 50; % input needed for 1/2-max firing
Idelta = 20; % determines steepness of sigmoidal f-I curve
wself = 1; % recurrent excitatory feedback

wcross = -0.2; % strength of cross-inhibition
trial = 0;
rhighstart = 0*rmax; % If nonzero, provides a range of starting points.
for trial = 1:Ntrials
if( trial <= Ntrials/2 )
r1init = 0.1+rhighstart*(Ntrials/2-trial)/(Ntrials/2-1);
r2init = 0.1;
else
r2init = 0.1+rhighstart*(Ntrials-trial)/(Ntrials/2-1);
r1init = 0.1;
end
% Implement initial conditions

r1(trial,1) = r1init;
r2(trial,1) = r2init;
% Produce independent random noise for each population

noise1 = sigma*sqrt(dt/tau)*randn(size(tvec));
noise2 = sigma*sqrt(dt/tau)*randn(size(tvec));
% Now integrate with basic Euler-Maruyama method using sigmoid f-I

% curves and current, I, linear in rate.
for i = 2:length(tvec)
r1(trial,i) = r1(trial,i-1) + ...
dt/tau*( rmax/(1+exp(-(Iapp1+wself*r1(trial,i-1)+ ...
wcross*r2(trial,i-1)-Ithresh)/Idelta))-r1(trial,i-1) )+ ...
noise1(i);
r2(trial,i) = r2(trial,i-1) + ...
dt/tau*(rmax/(1+exp(-(Iapp2+wself*r2(trial,i-1)+ ...
wcross*r1(trial,i-1)-Ithresh)/Idelta))-r2(trial,i-1) )+ ...
noise2(i);
end
end
end
Decision-Making, Models, Fig. 10 Models for deci- the sum of multiple decision variables. Right: extension of
sion making with multiple alternatives. Left: extension of the biophysical model, in which cells providing inhibitory
the leaky competing accumulator model to multiple alter- feedback are activated by the summed activity of multiple
natives, such that the negative feedback is proportional to groups of excitatory cells
For example, one could move a chess piece to Cross-References

produce the pattern of pieces that has led to most
games won in the past. ▶ Decision-Making Tasks
A wide literature in the field of reinforcement ▶ Decision-Making, Bias
learning (Barto 1994; Redish et al. 2007) and its ▶ Decision-Making, Motor Planning
possible neural underpinnings (Daw and Doya ▶ Decision-Making, Threshold
2006; Johnson et al. 2007; Lee and Seo 2007) ▶ Decision-Making: Overview
addresses how one can learn the value of states ▶ Perceptual Decision-Making
over multiple trials. This literature has influenced ▶ Reinforcement Learning in Cortical Networks
model-free biophysical models of decision mak- ▶ Speed-Accuracy Tradeoff
ing (Soltani and Wang 2008, 2010), with the
principal requirement being a need for enhanced
Hebbian learning when the resulting action leads References
to positive reinforcement, but not otherwise. The
neural activity underlying the decision precedes Balci F, Simen P, Niyogi R, Saxe A, Hughes JA,
the reinforcement signal, so in order for the Holmes P, Cohen JD (2011) Acquisition of decision
appropriate synapses to be potentiated when pos- making criteria: reward rate ultimately beats accuracy.
Atten Percept Psychophys 73:640–657
itive reinforcement arrives, authors have Barto AG (1994) Reinforcement learning control. Curr
suggested the activity itself remains in Opin Neurobiol 4:888–893
a persistent state of high firing rate (Soltani and Barto AG, Mahadevan S (2003) Recent advances in hier-
Wang 2006) or a molecular “eligibility trace” of archical reinforcement learning. Discrete Event Dyn
Syst Theory Appl 13:343–379
earlier activity persists through the time of the Beck JM, Ma WJ, Kiani R, Hanks T, Churchland AK,
reinforcement signal (Bogacz et al. 2007a; Roitman J, Shadlen MN, Latham PE, Pouget A (2008)
Izhikevich 2007). Probabilistic population codes for Bayesian decision
A model-based method for making a decision making. Neuron 60:1142–1152
Bellman R (1957) Dynamic programming. Princeton Uni-
is more flexible and can be more accurate, but in versity Press, Princeton
all but the simplest cases, the combinatorial Bogacz R, Brown E, Moehlis J, Holmes P, Cohen JD
explosion of alternatives becomes unwieldy and (2006) The physics of optimal decision making:
time-consuming to calculate. Hierarchical rein- a formal analysis of models of performance in
two-alternative forced-choice tasks. Psychol Rev
forcement learning renders such models more 113:700–765
manageable (Barto and Mahadevan 2003; Ito Bogacz R, McClure SM, Li J, Cohen JD, Montague PR
and Doya 2011; Botvinick 2012). (2007a) Short-term memory traces for action bias in
human reinforcement learning. Brain Res Gold JI, Shadlen MN (2001) Neural computations that
1153:111–121 underlie decisions about sensory stimuli. Trends
Bogacz R, Usher M, Zhang J, McClelland JL (2007b) Cogn Sci 5:10–16
Extending a biologically inspired model of choice: Gold JI, Shadlen MN (2007) The neural basis of decision
multi-alternatives, nonlinearity and value-based making. Annu Rev Neurosci 30:535–574
multidimensional choice. Philos Trans R Soc Lond Hanks TD, Mazurek ME, Kiani R, Hopp E, Shadlen MN
B Biol Sci 362:1655–1670 (2011) Elapsed decision time affects the weighting of
Botvinick MM (2012) Hierarchical reinforcement learn- prior probability in a perceptual decision task.
ing and decision making. Curr Opin Neurobiol J Neurosci Off J Soc Neurosci 31:6339–6352
22:956–962 Huk AC, Shadlen MN (2005) Neural activity in macaque
Brunton BW, Botvinick MM, Brody CD (2013) Rats and parietal cortex reflects temporal integration of
humans can optimally accumulate evidence for visual motion signals during perceptual decision
decision-making. Science 340:95–98 making. J Neurosci Off J Soc Neurosci
Cain N, Shea-Brown E (2012) Computational models of 25:10420–10436
decision making: integration, stability, and noise. Curr Ito M, Doya K (2011) Multiple representations and algo-
Opin Neurobiol 22:1047–1053 rithms for reinforcement learning in the cortico-basal
Churchland AK, Ditterich J (2012) New advances in ganglia circuit. Curr Opin Neurobiol 21:368–373
understanding decisions among multiple alternatives. Izhikevich EM (2007) Solving the distal reward problem
Curr Opin Neurobiol 22:920–926 through linkage of STDP and dopamine signaling.
Churchland AK, Kiani R, Shadlen MN (2008) Decision- Cereb Cortex 17:2443–2452
making with multiple alternatives. Nat Neurosci Joel D, Niv Y, Ruppin E (2002) Actor-critic models of the
11:693–702 basal ganglia: new anatomical and computational per-
Cisek P, Puskas GA, El-Murr S (2009) Decisions in spectives. Neural Netw Off J Int Neural Netw Soc
changing conditions: the urgency-gating model. 15:535–547
J Neurosci 29:11560–11571 Johnson A, van der Meer MA, Redish AD (2007) Integrat-
Daw ND, Doya K (2006) The computational neurobiology ing hippocampus and striatum in decision-making.
of learning and reward. Curr Opin Neurobiol Curr Opin Neurobiol 17:692–697
16:199–204 Lawler GF (2006) Introduction to stochastic processes.
Dayan P, Daw ND (2008) Decision theory, reinforcement Chapman & Hall/CRC, Boca Raton
learning, and the brain. Cogn Affect Behav Neurosci Lee D, Seo H (2007) Mechanisms of reinforcement learn-
8:429–453 ing and decision making in the primate dorsolateral
Dayan P, Niv Y (2008) Reinforcement learning: the good, prefrontal cortex. Ann N Y Acad Sci 1104:108–122
the bad and the ugly. Curr Opin Neurobiol 18:185–196 Ludwig CJ, Davies JR (2011) Estimating the growth of
Deco G, Rolls ET (2005) Attention, short-term memory, internal evidence guiding perceptual decisions. Cogn
and action selection: a unifying theory. Prog Neurobiol Psychol 63:61–92
76:236–256 Machens CK, Romo R, Brody CD (2005) Flexible control
Deneve S (2012) Making decisions with unknown sensory of mutual inhibition: a neural model of two-interval
reliability. Front Neurosci 6:75 discrimination. Science 307:1121–1124
Ditterich J (2010) A comparison between mechanisms of Miller P, Katz DB (2013) Accuracy and response-time
multi-alternative perceptual decision making: ability distributions for decision-making: linear perfect inte-
to explain human behavior, predictions for neurophys- grators versus nonlinear attractor-based neural cir-
iology, and relationship with decision theory. Front cuits. J Comput Neurosci 35:261–294
Neurosci 4:184 Miller P, Wang XJ (2006) Discrimination of temporally
Doya K (2008) Modulators of decision making. Nat separated stimuli by integral feedback control. Proc
Neurosci 11:410–416 Natl Acad Sci U S A 103:201–206
Drugowitsch J, Moreno-Bote R, Churchland AK, Shadlen Newsome WT, Britten KH, Movshon JA (1989) Neuronal
MN, Pouget A (2012) The cost of accumulating evi- correlates of a perceptual decision. Nature 341:52–54
dence in perceptual decision making. J Neurosci Off Niwa M, Ditterich J (2008) Perceptual decisions between
J Soc Neurosci 32:3612–3628 multiple directions of visual motion. J Neurosci Off
Furman M, Wang XJ (2008) Similarity effect and optimal J Soc Neurosci 28:4435–4445
control of multiple-choice decision making. Neuron Ratcliff R (1978) A theory of memory retrieval. Psychol
60:1153–1168 Rev 85:59–108
Gillespie DT (1992) Markov processes: an introduction Ratcliff R (2002) A diffusion model account of response
for physical scientists. Academic, San Diego time and accuracy in a brightness discrimination task:
Glimcher PW (2001) Making choices: the neurophysiol- fitting real data and failing to fit fake but plausible data.
ogy of visual-saccadic decision making. Trends Psychon Bull Rev 9:278–291
Neurosci 24:654–659 Ratcliff R, Hasegawa YT, Hasegawa RP, Smith PL,
Glimcher PW (2003) The neurobiology of visual-saccadic Segraves MA (2007) Dual diffusion model for single-
decision making. Annu Rev Neurosci 26:133–179 cell recording data from the superior colliculus in
Decision-Making, Motor Planning 957 D
a brightness-discrimination task. J Neurophysiol Sugrue LP, Corrado GS, Newsome WT (2005) Choosing
97:1756–1774 the greater of two goods: neural currencies for valua-
Ratcliff R, McKoon G (2008) The diffusion decision tion and decision making. Nat Rev Neurosci
model: theory and data for two-choice decision tasks. 6:363–375
Neural Comput 20:873–922 Thura D, Beauregard-Racine J, Fradet CW, Cisek
Ratcliff R, Smith PL (2004) A comparison of sequential P (2012) Decision making by urgency gating: theory
sampling models for two-choice reaction time. and experimental support. J Neurophysiol
Psychol Rev 111:333–367 108:2912–2930
Redish AD, Jensen S, Johnson A, Kurth-Nelson Z (2007) Usher M, McClelland JL (2001) The time course of per-
Reconciling reinforcement learning models with ceptual choice: the leaky, competing accumulator
behavioral extinction and renewal: implications for model. Psychol Rev 108:550–592 D
addiction, relapse, and problem gambling. Psychol Wald A (1947) Sequential analysis. Wiley, New York
Rev 114:784–805 Wald A, Wolfowitz J (1948) Optimum character of the
Romo R, Salinas E (2003) Flutter discrimination: neural sequential probability ratio test. Ann Math Stat
codes, perception, memory and decision making. Nat 19:326–339
Rev Neurosci 4:203–218 Wang XJ (2002) Probabilistic decision making by slow
Rorie AE, Gao J, McClelland JL, Newsome WT reverberation in cortical circuits. Neuron 36:955–968
(2010) Integration of sensory and reward information Wang XJ (2008) Decision making in recurrent neuronal
during perceptual decision-making in lateral circuits. Neuron 60:215–234
intraparietal cortex (LIP) of the macaque monkey. Wong KF, Wang XJ (2006) A recurrent network mecha-
PLoS ONE 5:e9308 nism of time integration in perceptual decisions.
R€uter J, Marcille N, Sprekeler H, Gerstner W, Herzog MH J Neurosci Off J Soc Neurosci 26:1314–1328
(2012) Paradoxical evidence integration in rapid deci- Wyart V, de Gardelle V, Scholl J, Summerfield C (2012)
sion processes. PLoS Comput Biol 8:e1002382 Rhythmic fluctuations in evidence accumulation dur-
Salinas E (2004) Fast remapping of sensory stimuli onto ing decision making in the human brain. Neuron
motor actions on the basis of contextual modulation. 76:847–858
J Neurosci 24:1113–1118 Zhou X, Wong-Lin K, Philip H (2009) Time-varying
Seymour B, O’Doherty JP, Dayan P, Koltzenburg M, perturbations can distinguish among integrate-to-
Jones AK, Dolan RJ, Friston KJ, Frackowiak RS threshold models for perceptual decision making in
(2004) Temporal difference models describe higher- reaction time tasks. Neural Comput 21:2336–2362
order learning in humans. Nature 429:664–667
Shadlen MN, Newsome WT (2001) Neural basis of
a perceptual decision in the parietal cortex (area LIP)
of the rhesus monkey. J Neurophysiol 86:1916–1936
Shankar S, Massoglia DP, Zhu D, Costello MG, Stanford Decision-Making, Motor Planning
TR, Salinas E (2011) Tracking the temporal evolution
of a perceptual judgment using a compelled-response
Terrence R. Stanford
task. J Neurosci Off J Soc Neurosci 31:8406–8421
Shea-Brown E, Gilzenrat MS, Cohen JD (2008) Optimiza- Department of Neurobiology & Anatomy,
tion of decision making in multilayer networks: the Wake Forest School of Medicine,
role of locus coeruleus. Neural Comput 20:2863–2894 Winston-Salem, NC, USA
Simen P, Contreras D, Buck C, Hu P, Holmes P, Cohen JD
(2009) Reward rate optimization in two-alternative deci-
sion making: empirical tests of theoretical predictions.
J Exp Psychol Hum Percept Perform 35:1865–1897 Synonyms
Smith PL, Ratcliff R (2004) Psychology and neurobiology
of simple decisions. Trends Neurosci 27:161–168
Action planning; Action selection; Motor inten-
Soltani A, Wang XJ (2006) A biophysically-based neural
model of matching law behavior: melioration by sto- tion; Motor preparation; Motor set; Preparatory
chastic synapses. J Neurosci 26:3731–3744 set; Response planning; Response selection
Soltani A, Wang XJ (2008) From biophysics to cognition:
reward-dependent adaptive choice behavior. Curr
Opin Neurobiol 18:209–216
Soltani A, Wang XJ (2010) Synaptic computation under- Definition
lying probabilistic inference. Nat Neurosci
13:112–119 “Motor planning” refers to the cognitive opera-
Stanford TR, Shankar S, Massoglia DP, Costello MG,
tions that are antecedent to the execution of
Salinas E (2010) Perceptual decision making in less
than 30 milliseconds. Nat Neurosci 13:379–385 a voluntary, goal-directed motor act or sequence.
D 958 Decision-Making, Motor Planning
Detailed Description formulation of a motor plan as a problem of

neural computation. With this in mind, it is
As the term implies, motor planning refers to the a simple matter to consider how the motor plans
mental operations that precede the initiation of for the aforementioned hypothetical scenarios
a voluntary motor act. Imagine, for example, the must differ on the basis of presumed input and
simple act of reaching to pick up a glass of wine desired output. Clearly, a motor plan developed
while dining at a restaurant. If you are like most to take a taste of wine is guided by very different
people, you have just envisioned yourself making sources of information than that for stabilizing
a deliberate reach to pick up the glass, presum- the unsteady glass. In the former case, the deci-
ably for the eventual purpose of taking a sip. sion, and consequently the plan, to reach for the
Now, instead imagine that this same glass of glass is informed primarily by a host of appetitive
wine is teetering due to an inadvertent bump of and social cues, while in the latter, visual
the table by a passing waiter. It is readily apparent (swaying motion of the glass) and tactile
that your reaching movement in this circum- (vibration of the table) motion cues would likely
stance, though requiring the same basic excur- predominate. In addition to receiving input from
sion, would be substantively different in both its vastly disparate sources, the ensuing motor
intended purpose and, accordingly, in the manner plans must evolve to produce outputs that opti-
in which it is executed. Intuitively, the most mize different movement parameters. Ideally,
likely distinction between a reach intended to a reach to drink the wine would be graceful, or
take a sip of wine and one designed to prevent it at least sufficiently precise, so as to avoid
from spilling is that the latter would need to be knocking over the salt shaker or dipping one’s
executed much more rapidly if it is to have any sleeve in the olive oil intended for dipping
chance of producing a successful outcome. From bread. In contrast, grace may be sacrificed in the
the perspective of neural computation, the differ- name of urgency when attempting to prevent
ence between the two movements would at first a spill. Here, in addition to moving with greater
glance seem rather trivial – reaching faster sim- speed, a more direct hand path to the glass (salt
ply requires that a “stronger” neural command be shaker be damned!) might be more effective.
provided to the relevant musculature – but, in It goes without saying that if the respective
fact, deeper consideration of what would consti- motor plans are to specify movements having
tute an effective motor plan in each case reveals different speeds and trajectories, their
that the problem is not at all simple from a neural corresponding neural correlates must carry infor-
coding perspective. mation that can be translated into the appropri-
As is generically true for any “plan,” a motor ately distinct dynamical signals at the effector
plan is a prescription for action that is informed level (e.g., force, joint torque).
by current circumstances and with an eye toward Along with providing a general sense for what
an expected, future outcome. If considered as a motor plan is, the juxtaposition of these real-
a series of mental events, the “plan” then is world reaching scenarios makes a crucial point:
what occurs after one has decided what to do motor planning, and the neural that implements
(e.g., take a sip of wine) but before one actually it, represents a flexible interface that ties internal
does it; by definition, a plan can be canceled, thus and external cues to motor effectors in a way that
resulting in no action. As detailed in another entry ensures that actions are appropriate to context.
in this section (see “▶ Perceptual Decision-Mak- Once again, if motor planning could be thought of
ing”), this depiction of mental events as discrete as a discrete computational stage, it is one that
and serial is unlikely to have a direct correspon- must be receptive to all manner of possible input
dence in neural processing stages (see also and capable of producing a widely varied range
Requin et al. 1988; Cisek and Kalaska 2010), of output – a formidable computational problem
but it is of heuristic value for considering the to be sure.
Experimental Investigation fraction of neurons display two temporally dis-
The preceding section provides a highly descrip- crete periods of activation, one linked to the ini-
tive account of motor planning, making use of tial stimulus indicating the movement goal
a concrete example to generate some intuition (sensory event) and another to the motor response
about its implications for behavior and the neural (saccade or reach). Interposed between these
activity that underlies behavior. In the laboratory, event-linked bursts of activity is the so-called
neurophysiological studies of motor planning delay period activity, a major focus of neurophys-
have a long history, and our understanding has iologists’ attempts to understand how motor plan-
relied extensively on behavioral tasks that incor- ning is represented in the activity of single D
porate what is known as an “instructed delay neurons. Consistent with the logic of the
period,” sometimes simply referred to as the instructed delay task, activity during this period
“delay period” or “instructed delay.” (The term demonstrates the hallmark features of that which
“delay period” in this context should not be con- would constitute a plan for action. First, such
fused with its use to refer to a memory retention activity reflects the vector of the upcoming move-
interval (e.g., see Funahashi et al. 1989). The ment well in advance of the imperative signal to
delay period is the interval of time that elapses actually produce it (i.e., the go signal). And sec-
between two task instructions. The first instruc- ond, as mentioned earlier, such delay period rep-
tion comes in the form of a sensory cue informing resentations of future movement need not lead to
the subject about what to do and the second (the execution – plans can be canceled – as has been
“go signal”) instructs the subject when to do it. In convincingly demonstrated with behavioral tasks
principle, the delay period is a window of time that later “countermand” the initial instruction of
during which the subject can perform the neces- “what to do.” That the specification of movement
sary perceptual evaluation of the instruction and parameters in neural activity does not inevitably
subsequently formulate aspects of the motor plan lead to the initiation of movement firmly supports
(e.g., select the appropriate response, specify the interpretation of such premovement activity
metrics), all in advance of receiving the instruc- as a motor plan and not a motor command.
tion to execute the action. Crucially, the time
allowed for advance preparation contrasts with Specificity of Motor Plans
standard reaction time tasks, in which the go The question of precisely what information is
signal and the informative sensory cue are simul- being represented in motor planning activity can
taneous, so perceptual judgment and motor plan- be a difficult one, particularly so for the
ning processes are all completed within the skeletomotor system where it is not entirely
reaction time interval. The temporal expansion clear what movement parameters are being opti-
of processing time that is afforded by the delay mized upon execution (e.g., see Shenoy et al.
period has been oft exploited in neurophysiolog- 2013). For example, returning to the example of
ical studies of the covert cognitive processes that how to plan movements of different speed, one
lead to an overt motor response. would not be particularly surprised if during
Traditionally, single-neuron electrophysiolog- movement execution, the concurrent activity of
ical studies of motor planning have concentrated motor-related neurons displayed different firing
on the activity associated with either saccadic eye rates in association with movements of different
movements or reaching movements of monkeys speed. But is there a neural correlate of the inten-
trained to direct movements toward visual goals. tion to move at a certain speed? More specifi-
Superficially at least, the neural activity recorded cally, could one predict from delay period
in association with performance of a “delayed activity whether an impending movement will
saccade” or “delayed reaching” task is strikingly be fast or slow? Indeed one can, as shown in an
similar. Whether recorded in saccade-specific or elegant study by Churchland and colleagues
reach-specific brain regions, a substantial (2006; see also Cisek 2006) that exemplifies the
way in which different intentions may manifest as introspection is not quite so valuable for consid-
differences in neural activity. In this experiment, ering how motor plans might differ when it comes
monkeys were trained on an instructed delay task to making saccades. Indeed, while making
to make either a high- or low-velocity reach approximately three saccades every second, the
depending on the color of the target to be complex decision and motor planning processes
acquired. Single-neuron recordings from dorsal that inform one’s choices of where and when to
premotor and primary motor cortex revealed that, look go largely unnoticed. Yet such decisions and
along with representing the vector of the upcom- the plans that embody them are essential to effec-
ing movement, the delay period activity of indi- tive interaction with the environment, whether
vidual motor cortical neurons was strongly that entails watching a movie, hitting a baseball,
modulated by the instructed speed of the reach. or just searching for one’s keys on a cluttered
Though clearly demonstrating a neural correlate countertop.
of planned speed, the relationship was not While it may be difficult to deduce from one’s
a simple one; the activity of individual neurons own eye movements, there is no argument that
could be positively or negatively modulated by cognitive strategy, or planning, plays a major role
the intention to make a faster movement, and in guiding saccade choices. In his pioneering
furthermore, this representation of intended studies of nearly a half century ago, Russian
speed was shown to interact with those of osten- psychologist Alfred Yarbus showed (Yarbus
sibly orthogonal spatial dimensions like move- 1967) that subjects who were viewing static
ment distance and direction. images could be induced to alter their patterns
The Churchland et al. study is but one example of saccadic landing points simply by changing
taken from more than three decades of work task instructions (and not the image itself).
demonstrating how movement parameters rang- Thus, the way in which saccades interrogate
ing from low (e.g., force) to high (e.g., trajectory) a visual scene depends heavily on prior knowl-
level can be observed in the preparatory activity edge of what information needs to be gleaned
associated with limb movements (e.g., Tanji and from it. While many subsequent insights have
Evarts 1976; Wise 1985; Kalaska et al. 1989; come from psychophysical studies of natural
Riehle and Requin 1989; Fu et al. 1995; visual scanning, studies attempting to expose
Crammond and Kalaska 2000). As is the case the neural correlates of visual decision making
for movement speed, it is not fully clear how the and saccadic motor planning have typically relied
varied and sometimes redundant information on more controlled task settings. As is the case for
concurrently present within motor preparatory limb movements, the neural signature of oculo-
activity is read out to yield the eventual motor motor planning in its most fundamental form is
commands. This is a rich topic and well beyond that of a representation of impending saccade
the scope of this short entry; nevertheless, it vector. Such is readily observed in single-neuron
should be appreciated that current models of recordings from monkeys trained to perform var-
how plans become action are necessarily iants of a delayed saccade task, and at least qual-
constrained by assumptions about how such sig- itatively similar spatial representations have been
nals would need to be transformed to yield the observed in visuo-oculomotor territories in fron-
requisite downstream signals. tal cortex, parietal cortex, midbrain, thalamus,
and basal ganglia (e.g., Thompson et al. 1996;
Planning of Eye Movements Snyder et al 1997; Glimcher and Sparks 1992;
Relative to the skeletomotor system, the oculo- Schlag and Schlag-Ray 1984; Schlag-Rey and
motor system is considerably less complex bio- Schlag 1984; Hikosaka et al. 1989). But more
mechanically, and as such, its study has been than simply predicting where the eyes will go,
instrumental to deciphering how cognitive pro- a correlate of when the eyes will move, or sac-
cesses influence the neural activity that leads to cadic reaction time can also be observed in the
movement. However, in contrast to reaching, preparatory activity of visuo-oculomotor
neurons, specifically those with ties to brainstem achieve the putative threshold criterion – i.e.,
circuits that are proximal to the extraocular mus- the neural correlate of a canceled motor plan.
cles (e.g., Hanes and Schall 1996; Dorris et al.
1997). Motor Planning Dynamics
The above-described examples from the
Planning Versus Executing skeletomotor and oculomotor systems convey
At what point does an oculomotor plan become the essence of what a motor plan is and how it is
a commitment to action and what is the neural manifested in the neural activity that precedes an
basis of this transition from the merely possible to action. Such an account, however, fails to capture D
the inevitable? In a compelling demonstration, the dynamism of how motor plans are informed
Hanes and Schall (1996) tested the hypothesis by perceptual events and changing internal states
that commitment to a particular saccadic choice in the milliseconds leading up to a motor choice.
occurs only when the preparatory activity of ocu- As noted above, the perception-action cycle has
lomotor neurons representing that choice (i.e., been conceptualized as consisting of a series of
a given saccade vector) exceeds a specific or discrete computational stages, a characterization
fixed threshold. They recorded from saccade- that is reinforced by the way in which its neural
related neurons in the primate frontal eye field basis has been studied. For the most part, the
while monkeys performed the simple task of neural activity associated with performing
looking toward a single visual target presented a perceptually guided motor response has been
at an eccentric location. Importantly, even when examined with tasks, like the instructed delay
looking toward a highly salient and unambiguous task, that promote or even mandate serial com-
goal, saccadic reaction times are known to dis- pletion of perceptual evaluation and motor plan-
play considerable variance. Hanes and Schall ning processes. But, unlike for many laboratory
demonstrated that the fixed-threshold mechanism tasks, perceptual information and behavioral
could account for this variance by verifying two goals are fluid in real life, and accordingly,
predictions about frontal eye field saccade- plans may evolve in ways that reflect this ever-
related activity. First, they observed that the changing reality. As discussed at length in
activity of individual neurons increased mono- another entry (see “▶ Perceptual Decision-Mak-
tonically leading up to saccade onset, and just ing”), the notion that accumulating sensory evi-
prior to saccade onset, activity consistently dence is manifested directly in developing motor
attained the same level. The second prediction is plans (e.g., Gold and Shadlen 2000) blurs the
a necessary consequence of the first one. If the distinction between perceptual and motor pro-
commitment to move occurs when the increasing cesses and hints at a more continuous interplay
motor planning activity crosses a fixed threshold, between perceptual evaluation and response
then the reaction time should be inversely related selection. In fact, studies of perceptual decision
to the rate at which the plan rises: shorter reaction making have consistently demonstrated that the
times should correspond to motor plans that build neurons that accumulate sensory evidence in
up quickly and longer reaction times should occur favor of a particular choice are very often the
when the accumulation is more gradual. Indeed, same ones that plan the action that serves as the
this is precisely what they observed. In an inter- ultimate expression of the decision/choice pro-
esting variation on this experiment, Hanes and cess (e.g., Requin et al. 1988; Gold and Shadlen
Schall also employed a countermanding version 2000; de Lafuente and Romo 2006; Hernández
of the task wherein, on some trials, monkeys were et al. 2010). Clearly, a direct and intimate rela-
cued to withhold the already instructed saccade. tionship between perceptual decision and motor
They found that trials in which the monkey was planning makes good sense if the objective is to
successful in withholding the eye movement base actions on the most currently available
corresponded to those in which the accumulating information. In fact, the immediacy of this rela-
activity of frontal eye field neurons failed to tionship is emphasized by recent studies of urgent
decision making with tasks that prioritize the Glimcher PW, Sparks DL (1992) Movement selection in
incorporation of incoming sensory information advance of action in the superior colliculus. Nature
355:542–545
to guide ongoing motor plans. Remarkably, Gold JI, Shadlen MN (2000) Representation of
these studies suggest that motor systems may a perceptual decision in developing oculomotor com-
simultaneously harbor multiple and sometimes mands. Nature 404:390–394
mutually exclusive motor plans until just milli- Hanes DP, Schall JD (1996) Neural control of voluntary
movement initiation. Science 274:427–430
seconds before a single alternative is committed Hernández A, Nácher V, Luna R, Zainos A, Lemus L,
to action – a conflict resolved on the basis of Alvarez M, Vázquez Y, Camarillo L, Romo R (2010)
a timely perceptual cue (Cisek et al. 2009; Cisek Decoding a perceptual decision process across cortex.
and Kalaska 2010; Stanford et al. 2010). Ulti- Neuron 66:300–314
Hikosaka O, Sakamoto M, Usui S (1989) Functional prop-
mately, understanding the neural basis of motor erties of monkey caudate neurons. I. Activities related
planning will require a more complete framework to saccadic eye movements. J Neurophysiol
for describing how perceptual and motor systems 61:780–798
interact to affect the right action at the right time. Kalaska JF, Cohen DA, Hyde ML, Prud’homme M (1989)
A comparison of movement direction-related versus
load direction-related activity in primate motor cortex,
using a two-dimensional reaching task. J Neurosci
Cross-References 9:2080–2102
Requin J, Riehle A, Seal J (1988) Neuronal activity and
information processing in motor control: from stages
▶ Perceptual Decision-Making to continuous flow. Biol Psychol 26:179–198
Riehle A, Requin J (1989) Monkey primary motor and
premotor cortex: single-cell activity related to prior
References information about direction and extent of an intended
movement. J Neurophysiol 61:534–549
Churchland MM, Santhanam G, Shenoy KV (2006) Prepa- Schlag J, Schlag-Rey M (1984) Visuomotor functions of
ratory activity in premotor and motor cortex reflects central thalamus in monkey. II. Unit activity related to
the speed of the upcoming reach. J Neurophysiol visual events, targeting, and fixation. J Neurophysiol
96:3130–3146 51:1175–1195
Cisek P (2006) Preparing for speed. Focus on “Preparatory Schlag-Rey M, Schlag J (1984) Visuomotor functions of
activity in premotor and motor cortex reflects the speed central thalamus in monkey. I. Unit activity related to
of the upcoming reach”. J Neurophysiol 96:2842–2843 spontaneous eye movements. J Neurophysiol
Cisek P, Kalaska JF (2010) Neural mechanisms for 51:1149–1174
interacting with a world full of action choices. Annu Shenoy KV, Sahani M, Churchland MM (2013) Cortical
Rev Neurosci 33:269–298 control of arm movements: a dynamical systems per-
Cisek P, Puskas GA, El-Murr S (2009) Decisions in spective. Annu Rev Neurosci 36:337–359
changing conditions: the urgency-gating model. Snyder LH, Batista AP, Andersen RA (1997) Coding of
J Neurosci 29:11560–11571 intention in the posterior parietal cortex. Nature
Crammond DJ, Kalaska JF (2000) Prior information in 386:167–170
motor and premotor cortex: activity during the delay Stanford TR, Shankar S, Massoglia DP, Costello MG,
period and effect on pre-movement activity. Salinas E (2010) Perceptual decision making in less
J Neurophysiol 84:986–1005 than 30 milliseconds. Nat Neurosci 13:379–385
de Lafuente V, Romo R (2006) Neural correlate of sub- Tanji J, Evarts EV (1976) Anticipatory activity of
jective sensory experience gradually builds up across motor cortex neurons in relation to direction
cortical areas. Proc Natl Acad Sci U S of an intended movement. J Neurophysiol
A 103:14266–14271 39:1062–1068
Dorris MC, Pare M, Munoz DP (1997) Neuronal activity Thompson KG, Hanes DP, Bichot NP, Schall JD
in monkey superior colliculus related to the initiation (1996) Perceptual and motor processing stages
of saccadic eye movements. J Neurosci 17:8566–8579 identified in the activity of macaque frontal eye field
Fu QG, Flament D, Coltz JD, Ebner TJ (1995) Temporal neurons during visual search. J Neurophysiol
encoding of movement kinematics in the discharge of 76:4040–4055
primate primary motor and premotor neurons. Wise SP (1985) The primate premotor cortex:
J Neurophysiol 73:836–854 past, present, and preparatory. Annu Rev Neurosci
Funahashi S, Bruce CJ, Goldman-Rakic PS (1989) Mne- 8:1–19
monic coding of visual space in the monkey’s dorso- Yarbus AL (1967) Eye movements and vision. Plenum
lateral prefrontal cortex. J Neurophysiol 61:331–349 Press, New York
Decision-Making, Threshold 963 D
Low thresholds produce speedier responses,
Decision-Making, Threshold whereas higher thresholds increase accuracy. In
particular, for the drift diffusion model, which is
Paul Miller based on a perfect integrator, noise accumulates
Department of Biology and Brandeis University, as the square root of time, whereas the signal
Waltham, MA, USA accumulates linearly in time. Therefore, with
Volen National Center for Complex Systems, large enough thresholds resulting in long
Waltham, MA, USA enough decision-making time, accuracy of
decisions can approach 100 %. Specifically, if D
the decision variable has an initial value of zero
Synonyms and thresholds are set at a and b, the two types of
error rate, a (probability of choosing + when the
Absorbing boundary; Decision space boundary signal, S, is ) and b (probability of
choosing when the signal, S, is +), are given
by (Ratcliff 1978):
Definition
2jSjb
exp 2 1
A decision-making threshold is the value of the s
a¼ and
decision-making variable at which the decision is 2jSjb 2jSja
made, such that an action is selected or a commit- exp 2 exp
s s2
ment to one alternative is made, marking the end
2jSja
of accumulation of information. 1 exp 2
s
b¼
2jSjb 2jSja
exp exp
s2 s2
Notably, if thresholds are symmetric (a = b) with
A decision-making threshold determines when
a starting point at the origin, then
a decision-making process is completed. It repre-
sents a value of the decision variable, which in
1
practice could be a linear combination of a set of a¼b¼
2jSja
neural firing rates, at which the accumulation of 1 þ exp
sensory evidence terminates and a response or s2
action is chosen. In two-alternative forced-choice
tasks, two thresholds exist, one for each of the which decreases exponentially with increasing
two alternatives. threshold, a.
In models of decision making, the Prior belief can be incorporated by setting the
threshold can be either static or dynamic two thresholds to be different, reducing a or b,
during an individual trial. Mathematically, respectively, if a positive or negative response is
a threshold is an absorbing boundary of the dif- preferable or, more likely, according to informa-
fusion process for the decision variable. The time tion acquired before stimulus onset.
for the decision variable to reach such an absorb-
ing boundary is the decision-making time
(Fig. 1). Response Times
Static Thresholds Assuming a positive stimulus (S > 0), the distri-

If the threshold is static, it determines where bution of response times for correct trials, P(T +),
one is operating in the speed-accuracy trade-off. is given by
D 964 Decision-Making, Threshold

þ a Sa
hT i ¼ hT i ¼ hT i ¼ tanh 2
S s
If thresholds are static, they can be set at an

optimal level, by making use of the above formu-
lae, if the statistics of the task and the relative
costs of different types of error are known. In
particular, if errors are costly, such as in lost
time due to a long intertrial interval, it pays to
use thresholds far from the decision variable’s
initial value.
Decision-Making, Threshold, Fig. 1 Two static thresh- Multidimensional Models

olds act as absorbing boundaries for the 1D Brownian In models with two decision variables, a static
motion of the decision variable, X. The boundary reached threshold can depend either on each variable sep-
in any particular trial determines the choice (or percept in
a perceptual categorization task) with the time for the arately, independent of the value of the other
decision variable to reach the boundary making the dom- variable, or on the difference between the two
inant contribution to response time variables (see Fig. 2). In terms of neural circuits,
the former corresponds to independent outputs
from the competing groups of neurons, whereas
ps2 the latter corresponds to a push-pull arrangement,
ð1 bÞPðT þ Þ ¼ es2
Sb
2 such that each group excites one response while

ð a þ bÞ
1S2 p2 k2 s2
inhibiting the other. It is notable that for a
X
inf
pka 2 2 þ 2 t two-variable integrator model with static inde-
k sin e s ðaþbÞ
k¼1
aþb pendent thresholds (Fig. 2b), the difference in
decision variables follows a single-variable inte-
and the distribution of response times for incor- grator model, but with thresholds which collapse
rect trials, P(T ), is given by toward zero as the sum of the two variables
increases.
ps2
bPðT Þ ¼
Sa
2
es2 Dynamic Thresholds
ð a þ bÞ
1 In models of decision making with a single deci-
X
inf
pkb 2 S2 p2 k2 s2
þ t
k sin e s2 ðaþbÞ2 sion variable, dynamic thresholds can improve
k¼1
aþb accuracy in tasks with fixed stimulus duration
and can optimize reward accumulation in other
This leads to a mean response time of (Bogacz situations where the cost for acquiring informa-
et al. 2006) tion is not constant in time (Drugowitsch
et al. 2012). If thresholds are initially high, they

aþb Sða þ bÞ can ensure the decision process is not terminated
hT i ¼ ð1 bÞhT þ i þ bhT i ¼ tanh
2S 2s2 too quickly and evidence has time to accumulate
h SðabÞ
i
ða þ bÞ 1 e s2 and outweigh the noise. However, if thresholds
ab
þ h SðaþbÞ SðaþbÞ
i þ : collapse to zero by the end of the stimulus dura-
2 2S
S e s e s
2
tion, they ensure a response is made and any
evidence accumulated up until then contributes
In the symmetric condition, with a = b, then to the determination of that response.
P(T+) = P(T) and the mean response times are Implicit in the calculation of performance of
given by models with collapsing threshold is an
Decoding Field Potentials 965 D
Decision-Making, Threshold, Fig. 2 In models of (Fig. 1) with static thresholds. (b) If thresholds correspond
decision making with two variables, each variable repre- to a constant absolute firing rate for each cell group, then
sents a simple function (like the mean) of neural firing a decision is made with smaller difference in rates as the
rates. (a) If the thresholds correspond to a constant differ- sum of rates increases. This can be equivalent to a 1D
ence in firing rate, the model can map to a 1D model model with collapsing thresholds
assumption that the system can distinguish very References

small differences in the decision variable as the
two thresholds converge on each other. An equiv- Bogacz R, Brown E, Moehlis J, Holmes P, Cohen JD
(2006) The physics of optimal decision making:
alent assumption is the system’s ability to
a formal analysis of models of performance in two-alter-
respond to very small differences in the times native forced-choice tasks. Psychol Rev 113:700–765
when the decision variable reaches one threshold Deneve S (2012) Making decisions with unknown sensory
versus the other, since the two thresholds reach reliability. Front Neurosci 6:75
Drugowitsch J, Moreno-Bote R, Churchland AK, Shadlen
zero simultaneously.
MN, Pouget A (2012) The cost of accumulating evidence
When the statistics of the inputs are unknown, in perceptual decision making. J Neurosci 32:3612–3628
one can optimally adjust the thresholds over the Ratcliff R (1978) A theory of memory retrieval. Psychol
course of the decision, according to whether the Rev 85:59–108
inputs are strongly informative or not. Perhaps
surprisingly, when the inputs are more ambigu-
ous in their evidence for one response or another, Decision Space Boundary
it is optimal for thresholds to collapse quickly,
enforcing a rapid response that depends little on ▶ Decision-Making, Threshold
information in the stimulus but more closely fol-
lows one’s prior beliefs (Deneve 2012).
Decoding Field Potentials
Cross-References Yan Tat Wong and Bijan Pesaran

Center for Neural Science, New York University,
▶ Accumulation of Evidence in Decision-Making New York, NY, USA
▶ Decision-Making Tasks
▶ Decision-Making, Bias
▶ Decision-Making, Models Synonyms
▶ Diffusion Equation
▶ Speed-Accuracy Tradeoff Local field potential decoding
D 966 Decoding Field Potentials
Definition (Donner and Siegel 2011; Fries 2005; Pesaran

et al. 2008).
The local field potential (LFP) is believed to Previous work has shown that the LFP can be
represent the aggregated neuronal activity of decoded off-line to reconstruct high-dimensional
populations of neurons. LFP activity encodes upper limb kinematics (Bansal et al. 2011;
information related to the control of movements Zhuang et al. 2010), 2-dimensional movement
as well as information about sensory and trajectories (Mehring et al. 2003), and the end
other cognitive processes. LFP decoding refers goal of movements (Hwang and Andersen 2013;
to the process of extracting features of these pro- Markowitz et al. 2011). The LFP can also be used
cesses from the measured LFP signals. LFP in conjunction with spiking activity to decode
decoding is typically performed for a brain- behavioral states (Aggarwal et al. 2013); how-
machine interface application to allow control ever, the information content of the LFP is not
of an external device such as a prosthetic limb necessarily dependent on spiking activity being
or a cursor on a screen. present in the vicinity of the electrode (Flint
et al. 2012a). In fact, the information conveyed
via spiking and LFPs can differ in how they vary
Detailed Description with cortical depth (Markowitz et al. 2011).
Work is now being undertaken to move LFP
Overview decoding online, with subjects controlling
Neuromotor prosthesis is a class of brain- a cursor on a screen (Flint et al. 2012b).
machine interfaces (BMIs) that aims to restore
lost motor function by decoding control informa- Applications
tion from neural signals. The local field potential Applications of LFP decoding include extracting
(LFP) is one such neural signal that can be uti- information regarding skeletal joint angles, bone
lized. The LFP is operationally defined as the segment kinematics, and movement end goal
low-frequency components (<~300 Hz) of neural states. This information can then be used to con-
activity recorded extracellularly using microelec- trol external actuators such as an upper limb
trodes and is generated via the summation of prosthetic, motorized wheelchairs, speech
transmembrane currents (Buzsáki et al. 2012). decoders, a computer cursor on a screen, or func-
There is great interest in decoding LFP activity tion electrical stimulators to restore movements
due to the relative ease of recording and long- or paralyzed muscles.
term stability of the signal in comparison with the
spiking activity of individual neurons (Andersen Methods of Decoding
et al. 2004; Hwang and Andersen 2013). Further, The LFP is a continuous time varying voltage
the LFP has been shown to be information rich signal and comprises a multivariate set of obser-
and contains temporally structured information vations when many electrodes are recorded
that reflects behavioral processes, such as move- simultaneously. Decoding can be performed by
ment planning and execution (Pesaran et al. 2002; using the amplitude of the time varying signal or
Scherberger et al. 2005). There is mounting evi- by transforming the signal into a set of features.
dence that specific frequency bands of the LFP Commonly used features involve applying time
may reflect neuronal processes at different spatial frequency spectral techniques and decoding in
scales (Siegel et al. 2012). Gamma band the frequency domain (Mitra and Pesaran 1999),
(30–90 Hz) and high gamma band (90–300 Hz) but other features include those based on wavelet
activity is hypothesized to reflect local encoding transformations and other sparse, over-complete
due to firing of neurons near the recording elec- basis sets. To estimate the frequency spectrum,
trode (Schomburg et al. 2012). Activity in lower- the LFP is generally assumed to be stationary
frequency bands, <30 Hz, is believed to reflect over a short time window, typically 100–500 ms
distributed processing across larger-scale circuits in duration. The window is then stepped in time
Decoding Field Potentials 967 D
allowing decoding across time. To minimize bias et al. 2005). In comparison, the LFP has proven
in the spectral estimates, the LFP can be first to be easier to acquire and the information con-
multiplied by one or many orthogonal data tapers tent more stable to tissue responses (Andersen
before being Fourier transformed. et al. 2004). The low-frequency, <~300 Hz,
Decoding in the frequency domain allows for nature of the LFP also minimizes the sampling
extraction of information carried at each fre- frequency required to acquire the signal com-
quency band; however, this greatly increases the pared to spiking activity. This greatly reduces
number of inputs to the decoder. When the power consumption and data transfer require-
decomposing N channels into M frequency ments of the end application BMI wireless D
bands, the number of potential inputs to the recording systems.
decoder increases to N*M. Due to the challenge
of fitting decoding algorithms in many dimen- Limitations
sional feature spaces, the real-time demands of The main limitation of decoding the LFP is that the
applications, and limitations in computing power, integrative nature of the signal results in reduced
dimensionality reduction such as principal com- spatial and temporal information. Currently, it is
ponent analysis can be applied to the frequency unclear whether enough information can be
transformed LFP to extract a set of modes decoded from the LFP to control higher-
(Markowitz et al. 2011). Feature selection can dimensional devices such as an upper limb
then be performed on these modes to obtain prosthetic. Finally, challenges presented by
a smaller number of features that are then used performing spectral analysis in real time via
for decoding. a fully implanted device also remain to be resolved.
Once the LFP has been preprocessed to give
features, the features are decoded to obtain con-
trol signals. Discrete control signals, such as Cross-References
selecting letters on a keyboard, can be decoded
using classification algorithms such as linear dis- ▶ Brain-Machine Interface: Overview
criminant analysis (Pesaran et al. 2002), the naı̈ve ▶ Local Field Potentials (LFP)
Bayes classifier (Bai et al. 2007), and support ▶ Spectral Methods in Neural Data Analysis:
vector classifiers. Continuous control signals, Overview
such as the trajectory of an arm movement, can
be decoded using regression algorithms, such as
linear regression (Wessberg et al. 2000), penal-
References
ized linear regression (Mulliken et al. 2008),
Kalman filters (Wu et al. 2006), and support Aggarwal V, Mollazadeh M, Davidson AG, Schieber MH,
vector regression (Shpigelman et al. 2008). Thakor NV (2013) State-based decoding of hand and
finger kinematics using neuronal ensemble and LFP
Advantages of the Approach activity during dexterous reach-to-grasp movements.
Local field potential decoding offers the advan- Andersen RA, Musallam S, Pesaran B (2004) Selecting
tage of improved long-term stability and ease of the signals for a brain-machine interface. Curr Opin
recording compared to decoding of action poten- Neurobiol 14:720–726
tials generated by individual neurons. Recordings Bai O, Lin P, Vorbach S, Li J, Furlani S, Hallett M (2007)
Exploration of computational methods for classifica-
of action potentials can be difficult to acquire as tion of movement intention during human voluntary
neurons need to be within ~100 um from the tip of movement from single trial EEG. Clin Neurophysiol
the recording electrode (Buzsáki and Draguhn 118:2637–2655
2004). Further, action potential recordings are Bansal AK, Vargas-Irwin CE, Truccolo W, Donoghue JP
(2011) Relationships among low-frequency local field
highly susceptible to changes in electrode char- potentials, spiking activity, and 3-D reach and grasp
acteristics due to local inflammatory responses kinematics in primary motor and ventral premotor
such as the buildup of scar tissue (Polikov cortices. J Neurophysiol 105:1603–1619
D 968 Deep Brain Stimulation
Buzsáki G, Draguhn A (2004) Neuronal oscillations in Wessberg J, Stambaugh CR, Kralik JD, Beck PD,
cortical networks. Science 304:1926–1929 Laubach M, Chapin JK, Kim J, Biggs SJ, Srinivasan
Buzsáki G, Anastassiou CA, Koch C (2012) The origin of MA, Nicolelis MA (2000) Real-time prediction of
extracellular fields and currents–EEG, ECoG, LFP and hand trajectory by ensembles of cortical neurons in
spikes. Nat Rev Neurosci 13:407–420 primates. Nature 408:361–365
Donner TH, Siegel M (2011) A framework for local cor- Wu W, Gao Y, Bienenstock E, Donoghue JP, Black MJ
tical oscillation patterns. Trends Cogn Sci 15:191–199 (2006) Bayesian population decoding of motor cortical
Flint RD, Lindberg EW, Jordan LR, Miller LE, Slutzky activity using a Kalman filter. Neural Comput
MW (2012a) Accurate decoding of reaching move- 18:80–118
ments from field potentials in the absence of spikes. Zhuang J, Truccolo W, Vargas-Irwin C, Donoghue JP
J Neural Eng 9:046006 (2010) Decoding 3-D reach and grasp kinematics
Flint RD, Wright ZA, Slutzky MW (2012b) Control of from high-frequency local field potentials in primate
a biomimetic brain machine interface with local field primary motor cortex. IEEE Trans Biomed Eng
potentials: performance and stability of a static 57:1774–1784
decoder over 200 days. Proc Ann Int Conf IEEE Eng
Med Biol Soc 2012:6719–6722
Fries P (2005) A mechanism for cognitive dynamics:
neuronal communication through neuronal coherence.
Trends Cogn Sci 9:474–480
Hwang EJ, Andersen RA (2013) The utility of Deep Brain Stimulation
multichannel local field potentials for brain-machine
interfaces. J Neural Eng 10:046005
Markowitz DA, Wong YT, Gray CM, Pesaran B (2011) ▶ Parkinson’s Disease: Deep Brain Stimulation
Optimizing the decoding of movement goals from
local field potentials in Macaque cortex. J Neurosci
31:18412–18422
Mehring C, Rickert J, Vaadia E, Cardosa de Oliveira S,
Aertsen A, Rotter S (2003) Inference of hand move-
ments from local field potentials in monkey motor Deep Cerebellar Nuclei
cortex. Nat Neurosci 6:1253–1254
Mitra P, Pesaran B (1999) Analysis of dynamic brain Dana Cohen
imaging data. Biophys J 76:691–708
Mulliken GH, Musallam S, Andersen RA (2008) Decoding Gonda Brain Research Center, Bar Ilan
trajectories from posterior parietal cortex ensembles. University, Ramat-Gan, Israel
J Neurosci 28:12913–12926
Pesaran B, Pezaris JS, Sahani M, Mitra PP, Andersen RA
(2002) Temporal structure in neuronal activity during
working memory in macaque parietal cortex. Nat Definition
Pesaran B, Nelson MJ, Andersen RA (2008) Free choice The deep cerebellar nuclei (DCN) consist of three
activates a decision circuit between frontal and parietal nuclei: the fastigial (medial) nucleus, the inter-
cortex. Nature 453:406–409
Polikov VS, Tresco PA, Reichert WM (2005) Response of posed nucleus, and the dentate (lateral) nucleus.
brain tissue to chronically implanted neural electrodes. Together they form the primary output from the
J Neurosci Methods 148:1–18 cerebellum. DCN excitatory neurons project to
Scherberger H, Jarvis MR, Andersen RA (2005) Cortical a variety of extra-cerebellar targets, whereas the
local field potential encodes movement intentions in
the posterior parietal cortex. Neuron 46:347–354 inhibitory neurons project exclusively to the infe-
Schomburg EW, Anastassiou CA, Buzsáki G, Koch rior olive. The activity of DCN neurons is shaped
C (2012) The spiking component of oscillatory extra- by massive inhibitory input from Purkinje cells in
cellular potentials in the rat hippocampus. J Neurosci the cerebellar cortex and excitatory input arriving
32:11798–11811
Shpigelman L, Lalazar H, Vaadia E (2009) Kernel-arma via the inferior olive axons – the climbing
for hand tracking and brain-machine interfacing dur- fibers – and mossy fiber collaterals carrying infor-
ing 3D motor control. Adv Neural Info Process Syst mation from the spinal cord, the cerebral cortex,
21:1489–1496 and several brainstem structures. Hyperpolariza-
Siegel M, Donner TH, Engel AK (2012) Spectral finger-
prints of large-scale neuronal interactions. Nat Rev tion induced by the inhibitory input actively con-
Neurosci 13:121–134 trols plasticity in the excitatory synapses in the
Deep Cerebellar Nuclei 969 D
cerebellar nuclei. Overall, DCN neurons partici- DCN Networks
pate in local and global feedback and
feedforward networks thus generating complex The Olivocerebellar Pathway: The DCN form
dynamics. a local closed-loop network with the inferior olive
(IO) and the cerebellar cortex. The climbing fibers
which are the axons of the inferior olive induce
complex spikes in Purkinje cells which in turn inhibit
Detailed Description the DCN. Then via their inhibitory projection neu-
rons, the DCN regulate the coupling strength within D
DCN neurons can be divided into three main IO neurons by activating GABAergic receptors near
subcategories: excitatory projection neurons the IO dendritic gap junctions. The strength of the IO
targeting a variety of extra-cerebellar structures, coupling influences the probability of the IO to gen-
including the cerebral cortex via the thalamus; erate an action potential in the climbing fibers.
inhibitory projection neurons targeting exclu- Climbing fiber collaterals innervate DCN regions
sively the inferior olive; and local inhibitory in a closed-loop manner with the Purkinje cells
interneurons. All types of neurons receive similar activated by them. The one-to-one relation between
ratios of converging afferents; however, the the climbing fiber activity and complex spike occur-
information can differentially be processed due rence in Purkinje cells and the strong synaptic con-
to cell-type-specific biophysical mechanisms. nection formed by the climbing fiber collateral in the
The main synaptic input to the DCN arrives DCN generate a unique excitation-followed-by-
from the Purkinje cells which provide 60–80 % inhibition activation sequence in the DCN.
of the total number of synapses formed on DCN The Mossy Fiber-Cerebellar Pathway: Infor-
neurons. Recent estimates suggest a convergence mation arriving via the mossy fibers is processed by
ratio of about 20–50 Purkinje cells per the granule cells and then distributed by the parallel
nuclear cell which is by an order of magnitude fiber system to the Purkinje cells from which it is
smaller than previously estimated (600–800 projected to the DCN. The mossy fiber collaterals
Purkinje cells per one DCN neuron). The show considerable ramification and a bilateral pro-
climbing fiber collaterals constitute about 5 % jection pattern in the DCN thus generating complex
of DCN input, and the number of interneuron convergence–divergence relation with the mf path-
and mf synapsing on the projection neurons way propagating through the cerebellar cortex. Thus
remains unclear. In addition to their far, information processing rules governing the
glutamatergic and GABAergic synaptic inputs, mossy fiber-cerebellar pathway remain to a large
the DCN receive a less studied serotonergic and extent an enigma.
cholinergic neuromodulatory input. Neurons in
the DCN tend to fire spontaneously at rates >10
spikes/s and can modulate their firing rates in The Shaping of DCN Activation
complex ways in response to either sensory or
motor activation. DCN activation is shaped primarily by massive
The DCN are classically considered relay sta- inhibitory input arriving from the Purkinje cells.
tions that linearly integrate incoming informa- Two neuronal mechanisms are responsible for
tion; however, recent advancement in cerebellar regulating the effectiveness of Purkinje cell inhi-
research suggests otherwise. Neuronal mecha- bition: the Purkinje cell–DCN synaptic charac-
nisms such as short-term depression at the teristics and DCN rebound response.
Purkinje cell–DCN neuron synapse, the occur- The Characteristics of the Purkinje Cell–DCN
rence of rebound response, and diversity in the Synapse: Two unique features characterize the
convergence–divergence ratios of DCN afferents dynamics of the Purkinje cell–DCN synapse – the
enhance the capacity of DCN neurons to perform short duration of its inhibitory postsynaptic current
complex computation. (IPSC) and short-term depression (STD).
D 970 Deep Cerebellar Nuclei
The synapse between the Purkinje cells and during nonrhythmic movement. The source and
the DCN has extremely fast inhibitory postsyn- computational relevance of these transient oscil-
aptic current kinetics (tdecay = 2.5 ms) which lations remains to be unraveled; however, recent
prevents accumulative effect of residual currents. findings suggest that together with the spinal cord
Consequently, DCN neurons are inhibited by and multiple cortical and subcortical brain
asynchronous Purkinje cell activity and are regions, DCN oscillations may underlie the
entrained to a synchronized subset of Purkinje 6–9 Hz discontinuities detected in muscle activ-
cells displaying similar firing rates. ity during smooth slow finger movements. By
The Purkinje cell–DCN synapse displays strong generating coherent oscillations at different
STD which ranges from 30 % for a 20 spikes/s phase lags, this global network may limit drive
input to 90 % for rates >140 spikes/s. As a result, to muscle at the oscillation dominant frequency,
the massive Purkinje cell–DCN projection is thereby reducing tremor and improving move-
counteracted by a nonlinear physiological mecha- ment precision. The concept of neuronal oscilla-
nism which controls the dominant influence of tions as a biological controller is intriguing and
Purkinje cells as indicated by anatomy. The STD should be further examined experimentally, com-
mechanism enhances the Purkinje cell–DCN syn- putationally, and theoretically.
apse sensitivity to dynamic than to steady Purkinje
cell activation. The DCN remain within their active
zone despite the significant depression at this syn- Summary
apse by intrinsically firing at high rates. The STD
together with the short-duration IPSCs generates Over the years the DCN have received little atten-
a neuronal mechanism that enables the DCN to tion in comparison to the plethora of studies
preferentially respond to low-rate, synchronized which focused on the olivocerebellar cortex sys-
Purkinje cell inputs within a high-rate, stochastic tem. The discovery of several complementary
background activity. nonlinear neuronal mechanisms for controlling
After-Hyperpolarization Rebound Response: DCN neuronal activity makes it attractive to
DCN neurons respond to a hyperpolarizing cur- computational and experimental neuroscientists
rent with a after-hyperpolarization rebound who are interested in adding the DCN to the
response, which plays a major part in most cere- cerebellar loop and thus develop a broader view
bellar theories; following hyperpolarization, the of its function (Beitz and Chan-Palay 1979;
membrane rapidly depolarizes and is capable of Teune et al. 1998; Shinoda et al. 2000; Telgkamp
producing spikes at membrane potentials lower and Raman 2002; Pedroarena and Schwarz 2003;
than prior to hyperpolarization. The strength of Baumel et al. 2009; Steuber et al. 2010; Williams
the rebound response increases with the hyperpo- et al. 2010; De Zeeuw et al. 2011; Jaeger 2011;
larization amplitude and duration. Rebound firing Person and Raman 2012; Uusisaari and Knopfel
endows DCN neurons with the ability to respond 2012).
to inhibitory inputs with a spiking output and also
to respond differently to the same input
depending on the history of synaptic input. The
occurrence of rebound response in normal in vivo References
conditions is still controversial.
Baumel Y, Jacobson GA, Cohen D (2009) Implications of
functional anatomy on information processing in the
deep cerebellar nuclei. Front Cell Neurosci 3:14
Oscillations in the DCN Beitz AJ, Chan-Palay V (1979) The medial cerebellar
nucleus in the rat: nuclear volume, cell number, den-
sity and orientation. Neuroscience 4:31–45
Occasionally, a subset of DCN neurons display
De Zeeuw CI, Hoebeek FE, Bosman LW, Schonewille M,
highly synchronized transient epochs of coherent Witter L, Koekkoek SK (2011) Spatiotemporal firing
oscillations at a dominant frequency of 6–9 Hz patterns in the cerebellum. Nature Rev 12:327–344
Delayed Rectifier and A-Type Potassium Channels 971 D
Jaeger D (2011) Mini-review: synaptic integration in the
cerebellar nuclei–perspectives from dynamic clamp Delayed Rectifier and A-Type
and computer simulation studies. Cerebellum (Lond)
10:659–666 Potassium Channels
Pedroarena CM, Schwarz C (2003) Efficacy and short-
term plasticity at GABAergic synapses between Paul B. Manis
Purkinje and cerebellar nuclei neurons. Otolaryngology/Head and Neck Surgery and Cell
Person AL, Raman IM (2012) Purkinje neuron synchrony and Molecular Physiology, University of North
elicits time-locked spiking in the cerebellar nuclei. Carolina, Chapel Hill, NC, USA
Nature 481:502–505 D
Shinoda Y, Sugihara I, Wu HS, Sugiuchi Y (2000) The
entire trajectory of single climbing and mossy fibers in Synonyms
the cerebellar nuclei and cortex. Prog Brain Res
124:173–186
Non-inactivating potassium current/conduc-
Steuber V, Schultheiss NW, Silver RA, De Schutter E,
Jaeger D (2010) Determinants of synaptic integration tances; Transient potassium currents
and heterogeneity in rebound firing explored with
data-driven models of deep cerebellar nucleus cells.
J Comput Neurosci 30:633–658 Definition
Telgkamp P, Raman IM (2002) Depression of inhibitory
synaptic transmission between Purkinje cells and neu- Voltage-gated potassium channels can be
rons of the cerebellar nuclei. J Neurosci 22:8447–8457
Teune TM, van der Burg J, de Zeeuw CI, Voogd J, grouped into two broad categories: the delayed
Ruigrok TJ (1998) Single Purkinje cell can innervate rectifiers and the “A-type” channels. Delayed
multiple classes of projection neurons in the cerebellar rectifiers are named after their delay before acti-
nuclei of the rat: a light microscopic and ultrastructural vation commences following a voltage change
triple-tracer study in the rat. J Comp Neurol
392:164–178 and their contribution to rectification of the
Uusisaari MY, Knopfel T (2012) Diversity of neuronal current-voltage relationship of cells once they
elements and circuitry in the cerebellar nuclei. Cere- are activated. Delayed rectifiers show little
bellum (Lond) 11:420–421 time-dependent inactivation. Inactivation is
Williams ER, Soteropoulos DS, Baker SN (2010) Spinal
interneuron circuits reduce approximately 10-Hz a process whereby the channels close without
movement discontinuities by phase cancellation. Proc a change in the driving force. A-type potassium
Natl Acad Sci USA 107:11098–11103 channels, in contrast, show prominent time and
voltage-dependent inactivation. A-type channels
may also activate with a delay. There are several
approaches to modeling these channels. The roles
Deep Electroencephalography that the different classes of channels may play in
neuronal excitability depend on their interactions
▶ Local Field Potential and Deep Brain Stimula- with other channels, the time course of voltage
tion (DBS) fluctuations in the cells, and their localization in
the membrane. A-type potassium channels in
particular can impart interesting time and
voltage-dependent regulation of spike timing
Deep Learning Architecture that can significantly affect the temporal repre-
sentation of information in the nervous system.
▶ Hierarchical Models of the Visual System
Degeneracy Background
Voltage-gated potassium channels are a highly
▶ Neuronal Parameter Non-uniqueness heterogeneous family. Voltage-gated potassium
D 972 Delayed Rectifier and A-Type Potassium Channels
channels appeared early in cellular evolution, and a-subunits, which have six transmembrane
homologous channels can be even found in plants domains, forms the pore-conducting portion of
(Cherel 2004). Early voltage clamp studies in the channels and assembles as tetrameric protein
invertebrate preparations first identified the complexes. The channel pore loop (the perme-
delayed rectifier conductances (Hodgkin and ation pathway through which potassium flows) is
Huxley 1952). Subsequent studies identified located between the fifth and sixth transmem-
potassium conductances that inactivated rapidly brane spans of each a-subunit, and the assembly
(Nakajima 1966; Connor and Stevens 1971), together forms a single pore. The b-subunits,
which were then termed A-type currents. which can modify channel function, associate
The discharge patterns of neurons result from with the a-subunits with a 1:1 stoichiometry.
the interplay between the membrane voltage and Diversity in Kv channel structure arises from
all of the space, time, and voltage dependence of two principal sources. One is the existence of
their ion conductances. The expression of potas- splice variants (in Kv1.5, Kv3.1, Kv3.2, Kv3.3,
sium conductances varies widely among neurons Kv3.4, and Kv4.3). The other is the way in which
and is one of the main determinants of excitabil- the tetrameric structure is assembled. The spe-
ity and discharge patterns. However, potassium cific assembly of the channel subunits is thought
conductances, even from a single class, play mul- to be governed by C-terminal protein interactions
tiple roles that depend on where they are located such that all of the subunits of a given channel
in the cell membrane (Jensen et al. 2012). Those must belong to a single family. Although there is
conductances located in the soma and axon hill- some evidence for specific rules regarding assem-
ock region are important for discharge pattern bly, the combinatorial possibilities of various
and spike rates. Conductances located in den- alpha and auxiliary subunits, as well as splice
dritic trees help to shape synaptic integration variants, can create channels with a wide diver-
(Johnston and Narayanan 2008) and can contrib- sity of voltage-dependent behavior and with the
ute to bursting behavior of cells. Conductances potential for specific targeting by intracellular
located in the presynaptic membrane can regulate messengers. There seems to be no evidence for
transmitter release. Thus, both delayed rectifier posttranslational editing of these channels in
and A-type potassium channels can confer very mammals. However, posttranslational modifica-
different spiking patterns, subthreshold dynam- tions of the channels, including glycosylation
ics, and shaping of synaptic integration and thus (Watanabe et al. 2003) and phosphorylation
are versatile tools for neural information (Jones and Kaczmarek, 1996; Song and
processing. Kaczmarek, 2006), can significantly alter channel
In mammals, versatility in part results from function.
the production of voltage-gated potassium chan- Although the details of channel assembly are
nels by multiple gene families. Each gene family not well understood, there is evidence that assem-
directs the production of proteins corresponding bly depends on targeting motifs in the protein that
to primary subunits, auxiliary subunits, and regulate channel trafficking to the membrane
channel-associated proteins, and each of these (Jensen et al. 2011), as well as specific interac-
can have quite distinct physiological functions tions among subunits during trafficking through
(reviewed in Coetzee et al. 1999; Jan and Jan the endoplasmic reticulum (Heusser and
2012; Rudy et al. 2009). The delayed rectifier Schwappach 2005). In addition, auxiliary sub-
and A-type channels arise from that is known as units (beta subunits) have been shown to act as
the Kv family of channels. There are four broad chaperones and influence the targeting and
Kv families, which in the literature are addressed assembly of the channels, as well as gating in
by multiple nomenclatures. Kv1 (KCNA), Kv2 assembled channels. Furthermore, the localiza-
(KCNB), Kv3 (KCNC), and Kv4 (KCND) are the tion of Kv channels can be regulated by motifs
primary alpha-forming families. Each of the that can target the channels to postsynaptic
densities. Channel localization in the membrane convenient analytic functions that capture the
is not random, but is, at least for some channels, voltage and time dependence of channel function.
quite focal. This allows the channels to interact In many cases, these functions represent an
with other proteins and may also facilitate local approximation of the underlying biophysical
modulation by short-range messengers, including basis for channel gating. Thus, the voltage depen-
calcium (An et al. 2000; Anderson et al. 2010). dence of the channel activation (and inactiva-
The classification of ion channels is complex, tion), as well as the rates that the channels open,
much as the classification of cell types in a given close, and inactivate at different voltages, needs
brain region. A variety of factors must be consid- to be measured. Such measurements can only be D
ered, including the voltage dependence of activa- made using voltage clamp methods and are best
tion, the kinetics of activation, the presence and made on cells with few processes to improve
speed of inactivation, and the voltage dependence space clamp, or using membrane patches, such
of inactivation. As each of these parameters usu- as outside-out patches. The voltage dependence
ally is represented as a continuum, the categori- and rate of activation (channel opening) are mea-
zation is necessarily fuzzy. Furthermore, the sured by holding the cell at a negative voltage and
categorization of delayed rectifiers and A-type then stepping to different voltage levels, while
channels does not map cleanly onto particular recording the time course of the current. The
Kv families. For example, Kv1.4 is often consid- voltage-dependent rate of deactivation
ered an “A-type” channel, while Kv1.1 and (channels returning to a closed state following
Kv1.2 are delayed rectifiers. It could be argued a step that reduces activation) that can be used
however that the entire Kv1 family is best is measured at voltages where activation is weak
described as delayed rectifiers, with varying or nonexistent by stepping the cell (or patch) to
amounts of slow inactivation. Similarly, Kv3.4 a voltage where the channels are substantially
is often considered “A-type,” while Kv3.1 is opened and then stepping the membrane potential
another kind of delayed rectifier; however the to a new level. The voltage dependence of inac-
same argument as for Kv1 might be used here, tivation (channels closing while the voltage is
in that Kv3.4 inactivation is slow and incomplete. maintained at a level that would otherwise keep
The three members of the Kv4 family all create the channels open) is usually measured by
channels that show rapidly inactivating currents activating the conductance and then measuring
and thus are “classical” A-type channels in this the time course of inactivation directly from
respect. The two members of the Kv2 family are the decrease in current over time with
delayed rectifiers and do not show inactivation. a constant driving force. An additional set of
measurements of the recovery from inactivation
Basis of Models also need to be made. Recovery can be measured
The classical models of Hodgkin and Huxley are by stepping the patch to a voltage that nearly fully
the most commonly used formulations to repre- activates the channels, holding long enough to
sent the responses of delayed rectifier ion chan- allow them to inactivate, and then using a test
nels. These models provide macroscopic pulse to measure the return of the current at
descriptions of conductances and have the advan- various times after the end of the inactivation
tage that they have been well studied and are step, for a variety of voltages during the interme-
readily implemented. diate period (Fig. 8).
The kinetics of the channel opening are mea-
Measurements Needed to Constrain Models sured by direct fits of equations to the activation
In order to construct a model for a particular phase. Typically, the goal is to extract the time
conductance, a number of parameters must be constants as a function of voltage, and the rising
measured. The measurements are usually phase is sigmoidal. Thus, the activation function
re-represented, through curve fitting, by will be
n
1
AðV, tÞ ¼ A0 1 et=t t¼
aþb
where A(V, t) is the current at voltage V, as To isolate alpha and beta, the voltage dependence
a function of time following the onset of a step of steady-state or peak activation can be mea-
voltage change; A0 is the amplitude of the current sured and fit to a Boltzmann function:
change; t is the time constant for activation at
voltage V; and n determines the shape of the 1
sigmoidal rise of the current and is usually an nð V Þ ¼
1þ eðVV 0 Þ=k
integer value between 1 and 4.
In some cases, multiple terms are needed, for where n(V) represents the fraction of channels
example, when the activation has both open at any given time, which is also a/a + b.
a sigmoidal rise and a slow component. Following some algebraic manipulations, a and b
Deactivation rates are determined from can then be computed from the t and n(V) curves
as
AðV, tÞ ¼ b et=t
a ¼ nðV Þ=t b ¼ ð1 t aÞ=t
Usually this single-exponential function is suffi-
cient. However, there are some conditions where These calculations can be done for the individual
the current decay associated with the channels voltage points. The resulting a(V) and b(V) are
turning off at a new voltage may have multiple then typically fit to exponential functions of volt-
components. For example, a cell might express age. Alternatively, the t(V) can be fit, combining
two populations of channels that cannot be phar- activation and deactivation kinetic measurements
macologically isolated. Alternatively, there to a more complex function, usually of the form
might only one set of channels that have multiple
energetically distinct open states or closed states. 1
In this case, the closing rate will depend on which tð V Þ ¼ þO
a eðVV 1 Þ=k1 þ b eðVV 2 Þ=k2
state a population of channels might be in at the
time of the step. In intact neurons with extensive although any equation that adequately describes
dendritic trees, multiple decay time constants the data can be used.
may also appear, but will have an ambiguous
interpretation. While such decays may represent Measuring Inactivation
multiple channels or states within channels, they The voltage dependence of inactivation is mea-
can also result from channels located at different sured by using a series of voltage steps, usually of
places in the cell that are not adequately con- fairly long duration, prior to a depolarizing step
trolled by the voltage clamp and so are not at that activates the channels. The peak current is
the same voltages or state when the step is then fit to a Boltzmann function to describe the
terminated. steady-state inactivation. In practice, it may be
These rates are, in all cases, aggregate necessary to also measure the voltage-
rates resulting from the equilibrium rates, which independent conductances (such as residual
are usually analyzed in terms of a two-state delayed rectifiers) using a depolarizing condi-
model: tioning voltage step and subtracting the two sets
of traces to isolate the transient component of the
a
CÐO current. Inactivation may be incomplete, so the
b function needs to include an offset parameter to
account for the minimum level of inactivation.
The t values are related to a and b as follows: The voltage dependence of the rate of
inactivation is measured by first holding at sufficient to model the excitability of cells, more
a negative potential to remove all inactivation detailed kinetic information may be obtained
and then stepping to various positive potentials. from single-channel measurements.
The decay phase of the current is then fit to
a single or double exponential function, and Additional Kinetic Components
these rates are used for the t curves. In some cases, the kinetic behavior of the chan-
nels appears to be more complex than a simple
Problems Encountered in Making These model may predict. For example, some channels
Measurements show additional state-dependent transitions that D
Technical appear as slow changes in open probability or
As already indicated, having adequate space multiple stages in recovery from inactivation. In
clamp is absolutely necessary. This can be this case more insight can be gained by examin-
achieved by using expression systems in small ing single-channel currents using the same pro-
cells without processes, performing acute isola- tocols described above. The incorporation of
tions of cells, which shears off most of the pro- additional kinetic components requires some
cesses, or by using outside out, inside out, or care and can be done either as an adjustment to
enucleated patches. In addition it is important to an existing model or by treating the kinetic com-
properly compensate the amplifiers so that the ponents as if they arose from a separate popula-
current measurements and time course are as tion of independent channels.
accurate as possible. The best measurements are
made by minimizing series resistance and elec- Computation of Models
trode capacitance and using a wide-band ampli- There are several approaches that can be used for
fier with minimum lag settings. the computation of the models. The first, and
most common, is to express the channel kinetics
Heterogeneity as a set of coupled differential equations and to
When working with native channels, it is quite solve those equations numerically for forcing
common to encounter heterogeneity in the functions that change the voltage. This is the
responses to voltage steps. There are two sources approach used in the examples that are shown
of heterogeneity in such measurements. The first below. A second, which can give insight into
is measurement error that arises from incorrect the structure of the model, is to use the Q-matrix
voltage clamp due to space clamp problems, or formulation (Colquhoun and Hawkes, 1995).
inadequate compensation for series resistance. Q matrices are useful for examining state changes
The second is true biological variability, such as in response to step perturbations. The simulations
modifications of voltage dependence or open shown in Figs. 1–10 were computed using NEU-
probability by phosphorylation. This increases RON (www.neuron.yale.edu), with Python. The
the scatter in the data. This can be addressed equations for the channels are from Rothman and
either by averaging results from many cells Manis (2003b). For simulations with action
(or fitting to the population data) or by identifying potentials, a model of the NaV1.1 sodium chan-
an underlying mechanism that causes the scatter, nel (Barela et al. 2006) was incorporated into the
such as phosphorylation, and attempting to con- model.
trol that factor. Another approach is to fit the data
from individual cells and to use the variability in Examples of Delayed Rectifier Models
subsequent modeling.
Example 1 Figure 1 shows an example of a volt-
Alternative Approaches age clamp simulation of a delayed rectifier
While measurements of the macroscopic behav- potassium conductance, using the description
ior of neuronal conductances are frequently of Rothman and Manis (2003a, b). This is
Delayed Rectifier and A-Type Potassium Channels, Fig. 1 Currents and conductances for delayed rectifier
potassium conductance
a “high-voltage” activated delayed rectifier. The a rapid activation (in red), and a small component
voltage clamp command voltages are shown in with slower activation (in blue). The activation
Fig. 1c, and the resulting currents are shown in rates, as a function of voltage, are plotted in
Fig. 1a, for steps up to 0 mV. The current grows Fig. 1f for this model.
nonlinearly with voltage. The channels activate This delayed rectifier is typical of the channels
rapidly, and current through the channels is that help repolarize action potentials in cortical
sustained during the 100 msec voltage step. interneurons and neurons of the auditory pathway
A tail current is visible at the end of the step and is associated primarily with the Kv3 (Kv3.1,
(at 120 ms). Figure 1b shows the conductance of Kv3.2) family of potassium channels. Note that
the potassium channels during the voltage step. the conductances do not activate until nearly
Figure 1d shows the steady-state current-voltage 40 mV. This high activation voltage suggests
relationship of the currents in Fig. 1a. The chan- that the conductances will not play a role in the
nels begin to open near 40 mV, and once acti- integration of synaptic events, but are only trig-
vated, the currents become large with increasing gered when action potentials occur.
voltage. Figure 1e shows the gating variables, Figure 2 shows the time course of the delayed
which range from 0 to 1. There are two activation rectifier channel activation during an action
gating variables in this model, corresponding to potential, in current clamp. The current begins
Delayed Rectifier and
A-Type Potassium
Channels,
Fig. 2 Activation time
course of delayed rectifier
channels during an action
potential
to activate near the peak of the action potential Example 2 The delayed rectifier class of channels
(indicated by the black dashed line) and reaches shows variability in the activation voltage range
its peak (indicated by the red dashed line) about and in the amount of inactivation that can occur.
halfway down the falling phase of the action The channels in Example 1 represent one part of the
potential. Thus, the delayed rectifier can help spectrum and have a relatively high-voltage thresh-
repolarize action potentials. old and no inactivation. However other channels
This conductance does not show inactivation have a much lower activation-voltage range and
and deactivates very rapidly. Therefore, it should show some inactivation. For example, the channels
be available to repolarize action potentials with that help repolarize action potentials at the nodes of
nearly the same conductance every time. This Ranvier in axons are largely composed of Kv1
ability is illustrated in Fig. 3, where a high- family channels (Rasband 2010). These channels
frequency train of action potentials is elicited by are also found in the cell bodies of some neurons,
a depolarizing current pulse. The potassium con- where they play a role in the integration of synaptic
ductance, plotted in Fig. 3b, is nearly the same for events and regulate discharge patterns. As an exam-
each action potential. The slightly smaller current ple of these channels from this group, simulations
associated with the first action potential is related are shown in Fig. 4 for voltage clamp of the
to initial conditions and the presence of other low-voltage activated potassium conductance
channels in the model that affected the action found in neurons of the auditory brainstem, which
potential width very slightly. are composed of various mixtures of Kv1 subunits.

A-Type Potassium
Channels, Fig. 3 Delayed
rectifier that activates in
nearly the same way for
every action potential in
a high-frequency train
There are several features of the voltage clamp with more depolarized steps. The activation and
currents in Fig. 4a that are important to identify. inactivation state variables are plotted in Fig. 4e.
First, with the model cell held at 60 mV, there is Note that near the resting potential (60 mV), the
a slight outward current (visible from 0 to 20 ms). activation curve is already at 5–10 % of the chan-
This current arises because these channels have nels “open,” while the conductance is fully acti-
a significant open probability at this voltage, vated by 20 mV. Figure 4f shows the activation
which is close to the resting potential of typical rate (in red), which is very fast, and the slower
neurons. Second, with depolarization, the current inactivation rate in blue. These curves have the
activates rapidly, but then begins to inactivate. typical voltage dependence, where the time con-
However, the inactivation is slow and incom- stants are slowest near the resting potential of the
plete. Third, hyperpolarizing pulses turn the con- cells.
ductance off. At the end of a hyperpolarizing These low-voltage activated conductances
pulse, the conductance activates again, and often prevent cells (and axons) from firing repet-
because some of the inactivation has been itively during depolarization. This feature is illus-
removed by the hyperpolarization, the current trated in Fig. 5a. A large depolarization can
shows a slight peak before settling to a new equi- initiate a single spike, and the rapid activation
librium. This rebound is visible in the slight out- of this conductance (Fig. 5b) contributes to the
ward current between 125 and 150 ms. These repolarization of the action potential. In addition,
features are also all visible in the conductance the activation of the conductance generates
traces in Fig. 4b. Figure 4d shows the peak (blue) a sustained shunting current that prevents subse-
and steady-state (red) currents as a function of quent action potentials. The voltage during the
voltage, showing how the inactivation increases current pulse in Fig. 5a represents an equilibrium
Delayed Rectifier and A-Type Potassium Channels, Fig. 4 Delayed rectifier with incomplete inactivation and low
activation-voltage range
between the depolarization caused by the current Finally, we can consider the states of the chan-
step (Fig. 5c) and the outward, hyperpolarizing nels, as a useful reminder that the total conduc-
current generated by the conductance (Fig. 5b). tance is an equilibrium condition across
This single-spike firing pattern is also typical of a population of channels and that the channels
models using the standard Hodgkin-Huxley exist in more than just two states (open and
delayed rectifier, which has a similar activation closed). In the Hodgkin-Huxley formulation, the
voltage close to the resting potential. Note that activation of a channel requires a series of ener-
the membrane depolarizes slightly during the getic transitions across a set of four closed
current pulse, as the slow inactivation of the states, culminating in the open state. At any
conductance proceeds. Cells that express high given voltage, the channels are distributed
levels of this conductance are sensitive to the across these states, in different proportions.
slope, not just the amplitude, of synaptic poten- Using the Q-matrix formulation (Colquhoun
tials, and the conductance can prevent action and Hawkes 1995), we can easily see at equilib-
potentials from being elicited by slow depolar- rium what this distribution is like for any given
izations. In addition, because the conductance is state, as shown in Fig. 6. This figure shows the
active at the resting potential, cells tend to have distribution of channel subunit states for the
short time constants. model:

A-Type Potassium
Channels, Fig. 5 Low-
voltage activation range of
some delayed rectifier
channels that can suppress
repetitive firing

A-Type Potassium
Channels,
Fig. 6 Residence of
channel states in
a multistate model that is
voltage dependent
C0 Ð C1 Ð C2 Ð C3 Ð O expected. This is particularly true at 60 mV,
where about 10 % of the channels are open,
where each of the Cn refers to a population of while just over 30 % are in the C3 state. The
subunits (e.g., four subunits in the “closed” state channels in the C3 state only have to go through
at C0, three at C1, etc.) and O refers to the open a single energetic transition to reach the open
(conducting) state. Thus, at 100 mV, nearly all state, and thus the activation will be immediate
of the subunits are in the C0 state, while at and take on a single-exponential shape. The chan-
20 mV, nearly all are in the O state. The volt- nels in the C2 state (~35 %) will also transit into
ages near rest (70 to 50) are interesting. Here, the C3 and O states, with those reaching the D
we see that the channel subunits are in a different O state doing so with a second-order exponential
equilibrium, with a high probability of subunits activation (sigmoidal) function. This state-based
being in the C2 or C3 state. Channels with subunit approach is conceptual but also very useful in
state distributions like this will activate very rap- thinking about how channels activate or deacti-
idly, and the activation may appear to be first vate and in understanding channel behavior in the
order, rather than sigmoidal as might be important voltage range at the cusp of activation.
Delayed Rectifier and A-Type Potassium Channels, Manis (2003b). This channel has two different voltage-
Fig. 7 Voltage clamp model of A-type (transient potas- dependent inactivation rates (visible in F with different
sium) currents, based on channels from Rothman and colors)

A-Type Potassium
Channels, Fig. 8 Time
course of recovery from
inactivation of A-type
potassium channel model
using a two-pulse paradigm
Examples of A-type Potassium Channel described for delayed rectifiers (or the Hodgkin-
Models Huxley sodium channel), taking into account the
The transient potassium conductances can be instantaneous rates as a function of voltage. This
modeled in the same way as the delayed rectifiers, formulation is equivalent to adding an
but with the addition of a term that describes the inactivating state, but does not ascribe
rate at which the channels inactivate. Inactivation a particular mechanism to inactivation. Nonethe-
was discussed briefly above with respect to less, it often provides an acceptable description of
low-voltage activated channels. Modeling inacti- channel behavior under a variety of conditions.
vation requires the addition of a state variable and The response of a voltage-clamped cell with
its rate constants as a function of voltage. Typi- a rapidly inactivating potassium current (from
cally, an inactivating conductance is represented Rothman and Manis 2003a, b) is shown in
as Fig. 7. Activation with depolarization is fast
(Fig. 7a), but inactivation begins immediately,
gðt, V Þ ¼ gmax aðt, V Þn bðt, V Þ shutting off the conductance (Fig. 7b) and
decreasing the current until the channels are
where gmax is the maximal conductance, nearly completely off. The peak activation is
a represents the activation state variable, and similar to that of the delayed rectifier (Fig. 7b),
b is the inactivation state variable. The time evo- but the steady-state current shows only a small
lution of a and b is computed in the same way as hump in the middle of the voltage range.
A-Type Potassium
Channels, Fig. 9 Time
courses of A-type
potassium conductance
during an action potential.
The activation is similar to
the delayed rectifier in
Fig. 2. Note that the action
potential is also being
repolarized by the delayed D
rectifier in this model
The steady-state activation and inactivation channels recover from the inactivated state, even-
curves are shown in Fig. 7e. Note that there is tually reaching the same conductance that they
overlap of the curves in the 60 to 40 mV did when activated from 60 mV. The recovery
range. This overlap creates a “window” current, is quite fast and essentially proceeds with a single-
which causes the hump in the steady-state rela- exponential time course, beginning at the instant
tionship in Fig. 7d. In this range, the channels of repolarization. Recovery proceeds more
both activate and inactivate, but there is a finite quickly with deeper hyperpolarization between
probability that some channels will be in the open the conditioning 100 ms step and the test step.
state at any time. The rate constants are shown in Like the delayed rectifier, the activation of the
Fig. 7f. This particular version of the A-current transient currents begins near the peak (Fig. 9a)
had a second slow inactivation component, of the action potential, after which the channels
whose unusual voltage dependence is shown by close quickly (Fig. 9b). During trains of action
the green curve. potentials, recovery from inactivation depends on
After inactivation, the channels require some the depth of the spike after hyperpolarization and
time at a hyperpolarized voltage to recover and the interspike interval. Thus, the contribution of
become available for activation again. This is transient potassium currents is highly dependent
illustrated in Fig. 8a. A 100 ms voltage clamp on multiple factors. In Fig. 10, a train of action
step to 10 mV is followed at a variable delay by potentials (Fig. 10a) is associated with variable
10 ms steps to 0 mV. As the delay increases, the activation of the transient current. The transient

A-Type Potassium
Channels,
Fig. 10 Activation of
A-type potassium
conductance during a train
of action potentials. Note
that the conductance
decreases during the train
because of inactivation
conductance decreases during the train, as chan- synaptic potentials and to regulate spike dis-
nels are inactivated and there is not sufficient charge patterns and spike latency (Kanold and
time for full recovery from inactivation between Manis 1999). Indeed, interesting nonlinearities
successive action potentials. and bifurcation behavior of the spike patterns
have been seen and analyzed (Meng
et al. 2011). As the relative timing of spikes in
Biological Significance
neurons, even on a millisecond scale, is thought
The variety of potassium conductances contrib-
to be important in the representation of informa-
utes to a rich repertoire of firing patterns among
tion, mechanisms that can modulate that timing
neurons. The details of the voltage dependence of
are likely to play important roles in how informa-
activation and inactivation, as well as the voltage
tion is integrated in downstream neurons.
dependence of rates, all play into the functional
contributions of the channels. While the repolar-
Acknowledgement This work was supported by NIH/
ization of action potentials is one important func- NIDCD grant DC004551 to PBM.
tion, the A-type channels in particular can have
a complex effect on spike latency in many cells.
Channels with a relatively negative activation
References
and inactivation that are significantly modulated
by small changes in membrane potential near and An WF, Bowlby MR, Betty M, Cao J, Ling HP,
below rest are especially poised to interact with Mendoza G, Hinson JW, Mattsson KI, Strassle BW,
Delta Rhythms: Models and Physiology 985 D
Trimmer JS, Rhodes KJ (2000) Modulation of A-type cochlear nucleus neurons. J Neurophysiol
potassium channels by a family of calcium sensors. 89:3097–3113
Nature 403:553–556 Rudy B, Maffie J, Amarillo Y, Clark B, Goldberg EM,
Anderson D, Mehaffey WH, Iftinca M, Rehak R, Engbers Jeong H-Y et al (2009) Voltage gated potassium chan-
JDT, Hameed S, Zamponi GW, Turner RW nels: structure and function of Kv1 to Kv9 subfamilies.
(2010) Regulation of neuronal activity by Cav3-Kv4 In: Squire LR (ed) Encyclopedia of neuroscience,
channel signaling complexes. Nat Neurosci 10th edn. Academic, Oxford, pp 397–425
13:333–337 Song P, Kaczmarek LK (2006) Modulation of Kv3.1b
Barela AJ, Waddy SP, Lickfett JG, Hunter J, Anido A, potassium channel phosphorylation in auditory neu-
Helmers SL, Goldin AL, Escayg A (2006) An epilepsy rons by conventional and novel protein kinase C iso-
mutation in the sodium channel SCN1A that decreases zymes. J Biol Chem 281:15582-15591. D
channel excitability. J Neurosci 26:2714–2723 Watanabe I, Wang HG, Sutachan JJ, Zhu J, Recio-Pinto E,
Cherel I (2004) Regulation of Kþ channel activities in Thornhill WB (2003) Glycosylation affects rat Kv1.1
plants. From physiological to molecular aspects. potassium channel gating by a combined surface poten-
J Exp Botany 55:337–351 tial and cooperative subunit interaction mechanism. J
Coetzee WA, Amarillo Y, Chiu J, Chow A, Lau D, Physiol. 550:51-66. Erratum in: J Physiol. 551:1059.
McCormack T, Moreno H, Nadal MS, Ozaita A,
Pountney D, Saganich M, Vega-Saenz de Miera E,
Rudy B (1999) Molecular diversity of K+ channels.
Ann N Y Acad Sci 868:233–285
Colquhoun D, Hawkes AG (1995) A Q-matrix cookbook. Delta Oscillations
In: Sakmann B, Neher E (eds) Single-channel record-
ing, 2nd edn. Plenum Press, New York, pp 589–633
▶ Low Frequency Oscillations (Anesthesia and
Connor JA, Stevens CF (1971) Voltage clamp studies of
a transient outward membrane current in gastropod Sleep): Overview
neural somata. J Physiol Lond 213:21–30
Heusser K, Schwappach B (2005) Trafficking of potas-
sium channels. Curr Opin Neurobiol 15:364–369
Hodgkin AL, Huxley AF (1952) Currents carried by Delta Rhythms: Models and
sodium and potassium ions through the membrane of Physiology
the giant axon of Loligo. J Physiol Lond 116:449–472
Jan LY, Jan YN (2012) Voltage-gated potassium channels
and the diversity of electrical signalling. J Physiol Roger Traub1 and Miles Whittington2
Lond 590:2591–2599 1
Physical Sciences, IBM T.J. Watson Research
Jensen CS, Rasmussen HB, Misonou H (2011) Neuronal Center, Yorktown Heights, NY, USA
trafficking of voltage-gated potassium channels. Mol 2
Cell Neurosci 48:288–297 Hull-York Medical School, Hull/York, UK
Johnston D, Narayanan R (2008) Active dendrites: color-
ful wings of the mysterious butterflies. Trends
Neurosci 31:309–316 Synonyms
Kanold PO, Manis PB (1999) Transient potassium cur-
rents regulate the discharge patterns of dorsal cochlear
nucleus pyramidal cells. J Neurosci. 19:2195–2208. Slow oscillation; Slow-wave oscillation (SWO)
Jones, EA, Kaczmarek, L (1996) Regulation of potassium
channels by protein kinases. Curr Opin Neurobiol.
6:318–323.
Meng X, Lu Q, Rinzel J (2011) Control of firing patterns Definition
by two transient potassium currents: leading spike,
latency, bistability. J Comput Neurosci 31:117–136 A high amplitude brain rhythm with frequency
Nakajima S (1966) Analysis of K inactivation and TEA between 1 and 4 Hz.
action in the supramedullary cells of puffer. J Gen
Physiol 49:629–640
Rasband MN (2010) Clustered K+ channel complexes in
axons. Neurosci Lett 486:101–106 Detailed Description
Rothman JS, Manis PB (2003a) Kinetic analyses of three
distinct potassium conductances in ventral cochlear
nucleus neurons. J Neurophysiol 89:3083–3096 History
Rothman JS, Manis PB (2003b) The roles potassium cur- The term “delta” was first introduced by Walter
rents play in regulating the electrical activity of ventral in 1936 to describe frequencies below the alpha
D 986 Delta Rhythms: Models and Physiology
(9–11 Hz) band while attempting to diagnose sleep is associated with consolidation of declara-
tumors using EEG signatures (Walter 1936). He tive memory – suggesting a role in modifying
subsequently, with Berger, subdivided this band synaptic plastic changes induced during wakeful-
to include theta rhythms (between 4 and 8 Hz). ness by sensory input. This has led to the “syn-
Since then the presentation of delta rhythms has aptic rescaling” hypothesis – the restoration of
been shown to be variable in frequency and a mean baseline of synaptic weights in the cortex
amplitude during sleep (Church 1975), and (Tononi and Cirelli 2003, 2006). The type of
some observers subdivide delta further into plasticity seen during deep sleep appears to be
delta1 (1–2 Hz) and delta2 (3–4 Hz); this may dominated by long-term depression. Dopaminer-
have some mechanistic validity given the differ- gic neuromodulation is very low during deep
ent potential origins of the rhythm (see below, sleep, and, specifically, D1 receptor activation
Weiss and Rappelsberger 2000). impairs long-term depression and enhances
spike timing-dependent long-term potentiation
Association with Brain State (Xu and Yao 2010). Computational models pre-
Delta rhythms (1–4 Hz) are predominantly asso- dict that as delta rhythms decline during sleep, so
ciated with deep sleep (stage N3). These sleep does overall cortical synaptic strength (Riedner
stages are manifest most strongly during the first et al. 2007), and, experimentally, GluR1 receptor
few hours of sleep in humans, occurring itera- phosphorylation levels are potentiated during
tively in epochs of ca. 0.5–1.0 h, with decreasing wakefulness and depressed after sleep
power as sleep progresses (Campbell 2009). They (Vyazovskiy et al 2008). From a synaptic
have been linked to the consolidation of memo- dynamic point of view, activity in the delta fre-
ries of consciously experienced events (Huber quency band in general facilitates depression
et al. 2004). However, it currently remains (Staubli and Lynch 1990).
unclear how such memory consolidation may If delta rhythms do chaperone a synaptic
take place. The correlation with deep sleep is depression process, which exact cortical repre-
reinforced with observations of very large sentations are being attenuated? Sleep-associated
increases in delta rhythms following sleep depri- slow rhythms in general have been proposed to be
vation and an association with narcolepsy. temporally discrete (hundreds of ms) fragments
Delta rhythms are also associated with certain of wakefulness (Destexhe et al. 2007). Thus it is
forms of cognition during wakefulness. They are perhaps not surprising that brief, iterative epochs
manifest during the processing of repetitive audi- of sensory-/memory-associated cortical represen-
tory sensory information (Lakatos et al. 2008), tations are somehow operated upon during delta
being precisely phase reset by entrainment to (Peyrache et al. 2009; Riedner et al. 2007). By
such inputs. This results in response gain “cortical representation,” we mean here “defined
enhancement and reaction time improves sets of synapses and the neurons they intercon-
(Lakatos et al. 2008). Such a process is seen nect.” Studies on delta rhythms generated exper-
predominantly when there is a pattern of sensory imentally in local circuits of association cortex
input to entrain to. Without such predictable have shown an iterative cycle of reciprocal inter-
input, the lower frequency delta “carrier wave” actions between superficial and deep cortical
becomes detrimental to cortical function and is laminae (Carracedo et al. 2013). This type of
actively suppressed in favor of a state of near- switching between descending and ascending
continuous vigilance manifest as continuous interlaminar interactions has been seen during
high-frequency rhythms (Schroeder and Lakatos wake states on a slower timescale and more per-
2009). sistently during memory tasks: Coding of sensory
inputs is associated with descending interlaminar
Function communication, whereas the subsequent memory
Little is known with certainty about what delta component of the task involves a switch to
rhythms do in the brain. In general, slow-wave ascending interactions (Takeuchi et al. 2011).
In addition, this pattern of activity closely individual thalamocortical cells in a manner
matches the “Helmholtz machine” proposed by dependent on membrane potential (McCormick
Hinton and colleagues as a general format for and Pape 1990). These authors showed
unsupervised learning (Hinton and Zemel 1994; a mechanism involving behavior of intrinsic
Hinton et al. 1995). The mechanistic key to any membrane conductances (Ih and IT, respectively,
selectivity in changing synaptic weights comes the anomalous rectifier and low-threshold,
from the interaction of delta rhythms with theta T-type, calcium conductance). Specifically,
oscillations. More active neurons generate out- a window current generated by T-type calcium
puts at the delta-nested theta frequency, which channels, formed by the overlap at around D
promotes or preserves synaptic weights (Larson 60 mV between their steady-state activation
et al. 1986), whereas weaker activated neurons and inactivation behaviors, has been well charac-
fire only at the synaptic depression inducing delta terized (Pirchio et al. 1997; Hughes et al. 1999,
frequency. 2002). Small perturbations in membrane poten-
tial around this window generate large signal
Mechanism amplifications and the rhythmic behavior typical
When considering experimental evidence for of thalamic delta rhythms (Williams et al., 2004).
mechanisms underlying delta rhythms, it is The conditions in which expression of this intrin-
important to clearly define the frequency band. sic delta behavior is favored in vitro match the
Oscillations faster than 4 Hz in the neocortex are neuromodulatory state of the in vivo brain during
mechanistically distinct (from delta) when they deep sleep. Cholinergic and monoaminergic tone
coexist with such delta rhythms – the faster oscil- are very low during deep sleep, abolishing a tonic
lations involve a different population of layer depolarized state in thalamocortical neurons
5 pyramidal neurons from those generating the (McCormick 1990) and facilitating activation of
delta rhythm (Carracedo et al. 2013). Frequencies the window current.
lower than 1 Hz are also seen with delta rhythms
during very deep sleep but, again, appear to have Neocortical Generator
different mechanisms (Steriade et al. 1990). In Some forms of delta rhythm have characteristics,
particular the dependence of very slow “up-/ suggesting a purely cortical origin (Amzica and
down-state” oscillations has been shown to be Steriade 1998; Fell et al. 2002). During deep
exquisitely sensitive to balance between fast syn- sleep, cortical delta rhythms can be generated
aptic excitation and inhibition, while delta locally (Vassalli and Dijk 2009), sometimes hav-
rhythms are more robust in this respect ing dynamic properties consistent with discrete
(cf Carracedo et al. 2013 with Haider cortical region specificity (Vyazovskiy
et al. 2006). This entry will consider delta rhythm et al. 2011). In general the expression of delta in
mechanisms only. It will focus on a brief descrip- cortex shows strong nonuniformity – favoring the
tion of what is known about the thalamic and frontal-parietal axis (Ioannides et al. 2009).
neocortical generators separately, though both Current source density measurements during
clearly interact to produce the rich dynamic sig- sleep in experimental animals and humans reveal
natures associated with noninvasive recordings a strong source in superficial layers, suggesting
of in vivo brain function during slow-wave oscil- that any synaptic drive underlying delta rhythms
lations (Crunelli and Hughes 2010). was located in these layers (Csercsa et al. 2010).
In contrast, targeted activation of deep layer pyr-
Thalamic Generator amids strongly modulated slow-wave sleep
A thalamic generator of delta (1–4 Hz), modified rhythms, whereas targeted activation of superfi-
by sensory input, is seen in the thalamus (Amzica cial neurons did not (Beltramo et al. 2013).
et al. 1992). This delta activity persists after dis- A potential resolution to this dichotomy came
connection of the thalamus from the cortex from in vitro models of local delta rhythms that
(Steriade et al. 1987) and can be seen in showed a critical role for intrinsically bursting
neurons in layer 5 of the parietal cortex labor appears to exist in layer 5. While intrinsi-
(Carracedo et al. 2013). These neurons have cell cally bursting neurons participate in circuits gen-
bodies in upper layer 5 but also have tufted distal erating delta, their outputs nest a theta generator
apical dendrites extending up through layers 1–3. centered around layer 5 regular spiking neurons.
Pharmacological manipulation of this model It is this nested theta output that provides the
demonstrated a mechanism whereby interplay substrate for at least one set of “weights” in the
between slow GABAB receptor-mediated inhibi- Helmholtz analogy – those dictating the degree of
tion and NMDA receptor-dependent synaptic reciprocal interlaminar interaction seen on each
excitation, and gap junctional excitation, was delta period.
vital to delta generation. Intrinsically bursting
neurons have high concentrations of these synap- Models
tic receptor subtypes out on their distal dendrites Thalamic Models
suggesting that whereas the synaptic connections The intrinsic delta generator in thalamocortical
that underlie the rhythm as an emergent property neurons has been modeled extensively based on
of local networks are located mainly in dendrites the interplay between a hyperpolarization-
in superficial layers, the intense burst outputs of activated nonspecific cation current (Ih) and the
these neurons during delta arise from the somata transient low-threshold calcium current (IT) (see
of these cells in deep layers. Tóth and Crunelli 1992). In addition most model
This experimental prediction for the core delta neurons possess a leak conductance and fast cur-
mechanism model is supported by observations rents underlying sodium spikes (INa and IK(DR),
showing that reduced gap junction conductance, Traub et al. 1991). In single-compartment models
or blockade of NMDA receptors, reduces slow- the steady-state values for activation and inacti-
wave sleep and layer 5 neuron bursting vation of IT and Ih were critical in generating both
(Armstrong-James and Fox 1988; Franco-Pérez the occurrence and frequency of delta rhythms
and Paz 2009). In addition, generation of intense and the transition from slow delta bursting to fast,
bursts from interneurons appeared to underlie the continuous firing (Williams et al. 1997, Fig. 1).
slow GABAB receptor-mediated IPSP separating Using multicompartment models capturing the
the active phase of each delta period, perhaps dendritic structure of thalamocortical cells, it
through volume conduction of GABA overspill has been shown that a distal dendritic location
from intensely active synapses and/or of these two currents was optimal for delta gen-
neurogliaform neuronal output (Oláh eration (Lytton et al. 1996) – a prediction with
et al. 2009). However, a remarkable division of interesting consequences when considering the
Delta Rhythms: Models

and Physiology,
Fig. 1 Comparison of
single-cell
electrophysiology with
computational model of
thalamic delta oscillations.
(a) and (b) show the upper
and lower limits of
existence of the window
current-generated delta
rhythm in thalamocortical
cells (experiment) and
computational model.
Frequency ranged from
ca. 0.6–1.6 Hz. Scale bars
20 mV, 4 s (Adapted from
William et al. 1997, Fig. 7)
Delta Rhythms: Models Model Experiment
and Physiology,
Fig. 2 Comparison of a
electrophysiology with
field
network computational
model of neocortical delta
oscillations. Extracts of
b
field (a) and single neuron
subtype behavior in
a columnar model of local
circuit interactions. The D
ca. 2 Hz rhythm in layer IB
5 field recordings was
accompanied by intense c
spike bursting in
intrinsically bursting (IB,
layer 5 pyramids (b))cells,
fast spike bursts in basket FS1
cells (FS1), and intense
spike bursts and theta-
frequency single spiking in
model neurogliaform cells
(FS2, c). In contrast layer
5 regular spiking cells
FS2
(RS) generated packets of
theta-frequency single
spike outputs phase locked d
to the delta rhythm (d).
Scale bars 30 mV, 600 ms
(Adapted for Carracedo
et al. 2013) RS
proximal vs. distal location of inputs to these specific conductance expression and (b) the man-
neurons from ascending sensory streams and ifestation of the neocortical delta rhythm as an
feedback corticothalamic neurons. emergent property of local networks requires
Further development of these thalamocortical a more complex model structure to capture bio-
models included interaction with GABAergic logical realism. The cortical column model of
neurons in thalamic reticular neurons. Neurons Traub et al. (2005) was modified to generate the
in the nucleus reticularis also show burst behavior delta computational model in Carracedo
at hyperpolarized membrane potentials with et al. 2013. The model predicted the experimental
dependence on IT. Coupling such neurons with derivation of the core delta rhythm circuit with
model GABAergic synapses generated robust high fidelity in a circuit consisting of 2000 layer
oscillations in the spindle frequency range 5 intrinsically bursting neurons, coupled via
(7–14 Hz) (Destexhe et al. 1998) and provides NMDA receptors and gap junctions, driving
a compelling basic structure, when coupled to GABAB receptor-mediated feedback dendritic
thalamocortical neurons, for interaction between inhibition through mixed glutamatergic excita-
spindling and delta rhythms seen in vivo. tion of “neurogliaform” interneurons (fast spik-
ing interneurons modified to be delayed spiking
Neocortical Models via somatodendritic “A”-type potassium
In contrast to the dominance of intrinsic single- conductance).
cell conductance interactions for the thalamic In this manner, the model reproduced domi-
generator, a combination of (a) compartment- nant, delta-locked spike outputs in intrinsically
bursting and neurogliaform cells. The behavior References

was initiated by predominantly NMDA receptor-
mediated recurrent excitatory connections Amzica F, Steriade M (1998) Electrophysiological corre-
lates of sleep delta waves. Electroencephalogr Clin
between intrinsically bursting cells and termi-
Neurophysiol 107:69–83
nated by a combination of the strong intrinsic Amzica F, Nuñez A, Steriade M (1992) Delta frequency
afterhyperpolarization (SK-type potassium (1–4 Hz) oscillations of perigeniculate thalamic neu-
channel-mediated) in this neuron subtype and rons and their modulation by light. Neuroscience
51:285–294
a phasic GABAB receptor-mediated slow inhibi-
Armstrong-James M, Fox K (1988) Evidence for a specific
tory postsynaptic potential. It also suggested the role for cortical NMDA receptors in slow-wave sleep.
layer 5 nested theta generator was mediated by Brain Res 451:189–196
a combination of regular spiking layer 5 neuron Bazhenov M, Timofeev I, Steriade M, Sejnowski TJ
(2002) Model of thalamocortical slow-wave sleep
intrinsic properties and their appropriately
oscillations and transitions to activated States.
weighted synaptic inputs (Fig. 2). J Neurosci 22:8691–8704
Beltramo R, D’Urso G, Dal Maschio M, Farisello P,
Combined Models Bovetti S, Clovis Y, Lassi G, Tucci V, De Pietri J,
Tonelli D, Fellin T (2013) Layer-specific excitatory
Currently no model has been published combin-
circuits differentially control recurrent network
ing the biophysical detail of both the cortical and dynamics in the neocortex. Nat Neurosci 16:227–234
thalamic delta generators highlighted above. Campbell IG (2009) EEG recording and analysis for sleep
However, higher-level simulations using inte- research. Curr Protoc Neurosci Chapter 10:Unit10.2
Carracedo LM, Kjeldsen H, Cunnington L, Jenkins A,
grate and fire neurons have been generated and
Schofield I, Cunningham MO, Davies CH, Traub RD,
used to understand the emergence of behavior Whittington MA (2013) A neocortical delta rhythm
typical of slow-wave sleep, again implicating facilitates reciprocal interlaminar interactions via
low-threshold bursting neuron in controlling the nested theta rhythms. J Neurosci 33:10750–10761
Church MW (1975) Changes in frequency and amplitude
global thalamocortical network state (Destexhe
of delta activity during sleep. Electroencephalogr Clin
2009). In general such models reveal much reci- Neurophysiol 39:1–7
procity in thalamocortical interaction: Cortical Contreras D, Destexhe A, Sejnowski TJ, Steriade
projections to thalamus play a vital role in stabi- M (1996) Control of spatiotemporal coherence of
a thalamic oscillation by corticothalamic feedback.
lizing thalamic output, as seen in experiment
Science 274:771–774
(Contreras et al. 1996), and conversely thalamic Crunelli V, Hughes SW (2010) The slow (<1 Hz) rhythm
projections to the cortex are vital in controlling of non-REM sleep: a dialogue between three cardinal
neuronal response state (Bazhenov et al. 2002). oscillators. Nat Neurosci 13:9–17
Csercsa R, Dombovári B, Fabó D, Wittner L, Eross L,
While increasingly biophysically accurate
Entz L, Sólyom A, Rásonyi G, Szucs A, Kelemen A,
models are being generated, similar basic princi- Jakus R, Juhos V, Grand L, Magony A, Halász P,
ples become apparent: The detailed biophysical Freund TF, Maglóczky Z, Cash SS, Papp L,
thalamocortical model of Hill and Tononi (2005) Karmos G, Halgren E, Ulbert I (2010) Laminar analy-
sis of slow wave activity in humans. Brain
reiterated the importance of hyperpolarization of
13:2814–2829
neurons (in their case via an increase in potas- Destexhe A (2009) Self-sustained asynchronous irregular
sium leak conductance) in stabilizing a c.2 Hz states and Up-Down states in thalamic, cortical and
delta rhythm. In addition, in the cortex and thal- thalamocortical networks of nonlinear integrate-and-
fire neurons. J Comput Neurosci 27:493–506
amus, a vital role for Ih was also evident.
Destexhe A, Neubig M, Ulrich D, Huguenard J (1998)
Dendritic low-threshold calcium currents in thalamic
relay cells. J Neurosci 18:3574–3588
Destexhe A, Hughes SW, Rudolph M, Crunelli V (2007)
Cross-References Are corticothalamic ‘up’ states fragments of wakeful-
ness? Trends Neurosci 30:334–342
Fell J, Elfadil H, Röschke J, Burr W, Klaver P, Elger CE,
▶ Low Frequency Oscillations (Anesthesia and Fernández G (2002) Human scalp recorded sigma
Sleep): Overview activity is modulated by slow EEG oscillations during
▶ Sleep, Neural Population Models of deep sleep. Int J Neurosci 112:893–900
Franco-Pérez J, Paz C (2009) Quinine, a selective gap EEG study of sleep slow waves in humans. Sleep
junction blocker, decreases REM sleep in rats. 30:1643–1657
Pharmacol Biochem Behav 94:250–254 Schroeder CE, Lakatos P (2009) The gamma oscillation:
Haider B, Duque A, Hasenstaub AR, McCormick DA master or slave? Brain Topogr 22:24–26
(2006) Neocortical network activity in vivo is gener- Staubli U, Lynch G (1990) Stable depression of potenti-
ated through a dynamic balance of excitation and inhi- ated synaptic responses in the hippocampus with
bition. J Neurosci 26:4535–4545 1–5 Hz stimulation. Brain Res 513:113–118
Hill S, Tononi G (2005) Modeling sleep and wakefulness Steriade M, Domich L, Oakson G, Deschênes M (1987)
in the thalamocortical system. J Neurophysiol The deafferented reticular thalamic nucleus generates
93:1671–1698 spindle rhythmicity. J Neurophysiol 57:260–273
Hinton GE, Zemel RS (1994) Autoencoders, minimum Steriade M, Gloor P, Llinás RR, Lopes de Silva FH, D
description length and Helmholtz free energy. In: Mesulam MM (1990) Report of IFCN committee on
Cowan JD, Tesauro G, Alspector J (eds) Advances in basic mechanisms. Basic mechanisms of cerebral
neural information processing systems 6. Morgan rhythmic activities. Electroencephalogr Clin
Kaufmann, San Mateo, pp 3–10 Neurophysiol 76:481–508
Hinton GE, Dayan P, Frey BJ, Neal RM (1995) The Takeuchi D, Hirabayashi T, Tamura K, Miyashita
“wake-sleep” algorithm for unsupervised neural net- Y (2011) Reversal of interlaminar signal between sen-
works. Science 268:1158–1161 sory and memory processing in monkey temporal cor-
Huber R, Ghilardi MF, Massimini M, Tononi G (2004) tex. Science 331:1443–1447
Local sleep and learning. Nature 430:78–81 Tononi G, Cirelli C (2003) Sleep and synaptic homeosta-
Hughes SW, Cope DW, Tóth TI, Williams SR, Crunelli sis: a hypothesis. Brain Res Bull 62:143–150
V (1999) All thalamocortical neurones possess Tononi G, Cirelli C (2006) Sleep function and synaptic
a T-type Ca2+ ‘window’ current that enables the homeostasis. Sleep Med Rev 10:49–62
expression of bistability-mediated activities. Tóth T, Crunelli V (1992) Computer simulation of the
J Physiol 517:805–815 pacemaker oscillations of thalamocortical cells.
Ioannides AA, Kostopoulos GK, Liu L, Fenwick PB Neuroreport 3:65–68
(2009) MEG identifies dorsal medial brain activations Traub RD, Wong RK, Miles R, Michelson H (1991)
during sleep. Neuroimage 44:455–468 A model of a CA3 hippocampal pyramidal neuron
Lakatos P, Karmos G, Mehta AD, Ulbert I, Schroeder CE incorporating voltage-clamp data on intrinsic conduc-
(2008) Entrainment of neuronal oscillations as tances. J Neurophysiol 66:635–650
a mechanism of attentional selection. Science Traub RD, Contreras D, Cunningham MO, Murray H,
320:110–113 LeBeau FEN, Roopun A, Bibbig A, Wilent WB,
Larson J, Wong D, Lynch G (1986) Patterned stimulation Higley MJ, Whittington MA (2005) Single-column
at the theta frequency is optimal for the induction of thalamocortical network model exhibiting gamma
hippocampal long-term potentiation. Brain Res oscillations, sleep spindles and epileptogenic bursts.
368:347–350 J Neurophysiol 93:2194–2232
Lytton WW, Destexhe A, Sejnowski TJ (1996) Control of Vassalli A, Dijk DJ (2009) Sleep function: current questions
slow oscillations in the thalamocortical neuron: and new approaches. Eur J Neurosci 29:1830–1841
a computer model. Neuroscience 70:673–684 Vyazovskiy VV, Cirelli C, Tononi G, Tobler I (2008) Cor-
McCormick DA, Pape HC (1990) Properties of tical metabolic rates as measured by 2-deoxyglucose-
a hyperpolarization-activated cation current and its uptake are increased after waking and decreased after
role in rhythmic oscillation in thalamic relay neurones. sleep in mice. Brain Res Bull 75:591–597
J Physiol 431:291–318 Vyazovskiy VV, Olcese U, Hanlon EC, Nir Y, Cirelli C,
Oláh S, F€ule M, Komlósi G, Varga C, Báldi R, Barzó P, Tononi G (2011) Local sleep in awake rats. Nature
Tamás G (2009) Regulation of cortical microcircuits 472:443–447
by unitary GABA-mediated volume transmission. Walter WG (1936) The location of cerebral tumors by
Nature 461:1278–1281 electroencephalography. Lancet 2:305–308
Peyrache A, Khamassi M, Benchenane K, Wiener SI, Weiss S, Rappelsberger P (2000) Long-range EEG syn-
Battaglia FP (2009) Replay of rule-learning related chronization during word encoding correlates with
neural patterns in the prefrontal cortex during sleep. successful memory performance. Brain Res Cogn
Nat Neurosci 12:919–926 Brain Res 9:299–312
Pirchio M, Turner JP, Williams SR, Asprodini E, Crunelli Williams SR, Turner JP, Hughes SW, Crunelli V (1994)
V (1997) Postnatal development of membrane proper- On the nature of anomalous rectification in
ties and delta oscillations in thalamocortical neurons of thalamocortical neurones of the cat ventrobasal thala-
the cat dorsal lateral geniculate nucleus. J Neurosci mus in vitro. J Physiol 505:727–747
17:5428–5444 Xu TX, Yao WD (2010) D1 and D2 dopamine receptors in
Riedner BA, Vyazovskiy VV, Huber R, Massimini M, separate circuits cooperate to drive associative long-
Esser S, Murphy M, Tononi G (2007) Sleep homeo- term potentiation in the prefrontal cortex. Proc Natl
stasis and cortical synchronization: III. A high-density Acad Sci 107:16366–16371
D 992 Dendritic Arithmetic
discuss several examples of dendritic integration

Dendritic Arithmetic focusing primarily on pyramidal neurons, as
these cells comprise the main processing class
▶ Dendritic Computation in the brain (Spruston 2008).
Factors That Influence Dendritic Integration

The ways in which incoming inputs are combined
Dendritic Computation within dendritic branches depend mainly on two
factors: (a) the morphology of the dendritic tree
Panayiota Poirazi and (b) the repertoire of local ionic conductances.
Foundation for Research and Technology-Hellas Dendritic morphology together with ionic mech-
(FORTH), Institute of Molecular Biology and anisms whose conductance does not depend on
Biotechnology (IMBB), Heraklion, Crete, membrane voltage shapes the “passive” proper-
Greece ties of neurons, while voltage-dependent mecha-
nisms contribute to the “active” properties of
neurons. The following sections discuss exam-
Synonyms ples of how passive and active mechanisms
shape dendritic properties and in turn the output
Dendritic arithmetic; Dendritic operations; of the neuron.
Linear/nonlinear dendritic integration
Passive Properties
The traditional view of dendritic processing con-
Definition sidered dendrites as passive devices whose pri-
mary function is to transmit signals towards the
Dendritic computations refer to the ability of indi- soma. The extended morphology of the dendritic
vidual dendritic branches to perform elementary tree often seen in multiple neuron types was sim-
computations on incoming signals. These compu- ply considered as a means of increasing the sur-
tations include addition, subtraction, multiplica- face area for receiving synaptic inputs. This
tion, and division and can also take the form of simplified view of dendrites resulted in their
logical operations (e.g., AND, OR, XOR). In this omission from early neuronal models, whereby
chapter, we discuss evidence regarding the type of all processing was done at the cell body. Specif-
computations that have been observed in dendritic ically, single neurons were treated as thresholded
branches of multiple neuron types as well as the summation devices that generated an axon poten-
impact of these computations on the function of tial when the additive combination of incoming
neural cells and circuits. signals reached a certain threshold (McCulloch
and Pitts 1943), completely disregarding the
function of dendrites.
Detailed Description This notion of dendrites as conduction wires
led to the formulation of “passive cable theory”
Dendrites are elaborate processes that extend which describes how electrical signals travel
from the cell body of neurons and form the from their point of origin in a particular dendrite
locus of synaptic communication between cells. towards other parts of the neuron. A few years
The biophysical mechanisms that reside in den- later, applications of the cable theory revealed
drites allow them to locally integrate synaptic that the complex geometry of dendrites in com-
inputs in a linear and/or nonlinear manner bination with their passive properties enables
enabling them to perform different types of neurons to integrate their inputs in highly
computations (London and H€ausser 2005; nonlinear ways. For example, when
Silver 2010). In the following paragraphs, we a postsynaptic potential propagates through
Dendritic Computation 993 D
a passive dendrite towards its integration site channels, whose conductance and distribution
(soma or axon), its amplitude is attenuated in differ among neuronal types, dendritic layers
a distance-dependent manner. This effect was (e.g., apical vs. basal dendrites), as well as with
first described theoretically by Wilfrid Rall back distance from the cell body. Therefore, the inte-
in 1967 (Rall 1967), whose work revealed that gration of synaptic inputs does not only depend
passive dendrites act as low-pass filters. The on the specific morphological features and pas-
diameter of dendritic branches affects the filter- sive properties of the dendritic tree but also on the
ing of synaptic inputs for two reasons: (a) Distal expression profile and kinetics of the various
dendrites are usually thinner compared to proxi- active conductances. D
mal ones, and as a result, synaptic potentials that The toolkit of active dendritic mechanisms
travel towards the soma are reduced in amplitude includes a variety of channel types, such as Na+,
(Rinzel and Rall 1974; Zador et al. 1995). This Ca2+, K+, and h channels and NMDA receptors.
attenuation can be more than 100-fold, in pyra- In general, Na+ and Ca2+ channels are considered
midal neurons (Stuart and Spruston 1998) or in as boosters of postsynaptic potentials, whereas
cerebellar Purkinje cells (Vetter et al. 2001), and both K+ and h channels are more likely to dampen
it appears to be similar for the apical and basal dendritic excitability and contribute to the control
trees, at least in layer V pyramidal neurons of action potential frequency or act as pacemaker
(Williams and Stuart 2002; Nevian et al. 2007). currents, respectively (Migliore and Shepherd
(b) A smaller dendritic diameter results in 2002; Larkum and Nevian 2008; Spruston
larger input resistance which in turn leads to 2008). Importantly, various ionic conductances
larger local EPSPs. Membrane capacitance is in the dendrites, such as sodium (Stuart and Sak-
also smaller due the reduced membrane area, mann 1994), calcium (Yuste et al. 1994), NMDA
inducing faster local depolarizations in distal (Schiller et al. 2000), or a combination of the
dendrites (Rinzel and Rall 1974; Agmon-Snir above (Wei et al. 2001; Ariav et al. 2003), have
and Segev 1993). been shown to facilitate the back- or forward-
Due to the abovementioned anatomical prop- propagation of dendritic spikes, allow the
erties, summation of multiple excitatory synaptic supralinear summation of synaptic inputs as
inputs within passive dendrites depends on the shown in Fig. 1a, b (Poirazi et al. 2003a, b; Polsky
degree of their spatial and temporal isolation: et al. 2004; Losonczy and Magee 2006), and
When inputs are spatially clustered or synchro- enhance LTP induction (Magee and Johnston
nously activated, they summate sublinearly, as 1997; Remy and Spruston 2007). In particular,
a consequence of the reduction in the ionic driv- NMDA receptors when sufficiently activated can
ing force. However, this conclusion is reversed lead to the generation of dendritic regenerative
when taking into account the presence of voltage- events termed NMDA spikes.
dependent (active) conductances that greatly Those spikes are characterized by larger local
influence local responses and counteract attenu- amplitude and duration, compared to the briefer
ating effects. Na+ spikes, but smaller amplitude compared to
Ca2+ spikes. NMDA spikes are highly localized
Active Properties events, confined to thin dendritic branches. For
Over the last decade, advances in technology led these reasons, NMDA spikes are considered an
to the discovery of an array of voltage-dependent attractive dendritic mechanism for the detection
ion channels with different densities and biophys- of incoming signals that arrive at close temporal
ical properties that lie along the axo-somato- and/or spatial proximity (Ariav et al. 2003;
dendritic axis (Spruston 2008). These findings Losonczy and Magee 2006). Interestingly, recent
revealed the key role of dendritic branches in work has assigned new roles to NMDA spike
shaping synaptic integration locally and in turn generation: (a) to extend the temporal window
the response of the neuron. Dendrites are for supralinear integration in CA1 pyramidal neu-
enriched with a variety of voltage-dependent ron dendrites in response to bursty signals
D 994 Dendritic Computation
a MODEL b EXPERIMENT
Within branch 20 Within branch

12
Measured peak EPSP [mV] [10mV]
Measured peak EPSP [mV]

10
B 15 AB
A
8
6 10
A
4
B 5 A B
2
Between branches Between branches
0 0
0 2 4 6 8 10 12 0 5 10 15 20
Expected peak EPSP [mV] [10mV] Expected peak EPSP [mV]
s(ni)
n1
αi n2 αi
y
ni n3 g
n37
Dendritic Computation, Fig. 1 Nonlinear dendritic Single-pulse stimulation of synapses in different branches
computations. (a) Summation of paired, single-pulse results in linear summation at the cell body (green
inputs in the apical dendrites of a CA1 pyramidal model squares). When g-aminobutyric acid (GABA)ergic inhi-
cell. Simulations predict a sigmoidal modulation of com- bition is blocked, the summation of within-branch EPSPs
bined excitatory postsynaptic potentials (EPSPs) within is modulated by a sigmoidal nonlinearity (all other sym-
a branch (red symbols) and a linear summation of EPSPs bols). (c) Schematic representation of a pyramidal neuron
between branches (green symbols). Red curves corre- as a two-layer neural network. Radial oblique dendrites
spond to within-branch data for dendrites at 92 mm provide the first layer of the network, each performing
(short dashes), 232 mm (solid), and 301 mm (long dashes) individually thresholded computations as shown in (a) and
from the soma. Due to differences in local compared with (b). The outputs of this layer feed into the cell body, which
somatic responses, axis values for the red curves are constitutes the second layer of the network model
scaled up by a factor of 10. (b) Experimental verification (Reproduced with permission from Sidiropoulou
of predicted summation rules in basal dendrites of a layer et al. (2006))
V pyramidal neuron as reported in Polsky et al. (2004).
(Gomez et al. 2011) and (b) to serve as glutamatergic inputs in layer V pyramidal neu-
a mechanism for the detection of “preferred” rons is received by their basal, oblique, and apical
sensory signals, namely, signals that will induce tuft dendrites that can give rise to regenerative
persistent activity in a single-layer V PFC pyra- NMDA events (Antic et al. 2010).
midal neuron (Sidiropoulou and Poirazi 2012). Although such regenerative events can be con-
Importantly, the majority of the excitatory fined to their dendritic site of origin (Schiller
Dendritic Computation 995 D
et al. 2000; Wei et al. 2001), physiologically types and add spatiotemporal specificity to their
relevant situations such as large, suprathreshold functional output (Stuart et al. 2008, 2010; Silver
synaptic stimuli in the distal dendrites or the 2010). In pyramidal neurons, synaptic inputs are
co-activation of several branches allow these integrated linearly when they are located
spikes to act globally and modulate neuronal on different dendritic branches and sigmoidally
output (Larkum et al. 2001). Thus, dendritic when located within the same dendrite (Poirazi
events emerge as key players in shaping neuronal et al. 2003a; Poirazi et al. 2003b; Polsky
responses. et al. 2004; Losonczy and Magee 2006), but
they could also be integrated sublinearly, D
Dendritic Morphology depending on the spatial distribution of incoming
While extensive experimental and theoretical signals (Oviedo and Reyes 2012). In CA1 pyra-
work has characterized the arithmetic of synaptic midal neurons, dendrites produce two distinct
summation in pyramidal neurons, the role of den- modes of integration: Asynchronous or spatially
dritic morphology in this process remains poorly distributed synaptic inputs are integrated linearly,
understood. Electrophysiological characteriza- while synchronous and/or clustered inputs are
tion with current injection has shown that neurons integrated in a sigmoidal manner (Cash and
with a more complicated morphology (in terms of Yuste 1999; Gasparini and Magee 2006). These
larger and more elaborate dendritic trees, more phenomena are mediated by the generation of
branches, extensive bifurcations, and more dendritic spikes that are locally confined.
spines) display bursting firing patterns, whereas The combination of a complex dendritic mor-
neurons with simpler morphologies tend to fire phology with a location-specific distribution of
tonic spikes (Bilkey and Schwartzkroin 1990; ionic mechanisms results in the formation of dis-
Chagnac-Amitai et al. 1990; Yang et al. 1996). tinct dendritic subunits which can integrate inputs
Simulations of reconstructed neurons are in in a quasi-independent manner (Poirazi et al.
accordance with the in vitro studies (Mainen 2003a, b; Polsky et al. 2004; Losonczy and
and Sejnowski 1996; Sheasby and Fohlmeister Magee 2006), thus enabling the cell to integrate
1999; Krichmar et al. 2002) and further propose inputs like a two-layer neural network. This abil-
that dendritic trees with asymmetric branch con- ity to integrate inputs like a two-layer neural
nections are associated with bursting (van Elburg network, graphically illustrated in Fig. 1c, boosts
and van Ooyen 2010). the computational capacity of the neuron by at
The branching pattern of the dendritic tree also least an order of magnitude compared to that of
influences both the forward and the a thresholding point neuron (Poirazi and Mel
backpropagation of signals (Vetter et al. 2001) 2001). Recent experimental evidence shows that
while at the same time shaping coincidence the distribution of synapses both along the
detection (Schaefer et al. 2003). Even the area somatodendritic axis and along the oblique
of the bifurcation tapering membrane surround- branches in CA1 pyramidal neurons favors a
ing a branch point can determine neuronal excit- two-stage integration model (Katz et al. 2009).
ability, through its effects on the generation, At a smaller scale, single dendrites of layer
propagation, and timing of dendritic spikes II/III and layer V pyramidal neurons exhibit
(Ferrante et al. 2013). Taken together, the elabo- nonuniform synaptic integration modes along
rate morphology of the dendritic tree seems to their length. Synaptic inputs in the distal end of
play a critical role in shaping dendritic as well as the dendrite summate sigmoidally with a low
neuronal responses. threshold for dendritic spike generation, while
inputs to the proximal end of the same dendrite
Importance and Functional Implications of summate linearly and exhibit a higher threshold
Dendritic Computations for dendritic spikes (Major et al. 2008; Branco
The various active properties of dendrites expand and Hausser 2011). These and previous data indi-
the computational capabilities of various neuron cate that the spatial arrangement of synaptic
D 996 Dendritic Computation
inputs within dendritic branches greatly influ- Bilkey D, Schwartzkroin PA (1990) Variation in electro-
ences local responses by differentially engaging physiology and morphology of hippocampal CA3
pyramidal cells. Brain Res 405:320–325
local conductances (Larkum and Nevian 2008). Branco T, Hausser M (2010) The single dendritic branch
For example, the co-activation of nearby synap- as a fundamental functional unit in the nervous system.
ses, also termed “synaptic clustering,” leads to Curr Opin Neurobiol 20(4):494–502
signal amplification via the induction of dendritic Branco T, Hausser M (2011) Synaptic integration gradi-
ents in single cortical pyramidal cell dendrites. Neuron
spikes, resulting in a much larger somatic 69(5):885–892
response than the one obtained when the same Cash S, Yuste R (1999) Linear summation of excitatory
number of synapses is stimulated in a distributed inputs by CA1 pyramidal neurons. Neuron
manner. As dendritic depolarization and den- 22:383–394
Chagnac-Amitai Y, Luhmann HJ, Prince DA (1990) Burst
dritic spikes can have a strong effect on LTP generating and regular spiking layer 5 pyramidal neu-
induction, clustered synapses are also more likely rons of rat neocortex have different morphological
to undergo stronger potentiation than diffusely features. J Comp Neurol 296:598–613
activated inputs (Hardie and Spruston 2009). Ferrante M, Migliore M, Ascoli GA (2013) Functional
impact of dendritic branch-point morphology.
Overall, these findings show that dendritic J Neurosci 33(5):2156–2165
arithmetic obeys complex rules that are often Gasparini S, Magee JC (2006) State-dependent dendritic
not only brain region but also branch-type computation in hippocampal CA1 pyramidal neurons.
specific. This ability of dendrites to perform com- J Neurosci 26(7):2088–2100
Gomez JM, Mel BW, Poirazi P (2011) Distinguishing
plex local computations has led to the proposal linear vs. nonlinear integration in CA1 radial oblique
that dendritic branches (as opposed to single syn- dendrites: it’s about time. Front Comp Neurosci 5:44
apses or whole neurons) should be considered as Epub 2011 Nov 14
the lowest-level processing unit of the nervous Hardie J, Spruston N (2009) Synaptic depolarization is
more effective than back-propagating action potentials
system (Branco and Hausser 2010). If proven, during induction of associative long-term potentiation
this hypothesis will change our current view of in hippocampal pyramidal neurons. J Neurosci
information processing and storage in the brain 29(10):3233–3241
and will open new avenues for experimental Katz Y, Menon V, Nicholson DA, Geinisman Y, Kath
WL, Spruston N (2009) Synapse distribution suggests
investigations regarding the cellular substrate of a two-stage model of dendritic integration in CA1
learning and memory. pyramidal neurons. Neuron 63(2):171–177
Krichmar J, Nasuto SJ, Scorcioni R, Washington SD,
Ascoli GA (2002) Effects of dendritic morphology
on CA3 pyramidal cell electrophysiology:
a simulation study. Brain Res 941:11–28
Cross-References Larkum M, Nevian T (2008) Synaptic clustering by den-
dritic signalling mechanisms. Curr Opin Neurobiol
▶ Cable Theory: Overview 18:321–331
Larkum ME, Zhu JJ et al (2001) Dendritic mechanisms
underlying the coupling of the dendritic with the axo-
nal action potential initiation zone of adult rat layer
5 pyramidal neurons. J Physiol 533(Pt 2):447–466
References London M, H€ausser M (2005) Dendritic computation.
Annu Rev Neurosci 28:503–532
Agmon-Snir H, Segev I (1993) Signal delay and input Losonczy A, Magee JC (2006) Integrative properties of
synchronization in passive dendritic structures. radial oblique dendrites in hippocampal CA1 pyrami-
J Neurophysiol 70(5):2066–2085 dal neurons. Neuron 50(2):291–307
Antic SD, Zhou W-L et al (2010) The decade of the Magee JC, Johnston D (1997) A synaptically controlled,
dendritic NMDA spike. J Neurosci Res 88(14): associative signal for Hebbian synaptic plasticity in
2991–3001 hippocampal neurons. Science 275:209–213
Ariav G, Polsky A et al (2003) Submillisecond precision Mainen ZF, Sejnowski TJ (1996) Influence of dendritic
of the input–output transformation function structure on firing pattern in model neocortical neu-
mediated by fast sodium dendritic spikes in basal den- rons. Nature 382:363–366
drites of CA1 pyramidal neurons. J Neurosci 23(21): Major G, Polsky A et al (2008) Spatiotemporally graded
7750–7758 NMDA spike/plateau potentials in basal dendrites of
Dendritic Spines 997 D
neocortical pyramidal neurons. J Neurophysiol Stuart G, Spruston N (1998) Determinants of voltage
99(5):2584–2601 attenuation in neocortical pyramidal neuron dendrites.
McCulloch W, Pitts W (1943) A logical calculus of the J Neurosci 18(10):3501–3510
ideas immanent in nervous activity. Bull Math Stuart G, Spruston N, Hausser M (2008) Dendrites.
Biophys 5:115–133 Oxford University Press, Oxford
Migliore M, Shepherd GM (2002) Emerging rules for the van Elburg RA, van Ooyen A (2010) Impact of dendritic
distributions of active dendritic conductances. Nat Rev size and dendritic topology on burst firing in pyramidal
Neurosci 3(5):362–370 cells. PLoS Comput Biol 6(5)
Nevian T, Larkum ME et al (2007) Properties of basal Vetter P, Roth A, H€ausser M (2001) Propagation of action
dendrites of layer 5 pyramidal neurons: a direct patch- potentials in dendrites depends on dendritic morphol-
clamp recording study. Nat Neurosci 10(2):206–214 ogy. J Neurophysiol 85(2):926–937 D
Oviedo H, Reyes AD (2012) Integration of subthreshold Wei DS, Mei YA, Bagal A, Kao JP, Thompson SM, Tang
and suprathreshold excitatory barrages along the CM (2001) Compartmentalized and binary behavior of
somatodendritic axis of pyramidal neurons. PLoS terminal dendrites in hippocampal pyramidal neurons.
One 7(3) Science 293(5538):2272–2275
Poirazi P, Mel BW (2001) Impact of active dendrites and Williams SR, Stuart GJ (2002) Dependence of EPSP effi-
structural plasticity on the memory capacity of neural cacy on synapse location in neocortical pyramidal
tissue. Neuron 29(3):779–796 neurons. Science 295(5561):1907–1910
Poirazi P, Brannon T et al (2003a) Pyramidal neuron as Yang CR, Seamans JK, Gorelova N (1996) Electrophysi-
two-layer neural network. Neuron 37(6):989–999 ological and morphological properties of layers V-VI
Poirazi P, Brannon T, Mel BW (2003b) Arithmetic of principal pyramidal cells in rat prefrontal cortex
subthreshold synaptic summation in a model CA1 in vitro. J Neurosci 16(5):1904–1921
pyramidal cell. Neuron 37:977–987 Yuste R, Gutnick MJ et al (1994) Ca2+ accumulations in
Polsky A, Mel BW et al (2004) Computational subunits in dendrites of neocortical pyramidal neurons: an apical
thin dendrites of pyramidal cells. Nat Neurosci band and evidence for functional compartments. Neu-
7(6):621–627 ron 13:23–43
Rall W (1967) Distinguishing theoretical synaptic poten- Zador AM, Agmon-Snir H, Segev I (1995) The
tials computed for different soma-dendritic distribu- morphoelectrotonic transform: a graphical approach
tions of synaptic input. J Neurophysiol to dendritic function. J Neurosci 15(3 Pt 1):1669–1682
30(5):1138–1168
Remy S, Spruston N (2007) Dendritic spikes induce
single-burst long-term potentiation. Proc Natl Acad
Sci USA 104(43):17192–17197
Rinzel J, Rall W (1974) Transient response in a dendritic Dendritic Operations
neuron model for current injected at one branch.
Biophys J 14(10):759–790
▶ Dendritic Computation
Schaefer A, Larkum ME, Sakmann B, Roth A (2003)
Coincidence detection in pyramidal neurons is tuned
by their dendritic branching pattern. J Neurophysiol
89:3143–3154
Schiller J, Koester H, Major G, Schiller Y (2000) NMDA
spikes in basal dendrites of cortical pyramidal neurons.
Dendritic Spines
Nature 404:285–289
Sheasby B, Fohlmeister JF (1999) Impulse encoding William R. Holmes
across the dendritic morphologies of retinal ganglion Department of Biological Sciences,
cells. J Neurophysiol 81:1685–1698
Ohio University, Athens, OH, USA
Sidiropoulou K, Poirazi P (2012) Predictive features of
persistent activity emergence in regular spiking and
intrinsic bursting model neurons. PLoS Comp Biol 8
(4):e1002489. doi:10.1371/journal.pcbi.1002489 Definition
Sidiropoulou K, Pissadaki EK et al (2006) Inside the brain
of a neuron. EMBO Rep 7(9):886–892
Silver RA (2010) Neuronal arithmetic. Nat Rev Neurosci Many types of neurons receive synaptic inputs on
11(7):474–489 dendritic spines which are short protrusions of
Spruston N (2008) Pyramidal neurons: dendritic structure membrane composed of a bulbous “head”
and synaptic integration. Nat Rev Neurosci 9(3):206–221
connected to the dendrite by a thin “stem” or
Stuart G, Sakmann B (1994) Active propagation of
somatic action potentials into neocortical pyramidal “neck” (Yuste 2010). Although spines are typi-
cell dendrites. Nature 367:69–72 cally small in size, they occur in high densities
D 998 Dendritic Spines
along dendrites and may compose 40–60 % of the implicitly modeled by multiplying the dendrite
total dendritic surface area. There are two Cm by 1.33 and dividing dendritic Rm by 1.33.
straightforward methods for including spines in Alternatively, one can keep Rm and Cm constant
computational models of neurons without but increase length and diameter in such a way as
representing every spine explicitly. to keep intracellular resistance constant (i.e.,
keep length/diameter2 constant). We call this
the l-d method. For example, if spines increase
Detailed Description membrane area by 33 %, let F = 1.33 and multi-
ply length by F2/3 and diameter by F1/3.
Spines are widely regarded to serve These two methods for including spines in
a biochemical function by providing an isolated models implicitly provide identical results with
compartment where highly localized calcium sig- passive models, but they make very different
nals and subsequent reactions can occur. An elec- assumptions when voltage-dependent conduc-
trical function was initially proposed by Rall who tances are present. The first method, where Rm
showed that the steady-state attenuation of and Cm are changed, assumes that spines have
a signal from the spine head to the dendrite passive membrane, while the second method,
could be expressed as VSH/VBI = 1 + RSS/RBI, where length and diameter are changed, assumes
where VSH and VBI are the voltages at the spine that the spines have the same densities of voltage-
head and dendrite, respectively; RSS is the spine dependent conductances as the dendrite. The cell
stem resistance; and RBI is the input resistance at should have a lower input resistance with the
the dendrite. Rall suggested that changes in spine second method but should also respond more
stem diameter might be important for plasticity nonlinearly with voltage perturbations. Needless
because of its effect on RSS, but studies searching to say, results may be very different.
for such changes produced mixed results. Others Nevertheless, it is straightforward to modify
have suggested that if voltage-dependent conduc- either method to include spines possessing differ-
tances are present on spines, then spines might ent densities of voltage-dependent conductances
amplify inputs and could otherwise increase the than present on the dendrite. With the first
computational possibilities of the cell. Many of method, one should change dendritic Rm and Cm
the more interesting possibilities suggested by as described above but also modify the maximum
theoretical studies require a large RSS, but
conductance density (g-bar)
for each conduc-
whether or not RSS is sufficiently large has been tance by gd-new ¼ gd-old Ad þ gspine As =Ad
controversial. where Ad is the membrane area of the dendritic
Because the proportion of membrane area segment and As is the membrane area of all spines
occupied by spines can be large, single-neuron on that segment. With the second method, one
models that ignore spines are bound to produce should modify the maximum conductance den-
erroneous results. There are two straightforward sity for every conductance on each dendrite by

methods for including spines in models without gd-new ¼ gd-old Ad þ gspine As =ðAd þ As Þ and
coding each and every spine, as long as these
implicitly included spines do not receive synaptic then change length and diameter of the dendrite
input. First, if dendritic and spine membranes are as described above. These expressions assume
assumed to have the same specific membrane spine head and neck have the same channel den-
resistance and capacitance, Rm and Cm, spines sities. If this is not true, one could calculate an
can be implicitly included in a model by increas- equivalent gspine to use.
ing Cm and reducing Rm according to the propor- These methods are easily incorporated into
tion of total membrane area contributed by single-neuron simulators such as NEURON or
spines. We call this the Rm–Cm method. For GENESIS and work well for voltage-dependent
example, if spines increase the membrane area conductances with fixed reversal potentials.
of a dendrite by 33 %, then spines can be However, these methods are subject to error
Dendritic Spines: Continuum Theory 999 D
when calcium and calcium-dependent conduc- classical cable theory for which the distribution of
tances are present and particularly so if these spines is treated as a continuum. The theory
conductances have different densities in spines applies when the interspine distance is much less
and dendrites. Calcium accumulation and than the length scale of the dendrite, for example,
removal can be different in spines and dendrites, when the dendrite is populated by a large number
and different intracellular calcium concentrations of spines. The formulation maintains the basic
will affect the calcium Nernst potential and acti- feature that there is no direct coupling between
vation of calcium-dependent conductances dif- neighboring spines; voltage spread along dendrites
ferently. The Rm–Cm method is preferred if is the only way for spines to interact. With the D
dendrites have calcium conductances and spines continuum theory, different spine morphologies,
do not. The l-d method is preferred if calcium multiple populations of spines, and distributed
conductances are on both dendrites and spines, physiological properties are represented explicitly
but activation of calcium-dependent potassium and compactly by relatively few differential equa-
conductances in spines remains problematic. It tions. The theory is general so that idealized or
is possible to compensate for these types of errors realistic kinetic models may be adapted.
by adjusting calcium pool parameters or by con-
sidering spine calcium and calcium-dependent
conductances as separate conductances with Detailed Description
their own parameters when conductances are
merged into dendrites. These errors are not likely Classical cable theory has provided valuable
to be significant in short-time simulations but can insights into structure-function relationships
be seen to accumulate over longer time periods. and guidance in the design of experiments. Per-
haps the best example is Rall and Shepherd’s
electrotonic model of the olfactory bulb,
References which predicted the existence of functional
dendrodendritic synapses (Rall and Shepherd
Yuste R (2010) Dendritic spines. MIT Press, Cambridge,
1968). Cable equations have been used success-
MA
fully to study tapered nerves and extensively
branched dendritic trees (Jack et al. 1975; Rall
1977), as well as cables with nonuniform electri-
Dendritic Spines: Continuum Theory cal parameters such as specific membrane resis-
tivity (Holmes and Woody 1989).
Steven M. Baer In the 1970s and 1980s, Rall, Rinzel, Miller,
School of Mathematical and Statistical Sciences, and Segev spearheaded efforts that accelerated
Arizona State University, Tempe, AZ, USA the theoretical investigation into the function of
dendritic spines (Rall and Rinzel 1971a, b; Miller
et al. 1985; Shepherd et al. 1985; Rall and Segev
Synonyms 1987; Segev and Rall 1988). Dendritic spines are
small knoblike evaginations of the dendritic sur-
Baer-Rinzel (BR) continuum model; Continuum face (see Fig. 1a, b), and a single neuron may
spine model; Continuum spine theory; Contin- have 105spines. These large numbers of spines
uum theory for active spines present a challenge for computational studies,
particularly when using compartmental model-
ing. To address this, Baer and Rinzel (1991)
Definition formulated an extension of classical cable theory
to model cables with large spine densities. They
The continuum theory for dendritic spines, devel- treated the distribution of spines as a continuum
oped by Baer and Rinzel (1991), is an extension of (see Fig. 1). This allowed a single cable with
D 1000 Dendritic Spines: Continuum Theory
Dendritic Spines:
Continuum Theory,
Fig. 1 Continuum of
dendritic spines. (a):
Cortical pyramidal cells
(Modified from Berkley
1897). (b): A schematic
magnification from (a)
shows the dendrite
(cylinder) studded with
spines. (c): If there are
many spines, the dendritic
electrical response can be
approximated by
considering a continuum of
spines (stippled envelope).
Although spines are treated
mathematically as
a continuum, the model is
constructed so that spines
interact only through
spread of dendritic
potential (From Baer and
Rinzel (1991), Fig. 1)
hundreds or thousands of spines to be modeled per unit electrotonic length. The spine stem cur-
with just a few differential equations. Baer and rent is denoted by Iss and represents the I R
Rinzel’s model is voltage drop across the spine stem resistance
Rss, as expressed in Eq. 3. The spine head is
@V d @ 2 V d modeled in Eq. 2 as an isopotential compartment
tm ¼ V d þ R1 ni I ss (1) with surface area Ash(mm2) and specific mem-
@t @X2
brane capacitance Cm(mF/cm2); individual spines
have a capacitance of Csh = AshCm (mF). Equa-
@V sh
Csh ¼ I ion I syn I ss , (2) tion 2 describes the membrane potential in a sin-
@t gle spine obtained from a current balance relation
for the capacitive, ionic, spine stem, and synaptic
where
currents.
The primary focus of Baer and Rinzel’s study
V sh V d
I ss ¼ : (3) was on threshold properties, e.g., minimum num-
Rss ber of synaptically activated spines, or minimal
density of spines, for the initiation and spread of
Here, Vsh and Vd are respectively the mem- activity. They found that propagation was pre-
brane potential in the spine head and the dendritic cluded for spine stem resistance (Rss) either too
base (or shaft). Equation 1 is the cable equation in large or too small (Fig. 2). Moreover, even if Rss
terms of dimensionless (electrotonic) length was in a suitable range for the local generation of
X (computed for the passive cable without an action potential (resulting from local synaptic
spines), tm is the membrane time constant, and excitatory input), the range was shown to be not
R1 is the input resistance for a semi-infinite suitable to initiate a chain reaction of spine firings
passive cable with circular cross section. The along the dendrite; success or failure of impulse
spine density n expresses the number of spines propagation depends on an even narrower range
Dendritic Spines: Continuum Theory 1001 D
Dendritic Spines: Continuum Theory, of range for the given input. Propagation also fails when
Fig. 2 Propagation failure for spine stem resistance the stem resistance is too high (long-dashed curves)
to large or too small. A dendrite of electrotonic length because large stem resistance results in a decreased shaft
3 has 60 uniformly distributed excitable spines. Four potential; the spreading dendritic depolarization becomes
spines are synaptically activated near X = 0 using an too weak to bring the spine heads in front of the wave to
alpha function stimulation. Electrode recordings in the threshold. However, for intermediate values of spine stem
spine head and dendrite are simulated at X = 0 and resistance (solid curves), a “spike-diffuse-spike” chain
X = 2. (a): Voltage transients in the head indicate propa- reaction occurs and the wave propagates. (b): Peak volt-
gation failure when the spine stem resistance is too low age versus stem resistance in spine head (dashed) and
(short-dashed curves) because dendritic loading effec- dendritic shaft (solid) (From Baer and Rinzel (1991),
tively raises the threshold for a local action potential out Fig. 3, with figure caption modified)
of Rss values. Also demonstrated was that appro- safety factor and therefore facilitate attenuating
priate clustering of spines can enhance synaptic waves and further enhance waves that are
amplification and spread of activity (Fig. 3). successfully propagating. In regions of
The Baer-Rinzel model led to some new and increased conductance load such as common
important insights into the electrical interplay branch points, attenuation of the potential
between passive and excitable membranes. For and propagation failure are more likely. Spine
example, regions of decreased conduction load densities (passive or excitable) also affect
(e.g., near sealed ends of nerves) have a higher conductive loading and safety factor. A suitable
D 1002 Dendritic Spines: Continuum Theory
Dendritic Spines: Continuum Theory, Fig. 3 Spine 3 spines and width 0.04 (n ¼ 75 within a cluster). Impulse
clusters. Impulse propagation is contingent on the spatial propagation is successful for this case. At X = 1 the
distribution of spines. Effects of 3 spatial distributions of simulated electrode recording is between spine clusters;
24 excitable spines are compared in a cable of electrotonic hence, membrane potential is shown for the dendrite only.
length 2. Simulated electrode recording positions are (c): The 24 spines are grouped into 4 clusters spaced 0.42
X = 0, 1, and 2. Solid curves denote dendritic potentials, apart. Each cluster has 6 spines and width 0.08 (again,
and dashed curves denote spine head potentials. Three n ¼ 75 within a cluster). Here the impulse propagates to
spines near X = 0 are activated synaptically. (a): With X = 1 but fails before reaching X = 2. Although the clus-
a uniform but low density of spines (n ¼ 12 spines per e.l.), ters have the same density as in (b), they are too far apart
an impulse fails to propagate. (b): The same 24 spines are for successful propagation (From Baer and Rinzel (1991),
grouped into 8 clusters spaced 0.34 apart. Each cluster has Fig. 4)
spine density could ensure propagation through spike model (Coombes and Bressloff 2003), orig-
regions where the safety factor is low. More inally intended as a computationally simple ver-
generally, propagation in regions of nonuniform sion of the Baer-Rinzel model.
loading may be complemented or counteracted
by variations in spine density; the dynamics of
electrical current flow in cables with spines are
References
dependent on spine distribution and spine
morphology. Baer SM, Rinzel J (1991) Propagation of dendritic spikes
The continuum model has been applied to mediated by excitable spines: a continuum theory.
a variety of problems, for example, the dynamics J Neurophysiol 65:874–890
Bell J, Holmes M (1992) Model of the dynamics of receptor
of receptor potentials in mechanoreceptors (Bell potential in a mechanoreceptor. Math Biosci 110:139–174
and Holmes 1992) and, more recently, the effect Coombes S, Bressloff PC (2003) Saltatory waves in the
of noise on spiny dendrites (Coutts and Lord spike-diffuse-spike model of active dendritic spines.
2013). The continuum theory of dendritic spines Phys Rev Lett 91:028102
Coutts EJ, Lord GJ (2013) Effects of noise on models of
has been extended to study spine plasticity (Verzi
spiny dendrites. J Comput Neurosci 34:245–257
et al. 2004), and it has spawned other computa- Holmes WR, Woody CD (1989) Effects of uniform and
tional approaches, including the spike-diffuse- non-uniform synaptic ‘activation-distributions’ on the
Determining Network Structure from Data: Nonlinear Modeling Methods 1003 D
cable properties of modeled cortical pyramidal neu-
rons. Brain Res 505:12–22 Determining Network Structure
Jack JJB, Noble D, Tsien RW (1975) Electric current flow
in excitable cells. Clarendon, Oxford from Data: Nonlinear Modeling
Miller JP, Rall W, Rinzel J (1985) Synaptic amplification Methods
by active membrane in dendritic spines. Brain Res
325:325–330 Bjoern Schelter1 and Marco Thiel2
Rall W (1977) Core conductor theory and cable properties 1
of neurons. In: Kandel ER, Brookhardt JM, Institute for Complex Systems and
Mountcastle VM (eds) Handbook of physiology, the Mathematical Biology, University of Aberdeen,
nervous system, cellular biology of neurons. American Aberdeen, UK D
Physiological Society, Bethesda, pp 39–97 2
Department of Physics, Institute for Complex
Rall W, Rinzel J (1971a) Dendritic spines and synaptic
potency explored theoretically. Proc Int Union Physiol Systems and Mathematical Biology, University
Sci XXV Int Congr IX:466 of Aberdeen, Aberdeen, UK
Rall W, Rinzel J (1971b) Dendritic spine function and
synaptic attenuation calculations. Prog Abstr Soc
Neurosci 1:64
Rall W, Segev I (1987) Functional possibilities for synap-
ses on dendrites and on dendritic spines. In: Edelman Definition
GM, Gall WE, Cowan WM (eds) Synaptic function.
Wiley, New York, pp 605–636 In this entry we focus on inference of network
Rall W, Shepherd GM (1968) Theoretical reconstruction
of field potentials and dendrodendritic synaptic inter- structures from data. One possible approach to
actions in olfactory bulb. J Neurophysiol 31:884–915 studying networks is to model the nodes, such as
Segev I, Rall W (1988) Computational study of an excit- neurons, and generate networks and run simula-
able dendritic spine. J Neurophysiol 60:499–523 tions and observe the network behavior. This
Shepherd GM, Brayton RK, Miller JP, Segev I, Rinzel J,
Rall W (1985) Signal enhancement in distal cortical approach requires on a priori assumptions about
dendrites by means of interactions between active den- the constituent parts; for instance, Hodgkin-
dritic spines. Proc Natl Acad Sci U S A 82:2192–2195 Huxley neurons may be coupled and the resulting
Verzi DW, Rheuben MB, Baer SM (2004) Impact of time- network behavior is investigated. The model
dependent changes in spine density and spine shape on
the input–output properties of a dendritic branch: behaviors can be compared to the measured neu-
a computational study. J Neurophysiol 93:2073–2089 ronal signals through statistical analysis. However,
this approach only provides indirect information
about the network structure. An alternative
approach to studying network structure is to use
parametric, semiparametric, and nonparametric
Depression analyses of the observed signals and reconstruct
the network connections. Such analyses are essen-
▶ Short-Term Synaptic Plasticity in Central tial tools for systems in which network structure is
Pattern Generators not known, such as when anatomical connections
are not known or information is not sufficient.
A particular challenge when inferring the net-
work structure from data lies in the fact that
analysis techniques to investigate interactions
Descending Control of Spinal are often bivariate, investigating pairwise con-
Circuitry in Voluntary Movements nections. Often, analyses cannot distinguish
between direct and indirect interactions. A good
▶ Coordinate Transformations, Role of Spinal measure of coupling between nodes should be
Circuitry in able to distinguish direct and indirect
D 1004 Determining Network Structure from Data: Nonlinear Modeling Methods
interactions. To achieve this, a multivariate data between brain areas and the periphery. The goal
analysis approach needs to be used. For linear is to detect changes in coupling that may be
systems multivariate data analysis approaches caused by the underlying diseases. Detection of
have been developed over the past decades. How- changes in coupling may lead to improved diag-
ever, for nonlinear systems, the development of nosis and treatment strategies especially for neu-
multivariate analysis techniques is in its infancy. rological diseases.
Nonlinear approaches typically require computa- The approach proposed here considers the
tionally intensive algorithms and large data sets brain as a dynamic system from which we mea-
which have limited their application. sure the activity through the electroencephalo-
In this entry, we provide a survey of gram (EEG), magnetoencephalogram (MEG), or
approaches for inferring the network structure another representation of the underlying neuronal
from nonlinear systems using nonlinear data- activity. By applying the network analysis to the
based modeling. data sets recorded from normal patients and those
suffering from neurological diseases, we hope to
gain an understanding of the underlying mecha-
Detailed Description nisms generating these dysfunctions (Volkmann
et al. 1996; Tass et al. 1998; Hellwig et al. 2001).
Introduction However, there is a wide variety of applications
Mathematicians and physicists often generate beyond neurosciences to which the linear as well
models of complex networks starting with first as nonlinear data analysis techniques presented
principles. However, for complex biological sys- here can be applied successfully.
tems found in neuroscience, the dynamics of neu- Various linear analysis techniques have been
ronal behavior underlying the measured signals is proposed to determine interdependency between
poorly understood, leaving an approach based on dynamic processes and to determine causal influ-
first principles ideal. Therefore, an understanding ences in multivariate systems (Schelter
of the behavior can only be based upon the anal- et al. 2006a). Using the Fourier transform,
ysis of measured data of the dynamics, i.e., time a signal can be converted to the frequency
series from EEG or other population level sig- domain and the interdependencies analyzed
nals, and point processes, from action potential using the cross-spectrum and coherence. But
timing. This approach is called data-based these tools are not sufficient to adequately
modeling. describe interdependence within a multivariate
Time series and point process analysis have system where correlation arises not because the
different roots originating in mathematics, phys- two are directly coupled but because they receive
ics, and engineering. The underlying assumptions common inputs. To enable a differentiation
of the sources of the signals are different resulting between direct and indirect influences in multi-
in different approaches to the analysis. In mathe- variate systems, multivariate linear approaches
matical statistics, the model has been based on have been developed (Dahlhaus 2000; Baccala
linear stochastic systems; in physics the models and Sameshima 2001; Ding et al. 2006).
have been of nonlinear deterministic systems. Furthermore, multivariate network analysis
While methodological developments evolved can uncover directed interactions, which enables
independently across disciplines, over the past deeper insight into the basic mechanisms under-
decade cross-fertilization has resulted in novel lying such networks. These tools can determine
approaches to data-based modeling of nonlinear not only if nodes are connected but the direction
stochastic systems (Schelter et al. 2006b). of the connection. In some cases in which com-
Reconstruction of networks depends on the munication is present in both directions, it is
detection of coupling between areas, particularly possible that they can communicate in distinct
causal influences between two different pro- frequency bands. To detect directed nonlinear
cesses, for instance, between brain areas or connections, appropriate analysis techniques are
needed. Granger causality (Granger 1996) is uti- of the network structure based on the observation
lized to determine causal influences when the of nonlinear processes utilizing nonlinear
coupling is linear. This probabilistic approach to data-based modeling approaches. The entry is
determining causality is based on the principle subdivided into Phase-Based Approaches,
that a cause precedes the effect in time and this is Recurrence-Based Approaches, Information
formulated in terms of predictability. Granger Theoretic Approaches, and Linear Approaches.
causality is done by fitting autoregressive models We should note that it is impossible to provide
to the signals and determining if the error in a complete survey of all approaches.
prediction of the next time point of one signal We begin this entry by introducing the general D
can be improved by the knowledge of the other. concept underlying most of the approaches to
A graphical approach for modeling Granger- provide a quick introduction into the topic.
causal relationships in multivariate processes
has been discussed (Eichler 2007). More gener- General Concept
ally, networks may be determined by the use of One challenge when inferring the network struc-
Granger causality between the nodes, but unlike ture from data lies in distinguishing direct and
linear cross-correlation, this approach tries to indirect interactions. Given a network with direct
address spurious causalities caused by and indirect interactions, a bivariate analysis can
confounding by unobserved processes (Eichler be applied to every pairwise combination of
2005). nodes of the network to detect all connections.
Nonlinear systems can show behaviors that are The resulting network could be fully connected.
impossible in linear systems (Pikovsky But the indirect connections are often weaker
et al. 2001); for example, nonlinear systems can than direct ones. By setting some finite statistical
synchronize. In the seventeenth century, Huy- power, we can hope to distinguish the direct ones
gens observed synchronization between two pen- from the indirect ones. But such an approach
dulum clocks on the same wall. These clocks are indeed relies on the assumption that indirect con-
coupled self-sustained oscillators. The process of nections are the weakest in the network and that
synchronization is an adaptation of certain char- the statistical power is actually too low to detect
acteristics of the two processes. The oscillations the interaction. In general, both assumptions are
between the clocks were always antiphase and not valid and bear the risk that erroneous conclu-
restored synchronization after being perturbed sions are drawn; this leads to several false-posi-
(Pikovsky et al. 2000); for general systems, dif- tive conclusions.
ferent phase relations are conceivable. A weaker Multivariate approaches attempts to correct
form of synchronization has been observed for these problems with bivariate approaches.
between two coupled chaotic oscillators. These The theory behind multivariate approach is that
oscillators can synchronize their phases while if all contributing processes have been observed,
their amplitudes stay almost uncorrelated it should be possible to infer whether or not
(Pikovsky et al. 2001; Boccaletti et al. 2002; a connection is indirect. This is done by
Wiesenfeldt et al. 2001). Several forms of syn- partializing the information of the third pro-
chronization have been described, ranging from cesses. Practically, there are several approaches
phase synchronization via lag synchronization to to do this. Here, we just provide an outline for
almost complete synchronization (Boccaletti how this is done. To test if two systems are direct
et al. 2002). Generalized synchronization is the or indirect, we will first assume for simplicity that
most general case where one signal is related to there are only the two systems in question and
the other through any arbitrary function. one other system that could potentially influence
While a battery of various analysis techniques them. Because we have observed all the systems,
is available for linear stochastic systems, for it is possible to determine the amount of informa-
nonlinear systems techniques are still in their tion transfer from the third system onto the two
infancy. This entry is dedicated to the inference systems of interest. If this extra information from
the third process suffices to fully explain the Self-Sustained Oscillators

information transfer between the first two sys- To discuss synchronization, we must start with
tems, then we can conclude that the connection a simple model of each node. One of the simplest
is indirect. If the third cannot explain all the behaviors with nonlinear dynamics is a self-
information transfer, then we conclude that they sustained oscillator. In general these oscillators
are directly connected. Whether the connection can be described using the very general differen-
between the two systems of interest is causal can tial equation:
be determined by only using past points of each
process to predict the future behaviors of the
X_ ðtÞ ¼ f ðXðtÞ, aðtÞ, UðtÞÞ,
other processes.
In the following, we present some concepts
where X(t) is a multidimensional variable to
how the partialization is performed in practice.
ensure an oscillatory behavior. The external
A complete coverage of all possible approaches is
influence U(t) as well as the parameters a(t) are
beyond the scope of this entry.
vectors. In this case, external driving is neglected
in the following and the parameters ai are
Phase-Based Approaches
assumed to be constant in time.
Synchronization analysis is a common approach
Two coupled oscillators can be written as
to detect interactions between nonlinear self-
follows:
sustained oscillators (Pikovsky et al. 2001). Fol-
lowing the observations and pioneering work of
Huygens, the synchrony has been observed in X_ 1 ðtÞ ¼ f 1 ðX1 ðtÞ, a1 Þ þ ϵ1, 2 h1 ðX1 ðtÞ, X2 ðtÞÞ
many different systems including systems with X_ 2 ðtÞ ¼ f 2 ðX2 ðtÞ, a2 Þ þ ϵ2, 1 h2 ðX2 ðtÞ, X1 ðtÞÞ
a limit cycle or a chaotic attractor. Several differ-
ent types of synchronization have been observed where the coupling from oscillator j onto
for these systems, ranging from phase synchroni- oscillator i is the coefficient. If ei,j is nonzero,
zation, as the weakest form of synchronization then the system is considered coupled. h1 (.) and
via lag synchronization, to generalized or com- h2 (.) are the coupling functions and can be any
plete synchrony (Rosenblum et al. 1996; arbitrary function. Usually, diffusive coupling is
Rosenblum et al. 1997; Kocarev and Parlitz assumed, i.e.,
1996; Pecora and Carroll 1990). h1(X1(t), X2(t)) = (X2(t) X1(t)) and for h2
Phase synchronization analysis is a powerful accordingly.
tool because it detects even weak coupling
between oscillators. For example, in some cha- Phase Synchronization
otic oscillators, very weak coupling between The phase of a limit cycle oscillator is
oscillators can synchronize their phases but not a monotonically increasing function with
their amplitudes (Rosenblum et al. 1996). To F(t)|t = pT = p2p = poT,
quantify the process of synchronization, different where p denotes the number of completed
measures have been proposed (Tass et al. 1998; cycles, T is the time needed for one complete
Mormann et al. 2000; Rosenblum et al. 2001). cycle, and o is the eigenfrequency of the oscilla-
The first approach presented here is a measure tor. But, more generally, we can consider the
based on circular statistics, which is the so-called phase of any oscillation from a differential
mean phase coherence (Mormann et al. 2000). equation
We will introduce it by first reviewing phase F_ i ðtÞ ¼ oi , i ¼ 1, . . . , N where the oi are the
synchronization that in weakly coupled self- frequencies of the uncoupled oscillators with
sustained oscillators. For a more detailed intro- i denoting the i-th oscillator.
duction to synchronization including phase syn- Differential equations describing the phase in
chronization, we refer to the literature, e.g., a coupled two-oscillator system with different
Pikovsky et al. (2001). frequencies n and m can be written as follows:
nF_ 1 ðtÞ ¼ no1 þ ϵ1, 2 nH 1 ðF1 , F2 Þ If there is a sharp peak in the phase difference
mF_ 2 ðtÞ ¼ mo2 þ ϵ2, 1 mH 2 ðF2 , F1 Þ, distribution of Fn,m1,2 mod 2p, it indicates that two
phases have a coherent motion; if the distribution
which can be written as the phase difference is flat, then it indicates that the two have indepen-
between the two oscillators dently evolving phases. Based on circular statis-
tics, the phase difference distribution can be
quantified as follows:
nF_ 1 ðtÞ mF_ 2 ðtÞ ¼ no1 mo2 þ ϵ1, 2 nH 1 ðF1 ,F2 Þ
ϵ2, 1 mH 2 ðF2 , F1 Þ D
n, m 1 X
T n, m

R ¼ iF
e 1, 2 ð tÞ
This equation represents the generalized phase
12 T t¼1
difference (Pikovsky et al. 2001).
In the case of e2,1 = e1,2 then Fn,m
1,2 = nF1 This quantity is normalized between 0 and
mF2 and Do = no1 mo2; the above differen- 1 and will be one for perfectly synchronized
tial equation can be rewritten as phases and zero for independent signals (Tass
et al. 1998; Mardia and Jupp 2000; Mormann
et al. 2000).
F_ 1, 2 ðtÞ ¼ Do þ e1, 2 H Fn;m
n;m
1, 2 Often the phase of a signal cannot be directly
measured. In which case, we can use the Hilbert
with a new periodic function H(.). transform to calculate a 90 phase shifted version
This differential equation has a fixed point of the time series measured, with which we can
characterized by the following equation: use to calculate the instantaneous phase

Do þ e1, 2 H Fn1,, m ¼0 ð
1
2 1 X ð tÞ
Xh ðtÞ ¼ P:V: dt
p tt
1
In this case, the phase difference is constant
with time. Thus, both phases maintain a fixed
where P.V. refers to Cauchy’s principal value.
relationship and the system is considered to be
This leads to
n:m phase synchronized.
If some stochastic influence is present in the
oscillators, which in turn alters the phase differ- V ðtÞ ¼ XðtÞ þ iXh ðtÞ ¼ AðtÞeiFðtÞ
ence dynamics, certain fluctuations are possible
but these are restricted by an appropriately cho- where i is the imaginary number and through
sen constant, i.e., Euler’s formula can be written as a complex
exponent where F(t) is the phase. Alternative
approaches to obtain the phases are conceivable
n, m
F1, 2 mod 2p const based on, e.g., wavelet transformations (Le van
Quyen et al. 2001; Bandrivskyy et al. 2004). Note
In this case there will be variability between that this transformation is applicable for signals in
the phases, but the phase synchronization will which a frequency is well defined; different
still be preserved. approaches are needed for broad band signals (see
below) or point processes (Smirnov et al. 2007).
The Mean Phase Coherence
If the weakly coupled self-sustained oscillators Multivariate Phase Synchronization
are phase synchronized, the above equation can If multiple signals are being analyzed, a multi-
be reformulated to yield a single number to quan- variate phase synchronization, also referred to
tify the phase synchrony. as partial phase synchronization, can be used.
For an N-dimensional process, the following with i,j = 1,. . .,3 and parameters are set to
matrix of pairwise interactions can be generated: a = 0.15, b = 0.2, c = 10, o1 = 1.03,
o2 = 1.01, and o3 = 0.99 yielding chaotic

R ¼ Rnij, m behavior in the deterministic case. The noise
i¼1, :, N , j¼1, :, N term, j, is Gaussian distributed with mean of
zero and standard deviation sj = 1.5. Both the
This is called the synchronization matrix bidirectional couplings between oscillators 1 and
containing all pairwise phase synchronization 3 and 1 and 2 are varied between 0 and 0.3. The
measures. The inverse PR = R1 of this matrix oscillators 2 and 3 are not directly coupled.
leads to the definition of the n:m partial phase The 1:1 mean phase coherence and the partial
synchronization index phase synchronization indices are depicted in
jPRkl j Fig. 1. The bivariate synchronization index
RkljY ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi mean phase coherence R12,as well as R13,
PRkk PRll
increases for increasing coupling strength, indi-
between oscillators k and l. It is conditioned on cating phase synchronization (Fig. 1a upper tri-
the remaining processes which are summarized angular). Once there is a sufficiently strong
as Y. It can be analytically shown (Dahlhaus coupling between oscillators 1 and 2 as well as
2000; Schelter et al. 2006b) that this matrix inver- 1 and 3, indirect coupling is seen between 2 and
sion partializes the information of the third pro- 3 with a nonvanishing bivariate synchronization
cesses, as introduced in the “General Concept” index R23 (Fig. 1a upper triangular). This high but
section. Indirect interactions are characterized by spurious phase synchronization is caused by the
a vanishing partial phase synchronization. If the common influence from oscillator 1. In the same
bivariate phase synchronization index Rkl is con- figure (Fig. 1a below the diagonal), the results of
siderably different from zero while the partial phase synchronization analysis are shown.
corresponding multivariate partial phase syn- While R12 | 3 and R13 | 2 are essentially unchanged
chronization index Rkl | Y is approximately zero, compared to the bivariate synchronization indi-
then the connection is most likely due to indirect ces, R23 | 1 stays almost always below 0.1 and
coupling between the processes k and l (Schelter therefore indicates this measure is due to spurious
et al. 2006b). A rigorous statistical test using synchronization, indicating that there is an
natural surrogates or twin surrogates (Nawrath absence of a direct coupling between oscillators
et al. 2010) can be used to determine significance. 2 and 3. Hence, the true underlying (multivariate)
network structure is correctly revealed by the
Example analysis (Fig. 1b).
Three coupled stochastic Roessler oscillators
(Roessler 1976) are an example of a system of Phase Dynamics Modeling
weakly coupled self-sustained phase-coherent Another approach in reconstructing networks is to
stochastic oscillators: estimate the coupling functions H1(F1, F2) and
" # H2(F2, F1) from observations of the dynamics.
X

X_j ¼ oj Y j Zj þ e j , i X i X j þ sj j These reconstructions can, for instance, be based
i, i6¼j on approximating the functions H1 and H2 with
trigonometric functions. With this approach, it has
Y_j ¼ oj Xj þ a Y j been shown that a reliable detection of interactions
(Smirnov et al. 2007 and references therein) and

Z_j ¼ b þ Xj c Zj therefore the reconstruction of the network topol-
ogy are possible (Kralemann et al. 2011).
Determining Network Structure from Data: Nonlinear Modeling Methods, Fig. 1 (a) Phase synchronization as
well as partial phase synchronization analysis. (b) Corresponding network as inferred from the observations
Recurrence-Based Approaches phase space was originally introduced by

If the processes are not phase-coherent, which is Poincare (1890). A representation of these recur-
typically the case for many observed systems, the rences is given by the recurrence matrix
approach based on the direct calculation of the (Eckmann et al. 1987; Marwan et al. 2007):
phases is not feasible because neither the phase
RPðeÞi, j ¼ Yðe jjXðiÞ XðjÞjjÞ ,
nor the onset of phase synchronization can be
explicitly defined. Bivariate analysis of i, j ¼ 1, . . . , nn
noncoherent systems has been approached by
testing synchronization using alternate notion of where X(i) and X(j) denote trajectories of length
the phase, i.e., a phase that builds on the general n in phase space, Y(.) is the Heaviside function,
idea of curvature (Osipov et al. 2003). However, | || is an appropriate norm, and e is a predefined
such a definition of the phase restricts the analysis threshold. If only a scalar time series has been
to systems where the phase trajectory corre- observed, the state of the system, i.e., the vector
sponds to a curve with a positive curvature on X(i), can typically be reconstructed by delay
some projection plane. An alternative approach is embedding (Packard et al. 1980; Takens 1981;
based on recurrence analysis. Sauer et al. 1991), where each dimension repre-
sents the data at some time lag t in the past:
Recurrence Analysis
A synchronization measure based on the recur- XðiÞ ¼ ½xðiÞ, xði tÞ, . . . , xði ktÞ
rences of a trajectory can also be used. This
approach considers phase synchronization in When a series of points stays within a small
a statistical sense. The notion of recurrences in distance apart, say within a small tube of radius e
around the other, that section of the trajectory coherent systems. Using the inverse CPR1, the
corresponds to a diagonal of 1’s in the recurrence generalized partial synchronization index can be
matrix. Thus, an estimate of the probability calculated:

1 X Lt CPR1
p^ðe, tÞ ¼ RRt ðeÞ ¼ RPðeÞi, iþt CPRkljY ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
kl
L t i¼1 CPR1 1
kk CPRll
that a system recurs to the e-neighborhood of This measure quantifies phase synchroniza-
a former point of the trajectory after t time steps tion between two oscillators k and l, conditioned
is given by the diagonal-wise calculated recur- on third processes Y of a multivariate possibly
rence rate RRt(e). non-phase-coherent and noisy oscillatory process
This measure can be considered as (Nawrath et al. 2009).
a generalized autocorrelation function (Romano
et al. 2005), since it also describes higher-order Information Theoretic Approaches
correlations among the points of the trajectory. It Information theory-based approaches comple-
measures the adaptation of the time scales of two ment the nonlinear approaches presented above.
interacting systems by comparing their The mutual information between time series
corresponding recurrence matrices; if the dis- X and Y can be calculated as follows:
tances between diagonal lines in their recurrence
matrices coincide, the time scales of the systems XX pðx, yÞ
adapt to each other. To quantify this, the cross- I ðX; Y Þ ¼ pðx, yÞlog
yY xX
pðxÞpðyÞ
correlation coefficient of the probabilities of
recurrence can be calculated as follows:
where p(x) represents the probability that the
CPRkl ¼ hp^k ðe, tÞ^
pl ðe, tÞi, CPRkl ½0, 1 signal X will have the amplitude x and p(x,y) is
the joint probability that the time series X will
In the case when the two systems have locked have the value x while simultaneously the time
phase dynamics, the probability of recurrence is series Y will have the value y. Transfer entropy
simultaneously high for both systems and the calculates the direction of information flow
CPRkl will differ from zero significantly. This between two systems by testing how the informa-
synchronization measure has been demonstrated tion from the past of the two systems predicts the
to be effective for detecting coupling in a general future of one of the signals:
class of non-phase-coherent and nonstationary X ðkÞ ðlÞ

systems and even for time series corrupted by T X!Y ðk, lÞ ¼ p ytþ1 , yt , xt
strong noise (Romano et al. 2005).
ðkÞ ðlÞ
p ytþ1 jyt , xt
log
ðkÞ
Multivariate Recurrence Analysis p ytþ1 jyt
Recurrence analysis can also be applied to
a multivariate system with N interdependent and
This approach utilizes the fundamental con-
non-phase-coherent processes. To treat the mul-
cepts of the information theoretic approaches
tivariate case can be generated as a matrix of the
(Kantz and Schreiber 1997). The p(yt+1|yt(k),xt(l))
bivariate CPR synchronization indices:
are the conditional probability density functions

where the probability of yt+1 given the previous
CPR ¼ CPRij i, j¼1, ..., N k points of the times series yt(k),xt(l).
The information and the transfer information
is used. This matrix has the same symmetry prop- as presented are bivariate; however multivariate
erties as the synchronization matrix for phase- extensions are straightforward as the
(conditional) probability density functions can Partial directed coherence |pi j(o)| provides
include more processes. The challenge in calcu- a measure for the directed, linear influences from
lating these measures in higher dimensional Xj(t) onto Xi(t) at frequency o. It is estimated by
densities is that they quickly become computa- fitting an N-dimensional VAR[p] model to the
tionally intensive to find all the conditional data and directly using the above equations with
probabilities. Solutions are possible and can, for the parameter estimates substituted for the true
instance, be found in Palus and Stefanovska parameters.
(2003), Runge et al. (2012), or Wibral Particularly important for nonlinear systems is
et al. (2012) and references therein. the corresponding pointwise a-significance level D
for the partial directed coherence (Schelter
Linear Approaches et al. 2006c)
Although linear approaches are strictly speaking
not part of the nonlinear data-based modeling 0 11=2
approaches, we include them here. They are îj ðoÞ w2
C
@ X 1, 1a A
indeed very powerful in revealing the true n A
^ mj ðoÞ 2
m
underlying network structure despite the fact
that they are technically speaking not applicable.
where w21,1a is the 1a quantile of the
Due to space limitations, we can only discuss one
w2-distribution with one degree of freedom. The
of the linear approaches, although many more "
X p
exist in the literature. We refer the reader, for values Cij ðoÞ ¼ Sii Hjj ðl, mÞð cos ðloÞ
instance, to Schelter et al. (2006a) and references l, m¼1
#
therein.
The concept of Granger causality (Granger cos ðmoÞ þ sin ðloÞ sin ðmoÞÞ :
1969) is based on the commonsense conception
that causes precede their effects in time. Partial can be calculated based on the inverse
directed coherence H = V1 of the covariance matrix V of the VAR
[p] process X(t), which is composed of the entries
Vij(l, m) = cov(Xi(t l), Xj(t m)) for i,
Aij ðoÞ
jpi j ðoÞj ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi j = 1,. . .,N and l,m = 1,. . .,p (Luetkepohl 1993).
X
Amj ðoÞ 2 We emphasize that the significance level
m depends on the order of the vector autoregressive
process; higher model orders lead to higher sig-
with nificance levels which in turn indicates that for
higher models the ability to detect weak cou-
X
p
plings decreases.
AðoÞ ¼ I aðr Þ eior Partial directed coherence by construction is
r¼1
a multivariate analysis technique. It has been
for an N-dimensional vector autoregressive pro- developed for linear stochastic processes. How-
cess of order p (VAR[p] process) ever, in neurophysiology, nonlinear stochastic
processes are generating the time series. In most
of these cases, the dependence structure is
X
p
XðtÞ ¼ aðr Þ Xðt r Þ þ eðtÞ, reflected in the linear second-order structure; for
r¼1 this reason the partial directed coherence discloses
the network structure also for nonlinear processes.
where e(t) is a multivariate Gaussian white noise As an example system for which to test these
process with covariance matrix S. This measure tools, we will introduce a system of coupled
was introduced by Baccala and Sameshima stochastic van der Pol oscillators (van der Pol
(2001) as a measure of Granger causality. 1922):

X€1 ¼ m 1 X21 x_1 o21 X1 þ s1 þ e12 ðX2 X1 Þ þ e13 ðX3 X1 Þ þ e14 ðX4 X1 Þ



4 4
where the oscillators are given coefficients bidirectional couplings between these four non-
o1 = 1.5, o2 = 1.48, o3 = 1.53, o4 = 1.44, identical oscillators are set to e12 = e21 = 0.2,
s = 1.5, and Gaussian distributed white noise e24 = e42 = 0.2, e31 = 0.2, and e34 = 0.2.
i. This system is simulated for ith n = 50,000 The causal influences are summarized in the
data points for each process to generate signals to graph in Fig. 2a. Estimated partial directed coher-
apply the tools for reconstructing the network ences as well as the spectra are given in Fig. 2b.
topology. Note that although the four oscillators The order of the vector autoregressive process is
are diffusively coupled and therefore their inter- chosen to be p = 200. This high model order is
actions are still linear, the system is nonlinear. required to reproduce the spectra with sufficient
The nonlinearity parameter m is fixed to m = 5, accuracy, as opposed to nonparametric spectral
leading to a highly nonlinear behavior of the van estimates. The corresponding 5 % significance
der Pol oscillators. The unidirectional and levels are indicated by gray lines. Partial directed
Determining Network Structure from Data: corresponding 5 % significance levels are indicated by
Nonlinear Modeling Methods, Fig. 2 Causal influ- gray lines. The simulated causal influences are reproduced
ences for the example of four coupled stochastic van der correctly since only those partial directed coherences that
Pol oscillators (a). Corresponding spectra (diagonal) and correspond to direct interactions are statistically signifi-
partial directed coherence (off-diagonal) (b). The cant at the oscillation frequencies
coherence correctly detects the causal influences activate tremor. Data were sampled at 1,000 Hz.
in the van der Pol system. Note that the The EEG data as well as the EMG data
significance level depends on the investigated were preprocessed applying a band-pass filter
frequency. At the peaks in the spectra of the between 30 and 200 Hz to avoid aliasing and
van der Pol oscillators, the significance level is movement artifacts; the EMG was rectified
slightly higher than at the remaining frequencies. afterward. The EEG was then high-pass filtered
Thus, only those partial directed coherencies above 0.5 Hz to avoid baseline fluctuations and
are statistically significant that correspond also anti-aliasing filtered. Scalp electrodes over
to a direct causal influences between the the left and right sensorimotor cortex and the D
oscillators. EMG of the left and right wrist extensor are
analyzed.
Application It is of importance to investigate whether the
So far, the multivariate analysis techniques cortex imposes its oscillatory activity on the mus-
presented here have been applied to simulated cles via the corticospinal tract or whether the
time series. To illustrate their performance in muscle activity is just reflected in the cortex via
physiological applications, examples of patients proprioceptive afferences. To get these deeper
suffering from essential tremor are presented. insights into the tremor generating mechanisms,
The pathophysiological basis of essential partial directed coherence is applied to data
tremor, a common neurological disease with recorded from these patients.
a prevalence of 0.4–4 % (Louis et al. 1998), is In Fig. 3, the results of the partial directed
not precisely known. Essential tremor manifests coherence analysis for the two EMG and the
itself mainly in the upper limbs, when the hands two EEG channels are shown for an exemplary
are in a postural outstretched position. Usually patient. Both directions from the cortex to the
the trembling frequency of the hands is 4–10 Hz. muscles and vice versa are observed.
To elucidate the tremor generating mechanisms A significant influence from the right EEG
in essential tremor, relationships between the onto the left EMG, from the left EMG onto the
brain and trembling muscles are of particular right EEG, from the left EEG onto the right EMG,
interest. For unilaterally activated tremor, and from the right EMG onto the left EEG is
tremor-correlated cortical activity to the detected. Especially, the partial directed coher-
contralateral tremor side has been revealed by ence from the right EMG to the left EEG is rather
magnetoencephalography (MEG) and electroen- large.
cephalography (EEG) for Parkinsonian tremor The graph summarizing the results of partial
(Volkmann et al. 1996; Hellwig et al. 1999) and directed coherence analysis for the essential
by electroencephalography for essential tremor tremor patient is presented in Fig. 4. The influ-
(Hellwig et al. 2001). In bilaterally activated ences between both EEGs are marked by dashed
essential tremor, however, a more complex inter- arrows. This is due to the fact that partial directed
relation structure has been observed by simulta- coherence is not significant exactly at the
neous electroencephalographic recordings from tremor frequency but over a range of frequencies
the scalp and electromyographic (EMG) record- close to the tremor frequency. Since causal
ings from the extensor muscles (Hellwig influences from both EEGs to the corresponding
et al. 2003). In addition to contralateral coher- contralateral EMGs are present, participation of
ences, also ipsilateral coherences between the the motor cortex in tremor generation is
sensorimotor cortex and the muscles have been strongly indicated. Moreover, there is also
detected. a significant partial directed coherence from the
In this study, patients were seated in EMG to the contralateral EEG at the tremor fre-
a comfortable chair having their forearms quency. This corresponds to a feedback from the
supported while their hands were outstretched to muscles to the somatosensory cortex. For both
Determining Network Structure from Data: onto the right EEG (1st row/3rd column), from the left
Nonlinear Modeling Methods, Fig. 3 Partial directed EEG onto the right EMG (4th row/2nd column), and from
coherence analysis for a patient suffering from essential the right EMG onto the left EEG (2nd row/4th column) is
tremor. A significant influence from the right EEG onto detected
the left EMG (3rd row/1st column), from the left EMG
patients, unexpected ipsilateral interrelations are

not detected by partial directed coherence
analysis.
Summary
Several approaches to data-based modeling are
conceivable. In this entry, we focused on a few
multivariate approaches that enable to distinguish
direct from indirect interactions. Some even pro-
vide directions for the interactions, for instance,
Granger-causality-based concepts. In applications
Determining Network Structure from Data: to observed nonlinear systems, the type of data, its
Nonlinear Modeling Methods, Fig. 4 Graph for partial dimension, noise contamination, and stationarity
directed coherence analysis of the tremor application of typically decides which technique can be applied
Fig. 3. The arrows indicate a direct and directed interrela- and what conclusions can be drawn. Typically,
tion at the tremor frequency. The dashed arrows indicate
possible interactions between the EEGs that could not be simulation studies tailored to the problem at hand
unambiguously shown should be performed prior to any analysis.
References Mormann F, Lehnertz K, David P, Elger CE (2000) Mean
phase coherence as a measure for phase synchroniza-
Baccala LA, Sameshima K (2001) Partial directed coher- tion and its application to the EEG of epilepsy patients.
ence: a new concept in neural structure determination. Physica D 144:358–369
Biol Cybern 84:463–474 Nawrath J, Romano MC, Thiel M, Kiss IZ,
Bandrivskyy A, Bernjak A, McClintock PVE, Wickramasinghe M, Timmer J, Kurths J, Schelter B
Stefanovska A (2004) Wavelet phase coherence anal- (2010) Distinguishing direct and indirect interactions
ysis: application to skin temperature and blood flow. in oscillatory networks with multiple time scales. Phys
Cardiovasc Eng 4:89–93 Rev Lett 104:038701
Boccaletti S, Kurths J, Osipov G, Valladares DL, Zhou CS Osipov GV, Hu B, Zhou C, Ivanchenko MV, Kurths
J (2003) Three types of transitions to phase synchroni- D
(2002) The synchronization of chaotic systems. Phys
Rep 366:1–101 zation in chaotic oscillators. Phys Rev Lett 91:024101
Dahlhaus R (2000) Graphical interaction models for mul- Packard N, Crutchfield J, Farmer D, Shaw R (1980)
tivariate time series. Metrika 51:157–172 Geometry from a time series. Phys Rev Lett 45:712
Ding M, Chen Y, Bressler SL (2006) Granger causality: Palus M, Stefanovska A (2003) Direction of coupling
basic theory and application to neuroscience. In: from phases of interacting oscillators: an information
Schelter B, Winderhalder M, Timmer J (eds) Hand- theoretic approach. Phys Rev E 67:055201
book of time series analysis. Wiley-VCH, Weinheim, Pecora LM, Carroll TL (1990) Synchronization in chaotic
pp 437–460 systems. Phys Rev Lett 64:821–824
Eckmann J-P, Oliffson Kamphorst S, Ruelle D (1987) Pikovsky A, Rosenblum M, Kurths J (2000) Phase syn-
Recurrence plots of dynamical systems. Europhys chronization in regular and chaotic systems. Int
Lett 4:973–977 J Bifurc Chaos 10:2291–2305
Eichler M (2005) A graphical approach for evaluating Pikovsky A, Rosenblum M, Kurths J (2001)
effective connectivity in neural systems. Phil Trans Synchronization – a universal concept in nonlinear
R Soc B 360:953–967 sciences. Cambridge University Press, Cambridge
Eichler M (2007) Granger-causality and path diagrams for Poincare H (1890) Sur les equations de la dynamique et le
multivariate time series. J Econom 137:334–353 probleme de trois corps. Acta Mathematica 13:1–270
Granger CWJ (1969) Investigating causal relations by Roessler OE (1976) An equation for continuous chaos.
econometric models and cross-spectral methods. Phys Lett A 57:397–398
Econometrica 37:424–438 Romano MC, Thiel M, Kurths J, Kiss IZ, Hudson JL
Hellwig B, Haeussler S, Schelter B, Lauk M, Guschlbauer B, (2005) Detection of synchronization for non-phase-coher-
Timmer J, Luecking CH (2001) Tremor correlated corti- ent and non-stationary data. Europhys Lett 71:466–472
cal activity in essential tremor. Lancet 357:519–523 Rosenblum MG, Pikovsky AS, Kurths J (1996) Phase
Hellwig B, Schelter B, Guschlbauer B, Timmer J, synchronization of chaotic oscillators. Phys Rev Lett
Luecking CH (2003) Dynamic synchronisation of cen- 76:1804–1807
tral oscillators in essential tremor. Clin Neurophysiol Rosenblum MG, Pikovsky AS, Kurths J (1997) From
114:1462–1467 phase to lag synchronization in coupled chaotic oscil-
Kantz H, Schreiber T (1997) Nonlinear time series analy- lators. Phys Rev Lett 78:4193–4196
sis. Cambridge University Press, Cambridge Rosenblum MG, Pikovsky A, Kurths J, Schaefer C, Tass
Kocarev L, Parlitz U (1996) Generalized synchronization, PA (2001) Phase synchronization: from theory to data
predictability, and equivalence of unidirectionally analysis. In: Moss F, Gielen S (eds) Handbook of
coupled dynamical systems. Phys Rev Lett 76:1816–1819 biological physics, vol 4, Neuroinformatics. Elsevier,
Kralemann B, Pikovsky A, Rosenblum M (2011) Amsterdam, pp 279–321
Reconstructing phase dynamics of oscillator networks. Runge J, Heitzig J, Marwan N, Kurths J (2012) Quantify-
Chaos 21:025104 ing causal coupling strength: a lag-specific measure for
Le van Quyen M, Martinerie J, Navarro V, Boon P, multivariate time series related to transfer entropy.
D’Have M, Adam C, Renault B, Varela F, Baulac M Phys Rev E 86:061121
(2001) Anticipation of epileptic seizures from standard Sauer T, Yorke J, Casdagli M (1991) Embedology. J Stat
EEG recordings. Lancet 357:183–188 Phys 65:579–616
Louis ED, Ford B, Wendt KJ, Cameron G (1998) Clinical Schelter B, Winterhalder M, Timmer J (eds) (2006a)
characteristics of essential tremor: data from Handbook of time series analysis. Wiley-VCH, Berlin
a community-based study. Mov Disord 13:803–808 Schelter B, Winterhalder M, Dahlhaus R, Kurths J, Timmer
Luetkepohl H (1993) Introduction to multiple time series J (2006b) Partial phase synchronization for multivariate
analysis. Springer, New York synchronizing system. Phys Rev Lett 96:208103
Mardia K, Jupp P (2000) Directional statistics. Wiley, Schelter B, Winterhalder M, Eichler M, Peifer M,
West Sussex Hellwig B, Guschlbauer B, Luecking CH,
Marwan N, Carmen Romano M, Thiel M, Kurths J (2007) Dahlhaus R, Timmer J (2006c) Testing for directed
Recurrence plots for the analysis of complex systems. influences among neural signals using partial directed
Phys Rep 438:237–329 coherence. J Neurosci Methods 152:210–219
D 1016 Deterministic Reaction-Diffusion Simulators
Smirnov D, Schelter B, Winterhalder M, Timmer J (2007) of a system governed by the diffusion and inter-
Revealing direction of coupling between neuronal action of species within a domain in
oscillators from time series: phase dynamics modeling
versus partial directed coherence. Chaos 17:013111 a deterministic fashion.
Takens F (1981) Detecting strange attractors in turbu-
lence. In: Rand DA, Young L-S (eds) Dynamical sys-
tems and turbulence (Warwick 1980), vol 898, Lecture Detailed Description
notes in mathematics. Springer, Berlin, pp 366–381
Tass PA, Rosenblum MG, Weule J, Kurths J, Pikovsky A,
Volkmann J, Schnitzler A, Freund HJ (1998) Detection In neuroscience, these species can be one of many
of n : m phase locking from noisy data: application classes of molecules: ions, enzymes, polypep-
to magnetoencephalography. Phys Rev Lett tides, globular proteins, microRNAs, etc. The
81:3291–3295
van der Pol B (1922) On oscillation-hysteresis in a simple interactions are chemical reactions, such as phos-
triode generator. Phil Mag 43:700–719 phorylation or binding, the synthesis of a new
Volkmann J, Joliot M, Mogilner A, Ioannides AA, Lado F, molecule out of substrates, or the breakdown of
Fazzini E, Ribary U, Llinas R (1996) Central motor loop a molecule. Unlike stochastic simulators, which
oscillations in Parkinsonian resting tremor revealed by
magnetoencephalography. Neurology 46:1359–1370 approximate these dynamics using pseudoran-
Wibral M, Wollstadt P, Meyer U, Pampu N, Priesemann V, dom processes, deterministic simulators solve
Vicente R (2012) Revisiting Wiener‘s principle of a system of partial differential equations
causality – interaction-delay reconstruction using transfer (PDEs). Thus, while stochastic simulators need
entropy. In: Proceedings of the 34th annual international
conference of the IEEE EMBS (EMBC 2012), San Diego to be run many times to identify the range of
Wiesenfeldt M, Parlitz U, Lauterborn W (2001) Mixed likely outcomes, deterministic simulators need
state analysis of multivariate time series. Int J Bifurc only be run once, as the result is unique.
Chaos 11:2217–2226
Theory
Further Reading Deterministic simulators are most appropriate
Hellwig B, Haeussler S, Lauk M, Koester B,
Guschlbauer B, Kristeva-Feige R, Timmer J, Luecking when they can rely on the law of large numbers
CH (2000) Tremor-correlated cortical activity (Kotelenez 1986), which implies that noise vari-
detected by electroencephalography. Electroen- ation of concentration becomes small as the num-
cephalogr Clin Neurophysiol 111:806–809 ber of molecules in the population increases. That
is, when a large number of molecules are present,
the dynamics of the population becomes well
approximated by the mean dynamics. This
Deterministic Reaction-Diffusion assumption underlies a key trade-off between
Simulators deterministic and stochastic simulators. Deter-
ministic simulations allow the use of a single
William W. Lytton1,2 and Robert A. McDougal3 number to represent the concentration in
1
Departments of Physiology & Pharmacology a given volume over a relatively large time step
and Neurology, SUNY Downstate Medical in the setting of large numbers of particles.
Center, Brooklyn, NY, USA Stochastic simulations are needed when the
2
Department of Neurology, Kings County number of molecules is small, but exact
Hospital, Brooklyn, NY, USA stochastic methods generally require either
3
Department of Neurobiology, Yale University a state variable for each molecule (which is inef-
School of Medicine, New Haven, CT, USA ficient if the number of molecules is large) and/or
small time steps (Rathinam et al. 2003; Kerr
et al. 2008).
Definition Fick’s second law of diffusion states that the
diffusion of a species with concentration u (x, t)
A deterministic reaction–diffusion simulator is at position x (in 1, 2, or 3 dimensions) at time
software designed to approximate the dynamics t within a domain is governed by
Deterministic Reaction-Diffusion Simulators 1017 D
ut ¼ D ∇ 2 u Chemesis (Blackwell and Kotaleski 2003) and
MOOSE (Ray and Bhalla 2008).
which states that the rate of change of concentra- The domain may be specified and discretized in
tion at a given point and time is proportional (with a variety of ways. Discretization is the process of
proportionality constant D, the diffusion constant) choosing finitely many locations to represent the
to the Laplacian of the concentration u, a measure infinitely many points in a domain. Virtual Cell
of its unevenness. A reaction–diffusion equation simulates on 1-, 2-, or 3-dimensional domains
for a single species is the diffusion equation except specified either analytically or via segmented
images. It then discretizes the domain using D
with an additional term f (u, t) added to the right-
hand side to account for the net effects of the a regular Cartesian mesh. A variation on this
reactions: approach would be to use unequally sized cubes
chosen at either the beginning or repeatedly
ut ¼ D∇2 u þ f ðu, tÞ throughout a simulation to focus computational
effort as needed. STEPS works with (possibly
When multiple species are present, the system unequally sized) tetrahedra, but the discretization
of PDEs consists of many equations of this form, must be done using another tool. NEURON works
where the reaction term may depend on all spe- on 1-dimensional branching geometries either
cies present. Initial and boundary conditions described programmatically or imported from
complete the mathematical description (Fife tracing programs such as Neurolucida (Glaser
1979). In neuroscience, Neumann boundary con- and Glaser 1990). This one-dimensional branched
ditions are common, as they describe the flux geometry is then discretized into short segments.
across the boundary, such as occurs through ion Reaction dynamics are typically specified as
channels. either kinetic schemes directly linked to the biol-
ogy or directly as terms in the underlying PDEs.
Practice Simulators that support kinetic schemes typically
Since every deterministic reaction–diffusion sim- internally translate them into PDEs so that they
ulator must provide tools to fully specify the can be integrated.
PDE, existing tools require the same conceptual Multiple integration options are possible.
steps, although the order may vary: 1. the domain The STEPS solver uses “Wmrk4,” an explicit
is specified (whole cell or part of cell) and Runge–Kutta method (Kutta 1901). NEURON
discretized, 2. the species are specified, 3. the supports first- and second-order implicit
reactions are specified, and 4. the equations are fixed-step integration via implicit Euler and
integrated. Results may then be analyzed, either Crank-Nicholson (Crank and Nicolson 1947),
within the simulator or with external tools. respectively. NEURON also supports variable-
NEURON (Carnevale and Hines 2006), step variable-order implicit methods from the
STEPS (Hepburn et al. 2012), and Virtual Cell SUNDIALS (Hindmarsh et al. 2005) library. Vir-
(Loew and Schaff 2001) – widely used tools tual Cell supports a variety of explicit, implicit,
in the computational neuroscience community and semi-implicit finite volume methods. Finite
that support deterministic reaction–diffusion – elements are an alternative strategy.
illustrate much of the variation possible within
the shared conceptual framework. NEURON and Acknowledgment This work was partially supported by
STEPS are primarily targeted at computational NIH R01MH086638 and NIH 2T15LM007056.
neuroscientists, while Virtual Cell is primarily
targeted at cell biologists but supports the spatial References
variation and membrane-associated ion channels
Blackwell KT, Kotaleski JH (2003) Modeling the dynamics
essential for computational neuroscience (Brown
of second messenger pathways. In: neuroscience data-
et al. 2011). Other simulators also provide facil- bases: a practical guide. Springer Science + Business
ities for deterministic diffusion, including Media, New York, pp 63–79
D 1018 Diaphragm Pacing
Brown SA, Moraru II, Schaff JC, Loew LM (2011) Virtual

NEURON: a strategy for merged biochemical and Diffusion Equation
electrophysiological modeling. J Comput Neurosci
31:385–400
Carnevale NT, Hines ML (2006) The NEURON book. Fidel Santamaria1 and Toma M. Marinov2
1
Cambridge University Press, Cambridge UTSA Neurosciences Institute, The University
Crank J, Nicolson P (1947) A practical method for numer- of Texas at San Antonio, San Antonio, TX, USA
ical evaluation of solutions of partial differential equa- 2
tions of the heat-conduction type. Math Proc Camb Department of Biology, University of Texas at
Philos Soc 43:50–67 San Antonio, San Antonio, TX, USA
Fife PC (1979) Mathematical aspects of reacting and
diffusing systems. Springer, Berlin
Glaser JR, Glaser EM (1990) Neuron imaging with
Neurolucida – a PC-based system for image combin- Synonyms
ing microscopy. Comput Med Imaging Graph
14:307–317 Fick’s second law; Heat equation
Hepburn I, Chen W, Wils S, De Schutter E (2012) STEPS:
efficient simulation of stochastic reaction–diffusion
models in realistic morphologies. BMC Syst Biol
6:36 Definition
Hindmarsh AC, Brown PN, Grant KE, Lee SL, Serban R,
Shumaker DE, Woodward CS (2005) The diffusion equation (DE) is a second-order
SUNDIALS: suite of nonlinear and differential/
algebraic equation solvers. ACM Trans Math Softw parabolic partial differential equation describing
31:363–396 mass transport phenomena due to thermal motion
Kerr RA, Bartol TM, Kaminsky B, Dittrich M, of particles (Crank 1975).
Chang JJ, Baden SB, Sejnowski TJ, Stiles JR
(2008) Fast Monte Carlo simulation methods
for biological reaction–diffusion systems in solution
and on surfaces. SIAM J Sci Computing Detailed Description
30:3126–3149
Kotelenez P (1986) Law of large numbers and central limit
Diffusion is an important transport mechanism in
theorem for linear chemical reactions with diffusion.
Ann Probab 14:173–193 neurons (Bloodgood 2005; Ehlers et al 2007;
Kutta MW (1901) Beitrag zur n€aherungsweisen Integra- Korkotian and Segal 2006; Makin and Malinow
tion totaler Differentialgleichungen. Zeitschrift Math 2009; Qian and Sejnowki 1988; Renner et al.
Phys 46:435–453
2009; Sabatini et al. 2002; Santamaria 2006;
Loew LM, Schaff JC (2001) The virtual cell: a software
environment for computational cell biology. Trends Schmidt 2007a. Smith et al. 1993; Soler and
Biotechnol 19:401–406 Sabatini 2006). It takes place in the cytosol, the
Rathinam M, Petzold LR, Cao Y, Gillespie DT cell membrane and the endoplasmic reticulum.
(2003) Stiffness in stochastic chemically reacting sys-
Subject to diffusion are both ions (Ca+, Na+,
tems: the implicit tau-leaping method. J Chem Phys
119:12784–12794 K+, Cl-, etc) and large complex molecules
Ray S, Bhalla US (2008) PYMOOSE: interoperable (lipids, proteins, DNA, RNA, etc.). The physical
scripting in Python for MOOSE. Front Neuroinform process of diffusion in absence of any chemical
2(6):1–16
reactions is described by the diffusion equation:
@Cðx, tÞ
¼ ∇ðDðx, tÞ∇Cðx, tÞÞ
@t
Diaphragm Pacing where D is the diffusion coefficient (diffusivity)

and C is the concentration.
▶ Peripheral Nerve Interface Applications, In the case of constant isotropic diffusion
Respiratory Pacing coefficient, the DE,
Diffusion Equation 1019 D
@C @2 @2 @2 Then the probability for the particle to be at
¼ D∇2 C∇2 ¼ 2 þ 2 þ 2 position x at time t + Dt will be
@t @x @y @z
is separable. 1 1
pðx, t þ DtÞ ¼ pðx Dx, tÞ þ pðx þ Dx, tÞ
In the one-dimensional case, the DE becomes 2 2
Combining the Taylor expansions of the left

@C @2C
¼D 2 and the right hand sides
@t @x D

For an initial amount of substance M released at pðx, t þ DtÞ ¼ pðx, tÞ þ pt ðx, tÞDt þ O Dt2
t = 0 and x = 0, assuming that C ! 0 as x ! 1 , pðx Dx, tÞ ¼ pðx, tÞ px ðx, tÞDx
the solution to the one-dimensional DE is given by 1

þ pxx xðx, tÞ þ O Dx3
2
M
Cðx, tÞ ¼ pffiffiffiffiffiffiffiffiffiffi eððx Þ=4DtÞ
2
4pDt in the limit Dt ! 0 and Dx ! 0 leads to

where
and the mean square displacement (MSD) of the
diffusing particles is a linear function in time
Dx2
MSD ~ t. D ¼ limDt!0, Dx!0
Dt
There are two approaches to deriving the DE:
the phenomenological and the atomistic Reaction-Diffusion Equation
(Philibert 2005). The phenomenological In the presence of chemical reactions, additional
approach employs Fick’s first law, stating that terms (usually nonlinear) reflecting the various
the flux of particles J is proportional to the gradi- chemical reactions in the system can be added to
ent of the concentration the diffusion equation. Such a generalized varia-
tion of the diffusion equation is known as the
@C reaction-diffusion equation:
J ¼ D
@x
Since particles cannot be created or destroyed @CA ðx, tÞ XN

¼ ∇ðDA ðx, tÞ∇CA ðx, tÞÞ þ f i ðCA , Ri Þ
in the process, the flux of particles into a region @t i¼1
must be equal to the particle flux out of the
region, i.e., particles must satisfy the continuity where Ri are the reactants interacting chemically
equation: with substance A.
@C Anomalous Diffusion
þ∇J ¼0 Random walks with biased jump probabilities
@t
give rise to the so-called anomalous diffusion.
Combining Fick’s first law with the continuity In that case, the MSD of the diffusing particles
equation yields @C/@t = ∇ (D∇C) is a power law, i.e., MSD ~ ta. The regime with
Another way of deriving the diffusion equa- 0 a 1 is called subdiffusion, and the regime
tion is by introducing the concept of the random with a > 1 is called superdiffusion (Metzler and
walk. Consider a particle moving along the real Klafter 2000).
line starting at t = 0 and x = x0. At each time step Anomalous diffusion can be described in
Dt, the particle can make a jump left or right at terms of fractional calculus, i.e., the anomalous
a distance Dx with equal probability. diffusion equation can be written as
D 1020 Diffusion Equation
@aC @2C In this case boundary conditions are solved by

¼ D assuming that particles collide and are reflected.
@ta @x2
Although highly accurate, this process could
where the fractional derivative of the Riemann- become computationally intractable for large
Liouville type for is given by sections of dendritic tree being modeled and is
not easily integrated into compartmental models
ðt
1 d f ð tÞ of electrical activity.
Dat f ðtÞ ¼ dt
Gð1 aÞ dt 0 ð t tÞ
Shell Approximation
Applications In compartmental modeling the electrical vari-
Diffusion in neurons has been modeled using 1D ables (conductances, voltages, and currents) are
to 3D approximations with stochastic and differ- homogeneous through the compartment (Jedlicka
ential equation models (Santamaria et al. 2006; et al. 2011). However, if calcium ions flow into
Santamria et al. 2010; Santamaria et al. 2011; the cell, they could diffuse to the center of the
Bresslof and Earnshaw 2009; Brown et al. 2008; dendrite or along the dendrite. To simulate this
Jedlicka et al. 2011; Schmidt Kunert et al. 2007b; mechanism, the problem is simplified to a 2D
Holcman 2006). A large number of studies have diffusion process. Within a single compartment,
concentrated in the diffusion of intracellular cal- it is assumed that the concentration of a molecule
cium ions. Calcium ions can activate potassium will be homogenous along the radial axis of the
channels or act as second messengers in biochem- compartment; thus, the volume is divided into
ical cascades activated by synaptic activity. concentric shells. Diffusion takes place between
contiguous shells with each shell potentially hav-
Differential Versus Stochastic Simulation ing immobile buffers. Diffusion from one com-
It could be confusing that the same physical pro- partment to its neighbors has been implemented
cess can be modeled with two frameworks. on shells of the same level across compartments.
Choosing between modeling using differential Thus, each compartment has the same number of
equations or a stochastic process depends on the shells, but not all shells have the same thickness
nature of the problem. The differential equation due to variations in the diameter of each
diffusion model is amenable to be applied to compartment.
compartments as those used in modeling the elec-
trical activity of neurons. The fundamental prop- Stochastic Approximation
erties of the equation are identical to those of The diffusion of molecules inside a dendrite can
the cable equation. When using more complex be modeled with a random walk. Each molecule
dendritic structures, solving the diffusion is characterized by a diffusion coefficient. From
equation becomes computationally expensive. this diffusion coefficient, a random number gen-
A reality check has to be performed when using erator chooses distances to move along a random
differential equation since the volumes being direction. The average movement of the particle
investigated could be very small and the concen- can be described with the stochastic version of
trations could be in the nanomolar range. If the the diffusion coefficient (MSD ~ t). If a particle
concentrations being modeled are equivalent to collides with a membrane, then the molecule is
fractions of molecules that would call into ques- reflected and continues its movement. This type
tion the results of the simulations, especially if of modeling can be used to study the diffusion of
there are biochemical reactions involved. ions and proteins in dendrites and spines. The
The stochastic simulation strategy is used shapes of dendrites can be modeled or obtained
when the model is tracking the position of indi- from serial electron microscope reconstruction.
vidual molecules or small groups of molecules. The same strategy has been used to study the
Diffusion Equation 1021 D
lateral mobility of glutamate receptors in and out Jedlicka P, Deller T, Gutkin BS, Backus KH (2011)
of a synapse. Activity-dependent intracellular chloride accumula-
tion and diffusion controls GABA(A) receptor-
mediated synaptic transmission. Hippocampus
Modeling Anomalous Diffusion 21:885–898
The fundamental assumption to use the stochastic Korkotian E, Segal M (2006) Spatially confined diffusion
or differential equation models of diffusion is that of calcium in dendrites of hippocampal neurons
revealed by flash photolysis of caged calcium. Cell
molecules are moving without any memory of Calcium 40:441–449
their past. This is called a Markovian process. Makino H, Malinow R (2009) AMPA receptor
However, the complex structure of dendrites can incorporation into synapses during LTP: the role D
impose correlations in the diffusing molecules. of lateral movement and exocytosis. Neuron
64:381–390
For example, when molecules move along den- Metzler R, Klafter J (2000) The random walk’s guide to
drites, they can enter a spine. Due to the spine anomalous diffusion: a fractional dynamics approach.
bottle-neck produced by the spine head to neck Phys Rep 339:1–77
diameter ratio, the molecule remains trapped in Philibert J (2005) One and a half century of diffusion:
Fick, Einstein, before and beyond. Diffusion Funda-
the spine for a random period of time. Once the mentals 2:1.1–1.10
molecule escapes one spine, it can then be Qian N, Sejnowski TJ (1988) Electro-diffusion model of
trapped by another spine. If the dwell time in electrical conduction in neuronal processes. In: Wood
each spine varies widely, then the position of CD, McGaugh JL, Alkon DL (eds) Cellular mecha-
nism of conditioning and behavioral plasticity.
the molecule along the dendrite can no longer Pergamon, London, pp 237–244
be assumed to be Markovian. Instead, the diffu- Renner M, Choquet D, Triller A (2009) Control of the
sion of the molecules becomes correlated due to postsynaptic membrane viscosity. J Neurosci
the density of spines and their shape. It has been 29:2926–2937
Sabatini BL, Oertner TG, Svoboda K (2002) The life
experimentally shown that this process causes cycle of Ca(2+) ions in dendritic spines. Neuron
anomalous diffusion in Purkinje and hippocam- 33:439–452
pal pyramidal cells (Santamaria 2006, 2011). Santamaria F, Wils S, De Schutter E, Augustine GJ
(2006) Anomalous diffusion in Purkinje cell dendrites
caused by spines. Neuron 52:635–648
Acknowledgment This work was partially supported by Santamaria F, Gonzalez J, Augustine GJ, Raghavachari
NSF IOS-1209029 and NSF EF-1137897 S (2010) Quantifying the effects of elastic collisions
and non-covalent binding on glutamate receptor traf-
ficking in the post-synaptic density. PLoS Comput
References Biol 6:e1000780
Santamaria F, Wils S, De Schutter E, Augustine GJ
Bloodgood BL, Sabatini BL (2005) Neuronal activity (2011) The diffusional properties of dendrites depend
regulates diffusion across the neck of dendritic spines. on the density of dendritic spines. Eur J Neurosci
Science 310:866–869 34:561–568
Bressloff PC, Earnshaw BA (2009) A dynamic corral Schmidt H, Arendt O, Brown EB, Schwaller B, Eilers
model of receptor trafficking at a synapse. Biophys J (2007a) Parvalbumin is freely mobile in axons,
J 96:1786–1802 somata and nuclei of cerebellar Purkinje neurones.
Brown SA, Morgan F, Watras J, Loew LM (2008) J Neurochem 100:727–735
Analysis of phosphatidylinositol-4,5-bisphosphate Schmidt H, Kunerth S, Wilms C, Strotmann R, Eilers
signaling in cerebellar Purkinje spines. Biophys J (2007b) Spino-dendritic cross-talk in rodent Purkinje
J 95:1795–1812 neurons mediated by endogenous Ca2 + -binding pro-
Crank J (1975) The mathematics of diffusion. Clarendon, teins. J Physiol 581:619–629
Oxford Smith SJ, Buchanan J, Osses LR, Charlton MP,
Ehlers MD, Heine M, Groc L, Lee MC, Choquet D (2007) Augustine GJ (1993) The spatial distribution of cal-
Diffusional trapping of GluR1 AMPA receptors by cium signals in squid presynaptic terminals. J Physiol
input-specific synaptic activity. Neuron 54:447–460 472:573–593
Holcman D, Triller A (2006) Modeling synaptic dynamics Soler-Llavina GJ, Sabatini BL (2006) Synapse-specific
driven by receptor lateral diffusion. Biophys plasticity and compartmentalized signaling in cerebel-
J 91:2405–2415 lar stellate cells. Nat Neurosci 9:798–806
D 1022 Digital Filtering
A digital filter changes the shape of the input

Digital Filtering signal in a way that can be fully characterized by
means of its impulse response. For practical pur-
Roemer van der Meij and Jan-Mathijs Schoffelen poses, however, it is more convenient to describe
Donders Institute for Brain, Cognition and filter performance in terms of its frequency
Behaviour, Centre for Cognition, Radboud domain response.
University Nijmegen, Nijmegen,
The Netherlands Frequency Domain Performance
A filter is designed to keep certain spectral content
and remove others. For example, a high-pass filter
Definition only passes spectral content from the input signal
above a certain cutoff frequency, and a band-stop
Digital filtering is the process of transforming filter suppresses spectral content in a certain fre-
a discretely sampled input signal into an output quency band. A more detailed account of filter
signal, such that certain spectral characteristics of performance is given by the filter’s amplitude and
the input signal are lost, while others are retained. phase response as a function of frequency (Fig. 1a,
In neuroscience, it is performed on time series b). This can be obtained by taking the discrete
that represent electrophysiological or hemody- Fourier transform of the impulse response (Fig. 1c).
namic signals measured over time. Whereas ana- The ideal amplitude response would be
log filters are applied online and implemented as a boxcar function, indicating that the amplitude
electronic circuits, digital filters are applied of the “pass” frequencies is unaffected and the
off-line and implemented in software. amplitude of the “stop” frequencies is fully
suppressed. In reality, however, the transition
from passband to stopband is in general much
Detailed Description more gradual, and the amount of attenuation in
the stopband is variable. Fig. 1a (blue) shows as
Introduction an example the amplitude response of a 30 Hz
A digital filter is an important signal processing low-pass filter, with a slow but deep roll-off. The
tool for the analysis of neuroscientific data. It is filter in Fig. 1a (green) has a steep but oscillating
used to increase sensitivity to aspects of the sig- roll-off. A filter can also have unwanted effects
nals that are of interest while suppressing noise. on the “pass” frequencies.
For the purpose of simplicity, we will focus on The phase response shows to what extent the
electrophysiological time series, e.g., temporal phase of the oscillatory components in the input
fluctuations of the electric potential measured at signal is affected by the filter. Filters that do not
the scalp (for an extensive treatment of digital introduce a phase shift, i.e., the peaks of the oscil-
filters see Smith, 2003). Yet, a digital filter can be lations in the output signal occur at the same time
applied to any discrete signal sampled as points as the peaks of the oscillations in the input
a function of time, e.g., hemodynamic time signal, are called zero phase, or acausal filters. In
series, but also to signals sampled across space contrast, nonzero phase or causal filters introduce
such as an anatomical image obtained with mag- a phase shift, which can be either linear as
netic resonance imaging (MRI). Typical applica- a function of frequency (Fig. 1b, green) or nonlinear
tions of digital filters in computational (Fig. 1b, blue). The desired phase response depends
neuroscience include the removal of (power-) on the aim of the analysis step. In most cases a zero
line noise, high-pass filtering prior to rectification phase or acausal filter is desired.
for the analysis of multiunit activity, and
low-pass filtering for the reduction of high- Affecting the Amplitude Response
frequency noise prior to event-related potential The amplitude response of a filter is determined
averaging. mainly by the filter order. Filter order can be
Digital Filtering 1023 D
Digital Filtering, Fig. 1 Example of two 30 Hz low-pass of the FIR filter (green) and the slow but deep roll-off of
(forward) filters. (a–c) blue lines are the 4th-order IIR the IIR filter (blue). (b) Phase response. The FIR filter has
forward filter (Butterworth), and green lines are the a linear phase for the first part of the stopband. The IIR
90th-order FIR forward filter (window method). (a) filter has a nonlinear phase throughout the stopband. (c)
Amplitude response. Note the fast but oscillating roll-off Impulse response
thought of as the “length” of the filter. It specifies (Fig. 2a, blue). In contrast, a low filter order
how much time-domain information from the results in a “slow” amplitude response (slow
input signal is used to construct each sample of roll-off) (Fig. 2a, green). Additionally, the longer
the output signal. The higher the filter order, the the impulse response, the further oscillatory com-
better individual frequencies can be distin- ponents are time shifted in the output signal, i.e.,
guished, thus leading to “faster” transitions in the bigger the phase shift in the phase response
the amplitude response (also termed fast roll-off) (Fig. 2b). Note that there is always a trade-off
Digital Filtering, Fig. 2 Example of a high and low response. The higher-order filter shows a considerably
order 30–60 Hz band-pass (forward) FIR filter. (a–c) quicker passband to stopband transition. Higher-order fil-
blue lines are the 360th-order FIR forward filter ters use more time-domain information and can thus more
(window method), and green lines are the 90th-order easily distinguish neighboring frequencies. (b) Phase
FIR forward filter (window method). (a) Amplitude response. (c) Impulse response
between frequency and time resolution (Fig. 2c). In applications where the phase of the oscillations
A higher filter order trades time resolution for is relevant, it should always be taken into
a sharper amplitude response. account. Practically, any causal filter can be
made acausal. This is achieved by filtering the
Affecting the Phase Response input signal twice, in opposite directions. The
Typically the phase response is considered to be phase shift introduced in the first filter step is
of less importance than the amplitude response. “undone” by the filter applied to the signal in
Digital Filtering 1025 D
Digital Filtering, Fig. 3 The effect of bidirectional fil- a considerable linear and nonlinear phase shift. The bidi-
tering with a 30–60 Hz band-pass 90th-order FIR filter rectional filter, however, has zero phase, indicating that all
(window method). (a–c) blue lines represent filtering only of the phase shifting of the forward filter is reversed. (c)
in the forward direction, and green lines represent forward Impulse response. In this example, the impulse was cen-
and reverse filtering. (a) Amplitude response. The bidi- tered at t = 0. Because the forward filter was causal, the
rectional filter (green) has a much stronger amplitude impulse response occurs after t = 0. The bidirectional
attenuation of the stopband but also has a faster roll-off. filter is acausal; however, the impulse response is centered
(b) Phase response. The original forward filter shows around t = 0
reverse (Fig. 3). These two steps are also called Two Classes of Filters
forward and backward filtering. Note that this Of the many types of filters that exist, two broad
bidirectional filtering attenuates all amplitudes classes can be distinguished: (1) convolution, or
twice as compared to unidirectional filtering. finite impulse response (FIR) filters, and
Digital Filtering, Fig. 4 Example of edge artifacts and immersed in the data, causing residual 60 Hz noise to
their prevention. Panel A shows 1.5 s of example data with present. In order to remedy this, a larger data segment is
a strong 60 Hz line noise component on top of it. In order filtered (200 ms data padding on each side; panel C; gray
to remove this, a 59–61 Hz band-stop fourth-order bidi- boxes). The edge artifacts now occur in the padded seg-
rectional IIR filter (Butterworth) is used (panel B). How- ments. After removal of the padded segments, the original
ever, at the edges of the signal, the filter is not fully 1.5 s of data is artifact-free
(2) recursive, or infinite impulse response (IIR) weights are given by the filter coefficients. As
filters. FIR filters operate by convolution of the a consequence the length of the impulse response
input signal with the filter coefficients. In short, is equal to the filter order, which gives the FIR
the result of the convolution operation is that each filter its name. The filter coefficients can be
sample in the output signal is constructed as constructed in various ways. A common
a weighted sum of N samples of the input signal, approach is the “window method,” in which an
where N represents the filter order and the ideal IIR filter is constructed and then truncated
2-Dimensional Optical Imaging Spectroscopy 1027 D
by a particular window function (such as such as edge artifacts (described in the next
a Hamming window) of N samples. Another section).
approach is the “least squares method,” in
which a filter of order N is found whose ampli- Edge Artifacts
tude response is the least squares approximation Apart from desired effects, a filter can also have
of the desired amplitude response. undesired effects. A common one is an edge
IIR filters operate in a different way. In this artifact, from which all filters suffer. When filter-
case each sample of the output signal is ing the beginning of a signal, the filter is not fully
constructed from a weighted combination of immersed in the data. Because only a part of D
N samples of the input signal and of N-1 previ- the filter is used, the performance is very poor,
ously obtained samples of the output signal. resulting in “filter ringing” at the edges
Because of this recursion, all previous samples (Fig. 4a, b). It takes an FIR filter at least N samples
in the input signal contribute to the output signal. of the input signal to stabilize, where N is the
The impulse response thus is infinitely long, filter order. For an IIR filter it is more compli-
which gives the IIR filter its name. Still, most of cated, as the other parameters play an equally
the energy of the impulse response is concen- important role. The easiest solution to edge arti-
trated in a limited number of samples, the length facts is data padding. By filtering a longer signal
of which depends on the filter order, the pass/stop than required, the artifacts end up in the
frequencies, and their range. Commonly used IIR uninteresting parts of the signal, which can be
filters are Butterworth and Chebyshev filters. cut off afterward (Fig. 4c). Other padding
Which class of filter is most optimal depends methods, like mirror/zero/mean padding, are
on the situation. In general FIR filters are pre- less optimal, because these may introduce sharp
ferred when temporal spread of the impulse transitions in the signal.
response is undesired. An example would be
band-pass filtering followed by a Hilbert trans-
form to estimate temporal fluctuations in the References
amplitude and phase of band-limited signal com-
ponents. In this case it is desired to have full Smith SW (2003) Digital signal processing: a practical
control over the temporal information that is guide for engineers and scientists. California Techni-
used to construct the filtered signal, because one cal Publishing, San Diego
would like to relate more or less precisely the
“instantaneous” amplitude to the time point at Further Reading
Mitra S (2010) Digital signal processing. McGraw-Hill
which that amplitude occurred. The FIR filter Science/Engineering/Math, New York, NY
order directly specifies the temporal information Nitschke JB, Miller GA, Cook EW (1998) Digital filtering
that is used. An IIR filter will only provide very in EEG/ERP analysis: some technical and empirical
indirect control of temporal spread and is there- comparisons. Behav Res Methods Instrum Comput
30(1):54–67
fore less ideal to use in this scenario. Note that Percival DB, Walden AT (1993) Spectral Analysis for
FIR filter bandwidth must be sufficiently large for Physical Applications. Cambridge University Press,
a well-behaved amplitude response. When the Cambridge
length of the impulse response is unimportant,
IIR filters are often the preferred choice, as they
are more computationally efficient than FIR fil-
ters. An example would be the suppression
(power-)line noise. In this case it is irrelevant 2-Dimensional Optical Imaging
how much temporal information is used, as long Spectroscopy
as the noise is maximally suppressed. Temporal
spread of the impulse response is important to ▶ Voltage Sensitive Dye Imaging, Intrinsic
keep in mind for other practical considerations Optical Signals
D 1028 Direct Current Nerve Block (DC Block)
may also require the directed information in the

Direct Current Nerve Block reverse direction to be small.
(DC Block) It is important, however, to observe that infer-
ring causality from observations is an ill-posed
▶ Peripheral Nerve Stimulation Technique, question. This has been discussed in depth in the
Nerve Block context of Granger causality (Granger 1969),
where a rich literature exposes several fundamen-
tal issues. The same qualitative issues apply to
any causality argument based on directed infor-
Directed Information Flow and mation measures. The difference is that Granger
Causality in Neural Systems causality uses correlation measures, whereas
here, we consider directed information.
Michael C. Gastpar
University of California, Berkeley, CA, USA
Ecole Polytechnique Fédérale de Lausanne Detailed Description
(EPFL), Lausanne, Switzerland
Directed Information
The usual mutual information can be defined
Definition between any two random variables. For directed
information, however, we need to consider
In the human experience, information typically ordered sequences of random variables.
flows from one place to another. By contrast, the Specifically, for two random sequences
notion of mutual information introduced by X = (X1, X2, . . ., XN) and Y = (Y1, Y2, . . ., YN)
Shannon (1948) is perfectly symmetric in its of equal length N, the directed information is
arguments and does not distinguish between defined as
“input” and “output.” In this sense, it is perhaps
surprising that this very measure of information X
N
indeed captures the capacity of any communica- I ðX ! Y Þ ¼ fI ð X 1 , . . . , X n ; Y 1 , . . . , Y n Þ

n¼1
tion channel – though we should recall that the
proof of this fundamental fact is not merely I ðX1 , . . . , Xn ; Y 1 , . . . , Y n1 Þg,
a simple consequence of Shannon’s definition.
In spite of Shannon’s strong and fundamental where I(;) denotes the regular Shannon mutual
results, it has been tempting to define a notion of information. Let us first record that this is always
directed information. This was first proposed in nonnegative: I(X ! Y) 0, which follows
(Marko 1973) for stationary processes. The more directly from the definition and the fact
general and useful definition was given in the that Shannon mutual information is always
brief and beautiful note by Massey (1990). More- nonnegative. Second, we note that the directed
over, Massey (1990), Kramer (1998) and subse- information is never larger than the regular Shan-
quent work revealed that directed information has non mutual information between the two
a natural place in the study of information trans- sequences.
mission with feedback from the output to the Moreover, it is tempting to ask about the rela-
input. tionship between the forward directed informa-
In the present note, a different aspect of tion I(X ! Y) and the reverse directed
directed information is illuminated: that of iden- information I(Y ! X). For general probabilistic
tifying causal relationships. The basic idea is that models, there does not appear to be an intuitively
if there is a causal relationship from one process pleasing connection between these two. How-
to another, the directed information in the forever, if one considers instead the following new
ward direction should be large. Additionally, one random sequence Ỹ = (0, Y1, Y2, . . ., YN1), then
Directed Information Flow and Causality in Neural Systems 1029 D
the following conservation law applies (Massey scenario just discussed involving the sequences
et al. 2005): X, Y, and Z. Then, even if both I(X ! Y) and
I(Y ! Z) are large, this does not imply any

I ðX ! Y Þ þ I Y~ ! X ¼ I ðX; Y Þ, (1) lower bound on I(X ! Z); in fact, the latter
might even be zero.
where the latter denotes the regular Shannon
mutual information between the two sequences Information Measure of Causality
X and Y. Directed information can be postulated to be
To have a brief and simple example, let us a measure of causality in the following sense: D
consider the situation where (X1, X2, . . ., XN) are One claims a causal relationship from X to Y if
independent normal (Gaussian) random variables the directed information I(X ! Y) is large. In the
of mean zero and variance 1. Let Y1 also be explicit example involving normal random vari-
a mean-zero unit-variance normal. For n 2, ables discussed above, this is easy to see: The
we set Yn = aXn1 + Wn, where Wn are indepen- directed information from X to Y is large, and
dent normal random variables of mean zero and indeed, for this example, we would expect any
variance 1 a2 and 0 a 1. Then, the rationale to conclude that the sequence X drives
directed information evaluates to the sequence Y in a causal fashion (at least as long
as a is close to one). Moreover, in this example,
I(Y ! X) = 0, ruling out any causal relationship
N1 a2
I ðX ! Y Þ ¼ log2 1 þ : (2) in the reverse direction.
2 1 a2
The remaining issue is to define a threshold on
In this example, the reverse directed informa- the directed information above which the rela-
tion vanishes: I(Y ! X) = 0. tionship is claimed to be causal. There is no
Finally, we also note that another 10 years intuitive a priori rule, and in most cases, the
later, directed information was again discovered threshold must be selected arbitrarily or using
(interestingly again with a stationarity assump- additional knowledge of existing causal connec-
tion) under the name of transfer entropy tions, e.g., from physiological insight into the
(Schreiber et al. 2000). considered connection. To make matters more
complicated, it is also important to notice that
Directed Information in Networks due to the nonnegativity of directed information,
The concept of information flow requires that most classical estimators must be expected to
information cannot be created simply by have a positive bias.
processing a signal in a particular way. For mutual It should also be noted that this is closely
information, this is ensured by the data processing related in spirit to the notion of Granger causality
inequality, stating that the mutual information can- (Granger 1969). In the latter, correlation (i.e.,
not be increased by further processing of the out- second-order statistics) is exploited to claim cau-
put signal. For directed information, the situation sality. Directed information, by contrast, is sen-
is a little more complex. Namely, a data processing sitive to the full probability distribution in the
inequality still applies but only to causal usual entropy sense. Note that when directed
processing of the output signal. That is, if Z is information is considered with normal distribu-
a sequence obtained by processing the sequence tions, it is closely related to Granger causality.
Y in a causal fashion, possibly involving additional
independent randomness, then we have that Experimental Studies
Directed information measures have been
I ðX ! Z Þ minfI ðX ! Y Þ, I ðY ! Z Þg: (3) recently applied to simultaneous recordings in
the primary motor cortex of rodents and macaque
As with all information measures, the reverse monkeys (Quinn et al. 2011; So et al. 2012), lead-
is not true. More precisely, let us reconsider the ing to conjectured causality maps (directed
D 1030 Directed Spectral Methods
graphs) between the observed neurons. An addi- a (linear) multivariate autoregressive (MVAR)
tional rich literature concerns the transfer entropy model of the data. The various methods quantify
mentioned above, but is discussed elsewhere. in different ways the strength of interaction terms
in the Fourier transform of the MVAR model.
While these methods are sometimes referred to
References as capturing “causal” or “effective” connectivity,
they are most properly described as reflecting
Granger CWJ (1969) Investigating causal relations by “directed functional connectivity” (Friston
econometric models and cross-spectral methods.
et al. 2013).
Econometrica 37(3):424438
Kramer G (1998) Directed information for channels with
feedback, vol 11, ETH series in information
processing. HartungGorre, Konstanz Detailed Description
Marko H (1973) The bidirectional communication theory
a generalization of information theory. IEEE Trans For a multivariate n-channel process
Commun 21:1345–1351 X(t) = [X1(t), X2(t), . . ., Xn(t)]T (with zero
Massey JL (1990) Causality, feedback and directed infor-
mation. In: Proceedings of the 1990 international sym- mean), the MVAR model is given by
posium on information theory and its applications,
Hawaii, pp 303–305 X
p
Massey JL, Massey PC (2005) Conservation of mutual XðtÞ ¼ Ak Xðt kÞ þ ϵ ðtÞ, (1)
and directed information. In: Proceedings of the 2005 k¼1
international symposium on information theory,
Adelaide, pp 157–158 where Ak are matrices of regression coefficients,
Quinn C, Coleman TP, Kiyavash N, Hatsopoulos NG
(2011) Estimating the directed information to infer p is the model order, and ϵ(t) are the residuals
causal relationships in ensemble neural spike train (L€utkepohl 2007). The coefficients of this model
recordings. J Comput Neurosci. doi:10.1007/s10827- are uniquely specified by imposing zero correla-
010-0247-2 tion between the residuals and the regressors
Schreiber T (2000) Measuring information transfer. Phys
Rev Lett 85:461–464 X(t – k), k = 1,. . ., p, and can be derived from
Shannon CE (1948) A mathematical theory of communi- the Yule-Walker procedure (Ding et al. 2006).
cation. Bell Syst Tech J 27:379–423, and 623–656 Ideally, the model order should be sufficiently
So K, Koralek AC, Ganguly K, Gastpar MC, Carmena JM high to obtain a good representation of the data,
(2012) Assessing functional connectivity of neural
ensembles using directed information. J Neural Eng but not so high that overfitting and poor parame-
9:026004 ter estimation occur. Commonly, the model order
is selected based on minimizing either the Akaike
information criterion or the Bayes information
criterion, both of which balance the residual var-
Directed Spectral Methods iance against the number of coefficients to be
estimated (Ding et al. 2006).
Adam B. Barrett and Anil K. Seth
Directed spectral measures are computed from
Sackler Centre for Consciousness Science and
the frequency-domain representation of the
Department of Informatics, University of Sussex,
MVAR model:
Brighton, UK
AðoÞ X~ðoÞ ¼ ϵ~ðoÞ, (2)

Definition
where explicitly in terms of the regression
Directed spectral measures quantify, in the fre- coefficients
quency domain, directed statistical interactions
X
p
between time-series variables. Most commonly, Að o Þ ¼ I z k Ak , (3)
the various measures are computed based on k¼1
Directed Spectral Methods 1031 D
z = e 2ipoDt, and Dt is the time between obser- S11 ðoÞ ¼ H 11 ðoÞs11 H 11 ðoÞ þ H 12 ðoÞs22 H 12 ðoÞ:
vations. (Let tildes denote Fourier transform and (6)
dagger the Hermitian conjugate.) The measures
are written in terms of A(o), the spectral density GGC is given (after the transformation) by
(or power) SðoÞ ¼: X~ðoÞ X~† ðoÞ, the covariance
matrix of the residuals S, and also the transfer
S11 ðoÞ
matrix H(o) =: A1(o), which yields X~ðoÞ when GGC2!1 ðoÞ ¼: ln
H 11 ðoÞs11 H 11 ðoÞ
: (7)
it acts on the Fourier transformed residuals:
D
To analyze a system of n variables, one can
X~ðoÞ ¼ H ðoÞ ϵ~ðoÞ: (4) compute the above pairwise GGC between each
pair of variables by fitting MVAR models suc-
Nonparametric Estimation. An alternative cessively to each pair of variables. There also
method for computing directed spectral measures exists (i) the more complicated GGCj!ijk1 ,..., k‘ ,
involves obtaining the transfer matrix directly i.e., the conditional GGC from Xj to Xi condi-
from Fourier and wavelet transforms of the data, tional on Xk1,. . ., Xk‘ (Ding et al. 2006; Geweke
obviating the need to explicitly fit an MVAR 1984), which computes in the same fashion the
model (Dhamala et al. 2008; Percival and Walden causal contribution of Xj to the power of Xi that is
2000). An advantage of this approach is that it independent of the causal contributions of Xkl,. . .,
bypasses the need for model order estimation; Xkl to the power of Xi and the intrinsic power of
however, this is replaced with the task of choos- Xi, and (ii) “multivariate GGC” from one group
ing the wavelet prototype and number of tapers of variables to another group of variables (Barrett
appropriately. et al. 2010; Geweke 1982).
The spectral GGC is related to its time-domain
Granger-Geweke Causality (GGC) formulation (Ding et al. 2006; Geweke 1982;
For a pair of variables X(t) = [X1(t), X2(t)]T, the Granger 1969). The time-domain GGC from X2
Granger-Geweke causality GGC2!1(o) from X2 to X1 quantifies the extent to which the past of X2
to X1 at frequency o quantifies the directed con- helps predict the future of X1 over and above the
tribution of X2 to the power of X1 at frequency o extent to which the past of X1 predicts its own
(Ding et al. 2006; Geweke 1982). It does this by future. Specifically, it is given by the logarithm of
expressing the total power of X1 as the sum of an the ratio of the residual variance in
“intrinsic” term and a “causal” term and compar- a restricted regression of X1 on its past to the
ing the logarithms of the total power and the residual variance s11 in the (unrestricted)
intrinsic power. From an MVAR model of the MVAR model of [X1, X2]T. The mean spectral
two variables, the power can be expressed via GGC over all frequencies up to the Nyquist fre-
Eq. 4 as quency is equal to the corresponding time-
domain GGC.
SðoÞ ¼ H ðoÞ S H † ðoÞ: (5)
Partial Directed Coherence (PDC)
The decomposition of this into an intrinsic term The partial directed coherence PDCj !i(o) is
and a causal term is obtained via the transforma- defined as (Baccalá and Sameshima 2001)
tion of variables X2 !X2 – (s21/s11) X1, which
diagonalizes S and hence removes the cross- Aij ðoÞ
terms from the RHS. (Since GGC is interested PDCj!i ðoÞ ¼: qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn : (8)
Akj ðoÞ 2
in the independent causal contribution from X2 to k¼1
X1, it should by definition be invariant under the
addition of a multiple of X1 to X2.) After this This measure represents the relative coupling
transformation, we have strength of the interaction from Xj to Xi, as
D 1032 Directed Spectral Methods
compared to all of Xj’s interactions as Linearity. The linear formulation of these

a source with other structures. The PDC ranks measures can obviously only give information
the relative interaction strengths with respect to about linear features of signals. Extension to
a given signal source, and the normalization is nonlinear cases can however present further
such that problems. Marinazzo et al. (2011) review some
nonlinear approaches to GGC. Note that for
2
0 PDCj!i ðoÞ 1, (9) (stationary) Gaussian variables, interactions are
necessarily linear and time-domain GGC is
Xn equivalent to transfer entropy, enabling an inter-
PDCj!i ðoÞ 2 ¼ 1: (10)
i¼1 pretation of GGC in terms of Shannon informa-
tion flow (Barnett et al. 2009; Barrett et al. 2010).
Directed Transfer Function (DTF)
The directed transfer function DTFj !i(o) is Neurobiological Application
defined as (Kaminski and Blinowska 1991) The above limitations should always be kept in
mind when applying these measures to neurobi-
H ij ðoÞ
DTFj!i ðoÞ ¼: qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xn : (11) ological data. A general principle is that shorter
2
k¼1
j H ik ð o Þ j data segments are more stationary than longer
data segments, but yield poorer parameter esti-
This measure represents the coupling strength of mates. Thus, a trade-off must be made in choos-
the interaction from Xj to Xi, as compared to the ing the segment length; the most appropriate
sum of those from all variables to Xi. Similar to choice will depend on whether the data are
PDC the normalization is such that “event related” or “steady state” and on the esti-
mated model order (Barrett et al. 2012; Cohen
2 2014). Typically, segment lengths range from
0 DTFj!i ðoÞ 1, (12)
500 to 2,000 ms (Cohen 2014). These measures
Xn are applicable to continuous variables and there-
DTFj!i ðoÞ 2 ¼ 1: (13) fore to M/EEG or local field potential data, as
i¼1
opposed to spiking data. The sampling rate
Since it makes use of the inverse of the regression should be sufficiently high for the Nyquist fre-
matrix, the DTF measure is a linear combination quency to be substantially greater than the
of direct and indirect couplings. The “direct highest frequency of interest, but not so high
DTF” (dDTF) variant is an alternative that that the model order becomes too large;
emphasizes direct connections (Korzeniewska 200–250Hz is typical (Cohen 2014). fMRI typi-
et al. 2003). Note that, unlike for GGC, there is cally has too slow a sampling rate to make use of
no corresponding time-domain connectivity map these measures, although novel ultrarapid acqui-
for PDC or DTF. sition sequences potentially enable access to
low-frequency interactions (Seth et al. 2013).
Limitations and Extensions For M/EEG, preprocessing should include some
Stationarity. These measures assume that the form of spatial filtering or source localization in
time series are covariance stationary. order to minimize volume conduction artifacts
Nonstationary data should be divided into win- (Bressler and Seth 2011; Cohen 2014). Band-
dows that by themselves are approximately sta- pass temporal filtering should be avoided as this
tionary (Bressler and Seth 2011). smears the data in time and can increase model
Dependence on Variables. The measures orders, thus leading to poorer estimates of the
reflect directed statistical dependencies amongst measures (although notch filtering to remove
the particular set of variables chosen and should line noise will not contaminate frequencies that
not be interpreted as directly reflecting physical are far away from the stopband; Barnett and Seth
causal chains. 2011).
Distributed Robustness 1033 D
Summary and Comparison of Measures In: Schelter S, Winterhalder M, Timmer J (eds) Handbook
GGCj!i is a measure of the directed functional of time series analysis. Wiley, Weinheim, pp 438–460
Friston K, Moran R, Seth AK (2013) Analyzing connec-
effect that Xj has on Xi and compares “intrinsic” tivity with Granger causality and dynamic causal
and “causal” contributions to the power in modelling. Curr Opin Neurobiol 23:1–7
a bivariate autoregressive model (conditional Geweke J (1982) Measurement of linear dependence and
versions of GGC are used if one wants to separate feedback between multiple time series. J Am Stat
Assoc 77:304–313
out direct and indirect causal effects). PDCj!i Geweke J (1984) Measures of conditional linear depen-
captures the relative strength of the direct dence and feedback between time series. J Am Stat
pairwise coupling from Xj to Xi, compared to the Assoc 79:907–915 D
total coupling from Xj to other variables in the Granger C (1969) Investigating causal relations by econo-
metric models and cross-spectral methods.
frequency-domain MVAR model for the set of Econometrica 37:424–438
variables. DTFj!i compares the coupling from Xj Kaminski M, Blinowska K (1991) A new method of the
to Xi with the total coupling from all variables to description of the information flow in the brain struc-
Xi in the transfer matrix for the set of variables; as tures. Biol Cybern 65(3):203–210
Korzeniewska A, Manczak M, Kaminski M, Blinowska K,
such it is a measure of the combined strength of Kasicki S (2003) Determination of information flow
direct and indirect couplings from Xj to Xi. These direction among brain structures by a modified
measures, of which GGC is the most common directed transfer function (dDTF) method. J Neurosci
and perhaps best founded, hold great promise in Method 125(1–2):195–207
L€
utkepohl H (2007) New introduction to multiple time
moving beyond functional localization towards series analysis. Springer, Berlin/Heidelberg
the important problem of dissecting the func- Marinazzo D, Liao W, Chen H, Stramaglia S (2011)
tional circuits underlying cognition, perception, Nonlinear connectivity by Granger causality.
and behavior. Neuroimage 58(2):330–338
Percival D, Walden A (2000) Wavelet methods for time
series analysis. Cambridge University Press,
Cambridge
Seth AK, Chorley P, Barnett LC (2013) Granger causality
References analysis of fMRI BOLD signals is invariant to hemo-
dynamic convolution but not downsampling.
Baccalá LA, Sameshima K (2001) Partial directed coher- Neuroimage 65:540–555
ence: a new concept in neural structure determination.
Barnett LC, Seth AK (2011) Behaviour of Granger cau-
sality under filtering: theoretical invariance and prac-
tical application. J Neurosci Method 201:404–419
Barnett L, Barrett AB, Seth AK (2009) Granger causality Directional Motion
and transfer entropy are equivalent for Gaussian vari-
ables. Phys Rev Lett 103:238701
Barrett AB, Barnett L, Seth AK (2010) Multivariate
▶ Optic Flow Processing
Granger causality and generalized variance. Phys
Rev E 81:041907
Barrett AB, Murphy M, Bruno M-A, Noirhomme Q, Boly
M et al (2012) Granger causality analysis of steady-
state electroencephalographic signals during propofol-
induced anaesthesia. PLoS One 7(1):e29072 Disparity Energy Model
Bressler SL, Seth AK (2011) Wiener-Granger causality:
a well established methodology. Neuroimage ▶ Spatiotemporal Energy Models
58(2):323–329
Cohen MX (2014) Analyzing neural time series data. MIT
Press, Cambridge, MA
Dhamala M, Rangarajan G, Ding M (2008) Analyzing
information flow in brain networks with
non-parametric Granger causality. Neuroimage
41(2):354–362
Distributed Robustness
Ding M, Chen Y, Bressler S (2006) Granger causality:
basic theory and application to neuroscience. ▶ Neuronal Parameter Non-uniqueness
D 1034 Distributed Speech Processing
altered in psychiatric disorders (Strange 2001). DA

Distributed Speech Processing neurons fire a burst of action potentials whenever
a subject or an animal receives an unpredicted
▶ Neural Coding of Speech Sounds reward or a reward-predicting stimulus (Schultz
2002, but see Redgrave and Gurney 2006). Do the
DA neurons synthesize the reward-related signal by
some type of neuronal computation, or do they
Dopaminergic Cell Models simply relay a signal generated elsewhere? This
question motivates the research on the DA neuron
Alexey Kuznetsov1 and Boris Gutkin2,3 and leads to the more specific questions addressed
1
Department of Mathematical Sciences, Indiana in experiments and modeling.
University & Purdue University Indianapolis,
Indianapolis, IN, USA Anatomy
2
Group for Neural Theory, Laboratoire de Midbrain DA neurons are positioned within three
Neurosciences Cognitives (LNC), Département cell groups in different brain areas: the ventral
d’Études Cognitives, École Normale Supérieure, tegmental area (VTA), the substantia nigra pars
Paris, France compacta (SNc), and the retrorubral area. There
3
National Research University Higher School of is no clear anatomical distinction between the
Economics, Center for Cognition and Decision groups, but heterogeneities in biophysical prop-
Making, Moscow, Russia erties of the DA cells both within and between
these groups are documented (see, e.g., Wolfart
et al. 2001; Neuhoff et al. 2002). Nevertheless,
Definition the characteristics listed below are common for
all three groups of midbrain DA neurons.
• Dopaminergic (DA) neurons are defined as
neurons synthesizing and containing the neu- Firing Patterns
rotransmitter and neurohormone dopamine. DA neurons show multiple activity patterns:
Such neurons release dopamine synaptically low-frequency regular (tonic) spiking and differ-
as well as dendritically (Bustos et al. 2004). ent types of bursting. Mechanisms of the firing
• The electrophysiological signatures of this patterns of the DA neuron are currently debated
neuron are broad action potentials (Fig. 1) in the literature. Here, we present most
and a low-frequency, regular spontaneous established results and the different views.
activity (Fig. 2a).
• A distinctive property of the DA neuron is that Background Firing
it differentially responds to different types of In isolation, DA neurons spontaneously display
excitatory synaptic inputs (Fig. 2a and b). low-frequency (1–5 Hz) tonic spiking. A less regular
• Models have explained these properties and pattern of a similar average frequency is most com-
connected them with one another and with par- monly seen in vivo. Together, these two
ticular current compositions (Figs. 3 and 4). low-frequency patterns are called background firing.
The major properties of background firing are:
• Injection of a depolarizing current can increase
Detailed Description the frequency of repetitive firing only up to
10 Hz (Fig. 2b; Richards et al. 1997).
Importance • Higher firing rates are possible during
The central dopamine system is involved with many a transient at the onset of the depolarization.
behavioral and cognitive tasks (Carlson 1999). • Modeling suggests that the frequencies
Dopamine neurotransmission is found to play remain low due to very weak spike-producing
a role in addictive behavior (Wise 2004) and is currents:
Dopaminergic Cell Models 1035 D
et al. 2009; Blythe et al. 2009; Khaliq and
Bean 2010).
• Modeling has shown that the sodium and cal-
cium currents contribute to the same oscillatory
mechanism responsible for both spiking and
subthreshold oscillations (Drion et al. 2011).
• The calcium-based oscillatory mechanism is
distributed along dendrites (Wilson and
Callaway 2000), and a special type of syn- D
chrony among dendrites and the soma sets in
due to electrotonic coupling (Medvedev and
Kopell 2001; Medvedev et al. 2003).
Burst Firing
The stereotypic response of the DA neuron to an
unpredicted reward is a burst of activity with
firing frequencies above 20 Hz. Different exper-
iments result in bursts that are significantly dif-
Dopaminergic Cell Models, Fig. 1 A single spike of ferent from one another (see, e.g., Johnson and
a DA neuron
Wu 2004; Overton and Clark 1997). A central
– The current initiating the spike is weak: the question is how similar is in vitro bursting to
maximal conductance of the fast sodium that observed in vivo. The main features of each
current is drastically lower than in fast- type of bursting in the DA cell are described
spiking neurons (Kuznetsova et al. 2010). below with the following common properties to
– The spike-initiating current remains be emphasized:
inactivated: the fast-activating potassium • In both in vivo and in vitro, bursting is depen-
currents causing hyperpolarization after dent on the activation of NMDA receptors
a spike (afterhyperpolarizing or AHP cur- (Fig. 2a).
rents) cannot fully remove inactivation of • Activation of AMPA receptors alone does not
the fast sodium current (Kuznetsov elicit bursting (Fig. 2b; Chergui et al. 1993;
et al. 2006; Drion et al. 2011). Overton and Clark 1997; Tong et al. 1996;
This has two consequences, which contrast Morikawa et al. 2003; Deister et al. 2009).
the DA neuron with other neuron types: • Modeling has shown that the failure to burst in
1. The action potentials generated by these response to AMPAR stimulation follows
currents are broad (Fig. 1). immediately from the limited frequency of
2. Background firing must rely on the response to applied depolarization (see
a subthreshold oscillation to be robust. above and Kuznetsov et al. 2006).
• Two types of the subthreshold oscillations
were found: Bursts In Vivo
1. Calcium-based (Fig. 3; see, e.g., Harris Firing rates of 20 Hz can be achieved for irreg-
et al. 1989); the corresponding minimal ularly distributed short time intervals in the DA
model includes only two currents, L-type neuron in vivo (Hyland et al. 2002). The main
calcium and SK-type calcium-activated properties of these bursts are:
potassium (Amini et al. 1999; Wilson and • The bursts are singular events as opposed to
Callaway 2000). repetitive bursting (but see Freeman et al. 1985).
2. Sodium-based (calcium-independent; • Bursts are not superimposed on a depolarizing
Yung et al. 1991; Nedergaard and Green- plateau and may not be followed by a pause in
field 1992; Chan et al. 2007; Guzman activity.
D 1036 Dopaminergic Cell Models
subthreshold
oscillations
NMDAR activation
Current injection AMPAR activation
c Blockade of the calcium−dependent potassium current
plateau
potentials
Dopaminergic Cell Models, Fig. 2 Firing patterns of depolarization nor activation of the AMPA receptor
DA neurons. (a) Low-frequency background firing is evokes a burst. (c) The blockade of the calcium-dependent
replaced by a burst during NMDA receptor activation potassium current (by apamin) reveals subthreshold pla-
(e.g., by glutamate iontophoresis). The blockade of the teau potentials. The dashed line shows the spike threshold
spike-producing fast sodium current (with TTX) reveals of 40 mV
subthreshold oscillations. (b) Neither applied
• Within the burst, the spike amplitude progres- • NMDA but not AMPA receptor activation is
sively decreases, and the spike duration and critical for the burst (Chergui et al. 1993;
interspike interval progressively increase (but Overton and Clark 1997; Tong et al. 1996;
see Hyland et al. 2002). Meltzer et al. 1997).
• Burst evocation relies on excitatory synaptic • The frequency of burst occurrence can be reg-
inputs. ulated differentially from the average firing
Dopaminergic Cell
SK K+ channel
Models, Fig. 3 Calcium
processes generating the Ca2+ pump
subthreshold oscillations: L Ca2+ channel
the persistent L-type
calcium current is
responsible for
depolarization and activation
simultaneous accumulation
of intracellular calcium, membrane
whereas the SK-type D
calcium-activated
potassium current produces Ca2+
the repolarization phase of
the oscillation. These two
currents constitute Ca2+
a calcium-potassium
mechanism for pacemaking
in DA neurons Ca2+ buffering
soma GABAA receptor
NMDA receptor
Dopaminergic Cell Na+− K+ exchanger

Models,
Fig. 4 Generation of the
dendrite
oscillatory plateau
potentials that underlie
bursting induced by bath
application of NMDA
relies on the NMDAR
current and Na+–K+ Na+
exchanger
rate in vivo (Grace and Bunney 1984; Tepper oscillatory mechanism and increases the fre-
et al. 1995), which was explained in modeling quency of oscillations (Fig. 2a). This provides
(Canavier and Landry 2006). the distinction between responses to AMPA
• The burst is critically dependent on calcium and NMDA receptor activation.
(Grace and Bunney 1984). • In Komendantov and Canavier (2002) and
• SK channel blockers induce burst firing in vivo Canavier and Landry (2006), the role of the
(Ji and Shepard 2006; Waroux et al. 2005). NMDAR current is identified as opposing the
• Blockade of GABAA receptor currents has SK current and therefore reducing the ampli-
been shown to induce bursting in DA neurons, tude and duration of the AHP.
suggesting a disinhibition hypothesis of DA
burst firing in vivo (Tepper et al. 1995). In Vitro Bursting Induced by Stimulation
Modeling proposed two distinctly different of NMDA Receptors
mechanisms for in vivo bursting. Repetitive bursting can be elicited in DA
• In Kuznetsov et al. (2006) and Ha and neurons in vitro via bath application of NMDA
Kuznetsov (2009, 2013), the NMDAR current (Johnson and Wu 2004). The properties of this
directly interacts with the subthreshold bursting are:
• This is a repetitive bursting pattern. • The nonlinear voltage dependence of the

• The high-frequency firing is superimposed on NMDAR current is critical for evoking pla-
a depolarizing phase of a subthreshold oscil- teau potentials (Li et al. 1996; Canavier 1999;
lation called a plateau potential. Komendantov et al. 2004), as well as for the
• This bursting is Ca2+-independent: bursting high-frequency oscillations (Kuznetsov
persists in Ca2+-free solution or upon chela- et al. 2006; Ha and Kuznetsov 2013).
tion of intracellular Ca2. • By contrast, activation of linear currents, such
• This bursting is critically dependent on extra- as AMPAR or GABAR, suppresses plateau
cellular sodium (Johnson et al. 1992). potentials as well as the high-frequency firing.
These properties contrast with those for burst-
ing in vivo (see above). In Vitro Bursting Induced by Apamin
Other in vitro stimulation techniques evoked Repetitive bursting of the DA neuron was
bursts more similar to those found in vivo. A burst observed during the bath application of an SK
has been observed after extracellular electrical current blocker apamin (Fig. 2c; Shepard and
stimulation (a train of short pulses) or iontopho- Bunney 1991). The main properties of this burst-
resis of a glutamate receptor agonist onto the DA ing pattern are:
cell (Morikawa et al. 2003; Blythe et al. 2009; • This is also a repetitive bursting pattern.
Deister et al. 2009). • The high-frequency firing is superimposed on
Glutamate influences the firing through depolarizing plateau potentials.
NMDA receptors: • Blockade of spike generation preserves the
• Subsequent bath application of an NMDA oscillatory plateau potentials underling bursts
receptor antagonist inhibited the burst. (Fig. 2c; Ping and Shepard 1996).
• In dynamic clamp experiments, injection of • Modeling reproduces the transition from the
a virtual NMDA current elicited a burst low-frequency subthreshold oscillation to the
(Deister et al. 2009; Putzier et al. 2009). plateau potential oscillation upon simulated
• Removing the voltage dependence of the vir- blockade of the SK current (Amini
tual NMDAR current, which converts it into et al. 1999; Canavier et al. 2007; Oster and
AMPAR-like current, abolished the burst. Gutkin 2011).
The properties contrasting these bursts from • Without the SK current, DA cells become
bursting elicited by other methods in vitro and prone to depolarization block, and each burst
resembling bursts in vivo are: collapses to a plateau (Fig. 2c).
• Spikes within a burst are not superimposed on • Apamin-induced bursting is critically depen-
a depolarizing plateau. dent on calcium: L-type Ca2+ channel blocker
• The hyperpolarization after the burst origi- abolishes the plateau potential oscillations, and
nates from a separate pathway, which can be this bursting is abolished in Ca2+-free medium.
blocked or desensitized without affecting the • Modeling and modeling-driven experiments
burst. show that the ether-a-go-go (ERG) current
• Switches between the low-frequency back- contributes to the repolarization phase of the
ground firing and the high-frequency bursts plateau potential wave (Canavier et al. 2007;
follow the on-off pattern of the stimulation. Ji et al. 2012).
Thus, the high-frequency firing cannot be • Modeling suggests that the ERG current also
a consequence of the depolarizing phase of the provides repolarization after each spike and,
underlying plateau potentials. therefore, is involved in generation of low-
Modeling has shown that the high-frequency and high-frequency firing (Ha and
firing can be elicited independently of the plateau Kuznetsov 2013).
potentials (Kuznetsov et al. 2006). However, the Together, these characteristics suggest that
plateau potentials and the high-frequency firing this bursting pattern is different from in vivo
share certain properties: bursting (see above).
Dopaminergic Cell Models, Table 1 Comparison of different ways to induce a burst in the DA neuron
In vitro
Property\burst type In vivo NMDA iontophoresis Bath NMDA SK block Bath NMDA + SK block
Repetitive pattern No No Yes Yes Yes
Plateau potentials No No Yes Yes Yes
Ca2+ dependence Yes Yes No Yes No
Na+ dependence No No Yes No Yes
Bursting has been shown to be most robust in dependent potassium current (SK type) has the
½Ca2þ
4
simultaneous bath application of NMDA and
form I KCa ¼ gKCa 2þ 4 4 ðEK vÞ, which also
apamin. Properties of bursting in this case are ½Ca þ k
similar to those for whole bath NMDA-induced assumes instantaneous gating.
burst firing (see above) (Johnson and Wu 2004).
This suggests that apamin, by blocking Spike Generation
the SK current, can facilitate two different The delayed rectifier potassium current is taken
bursting mechanisms, one of which is intrinsic in a standard form I DR ¼ gDR n4 ðEK vÞ. Here,
and the other based on the NMDA receptor n is a gating variable, which obeys a standard
activation. Modeling reproduced the facilitation equation dn/dt = (ninf(v) n)/tn(v). The gating
of high-frequency firing during simulated function ninf(v) and time constant tn(v) in this
NMDAR activation combined with partial block- equation are calibrated to reflect high half-
ade of the SK current (Oster and Gutkin 2011; activation and slower activation kinetics of the
Table 1). current. This calibration and the current’s low
maximal conductance weaken the hyperpolariza-
Summary of Models tion produced by the delayed rectifier and lead to
All models of DA neurons have a conductance- a low safety factor or robustness of spike
based form generation.
The fast sodium current also takes a standard
dv X form with instantaneous activation
cm ¼ Ij
dt j I Na ¼ gNa m3inf ðvÞhðENa vÞ, where the function
minf(v) reflects the spike threshold of
where v is voltage and Ij are various transmem- 40 mV. The maximum conductance of this
brane currents. The following firing patterns were current is quite low, around 550 uS/cm2, which
reproduced by the inclusion of different currents gives a moderate slope of the spike upstroke and
in various models. broad action potentials (Fig. 1).
Subthreshold Calcium-Driven Oscillations Subthreshold Plateau Potentials

A calcium and a calcium-dependent potassium Slow subthreshold oscillations called plateau
current constitute a mechanism for subthreshold potentials are observed after the blockade of the
oscillations (Figs. 2a and 3). The calcium spike-producing sodium current in two
current has the form I Ca ¼ gCa c4inf ðvÞðECa vÞ, conditions:
which assumes instantaneous gating. The calibra- • The blockade of the calcium-dependent potas-
tion of the gating function cinf(v) reflects an sium current (Fig. 2c) or
activation threshold of an L-type current, which • The activation of a synaptic NDMA receptor
is significantly lower in DA cells than in current by bath application of an NMDAR
other neurons (~50 mV). The calcium- agonist
In the latter case (Fig. 4), voltage rises due to Weak Spike-Producing Currents Determine
the NMDAR current, and the subsequent voltage Low-Frequency and Broad Action Potentials
decline is due to a sodium pump current (Na+–K+ First, the high-frequency firing within bursts must
exchanger): be considered in the context of limited firing
frequencies in response to applied depolarization

1:5
I NP ¼ iNP = 1 þ K mNa =½Nai , (but see Blythe et al. 2009). The high frequency is
not a simple consequence of a depolarizing pla-
teau potential. The spike-producing currents can-
where [Na]i is the intracellular sodium
not reliably support repetitive firing, and both
concentration.
high- and low-frequency firing must rely on the
During the blockade of the calcium-dependent
subthreshold oscillations to be robust. The main
potassium current, the subthreshold depolariza-
reason is that the fast AHP currents are weak, as
tion is produced by the calcium current.
suggested by modeling studies. Additionally, the
Two mechanisms were proposed for the repolar-
fast sodium current has a drastically lower
ization phase: the calcium pump current

maximal conductance compared to that in
I CP ¼ iCP = 1 þ K mCa =½Cai or the ether-a-go-go fast-spiking neurons. This results in even weaker
(ERG) current I ERG ¼ gERG eðEK vÞ, where e is activation of the fast AHP currents and simulta-
its activation variable. neously ensures the broad action potentials,
which are characteristic for DA neurons.
Responses to Synaptic Inputs
An NMDA receptor current elicits high- Calcium-Dependent and Calcium-Independent
frequency firing in the DA neuron. The nonlinear Oscillations
voltage dependence of the current is assumed to Modeling suggests that a single oscillatory
be instantaneous: mechanism combines multiple currents that
support subthreshold oscillations in different
gNMDA
I NMDA ¼ ðENMDA vÞ, experimental conditions. Both calcium and
1 þ 0:1½Mgeme v subthreshold sodium currents contribute to
the depolarizing phase of the oscillation. Sim-
where [Mg] is the magnesium concentration. The ilarly, both calcium-dependent and voltage-
low slope of the voltage dependence dependent currents must contribute to the
(me = 0.062) is critical for the increase of the repolarization phase of the oscillation. The
frequency of spikes or subthreshold oscillations SK current is known to provide the calcium-
during NMDA application. dependent component. The ERG current is
An AMPA receptor current differs by its line- suggested by modeling studies as the
arity: I AMPA ¼ gAMPA ðEAMPA vÞ . Constant voltage-dependent component.
(tonic) AMPAR current does not elicit high-
frequency firing. Distinct Responses to AMPA and NMDA
Complete description of these models can be Receptor Stimulation
found in the literature cited above. The delicate oscillatory mechanism of the DA
neuron results in its distinct responses to different
Conclusions and Future Directions stimuli. In general, a passive (linear) or injected
Our understanding of the electrophysiological (constant) current can only impair oscillations.
properties of the DA neuron and the underlying To complement the oscillatory mechanism, the
mechanisms is not complete. However, modeling current must be active (nonlinear). This is the
studies have clarified some relations between its difference between AMPAR and NMDAR cur-
characteristics. rents: the NMDAR current contributes to the
regenerative depolarization phase of the oscilla- by different mechanisms in a model of a dopamine
tion. In various other types of neurons, the spike- neuron in vivo. J Neurophysiol 96:2549–2563
Canavier CC, Oprisan S, Callaway J, Ji H, Shepard PD
producing oscillatory mechanism is so powerful (2007) Computational model predicts a role for ERG
that the type of stimulation does not make much current in repolarizing plateau potentials in dopamine
difference. Taken together, because the oscilla- neurons: implications for modulation of neuronal
tions in the DA neuron are much less robust than activity. J Neurophysiol 98(5):3006–3022
Deister CA, Teagarden MA, Wilson CJ, Paladini CA
in fast-spiking neurons, the ability of an input to (2009) An intrinsic neuronal oscillator underlies dopa-
contribute to the oscillation becomes critical, and minergic neuron bursting. J Neurosci 50:15888–15897
the difference between NMDA and AMPA Drion G, Massotte L, Sepulchre R, Seutin V (2011) How D
receptor currents is revealed. modeling can reconcile apparently discrepant experi-
mental results: the case of pacemaking in dopaminer-
gic neurons. PLoS Comput Biol 7(5):e1002050
Major Future Directions for Modeling Firing Freeman AS, Meltzer LT, Bunney BS (1985) Firing prop-
Patterns of the DA Neuron erties of substantia nigra dopaminergic neurons in
• Combine mechanisms for all types of oscilla- freely moving rats. Life Sci 36(20):1983–1994
Grace AA, Bunney BS (1984) The control of firing pattern
tions in a single model to examine their inter- in nigral dopamine neurons: burst firing. J Neurosci
action and establish the connections among 4:2877–2890
different experiments. Harris NC, Webb C, Greenfield SA (1989) A possible
• Model burst evocation in realistic in vivo-like pacemaker mechanism in pars compacta neurons of
the guinea-pig substantia nigra revealed by various ion
conditions, which includes accurate calibra- channel blocking agents. Neuroscience 31:355–362
tion of the glutamate and GABA synaptic Hyland BI, Reynolds JNJ, Hay J, Perk CG, Miller
inputs and dynamic gating of receptors. R (2002) Firing modes of midbrain dopamine cells in
• Study the influence of dopamine autoreceptors the freely moving rat. J Neurosci 114(2):475–492
Ji H, Shepard PD (2006) SK Ca2 + -activated K + chan-
on firing of the DA neuron. nel ligands alter the firing pattern of dopamine-
• Investigate the functional consequences of the containing neurons in vivo. Neuroscience
distinctive morphology of the DA neuron, 140(2):623–633
such as the separation each spike into Johnson SW, Wu Y-N (2004) Multiple mechanisms
underlie burst firing in rat midbrain dopamine neurons
somato-dendritic (SD) and initial segment in vitro. Brain Res 1019:293–296
(IS) components due to the distant location Johnson SW, Seutin V, North RA (1992) Burst firing in
of the axon hillock from the soma. dopamine neurons induced by N-methyl-D-aspartate:
role of electrogenic sodium pump. Science
258:655–657
Komendantov AO, Canavier CC (2002) Electrical cou-
References pling between model midbrain dopamine neurons:
effects on firing pattern and synchrony.
Amini B, Clark JW, Canavier CC (1999) Calcium dynam- J Neurophysiol 87:1526–1541
ics underlying pacemaker-like burst firing oscillations Komendantov AO, Komendantova OG, Johnson SW,
in midbrain dopaminergic neurons: a computational Canavier CC (2004) A modeling study suggests com-
study. J Neurophysiol 82:2249–2261 plementary roles for GABAA and NMDA receptors
Bustos G, Abarca J, Campusano J, Bustos V, Noriega V, and the SK channel in regulating the firing pattern in
Aliaga E (2004) Functional interactions between midbrain dopamine neurons. J Neurophysiol
somatodendritic dopamine release, glutamate recep- 91:346–357
tors and brain-derived neurotrophic factor expression Kuznetsov AS, Kopell NJ, Wilson CJ (2006) Transient
in mesencephalic structures of the brain. Brain Res high-frequency firing in a coupled-oscillator model of
Rev 47(1–3):126–144 the mesencephalic dopaminergic neuron.
Canavier CC (1999) Sodium dynamics underlying burst J Neurophysiol 95:932–947
firing and putative mechanisms for the regulation of Li Y-X, Bertram R, Rinzel J (1996) Modeling N-methyl-
the firing pattern in midbrain dopamine neurons: D-aspartate-induced bursting in dopamine neurons.
a computational approach. J Comput Neurosci 6:49–69 Neuroscience 71:397–410
Canavier CC, Landry RS (2006) An increase in AMPA Medvedev GS, Kopell N (2001) Synchronization and
and a decrease in SK conductance increase burst firing transient dynamics in chains of electrically coupled
D 1042 Dorsal Column Stimulation
FitzHugh-Nagumo oscillations. SIAM J Appl Math Further Reading

61:1763–1801 Kotter R, Feizelmeier M (1998) Species-dependence and
Medvedev GS, Wilson CJ, Callaway JC, Kopell N (2003) relationship of morphological and electrophysiologi-
Dendritic synchrony and transient dynamics in cal properties in nigral compacta neurons. Prog
a coupled oscillator model of the dopaminergic neu- Neurobiol 54:619–632
ron. J Comput Neurosci 15:53–69
Meltzer LT, Christoffersen CL, Serpa KA (1997) Modu-
lation of dopamine neuronal activity by glutamate
receptor subtypes. Neurosci Biobehav Rev
21(4):511–518 Dorsal Column Stimulation
Morikawa H, Khodakhah K, Williams JT (2003) Two
intracellular pathways medicate metabotropic gluta-
mate receptor-induced Ca2+ mobilization in dopamine ▶ Spinal Stimulation for Parkinson Treatment
neurons. J Neurosci 23:149–157
Neuhoff H, Neu A, Liss B, Roeper J (2002) Ih channels
contribute to the different functional properties of
identified dopaminergic subpopulations in the mid-
brain. J Neurosci 22(4):1290–1302 Dorsal Root Stimulation
Overton PG, Clark D (1997) Burst firing in midbrain
dopaminergic neurons. Brain Res Rev 25:312–334 ▶ Finite Element Models of Transcutaneous Spi-
Ping HX, Shepard PD (1996) Apamin-sensitive Ca2 + nal Cord Stimulation
-activated K + channels regulate pacemaker activity
in nigral dopamine neurons. Neuroreport 7:809–814
Putzier I, Kullmann PH, Horn JP, Levitan ES
(2009) Cav1.3 channel voltage dependence, not Ca2+
selectivity, drives pacemaker activity and amplifies
bursts in nigral dopamine neurons. J Neurosci
49:15414–15419
Down Under Neural Population
Redgrave P, Gurney K (2006) The short-latency dopamine Models
signal: a role in discovering novel actions? Nat Rev
Neurosci 7(12):967–975 David T. J. Liley
Richards CD, Shiroyama T, Kitai ST (1997) Electrophys-
iological and immunocytochemical characterization of
Brain and Psychological Sciences Research
GABA and dopamine neurons in the substantia nigra Centre, Swinburne University of Technology,
of the rat. J Neurosci 80(2):545–557 Hawthorn, VIC, Australia
Schultz W (2002) Getting formal with dopamine and
reward. Neuron 36:241–263
Shepard PD, Bunney BS (1991) Repetitive firing
properties of putative dopamine-containing neurons Synonyms
in vitro: regulation by an apamin-sensitive Ca2 +
-activated K + conductance. Exp Brain Res Liley model; Robinson model; Steyn-Ross model
86:141–150
Tepper JM, Martin LP, Anderson DR (1995) GABAA
receptor-mediated inhibition of rat substantia nigra
dopaminergic neurons by pars reticulata projection Definition
neurons. J Neurosci 15:3092–3103
Waroux O, Massotte L, Alleva L, Graulich A, Thomas E,
Liégeois JF, Scuvée-Moreau J, Seutin V (2005) SK
Australia and New Zealand, despite their rela-
channels control the firing pattern of midbrain tively small academic communities, have been
dopaminergic neurons. Eur J Neurosci 22(12): vibrant sites for the development and application
3111–3121 of a number of important approaches to the
Wilson CJ, Callaway JC (2000) A coupled oscillator
model of the dopaminergic neuron of the substantia
mesoscopic modeling of neural populations,
nigra. J Neurophysiol 83:3084–3100 developed initially to account for the physiolog-
Wolfart J, Neuhoff H, Franz O, Roeper J (2001) Differen- ical genesis of rhythmic activity in the mamma-
tial expression of the small-conductance, calcium- lian electroencephalogram. These models are
activated potassium channel SK3 is critical for
pacemaker control in dopaminergic midbrain neurons.
collectively referred to as “down under” neural
J Neurosci 21(10):3443–3456 population models.
Down Under Neural Population Models 1043 D
Detailed Description the propagation of axonal pulse activity in con-
tinuously distributed planar cortical populations
History of excitatory and inhibitory neurons and was later
The work of Wright (Wright and Kydd 1984), extended to include thalamocortical feedback
which aimed theoretically to account for the reg- (Robinson et al. 2002). The model of Liley
ulatory effects of lateral hypothalamic activity on et al. (1999, 2002), which also theoretically
electrocortical activity, can be seen as the pro- incorporated spatially continuous axonal pulse
genitor of the later “down under” models that propagation, focused on the topology of
were developed to account for electrocortical intracortical and cortico-cortical connectivity D
rhythmogenesis. In this approach telencephalon and differentiating the kinetics of excitatory and
was functionally modeled as a mass of coupled inhibitory neurotransmission. Finally Steyn-Ross
linear oscillators. While largely phenomenologi- et al. (2007) developed an approach based on
cal and not formulated in terms of the structure Liley et al. (1999, 2002) that included the effects
and composition of cortical tissue, this theory of electrical gap junction synapses. All models
nevertheless gave mathematical form to the role represent examples of continuum neural field
that a specific anatomical pathway played in models (see entries “▶ Neural Field Model, Con-
modulating electroencephalographic activity. tinuum”; “▶ Neural Population Model”).
The ambitions of Wright’s approach significantly
influenced the subsequent development of Robinson et al.: Thalamocortical Feedback
a number of neural population models. The model of Robinson et al. (2002) originally
described cortical activity in terms of a spatially
Primary Models homogeneous field quantity f e (propagated
Three principal neural population models form excitatory neuronal pulse rate) evolving
the core of “down under” approaches. The popu- according to the following set of delay ordinary
lation model of Robinson et al. (1997) considered differential equations:

Lax fe ðtÞ ¼ S½V e ðtÞ ,
cortex (1)
Ldend V e ðtÞ ¼ aee fe ðtÞ þ aai S½V e ðtÞ þ aes S ½V s ðt t=2Þ,

Ldend V r ðtÞ ¼ are fe ðt t=2Þ þ ars S½V s ðtÞ ,
thalamus (2)
Ldend V s ðtÞ ¼ ase fe ðt t=2Þ þ asr S½V r ðtÞ þ asn fn ðtÞ,
where e, i, r, s index, respectively, cortical excit- l2 d2 l d

Lax ¼ þ 2 þ 1 ; Ldend
atory, cortical inhibitory, thalamic reticular, and u2 dt2 u dt

thalamic relay neuronal populations, alk nominally 1 d2 1 1 d
¼ þ þ þ 1, (3)
represent population synaptic weights, Vl repre- ab dt2 a b dt
sents a mean population membrane potential, S a
sigmoidal firing rate function, t a fixed delay, and where l is the mean range of excitatory axons of
fn an external pulse input. The fixed delay t velocity u. This model was subsequently gener-
represents the return corticothalamic conduction alized to the spatially inhomogeneous case
time. Lax, Ldend define second-order linear ordi- (Robinson et al. 2004). This model has been
nary differential operators that account for passive used to investigate the genesis of the resting
neuronal population cable delays and pulse prop- alpha rhythm, the dynamical instability of epi-
agation via cortico-cortical fibers, respectively: lepsy, and the oscillatory architecture of sleep.
D 1044 Down Under Neural Population Models
Liley et al.: Kinetics of Excitatory and Steyn-Ross et al.: Electrical Gap Junctions
Inhibitory Neurotransmission In a novel departure from the usual axo-synaptic
The neuronal population formulation of Liley connectivity of most neural population models,
et al. (2002) is an example of a conductance- Steyn-Ross et al. (2007), on the basis of substan-
based model. The defining equations distinguish tial empirical evidence, added the bulk effects of
between the kinetics of excitatory and inhibitory gap junction currents. Gap junctions are
neurotransmission and the topology and proper- intercellular protein channels that can form spon-
ties of intracortical and cortico-cortical connec- taneously at the point of contact between the den-
tivity. It describes cortical activity in terms of the dritic trees of two adjacent neurons and enable the
mean soma membrane potential hk(r,t) of spa- free diffusion of ions (current) along intercellular
tially continuously distributed cortical electrochemical gradients. By assuming homoge-
populations of excitatory (k = e) and inhibitory neous cellular ionic environments, such free diffu-
(k = i) neurons: sion will be determined by intercellular potential
differences. On this basis Steyn-Ross et al. (2007)
tk @t hk ðr, tÞ ¼ hrest
k hk ðr, tÞ
X defined the mean soma membrane potential hk
þ cl ½hk ðr, tÞ I lk ðr, tÞ, (4) (r, t) of a spatially continuously distributed cortical
l neuronal population of type k in terms of
a modified conductance-based neuron:
ð@t þ glk Þ2 I lk ðr, tÞ ¼ exp ð1Þ glk Glk
n o tk @t hk ðr, tÞ ¼ hrest
k hk ðr, tÞ
N blk Sl ½hl ðr, tÞ þ flk þ plk ðtÞ , X
þ cl ½hk ðr, tÞ I lk ðr, tÞ (8)
(5) l
þ Dkk ∇2 hk ðr, tÞ
3
ð@t þ uek Lek Þ2 fek ðr, tÞ þ u2ek ∇2 fek ðr, tÞ
2 where Dkk is a diffusive coupling strength
¼ N aek u2ek L2ek Se ½he ðr, tÞ, (6) between adjacent neurons of type k. The
remaining equations defining the model are sim-
fik ðr, tÞ ¼ 0, (7) ilar to that of Liley (see above). In general,
because the density of gap junction coupling
where r
(x,y), tk is a membrane time constant between inhibitory neurons is much greater than
and cl [hk (r,t)] incorporates the effect of the excitatory neurons, Dii Dee. All other symbols
reversal potential on the respective postsynaptic are defined as before. This model has been
potential of type l. @ t and ∇2 are the temporal invoked to account for the low-frequency oscil-
derivative and 2D Laplacian operator, respec- lations seen in resting-state functional MRI.
tively. The amplitude and characteristic time-
scale of the unitary postsynaptic potential of
type l are given by Glk and glk, respectively. The Cross-References
topology of neuronal population connectivity is
defined in terms of the intracortical (Nlk b
) and ▶ Anesthesia, Neural Population Models of
a
cortico-cortical (Nlk) connectivity between ▶ Gap Junctions in Small Networks
a presynaptic population l and a postsynaptic ▶ Mesoscopic Anatomy and Neural Population
population k. Cortico-cortical connectivity is Models
assumed to exclusively propagate pulses via ▶ Neural Field Model, Continuum
axons of excitatory neurons having a conduction ▶ Neural Mass Action
velocity ulk and a characteristic length 1/Llk . This ▶ Neural Population Model
model has been utilized to account for the emer- ▶ Sleep, Neural Population Models of
gence of the resting EEG spectrum and its pertur- ▶ Synaptic Connectivity in Neural Population
bation during anesthesia. Models
Dynamic Causal Modelling with Neural Population Models 1045 D
References
Dynamic Causal Modelling with
Liley DTJ, Cadusch PJ, Wright JJ (1999) A continuum Neural Population Models
theory of electro-cortical activity. Neurocomp
26–27:795–800
Liley DTJ, Cadusch PJ, Dafilis MP (2002) A spatially Rosalyn Moran
continuous mean field theory of electrocortical activ- Virginia Tech Carilion Research Institute,
ity. Network 13:67–113 Roanoke, VA, USA
Robinson PA, Rennie CJ, Wright JJ (1997) Propagation
Bradley Department of Electrical and Computer
and stability of waves of electrical activity in the D
cerebral cortex. Phys Rev E 56:826–840 Engineering, Virginia Tech, Blacksburg,
Robinson PA, Rennie CJ, Rowe DL (2002) Dynamics of VA, USA
large-scale brain activity in normal arousal states and Department of Psychiatry and Behavioral
epileptic seizures. Phys Rev E 65:041,924
Medicine, Virginia Tech Carilion School of
Robinson PA, Rennie CJ, Rowe DL, O’Connor SC
(2004) Estimation of multiscale neurophysiologic Medicine, Roanoke, VA, USA
parameters by electroencephalographic means. Hum
Brain Mapp 23(1):53–72
Steyn-Ross ML, Steyn-Ross DA, Wilson MT, Sleigh JW
(2007) Gap junctions mediate large-scale Turing struc-
Definition
tures in a mean-field cortex driven by subcortical
noise. Phys Rev E Stat Nonlin Soft Matter Phys Dynamic causal models with neural populations
76(1 Pt 1):011,916 are state-space models that impose a specified
Wright JJ, Kydd RR (1984) A linear theory for global
dynamical systems form to the generation of neu-
electrocortical activity and its control by the lateral
hypothalamus. Biol Cybern 50(2):75–82 roimaging and electrophysiological data. The
forms of these models constitute neurobio-
logically motivated and identifiable parameteri-
zations, where empirical observations offer
Drift-Diffusion Equation conditional descriptions of parameter space fol-
lowing application of a Bayesian inversion
▶ Fokker-Planck Equation scheme. They comprise separate generative pro-
▶ Nernst-Planck Equation cesses at the neuronal level and at the observation
level through a set of deterministic or stochastic
differential equations and an observer functional,
respectively.
Drug Effect
▶ Neuropharmacological Modeling, Pharmaco- Detailed Description

genomics and Ion Channel Modulation
Dynamic causal models were first invented in
2003 for the purpose of estimating human brain
connectivity and task-dependent functional inte-
Dual-Generator Concept gration using functional magnetic resonance imag-
ing (fMRI) (Friston et al. 2003). The framework
▶ Coupled Oscillations in Neural Control of was extended for electrophysiological domains
Breathing including noninvasive electroencephalography
(EEG) and magnetoencephalography (MEG) and
later for invasive local field potential recordings.
Application domains share a common generative
Dynamic Attending and inversion framework (Stephan et al. 2010). At
the neuronal level, a network of “nodes” are
▶ Rhythm Perception: Pulse and Meter connected with intrinsic (within-region) and
D 1046 Dynamic Causal Modelling with Neural Population Models
extrinsic (region-to-region) connections. These depolarization changes at different inhibitory

effective or model-based connections constrain and excitatory cell ensembles), owing to their
the output of a set of differential equations and higher temporal resolution. For example, DCM
can be modulated by experimental perturbation. for EEG employs (among other models) the Jan-
These are modality-dependent, mesoscale descrip- sen and Rit formulation (Jansen and Rit 1995) to
tions of neuronal ensemble activity and are sum- describe layer-specific synaptic responses, where
marized through mean-field reductions. Different inputs, u, are convolved with postsynaptic ker-
connectivity architectures can be considered as nels, h, which can depolarize (e) or hyperpolarize
different models, embodying competing hypothe- (i) the postsynaptic neuronal ensemble to produce
ses to be formally tested. At the observation level, an average membrane potential, v:
contributing neuronal states form inputs to
N
a second dynamic process or static function (lead u¼he=i u
(2)
field), transforming neuronal activity to the he=i ¼ He=i ke=i t exp tke=i
required measurement space, e.g., blood oxygen
level dependent – BOLD fMRI responses. The parameter H tunes the maximum amplitude
Together, these form a complete forward model of PSPs and k is a lumped representation of the
that is inverted, given real data, using Bayesian sum of the rate constants of passive membrane
approaches. and other spatially distributed delays in the den-
dritic tree. The convolution leads to a second-
Modality-Dependent Population Models order differential equation (David and Friston
The neural population model describes the time 2003) that can be decomposed into normal form
evolution of unobservable neuronal states. For as
fMRI, x is an n 1 column vector representing
neuronal activity at n brain regions. This activity u_ ðtÞ ¼ iðtÞ
is under the control of inhibitory rate parameters, (3)
i_ðtÞ ¼ H e=i ke=i uðtÞ 2ke=i xðtÞ k2e=i uðtÞ
aij 8 j = i, describing the influence of regional
activation on its own output activity and the
A sigmoidal operator, S, then transforms the aver-
excitatory or inhibitory influence, aij 8 j 6¼ i,
age membrane potential into a population firing
from other brain regions. This n n matrix A is
rate and allows us to substitute direct (thalamic)
referred to as the endogenous or intrinsic connec-
input, u, with afferent input from other cell
tivity matrix. A bilinear approximation B( j) is an
populations. These inputs are parameterized
n n matrix used to capture effects from m,
with a connectivity parameter g and can arise
experimental input modulations, and C, an n x k
from within the local layered neural mass or
matrix, represents the direct perturbation from
from pyramidal cells of other connected brain
k experimental inputs, giving
regions (both having identical mathematical
! form), giving, for example,
X
m
ðjÞ
x_ ¼ Aþ uj B x þ Cu (1)
j¼1 u_ ¼ i

(4)
i_ ¼ ke H e gS uaff þ u 2ke i k2 u e
Here, each state is a one-dimensional neural mass
model representing the mean neuronal activity Modality-Dependent Observers
within a region encompassing approximately Feature extraction from these neural population
10–100 voxels. Such a mean-field reduction is models is specified through an appropriate trans-
also applied to electrophysiological modalities form from the neural state space to measurement
where multiple-state models are used to represent space. In the case of fMRI, this transform can be
several dimensions of unobservable dynamics parameterized by a nonlinear “balloon model”
within a single brain region (e.g., membrane (Buxton et al. 1998) describing the neuronal
Dynamic Causal Modelling with Neural Population Models 1047 D
activity-induced dynamic alterations in vasodila- Applications of DCM
tion, blood flow, and subsequent volume and The DCM approach has been applied to many
deoxyhemoglobin content – to predict the experimental preparations and imaging modali-
observed BOLD time series. For electrophysiolog- ties (a PubMed search reveals 400 articles asso-
ical modalities, a static lead field transforms ciated with dynamic causal modeling). As well as
a mixture of cell membrane potentials from each offering insights into the fundamental role of
source to scalp-level-evoked, scalp-level-induced, effective brain connections, e.g., how callosal
or correlational data features (Kiebel et al. 2006): connections support interhemispheric integration
in a task-dependent manner (Stephan et al. 2005) D
h ¼ L:Kx (5) or how feedback loops are necessary to support
prototypical event-related potentials (Garrido
where K describes the contribution from different et al. 2007), specific network topologies have
cell ensemble potentials x, which is used to weigh, been elucidated for cognitive and motor phenom-
primarily, pyramidal cell contributions, and L is ena in health and disease. DCM for fMRI was
a lead-field matrix that accounts for passive con- used in one study to ascertain which regions in
duction of their electromagnetic field and can be a prefrontal-subcortical network drive the
constrained along three orthogonal dimensions, so-called “status quo” motor-bias behavior
when dipole orientations are not known. (Fleming et al. 2010), finding a prefrontal-
Together with the neuronal state space, these causality the most likely. During learning, similar
transforms produce a full generative model with networks have been shown to adapt along bottom-
which mechanistic hypotheses regarding the up axes, with changes in connections that support
physiological processes underlying empirical motor responsivity driven by modulation from
data observations can be performed. subcortical structures (den Ouden et al. 2010). In
disease, support for disconnection-related mecha-
Bayesian Inversion, Parameter Estimation, nisms has been observed in patients with schizo-
and Model Selection phrenia (Dima et al. 2009) but also in neurological
The models described above produce the data disorders more traditionally associated with focal
likelihood and when combined with a set of priors lesions, such as Parkinson’s disease (Marreiros
on parameter space allow for a full Bayesian et al. 2012). As a “mathematical microscope,”
inversion. The inversions employed in DCM electrophysiological DCMs have been applied to
mostly use a variational expectation maximiza- small, one-node networks to uncover particular
tion routine (Šmı́dl and Quinn 2006) (stochastic synapses and receptors subtending memory per-
routines employ generalized filtering). Here, formance (Moran et al. 2011). Overall, the meth-
a partition between the multivariate Gaussian odology provides a principled Bayesian
model parameter space and the observation framework to address the mechanisms of func-
noise parameter (known as the hyperparameter) tional brain architectures that support our mental
provides a scheme over which gradient ascent on repertoire.
a free energy objective function is amenable:
F ¼ ln: 25empðyjmÞ KLðqðyÞjjpðyjy, mÞÞ (6) Cross-References
This quantity F thence returns an approximation to ▶ Determining Network Structure from Data:
the model evidence, which is exact for linear Nonlinear Modeling Methods
models as well as an approximate conditional ▶ Forward and Inverse Problems of MEG/EEG
parameter density q. These two quantities can ▶ Imaging Analysis, Bayesian
then be used to assess competing model architec- ▶ Jansen-Rit Model
tures and the particular conditional (data: y, depen- ▶ Neural Mass Action
dent) distributions on parameters, respectively. ▶ Neural Population Model
D 1048 Dynamic Clamp Technique
▶ Neuroimaging, Neural Population Models for

▶ Pattern Formation in Neural Population Dynamic Clamp Technique
Models
▶ Population Density Model Thomas Nowotny1 and Pablo Varona2
▶ Stochastic Neural Field Theory 1
Centre for Computational Neuroscience and
▶ Synaptic Connectivity in Neural Population Robotics, School of Engineering and Informatics,
Models University of Sussex, Brighton, UK
2
Departamento de Ingenieria Informatica,
Universidad Autónoma de Madrid,
Madrid, Spain
References
Buxton R, Wong E, Frank L (1998) Dynamics of blood Definition

flow and oxygenation changes during brain activation:
the balloon model. Magn Reson Med 39:855–864
David O, Friston KJ (2003) A neural mass model for a Virtual Object Screen
MEG/EEG: coupling and neuronal dynamics.
Neuroimage 20:1743–1755 Animal
den Ouden HE, Daunizeau J, Roiser J, Friston KJ, Stephan lscreen Eye
l
KE (2010) Striatal prediction error modulates cortical q
coupling. J Neurosci 30:3210–3219 v
Dima D, Roiser JP, Dietrich DE, Bonnemann C,
Lanfermann H, Emrich HM, Dillo W (2009) Under-
standing why patients with schizophrenia do not per- xeye-screen
ceive the hollow-mask illusion using dynamic causal x(t)
modelling. Neuroimage 46:1180–1186
Fleming SM, Thomas CL, Dolan RJ (2010) Overcoming
status quo bias in the human brain. Proc Natl Acad Sci Initial Distance
107:6005–6009
Friston K, Harrison L, Penny W (2003) Dynamic causal b 60 120
modelling. Neuroimage 19:1273–1302 l/v=479ms

Garrido M, Kilner J, Kiebel S, Friston K (2007) Evoked 50 θo=3.9° 100
brain responses are generated by feedback loops. Proc
Edge Velocity ψ (°/s)

Angular Size θ (°)
Natl Acad Sci 104:20961 40 Angular Size 80

Jansen B, Rit V (1995) Electroencephalogram and visual
Edge Velocity
evoked potential generation in a mathematical 30 60
model of coupled cortical columns. Biol Cybern
73:357–366
20 40
Kiebel S, David O, Friston K (2006) Dynamic causal
modelling of evoked responses in EEG/MEG with
lead field parameterization. Neuroimage 30:1273–1284 10 20
Marreiros AC, Cagnan H, Moran RJ, Friston KJ, Brown
P (2012) Basal ganglia-cortical interactions in Parkin- 0 0
sonian patients. Neuroimage 66:301–310 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 0
Moran RJ, Symmonds M, Stephan KE, Friston KJ, Dolan Time to collision (sec)
RJ (2011) An in vivo assay of synaptic function medi-
ating human cognition. Curr Biol 21:1320–1325
Šmı́dl V, Quinn AP (2006) The variational Bayes method Dynamic clamp is an electrophysiological tech-
in signal processing. Springer, New York nique for introducing simulated electrical com-
Stephan KE, Penny WD, Marshall JC, Fink GR, Friston
KJ (2005) Investigating the functional role of callosal ponents into biological cells using a real-time
connections with dynamic causal models. Ann N Y closed loop between the cell and a computer or
Acad Sci 1064:16–36 another electronic device. Classic dynamic clamp
Stephan K, Penny W, Moran R, den Ouden H, protocols build a voltage-dependent current
Daunizeau J, Friston K (2010) Ten simple rules
for dynamic causal modeling. Neuroimage injection cycle to implement artificial membrane
49:3099–3109 or synaptic conductances in the cell membrane of
Dynamic Clamp Technique 1049 D
biological neurons. These protocols are Figure 1 illustrates a typical dynamic clamp
employed to assess a large variety of neuronal configuration.
computational properties and are widely applied
for studying the physiology of neural systems at Applications and Implementations
the cellular and circuit levels. The dynamic clamp technology for in vitro and
in vivo electrophysiology has produced many
examples of successful closed-loop interactions
Detailed Description with neural systems, which include adding or
cancelling ionic channels, adding or cancelling D
History electrical or chemical synapses (including learn-
The use of closed-loop feedback interactions with ing and plasticity protocols), and implementing
living neurons for observation and control pur- hybrid circuits of real and artificial neurons and
poses goes back to the beginnings of electrophys- networks (for reviews see Prinz et al. (2004),
iology when the voltage clamp technique was Goaillard and Marder (2006), Destexhe and Bal
developed (Marmont 1949; Cole 1955). The volt- (2009), and Economo et al. (2010)).
age clamp technique measures currents across the Different dynamic clamp software
membrane of excitable cells while holding the implementations have increasingly become
membrane potential at a constant level. In current available under both Windows (Pinto
clamp, on the other hand, the current passing into et al. 2001; Kullmann et al. 2004; Nowotny
and out of the neuron is tightly controlled, while et al. 2006; Kemenes et al. 2011) and real-time
the membrane potential is being recorded. Both Linux operating systems (Butera et al. 2001;
voltage and current clamp protocols are routinely Dorval et al. 2001; Muniz et al. 2009; Lin
used to study the electrical properties of neuronal et al. 2010) and contribute to expand the use of
membranes and contribute to the design of dynamic clamp protocols by implementing
conductance-based neuron models. In the early a realistic cycle update in soft or hard real-time
1990s, the first hybrid circuit experiments cou- approaches. Novel active electrode compensation
pling living and artificial cells through synthetic techniques also allow experimenters to conduct
electrical connections were carried out in heart dynamic clamp experiments with a single high-
cells (Tan and Joyner 1990) and neurons (Yarom resistance electrode (Brette et al. 2008; Samu
1991). et al. 2012).
In 1992, Sharp et al. and Robinson and Kawai
independently developed the dynamic clamp Modern Developments
concept (Sharp et al. 1993; Robinson and Kawai The dynamic clamp technique has been general-
1993). The dynamic clamp technique operates in ized in the context of activity-dependent stimula-
a real-time closed-loop manner: an intracellular tion protocols beyond electrophysiological input/
electrode measures the membrane potential of output by using chemical and mechanical stimu-
a neuron, which is fed into a computer or an lation and optical recording methods (Chamorro
electronic device that calculates a current to et al. 2012). Dynamic clamp-inspired closed
inject into a postsynaptic neuron. The injected loops can also be used in novel assisted
current is computed after solving a model that neurofeedback techniques by implementing
represents ionic or synaptic conductances or sim- goal-driven stimuli that not only depend on
ulates the dynamics of individual neurons or neu- instantaneous measurements but also on an
ral networks. The postsynaptic cell can be the adequately long history of the recordings
identical to or different from the one from (Wallach et al. 2011; Chamorro et al. 2012;
which the membrane potential is being recorded. Fernandez-Vargas et al. 2013). Dynamic clamp
This activity-dependent closed loop is repeated and novel activity-dependent stimulation tech-
indefinitely and has to be implemented at a high niques reveal neural dynamics otherwise hidden
enough frequency to achieve a realistic effect. under traditional stimulus–response protocols,
D 1050 Dynamic Clamp Technique
Dynamic Clamp Technique, Fig. 1 Typical dynamic simulated and issues a current command that is converted
clamp configuration. A cell is impaled by two electrodes, to an analog command and injected as a current through
one for monitoring the membrane potential and one for the current injection electrode. This cycle is repeated at
injecting currents. The signal from the recording electrode a high frequency on the order of kHz. A second recording
is amplified, converted to a digital signal, and fed into the computer is often used to obtain an unbiased record of the
dynamic clamp software. The software calculates the experiment
appropriate current according to the model that is
provide control of regular and pathological states, Destexhe A, Bal T (eds) (2009) Dynamic-clamp: from
induce learning processes, bridge between dis- principles to applications. Springer, New York
Dorval AD, Christini DJ, White JA (2001) Real-time linux
tinct levels of analysis, and lead to a further auto- dynamic clamp: a fast and flexible way to construct
mation of experiments in neuroscience research. virtual ion channels in living cells. Ann Biomed Eng
29:897–907
Economo MN, Fernandez FR, White JA (2010) Dynamic
clamp: alteration of response properties and creation of
References virtual realities in neurophysiology. J Neurosci
30:2407–2413
Brette R, Piwkowska Z, Monier C, Rudolph-Lilith M, Fernandez-Vargas J, Pfaff HU, Rodriguez FB, Varona
Fournier J, Levy M, Frégnac Y, Bal T, Destexhe P (2013) Assisted closed-loop optimization of
A (2008) High-resolution intracellular recordings SSVEP-BCI efficiency. Front Neural Circuits 7:27
using a real-time computational model of the elec- Goaillard J-M, Marder E (2006) Dynamic clamp analyses
trode. Neuron 59:379–391 of cardiac, endocrine, and neural function. Physiology
Butera RJ, Wilson CG, Delnegro CA, Smith JC (Bethesda) 21:197–207
(2001) A methodology for achieving high-speed rates Kemenes I, Marra V, Crossley M, Samu D, Staras K,
for artificial conductance injection in electrically Kemenes G, Nowotny T (2011) Dynamic clamp with
excitable biological cells. IEEE Trans Biomed Eng StdpC software. Nat Protoc 6:405–417
48:1460–1470 Kullmann PHM, Wheeler DW, Beacom J, Horn JP
Chamorro P, Muñiz C, Levi R, Arroyo D, Rodrı́guez FB, (2004) Implementation of a fast 16-Bit dynamic
Varona P (2012) Generalization of the dynamic clamp clamp using LabVIEW-RT. J Neurophysiol
concept in neurophysiology and behavior. PLoS ONE 91:542–554
7:e40887 Lin RJ, Bettencourt J, Ite JW, Christini DJ, Butera RJ
Cole KS (1955) Ions, potentials and the nerve impulse. In: (2010) Real-time experiment interface for biological
Shedlovsky T (ed) Electrochemistry in biology and control applications. Conf Proc IEEE Eng Med Biol
medicine. Wiley, New York, pp 121–140 Soc 2010:4160–4163
Dynamic Diseases of the Brain 1051 D
Marmont G (1949) Studies on the axon membrane; a new control mechanisms. Whereas traditional
method. J Cell Physiol 34:351–382 approaches for classifying neurological diseases
Muniz C, Rodriguez FB, Varona P (2009) RTBiomanager:
a software platform to expand the applications of real- are based on (static) anatomical, cellular, and
time technology in neuroscience. BMC Neurosci 10: molecular abnormalities, the focus here is on
P49 dynamics, namely, the variation of signs and
Nowotny T, Szucs A, Pinto RD, Selverston AI symptoms of disease as a function of time. The
(2006) StdpC: a modern dynamic clamp. J Neurosci
Methods 158:287–299 hallmarks of dynamic diseases are sudden, qual-
Pinto RD, Elson RC, Sz€ ucs A, Rabinovich MI, Selverston itative changes in the temporal pattern of clinical
AI, Abarbanel HD (2001) Extended dynamic clamp: signs. Identifying a neurological disorder as D
controlling up to four neurons using a single desktop a dynamic disease has two major implications:
computer and interface. J Neurosci Methods 108:39–48
Prinz AA, Abbott LF, Marder E (2004) The dynamic (1) the observed dynamics and their responses to
clamp comes of age. Trends Neurosci 27:218 various manipulations provide important insights
Robinson HP, Kawai N (1993) Injection of digitally syn- into the nature and abnormality of neural control,
thesized synaptic conductance transients to measure and (2) based on computational models of the
the integrative properties of neurons. J Neurosci
Methods 49:157 abnormalities, it may be possible to devise
Samu D, Marra V, Kemenes I, Crossley M, Kemenes G, novel treatment strategies for dynamic diseases
Staras K, Nowotny T (2012) Single electrode dynamic of the brain.
clamp with StdpC. J Neurosci Meth 211:11–21
Sharp AA, O’Neil MB, Abbott LF, Marder E (1993)
Dynamic clamp: computer-generated conductances in
real neurons. J Neurophysiol 69:992–995 Detailed Description
Tan RC, Joyner RW (1990) Electrotonic influences on
action potentials from isolated ventricular cells. Circ Historical Perspectives
Res 67:1071–1081
Wallach A, Eytan D, Gal A, Zrenner C, Marom S (2011) The concept of a dynamic disease arises out of the
Neuronal response clamp. Front Neuroengineer 4:3 concept of a dynamic disease introduced by
Yarom Y (1991) Rhythmogenesis in a hybrid system- Mackey and Glass (1977; Glass and Mackey
interconnecting an olivary neuron to an analog net- 1979) and expanded by their co-workers
work of coupled oscillators. Neuroscience 44:263–275
(Mackey and an der Heiden 1982; Glass and
Mackey 1988; Bélair et al. 1995). A dynamic
disease was originally defined as a disease that
Dynamic Diseases of the Brain occurs in a normal feedback control mechanism
operating in a range of parameters that produced
Gerold Baier1 and John Milton2 qualitatively altered, and presumably unhealthy,
1
Cell and Developmental Biology, Faculty of dynamic behaviors. The concept was illustrated
Life Sciences, University College London, by applications to the respiratory and hematopoi-
London, UK etic systems where the onset of disease was
2
W.M. Keck Science Center, The Claremont marked by a change involving an oscillatory
Colleges, Claremont, CA, USA physiological pattern: in Cheyne-Stokes respira-
tion, a regular breathing transforms to
a periodically modulated breathing pattern, and
Synonyms in periodic chronic myelogenous leukemia,
a normally constant level of leukocytes switched
Dynamical diseases; Periodic diseases first to a periodically and then to a chaotically
varying level. In each case, the qualitative
changes related to the onset of disease were asso-
Definition ciated with an analogous transition in the dynam-
ics of a mathematical model of an underlying
A dynamic disease of the nervous system is feedback control mechanism. Since time delays
a disease that arises from abnormalities in neural are essential components of feedback controllers,
D 1052 Dynamic Diseases of the Brain
the models took the form of delay differential Parkinsonism), mood (e.g., affective and psy-
equations (see also “▶ Time-Delayed Neural chotic disorders), and consciousness (e.g., epi-
Networks: Stability and Oscillations”). To induce lepsy, migraines, sleep-wake cycles). The large
the transition in the models, only quantitative number of articles in this encyclopedia that
changes of parameters were required. The address the computational aspects of disease
changes in dynamics arose when the parameter attests to the importance of adopting a dynamic
values crossed so-called stability boundaries, perspective in order to understand and treat
corresponding mathematically to bifurcations. It dynamic diseases of the nervous system. Here,
was hypothesized that pathological oscillations in we focus on a number of challenges and unsettled
the selected diseases arose because normal phys- issues that concern the application of computa-
iological control mechanisms were operating in tional modeling to clinical practice. Our discus-
an abnormal parameter range. A consequence of sion of dynamic disease states is along a similar
the computational studies is that the understand- vein that Izhikevich has made for the mathemat-
ing of the mechanism can lead to the prediction of ical classification of neuronal bursting patterns
novel treatments: in the case of a rare hematolog- (Izhikevich 2007). Namely, an understanding of
ical disease, cyclic neutropenia, the oscillations the basic mathematical mechanisms is mandatory
in neutrophils can be stopped by the administra- to properly analyze the computational studies and
tion of granulocyte colony-stimulating factor to use these studies for the interpretation of clin-
(G-CSF) (Foley and Mackey 2009), an interven- ical recordings of human brain activity. We there-
tion resulting in a bifurcation (specifically, fore introduce a basic mathematical classification
a reverse Hopf bifurcation). of transitions between dynamic states and give
However, changes in dynamics can have var- examples where these concepts have been
ious origins and are not necessarily associated applied to disease states of the human brain.
with bifurcations. This is especially true in the Finally, we discuss the unresolved computational
nervous system in which sudden changes in neu- challenges for the modeling of dynamic diseases
ral dynamics can arise because of the responses of of the nervous system and the development of
multistable neural networks to perturbations, the novel, dynamically inspired treatment strategies.
transient synchronization of neural population
activity, the formation of traveling waves of elec- Dynamic Transitions
trical potential, and the response to random fluc- We first restrict the discussion to qualitative
tuations in the vicinity of critical phenomena changes in the temporal dynamics of some phys-
such as phase transitions. The more inclusive iological variable. Taking the simplest case
term dynamic disease includes all disorders char- where there is the sudden appearance or disap-
acterized by abnormal dynamic behaviors. Thus, pearance of an oscillatory state, there are four
for example, epileptic seizures are the manifesta- major mathematical concepts that can explain
tions of a dynamic disease (Milton and Black this observation (Baier et al. 2012; Fig. 1):
1995; Milton and Jung 2003) where features (1) bifurcation, (2) bistability, (3) excitability,
other than bifurcations have been proposed as and (4) spontaneous bursting behavior
underlying mechanisms. (intermittency). These are illustrated in Fig. 1.
In retrospect, it is surprising that diseases of Bifurcation (Fig. 1a): A quantitative parame-
the nervous system received only scant mention ter change results to transition from one dynamic
in early discussions of the dynamics of diseases state to a qualitatively new state. The new state
(Glass and Mackey 1979, 1988; Mackey and an persists as long as the parameter value remains
der Heiden 1984; Mackey and Milton 1987). altered. Both states are robust to noise
Abnormal dynamic behaviors are frequently perturbations.
associated with diseases of the nervous system Bistability (Fig. 1b): Two qualitatively differ-
(Milton and Black 1995) and characterize, for ent states coexist under the same conditions
example, disorders of movement (e.g., tremors, (same parameter values). Transitions into and
Dynamic Diseases of the Brain, Fig. 1 State transi- required only to induce the first transition. The return to
tions in dynamic systems as a function of control param- the original state is spontaneous. (d) In a system which
eter. (a) In a monostable system, a transition is induced exhibits spontaneous bursting behavior (deterministic
when a parameter crosses a bifurcation point. (b) In intermittency), the transitions do not require change of
a bistable system, a suprathreshold stimulus (e.g., a short parameter or external input to occur (Figure originally
change of parameter) is required for each state transition. published in Baier et al. (2012). doi:10.3389/
(c) In an excitable system, a suprathreshold stimulus is fphys.2012.00281, modified)
out of the abnormal state are realized via pertur- Spontaneous bursting behavior (Fig. 1d):
bations that are strong enough to cross the bound- This globally stable state consists of two qualita-
ary separating the two states (the separatrix). tively different substates that alternate spontane-
Both states are vulnerable to noise perturbations. ously and irregularly and last for irregular periods
Excitability (Fig. 1c): A stable state is of time. Consequently, both normal and abnormal
interrupted temporarily by a qualitatively differ- dynamics are part of the overall state. Mathemat-
ent abnormal state due to a suprathreshold per- ically, this is equivalent to the type of dynamics
turbation. Both states are vulnerable to noise called intermittency. However, whereas most of
perturbations, but the return to the original state the mathematical literature on intermittency
is spontaneous (i.e., does not require perturba- deals with low-dimensional models, actual brain
tion). Therefore, the abnormal state is only dynamics will more generally display spatiotem-
a transient. Note that in principle the transient poral intermittency which at present is only
can be arbitrarily complex. incompletely understood.
This is the basic classification assuming that circuits in which identified parameters have
there is only one well-defined normal state. How- been changed systematically (Wilson 1999;
ever, an organism can also transit from one state Gerstner and Kistler 2002; Izhikevich 2007).
to another normal state, both characterized by The changes of neural activity here correspond
different dynamics (e.g., during development or to bifurcations in the appropriate mathematical
when transiting from sleep to the awake state). models. In the context of brain disease, it has
Pathological dynamics can then also occur been difficult to demonstrate this phenomenon
(transiently) during the transition between two convincingly. A major problem is that the rele-
normal states. The classification of such transi- vant control parameter(s) cannot be directly iden-
tions is presently incomplete, but we will discuss tified from the observation of a dynamic
an example below. transition. Indirect evidence can be used in com-
Finally, the basic classification looks exclu- putational studies to argue for certain parameter
sively at temporal changes, whereas brain changes and corresponding bifurcation patterns
dynamics involves spatiotemporal changes of with some plausibility. The indirect evidence
(neuro-)physiological variables. The theory of comes in various forms, for instance, by studying
spatiotemporal systems again allows for more the phenomenology of waveforms (Ebersole and
complex pathodynamics but is presently incom- Milton 2003), by doing systematic parameter
plete and will be illustrated only via examples scan and sensitivity analysis, and by comparison
below. with experimental results from animal models of
a specific disease.
The evolution of complex partial epileptic sei-
Pathodynamics in the Human Brain zures has been considered to be an example
where different states arise through transitions
In order to illustrate dynamic disease approaches via bifurcations (Lopes Da Silva et al. 2003).
to neurological disease, we focus our discussion An analysis of clinical data from subdural and
on epileptic seizures. Topologically, seizures are depth electrode recordings identifies a number of
characterized by their mode of onset. The term abnormal states characterized by distinct rhyth-
partial is used to indicate a seizure which begins mic properties: frequency, amplitude, and wave-
in a focal area of the brain. Partial seizures are form. One model of focal seizures proposed by
termed simple if there is no alteration in con- Wendling et al. is based on the mean behavior of
sciousness and complex if there is an alteration large neural populations of principal and inhibi-
in consciousness. A generalized seizure may be tory neurons in the cortex (Wendling et al. 2002).
generalized from the onset (primary generalized A study of the generic bifurcation structure of this
epilepsy) or represent the result of spreading from model identifies potential transition points where
an epileptic focus (secondarily generalized epi- the dynamics change from background to
lepsy). Dynamically, seizures are characterized pre-seizure, seizure, and post-seizure activity in
by the sudden onset of an uncontrolled discharge accordance with the clinical data. Interestingly,
pattern in very large populations of neurons, and the exploration of the model leads the authors to
the diagnosis of epilepsy depends (in addition to identify dynamics corresponding to a significant
the clinical history and description of the sei- proportion of the highly heterogeneous rhythms
zures) on the qualitative electroencephalographic during partial complex seizures with hippocam-
(EEG) changes associated with seizures. Epi- pal origin. The computational study thus supports
lepsy, with its recurrent seizures, can therefore the hypothesis that patients with partial seizures
serve as the epitome of a dynamic disease. share the basic neural circuitry but vary in the
settings of their individual parameters. The
Bifurcations model also explains how abnormal epileptic
It is well established at the benchtop that dynamic activity results from the unbalanced cooperation
transitions occur in neurons and simple neural between excitatory and inhibitory neural
populations. The major insight from the compu- seizures is the distance of the separatrix from
tational study is that focal-onset seizures tend to the normal behavior in phase space. The smaller
follow a certain route in parameter space during the distance, the more likely it is to spontaneously
the course of the seizure (Wendling et al. 2005). switch from one state to another. A biophysical
computational model of this dynamics was pro-
Bistability posed by Suffczynski et al. (2001). The model
As is the case for bifurcations, it is well was validated using animal experimental and
established that neurons (Tasaki 1959; Guttman clinical data and a numerical bifurcation analysis.
et al. 1980) and simple neural circuits (Kleinfeld The model could specifically explain how the D
et al. 1990; Foss et al. 1997; Foss and Milton interaction between cortical neural populations
2000; Ma and Wu 2007) can exhibit bistability. and populations in the thalamus interacts to pro-
It is to be expected that similar phenomena also duce the noise-induced transition to large-
occur on the level of the brain in vivo. Physicians amplitude, slow-frequency rhythms. In addition,
have known for a long time that it is sometimes an analysis of simulations of external stimulation
possible to abort pathological dynamics like car- suggested that the termination of the epileptic
diac arrhythmia or an epileptic seizure using state might be favorable at certain phases of the
a brief mechanical, electrical, or sensory stimulus epileptic rhythm only whereas the desired result
(see Rajna and Lona 1989; Lopes Da Silva might be missed for stimulations at other phases.
et al. 1994; Milton and Foss 1997; Milton 2000; The concept of bistability also arises in studies
Lopes Da Silva et al. 2003). This points to the concerning other neurological diseases, including
possibility that a seizure state coexists with migraine (see also “▶ Migraines and Cortical
a normal brain state and can be reached and left Spreading Depression”) and bipolar affective dis-
via specific perturbations (cf. Fig. 1b). The like- orders (see also “▶ Computational Psychiatry”).
lihood of causing a switch between coexisting Analysis of data of the spreading depression in
attractors by choosing the magnitude and timing migraine suggests a switch from the normal
of the stimulus randomly is very low (see also polarized state of the cortex to a totally
“▶ Multistability: Stopping Events with Single depolarized state (Dahlem et al. 2008). This tran-
Pulses”). Thus a great deal of attention has been sition is supported by computational models of
focused on determining the basin of attraction of the cortical polarization state. However, the full
the normal state (Lopes Da Silva et al. 2003; dynamics of spreading depression is only given if
Zakynthinaki et al. 2010), that is, the set of state the depolarization can propagate in space (see
variables that will result in normal behavior. below). In the context of psychiatric disorders,
Noise-induced transitions across the separatrix high-level considerations have led to the proposal
(additive noise) provide another possible mecha- of a bistable mechanism for the occurrence of
nism for epilepsy in the setting of bistability since disorders of the bipolar spectrum. The interaction
they can be accompanied by the occurrence of of a manic and a depressed state can lead to
delay-induced transient oscillations (see also a bistable situation which exhibits both the pos-
Pakdaman et al. 1998). Delay-induced transient sibility for regular transitions between the states
oscillations may be related to the tendency of (manic-depressive illness) and the possibility for
seizures in patients with nocturnal frontal lobe perturbation-induced transition (e.g., the known
epilepsy to occur at transitions between sleep phenomenon that antidepressants are able to
stages (Quan et al. 2011). induce the transition to the manic state
In the case of primary generalized epileptic (Goldbeter 2011)).
seizures, analysis of the clinical data has
supported the hypothesis that the epileptic state Excitability
coexists with the normal brain dynamics. From Neurons are excitable and excitability can also be
this point of view, the main difference between observed in large neural populations, e.g., in
a normal person and a patient experiencing the human cortex. In an individual neuron,
a well-defined depolarization threshold is found EEG. Abnormal wave propagation arises in

beyond which an action potential can be induced. a number of neural conditions including flicker
In neural populations within the living brain, it is phosphenes (see also “▶ Flicker-Induced Phos-
difficult to do quantitative experiments, but clin- phenes”), drug-induced visual hallucinations (see
ical observations point to similar mechanisms in also “▶ Visual Hallucinations and Migraine
some pathological cases. Notably, in patients Auras”), migraine (see also “▶ Migraines and
with epilepsy, the so-called interictal spike is Cortical Spreading Depression”), stroke, and the
a candidate for suprathreshold response to stimuli propagation of epileptic seizures (see Milton and
from the surrounding brain tissue. Model studies Jung (2003), Baier et al. (2012) and also “▶ Epi-
support the idea that in some location on the lepsy: Computational Models”). However, sei-
cortex or within the hippocampus, the excitability zure propagation from an epileptic focus is
threshold is lowered such that the surrounding exceedingly complex (see, e.g., Chkhenkeli and
activity occasionally reaches the excitation Milton 2003; Milton et al. 2007). Moreover, it is
threshold and induces an abnormal response in difficult to obtain data on seizure propagation in
the form of a single spike or sharp wave. In humans since pathways involving subcortical and
contrast to action potentials (which last on the brainstem nuclei, for example, thalamocortical
order of a few milliseconds), the epileptic spike circuits, play major roles. The subject of seizure
or sharp wave lasts on the order of 60–200 ms. propagation represents an opportunity for com-
Cortical excitability has been proposed to be of putational neuroscience and may potentially
clinical relevance to distinguish normal from point to treatment strategies designed to prevent
abnormal tissue response. Valentı́n et al. (2005) or limit seizure propagation.
studied the response of different cortical loca-
tions to single-pulse stimulations and concluded Spontaneous Bursting Behaviors
that simple excitability, a single population spike, In dynamic systems, the subject of bursting
indicated normal tissue. In the case of abnormally behaviors falls under the broad subject heading
altered tissue, they frequently saw rhythmic of intermittency. There are several mechanisms
responses that consisted of several oscillations that can generate intermittent (bursting) behav-
before they disappeared. In addition, the epilep- iors. In contrast to the bursting behaviors
tiform responses appeared to recruit wider corti- exhibited by single neurons (Izhikevich 2007),
cal areas, i.e., they had a tendency to spread. intermittency refers patterns of burstiness in
A computational modeling study confirmed that which both the size and timing of the bursts
spatially extended neural population models of vary unpredictably. First, intermittency can be
human cortex gave single spike responses when generated by purely deterministic mathematical
the balance between excitation and inhibition models (Pomeau and Manneville 1980). This
was intact (Goodfellow et al. 2012). In cases type of intermittency belongs to the class of cha-
where locally diminished inhibition was otic systems, and in low-dimensional systems, it
assumed, the response to simulated perturbations is associated with inverse tangent, subharmonic,
was prolonged and rhythmic as in the or secondary Hopf bifurcations. Despite the
clinical data. deterministic nature of the model, the timing
In an excitable neural medium like the neocor- (and size) of the laminar phases cannot be
tex, propagation of a disturbance can take the predicted. However, the probability distribution
form of synchronized waves of neural activity. of the duration of laminar phases is often associ-
Synchronization of neural population activity is ated with a power law (as is observed for the other
involved in binding of sensory stimuli, and orga- mechanisms discussed below). Following some
nization of large numbers of neurons into syn- evidence from animal model data (Hramov
chronized waves lies at the basis of phenomena et al. 2006), it has been shown that the occurrence
such as consciousness, cognition, and sleep and of seizures in primary generalized epilepsy may
thought to underlie the generation of the be related to deterministic intermittency
(Goodfellow et al. 2011). Interestingly, the model represented by delayed antagonistic dynamics
resulted from the coupling of bistable local ele- (Cabrera 2005).
ments (cortical compartments) and introducing
a certain degree of heterogeneity in the parameter
for different cortical locations. The model pre- Computational Challenges
dicts the spontaneous transition to the seizure and
spontaneous termination and the essentially Dynamic behaviors can be combined with
unpredictable nature of typical absence seizures. molecular and structural information to make
A second mechanism that generates intermit- a complete classification of neural disease. The D
tency arises in slowly driven dissipative systems modern day’s reductionist approach of associat-
that self-organize themselves, without an exter- ing with each disease a molecular defect is often
nal fine-tuning of control parameters, into not sufficient to base the development of thera-
a critical state near a state transition. The critical peutic approaches. While temporal structure is
state is characterized by a propensity to generate obviously important for clinicians to diagnose
transient phenomena which occur on all length and treat pathological states such as ventricular
scales (see Bak 1996; Sornette 2004). In this tachycardia, its role in diseases for which mea-
situation, abnormal discharges in the form of surements with appropriate temporal resolution
transient oscillations continue to appear and are not routinely available as, e.g., pulsatile hor-
abort spontaneously even though the parameters mone release is less obvious to the practicing
are fixed and there are no external inputs. The physician. On the other hand, current dynamic
transients are induced by internal (e.g., thermal) disease approaches shed little light on the genesis
fluctuations. The key point is that the length of of diseases such as epilepsy. In other words, what
time between successive oscillatory bursts, or is the dynamic basis of epileptogenesis, namely,
laminar phases, obeys a power law distribution. the process which describes why an initially
This means that laminar phases of all length seizure-free individual, such as a child, develops
scales are possible (Stead et al. 2010). Evidence over the course of many years medically intrac-
for such power law behaviors has been obtained table epilepsy. Presumably, the missing link is
in the case of epileptic seizure recurrences related to the self-organization of complex
(Osorio et al. 2010) and neuronal avalanches dynamic systems. The nonlinear dynamics of
(see Beggs and Plenz 2003; Milton 2012). physiological processes assumes that macro-
Although no general method is known that can scopic behavior is in the form of dissipative struc-
prevent these recurrences, it is currently thought tures, namely, structures which are maintained by
that some degree of predictability may be possi- a supply of energy. This implies that although the
ble (Sornette and Ouillon 2012). generating anatomical-physiological mecha-
Finally, noise-induced type of intermittency nisms may be rather complex, structural features
referred to as on-off intermittency can arise in can be studied properly in simplified models that
dynamic systems that are tuned close enough to omit certain details of the biological mechanisms.
a stability boundary (“edge of stability”) so that We stress that only in rather simple and abstract,
noisy perturbations cause transient switches nonlinear models can current mathematical tools
between two states (see also “▶ Human allow a thorough explanation of the basic mech-
Balancing Tasks: Power Laws, Intermittency, anism of change between patterns that are asso-
and Lévy Flights”). In the case that the noise ciated with a transition to disease.
enters through a parameter (multiplicative Questions of scale are central to dynamic
noise), power law behaviors arise (Cabrera and approaches to disease. In other words, “How are
Milton 2002). These phenomena would be antic- dynamics at the length scale of molecules trans-
ipated to arise generically in situations in which lated into behaviors that are observable at the
instability, such as manifested by epileptic cor- bedside?” Our emphasis on simple models to
tex, is controlled by cortical and subcortical investigate complex phenomena is based on the
observation that complex dynamic systems can incorporate the PRC have been quite successful
exhibit phenomena which are largely indepen- in describing, for example, bistability in simple
dent of scale. For example, limit cycle oscilla- neural circuits (Foss and Milton 2000). Popula-
tions can be generated at the level of molecules, tion models using the RC approach exhibit many
single neurons, small neural circuits, large neural complex and counterintuitive dynamic properties
populations, the human nervous systems, and (Timme and Wolf 2008).
human populations. Despite differences in the The second approach uses neural mass
details and scales of the underlying mathematical models, such as that developed by Wilson and
models, limit cycles at all levels of organization Cowan (see also “▶ Wilson-Cowan Model,”
share common properties with respect to the “▶ Visual Hallucinations and Migraine Auras,”
response to external perturbations including sta- “▶ Epilepsy: Computational Models,” “▶ Large-
bility, similar responses to perturbations, and the Scale Neural Networks: Vision”), to make
capacity to synchronize with other oscillations. a compromise between an abstract mathematical
Most nonlinear dynamic systems when tuned construct and a mechanistic biophysical model.
at the edge of stability exhibit low-dimensional In particular, it is possible to develop a statistical
dynamics (Guckenheimer 1995). In particular, mechanics for neurodynamics (Buice and Cowan
the center manifold theorem states that the 2009) and then analytically derive the nature of
dynamics at the edge of stability reduces to a the emergent properties that arise as a function of
low-dimensional normal form on a center mani- scale. However, the development of these statis-
fold that can be written in terms of first- or tical approaches requires that the effects of neu-
second-order differential equations referred to ronal heterogeneity and connectivity pathways be
as the normal forms. Hence, one would anticipate suitably averaged. Thus, it can become difficult
that one might be able to capture the dynamics for to translate predicted dynamic behaviors into
complex neural systems by using models based treatment strategies based on known molecular
on the normal forms. An example occurs in the and neurophysiological mechanisms. Instead,
spring-thumb compression task in which these models are more suited to capture the
a subject tries to compress a spring held between mean or population output of large numbers of
their thumb and index finger without causing the neurons. They can also be used to investigate the
spring to buckle (Venkadesan et al. 2007). nature of the underlying bifurcations and com-
Mechanical considerations suggest that the likely plex dynamic properties (Wang et al. 2012). By
normal form is the one that describes a pitchfork combining mathematical insight with mechanis-
bifurcation, i.e., tic predictions, it is possible to explore, for exam-
ple, the impact of inhibitory processes on a given
dx pathophysiological phenomenon.
¼ mx þ x3 x5 (1)
dt A further computational question concerns the
role of time delays. Time delays, t, are intrinsic
where m is a bifurcation parameter. This simple components of the nervous system and arise
model is sufficient to capture the dynamics of this because neural processing times and axonal con-
task observed in the laboratory. duction velocities are finite. The inclusion of time
A disadvantage of the normal form approach delays in computational models may become
is that it does not necessarily capture the respon- increasingly important as the spatially extended
sible mechanism(s). Two approaches have been nature of neural control mechanism is taken into
developed to try to close this gap. The first account. There is a 10- to 20-fold difference in
approach makes use of the phase resetting curve seizure propagation velocity that arises because
(PRC) for periodically spiking and bursting neu- of differences in the conduction velocity between
rons (see also “▶ Phase Response Curves: Over- unmyelinated thin and larger diameter, myelin-
view”). The advantage is that the PRC can be ated axons of cortical neurons (Milton 2003).
measured experimentally. Models which Thus, fast transfer of seizure activity occurs
initially in the direction perpendicular to the gray appropriate stimulus. Devices of this type have
matter via corticothalamic and intracortical and already been developed for treating epilepsy
interhemispheric commissural fibers, whereas (Osorio and Frei 2009). An exciting development
slower seizure transmission occurs in parallel is the increasing availability of noninvasive,
directions within the gray matter. These observa- physiological monitoring devices suitable for
tions point to computational models of cortical long-term uses (e.g., weeks, months) (Cook
dynamics which take the form of delay, or partial et al. 2013). Such data sets will likely prove
delay, differential equations (see also “▶ Time- invaluable since they will provide the first
Delayed Neural Networks: Stability and Oscilla- detailed insights into the nature of D
tions”) that include either multiple discrete neurodynamics during a patient’s daily life, as
delays or a distribution of delays. distinct from the behaviors observed in the phy-
sician’s office and in the hospital bed. The avail-
ability of such information will impact the goals
The Future of computational neuroscience: (1) should more
efforts be directed towards the prediction of rare
The identification of the dynamic aspects of neu- events and (2) can computational neuroscience
rological disease motivates the development of techniques be used to develop more efficient
dynamic treatment strategies. It is quite likely implementations of a control strategy, e.g., strat-
that proposed therapeutic strategies will be based egies that maximize battery life (changing
on the use of external stimuli to control dynamics. a battery requires an operation).
The stimuli could be chemical, electrical, or The remaining possibility for treating neuro-
mechanical (e.g., vibrations) and take the form of logical disease is to simply replace the broken
brief pulses or pulse trains. Even noisy signals can part with a new one. It may be eventually be
be of benefit (see also “▶ Stochastic Resonance: possible to use stem cell techniques to produce
Balance Control and Cochlear Implants”). Most the replacement part. However, it is more likely
present-day implantable devices operate in an to be the case in diseases of the nervous system
open loop mode, namely, the same stimulus or that the replacement part will take the form of an
stimulus protocol is delivered regardless of the electronic circuit that can directly interface with
state of the patient’s nervous system. Examples the patient’s nervous system. Existing examples
include the vagal nerve stimulator for epilepsy, include brain motor prosthesis (see also “▶ Func-
deep brain stimulators for Parkinson’s disease tional Neuroscience: Cortical Control of Limb
(see also “▶ Parkinson’s Disease: Deep Brain Prosthesis” and “▶ Cortical Motor Prosthesis”),
Stimulation” and “▶ Computational Models of exoskeletons, and insect robots. In some sense
Deep Brain Stimulation (DBS)”), and spinal cord computational neuroscience may be well suited
stimulation for pain control and spasticity. An for the design of such replacement neural circuits.
important observation is that all of these After all, the nervous system appears to be plastic
implementations do not benefit from insights enough to adapt to whatever errors the circuits
potentially available from modeling efforts. may have introduced. Thus, we anticipate that
Thus, from the computational neuroscience point computational approaches may play an important
of view, the good news is that there is a lot of room role in this area.
for improvement.
Closed loop treatment strategies have the
advantage that the therapy is delivered based on Cross-References
the state of the patient’s nervous system (Milton
and Foss 1997; Milton and Jung 2003). Such ▶ Computational Psychiatry
devices would likely take the form of implantable ▶ Computational Models of Deep Brain
electronic computer devices which both monitor Stimulation (DBS)
the important variables and then deliver the ▶ Cortical Motor Prosthesis
▶ Epilepsy: Computational Models spreading depolarizations in migraine and stroke.

▶ Flicker-Induced Phosphenes Chaos Interdiscip J Nonlinear Sci 18(2):026110
Ebersole JS, Milton J (2003) The electroencephalogram
▶ Functional Neuroscience: Cortical Control of (EEG): a measure of neural synchrony. In: Milton J,
Limb Prosthesis Jung P (eds) Epilepsy as a dynamic disease. Springer,
▶ Human Balancing Tasks: Power Laws, Berlin/Heidelberg/New York, pp 51–68
Intermittency, and Lévy Flights Foley C, Mackey MC (2009) Mathematical model for
G-CSF administration after chemotherapy. J Theor
▶ Large-Scale Neural Networks: Vision Biol 257(1):27–44
▶ Migraines and Cortical Spreading Depression Foss J, Milton J (2000) Multistability in recurrent neural
▶ Multistability: Stopping Events with Single loops arising from delay. J Neurophysiol 84:975–985
Pulses Foss J, Moss F, Milton J (1997) Noise, multistability, and
delayed recurrent loops. Phys Rev E 55:4536–4543
▶ Parkinson’s Disease: Deep Brain Stimulation Gerstner W, Kistler WM (2002) Spiking neuron models:
▶ Phase Response Curves: Overview single neurons, populations, plasticity. Cambridge
▶ Stochastic Resonance: Balance Control and University Press, Cambridge
Cochlear Implants Glass L, Mackey MC (1979) Pathological conditions
resulting from instabilities in physiological control
▶ Time-Delayed Neural Networks: Stability and systems. Annu NY Acad Sci 316:214–235
Oscillations Glass L, Mackey MC (1988) From clocks to chaos: the
▶ Visual Hallucinations and Migraine Auras rhythms of life. Princeton University Press, Princeton
▶ Wilson-Cowan Model Goldbeter A (2011) A model for the dynamics of bipolar
disorders. Prog Biophys Mol Biol 105(1):119–127
Goodfellow M, Schindler K, Baier G (2011) Intermittent
spike-wave dynamics in a heterogeneous, spatially
References extended neural mass model. Neuroimage 55:920–932
Goodfellow M, Schindler K, Baier G (2012)
Baier G, Goodfellow M, Taylor PN, Wang Y, Garry DJ Self-organised transients in a neural mass model of
(2012) The importance of modeling epileptic seizure epileptogenic tissue dynamics. Neuroimage
dynamics as spatio-temporal patterns. Front Physiol 59:2644–2660
3:281 Guckenheimer J (1995) A robust hybrid stabilization strat-
Bak P (1996) How nature works: the science of self- egy for equilibria. IEEE Trans Autom Control
organized criticality. Copernicus, New York 40:321–326
Beggs JM, Plenz D (2003) Neuronal avalanches in neo- Guttman R, Lewis S, Rinzel J (1980) Control of repetitive
cortical circuits. J Neurosci 23:11167–11177 ring in squid axon membrane as a model for a neuron
Bélair J, Glass L, an der Heiden U, Milton J (1995) oscillator. J Physiol 305:377–395
Dynamical disease: mathematical analysis of human Hramov A, Koronovskii AA, Midzyanovskaya IS,
illness. American Institute of Physics, Woodbury Sitnikova E, Van Rijn CM (2006) On-off intermittency
Buice MA, Cowan JD (2009) Statistical mechanics of the in time series of spontaneous paroxysmal activity in
neocortex. Prog Biophys Mol Biol 99:53–86 rats with genetic absence epilepsy. Chaos Interdiscip
Cabrera JL (2005) Controlling instability with delayed J Nonlinear Sci 16(4):043111
antagonistic stochastic dynamics. Phys A Stat Mech Izhikevich EM (2007) Dynamical systems in neurosci-
Appl 356(1):25–30 ence: the geometry of ex-citability and bursting. MIT
Cabrera JL, Milton JG (2002) On-off intermittency in Press, Cambridge, MA
a human balancing task. Phys Rev Lett 89:158702 Kleinfeld D, Raccuia-Behling F, Chiel HJ (1990) Circuits
Chkhenkeli SA, Milton J (2003) Dynamic epileptic sys- constructed from identified aplysia neurons exhibit
tems versus static epileptic foci? In: Milton J, Jung multiple patterns of persistent activity. Biophys
P (eds) Epilepsy as a dynamic disease. Springer, Ber- J 57(4):697–715
lin/Heidelberg/New York, pp 25–36 Lopes Da Silva FH, Pijn JP, Wadman WJ (1994) Dynam-
Cook MJ, O’Brien TJ, Berkovic SF, Murphy M, ics of local neuronal networks: control parameters and
Moroko A, Fabinyi G, D’Souza W, Yerra R, state bifurcations in epileptogenesis. Prog Brain Res
Archer J, Litewka L, Hosking S, Lightfoot P, 102:359–370
Ruedebusch V, Sheffield WD, Snyder D, Leyde K, Lopes Da Silva FH, Blanes W, Kalitzin SN, Parra J,
Himes D (2013) Prediction of seizure likelihood with Suczynski P, Velis DN (2003) Dynamical diseases of
a long-term, implanted seizure advisory system in brain systems: different routes to epileptic seizures.
patients with drug-resistant epilepsy: a first-in-man IEEE Trans Biomed Eng 50:540–548
study. Lancet Neurol 12(6):563–571 Ma J, Wu J (2007) Multistability in spiking neuron models
Dahlem MA, Schneider FM, Schöll E (2008) Failure of of delayed recurrent inhibitory loops. Neural Comput
feedback as a putative common mechanism of 19(8):2124–2148
Dynamical Diseases 1061 D
Mackey MC, an der Heiden U (1982) Dynamical diseases Stead M, Bower M, Brinkmann BH, Lee K, Marsh RW,
and bifurcations: understanding functional disorder in Meyer FB, Litt B, Van Gompel J, Worrell GA
physiological systems. Funkt Biol Med 1:156–162 (2010) Microseizures and the spatiotemporal scales
Mackey MC, an der Heiden U (1984) The dynamics of of human partial epilepsy. Brain 133(9):2789–2797
recurrent inhibition. J Math Biol 19:211–225 Suffczynski P, Kalitzin S, Lopes Da Silva FH
Mackey MC, Glass L (1977) Oscillation and chaos in (2001) Dynamics of nonconvulsive epileptic phenom-
physiological control systems. Science 197:287–289 ena modeled by a bistable neuronal network. Neuro-
Mackey MC, Milton JG (1987) Dynamical diseases. Annu science 126(2):467–484
NY Acad Sci 504:16–32 Tasaki I (1959) Demonstration of two stable states of the
Milton JG (2000) Epilepsy: multistability in a dynamic nerve membrane in potassium-rich media. J Physiol
disease. In: Walleczek J (ed) Self-organized biological (Lond) 148:306–331 D
dynamics and nonlinear control. Cambridge Univer- Timme M, Wolf F (2008) The simplest problem in the
sity Press, New York, pp 374–386 collective dynamics of neural networks: is synchrony
Milton J (2003) Insights into seizure propagation from stable? Nonlinearity 21:1579–1599
axonal conduction times. In: Milton J, Jung P (eds) Valentı́n A, Alarcón G, Honavar M, Garcı́a Seoane JJ,
Epilepsy as a dynamic disease. Springer, Berlin/Hei- Selway RP, Polkey CE, Binnie CD (2005) Single
delberg/New York, pp 15–23 pulse electrical stimulation for identification of struc-
Milton J (2012) Neuronal avalanches, epileptic quakes tural abnormalities and prediction of seizure outcome
and other transient forms of neurodynamics. Eur after epilepsy surgery: a prospective study. Lancet
J Neurosci 35:2156–2163 Neurol 4(11):718–726
Milton J, Black D (1995) Dynamic diseases in neurology Venkadesan M, Guckenheimer J, Valero-Cuevas FJ
and psychiatry. Chaos 5(1):8–13 (2007) Manipulating the edge of stability. J Biomech
Milton JG, Foss J (1997) Oscillations and multistability in 40:1653–1661
delayed feedback control. In: Othmer HG, Adler Wang Y, Goodfellow M, Taylor PN, Baier G (2012) Phase
FR, Lewis MA, Dallon JC (eds) The art of mathemat- space approach for modeling of epileptic dynamics.
ical modeling: case studies in ecology, physiology and Phys Rev E 85:061918
cell biology. Prentice-Hall, Englewood Cliffs, Wendling F, Bartolomei F, Bellanger JJ, Chauvel P (2002)
pp 179–198 Epileptic fast activity can be explained by a model of
Milton J, Jung P (2003) Epilepsy as a dynamic disease. impaired gabaergic dendritic inhibition. Eur
Springer, Berlin/Heidelberg/New York J Neurosci 15:1499–1508
Milton JG, Chkhenkeli SA, Towle VL (2007) Brain con- Wendling F, Hernandez A, Bellanger JJ, Chauvel P,
nectivity and the spread of epileptic seizures. In: Jirsa Bartolomei F (2005) Interictal to ictal transition in
VK, McIntosh AR (eds) Handbook of brain connectiv- human temporal lobe epilepsy: insights from
ity. Springer, Berlin/Heidelberg/New York, pp 477–503 a computational model of intracerebral EEG. J Clin
Osorio I, Frei MG (2009) Real-time detection, quantifica- Neurophysiol 22:343–356
tion, warning, and control of epileptic seizures: the Wilson HR (1999) Simplified dynamics of human and
foundations for a scientific epileptology. Epilepsy mammalian neocortical neurons. J Theor Biol
Behav 16(3):391–396 200:375–388
Osorio I, Frei MG, Sornette D, Milton J, Lai YVC Zakynthinaki MS, Stirling JR, Cordent Martinez CA,
(2010) Epileptic seizures: quakes of the brain? Phys Diaz L, de Durana A, Quintana MS, Romo GR,
Rev E 82:021919 Molinueve JS (2010) Modeling the basin of attraction
Pakdaman K, Grotta-Ragazzo C, Malta CP (1998) Tran- as a two-dimensional manifold from experimental data:
sient regime duration in continuous-time networks applications to balance in humans. Chaos 20:013119
with delay. Phys Rev E 58:3623–3627
Pomeau Y, Manneville P (1980) Intermittent transition to
turbulence in dissipative dynamical systems. Commun
Math Phys 74:189–198 Dynamic Models
Quan A, Osorio I, Ohira T, Milton J (2011) Vulnerability
to paroxysmal oscillations in delayed neural networks:
a basis for nocturnal frontal lobe epilepsy? Chaos ▶ Metabotropic Receptors (G Protein Coupled
21:047512 Receptors)
Rajna P, Lona C (1989) Sensory stimulation for inhibition
of epileptic seizures. Epilepsia 30(2):168–174
Sornette D (2004) Critical phenomena in natural sciences:
chaos, fractals, self-organization, and disorder: con- Dynamical Diseases
cepts and tools. Springer, Berlin/Heidelberg
Sornette D, Ouillon G (2012) Dragon kings: mechanism,
statistical methods, and empirical evidence. Eur Phys ▶ Dynamic Diseases of the Brain
J Spec Top 205:1–26 ▶ Dynamics of Disease States: Overview
E
Echo State Networks Efficient Population Coding
▶ Recurrent Network Models, Reservoir Matthias Bethge

Computing Werner Reichardt Centre for Integrative
Neuroscience, University of T€ubingen and Max
Planck Institute for Biological Cybernetics,
Echoic Memory
Synonyms
▶ Auditory Memory
Coding accuracy of neural populations; Coding
efficiency of neural populations; Information
transmission by neural populations
EEG, Neural Population Models of

Definition
▶ Neuroimaging, Neural Population Models for
Natural stimulations caused by objects in the
surrounding world do not stimulate single sen-
sory receptors in isolation but lead to the activa-
tion of large numbers of neurons simultaneously.
Effective Connectivity, From fMRI Thus, typical stimulus variables of interest are
Data represented only implicitly in activation patterns
across large neural populations. These patterns
▶ Statistical Analysis of Neuroimaging Data are statistical in nature since repeated

E 1064 Efficient Population Coding
presentation of the same stimulus usually leads to a

highly variable responses. The large dimension-
ality and randomness of the neural responses
make it difficult to assess how well different
stimuli can be discriminated. Depending on how
effectively neurons share the labor of encoding,
the accuracy with which stimuli are represented b
can change dramatically. Thus, studying the effi-
ciency of population codes is important for our
understanding of both which information is
discrimination
encoded in neural populations and how it is
error
encoded.
Efficient Population Coding, Fig. 1 In each trial of an
The question for the efficiency of population orientation discrimination task two gratings are presented
codes naturally arises if one seeks to relate the as depicted in (a) where one of the two gratings is tilted to
psychometric performance in a sensory discrim- the right and the task is to say which of the two gratings is
tilted. The task performance can be quantified by measur-
ination task to the amount of information avail-
ing the error frequency as a function of the tilt angle. The
able from the activation of neural populations. resulting neurometric function is sketched in (b)
For illustration, we consider an orientation dis-
crimination task (Fig. 1a) for which the behav-
ioral performance can be summarized by the neurons in this population (Zohary et al. 1994).
psychometric function (Fig. 1b). The latter quan- In case of suboptimal decoding, the sensory pop-
tifies the discrimination error as a function of the ulation carries more information about the task
stimulus parameter of interest. If one records than its individual neurons, but the readout of
from a neural population, one can use decoders downstream neurons is unable to extract this dis-
to solve the same task on the basis of the neural tributed information (Beck et al. 2012).
responses and compare the resulting neurometric The relationship between neural representa-
performance to the psychometric one. In addition and behavioral performance can be studied
tion, one can try to predict the behavioral choice using probabilistic modeling. In the beginning we
from the neural responses which is quantified will assume that the stimulus entity of interest is
using the so-called choice probabilities . parameterized by a one-dimensional variable
For several tasks it has been observed that the x (e.g., contrast, orientation, etc.). Later, we will
neurometric function obtained for a single neuron also discuss the encoding of high-dimensional
or few is sufficient to explain the behavioral, stimuli such as natural images in the context of
psychometric performance. Given that the popu- normative approaches to population coding that
lation of sensory neurons encoding for the stim- investigate optimal encoding strategies from an
ulus contains many more neurons than just the engineering perspective.
recorded ones, this observation indicates that the Throughout this entry neural responses will be
encoding or the decoding of the task-relevant denoted by a vector r which oftentimes describes
information from this sensory population is the spike counts of the different neurons but could
highly suboptimal: In case of suboptimal similarly describe other properties such as
encoding, the information conveyed by each of response latencies, interspike intervals, or alike
these neurons is extremely redundant such that as well. Since the same stimulus usually leads to
the entire population does not convey more infor- variable responses, the relationship between stim-
mation about the task than small subsets of ulus and population response is determined by the
Efficient Population Coding 1065 E
Efficient Population Coding, Fig. 2 The decision boundary of the MAP estimator is located at the intersection of the
two curves. Correspondingly, the minimum discrimination error is given by the shaded area
conditional probability distribution p(r|x). This 0 ð d 0 E

d 2
conditional probability function specifies the F ¼ N ðzj0, 1Þdz, where d 0 :
2 1
encoding of the task-relevant stimulus variable. rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
! ! ⊤ 1 ! !
¼ m1 m0 C m1 m0 :
Ideal Observer Analysis (Minimum
Discrimination Error)
In order to assess how much information in prin- In the absence of noise correlations, the noise
ciple can be extracted from a neuron or covariance matrix C is diagonal. Its dimensionality
0
a population, it is necessary to determine the pffiffiffiffithe number of neurons N and d is of the order
equals
performance of an ideal observer. In a binary of N . Thus, for uncorrelated noise the minimum
discrimination task with prior probabilities discrimination error converges to zero in the limit
p(x = x1) = l and p(x = x2) = 1 –l for the two of large N. In the presence of certain types of noise
possible stimuli x1, x2, the discrimination error is correlations, however, the error may saturate at
minimized by the maximum a posteriori (MAP) a nonzero level. For example, this can happen in
estimator x^MAP ¼ x1 if pðx1 jrÞ > pðx2 jrÞ and case of input noise which obviously cannot be
x^MAP ¼ x2 if pðx2 jrÞ > pðx1 jrÞ . Correspond- removed by subsequent signal transmission (unless
ingly, the minimum discrimination error (MDE) by the use of contextual side information).
is given by (see Fig. 2) As an illustration, consider the following lin-
ð
ear encoding r5f ðx2 þ nx Þ þ nr with f ðxÞ ¼ wx,
where nx N 0, sx denotes the input noise
MDE ¼ minflpðrjx1 Þ, ð1 lÞpðrjx2 Þgdr and nr N 0, s2r I constant additive output
noise. In this case, the total noise covariance
In case of Gaussian distributions matrix reads ws2x w⊤ + s2r I, and its largest eigen-
pðrjxÞ ¼ N ðrjf ðxÞ, CÞ with fixed noise covari- value is given by s2x + s2r with eigenvector w. The
ance matrix C, one can show that the optimal remaining (n–1) dimensions belong to the same
classifier is linear and takes the following form: degenerated eigenvalue s2r . Using the Sherman-
Morrison formula, it is possible to determine the
D E
! ! ! ! inverse of this covariance matrix
x^MAP ¼ x1 if m 1 m 2 C1 x m

l 1 1 s2x ww⊤
> log , x^MAP ¼ x2 otherwise C ¼ 2 I 2
1l sr sr þ s2x w⊤ w
D E
! and
with m1 ¼ f ðx1 Þ, m2 ¼ f ðx2 Þ, and m
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
A w⊤ w
1 ! ! 0
¼ m1 þ m2 : d ¼ Dx with A :¼ ,
2 1 þ s2x A s2r
Furthermore, if l = 1 l = 0.5, the MDE is where A is of the order of N. Accordingly, for

given by finite input noise sx > 0, the population
Efficient Population Coding, Fig. 3 The sensitivity

index d0 as a function of the number of neurons N in the
presence (red) or absence (blue) of input noise
Efficient Population Coding, Fig. 4 The minimum
discrimination can be bound from above and below
(Berens et al. 2009) by the Jensen-Shannon information
sensitivity d0 converges to Dx/sx in the limit N !
H 1
B ð1 I JS Þ MDE 2 ð1 I JS Þ where HB(u) = u
1
1 to. This can be one reason why behavioral log u (1u) log (1u) and the Jensen-Shannon infor-
performance may be not much better than one mation is equal to the mutual information IJS = I [x : r]
would predict on the basis of the neurometric between the stimulus x and the responses r
performance of very few neurons. By estimating
(Dx/d0 )2 as a function of N for empirical data, one lower bound on the minimum discrimination
can use the following relationship error (see Fig. 4) but is not related to it in
2 a one-to-one fashion (Berens et al. 2009). For
Dx 1 N!1 2 Gaussian stimulus distributions, if we still
¼ s2x þ ! sx
d0 A assume a linear encoding and a Gaussian noise
model, the joint distribution of stimulus and
for estimating the effective amount of input noise response is Gaussian, and mutual information as
(see also Fig. 3). well as the minimum mean square error can be
Further reading: (Dayan and Abbott 2001; computed analytically. For nonlinear encodings
Salinas and Abbott 1994; Abbott and Dayan and non-Gaussian noise models, these quantities
1999; Shamir and Sompolinsky 2004, 2006; are much more difficult to assess and approxima-
Ecker et al. 2011; Haefner et al. 2013; Shamir tion approaches are important.
2014).
Fine Discrimination and Fisher Information
Discrimination, Mutual Information, and We cannot expect that neural encoding is accu-
Minimum Mean Square Estimation rately captured by a linear Gaussian model. Nev-
The minimum discrimination error is only one ertheless, in the case of fine discrimination, when
way to evaluate the coding accuracy of a neural |x2–x1| is sufficiently small, d0 can often be used to
population. Other popular measures are based on quantify approximately how well the two stimu-
the mutual information between stimulus and lus alternatives can be discriminated.
response or minimum mean square estimation. In a fine discrimination task, we study how
In contrast to the discrimination approach, these well a stimulus at any location x1 = x can be
other measures make assumptions about the stim- discriminated from a stimulus with a slightly
ulus distribution in order to obtain an average changed parameter value x2 = x + Dx. Assuming
performance. If one assumes a binary stimulus that the tuning functions are smooth and the noise
distribution, the mutual information between variance is finite, the discrimination performance
stimulus and response is also called the Jensen- approaches chance level MDE ! 0.5 in the
Shannon divergence which provides an upper and limit of small differences in the stimulus
parameter Dx ! 0. The asymptotic behavior in the often requires additional considerations that go
limit of fine discrimination is more amenable to an beyond the framework of discrimination. While
analytical derivation of coding efficiency. Specif- for each pair of stimuli tested for in
ically, we can bound the minimum discrimination a discrimination task the prior distribution is
by the Jensen-Shannon information (see Fig. 2) reduced to two discrete alternatives, for continu-
which for small Dx can be approximated by the ous stimulus parameters in principle, any pair of
Fisher information: stimuli could be picked. Therefore, one often
rather wants to assess the coding efficiency for
J ðx Þ 2 a prior distribution that describes how the stimu-
I JS ðDxÞ Dx :
8 lus parameter is distributed under natural condi-
E
tions. For example, if the stimulus parameter is
Fisher information is defined as the variance contrast or speed, one would assume that the
of the score V = @ x log p(r|x) conditioned on the natural distribution has a unimodal density over
stimulus x: the positive real axis which takes its maximum at

zero and is monotonically decreasing.
J ðxÞ :¼ Var ½Vjx ¼ E V 2 jx To be able to compare the performance of two
coding schemes, it is necessary to use a measure
where the second equality holds because it gen- that provides a unique ordering relation. A direct
erally holds E[V|x] = 0. Many papers on neural ranking of neurometric functions, for example, is
coding refer to the Cramer-Rao bound to moti- not possible unless in the special case if one
vate the evaluation of Fisher information which function lies uniformly above or below the other
gives a lower bound on the mean squared error of one. If one was interested exclusively in fine
any unbiased estimator: discrimination, one could compare just the slope
of the neurometric function at chance level
h i 1 describing how fast the MDE drops from chance
E ðxûnbiased ðrÞ xÞ2 jx :
J ðx Þ level and ignore the discrimination performance
for larger differences in the stimulus parameter.
This bound, however, is only of intuitive use In most cases however, one would rather be inter-
for the evaluation of encodings as it does not ested in the average performance when the stim-
account for the fact that the signal x must be uli are drawn from the natural distribution. The
a random variable and the precision with which corresponding notions and assessments of coding
x can be decoded strongly depends on the stimu- efficiency can be very different. The optimization
lus distribution p(x). Similarly, the local expan- for fine discrimination is tightly related to opti-
sion of the Jensen-Shannon divergence must be mizing Fisher information, whereas an optimiza-
wrong whenever IJS exceeds 1 bit since it gener- tion of the average discrimination error (Berens
ally holds IJS 1 bit. Thus the approximation for et al. 2011) is more similar (but not equivalent) to
fine discrimination only holds if optimizing mutual information (Yarrow et al.
2012) or the minimum mean squared error.
8 There is a special condition for which the
Dx2 Dx2c
:
J ðx Þ minimum mean squared error is determined by
Fisher information. If the responses of the differ-
In the next section, we will discuss the evalu- ent neurons can be described as independent,
ation of population codes if the stimulus distribu- identically distributed (i.i.d.) observations or
tion p(x) is given. more generally if the conditions of the central
limit theorem are fulfilled, then Fisher informa-
Coding Efficiency and Optimal Encodings tion can be used to determine an asymptotic
The question of coding efficiency critically approximation of the decoding error (Bethge
depends on the stimulus distribution p(x) which et al. 2002). Optimal codes, however, generically
exhibit highly heterogeneous encodings which the interpretation of the available experimental
preclude the use of the central limit theorem. In data. As mentioned in the beginning of this entry,
particular, one should note that the linear increase it is typically observed that the neurometric func-
of Fisher information in case of i.i.d. observations tion obtained for a single neuron or few is suffi-
corresponds to a very slow logarithmic growth in cient to explain the behavioral, psychometric
mutual information. In contrast, encodings that performance and that this may hint to highly
are optimized for mutual information usually inefficient encoding or decoding. If the encoding
achieve a much faster, linear growth of the is fixed (i.e., independent of the task), it is
mutual information. The two situations can be straightforward to show that the efficiency of
illustrated with the following two prototypical the encoding will depend on the task. For illus-
examples: tration, let us assume we have the same number of
neurons as stimulus dimensions and the follow-
N
ing encoding with independent noise model:
i:i:d:coding : piid ðr, xÞ ¼ pðxÞ ∏ pðrk jf ðxÞÞ vs
k¼1
N
N
factorial coding : pfac ðr, xÞ ¼ pðxÞ ∏ pðr k , f k ðxÞÞ: penc ðrjxÞ ¼ ∏ pðr k jxk Þ:
k¼1
k¼1
Depending on which stimulus distribution

It is easy to derive (Brunel and Nadal 1998)
p(x) we use, one can obtain either a highly redun-
that for the left example of i.i.d. coding mutual
dant i.i.d. coding scheme or a maximally efficient
information grows only logarithmically, I[r, x] /
factorial code. Specifically, if we set
log(N), whereas for the right example of
p(x) = ∏ Nk = 1p(xk) we obtain the maximally
a factorial code (Nadal and Parga 1994), mutual
efficient factorial code
information grows linearly, I[r, x] / N. Thus, for
large populations of neurons, i.i.d. coding is N
highly inefficient compared to factorial coding. penc ðrjxÞpðxÞ ¼ ∏ penc ðr k jxk Þpðxk Þ
In cases other than i.i.d. coding (or, more k¼1
generally, whenever the conditions for the central N

¼ ∏ pðr k , xk Þ,
limit theorem are not fulfilled), the behavior of k¼1
Fisher information can be very different from that
of mutual information. For example, it is possible whereas if we choose a stimulus distribution
to build factorial codes for which Fisher informa- p(x) = p(x1) ∏Nk = 2 d(xk x1), for which it
tion grows only linearly in N just like in the case always holds x1 = x2 = = xN, we then obtain
of i.i.d. coding, but an exponential or super- the highly redundant i.i.d. coding scheme. The
exponential growth of Fisher information is pos- first example roughly corresponds to a task where
sible as well. Conversely, it is easy to construct any two images are to be discriminated, while in
encodings for which Fisher information grows the second case only full-field uniform gray
arbitrarily fast with N, but this does not imply images need to be discriminated.
that mutual information grows linearly (Bethge The mutual information between stimulus
et al. 2002). Fisher-optimal codes are exclusively x and population response r cannot be larger
optimized for fine discrimination and may exhibit than the entropy of the stimulus distribution. For
extremely poor performance for any discrimina- good reasons, typical laboratory tasks are inher-
tion for which Dx2 Dx2c . ently biased toward low entropy stimuli. More
specifically, most experiments are designed
Task Dependence of Coding Efficiency such that it is not critical which neurons exactly
While it is obvious that coding efficiency is are recorded. In vision, this is achieved by using
highly task dependent, it is important to under- global homogeneous stimuli such as global
stand the implications of this task-dependency for motion patterns or gratings. By varying only
one parameter of these stimuli like the net veloc- Cross-References
ity or orientation, it is achieved that all neurons
encode the same parameter in an i.i.d. fashion. ▶ Bayesian Inference with Spiking Neurons
Correspondingly, mutual information cannot ▶ Choice Behavior
grow faster than logarithmically with the number ▶ Cortical Function, Normative Models of
of neurons yielding a highly suboptimal popula- ▶ Decision-Making, Threshold
tion coding strategy. ▶ Estimating Information-Theoretic Quantities
For many tasks, however, it is likely that this
type of suboptimality is not sufficient to explain
why behavioral performance does not utilize References E
more information than provided by a single or
few neurons. For spatially homogeneous stimuli, Abbott L, Dayan P (1999) The effect of correlated vari-
the information of neurons with different recep- ability on the accuracy of a population code. Neural
Comput 11(1):91–101
tive field locations should still accumulate such
Beck J, Ma W, Pitkow X, Latham P, Pouget A (2012) Not
that doubling the number of neurons should noisy, just wrong: the role of suboptimal inference in
reduce the squared error by a factor of two as behavioral variability. Neuron 74(1):30–39
predicted by an i.i.d. coding scheme. This fact, Berens P, Gerwinn S, Ecker AS, Bethge M (2009)
Neurometric function analysis of population codes.
however, is not consistent with the observation
In: Advances in neural information processing sys-
that the information of a local population of neu- tems, vol 22. Vancouver, BC, Canada
rons reaches behavioral performance. Thus, it Berens P, Ecker AS, Gerwinn S, Tolias AS, Bethge M
seems that for typical low-level tasks, the readout (2011) Reassessing optimal neural population codes
with neurometric functions. Proc Natl Acad Sci
cannot integrate information from different local
USA 108(11):4423–4428
regions. Bethge M, Rotermund D, Pawelzik K (2002) Optimal
In conclusion, both encoding and decoding are short-term population coding: when fisher information
likely to render behavioral performance highly fails. Neural Comput 14(10):2317–2351
Bethge M, Rotermund D, Pawelzik K (2003a) Optimal
suboptimal in case of common laboratory tasks
neural rate coding leads to bimodal firing rate
for which the decision space has low entropy and distributions. Network Comput Neural Syst 14(2):
for which the stimulus size is rather large. In the 303–319
future, important new insights about principles of Bethge M, Rotermund D, Pawelzik K (2003b) Second
order phase transition in neural rate coding: binary
neural coding and sensory decision making may
encoding is optimal for rapid signal transmission.
be obtained by choosing more ecologically rele- Phys Rev Lett 90(8):088104
vant tasks such as natural scene classification. Brunel N, Nadal J-P (1998) Mutual information, fisher
The importance of studying natural tasks may information, and population coding. Neural Comput
10(7):1731–1757
be corroborated by the example of grid cell
Dayan P, Abbott LF (2001) Theoretical neuroscience,
encoding in hippocampus. Despite the fact that vol 31. MIT Press, Cambridge, MA
the stimulus parameter is only two Ecker AS, Berens P, Tolias AS, Bethge M (2011) The
dimensional – namely, the x-y position of the effect of noise correlations in populations of diversely
tuned neurons. J Neurosci 31(40):14272–14283
animal – the mutual information between this
Haefner R, Bethge M (2010) Evaluating neuronal codes
stimulus parameter and the grid cells scales pro- for inference using fisher information. In: Advances in
portional to the number of neurons (Sreenivasan neural information processing systems, vol 23.
and Fiete 2011; Mathis et al. 2012). Similarly, Vancouver, BC, Canada
Haefner RM, Gerwinn S, Macke JH, Bethge M (2013)
one would expect a linear growth of mutual infor-
Inferring decoding strategies from choice probabilities
mation for the encoding of natural images in the in the presence of correlated variability. Nat Neurosci
visual system. 16:235–242
Further reading: (Brunel and Nadal 1998; Huys QJ, Zemel RS, Natarajan R, Dayan P (2007) Fast
population coding. Neural Comput 19(2):404–441
Bethge et al. 2003a, b, Huys et al. 2007;
Klam F, Zemel RS, Pouget A (2008) Population coding
McDonnell and Stocks 2008; Klam et al. 2008; with motion energy filters: the impact of correlations.
Haefner and Bethge 2010). Neural Comput 20(1):146–175
E 1070 Electrical Nerve Block
Mathis A, Herz AVM, Stemmler MB (2012) Optimal Definition

population codes for space: grid cells outperform
place cells. Neural Comput 24(9):2280–2317
McDonnell MD, Stocks NG (2008) Maximally informa- ECoG refers to the invasive technique of record-
tive stimuli and tuning curves for sigmoidal rate- ing brain electrical field potentials with elec-
coding neurons and populations. Phys Rev Lett trodes placed directly on the cortical surface.
101(5):058103 ECoG electrodes are typically placed subdurally,
Nadal J-P, Parga N (1994) Nonlinear neurons in the
low-noise limit: a factorial code maximizes informa- above the pia mater. In humans, ECoG is often
tion transfer. Network Comput Neural Syst conducted in patients with medically intractable
5(4):565–581 epilepsy in order to monitor their seizures.
Salinas E, Abbott L (1994) Vector reconstruction from Because a craniotomy (a surgical incision into
firing rates. J Comput Neurosci 1(1–2):89–107
Shamir M (2014) Emerging principles of population cod- the skull) is required to implant the electrode
ing: in search for the neural code. Curr Opin Neurobiol grid, ECoG is an invasive procedure. Clinical
25:140–148 ECoG offers a rare and invaluable research
Shamir M, Sompolinsky H (2004) Nonlinear population opportunity for obtaining invasive electrophysi-
codes. Neural Comput 16(6):1105–1136
Shamir M, Sompolinsky H (2006) Implications of neuro- ological signals in awake, behaving humans.
nal diversity on population coding. Neural Comput Another avenue of research that is beginning to
18(8):1951–1986 be explored is cortical stimulation using ECoG
Sreenivasan S, Fiete I (2011) Grid cells generate an analog electrodes (e.g., Parvizi et al. (2013)). The elec-
error-correcting code for singularly precise neural
computation. Nat Neurosci 14(10):1330–1337 trode grids often cover much wider areas than the
Yarrow S, Challis E, Series P (2012) Fisher and shannon epileptogenic zone, allowing recordings from
information in finite neural populations. Neural healthy brain tissue. A postoperative CT scan
Comput 24(7):1740–1780 allows visualization of the electrodes and docu-
Zohary E, Shadlen MN, Newsome WT (1994) Correlated
neuronal discharge rate and its implications for mentation of their positions in relation to skull
psychophysical performance. Nature 370(6485): geometry (Fig. 1a). By co-registering the CT scan
140–143 with a preoperative MRI scan, the precise loca-
tions of the electrodes on the cortical surface can
be determined (Fig. 1b). Such co-registration
allows the comparison between ECoG signals
and functional magnetic resonance imaging
Electrical Nerve Block
(f MRI) data obtained in the same subject. As
with other types of brain field potentials, such as
▶ Peripheral Nerve Stimulation Technique,
local field potentials (LFPs), ECoG signals are
Nerve Block
a mixture of arrhythmic activity and brain oscil-
lations (Fig. 2a). In this entry, we will specifically
address computational modeling of arrhythmic
ECoG activity.
Electrocorticogram (ECoG)
Biyu J. He Detailed Description

National Institute of Neurological Disorders and
Stroke, National Institutes of Health, Bethesda, Brain Oscillations Versus Arrhythmic Brain
MD, USA Activity
Brain oscillations are recurring patterns of brain
activity that follow a particular temporal beat. For
Synonyms example, the first discovered EEG rhythm, the
occipital alpha wave, proceeds at roughly
Modeling intracranial electroencephalography 10 cycles per second (Berger 1929). Thus, brain
(iEEG) signals; Modeling the electrocorticogram oscillations (arrows in Fig. 2a) are most easily
Electrocorticogram (ECoG) 1071 E
Electrocorticogram (ECoG), Fig. 1 Spatial topogra- area based on median nerve somatosensory evoked poten-
phy of electrode coverage in an example patient. (a) tial (SSEP); yellow, hand sensory area based on SSEP;
X-ray showing electrode placements. (b) 3D rendering of blue, facial twitching in response to cortical stimulation;
anatomical MRI and projection of electrode locations onto green, hand grasp in response to cortical stimulation. L left
the 3D surface. Clinical mapping of the sensorimotor hemisphere (Adapted from He et al. (2008))
cortex is indicated by color patches. Red, hand motor
a Electrocorticography (ECoG) data in Human b ECoG signal power spectra

106 #1
#1
#2
#2 #3
102
Power (μV2)
#3
0 10
#1 10−2
#2
10−6
#3 Frequency (Hz)
0 10 0.01 0.1 1 10 100 (Hz)
Time (sec)
Electrocorticogram (ECoG), Fig. 2 Sample human Arrows point to examples of oscillations. (b) Power spec-
ECoG recordings and power spectra. (a) Raw activity tra for the three electrodes shown in a, averaged over
traces from three ECoG electrodes in a neurosurgical 83-min recording during the awake state. The spectra are
patient. Electrodes #1 and #2 were over the left frontal presented in log-log scale, under which a power-law dis-
cortex, electrode #3 over left temporal cortex. Data were tribution manifests as a roughly straight line
recorded using DC-coupled amplifier with a 500-Hz sam- (log(P) / b log(f)) (Adapted from He et al. (2010))
pling rate. Two 10-s-long segments are shown.
E 1072 Electrocorticogram (ECoG)
identified in the frequency domain, as their power et al. 2009). Similar power-law exponents
spectra contain a peak at the corresponding fre- were found in the power spectrum of neuronal
quency (Fig. 2b). There are a number of brain membrane potential fluctuations (Destexhe
oscillations at different frequencies, each with et al. 2003; El Boustani et al. 2009). In particular,
their own underlying mechanisms and cognitive two studies using human ECoG, focusing on the
functions (Buzsaki et al. 2013). relatively low-frequency (from <0.01 to 100 Hz)
At the same time, the power spectrum of brain (He et al. 2010) and relatively high-frequency
field potentials contains a predominant “1/f ” (10–500 Hz) (Miller et al. 2009) ranges, respec-
component; that is, power tends to fall off with tively, have reported strikingly similar exponents
increasing frequency following a power-law in the middle-frequency range: b = 2.44 in the
function: P / 1/f b, where P is power, f is fre- range of 1–100 Hz in the first study, and
quency, and b is a parameter (typically in the b = 2.46 in the range of 15–80 Hz in the second
region of 0–3) named the “power-law exponent” study. The close alignment of these numbers
(Fig. 2b). A power-law function is indicative of exemplifies the robustness of the scale-free
scale invariance, suggesting that no particular distribution.
time scale or frequency dominates the dynamics. In addition to this middle-frequency range,
Hence, the brain activity contributing to this 1/f Miller et al. (2009) found a transition to b 4
slope in the power spectrum is devoid of period- in the high-frequency range above 75 Hz. They
icity (i.e., being arrhythmic or “scale-free”). proposed a simple model that can explain the
Note that white noise (including Poisson firing power spectral shape, which utilizes the convo-
patterns) is a special case of arrhythmic activity, lution of two exponentially decaying functions
in which case b equals 0. A power-law distribu- representing, respectively, the postsynaptic cur-
tion of the power spectrum is characteristic of rent and the membrane leak. An exponentially
the temporal dynamics of brain activity at many decaying function in the time domain is charac-
different observational levels: It has been terized by a “Lorentzian” function in the fre-
described in the fluctuations of neuronal mem- quency domain of the form P constant for
brane potentials, LFP, ECoG, EEG and magne- f <<f0 and P / 1/f 2 for f >> f0, where f0 is
toencephalography (MEG) recordings, and the “knee” frequency. The “knee” frequency is
fMRI signals. In addition, the amplitude fluctu- directly related to the time constant t of the expo-
ations of narrowband brain oscillations in nential decay such that f0 = 1/(2pt). Recalling
EEG/MEG recordings exhibit prevalent scale- that convolution in the time domain is equivalent
free dynamics (Linkenkaer-Hansen et al. 2001). to multiplication in the frequency domain, the
A power-law distribution has also been reported resulted power spectrum from this model is thus
in the statistics of neurotransmitter release the multiplication of two Lorentzians, following
(Lowen et al. 1997) and neuronal firing (Lowen the form P constant for f << f1, P / 1/f 2 for
et al. 2001), where it has been more controver- f1 << f << f2, and P / 1/f 4 for f >> f2, where
sial. Lastly, fluctuations of human behavioral f1 and f2 are two “knee” frequencies determined
output such as reaction times, hit rate, and by the time constants.
force have often been found to exhibit a 1/f-like The locations of f1 and f2 can be determined
power spectrum as well (Gilden 2001; Kello empirically. While Miller et al. found f2 to be
et al. 2010). around 75 Hz, their data do not reveal the location
of f1 in the lower-frequency range. This informa-
Modeling the Arrhythmic ECoG Signals tion is provided in He et al. (2010), which found
Studies investigating the power spectrum of f1 to be around 1–2 Hz. Together, these results
ECoG and LFP signals have reported very simi- suggest the existence of two time constants: one
lar values for the power-law exponent, which is around 2–3 ms and another around 100 ms. At
typically in the range of 2–3 (Freeman and Zhai present, the origins of these time constants
2009; He et al. 2010; Miller et al. 2009; Milstein remain a speculation. One possibility is that the
Electrocorticogram (ECoG) 1073 E
2–3 ms time constant originates from synaptic play in generating the low-frequency ECoG sig-
current and the ~100 ms time constant from mem- nals. The effect of different network topology on
brane leak (Miller et al. 2009). An alternative the power spectrum is an active topic of current
possibility is that the ~100 ms time constant investigation. Two recent modeling studies found
comes from the slow NMDA synaptic current that network multistability caused by clustered
(Gibb and Colquhoun 1991; Wong and Wang connections or inhibitory neurons is conducive
2006) and the membrane time constant can be to producing slow fluctuations in population
as short as 2–3 ms during active synaptic activity activity (Ho et al. 2012; Litwin-Kumar and
(Koch et al. 1996). Adjudication between these Doiron 2012). Potential slow cellular mecha-
two scenarios will require future experiments, nisms include metabotropic glutamate receptors
E
such as blocking NMDA signaling and examin- (Zhang and Seguela 2010), endocannabinoid sig-
ing the resulting change in the power spectrum of naling (Carter and Wang 2007), and the cholin-
field potentials. In addition, the contribution of ergic pathway (Letzkus et al. 2011). The
dendritic filtering to the ECoG power spectrum contributions of these different mechanisms to
should be further explored (Dehghani et al. 2010; low-frequency ECoG signals should be investi-
Einevoll et al. 2013). gated by future experimental studies.
At the very low-frequency range, below
a “shoulder” of 0.1–1 Hz where power is rela-
tively flat, the power spectrum again follows
a form close to P / 1/f 2 (He et al. 2010). Poten- References
tial generative mechanisms for this
Berger H (1929) Uber das Elektroenkephalogramm des
low-frequency behavior have recently been
Menschen. Arch Psyhiatr Nervenkr 87:527–570
explored in a recurrent network model Brody CD, Romo R, Kepecs A (2003) Basic mechanisms
(Chaudhuri et al. 2014). It was found that for graded persistent activity: discrete attractors, con-
a recurrent network of rate-based nodes with lin- tinuous attractors, and dynamic representations. Curr
Opin Neurobiol 13:204–211
ear couplings and random connectivity, when
Buzsaki G, Logothetis N, Singer W (2013) Scaling brain
poised near criticality (i.e., with balanced excita- size, keeping timing: evolutionary preservation of
tion and inhibition), could reproduce the brain rhythms. Neuron 80:751–764
low-frequency behavior of the ECoG power Carter E, Wang XJ (2007) Cannabinoid-mediated disinhi-
bition and working memory: dynamical interplay of
spectrum. In such a network, the network recur-
multiple feedback mechanisms in a continuous attrac-
rent activity produces one or more very slow time tor model of prefrontal cortex. Cereb Cortex 17(Suppl
constants, whose corresponding “knee” frequen- 1):i16–i26
cies are below that so far investigated empirically Chaudhuri R, He BJ, Wang XJ (2014) The temporal struc-
ture of a random network near criticality and human
(0.003 Hz). Moreover, reducing the spatial cor-
ECoG dynamics. Cosyne, Salt Lake City
relation of the inputs to different nodes resulted in Dehghani N, Bedard C, Cash SS, Halgren E, Destexhe A
a flattening of the power spectrum, similar to that (2010) Comparative power spectral analysis of simul-
observed in ECoG and fMRI data during task taneous electroencephalographic and magnetoence-
phalographic recordings in humans suggests
performance (He 2011; He et al. 2010). Thus,
non-resistive extracellular media: EEG and MEG
this network naturally converts spatial correlation power spectra. J Comput Neurosci 29(3):405–421
into temporal correlation, providing support for Destexhe A, Rudolph M, Pare D (2003) The high-
a previous proposal that decoupling among neu- conductance state of neocortical neurons in vivo. Nat
ronal groups might underlie the reduction of Einevoll GT, Kayser C, Logothetis NK, Panzeri S (2013)
power-law exponent during task (He 2011). Modelling and analysis of local field potentials for
This low-frequency range of the ECoG activ- studying the function of cortical circuits. Nat Rev
ity bears significant resemblance to persistent Neurosci 14:770–785
El Boustani S, Marre O, Behuret S, Baudot P, Yger P,
neural activity (Brody et al. 2003; Major and
Bal T, Destexhe A, Fregnac Y (2009) Network-state
Tank 2004). As with persistent neural activity, modulation of power-law frequency-scaling in visual
both network and cellular mechanisms may be at cortical neurons. PLoS Comput Biol 5:e1000519
E 1074 Electroencephalogram, Relation to Unit Activity
Freeman WJ, Zhai J (2009) Simulated power spectral Wong KF, Wang XJ (2006) A recurrent network mecha-
density (PSD) of background electrocorticogram nism of time integration in perceptual decisions.
(ECoG). Cogn Neurodyn 3:97–103 J Neurosci 26:1314–1328
Gibb AJ, Colquhoun D (1991) Glutamate activation of Zhang Z, Seguela P (2010) Metabotropic induction of
a single NMDA receptor-channel produces a cluster persistent activity in layers II/III of anterior cingulate
of channel openings. Proc Biol Sci 243:39–45 cortex. Cereb Cortex 20:2948–2957
Gilden DL (2001) Cognitive emissions of 1/f noise.
Psychol Rev 108:33–56
He BJ (2011) Scale-free properties of the functional mag-
netic resonance imaging signal during rest and task. Electroencephalogram, Relation to
J Neurosci 31:13786–13795
He BJ, Snyder AZ, Zempel JM, Smyth MD, Raichle ME Unit Activity
(2008) Electrophysiological correlates of the brain’s
intrinsic large-scale functional architecture. Proc Natl ▶ Local Field Potential, Relationship to Unit
Acad Sci U S A 105:16039–16044 Activity
He BJ, Zempel JM, Snyder AZ, Raichle ME (2010) The
temporal structures and functional significance of
scale-free brain activity. Neuron 66:353–369
Ho EC, Struber M, Bartos M, Zhang L, Skinner FK
(2012) Inhibitory networks of fast-spiking interneu- Electron Microscopic Imaging
rons generate slow population activities due to excit-
atory fluctuations and network multistability. ▶ Imaging, Electron Microscopy
J Neurosci 32:9931–9946
Kello CT, Brown GD, Ferrer ICR, Holden JG,
Linkenkaer-Hansen K, Rhodes T, Van Orden GC
(2010) Scaling laws in cognitive sciences. Trends
Cogn Sci 14:223–232 Electron Microscopy
Koch C, Rapp M, Segev I (1996) A brief history of time
(constants). Cereb Cortex 6:93–101 ▶ Imaging, Electron Microscopy
Letzkus JJ, Wolff SB, Meyer EM, Tovote P, Courtin J,
Herry C, Luthi A (2011) A disinhibitory microcircuit
for associative fear learning in the auditory cortex.
Nature 480:331–335
Linkenkaer-Hansen K, Nikouline VV, Palva JM, Electronic Retinal Implants
Ilmoniemi RJ (2001) Long-range temporal correla-
tions and scaling behavior in human brain oscillations. ▶ Visual Prosthesis, Subretinal Devices
J Neurosci 21:1370–1377
Litwin-Kumar A, Doiron B (2012) Slow dynamics and
high variability in balanced cortical networks with
clustered connections. Nat Neurosci 15:1498–1505
Lowen SB, Cash SS, Poo M, Teich MC (1997) Quantal
Electrophysiological Indices of
neurotransmitter secretion rate exhibits fractal behav- Speech Processing
ior. J Neurosci 17:5666–5677
Lowen SB, Ozaki T, Kaplan E, Saleh BE, Teich MC Sonja A. Kotz
(2001) Fractal features of dark, maintained, and driven
School of Psychological Sciences, University of
neural discharges in the cat visual system. Methods
24:377–394 Manchester, Manchester, UK
Major G, Tank D (2004) Persistent neural activity: preva- Max Planck for Human Cognitive and Brain
lence and mechanisms. Curr Opin Neurobiol 14:675–684 Sciences, Leipzig, Germany
Miller KJ, Sorensen LB, Ojemann JG, den Nijs M (2009)
Power-law scaling in the brain surface electric poten-
tial. PLoS Comput Biol 5:e1000609
Milstein J, Mormann F, Fried I, Koch C (2009) Neuronal Some Core Facts About Event-Related
shot noise and Brownian 1/f 2 behavior in the local Brain Potentials (ERPs)
field potential. PLoS One 4:e4338
Parvizi J, Rangarajan V, Shirer WR, Desai N, Greicius
MD (2013) The will to persevere induced by electrical
Speech and its acoustic and linguistic properties
stimulation of the human cingulate gyrus. Neuron essentially evolve in time (Kotz and Schwartze
80:1359–1367 2010). Therefore, high-temporal-resolution methods
Electrophysiological Indices of Speech Processing 1075 E
such as the electroencephalography (EEG) or mag- how they are integrated with high temporal pre-
netoencephalography (MEG) are well suited to trace cision. Therefore, a number of exemplar ERP
the temporal unfolding of the speech signal. In par- studies that have explored the sensory-cognitive
ticular, ERPs embedded in the EEG permit speech interface in speech processing will be reviewed in
to be monitored with millisecond resolution. By the following.
time-locking and averaging ERPs to a large number A good way to test how early sensory and later
of specific and similar speech events, responses to cognitive processes in speech processing interact
acoustic and linguistic properties of these events can is to study speech in noise. In principle there are
tell us how an event is perceived and understood as it at least three sources of noise that can impact
occurs. The resulting wavelike pattern consists of an successful speech comprehension: noise (i) in
E
alteration of positive and negative peaks that, when the environment, (ii) in the hearing system, and
compared to a control condition, leads to the emer- (iii) in the speech signal itself. All three types of
gence of components that are defined by their polar- noise are considered to disrupt the acoustic/pho-
ity (positive or negative), by the delay after the onset netic interface and consequently successful
of an event of interest (latency), and by their distri- lexico-semantic integration (e.g., Boulenger
bution across the scalp (topography). For example, et al. 2011). A number of ERP investigations
the N400 component is a negative deflection (hence have therefore systematically manipulated
“N”) that peaks around 400 ms post-stimulus onset. speech signal quality to look at its consequences
Note that while the N400 has a relatively global on lexico-semantic integration and its impact on
scalp distribution in the auditory modality, we can- successful speech comprehension. In this
not relate cognitive processes tied to the N400 scalp research two ERP components feature promi-
topography to specific underlying neural sources in nently: the N100 and the N400.
the brain. The reason is that EEG activity is oriented
orthogonally to the sulcated cortex surface and not to
the skull surface. Consequently, ERP patterns Two Components of Interest: N100 and
recorded at the surface of the skull likely reflect the N400
summation of infinite large numbers of differently
oriented sources (generators). Early ERP components such as the N100 have
been linked to the physical properties of
a stimulus and are considered to reflect early
What Can ERPs Tell Us About Successful obligatory sensory processing (e.g., Luck 2005;
Speech Comprehension? Steinschneider and Dunn 2002). In audition the
N100 is considered to encode a sound’s onset
When listening to speech, our primary goal is to (e.g., Martin et al. 2008) and to index early per-
understand what is said. This necessitates the ception formation (e.g., N€a€at€anen 2001). It is also
integration of sensory, perceptual, and cognitive attuned to task demands in speech processing
processes to ensure successful speech compre- (e.g., Bonte et al. 2009; Obleser et al. 2004;
hension (Obleser and Kotz 2011). While there Poeppel et al. 1996 for N100m magnetoencepha-
has been ample research identifying the neural lographic evidence) and the degradation of the
sources of such integration in neutral (for sound signal (Miettinen et al. 2010).
a review, see Scott and Johnsrude 2003) and N400 research looks back onto a long tradition
emotional speech (for a review, see Schirmer and has tested semantic anomalies (whether
and Kotz 2006), there is less evidence from ERP a word fits into a sentence context or not; e.g.,
studies. This is surprising in light of the fact that Kutas and Hillyard 1980), semantic cloze proba-
ERPs monitor the unfolding speech signal from bility (how likely a word fits a sentence context;
the sensory to the cognitive stage at millisecond e.g., Kutas and Hillyard 1984), context manipu-
resolution and may thus give more precise insight lation (how expected a word is in a sentence
int

Encyclopedia of Computational Neuroscience

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Encyclopedia of Computational Neuroscience

Загружено:

Авторское право:

Доступные форматы

Dieter Jaeger

With 1109 Figures and 71 Tables

ISBN 978-1-4614-6674-1 ISBN 978-1-4614-6675-8 (eBook)

# Springer Science+Business Media New York 2015

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

This edition comprises 4 Volumes:

Computational Neuroscience has emerged in the last three decades as an

online SpringerReference enterprise, a hot-linked version of this encyclope-

James Bednar uses computational modeling to understand how visual cortex

Ulrik Beierholm’s research focuses on developing machine-learning-

Sliman J. Bensmaia is an assistant professor in the Department of Organis-

Kim T. Blackwell is a Professor in the Department of Molecular Neurosci-

approach interlaces experiments and model development on the network

Alla Borisyuk is an Associate Professor of Mathematics and an associate

Steven Bressler is a cognitive neuroscientist at the Center for Complex

Robert Butera is a Professor jointly appointed in the School of Electrical and

electrophysiology experiments, and general mechanisms of rhythm central

Carmen Canavier’s major research interest is the nonlinear dynamics of

Yoonsuck Choe is a professor and director of the Brain Networks Laboratory

Diego Contreras’ main focus of research is on how the nervous system

Sharon Crook holds a joint appointment between the School of Mathemat-

Gennady S. Cymbalyuk is an associate professor in the Neuroscience

Alain Destexhe is Research Director at the CNRS in the research campus of

Alexander Dimitrov’s main research interests involve the study of neural

Fabrizio Gabbiani is a Professor in the Department of Neuroscience at

Marc-Oliver Gewaltig is codirector of the Neurorobotics subproject in the

and is coauthor of the neural simulation tool NEST (www.nest-inititive.org).

Padraig Gleeson is a postdoc in the lab of Angus Silver at University College

Joshua Goldberg works on preclinical models of Parkinson’s disease and

Sonja Gr€ un’s research goal is to gain an understanding of the relevant

Christian Hauptmann works in the field of computational neuroscience,

University Hospital of Cologne, Germany. He has published more than

William Holmes is a Professor in the Department of Biological Sciences at

Dieter Jaeger is a Professor in the Department of Biology at Emory Univer-

Devin Jindrich is an Assistant Professor at the California State University,

Jeanette Hellgren Kotaleski’s main focus of research has been to use

more abstract system-level models down to kinetic models of subcellular

Michel Lemay obtained is Ph.D. in biomedical engineering from Case

Christiane Linster is a Professor in the Department of Neurobiology and

William Lytton is a Professor in Physiology, Pharmacology, and Neurology

Americo Migliaccio is a vestibular researcher with a broad range of skills,

(1) investigation of the physiological basis for the development of a human

John Milton’s research interest is in the interplay between sensory uncer-

Kendall F. Morris is a Professor in the Department of Molecular Pharma-

Farzan Nadim is a Professor of Neurobiology in the Federated Department

Theoden I. Netoff is an Associate Professor in the Department of Biomedical

Klaus Obermayer received his Ph.D. in 1992 from the Department of

Astrid Prinz is an Associate Professor in the Department of Biology at

Sylvie Renaud is a Professor at ENSEIRB-MATMECA at the University of

Jorge Riera’s scientific interest is to develop methods for the integration of

cells in the cortex, dynamics of neuronal masses and brain connectivity

Patrick Roberts is an affiliate Associate Professor in the Department of

Jonathan Rubin is a Professor of Mathematics at the University of Pitts-

Ilya A. Rybak received his M.S. in Electrical Engineering from Odessa

Emilio Salinas grew up in Mexico City. He obtained a B.Sc. degree in

correlated synaptic inputs on the responses of single model neurons. Emilio

Lars Schwabe aims at understanding cortical information processing in

he is currently applying core computer science concepts such as process

Frances Skinner is a Senior Scientist in the Division of Fundamental Neu-

Volker Steuber is a Reader (Associate Professor) in Biocomputation and

Peter Tass develops stimulation techniques with computational methods. He

Matthew Tresch is Associate Professor in Biomedical Engineering, Physical

Sharmila Venugopal is a Research Faculty in the Department of Integrative