Readings in

NEUROETHOLOGY

BioNB4240
BY CARL D. HOPKINS

Fall 2011

This material is copyrighted. It may not be reproduced under

any circumstances. Infringement of copyright is a violation of
civil law and may result in the imposition of penalties or even in
criminal prosecution.
 Preface
These readings are intended to supplement the textbook for an introductory undergraduate course in
Neuroethology, BioNB424 at Cornell University. The textbook for Fall semester, 2011 is Peter Simmons
and David Young (2010), Nerve Cells and Animal Behaviour 3rd ed, Cambridge University Press, New
York.. The book provides an introduction to a number of key model systems in Neuroethology but does
not cover all important fields. These readings are intended to add to this list and to provide additional
background in key subject areas that will be discussed during the semester. In addition, we have provided
reprints of articles intended for discussion sections.
Neuroethology is a sub-discipline of Neuroscience that focuses attention on the neural basis of natural
behavior of animals. In short, Neuroethology traces its origin to the science of Ethology, or Animal
Behavior. In this course we intend to concentrate on different aspects of the field: release mechanisms,
fixed action patterns, sensory processing in vision, audition, and electroreception, plasticity, and
communication. We shall focus on sensory worlds and sensory function highlighting echolocation in
bats, sound localization in owls, feature analysis in toad vision and frog hearing, and electroreception and
the jamming avoidance response of electric fish. In the final section of the course, we discuss
computational Neuroethology, with a discussion of neural modeling and neural networks.
Neuroethology is a rapidly changing field, still in its infancy as a scientific discipline, and very much in the
early stages of scientific development. There are few theories uniting the field. Instead, there are a
number of model systems that serve as examples of how neural systems are organized. In this course we
will consider a number of different approaches that are important to the integrative approach used in
Neuroethology: first, and foremost, we review the importance of good behavioral observation and
experimentation as a prelude to neural systems exploration. Second, we consider the evolution of the
behavior and the phylogeny of the behavioral and neural system being studied. Third, we explore the
neuroanatomy and neurophysiology of the system, and ask how the system functions physiologically.
Finally, we examine how the system can be modeled using computational methods. These methods are
but a few of the many approaches to the study of neuroetholgy.
It is my hope that you will enjoy the readings selected for this course, that they will provide examples of
excellent and stimulating work that has been done in the field, and that these will serve as a guide to this
rapidly evolving discipline.
Carl D. Hopkins
August, 2011
Ithaca, New York

Cover. The motion-sensing neuron, H1, in the lobula plate of the fly (above left), responds to movement of visual stimuli
from rear to front in the horizontal plane. The Hassenstein/Reichardt model of a generalized motion detector (upper right)
in the eye is based on correlation of two visual inputs separated in space. The figures below show the response of the H1
cell (lower left) and the model (lower right) to optomotor inputs at different rotational velocities. LP and HP refer to low pass
and high pass filters; M is the multiplier step. Adapted from:: Borst, A., V.L. Flanagin, and H. Sompolinsky, Adaptation
without parameter change: Dynamic gain control in motion detection. Proc Natl Acad Sci U S A, 2005. 102(17): p. 6172-6.
The motion sensing model for the visual system was an early example of computation in a neural circuit.

Fall 2011 BioNB 4240

C. D. Hopkins Cornell University
 Table of Contents

Greenspan, R. J. (2007) An Introduction to Nervous Systems. Cold Spring

Harbor Press, Cold Spring Harbor, NY. Chapter 7 ........................................................ 7-16

Kaas, J. H. and Catania, K. C. (2002) How do features of sensory

representations develop? BioEssays 24:334-343 ........................................................ 17-26

Catania, K. C. (1999) A nose that looks like a hand and acts like an eye: the
unusual mechanosensory system of the star-nosed mole. J Comp Physiol A 185:
367-372 ........................................................................................................................ 27-32

Bullock, T. H., Orkand, R., and Grinnell, A. (1977) Introduction to Nervous

Systems. W. H. Freeman & Co. San Francisco. Chapter 7 Integration at the
Intermediate Levels. pp. 242-290................................................................................. 33-82

Yuste, R. and Tank, D. W. (1996) Dendritic Integration in Mammalian Neurons a

Century after Cajal. Neuron 16: 701-716 ..................................................................... 83-98

Cator, L. J., Arthur, B. J., Harrington, L. C., and Hoy, R. R. (2009) Harmonic
Convergence in the Love Songs of the Dengue Vector Mosquito. Science 323:
1077-1079 .................................................................................................................. 99-102

Neuweiler, G. (2003) Evolutionary aspects of bat echolocation. J Comp Physiol A

189: 245-256 ............................................................................................................ 103-114

Hiryu, S., et al (2008) On-board telemetry of emitted sounds from free-flying bats:
compensation for velocity and distance stabilizes echo frequency and amplitude.
J Comp Physiol A 194: 841-851............................................................................... 115-126

Hiryu, S., Hagino, T., Riquimaroux, H., and Watanabe, Y. (2007) Echo-intensity
compensation in echolocating bats (Pipistrellus abramus) during flight measured
by a telemetry microphone. J. Acoust. Soc. Am. 121: 1749-1757 .......................... 127-136

Suga, N. (1990) Biosonar and Neural Computation in Bats. Scientific American

pp. 60-68 .................................................................................................................. 137-144

Konishi, M. (1993) Listening with Two Ears. Scientific American pp. 66-73 ............ 145-152

Harmening, W. M., Orlowski, J., Ben-Shahar, O., and Wagner, H. (2011) Overt
attention toward orented objects in free-viewing barn owls. PNAS 108: 8461-
8466 ......................................................................................................................... 153-158

Ewert, J. (1974) The Neural Basis of Visually Guided Behavior. Scientific

American pp. 34-43 .................................................................................................. 159-168

3
 Table of Contents

Kay, J. N., et al (2011) Retinal Ganglion Cells with Distinct Directional

Preferences Differ in Molecular Identity, Structure, and Central Projections. The
Journal of Neuroscience. 31: 7753-7762 ................................................................ 169-178

Heiligenberg, W. (1989) Coding and Processing of Electrosensory Information in

Gymnotiform Fish. J. Exp. Biol. 146: 255-275......................................................... 179-200

Xu-Friedman, M. A., and Hopkins, C. D. (1999) Central Mechanisms of Temporal

Analysis in the Knollenorgan Pathway of Mormyrid Electric Fish. The Journal of
Experimental Biology 202: 1311-1318 .................................................................... 201-208

Carlson, B. A., and Hopkins, C. D. (2004) Central control of electric signaling

behavior in the mormyrid Brienomyrus brachyistius: segregation of behavior-
specific inputs and the role of modifiable recurrent inhibition. The Journal of
Experimental Biology 207: 1073-1084 .................................................................... 209-220

Carlson, B. A., et al (2011) Brain Evolution Triggers Increased Diversification of

Electric Fishes. Science 332: 583-586 .................................................................... 221-225

Carlson, B. A., et al (2011) (Supporting Material) Brain Evolution Triggers

Increased Diversification of Electric Fishes. Science 332: 583-586........................ 226-246

Suthers, R. A. (2004) How birds sing and why it matters. In Natures Music: the
science of birdsong. Marler, P. and Slabbekoorn, H. eds. Elsevier. New York.
Chapter 9 pp. 272-295 ............................................................................................ 247-270

Fiete, I. R., and Seung, H. S. (2009) Birdsong Learning. In: Encyclopedia of

Neuroscience. Squire, L., Editor. Elsevier, New York pp. 227-239 ......................... 271-284

Nottebohm, F. (2005) The Neural Basis of Birdsong. PloS Biology 3: 759-761 ...... 285-288

Leonardo, A., and Konishi, M. (1999) Decrystallization of adult birdsong by

perturbation of auditory feedback. Nature 399: 466-470 ......................................... 289-294

Brainard, M. S. and Doupe, A. J. (2000) Interruption of basal ganglia-forebrain

circuit prevents plasticity of learned vocalizations. Nature 404: 762-766................ 295-300

Olberg, R. M. (1986) Identified target-selective visual interneurons descending

from the dragonfly brain. J Comp Physiol A 159: 827-840...................................... 301-314

Olberg, R. M. et al (2007) Eye movements and target fixation during dragonfly

prey-interception flights. J Comp Physiol A DOI 10.1007/s00359-007-0223-0 ....... 315-324

4
 Table of Contents

OMalley, D. M., Kao, Y., and Fetcho, J. R. (1996) Imaging the Functional
Organization of Zebrafish Hindbrain Segments during Escape Behaviors. Neuron
17: 1145-1155 ......................................................................................................... 325-336

Liu, K. S., and Fetcho, J. R. (1999) Laser Ablations Reveal Functional

Relationships of Segmental Hindbrain Neurons in Zebrafish. Neuron 23:
325-335 ................................................................................................................... 337-348

Kinkhabwala, A., et al (2011) A structural and functional ground plan for neurons
in the hindbrain of zebrafish. PNAS 108: 1164-1169 .............................................. 349-354

Koyama, M., et al (2011) Mapping a sensory-motor network onto a structural and

functional ground plan in the hindbrain. PNAS 108: 1170-1175 ............................. 355-360

Greenwood, A. K., Peichel, C. L., and Zottoli, S. J. (2010) Distinct startle

responses are associated with neuroanatomical differences in pufferfishes. The
Journal of Experimental Biology 213: 613-620 ....................................................... 361-368

Crapse, T. B., and Sommer, M. A. (2008) Corollary discharge across the animal
kingdom. Nature Reviews/Neuroscience 9: 587-600 .............................................. 369-382

Poulet, J. F. A. (2005) Corollary discharge inhibition and audition in the

stridulating cricket. J Comp Physiol A 979-986 ....................................................... 383-390

Poulet, J. F. A., and Hedwig, B. (2006) The Cellular Basis of a Corollary

Discharge. Science 311: 518-522 ........................................................................... 391-396

Cullen, K. E. (2004) Sensory signals during active versus passive movement.

Current Opinion in Neurobiology 14: 698-706 ......................................................... 397-406

Bell, C. C. (2001) Memory-based expectations in electrosensory systems.

Current Opinion in Neurobiology 11: 481-487 ......................................................... 407-414

Caporale, N., and Dan, Y. (2008) Spike Timing-Dependent Plasticity: A Hebbian

Learning Rule. Annu. Rev. Neurosci. 31: 25-46 ...................................................... 415-438

Requarth, T., and Sawtell, N. B. (2011) Neural mechanisms for filtering self-
generated sensory signals in cerebellum-like circuits. Current Opinion in
Neurobiology 21: 602-608 ....................................................................................... 439-446

Meitzen, J., and Thompson, C. K. (2008) Seasonal-like growth and regression of

the avian song control system: Neural and behavioral plasticity in adult male
Gambels white-crowned sparrows. Elsevier 157: 259-265 .................................... 447-454

5
 Table of Contents

Bell, C. C., Han, V., and Sawtell, N. B. (2008) Cerebellum-Like Structures and
Their Implications for Cerebellar Function. Annu. Rev. Neurosci. 31: 1-24 ............ 455-480

Land, M. F., and Collett, T. S. (1974) Chasing Behaviour of Houseflies (Fannia

canicularis). J. Comp Physiol. 89: 331-357 ............................................................. 481-508

Gegear, R. J., et al (2008) Cryptochrome mediates light-dependent

magnetosensitivity in Drosophila. Nature 454: 1014-1019..................................... 509-514

Borst, A., and Euler, T. (2011) Seeing Things in Motion: Models, Circuits, and
Mechanisms. Neuron 71: 974-994 .......................................................................... 515-536

Dombeck, D. A., et al (2010) Functional imaging of hippocampal place cells at

cellular resolution during virtual navigation. Nature 13: 1433-1443 ........................ 537-548

Harvey, C. D., et al (2009) Intracellular dynamics of hippocampal place cells

during virtual navigation. Nature 461: 941-949 ....................................................... 549-558

Barlow, H. B., and Levick, W. R. (1965) The Mechanism of Directionally Selective

Units in Rabbits Retina. J. Physiol 178: 477-504 ................................................... 559-586

Passaglia, C., et al (1997) Deciphering a neural code for vision. Proc. Natl. Acad.
Sci. 94: 12649-12654 .............................................................................................. 587-592

Passaglia, C. L. et al (1998) Cell Based Model of Limulus Lateral Eye. J

Neurophysiol 80: 1800-1815 ................................................................................... 593-610

6
Greenspan, R. J. (2007) An Introduction to Nervous Systems.
Cold Spring Harbor Press, Cold Spring Harbor, NY. Chapter 7

T E R

Love on the Fly

The wren goes to't, and the small gilded fly

Does lecher in my sight.
Let copulation thrive...
from King Lear by William Shakespeare

W hen a fly makes its way to a fallen, rotting peach, or to the rim of a vat of
crushed grapes, what is it looking for? Perhaps food, but adult flies are not big
eaters. They have already spent their carefree larval youth eating nonstop, day and
night. As adults, they may occasionally snack on yeast, but they have more important
business to attend to-finding a mate. So when a male fly heads for a rotten peach,
he is looking for females. When a female heads for that same peach, she is either
looking for males or a suitable place to lay her eggs once having mated. The next time
you see a bruised piece of fruit lying on the ground, be aware that you are looking
at an active singles scene.
In the world of fruit flies, females generally have the upper hand when it comes to
mate choice. Males must prove themselves worthy before a female will consent. Flies
are certainly not monogamous in the sense of forming some kind of lasting affiliation
between mates, but a female may mate only once or twice in her life, because she will
fertilize many eggs with sperm from a single mating. Therefore, the choice matters.

Recognizing a Potential Mate

Fallen pieces of fruit are social centers for many different creatures. Various species
of fruit fly may end up together on the same peach, but they will not make the mis-
take of mating with a fly of the wrong species. Males are not always so successful at
making such discriminations at first, and although a male Drosophila melanogaster
may initially try to court a female Drosophila virilis or Drosophila simulans, they gen-
erally will not get very far. Recognition is crucial and much of it is based on chemi-
cal sensing.

7
108 CHAPTER 7
L O V E O N THE FLY 109

Female fruit flies produce a mixture of secreted chemicals on the surface of their
external cuticle that acts as a perfume for males (Fig. 7.1). These pheromones are
attractive to males and all% them to determine if this other fruit fly is of the same
species and the opposite sex. Detection of pheromones takes place on chemoreceptor
sensory cells found on the male's forelegs, antennae, and mouthparts (Figs. 7.2 and
7.3). These sensory cells have chemoreceptor proteins and associated ion channels that
respond to pheromones. Mutant males that lack these proteins have a hard time court-
ing females in the dark, where chemosensory cues are essential. (In broad daylight, the
visual perception of a female arouses a male's interest enough to begin courting.)
The fly's chemoreceptors are of the same family as the opsin proteins in pho-
toreceptor cells (see Chapter 2). They are proteins with seven transmembrane seg-
ments, but instead of absorbing quanta of light, they bind pheromones. The binding
triggers a conformational change in the protein so that it interacts with a G protein
and initiates a cascade of enzymatic reactions. Ultimately, the cascade leads to the
opening of ion channels in the membrane of the chemoreceptor cells. Figure 7.2. Chemosensory organs on the fly's
If the male is in the right mood (i.e., if he is old enough but not too old, the head: antennae, maxillary palps, and labial palps.
humidity is right, and he has not been cooped up with other males-where he would
have been exposed to a high dose o f a a l e pheromones, some of which inhibit other

7,l 1-HD
Figure 7.1. Differences in pheromones between male (top) and female (bottom) fruit flies. (Left)
Chemical profiles from extracts of male and female cuticular hydrocarbons detected by gas chro- Figure 7.3. Chemoreceptor (arrows pointing to blue
matography. (Right) Chemical structures of principal peaks from profiles. Arrows indicate key dou- areas) differences between male (left) and female
ble bonds. (right) forelegs.

8
 LOVE O N THE F L Y 111

ly and usually follows by vibrating that wing to produce a courtship song). Only one
side of the bilaterally symmetrical brain needs to develop as male, and only the dorsal
part at that. Sexually mixed flies do not perform fully normal courtship, but they will
display these initial steps. (Some investigators have interpreted this finding as confir-
mation of their suspicion that males use very few neurons when pursuing females.)
Later steps in courtship (Fig. 7.5) require male development in other parts of the
nervous system: Production of the song requires male neurons in the thoracic gan-

Figurt ,.-I.
Chemosensory circuitry, necessary for male courtship, that differs between males (left)
and females (right). A representative sample of neurons (light green) is shown from antennae and
palps projecting into the antennal lobes (AL), whose neurons then project to the lateral protocere-
brum (LPR) and to the mushroom bodies (MB).

males from courting them), he will sHte up to a female and tap her abdomen with his
foreleg. The stimulation of his various chemoreceptors by the female's pheromones
will then (usually) trigger his performance of the fruit fly courtship ritual. The signals
from the male's chemoreceptors are relayed through the antennal lobes and into the
major lobes of the brain, the protocerebrum (Fig. 7.4). When neurons in the lateral
protocerebrum are sufficiently excited, the male begins to court vigorously. In males
in which these cells have been genetically engineered to be unresponsive, courtship
will not begin, and conversely, if these cells are genetically engineered to be more Dorsal
hyperexcitable, the male will start courting at the drop of a hat. ~rotocerebrum

Sex-specific Development in the Nervous System

Some of the circuitry for male courtship is the product of differences in neuronal
development in the male brain as compared to the female brain-differences that
are both physiological and anatomical. Many of the sex-specific brain regions have
been revealed in flies whose brains have been genetically engineered to develop as
part male and part female. For example, the fly's antennal lobes are grouped into
neuron clusters (glomeruli). Each glomerulus receives a unique combination of
synapses from the various types of olfactory receptor neurons. Several of these
glomeruli are specifically involved in pheromone recognition. When these glomeruli
are engineered to develop as female in an otherwise male brain, the flies fail to dis-
tinguish pheromonal differences between males and females. Tarsus
Surprisingly little of the brain needs to develop in male-like fashion for the trig- Figure 7.5. (Top) Steps in male courtship: orienting, tapping, wing vibration, licking, attempted cop-
gering of the first steps in courtship: orienting toward the female, tapping her ulation, and copulation. (Bottom) Portions of the nervous system that develop differently in males
abdomen, and engaging in wing display (in which the male extends one wing lateral- and females.

9
 112 CHAPTER 7

glion, licking of the female's genitals requires male brain tissue in the midbrain,
attempted copulation requires male neurons in the antenna1 lobes and the abdomi-
nal ganglion, and copulatbn requires additional male neurons in the abdominal gan-
glion. Much more of the circuitry for courtship is not sex-specific, however, but,
instead, simply makes use of circuits common to many behaviors. These common
circuits are present in both sexes and have been incorporated into the courtship rou-
tine in males. Courtship is likely to involve a substantial portion of the fly's nervous
system.
A gene called fruitless is essential for much of this male-specific development. Figure 7.7. The bl muscle used in the fly's courtship song.
The gene encodes multiple forms of a transcription factor, one of which is male spe-
cific. In males that are missing the male-specific product, courtship does not occur.
Conversely, when this male-specific product of the fruitless gene is expressed in a The female "hears" the song through her antennae. The tip of the antenna con-
developing female brain, the adult female will perform male-like courtship behavior tains a feather-like structure called the arista that is induced to vibrate by the sound
that is never otherwise seen in females. The male-specific product of the fruitless waves from the male courtship song. These vibrations stimulate stretch receptors in
gene is expressed in most of the brain regions shown to undergo sex-specific devel- the hinge at the base of the antenna, where it attaches to the head, and these stretch
opment for courtship (see Fig. 7.4). receptors convey impulses into the region of her brain required for mechanosensa-
tion (Fig. 7.8). The stretch receptor cells contain mechanosensitive channels in their
membranes, similar to those in jellyfish (Chapter 3) and microorganisms (Chapter 1).
The Love Sow of D. melunogaster
One of the major functions of courtship is for the male to advertise himself to a Flies Have Rhythm
prospective mate. The male does this by extending one wing and vibrating it to pro-
duce a courtship song. This song does not sound very musical to our ears (it resem- One of the ways that females recognize males of the various species is through the
bles a monotonous Morse code of rapid bumps), but it is music to the female (Fig. males' courtship song, and a key characteristic of the song that varies between
7.6). When recordings of simulated courtship song are played to potentially recep- species is its rhythmic timing. The interval between pulses (known as the interpulse
tive females, before placing them in the company of a male, the preexposure makes
them more receptive, that is, it takes them less time to begin copulating.
When flies that are part male and part female (see above) contain male neurons
in the dorsal brain, they will perform the initial stages of courtship and extend one
wing, but they will not produce a proper courtship song unless they also have the
necessary male neurons in their thoracic ganglion (see Fig. 7.5). The inference is that
they will extend their wing if a signal is sent from the brain down into the thorax (sim-
ilar to the signal sent to initiatk flight; see Chapter 6), but that proper control of the
wing's vibration requires fine-tuning from local neurons in the thoracic ganglion, as
is also seen in flight control. In fact, one of the flight-steering muscles ( b l ) beats in
time with the courtship song, firing once with every pulse (Fig. 7.7), and may also be
the trigger for each pulse.
'1
Johnston's
organ

Figure 7.6. Pulses of courtship song produced by male wing vibrations. Each pulse consists of two Figure 7.8. Stretch receptors (Johnston's organ) inside the fly's antenna. The arista is on the right
to three cycles separated by an interval of 32-38 msec. The fly's head would be on the left.
10
114 CHAPTER 7 LOVE O N THE FLY
IPI (msec) ,
per S
7 = 54
7 = 43
sec
sec
I Figure 7.10. Blocking action potentials interrupts the song-rhythm pacemaker. Transient suppression
2 4 6 of action potentials was achieved in a male fly whose neurons contain a temperature-sensitive
Time (min)
Figure 7.9. Song rhythms parallel circadian rhythms in mutants. Rhythmic oscillation of interpulse
interval (!PI) in the fly's courtship song in normal males (per') and period mutant males: short day
0 0
(pe?), long day @eS), and arrhythmic (per ). a e a s u r e s the period of the song rhythm (per has no
discernible rhythm). (Dashed lines) Sinusoidal waves that best fit the data.
interval, or IPI) averages 34 msec in Drosophila melanogaster. But closer inspection
reveals that this interval actually oscillates in a very regular way. Over the course of
approximately 55 seconds, the IPI will gradually increase to 38 msec and then grad-
ually decrease to 32 msec, following a regular, sinusoidal wave (Fig. 7.9).
The physiology underlying this rhythm is not well understood, but i t shows a
remarkable dependence on the same gene (period) that regulates circadian rhythms.
Males mutant for alleles of the period gene that produce short-day (19-hr), long-day
(28-hr), and arrhythmic circadian periods sing courtship songs with correspondingly
abnormal rhythms. The short-day flies sing with a 46-second rhythm, the long-day
flies sing with a 96-second rhythm, and the arrhythmic flies have no discernible
rhythm (Fig. 7.9).
Male flies generally sing their courtship song in bouts of 30 seconds to 1 minute at
a time. -When a male pauses in his singing, however, he continues to maintain the same
rhythm ko that when he resumes, the oscillation is right on the beat as if he had been
humming it all along. This humming-like behavior requires ongoing firing activity of
neurons, as shown by its ability to be interrupted and delayed when action potentials
are suppressed (Fig. 7.1 0). The suppression of action potentials is achieved by means
of a temperature-sensitive mutation in the voltage-gated sodium channel of neurons,
so that at normal temperature the neurons fire and at higher temperature they become
silent. (In the experiment, the fly's entire nervous system was mutant, so it is not known
which neurons were specifically responsible for maintaining the rhythm.)
115
P
Idelay = 39 sec
heat pulse
0 5 10 15
Time (min)
mutation in the voltage-sensitivesodium channel. (Solid blue line) Actual singing; (dashed line [long
dashes]) normal sinusoidal oscillation in the interpulse interval (IPI) up until the moment (blue dash
and inset) when a 30-sec pulse of high temperature is delivered. During the 30-sec pulse, singing
stops and no action potentials can occur. When the male resumes singing, several minutes later, the
period of the oscillation is normal (55sec), but the phase of oscillation of his song is delayed by 39
sec (shaded box in inset). (Dashed line [short dashes]) Phase-shifted sinusoidal oscillation due to
blockage of action potentials. When males not harboring temperature-sensitive sodium channels
are similarly treated, the temperature pulse does not delay the phase of their song. When they
resume singing, they are "in time," as if they have been keeping the beat all along.
Sequencing in Behavior
Orienting, tapping, wing extension, and singing the courtship song comprise.the first
steps in a larger sequence of courtship actions. If the female does not run away or
kick the male in his face, he will continue to engage in these early steps. Eventually,
if she permits it, he will go on to the next steps: licking the female's genitals, mount-
ing her, and copulating. These actions constitute a stereotypical behavioral sequence
that is usually performed in order. It is exceedingly rare for a male to skip the early
steps and immediately attempt to mount the female. However, when flies are genet-
ically engineered to be mutant for fruitless in a central part of the brain called the
median bundle, they violate the sequencing rule and skip directly to attempting cop-
ulation (Fig. 7.11). The experiment suggests that neurons in the median bundle of
males normally inhibit the performance of the late steps in courtship.
Motor Routines and Beyond
Each step in courtship consists of a routine analogous to those described earlier in
simpler organisms, such as swimming in the jellyfish (Chapter 3) or flying in the fly
11
116 CHAPTER 7 L O V E O N T H E FLY 117

next. Examples of this kind of analysis were presented in the context of geographical
differences in temperature compensation or light sensitivity due to the variations in
the period gene (Chapter 5).

Pheromone Evolution
The fly's chemical signature varies considerably from place to place. Within D.
melanogaster, the ratios of the different hydrocarbons on the cuticle surface (Fig. 7.1)
differ worldwide (Fig. 7.12), and also as locally as between adjacent vineyards near
Wiirzberg, Germany. Differences between species. can be even greater, but they do
not always reflect phylogenetic distance. In other words, D. melanogaster
lobe lobe
pheromones show greater divergence from those of the closely related species D.
Figure 7.1 1. The median bundle, the region of the fly brain that causes male flies to skip steps in simulans and Drosophila erecta than from those of the more distant species
the courtship sequence when mutant for fruitless.
Drosophila affinis, Drosophila elegans, and Drosophila serrata. Such deviations from
expectation are usually indicative of strong selective pressure, suggesting that
pheromonal signatures play a major role in speciation.
Fruit flies can be confused when mischievous humans paint them with foreign
(Chapter 6), but the physiological basis for these steps is far less understood. It is the
pheromones. Such treatment is sufficient to induce a Drosophila mauritiana male to
sequencing of routines such as these that makes behavioral repertoires elaborate and
take the unprecedented step of courting a D. melangoaster female, which he would
sophisticated. The stereotypical nature of the sequence is a product of natural selec-
never do otherwise. He will even court a dummy fly, if the dummy is painted with
tion and has been co-selected in the species along with the female's response to it.
the appropriate pheromones.
Understanding the neurobiological basis for the sequencing of behaviors represents
Within the D. melanogaster species, some of the geographically diverse strains
a major unanswered question, and its significance may reach far beyond the stereo-
will not court one another. In one case, their reproductive isolation correlates with
typical courtship of fruit flies. There is serious speculation that the origin of thinking
lies in the sequencing of motor routines. In this perspective, the evolutionary leap to
thought came with the ability to string together sequences in the brain such as those
controlling motor routines, but without the actual performance. Thus, a behavior as
apparently stereotypical as the fruit fly's courtship, with its orderly progression
through different actions and its capacity to maintain rhythmic time in the absence
of motor output, may have much to tell us about how brains accomplish more inno-
vative and creative tasks, such as thinking things through or telling a story.

Courtship Evolving
In the fly world, choosing the right partner is not a matter of whether the relationship
will last; one-night stands are the norm. Instead, mate choice-particularly, picking
the right species-means the difference between having offspring or not. As men-
tioned above, olfactory cues are one means of species discrimination and song is
another. Variations in these signals and in the response to them may provide raw
material for species divergence. In trying to infer evolutionary mechanisms, one
source of clues is to examine the divergence in gene function from one species to the

12
118 CHAPTER 7 L O V E O N THE FLY 119

the only D. simulans feature of their song. The average IPI remains unchanged, show-
variation in a gene for one of the enzymes involved in pheromone synthesis, desat2,
ing that it is not affected by the period gene. Because both species have the same 24-
that sets the position of the double bond (see Fig. 7.1). A variant form found in Africa
hour circadian rhythm, this too remains unchanged, indicating that selection on the
and the Caribbean produzs much less of the hydrocarbon 7,Il-HD that is other-
period gene has been confined to its song rhythm properties, allowing it to retain its
wise predominant in D. melanogaster females (Fig. 7.1). The allelic difference
24-hour timekeeping properties. It may be significant that the segment of the D. sim-
between this form of the desat2 gene and the more distributed form is a deletion of
ulans period gene that confers this song characteristic is approximately the same as
16 base pairs from the promoter region that prevents transcription of the mRNA for
the segment affecting temperature compensation among geographically dispersed
this enzyme.
strains of D. melanogaster (see Chapter 5). It is not unusual to find that some portions
To test directly for the role of this gene in reproductive isolation, flies were genet-
of a gene are freer to vary than others, presumably because the functions they affect
ically engineered so that the only difference between them was in the desat2 gene.
are also freer to vary.
These two laboratory strains showed the expected difference in hydrocarbon ratios
Song is not the only aspect of courtship affected by rhythms. Different species
and a mating preference for their own kind. That is, the pheromonal difference was
have distinct preferences for the time of day when they are most likely to mate. D.
capable of producing a certain degree of reproductive isolation, although not nearly
melanogaster tends to prefer late in the day and even seems to mate quite well in dark-
as complete as the original geographical strains' isolation. The incompleteness of
ness, whereas D. pseudoobscura has two different peak mating times, once around
effect is actually consistent with genetic mapping studies between the two geo-
dusk, a few hours later than D. melanogaster, and again later in the night (the rogues!).
graphical strains, showing that multiple genes on different chromosomes are collec-
Using a similar approach to the one described above, D. melanogaster males were
tively responsible for the full-blown isolation between the strains.
engineered to contain a period gene from D. pseudoobscura. As you might expect by
now, the new strain suddenly developed a predilection for mating rather later than
Sorrg Evolution their counterparts who carried an engineered D. melanogaster gene (Fig. 7.1 3).
Is a preference for time of day in mating important for segregating different
As a song characteristic, the rhythmic oscillation in IPI accounts for a substantial
species? The D. pseudoobscura-engineeredD. melanogaster flies offer a partial test
amount of the species specificity of the song. D. simulans, as its name implies, is very
case. They differ only in their period genes. If males and females of both types (nor-
closely related to D. melanogaster, so much so that they can form viable and some-
mal and D. pseudoobscura-engineered D. melanogaster) are mixed together, will the
times fertile offspring. This makes the need for correct recognition all the more press-
engineered flies prefer each other, and likewise, will the nonengineered flies prefer
ing (from an evolutionary standpoint). The average IPI of D. simulans is approxi-
each other? So it seems.
mately 50 msec, in contrast to 35 msec in D melanogaster, and the rhythmic
Because we already know that the period gene also affects courtship song, one
oscillation in D. simulans is 35 seconds, in contrast to 55 seconds in D. melanogaster.
might imagine that the observed mating preference was not related exclusively to
The importance of the two different parameters can be tested by playing synthetic
time of day. By removing the wings from both sets of males, however, any contribut-
courtship songs to virgin females of either species. These females are then placed
ing effect of courtship song as influenced by the period gene was ruled out. (Wing-
with males of the same species and timed for how much more quickly they mate
less males can still court successfully, although it takes them longer.)
than if no song had been played. When synthetic songs are played to D. simulans
What would happen if these two strains were left alone in the laboratory for hun-
females, their own species' rhythm is as effective in stimulating them to mate faster
dreds of generations? Their temporal separation could lead to the buildup of other
with D. simulans males as their species' average IPI. Likewise, D. melanogaster
differences between them by mutation, and an incipient speciation process could be
females care just as much about the rhythm of their own species' song as its average
initiated. Thus, even though they share the same culture bottles, sympatric speciation
IPI, so both have to be right to enhance their mating. Needless to say, a song that has
could occur.
the correct range of IPls, but is arrhythmic, is not very effective with these very dis-
Song variation among species goes beyond merely rhythm and IPI. Some species
criminating females.
differ in the notes (i.e., the pulses) themselves. D. melanogaster pulses have only two
Having already seen that mutations of the period gene in D. melanogaster alter
or three peaks that are of relatively small amplitude (Fig. 7.14). D. virilis song pulses
song rhythm as well as circadian rhythm, it stands to reason that variation in the peri-
have more peaks and they are of greater amplitude. A mutant in D. melanogaster
o d gene between D. simulans and D, melanogaster might affect the species-specific
produces an abnormal song whose pulses resemble those of D. virilis. The affected
song rhythm. This turns out to be true. When D. melanogaster males are genetically
gene is known as dissonance, and it encodes a protein involved in RNA processing.
engineered to have the D. simulans version of the period gene instead of their usual
(The dissonance gene is by no means dedicated exclusively to courtship song. The
version, they sing a song that has the rhythmic oscillation of D. simulans. But this is

13
120 CHAPTER 7 L O V E ON T H E F L Y 121

D. melanogaster with its own

period gene Normal
0.7 D.melanogaster
D. melanogaster

D. melanogaster
dissonance mutant

0.3~ D. melanogaster with period gene

from D. pseudoobscura
D. pseudoobscura

0.1J , , ,o.lJ , I , 4 o I
Figure 7.14. Courtship song pulses of normal Drosophila melanogaster (top), mutants of D.
0 3 6 9 12 15 18 21 0 3 6 9 12 15 18 21
melanogaster in the dissonance gene (middle), and normal Drosophila virilis (bottom). The disso-
I
Circadian time Circadian time nance mutants produce polycyclic pulses that resemble pulses in D. virilis.

Figure 7.13. Time of day for mating preference in Drosophila melanogaster and Drosophila
pseudoobscura (left) and genetically e n g i i r e d D. melanogaster (right). Flies were placed in con-
stant darkness so that their circadian rhythms would be governed by their internal molecular clock.
Graphs show proportion of pairs mating at various times of day. (Right) Strains are genetically engi-
neered D. rnelanogaster with either their same species' normal period gene (upper right) or the D.
pseudoobscura period gene (lower right). Genetically engineered flies carrying the D. pseudoobscu-
ra period gene show a peak of mating during the night that is absent in normal D. melanogaster and
oresent in normal D. pseudoobscura.
protein is expressed all over the nervous system, and other alleles of dissonance affect
visual sensitivity as well as song. The mechanisms by which it affects either song or
vision are not understood.)
Genetic engineering was brought to bear once again and the normal dissonance
gene from D. virilis was substituted into a strain of D. melanogaster in place of its own
gene. The song in these flies clearly resembles that of the dissonance mutant but not
that of the normal D. melanogaster (Fig. 7.14). Reminiscent of the period story, the
result suggests that selection for subtle variation in this gene may account at least par-
tially for the species difference in courtship song. The variation must necessarily be
subtle so that other functions of the gene, such as those required for visual sensitivi-
ty, are left intact.
Neuromodulators and "Instincts"
Behaviors such as fly courtship are sometimes referred to as "hard wired," reflecting
the idea that they are innate or instinctual. In fact, male flies raised in total isolation
from the egg onward are capable of performing the full repertoire of courtship
toward a female on their first try. Their behavior may not be quantitatively normal in
all respects, but it is qualitatively normal. That does not mean that courtship behav-
ior is entirely impervious to experience (see Chapter 8), but it suggests that, like fly-
ing, there has been selection pressure for fruit flies to "know" how to court as part
of their normal developmental maturation.
If experience is not essential, then what role, if any, do neuromodulators play in
a behavior like courtship? Does a male register the presence of a female with the
same internal alarm system (dopamine) that he uses when he begins tracking toward
a goal during flight (see Chapter 6)?Apparently it is not the same, because male flies
that are depleted of dopamine (after being consistently fed with an inhibitor of its
synthesis) will still court females. The salience of a female to a male is thus not depen-
dent on the same neuronal circuitry, a fact that may reflect the species' selection his-
tory. The same selection pressures that have produced males capable of courting at
first blush have also produced a perceptual response in them that does not need any
reinforcement from the neuromodulatory system.
The interplay between "hard-wired" and "plastic" behavior is a recurrent theme
in evolution. Most behaviors are a mixture of the two, whereas some are more heav-
ily weighted one way or the other. What is flexible in one species may be stereotyp-
ical in another, as seen in the previous discussion of sensitization in Aplysia and its
relatives (see Chapter 4). In all cases, the behavioral differences reflect variations in
the underlying neurons-in their anatomy, physiology, or both. Variation in genes
such as period or dissonance ultimately produce their behavioral effects by acting on

14
122 CHAPTER 7

one or more of these properties. At present, we still have only a faint picture of very
few steps in the mechanisms through which such genes affect the nervous system and
--.
alter behavior. .-

Further Reading
Greenspan R.J. and Ferveur 1.-F. 2000. Courtship in Drosophila. Annu. Rev. Cenet. 34: 205-232.
Manoli D.S., Meissner C.W., and Baker B.S. 2006. Blueprints for behavior: Genetic specification of
neural circuitry for innate behaviors. Trends Neurosci. 29: 444-451.
Peixoto A.A. 2002. Evolutionary behavioral genetics in Drosophila. Adv. Cenet. 47: 117-1 50.

15


16
Review articles
How do features of sensory
representations develop?
Jon H. Kaas1* and Kenneth C. Catania2
Summary
Sensory representations in the brainstem and cortex
have a number of features that support the idea that
neural activity patterns are important in their develop-
ment. Many of these features vary across species in ways
that could result from perturbances in the balance of the
effects of activity patterns and position-dependent gene
expression. (1) Most notably, disruptions or septa in
sensory maps often reflect actual discontinuities in the
receptor sheet, and the discontinuities may be reflected
in a series of interconnected maps. Species with different
disruption patterns in sensory sheets have different
matching disruption patterns in the sensory maps and
variant individuals and strains of the same species have
matching variations in the receptor disruption patterns
and their sensory maps. (2) In addition, mutations that
misdirect some of the retinal afferents from one side of
the brain to the other create new sensory maps that
preserve continuities in the altered pattern of input, while
creating new structural discontinuities. (3) Furthermore,
functionally different classes of afferents that are mixed
in the receptor sheet often segregate to activate separate
populations of target cells. (4) Finally, early developing
portions of receptor sheets may gain more than their
share of territory in sensory maps. These and other
variable features of sensory maps are most readily
accommodated by theories that involve roles for in-
struction by evoked and spontaneous neural activity
patterns. BioEssays 24:334343, 2002.
2002 Wiley Periodicals, Inc.
Introduction
The brains of mammals are characterized by systems of
interconnected subcortical and cortical representations of the
sensory surfaces of the body, eye and ear. In all mammals,
sensory representations occupy an important portion of the
brain,(1) but larger brains contain more sensory representa-
tions than small brains. Sensory representations are typically
isomorphs of the sensory surface in that neurons in different
1
Department of Psychology, Vanderbilt University, Nashville, TN.
2
Department of Biological Sciences, Vanderbilt University, Nashville,
TN.
*Correspondence to: Jon H. Kaas, Department of Psychology,
Vanderbilt University, 301 David K. Wilson Hall, 111 21st Avenue
South, Nashville, TN 37203. E-mail: Jon.Kaas@vanderbilt.edu
DOI 10.1002/bies.10076
Published online in Wiley InterScience (www.interscience.wiley.com).
334 BioEssays 24.4
parts of the representation are activated by stimuli in a pattern
that matches that of the sensory surface (see Fig. 1). The
matching order of representation is created and maintained by
matching patterns of connections of afferents from the re-
ceptor sheet to the brain, and from one brain structure to
another. Such representations or maps of sensory surfaces
have long been of interest to developmental neurobiologists
because an understanding of how order emerges in the
development of such maps and their connections may apply
broadly to other parts of the brain and provide a general under-
standing of the development of topographic patterns in brains.
There have been two general types of explanations for the
development of order in sensory maps. One, stemming from
the chemospecificity theory of Sperry,(2) is that molecular
gradients in two parts of the brain complement each other to
form patterns of molecular cues that guide the formation and
maintenance of orderly arrangements of connections. Pre-
sumably the molecular gradients would be the consequence of
position-dependent gene expression. The other type of ex-
planation is that patterns of neural activity are postulated to
shape important aspects of sensory representations. This is
founded on the findings of Hubel and Wiesel(3) that basic
features of sensory maps can be altered by experience during
development.
Most current investigators hold that the organizations of
subcortical and cortical maps are shaped by a combination of
both factors; that is, by some sort of patterning that is intrinsic
to cortical areas and subcortical nuclei, and by some sort of
self-organization(4) derived from patterns of induced or spon-
taneous neural activity.(5) Nevertheless, opinions vary on the
relative roles of these two modes of inducing organization,
and considerable uncertainty remains, especially in the face of
important recent findings that challenge traditional assump-
tions about the role of experience and neural activity.(6,7) Be-
cause of this uncertainty, it may be useful to review some of
the characteristics of sensory representations that appear to
be molded by these two factors. We draw special attention to
features of sensory maps that seem to require information
from the receptor sheet.
Sensory maps vary in internal order
Each sensory system is characterized by topographic maps of
the receptor array. Thus, visual maps are retinotopic, somato-
sensory maps are somatotopic, and auditory maps are tono-
BioEssays 24:334343, 2002 Wiely Periodicals, Inc.
17

Figure 1. A simplified schematic of how axon projections
from the receptor distribution on a sensory surface terminate
in a topographic pattern in brainstem nuclei, and how this
order is maintained over several steps of processing by simi-
lar topographic patterns of connections between subsequent
nuclei and cortical areas. A longstanding question in neuro-
science is how such order is created in development. Position-
dependent molecular signals are undoubtedly important in
forming the basic patterns, but here we argue that neural
activity patterns shape important details of such patterns.
topic. This means that the central representations reflect
the spatial arrangement of receptors across the retina, the
contralateral body surface, or along the length of the cochlea,
the auditory receptor organ. Yet, within this general frame-
work, considerable variability exists both for sensory represen-
tations across species and for different representations in a
system within a species. This is most obvious in the somato-
sensory system where a number of different ways of represent-
ing the contralateral body surface have been described. The
trunk of the body, for instance, is generally represented in a
continuous manner in a part of the primary somatosensory
cortex (S1 or area 3b) that is between the representations
of the forelimb and hindlimb.(8) However, in tree shrews (a
small mammal related to primates), only the ventral trunk is
represented in the expected location lateral to the hindlimb
and the dorsal trunk is represented in cortex medial to the
hindlimb.(9) Within primates, the continuous representation of
the trunk from dorsal to ventral appears to have opposite
orientations in S1 of different species of monkeys.(10) In many
monkeys, the arm is represented in cortex medial to cortex for
the hand. However, in prosimian galagos the anterior arm is
represented lateral to the hand while the posterior arm is
medial to the hand.(11) Subcortically in the cuneate nucleus of
Review articles
the brainstem, which receives direct cutaneous afferents
from the forelimb, digits are represented dorsally in squirrel
monkeys and most studied mammals, but ventrally in macaque
monkeys.(12) In S1 of bats, the representation of the digits of
the forelimb is reversed in orientation from that in S1 of other
studied mammals.(13) While somatosensory areas share
many broad topographic features in common, such as a medial
cortical representation of the tail and hindlimb and a lateral
representations of the face, the details of the internal order
vary widely across even closely related species.
Visual representations also vary in retinotopic organization.
We commonly think of retinotopic maps as having a conti-
nuous representation of the contralateral half of the visual field
via a superimposition of the inputs from the two matching hemi-
retinas. Yet, clear variations of this prevalent pattern exist.
Most notably, the superior colliculus of the visual midbrain
represents the complete retina of the contralateral eye in most
mammals, while the superior colliculus of all studied primates
represents only the contralateral hemiretina.(14 16) Even in
visual cortex where representation of the contralateral visual
hemifield seems to be the rule, some partial representation of
the ipsilateral visual hemifield near the zero vertical meridian
does occur, and the extent of this representation appears to
be variable across various species.(17 19) In extrastriate visual
cortex, representations include both continuous maps of the
contralateral hemifield and maps in which the representation is
partially split in various ways. The split in the map is commonly
along the representation of the zero horizontal meridian so that
the representation of the upper and lower visual quadrants are
largely separate from each other,(20) but the split may instead
be along a line well into the lower visual quadrant in some
maps.(21,22) A split in a representation allows one visual area to
wrap around another to form a retinotopically matched border,
and shorter interconnections between the fields may result,
but it is not clear why the location of the split is so variable.
These examples demonstrate the extensive variation in the
internal organization of sensory representations and it is not
obvious that any of these rearrangements have important
functional consequences. Yet, their existence does suggest
that small alterations in the balance of developmental mechan-
isms can have major consequences. While the basic patterns
of internal organization in at least primary sensory maps
appears to be generated by gradients of signaling proteins,(23)
activity patterns related to species differences in the number
and arrangement of sensory afferents could be a significant
source of the alterations in these patterns.
The modular and laminar subdivisions of
representations are variable
Nuclei and cortical areas are commonly subdivided into layers
and modular patchworks of two or more functionally distinct
classes of neurons. The ways nuclei and areas of different
mammals vary in laminar and modular patterns has not been
BioEssays 24.4 335
18
Review articles

extensively studied, but there is clear evidence for much

variability.(24,25) For instance, the mammalian retina has two
basic classes of ganglion cells, one that responds to an
increase in illumination in the receptive field center (ON-
center) and one that responds to a decrease (OFF-center).
The projections of these two classes of ganglion cells usually
have mixed terminations in the same layers of the lateral
geniculate nucleus of the thalamus and in the subsequent
relay to layers of visual cortex, but not in all mammals. In tree
shrews, for example, ON and OFF ganglion cells project to
different geniculate layers,(26) and these geniculate layers
relay to different sublayers in visual cortex.(27) Even closely
related species may differ in how they segregate and merge
ON and OFF inputs. Cats have overlapping gradients of ON
and OFF cell inputs in the same geniculate layers,(28) while
ferrets and mink have separate geniculate layers for ON and
OFF inputs.(29,30) Species differences in the laminar segrega-
tion of other types of retinal ganglion cell projections (``X'' and
``Y'' cells in cats and primates) also exist.(31) These two types
(X and Y) are mixed in the same layers in cats and segregated
in the ``P'' and ``M'' layers in primates. In some species of
monkeys, inputs to layer 4 of primary visual cortex are highly
segregated within eye-specific ocular dominance columns
(groups of neurons activated exclusively or mainly by the right
or left eye), but this is not true for other species of monkeys Figure 2. The barrel field of primary somatosensory cortex
of rats. Each whisker on the side of the face (a) activates a
where only a weak or transient segregation occurs, and no
specific cluster of cells or ``barrel'' in cortex (b). A portion of a
segregation is apparent in most other mammals.(32,33) Yet, brain section through primary somatosensory cortex has been
such segregations are apparent in cats(34) and several other cut parallel to the surface to reveal the barrels as ovals of
species. Variable modular arrangements are also common in tissue densely expressing the metabolic enzyme, cytochrome
the second visual area (V2) of tree shrews, monkeys and oxidase (CO). The ovals are separated by CO light septa. See
Woolsey et al.(39) for further details.
squirrels.(35)
Similar variations in the segregation of the inputs from
different receptor classes are found in the somatosensory
system. For example the duck-billed platypus has both (Fig. 2). More recently, it has become apparent that every
mechanoreceptors and electroreceptors in its bill and the distinct part of the rat's body, for instance the digits and pads of
cortical representation of the bill is subdivided into sets of the hands and feet, can be seen as a highly metabolically
alternating tactile or electroreceptive modules.(36) In mon- active island of densely packed cells separated by metaboli-
keys(37) and cats,(38) slowly adapting (SA) and rapidly adapting cally weak, cell-poor septa.(40) When the pattern of projections
(RA) cutaneous afferents are represented in different modules from the somatosensory thalamus to cortex is considered,
or columns in primary somatosensory cortex, but in rats and these islands receive inputs from matching subdivisions of the
most other mammals no such segregation is apparent. How is ventroposterior nucleus, while the septa are devoid of such
this variability generated during development? A variable role inputs.(41) An additional matching pattern of modules exists in
for activity-based segregation is certainly suggested by the the brainstem where afferents from the face and other parts of
common isolation of inputs by response type. the body are segregated by body part.(42) Thus, sensory inputs
related to each whisker, digit, or pad of the foot or forepaw are
Disruptions in the sensory sheet are reflected isolated from each other in anatomically distinct modules in the
by septa in sensory representations brainstem, thalamus, and cortex. Such architectonically identi-
Sensory representations commonly have visible septa be- fiable groups of cells, ``barrels'' or ``barreloids'', have been vari-
tween groups of cells that are activated by separated arrays of ably identified in several other rodents, some marsupials, and
peripheral receptors. Perhaps the ``barrel field'' of somatosen- rabbits. Woolsey et al.(38) concluded from such observations
sory cortex of rats and mice is most well known.(39) In these that the organization of the barrel field of cortex in these
rodents, a distinct barrel-like structure in primary somatosen- mammals is ``dictated by the organization of the sensory
sory cortex (S1) can be identified for each whisker of the face periphery.''

336 BioEssays 24.4

19

Figure 3. The brain of the star-nosed mole ismorphically
represents the unique fleshy appendages of the tactile nose.
A: A face-on view of the mole with 11 rays on each side of the
nose. They detect and locate food by repeatedly contacting
the ground with their nose. B: A rotated view of the left half of
the nose. Ray 11 is just over the mouth. Scale bar 1 mm. C:
The cytochrome oxidase (CO) dense bands that represent
each ray of the nose in S1. Light, CO-poor septa separate the
representations of the rays. A similar array of bands for each
ray is apparent in the second somatosensory area, S2, and a
third less distinct pattern is found in a third somatosensory
area. Scale bar 500 mm. See Catania and Kaas(44) for
review.
Comparable subdivisions in the brain exist for each of the
11 fleshly appendages on each side of the face of the star-
nosed mole (Fig. 3).(43 45) Cell-dense bands of neurons for
each appendage (or ray) of the nose are separated from each
other by narrow septa in each of three different cortical re-
Review articles
presentations of the face, as well as in subcortical represen-
tations in the thalamus and brainstem. The representations of
the digits of the forepaw in these moles are also separated by
septa (Fig. 4). In monkeys, narrow septa separate the re-
presentations of digits of the hand, and a wider septum sepa-
rates the representation of the hand from that of the face; other
septa separate areas of cortex devoted to the upper lip, lower
lip and chin, teeth, and tongue.(46,47)
The common finding in rats, moles, monkeys and other
mammals is that disruptions in the receptor sheet resulting
from protuberances like digits, folds between pads of the palm,
and separations in the skin between the lips and between the
hand and the face (when the hand is represented next to the
face) are reflected in representations by narrow septa between
homogenous regions of neural tissue that are activated by
continuous parts of the receptor sheet. These separating septa
are variably visible in somatosensory nuclei and cortical areas,
according to the species, the level in the somatosensory sys-
tem and the body features.
Such subdivisions are unknown in the auditory system
where the receptor surface for each ear comprises a single
continuous row of hair cells. The visual system, however, has a
small disruption of the retina, the optic disc or nerve head,
where axons exit from the center of the retina. This receptor-
free portion of the retina is commonly reflected as a rod-like
septum through the layers of the lateral geniculate nucleus
with inputs from the contralateral hemiretina, which is the
hemiretina that contains the optic disc.(48) In ground squirrels,
the optic disc is horizontally elongated across much of the
retina. Presumably, this reduces the impact of having a blind
spot, as predators are less likely to be masked by a strip than
an oval. This elongated horizontal optic disc is represented by
a narrow septum across much of the primary visual cortex.(49)
The varying patterns of lamination in the lateral geniculate
nucleus are more well known. Septa commonly separate
layers devoted to inputs from each eye, but they also often
separate classes of ganglion cells.(48) Why do such septa
occur, and why are they variably apparent? When they occur,
they always separate groups of neurons with discorrelated
activities as a result of receiving inputs from separated loca-
tions on the receptor sheet or from receptors with different
transduction properties.
Errors and alterations in the receptor sheet
are transmitted through the system to cortex
Typically, the pattern of sensory input from a receptor sheet
is very consistent across individuals within a species. How-
ever, gene mutations may alter this pattern so that strains of
species differ, and individuals may also occasionally vary
from the population in this regard for unknown reasons. In an
innovative study, Van der Loos and Dorfl(50) took advantage
of the rare occasion of finding mice with an atypical number of
vibrissae on their muzzle. As noted above, normal mice have
BioEssays 24.4 337
20
Review articles
a morphologically distinct structure, the ``barrel,'' in cortex
for each whisker on the side of the face (Fig. 2), as well as
equivalent ``barreloids'' for each whisker in the thalamus and
brainstem. When they examined those individuals with one
or two extra or less whiskers, they found the same changes
in the number of barrels in cortex and barreloids at each
subcortical level.(51) Similarly, when star-nosed moles have
an extra appendage, or one less, on their tactile nose, a series
of separate bands in cortex always matches the number of
appendages (Fig. 5) .(52) Van der Loos and Dorfle(50) con-
cluded from their observations on mice that the cortical
array of barrels is ``slaved'' to the peripheral array of whiskers
``as a result of a cascaded induction over three synaptic
stations'' (brain stem, thalamus, and primary somatosen-
sory cortex). In star-nosed moles, an altered pattern of bands
is also obvious in the second somatosensory area, S2, so
the cascade of induction occurs over at least four levels of
processing.
To us, the most parsimonious interpretation of the matching
changes in the number of whiskers or rays on the nose and the
number of related neural structure at three or four levels of
338 BioEssays 24.4
Figure 4. The modular representation of the digits of the
forepaw of the star-nosed mole in the second somato-
sensory area of cortex (S2). A. The ventral glabrous surface
of the forepaw with the digits from the ``thumb'' numbered
D1D5. B. In cortex, bands of cytochrome oxidase (CO)
dense tissue that are separated by narrow CO light septa
are activated by touching each digit, as numbered D1D5.
The palm activates a more caudal portion of the hand
representation without septa. Rostral (R) and dorsal (D) on
the brain are indicated. Based on Catania and Kaas.(43)
processing in the somatosensory systems is that a genetic
change or environmental event during development altered
the number of whiskers or rays on the face. Information
about this change, most likely in the form of a neural activity
pattern, traveled from the face to the brainstem, then the
thalamus, and then to one or two levels of cortex, altering the
development at each level to produce a matching pattern.
Similar alterations occur in the visual system. Siamese cats
result from a mutation that changes coat color and eye color by
reducing pigmentation in parts of the body with normal body
temperatures. The reduced pigment in the retina has the un-
explained developmental consequence of misdirecting most
of the axons from the temporal half of the retina that normally
project to the ipsilateral lateral geniculate nucleus to the con-
tralateral lateral geniculate nucleus (Fig. 6).(53) This causes
the end of the A1 layer, now abnormally innervated from the
contralateral eye, to fuse with the end of the A layer with normal
inputs from the contralateral eye. This results in a continuous
representation that has been enlarged by including an extra
208 of contralateral retina.(54) The enlarged representation
is preserved in a single enlarged pattern in the projection to
21

cortex in some Siamese cats,(55) but in others the A1 layer
inputs to cortex are suppressed.(56)
Related alterations in the extent of geniculate and corti-
cal maps of the retina have been reported for two different
mutant phenotypes of Belgian sheepdogs. In one type, most
of the retinal ganglion cells of both eyes project to the lateral
geniculate nucleus of one hemisphere and not the other.(57)
In the other phenotype, retinal axons of each eye project only
to the ipsilateral hemisphere.(58) Nevertheless, topological
maps of the altered inputs were found in visual cortex. Thus,
when the pattern of inputs from the retina is abnormal, the
organizations of the relay nuclei in the thalamus and their
sensory areas in the cortex adjust to accommodate the
abnormal pattern. Such adjustments strongly suggest a role
for activity patterns that are altered with the misdirected
inputs.
Some types of modular organization have
evolved more than once
Primary visual cortex in cats and monkeys is characterized
by a pattern of alternating ocular dominance bands (OD)
or columns, and a distribution of dot-like modules that are
Review articles
Figure 5. Star-nosed moles with extra or fewer
rays on the nose have matching alterations in the
numbers of isomorphic structures in brain stem,
thalamus and cortex. Based on Catania and
Kaas(52) and subsequent, unpublished research.
marked by high concentrations of the metabolic enzyme,
cytochrome oxidase (the CO-blobs). However, ocular dom-
inance (OD) bands and CO-blobs are not found in most
mammals.(32,33,59,60) Such distributions lead to the conclusion
that the common ancestors of cats and primates had neither
CO-blobs or OD bands in cortex.(16,25) Thus, these highly
similar patterns of modular organization evolved independen-
tly. In addition, alternating OD bands have even been created
in the optic tectum of frogs, normally with input from only
the contralateral eye, by artificially providing an extra eye to
equally innervate the structure.(5) Other examples include the
independent evolution of the SA and RA modules in S1 of cats
and monkeys. The independent evolution of similar types of
modules in sensory nuclei and areas likely reflects similar
expressions of basic self-organizing factors that have the
potential to be revealed in all sensory representations. Com-
puter models indicate that alternating bands can result from
activity-based competition that is balanced, while an im-
balance creates a dot (blob) and surround pattern.(61) The
variability that exists across species in the expression of
activity-based modules may relate to factors that alter the
period of susceptibility.(62)
BioEssays 24.4 339
22
Review articles
Figure 6. Retinal projections and lateral geniculate organi-
zation in normal and Siamese cats. A: In normal cats, inputs
from the nasal hemiretina of the contralateral eye (above
right) terminate in a retinotopic sequence in the dorsal ``A''
layer of the lateral geniculate nucleus (LGN), while the inputs
form the temporal hemiretina of the ipsilateral eye (above left)
terminate in the ``A1'' layer. The thin part of the arrows in the
right eye and the A layer corresponds to the monocular visual
field. The A and A1 layers are isolated from each other by a
narrow cell-poor septum. B: In Siamese cats, about 208 of the
temporal hemiretina projects abnormally to the contralateral
LGN instead of the ipsilateral LGN. These misdirected inputs
still terminate in the correct layer for the temporal retina (the
A1 layer), but the end of the A layer now fuses with the end of
the A1 layer, as these ends are now activated by adjoining
portions of the contralateral retina. However, the A1 layer also
develops an abnormal septum separating a medial segment
of the layer with input from the contralateral eye from a lateral
segment of the layer with input from the ipsilateral eye. This
altered pattern indicates that the septa are not based on
factors intrinsic to the LGN but depend instead on the nature
of the input. However, misdirected inputs go to the correct
layer, and this likely depends on intrinsic guidance factors.
Based on Guillery and Kaas.(54)
Early maturing parts of receptor sheets
capture more representational space
In general, the proportions of sensory representations that are
devoted to different parts of receptor arrays seem to reflect the
varying densities of receptors in the arrays.(63,64) Thus, finger
tips with high numbers of receptors activate more cortical
space than equal sectors of the backs of the digits with far
340 BioEssays 24.4
fewer receptors. Such proportional activation could result from
afferents competing for equal space as they grow into brain-
stem nuclei, and similar competitions by afferents in each relay
to cortex. The time of development might be an additional
factor. The central or foveal region of the retina is the first to
develop, and the ganglion cells of the fovea activate a greater
proportion of primary visual cortex than would be predicted by
their number.(65) In a similar manner, the tactile appendage or
ray of the nose of star-nosed moles that is just over the mouth
(ray 11 of each side of the face) develops sooner than the other
rays and captures more of somatosensory cortex than predict-
ed by the number of afferents from the ray.(66) Early maturing
receptors and afferents could have an advantage in activity-
based competition for representational territory,(67) or this
activity could even promote local brain growth.(68)
Adjoining sensory representations have
matching borders
Typically, cortical representations are precisely matched
along their common borders so that receptive fields overlap
for adjoining neurons on each side of the border. Such
congruent borders(20) are clearly apparent for S1 and S2 of
somatosensory cortex of the star-nosed mole, as the two
cortical isomorphs of the rays of the nose are precisely aligned
(Fig. 5). Other examples include the borders between the four
somatosensory representations (areas 3a, 3b, 1 and 2) of
anterior parietal cortex of primates, the border between visual
areas 1 and 2 , and the borders between core auditory areas in
primates. Such precise alignments suggest that some
organizing factor (possibly activity) is communicated across
the border during development.
Experimental changes in sensory activation
and innervation patterns alter the
course of development
A great deal of evidence has accumulated about how sensory
loss or deprivation alters the course of the development of
sensory systems, and many of these changes have been used
as evidence that neural activity patterns play an important role
in shaping the detailed organization of sensory representa-
tions.(69) For example, the two sets of ocular dominance bands
in primary visual cortex of monkeys and cats develop un-
equally if one eye is deprived of normal vision. As a result,
the bands are narrower for the deprived eye. Impressive
recent evidence for an instructive role of neural activity in
brain development also comes from experiments where
selectively blocking the activity of retinal ganglion cells
prevents the development of orientation selective cells in
visual cortex.(70,71)
Sensory loss of one type may even result in the deprived
brain regions being activated by other types of sensory inputs.
In the anterior ectosylvian region of cortex in cats, visual
deprivation early in development leads to an expansion of
23

neighboring auditory and somatosensory areas into the depriv-
ed visual region.(72)
Quite dramatic alterations in activity patterns have been
produced by allowing retinal axons to grow into the auditory
thalamus after removing auditory inputs in developing mam-
mals so that auditory cortex becomes activated by visual
stimuli.(73) This manipulation changes the course of develop-
ment of auditory cortex to create neurons with orientated visual
receptive fields that resemble those normally found in visual
cortex. Thus, auditory cortex has the potential to create some
of the basic features of a visual representation when it is in-
duced to receive visual information. These and related results
seem to indicate that alterations in the nature of activity patterns
in the sensory inputs from the thalamus have an organizing
influence on cortex.
Related experiments also show that visual maps can be
induced to form in non-visual cortex. When the portion of
posterior neocortex that normally develops into visual cor-
tex is largely or completely removed before innervation by
the visual thalamus, the visual inputs from the thalamus
grow instead to remaining portions of cortex to form visual
maps.(74)
Implications of normal and induced features of
sensory maps for theories of map formation
Many of the features of sensory maps that we have outlined
here are most easily explained by the assumption that informa-
tion is transported from the receptor sheet to sequentially
inform and guide the development of each subsequent station
in the processing sequence. This is possible since representa-
tions do develop according to position in the processing se-
quence. While it seems possible that factors intrinsic to an area
of cortex could code for all the major organizational features
of a sensory representation, it seems more likely that major
aspects of internal organization are determined by the nature
of the sensory input. For example, in mice with an extra whisker
on the face or moles with an extra ray on the nose, this change
in the parcellation of afferents results in matching changes in at
least three or four sensory maps in the brain. It seems extre-
mely unlikely that a number of genes have mutated in consort
so that independent intrinsic guidance factors in the face and at
each of several levels of the nervous system have all been
changed to precisely match modular components in a single or
a few generations. Instead, it seems more reasonable to pro-
pose that information from the sensory sheet instructs the
formation of central sensory representations in a cascading
fashion. The nature of this instruction could be chemical, as
neurons do effectively transport substances, but neurons are
specialized to send information in discharge patterns. Neural
activity patterns have the potential to alter gene expression
and other cellular processes that mediate neural growth and
synapse formation, as well as axon retractions and synaptic
loss.(75)
Review articles
We have argued here and elsewhere(49) that neural activity
patterns likely create the detailed order within representations,
the general proportions devoted to sensory surfaces, the varia-
tions in representational order, the septa between adjoining
representations of separate regions in the receptor sheet,
various types of modules within areas, and possibly congruent
borders between representations. The commonly expressed
view is that ``neurons that fire together wire together''.(76,77)
According to this model of development, neurons that are acti-
vated at the same time maintain and enlarge shared connec-
tions, while neurons that are activated at slightly different
times reduce or lose shared connections, apparently to the
extent that isolating septa sometimes form between differently
activated populations of neurons. Co-activations can depend
not only on sensory stimuli and on the spatial arrangement of
receptors, but also on the transduction and response proper-
ties of receptors and neurons. The spontaneous activity within
brain structures can also have an important role as adjacent
neurons respond together.
Conclusions
Position-dependent differences in gene expression and the re-
sulting gradients of signaling molecules appear to have a critical
role in the positioning of cortical and subcortical representa-
tions of sensory surfaces. They likely guide the formation of
overall patterns of connections between such representations,
and shape at least a crude pattern of internal organization. We
have reviewed some of the evidence that seems to indicate that
information from the receptor sheet is passed on from level to
level in sensory systems to modify the crude patterns of
organization into detailed and often modular representations.
Part of the evidence comes from experimental manipulations
of the sensory environment during development that alter
developmental outcomes. These experiments have been more
extensively reviewed elsewhere. Here we concentrated on
describing variations in the internal organizations of sensory
maps within and across species that seem most parsimo-
niously explained by assuming that information from the sen-
sory sheet guides the construction of sensory maps. This
information would most likely transfer from level to level via
patterns of action potentials, but the transfer of chemical signals
by axon transport or the release of chemical signals in axon
arbors as a result of non-spiking neural activity are other
possibilities. While further comparative studies of variations
in map structure may usefully suggest classes of mechan-
isms, the roles of such proposed mechanisms in the course
of development need to be evaluated with experimental
manipulations.
References
1. Kaas JH. The evolution of isocortex. Brain Behav Evol 1995;46:187196.
2. Sperry RW. Chemoaffinity in the orderly growth of nerve fiber patterns
and connections. Proc Natl Acad Sci USA 1963;50:703710.
BioEssays 24.4 341
24
Review articles
3. Hubel DH, Wiesel TN. The period of susceptibility to the physiological
effects of unilateral eye closure in kittens. J Physiol 1970;206:419436.
4. Shatz CJ. Emergence of order in visual system development. Proc Natl
Acad Sci USA 1996;93:602608.
5. Constantine-Paton M. The retinotectal hookup: The process of neural
mapping. In: Subtelny S, editor. Developmental Order: Its Origin and
Regulation, New York: Alan R. Liss; 1982, pp 317349.
6. Crowley JC, Katz LC. Early development of ocular dominance columns.
Science 2000;290:13211324.
7. Chiaia NL, Fish SE, Bauer WR, Bennett-Clarke CA, Rhoades RW. Post-
natal blockade of cortical activity by tetrodotoxin does not disrupt the
formation of vibrissa-related patterns in the rat's somatosensory cortex.
Dev Brain Res 1992;66:244250.
8. Kaas JH. What, if anything, is S-I? The organization of the ``first somato-
sensory area'' of cortex. Physiol Rev 1983;63:206231.
9. Sur M, Weller R, Kaas JH. Physiological and anatomical evidence for a
discontinuous representation of the trunk in S-1 of tree shrews. J Comp
Neurol 1981;201:135147.
10. Sur M, Nelson RJ, Kaas JH. Representations of the body surface in
cortical areas 3b and 1 of squirrel monkeys: Comparisons with other
primates. J Comp Neurol 1982;211:177192.
11. Sur M, Nelson RJ, Kaas JH. The representation of the body surface in
somatic koniocortex in the prosimian (Galago senegalensis). J Comp
Neurol 1980;180:381402.
12. Florence SL, Wall JT, Kaas JH. Central projections from the skin of the
hand in squirrel monkeys. J Comp Neurol 1991;311:563578.
13. Calford MB, Graydon ML, Huerta MF, Kaas JH, Pettigrew JD. Altered
somatotopy in the brain of a flying mammal. Nature 1985;313:477479.
14. Lane RH, Allman JM, Kaas JH, Miezin FM. The visuotopic organization of
the superior colliculus of the owl monkey (Aotus trivirgatus) and the bush
baby (Galago senegalensis). Brain Res 1973;60:335349.
15. Kaas JH, Harting JK, Guillery RW. Representation of the complete retina
in the contralateral superior colliculus of some mammals. Brain Res
1973;65:343346.
16. Kaas JH, Preuss TM. Archontan affinities as reflected in the visual system.
In: Szalay F, Novacek M, McKenna M, editors. Mammal Phylogeny, New
York: Springer-Verlag; 1993, pp 115128.
17. Payne BR. Representation of the ipsilateral visual field in the transition
zone between areas 17 and 18 of the cat's cerebral cortex. Vis Neurosci
1990;4:445474.
18. Bosking WH, Kretz R, Pucak ML, Fitzpatrick D. Functional specificity of
callosal connections in tree shrew striate cortex. J Neurosci 2000;20:
23462359.
19. White LE, Bosking WH, Williams SM, Fitzpatrick D. Maps of central visual
space in ferret V1 and V2 lack matching inputs from the two eyes.
J Neurosci 1999;19:70897099.
20. Allman JM, Kaas JH. The organization of the second visual area (VII) in
the owl monkey: A second order transformation of the visual field. Brain
Res 1974;76:247265.
21. Rosa MGD, Casagrande VA, Preuss TM, Kaas JH. Visual field represen-
tation in striate and pre-striate cortices of a prosimian primate (Galago
gainetti). J Neurophysiol 1997;77:31933217.
22. Pinon MC, Gattass R, Sousa APB. Area V4 in cebus monkey: Extent and
visuotopic organization. Cerebral Cortex 1998;8:685701.
23. Fukuchi-Shimogori T, Grove EA. Neocortex patterning by the secreted
signaling molecule FGF8. Science 2001;294:10711074.
24. Purves D, Riddel DR, LaMantia AS. Iterated patterns of brain circuitry (or
how the cortex gets its spots). Trends Neurosci 1992;15:362368.
25. Preuss TM. Taking the measure of diversity: Comparative alternatives to
the Model-Animal paradigm in cortical neuroscience. Brain Behav Evol
2000;55:287299.
26. Conway JL, Schiller PH. Laminar organization of tree shrew dorsal lateral
geniculate nucleus. J Neurophysiol 1983;50:13301342.
27. Kretz R, Rager G, Norton T. Laminar organization of ON and OFF regions
and ocular dominance in the striate cortex of the tree shrew (Tupaia
belangeri). J Comp Neurol 1986;251:135145.
28. Thurlow GA, Bowling DB, Cooper RM. ON and OFF activity gradients in
the lateral geniculate nucleus of the cat: A combined 14C2-deoxyglucose
and D,L-2-amino-4-phosphonobutyric acid study. Visual Neurosci 1993;
10:10271033.
342 BioEssays 24.4
29. LeVay S, McConnell SK. ON and OFF layers in the lateral geniculate
nucleus of the minK. Nature 1982;300:350351.
30. Stryker M, Zahs K. ON and OFF sublaminae in the lateral geniculate
nucleus of the ferret. J Neurosci 1983;3:19431951.
31. Sherman SM, Wilson JR, Kaas JH, Webb SV. X- and Y-cells in the dorsal
lateral geniculate nucleus of the owl monkey (Aotus trivirgatus). Science
1976;192:475477.
32. Casagrande VA, Kaas JH. The afferent, intrinsic, and efferent connec-
tions of primary visual cortex in primates, In: Peters A, Rockland K,
editors. Cerebral Cortex, Volume 10, Primary Visual Cortex in Primates.
New York: Plenum Press; 1994, pp 201259.
33. Horton JC, Hocking DR. Anatomical demonstration of ocular dominance
columns in striate cortex of the squirrel monkey. J Neurosci 1996;16:
55105522.
34. Shatz CJ, Lindstrom SH, Wiesel TN. The distribution of afferents re-
presenting the right and left eyes in the cat's visual cortex. Brain Res
1977;131:103116.
35. Lyon D, Jain N, Kaas JH. Cortical connections of striate and extrastriate
visual areas in the tree shrew. J Comp Neurol 1998;401:109128.
36. Krubitzer L. The organization of neocortex in mammals: Are species dif-
ferences really so different? Trends in Neurosciences 1995;18:408417.
37. Sur M, Weller R, Kaas JH. Modular segregation of functional cell classes
within the postcentral somatosensory cortex of monkeys. Science 1981;
212:10591061.
38. Sretavan D, Dykes RW. The organization of two cutaneous submodalities
in the forearm region of area 3b of cat somatosensory cortex. J Comp
Neurol 1983;213:381398.
39. Woolsey TA, Welker C, Schwartz RH. Comparative anatomical studies of
the Sm1 face cortex with special reference to the occurrence of ``barrels''
in layer IV. J Comp Neurol 1975;164:7994.
40. Dawson DR, Killackey HP. The organization and mutability of the forepaw
and hindpaw representations in the somatosensory cortex of neonatal
rat. J Comp Neurol 1987;256:246256.
41. Lu SM, Lin RCS. Thalamic afferents of the rat barrel cortex: A light and
electronmicroscope study using Phaseolus valgaris leucoagglutinin as
an anterograde tracer. Somatosens Mot Res 1993;10:116.
42. Henderson TA, Jacquin MF. What makes subcortical barrels? Requisite
trigeminal circuitry and developmental mechanisms. In: Jones EG,
Diamond IT, editors. Cerebral Cortex, Volume 11, The Barrel Cortex of
Rodents. New York: Plenum Press; 1995, pp 123187.
43. Catania KC, Kaas JH. The organization of the somatosensory cortex of
the star-nosed mole. J Comp Neurol 1995;351:549567.
44. Catania KC, Kaas JH. The unusual nose and brain of the star-nosed
mole. BioScience 1996;46:578586.
45. Catania KC. A nose that looks like a hand and acts like an eye: The
unusual mechanosensory system of the star-nosed mole. J Comp Physiol
1999;185:367372.
46. Jain N, Catania KC, Kaas JH. A histologically visible representation of the
fingers and palm in primate area 3b and its immutability following long-
term deafferentations. Cerebral Cortex 1998;8:227236.
47. Jain N, Qi HX, Catania KC, Kaas JH. Anatomical correlates of the face
and oral cavity representations in somatosensory area 3b of monkeys.
J Comp Neurol 2001;429:455468.
48. Kaas JH, Guillery RW, Allman JM. Some principles of organization in
the dorsal lateral geniculate nucleus. Brain Behav Evol 1972;6:253
299.
49. Kaas JH. Organizing principles of sensory representations. In: Novartis
Foundation Symposium 228, Evolutionary Developmental Biology of the
Cerebral Cortex. New York: John Wiley & Sons; 2000, pp 188205.
50. Van der Loos H, Dorfl J. Does the skin tell the somatosensory cortex how
to construct a map of the periphery? Neurosci Lett 1978;7:2330.
51. Welker E, Van der Loos H. Quantitative correlation between barrel-field
size and the sensory innervation of the whisker pad: A comparative study
in six strains of mice bred for different patterns of mystacial vibrissae.
J Neurosci 1986a;6:33553373.
52. Catania KC, Kaas JH. The mole nose instructs the brain. Somatosem.
Motor Res 1997;14:5658.
53. Guillery RW. Why do albinos and other hypopigmented mutants lack
normal binocular vision, and what else is abnormal in their central visual
pathways. Eye 1996;10:217221.
25

54. Guillery RW, Kaas JH. A study of normal and congenitally abnormal
retinogeniculate terminations in cats. J Comp Neurol 1971;143:71100.
55. Hubel DH, Wiesel TN. Aberrant visual projections in the Siamese cat.
J Physiol 1971;218:3362.
56. Kaas JH, Guillery RW. The transfer of abnormal visual field representa-
tions from the dorsal lateral geniculate nucleus to the visual cortex in
Siamese cats. Brain Res 1973;59:6195.
57. Hogan D, Garraghty PE, Williams RW. Asymmetric connections, dupli-
cate layers, and a vertically inverted map in the primary visual system.
J Neurosci 1999;19:RC38.
58. Hogan D, Garraghty PE, Williams RW. Lamination and visual topography
in the lateral geniculate nucleus of normal and achiasmic dogs. Eur
J Anat 1996;1:311.
59. Florence SL, Kaas JH. Ocular dominance columns in area 17 of Old
World macaque and Talapoin monkeys: Complete reconstructions and
quantitative analyses. Visual Neurosci 1992;8:449462.
60. Preuss TM, Kaas JH. Cytochrome oxidase ``blobs'' and other character-
istics of primary visual cortex in a lemuriform primate, Cheirogaleus
medius. Brain Behav Evol 1996;47:103112.
61. Tanaka S. Theory of ocular dominance column formation. Biological
Cybernetics 1991;64:263272.
62. Hensch TK, Fagiolini M, Mataga N, Stryker MP, Baekkeskov S, Kash SF.
Local circuit control of experience-dependent plasticity in developing
visual cortex. Science 1998;282:15041508.
63. Lee KJ, Woolsey TA. A proportional relationship between peripheral
innervation density and neuron number in the somatosensory system of
the mouse. Brain Res 1975;99:349353.
64. Welker E, Van der Loos H. Is the areal extent of sensory cerebral cortex
determined by peripheral innervation density? Exp Brain Res 1986b;63:
650654.
65. Azzopardi P, Cowey A. Preferential representation of the fovea in primary
visual cortex. Nature 1993;361:719721.
Review articles
66. Catania KC, Kaas JH. A somatosensory fovea in the star-nosed mole.
J Comp Neurol 1997a;387:215233.
67. Guillery RW. Competition in the development of visual pathways. In:
Parnavelas JG, Stern CD, Stirling RV, editors. The Making of the Nervous
System. Oxford University Press; 1988, pp 356379.
68. Zheng D, Purves D. Effects of increased neural activity on brain growth.
Proc Natl Acad Sci USA 1995;92:18021806.
69. Casagrande VA, Wiencken-Barger AE. Developmental plasticity in the
mammalian visual system. In: Kaas JH, editor. The Mutable Brain.
Amsterdam, The Netherlands: Harwood Academic Publishers; 2001,
pp 148.
70. Chapman B, Styker MP. Development of orientation selectivity in ferret
visual cortex and effects of deprivation. J Neurosci 1993;13:5251
5262.
71. Chapman B, Godecke I, Bonhoeffer T. Development of orientation pre-
ference in the mammalian visual cortex. J Neurobiol 1999;41:18
24.
72. Rauschecker JP. Compensatory plasticity and sensory substitution in the
cerebral cortex. TINS 1995;18:3643.
73. Angelucci A, Sharma J, Sur M. Modifiability of neocortical connections
and functions during development. In: Kaas JH, editor. The Mutable
Brain. Amsterdam, The Netherlands: Harwood Academic Publishers;
2001, pp 351392.
74. Huffman KJ, Molnar Z, VanDellen A, Kahn DM, Blakemore C, Krubitzer L.
Formation of cortical fields on a reduced cortical sheet. J Neurosci
1999;19:99399952.
75. Cowan WM, Fawcett JW, O'Leary DD. Regressive events in neurogen-
esis. Science 1984;225:12581265.
76. Shatz CJ. Impulse activity and the patterning of connectivity during CNS
development. Neuron 1990;5:745756.
77. Katz LC, Shatz CJ. Synaptic activity and the construction of cortical
circuits. Science 1996;274:11331138.
BioEssays 24.4 343
26
J Comp Physiol A (1999) 185: 367372
REVIEW
K. C. Catania
A nose that looks like a hand and acts like an eye:
the unusual mechanosensory system of the star-nosed mole
Accepted: 31 May 1999
Abstract The star-nosed mole (Condylura cristata) has a
snout surrounded by 22 fleshy and mobile appendages.
This unusual structure is not an olfactory organ, as
might be assumed from its location, nor is it used to
manipulate objects as might be guessed from its ap-
pearance. Rather, the star is devoted to the sense of
touch, and for this purpose the appendages are covered
with thousands of small mechanoreceptive Eimers or-
gans. Recent behavioral studies find that the star acts
much like a tactile eye, having a small behavioral focus,
or fovea at the center used for detailed explorations the sensory system, including the moles behavior, the
of objects of interest. The peripheral and central nervous
systems of the mole reflect these behavioral specializa-
tions, such that the small behavioral focus on the nose is
more densely innervated in the periphery, and has a
greatly enlarged representation in the somatosensory
cortex. This somatosensory representation of the tactile
fovea is not correlated with anatomical parameters (in-
nervation density) as found in other species, but rather is
highly correlated with patterns of behavior. The many
surprising parallels between the somatosensory system
of the mole, and the visual systems of other mammals,
suggest a convergent and perhaps common organization
for highly developed sensory systems.
Key words Cortical magnification Somatosensory
cortex Development Evolution Behavior
Introduction
The star-nosed mole is renowned for its unusual nose,
which consists of 22 mobile, fleshy appendages that
surround the nostrils. This peculiar nasal specialization
K.C. Catania
Department of Psychology, Vanderbilt University,
301 Wilson Hall, 111 21st Ave South,
Nashville, TN 37240, USA
Tel.: +1-615-322-7491; Fax: +1-615-343-8449
Springer-Verlag 1999
is unique among mammals and raises several basic
questions about star-nosed mole biology. What is the
star and how is it used by the mole to explore its dark
underground environment? What kinds of sensory re-
ceptors are associated with the star? How is information
from the star represented in the neocortex? Such ques-
tions have been the subject of a number of recent studies
exploring the sensory biology of this species and its
relatives (Catania et al. 1993; Catania 1995a, b, c, 1996;
Catania and Kaas 1995, 1996, 1997a, b). Each aspect of
structure of the nose and sensory receptors, and the
physiology and anatomy of the cortex, has provided
some important insight into these basic questions. But
the greatest insights come from combining these studies.
Unusual specializations in the cortex are explained by
behavioral observations. Anatomical comparisons
between the star and its cortical representation reveal
unexpected parallels between the moles somatosensory
system and the primate visual system. These results
provide an example of how a specialized species can
reveal general principles of brain organization.
Star-nosed mole behavior reveals a tactile fovea
Star-nosed moles live in wetlands in the eastern United
States and Canada. They build extensive tunnel systems
with their heavily clawed forelimbs, and seldom come to
the surface. Moles are insectivores with a high rate of
metabolism, and they must find and eat a large number
of invertebrate prey each day. As would be predicted
from their fossorial lifestyle, they have small eyes and a
tiny optic nerve. Thus a major challenge to their survival
is finding sufficient quantities of often small prey in their
dark underground tunnels. The star seems to have
evolved mainly for this purpose, and may be the most
sensitive touch organ among mammals.
When moles search for food or explore their envi-
ronment, the star is in constant motion and is repeatedly
touched to the substrate or objects of interest. Each
27
368
touch consists of raising the nose upward while the nasal
rays swing backward, then swinging the rays forward
while the star is brought into contact with the substrate.
This behavior is very rapid; star-nosed moles may touch
ten or more dierent places each second. Similar be-
havior in other moles without a star has been de- 11, and take the prey into the mouth, all in about 400 ms
scribed as tapping the ground with the nose (Nagorsen
1996). When a prey item such as an earthworm is
encountered, the rays are repeatedly touched to the prey
as it is bitten, torn up, and eaten. When prey is missed by
the star, by even a millimeter or two, the mole proceeds
past without apparent notice (whether searching in wa-
ter or soil), arguing against the hypothesis that the star
functions to detect prey from a distance through elec-
troreception (Gould et al. 1993).
With the naked eye, the details of foraging and prey
encounters are difficult to discern, but slow motion
videotaped behavior of moles locating small prey reveal
a stereotyped sequence of nose movements and a be-
havioral focus or fovea on the star (Fig. 1). Ray 11,
the center-most ventral ray (Fig. 2) is preferentially used
to explore prey items, acting in a manner analogous to
the fovea in the visual system of other mammals. For
example, whenever a prey item is first contacted with the
peripheral rays (rays 1 through 10) the nose is then
shifted so that further detailed explorations are made
with ray number 11 (Fig. 1A). No food item is eaten
without first being explored with ray 11. But since ray 11
is one of the smallest rays on the nose, food is usually
Fig. 1AC The use of the nose while foraging for food. A A
schematic representation of the nose with the rays numbered from 1 to
11, shown in relation to a prey item (gray oval). The first touch made
contact with rays 4 and 5 as the mole discovered the food item. The
second touch was centered on the 11th central pair of rays, after which
the food was eaten. For brevity, the prey is shown changing position
relative to the star but in reality, the star is moved relative to the
prey. This sequence of movements is characteristic of prey encounters,
which often include multiple touches with the 11th pair of rays. B A
flow diagram illustrating the progression of touches from the lateral
rays to the central 11th ray before the food is eaten. C An example of
the distribution of touches across the nose for ten consecutive preys
encounters, showing the preferential use of the 11th and surrounding
rays. B and C from Catania and Kaas 1997b
first encountered by the larger peripheral rays which
together make up most of the surface area of the star.
This stereotyped sequence of movements is very rapid.
Moles can touch a small prey item with the peripheral
rays, shift the star for several additional touches with ray
(Catania and Kaas 1997b).
The division of the star into peripheral touch and
central touch seems analogous to the retina in the visual
system of other mammals. In primates, for example,
photoreceptors in the peripheral retina typically detect
potentially important stimuli first, and then a saccade
brings the image onto the higher resolution, denser re-
gion of photoreceptors of the retinal fovea. Does the
anatomy of the nasal rays also reflect the dierent roles
they play in behavior?
The peripheral anatomy and sensory receptors of the star
Although the star is a specialization of the distal portion
of the nose, it is obviously not an olfactory structure. At
first glance, one might imagine the nose acting as an
extra hand. But the rays do not contain muscles or bones
and are not used to manipulate objects or capture prey.
They are controlled through tendons by a complex series
of muscles that attach to the skull, and their role seems
to be purely mechanosensory. For this purpose, they are
composed almost entirely of a series of small mechano-
sensory organs called Eimers organs. Under the
scanning electron microscope, these organs are visible in
a honeycomb pattern of epidermal domes on the surface
of each nasal ray (Fig. 2). Eimers organs are very sen-
sitive to light touch (Catania and Kaas 1995, 1997a) and
they are found on the snout of every member of the
family Talpidae that has been examined (Catania
1995b). However, the star-nosed mole has many more
Eimers organs than any other species and their distri-
bution across the nasal rays is unique: while most moles
have at most a few thousand Eimers organs surround-
ing their nostrils, the star-nosed mole has over 25 000 on
the star. Each organ is supplied by a number of primary
aerents, thus the star is very densely innervated. There
28

2
A. 1
N
11
10
8
9 Eimer's Organ
Fig. 2AC The nose and sensory organs of the star-nosed mole. A
The nose is composed of 22 appendages, or rays that surround the
nostrils (N). There are 11 symmetric pairs, numbered from 1 to 11 on
each side. The rays are used by the mole to explore its dark
underground environment through touch. The rays can be moved
through tendons that connect to facial muscles, and they are
repeatedly touched to the substrate as the mole moves through its
system of tunnels. B Each ray is covered with many hundreds of
domed sensory organs called Eimers organs. These are apparent in
a honeycomb distribution on the skin surface at high magnification.
The rays are essentially composed of these sensory organs which are
densely innervated by a branch of the infraorbital nerve. C Each
Eimers organ is a specialization of the epidermis that contains an
array of sensory receptors, including an encapsulated corpuscle, a
merkel cell-neurite complex, and a series of intra-epidermal nerve
endings that terminate in swellings at the apex of the central cell
column. Plate A from Catania and Kaas 1995
are over 100 000 myelinated fibers innervating the rela-
tively small star, and a substantial cross section of each
ray is taken up by a large nerve branch (Catania 1995a).
The internal structure of the star-nosed moles Ei-
mers organs is illustrated in Fig. 2C. Each organ is a
roughly 40-lm swelling of the epidermis that contains a
central column of keratinocytes. At the bottom of each
organ, in the connective tissue of the dermis, there is a
single encapsulated corpuscle. Just above this, among
the keratinocytes that form the base of the central cell
column, there is a single merkel cell-neurite complex.
Within the cell column, a series of nerve fibers branch
from three myelinated fibers in the dermis and ascend
along the margins of the column to terminate in a cir-
cular arrangement of swellings just below the outer layer
of epidermis. This concise geometric arrangement of
terminal nerve swellings is entirely enclosed and encap-
sulated by a single circular keratinocyte at the apex of
the cell column and the overlying outer layer of epi-
dermis is tightly sealed by many desmosomal adhesions
(Catania 1996).
While the response properties of the separate nerve
terminals at the apex of Eimers organ have not been
369
Terminal Swellings of
B. C. Free Nerve Endings
Epidermis
5
6 Merkel Cell
Neurite Complex
Dermis
Encapsulated
Corpuscle
7
Myelinated 10m
Fibers
characterized, recordings from the cortex show the star to
be highly responsive to several dierent features of fine
tactile stimulation (Catania and Kaas 1995). The super-
ficial nerve terminals in each Eimers organs are ideally
positioned to be stimulated by pressure to the top of the
cell column when the rays are touched against an object.
The Eimers organs have the same basic structure
across the dierent rays on the star. Ray 11, the be-
havioral focus of the nose, does not have a higher den-
sity of organs, or more organs than other rays, as might
be guessed from its important role in search behaviors.
In fact, because it is relatively short and has little caudal
surface, it has fewer Eimers organs than almost any
other ray. Ray 11 has about 900 Eimers organs on its
surface while some of the lateral rays have well over
1500 (Catania and Kaas 1997b).
Rather than having more sensory organs, a dierent
way in which a skin surface may be more sensitive to
mechanoreceptive input is by increasing its innervation
density and decreasing each aerents receptive field. To
test for this possibility, the number of Eimers organs
and the number of myelinated fibers innervating each
ray were counted and compared (Fig. 3). Rays 1
through 9 each had about 4 fibers per Eimers organ,
while rays 10 and 11 had significantly higher innervation
densities of 5.6 and 7.1 fibers per organ, respectively.
Thus there is a peripheral specialization of the rays most
used for exploratory behaviors, such that ray 11 has a
much higher innervation density than more lateral rays.
Previous studies of the organization of the somato-
sensory cortex, where touch information in mammals is
processed, have found a correlation between peripheral
innervation density and the size of the corresponding
representation or map of the sensory surface in cortex.
To investigate this relationship in moles, we examined
the organization of the moles somatosensory cortex and
related these areas to the anatomy of the star (Catania
and Kaas 1997b).
29
370
Average Fibers per Eimer's Organ 8 Fibers per Organ
7
6
5
4
3
2
1
0
1 2 3 4 5 6 7 8 9 10 11
Ray Number
Fig. 3 The ratio of myelinated fibers to Eimers organs from the 11
rays of the nose (from four moles). While rays 1 through 9 have an
average of roughly 4 fibers per Eimers organ, the 10th and 11th have
significantly higher innervation densities of 5.6 and 7.1 fibers per
organ, respectively (from Catania and Kaas 1997b)
The representation of the star in cerebral cortex
To explore the organization of the somatosensory cortex
in star-nosed moles, neuronal responses were recorded
using microelectrodes in anesthetized moles, while the
nose and body were gently stimulated with small probes
(Catania et al. 1993; Catania and Kaas 1995). A large
area of cortex responded to stimulation of the star, while
smaller areas responded to the rest of the body and the
limbs (Fig. 4). After mapping the locations of receptive
fields from the star and body of the mole onto the so-
matosensory areas, the moles were perfused and their
cortex was flattened, sectioned parallel to the cortical
surface, and stained for the metabolic enzyme cyto-
chrome oxidase, which often reveals cortical subdivi-
sions.
In the area of cortex that contained an electrophysi-
ological map of the star, a set of dark stripes was found
in brain sections processed for cytochrome oxidase
(Fig. 5). Each stripe corresponded precisely in location
to an area that responded to a single nasal ray in the
microelectrode recordings. Thus the representation of
the star in somatosensory cortex is visible in brain sec-
tions, much as the representation of whiskers in rodents
is visible as a set of cortical barrels in their cortex
(Woolsey and Van der Loos 1970). However, dierences
between the barrel cortex in rodents and the star rep-
resentation in moles were immediately apparent.
First, multiple maps of the star were seen in sections
of the moles cortex. As in all mammals, each half of the
body is represented in the opposite cerebral hemisphere,
but in each mole hemisphere there were at least two
clearly visible sets of 11 stripes representing the contra-
lateral star. In some favorable cases, a smaller third set
of stripes was also apparent (Catania and Kaas 1995,
1996). The most prominent set of stripes corresponded
to the primary somatosensory cortex, S1. The second set
Anatomical proportions
Cortical proportions
Fig. 4 Cortical magnification in star-nosed moles. The upper drawing
shows the anatomical proportions of the various body parts. The
lower drawing of a moleunculus shows the relative size of each body
part as it is represented in somatosensory cortex. As would be
predicted from its high innervation density and behavioral impor-
tance, the nose dominates the cortical representation. The forelimb
also has a large cortical representation, probably reflecting its
important sensory role in the excavation of tunnels, rather than in
locating prey (from Catania and Kaas 1996)
of stripes was a mirror image of the first, and corre-
sponded to the second somatosensory area, S2 (Catania
and Kaas 1995). S1 and S2 are somatosensory areas
found in all mammals (Kaas 1987, 1995), including ro-
dents, but there is not a visible representation of each
whisker representation in S2 of rodents. The third set of
stripes also responds to tactile stimulation of the nose,
but it has not yet been fully mapped because of its small
size and lateral location in the brain. The presence of
two and perhaps three visible maps of the star in dif-
ferent somatosensory areas of the mole is unique, and
may be a specialization related to the very high inner-
vation density and large number of mechanoreceptors in
the star.
Another unusual but not entirely unexpected finding
was a hugely enlarged cortical representation of the 11th
nasal ray. This was seen in electrophysiological record-
ings (Catania and Kaas 1995, 1996), and is also strik-
ingly clear in the cytochrome oxidase pattern of stripes
visible in brain sections (Fig. 5). This ray often takes up
over 25% of the entire S1 cortical representation of the
nose, despite its relatively small size on the star. Be-
haviorally important sensory surfaces often have greatly
enlarged cortical representations relative to their ana-
tomical size. But these enlargements are generally at-
tributed to their higher innervation densities, rather than
specific influences of behavior. For example, a direct
linear correlation has been found between the size of
cortical barrels in rodent somatosensory cortex and the
number of aerents innervating the corresponding
whisker on the face (Welker and Van der Loos 1986). To
30
 371

b
4
A. 3 5 Fig. 5AC The cortical representation of the star. A A rotated view
of the left side of the nose (dorsal is to the left, lateral is up) oriented to
match the nose representation in the primary somatosensory cortex
6 (S1) of the right hemisphere (below). B A tangential section through
layer 4 of the primary somatosensory nose representation in cortex,
7 processed for the metabolic enzyme cytochrome oxidase. The pattern
of 11 rays on the nose is mirrored in the brain as a series of 11 dark
stripes. Note, however, the 11th subdivision in the brain is greatly
2 enlarged relative to the size of the 11th ray on the nose. The
disproportionate representation of the 11th ray, and to a lesser extent
8 the adjacent rays, reflects their behavioral importance (Fig. 1). C The
average area of cortex in S1 per primary aerent for each nasal ray.
The representations of the rays in cortex are not proportional to their
respective innervation densities. Rather, the 11th and surrounding
1 9 rays have greater areas of cortex per aerent. The pattern of aerent
10 magnification is very similar to the distribution of touches across the
11 nose during feeding behaviors (Fig. 1C), suggesting a possible role for
behavior in shaping these cortical areas during development. Plate C
from Catania and Kaas 1997b
B.
8 9 innervation density did not account for the size of the
6 7 10 11th subdivision in somatosensory cortex (Catania
5 1995c; Catania and Kaas 1997b). Ray 11 contained
4 about 7% of the Eimers organs on the star, received
about 11% of the nerve fibers innervating the star
3 (accounting for the higher innervation density per or-
gan), but took up about 25% of the cortical represen-
2 tation of the star. Thus ray 11 is specialized both in the
periphery and in the cortex, by a higher innervation
1 11 density and a larger cortical representation per aerent,
respectively.
These findings are dierent from the results of similar
investigations in rodents (Welker and Van der Loos
1986), but nevertheless may be a common condition in
mammalian somatosensory cortex that is simply difficult
C to quantify in most species where the areas of cortical
150 S1 Cortex per Fiber representations and corresponding innervation densities
(Afferent Magnification) of skin surfaces are difficult to measure. This preferential
Average Area of Cortex

120 magnification of primary aerents has been termed

aerent magnification, to distinguish it from the tra-
Per Afferent (m2)

90 ditional term cortical magnification (Catania and

60
30
0 ploring prey (Fig. 1C).
1 2 3 4 5 6 7 8 9 10 11
Ray Number
explore this relationship in moles, the area of each rays
cortical representation was measured and compared to
its corresponding peripheral innervation (Fig. 5C).
Cortical representations reflect behavior,
rather than innervation density
While ray 11 was found to have a higher innervation
density than the lateral rays (Fig. 3), this increased
Kaas, 1997b). The latter is used to describe enlarged
representations of a skin surface in the corresponding
cortical representation, but does not take into account
innervation density. The pattern of aerent magnifica-
tion of the rays (Fig. 5C) is strikingly similar to the
distribution of touches across the moles nose when ex-
The relationship between peripheral innervation
density and cortical representational area has long been
debated for the visual system of primates (Malpeli and
Baker 1975; Drasdo 1977; Myerson et al. 1977; Perry
and Cowey 1985; Silveira et al. 1989; Wassle et al. 1989,
1990). The findings in moles support the most recent
findings in primates (Azzopardi and Cowey 1993) which
report a disproportionately large cortical representation
of ganglion cells from the retinal fovea. The relative
degree of aerent magnification is similar in the two
sensory systems (Catania 1995c). Thus there seem to be
parallels between visual system organization in primates
and somatosensory organization in moles, perhaps re-

31
372
vealing a general aspect of the cortical representations of
important inputs across dierent sensory systems.
All of these findings raise a number of additional
questions about star-nosed mole sensory biology. For
example, why have multiple cortical representations of
the nose, rather than one large representation? Does
adding new representations of the nose increase com-
putational efficiency through parallel processing of dif-
ferent aspects of touch? These representations are also
topographically interconnected both in the same hemi-
sphere and across hemispheres to the representation of
the contralateral star. The ability to identify each re-
spective ray representation in multiple maps will allow
future studies to determine the circuitry of this system in
detail (with the aid of anatomical tracers) to determine
the degree of parallel or hierarchical organization of the
cortical representation.
Another area of future research includes the func-
tioning of Eimers organs. The concise geometric ar-
rangement of nerve terminals at the apex of each organ
may be used to detect the surface features of objects in the
10- to 100-lm size range. If so, the sensory world of the
star-nosed mole could include a new realm of perceptions
not usually considered salient to mechanosensation, such
as microscopic textures and surface features previously
considered impossible to detect by mammalian touch.
However, the most striking result of these studies is the
similarity between the patterns of behavior in moles
(Fig. 1C) and the pattern of aerent magnification of the
rays (Fig. 5C). The sizes of the cortical modules repre-
senting the rays are not correlated with the anatomical
parameters traditionally assumed to be driving these
specializations, but rather are highly correlated with the
patterns of mole behavior (see Catania and Kaas 1997b
for more details). Here we may be able to examine the role
of behavior in brain development to see if it is in fact
driving the enlargement of cortical areas, or whether these
behavioral and anatomical parameters are somehow in-
dependently matched to one another during development.
Acknowledgements Thanks to Jon Kaas and Glenn Northcutt for
their support, guidance, and many fruitful collaborations in the
course of these studies. Neeraj Jain, Christine Collins and Melanie
Catania provided helpful comments on the manuscript. Special
thanks to Hrathkus and Bill Catania for their help capturing star-
nosed moles.
All experiments and animal care procedures were approved by
the Vanderbilt University Animal Care Committee and follow the
National Institute of Health guide for the care and use of labora-
tory animals.
References
Azzopardi P, Cowey A (1993) Preferential representation of the
fovea in primary visual cortex. Nature (Lond) 361: 719721
Catania KC (1995a) The structure and innervation of the sensory
organs on the snout of the star-nosed mole. J Comp Neurol 351:
549567
Catania KC (1995b) A comparison of the Eimers organs of three
North American moles: the hairy-tailed mole (Parascalops
breweri), the star-nosed mole (Condylura cristata) and the
eastern mole (Scalopus aquaticus). J Comp Neurol 354: 150160
Catania KC (1995c) Magnified cortex in star-nosed moles. Nature
(Lond) 375: 453454
Catania KC (1996) Ultrastructure of the Eimers organs of the star-
nosed mole. J Comp Neurol 365: 343354
Catania KC, Kaas JH (1995) The organization of the somatosen-
sory cortex of the star-nosed mole. J Comp Neurol 351:
536548
Catania KC, Kaas JH (1996) The unusual nose and brain of the
star-nosed mole. Bio Science Rep 46: 578586
Catania KC, Kaas JH (1997a) The organization of somatosensory
cortex and distribution of corticospinal neurons in the eastern
mole (Scalopus aquaticus). J Comp Neurol 378: 337353
Catania KC, Kaas JH (1997b) Somatosensory fovea in the star-
nosed mole: behavioral use of the star in relation to innervation
patterns and cortical representation. J Comp Neurol 387:
215233
Catania KC, Kaas JH, Northcutt RG, Beck PD (1993) Nose stars
and brain stripes. Nature (Lond) 364: 493
Drasdo N (1977) The neural representation of visual space. Nature
(Lond) 266: 554556
Gould E, McShea W, Grand T (1993) Function of the star in the
star-nosed mole, Condylura cristata. J Mammal 74: 108106
Kaas JH (1987) The organization and evolution of neocortex. In:
Wise SP (ed) Higher brain functions. Wiley, New York, pp 347
378
Kaas JH (1995) The evolution of isocortex. Brain Behav Evol 46:
187196
Malpeli JG, Baker FH (1975) The representation of the visual field
in the lateral geniculate nucleus of Macaca mulatta. J Comp
Neurol 161: 569594
Myerson J, Manis PB, Miezin FM, Allman JM (1977) Magnifica-
tion in striate cortex and retinal ganglion cell layer of owl
monkey: a quantitative comparison. Science 198: 855857
Nagorsen DW (1996) Opossums shrews and moles of British
Columbia. UBC Press, Vancouver, British Columbia
Perry VH, Cowey A (1985) The ganglion cell and cone distributions
in the monkeys retina: implications for central magnification
factors. Vision Res 25: 17951810
Silveira LCL, Picanco-Dinz CW, Sampaio LFS, Oswaldo-Cruz
E (1989) Retinal ganglion cell distribution in the cebus monkey:
a comparison with the cortical magnification factors. Vision
Res 29: 14711483
Wassle H, Grunert U, Rohrenbeck J, Boycott BB (1989) Cortical
magnification factor and the ganglion cell density of the primate
retina. Nature (Lond) 341: 643646
Wassle H, Grunert U, Rohrenbeck J, Boycott BB (1990) Retinal
ganglion cell density and cortical magnification factor in the
primate. Vision Res 30: 18971911
Welker E, Van der Loos H (1986) Quantitative correlation between
barrel-field size and the sensory innervation of the whiskerpad:
a comparative study in six strains of mice bred for dierent
patterns of mystacial vibrissae. J Neurosci 6: 33553373
Woolsey TA, Van der Loos H (1970) The structural organization
of layer IV in the somatosensory region (SI) of mouse cerebral
cortex: the description of a cortical field composed of discrete
cytoarchitectonic units. Brain Res 17: 205242
32
 Bullock, T.H., Orkand, R., and Grinnell, A. (1977) Introduction
to Nervous Systems. W. H. Freeman &Co. San Francisco. Chapter 7
Integration at the Intermediate Levels. pp. 242-290.

I. INTRODUCTION, DO MA INS IN to the afferent centers a coded Form of

THE INTERMEDIATE LEVElS the sense organs' selection, or fil trate, of
the act ual sti muli. The aHerent net-
We are concerned in this chapter with works-meani ng the interconnected
the transactions that occur in organized second-, th ird- and higher-order neurons
subsystems of receptors, neurons, and that receive thei r input primarily fro m
effectors. T heir mechan isms wi ll be cen- the receptors, though also from ot he r
tral to the issues of how the nervoussysttm sources-do not merely pass on this
works as it communications machine information. They use convergence
that recognizes, decides, and commands. of separa te cha nnels (comparison),
Of Ihe va rious domai ns of ope rations divergence of each channel (pa rallel
thai are ava ilable for our scrutiny. we processing), later.ll inh ibi tion (enhancing
include here fi ve, and these form the contrast), and other processes to modify
headings of sections II to VI. the signal. Special Attribu tes of the
T his is the rirsl encQunte r, between original input a re passed on (recogni-
these covers, w it h the phys iology of tion), a nd much in form.l lion is dis-
useful arrays of neurons. It wil l help to carded . The structure and coupling
introduce two dassrs of function that Functions of the network determine what
e merge from a nd pervddc the activities gets through .
of such arrays. Compara ble networks exist on the
Each pari of th e nervous system-but out put side. They are also fil te rs, but
especia lly the receiv ing s ide, fr om Ihe since they Formulate and send to the
battery of receptors to the networks of eHectors comma nds thill are crucia lly
higher-order neurons of the aHerent patterned in sp.1Ce a nd ti me, they are
systems-ca n be thought of as a filter . oft en thought of as palle rn genera tors.
The sense organs send a pa tt erned T hey convert triggeri ng inpu t or steering
stre.1m of impulses, in space a nd t ime, to input from receptors or their own spon-
the affe rent centers, and these represent taneous discharge or a mixture of these

33
242
into ad.lplively patterned ("coordil1<1ted") second means is to vary the frequency of
Chapter 7
sire.llTIS of impulses in the output chan- impulses in each unit, since either the
Integrdlion at the iniermediat(' Levels
nels. Again the connectivity and dynamic aver,'ge tension in a series of twitches or
properties of the network determine its the teta nic tension is a function of fre -
Figure 7.1 outpul, but not merely by passive filter - quency. You can observe frequency and
Rf(rrlilm(ll/ ill /I muscle wilh ing. Often there are intrinsic rhythms of recruitment;control phenomena by plac-
IIlru molor III1i/S.
spontaneous impulse bursts, and these ing a stethoscope over your eyelid and
Silent mOlor result in the generation of specified con- listen ing to the twitches of its muscle
stellations and sequences of activity in while you control them willfully. Re-
populations of units and finally in the cruitment may be important in verte-
effectors. Let us look first at the final brate skeletal muscle only at low ten-
link. sions.
Since there are many motor units in
most vertebrate skeletal muscles, oper-
II. NERVOUS CONTROL IN ating these unit s out of phase with each

r
EFFECTORS: DIVERSITY OF other can give a smooth overall contrac-
PERIPHERAL INTEGRATION tion even when the frequency in each is
so low that it contracts in a series of
The best known vertebrate skeletal twitches. Another possible utility of
muscles consist of muscle fibers whose having many units is that there could be
cell membrane is capable of producing a rotation of activity during low or
propagated all-or-none impulses like medium work loads, sa..that units could
those of nerve fibers. Each muscle fiber is rest. But this old notion is apparently not
innervated, in general, by only one axon confirmed in the best studied materials.
with one terminal at the speci"lized Other muscle fibers in vertebrates arel
end-plate. A motor nerve impulse arriv- incapable of producing propagated
Muscle fibers '1. inactive
ing there gives rise to an end-plate or Mtion potentials. They are innervated
Low ~sjon junctional potential, like an e.p.s.p., not by a single end-plate, but by numer-
Axons with
impuls.!;. trains which is usually suprathreshold for the ous spatially distributed motor nerve
initiation of a propagated muscle action terminals. Such multiterminally in-
potential. Each motor neuron innervates nervated muscle fibers (Fig. 7.2) respond
many muscle fibers. The motor neuron electrically with junctional potentials
and its muscle fibers comprise a motor only, but since these occur at many sites,
unit. Each muscle contains from one to eJectrotonically spreading depolarization
several thousands of motor units. An can activate the whole fiber. Muscle
impulse in one unit, or a synchronous fibers of this type m"y be mixed with
volley in many, causes a brief contraction others, as in some frog muscles. They are
to summate if they overlap, or to fuse usually innervated by small diameter
into a smooth contraction called a teta- motor axons.
j
nus if the frequency is high. When there Since individual vertebrate muscles
are many motor units per muscle, a are excited by many axons, we think of
means of grading the strength of con- them as being driven by pools of motor
traction of the muscle is to vary the neurons. There is a diversity of s ize
number of active un its. This method of within the pool of motor neurons. Gen-
~-v--
High tension control is called recruitment (Fig. 7.1). A er<llly, the smaller ones have lower

34

Fig ure 1.2
T!lI'''s of s1 riR /fd /H1/S(lf illlUrvRlio ,l,
""""""""""""""" ; """""""'""""',:
Unil erminal innervatio n
Mul1i ' lIrminal innllrvalion
";::Z""""""\C""",,z ',,,n, ;:;$';,
Dinou ronal innerva,i on
! I , ~ I " I " :m , ,:::25tii,
Polynooronal innervation
thresholds for natural stimulation and
are more tonic in their discharge. They
innervate relatively fewer mu scle fibers
and produce wea ker contractions, but
are probably principals in th e ord inary
load of muscle work. T he larger phasic
fibe rs are called forth by stronger stim-
ulat ion and produce vigorous action.
T his relation among threshold, fi ber s ize,
and tonic versus phasic mode of dis-
cha rge is probably of gener,J\ s ignifi c.mce
in neurophysiology. 11 is easily demon-
strated in sensory system s as well.
Many whole arthropod muscles, and
some in an nelids, are in nervated by only
one or a very few motor axons. [n
muscles innervated by only one axon,
recruitment cannot be li ke that in verte-
brates, which add motor units by centra l
enl istment of motor neurons. Bul a
s ingle axon can recru it incre"sing Ilu m-
243
bers of the muscle fi bers it supplies
Se<:tiol'l II
because diffe rences in the facili talion Nervous Con'rol In Effec,o rs:
properties of the synapses allow trans- O ivers il y of Peri pheral In. egra Uo n
mission to be effecti ve "I different fre-
q uencies of arriving nerve impu lses in
different fibers. There are also differ-
ences in excitati on-cont raction coupling
threshold in some muscles.
Innerva tion of all th ese muscles is of
the mul tilerminal type. At norma l fre -
quencies of "rriv ing nerve impulses, the
junctions commonly ex hibit the t ime-
de pendent properlies of facilitation or
antifacilitation. Contract ion strength can
be controlled not only by average fre -
quency of moior axon impu lses, but, in
some muscles, also by deta iled tempora l
structu re of the impu lse train. For
example, the opener muscle of the cray-
fi sh claw is innerva ted by only one
excitatory neuron, and the myoneural
junction shows sl rong faci litation. A
steady rhythmic Irai n of impulses causes
a contraction of a given strength. But if
the same lotal number of impulses in a
given time is delivered grouped in pairs,
the contraction is stronger. The junction
is considered to be pattern sensitive in
that it responds to delai ls of innervating
tempora l p,1 I1ern, nol just ave rage fre-
quency (Fig. 7.3,A). Recordings from
intact moving animals s how Ihat the
eNS sometimes iss ues molar comm.1 nds
in impulse doublets, but il is not yel
certa in how this potential code is em-
ployed, wheth er independently of or
correlated wit h mea n frequency.
Muscles may be polyneuronally
innervated in that si ngle fibers can
receive input fr om more than one motor
axon (Fig. 7.2). Each innervating axon
di ffers in the properties of its endings on
the same muscle fi ber. O ne of the axons
is usually d so-c<l lled "("sl" or phasic
axon that normally car ries short bursts of
35
244
impulses at high frequen cy for quick
A SLOW movements and produces la rge junc-
tional potent ia ls in the muscle fi ber (Fig.
7.3,8), .md these elicit spike-shaped loca l
response potentials and rapid cont rac-
tions. Anotller axon is a "slow." or tonic,
nerve fiber that normally carries long
trains of impulses at low frequency;
these impulses decrement severely as
they approach the nerve ending, and
B FAST
prod uce smaller and more facilitating

20 mV I junct ional potentials that cause slowly

developing con tractions. Common ly
there is also an inhi bitory axon. Each

0.25 9 I - ltm\\\~~I~II'lI~\I'I\I\I\I\\"\I'I.
~

~
innervating axon may end upon a large
fraction of all th e fibe rs of the muscle,
the fast axon preferentially reaching the
2 sec fa st (short sa rcome re) and intermediate
C
muscle fibe rs; the slow axon, the s low
ISO (long sarcomere) and intermediate mus-
cle fibe rs (Fig. 7.4). Th e presynaptic
2 endings of the same axon upon muscle
~

.,
0 _
0-
o 100
"0
_ -
0
fibe rs of d ifferent sarcomere length and
hence contraction properties differ in
effectiveness, in arousi ng post junctional
,"
0 -

-=
oE
potentia ls, and in degree of faci litation.
An extr.lord inary heterogeneity of mus-
'~a 50 cles and actions and thei r smooth grada -

~
0
tion ca n now be understood, even in
such animals as arthropods, in which
U
only a few axons s upply a whole muscle.
o ~O--:2C--:.--:,C--~8-:-:I~O--:I2;--:I';'--:'~6 ~.3 Th is heteroge neity is the resu lt of the
di ffer ent combinat ions of phasic, tonic,
Excita tory spaci ng interval (msec)
Fig ure 7.3 and inhib itory axons, th e different de-
Differences .llllong mu sc le s in response 10 stimulus inlerv"L A. A mu S( le grees of polyneu ronal innerva tion, th e
fiber in the c!.1\V. cl 05cr musc le o f " cTab res po nd s to repetit ive sl imu l,l!io n d iversity of type of terminal of each axon
of " "slow" ~ )(on with 1.ITge detion potentials (upper Ir",e) hav ing slow upon dirfel'ent muscle fi bers, the diver-
(,lci lit,ltio n, and wi th gr,u(u,llly SU lllllldling smooth tetani c contTdc! ;o n
s ity of type of muscle nbe r, and the
(lower I rd CC shows tenSio n). B. T he Sdme Illu scle fib er exc ited vi .. a " foist"
,,"o n shows 5111.111, qllickly pl,l\cduing action pote ntia ls and small, cloni c . diversity of rise and fa ll times of fa cilita-
( :::; nonfuse d) co nlrdclions. IAtwood, 1967. ] C. Pd ttern -se nsilive (fi lled cir- ti on.
cles) ,1lId p,lttcrn-i n sensi livc (open c ircles) mu sc le of (ray fish. At a con- Inhi bitor axons exert two distinct
st,lnt ,,,~n,, freq ue 'lcy o f 30 shocks/ sec (== 33.3 msec {ntervals), trains o f effect s, presynapti c and postsynaptic
~ hock s Me d elivered ei ther ,1 t uniform intervals (extreme right) or at alt er-
(Fig. 7.5). Postsynaptic inhibition causes
ndtely short (d b5d ss~ v~ l ucs) ~n d long illter\'~ls; respon se plo tted as a per_
centd ge of the e\'en ly sp~C(!d cO llt ra ction. Refrdctori ness red uces responses hyperpolariz..ll ion and increased con-
at the shortest int erv.ll s. IRipley .I nd Wiersmd, 1953. 1 ductance of the muscle fi ber membrane;

36
 245

"Fast"" a~on (phasic) Sec tion 11

Nervous Control in Ef(ectors:
Diversity of Peripheral Integration
Large
" fas!" '
e.p.s,p.
+ spike

Fast conlraction
\LliLJ:LLLL1.LLLL.'.LLJ..LlllLLllJJ.J

More
facilitat ion
1

Figure 7.5

Less
fac ilitalion
j Di~grtHlwmli(

of a (msla(e~tI
re"restllialioll of
Hcitatory a"d i"/Iihitory imlervaliml
mllsdt fiber.
III ~

Nerve
/termina I S~
Excitatory Inhibitory

Slow contraction Presyn. inhib.

Postsyn.

m ))) ) ) ) )6
inhib.

Large "slow" e.p .s,p. Muscle

ExdlMory urn'e It l'ltliuui, form rleIHO

"Slow" a~on (tonic) mlisO/Ia. ' )I>1apse, . Inlribilory la mi
to msec Iluis forlll 'Ifllrorllllmdar sy'W/'SfS
Figure 7,4 (/wslsyrlapli( illhihilioll) I1Ild nxo axo
Crustacean neuromuscular innervation. The diagram illustrates the matching of phdsic ("f.1St /lal syrla/lSf, H/wrl fxril~lory lerll/illais
axon"') a nd tonic ("slow axon") in nervation with different ca tegories of muscle fiber (indicdted (llre, y>1aplic illllibiliOlI). [LI1IIS IlUd
by sarcomere le ngth ) in a crab muscle. [Atwood, 1973.] Alwood,1973 1

37

246

Chapter 7
,
Integration at Ihe Intermediate levels

,
~

D F
Fig ure 7 .6
The ,tn"tomical distribution of ,1XOnS to the dista l thor"cic limb muscles of several groups of
Mabcostrac ... BI.lC k lines rep resent motor excitor " xons; colored lines represent inhibitory axons.
Br,Kke!s in dicate tht: distribution ofaxons behveen nerve blllldies. A. Br,lChyu,,,.,.u. AIlDmur" . C.
Stom"lopod,,- D. Palinu,a. E. Astaell'''. F. Stenopodide,' (N"I",,!i.,). The boxes rep,esent the muscles
in the three di stal segments o f the legs, .IS indicated by thei r COllllllOn IMllle5: Il. ,1Ccessory nexor; b,
bender; (, closer; f, exten so r; f. nexor; 0, opener; r, rol,l\or; ~. stre tcher. IWiersll1.l and Ripley, 1952.)

presynaptic inhibition redu ces mi niature
e.p.s.p.'s and the amount of excitatory
transmitter released per impulse. Both
depend on the arrival time of the inhibi-
tory impulse relative to the excitatory,
but in different ways. They are widely
different in importance. Inhibition in
some muscles is mainly postsynaptic; in
others, presyna ptic. In some mu scles the
response to a slow excitor is inhi bited
presytl<1ptically, and the response to a
fast excitor in the same muscle is 'Iess
inh ibited . In the crayfi sh claw opener
muscle, inh ibit ion cau sed by 10 impulses
per second is 90% presynaptic, but a t 40
impulses per second it is only 50% pre-
synaptic. Inhibitory trans mitter is re-
leased with facilitation, which varies
among junctions.
An additional degree of complexity of
muscle innervation in arthropods can be
appreciated if one examines the innerva-
tion of the muscle of a whole limb, such
as the walking leg of a crayfish. Several
muscles may be innervated by a single
motor axon (Fig. 7.6). The overlapping
but nonidentical innervation patterns
still .1110w independent action. For
example, the opener and stretcher mus-
cles are innervated by the same excitor
axon but by different inh ibitors. Never-
theless, the whole limb is innervated by
only 12 excitatory and J inhibitory motor
fibers, and this fact of s mall numbers

38

makes it seem possi ble that we will be
ab le to understand fully ,l t the level of
s ingle nervous units the normal control
of movement in arthropods.
The important lesson from the
exa mples chosen is that even ,lt th e last
lin k, the nervous control of effectors,
th ere is diversity in the principles em-
ployed, some cases ex hi biti ng a sub-
stantial degree of in tegration in the
periphery. Wags have sa id crabs can
think (evaluate) in their legs. We h,we
not exha usted the diversity: fast insect
flight muscles, a va riety of smooth mus-
cles (Box 7.1), electric organs, glands,
cilia, chromatophores, a nd other effec-
tors manifest additional principl es of se-
ries and parallel control (Fig. 7.7).
III. ANALYSIS OF
S ENSORY INPUT,
PARALLEL AND SERIES
PROCESSING
The general principles of sensory recep-
tion are common to systems as different
as vision and taste and to inputs that
rarely re,lch consciousness, such as those
from pressure receptors and che mo-
receptors in the grea t Mteries, as well as
those that do. Foremost is the principle
of parallel cha nnels: many receptors in-
dependently sample the sti mulu s world
a nd send their imp ulse-coded reports to
the centra l nervous system in parallel
afferent fiber s. The basic independence
may in some organs be abridged by
superimposed in fluence fro m the centra l
nervous system (centrifugal or efferent
infl uence) or from neighboring receptors
(lateral inhibition).
Inti mate ly rela ted is the second
principle, that of overlapping receptive
fields: neighboring receptors are gener-
247
Figure ??
all y stimulated by a fract ion of the im- Various Iyl"'s of ,!f,c/ors.
pi ngi ng world that is large but circum-
scribed (the excita tory receptive fi eld,
ERF) and which overlaps substantially , ,
:,
~
- :
with that of cont iguous ch,lllneis. T his
, , -,
applies in some systems to topographic
oV('flap and field s, s uch as areas of the , , ... ,
/ '\
skin, retina, a nd basilar membrane --,
Fly wing muscles
(which distributes sound frequencies to
aud itory afferent s). In some systems the
overlap involves ambiguity in modality
(Chapter 6, p. 214).
The third principle relevan t to the Snail heart
present section states that sensory input
typically exhibits divergence and con-
vergence in th e central nervous system,
the former d istributing information to
analyzers concerned with different as-
pects of the sti mu lus world, the la tter Bladder. ureter. u reth ra
permitt ing resolution of ambigu ities,
sha rpening of field s, and abstraction of
0 'VIJ\iIfWWVl M!l!\ilflfV\f\']
special features. Generally, divergence
predominates in sensory pathways, so rlJlit/lJlJlflJVWW\IWIJV\I)
that even though each neuron still re- ~'1NlfWV\N\I\I\fV\.I\i\iUvt,'j
ce ives synapses from ma ny different
presynaptic cells, there are increasing e""""'VWINVlIVWWVW]
nu mbers of cells at each higher neu ral Elect ric organ
level enroute to the sensory cortex . One
c~,
characteristic exam ple is the first- to
second-order neuron stage in the audi-
tory system, already mentioned (p. 108).
The value of such wide divergence is that
different postsynaptic cells, receiving Salivary gland
in put from subtly different, overlapping,
presynapti c populations, can perform
d ifferent types of information-process ing
operations in parallel, more fully ex-
tracting all available in for m'ltion. Thus a
visua l signal ca n be analyzed by diffe rent
cell populations for brigh tness, shape,
color, distance, and movement; an audi-
tory signal, for frequen cy, rate and di-
rection of change of frequen cy, duration,
in tens ity, presence of overtones and
temporal patterns, localiza tion of the
Chromatophore
39

248

Box 7.1 T he Nervous Control o~

,.------
We will briefly revicw the sl.lte of knowledge with
respect to vcrtebr<ltcs. Prosser (1 973), from whom we
take muc h of this .lecount, divides 5moolh muscles of
vertebra tes into two kinds, unit<ll'y and multiunitary,
with some intermedia tes. Unitary muscles include Iho~
of the ma'or viscera - uterus ureter, and gastro-
lnies!in .. ! IT"'t. These s ow spontdllcOUS rhythmicil ,
and distr ibution of excitation IS tned ro m muscle
fi ber to fibe r. They ca n be stimulated by st ret ch and
mo(lii1;ueaby nerves. Multi u n itd ry muscles ii1c"l ude
niclildlJ ng membrane an pi omolor, cilia!),. <tnd iris
nruscles. T hese are norma lly aclivdted not spont.lne-
o~y or by stretch, b'"Ut by nerves or ho rmones; d is -
tri bulion of excitation within the muscle is normally by
nerves. I here may be 1 ,lcilitatio n of p.s.p.'s; severdl
nerve- fibers may influence one muscle fiber.
Unitary muscles ,lI1d some o thers ,ue typically
arrJngcd in bund les of fibers glnncctcd to each other by
" nexuses," areas of dose ilpposi tion .1Ild e lecirica l low
resisl,mce. The motor nerve termina ls release trans-
mitter from large numbers of varicosities (swellings).
Only some muscle cells receive d irect inncrvdtion (see
figure). O the rs are excited by their d irect electrotonic
coupling 10 these. Still others aTe excited ind irectl y,
pcrll<lps by a combinatio n o f e lectronic coupling .Iud
diffu se transmitter. In different viscer.l the smooth
muscles vary widely in the proportio ns of these th ree
types,.lS well as in cable prope rties, ion dependence.
the effects o f transm itte rs and drugs, sym p.lthetic ilnd
pa r.lsym p.lthetic innervation, a nd Spont.lIlCOUS activity
patte rn .

~ "Di reclly innervated "" cell with close (200 A)

neu romu scular junc tions

C iii,' . :J "Coupled"" cell exhibits junct ion po te nt ials

Carried by e tectrot onic coupl ing
c ~ "" tndirectly coup led' cell ex hibits only ac tion poten tials
l ow-resistance pathway

Varicose nerve fiber

Schelnatic represenldtio n of the types of autono mic innervation of s mooth mu scle. All th e smooth -
muscle fibers are In terconnected by l ow-re5 i 5 t~ n ce electrotonic junc tio ns, shown here as bridges.
Some receive direct nerve endings (ddrk shading): othe r.; (light shading) do not, but Me clo~
enough to the forego ing to show junction potentidls th~t h,we spread elect rotonical1y. The mo st
remote (unsh ..ded) show no junction potentials; they n",y be elicited by electroton ic sprNd of
aClion potentials from the preceding d"ss and by tr.. nsmitter rele.. sed fro m the varicose nerve fibers
at some disl"nce. [BurnstO( k and IWJ~ama, 1971.1

40

source in space, etc. After bu ilding up
s harply specillc response requirements,
these elements can again converge to
combine their specificities.
Ani mals abstract from their total sen
sory input specia l qua lities of that input
before formulating a motor command. A
frog jumps and snaps at any sma ll, dark,
moving object wit hi n range if it is in the
mood. A male stick leback fi sh attempts
to CQmt with any oval, silvery, red-
bottomed object of su itable size, whether
it be a female stickleback or a crude
mode l. A few ChM.lclers of Ihe whole
constellation of visual inputs associated
with the female stickleback seem to be
the only relevant ones to the ma le (sec
Chapter 8). How might fil tering of this
quality (recogni tion) go on in the
nervous system?
We could discuss mtering networks
for any sensory modality, but space does
not permit a survey. Instead we will
concent rate on visua l filteri ng. Since
notions of moda lity, submodality,
labeled lines, and tempora l coding have
been dealt with in Chapter 6, we can go
directly to a consideration of relatively
complex network functions.
The histological structure, cell types,
and connectivity of the vertebrate retina
have already been descri bed (pp. 121-
126), together with some of the electro-
physiological properties expressi ng the
coupling fun ctions of the connections.
At least thirteen fun ctiona lly distinct
types of optic nerve IIbers carry infor-
mation to Ihe brain from the eye in the
ca t. Ta ble 7.1 gives ,I recent classificat ion.
Note that 92% have exci tatory receptive
fie lds (ERF's) that are concentrically
organized. These arc either brisk or
sluggish, referring to the promptness and
vigor of their responses, and either
tr.lnsienl or sus tai ned, referri ng to thei r
249
mainly phasic or mainly tonic char.lcter.
5lIon III
For each of the four combinations of AnJiY5iJ of ~n'iOry Inpul:
these properties there are two types, I'Molliel ~nd Series ProcesSi ng
distingu ished by the sign of their COIl-
centric orga niza tion. There are ON-
center-OfF-surround units and Ihe con- Overlap 01
receptive lields
verse, OFF-center-ON-surround units.
Units of the former sorl are excited by
the ON of a sma ll spot of light in the
centra l zone of the ERF or by the OFF of
an annular illu mi nation of the s urrou nd-
ing zone. Units of the latter sort arc
converse in character. The table lists all
types of uni ts in the order of their axon
diameter, the largest first. T he largest
and fa stest un its, those in the brisk-
transient group, are also ca ll ed
V-neuro ns; the next la rgest un its, those
in the brisk-s ustained group, arc also
Convergence
called X-neurons. (All the rest, both
concentric sluggish and nonconcentric,
are coIiectively ca lled W.neurons, but
this term embrilces too heterogeneous a
set to be useful. )
For ma ny years the cat was descri bed
as havi ng esse ntially only two types of
optic nerve fibers, ON-center and OFF-
center. In contrast, the rabbit, ground
squirrel, and gray sq uirrel were de-
scribed as more frog-l ike in having
several com mon nonconcenl ric and more
complex types-for example, those pre-
ferring movement, some specillc for
direction of movement, some even for
orientation. It now appea rs that the cal
has about the same v.uiely of Iloncon-
centric units, including th ese complex
types, but in much small er propo rtion,
8% compared 1034% in the rllbbi t. In the
cal, "loca l edge detectors" are fi ve ti mes
as com mon as "direction selective" units;
in the rabbit the lalter predominate,
except in the viSual st reak (the equiva-
lent of the are.' centralis, for high resolu -
tion) .
41
250
All of these optic nerve fi be r Iypes
Clupler 1
Intcgr.ltio n at the I nter l11ed l~ 11!' leve ls bespeak transa ctions tha t process the
information encoded by rods and cones.
La tcra l interact ions, convergence, and
highly specified connect ivity (see
Chapler 3) among receptor, horizontal,
bipola r, amacrine, and ga ng lion cells are
a ll indicated. This is usua lly referred to
as early extraction of fea tures to dis-
tinguish it from further changes in the
mea ning of cell discha rge at laler stages
in the visual pathways. It also evidences
paralle l processing for centra l destina-
tions of q uite differe nt fu nction. The
c>l1lr.11 targets of these o ptic nerve fiber
types are known o nly in parI. In the ca l,
X- a nd V-fibers go ma inly to the dors.l l
late ral geniculate, e.lch to its priVate class
of geniculate ne urons; these in turn
project 10 the vi sual cortex, X-fibers to
the so callea simple cell s and Y-fibers to
the complex cells. W-fi bcrs go to the
midbr.l in, largely to the tectum. It sho uld
be pointed ou t here that the re are a t least
s ix cent ral targets of optic nerve fibers, of
whi ch the dorsa l later.ll genicu la te and
th e tectum are on ly two. Evidence sug-
gests that they have d ifferent functions.
T he retinas of frogs. lizards. and
pigeons have even fewer concentrica lly

Tabte 7. 1
Recepti ve fi eld types of 960 cal relina l ganglion cell s, Latencies are the ranges of a nti-
dromic conduc tio n ti mes (rom the optic tract stimulus site, and can be t,lke n as pro-
portional 10 Ihe reciproca l of axon diameter.

Types Number Pereen/age

- Larency
(msee)

Concentrically organized 887 92

Brisk m BO
Transient 243 25 1.0-2.4 ) y."I1,
ON -center 115
OFF-center 128
Sustained
ONcenter
531
271
55 2.5-5.9 ] X-cells
OFF-cenler 260
Sluggish 113 12
Sustained 44 4.6 - 2 4.0
ON-cenler 22
OFF-cenler 22
Transient" 27 6.1-18.7
ON -center 13
OFF-cenler 14 Wcells
Nonconcentrically organized 73 B
local edge detector 45 5 6.6-15.9
Direction selecti . . . e 11 1 6.1- 12.4
Color-coded 6 <1 3.8-14.2
Unilormily detector 5 <1 8.7-13.9
Edge.lnhibitory OFF-center 3 <1 3.9-6.6
Unclassilied J 3 <,
Sou"e. Cleland and Levl&k. 1914
' In e lurtl'lOr 42. Ihere wera insuiliciont obserYlllLorlS to distinguish whelher sustained Or transie nt (21 ON-center.
22 OFFcenter)
I Insu1liCIIlnt ObServalions 10 reach e conc lu sion.

42

organized ganglion cells and many more
movement-specific cells. This is prob-
ably not a phylogenetic trend; it may
have some relation to habit of life and
the roles of vision. frogs are known best
from ~onee ring work by ~et vi n,
Maturana, NrccU11iJcI~s, who
SeCI-Aim:u! h5f more Itah.lf<frk1ilOs than
flashes of diffuse fields or focused
beams, and from the quantitative work
of the Griissers (Fig. 7.8). In contrast to
the cat, in which all optic nerve fibers are
myelinated, the vast majority in frogs are
very fine and unmyelinated. The five
types of optic nerve afferenls may be
summarized in the following way. (a)
Type 1 fibers, called "susta ined-edge
detectors," respond to a sharply focused l
edge of an ob ject, light or dark, moving I
or recently having moved into the 1 -30l
ERF. (b) Type 2 fibers, called "convex-
edge detectors" (Fig. 7.9) respond only to
a sma ll object, darker than the back-
ground, that moves into or has recently
moved into the field; they must detect
not only the change of position in the ca.
3 ERF but also tha t there is little or no
change of position in the ca. 15 sur-
rounding inhibitory receptive field (IRF).
(c) Types 3 fibers, called "changing con-I
trast detectors:' respond to any edges in
motion, large 01' small, dark-an-light, or
light-an-dark, if the contrast and rate are
adequate and the edge is not too fuzzy.
These fibers give a weak response to
nonmoving tempor.l[ changes in light;
they aTe classical ON-OFF units, but
much prefer motion. (d) Type 4 fibers,
c.llled "dimming detectors" respond to
any dimming or darkening, whether
caused by motion or not; these are OFF
fibers but not like the concentric OFF-
center units of the C.1t. (e) Type 5 fibers
respond to any brightening; these are ON
fibers, but not concentrically organized,
251
with a n OFF-sensitive surrounding ex-
Seclion III
citatory field, as in the cat; they are more Analysis of Sensory Input:
sensitive to blue light than to white or I'arallel and Series I'rocessi ng
other colors.
Types 1 and 2 are fine, unmyelinated
axons in the optic nerve and by far the
most numerous. Types 3, 4, and 5 are
myelinated. Types Ito 4 go to the optic
tectum of the mesencephalon, type 5 to
the dorsal [ateral geniculate of the
diencephalon. Histo[ogica [ types of
ganglion cells with distinctive dendrite
branching patterns are probably asso-
ciated with these fiber types (see
Fig. 3.4).
Types 1 and 2 do not respond to gen-
eral illumination, and their responses to
movement are not influenced by the
[evel of ill um ination over a range of
intensity of mOTe than a hundredfold.
Given their req uirements they di scharge
at a rate tha t encodes contrast of the
moving object (but not light [evel) and
rate of motion of the object minus that of
background objects in the IRF (but not
relative motion). The discharge is
ambiguous for certain severely con-
stra ined combinations of object size,
contrast, speed, and amplitude of move-
ment.
Change of position is as good a stim-
ulus .15 visible movement; th us a good
res ponse is elicited by briefly illuminat-
ing a sta tic scene (no response) and th en,
during the dark period between flashes,
moving an object within the ERF. The
"memory" of the position of objects
during the first flash lasts through at
least one second of darkness (Fig. 7.10).
An additional and remarkable property
of type 2 units is erasability. When such
a unit fires upon motion of a sufficiently
convex edge toward the ERF center, it
continues to fire at a [ower rate for some
seconds after the motion stops; but it
43
252

Oitfuse light Movement of a 2' black spot

---I
00
1 sec I
Of<
J f" ERFsize
Class t Class 1 -- tttHtIHHlHtttttHtttttttt--~ 2- 4'
Class 2 Class 2 11111 1111 I II 25-4'
Class 3 I Class 3 11111 6_8'
Class 4 I II 111 11111111 Class 4 - - - - - - -- ~ 10'

Horizontal movement Location of optic Horizontal movement

of a 2 x 10' vertical nerve fiber endings of a 2 x 10' horizontal
black bar at tacta l cell black bar
_ __ _ _ _ __ dendrites
. /- ~ ./
Class!
Class 2
11 11111 11 ~~~
....:::::::::
Class 1
Class 2
61111 111111 11 I I I I II I I
111111 1 I I I
. Class 3 11111 ~ ,.,.".g05. Class 3 1111
Class 4 11111 I I II ----"'- . Class 4 111 1 I I I I I

_ Excitatory synapses

----l Postsynaptic inhibitory synapses

~I Presynaptic inhibitory synapses

Fi g ure 7.S
Types of ganglion cells and their optic nerve fibers in the frog. Above. The four types found in the
optk tectum, distinguished by thei r responses to four kinds of tests. ER f, excitatory receptive field;
s.g.r., stra tum griscum ce ntrdle of the tectum; . g.s., stratum griseum superficiale. Below. Diagram of
the connections between the elements of the retina converging on a ganglion cell; presumably differ-
ences in the details of these connect ion s and their transfer functions account for the types. II. ama-
crine cells; B, bipolar cclls; G, ganglion cells; fi , horizontal cells; R, receptor cells. [G rilsser and
GUsseT-Cornehls, 1972.]

44
 253
Type 1
tlbor

No spikes in response 10 room light ON

or OFF or to e large moving bar

Flash Flash
No
Light 0 response
inion sily

8
or to illuminating or moving
a natural-looking image.

-'
~
8 Stationary obloct

Good re5ponse to small, dark

mOving object In II 3" field, Flash Flash
light 0 Response
inten sily

even if th e illumination is
1p::..J '~. 8
very taint. but
Change in position
during dark

Figure 7.10
"Movement neurons" mdY respond to chJngc of
pos iti on, with ~ forgetting time.

not ilthe edges ere too fUllY ,

or to reversed contrast.

Figure 1.11
Abstrdctlon e.. rly in ,m .1(ferellt pJthway. 1m
pulses recorded from .. frog oplk nerve fibe r
in the roof of the midbr.. in. [Based on d .....
of M"t ur"nil al aI. , 1960. )

45
254
promptly (eases if the light is turned off
Chapter 7
Integr~lion at the Intermediate levels
briefly, and does not resume after the
light is back on.
Notice that units of type 2 respond to
any sma ll, dark, moving object, espe-
(i,llly if the motion is jerky. A hungry
frog will jump at .my such object.
Normally any object having these char-
acteristics in the environment of a frog
will be a bug, an edible object. Since
there are many detectors of type 2 spread
over the visual field, the activity of one
or morc fibers of that type can both
signal the presence of a bug and give its
spatia l coordinates, hence both activate
and steer a jump and a tongue flick.
Tadpoles may lack some of the Iypes
more important for the metamorphosed,
hunting frog. Type 4 units in the frog's
optic nerve signal general dimming,
perhaps the approach of a predator or
any large object.
The known visually oriented behavior
of frogs in a natural environment in~
cludes finding food and avoiding ob-
stacles and large, threatening objects.
The fi ltering processes that occur in the
retina in frogs can abstract the relevant
aspects of the whole visual input signal
for those behaviors. Studies by Ingle and
by Ewert, us ing single-unit recording,
ablation, and stimulation suggest that
central processing builds on the retinal
filtering by separately analyzing optic
input for different behavioral meanings
in different structures. The frog's attrac-
tion to blue depends on the dorsal
geniculate; its avoidance of large ob-
stacles in jumping and its retreat from
large threatening objects may depend on
distinct, overlapping parts of the pretec-
tum, and its approach to food upon the
tectu m. We may speculate that the phase
locking of its circadi,'11 rhythm with the
environment.,1 photoperiod depends on
a hypoth.1lamic center, the suprach ias-
matic nucleus, .'s has been shown in rats.
Each of these four central structures re-
ceives its own optic input, consisting
la rgely of a distinct mix of the ganglion
cell types.
As activity proceeds through first-,
second-, third-, and nth-order neurons in
a sensory pathway, the me.1ning of the
impulse activity changes. We have been
considering differences in meaning in
parallel neurona l pathways as a result of
divergence; we should note the differ-
ences in meaning in successive neurons
as a result of convergence. When the
criteria for firing include .1 significant
fraction of the complex features of .,
stimulus that releases norma l behavior,
we m.,y speak of recognition cells; th is is
a matter of degree. Higher-order ce ll s
may add dimensions to the criteria, such
as novelty or familiarity. A novelty un it
deep in the frog tectum may fire in
response to a small, dark, moving object
anywhere in a large (30) field, but will
soon cease, only to resume if th e same or
another object wiggles in a fresh part of
the field. A familiarity unit fires in
response to a similar object, but, instead
of ceasing, continues, even maintaining a
low rate of discharge ("muttering") for
many seconds if the object stops moving.
It now ignores fresh, moving objects
within its la rge (30) field. It will fla re up
if "its" object s lowly moves about within
Ihe field, but will lose it and go silent if
the object jumps too far-to a fresh part
of the field!
In the auditory sphere too, we know of
46

units of a wide r,lllge of co mplexity of
criteri;!., In bats, for ins tance, a series is
found leading to ce lls that do not fire in
rcsponse to any pure tone at any
intensity but on ly to frcq uc ncy-
modulated to nes with a certain range,
rate, a l.l d direct ion of modulation, like
the bat's echo-ranging cry. In squ irrel
monkeys, cells arc found that s tro ngly
pre fer a certain one out of some 20
tape-recorded sounds chosen from the
35 or so natural vocalizations in the
species' reperto ire.
Some workers distinguish between
recognition of natural sti muli and feature
extraction; thc form er is potentially more
complex a nd may result from con-
vergence of cells of th e laller type. The
term "fea ture" in this context means
lim ited aspects of a complete natural
stimulus, such as duratio n or frequen cy
modulation . Our information is still too
lim ited to decide whcthcr some sub-
systems in some a nimals work differ-
ently from o thers in a fundamental
sense. It is oft e n s upposed, fo r example,
that some subsystems fu nnel s uccessive
featu re detectors down to a s ingle recog-
nition unit, like a " bug detector,"
whereas o thers never quite converge the
set of re leva nt feat ure detccto rs. We d is-
cussed the samc problc m from ano ther
di rection o n pp. 238-240. By whatever
means, complex natura l stimulus recog-
niti on mu st occur widely in nervous sys
tems-sometim es early in the pathw.1YS,
sometim es late.
M odulation of input by ce ntral in-
fluence via centrifugal (efferent) fibers is
a potentia lly important part of the active
fil tering in many sense o rga ns, as was
noted on pages 170 and 225. Inh ibito ry
255
, d,
Secllo n IV
Elemen' MY Ne uronal Nelworks:
Eme rgen t P,o!,(, rties of Ci rcuitry
"
f/ ., ga
Figure 1.11
Efferent fibers ' 0 <l sense o rgd n. Axons from the
brain to the ret in~ in the pigeon. Golgi prepdfa-
tion. ig, displaced gang lion cell; f, n~t ~macr ine
cell; gil, ganglion celt I~yer; ir, inner nurleM
I~ yer; ip, inn er plexiform I~ yer; 5, Snl,l 11 pM,Isol
amacrine cell. [Maturana and Frenk, 1965.J
fibers to m uscle receptor orga ns in cray-
fish are illustrated in Fig ure 2.75. The
y-effcrents to muscl e spindles a rc trea ted
I
below (p. 267 el seq. ). T he mammalian
coch lea (Fig. 2.80,C) and the avian retina
(Fig. 7.11) arc also well-known exam ples.
but the fu nctio nal Significance of the
effercnts is no l yel adeq uately under-
stood.
IV. ElEMENTARY NEURONAL
NETWORKS, EMERGENT
PROPERTIES OF CIRCU ITRY
In Chapter 3 we introduced some well-
stud ied examples of connectivity and
ccrtain general prin ciples of circuits of
neuron-like units. Here we extend the
di scussion to emphas ize th e physio log-
ica l consequences. Three kinds of ,Hrays
will be chosen; these may be c.lllcd
" networks," fo llowing th e usage in th e
literature o n neural mode ling, me,m ing
any assemblage of con nected ne uro ns .
47

256
A. Mutually Exdtatory or
Ch3pter 7
Positive-feedback Networks
Integration at the Intermediate levels
If two or morc neurons are capable of
exciti ng each other, then input tha t ex-
ceeds th e threshold of one will likewise
excite the others (Fig. 7.12A), Should the
others exceed threshold as it result, they
feed excitation back to th e first, a nd a
runaway process ensues. The whole
Inetwork may come into a state of max-
imum activity a nd, without lim iting
/ processes, might stay in that condition.
Most neurons have relatively long-term
sel f-inhibitory processes, such as adapta-
tion, accommodation, or fati gue. As the
neurons in the positive-fe ed back net-
work begin to fatigue, one or more of
them will decrease in frequency. They
then excite the others less, and hence
I I
Figure 7.12
Simple networks wilh (AI mulua1 excitMion and
(B) muhIJ I inhibition. lnpu l (omes from presyn-
dplic axons. Excitatory synapses shown by
forked axon terminals, inhibilory by te rmi nal
balls.
also receive less excitatory Feed back. Just
as the whole network was able to rUIl
away to maximum activity, it now ru ns
away negatively to a min imum state.
Once the adaptation or fa tigue has worn
a way, a n~ cycle C,lll begin. Networks
of cells, each of which may not be capa-
ble of rhyth mic bursts of activi ty, can
produce bursts of more-or-less syn-
chronous activity in all mem bers.
Networks of this kind have been
demonst rated in severa l cases and may
be widespread. Inspiratory interneurons
in the medulla are probably so con-
nected, perhaps leading to their rhythm ic
bursting. Certain cells in the brai n in
several gastropod molluscs have been
found to be positively coupled. They
produce synchronous rhythm ic bursts ')
that act as triggers in the control of
feeding and other activities. D eca pod
crustacean hearts al'e controlled by t,ri\
ga nglia containing only nine cells. -; ~ ~
baweerr- some of these is known at
least to aid in the build up of the heart-
beat.
B, Mutually Inhibitory or
Negative-feedback Networks
Two cells, or two clusters of cells that
inh ibit each other, may produce alter-
nating single impulses or rhythmiC
bursts. This arrangement, ca ll ed
reciprocal inhibition, is a common fea-
ture of the control of antagonistic effec-
tors or actions. The activation of motor
neurons of one group of muscles is often
coupled, both by feedback and feed for -
ward, to inhib ition of the motor neu-
rons controlli ng a ntagon istic muscles (see
Section V, p. 266) . This ki nd of reciprocal
inhibition is di,lgramma tically simple in
48
 257

Vest>bular l'Iuclel
VIii n(trve roolS Secllon tV
Eleillelltuy Nr urorrdl Ne tworks:
Emergent Properlie5 of CIrc uit ry

Figure 7.13
A reciproedlly inh ibi ting pair of neuron s. The gi.1l1 t cells of Mduthner in the medull~ 0' nsh, sc he
'll~ticdlly shown, with the inhib itory colldlerdl (/I) Indicd tcd only on the left, the indirect VIII nerve
afferenls (8) only on the lefl, ~rrd the dired VJJJ netve dffercrrts (e) CO.lllrlg orrly from the right,
~lthough .III the5e components ~re ru lly bildleul. I Rct~l.lff dnd Fontdlne, 1960. 1

the Cdse of the paired giant Mauthner's
fibers of teleost fish and tailed amphibia
(Fig. 7.13). T he cell bodies and large
dendrites ofl hese neurons are situated in
the medulla, where they collect input,
including particu larly Ihat fro m Ihe
eighth crania l nerve (sec also Fig. 2.17),
The axons cross over before descending
the spinal cord to synapse on the motor
neurons of the longi tudinal musculature.
Each axon has a bra nch that ends in an
inhili,ilQry: 5 na se on til contra I ral
Maulhner's cell axon hil , nca r the
spike- initiating site. (This is an electricill
synapse. ) An impulse in one cell pre-
vents a simu ltaneous one in the other.
Each descending impulse activates,
nearly synchronously, an extensive lon-
gitudinal body wa ll musculature on one
side, causing a twi tch-li ke curvature of
the posterior body region or tail. This
startle reaction begins the familiar, sud-
den "jump" of some fish when the
aquari um glass is struck. The vibrations
excite the sensory endings in the vest ib-
ular apparatus of the ear. Although these
cells have the largest axons in the body,
and although each has thousands of
input terminals and must be often bom-
barded with very many impulses per
second, they produce on ly the occasional
output necessary for start le reactions.
Another system of one-to-one alterna-
tion is found in the neurons driving the
tymbal muscles in some cicad.l s. A pace-
maker interneuron firing about 200 im-
pul ses per second drives two mola l'
neurons, which each fire at halF this rale
in exact alternation and precisely phased
with respect 10 the pacemaker. The
alternating clicks of the two tymba ls
during song double the sound frequency
poss ible if they were synchronous.
Reciprocal inhibition Cdn lead to

49
,
258

C hapter 1
Thresho ld 20 mV I
[tlt cgr~tlon at th e [nlermedl.t!e le ... el~

Rebound "'lI f table (Iorma l)

Stimuli Inhibitory spike s in

,
II III
I Spikes out 1
I IIII I
A

1111 1 1111+1--111111 1 1111111 1I1IIH-lH+1II1-1--+
m


------

~II I 11111 1 111111 1111111 1111111 III III I
3 iIIl ~III IIUIII!llIIllllllyllmllYilll
Briel I.p.s.p. barrage sta rlS long barrage
altern at ing bursts sl ops
8

O>---~
'--IlIII----!IIIUIIIIII- lIlIlIlIllI--IIa8111f--IIIII-~IIII'BIIIIllHlIlIlI----III!IIIiIIIIIII--II~1
I -IIIllllIIII 1l1li
'-.
'- --+1111111111 1I11 1 1 111l---111111111 lI ;mllll- 11IIlI1I1I 1I1"1111U1l 111111111 1111 RIll--
3 I I I I I I
Single I.p.s. p. l a tephsse La ler
triggers paUern intens ifies pha se resets
c
Figu re 1.14
Rccip rOC.t1 il1hibilion .. nd pastinhibitory rebound provide ncx ible mechanisms (or ~cnerJling bursts,
A. POSlinhibitory rebound in a model neuro n. A neuron 11t.11 is not spont.lneously ,lcli"e receive s dn
inhibitory input and produces spike o utput by rebound . lJ. Two such neurons, reci procally in hibi.
tory. c;on giW! a long 5eries o f altern ating burst s to ~ brief input bArr~ge from ~ third neuron. C. A
single inp ut impu lse has d iffe rent effects .according to t he phase of the .altern ation when It url\"!'s.
[perke l and M ulloney, 1974.1

alternating bursts of act ivity in otherwise
nonbursting cells. The networks shown
in Figure 7.14, B and C cons ist of only
two interconnected cells, but they could
represen t two populat ions. Input to the
network is inhibitory in this exam ple,
.1nd reaches only one of the cells. Th e
1
com mon neurona l property of post-
inhibitory rebound (Fig. 7.14,A) is
invoked in this model, and causes bursts
that suppress the ot her cell. As the re-
bound burst in the first cell s lows down,
the second is dis inhibited. Released and
rebound ing in its tum, the second fires

50

and inhi bits th e first. As the seco nd ce ll
s lows, th e fir st recovers, and the whole
cycle repc.l ts.
Whether dependent on rebound or
not, some such si mple mecha nis m for
alterna ting burst acli vity, though difficult
to demonstrate, see ms to operate in
f.wo rable mate rials, like the lobsler
stomatog.lstric ganglion. We believe it to
be an important mechanism for many
kinds of rhythmic and alternati ng be-
havior. Ins piratory and expi ratory inter-
neurons in the mam malia n medulla
probably inhibit each other. Locomotory
syste ms in ma ny a nimals may involve
r~i p rocal inhibition between pace-
makers for antagon istic muscle sets.
Ikciproca ll y inhi biting networks can
pe rform a variety of fu nct ions, according
to the pa rt icular trans fer functions of the
' synapses and the input and output con-
nections. T he following examples are
theoretica l and q ualita tive; proof that
reciprocal inhi bition is the mech.1 nism
that operates in living an imals is incom-
plete. G ive n certain properties, recipro-
cally inhibiting networks can act as
gates, switching ra pid ly fro m control of
the output by one input line to another.
This can recur at a steady r.lIe, for ,1
steady-s tate input, providing a pace-
maker in which no s ingle cell is the
essential clement, as discussed above.
With cert ain dynam ic properties the
same simple circuit C,ln act as a n
intensity-lo-Ii me con verier that m,l Y be
useful in comparing the strength of two
inputs by their relative dura tion of con-
trol of some downstre,lm system. Slight
changes could provide sensory sca nning
by periodically or irregularly sampling
e,lch of a number of input li nes a nd
giving them control, in turn, of some
later elements. In a cell w ith a large
number of input li nes converging on it, a
,
259
drastic increase of acti vity in one li ne
Seellon IV
would ord ina rily be drowned in the Etrmt' nluy Ne uron,,] Ndworks:
background activity of the ot he rs; but EIlIt'rgt'nl t'rop"'rlies of C ircuitry
reciprocal inh ibi tion in the same cell
might all ow one line to dom inate if its
activity were to rise above some level,
and thus serve to d irect allention or to
switch control. Still a nother possible use
of such circuits, with only slight cha nges
in the coupling functions, might be as an
ala rm system. Time- a nd load-sharing /,1
delay lines. null detection, and filt ering !
are othe r theo retically availa ble con- I
seq uences of such ne tworks.
An array of inhibitory cross-con-
nections can be thought of as a n arelM
in which there is competiti on be-
tween par.ll1el streams of impulses.
Both the nonlinearity of dependence on
act ivity levels a nd the cri tical inflections
in the input -output functions wou ld
all ow one strea m to win control. A
spectrum of properties is possible from
democratic to oligarchical to dictatorial.
The reliability of performance of
networks is considered on p. 233.
C. La teral Inhibition Networks
Th is term differs from the precedi ng
head ing in directing a tte ntion more to
l'l yers or arrays than to alternate cells or
groups. Many neural tissues consist of
layers of similar neurons that inh ibit
e.lch other eithe r direct ly or ind irect ly.
The inhibitory connections spread from
a ny part icular cell to make conta ct wit h
neighboring and more distant members
of the layer, but wilh decreas ing de ns ity,
aSwe saw in Chapter 3 (p. 11 1). Hence the
effect iveness of inhibition decreases with
dista nce.
Some effects of latera l inhibition a re
ill ustrated schematically in Figure 7.15.
51
260

-
Inhibitory sensory nOl,l ro ns
Interneu ron s

5 /0
O..;'-------~,,?--~J:0"''----
0..,/
+ $ -
,s'---------<<E;"-
0'~ - a.."~" s
__(0,:'-----
'"

o,:s'--------~C~"-'-<'~-(o,:S'--C+~1
"Yo"v/'"

~ +"E'--_ _ _ _~O"-'-'-"J
O,S'-" - S + E+ I
(0':'-'--"-''-'
;<,',).

0;s'-+:':.!E'--_ _ _ _~<::~"~,-,-:'~,1os + E + 21

O"S'-'+"E"--____~cr:c"-'-'-'~(OS + E + 21
0;;;;:0..,,/"

"S'-+:':.!E'---_ _ _ _~C:"-'-
0- ~ ''--- ', (O~'-"'-'--"
5 + E + 21
Re ceptors 0 Second-order
Inhibitory sensory neurons
Interneurons
" Darker" "Brighter"
~----.
Stimuh.l! Physiological
in tensity ellect
A B c o
1

52
 261
Flgu.~ 1. IS I/llriHg r~SO')
0'
l .ltl'r.ll in hibillon. A. A p.lUl'rn unifolmly gr.y ,"r.lS with shup f>dgn. rrpr~rnlinR .I "isu.1 fit'ld
_n by .I n eye. (M.lch b.nds .In illusion in ",hlch .Ipp.onmtly d.lrk er .nd lightl'r lunds .Irt' _ n El tm~nl.lry
51"<1Ion IV
Neuronal Ntlworks:
.I round t'.tch t'dgt', Mt' not evidl'nt to us in this conrrguration.) B. The sllmulus ploul'd <IS inlensity Emt'rgtnt Propt'Tl ies of Circu itry
(horizontlll) .Ivinsl lpo1 lial e>:tt'n t (,-ert ic", IL emphllsizi ng the unifonn ily of the pnysiclll intt'Mily
wit hin ('IIC"h IIrt'll. C. A network of rtc~ptOfS ",nd S('cond ~rd('r neurons with rt'ciprocal Inhibitory
C"o nntelions vi. inll'rneurons (brokr n linrs). Spont.l nrous act i\~ l y in the replors (5) is .I ugmenlOO by
ucildtion (E ) d ue 10 light. Thl' network CdU$e5 Ihe sl"<ond ~rder nl'u.o ns 10 show 5 dC"ti vity dU g-
1111'ntl'd by E . nd/ or rt'd ucl'd by inhibition (I ) in single or doubll' (11) dose. D. Th(' output, ('(j Ui"d'
len l to our $ensdlion, piatti'd .IS dMker or ligh ter th.n tht' bolCkground dut' to S.

First consider Ihe behavior of one spon-
taneously active cell in the network
while a stimu lus is moved about its input
field . As the stimulus approa ches, it first
excites neighbors of the recorded ce ll.
Si nce they inh ibit Ihat cell, it res ponds
by a decrease in firing frequency. If the
stimulus passes directly over the re-
corded cell, it is excited 10 fire above its
normal ra te, but as the stimulus moves
on il is again depressed by ils neighbors.
Any cell in Ihe network may be char-
acterized as having a recept ive fi eld that
has an excita tory center and an annular
inhibitory s urround. Th is is compa r.lble
to the ON-center ganglion cells of the cat
ret ina.
If, ins te.ld or looking at the response
made by a single cell in the s patial array
or Figure 7.15 to a moving sti mulus, we
examine the ou tput of a whole line of
cells in the array while one hal f of the
array is stimulated more than the other
half, lVe see an abstracting function of
the network. Cells in either uniformly
stimula ted half of the fi eld all in hi bit
each other symmetrically, but th ose at
the edge do not. Ce lls on the strongly
stim ulated side of the edge are inhibited
wcak ly by their nCighbors across the
edge whi le they strongly inhibit those
!lame neighbors. T he result is especially
high .llld low firi ng frequencies CI t the
stimulus edge. In terms of the fir ing
frequencies or Ihe cells in the network,
the stimulus edge has been enhanced;
there has been a spatial di ffere ntiation of
the input signal. A second layer of
neurons cou ld be so constructed tha t it
lVould detect the edge only. Latera l in-
hibition probably explains our psycho-
physical illusion known as "Mach
bands" (Fig. 7.15,0), and perhaps oper-
ates widely to en hance the sensitivity to
cont rasts.
Since the lateral spredd and ex trd
synapse take time, there is a delay in the
inhibi tion. Th is confers a tem poral prop-
erty of freq uency se lect ion or fil tering
that acts to attenu.l te ra pid changes as a
function of dista nce rrom the center and
to prolong the exaggerat ion of the pri-
mary response at the center, again fa vor-
ing low frequenci es.
It Cdn also be seen that the inhibition
exerted on a given cell by those sur-
rounding it can be redu ccd by stilllulation
of a slightly more distant popu lation.
T he di stant popu latio n red uces the lcvel
of activity of the nearer one, wh ich
"d isi nhibils" the center.
These phenomena were first described
by Hartline and his .lssocia tes in a se ries
of elegant ex periments on the compound
eye of Lil/IIIIIIs. The same latera l inter-
action exist, however, in virtua lly "II

53
,
262

sensory syslems in both vertebra tes a nd known <l nd presu mably main ly che mica l.
Chapter 1
Integr.lHo n at the Inlerm~d l~te l eve ls invertebrates and may o peT,l le at the All the fu nctiona lly established connec-
earliest stage in the pathw,lY a nd / or at tions can be an.l torn ica ll y justified by
later stages, even in the cortex (Fig. 7.16). fibers vi sualized by inj ection of Procion
We h,wc seen in Chapter 3 how lateral yell ow. How mu ch spontaneity there is
inh ibition might result from recurrent cannot be sf'ated, but it must be con-
collalera ls of the axon e nding on neigh- s ide rable. Among these junctions, .,
bori ng cells (Fig. 7.17). We saw holV such v,uiety of integrative input-output prop-
inhibition ca n be complicated in the erties are found. It seems like ly tha t if we
cerebellar cortex by th e inhibitory influ- could unrave l the web of influences,
ence of recurrent coJlalerals of Purkinje there wou ld be bot h excitatory a nd in-
cell axons, not only on neighboring hibitory reciprocity, latera l effects, a nd
Purkin je cells bu t on basket cells, mixtures of spontaneous rhythmicity
thereby disinhi bit ing Purki njes in a cer- with imposed bu rst-sha ping effects. The
tain geome tric pattern. Re nshaw cell norlllal activity is largely a rhythm ic
laler,l] inhibition via molor neu ron re- se ries of bu rsts of repeatable but la bile
curren t collaleraJs is trea ted on page 269 s pike pattern; the known connections
(Fig. 7.24). ex plain much of the detail of the bursts.
Alth ough many of the connections of
this network are known (nearly all;
D. Mixed Networks hence far more in proportion than for
a ny ot he r known system of some com-
If both excitatory and in hibitory connec- plexit y), it is difficult to assign causes to
tions exist in homogeneously in a set of th e bursting phenomenon itself. Does
cell s, the variety of possible outputs th e pos itive feedba ck, by itself, cause
expand s to the degree that it is useless to one group of cell s to burst, or is the
make (l priori or genera lized, uncon- reci proc.ll inhibition relationship with
strained models. A recently studied real a not her group necessary, or even s uffi-
exa mple is, however, worth describing cient? In theory, ei ther mechanism could
(Fig. 7.18). T he stomatogastric ganglion prod uce bursting, but physiological evi-
of crustacea ns controls the stomach, one dence suggests that both kinds of con-
part of which contains the gastri c mill nect ions exisl. Perha ps both mechanisms
used to grind food. The gangli on con- oper.lle synergistica lly to make th e
ta ins 30 cells, almost all of which a re whole system more stable.
identifiabl e ,md consta nt in con nections
a nd influe nce. Te n a re motor neurons to
the g.lstric mill part and 14 to the pyloric E. Connections Ensuring
part of the stomach, but both groups also Synchrony of Activity
have direct infl uence upon each other.
Furthermore, two interneurons are pres- Although exactly synchronous activity of
ent with connections to both sets of neu rons is not known to be req ui red in
motor neurons. There are 123 known m.1 ny insta nces, it is importa nt in a few,
inhibitory connections and only 6 ex- and these a re interesti ng in s howing the
ci tatory junctions. Twenty-nine of the fl ex ibility of design with whi ch cells ca n
junctions are electrotonic, the rest un- be connected, among th emselves .l nd to

54
 ",g ill
263

0
:g 20
.,.30
'~ 40
- 20 "'.
~ SO 20 30 50 100 -"
i ~~1iI ,, - 60
Lower
- 80 aUditory
30
~ 40 m ce nter
.~ - '00
50 20 30 50 100 oE
A Vi b rat ion freq uency (Hz) ."
,-
~
0.2 0.5 2 5 10 20

0> 0
-
00_

.. t
< !
Fig u re 7.16
Sh~ rjX'n i ng by suppressing sensitivity o n e.lch

<
~
- 20

-" V
side of th e best frequency, with converging
inpu t. A. Th resholds for sensMion as the e nd -
point; vibr<ltion felt by 2 or 3 fi ngers. 8. Th resh-
olds for si ngle neuron firi ng as the end poi nt;
un it respo nses to sou nd at lower and higher ~ udi
- 60

- 80
Higher
audito ry
cen ter
..\1
\
H
tory centers. It should be noted t h ~ 1 such narrow - 100
cu rves at h ighe r cen ters are no t common; units
0.2 0.5 2 5 10 20
wi th m.my types of cu rves are found. {Von
Ilckesy, 1967.)
B Sou nd Irequency (k Hz)

~i

f igure 7.11
Recun en! (011,,11'. ,,15, The li ne fibers are the array of colliller.. ls of th ree pyra middl
cells in the corte .. of the killen. IScheibel and Scheibel. 1970d.)

55

264

ALN '1 $ " hO' ! .. t! !::; 'J ' Io\;" ,IHI I!

"" ~~~
LGN
ifjirgii"1'. k':: :: . -".,If:;'; 'I:;:: }ff
j ' " . ," "' " .
PNI ~ I , I I , I , , I I I ! , , I I I I I I I I I I I , I
d-l VN : ' . . 'r .:!",': . .. ... ..... I.'
!. ' ~,. - .~

"
GASTRIC MILL
'"
REGION
~ ____ r,
: E I
I i ibe.s I
: (2) leG's
of 0
o
.... n ___ J0
PYLORIC REGION
A
Inl 1

E lPG N MGN LGN

al (2)

gm l .2,3a gm 4 c7

,,' cpv l a. lb. 2b ,2. p2b- 13

56
 265
the periphery, to ,lccomplish precise com n1.1nd or p.l ccmaker neurons ,1rC
Sectio n IV
ti ming. These can be rcg,lrded as s pecia l usually located in the medull'l 01' mid- Elfnlent.lrY Neurona l Ne two rks:
cases of the gene r,\l problem of ass uring brain and act th rough interne urons and Emergent Pro perties of Circ uitry
precise lim ing of sequences of activ ity. motor neurons on the electrocytes. In all
Among the best exa mples of systems C,lSes, these neurons are electrica lly
requ iring nearly exact synchrony are the coupled to each other via gap junctions
electric organ discharges of electric fi sh. (see p. 333). The electrotonic connections
(Ot her exa mples are the oculomotor may be between celi bod ies, or den-
neurons and the neurons innervating drites, or via presyn,l ptic fibers that form
sound-producing muscles and wings in electrical synapses on Ill.lny of the cells.
certa in ani ma ls.) The electric organ celis T hese synapses, by equa lizing the level
(electrocytes) are in most cases com- of depolarization between different celi s,
posed of modified muscle cells thai are are both excitatory and inh ibitory, but
oriented in seri es so th,ll Iheir depola ri- ensure synchronous acti vity of all cou-
zalions ca n be sum med. It is necessary in pled cells. The simplest exampl e of such
Ihese orga ns Iha t all the elect rocytes be coupling is in th e electric catfish,
activated nearly sim uli,lneously, and in Mn/npierllrJIs, in which there are two
some species wit hin 0: 1 Insec. Th e electromotor neurons, one on ei th er s ide

Figure 7.18 URdHS P~8t)

A si mple sys tem of so me th irty neurons; the stomdtog<lslric gdnglion of .I lobster. A. Side view 01 lobster stom-
..ch. T he two princi pal funct ion~ 1 divisions, the g.1strk mill region .Ind the pyloric region, are $Cpdr.. ted by ..
broken li ne. PMt of the st Ont.l t og~ st ric nervous syste m is shown togethe r with. few of the stom ..ch muscles
th .. t it innerva tes. T he stomdlog..stric g.1ngl ion (SIG) ",n be seen on the dOrs.l1 su rf..ce 01 the stom ..ch just
dbove g.>stric mill muscle I (gm l). Other lettering identifies muscles .. nd nerves. B. The two bdsic rhythms pro-
duced by the isolJ ted g~ngl ion Cd n be $Cen in the extr..cellul.. r recordings from neryes supplying the hvo
different regions of the stomdch. T he top thru nen'es (/U N, LGN, DGN), not all shown in pa rt /I.. supply
muscles th.. t oper.. te the g.lstric mill. Note thdt the bUr5ts of dct ivity be.. r a particuldr ..se rela tionship wi th
e..ch o ther . nd th .. t the d ura tion of cadi burst is ~lseConds. -The th ru lowe"i=tr..ces cont ~ in axons of
mot~ns (MVN, PN, dol VN) suppl ying p.1!2ric musclt!"s. The bunts are much shorter in d ur... ion, and the
o vera ll frequency is bou t seven times t ha~U!!.e g_~st~1. Note ~ I so t hat the bunts of ~ctivity ,;;afnt ~in d
pJTticu lar phJse rel ... ionship. The d.LVN t r~ce cont.. ins dxons innervating both regions (sec 'lbo\'e) and the long
bu n ts sun in this tr,lce ,I re from .. xons to muscle 8111 3 ~ . WoIrts A ~nd H, Selverston and Mulloney, 1974. 1 C.
Neurond l con nect ivity di agrdm for the lobster sto m.. tog.ilstric ganglion. All th e cells except int erneurons 1 and 2.
Me motor neuro ns. The top ten neu ro ns control the g..st ric mill cycle, and th e bo tto m fOllrtee n ce lls cO'ltrol the
pyloric rhythm. The dxo na l pathw"ys, ~s well as th e musc les inn ervat ed by the celis, Me kno wn. So mol, neu ro-
pile, and ~ xon .. 1 pMtS of the celis Me Indi cated o n the left. Broken li nes Mound so me of the neuropile Meas in_
di cate that ce lls of tl1.l1 group Me elec troton ically conilCc ted and can be co nsidered together. Kn own con nec-
tions wit h the ce ntrdl nervous system Me shOWIl at the top. Ro und dots represelll che mi c.ll inhibitory syll.lpse5;
tridngles represe nt chem icoll excitatory synapses and resistors (= electrotonic junctions): F, functional sy napse
wi th strong effec t but no clear uni t"ry po5t syn ~pt ic poten tial; f, exci tatory fiber input from commissuroll 8<ln-
gli... LPGN, Iolterol l post('rior g.st ric neu ro n; MGN, medi .. n gas tric neuron; LGN, latera! gastr ic neuron; ' "' I
,md 2. interneuro n neuron I .. nd 2; GM, g..stric mill neuron; DGN. dorS<l1 golstric ne\l ro n; AMN. ,Interior me-
di.lI1 neUTOn; IC, infe rior cMd iolc; VD, ventr iculM dilator; PD. pyloric dil .. tor; A8, .Interior burster; LP, l"'er .. 1
pyloric; PY. pyloric; eG, conlmissu r.. 1 8"ngli.; STGN. stomd tog.lstric nerve. ICourtesy of A. Seiv('r5ton. j

57

266
of the /lrst spina l segment, lightly sponse to IMturaJ stimuli . What can we
Clupl t'r 7
Integrdtlon .at Ihe Inlerp.edi.alt' uvels
coupled to each other. In ot her fi sh, the re say is the s implest ne rvously mediated
may be as many as 20- 50 p.lccma hr response?
cells coupled to each other, all of which The verlebrate stretch reflex, while
electrically excite internuncial relay cells specialized for simplicity and speed
that are also electrotonica lly coupled. rather thaI!, be ing primitive, is a good
G iven a synchron ized command, how starting poi nt because it is mono-
aTC electrocytes at di fferen t distances synaptic-that is, no interneurons a rc
from the brain activated synchronously? interpolated between afferent fib ers and
Thi s is accomplished, in most cases, motor neurons (Fig. 7.19). Skeletal mus-
either by systematically shortening the cles contain numerous sense organs,
lengt h of branches to the progressively called muscle s pind les, that a re se nsitive
mOTe dista nt electrocyles, or by altering to s tre tch in the ax is of the muscle fibers
their diameter so that the s lower con- (Fig. 7.20). Passive stretch, as by the
duction velocities ofaxons inne rva ting action of gravity or of other muscles,
the nearer electrocytes compe nsa te for causes a train of im pulses to arise in the
Ihe shorter distance. In several s pecies term inals of a sensory axon in a muscle
the gra ded delay is built into th e electro- spindle, and these are conducted via the
cytes. Dista nt ones a re innervated nea r dorsal root to the spinal cord, the re 10 be
their principal surfa ce, nearer ones at the distributed in axon collaterals to the
end of long, slow-conducti ng stalks of dorsal and ventral horns of the same
the elect rocyte. segment on both s ides of the cord, 10
The ex istence of these mechanis ms for nearby segments up a';;'d down the cord,
building compensating delays into neu- a nd to the dorsa l colum ns ending in
ra l ci rcuits opens the poss ibility tha t nuclei in Ihe medulla. Of these destina-
similar refinements of st ructure may be tions one is the large alpha (a ) motor
involved in other systems that are sensi- neurons of Ih e sa me muscle, wh ich a re
tive to the precise tim ing of inputs, such excited. This excitation is distribut ed by
as the parts of the auditory system re- the axon branches of the motor neuron
sponsible for sound localization or the to a group of mu scle fibers, usually
motor systems controll ing speech, eye between WO and 1000, called a molor
movements, midd le ear muscles, and the unit. Conlrdction of the molar unit tends
li ke. to ca ncel the s tretch. This proprioceptive
refl ex acts as a tonic muscle-length servo
(like a gyrocom pass), tending to main-
V. STIMULUS-TRIGGERED tain the length aga inst any change in
REACTIONS, THE load either way. Gravity is an import.lnt
ORGANIZATION OF norma l stimulu s; th is re Aex arc is the
REFLEX ES principal a nt igravity circuit.
AI th e sa me ti me, coll alcrals of the
With some of the principles of effector spindle afferents fil'c interneurons thai in
control. of sensory input, a nd of ele- turn exci te synergistic motor neurons
mentary circuits in mind, we can now a nd inhibit motor neurons of a ntag-
turn to the lowest levels of motor re- onistic muscles on the sa me side whi le

58
 FIgure 7.20
267
Til, mll"multiaH /IlI/SlI, sp;/Jdl,.

1111
'-H -t-7 raxons
J
Molor

Type Aa
Nuclea r affe ren t
c hain axon
Figure! 7.19
fibe r
T he monosYn.lpli<: slrttch rrAu p.ilhlV"Y in IllAmm.. ts, si m pl ifi~ by
oml .. ing Ihr ol her dtslindlions of Ihr s.am(' ..ffrrr ni n('U ron, Ihr oIher Type All
Annulo al/eren!
inputs 10 Ihe s.amt mo lo r nturon, ils oulpul rt.'Currenl coll.. ler~ ls, .md splrlll aKon
Ihe molor conlrol of Ihe spindle!. IGlrd uer, 1963.[ endings
Nuclear bag
Flower -'.-,...,~ fiber
s pray
endings
motor neurons of the homologous mus-
cle and its synergists on the contralater.ll ..... Sill S' orgjHl ,x(;I,d by j1fl$S illt 51ft/ell , Exl ralusal
side are inhi bited and antagonists ex- 0' IIdi vII/ioli 01 ;/$ 01011 ;III,i,Jsir ml/Sllt skelelal
fibtr5 by ga"",111 m% r axo/IS. m uscle fiber
cited. The phenomenon of reciproca l
innervation contributes to coordination
by prevent ing antagonists from working
against e.lcn other. noncontractile enlclrgement filled with
Given this iength- maint.lining re Aex. nucle i, and it is to this zone that the
how ca n the organ is m walk or volun- stretch receptor fibe rs come (Fig. 7.20).
tari ly ch.l nge muscle length? If higher The end regions of the s pind le fibers are
cen ters were simply to comma nd "" - contractile and are innervated by the
motor neurons to greater or less activity. y-efferents. W hen the y-effe rents in-
the stretch renex wou ld quickly cancel crease their activity. the spindle fibers
the effect. Somehow. the set point shorten, the cen tral region is st retched,
(SoHwert) must be changed. This c.lIl be and the stretch receptor fi bers are ex-
done by adjusti ng tension in the muscle ci ted. It is as if the muscle had been
fi bers within the spi ndle. ca lled in tra- passively s tretched. T he opposite effects
fu sal fibers. obtain if the efferents have reduced
There is another set of motor neurons activity or the main muscle has increased
in th e s pinal eOI'd that also se nd th eir activity.
e(ferents to th e mu scle. These are the Th e set point of th e postur.ll renex ca n
gamma (y) effcrents. emana ting fro m be affected by tens ion changes in the
small motor neurons (Fig. 7.21). T hey spi ndle muscle fi bers. When these con-
innerv.lte the muscle fi bers in the muscle tract they stretch the muscle receptors
spindle and alter its sensitivity. In the but do not change the lengt h of the
centr.ll region of the spindle fi bers is a muscle. T he resulting increase in stretch

59

>68
From brain
Chapter 7 2
Integr~tion at the Interm~dlale levels Spinal
cord
axon
Muscle
tiber
Figure 7.21
The yloop servomechanism. Commands (rom the brain may operate by contracting the mu~le fibers
of the spindle via the y-efferents in th e venlral root (2). This excites the stretch reflex (3) Jnd hence
the main muscle. There are also direct conllections (not shown) from the brain to the Ill,,;n motor
neurons. [MNton, 1972. J
receptor firing f<lle excites a -effercnts,
which elicit greater tension and hence
muscle shortening. Shortening continues
until stretch receptor firing rate is re-
duced again to normal values. The cen-
tral nervous system can command a
long-lasting new length by varying fre-
quency in the ),-efferents and thus
changing the sensitivity of the stretch
receptors with respect to muscle length.
The setting of muscle length by way of
the y-efferent is called y-loop activation.
The motor neurons (both y and n) for
one muscle are loosely grouped in the
ventral ho rn of the spinal cord. T hey
receive many inputs, usu.1lly in parallel.
The stretch receptor afferents excite only
the n-efferents. If they excited y-efferent s
the reflex would comprise a positive
feedback loop, which would run away to
I
maximum or minimum tension . Descend-
ing excitation or inhibition from the
brain impinges on both y- and Ct,-fibers.
We have already encountered the gen-
eral rule- that small units are more tonic
and have lower thresholds for n.1tural
stimulation. The y-neurons are smaller
and more sensitive than the Ct"s. A weak
command from the brain excites the y's,
which in turn change the set point of the
muscle-rece ptor-Ct'-efferent reflex, and
muscle length changes to compensate for
this. Strong descending-movement com-
mands excite both y's and n's. The
faster-conducting a's initiate a move-
ment that wou ld be later cancelled by
reflex function were it not for the more
slowly developing effect of y-excitation,
which changes the set point of the reflex.
Thus volunt.wy and other brain control
60

Go igi
tendon
recepto r
Figure 7.22
The tendon recep tor for st retch. Anothe r se nse
org~ n, besides the spindle, is in th e tendon,
which differs by being stretched (excited) when
the muscle con lr~cts as well .1$ when it is
lo ..ded. T he sign of its innucnce is such as to
preven t overconl r.l(lion.
of movement oper.l tes largely via the
y.loop balanced in vc1rious degrees with
coactiv<l tiol1 of a's.
Anoth er set of receptors is on the
muscle tendons, and these respond to
stretch of the tendon (Fig. 7.22). In can
tr<lst to s pind les, therefore, they respond
in the same d irection to im posed load
and to contr.1Ction of the mu scle (Fig.
7.23) . These receptors (Colgi tendon
organ receptors) have no monosyna ptic
endings, but, via inlerneurons, they send
(a) inh ibitory input to motor neurons of
the muscle they innerv.1 tc and to its
synergists, (b) excita tory input to molar
neurons of antagon ist muscles, and (c)
opposite inputs to motor ncurons of the
correspond ing muscles on the opposite
side of the body. T hese inputs might be
said to perform a tension- regulating
269
fun ction, restraining th e motor neuron
Seello ll V
from causi ng the muscle fibers to COll- Stimu lus-Triggered RNc tion5:
tract too violently. T he Organization of HeOu t'S
O ne more automatic s ubsystem is
important in hel ping to determine motor figure 7.23
Spimflt IIfrsrl5 I"rdorr 't(flrlor~.
neuron activity loca ll y; this is thc
j Pll llorr , 1965.J
Rens haw cell nega tive-feedback loop,
acting on motor neurons in the s pi nal
cord (Fig. 7.24). Situated in the ventral
roots there are branches of a-motor
axons called recurrent collalera ls because
they turn back and reenler the ventra l
horn to synapse with small interneurolls,
na med for Renshaw, who discovered
them physiologically. These cells fire at
high frequency when excited by motor
neuron output and have the effect of
inhibiting the same and neighboring
motor neurons. The roles of this inh ibi-
tion may include preventing motor neu-
rons from excessive activity, focusing
activity upon certain cells and perhaps
su ppressing phasic responses more than
ton ic ones. Muscle con tracted
Many refle xes are elicited by cxtra- SPINDLE RECEPTORS
" IN PARALLEL"'
muscu lar stimuli. For exa mple, painful
cutaneous s timulation often gives rise to
fl ex ion of a limb (Fig. 7.25). The fl exor
refl ex to nocicept ive stim uli is not ~
sharply loca lized; it may affecl all the Discharge
muscles of a li mb. Both strength of re-
s ponse and degree of s pread of response
are related to stimulus strength. Input
fibers do not impinge direct ly upon Musc te stre tched
motor neurons, but on interneurons; the
reflex pathw.1Y is polysynaptic. The
input excites flexor action and at the
same tim e inhibits ext ensor motor neu-
rons. If the stimulu s is quite strong its Acceterated
discharge
effects may s pread even to the conlra-
lateral limb, where they are opposite in
sign. The crossed extcns ion refle x pre- Musc te contrac' ed
pares one limb 10 bear the extra weight TENDON RECEPTORS
s hifted to it when the painfully sti m- ""IN SERIES'
ulated one flexes . Similar refl ex con-
61
270
+eo
Motor neuron
+<0

0

- 40

- eo
,-
.s,
+ 80
'~

~n +<0


0
0
D

, E -40

- eo
- 120

Motor

~"'"W" -60~
-70 _ __ :_p_s_p,

-80 I a!
0 30
5
" msec
15 20 25

Figure 7.24
The Renshaw type of inhibitory interneuron. Axon collaterals from motor neurons activate
Renshaw cells to high frequency discharge, which set~ up summa ting inhibitory postsyfl<lp-
tic potentials in neighboring motor neurons of the same pool. [Eccles, 1964.)

ncctions are seen for cutaneous touch,
pressure, and temperature receptors.
We have seen now some qui te genera l
contrasts between flexion ,lnd ex tension
reflexes, nociceptive .1nd proprioceptive,
pain and postural, cutaneous and
muscle-afferent, phasic and Ionic re-
flexes . These are overlapping but not
synonymous d ichotomies. Anothet use-
ful one, referring to their roles in causing
normal behavior is the distinction be-
tween elementa l and tuning reflexes; the
fonner cause the basic sequences, th e
laller adjust to momentary conditions.
Sherrington sta ted the concept of the
reflex, one of the truly great and fruitful
abstractions in biology, with these
words: "The Ullil reae/iOIl hi l1erVOU5 integra-
liOlI is the reflex. because every refl ex is an
integrative reaction and no nervous
action short of a reflex is a complete act
of integration" (1906, emphasis his).
Many authors have pointed out the
limitations of the concept or criticized it
as art ificia l, and Sherri ngton, as clearly as
anyone else, emphasized the non-
existence of a discrete circuit insu lated
from others. Nevertheless, the funda-

62

Nociceptor
/ in skin
Figure 7.25
Nociceptors .. nd the polysy n.. ptic p.. thw.. y. An
other !OCt of inputs impinging on th e motor neu-
ron comc fro m pain and sirn ila r receptors. T he
sign of their aclion is generally excitatory to the
ipsilate ral Aexor a nd con tra lateral exten so r mus
cles ~nd inhi bito ry to th e ips il ate ral extensors
.m d con tral ateral nexors.
mental usefu lness of recognizing this
category of responses has been amply
proved by th e ins ight that experi ments
based upon it have provided. The refl ex
is a useful abstraction, but we qualify the
bro.."ld sta tement that it is the unit of all
nervous integration: (a) the arousal func-
tions of the reticular activa ting system in
ma mmals cannot be resolved into re-
fl exes, nor can (b) mere sensing, (c)
autochthonous action (arising from
within), or (d ) many instincts.
Familiar examples of the phenomena
under consideration are the stretch
refl ex, the fl exion, crossed extension,
sha ke and scratch (dog) refl exes, the
refl exes of micturition and defecat ion,
,
271
the pi nna, swallowing, stepping, s neez-
Sec tio n V
ing, sal iv.ltion, blink, accommodation, Stimu lus-Tri ggered RCd cllo ns,
and tonic neck refl exes, placi ng, hop- T he O rga nlutio n or Hrncxe5
ping, and righting reflexes. T hey arc
re.ldy- mad e, un leamed, adapti ve move-
ments, prompt and coordinated . The
coordination is not observed just withi n
each refl ex, which we might ex plai n as a
fi xed pattern, but is observed even when
conflicting refl exes are sti mulated simul-
taneously, for th ere is al most in va riably a
resolution of the pote nt ia lly maladaptive
conflict or intermed iate action in fa vor of
adaptive selection .lnlOng them. Cooper-
ative interplay is also marked between
sim ple segmental refl exes, long spinal
intersegmental refl exes of foreli mbs and
hind limbs, and coord ination of body,
li mbs, neck, hea d, and eye refl exes, as in
visually guided wa lk ing.
T hough the present account of re-
flexology necessa rily draws mainly from
ma mmalian literature because lower
forms have been less studied in this
respect, we believe the princi ples
en unciated are probably gene ra I. Jm pl ici t
in the conce pt of refl exes-si nce there is
almost invariably more than one refl ex
util izing a given muscle, and hence more
than one cent ra l mechanism converging
on the sa me motor neurons- is the con-
cept of the fin al common path. The very
ex press ion em phasizes the int egrative
function.
The propert ies of refl exes (see Box 7.2)
and the ru les by which they are used
ma ke the best case for their reali ty and
importance. We have already learned
ma ny of these rules in the exam ples
detailed above. Let li S look furth er at the
ways they com bine.
Separate reflexes may be either com-
patible or incom patible. The fortner
combine adaptively into compound re~
flexe s. For exa mple, a grav ity refl ex that
63
272
keeps a fish u prigh t adds to a ligh t reflex
Chap ter 7
that keeps the dorsal side toward the
In tegration at the Intermediate Levels
light, so that if light comes from the side,
cert.lin fish tilt to a degree graded
according to the ligh t intensity, the
strength of gravity, and a central evalua-
tion thai multiplies each in put by a
weight ing faclor. The weighting depends
on time of day, temperature, hunger, and
other inputs, such as chemical signal s
associated with food, m echa nica l inputs
that ind icate a 5ubslr,liul1l, and visual
input su fficiently imaged and "under-
stood" to represent a substratum.
Incompatible reflexes are those that
cannot be accomplished at the same
time. If the hind leg reflex to scratch the
back is elicited on one side in a dog and
at the same time the sti m ul us for an
extensor thrust of the same sid e is given,
the nervous system d oes not release
them both, One o r the other is sup-
pressed, at each moment. Incom pa ti ble
reflexes do not add algebraically or com -
bine li nearly; instead, s witches or pat-
terned contro l ins ure normally adaptive
interaction. This may show either in-
hibition or facilitat ion . It is as though the
system were preorganized for useful
m ovements.
The tonic nec k reflexes and a number
of related postural reflexes d ue to
vestibular and proprioceptive input
interact with each other and with phasic
movements as though add ing a bias o r
"tuning" the res ponse to the conditions
o f th e m oment. For example, if a load is
li fted by wrist flex ion, mo re wor k c.1I\ b e
done (stretch reflexes faci lit.1ted) wi th
the head bent down or turned away from

Box 7.2 Properties of Reflexes
WIMt are the properties of reflexes that mani fest
integration? (a) The th reshold st imulus is very much
de pendent on conditions. (b) Above the th reshold,
gradation of response does not closely correspond with
grJd,ltion of st imulus, (c) If the stimulus is repetit ive,
there is USUJlly a poor correspondence between its
rh ythm and that of the reflex response. (d) Single
afferen t impulses are U511,11ly not adequate; temporJl
summation is usually necessary to el icit a response. (e)
A depressed excitability typically follows a reflex and is
oft en qui te long. (f) Afterdisch.uge, or the prolong.dion
of the motor neuron activity after the cessation of the
stimulus, is ,1 prom inent fea tu re of nl.lny reflexes, as
though the mecha nis m were org,lIlized to com plete a
certai n movement in a controlled W<1y. (g) Spa ti.l lly and
temporally patterned con trol of severa l muscles is
probably involved in .111 reflexes.
The ti ming of inhibition is coord inated with that of
excit<1tion, as is cleMly seen in the alternating reflexes,
s haking, stepping, .lnd scratching. Even sim ple flexion
and crossed extension reflexes show ch.lfacteri stic
tem poral patterning. (h) Irradiation wi th increasing
intensity of stimulus occurs in some reflexes-for
example, the protective flexion reflex. At th reshold <1
response lll<1y involve a lim ited part of a synergic
muscle group <1cross one joint, but with irrJd i,ltion it
may spre.ld to other joints of the same append,lge, to
other appendages and segmcnt.llievels, to the he,ld and
neck. The spre<1d is generally salt,ltory and is confi ned
strictly to certain Jines or muscle groups. However,
apparently uninvolved muscles m,lY in fact be involved
as objects of inhibition. The possible movements Me
thus cirnllTIscribed in a chM<1cteristic p<1ltern. These
properties help to define reflexes, to bring out thei r
integrative nature, and to emphasize th e centr.ll deter-
mination of detai ls of form and timing.

64

tha t Mill, because of the tonic neck re-
fle xes; if the Io.ld is met by wrist ex-
tension, the opposite head movements
e nh'lIlce ou tput.
Refl exes Me normally \'/Oven into a n
integra ted f" brie, without s harp li nes.
They Me gr"ded in amplitude by influ -
ences descend ing from highe r cente rs
and comb ined unde r the ru les of ,l
hie ra rchy. We ca nnot help wondering
whelher the sa me ci rcuits Ihal a re re-
fl ex ly t riggered from Ihe periphery might
be centrally triggered in patterned
sequences.
Hav ing erected this edifice of plausible
ass umptions and conclus ions from
studies of reflexes, we should now raise
the questi on as to wh.,t evidence th ere is
for the centr"l origin of patterned im-
pul se discharge.
VI. CENTI!AllY SCORED
BEHAVIOR: PATTERNING IN
SPACE AND TIME
One way to sta te Ihe fu nction of the
n('rvous system is that it formulales
appropriately pa tte rned messages to
drive the effectors. A core qu('stion in the
study of intermedia te level integrAtion is:
" How is this doner' T he patterni ng in
tim(' may be treated as a more serious
issue than tha i in space. Theoret ica lly it
could a rise in either or both of two W.1YS:
by followi ng (a) timing cues from
peripheral se nsc organs or (b) liming
cues frOIll central pace makers or patt el'll
generators. We may call the second
mecha nism ,I central scor(' to empha size
its potential complexity of detail, its
mod ifiability from outs ide on any given
occas ion, and its reali ty as a slored
progr.un dPdft from external inputs. A
score hdS ti mi ng built in, though subject
273
to exterlld l influen ce; a progr.1Il1 might
5~dlon VI
have the same, but the te rm ,' pplies also C~nt '~ lI y 5c:ored 8~h~y lor:
to cases tha t merely ca ll for reactions, I'.! tln ning in 5p~c~ .and Time
leaving the tim ing to effectors, trdns-
ducers, loops, .lnd largely peripher.l l
events.
Instead of a s imple d ichotomy we may
disti nguish several poss ible mecha nisms
(Fig. 7.26). In A, we have a s imple reflex, .
such as a n eye blink, a sWdllow, or a
cough, triggered by a stimulus. Sla rtle
responses media ted by gia nt ribers
belong in this category (see Box 7.3).
Even here the t('mpora l pa Hern 0
messages to various muscles is deter-
I
mined by central pathways and integra-
tive junctions. In B, the re is sensory
feedback from proprioce ptors ea rly
enough to determin e a rh yt hm of recur-
re nce; a chain refl ex aCCOli nts for fre-
quency, phasing, and amplitude. The
cxtreme case, in E, is purely centra lly
timed, without any immediate feedback.
C a nd 0 are combi ned mechanis ms in
which the central s ponta neity ca n deter-
mine the bas ic rhythm but feed back may
alter either the rhythm (C) or only the
details of the expression of the rhythm
(D).
Probably all five mechanis ms a re com-
ilion, though us ually there has not been
enough analysis to be s ure which class
a n activity belongs to. In this section we
are concerned particul.uly with some
examples of C, 0 and E. In each
exa mple, there is permiss ive or essentia l
input from sense org,lIlS that may start or
stop the whole pattern or innu ence the
overall "central exci tatory state," to lise
ShNrington's phrase. This is what lVe
mean by spontan eit y, not that there is
independence of the envi ronment for
permissive conditions, bu t only for trig-
gering the s uccession of ,1(tions. S pon-
taneous rhyth ms ca n be in nue nced in
65

A B c o E
Fig ure 7.26
Five mechanisms of pattern fOTmulat ion. The three levels of neurons are understood to represent
brJnch ing cha ins in whose functions integrat ive p roperties may a lter the actua l impulses ,md distr ib_
ute them spat ially as well as temporilily to the effectors (bollo"I). A and B are shown with receptors;
C, 0, and E, with spontaneous p<lcema kers giving simple or group dischMges. Band C have proprio -
cept ive feedback acting on the trigger neuron; D, o nly on the shaping of the pattern. [Bullock, 1961a.)

Box 7.3 Giant Fi bers and Startl e Responses

Striking among the behaviors tha t are merely trigge red
by environmenta l st imuli, but not further guided by
either environmental or propriocept ive inputs, arc the
responses med iated by giant fibers . Giant fibers are
Found in many invertebrate .m imals and fish (see
ChJpter 10). T heir phyletiC d istri bution is scattered Jnd
their st ructures are diverse; hence they a rc likely
correl<1 ted only in func tion, not through evolutionJry or
developmental relationships. BecJuse of this diversity,
the follOWing generalizat ions l1lily have exceptions.
Beciluse of their size they conduct impulses TJpid ly.
Perhaps related to size is the fact that they can have ,1
large divergence rJtio-that is, one giant fiber can excite
mJny other neu rons or muscle fibers . Moreover, the
curren t available for electricJI trJnsmission of the
act ion potenti<l l to downstre,ml neurons is large; elec-
trica l trans mission may be common in these systems. In
all of the adequ,ltely studied C<1ses gi<1nt fiber ac tion
results in rapid, neJrly synchronous, widespread mus-
cle act ivity. Behaviora lly, g iant fibers ordina ri ly fire
only with rather special input requirements; these often
have the ch'Hacteristics we call startle.
The giant axons of sq uid arc the best known of all
nerve fibers. We may briefly review their funct ional
ana tomy (see <1lso Ch,lpter 10, p. 434) and behaviora l
role. The muscles of the mantle of the squ id a re doubly
innerv,lted. Many small mo tor axons From the stellate
g,l1lglion excite rela ti vely few muscle fibers each, and
these slnJIl neurons control the slower movements
involved in respir,l\ion and ordinary swim mi ng. The
single giant fibe r in each of the 6 or 8 stellar nerves on
e,1(h s ide together innervate most or all of the muscle
fibers of the m'11l tle. EJch giant fiber has In,my cell
bodies in a lobe of the g,lnglion and is therefore a fused
syncytium of many Illotor neurons. "The" giant fiber of
the squid is the I<lst, longest, ,md I,lrgest of t he 6 or 8 on

66
 275

Box 7.3 (wlI/'-mud )

c.leh side; these fibers arc properl y thc third~ordcr
g ianis. Their input is from two second-order giant fibers
Ih.lI arise in the viscera l lo be of the bra in, enter the Tactile rBceptors
stell.ltc g.lnglion, and synapse all each third-order giant
fib er in tha t gilnglion. The 8 i.1nl synapses are one-
to-one re lays; every input impulse causes an all-or- Anlilacilitating
no ne twitch throughou t the mantic, nearly simulta- 'h~'r".p""
neously, .md a vigorous ejection of water through the
funnel . When a squid is slilflled. it aet iv.lles Ihe g iant
fiber system probolbly Viol the 5ingle bilaterally fused
~, ","W", ,,,""
firs t-order giant unit in the ped.d lo be of Ihe brain. This
unit is .J cornm,llld unit. as ddlncd o n p,lge 279. There is
pro bably no immediate feedback influcndng Ihe oper-
1L / B
'"""'''00'
a tion of the giant system.
LG ;---electncal synapses
The gia nt fiber syste m of earthworms consists of two
pd ra llel cha ins of elec tric~!ly coup led segmental COl11-
Lateral giant fiber /
"
mand units: a median fiber and a latera l eledrotonically en and iJ responses) / Rectifying synapse
coupled pair of fibe rs (see p. 405). These premotor
inle reurons control largely overlapping musculature,
the longitudina l or shortening muscles, plus separate Fast lIexol motor
neulons MG MOlor gian t cell
muscles for the setae of a nterior .lnd poste rior seg-
ments. The inputs h.we some labi le overlap. The
median giant is activated by stilrtling stimul i, such as a
vibration or a tap anywhere o n the an terior third of the
worm, and causes anchoring by pro trus ion of setae in
the tail, plus shorten ing, which therefore retracts the FleJ;or musculature
head. The lateral giant pair is activ<l ted by mecha nica l
stim uli in the posterior two-th irds, and causes anchor-
ing of the head end, hence pulling up the tail. Each giant Zucker's dwit for the Tdpid tail nexion of the crayfish.
Schem~ti(' diagrolm of the known elements and ronnl"Ctions
is the re fore a unique, consistent chain of neurons acting
as a decision unit. The adeqUAte input is from m,my for ph.sic mec hanical 51imuli to abdomen. [Zucker, 1972.1
receptors, dnd redches thres hold only when some su btle
criterio n is met thdt involves ra te of rise, recent history,
s p.ltial pattern, and a centrill excitato ry st.lte.
The crayfish giant syste m hols been more completely initial body bend of a s tdrile res ponse. The wea lth and
studied, and its circuitry is diagr,ul1l11ed in the sketch. v<1rie ty of synapses known-on the so ma, axon hillock,
Ma uthner's fibers in teleosts and .1Cj uatic amphibia ,lI1d large dendrites-sugges t tens of thousands of im-
are likewise premotor COllltll<lnd units receiving a large pulses may arrive per second, during the reaction time
input from many sources, especially vibrdt ion receptors of a single afferent impu lse, representing a high degree
(see pp. 30, 106,257, 440). They normally fire only once (but perhaps not dtypical for neurons) of integrative
o r twice, firsl one side and Ihen the olher, causing the filtering, recognizing. decid ing, .1l1d com manding .

67

276
occurrence and Frequency both by tonic
input and by higher central levels (e.g.,
by changes of mood).
A. Central Rhythms
and Reflex Modulation
The peripheral or reflex hypothesis fails
to account adequately for respiratory
control of vertebrates. A crucial test, total
deafferentation, leaves a functioning
central system. Similar tests have shown
that many other rhythmic control sys-
tems have built-in central scores. The
motor neuron discharge seq uence, which
withdraws the mantle and closes the
valves in a clam, Mya, occurs after de-
afferentation. The copu latory move-
ments of a praying mantis, the flight
rhythm of insects, the walki ng patterns
of insects and amph ibia, respiratory
movements in insects, beating of the
swimmerets in decapod crustacea, strid-
ulatory singing in crickets and grass-
hoppers, side-la-side alternation of
longitud inal muscle activity in sharks,
heartbe.1t and gastric mill contro l in
crustacea, and the swimmi ng beal of
jellyfish have been shown with varying
degrees of rigor to be centrally con-
trolled.
A few of these examples deserve more
detailed discussion. But first we should
poi nt out that proprioceptive feedback
functions have also been demonstrated
in nearly all of th em, and in the dis-
cussion of centrally patterned motor
output we should attend to the role of
reflexes as well. In studying cases of
oscillatory behavior, as in all considera-
tions of oscillatory phenomena, three
measures always merit notice: frequency
(or the reciprocal of frequency, th c pe-
riod) of the osci lla tion, amplitude, and
phase of one element of the oscillatory
pattern relative to another. In severa l of
the known centra lly driven beh.wiors,
we can relate the functions of peripheral
feedback to one or more of these pM-
ticular parameters.
Lomsl Flight . One of the most decisive
studies on central rhythms is tha t of
Wilson on wingbeat control in grass-
hoppers. These ani ma ls will sometimes
fly when the g.1nglia of the head and
abdomen are removed, so that the pat-
tern generator must be present within
the thoracic segments. Th e normal motor
score has been analyzed in such detail
that the temporal sequence and phasi ng
of every motor neuron impulse during
Aight is known. This pattern of motor
neuron discharge can be recognized in
the central stumps of the thoracic nerves,
even after all those"'11erves have been
severed. Isolated thoracic nerve cord
preparations are not spontaneously ac-
tive in the flight rhythm, but random
electrical st im ulation of the nerve cord
ca n elicit nearly the same pattern as is
found in intact animals during Aight. The
output pattern of deafferented prepara-
tions is defi cient in one major respect; it
is low in frequency. The decre.1sed fre-
quency can be ascribed to lack of input
from four stretch receptor cells, one in
the hi nge of each wing. These stretch
receptors discharge when the wing is
elevated. During flight they fi re one to a
few impulses toward the end of the
upstroke in each wingbeat cycle (Fig.
7.27). The number of impulses in the
burst is correlated with wingbeat ampli-
tude, the timing of the burst with wing-
beat phase, and the burst repetition rate
with wingbeat frequency -all the meas-
ures of a n oscillation. This information,
68
 277

Sec tlo n VI
Cenerall y Scored Behav ior:
l'.IU ern ing in Space and T ime

I I I I I I I I I , I ' I I I I I I I I I I I I I I I I I , I ' I I I I I I I ,I
o '00 '00 300 400
Tima (msec)

Fig ure 7.27

Propr iocep tive feedbac k in a fly ing insect. ~nsory dischMges in nerves from the wing and
wing h inge in a locus t, recorded with wires manipul,l ted into the largely eviscerated tho-
racic cavity of a locust. The top record is of downstroke muscle potentials, which are re-
pea ting ,It the wing-beat frequency. The bottom record is of a sensory (stretch) receptor
from one wing, firing one or two times per wing beat. [Wilson, 1968.J

upon enteri ng the eNS, would be ade-
q uate to trigger and control the events of
th e next cycle. Appa rently, however, it
does not even signi fi can tly affect the next
cycle. In spite of th is rich detail of in-
formation a bout wing position, there is
li tt le input/output correlation except a
relatively long-termed one relati ng aver-
age input frequency in th e stretch re-
ceptofs to average wingbeat freq uency
and ampli tude. Th e ganglion integrates
and smooths the input over several
wingbeat cycles, thereby almost bu t not
quite los ing the phasic informa ti on. The
fil tering process is analogous to th e
smoothing and integration in a re-
sistance-capacitance electrica l net work.
In su m, locust fli ght reaffercnce primar-
ily plays a role in con trolling the average
excitation of the motor p.l t1ern genera tor,

69

278
thus affecting wing beat frequency and
Chapter 7
integrdti OI1 at the intennedidte levels
power. It has a very we<lk effect on
wingbeat phase.
An impo rt.l llt lesson to be Je.1!"ned
from this exam ple is thai, even though
one can demonstrate that proprioceptive
input in a rhythm ic syste m fits the re-
qu irements for a refl ex feedback model,
that input may not be necessary for th e
normal pa ttern, and significant infor-
mation parameters for the peripheral
hypothesis may not even be used in the
normal operation of intact anima ls. They
may be discarded in a filtering process.
Swimming ill Slzarks. When many fi shes
swim, the longitudinal musculature of
the two sides contracts in alternate
metachronal waves. If a shark is curar-
ized to the point of total paralysis, motor
output in the segmental nerves may still
result when the anima l is stimulated, but
since no movement occurs there can be
no correlated proprioceptive feedback . In
th is circumstance, bursts of impulses still
issue alternately through contralateral
nerves, but at unusually low frequency.
Proprioceptive input may, as in locusts,
be necessary for ton ic excitation of cen-
tral state. But in sharks, compMison of
the outputs on the same side in different
segments shows them all to be syn-
chronous. The metachronicity is lost. On
the basis of presently available evidence,
it appears that proprioceptive feed back
is necessary for the phasi ng of outputs in
proper segmental seq uence as well as for
the ma intenance of normal frequency
(see Fig. 10.63).
Crayfisl! find Lobster Swimmerei Beal. The
abdomina l appendages of decapod
crustaceans also have a metachronal
rhythmic beat. The completely isola ted
abdominal nerve cord can be stim ulated
to produce th e swimmeret beat com-
mand, and the output of a deafferented
or isolated preparation is normal in both
frequency a nd segmental phasing (Fig.
7.29). The known proprioceptive reflexes
cannot even modulate these parameters.
They do, however, modulate th e force of
each st roke, affe cting the velocity a nd
amplitude by influencing the number
and repetition rate of motor impulses
during each cycle-in other words, the
magnitude of the motor discharge.
B. Command Cells
A superficially quite different category of
centra lly scored pattern is that called up
by com nl.1nd cells, actually just the e nd
of a spectrum of mechanisms that occur
in all degrees. Command units, first
discovered in crayfish, are now known in
many arthropods, annelids, molluscs,
and vertebrates (see Box 7.3, p. 274 ).
Com mand units, or redundant clusters of
similar units, may turn out to be q uite
general. The concept and the term come
from the observation that certain single
un its, upon sti mulation, are capab le of
causing actions rese mbl ing major pieces
of normal behavior. Some c<luse static
postu re, others phasic sequences. In
higher invertebrates, in which they Me
prob<lbly all potentially identifiable, non-
identical s ubsets of com mand cells may
act together to determine the form of the
behavior. Th ey are presumed to trigger
or release the action, not to instruct the
motor neurons in the te mporal patte rn of
their firing; that pattern is alre.1dy pre-
formu lated by other neurons.
The best-known command fi bers in
cf<lyfish are not of extraordinary size, are
not motor neurons themselves, and in
general do not even synapse d irectly on
motor neurons. They are high er-order
70
 279
interneurons that run through several CM10

CJ
62 / 1
ganglia or even the whole length of the , 60
neuraxis. They excite whole motor 0", ,
- r 63 ,
' '75
,; ;
systems, small or large, controlling pos- L. __ ,. ~; 1 74

ture or locomotion (Fig. 7.28). They are

found repeatably in different anima ls.
--"\ --7,. . . ---}- --
,
4 1 65 "\ - - ;
, 67 , , , 73 ,
61

66 ,, /69 , 71 '>' 72
Each is uniquely characterized by posi- --.J ~- ,
tion in the nerve cord, axon diameter, 68 70
output function, and other properties.
The pattern of output does not depend A
much upon command fiber frequen cy or
pattern, and the output may continue
well after the stimulation of the com-
mand fiber has ceased.
The command fibers are obviously
labeled lines. The temporal pattern of
activity in each is relatively unimportant.
What is important is which command
fibers are active, since they drive diverse
behaviors. Each command fiber con-
trolling posture of the crayfish abdomen
seems to have unique but overlapping
output fields. Perhaps the severa l fibers
commanding swimmeret beat are also
unique and produce somewhat different
B
outputs (Fig. 7.29). Whether they do or
not, swimmeret movements can be
differentially controlled by combining
the action of fibers driving the OScillatory
mechanis m with those affecting posture.
Command fibers m.1Y turn motor
system s on or off or bias them for pur-
poses of steering or orientation.
Some giant fibers (see Box 7.3) are
command units- namely, those that are
not motor neurons but interneurons re-
ceiving from many small cells. They are
specialized for prompt, synchronous, and
brief action s that do not continue after Figure 1.28
the giant fiber stops firing. local sites in A. A comm and fibe r for posture. This fiber, called CMlO, travels in area
the hypothalamus of mammals, where 75 of th e circ um eso phageai conne cti ve of the cray fi sh. Wh en stimul ated at
stimulation rather reliably triggers char- a minimum of 20 sho ch per sec, it release s the "defensive postu re"
shown in H, invol ving stereotyped positions o f all appendage s. This is one
acteristic behavior, may represent a of the stati c, postura l respo nses trigge red by anyone of a small numbe r
cluster of cells acting as a kind of com- (3-5) of sp ecifi c co mmand neu rons; othe rs are dynam ic, rh ythmi c move-
mand cente r (pp. 316-319). ment s, such as swimmere! beat ing (e). [l'art A, Wiersma, 1958.J

71

280
Command neuron
Ch"pter 7
Inte gration at the Inte rmed iate Levels

Coordinat ing neuron

~~r-,--~(--___~

Fig ure 7 .29

Circu it for swimmer,,! beating in crdyfls h. T he segmentally repeated circuit is connected
both by th e multisegmental command neuron dnd by the intersegmental coordirMt ing neu-
ron tha t controls the ,lelual time relatio ns 0( the wave of beating. [Stein, 1971.]

What sets off a command cell? In
general, we do not knolV, but an in-
formed guess might be that it usually
requires a number of backgrou nd condi-
tions plus some triggering input. This
means it is highly integrative, acts as a
recognit ion un it (p. 236) to detect these
criteria, a nd is therefore a decision unit
(p. 238) in a significant sense. The input
it requires may come fro m sense organs,
but in some cases it seems likely that
centrally a risi ng changes in mood or
re.ldiness might do the same th ing (see
vacuum activity, p. 315, central spon-
ta neity, p. 325).
C. Hierarchical Structuring
of Molor Systems
If rhythmic motor output can be driven
by nonoscillatory impu lse trains in the
command fibers, where does the oscilla-
tion arise? Cou ld it be due to interactions
between the motor neurons? Several
sorts of motor neuron interaction a re
known . Mutual excitation and reciprocal
inhi bition between motor neurons in the
stomatogastric ganglion of crabs and
lobsters are respons ible, at least in part,
for the way in which the output pattern
of that ganglion rein forces a nd ad justs a
spon taneous rhythm or command fi ber
response (p. 279). Some motor neurons
exciting the same muscle in insects are
electrotonically coupled and tend to fire
together. The sa me is true for contra-
laterally paired inhi bitory motor neurons
in nervati ng the a bdominal postural
muscles in crayfi sh . Motor neu rons in-
nervati ng th e same flight m uscle in cer-
tain flies inh ibit each other. Synergistic
motor neurons in vertebrate spina l cords
are a lso inhibitorily lin ked, throug h
known short-.lxon inte rne urons, the Ren-
shaw cells (p. 269).

72

If interaction between motor neurons
were genera lly responsible for thei r
rhyth mic activity, one might expect that
ant idromic stim ulation of sets of motor
neurons could reset or modify the output
of the whole network. In genefil!, this is
not true. At a more detailed level of
analys is one wou ld expect to find th.lt
intracellular stimu lati on of one motor
neuron would give rise to synaptic po-
tentials in functionally related ones, but
aga in this is not generally true, or else
the effects are quite weak. Current
thought on the matt er is that motor
neuron interactions are usually in-
adequate to account for the patterns of
output in motor systems.
This concl usion leaves us look ing for
a process, or even a structure, that
mediates between the com mands and the
motor d ischarge itself. The search ison for
interneurona l pacemaker cells or net-
works th"t produce rhythmicity(Fig. 7.29}.
Cons istent with the notion that there
is a hierMchy in molor systems, with
input co mmands driving osci ll ators that
drive motor neurons, is the fact that the
same set of motor ne urons ca n be used
in more tha n one behavioral pattern.
Frogs swim or jump with synchronous
ou tput to homologous cont ralateral
muscles, bu t alterna te them d uring
walking. Insects use some of the same
muscles to move the wings and legs, but
they do so according 10 different
synergistic/ antagonist ic relationships,
depend ing upon whether the com mand
says "walk," "jump," "sing," or "fly."
We are pu shed into thinking th"t even in
the lower ga nglia or s pina l cord there is
a mu ltiplicity of p"Uern gener.ltors or
oscillators that can each be turned on or
off or be modul.1ted in frequen cy or .1mp-
lilude by command input, and thai
these pattern generators each converge
281
upon identical or overlapping pools of
Section VI
motor neurons. Thus the motor neurons C~"'tr.l.lly Sco<('d Bth~Yior'
Me, to use Sherringlon's phr,1se, the r~tt (' rnlng In S p~c ... ~nd Timt
"final com mon paths" transm itting to the
muscles signa l piltt em s th"t are pro-
duced at a higher level.
If we turn from a preoccupat ion with
the genesis of rhythm to the question of
how alternative motor patterns that in-
volve higher level switching are selected
and programmed, we have to dea l
mai nly with conce ptual models that
seem compatible with principles of
physiologica l organization.
The rather wide ly accepted notion
today is that actions commanded by
higher centers are not specified in detail
by those centers, nor primarily deter-
mined by peripheral stimuli, but trig-
gered in a preprogrammed language
ca lling up combinations of elementary
acts in a hierarchy of levels. The highest
command, as well as the lower-level
instructions to still lower levels, may
simply specify the sequence, strength,
" nd duration of the nex t lower compo-
nents, finally elicit ing moveme nts as
though the adequate periphera l stimuli
had occurred that can reflexly elicit
them . Ana lysis of six ga its of horses
s hows that they could be produced
simply by ca lling up components eq ual
to certain loca l spinal and long spinal
reflexes in formulat ed sequences and
durations, given a few fixed rules
(Easton, 1972). Whether the horse
actually works this W,lY, we ca nnot tell,
but the !'ul es are simple, and the model
might work. Of course, superimposed on
wha tever "c" lIing up" the bra in initiates,
proprioceptive input as well as visual
and s01llesthelic ]'eafference (sensory
input c"used by one's own actions) wi ll
be important in shaping and correcti ng
the centrally patterned prograJll, by
73
2.2

!'DI---==-
l
Comma nd

+.-----...----''-<G)I----==- L""".
Ton ic

Depressors

Ph asic
(campanlfo rm
E)tci lalory se nsi ll a)
Inhibitory
A

Feeding
pacemaker
Com po nen ts 01
intri nsic
program

Positive leedback
c~' ..'''~~
ConUllcUon loop sustaining

j re tra cto r burst

S tress ing 01

J
mechano receptor,
Ret ra cti on 01

An tagonist ~j,,.;..,
+ .,.... buccal mass

inhi bi tion Mechanore<:ep tor

activity
B
I-
Destressi ng 01
mechan'o recepto rs

~----~~----
Delayed nega tive l eedback loop
ending mechanoreceptor bu rst

74

s imple adjustments of strengt h and
ph.1Se of pdrticular components.
Motor patterning systems have been
so incompletely s tudied that we kn ow
velY little of th eir actua l mechanisms.
Parti,ll models with some s upporting
evidence ca n be made in many cases.
(Examples are s hown in Figure 7.30 and
in the figure in Box 7.3, p. 275; see also p.
333.) Since it has been so diffi cul t to
ma ke an ana lys is of even relatively sim -
ple motor syste ms in terms of individual
neuron activities, perha ps we shou ld ex-
pect 10 fin d eventua lly Ihat presen lly
un known concepts of neural function are
involved.
283
D. Centrall y Scored Pattern
Se(tlo n VI
by Sensory Tape Cen trall y Scored Dehulor:
Patterning in Sp.ce a nd Time
Another possible mecha nism of central
programming, besides the motor score,
has been suggested by Hoyle. He calls it
a sensory la pe, or 10 use the phrase of
the ethologists, a sensory templa te (sec
p. 309). Sensory tapes or templates have
not been demonstrated, though strongly
inferred for the control of some bird
song. T he idea is worth more discuss ion.
Suppose the CNS contained an instruc-
tion tha t sa id, " Produce a motor output
that results in a s peci fi ed feedback from

Figure 1.30 (f~,iQg ,.~gt )

Circu its for inso:><:t w,ll k ing (A I and sn~i l feedi ng (8). A. Hypothetic .. 1 scheme for the ob5erved d is
charge pdUerns of levator (5 and 6) and d epressor motor axon s (D). A bursting interneuro n (bi) is
excited by the comm~nd neuron and in lurn excites the levators whi le inhibiting depressor motor
neuron s. The comm,md nber is believed to exc it e the depressor, so t h ~1 ~n incre,lse in command
input decreases int erburst inte rvdl while producing ~ less marked decru se in burst durdtion. Ce rl.lin
se nsory in put toniCollly (.. cilitates the bi ~ nd inhibi ts D; other input ph ..sic.Uy excites D dnd in hib its
bi. [pe.. rson .. nd lies, 1913. 1 D. Diolgr.. m showing the funct iondl interco nnec tions that give rise to Ihe
tempor.. 1 re l.. tions o( dCl ivity in the retr.. ctor .. nd protr.. ctor elements o f the 2S pairs of lIIust tes re-
s ponsible fo r feed ing in the s n.. il Ht/isortllf. The neu rons of the p.Ktm"ker generdte dn .l utonomous
rhythmic output th~t olppeMS with differen t "mounts o f ph ..se shift in the \'o1rio us motor neuron s.
The rhythm is symbo li ~ed by the saw tooth Wolve ~ nd the two represen t,ltive populations of mo tor
neuro ns at the lop of the di,lgr,lIn (upper bra ce, right -hand side of diagram). T he retrdcto r neurons
drive the retractor mu scles. Co nt ract ion in th ese mu scles exo: it es mech~norece ptors, ~nd their ou lput
is fed back pos itively 'lid exc it atory syna pses on the retractor motor neuron s. T his posi tive- feedba<:k
loop (Io\ve r br ..ee, righ t-ha nd side) sust" ins the re tr.. cto r bu rst. Contrdction o f the retractor muscles
p rod uces, ..fter .. de l.. y (due 10 exci t.. tion-contract ion coupling .. nd the Vi5Coel.. ~ic .. oo inert i.. 1 prop-
erties of the system), .. retrol("t ion of the bucc.. 1 mdSS. W hen it is retr.acted, cont r.. ction of Ihe muscles
no longe r stresses the mech,Uloro:><:eptors, .. nd th ese shut off; Ihis series of events constitu tes a nega-
ti ve-fCi'dback loop wi th deldY, ,md it limit s th e retrdctor burst by ope ning the positive-feedbd ck loop
,1(ler relra ction is comple le (low brdce). The protra clo r motor neurons and mu scles dre excited 011 Ihe
op pos ite phase of the cycle from Ihdt o f retractors. They fire un til in h ibited by inpu t (rom mecl1dIlO.
re cep tors, whi ch Me stretched by the contracting re tractor muscles. This accomp lishes a unidlrec
Uondl ,m tdgon ist in hibition with de lay, wh ich .. lIows the protract ion ph .. se to be sustained u nt il re-
tr.. ctor tension is developed. Ant.agonist inh ib ition in the re\"('rse d irect ion (i. e., protr.actors in hibiting
retr.lctors) is absent, ~nd thi s allows the O\"erlap ping activity in the two groups of muscles .It the
beginning of the retraction phdse. (Kater and Ro well, 1913.1

75
,
264

MOlor ta pes SenlOlY tapes

In ternal stal e ,..'--''--',

Tape "clues"
leleC lo r

Compule ' rom

-
Compare
di fference

Genera! Specillc
d river driver

Propriocep tive
afference
Spec ilic
driver

Othe r
"","U

Xp".,
2nd- o rder
driver

Extenso r

Jo int
A

76

proprioceptors or other sense organs."
This ins tructi on would not lead inevit-
ably to a sing le s tereotyped motor out-
put, as from a molar score ge nerator, but
it could guide motor out put to achieve a
goal (Fig. 7.31). Ou tput might at first give
in correct results, but feedb<1Ck cou ld
modulate th e output on successive cycles
of loop operation . Consistent with th is
notion of central progra mming by com-
parison of sensory feedback with a
centrally slored goa l pattern is the fa ct
that diverse motor oUlputs may be asso-
ciated wit h a pparently identical leg
movements during walk ing in insects. In
some cases the flexors and extensors
alternate. In others, one muscle contracts
tonically whil e the other oscil lates. The
resulting movemcnt is thc same.
T he only reasonably strong case of a
sensory template is found in bird song
control (see Fig. 8.18). In the invertebrate
cases in wh ich a sensory principle may
be operat ive, it seems to be supe r-
imposed upon a motor score type of
central generator. Insect fli ght is basically
programmed by .1 motor score, but ex -
teroceptive as well as proprioceptive
inputs can modify that score for pur-
poses of s tability, s teeri ng, or com pensa-
tion for inherent error or damage to bod y
paris. Perhaps in these systems the no-
285
tion of a sensory template n~ ally redu ces
Seell a n VI
to reflex modulation of a motor score. Celltrd ll y Score.! Beha vior:
r dtt ern ing in Sp.lce ~ nd Ti me
E. Coordinated Movement to Gross
Stimulation of the Brain
A clear progression is ev ident if one
compa res the responses to crude elec-
trica l stimulation of s tructures at succes-
sive neural levels. Ventral roots give a
segmental, loca l contraction closely re-
lated to the duration and strenglh of
stimula tion. Lower motor centers in the
cord or bra in stem give little more,
though the d istribution m.1Y be mOfe
functional (e.g. flexion of certain joints).
The responses relevant to this section are
the quite normal actions involving
sequences of 'movements, s uch as ca n be
elicited from the hypothalamus of
mammals (p. 471). These are so natural as
10 suggest a centra l pattern, si nce the
stimuli are like lightning bol ts. A pocket
mouse may s tuff invisible seeds inlo its
cheek pouches at a high rate, a cal may
arch its back, hiss, unshealh its claws,
erect its hair. The value for our purposes
is the same whether we assume that the
stimuli trigger motor patterns or sensory
" hall ucinations": the patterns are cenlral
and need only an adequate trigger.

Figu re 1.31 Ullting ,111<')

Two types of cont rol by centrally determined seq u('nces. A. System dri ven by sequellces Ih ~t deter-
mine motor outp ut directly. Driving comm ~ nds c~n be gener~ 1 or s pecific to dny degree of det.!il,
,md .It a lower level th ey Cd ll be modified by prop ri ocepto rs. B. System drive n by tdpes of se nso ry
feedbdc k th .. t must be expected. A com pd(~to r m .. kcs th e dctl1~ 1 (OIllIllJnds on the bdsis of the
differences between the proprioce ptive dfferenee dnd the inst ruction from the tJpe. Th ere c.. n dlso be
proprioceptive (onlrol lower down, ..s before. Thi~ syslem is more "d"ptdble, bu t requires much
more circui try in dddit ion to d COlll pd ril tor ~nd .. COlllllldnd center, which th('m~I"es musl produce
high ly compleK .Id.l lJlil'e sequences of impulses. Most inverlebr.lte responses inveSlig" tcd usc .I n " in-
line'" syslem, ,15 in A. IHoyle, 19M. 1

77
286

Box 7. 4 Chronology and Background of Ideas on Ihe Ph ysiology of Ihe Nervous System

We present here i\ selection of highlights in the history
o f ide,IS, from the ('.Hlies! till1es up to 1929. The inte r-
pretation of the brain in terms of cells is highlighted in
Box 3. 1, p. 102. The roots of bra in chemistry, membrane
bio physics, ph,umacology, sensory and psycho-
physiology, .md be havioral a nalys is arc not attempted
he re.
The redder is urged to look furthe r, fo r more balance
a nd adcquille representation. Some useful, more-or- Iess
conde nsed accounts aTC by Nordcns kiold (1935),
Dampier ( 1948), Singe r (1959), Brazier (1961). Sirks and
Zirkle (1964), Gard ner (1965), Clarke and O'Millley
(1968), and McHe nry (1969). Specia l as pects are dealt
with in FeiHing (1 970), Brazier (1961), and Swazey and
Worden (1976).
1700 B.C. An Egyptian document, transl,l ted cen-
turies later ,lnd published ,lS The Edwin Smi th Surgicdl
Papyrus, includes 13 Cdse descriptions of head injuries.
Aphasia, paralySiS, and seizures were descri bed, and
suggested the fun ct ions of the brain. Nevertheless,
diseolse continued to be generally oltiributed to suprd-
naturoll influences olnd whims of the gods.
800 B.C. Homer's works and the flowering of Greek
iHt and intellectua l life led slowly and incomple te ly to
the idea thilt the world is knowilble.
500 B.C. Alcmaeon performed the first recorded
d issection of ,l human body. He p,l id some ,ltiention to
the brain and discovered the optic nerves. His teache r,
Pythagoras, tilught th.t t the brain is conc::erned wi th
reasoning. In the next century more dissections and
similar speculations were made by others.
400 8.C. Hippocr,ltes of Cos countered the mys tics
and the entrenched supranaturalists to introduce ra-
tiona l medicine. This requ ired the systematic accumu-
lation of cl inic::al experience, and his desc::ription of
e pilepsy went unsurpassed un til the work of H ughlings
Jackson. However, Hippocratic teaching was dominil ted
by the idea tha t fun ction derives from the combination
of four humors; blood, phlegm, ,lnd black and yellow
bile. The brilin is the org.tn of intelligence .md d re,lms,
but it a lso secretes phlegm OWd cools the blood. Even in
the Gold en Age the Greeks- d id no l easily fo llow his
exam ple. The lack of .lutopsies de l'lyed progress.
340 B.C. Aristotle syste molt k.llly pursued compar,l-
live an.t to my and in his 19 books set a high wolter mark
of natura l knowledge thilt l.lsted until the RenaisS.lnce,
but he did litt le to change ideas on the bra in.
300 8.(;. Herophilus and the greilt school of
Alexandria in Egypt dissected many cadavers Jnd really
founded anatomy. He distinguis hed sensory and mo tor
nerves and sho wed tha t they connect from spinal cord
to periphery.
250 8.C . Erisistrat us postulated a mechanis m of the
brain function: blood and two kinds of air arc carried in
the veins, arteries, and nef\'es; air is changed to vita l
spi rits in the hear t and these to animal spirits in the
brain ventricles, whence they go via the nerves to
distend and sho rte n the muscles. -.
200 6 .C. Galen culminated the classic period, writing
more than 400 works that were dennitive for a
mille nium. By now much of the naked-eye anatomy of
the nervous system had been d iscovered, including
most of the cranial nerves. Among the few advances in
the unde rstanding of function were t he descriptions of
symptoms fo llowing section OInd hemisection of the
spi na l cord in lower mammol ls.
400- 800 The D.uk Ages lasted more than 12
generat ions. Greek knowledge WolS forgotten in Europe.
Men did not ask to unde rstolnd themse lves or nature but
to be told the supra natura l or religious meanings of
things.
600-1200 Islam s pread from Asia minor to Spain,
carrying Gree k knowl edge unknown in the Christian
wo rld. Jewish tfolders introduced Arabic transla tions to
Europe. Long-lost Latin versions of Greek writings were
rediscovered in monastery storeroo ms. Men d id nol ye t
ask abou t nature but Jbout thei r heritage.
]200- 1300 This interest in book learning and the
new wave of ideas induced schol.lSticis m, which in turn

78

110x 7.4 (n",'hilltd)
brought on huma nis m as a redetian. Univers ities sprang
up widely du ring the last period of thc crusades.
HOO- 15OO The inve ntio n of movable ty pe stimu-
la ted p rinting. Voyages o f ex plo ra tion expressed the
new .ttti tud c IOWo1rd d iscovery.
1478 Mo ndino represents the height of dassicill,
dulhoritar i.J[l (Ga lenic) anatomy. His manual, illustra ted
crudely and ascribing func tions such as f,l1l tasy to the
anterior part of the la teral ven tricle, was little influenced
by actual d issection. Nevertheless. it WdS used for 200
Y C<lTS.
1500 l eonardo da Vinci manifested the new curios-
ity by making his own dissections and drawing more
accurately th,1n anyone before; but, fai li ng to publish,
he had litt le influence on the progress of .-millomy.
1543 Vesalius broke wilh the tradition of
Aris totelian J nd c..lenic .Julhority, inaugu rating the
mode rn id e.J of the a utho rity of o rigina l observations.
His landmark work, Dr H uma'ii Corporis Fllb rica, con-
ta ins many plates of brain dissectio ns showing nume r
ous fe.Jture s fo r the fiTst ti me.
1608 Ha rvey WJS the first to reason that the blood
circul.Jtes. He multiplied the c.Jpolcity of the heMt (with
its o ne-w.JY va lves) by the heart rate, both long-known
q uantities. T he met hod of inductive reasoning, lost
since the G reeks, was re-established .
1662 Descartes, the lCiider of 17th-century physi-
o logical thoug ht, crude ly conceived the idea of reflex
action powered by a Galenic mechanism. He broke new
g round also in the mind-body pro ble m, placing the seat
of the soul in the pineal.
1664 T ho mols W illis published o ne of the first
separate wo rks o n the brain, the most complete and
accura te so far, introduci ng several of our current terms.
He suggested tha t the cerebrum presides over voluntary
motions and the cerebellum over involuntary move-
me nts; he manipu lated the cerebellum in a living
m.Jmm.J I itnd no ted tha t the hE'oIrt s to pped.
1691 Ro be rt BoylE' pointed to the existence of a
motor cortex. A kn ight suffering a depressed skull
287
fr,Kture sho we d a s ustained para lys is of a rm and leg;
the pOlrOl lysis disappe.ued promptly after surgical re-
moval of iI spicule of bone.
1.730 StE'p hE'n Hales openE'd the d oor to reflex
physiology by noling that the legs of a decapitated frog
would withdraw upon pinching but that s uch "re-
actions" d isappeared whe n the spinal cord was de-
stroyed. T he terms "stimu lus," "response," "reflex,"
"afferent," a nd "efferent" came into use by the 1770's.
Rohert Whytt (175 1) played an important role in the
drama.
]740 Swede nborg conside red the basal gitl\glia the
seat of primary sensibility of bod y and soul and the
route of "all determinations of the wilL" He d istin-
guished u pper and lower motor cen ters a nd co rrectly
subdivided the mo to r cortex.
1791 Galvani starled electrophysiology by il\-
adve rten lly stimula ting the muscles of d issected frog
legs when they completed a circuit with two d issimilar
metals. Volta used the d iscovery of a source of electric
potentia l to develop the battery and voltaic pile.
Ga lvani, mistake nl y believing tha t the poten tial came
from t he tissue, wen t o n to discover bioelectricity by
observing tha t a ne rve is excited when it completes a
circuit betwee n an injured and an un in jured tissue. The
use of a nerve-muscle preparation as a biological
de tecto r, amplifier, and ind icator was an ingenious
physiological tech ni(IUe that permitted the d iscovery of
mi llivolt level bioelect ricity many years before the
g<1lvanometer was invented. Elect ricity came just in ti me
to fill the gap as improved anatom y excluded the
hyd raulic mo del of ne rves and the new physics and
che mistry ra ised do ubts about "a nima l s pirits." T he
s t<1ge was set fo r a r<1tiollal, mechanistic physiology.
1809 Rolando removed the cerebellum in fish, rep-
tiles, and mamma ls and saw disturbances in volun tary
move ments witho ut influencing sensa ti on .
1822 Francois Magendie made fi rm an eMlie r claim
by Charles Be ll that do rsa l roots are sensory a nd a lso
s howed unequivoca lly that ventral roots are moto r. like
a ll s uch experiments in these pre-anesthesia days, his
work was based on vivisection.
(ColllillJlfd QII II fXl l!IIg~.)
79
288
UOI( 7.4 (rOlllj'HIt<I)
1823 Pierre Flourc ns showed that vision depends on
th e cortex; ablatio n on one s ide in pigeons, T,lbbits, and
dogs was found 10 cause contrali1ter,l l blindness.
1826 Johannes Mli ller, wide- rangingGcrmilll natu ra l
philosopher, sensory phys iologist, and ("omp,-u .ltive
anatomist, enunci,lied the " law of specific nerve
energies," which s tates that each sensory nerve gives
rise to its own charilderistic sensatio n, however it is
s timul<l ted . For example, electrical, mechanicdl, or
chemical stimulation of the optic nerve (.luses a sensa-
tion of light.
1833 Marshall Hall recogniled segmcntill, intcr-
segment<ll, and s uprasegmenlal reflexes. '1"he spinal
cord is a chain of segments whose function.ll units iH C
sep.H.lte reflex arcs which interact with o ne another and
with the highe r centres of the nervous system to secure
coo rdina ted movement" He recognized the tempor.uy
de pressio n of reflexes below a spinal transection and
c,tl le d it s pinal shock.
1848 Du Bois Reymond showed that activity in a
nerve is invMiably accomp.mied by an e lectrica l change
(" negative v<lria tion"). He described the properties of
neurOlnuscul<lr transmissiOn and opposed the prevail-
ing doctrine of vitalism.
1850 Helmho ltz measured the velocity of conduc-
tion in nerve tissue and began what would become a
continuing effort to improve the instruments of electro-
physiology.
1651 CI.lude BernMd de veloped the la ndm.uk con-
cept tha t the body maintains a consta nt inte rnal e n-
vironment for the ce lls by means of the extracellular
fluids . Among other things related to sympa thetic func-
tion, he described the vasomotor ne rves, which pldy a
ke y part ill this regulation, late r ca lled ho meostas is. An
en thus ias tic e xperimen talis t, he wrote influentially o n
the experilllentdi method, advocating rigo r, con tro ls,
and fo rm uldtion of testable pred ictions.
1861- 1898 H ughlings Jackson, British neu rologist,
developed concepts on the underlying principles of
brain function from cl inica l observa ti ons. One was the
concept of "release" to olccount fo r va rio us signs of
in jury to higher p.uts of the brain, s llch itS s pasticity,
thai a rc more pos itive th,ln negative; the resulting
OVC Iact ivi ty of the s urviving lower centers bespeaks .1
normal restrolint imposed by the higher. A seco nd
co ncept lVds that although e \'olution has been a process
of increased differentioltion and he terogcneity, wi th
integration keeping pace, d isedse reverscs this, such
that higher parts go first and the lower take control. A
th ird concept, growing out of intense stud y of patien ts
with speech d iso rders, those wi th sensory, motor, or
psychic epilepsy ilnd hemiplegi.Js, was that the cortex
has Illany locollized func tions.
1863 Sechenov stud ied " reflexes of the brol in,"
me.lning cerebrol l activity thai .lTises from sensory
st imulat io n and med iates psychic experience and causes
volun tMy action, s ubject to modulation by other bra in
cente rs, incl ud ing inhi bition by the midb rain. This he
obtained by placing salt on the optic lobes. He recog-
nized temporal s ummation of subth reshold stimuli; also
muscle sense, later ca lled proprioception . He empha-
s ized the physicochemiCdI anollysis of metolbolism and
excitoltion. Trained with CI,lude Bernol rd in Paris dnd Ou
Bo is Rcymond in Be rl in, he is reg.lTded as the father of
Russian physiology.
1865 Pfluge r systcmoltically investi g~t ed inhibition,
mol inly vid a utonomic nerves. Searching fo r inhibitory
nerves to skele ldl muscle, Pa ylov (1885) found the m [n
the fres h-water mussel, AIIQdQrlili. a biva lve mollusc;
Biedermann (1887) found them in the crayfish. They ~re
still unknown in vertebrJ.tes.
1870 Gudde n's find ing th itt s pecific thalamic nuclei
degencrate when certilin areas of the cerebral corte x arc
dcstroyed was a milestone in experimental anatomy, as
well as col iling attention to retrograde degenera tion and
o pening the mode rn period of stud y of the thalamus.
t874 Bartholow, in the U.s.A., stimu),lted and
mapped the motor cortex in man. He fo und, incident-
a lly, ihat the brain itself is inse nsitive to manipulolt ing
and cutting.
1875 Richard Caton, in Engl.lIld, observed clec tr ic~l
waves from t he exposed brains of rolbb its dnd monkeys;
80

80)( 7.4 (Wllf;,mrd)
his finding was overlooked, but the waves were re-
discovered later in Russi,1 (1877), Poland (1890), .lnd
Austria (1890). Caton IV.lS looking for action potcnti,ll s
ill the ur,lin, inspired by Du Bois Reymond's in nerve,
hoping they would provide a Inelhod for loc.tl izing
sensory arCilS. [n this he succeeded, d iscovering evoked
potentiolls and, inddenlally, DC s hifts with activity, as
well as the ongoing EEG
1898 Lmgley introduced the term "autonomic" ilnd
7 yc,lrS lale r, "sympathetic" and " p.Holsympathelic."
1902 Pavlov, investigating the phys iology of d iges-
tio n, saw clearly the rOild he would follow for 34 years,
analyzing the psychologica] propert ies of conditioned
reflexes. His influence on neurophysiology was "" Imost
nil" (Fu lton, 1949) until recen t yeMs.
1903 Brodm,lnn, Vogt, and Ca mpbell e<1ch made
their first communiutions on the archi tectonics of the
cerebr<1l cortex, m<1pping the dist ribu tion of d ifferent
types of cortiC<11 struclure. It lYas some ye.us, however,
before the 6-layered str<1tificoiltio n lYas fully exploited;
then its embryologic<11 origin lYoiIS e mphasized. By 1929
Ariens K<1ppers and others had ildded a n evolutionary
origin from a primitive 3.I<1yered mantle. Enthusiasm
for the 6Ja yered structure long delayed a concern for
the neurona l organ iz.ltion and connecti o ns.
1906 Sherrington, in England, published his land
m.uk treollise, Thf IlIlfgrali vt ATlioll of illt Navolls 51/SItIII,
in which he systematic.llly .malyzed holY the nervous
system works, by close eXolm inoltion of si mple a nd
compound reflexes. Primarily he used mechanical re
cording of cont raction of individual muscles. Most of
thc ideas on pp. 271-27Z M e his.
1909 Karplus .Ind Kreidl beg.ln the first experi.
ment.ll study of the hypothalamus, their resu lts .lppe.u
ing in iI long series of papers. Many othe rs joined in,
including Ihe ce lebr.lIed surgeon Harvey Cush ing, who
in 1912 discovered that removal of the pituit.uy or
merel y tr<1nsecting its st.llk c.luses an adi posogenit<11
dystrophy. Nevertheless. the modern period of re
seMch, in IYhich the hypotha lamus is related to
"utonomic, emotio nal, a nd instinctive functions and the
,
289
control of the pituitary, did not really begin until the
1930's.
19"1"1 Henry Head and Gordon Holmes, using
psychologic" l concepts ilnd testing, studied sensory
deficits after clinical lesions. They were more concerned
lYi th thc [ldture rather than the locus of cortical sensory
processes in man. They s holYed th.l t the cortex is
especi.,11y involved in discri minat ive "nd higher as pects
of perception. Head is remembered a lso for severing a
nerve in his own ,J.rm to stud y the loss "nd return of
sensation with rcgener,J. tion.
1917 Keith Lucas firmly established the allor none
law and quantitative relations in exci tation, such as the
minimal s lope of a slowly riSing current necessary for
excitation.
1924 Kato, in J"P<1I1, sett led a con troversy by show
ing nondecremental conduction in nerve.
1924 G,J.sser and Erlanger, in the U.s.A., used the
cathode fay osci lloscope to describe the components of
the compound action potential of the whole nerve.
1924 Rudo lph Magnus published his la ndmark
monogrdph on post ure. Starting his investigations 16
years before with Sherrington, he edrly realized the
importa nce of twisting the head on the neck, IYhich
reduces the tonus of posturalmusde and thei r reAexes
o n the side of the forward car in the human or the lower
ear in the dog.
1926 Adridn and Zo\terman recorded from si ngle
scnsory nerve fibe rs and found Iha t, as in motor fibers,
the repet ition rate of impulses is graded. Stronge r
stimuli result in a g reater frequ ency of s pikes.
1929 Denny. Brown recorded from a single motor
neuron ,lctivdled by a n0T1n~1 reflex stimu lus; he used
the stre tch reflex in a decerebrate Cdt.
[929 Berger reported that brJin W.lVes could be
recorded through the skull in humans. After" general
scepticism. this was confirmed by Adri.ln in England
and t,lken up by Davis and othe rs in the U.s.A.
81
290

Chapler 1
Inleg ra tion al the Inter mediate Levels
SUGGESTED READI NGS

Autrum, H., R. Jung, W. R. Loewenstein, D. M. MacKay, il,o.d H. L Teuber. 1911. Hllndboak of

5'"50/,y Ph ysiology. Sp ringer-Ve rlag. New York. fProjectea to make up e ight volumes, some
in paris, th is will be a comp rehE'ns ive treatise.]
Bach-y-Rita, P., C. C. Collins, and j. E. H yde. 197 1. Tht (,mlral of Eyt MClllfllltIJ/S. Academic
Press, New York. [A symposium on a system es pecial ly favo rable for revea ling principles.
Other mo re recent symposia and reviews can be fou nd on this system. ]
Gr,mit, R. 1970. TIl t BlI5is of M olor Conlro l. Academi c Press, New York. [A systematic ana lys is
from sensory receptors to higher cerebr'!l level s controlling motor neurons.]
G rod ins. F. S. 1963. CD lli rol Tlltory ,wd BitJlogiltl/ SyS/lm $. Colum bi(l Univ. Press, New Yo rk. [A
compact textbook, useful both as an introduction to this (lp prO<lc h a nd to some ins ightfu l
exa mples, which .ue partly neur.ll.!
Horridge, C. A. 1968. 'll/mlll.rou s. W. H. Freeman and Company, S.lIl Francisco. [A collection
of diverse exa mples of neuf;11 mec ha nisms that involve interneurons a t several levels of
analysis. ]
Reiss, R. F. 't964. NlUrtll TlltDI'y ami M odtling (Proc. t962 Ojai Synlpos ium ). Stanford Univ.
Press, Sta nfo rd. [A sym posium combining ex peri me ntal a nd theore tical trea tmen ts o f
examples of sensory and motor beh.lvior. [ ""'-
Schmi tt, F. 0 ., and F. G. Worden. 1974. Tlt t N furDsrinl(fs: TI,jrd Stu dy PrDgrll lll . (See Suggested
Read ings in C hap ter 8 fo r comments on this and two preced ing volumes in the series.]
Stein, It 8., K. G. Pe.lTSOn, R. S. Smith, ,m d J. B. Redford. 1973. CO Il/rol of pos/u,-e alld LOCDrIl O/iDli .
Plenum Press, New York. ]A symposium includ ing invertebrate and vertebrate sensory,
ccntr,ll, reflex, and rhy thmic control.!
St.uk, L 1968. NfrfrDIDgirlll CDII/ ru/ Sys/rllls: Slrulits ill BiolHgillrniag. Plenum Press, New York.
[Exemplary ,111,l lyses from an engineering ilpproach o f some subsystems controll ing
seei ng and manipulating.]
Wiener, N., and J. P. 5chad~. 1965. Cybmltiirs of Olt Nrrvolls 51/51t lll (Progress in the Br.l in
Resea rch, vol. 17). Elsevier Pu blishi ng Company, Amsterdam. IA lt ho ugh ma n y cha pters
(Ire d ated or mainly of historical interest, the collection exe mplifies how d iverse are the
problems and stages o f p rogress under this rubric. Many modern symposia and articles
are read ily fo u nd, in addition to the journ.l ls whose ti lles Cil rry Wie ner's term.)
Wie rs nlol, C. A. G. 1967. IIII'tr/tbral, Nuvcllts SysltJIl5; Tluir SigllifiCIIll ft fDr MllmmR/iali Nfur".
Vll 1/sivlQgy. Un iv. Chicago Press, ChicJgo. [A sym posium deal ing a t many levels wi th
results relevan t to Ill<lmma li,11l neuro physiology.]

82
Neuron, Vol. 16, 701716, April, 1996, Copyright 1996 by Cell Press
Dendritic Integration
in Mammalian Neurons,
a Century after Cajal
Rafael Yuste* and David W. Tank
*Department of Biological Sciences
Columbia University
New York, New York 10027
Biological Computation Research Department
AT&T Bell Laboratories
Murray Hill, New Jersey 07974
Introduction
The role of dendrites in neuronal signal processing has
been a topic of research for more than a century. Early
work started with Cajals original postulate that den-
drites constitute the input side of the neuron and culmi-
nated in the idea that dendrites are passive cables. Later
studies, however, showed that dendrites could produce
action potentials.
Here, we review within a historical perspective recent
experimental work on the function of mammalian den-
drites that has used novel approaches such as imaging
of calcium concentration and dendritic patch recording.
We do not intend an exhaustive review of the field of
dendritic physiology nor cover appropriately progress
in invertebrate preparations (Borst and Egelhaaf, 1994),
but instead focus on a few experimental studies from
pyramidal and Purkinje neurons.
These new findings have provided direct evidence for
active conductances in dendrites and chemical com-
partmentalization in spines. Although the logic of den-
dritic processing is still unclear, new results show that
mammalian dendrites have a rich repertoire of electrical
and chemical dynamics, suggesting that individual neu-
rons are capable of sophisticated information pro-
cessing. Combining these novel technologies with stud-
ies in behaving animals will greatly help in understanding
dendritic integration.
In any case, contemplating the form of the cells
was one of my most beloved pleasures. Because
even from an aesthetic point of view the nervous
tissue has fascinating beauty. Are there in our
parks any more elegant and lush trees than the
Purkinje neuron in the cerebellum or the so-called
psychical cell, that is, the famous cortical pyrami-
dal neuron? (Ramon y Cajal, 1923)
Brief History of Dendritic Research
In the Beginning, There Was Cajal
Deiters is credited as the first to realize that neurons
had two morphologically different types of neuronal pro-
cesses (Deiters, 1865): a long, thin, and unbranched
process, named axon after the Greek word for axle (Kol-
liker, 1889), and the protoplasmic processes, multiple
and branched, and named dendrites after the Greek
word for tree, dendron (His, 1893). These anatomical
descriptions led to the question of the propagation of
the nervous impulse. As Ramon y Cajal (1923) put it,
Does the nervous impulse propagate in all directions,
Review
like sound or light, or does it constantly flow in one
direction, like water running in a water-mill?
In 1889, Cajal noticed that in both retina and olfactory
bulb dendrites are always oriented toward the external
world, while axons are always oriented toward more
central neural centers. Based on this observation, he
proposed that dendrites conduct current toward the cell
body (i.e., centripetally), while axons conduct current
away from the cell body and toward other nervous cen-
ters (Ramon y Cajal, 1889). Nevertheless, to his dismay,
he realized that centripetal current conduction in den-
drites could not constitute a general rule of operation
in the nervous system because of two contradictory
pieces of evidence from his own work: the first one in
the optic tectum, where neuronal fields were composed
exclusively of dendrites, and the second in sensory gan-
glia, where axonal-looking processes were oriented to-
ward the periphery. These contradictions caused him
to abandon his hypothesis (Ramon y Cajal, 1923).
However, Cajal reexamined this issue 2 years later in
response to a severe criticism of his centripetal theory
(Van Gehuchten, 1891). Fortunately for him, his brother
Pedro had just finished, in his doctoral thesis, a more
complete study of the optic tectum, finding that the
regions previously considered by Cajal to be composed
only of dendrites also had rich axonal terminations (Ra-
mon, 1890). Cajal also then realized that the processes
in sensory ganglion neurons that appeared to be axons
were, in fact, dendrites (Ramon y Cajal, 1923). Therefore,
the two major obstacles to the theory were removed
and he proposed the law of functional polarization:
The transmission of the nervous movement al-
ways occurs from the protoplasmic branches
(dendrites) and cell body towards the axon or
functional expansion. Every neuron, then, has a
receptive apparatus, which is the soma and proto-
plasmic prolongations (dendrites), an emitting ap-
paratus, the axon, and a distributing apparatus,
the terminal axonal arbors. (Ramon y Cajal, 1891a)
This law, however, did not explain some instances in
which the axon originates directly from the dendrite, so
he coined a new, final, law in 1897, the so-called axip-
etal polarization law:
The soma and dendrites have axipetal conduc-
tion, i.e., they transmit the nervous waves towards
the axon. Inversely, the axon has somatofugal or
dendrofugal conduction, propagating the im-
pulses received by the soma or dendrites towards
the terminal arborizations. (Ramon y Cajal, 1897)
Thus, dendrites constituted the input side of the neuron,
while the axon constituted its output side. It is a tribute
to Cajals foresight that his law basically still stands, over
a hundred years from its first proposal. Nevertheless, it
is now apparent that there are several important excep-
tions to his rule, in cases where information is conveyed
from the soma or axon back to the dendritic tree and
where dendrites also serve as the output side of the
neuron (see below).
83
Neuron
702
Passive Dendrites: Motoneurons
and Cable Equations
After this initial progress, dendritic research remained
in the background of neurobiological studies until the
middle of this century, when a golden age for ideas
of passive dendrites flourished. This renaissance was
fueled by two technical advances: intracellular recording
(Hyde, 1921; Ling and Gerard, 1949) and cable theory
(Hermann, 1905; Hodgkin and Rushton, 1946; Rall,
1995).
Intracellular recordings with glass microelectrodes
enabled investigators for the first time to look into the
internal electrophysiology of the neuron. This approach
was systematically developed in studies of spinal cord
motoneurons (Brock et al., 1952). Two key findings were
the existence of excitatory and inhibitory postsynaptic
potentials (EPSPs and IPSPs) (Coombs et al., 1955a,
1955b) and the generation of the action potential at
the axon initial segment (IS) (Fuortes et al., 1957). An
integrative model was proposed: dendrites passively
summated the net current resulting from EPSPs and
IPSPs and electrotonically conveyed it to the IS, where
it would trigger an action potential that would then travel
down the axon (Fatt, 1957; Coombs et al., 1957). Al-
though investigators did not negate that dendrites also
had active properties and in fact proposed that the IS
spike secondarily triggered a somaticdendritic (SD)
spike (Eccles, 1964), the passive model dominated den-
dritic physiology and still today is taught in textbooks.
The behavior of passive dendritic trees was set on a
firm theoretical foundation by the application of cable
theory to dendrites by Rall (1959) (for a compilation of
Ralls work, see Rall, 1995). He derived and, in some
cases, analytically solved the equations describing the
flow of electric currents in model dendritic trees receiv-
ing different spatiotemporal patterns of synaptic input.
Rall introduced the concept of electrotonic structure of
dendritic trees and defined the basic cable parameters:
Rm, membrane resistance; Ri, internal resistance; Cm,
membrane capacitance; l, cable length; r, dendritic-to-
soma conductance ratio. His results characterized the
attenuation of inputs produced by dendrites, illustrating
the asymmetric electrical behavior of the dendritic tree
produced by the lower somatic impedance. Ralls ex-
plorations produced a quantitative picture of dendritic
physiology that would influence experimental and theo-
retical dendritic research to this date.
Decremental Conduction:
Dendritic Boosting
In this golden age of passive dendrites, Lorente de No
and Coundouris (1959) proposed that dendrites could
actively amplify synaptic inputs. In their influential paper,
they reexamined the theory of decremental conduction
in nerve fibers, which stated that nerve fibers conducted
impulses of various magnitudes and that impulses de-
crease (or increase) gradually as they spread throughout
the axons. They suggested that dendrites also showed
decremental conduction and that their nerve impulses
differed from those in axons only in their slower speed.
Motivated by previous results from chromatolyzed mo-
toneurons showing spike-like partial responses (Eccles
et al., 1958) and spread of axonal impulses increased
by excitatory synaptic impulses (Renshaw, 1942), they
suggested two mechanisms for enhancing EPSPs:
boosting of synaptic impulses produced by previous
synaptic excitation and summation of synaptic currents
taking place at dendritic branch points. Thus, the den-
dritic branching pattern and precise location of the in-
puts could make neurons particularly tuned to specific
spatiotemporal patterns of excitation. In spite of this
boosting, they argued against dendritic action potentials
by pointing out that all-or-none events impoverish den-
dritic physiology by annihilating subtle temporal sum-
mations by subthreshold events.
Active Dendrites: Prepotentials
and Dendritic Spikes
The time of active dendrites had arrived. Influential work
came from experiments in chromatolyzed motoneurons
(Eccles et al., 1958), Purkinje neurons (Llinas and Nichol-
son, 1971; Llinas and Hess, 1976), and pyramidal neu-
rons (Spencer and Kandel, 1961; Andersen et al., 1966;
Wong et al., 1979). These results introduced the idea
of dendritic action potentials into the mainstream of
dendritic research.
In an intracellular study of hippocampal neurons,
Spencer and Kandel (1961) described fast prepotentials
(FPPs), which were small potential steps immediately
preceding full-blown spikes. They concluded that FPPs
were spikes because of their all-or-none character and
because their repolarization was faster than the mem-
brane time constant. In their discussion, Spencer and
Kandel proposed that FPPs were distant dendritic
spikes produced in a trigger zone, presumably associ-
ated with a bifurcation of the apical dendrite.
Although influential conceptually, prepotentials were
not unambiguous experimental evidence for active den-
drites (see below). Nevertheless, in a series of seminal
papers, Llinas and collaborators ushered in a new era
in dendritic physiology by recording intracellularly from
dendrites of Purkinje neurons and demonstrating di-
rectly the existence of dendritic spikes (Llinas and Nich-
olson, 1971; Llinas and Hess, 1976; Llinas and Sugimori,
1980). Their main findings can be summarized in a pic-
ture of the Purkinje neuron in which dendrites produce
plateau potentials and calcium spikes while the somatic
region produces sodium spikes. Purkinje neurons would
then have two distinct functional areas, the soma and
the dendritic tree, with different active electrophysiologi-
cal properties that completely dominate the passive ca-
ble parameters. In addition, to explain the different am-
plitudes of dendritic spikes, they suggested that multiple
spike generation zones (hot spots) existed in the den-
dritic tree, thus conferring functional independence to
individual dendritic branches. Finally, this work intro-
duced to the field of dendritic research the new technol-
ogy of brain slices, making possible high quality intra-
cellular recordings and thorough pharmacological
manipulations (Yamamoto, 1972).
In later work, intradendritic recordings were also
achieved in hippocampal pyramidal neurons (Wong et
al., 1979). In response to current injection or spontane-
ous synaptic stimulation, a rich dendritic physiology of
sodium and calcium spikes was found. These results
were confirmed in surgically isolated dendrites (Ber-
nardo et al., 1982).
84
Review
703

Novel Imaging and Electrophysiological

Techniques: A New Era
for Dendritic Research
In the last decade, several new techniques have revolu-
tionized dendritic research by offering more direct ways
of probing dendritic function and answering questions
about the spatial distribution of chemical or electrical
activity. The first breakthrough was the synthesis of sen-
sitive optical indicators of calcium concentration (re-
viewed by Tsien, 1989). This technology enabled mea-
surements of the intracellular concentration of free
calcium ion ([Ca21]i) in dendrites and spines. These indi-
cators would not have been so successful without new
imaging technologies. Although some initial efforts were
carried out with photodiode arrays (Ross et al., 1986;
Ross and Werman, 1987), it was the advent of digital
imaging using cooled charge-coupled device cameras
(cCCD) that finally allowed systematic quantitative mea-
surements of [Ca2 1]i dynamics in neurons (Connor, 1986;
Tank et al., 1988; Jaffe et al., 1992).
Another crucial development has been the combina-
tion in brain slices of patch-clamp recording (Sakmann
and Neher, 1983; Blanton et al., 1989; Edwards et al.,
1989) with infrared video microscopy (Dodt and
Zieglgansberger, 1990; Figures 1 and 2). Even though
dendritic recordings (Llinas and Hess, 1976), and even
simultaneous dendritic and somatic recordings (Sugi-
mori and Llinas, 1984, Soc. Neurosci., abstract; Wong
and Steward, 1992), can be carried out with microelec-
trodes, infrared video microscopy makes it possible to
patch and record systematically from identified cells or
parts of a cell in slices (Stuart et al., 1993).
Finally, two new forms of microscopy have enabled
high resolution imaging of dendrites in brain slices. The
first one, confocal microscopy (Minski, 1961; Wilson and
Sheppard, 1984; Fine et al., 1988), uses pinholes to ex-
clude out of focus light, while the second one, two-
photon microscopy (see Figure 4) (Denk et al., 1990),
takes advantage of the restricted spatial profile of two-
photon absorption (Goeppert-Mayer, 1931) to excite
only fluorophores located at the focal point (Denk et al.,
1994).
We will now review the contribution of these new tech-
niques to central questions of dendritic function. Again,
we will focus specifically on experimental research on
Purkinje and pyramidal neurons.
Purkinje Cells Have Active Dendrites
Recent results have demonstrated the existence of ac-
tive conductances in dendrites from Purkinje and pyra-
midal neurons, confirming proposals from pharmaco-
logical (Wu et al., 1986) and immunocytochemical
Figure 1. Infrared Video Microscopy Imaging of a Pyramidal Neuron
localization studies (Westenbroek et al., 1990, 1992; Hill-
in a Neocortical Slice
man et al., 1991; Angelides, 1994, Soc. Neurosci., ab-
Reprinted with permission from Dodt and Zieglgansberger (1994).
stract).
The photomicrograph shows a layer 5 pyramidal neuron in a 400
In Purkinje neurons, intracellular recordings with mm thick slice of rat visual cortex. The apical dendrite is clearly
sharp microelectrodes showed dendritic calcium spikes visible for approximately 250 mm. A patch pipette is used to record
and somatic sodium spikes (Llinas and Sugimori, 1980). from the soma in whole-cell configuration. Picture taken with infra-
This strongly suggests that calcium channels are pres- red gradient contrast imaging.
ent in dendrites, while sodium channels may be re-
stricted to the soma. Calcium imaging results have con- these studies, calcium accumulations appeared
firmed that voltage-sensitive calcium channels (VSCCs) throughout the dendrites whenever calcium spikes or
are indeed present in the dendrite, as well as in the plateau potentials took place. The fact that dendritic
soma (Ross and Werman, 1987; Tank et al., 1988). In calcium accumulations occur immediately after calcium

85
Neuron
704
Figure 2. Photomicrograph of a Purkinje Neuron in Cerebellar Slice
Filled during Simultaneous Somatic and Dendritic Recording with
Two Patch Pipettes in Whole-Cell Configuration
Reprinted with permission from Stuart and Hausser, (1994). Each
pipette has a different fluorescent dye, Cascade blue in the somatic
one and Lucifer yellow in the dendritic one, which mix in the dendrite
of the Purkinje cell.
spikes are triggered, together with the slow diffusion
of calcium intracellularly (Hodgkin and Keynes, 1957;
Allbritton et al., 1992), indicates that the calcium chan-
nels responsible for the spikes must be dendritic (Ross
and Werman, 1987; Tank et al., 1988).
Are there sodium channels in dendrites of Purkinje
cells? In the original recordings of Llinas and Sugimori
(1980), small sodium spikelets were present in dendrites,
and outside-out patches from dendrites have directly
revealed dendritic sodium channels, although at much
smaller densities than the soma (Stuart and Hausser,
1994). From the persistence of sodium spikes after so-
matic voltage clamp, it has been further argued that
dendritic sodium channels can generate full-blown ac-
tion potentials (Regehr et al., 1992). Nevertheless, com-
bined intra- and extracellular recordings from the same
neuron suggest that sodium and calcium spikes are
generated by different compartments of the cell (Llinas
and Sugimori, 1992). Finally, when the intracellular so-
dium concentration was imaged during synaptic stimu-
lation, a small sodium influx was detected in dendrites,
but was attributed to residual sodium influx through
glutamate receptors (Callaway et al., submitted). Taken
together, these results suggest that there are indeed
dendritic sodium channels in Purkinje cells, but that
their usual mode of operation may not be regenerative,
although a persistent type of dendritic sodium channel
may still significantly contribute to synaptic boosting
(Llinas and Sugimori, 1980).
Pyramidal Neurons Have Active Dendrites
In hippocampus and neocortex, dendrites from pyrami-
dal neurons have been shown to have both sodium and
calcium channels by electrophysiological and imaging
experiments.
Since the Spencer and Kandel experiments, somatic
recordings of cortical pyramidal neurons have con-
firmed the existence of fast prepotentials in vivo after
synaptic or sensory stimulation (Purpura and Shofer,
1964; Deschenes, 1981; Hirsch et al., 1995). These pre-
potentials are resistant to the sodium channel blocker
QX-314, suggesting that they are calcium action poten-
tials (Nunez et al., 1993; Hirsch et al., 1995). Although
prepotentials may reflect neuronal coupling (MacVicar
and Dudek, 1981; Nunez et al., 1990), intradendritic re-
cordings from hippocampal and neocortical neurons
have indeed demonstrated full-blown sodium and cal-
cium spikes both in vitro (Schwartzkroin and Slawsky,
1977; Wong et al., 1979; Amitai et al., 1993; Kim and
Connors, 1993; Stuart and Sakmann, 1994; Spruston et
al., 1995) and in vivo (Houchin, 1973; Pockberger, 1991).
Confirmation of the existence of dendritic calcium
channels by cCCD imaging has come in CA1 neurons
(Regehr et al., 1989; Jaffe et al., 1992) and layer 5 cells
(Yuste et al., 1994). This conclusion rests upon the same
argument as the one developed above for cerebellar
cells, i.e., the absence of delay between calcium accu-
mulations and electrophysiological responses. Using
confocal and two-photon imaging, calcium channels
have even been specifically localized to dendritic shafts
(Markram and Sakmann, 1994) and spine heads (Yuste
and Denk, 1995). Therefore, it appears that the entire
dendritic tree of pyramidal neurons is covered with cal-
cium channels. Because a sodium spike is necessary
to open those channels antidromically (Jaffe et al., 1992;
Regehr and Tank, 1992; Spruston et al., 1995; Yuste and
Denk, 1995), sodium channels must also be distributed
throughout the dendritic tree, although it is unclear if
they are present in spines (Angelides et al., 1995, IBRO,
abstract). Finally, direct evidence comes from experi-
ments demonstrating sodium and calcium channels in
membrane patches from apical dendrites (Huguenard
et al., 1989; Magee and Johnston, 1995a, 1995b). Thus,
dendrites from the two populations of pyramidal neu-
rons examined so far have sodium and calcium chan-
nels. In contrast with Purkinje cells, in pyramidal neurons
both populations of channels appear to produce regen-
erative spikes in the dendrites.
What Are the Cable Properties
of Dendrites?
Although Purkinje and pyramidal neurons have active
dendrites, the passive cable properties of dendrites,
determined by the three basic cable parameters (Cm,
Rm, and Ri), are still necessary to understand dendritic
function, since in the subthreshold regime dendrites
could behave passively and since in the suprathreshold
regime Ri and Cm and the dendritic geometry can affect
the spatiotemporal properties of the spikes. Unfortu-
nately, measurements of Cm and Ri in dendrites are diffi-
cult and present estimates are based on values that
best fit numerical simulations of experimental data. Even
Rm, which can be determined experimentally by fitting
curves to the decays of hyperpolarizing current pulses
(Rall, 1959), is influenced by conductances that are ac-
tive at rest (Spruston et al., 1994) and may not even be
constant throughout the cell (Shelton, 1985; Rapp et al.,
1994).
Nevertheless, progress toward the estimation of the
passive properties of the neuron has been made on
86
Review
705
many fronts. First, whole-cell recordings suggest that
the neuronal time constant may be much larger (and
thus, the Rm much higher) than previously assumed from
microelectrode recordings (Spruston et al., 1994; Major
et al., 1994). This discrepancy is attributed to the shunt
introduced by the somatic leak induced by the micro-
electrode impalement. Second, another significant re-
sult has been the study of cable models of branched
dendritic trees that cannot be approximated by equiva-
lent cylinder models (Stratford et al., 1989; Major et al.,
1993; Major, 1993). This work has confirmed the impor-
tance of the shunt created by the electrode and has
suggested the impossibility of good voltage-clamp con-
trol of dendrites from the soma. Finally, two new ap-
proaches have been used to estimate the length con-
stant of the main dendrites: the measurement of voltage
attenuation between dendritic and somatic electrodes
for injected DC pulses of current (Sugimori and Llinas,
1984, Soc. Neurosci., abstract; Stuart and Hausser,
1994) and the use of the voltage ratio between somatic
and dendritic calcium plateaus, as recorded at the soma
(Reuveni et al., 1993; Yuste et al., 1994).
In spite of this, the basic cable properties of dendrites
are unknown, and it is even possible that they may not
be constant throughout the dendritic tree. Still, the past
few years have seen an explosion of modeling studies
of neurons, triggered by the availability of computer
programs that implement cable equations in multicom-
partmental models (reviewed by De Schutter, 1992). Re-
sults from different modeling approaches based on
careful fitting of experimental data converge on similar
values for Cm (0.72 mF/cm), Rm (50200 kVcm2 ), and Ri
(200400 Vcm), suggesting that these values may be
close to the real ones (e.g., Stratford et al., 1989; Sprus-
ton et al., 1994; Rapp et al., 1994; Major et al., 1994).
Where Is the Spike Initiation Zone?
In the classical work from the fifties, it was concluded
that motoneurons had two thresholds for spike initiation,
a lower one (z10 mV) in the axonal initial segment (IS)
and a higher one (z30 mV) in the somadendrites
(Coombs et al., 1957). This conclusion was reexamined
in CA1 pyramidal neurons in vitro with a combination of
extracellular, intrasomatic, and intradendritic microelec-
trode recordings. In response to synaptic stimulation at
threshold, the site of spike initiation was located at or
near the soma (Richardson et al., 1987), but shifted to the
proximal dendrites when synaptic strength is increased
(Turner et al., 1991). Therefore, the spike initiation zone
for CA1 neurons is dependent on the stimulation par-
adigm.
Experiments with patch electrodes have recently ad-
dressed this question in neocortical pyramidal neurons.
Under somatic voltage clamp, Regehr et al. (1993) found
unclamped sodium spikes triggered by synaptic stimu-
lation and suggested that the spike initiation zone was
dendritic (Regehr et al., 1993). In contrast, Stuart and
Sakmann (1994) used dual recordings from soma and
proximal apical dendrites to show that, after synaptic
stimulation, dendritic sodium spikes were always back-
propagated from the soma (Figure 3). The authors then
patched the axon and showed that, in response to so-
matic current injections, the axon always fired ahead of
the soma. They proposed that in neocortical pyramidal
neurons the axon is the spike initiation zone and sug-
gested that a major function of dendritic sodium action
potentials is to convey information about the state of
the soma to the dendrites. This centrifugal, antidromic,
or backpropagating role for the action potential has also
been proposed in olfactory bulb neurons (Rall and Shep-
herd, 1968), Purkinje neurons (Llinas and Nicholson,
1971), and CA1 pyramidal neurons (Richardson et al.,
1987) and constitutes a blatant contradiction to Cajals
axofugal law.
Similar experiments with dual somatic and dendritic
patch recordings have been carried out in other neuron
types. While results from Purkinje cells indicate a so-
matic initiation zone for sodium action potentials (Stuart
and Hausser, 1994), in CA1 pyramidal cells the spike
initiation was somatic but shifted to proximal dendrites
with increased synaptic stimulation (Spruston et al.,
1995). Finally, many neurons in substantia nigra had
dendritic initiation; interestingly, this was shown to be
due to the axon emerging from the dendrite (Hausser
et al., 1995).
Taken together, these studies lend support to the idea
that in pyramidal and Purkinje cells the spike initiation
zone is primarily axonal/somatic. However, several im-
portant caveats preclude a definite answer. First, like in
CA1 neurons, the location may depend on the intensity
of the stimulation. Second, these recent studies do not
address the role of calcium spikes in this process, due
perhaps to the young age of the cells and the experimen-
tal conditions at which experiments were performed.
Because calcium spikes are prominent in recordings
from mature animals, both in vivo (Pockberger, 1991;
Hirsch et al., 1995) or in vitro (Amitai et al., 1993; Sak-
mann, 1995, Soc. Neurosci., abstract), their existence
may change the balance of excitation between dendrites
and soma. Third, dendrites in vivo may sustain their
own spike-generating zones, isolated from the soma,
because of the electrical shunt produced in dendrites by
their constant bombardment with excitatory or inhibitory
synaptic inputs (Bernarder et al., 1991; see below) and
because significant changes in the length constant
could be produced by intrinsic electrophysiological
states of the neuron (see below). In fact, stimulus-
evoked calcium spikes in vivo are more frequent during
epochs of strong synaptic barrages (Hirsch et al., 1995).
Inhibition Can Regulate Dendritic Function
Approximately 20% of the synaptic contacts in the brain
are inhibitory (Peters et al., 1991); moreover, electron
microscopic studies indicate that the location of inhibi-
tory inputs in dendrites can be very specific (Peters
et al., 1991; Buhl et al., 1994). The influence of those
inhibitory inputs on the physiology of dendrites was
demonstrated by the inhibition of dendritic spikes in
Purkinje cell by stimulation of inhibitory interneurons
(Llinas et al., 1968).
The role of inhibition in controlling dendritic excitation
has been recently explored in pyramidal and Purkinje
neurons with calcium imaging and dendritic recordings
(Callaway et al., 1995; Kim et al., 1995). Callaway et al.
(1995) have studied the effect of inhibition on calcium
accumulations of Purkinje neurons and find that IPSPs
reduce the dendritic depolarizations and calcium accu-
mulations produced by climbing fiber activation. This
87
Neuron
706

Figure 3. Somatic Spike Initiation Zone in Pyramidal Neurons

Reprinted with permission from Stuart and Sakmann (1994).
(a) Simultaneous recording from the soma and apical dendrite of the same layer 5 pyramidal neuron. An action potential is evoked in response
to a depolarizing current pulse injected in the soma (top) or the apical dendrite (bottom). In both cases, the soma fires ahead of the dendrite.
(b) In response to synaptic stimulation in layer 1, the action potential occurs first in the soma.
(c) Latency distribution of the action potential in the apical dendrite as a function of the distance from the soma, in response to somatic
current pulses (closed circles) or distal synaptic stimulation (open circles). Note the linear relation between the time of onset of the dendritic
action potential and the distance from the soma, indicating that the action potential originates in the soma and propagates back into the
apical dendrite.

effect is graded and can be restricted to individual den-
dritic branches. In intradendritic recordings from neo-
cortical pyramidal neurons, Kim et al. (1995) have shown
that IPSPs can reduce, and eventually block, the den-
dritic sodium and calcium spikes generated by current
injections. They also report that the IPSP timing controls
the timing of dendritic spiking, producing phase-ad-
vance or phase-delay effects. These two studies demon-
strate that inhibition can play a fundamental role in the
control of dendritic spiking. Besides its traditional role
of blocking excitatory inputs, inhibition may also control
temporal aspects of the dendritic response.
The mechanisms by which inhibitory inputs produce
these effects are likely to be multiple. The increases in
chloride permeability produced by g-aminobutyric acid
(GABA) could lead to a hyperpolarization of the neuron
or to a shunt of depolarizing current produced by the
concomitant decrease in Rin. Also, an inhibition of the
calcium current by GABA independent from changes in
chloride permeability has been described (Dunlap and
Fischbach, 1978; Sugimori and Llinas, 1985, Soc. Neu-
rosci., abstract) and could be mediated by cisternal or-
ganelles (Benedeczky et al., 1994).
Dendrites Can Be Turned off by Spikes
One noteworthy recent result has been the discovery
that dendritic function can be intrinsically modulated by
activity-dependent mechanisms.
The first evidence came from imaging experiments in
CA1 neurons (Jaffe et al., 1992), where, in response to
a train of spikes generated at the soma by an intracellular
current injection, the authors detected a difference in
calcium and sodium dynamics between proximal and
distal dendrites. While calcium and sodium accumulated
in proximal dendrites during the entire spike train, the
accumulations in distal dendrites were limited to the
first few spikes, indicating that the sodium and calcium
channels responsible for the distal accumulations had
turned off. Hence, sustained firing of the cell produced
higher calcium accumulations in proximal dendrites (Re-
gehr et al., 1989). A similar result was found in neocorti-
cal intrinsically bursting neurons, where a train of anti-
dromic spikes generated a dissociation in the calcium
dynamics of the proximal versus distal apical dendrite
(Yuste et al., 1994). Again, while calcium accumulated
in the proximal compartment of the cell during the entire
train, the distal apical dendrite only increased its calcium
at the beginning of the train. It was hypothesized that
calcium-activated hyperpolarizing conductances medi-
ated the switching off of the distal dendrite.
More direct evidence for an activity-dependent
switching off of CA1 dendrites was obtained recently
with intradendritic recordings (Spruston et al., 1995; Cal-
laway and Ross, 1995). In both studies, the first spike
of an antidromic spike train propagated throughout the
entire dendritic tree, while subsequent spikes were fil-
tered to different degrees. In most distal dendrites, a

88
Review
707
gradual decrement in the amplitude of the action poten-
tials during the train was observed. This activity-depen-
dent filtering was triggered by action potentials and had
a slow time course that produced frequency depen-
dency, i.e., faster firing rates resulted in more attenua-
tion (Callaway and Ross, 1995). Interestingly, all-or-none
filtering was seen at some branch points, producing
spike invasion failures (Spruston et al., 1995). Finally,
the recovery from this activity-dependent inactivation
of a dendrite was accelerated by hyperpolarization
(Spruston et al., 1995).
Thus, although the mechanisms responsible are still
unknown, the intrinsic properties of neurons can turn
off certain dendrites in an activity-dependent fashion.
This produces a situation where the specific location of
the inputs matters, as opposed to a dendritic tree where,
like a computer network, the physical location of the
connection is irrelevant. For instance, synapses located
in distal dendrites may only be sensitive to the first few
spikes from a train of action potentials produced by
the soma, while those located more proximally may be
sensitive to every action potential. This filtering would
make distal synapses sensitive only to changes in the
rate of output of the cell, a mechanism that might be
used to implement synaptic algorithms gated by the
derivative of postsynaptic activity, similar to the delta
learning rules explored in artificial neural networks.
What Is the Minimal Functional Compartment
of the Dendrite?
A central question in dendritic integration is whether
the dendritic tree works as one unit, or whether it has
different functional compartments. We can define a
functional compartment as the morphological unit of
the cell that carries out a determined functional role.
Functional compartments do not necessarily have to
be electrical units, since chemical compartmentalization
may be of fundamental importance to the function of
the neuron. Moreover, certain computations may be im-
plemented chemically, rather than electrically (Sobel
and Tank, 1994).
In 1891, Cajal first described the existence of dendritic
spines (Ramon y Cajal, 1891b) and speculated that their
function was to increase the receptive surface of the
dendrites and enable intimate contacts between axons
and dendrites (Ramon y Cajal, 1904). The idea that
spines were sites of synaptic contacts was confirmed
with electron microscopy (DeRobertis and Bennett,
1955; Palay, 1956). Hence, anatomically, spines are a
basic subunit of the dendrite. Still, no functional informa-
tion existed about dendritic spines until the advent of
calcium imaging techniques. Initial experiments in hip-
pocampal pyramidal neurons showed that individual
spines could have different calcium dynamics than their
parent dendrites, behaving as independent calcium
compartments (Muller and Connor, 1991; Guthrie et al.,
1991). In contrast, experiments imaging spines in cul-
tured hippocampal neurons and Purkinje cells in slices
suggested that spines did not become activated inde-
pendently, but rather as part of a branchlet, composed
of a small segment of a dendrite with all its spines (Mur-
phy et al., 1994; Eilers et al., 1995). The branchlet then
would constitute the minimal compartment of the den-
drite. Recently, this issue has been reexamined with
two-photon microscopy (Denk et al., 1990), which en-
ables imaging of spines in brain slices for extended
periods of time (Figure 4). New results in CA1 and Pur-
kinje neurons show that individual spines are activated
under subthreshold synaptic stimulation, but that once
the cell fires an action potential, every spine in the cell
becomes activated (Yuste and Denk, 1995; Denk et al.,
1995). These findings therefore show that spines behave
as chemical compartments that could constitute the
minimal functional compartment of the dendrite.
What function might spines serve? The restriction of
calcium accumulations during subthreshold synaptic
stimulation to a single spine indicates that spines com-
partmentalize biochemical changes (Gamble and Koch,
1987; Wickens, 1988; Koch and Zador, 1993). This com-
partmentalization could segregate plastic changes in
synaptic transmission (Rall, 1974; Lisman, 1989). In fact,
electrophysiological experiments in vivo have implied
that individual spines may be specifically potentiated
because granule cell spines contacted by the contralat-
eral perforant pathway, but interspersed among spines
contacted by the ipsilateral pathway, are not potentiated
together with the rest of the spines (Levy and Steward,
1979; Levy and Desmond, 1985). In agreement with this,
supralinear calcium accumulations occurred in hippo-
campal spines when synaptic and spike activation oc-
curred simultaneously, indicating that spines may be
able to function as coincidence detectors (Yuste and
Denk, 1995). An implication from this argument is that
smooth neurons, with few or no spines, cannot compart-
mentalize plastic changes in synaptic transmission like
spiny neurons (Wickens, 1988), if plasticity is not pro-
duced by highly localized, submembrane calcium con-
centrations. In preliminary experiments from CA1 stel-
late cells, changes of synaptic efficacy in one input do
influence a second distant input, suggesting that spines
are necessary to compartmentalize plasticity (S. J. Red-
man, personal communication).
Interaction among Dendritic Compartments
Even if spines are the smallest functional compartment
of a dendrite, a more pertinent question is what the
typical functional compartments are under the condi-
tions in which dendrites are integrating inputs. In fact,
there is clear evidence that dendritic compartments of
a larger scale than a dendritic spine may be functional
units. Whereas anatomical evidence suggests that sev-
eral functional areas exist in dendrites, physiological
data show that different dendritic compartments can
interact to generate intrinsic functional dynamics in the
neuron.
While studying the development of synaptic connec-
tions in cilliary ganglion cells, Purves and Hume found
that developing neurons were innervated by a similar
number of axons, but in adult neurons the number of
innervating axons precisely matched the number of pri-
mary dendrites (Purves and Hume, 1981; Hume and
Purves, 1981). They concluded that cilliary ganglion cells
had a number of separate spatial dendritic domains,
each one receiving input from a single axon. This implied
that dendrites could serve to isolate inputs from one
another and suggested that a single axon would only
innervate one dendrite. Nevertheless, in a follow-up
study with electron microscopy, single dendrites were
89
Neuron
708
shown to be innervated by multiple axons, although a
spatial segregation of competing inputs was still present
(Forehand and Purves, 1984). Altogether, these results
suggested the existence of different dendritic compart-
ments dominated by different inputs. A similar relation-
ship between dendritic domains and functional com-
partments was suggested by reconstructions of neurons
from the tecta of three-eyed frogs (Katz and Con-
stantine-Paton, 1988). In these animals, the tectum is
divided into interdigitating stripes, alternatively inner-
vated by each eye. In some tectal neurons individual
dendritic branches were confined to the stripe of a single
eye, although the entire dendritic tree occupied more
than one stripe. This result suggested that dendritic
branches were functional units, because they received
only one type of input.
The first demonstration of a functional interaction by
separate compartments in a neuron came from work in
inferior olive cells (Llinas and Yarom, 1981). From the
mismatch in the sign of the spikes between intra- and
extracellular recordings, the authors deduced that so-
dium and low threshold calcium conductances were
located in the soma, while high threshold calcium con-
ductances were dendritic. The spatial separation of the
conductances into somatic and dendritic compartments
generated an intrinsic oscillatory dynamics of these neu-
rons, following this sequence of events: a somatic so-
dium spike (1) triggers a dendritic high threshold calcium
Figure 4. Two-Photon Imaging of Dendritic
Spines in Purkinje Neurons
Reprinted with permission from Denk et al.
(1995). The top panel (a) shows the complete
dendritic tree of a Purkinje neuron in a 400
mm guinea-pig cerebellar slice, filled with the
calcium indicator Calcium Green. In the lower
panels (bd), progressively higher magnifica-
tion images from the dendritic tree of a differ-
ent neuron are shown. Dendritic spines are
clearly visible. Image taken with a custom-
made laser scanning microscope using two-
photon excitation.
spike (2), which activates a calcium-dependent potas-
sium current (3) that deinactivates a somatic low thresh-
old calcium spike (4), which finally triggers again a so-
matic sodium spike (1).
A second case in which the interaction between den-
dritic and somatic compartments produces intrinsic
electrophysiological responses is the generation of
spike bursting by CA3 neurons. There is evidence that
high threshold calcium conductances, present in den-
drites from CA3 pyramidal neurons, are necessary for
burst responses (Wong and Prince, 1978). While single
compartment models of CA3 pyramidal neurons have
been unable to reproduce adequately the electrophysi-
ology, the nonuniform distribution of calcium and so-
dium channels into at least two compartments can gen-
erate burst responses (Traub et al., 1991; Pinsky and
Rinzel, 1994; Rhodes and Gray, 1994). Consistent with
this, imaging of CA3 neurons during bursts and single
spike firing shows that normalized calcium accumula-
tions are larger in dendrites than in the soma during
bursting (Strowbridge and Tank, 1994, Soc. Neurosci.,
abstract). Thus, the sequence of events during a burst
would be as follows: somatic sodium spike (1) that trig-
gers a dendritic calcium spike (2) that depolarizes the
soma producing the envelope of the burst and triggering
additional somatic sodium spikes (3). Finally, a recent
simulation study has extended these ideas by sug-
gesting that even the detailed dendritic morphology of
neocortical pyramidal neurons could influence many of
90
Review
709
their intrinsic electrophysiological properties (T. J. Sej-
nowski, personal communication).
Dendrites Can Serve as the Output Side
of the Neuron
Dendrites in many invertebrate nervous systems serve
as both input and output side of the neuron (Bullock,
1977; Watson and Burrows, 1988), and this motivated
the use of the term neurite instead of dendrite to ex-
press the lack of specialization among the processes.
Since Cajal, however, dendrites in vertebrate systems
have been considered the input side of the neuron. Nev-
ertheless, the existence of dendrodendritic synaptic in-
teraction and dendrodendritic gap junctions are clear
contradictions of Cajals law and may constitute general
forms of neuronal communications present throughout
the vertebrate central nervous system.
Dendrodendritic synapses have been clearly docu-
mented in mammalian olfactory bulb and retina. In the
bulb, the antidromic excitation of mitral cells is followed
by a long-lasting inhibition (Phillips et al., 1963). In a
computational and electrophysiological study, it was
proposed that secondary dendrites of mitral cells ex-
cited granule cells, which would then mediate the inhibi-
tion of mitral cells (Rall and Shepherd, 1968). Because
granule cells lack axons, their output had to be dendritic.
Finally, reciprocal dendrodendritic synapses between
mitral and granule cells were found with electron micros-
copy (Rall et al., 1966). The circuitry of the retina also
provides, in the amacrine cell, one clear example of a
neuron in which dendritic spines and axonal terminals
are not morphologically segregated to different pro-
cesses (Dowling, 1968; Dubin, 1970). Finally, a different
form of dendrodendritic synaptic interactions is the den-
dritic release of neurotransmitters by substantia nigra
neurons (Cheramy et al., 1981; Llinas and Greenfield,
1987). This release is calcium sensitive and may require
exocytosis from subcisternal systems.
An issue that has reemerged in the last decade is
whether mammalian neurons are coupled via dendritic
gap junctions, as was first established in olivary neurons
(Sotelo et al., 1974; Llinas et al., 1974). The existence of
dendrodendritic gap junctions was suggested to explain
dye coupling in neocortical and hippocampal slices
(Gutnick and Prince, 1981; MacVicar and Dudek, 1981).
This coupling is particularly dramatic in developing cor-
tex (Connors et al., 1983; Lo Turco and Kriegstein, 1991;
Peinado et al., 1993), where it can lead to coactivation
of large groups of neurons (Yuste et al., 1992, 1995).
These functional results indicate that in development
dendrites can serve as the output side of the neuron.
In the adult brain, the extent of dendritic gap-junctional
coupling remains a matter of controversy. While inferior
olive neurons (Sotelo et al., 1974), granule cells in the
olfactory bulb (Reyher et al., 1991), and many neuronal
types in the retina (Vaney, 1991) have dendrodendritic
gap junctions, in adult pyramidal neurons dye neuronal
coupling practically disappears (Connors et al., 1983;
Peinado et al., 1993; cf. Michelson and Wong, 1994).
Evidence for gap junctions in adult cortex and hippo-
campus from electron microscopy is scant (Sloper,
1972; Powell, 1981; Kosaka, 1983) and could be due to
the difficulty of detecting them. Nevertheless, recent
studies suggest that there are, in fact, dendritic gap
junctions in adult neocortex (Parnavelas et al., 1994,
1995, Soc. Neurosci., abstracts) and nonpyramidal neu-
rons in hippocampus (Gulyas et al., submitted; Acsady
et al., submitted). Finally, dendritic gap junctions among
interneurons may be necessary to generate hippocam-
pal oscillations in the presence of synaptic blockers
(Michelson and Wong, 1994; Traub, 1996).
Future Issues: The Time
of Dendritic Computation
In this last section, we review experimental and theoreti-
cal findings that address more speculative issues of
dendritic integration. A review covering many theoretical
results has been recently published (Mel, 1994).
Direct Measurements of Cable Parameters
Simultaneous somatic and dendritic recordings can be
used to measure the cable attenuation produced in test
voltage pulses (Stuart and Hausser, 1994; Stuart and
Sakmann, 1994; Sugimori and Llinas, 1984, Soc. Neu-
rosci., abstract). This approach could be used to esti-
mate Ri for the dendritic segment between the two elec-
trodes and establish the basic cable parameters of the
main dendritic shafts. The difficulty in obtaining whole-
cell recordings from small dendritic branches, however,
makes this strategy less optimal when trying to estimate
the values of cable parameters in secondary and tertiary
dendrites, which are target regions for the majority of
the inputs (Larkman, 1991).
As an alternative approach, voltage-sensitive dyes
(VSDs) have been used to measure voltage transients
in small processes from cultured invertebrate neurons
(Grinvald et al., 1981). In a recent study, cable equations
were fitted to measured spatiotemporal voltage pat-
terns, and tentative values for Rm (z22 kVcm2 ) and Ri
(z250 Vcm) were concluded (Fromherz and Muller,
1994). These values, incidentally, are not that different
from those found in numerical simulation studies (see
above).
It should be possible to apply optical recording with
VSDs to cultured mammalian neurons, or even to neu-
rons in a slice, and estimate Ri and Rm more directly.
Efforts in that direction have already started, using tradi-
tional dyes (Borst et al., 1995, Soc. Neurosci., abstract)
or injectable VSDs that may obviate poor signal-to-
background ratio when recording optically from den-
drites in a slice (Antic and Zecevic, 1995; Kogan et al.,
1996). Also, ratio imaging may enable accurate quantita-
tive measurements of dendritic voltages (Montana et al.,
1989; Gonzalez and Tsien, 1995). Optical approaches
seem ideally suited to investigate questions such as
whether cable parameters are constant in the dendritic
tree, or what the behavior of branch points is during the
propagation of voltage signals.
What Is the Role of the Active
Conductances in Dendrites?
Although it is clear that there are active conductances
throughout the dendritic trees of both pyramidal and
Purkinje cells, the role of these conductances remains a
mystery. Because they are voltage sensitive, they could
amplify synaptic inputs and help relay distal EPSPs to
the soma (Lorente de No and Coundouris, 1959; Shep-
herd et al., 1985). In fact, many recent studies demon-
strate that synaptic inputs are indeed amplified by the
91
Neuron
710
dendritic tree. For example, in neocortical neurons, de-
polarization of the soma increases the amplitude and
duration of EPSPs (Stafstrom et al., 1985; Thompson et
al., 1988; Sutor and Hablitz, 1989; Hirsch and Gilbert,
1991; Cauller and Connors, 1994). This subthreshold
amplification is gradual and has been attributed to either
sodium, calcium, or NMDA receptor conductances. A
recent report has suggested that it may preferentially
occur near the soma (Stuart et al., 1995, Soc. Neurosci.,
abstract). In addition, EPSPs can become amplified into
the suprathreshold regime, evoking dendritic spikes. In
fact, in Purkinje cells, climbing fiber input triggers den-
dritic spikes (Llinas and Nicholson, 1971), and a similar
situation occurs in neocortex with strong layer 1 inputs
to apical dendrites (Cauller and Connors, 1992). Active
conductances in dendrites may act regeneratively under
some conditions and in a subthreshold regime under
others, thus providing dendrites with two different
modes of operation.
Another possible function for active conductances is
the linearization of synaptic inputs, i.e, the progressive
changes in electrical properties induced to avoid satura-
tion by synchronous inputs (Bernarder et al., 1994). This
idea builds upon research proposing that active conduc-
tances control the gain of synaptic inputs (Laurent and
Burrows, 1989). Finally, as explained above, dendritic
voltage-activated conductances may also regulate the
intrinsic temporal dynamics of neurons (Llinas, 1988).
What Is the Role of Antidromic Spikes?
In many neurons, action potentials originating at the
soma can invade the dendritic tree (see above) and, in
CA1 neurons, it even reaches every spine in the cell
(Yuste and Denk, 1995). What functional significance
could this antidromic or backpropagating spike have?
The temporal coincidence of synaptic inputs with post-
synaptic spikes (i.e., the output of the neuron) could
trigger a change in synaptic efficacy, analogous to the
Hebb rule (Hebb, 1949). Thus, the spike might be the
temporal signal used for coincidence detection by the
spines. This idea has been corroborated in experiments
showing that action potential blockade by local perfu-
sion of TTX blocks long-term potentiation (LTP) (Scharf-
man and Sarvey, 1985), although LTP persists under
sodium (but not calcium) action potential blockade with
intracellular QX-314 (Gustafsson et al., 1987). Further-
more, recent results show that the temporal coincidence
of synaptic inputs and spikes produces cooperative cal-
cium accumulations in dendritic spines (Yuste and Denk,
1995). Studies of LTP in identified connections could
help resolve this issue (Markram and Sakmann, 1995,
IBRO, abstract).
Do Different Dendritic Compartment
Carry out Different Computations?
Anatomical results suggest the possibility of a relation-
ship between dendritic morphology and functional com-
partments (see above). In agreement with this, Purkinje
cells have clear functional differences in the action of
climbing and parallel fibers (Llinas and Sugimori, 1992),
inputs that impinge onto different areas of the dendritic
tree (Ramon y Cajal, 1904), although the exact location
of the parallel fiber input may not matter (De Schutter
and Bower, 1994). Therefore, it is possible to speak of
two functional compartments in the dendritic tree of
Purkinje cells, which might be specialized for different
computations. For instance, timing information (when)
could be provided by climbing fibers and preserved in
the complex spike, while other types of information
(what) might be conveyed by parallel fibers and plateau
potentials (Llinas and Sugimori, 1992).
Pyramidal neurons also have segregated dendritic
fields for specific inputs. In hippocampus, CA3 pyrami-
dal neurons are contacted by mossy fiber terminals se-
lectively at the base of apical dendrites (Claiborne et
al., 1986), while different populations of interneurons
contact specific areas of dendritic trees of pyramidal
neurons (Buhl et al., 1994). In neocortex, feedback axo-
nal projections make specific contacts with the tips of
apical dendrites, located in layer 1 (Rockland and Virga,
1989), while stellate cells can project specifically to the
base of the apical dendrite (Kisvarday et al., 1987). In
fact, the laminar structure of the cortex is a direct reflec-
tion of specific input/output relations, and particular lay-
ers can be thought of as having a functional identity
(Gilbert and Wiesel, 1979; Zeki, 1993), which could be
preserved in the dendritic tree. Nevertheless, it is still
unknown whether neurons can distinguish between in-
puts coming from different layers. Also, the reason be-
hind the morphological differences between apical and
basal dendritic trees is still mysterious. Electrophysio-
logical and imaging data indicates that distal apical den-
drites are a separate functional compartment from the
rest of the cell (Yuste et al., 1994), and data from awake
behaving monkeys suggests that they could mediate
specific behavioral responses (Cauller and Kulics, 1991).
Do Dendrites Translate Spatial
Patterns of Inputs?
The use of the spatial properties of the dendritic tree to
filter directional differences in input was proposed by
Rall (1964). In his model with a passive dendritic tree,
excitatory inputs located in a line would preferentially
excite the soma if they were activated in an ordered
sequence, with distal inputs excited ahead of proximal
ones. This resulted from the time delays associated with
the transmission of EPSPs to the soma and suggested
that dendrites could serve to detect spatiotemporal pat-
terns of inputs. A similar idea, incorporating inhibitory
inputs, was used to explain direction selectivity by the
retina (Koch et al., 1983). In this model, the topographic
map of visual inputs into a dendritic tree produced a
direction-selective response if excitatory inputs were
distal to inhibitory ones. Although recent results suggest
that the extent of the dendritic tree is the same as the
extent of the receptive field of rabbit ganglion cells (Yang
and Masland, 1992), the location of excitatory and inhibi-
tory inputs into ganglion cell dendrites does not seem
to follow any spatial rule (T. Famiglietti, personal com-
munication).
Do Dendrites Read Temporal Codes?
A closely related issue is whether dendrites are built to
detect temporal input patterns. This issue is actively
debated because of the increased awareness that pre-
cise timing may be used by the nervous system to code
information (Abeles et al., 1993; Softky and Koch, 1992;
Hopfield, 1995). In cerebellum, recent results show that
submillisecond timing differences from the olive are pre-
served by the olivocerebellar projection by subtle ad-
justments in the conduction velocity of climbing fiber
92
Review
711
axons (Sugihara et al., 1993). This striking finding sug-
gests that precise timing is essential for cerebellar com-
putations. In neocortex, however, it is controversial
whether precise timing is utilized (Schadlen and New-
some, 1994). On the one hand, repeated temporal firing
patterns can be detected in cortical neurons in vivo
(Abeles et al., 1993), and millisecond temporal syn-
chrony has been proposed to underlie the solution of
the binding problem by primary visual cortex (Gray et
al., 1989). In addition, the firing of neocortical neurons
in vitro appears surprisingly reliable in response to injec-
tions of noisy current waveforms (Mainen and Sejnow-
ski, 1995). On the other hand, the behavior of awake
monkeys performing perceptual tasks can be predicted
from the average firing rate of MT neurons (Zohary et
al., 1994; Schadlen and Newsome, 1994). In addition,
the time constants of many of the biophysical events in
neocortical neurons appear too slow for implementation
of millisecond timing detection, while, for example, neu-
rons from the cochlear nucleus, specialized in temporal
detection, have dendritic morphologies and receptors
that enable fast transmission (Rubel and Parks, 1988;
Raman and Trussell, 1992).
Experimental evidence of temporal selectivity by den-
dritic trees is scarce. An intriguing study of neurons in
the torus semicircularis in the electric fish has recently
shown two populations of neurons with differential fre-
quency sensitivity to sensory stimulation (Rose and Call,
1993). The cell type that responds better to lower fre-
quencies has dendritic spines, while the type that re-
sponds better to higher frequencies lacks them. The
authors have suggested that spines serve as frequency
filters.
Do Dendrites Implement Logical Elements?
Imaginative proposals have suggested that dendrites
are small logical elements (Koch et al., 1983; Shepherd
and Brayton, 1987; Rall and Segev, 1987). The basic
idea is that summation or shunting occurs when excit-
atory and inhibitory inputs interact in a dendritic tree. If
there is a nonlinear threshold mechanism afterward,
such as an action potential, these interactions can be
translated into a binary decision. For example, an AND
gate would result if two excitatory inputs are simultane-
ously active, an OR gate would result if there is satura-
tion by either one, and an AND/NOT gate would result
if there is a response to an excitatory input, provided
that a shunting inhibitory one is not active. Although
intriguing, there is at present no experimental evidence
to support or contradict these ideas.
Studying Dendrites during Behavior:
A Lesson from Invertebrate Work
Perhaps the major obstacle in the study of the function
of mammalian dendrites is our lack of understanding of
the particular computational tasks that these cells are
solving. This problem has been tackled in invertebrate
systems, where dendrites can be recorded optically or
electrically during behavioral tasks (Borst and Egelhaaf,
1994). Although conclusions reached from invertebrate
work on dendrites may not be applicable to understand
mammalian neurons, the general approach of identifying
the dendritic computation by studying the dendrites dur-
ing the performance of a behavior may be key toward
revealing the logic behind the design of dendrites.
An example of this behavioral approach is the study
of the optomotor response in the fly (Borst and Egelhaaf,
1992). Using calcium imaging, the visual field was shown
to map topographically into the dendritic field of the
neuron. This topographic arrangement of inputs could
enable local amplification among inputs coming from
the same area of the visual space. A different strategy
appears to be employed by dendrites from auditory in-
terneurons in the female cricket (Sobel and Tank, 1994;
Figure 5). With in vivo intracellular recording and calcium
imaging from identified neurons, it was demonstrated
that the dendrite becomes progressively inhibited in re-
sponse to loud chirps. This inhibition is mediated by
a calcium-activated hyperpolarizing current and could
contribute to the forward masking of auditory stimuli
(Pollack, 1988). An additional example comes from the
recent study of collision-sensitive neurons in the visual
system of the locus (Hatsopoulos et al., 1995). In this
system, the dendritic tree of a single neuron is thought
to carry out the multiplication of two independent input
signals that enable the neuron to calculate the time of
collision of an approaching object. Finally, the role of
active conductances have been investigated in visual
interneurons in the fly, by comparing the responses of
a type of neuron with dendritic sodium action potentials
with another type that lacks them (Haag and Borst,
1996). In this study, a correlation was found between
synaptic responses to higher temporal frequencies and
active dendrites, suggesting that dendritic action poten-
tials might be used for amplification of transient inputs.
A similar research program studying dendrites from
vertebrate neurons in vivo is feasible. Intracellular re-
cordings from mammalian dendrites can be carried out
in anesthetized animals (Llinas and Nicholson, 1971;
Houchin, 1973; Pockberger, 1991; B. W. Connors, per-
sonal communication), while intracellular recordings
from awake behaving monkeys have been achieved
(Matsumura et al., 1988; Glenn et al., 1988). In addition,
two-photon microscopy can be used to image individual
neurons in anesthetized rats (Denk et al., 1994). More-
over, the study of nonmammalian vertebrate systems
in which the basic cellular machinery may be highly
homologous to that of mammalian neurons, but in which
experimental access to behaving preparations is easier,
should speed up these investigations (Connors and
Kriegstein, 1986; Heiligenberg, 1991; Fetcho and OMal-
ley, 1995). It appears to us that the combination of the
new generation of electrophysiological and imaging
techniques with the biological relevance of studies in
behaving animals holds many fruitful experimental ave-
nues for those who wonder, like Cajal did a century
ago, what are the functions hidden behind the intriguing
morphologies of the dendrites.
Acknowledgments
We thank A. Borst, B. W. Connors, T. Freund, S. J. Redman, W. N.
Ross, T. J. Sejnowski, and R. D. Traub for sharing their results before
publication, A. Borst, S. Cash, B. W. Connors, W. Denk, S. R. Golob,
R. Llinas, W. Regehr, P. A. Rhodes, I. Segev, and K. Svoboda for
helpful comments, and A. Gelperin, J. Lichtman, Z. Mainem, V. Mur-
thy, G. M. Shepherd, L. Trussell, and R. O. L. Wong for their help.
R. Y. was supported by the Office of Naval Research.
93
Neuron
712

Figure 5. In Vivo Calcium Imaging from a Cricket Auditory Interneuron during a Behavioral Response
Courtesy of E. Sobel, Harvard University. Pseudocolor fluorescence images depicting increases in [Ca2 1]i in an omega neuron from a live
female cricket, in response to a simulated male song. The neurites that receive input from auditory fibers are located in the lower left quadrant
of the cell. The cell was filled with the calcium indicator fura-2 using intracellular iontophoretic injection. Violet/blue corresponds to resting
[Ca2 1]i, while red corresponds to high [Ca21]i. The four images were taken with a cooled CCD camera over a period of 20 s, and they are
displayed in the following order: top left, bottom left, top right, bottom right. Scale bar is 50 mm.

References Borst, A., and Egelhaaf, M. (1992). In vivo imaging of calcium accu-
mulation in fly interneurons as elicited by visual motion stimulation.
Abeles, M., Bergman, H., Margalit, E., and Vaadia, E. (1993). Spatio- Proc. Natl. Acad. Sci. USA 89, 41394143.
temporal firing patterns in the frontal cortex of behaving monkeys.
Borst, A., and Egelhaaf, M. (1994). Dendritic processing of synaptic
J. Neurophysiol. 70, 16291638.
information by sensory interneurons. Trends Neurosci. 17, 257273.
Allbritton, N.L., Meyer, T., and Stryer, L. (1992). Range of messenger
Brock, L.G., Coombs, J.S., and Eccles, J.C. (1952). The recording
action of calcium ion and inositol 1,4,5-trisphosphate. Science 258,
of potentials from motoneurones with an intracellular electrode. J.
18121815.
Physiol. 117, 431460.
Amitai, Y., Friedman, A., Connors, B.W., and Gutnick, M.J. (1993).
Buhl, E.H., Halasy, K., and Somogyi, P. (1994). Diverse sources
Regenerative activity in apical dendrites of pyramidal cells in neo-
of hippocampal unitary inhibitory postynaptic potentials and the
cortex. Cerebral Cortex 3, 2638.
number of synaptic release sites. Nature 368, 823828.
Andersen, P., Holmquist, B., and Voorhoeve, P.E. (1966). Excitatory
synapses on hippocampal apical dendrites activated by entorhinal Bullock, T.H. (1977). Introduction to Nervous Systems (San Fran-
stimulation. Acta Physiol. Scand. 66, 461472. cisco: Freeman).

Antic, S., and Zecevic, D. (1995). Optical signals from neurons with
internally applied voltage-sensitive dyes. J. Neurosci. 15, 1392
1405.
Benedeczky, I., Molnar, E., and Somogyi, P. (1994). The cisternal
organelle as a Ca21-storing compartment associated with GABAer-
gic synapses in the axon initial segment of hippocampal pyramidal
neurones. Exp. Brain Res. 101, 216230.
Bernarder, O., Douglas, R.D., Martin, K.A., and Koch, C. (1991).
Synaptic background activity influences spatiotemporal integration
in single pyramidal cells. Proc. Natl. Acad. Sci. USA 88, 11569
11573.
Bernarder, O., Koch, C., and Douglas, R.J. (1994). Amplification and
linearization of distal synaptic input to cortical pyramidal cells. J.
Neurophysiol. 72, 27432753.
Bernardo, L.S., Masukawa, L.M., and Prince, D.A. (1982). Electro-
physiology of isolated hippocampal pyramidal dendrites. J. Neu-
rosci. 2, 16141622.
Blanton, M.G., Lo Turco, J.J., and Kriegstein, A.R. (1989). Whole-
cell recording from neurons in slices of reptilian and mammalian
cerebral cortex. J. Neurosci. Meth. 30, 203210.
Callaway, J.C., and Ross, W.N. (1995). Frequency-dependent propa-
gation of sodium action potentials in dendrites of hippocampal CA1
pyramidal neurons. J. Neurophysiol. 74, 19.
Callaway, J.C., Lasser-Ross, N., and Ross, W.N. (1995). IPSPs
strongly inhibit climbing fiberactivated [Ca2 1] in the dendrites of
cerebellar Purkinje neurons. J. Neurosci. 15, 27772787.
Cauller, L.J., and Connors, B.W. (1992). Functions of very distal
dendrites: experimental and computational studies of layer I syn-
apses on neocortical pyramidal cells. In Single Neuron Computation,
T. McKenna, ed. (San Diego: Academic Press), pp. 199229.
Cauller, L.J., and Connors, B.W. (1994). Synaptic physiology of hori-
zontal afferents to layer 1 in slices of rat S1 neocortex. J. Neurosci.
14, 751762.
Cauller, L.J., and Kulics, A.T. (1991). The neural basis of the behav-
iourally relevant N1 component of the somatosensory-evoked po-
tentials of S1 cortex of awake monkeys: evidence that backward
cortical projection signals conscious touch sensation. Exp. Brain
Res. 84, 607619.
Cheramy, A., Leviel, V., and Glowinski, J. (1981). Dendritic release
of dopamine in the substantia nigra. Nature 289, 537542.

94
Review
713
Claiborne, B.J., Amaral, D.G., and Cowan, W.M. (1986). A light and
electron microscopic analysis of the mossy fibers of the rat dentate
gyrus. J. Comp. Neurol. 246, 435458.
Connor, J.A. (1986). Digital imaging of free calcium changes and of
spatial gradients in growing processes in single, mammalian central
nervous system cells. Proc. Natl. Acad. Sci. USA 83, 61796183.
Connors, B.W., and Kriegstein, A.R. (1986). Cellular physiology of the
turtle visual cortex: distinctive properties of pyramidal and stellate
neurons. J. Neurosci. 6, 164177.
Connors, B.W., Bernardo, L.S., and Prince, D.A. (1983). Coupling
between neurons of the developing rat neocortex. J. Neurosci. 3,
773782.
Coombs, S.J., Eccles, J.C., and Fatt, P. (1955a). Excitatory synaptic
actions in motoneurons. J. Physiol. 130, 374395.
Coombs, S.J., Eccles, J.C., and Fatt, P. (1955b). The specific ionic
conductances and the ionic movements across the motoneuronal
membrane that produce the inhibitory post-synaptic potential. J.
Physiol. 130, 396413.
Coombs, S.J., Curtis, D.R., and Eccles, J.C. (1957). The interpreta-
tion of spike potentials of motorneurones. J. Physiol. 139, 198231.
Deiters, O. (1865). Untersuchungen uber Gehirn und Ruckenmark
des Menschen und der Saugethiere. (Braunschweig).
Denk, W., Strickler, J.H., and Webb, W.W. (1990). Two-photon laser
scanning fluorescence microscopy. Science 248, 7376.
Denk, W., Delaney, K.R., Gelperin, A., Kleinfeld, D., Strowbridge,
B.W., Tank, D.W., and Yuste, R. (1994). Anatomical and functional
imaging of neurons using 2-photon laser scanning microscopy. J.
Neurosci. Meth. 54, 151162.
Denk, W., Sugimori, M., and Llinas, R. (1995). Two types of calcium
response limited to single spines in cerebellar Purkinje cells. Proc.
Natl. Acad. Sci. USA 92, 82798282.
DeRobertis, E.D.P., and Bennett, H.S. (1955). Some features of the
submicroscopic morphology of synapses in frog and earthworm. J.
Biophys. Biochem. Cytol. 1, 4758.
Deschenes, M. (1981). Dendritic spikes induced in fast pyramidal
tract neurons by thalamic stimulation. Exp. Brain Res. 43, 304308.
De Schutter, E. (1992). A consumer guide to neuronal modeling
software. Trends. Neurosci. 15, 462464.
De Schutter, E., and Bower, J.M. (1994). Simulated responses of
cerebellar Purkinje cells are independent of the dendritic location
of granule cell synaptic inputs. Proc. Natl. Acad. Sci. USA 91, 4736
4740.
Dodt, H.-U., and Zieglgansberger, W. (1990). Visualizing unstained
neurons in living brain slices by infrared DIC. Brain Res. 537,
333336.
Dodt, H.-U., and Zieglgansberger, W. (1994). Infrared videomicros-
copy: a new look at neuronal structure and function. Trends Neu-
rosci. 17, 453458.
Dowling, J.E. (1968). Synaptic organization of the frog retina: an
electron microscopic analysis comparing the retinas of frogs and
primates. Proc. R. Soc. Lond. (B) 170, 205228.
Dubin, M.W. (1970). The inner plexiform layer of the vertebrate retina:
a quantitative and comparative electron microscope analysis. J.
Comp. Neurol. 140, 479506.
Dunlap, K., and Fischbach, G.D. (1978). Neurotransmitters decrease
the calcium component of sensory neuron action potentials. Nature
276, 837838.
Eccles, J.C. (1964). The Physiology of Synapses (Berlin: Springer),
pg. 104.
Eccles, J.C., Libet, B., and Young, R.R. (1958). The behavior of
chromatolysed motoneurons studied by intracellular recording. J.
Physiol. (Lond.) 143, 1140.
Edwards, F.A., Konnerth, A., Sakmann, B., and Takahashi, T. (1989).
A thin slice preparation for patch recordings from neurones of mam-
malian central nervous system. Pflugers Arch. 414, 600612.
Eilers, J., Augustine, G.J., and Konnerth, A. (1995). Subthreshold
synaptic Ca2 1 signalling in fine dendrites and spines of cerebellar
Purkinje neurons. Nature 373, 155158.
Fatt, P. (1957). Sequence of events in synaptic activation of a moto-
neurone. J. Neurophysiol. 20, 6180.
Fetcho, J.R., and OMalley, D.M. (1995). Visualization of active neural
circuitry in the spinal chord of intact zebrafish. J. Neurophys. 73,
399406.
Fine, A., Amos, W.B., Durbin, R.M., and McNaughton, P.A. (1988).
Confocal microscopy: applications in neurobiology. Trends Neu-
rosci. 11, 345351.
Forehand, C.J., and Purves, D. (1984). Regional innervation of rabbit
ciliary ganglion cells by the terminals of preganglionic axons. J.
Neurosci. 4, 112.
Fromherz, P., and Muller, C.O. (1994). Cable properties of a straight
neurite of a leech neuron probed by a voltage-sensitive dye. Proc.
Natl. Acad. Sci. USA 91, 46044608.
Fuortes, M.G.F., Frank, K., and Becker, M.C. (1957). Steps in the
production of the motoneuron spikes. J. Gen. Physiol. 40, 735752.
Gamble, E., and Koch, C. (1987). The dynamics of free calcium
in dendritic spines in response to repetitive input. Science 236,
13111315.
Gilbert, C., and Wiesel, T.N. (1979). Morphology and intracortical
projections of functionally characterised neurons in the cat visual
cortex. Nature 280, 120125.
Glenn, L.L., Whitney, J.F., Rewitzer, J.S., Salamone, J.A., and Mari-
ash, S.A. (1988). Method for stable intracellular recordings of spinal
a-motoneurons during treadmill walking in awake, intact cats. Brain
Res. 439, 396401.
Goeppert-Mayer, M. (1931). Ueber Elementareakte mit zwei Quan-
tensprungen. Ann. Physik 9, 273283.
Gonzalez, J.E., and Tsien, R.Y. (1995). Voltage sensing by fluores-
cence resonacne energy transfer in single cells. Biophys. J. 69,
12721280.
Gray, C.M., Konig, P., Engel, A.K., and Singer, W. (1989). Oscillatory
responses in cat visual cortex exhibit inter-columnar synchroniza-
tion which reflects global stimulus properties. Nature 338, 334337.
Grinvald, A., Ross, W.N., and Farber, I. (1981). Simultaneous optical
measurements of electrical activity from mutiple sites on processes
of cultured neurons. Proc. Natl. Acad. Sci. USA 78, 31453249.
Gustafsson, B., Wigstrom, H., Abraham, W.C., and Huang Y.-Y.
(1987). Long-term potentiation in the hippocampus using depolariz-
ing current pulses as the conditioning stimulus to single volley syn-
aptic potentials. J. Neurosci. 7, 774780.
Guthrie, P.B., Segal, M., and Kater, S.B. (1991). Independent regula-
tion of calcium revealed by imaging dendritic spines. Nature 354,
7680.
Gutnick, M.J., and Prince, D.A. (1981). Dye coupling and possible
electrotonic coupling in the guinea pig neocortical slice. Science
211, 6770.
Haag, J., and Borst, A. (1996). Amplification of high-frequency syn-
aptic inputs by active dendritic membrane processes. Nature 379,
639641.
Hatsopoulos, N., Gabbiani, F., and Laurent, G. (1995). Elementary
computation of object approach by a wide-field visual neuron. Sci-
ence 270, 10001003.
Hausser, M., Stuart, G., Racca, C., and Sakmann, B. (1995). Axonal
initiation and active dendritic propagation of action potentials in
substantia nigra neurons. Neuron 15, 637647.
Hebb, D.O. (1949). The Organization of Behaviour (New York: Wiley).
Heiligenberg, W. (1991). Neural Nets in Electric Fish: Computational
Neuroscience Series (Cambridge, Massachusetts: MIT Press).
Hermann, L. (1905). Beitrage zur Physiologie und Physiks des
Nerven. Pflugers Arch. 109, 95120.
Hillman, D., Chen, S., Aung, T.T., Cherksey, B., Sugimori, M., and
Llinas, R.R. (1991). Localization of P-type calcium channels in the
central nervous system. Proc. Natl. Acad. Sci. USA 88, 70767080.
Hirsch, J.A., and Gilbert, C.D. (1991). Synaptic physiology of hori-
zontal connections in the cats visual cortex. J. Neurosci. 11, 1800
1809.
95
Neuron
714
Hirsch, J.A., Alonso, J.-M., and Reid, R.C. (1995). Visually evoked
calcium action potentials in cat striate cortex. Nature 378, 612616.
His, W. (1893). Uber den Aufbau unseres Nervensystems. Berl. Klin.
Wochenschr. 40, 41.
Hodgkin, A.L., and Keynes, R.D. (1957). Movements of labelled cal-
cium in squid giant axons. J. Physiol. 138, 253281.
Hodgkin, A.L., and Rushton, W.A.H. (1946). The electrical constants
of a crustacean nerve fiber. Proc. R. Soc. Lond. (B) 133, 444479.
Hopfield, J.J. (1995). Pattern recognition computation using action
potential timing for stimulus representation. Nature 376, 3336.
Houchin, J. (1973). Procion Yellow electrodes for intracellular re-
cording and staining of neurons in the somatosensory cortex of the
rat. J. Physiol. 232, 6769.
Huguenard, J.R., Hamill, O.P., and Prince, D.A. (1989). Sodium chan-
nels in dendrites of rat cortical pyramidal neurons. Proc. Natl. Acad.
Sci. USA 86, 24732477.
Hume, R.I., and Purves, D. (1981). Geometry of neonatal neurons
and the regulation of synaptic elimination. Nature 293, 469471.
Hyde, I.H. (1921). A micro-electrode and unicellular stimulation. Biol.
Bull. 40, 130133.
Jaffe, D.B., Johnston, D., Lasser-Ross, N., Lisman, J.E., Miyakawa,
H., and Ross, W.N. (1992). The spread of Na spikes determines the
pattern of dendritic Ca entry into hippocampal neurons. Nature 357,
244246.
Katz, L.C., and Constantine-Paton, M. (1988). Relationships between
segregated afferents and postsynaptic neurons in the optic tectum
of three-eyed frogs. J. Neurosci. 8, 31603180.
Kim, H.G., and Connors, B.W. (1993). Apical dendrites of the neocor-
tex: correlation between sodium- and calcium-dependent spiking
and pyramidal cell morphology. J. Neurosci. 13, 53015311.
Kim, H.G., Beierlein, M., and Connors, B.W. (1995). Inhibitory control
of excitable dendrites in neocortex. J. Neurophysiol. 74, 18101815.
Kisvarday, Z.F., Martin, K.A.C., Friedlander, M.J., and Somogyi, P.
(1987). Evidence for interlaminar inhibitory circuits in the striate cor-
tex of the cat. J. Comp. Neurol. 260, 119.
Koch, C., and Zador, A. (1993). The function of dendritic spines:
devices subserving biochemical rather than electrical compartmen-
talization. J. Neurosci. 13, 413422.
Koch, C., Poggio, T., and Torre, V. (1983). Nonlinear interactions in
a dendritic tree: localization, timing, and role in information pro-
cessing. Proc. Natl. Acad. Sci. USA 80, 27992802.
Kogan, A., Ross, W.N., Zecevic, D., and Lasser-Ross, N. (1996).
Optical recording from cerebellar Purkinje cells using intracellularly
injected voltage-sensitive dyes. Brain Res., in press.
Kolliker A. (1889). Handbuch der Gewebelehre (Leipzig).
Kosaka, T. (1983). Neuronal gap junctions in the polymorph layer of
the rat dentate gyrus. Brain Res. 277, 347351.
Larkman, A.U. (1991). Dendritic morphology of pyramidal neurons
of the visual cortex of the rat. III. Spine distributions. J. Comp.
Neurol. 306, 332343.
Laurent, G., and Burrows, M. (1989). Intersegmental interneurons
can control the gain of reflexes in adjacent segments of the locust
by their action on nonspiking local interneurons. J. Neurosci. 9,
30303039.
Levy, W.B., and Desmond, N.L. (1985). Associative potentiation/
depression in the hippocampal dentate gyrus. In Electrical Activity
of the Archicortex, G. Buzsaki and C.H. Vanderwolf, eds. (Budapest:
Akademiai Kiado), pp. 359373.
Levy, W.B., and Steward, O. (1979). Synapses as associative mem-
ory elements in the hippocampal formation. Brain Res. 175, 233245.
Ling, G., and Gerard, R.W. (1949). The normal membrane potential
of frog sartorius fibers. J. Cell Comp. Physiol. 34, 383396.
Lisman, J. (1989). A mechanism for the Hebb and anti-Hebb pro-
cesses underlying learning and memory. Proc. Natl. Acad. Sci. USA
86, 95749578.
Llinas, R. (1988). The intrinsic electrophysiological properties of
mammalian neurons: insights into central nervous system function.
Science 242, 16541664.
Llinas, R., and Greenfield, S.A. (1987). On-line visualization of den-
dritic release of acetylcholinesterase from mammalian substantia
nigra neurons. Proc. Natl. Acad. Sci. USA 84, 30473050.
Llinas, R., and Hess, R. (1976). Tetrodotoxin-resistant dendritic
spikes in avian Purkinje cells. Proc. Natl. Acad. Sci. USA 73, 2520
2523.
Llinas, R., and Nicholson, C. (1971). Electroresponsive properties of
dendrites and somata in alligator Purkinje cells. J. Neurophysiol. 34,
532551.
Llinas, R., and Sugimori, M. (1980). Electrophysiological properties
of in vitro Purkinje cell dendrites in mammalian cerebellar slices. J.
Physiol. 305, 197213.
Llinas, R., and Sugimori, M. (1992). The electrophysiology of the
cerebellar Purkinje cells revisited. In The Cerebellum Revisited, R.
Llinas and C. Sotelo, eds. (New York: Springer-Verlag), pp. 167181.
Llinas, R., and Yarom, Y. (1981). Properties and distribution of ionic
conductances generating electroresponsiveness of mammalian in-
ferior olivary neurones in vitro. J. Physiol. (Lond.) 315, 569584.
Llinas, R., Nicholson, C., Freeman, J.A., and Hillman, D.E. (1968).
Dendritic spikes and their inhibition in alligator Purkinje cells. Sci-
ence 160, 11321135.
Llinas, R., Baker, R., and Sotelo, C. (1974). Electrotonic coupling
between neurons in cat inferior olive. J. Neurophysiol. 37, 560571.
Lorente de No, R., and Coundouris, G.A. (1959). Decremental con-
duction in peripheral nerve integration of stimuli in the neuron. Proc.
Natl. Acad. Sci. USA 45, 592617.
Lo Turco, J.J., and Kriegstein, A.R. (1991). Clusters of coupled neu-
roblasts in embryonic neocortex. Science 252, 563566.
MacVicar, B.A., and Dudek, F.E. (1981). Electrotonic coupling be-
tween pyramidal cells: a direct demonstration in rat hippocampal
slices. Science 213, 782785.
Magee, J.C., and Johnston, D. (1995a). Characterization of single
voltage-gated Na1 and Ca21 channels in apical dendrites of rat CA1
pyramidal neurons. J. Physiol. 487, 6790.
Magee, J.C., and Johnston, D. (1995b). Synaptic activation of volt-
age-gated channels in the dendrites of hippocampal pyramidal neu-
rons. Science 268, 301304.
Mainen, Z.F., and Sejnowski, T.J. (1995). Reliability of spike timing
in neocortical neurons. Science 268, 15031506.
Major, G. (1993). Solutions for transients in arbitrarily branching
cables. III. Voltage clamp problems. Biophys. J. 65, 469491.
Major, G., Evans, J.D., and Jack, J.B. (1993). Solutions for transients
in arbitrary branching cables: I. Voltage recording with a somatic
shunt. Biophys. J. 65, 423449.
Major, G., Larkman, A.U., Jonas, P., Sakmann, B., and Jack, J.J.
(1994). Detailed passive cable models of whole-cell recorded CA3
pyramidal neurons in rat hippocampal slices. J. Neurosci. 14, 4613
4638.
Markram, H., and Sakmann, B. (1994). Calcium transients in apical
dendrites evoked by single sub-threshold excitatory post-synaptic
potentials via low voltage-activated calcium channels. Proc. Natl.
Acad. Sci. USA 91, 52075211.
Matsumura, M., Cope, T., and Fetz, E.E. (1988). Sustained excitatory
synaptic input to motor cortex neurons in awake animals revealed
by intracellular recording of membrane potentials. Exp. Brain Res.
70, 463469.
Mel, B.W. (1994). Information processing in dendritic trees. Neural
Comp. 6, 10311085.
Michelson, H.B., and Wong, R.K.S. (1994). Synchronization of inhibi-
tory neurons in the guinea pig hippocampus in vitro. J. Physiol. 477,
3545.
Minski, M. (1961). Microscopy apparatus. US Patent 1961, 3013467.
Montana, V., Farkas, D.L., and Loew, L.M. (1989). Dual-wavelength
ratiometric fluorescence measurements of membrane potential.
Biochemistry 28, 45364539.
Muller, W., and Connor, J.A. (1991). Dendritic spines as individual
96
Review
715
neuronal compartments for synaptic Ca21 responses. Nature 354,
7376.
Murphy, T.H., Baraban, J.M., Gil Wier, W., and Blatter, L.A. (1994).
Visualization of quantal synaptic transmission by dendritic calcium
imaging. Nature 263, 529532.
Nunez, A., Garca-Austt, E., and Buno, W. (1990). In vivo electrophys-
iological analysis of Lucifer Yellowcoupled hippocampal pyramids.
Exp. Neurol. 108, 7682.
Nunez, A., Amzica, F., and Steriade, M. (1993). Electrophysiology of
cat association cortical cells in vivo: intrinsic properties and synaptic
responses. J. Neurophys. 70, 418430.
Palay, S.L. (1956). Synapses in the central nervous system. J. Bio-
phys. Biochem. Cytol. 2, 193201.
Peinado, A., Yuste, R., and Katz, L.C. (1993). Extensive dye coupling
between rat neocortical neurons during the period of circuit forma-
tion. Neuron 10, 103114.
Peters, A., Palay, S.L., and Webster, H.D. (1991). The Fine Structure
of the Nervous System (New York: Oxford University Press).
Phillips, C.G., Powel, T.P.S., and Shepherd, G.M. (1963). Responses
of mitral cells to stimulation of the lateral olfactory tract in the rabbit.
J. Physiol. 168, 6588.
Pinsky, P.F., and Rinzel, J. (1994). Intrinsic and network rhythmogen-
esis in a reduced Traub model for CA3 neurons. J. Comput. Neurosci.
1, 3960.
Pockberger, H. (1991). Electrophysiological and morphological
properties of rat motor cortex neurons in vivo. Brain Res. 539,
181190.
Pollack, G.S. (1988). Selective attention in an insect auditory neuron.
J. Neurosci. 8, 26352639.
Powell, T.P.S. (1981). Certain aspects of the intrinsic organization
of the cerebral cortex. In Brain Mechanisms and Perceptual Aware-
ness, O. Pompeiano and C. Ajmone Marsan, eds. (New York: Raven
Press), pp. 119.
Purpura, D.P. (1967). Comparative physiology of dendrites. In The
Neurosciences: A Study Program, G.C. Quarton, T. Melnechuk, and
F.O. Schmitt, eds. (New York: Rockefeller University Press), pp.
372875.
Purpura, D.P., and Shofer, R.J. (1964). Cortical intracellular poten-
tials during augmenting and recruiting responses. I. Effects on in-
jected hyperpolarizing currents on evoked membrane potential
changes. J. Neurophysiol. 27, 117132.
Purves, D., and Hume, R.I. (1981). The relation of postsynaptic geom-
etry to the number of presynaptic axons that innervate autonomic
ganglion cells. J. Neurosci. 1, 441452.
Rall, W. (1959). Branching dendritic trees and motoneuron mem-
brane resistivity. Exp. Neurol. 1, 491527.
Rall, W. (1964). Theoretical significance of dendritic trees for neu-
ronal input-output relations. In Neural Theory and Modeling, R.F.
Reiss, ed. (Stanford: Stanford University Press), pp. 7397.
Rall, W. (1974). Dendritic spines, synaptic potency and neuronal
plasticity. In Cellular Mechanisms Subserving Changes in Neuronal
Activity, C.D. Woody, K.A. Brown, T.J. Crow, and J.D. Knispel, eds.
(Los Angeles: Brain Information Services), pp. 1321.
Rall, W. (1995). The Theoretical Foundation of Dendritic Function,
I. Segev, J. Rinzel, and G.M. Shepherd, eds. (Cambridge, Massachu-
setts: MIT Press).
Rall, W., and Segev, I. (1987). Functional possibilities for synapses
on dendrites and on dendritic spines. In Synaptic Function, G.E.
Edelman, W.F. Gall, and W.M. Cowan, eds. (New York: Wiley), pp.
605637.
Rall, W., and Shepherd, G.M. (1968). Theoretical reconstruction of
field potentials and dendro-dendritic synaptic interactions in olfac-
tory bulb. J. Neurophys. 31, 884915.
Rall, W., Shepherd, G.M., Reese, T.S., and Brightman, M.W. (1966).
Dendrodenritic synaptic pathways for inhibition in the olfactory bulb.
Exp. Neurol. 14, 4456.
Raman, I.M., and Trussell, L.O. (1992). The kinetics of the response
to glutamate and kainate in neurons of the avian cochlear nucleus.
Neuron 9, 173186.
Ramon, P. (1890). Investigaciones de histologa comparada en los
centros opticos de los vertebrados. PhD thesis, University of Madrid.
Ramon y Cajal, S. (1889). Conexion general de los elementos ner-
viosos. Medicina Practica.
Ramon y Cajal, S. (1891a). Significacion fisiologica de las expansi-
ones protoplasmicas y nerviosas de la sustancia gris. Rev. Ciencias
Med. 22, 23.
Ramon y Cajal, S. (1891b). Sur la structure de lecorce cerebrale de
quelques mamiferes. Cellule 7.
Ramon y Cajal, S. (1897). Las leyes de la morfologa y dinamismo
de las celulas nerviosas. Revista Trim. Microgr. 1.
Ramon y Cajal, S. (1904). La Textura del Sistema Nerviosa del Hom-
bre y los Vertebrados (Madrid: Moya).
Ramon y Cajal, S. (1923). Recuerdos de mi Vida: Historia de mi
Labor Cientfica (Madrid: Alianza Editorial).
Rapp, M., Segev, I., and Yarom, Y. (1994). Physiology, morphology
and detailed passive models of guinea-pig Purkinje cells. J. Physiol.
(Lond.) 474, 101118.
Regehr, W.G., and Tank, D.W. (1992). Calcium concentration dynam-
ics produced by synaptic activation of CA1 hippocampal pyramidal
cells. J. Neurosci. 12, 42024223.
Regehr, W.G., Connor, J.A., and Tank, D.W. (1989). Optical imaging
of calcium accumulation in hippocampal pyramidal cells during syn-
aptic activation. Nature 341, 533536.
Regehr, W.G., Konnerth, A., and Armstrong, C.M. (1992). Sodium
action potentials in the dendrites of cerebellar Purkinje cells. Proc.
Natl. Acad. Sci. USA 89, 54925496.
Regehr, W.G., Kehoe, J., Ascher, P., and Armstrong, C. (1993). Syn-
aptically triggered action potentials in dendrites. Neuron 11,
145151.
Renshaw, B. (1942). Effects of presynaptic volleys on spread of
impulses over the soma of the motoneurons. J. Neurophysiol. 5,
235243.
Reuveni, I., Friedman, A., Amitai, Y., and Gutnick, M.J. (1993). Step-
wise repolarization from Ca21 plateaus in neocortical pyramidal
cells: evidence for nonhomogeneous distribution of HVA Ca21 chan-
nels in dendrites. J. Neurosci. 13, 46094621.
Reyher, C.H.K., Lubke, J., Larsen, W.J., Hendrix, G.M., Shipley, M.T.,
and Baumgarten, H.G. (1991). Olfactory bulb granule cell aggre-
gates: morphological evidence for interperikaryal electrotronic cou-
pling via gap junctions. J. Neurosci. 11, 14851495.
Rhodes, P.A., and Gray, C.M. (1994). Simulations of intrinsically
bursting neocortical pyramidal neurons. Neural Comp. 6, 10861110.
Richardson, T.L., Turner, R.W., and Miller, J.J. (1987). Action-poten-
tial discharge in hippocampal CA1 pyramidal neurons. J. Neurophys.
58, 981996.
Rockland, K.S., and Virga, A. (1989). Terminal arbors of individual
feedback axons projecting from V2 to V1 in the macaque monkey:
a study using immunocytochemistry of anterogradely transported
Phaseolus vulgaris leucoagglutinin. J. Comp. Neurol. 285, 5472.
Rose, G.J., and Call, S.J. (1993). Temporal filtering properties of
midbrain neurons in an electric fish: implications for the function of
dendritic spines. J. Neurosci. 13, 11781189.
Ross, W.N., and Werman, R. (1987). Mapping calcium transients in
the dendrites of Purkinje cells from guinea-pig cerebellum in vitro.
J. Physiol. 389, 319336.
Ross, W.N., Lewenstein Stockbridge, L., and Stockbridge, N.L.
(1986). Regional properties of calcium entry in barnacle neurons
determined with Arsenazo III and a photodiode array. J. Neurosci.
6, 11481159.
Rubel, E.W., and Parks, T.N. (1988). Organization and development
of the avian brain-stem auditory system. In Auditory Function, G.M.
Edelman, W.E. Gall, and W.M. Cowan, eds. (New York: John Wiley
and Sons), pp. 393.
Sakmann, B., and Neher, E. (1983). Single Channel Recording (New
York: Plenum).
97
Neuron
716
Schadlen, M.N., and Newsome, W.T. (1994). Noise, neural codes
and cortical organization. Curr. Opin. Neurobiol. 4, 569579.
Scharfman, H.E., and Sarvey, J.M. (1985). Postsynaptic firing during
repetitive stimulation is required for long-term potentiation in hippo-
campus. Brain Res. 331, 267274.
Schwartzkroin, P.A., and Slawsky, M. (1977). Probable calcium
spikes in hippocampus neurons. Brain Res. 135, 157161.
Shelton, D.P. (1985). Membrane resistivity estimated for the Purkinje
neuron by means of a passive computer model. Neuroscience 14,
111131.
Shepherd, G.M., and Brayton, R.K. (1987). Logic operations are
properties of computer-simulated interactions between excitable
dendritic spines. Neuroscience 21, 151165.
Shepherd, G.M., Brayton, R.K., Miller, J.P., Segev, I., Rinzel, J., and
Rall, W. (1985). Signal enhancement in distal cortical dendrites by
means of interactions between active dendritic spines. Proc. Natl.
Acad. Sci. USA 82, 21922195.
Sloper, J.J. (1972). Gap junctions between dendrites in the primate
neocortex. Brain Res. 44, 641646.
Sobel, E.C., and Tank, D. W. (1994). In vivo Ca21 dynamics in a cricket
auditory neuron: an example of chemical computation. Science 263,
823826.
Softky, W.R., and Koch, C. (1992). The highly irregular firing of corti-
cal cells is inconsistent with temporal integration of random EPSPs.
J. Neurosci. 13, 334350.
Sotelo, C., Llinas, R., and Baker, R. (1974). Structural study of inferior
olivary nucleus of the cat: morphological correlates of electrotonic
coupling. J. Neurophysiol. 37, 541559.
Spencer, W.A., and Kandel, E.R. (1961). Electrophysiology of hippo-
campal neurons: IV. Fast prepotentials. J. Neurophysiol. 24,
272285.
Spruston, N., Jaffe, D.B., and Johnston, D. (1994). Dendritic attenua-
tion of synaptic potentials and currents: the role of passive mem-
brane properties. Trends Neurosci. 17, 161166.
Spruston, N., Schiller, Y., Stuart, G., and Sakmann, B. (1995). Activ-
ity-dependent action potential invasion and calcium influx into hip-
pocampal CA1 dendrites. Science 286, 297300.
Stafstrom, C.E., Schwindt, P.C., Chabb, M.C., and Crill, W.E. (1985).
Properties of persistent sodium and calcium conductances of layer
5 neurons from cat sensory motor cortex in vivo. J. Neurophysiol.
55, 153170.
Stratford, K., Mason, A., Larkman, A., Major, G., and Jack, J.J. (1989).
The modelling of pyramidal neurons in the visual cortex. In The
Computing Neuron, R. Durbin, C. Miall, and G. Mitchinson, eds.
(Workingham, England: Addison-Wesley), pp. 296322.
Stuart, G.J., and Hausser, M. (1994). Initiation and spread of sodium
action potentials in cerebellar Purkinje cells. Neuron 13, 703712.
Stuart, G.J., and Sakmann, B. (1994). Active propagation of somatic
action potentials into neocortical pyramidal cell dendrites. Nature
367, 6972.
Stuart, G.J., Dodt, H.-U., and Sakmann, B. (1993). Patch clamp re-
cording from the soma and dendrites of neurons in brain slices using
infrared video microscopy. Pflugers Arch. 423, 511518.
Sugihara, I., Lang, E.J., and Llinas, R. (1993). Uniform olivocerebellar
conduction time underlies Purkinje cell complex spike synchronicity
in the rat cerebellum. J Physiol. (Lond.) 470, 243271.
Sutor, B., and Hablitz, J.J. (1989). EPSPs in rat neocortical neurons
in vitro. II. Involvement of N-methyl-D-aspartate receptors in the
generation of EPSPs. J. Neurophysiol. 61, 621634.
Tank, D.W., Sugimori, M., Connor, J.A., and Llinas, R.R. (1988). Spa-
tially resolved calcium dynamics of mammalian Purkinje cells in
cerebellar slice. Science 242, 773777.
Thompson, A.M., Girdlestone, D., and West, D.C. (1988). Voltage-
dependent currents prolong single-axon postsynaptic potentials in
layer III pyramidal neurons in rat neocortical slices. J. Neurophysiol.
60, 18961907.
Traub, R.D. (1996). Model of sustained population bursts in electri-
cally coupled inerneurons containing active dendritic conductances.
J. Comput. Neurosci., in press.
Traub, R.D., Wong, R.K., Miles, R., and Michelson, H. (1991). A model
of a CA3 hippocampal pyramidal neuron incorporating voltage-
clamp data on intrinsic conductances. J. Neurophysiol. 66, 635650.
Tsien, R.Y. (1989). Fluorescent probes of cell signaling. Annu. Rev.
Neurosci. 12, 227253.
Turner, R.W., Meyers, E.R., Richardson, D.L., and Barker, J.L. (1991).
The site for initiation of action potential discharge over the somato-
sensory axis of rat hippocampal CA1 pyramidal neurons. J. Neu-
rosci. 11, 22702280.
Vaney, D.I. (1991). Many diverse types of retinal neurons show tracer
coupling when injected with biocytin or neurobiotin. Neurosci. Lett.
125, 187190.
Van Gehuchten, A. (1891). La moelle epiniere et le cervelet. Cellule
7.
Watson, A.H.D., and Burrows, M. (1988). Distribution and morphol-
ogy of synapses on nonspiking local interneurons in the thoracic
nervous system of the locust. J. Comp. Neurol. 272, 605616.
Westenbroek, R.E., Ahlijanian, M.K., and Catterall, W.A. (1990). Clus-
tering of L-type Ca2 1 channnels at the base of major dendrites in
hippocampal pyramidal neurons. Nature 347, 281284.
Westenbroek, R.E., Hell, J.W., Warner, C., Dubel, S.J., Snutch, T.P.,
and Catterall, W.A. (1992). Biochemical properties and subcellular
distribution of an N-type calcium channel a1 subunit. Neuron 9,
10991115.
Wickens, J. (1988). Electrically coupled but chemically isolated syn-
apses: dendritic spines and calcium in a rule for synaptic modifica-
tion. Prog. Neurobiol. 31, 507528.
Wilson, T., and Sheppard, C. (1984). Theory and Practice of Scanning
Optical Microscopy (New York: Academic Press).
Wong, R.K.S., and Prince, D.A. (1978). Participation of calcium
spikes during intrinsic burst firing in hippocampal neurons. Brain
Res. 159, 385390.
Wong, R.K.S., and Steward, M. (1992). Different firing patterns gen-
erated in dendrites and somata of CA1 pyramidal neurones in
guinea-pig hippocampus. J. Physiol. 457, 675687.
Wong, R.K.S., Prince, D.A., and Basbaum, A.I. (1979). Intradendritic
recordings from hippocampal neurons. Proc. Natl. Acad. Sci. USA
76, 986990.
Wu, K., Sachs, L., Carlin, R.K., and Siekevitz P. (1986). Characteris-
tics of a Ca21/calmodulin-dependent binding of the Ca2 1 channel
antagonists nitrendipine, to a postsynaptic density fraction isolated
from canine cerebral cortex. Mol. Brain Res. 1, 176184.
Yamamoto, C. (1972). Activation of hippocampal neurons by mossy
fiber stimulation in thin brain sections in vitro. Exp. Brain Res. 14,
423435.
Yang, G., and Masland, R.H. (1992). Direct visualization of the den-
dritic and receptive fields of directionally selective retinal ganglion
cells. Science 258, 19491952.
Yuste, R., and Denk, W. (1995). Dendritic spines as basic units of
synaptic integration. Nature 375, 682684.
Yuste, R., Peinado, A., and Katz, L.C. (1992). Neuronal domains in
developing neocortex. Science 257, 665669.
Yuste, R., Gutnick, M.J., Saar, D., Delaney, K.D., and Tank, D.W.
(1994). Calcium accumulations in dendrites from neocortical neu-
rons: an apical band and evidence for functional compartments.
Neuron 13, 2343.
Yuste, R., Nelson, D., Rubin, W., and Katz, L.C. (1995). Neuronal
domains in developing neocortex: mechanisms of coactivation.
Neuron 14, 717.
Zeki, S. (1993). The modularity of the brain. In A Vision of the Brain,
S. Zeki, ed. (Oxford: Blackwell), pp. 197296.
Zohary, E., Shadlen, M.N., and Newsome, W. T. (1994). Correlated
neuronal discharge rate and its implications for psychophysical per-
formance. Nature 370, 140143.
98

References and Notes
1. G. Kempermann, H. van Praag, F. H. Gage, Prog. Brain
Res. 127, 35 (2000).
2. G. L. Ming, H. Song, Annu. Rev. Neurosci. 28, 223
(2005).
3. T. M. Madsen et al., Biol. Psychiatry 47, 1043 (2000).
4. R. Jaenisch, A. Bird, Nat. Genet. 33 (suppl.), 245 (2003).
5. Materials and methods and supporting data are available
on Science Online.
6. J. E. Ploski, S. S. Newton, R. S. Duman, J. Neurochem. 99,
1122 (2006).
7. H. J. Jung et al., Oncogene 26, 7517 (2007).
8. H. Tran et al., Science 296, 530 (2002).
9. M. C. Hollander, A. J. Fornace Jr., Oncogene 21, 6228
(2002).
10. B. Lu, A. F. Ferrandino, R. A. Flavell, Nat. Immunol. 5, 38
(2004).
11. G. Barreto et al., Nature 445, 671 (2007).
12. V. Ramirez-Amaya et al., J. Neurosci. 25, 1761 (2005).
13. S. W. Flavell, M. E. Greenberg, Annu. Rev. Neurosci. 31,
563 (2008).
14. S. Ge et al., Nature 439, 589 (2006).
15. S. G. Jin, C. Guo, G. P. Pfeifer, PLoS Genet. 4, e1000013
(2008).
16. T. Aid, A. Kazantseva, M. Piirsoo, K. Palm, T. Timmusk,
J. Neurosci. Res. 85, 525 (2007).
17. K. Y. Alam et al., J. Biol. Chem. 271, 30263 (1996).
18. Y. Zhang, F. Madiai, K. V. Hackshaw, Biochim. Biophys.
Acta 1521, 45 (2001).
19. N. Cervoni, M. Szyf, J. Biol. Chem. 276, 40778 (2001).
20. M. Gehring et al., Cell 124, 495 (2006).
21. R. Mtivier et al., Nature 452, 45 (2008).
22. R. Holliday, J. Theor. Biol. 200, 339 (1999).
23. K. Martinowich et al., Science 302, 890 (2003).
24. I. C. Weaver et al., Nat. Neurosci. 7, 847 (2004).
25. F. D. Lubin, T. L. Roth, J. D. Sweatt, J. Neurosci. 28,
10576 (2008).
26. K. Garbett et al., Neurobiol. Dis. 30, 303 (2008).
27. V. M. Porterfield, H. Piontkivska, E. M. Mintz, BMC
Neurosci. 8, 98 (2007).
28. D. Hevroni et al., J. Mol. Neurosci. 10, 75 (1998).
29. M. Majdan, C. J. Shatz, Nat. Neurosci. 9, 650 (2006).
REPORTS
30. We thank D. Ginty, S. Synder, and members of Ming and
Song laboratories for help and critical comments and
L. Liu and Y. Cai for technical support. This work was
supported by NIH, McKnight, and NARSAD (to H.S.) and
by NIH, March of Dimes, and Johns Hopkins Brain
Science Institute (to G-l.M.). R.A.F. is an investigator with
the Howard Hughes Medical Institute.
Supporting Online Material
www.sciencemag.org/cgi/content/full/1166859/DC1
Materials and Methods
SOM Text
Figs. S1 to S19
Tables S1 and S2
References
6 October 2008; accepted 16 December 2008
Published online 1 January 2009;
10.1126/science.1166859
Include this information when citing this paper.

Downloaded from www.sciencemag.org on February 19, 2009

Harmonic Convergence in the Love attracted by the sounds of their singing hosts (20).
The auditory physiology on A. aegypti provides
direct evidence that females can hear and puts to
Songs of the Dengue Vector Mosquito rest textbook wisdom that females are deaf (8, 9).
For behavioral experiments, we tethered each
Lauren J. Cator,1* Ben J. Arthur,2* Laura C. Harrington,1 Ronald R. Hoy2 mosquito to the end of an insect pin. When sus-
pended in midair, flies initiated bouts of wing-
The familiar buzz of flying mosquitoes is an important mating signal, with the fundamental flapping flight. We recorded flight tones with a
frequency of the females flight tone signaling her presence. In the yellow fever and dengue vector particle velocity microphone. Acoustic interaction
Aedes aegypti, both sexes interact acoustically by shifting their flight tones to match, resulting in a was demonstrated by moving a tethered flying
courtship duet. Matching is made not at the fundamental frequency of 400 hertz (female) or mosquito past a stationary tethered flying partner
600 hertz (male) but at a shared harmonic of 1200 hertz, which exceeds the previously known (movie S1 with audio). Females were brought in
upper limit of hearing in mosquitoes. Physiological recordings from Johnstons organ (the and out of the male hearing range (2 cm) for 10-s
mosquitos ear) reveal sensitivity up to 2000 hertz, consistent with our observed courtship fly-bys. Recordings revealed acoustic interaction:
behavior. These findings revise widely accepted limits of acoustic behavior in mosquitoes. In 14 of 21 (67%) pairs, both sexes altered their
flight tones so that the males second harmonic
[fundamental (F0) = 636.7 T 15.1, second har-

M
osquito-borne diseases such as ma-
laria, yellow fever, and dengue con-
tinue to afflict millions, even after
decades of work to control vector populations.
Despite this effort, basic aspects of mosquito bi-
ology are not fully understood, including mating
behavior, an important target for vector control.
We describe investigations in Aedes aegypti that
require revision of the current understanding of
mosquito mating behavior. Since Johnston (1)
first suggested in 1855 that mosquitoes could
perceive sound, over 14 studies have been pub-
lished on sound production and hearing in A.
aegypti (217) (table S1). The buzz of a flying
female mosquito acts as a mating signal, attracting
males. Typically, the behaviorally salient frequen-
cy component of flight tone is the fundamental
frequency of wing beat, which is between 300 to
600 Hz depending on species (8). However, mate
attraction is not simply a matter of a male pas-
sively hearing and homing in on a 400-Hz tone.
For example, males and females of the nonblood-
feeding mosquito Toxorhynchites brevipalpis mod-
1
Department of Entomology, Cornell University, Ithaca, NY
14853, USA. 2Department of Neurobiology and Behavior,
Cornell University, Ithaca, NY 14853, USA.
*These authors contributed equally to this work.
To whom correspondence should be addressed. E-mail:
rrh3@cornell.edu
ulate their 300- to 500-Hz wing beat frequencies
to match each other (18). Thus, acoustically me-
diated mate attraction involves active modulation
by both sexes, creating a duet.
We show that males and females of the den-
gue and yellow fever vector A. aegypti also mod-
ulate their flight tones when brought within a few
centimeters of each other. This modulation, how-
ever, does not match the fundamental wing beat
frequency of around 400 Hz (female) or around
600 Hz (male) but a shared harmonic frequency
of around 1200 Hz (Fig. 1). Consistent with this,
a neurophysiological examination of the ears of
A. aegypti shows response in both males and fe-
males up to 2000 Hz (Fig. 2). These results are
unexpected because over 5 decades of behavioral
and physiological studies had concluded that male
mosquito ears (antennae and associated Johnstons
organ) are tuned to 300 to 800 Hz and deaf to
frequencies above 800 Hz (8, 19). The present
study also directly addresses the issue of audi-
tory competence in female mosquitoes. Acoustic
duetting behavior in the nonvector mosquito
T. brevipalpis (18) would seem to imply active
audition in both sexes, and laser vibrometry
studies of the Johnstons organ in that species
and A. aegypti (16, 17) indicate that they re-
spond mechanically to salient sounds. Moreover,
female frog-biting mosquitoes are reported to be
monic (F1) = 1238.3 T 31.0 (SEM) Hz] matched
the females third harmonic [F0 = 430.6 T 10.8,
F2 = 1356.2 T 29.2 (SEM) Hz] (Fig. 1, A to C).
The period of synchronization lasted an average
of 9.71 T 1.05 (SEM) s with the synchroniza-
tion frequency averaging 1354.5 T 31.5 (SEM)
Hz. A. aegypti do not shift their flight tones in
the absence of acoustic stimulation, as tested both
by deafening the mosquitoes [and stimulating
with tones, Fisher exact test, males P = 0.02,
females P = 0.04 (21)] and flying intact control
subjects in silence [Fishers exact test, males P =
0.02, females P = 0.04 (21)].
The presence of the fundamental frequency
tone was not necessary for harmonic matching.
We stimulated tethered mosquitoes with electron-
ically generated pure sinusoidal tones as well as
with harmonic combinations of pure tones lack-
ing the fundamental frequency (Fig. 1, D to F).
The intensity of the pure tones was set at a particle
velocity of 0.024 mm/s corresponding to 54 dB
sound pressure level (relative to 20 mPa) when
played through an ear bud speaker positioned 1.5
cm in front of the test mosquito. This intensity
is well within the response range of A. aegyptis
Johnstons organ, as measured by Doppler
vibrometry (17). Stimuli were played in 10- to 15-s
bursts with 5- to 20-s recovery periods. Play-
back experiments with pure-tone combinations

www.sciencemag.org SCIENCE VOL 323 20 FEBRUARY 2009 1077

99
 REPORTS
harmonic. In contrast, only 2 of 18 (11%) pre-
A viously mated females (confirmed by the pres-
ence of sperm in the females sperm storage organs)
performed a frequency match to playback of
B complete male songs, suggesting that mating
2500 decreases sensitivity to male stimuli.
Physiological data from the mosquitos audi-
tory organ were obtained by impaling Johnstons
2000
organ with tungsten electrodes and recording
frequency (Hz)

acoustically evoked field potentials. The Johnstons

1500 organ of both males and females responded to 0.5-s
cosine-enveloped pure-tone pulses at all frequen-
cies tested (125 to 2000 Hz), including the shared
1000 harmonic of their flight tones at 1200 Hz (Fig. 2).
The response consists of a sustained voltage de-
500 flection, which is negative with respect to the tho-
racic ground electrode, and a concurrent periodic
oscillation at the stimulation frequency and its har-

Downloaded from www.sciencemag.org on February 19, 2009

10 20 30 40 50 60 70 80 90 100 monics, a form of neural encoding that bears re-
time (s) semblance to that seen in the mammalian cochlea
(22). It is the amplitude of the sustained deflection
1200 paired flight that remains significantly higher than prestimulus
C background noise and thoracic control recordings
up to 2000 Hz (t test, P < 0.001 in both males and
1100
females); the amplitude of the periodic oscillation
35 40 45 50 55 is about an order of magnitude smaller and remains
higher than background and controls only up to
1300 800 & 1200 Hz 1000 Hz, in the case of the stimuluss fundamental,
D and 700 Hz, for its second harmonic (t test, P <
0.01). These recordings confirm earlier studies that
1200 the mosquito Johnstons organ is sensitive to 100-
10 20 30 40 50 60 70 80 90 100 110 to 500-Hz tones and extend the upper limit of
hearing to at least 2000 Hz. Earlier physiological
1400 studies of Johnstons organ that failed to report
a high-frequency response are likely due to
E 1300
1400 Hz only filter bandwidths set at the time of recording. We
1200 recorded high-frequency responses only when
10 20 30 40 50 60 70 80 we set the high-pass filter to 1 Hz or less, not
the customary 100 Hz or higher used in extra-
cellular recording.
1400 Tone-matching behavior does not require that
F
1400 Hz only
a tethered, flying mosquito hear a live partner.
1300 Either a male or a female A. aegypti can modulate
5 10 15 20 25 30 35 40 45 its flight tone harmonics to match electronically
time (s) generated pure-tone probes (Fig. 1). Frequency
Fig. 1. (A) An oscillogram from a sound clip of a tethered male and female duetting. (B) A match is elicited to a probe that simulates a nat-
spectrogram depicting the harmonic stack of the same sound clip. The male was held in a fixed ural flight tone stripped of its fundamental fre-
position and sang continuously for nearly 2 min. The female was brought within 2 cm of the male quency and even to a probe that contains only a
on three separate occasions. (C) An expanded view of (B) showing the synchronization of the flight single harmonic. Moreover, mosquitoes can mod-
tone at the second harmonic of the male (blue) and the third of the female (red). At t = 45 s, the ulate their flight tone harmonics to match the
two tones converge to the extent that they cannot be readily distinguished. (D) A separate probe tone whether the probe frequency is set
recording of a tethered male (blue) flying solo and synchronizing to a loud speaker stimulus above or below the flys actual flight tone. This
consisting of a simulated female flight tone with a missing fundamental (red). (E) A recording of a directly demonstrates that both sexes of this spe-
tethered male flying solo (blue) modulating his flight tone to match the playback of a simulated cies (and likely other mosquito species) can hear
third harmonic of a female. (F) A female flying solo (red) matches a simulated second harmonic of and respond to high-frequency tones alone. These
a male (blue) playback tone. behavioral findings are supported by sensory phys-
iology (Fig. 2). Extracellular recordings of acous-
demonstrated that 11 of 28 (39%) males could components. We also tested the ability of females tically evoked field potentials from the Johnstons
synchronize their 1200-Hz second harmonic to to synchronize to playback of pure tones organ elicited clear responses not only to frequen-
the simulated females 1200-Hz third harmonic in mimicking male sound. Six of 20 (30%) unmated cies of the fundamental (400 to 600 Hz), as ex-
the absence of the 400-Hz fundamental tone of an females matched the second harmonic of their pected, but into the kilohertz range, where males
actual female. Furthermore, 12 of 54 (22%) males flight tone to a complete stack of tones (F0 to F3, and females perform active acoustic modulation
could match a pure 1200-Hz tone (the third 700 to 2800 Hz). When unmated females were of their flight tone harmonics. Thus, both behav-
harmonic of female flight tone) in the absence stimulated with a pure 1400-Hz tone, 7 of 20 ioral and physiological experiments establish that
of the fundamental and any other harmonic (35%) responded by matching with their third A. aegypti signal to each other by using frequen-

1078 20 FEBRUARY 2009 VOL 323 SCIENCE www.sciencemag.org

100
 REPORTS
A B 3. P. Belton, in Insect Flight: Dispersal and Migration,
1200 Hz W. Danthanarayana, Ed. (Springer-Verlag, Berlin, 1986),
SD F1 pp. 6070.
F0 F2 4. P. Belton, Can. J. Zool. 67, 2625 (1989).
F3 5. P. Belton, J. Am. Mosq. Control Assoc. 10, 297 (1994).
SD F4 6. P. Belton, R. A. Costello, Entomol. Exp. Appl. 26, 105 (1979).
0.2 mV

7. K. S. Boo, A. G. Richards, J. Insect Physiol. 21, 1129 (1975).

8. A. N. Clements, The Biology of Mosquitoes: Sensory
Reception and Behaviour (CABI, Wallingford, UK, 1999),
10 100 1000 p. 740.
9. J. A. Downes, Annu. Rev. Entomol. 14, 271 (1969).
400 Hz response frequency (Hz)
10. P. T. Haskell, in The Physiology of Insecta, M. Rockstein,
1 mV

F0-4 Ed. (Academic Press, New York, ed. 2, 1974), vol. 2,

C
pp. 353410.
1000 11. M. C. Kahn, W. Offenhauser Jr., Am. J. Trop. Med. Hyg.
SD 29, 811 (1949).
12. A. M. Mayer, Am. Nat. 8, 577 (1874).

response amplitude (V)

13. L. M. Roth, Am. Midl. Nat. 40, 265 (1948).
100 14. J. Schwartzkopff, in The Physiology of Insecta,
M. Rockstein, Ed. (Academic Press, New York, ed. 2,
1974), vol. 2, pp. 273352.
0.5 1 1.5 15. M. C. Gpfert, D. Robert, Proc. R. Soc. London Ser. B 268,
10

Downloaded from www.sciencemag.org on February 19, 2009

time (s) 333 (2001).
F1 F0 16. M. C. Gpfert, D. Robert, Proc. R. Soc. London Ser. B
267, 453 (2000).
0 17. M. C. Gpfert, H. Briegel, D. Robert, J. Exp. Biol. 202,
2727 (1999).
100 300 1000 18. G. Gibson, I. Russell, Curr. Biol. 16, 1311 (2006).
stimulus frequency (Hz) 19. G. Wishart, C. F. Riordan, Can. Entomol. 91, 181 (1959).
20. A. Borkent, P. Belton, Can. Entomol. 138, 91 (2006).
Fig. 2. (A) Acoustically evoked field potentials recorded from Johnstons organ exhibit periodic 21. Materials and methods are available as supporting
material on Science Online.
oscillations (inset, F0 to F4) riding on top of a sustained deflection (SD). Shown are averages of 10
22. A. R. Palmer, I. J. Russell, Hear. Res. 24, 1 (1986).
and 5 repetitions to 1200 and 400 Hz, respectively, in a male. Thoracic control recordings are in 23. R. Williams, A. Berger, Mosq. News 40, 597 (1980).
black; the stimulus envelope is at bottom. (B) Spectral analysis of the response to the 400-Hz tone 24. A. Young, A. Downe, Physiol. Entomol. 7, 467 (1982).
in (A) shows substantial power at low frequencies (SD) during the stimulus (blue) compared with a 25. M. J. B. R. Vreysen, A. S. Hendrichs, in Area-Wide Control
prestimulus background period (black), as well as peaks at multiple harmonics (F1 to F4) of the of Insect Pests from Research to Field Implementation,
stimuluss fundamental frequency (F0). (C) Averaged across 12 males (blue) and 15 females (red), M. J. B. R. Vreysen, A. S. Hendrichs, Eds. (Springer,
Dordrecht, Netherlands, 2007), p. 792.
the amplitude (mean T SEM) of the sustained deflection (SD, solid squares) remains higher than 26. E. S. Krafsur, J. Agric. Entomol. 15, 303 (1998).
that of prestimulus background noise (dashed line at zero) up to 2000 Hz, whereas the amplitudes 27. R. C. Bushland, A. W. Lindquist, E. F. Knipling, Science
of F0 and F1 (open and solid circles, respectively) are substantially smaller. 122, 287 (1955).
28. M. Q. Benedict, A. S. Robinson, Trends Parasitol. 19, 349
cies in the kilohertz range and can detect these trolled by the release of sterile males (2527). (2003).
29. T. W. Scott, L. C. Harrington, B. G. J. Knols, W. Takken,
with their auditory organs. Taken together, these Our investigation connects control to signal func- in Transgenesis and Management of Vector-Borne
data call for revision of our understanding of acous- tion. We hypothesize that the ability of males to Disease, S. Aksoy, Ed. (Landes Bioscience, Austin, TX, 2008),
tically mediated mating behavior in mosquitoes modulate their flight tones to females is the result pp. 151168.
and, in particular, removal of the long-accepted of sexual selection. Hence, harmonic convergence 30. We thank B. Wyttenbach for the suggestion to look for
sustained deflections in the neural response and are grateful to
benchmark ceiling for hearing in mosquitoes. could be a measure of a males reproductive fit- L. Ristroph and I. Cohen for high-speed video of males and
Lastly, once mated, females are much less ness and the ability of lab-reared, sterilized or ge- females in free flight. Support was provided (to L.C.H.) from
responsive to male flight tones and less likely netically modified males to modulate their flight the Foundation for the National Institutes of Health through
to perform tone matching. These results are tones could be a useful behavioral bioassay for the Grand Challenges in Global Health Initiative and Hatch
Project NYC-139432.
consistent with the observation that, once mated, the sterilization program (28, 29). At the very
A. aegypti females are not responsive to addi- least, our findings open the door to a new under- Supporting Online Material
tional matings for the duration of one or more standing of their mating behavior, one that stresses www.sciencemag.org/cgi/content/full/1166541/DC1
egg-laying cycles (23, 24). This implies that an acoustic interactivity between the sexes at frequen- Materials and Methods
Fig. S1
initial mating depresses the likelihood of subse- cies thought previously to be beyond their range Table S1
quent matings. Thus, releasing sterile male mos- of hearing. References
quitoes into the wild might adversely affect the Movie S1
reproductive potential of virgin females, provid- References and Notes
1. C. Johnston, Q. J. Microsc. Sci. 3, 97 (1855). 29 September 2008; accepted 22 December 2008
ing a rationale for controlling these disease vec- 2. P. Belton, in Experimental Analysis of Insect Behavior, Published online 8 January 2009;
tors through diminishing mating potential. Other L. B. Browne, Ed. (Springer-Verlag, Berlin, 1974). 10.1126/science.1166541
vector and pest species of flies have been con- pp. 139147. Include this information when citing this paper.

www.sciencemag.org SCIENCE VOL 323 20 FEBRUARY 2009 1079

101


102
J Comp Physiol A (2003) 189: 245256
DOI 10.1007/s00359-003-0406-2

K AR L V ON FR I SCH LE C TU R E

G. Neuweiler

Evolutionary aspects of bat echolocation

Received: 9 September 2002 / Revised: 17 February 2003 / Accepted: 18 February 2003 / Published online: 28 March 2003
Springer-Verlag 2003

Abstract This review is yet another attempt to explain

how echolocation in bats or bat-like mammals came into
Introduction
existence. Attention is focused on neuronal specializa-
Biology has remained an integral and unied science in
tions in the ascending auditory pathway of echolocating
spite of the productive and continuing proliferation of
bats. Three dierent mechanisms are considered that
new disciplines incorporating a multitude of highly di-
may create a specic auditory sensitivity to echos: (1)
verse methods from biochemical protein identication to
time-windows of enhanced echo-processing opened by a
computer models of ecosystems. It owns its identity to the
corollary discharge of neuronal vocalization commands;
Darwinian theory of evolution that has been, and still is,
(2) dierentiation and expansion of ensembles of
veried by countless experimental evidence from the dy-
combination-sensitive neurons in the midbrain; and (3)
namics of complex ecological interactions to the level of
corticofugal top-down modulations. The second part of
molecular interactions in cells and genomes. From petri-
the review interprets three dierent types of echoloca-
ed records we learn how our extant fauna came into
tion as adaptations to ecological niches, and presents the
existence, e.g. how reptilian-like creatures may have
sophisticated cochlear specializations in constant-
learned to y and eventually became birds, and how Homo
frequency/frequency-modulated bats as a case study of
sapiens has evolved from anthropoid primates.
nely tuned dierentiation. It is briey discussed how a
Among mammals bats are unique since they feature
resonant mechanism in the inner ear of constant-
two very special achievements: ight and echolocation.
frequency/frequency-modulated bats may have evolved
Unlike in birds, however, there exists no archechirop-
in common mammalian cochlea.
terix that may show which mammalian precursors de-
veloped the ability to y and to echolocate. The oldest
Abbreviations AVCN anteroventral cochlear
chiropteran skeletons (Icaronycteris index) come from
nucleus CF constant frequency element of an
Eocenic deposits in Wyoming. This record contains a
echolocation signal DPOAE distortion product
perfect bat with only minor dierences to our extant bat
otoacoustic emission FM frequency-modulated
fauna. It appears that as far back as 50 million years ago
echolocation signal IC inferior colliculus IHC inner
bat species ew and echolocated as eciently as our
hair cell INLL intermediate nucleus of the lateral
extant microchiropteran bat fauna of about 650 species
lemniscus MSO medial superior olive OAE
(Neuweiler 2000).
otoacoustic emission OHC outer hair cell PVCN
Paleontological records do not disclose how ight
posterior ventral cochlear nucleus SI sparsely
and echolocation in bats came into existence. Therefore,
innervated section of the cochlea SOC superior olivary
one has to compare extant echolocating bats with non-
complex VNLL ventral nucleus of the lateral lemniscus
echolocating mammals to understand the mechanisms
that enable a brain to interpret the outer world by means
of acoustical reections. Are there any structural and/or
physiological specializations that distinguish an echolo-
cating mammal from a non-echolocating one?
An obvious organ to look at is the ear. However,
G. Neuweiler apart from specic modications in a few highly spe-
Department of Biology II, University of Munich (LMU),
Luisenstrasse 14, 80333 Munchen, Germany cialized species (e.g., auditory foveae) the functional
E-mail: neuweil@zi.biologie.uni-muenchen.de structure of the peripheral auditory system including
Fax: +49-89-5902450 the cochlea and the auditory nerve has been highly

103
246

conserved and is uniform in all mammals including bats
and man. This is not the case for the neuronal aspects of
audition.
Adaptations to echolocation
Neuroanatomical specializations in echolocating bats
In echolocating bats, the lower brainstem nuclei of the
ascending auditory pathway, including the inferior colli-
culus of the midbrain, are hypertrophied and some of
them are dierentially structured (Covey and Casseday
1995). First, the anteroventral cochlear nucleus (AVCN)
is extremely large. In Pteronotus, there is a unique group
of large multipolar cells in a marginal zone. Second, un-
like in other mammals, there is an additional, direct input
from the posterior ventral cochlear nucleus (PVCN) to
the superior olivary complex (SOC). Third, in non-
echolocating mammals, the medial superior olive (MSO)
receives identical inputs from ipsi- and contralateral
AVCN. This is considered a prerequisite for processing
interaural time dierences. In echolocating bats, howev-
er, the ipsilateral input is sparse, and in one bat species,
Pteronotus parnellii, the large majority of MSO-neurons
only receives monaural input from the contralateral
AVCN (Grothe and Neuweiler 2000). Finally, the mon-
aural ventral and intermediate nuclei of the lateral
lemniscus (VNLL and INLL) are unusually large and
conspicuously organized (Fig. 1): the VNLL is dieren-
tiated into a multipolar and a columnar section. Colum-
nar neurons receive tonotopically arranged inputs via
large calyx-synapses on their cell bodies. This unique
columnar organization of neurons with calyx-synapses is
also found in echolocating dolphins. The broadly tuned
neurons of the columnar VNLL process specically fre-
quency-modulated (FM) components of echolocation
signals. Columnar neurons are intensity invariant and
respond with one single spike at constant latencies over a
wide range of intensities and frequencies (Fig. 1). Thus,
they are ideally suited as time markers. Except for the few
cases mentioned above the specic auditory implications
for echolocation of these neuroanatomical specialities are
not yet known.
Echo suppression in non-echolocating mammals
In natural situations a human subject or an animal will
not only listen to a sound source but also to its multiple
reections (clutter) from reverberant surfaces. In most
experimental studies the natural clutter is simplied into
a singular lagging reection following a leading direct
sound (Litovsky et al. 1999). The auditory system en-
sures that reections fuse with the sound source to one
percept. Such perceptual fusion occurs for delays up to

Fig. 1ad Auditory brain stem

nuclei and neurons specialized
for precise time-coding in the
columnar part (Vc) of the
ventral lemniscal nucleus
(VNLL) of the echolocating bat
Eptesicus fuscus. a Schematic
map of auditory nuclei in the
brainstem. CG central gray;
DNLL dorsal, INLL
intermediate, Vc columnar part,
Vm multipolar part of the
VNLL. IC inferiuor colliculus.
b Frequency-modulated (FM)-
stimulated neurons in Vc are
intensity invariant and respond
with one spike at the same
latency. Inset: FM echolocation
sound of E.fuscus. c Latencies
of single unit responses in Vc
show no time-jitter. d Tuning
curves of Vc-units are very
broad and integrate over most
parts of the FM echolocation
signal (After Haplea et al. 1994)

104

5 ms between leading and lagging stimulus. When the
delay increases to 810 ms for brief sounds, e.g. for
clicks, a human listener begins to hear two separate
signals. This critical delay is called echo threshold. Ap-
parently, for brief delays perception of echoes as distinct
signals is suppressed, and the leading signal dominates
the heard sound image. It is hypothesized that a neural
gate closes after the rst elicited spike, and reopens after
a few milliseconds (echo threshold). The gate could be
implemented by self- or lateral inhibition (Litkovsky
et al. 1999). In cats (Litovsky and Yin 1998), rabbits
(Fitzpatrick et al. 1995) and barn owls (Keller and
Tagahashi 1996) a majority of single units from the
inferior colliculus (IC) showed lag suppression within
variable time delays. In these samples the median time
delay for lag suppression matches the behaviourally
measured echo thresholds. However, since in newborn
cats lag suppression is present in the IC, but not
behaviourally, it is also suggested that cortical events
may be involved (Litovsky 1998).
Echo suppression builds up even more when trains of
lead-lag signal pairs are heard, and echo thresholds rise
by several milliseconds. Echo thresholds additionally
increase when each leading signal is succeeded not by
one but by several echoes of various time delays (Yost
and Guzman 1996).
Since echolocating bats listen to trains of lead signals
(vocalized echolocation sound) followed each by several
highly correlated but non-identical lag signals (echoes),
they should be heavily aicted by echo suppression
mechanisms as described in humans and a few experi-
mental mammals. A bat would be neuronally deaf to
echoes returning within about 2.5 ms, i.e. reected from
objects up to at least 40 cm away.
Echo sensitivity locked to sound emission
Obviously, echolocation requires not only an elimina-
tion but a reversal of echo suppression into echo facili-
tation. Do such echo-facilitating neurons exist? Indeed,
Grinnell (1963) discovered echo-facilitation in an evoked
potential study of the IC in the bat Myotis l. lucifugus.
Theoretically, the most ecient way to implement
echo-sensitivity would be the opening of facilitative,
auditory gates by a corollary discharge of the neuronal
command that triggers vocalization of an echolocation
sound. By such a corollary command to the auditory
pathway echo-sensitivity would be specically locked to
an emitted echolocation call, and could be limited for a
dened time window. Schuller (1979) reported units in
the IC of horseshoe bats (Rhinolophus ferrumequinum)
that vigorously responded to phantom echoes only when
the awake bat had vocalized (Fig. 2). In these units,
playbacks of echolocation sounds were ineective
without a preceding sound emission by the bat. The
specic echo-sensitivity triggered by sound emission
vanished for phantom echoes with delays larger than
60 ms (Fig. 2c). Thus, these echo-neurons will respond
247
to echoes returning from objects up to a distance of
10 m, and will exclude echoes from targets further away.
This limitation nicely matches the foraging behaviour
of horseshoe bats which preferably utilize a sit-and-wait
strategy when foraging. From a vantage point within
vegetation horseshoe bats initiate pursuit of ying in-
sects passing by at distances of not more than about 7 m
(Neuweiler et al. 1987).
Unfortunately, recordings in the ascending auditory
pathway of bats that actively emit echolocation signals
have been discontinued. Thus, the few IC units reported
by Schuller (1979) are the only evidence that auditory
units sensitized for echoes by a corollary motor com-
mand may exist. These sparse neural data are, however,
corroborated by comparative behavioural studies
(Roverud and Grinnell 1985; Roverud 1993). Echolo-
cating bats that were trained for a distance-discrimina-
tion task failed when they were jammed by intense
playbacks of the species echolocation sounds. Distance
discrimination could only be disrupted by species-specic
echolocation sounds or signals carrying specic elements
of the echolocation sound. Jamming was time-locked to
the beginning of vocalizations, and lost its eectiveness
beyond echo delays of more than 30 ms. These results
from dierent bat species strongly suggest that a time
window of enhanced echo processing locked to sound
emission indeed exists.
Echo sensitivity by combination sensitive neurons
n contrast to the tiny sample of vocalization-triggered
auditory echo units, another type of echo-sensitive
neuron has been extensively described in large portions
Fig. 2ac Echo-sensitive neuron only triggered by a vocalized
echolocation sound recorded from the inferior colliculus of a
horseshoe bat. a The neuron vigorously responds to a phantom
echo (wavy line) played from a loudspeaker when the bat at the
same time emits an echolocation sound (VOC). b The same neuron
is unresponsive when the bat does not vocalize, and both signals are
played from a loudspeaker. c The specic echo-sensitivity triggered
by sound emission vanishes when the echo is delayed by more than
60 ms (time-window for echoes). After Schuller (1979)
105
248
of the IC and higher auditory centres: so called com-
bination-sensitive neurons. These neurons preferably or
exclusively answer to a combination of two auditory
stimuli (Fig. 3a), e.g. an initial loud FM signal that
corresponds to an emitted echolocation signal, and
after a specic delay a fainter second FM signal that
corresponds to an echo (FM/FM neurons). The so-
called best delay between rst and second signal to
which the neuron responds best is specic for each
neuron and ranges between 0 and 20 ms with a ma-
jority of best delays below 10 ms (Fig. 3b, c; Portfors
and Wenstrup 1999). In terms of a foraging, echolo-
cating bat these data indicate that most of these FM/
Fig. 3ac Combination-sensitivity in the auditory cortex of
horseshoe bats. a A FM1/FM2 neuron that only responds to a
combination of the FM elements of the rst harmonic (FM1 marks
the emitted echolocation sound) and the second harmonic (FM2
marks the returning echo) with a characteristic delay (3 ms). Upper
and lower graph: CF constant-frequency component. The neuron
was stimulated by a CF/FM signal; however, it only responded to
the FM component. b Delay-tuning curves of ve dierent units in
the FM/FM area of the auditory cortex. c Arrangement of best
delays of FM/FM units in a rostro-caudal chronotopic order. Bold
line: units with best delays between 2 and 3 ms, corresponding to
target distances of 3552 cm, are overrepresented. After Schuller
et al. (1991)
FM echo-sensitive neurons should be active when the
bat closes in on a target, for instance when pursuing
prey at distances shorter than about 2 m. Due to its
specic best delays each FM/FM unit will respond only
to an echo returning from the appropriate distance;
hence, these neurons have been also called range-
nding neurons (ONeill and Suga 1982).
According to a recent study in the nuclei of the lateral
lemniscus (Portfors and Wenstrup 2001), combination
sensitivity seems to be generated in the IC where about
75% of units recorded are combination sensitive; of these
70% are facilitated and 30% inhibited by the rst signal.
Echo-sensitive neurons often respond to a rst sig-
nal mimicking specic parameters of an emitted echo-
location sound only after long latencies of ca. 13 ms
compared to ca. 7 ms for the second signal (echo).
These long latencies to the rst stimulus may result
from inhibition initiated by the beginning of the rst
stimulus. These ndings support the hypothesis that
echo-sensitive neurons are coincidence detectors (Olsen
and Suga 1991; Portfors and Wenstrup 2001). Coinci-
dence is most probably achieved by rebound from in-
hibition elicited by the emitted sound. Duration of
inhibition varies and corresponds to the best delay of
the units. Combination sensitivity disappears when in-
hibition is blocked in the inferior colliculus.
106

Combination-sensitive neurons that are inhibited will
be silent during sound emission; however, by rebound
from inhibition they may feed additional excitation into
facilitated FM/FM neurons and thus enhance echo-
sensitivity (Portfors and Wenstrup 2001).
FM/FM or other echo-sensitive neurons belong to
the broad class of combination-sensitive neurons that
have been also recorded in higher auditory centres of
non-echolocating mammals and of birds. Combination-
sensitive neurons respond best to specic temporal re-
lationships between spectrally separate sound elements,
and are considered to be coders of species-specic
communication signals. A neuron is classied as com-
bination sensitive when its response to the combined
stimuli is dierent by at least 20% from the algebraic
sum of responses to each separate element of the
combined stimulus (Portfors and Wenstrup 1999).
Apparently, combination sensitivity is a general fea-
ture of auditory reception in all mammals and birds, and
not specic to echolocating animals. In evolution, echo
sensitivity, and hence, echolocation may have been es-
tablished more by quantitative than by qualitative
modications: combination-sensitive neurons are found
in very large aggregations in midbrain and higher au-
ditory centres of bats, whereas in non-echolocating
mammals and birds combination sensitivity is far less
frequently recorded and is narrowly conned to specic
subregions of auditory brain centres. Transformation of
combination sensitivity from coding communication
sounds to echo coding also implies that temporal rela-
tionships are shortened from hundreds to only a few
milliseconds, and neuronal spectral lters are tuned to
specic echolocation sound elements.
It is unlikely that combination sensitivity in echolo-
cating bats is achieved by the reversal of echo-suppres-
sion mechanisms in non-echolocating mammals from an
inhibitory to a facilitatory mode, since combination
sensitivity and echo suppression coexist in the auditory
system of non-echolocating mammals.
Locking time windows of enhanced audition to neu-
ronal commands of sound emission would be the perfect
way to generate echo sensitivity. Only one neural study
(Schuller 1979) supports such a mechanism. Apparently,
listening to the tight time relationship between heard
emitted sound and returning echoes is good enough to
generate echo sensitivity by modifying combinatory
neuronal circuits common to neuronal coding of species-
specic communication signals in all mammals and in
birds (Margoliash and Fortune 1992; Gehr et al. 2000).
Topdown modulation
Theoretically, echo sensitivity could be generated or
enhanced in a more exible way by corticofugal modu-
lation of auditory units. Recent studies (Ma and Suga
2001) have shown that corticofugal eerent innervation to
the medial geniculate and IC modulates auditory signal
processing in three domains: frequency, time, and
249
direction. Electrical stimulation of cortical FM/FM
neurons enhances the auditory responses of subcortical
combination-sensitive neurons matched in best delay to
the activated cortical neurons, and sharpens delay tuning
without shifting their best delays, whereas it suppresses
the auditory responses and shifts the best delays of un-
matched subcortical delay-tuned neurons (Yan and Suga
1996).
An earlier study (Kossl and Vater 1989) has demon-
strated that norepinephrine markedly diminishes latency
jitters in units of cochlear nuclei. This eect was inter-
preted as a more precise coding of echo timings in alert
bats due to the attentive eect of norepinephrinergic
innervation from locus coeruleus.
Thus, corticofugal modulation activated by echoes,
and general modulatory mechanisms of attention such
as norepinephrinergic or cholinergic inputs from brain
stem nuclei primarily serve to focus attention and per-
ceptive power to a sector of the outer world that is of
momentary behavioural interest. Auditory sensitivity to
echoes will be enhanced, and frequency and time lters
will be quickly shaped to actual requirements. Top
down modulation may prove to be a powerful way to
adjust the auditory system to actual demands of audi-
tory cognition in echolocation. As in vision, it might also
provide learned cognitive templates stored in memory
from previous experiences in echolocation. Therefore,
studies in topdown modulation in the neural auditory
pathway (Yan and Suga 1996) will become a central
issue in echolocation research.
Echolocation is a neuronal and not a cochlear
achievement. Mechanisms involved in generating echo-
sensitivity are still under discussion and include:
1. Time windows of echo-sensitivity triggered by vocal
command centers (Fig. 2).
2. Combination sensitivity (Fig. 3) derived from coding
species-specic communication signals.
3. Cortical topdown modulation.
In my opinion, echolocation phylogenetically was an
adaptation of already existing mechanisms in auditory
neuronal processing. In pre-bats, the transition from
neural combination circuits for sound communication to
those for echolocation was probably elicited by noctur-
nal pursuits of small ying insects. The capacity of ight
in turn may have evolved from small insectivorous and
arboreal mammals catching nocturnal insects while
running along branches and hopping from twig to twig
in trees (see also Simmons and Stein 1980; Padian 1985).
However, in another line of thinking it is supposed that
echolocation evolved in cave dwelling animals.
Specialization in evolution: the conquest
of ecological niches
Once bats came into existence their functional bauplan
did not substantially change for some 50 million years.
107
250
Yet, ight combined with echolocation was, and still is,
so advantageous to a nocturnal mammal that bats
radiated into many ecological niches. Microchiroptera
adapted echolocation systems, wing shapes and ight
styles (Rayner 1991) to all possible food sources. Some
species pursue ying insects at high speeds, others pick
up insect prey from the ground, tree bark and twigs, and
still others search for frogs, lizards, mice and birds.
Among the neotropical phyllostomids there is a large
group of bats species specialized for visiting nocturnal
owers for pollen and nectar, or for collecting fruits. The
infamous vampire bats can only subsist on blood from
mammals and birds. There is no other mammalian order
that has tapped such a rich variety of food resources.
Bat species may be dierentiated and identied by the
frequency range and the time structure of their echolo-
cation calls. These species-specic echolocation systems
may be grouped under three broad categories (Fig. 4;
Neuweiler 1984).
Fig. 4 Echolocation systems in bats. CF/FM: horseshoe bats,
hipposiderids, and the neotropic moustached bat emit multihar-
monic echolocation signals consisting of a long pure tone
terminated by a brief FM component. CF/FM bats often forage
within or close to dense vegetation FM. Most insectivorous bats
foraging on the wing emit brief, downward-FM echolocation
signals when approaching and catching a prey (FM). They often
emit longer and only shallowly modulated signal when they search
for prey (not shown) Click-like: gleaning bats and some ower
visiting bats emit very brief signals over a broad frequency band
(click-like). These echolocation sounds are far less intense than
those of FM and CF/FM bats
FM echolocation
Most bat species pursue insects on the wing. During the
approach to and capture of prey these bats emit
sequences of brief (several milliseconds) downward FM
sounds. FM signals serve as good time-markers in the
ascending auditory pathway, and hence are thought to
guarantee a precise distance perception in echolocation.
Constant-frequency/frequency-modulated echolocation
The Old World families of horseshoe bats (rhinolophids)
and hipposiderids, and one neotropical species, Pteron-
otus parnellii, emit a combined echolocation sound
consisting of a rather long (860 ms) constant frequency
(CF) invariably terminated by a brief FM element. In
many species the echolocation sounds start with a brief
frequency upward-modulated element that is especially
pronounced in commuting ights of the bats (Neuweiler
et al. 1987). Thus, the correct signature for these echo-
location systems should be FM/CF/FM. However, since
no specic function could be assigned to the initial FM
component this designation has not been endorsed.
CF/FM echolocation sounds consist of two or more
harmonics. In the moustached bat and in horseshoe bats
the second harmonic of the CF component (CF2 in
Fig. 4) is the most intense element of the echolocation
signal to which audition is specically adapted.
CF/FM bats preferably, but in no way exclusively,
forage wing-beating insects in dense vegetation. They
possess an auditory fovea in the cochlea and compensate
108

Doppler shifts of complete echo signals in order to keep
audition of CF2 echoes within an auditory fovea (see
below).
Click-like echolocation
A number of bat species that pick up prey from sub-
strates or visit owers and fruiting vegetation emit faint
and very brief signals of submillisecond durations. The
ultra-brief sounds consist of several shallowly FM har-
monics that add up to a click-like, broadband signal
covering a frequency band of up to 80100 kHz. Based
on behavioural experiments it has been suggested that
broad-band signals may serve to dierentiate textures of
targets by the interference pattern (colouration) reected
in broad band echoes (Schmidt 1988; Neuweiler and
Schmidt 1993).
In ethoecological studies of neotropical and European
bat communities the correlation between foraging niches
and type of echolocation signals has been subdivided in a
much more detailed way (Schnitzler and Kalko 1998;
Kalko and Schnitzler 1998).
Perfect adaptation by auditory specialization:
a case study in CF/FM bats
Auditory adaptations to the three categories of echolo-
cation have not been systematically studied with one
notable and striking exception, the CF/FM bats, i.e. the
moustached bat (Pteronotus parnellii) and the rufous
horseshoe bat (Rhinolophus rouxi). Over the last decades
these two bat species have become case studies of
sophisticated adaptation of an auditory system to a
specic foraging strategy.
Foraging behaviour
The foraging behaviour of the rufous horseshoe bat
exemplies foraging strategies in CF/FM bats (Neuwe-
iler et al. 1987). When these bats leave their caves in the
evening they immediately take cover within bushes and
commute under the canopy of trees and bushes into the
jungle. After a brief period of foraging on the wing the
bats spend the night foraging in a sit-and-wait strategy.
Each bat occupies its own foraging area and suspends
itself from slender twigs. The bat scans the surrounding
by emitting continuous sequences of CF/FM signals
throughout the night. Whenever an echolocation sound
hits a wing-beating insect the bat takes o for a catch
and returns with its prey to the very same spot or nearby
twigs. The horseshoe bats detect the ying insects
against the dense echo clutter reverberated from foliage
by echo-glints imposed on the long pure tone component
(CF) of the echo by each wing beat of the target. Echo-
glints are produced by frequency Doppler shifts induced
by the speed of the insect wing, and by an intense mirror
251
reection of the impinging sound waves when the
moving wing passes through a position perpendicular to
the incoming sound waves. Behavioural experiments
testied (Link et al. 1986) that horseshoe bats are indeed
specialized to wing beat detection, and deaf to non-ying
insects. Insects sitting or moving on a substrate do not
elicit attacks of horseshoe bats even at close distances.
Apparently, the long CF component of the echolocation
sound serves as a carrier for glints induced by uttering
prey. However, the modulation depths of such glints
amounts to less than 0.25% of the carrier frequency, for
instance to maximally 185 Hz on a CF component of
75 kHz emitted by a rufous horseshoe bat.
In order to detect even very tiny glints CF/FM bats
have implemented a very narrow cochlear frequency
lter tuned to the species-specic and even individual
frequency of the CF2 echo component (Fig. 5). We have
called this individually tuned frequency lter an auditory
fovea.
Mechanisms of auditory foveae
Structural and functional specialization that produce the
extremely narrow frequency lter of the auditory fovea
have been most thoroughly studied in Pteronotus
parnellii (Henson and Henson 1991; Russell and Kossl
1999; for a review see Kossl and Vater 1995) because in
the moustached bat the cochlea is experimentally more
accessible than in horseshoe bats. The functional prin-
ciples of an auditory fovea discovered in the moustached
bat are assumed to apply with moderate modications
to all CF/FM bats (rhinolophids and hipposiderids).
A complete and precise cochlear frequency map in
moustached bats (Fig. 6c) has been obtained on the
basis of an inner hair cell frequency map produced by
dye-labelling auditory neurons characterized by their
best frequency (Kossl and Vater 1985).
The foveal frequency lter in the inner ear of the
moustached bat is housed in the extended, large basal
cochlear turn and consists of two parts (Fig. 6a): (1) a
basal sparsely innervated (SI) zone characterized by
structural specializations of the basilar and tectorial
membrane. In the SI region the basilar membrane is
tuned from 72 to 62 kHz. The frequency of about
62 kHz (CF2 frequency), is represented at the transition
of the SI zone to (2) an adjacent long section of the
basilar membrane (CF2 zone) where the narrow fre-
quency band of the CF2 echo and its glints induced by
wing-beating insects are represented in a widely
expanded way.
In the SI zone, the thickness of the basilar membrane
is markedly increased by longitudinal bres that provide
mechanical coupling over its length. The tectorial
membrane is also highly modied, and shows a beam-
like structure (club-shaped in cross-sections) that may
easily vibrate since it is only loosely attached to the
spiral limbus. Both structural specializations in the
basilar and tectorial membrane abruptly disappear at
109
252
Fig. 5 The auditory fovea lter in the moustached bat. The
audiogram is derived from best thresholds of cochlear nucleus
units. EOAE: the frequency at which an otoacoustic emission may
be evoked or may occur spontaneously. EOAE frequency is
identical with the frequency of the resonator in the cochlea. EOAE
frequency is slightly dierent in each specimen. In the audiogram
EOAE frequency is normalized to 62.0 kHz (dotted line). After
Kossl (1994)
the transitory location between SI and CF2 zones where
CF2 frequencies are represented.
The frequency tuning of the tectorial membrane
In all mammals, tectorial and basilar membrane are
mechanically coupled by the hair bundles of the outer
hair cells (OHCs) that slip into depressions of the
tectorial underside. OHCs are electromechanical and
amplifying transducers with a strong nonlinearity. It
produces distortions that may show up as faint sounds
emitted from the inner ear into the outer ear canal where
they may be recorded as otoacoustic emissions (OAE).
Distortion product otoacoustic emissions (DPOAE)
Fig. 6ac Auditory fovea: the specialized basal turn of the cochlea
in moustached bats. a A camera lucida drawing of the basal turn.
Dashed lines with arrows demarcate the specialized zones SI
(sparsely innervated) and CF2 (fovea). Grey band: spiral ligament,
black band basilar membrane; radial bres: innervation of organ of
Corti. b Structural specializations for a resonant system in the SI
zone. Note the abrupt changes in thickness of basilar and tectorial
membrane, and in the limbal attachment of the tectorial membrane
at the transition between SI and CF2 zones of the basal cochlear
turn. c Frequency place map of the basilar membrane (BM) and
tectorial membrane (TM) of the complete cochlea. Note the widely
expanded representation (fovea) of the narrow frequency band
around the CF2 echolocation frequency which corresponds to that
of the evoked or spontaneous otoacoustic emission (CF2 echo,
EOAE). From Russell and Kossl (1999)
110
 253

occur when the ear is stimulated with two closely spaced
frequencies f1 and f2. The most pronounced distortion is
2f1f2. DPOAEs are ltered by the tectorial membrane
and, therefore, provide a convenient and noninvasive
way of measuring the frequency characteristics of
tectorial membranes (Kossl and Vater 1996b). In non-
echolocating mammals and in non-specialized regions of
the bat cochlea there is a mismatch of the mechanical
tuning between basilar and tectorial membrane by about
a quarter of an octave at each location (Kossl and Vater
1996b). It is assumed that this slight mismatch contrib-
utes to the sharpening of cochlear frequency lters by
reducing low-frequency tails in the tuning curves.
DPOAE measurements in FM bats disclosed the same
frequency-tuning pattern as in non-echolocating mam-
mals. In CF/FM bats, however, DPOAE-recordings
showed a functionally highly specialized tectorial mem-
brane (Fig. 6b, c; Kossl and Vater 1996a). Throughout
the SI zone of the cochlea in P. parnellii the tectorial
membrane is only tuned to the frequency of the CF2
echolocation sound around 62 kHz. This indicates that a
specialized tectorial membrane resonance plays an inte-
gral role in enhancing cochlear tuning to CF2 beyond
values normally encountered in non-CF/FM bats and
non-echolocating mammals.
operate with a basilar membrane tuned from 72 to
62 kHz and an overlaying tectorial membrane that only
resonates at about 62 Hz?
Measurements of basilar membrane vibrations in the
SI zone by noninvasive laser interferometry (Fig. 7;
Russell and Kossl 1999) disclosed that (1) each location
on the basilar membrane is indeed tuned to a charac-
teristic frequency between 72 kHz at the base of SI zone
and 62 kHz at the transition from SI to CF2 zone; and
(2) when the ear was acoustically stimulated by sounds
of about 62 kHz the SI zone resonated over its full
length.
These pioneering experiments show prove that the SI
zone in the cochlea of the moustached bat acts as a res-
onator tuned to CF2. The resonator consists of a tectorial
membrane as a driving element, and the longitudinal
coupling of the mechanics of the basilar membrane. Both
mechanical components are interconnected by the non-

The cochlear resonator in the basal turn of CF/FM bats

This resonating system is also a source for very loud

stimulus-evoked or even spontaneous OAEs that are
restricted to the CF2 frequency. In moustached bats
OAEs can not be evoked by any other stimulus
frequency. When driven by the resonator frequency, a
standing wave is generated in the SI zone. The resonant
CF2 lter is also characterized by a minimal threshold
and long lasting ringing of cochlear microphonics.
In the moustached bat, the frequency of spontaneous
or stimulus-frequency-elicited OAEs is considered as the
resonator frequency. In the auditory fovea this reso-
nance frequency is situated about halfway in the very
steep sensitivity slope from a high threshold peak at the
low-frequency side of the foveal lter to the best fre-
quency of the fovea (CF2 echo frequency; Fig. 5). The
frequency distance between maximal insensitivity and
maximal sensitivity only amounts to 1 kHz or less. The
insensitivity peak coincides with the so-called resting
frequency around 61 kHz, that is the frequency of the
CF2 component emitted by a non-ying moustached Fig. 7 Displacement measurements of the basilar membrane at a
bat. The best frequency of the fovea corresponds to the location where 64.2 kHz are represented in the moustached bat.
CF2 echo frequency (Fig. 5) that ying moustached bats Vibrations were recorded by laser interferometry in the intact
actively maintain by an audio-vocal feedback system cochlea through the round window while the ear was stimulated by
tone pulses of frequencies from 20 to 90 kHz (abscissa) and sound
(Doppler shift compensation) that eliminates Doppler levels indicated on the right vertical axis. When stimulated with the
shifts of the complete echo caused by the bats own ight individual CF2 frequency of the specimen (61.125 kHz) the 64.2-
speed. The closely spaced peaks of insensitivity and kHz place of the BM vibrates at even lower sound levels than by
sensitivity are both considered to be generated by the stimulations at the place frequency of 64.2 kHz. R resonant
vibrations; arrows only: vibrations elicited by 64.2-kHz stimuli.
resonating SI zone (see below). This is strong evidence for the existence of a resonator tuned to the
If a tectorial resonance is the source for the extremely CF2 echolocation signal in the SI zone (From Russell and Kossl
narrowly tuned auditory fovea, how does the SI region (1999)

111
254
linear dynamics of OHCs. When OHCs are functionally
eliminated, evoked OAE at the resonance frequency
surprisingly increases in level or convert to spontaneous
OAE (Faulstich and Kossl 1997; Kossl and Vater 2000).
It seems that the OHCs exert dampening and stabilizing
eects in order to prevent damages by an easily over-
driven resonator.
The discovery of the resonator in the SI zone of the
moustached bats cochlea generates more problems than
it solves:
1. Why is the row of IHCs in the SI region not excited
when this complete cochlear section is vibrating at
62 kHz?
2. How can the resonator generate at the same time a
closely spaced minimum and maximum of sensitivity?
3. How is resonant power fed into the adjacent CF2
zone of the auditory fovea?
In order to answer such questions Russell and Kossl
(1999) developed a model for the generation of standing
wave resonance (Fig. 8). They view the tectorial and
basilar membrane as longitudinally spanned strings. The
tectorial membrane is xed at the basal and apical ends of
the SI zone (nodes in Fig. 8), whereas the basilar mem-
brane acts as a sturdy string that may freely vibrate at the
basal end of the SI zone and at the transition to the foveal
CF2 zone. These are the locations were the longitudinal
bres in the thick basilar membrane section terminate.
The authors suggest that the standing wave at 62 kHz is
generated by reections at the two ends of the SI zone.
Thus, the response of the cochlea to CF2 echo components
will be sharpened and stabilized by the standing wave
resonance. The standing wave will create large cochlear
pressure changes at the 62-kHz location of the cochlea, i.e.
it will pump energy into the CF2 zone where the narrow
frequency band around 62 kHz is represented in a vastly
expanded frequency map of high sensitivity.
It is assumed that the SI zone has two modes of
vibration. First, when stimulated by a frequency of
62 kHz the tectorial membrane will strongly resonate
and the resonant energy will be released more apically in
the CF2 zone. In the SI zone, due to the loose attach-
ment of the tectorial membrane to the spiral limbus,
tectorial and basilar membrane will vibrate together in
the transverse plane as a single mass, and the vibration
of the tectorial membrane enforces the same movements
in the basilar membrane. Since both membranes move as
a single entity, only minimal shearing of the sensory hair
bundles between the two membranes will occur. There-
fore, the inner hair cells of the SI region are not excited
by 62 kHz. For frequencies a few hundred Hertz below
the resonance frequency of 62 kHz, basilar and tectorial
membrane are still locked in their movement and strong
oscillations emerge. However the phase of the oscillation
has shifted by 180 (as experimentally demonstrated)
with the result that now the resonator is in antiphase to
the input stimulus and absorbs acoustic energy similar
to the situation in commercial muers used, e.g. in cars.
Fig. 8 The Russell-Kossl model of a resonating system in the SI
zone as a constituting element of the auditory fovea in the
moustached bat. Upper graph shows the attachment of the TM to
the spiral limbus in the dierent zones of the basal cochlear turn.
TM bars symbolize the size of the attachment areas of the TM that
is narrow in the SI zone. BM bars symbolize thickness of the BM
that is thickened by longitudinal bres in the SI zone. Lower graph
shows the assumed vibration behaviour when both membranes in
the SI zone resonate as a standing wave at 62 kHz. After Russell
and Kossl (1999)
In the audiogram this results in a very narrowly tuned
peak of insensitivity just below the CF2 center frequency
of the fovea-lter (Fig. 5). Actually, the narrowness of
the foveal frequency lter is not owned to very low
thresholds but to this insensitivity peak that results in an
extremely steep slope at the low frequency side of the
lter. Thus, the SI resonator generates a peak of insen-
sitivity and at the same time a peak of sensitivity by
pumping energy into the adjacent CF2 zone of the
cochlear basal turn.
2) When the SI zone is driven by other frequencies
represented in the SI region (higher frequencies in
Figs. 6 and 8) that do not initiate resonant vibrations of
the complete tectorial membrane the SI region locally
vibrates as a two-mass system in the radial plane as in all
other regions of the cochlea and will shear the hair
bundles of the hair cells.
These two modes of vibration plausibly explain why
the IHCs of the SI zone are not excited by the 62-kHz
resonance, yet correctly respond to stimulations at their
best frequencies above 62 kHz up to 72 kHz.
It is assumed that this Russell-Kossl model of a
cochlear resonator derived from experiments in P.
parnellii may apply to all CF/FM bats. However, in
horseshoe bats resonator-characteristics are not as dis-
tinct as in moustached bats: the OAE at the foveal fre-
quency is far less intense, and ringing or the sensitivity
peak of microphonics are far less pronounced. These
dierences might be due to a strong damping of reso-
112

nator systems in horseshoe bats (Henson et al. 1985;
Kossl 1994).
Where do the resonators in the cochleae of CF/FM bats
come from?
The literature states that in inner ears resonance must
not occur. Where, then, do the unique resonators in the
basal turns of CF/FM bats phylogenetically come from?
In the New World, P. parnellii is the only CF/FM bat
so far known. A comparative study in three mormoopid
species closely related to the moustached bats disclosed
no structural or functional intermediate stages between
FM echolocation in related species and the CF/FM
system in the moustached bat (Kossl et al. 1999). The
only distinct dierence between them is the large basal
turn of the cochlea with its mechanical specializations
described above.
Molecular data suggest that the moustached bat only
recently evolved about 49 million years ago (Kossl et al.
1999). In rhinolophids and hipposiderids CF/FM echo-
location is phylogenetically much older, and this may be
the reason why there are about 150 CF/FM species in
the Old World and only one in the neotropics. In any
case, the auditory foveae in P. parnelli and in horseshoe
bats are one of the most striking examples of convergent
evolution.
It challenges our imagination to conceive evolution-
ary driving forces initiating mechanical modications
that turned the basal turn of the cochlea not only into a
place of high-frequency representation but also into
a sharply tuned resonator. One idea postulates that the
CF resonators are not the result of specic evolutionary
driving forces but were invented by an accident which
may consist of a change in ontogenetic programs con-
trolling the development of the basilar and tectorial
membranes and produce a strong mechanical disconti-
nuity. The consequence could be abrupt changes in
acoustical impedance, acoustical reections and resonant
oscillations which could produce a frequency-specic
cochlear insensitivity. This speculation is based on a
study of OAE-generating mechanisms in other mammals
(Shera and Guinan 1999). OAEs are not only measurable
in the form of distortion products but may be also elic-
ited by distinct stimulus frequencies or they may even
occur spontaneously. Spontaneous or frequency driven
OAEs mainly result from reections of travelling waves
at acoustical impedance irregularities occurring haphaz-
ardly along the cochlear partition. Coherent reections
from such mechanical discontinuities in the cochlea may
create narrow-band cochlear standing waves showing up
as OAEs that are recordable in the ear canal. This applies
to the resonator in P. parnellii: exactly such a very
prominent and punctuate mechanical discontinuity at the
transition from the SI to the CF2 zone, and an intense
standing wave are the hallmarks of the resonator in the
basal cochlear turn of moustached bats.
255
According to this concept the prominent mechanical
discontinuity and its resulting standing wave may have
occurred accidentally in one of the forbears of CF/FM
bats. The bats may have made the best of it by imple-
menting into their echolocation system the accidental
sharp-frequency ltering that results from the resonating
wave. In this case, function would follow structure and
the bats would have added a CF component to their
echolocation call with a frequency dened by the audi-
tory foveal frequency. Vocalization would adapt to an
auditory accident.
If this scenario were correct, this would be a re-
markable process in which an accidental non-functional
mishap would be turned into an evolutionary benecial
and highly sophisticated specialization.
In whatever way the auditory foveae with its narrow
frequency ltering may have evolved, the cochlear
resonators have made CF/FM bats specialists for de-
tecting uttering targets. In a highly cluttering envi-
ronment this specialization may be an advantage.
However, such sophisticated specializations are evolu-
tionary deadlocks. If, for instance, dense vegetations
disappear, the advantage of uttering target detection
disappears as well, and species that predominantly rely
on uttering target detection by CF/FM echolocation
may become extinct. Even without anthropogenic im-
pact ecosystems and ecological niches change continu-
ously. Hence, species with sophisticated adaptations to
specic niches will disappear and new species answering
to new challenges will evolve. As long as there is time
enough for newly evolving species to adapt to new
environments the turnover of specialized species is a
natural process and a driving force of biodiversity.
Obviously, specialized animals naturally have a limited
time of existence.
Acknowledgements I thank Prof. Manfred Kossl, Frankfurt for
critically reading the manuscript, and S. Peisker for preparing the
gures. I am also grateful to an anonymous reviewer who helped to
clarify the text.
References
Covey E, Casseday JH (1995) The lower brainstem auditory
pathways. In: Popper AN, Fay RR (eds) Hearing in bats.
Springer handbook of auditory research. Springer, Berlin
Heidelberg New York, pp 235295
Faulstich M, Kossl M (1997) Thalamonal alters cochlear me-
chanics in the mustached bat. Proc 20th Midwinter Res Meet
Assoc Res Otolaryngol, p 11
Fitzpatrick DC, Kuwada S, Batra R, Trahiotis (1995) Neural re-
sponses to simple, simulated echoes in the auditory brainstem of
the unanesthetized rabbit. J Neurophysiol 74:24692486
Gehr DD, Komiya H, Eggermont JJ (2000) Neuronal responses in
cat primary auditory cortex to natural and altered species-
specic sounds. Hear Res 150:2742
Grinnell AD (1963) The neurophysiology of audition in bats:
temporal parameters. J Physiol (Lond) 167:6796
Grothe B, Neuweiler G (2000) The function of the medial superior
olive in small mammals: temporal receptive elds in auditory
analysis. J Comp Physiol A 186:413423
113
256
Haplea S, Covey E, Casseday JH (1994) Frequency tuning and
response latencies at three levels in the brainstem of the echo-
locating bat, Eptesicus fuscus. J Comp Physiol A 174:671683
Henson MM, Henson OW (1991) Specializations for sharp tuning
in the mustached bat: the tectorial membrane and the spiral
limbus. Hear Res 35:237258
Henson OW, Schuller G, Vater M (1985) A comparative study of
the physiological properties of the inner ear in Doppler shift
compensating bats (Rhinolophus rouxi and Pteronotus parnellii).
J Comp Physiol A 157:587597
Kalko EKV, Schnitzler HU (1998) How echolocating bats
approach and acquire food. In: Kunz TH, Racey PA (eds)
Bat biology and conservation. Smithonian Institution Press,
Washington, DC, pp 197204
Keller CH, Takahashi TT (1996) Response to simulated echoes by
neurons in the barn owls auditory space map. J Comp Physiol
178:499512
Kossl M (1994) Evidence for a mechanical lter in the cochlea of
the constant-frequency bats, Rhinolophus rouxi and Pteronotus
parnellii. Hear Res 72:7380
Kossl M, Vater M (1985) The cochlear frequency map of the
mustached bat, Pteronotus parnellii. J Comp Physiol A 157:687
697
Kossl M, Vater M (1989) Noradrenaline enhances temporal audi-
tory contrast and neuronal timing in the cochlear nucleus of the
mustached bat. J Neurosci 9:41694178
Kossl M, Vater M (1995) Cochlear structure and function in bats.
In: Popper AN, Fay RR (eds) Hearing by bats. Springer, Berlin
Heidelberg New York, pp 191234
Kossl M, Vater M (1996a) Further studies on the mechanics of the
cochlear partition in the mustached bat. II. A second cochlear
frequency map derived from acoustic distortion products. Hear
Res 94:7886
Kossl M, Vater M (1996b) A tectorial membrane fovea in the
cochlea of the mustached bat. Naturwissenschaften 83:8991
Kossl M, Vater M (2000) Consequences of outer hair cell damage
for otoacoustic emissions and audiovocal feedback in the
mustached bat. J Assoc Res Otolaryngol 1:300314
Kossl M, Mayer F, Frank G, Faulstich M, Russell IJ (1999)
Evolutionary adaptations of cochlear function in Jamaican
mormoopid bats. J Comp Physiol A 185:217228
Link A, Marimuthu G, Neuweiler G (1986) Movement as a specic
stimulus for prey catching behaviour in rhinolophid and
hipposiderid bats. J Comp Physiol A 159:403413
Litovsky RY (1998) Physiological studies on the precedence eect
in the inferior colliculus of the kitten. J Acoust Soc Am
103:31393152
Litovsky RY, Yin TCT (1998) Physiological studies of the prece-
dence eect in the inferior colliculus of the cat. I. Correlates of
psychophysics. J Neurophysiol 80:12851301
Litovsky RY, Colburn HS, Yost WA, Guzman SJ (1999) The
precedence eect. J Acoust Soc Am 106:16331653
Ma X, Suga N (2001) Corticofugal modulation of duration-tuned
neurons in the midbrain auditory nucleus in bats. Proc Natl
Acad Sci USA 98:1406014065
Margoliash D, Fortune ES (1992) Temporal and harmonic com-
bination-sensitive neurons in the zebra nchs HVc. J Neurosci
12:43094326
Neuweiler G (1984) Foraging, echolocation and audition in bats.
Naturwissenschaften 71:446455
Neuweiler G (2000) Biology of bats. Oxford University Press, New
York
Neuweiler G, Schmidt S (1993) Audition in echolocating bats. Curr
Opin Neurobiol 3:563569
Neuweiler G, Metzner W, Heilmann U, Rubsamen R, Eckrich M,
Costa HH (1987) Foraging behaviour and echolocation in the
rufous horseshoe bat of Sri Lanka. Behav Ecol Sociobiol 20:53
67
Olsen JF, Suga N (1991) Combination-sensitive neurons in the
medial geniculate body of the mustached bat: encoding target
range information. J Neurophysiol 65:12751296
ONeill WE, Suga N (1982) Encoding of target range and its
representation in the auditory cortex of the mustached bat. J
Neurosci 2:1731
Padian K (1985) The origin and aerodynamics of ight in extinct
vertebrates. Palaeontology 28:413433
Portfors CV, Wenstrup JJ (1999) Delay tuned neurons in the in-
ferior colliculus of the mustached bat: implications for analyses
of target distance. J Neurophysiol 82:13261338
Portfors CV, Wenstrup JJ (2001) Responses to combinations of
tones in the nuclei of the lateral lemniscus. JARO 2:104117
Rauschecker JP, Tian B, Hauser M (1995) Processing of complex
sounds in the macaque nonprimary auditory cortex. Science
268:111114
Rayner JMV (1991) Complexity in a coupled system: ight, echo-
location and evolution in bats. In: Schmidt-Kittler, Vogel K
(eds) Constructional morphology and evolution. Springer,
Berlin Heidelberg New York, pp 173191
Roverud RC (1993) Neural computations for sound pattern
recognition: evidence for summation of an array of frequency
lters in an echolocating bat. J Neurosci 13:23062312
Roverud RC, Grinnell AD (1985) Echolocation sound features
processed to provide distance information in the CF/FM bat,
Noctilio albiventris: evidence for a gated time window utilizing
both CF and FM components. J Comp Physiol A 156:457469
Russell IJ, Kossl M (1999) Micromechanical responses to tones in
the auditory fovea of the greater mustached bats cochlea. J
Neurophysiol 82:676686
Schmidt S (1988) Evidence for a spectral basis of texture perception
in bat sonar. Nature 331:617619
Schnitzler HU, Kalko EKV (1998) How echolocating bats search
and nd food. In: Kunz TH, Racey PA (eds) Bat biology and
conservation. Smithonian Institution Press, Washington, DC,
pp 183196
Schuller G (1979) Vocalization inuences auditory processing in
collicular neurons of the CF-FM bat, Rhinolophus ferrumequi-
num. J Comp Physiol A 132:3946
Schuller G, ONeill WE, Radtke-Schuller S (1991) Facilitation and
delay sensitivity of auditory cortex neurons in CF-FM bats,
Rhinolophus rouxi and Pteronotus parnellii. Eur J Neurosci
3:11651181
Shera CA, Guinan JJ (1999) Evoked otoacoustic emissions arise by
two fundamentally dierent mechanisms: a taxonomy for
mammalian OAEs. J Acoust Soc Am 105:782798
Simmons JA, Stein RA (1980) Acoustic imaging in bat sonar:
echolocation signals and the evolution of echolocation. J Comp
Physiol A 135:6184
Yan J, Suga N (1996) Corticofugal modulation of time-domain
processing of biosonar information in bats. Science 273:1100
1103
Yost WA, Guzman SJ (1996) Auditory processing of sound
sources: Is there an echo in here? Curr Direct Psychol Sci 5:125
131
114
J Comp Physiol A (2008) 194:841851
DOI 10.1007/s00359-008-0355-x
ORIGINAL PAPER
On-board telemetry of emitted sounds from free-Xying bats:
compensation for velocity and distance stabilizes echo frequency
and amplitude
Shizuko Hiryu Yu Shiori Tatsuro Hosokawa
Hiroshi Riquimaroux Yoshiaki Watanabe
Received: 21 April 2008 / Revised: 16 July 2008 / Accepted: 20 July 2008 / Published online: 29 July 2008
Springer-Verlag 2008
Abstract To understand complex sensorymotor behavior
related to object perception by echolocating bats, precise
measurements are needed for echoes that bats actually lis-
ten to during Xight. Recordings of echolocation broadcasts
were made from Xying bats with a miniature light-weight
microphone and radio transmitter (Telemike) set at the
position of the bats ears and carried during Xights to a
landing point on a wall. Telemike recordings conWrm that
Xying horseshoe bats (Rhinolophus ferrumequinum nippon)
adjust the frequency of their sonar broadcasts to compen-
sate for echo Doppler shifts. Returning constant frequency
echoes were maintained at the bats reference frequency
83 Hz during Xight, indicating that the bats compensated
for frequency changes with an accuracy equivalent to that
at rest. The Xying bats simultaneously compensate for
increases in echo amplitude as target range becomes
shorter. Flying bats thus receive echoes with both stabilized
frequencies and stabilized amplitudes. Although it is widely
understood that Doppler-shift frequency compensation
facilitates detection of Xuttering insects, approaches to a
landing do not involve Xuttering objects. Combined fre-
quency and amplitude compensation may instead be for
S. Hiryu Y. Shiori T. Hosokawa H. Riquimaroux Y. Watanabe
Faculty of Engineering, Doshisha University,
Kyotanabe 610-0321, Japan
H. Riquimaroux Y. Watanabe
Bio-navigation Research Center,
Doshisha University, Kyotanabe 610-0321, Japan
Present Address:
S. Hiryu (&) H. Riquimaroux Y. Watanabe
Faculty of Life and Medical Sciences,
Doshisha University, Kyotanabe 610-0321, Japan
e-mail: shiryu@mail.doshisha.ac.jp
optimization of successive frequency modulated echoes for
target range estimation to control approach and landing.
Keywords Doppler-shift compensation Echo-intensity
compensation Rhinolophus ferrumequinum nippon
CFFM bats
Abbreviations
BF Best frequency
CF Constant frequency
CM Cochlear microphonic
DSC Doppler-shift compensation
FM Frequency modulated
IPI Interpulse interval
RF Resting frequency
Introduction
Echolocating animals perceive objects in their environment
with great accuracy by comparing emitted pulses to return-
ing echoes (GriYn 1958). Among echolocating bats, the
Rhinolophidae, Hipposideridae and Mormoopidae all emit
biosonar pulses consisting of a constant frequency (CF)
component followed, by a downward-sweeping frequency
modulated (FM) component. Often the CF component also
is preceded by, a short, upward-sweeping FM component
(GriYn 1958; Neuweiler 2000; Thomas et al. 2003). In
CFFM bats, the strongest part of the sound is the second
harmonic (CF2 and FM2). When a CFFM bat approaches a
target, the frequency of returning echoes is Doppler-shifted
upward depending on the relative velocity between bat and
target. Schnitzler (1968) demonstrated that greater horse-
shoe bats (Rhinolophus ferrumequinum) decreased the fre-
quency of the CF component (CF2) of emitted pulses to
123
115
842
compensate for the Doppler shift of echoes induced by
approaching Xight, thus regulating echoes to a stable
received frequency regardless of Xight velocity. In subse-
quent studies (Schuller et al. 1974; Simmons 1974), sta-
tionary greater horseshoe bats changed the frequency of the
CF component of their pulses when these bats were
exposed to artiWcial Doppler-shifted echoes. Further studies
revealed that CFFM bats regulated CF echoes to a very
narrow frequency range that corresponded to the narrowly
tuned frequency range in which the bats hearing is most
sensitive (Suga and Jen 1976; Ostwald 1984). Doppler-
shift compensation (DSC) is an important behavioral
adaptation of echolocation in these taxa and the sensory
motor interactions of DSC behavior are considered to help
our understanding of audiovocal control in mammalian
auditory system (Suga 1984; Gaioni et al. 1990; Riquima-
roux et al. 1991; Metzner 1993; Smotherman and Metzner
2003).
Other studies have reported that some species of bats
decrease the amplitude of their emitted pulses when they
approach prey or Xy towards an obstacle (GriYn 1958; Jen
and Kamada 1982; Vogler and Neuweiler 1983; Kick and
Simmons 1984; Kobler et al. 1985; Hartley et al. 1989;
Gaioni et al. 1990; Tian and Schnitzler 1997; Boonman and
Jones 2002). For example, when the CFFM mustached bat
(Pteronotus parnellii) is attached to a pendulum and swung
toward a large Wxed target, the bat signiWcantly decreased
the amplitude of its emitted pulse during the forward swing
(Kobler et al. 1985; Gaioni et al. 1990). This behavior rep-
resents another compensation mechanism, echo-intensity
compensation, in which pulse amplitude is adjusted in
relation to the distance to a target, resulting in stabilization
of echo amplitudes during the bats approach (Kobler et al.
1985; Gaioni et al. 1990). Frequency-adjusting DSC
responses have also been recorded while swinging bats on a
pendulum that carries the recording microphone (Henson
et al. 1982; Kobler et al. 1985; Gaioni et al. 1990), but oth-
erwise frequency and amplitude compensations by Xying
bats have thus far only been estimated from call parame-
ters recorded from bats Xying at a distance from the micro-
phone. To better understand complex sensorymotor
behavior and its role in object perception by echolocating
bats, more precise measurements are needed for the fea-
tures of echoes that the bats actually listen to during Xight.
We report here direct recordings of free-Xying CFFM
horseshoe bats, Rhinolophus ferrumequinum nippon, made
without the variability inherent in recordings made with
remote microphones. To obtain unimpeded measurements
of the dual compensation behaviors for frequency and
amplitude, we employed a miniature light-weight micro-
phone and radio transmitter (Telemike) attached to the
bat with the microphone at the position of the bats ears and
carried by the Xying bat (Henson et al. 1987; Lancaster
123
J Comp Physiol A (2008) 194:841851
et al. 1992; Riquimaroux and Watanabe 2000; Hiryu et al.
2007). When bats exhibit compensation behaviors, they are
assumed to regulate their call parameters (behavioral
response) in relation to the echoes returning from objects
they are attending to (acoustic stimuli). In this study, we
experimentally investigated the relation between echoloca-
tion calls and returning echoes during compensation behav-
iors which were induced from natural Xight.
Methods
Animals
Four adult Japanese horseshoe bats (R. ferrumequinum
nippon, all males; mass 2023 g) were used in this study. They
were captured from a natural cave in Hyogo prefecture in
Japan under license and in compliance with current Japa-
nese laws. The animals were housed in a temperature- and
humidity-controlled colony room (3 2 2 m) at Dosh-
isha University in Kyoto, Japan. In the colony room, the
bats were allowed to Xy freely and given access to food and
water. The day and night cycle of the room was controlled
at 12 h dark and 12 h with light. Once every three days, the
bats were released for an hour in a large Xight chamber (see
below for details) for exercise, and they were weighted to
check their condition.
Recording procedure
The experiments were conducted in a Xight chamber 8 (L)
3 (W) 2 m (H) under long wavelength infrared lighting
with red Wlters (>650 nm) to avoid visual eVects. The
chamber was made of steel plates to minimize interference
from external electromagnetic waves used by FM radio sta-
tions. Each of the four bats was trained to Xy from one end
of the Xight chamber directly toward the 3 2 m wall at the
other end, and to land on a landing mesh [1 (W) 1 m (H)]
was attached in the center of the wall. Echolocation sounds
emitted during each Xight were recorded using a custom-
made telemetry microphone (Telemike) mounted on the
bat. The recording procedure of the Telemike was the same
as in a previous study (Hiryu et al. 2007). The Telemike
consisted of a 1/8-inch omni-directional condenser micro-
phone (Knowles, Model FG-3329, Itasca, IL, USA), a min-
iature custom-designed FM transmitter unit, a 1.5-V
hearing-aid battery (Sony, Type SR421SW, Tokyo, Japan)
and a transmitting antenna. The Telemike was attached to
the back of the bat with a piece of double-sided adhesive
tape, with the microphone pointed forward and positioned
approximately 1 cm above the noseleaf on the bats face.
Because the Telemike weighed less than 0.6 g, including
the battery, it was light enough to be carried by Xying bats.
116
J Comp Physiol A (2008) 194:841851 843

Experimental sessions were less than an hour each, and the
bats did not exhibit any fatigue during the experiments as
veriWed by continuous visual inspection and vigor of Xight
(see below for details). The removal of the Telemike from
the back of a bat after each experiment was facilitated by
the use of a parting agent to avoid skin irritation. The
Telemikes transmitter produced radio signals with a carrier
frequency between 100 and 105 MHz that were received by
a wire antenna consisting of several back-and-forth loops
attached to the ceiling of the Xight chamber. The received
signals were demodulated to recover the bats ultrasonic
broadcasts using a custom-made FM receiver. The signals
from the receiver were then high-pass Wltered at 20 kHz
(NF Corporation, Model 3625, Yokohama, Japan), digi-
tized by a DAT recorder (SONY, Model SIR-1000W,
Tokyo, Japan, 16-bit, 384 kHz), and stored as Wles on the
hard disk of a personal computer. Prior to recording, the
output of the Telemike system was calibrated using a loud-
speaker (Pioneer, Type PT-R7, Tokyo, Japan) and a Brel
and Kajer Model 4138 (1/8-in) condenser microphone.
The total frequency response of the Telemike system was
Xat to within 4 dB between 20 and 100 kHz. Since echo-
location sounds were recorded directly by the microphone
and bats Xight velocity were determined in conjunction
with the frequency, amplitude and other parameters of the
bats echolocation sounds. Due to the size of the Xight
room and the placement of the video cameras, 3D recon-
struction of the bats Xight path was limited to positions
within 56 m of the target wall.
Sound analysis
The acoustic characteristics of the Xying bats broadcast
sounds and reXected echoes were analyzed from spectro-
grams of the Telemike recordings using custom Matlab
routines on a personal computer. As illustration of the
method, Fig. 1a shows typical echolocation sounds
recorded by the Telemike when the Xying bat approached
the target wall. Each pulse or echo was extracted from the
recording, and then the second harmonic component of the
pulse-echo pair was analyzed. The spectrogram exhibited a
peak in energy at the CF2 component in each sound, and
this energy maximum was measured in each pulse and each
A
Amplitude [V]

placed stationary above the head of the Xying bat, acoustic

parameters such as amplitude could be measured precisely
without unknown sound-to-sound variations from sound
propagation losses in the air or from directional properties
Frequency [kHz]

of the emitted pulse. 100

The Xight behavior of the bats was also recorded using 80

two digital high-speed video cameras (NIPPON ROPER
60
Co., Ltd., CR Imager model 2000s, Chiba, Japan) located
on the left and right sides of the Xight chamber, at the wall 40

behind the bat so as not to interfere with bats Xight path. 50ms
The video cameras recorded 125 frames per second, and B
three-dimensional coordinates of the Xying bats were
reconstructed from these video images using motion analy- 40
sis software (DITECT, Dipp-Motion 2D v 2.1). Prior to
recording bat Xights, a three-dimensional (3D) reference 30
frame with known coordinates was positioned in the center
of the Xight chamber and brieXy recorded by the two video
ED [ms]

20
cameras. The analysis software calibrated the 3D Xight path
reconstruction system using the cameras stereo view of the
10
reference frame. Based on a direct linear transformation Sound data
3D coordinate
technique from the reference frames coordinates, succes-
sive positions of the Xying bat as well as locations of other 0
5 4 3 2 1 0
objects were reconstructed from video-scene coordinates Target distance [m]
measured oV the pair of 2D video images. The control sig- Fig. 1 Representative echolocation sounds of R. ferrumequinum
nal that triggered and synchronized the frames of the video nippon, recorded with the Telemike during Xight. a Amplitude patterns
cameras with each other was digitally stored using the DAT of recorded sound and the corresponding sonogram. b Changes in the
recorder so that Xight coordinates could then be synchronized echo delay from the target wall determined from sound data (solid circle)
and three-dimensional coordinate data of the bats location (line). The
with Telemike sound recordings. Using three-dimensional sonogram indicates typical pulse-echo pairs in the second harmonic
coordinate data, the Xight trajectory of the bat, the dis- portion recorded by the Telemike during Xight. Echoes were observed
tance from the target wall (referred to as the target distance) following the second harmonic component of an emitted pulse

123

117
844 J Comp Physiol A (2008) 194:841851

echo to quantify changes in intensity of echoes in relation
to the pulses. For each Xight session, the measured energy
values were normalized relative to the maximum energy
across all that sessions pulses. The CF2 frequency was
determined from the frequency at the same peak energy
location in the spectrogram with a frequency resolution of
46 Hz using a fast Fourier transform over 8,192 sample
points identiWed from the spectrogram.
The representative spectrogram segment in Fig. 1b
shows both pulses and echoes from the target wall. The
accompanying graph in Fig. 1b compares the values of
echo delay (ED) determined from spectrograms (data-
points) with the delay expected for diVerent target distances
determined from the 3D video reconstruction of one land-
ing Xight as the bat approached the target wall (line). The
delay of successive echoes from the target wall (t) was
calculated from the target distance (d) using the formula
t = 2d/c, where c is the sound velocity in air (344 m/s). By
comparing the EDs obtained from the acoustic data with the
expected delays for the target wall from the video recon-
struction, the echoes from the target were identiWed in the
spectrograms for frequency and amplitude analysis. Statis-
tical analysis of data was conducted using either Students
t test or the Wilcoxon rank sum test.
Results
Echolocation sounds of bats during Xight
Four individual horseshoe bats made repeated Xights carry-
ing the Telemike. These Xights yielded complete simulta-
neous video and acoustic recordings for the type of analysis
shown in Fig. 1 from a total of 33 Xights from the four bats,
with 79 Xights per bat. From the video track reconstruc-
tions, the Xight speed of the bat usually reached a maximum
of 34 m/s during its Xight. The bats produced trains of echo-
location sounds while Xying along the room, during landing,
and then while stationary after landing on the mesh attached
to the wall. During Telemike-recorded Xights, R. ferrumequi-
num nippon used compound echolocation signals, each
consisting of a CF component with the 2nd harmonic around
6870 kHz being strongest, plus an accompanying initial
short upward FM sweep (28 kHz, ending at 6870 kHz) and
a terminal short downward FM sweep (beginning at 6870
kHz and extending 812 kHz lower) (Fig. 1a). The resting
frequency (RF; the CF2 frequency of pulses emitted by bats
at rest) was measured for the stationary bat on the landing
mesh just after each Xight using the Telemike. Table 1
shows the mean RFs of the signals emitted by the four
R. ferrumequinum nippon in the 33 Xights, together with
number of sounds contributing to the averages and their stan-
dard deviations (SD). RFs estimated for each Xight were fur-
ther averaged across Xights to obtain an overall mean RF for
each bat (Table 1). Overall mean values of RF ranged from
68.86 to 69.23 kHz, with SDs from 0.07 to 0.1 kHz.
Mean RFs were signiWcantly diVerent across the four bats
(ANOVA, P < 0.05), while the RF showed no signiWcant
intra-individual changes within a day for each bat.
Figure 1a illustrates the temporal pattern of broadcasts
emitted by a bat during a representative Xight. The sounds
were emitted not at a constant rate but in groups of double
and triple pulses (two or three successive pulses) having
shorter interpulse intervals (IPIs) within the group and
longer IPIs between groups. These clusters of sounds have
been called strobe groups (Moss and Surlykke 2001). Just
before landing the bats usually emitted four or more (<10)
multiple pulses in a single larger strobe group.

Table 1 Variations of the resting frequency for the four bats

No. Bat A Bat B Bat C Bat D

Day RF Day RF Day RF Day RF

1 8/28/08 69.17 0.03 (74) 8/28/06 69.14 0.02 (63) 9/9/06 68.98 0.05 (71) 10/24/06 68.97 0.02 (95)
2 8/28/06 69.18 0.03 (81) 10/3/06 69.18 0.05 (198) 9/9/06 69.02 0.06 (31) 10/24/06 68.96 0.03 (79)
3 8/28/06 69.16 0.03 (51) 10/12/06 69.16 0.02 (24) 10/12/06 69.03 0.06 (48) 10/24/06 68.99 0.03 (72)
4 8/31/06 69.22 0.03 (81) 10/12/06 69.16 0.02 (24) 10/24/06 68.96 0.07 (106) 11/16/06 68.91 0.04 (83)
5 8/31/06 69.20 0.03 (65) 10/12/06 69.20 0.04 (124) 10/24/06 68.99 0.06 (64) 11/30/06 68.80 0.04 (172)
6 8/31/06 69.18 0.04 (38) 10/12/06 69.22 0.04 (111) 11/16/06 68.93 0.07 (52) 11/30/06 68.80 0.04 (114)
7 8/31/06 69.19 0.03 (51) 12/8/06 68.96 0.06 (90) 11/16/06 68.96 0.06 (63) 11/30/06 68.80 0.05 (162)
8 9/5/06 69.31 0.03 (75) 11/16/06 68.96 0.07 (53) 11/30/06 68.81 0.04 (95)
9 9/5/06 69.30 0.04 (209) 11/16/06 68.98 0.05 (13)
Total RF 9 7 9 8
(kHz) 69.23 0.07 (725) 69.15 0.10 (610) 68.97 0.07 (503) 68.86 0.09 (872)
Values are means SD. The RFs are signiWcantly diVerent among four bats (KruskalWallis; P < 0.05). There is no signiWcant intra-individual
diVerence of RFs within a day for each bat. Number of pulses given in parentheses

123

118
J Comp Physiol A (2008) 194:841851
As is the case for other subspecies of Rhinolophus
ferrumequinum (Gustafson and Schnitzler 1979; Tian and
Schnitzler 1997; Hiryu et al. 2005), R. ferrumequinum
nippon modiWed its emission pattern as it came nearer to
the landing place. The bats started to decrease the IPIs and
the pulse durations when they were approximately 23 m
from the target wall. As illustrated in Fig. 2, during Xight
the IPIs alternated between shorter and longer intervals as a
reXection of the organization of the pulse train into strobe
groups. The long intervals in this alternating pattern gradu-
ally shortened from 80 to 60 ms, while the intervening
short intervals shortened from 30 to 40 ms. When these
alternations are taken into account, however, on average,
the IPIs did not shorten appreciably while the bat was in
Xight until it had approached to within 2 m from the wall
(Fig. 2). Alternations in IPIs persisted until the bat had
approached to within about 1 m from the wall, at which
point the IPIs abruptly shortened to 20 ms and then con-
tinued to shorted to 12 ms when landing occurred. Alter-
nating IPIs were less frequent during this terminal-stage
landing buzz, when the pulse emission rate reached
7080 pulses/s. As shown in Fig. 2, pulse durations for the
100
6 Spectrograms of echoes recorded by the Telemike as they
5 returned to the Xying bat only showed the second harmonic,
IPI [ms]
4 CF2. As in other horseshoe bats, CF2 was emitted with the
3 greatest energy. The CF1 component of the emitted pulse
2 ike. In these Xight experiments, the Telemike was suY-
10
5 4 3 2 1 0
50
40
Pulse duration [ms]
30
20
10
5 4 3 2 1 0
Target distance [m]
Fig. 2 Representative changes in a interpulse interval (IPI) and
b pulse duration as a function of target distance. IPI is the time interval
between the beginning of the pulse preceding the interval and the
beginning of the current pulse. The dashed line indicates the echo delay
from the target wall, which was obtained from 3D coordinated data of
the target distance
845
combined CFFM sounds ranged from 30 ms when the
bat was 5 m from the wall to 10 ms at the moment of
landing. Decreases in duration were achieved by shortening
the CF component because the durations of the initial/ter-
minal FM components remained approximately Wxed at 1
ms (Fig. 1a).
As the bat Xew along the room, the IPIs were always
longer than the ED from the target wall, even when the bat
produced pulses with shorter IPIs (ED shown by dashed
lines in Fig. 2). Consequently, the bat always waited for
echoes to return from the wall before it emitted the next
sound. This not to say that pulses did not overlap echoes,
however. Although the bat consistently shortened the dura-
tion of the CF portion of the pulses as it approached the tar-
get wall, the relation between pulse duration and the ED
(compare plots in Fig. 2) indicates that the echo from the
target wall deeply overlapped the corresponding emitted
pulse (up to 10 ms overlap duration) when the bat Xew to
within about 34 m of the wall. The short FM components
of the pulses did not overlap with the corresponding FM
components of echoes because they were only about 1 ms
in duration, but the CF components overlapped very much.
Doppler-compensation by bats during Xight
was attenuated by 3040 dB relative to CF2 at the Telem-
ciently sensitive that it recorded useable echoes from the
wall with delays in the range of approximately 0.530 ms,
corresponding to targets 0.15 m away from the bat. Ech-
oes from the target wall located directly in the Xight path of
the bat were clearly separated in the spectrograms from
their corresponding pulses and from other echoes (Fig. 1b).
Figure 3 shows representative spectrograms with the verti-
cal frequency axis expanded around frequencies of the CF2
components in echolocation sounds continuously recorded
during the bats Xight. Here, the bat Xew along the room to
land on the wall (L) and then took Xight again to circle two
times (U-turns) before landing on the wall a second time.
During these episodes of Xight, the Telemike recorded not
only the bats emitted sounds but also echoes reaching the
bat from the wall immediately toward the bats front. These
spectrograms clearly show that the bat dynamically
changed the CF2 frequency of the emitted pulse downward
by 1.5 kHz from the RF, while the CF2 frequency in the
echoes remained relatively stable. The bats Xight would
have produced upward Doppler shifts in the echoes that the
controlled reductions in emitted CF2 frequency compen-
sated. The Telemike thus successfully registered the conse-
123
119
846 J Comp Physiol A (2008) 194:841851

Fig. 3 Sonogram of the second 74

harmonic components of echolo-
72
cation sound of R. ferrumequi-
num nippon, recorded using a 70
Telemike during a continuous
68
Xight in the chamber. Top 13 s
sequence of pulse-echo pairs. U 66
and L indicate U-turn and land-

Frequency [kHz]
ing events, respectively. Bottom 64
magniWed view of pulse-echo
pairs before landing. Echoes
were observed following an
74
emitted pulse, and the frequency
of the echo received was shifted 72
slightly upward from that of the
70
pulse
68

66

64

quences of the bats DSC responses that occurred
throughout the entire sequence of Xights and landings
shown in Fig. 3. The CF2 frequencies of echoes that the bat
listened to during its Xight, including the landings and the
U-turns, were stabilized by the bats DSC to a frequency
near 69 kHz.
For quantitative analysis, we identiWed echoes returning
from the target wall by measuring EDs from the spectro-
grams (Methods), and then we measured the CF2 fre-
quencies of pulse-echo pairs emitted by the bat while it
intended to land. The spectrogram in Fig. 4a illustrates the
bats DSC response to the target wall during one Xight. In
this typical example, the bat smoothly varied the pulse CF2
frequency as it turned and approached the wall for landing.
Because echoes from the wall are clearly visible in this
Telemike recording, the resulting stability of echo CF2 fre-
quency is evident. As shown by the graph in Fig. 4b, the bat
adjusted the pulse CF2 frequency depending on its Xight
speed toward the wall so that the CF2 frequency of the ech-
oes from the target wall was maintained near the RF. As
further conWrmation of a result that has been shown in
many previous studies (Schnitzler 1968; Schuller et al.
1974; Gaioni et al. 1990), Xying bats adjusted the CF2 fre-
quency of echoes to a value slightly higher than the RF
measured separately (Methods) (Students t test, P < 0.05).
The estimated undercompensation (the frequency diVerence
between the echo CF2 and the RF of each bat) was 56.4
83.2 Hz (mean SD) for all Xight sessions for four bats.
Fully 50% of the observed echoes from the target wall were
maintained between 10 and 110 Hz above the RF; for
these echoes, the mean value was 59.3 Hz (Fig. 4c). The
frequency regulation thus was within 0.1% of the RF of
R. ferrumequinum nippon (
f/RF = 0.0593/69 kHz), indi-
cating that the compensation depth achieved by free-Xying
horseshoe bats approached very close to 100%.
As an essential aspect of DSC, successive pulses emitted
by the bat during Xight changed up or down in CF2 fre-
quency to keep the overall frequency proWle roughly Xat
(Fig. 4a). Figure 4d shows how CF2 frequency changed
between successive pulses. The maximum pulse-to-pulse
increment in CF2 varied from 281 to 422 Hz among four
bats, and the maximum decrement varied from 234 to
375 Hz. The mean increment and decrement were 61.0
and 83.2 Hz, respectively.
Parallel compensation for frequency and amplitude
In addition to changing the emitted CF2, the bats gradually
decreased the pulse amplitude as they approached the target
wall (Fig. 5a). This decrease began at a target distance
between 2 and 3 m before landing. In the search phase, the
maximum pulse amplitude was approximately 130 dB SPL
(re 20 Pa) peak to peak, measured at the Telemike above
the bats head. Overall, the pulse amplitude was reduced
progressively by approximately 40 dB before landing
(Fig. 5b). The reduction rate was 6.38.2 dB (average 7.6
dB) per halving of the target distance for the four bats in the
last 2 m of distance. As a consequence of decreasing pulse
amplitude, the amplitude of echoes returning from the tar-
get wall to the bat changed by 35.6 8.1 dB (mean
SD). Fully 50% of the observed echoes from the target wall
were stabilized at amplitudes between 30.1 and 41.2 dB
weaker than the pulse (Fig. 5c). Thus, during Xight towards
the wall, the bats compensated for increases in echo ampli-
tude as the target range became shorter by decreasing the
amplitude of their pulses. This behavioral adaptation is
interpreted as echo-intensity compensation (Kobler et al.
1985; Gaioni et al. 1990).
Figure 6a shows the combined frequency (DSC) and
amplitude compensation during a single landing Xight.

123

120
J Comp Physiol A (2008) 194:841851 847

Fig. 4 Characteristics of pulses A

and echoes as a function of the
target distance during the land-
ing Xight. a Spectrogram of the 73
sequence of pulse-echo pairs in

Frequency [kHz]
the landing Xight. L indicates the 71
landing point. b CF2 frequencies
of pulse-echo pairs as a function 69
of target distance. Open and sol-
id circles indicate the CF2 fre-
67
quency of emitted pulse and
returning echo from the target
wall. A dashed line indicates the 65
resting frequency (RF). A solid
line indicates echo CF2 fre- 63
quency which, when DSC was
absent, was estimated from the
velocity of the bat. c The box B C
71.5 200
plot of CF2 frequency of the
returning echo from the target
wall during the landing Xight. 71.0
Data were taken from all Xight
sessions for all four bats (33 70.5

Normalized echo CF2 [Hz]

100
Xights, n = 958) and the value
CF2 frequency [kHz]

was normalized to the resting 70.0

frequency of each bat. The bot-
tom and top boundaries of the 69.5
box indicate 25 and 75% of the
distribution of the data. The 0
69.0
whiskers indicate the 10th and
90th percentiles, and the lower
and upper dots indicate 5th and 68.5
95th percentiles. The dashed and
solid lines within the box repre- 68.0 -100
sent the mean (56.4 83.2 Hz;
mean SD) and median (59.3 67.5
Hz) values, respectively. d Fre-
quency change between succes- 67.0
sive pulses during the landing -200
Xight. Data were taken from all 5 4 3 2 1 0
Xight sessions for four bats (33 Target range (m)
Xights) that occurred while the
bats were Xying within 5 m of D
the target distance. Left Histo- 700
gram of all pulses (n = 1,415).
Right The frequency change 600
Number of pulses

when the bat decreased (top,

500
n = 377) or increased the pulse
CF2 frequency (bottom, Decreasing
400
n = 1,038). All values outside
the 10th and 90th percentiles 300
are plotted as dots
200
Increasing
100

0
-400 -200 0 200 400 -400 -200 0 200 400
Frequency change of successive pulses [Hz]

Using these two pulse compensation mechanisms in
tandem (blue data-points), R. ferrumequinum nippon received
echoes with both stabilized frequencies and stabilized
amplitudes (red data-points). Figure 6b shows the distribu-
tions of pulse CF2 frequencies and pulse amplitudes from
all analyzed Xights of four bats (blue data-points). The bats
received echoes from their intended target (i.e. the landing
wall) that were kept within a certain limited amplitude
range (35.6 8.1 dB of the maximum pulse amplitude).
As a result of their combined compensation responses, the

123

121
848 J Comp Physiol A (2008) 194:841851

B
Fig. 5 Amplitude change in pulse-echo pairs by bats intending to land
on the target wall. a Three-dimensional spatio-temporal reconstruc- Pulse
0
tions of echolocation behavior during the landing Xight. Coordinate
grids (1 m2) show the dimensions of the Xight chamber (8 3 2 m).
Relative sound pressure (dB)

Echolocation pulses recorded using the Telemike are shown alongside -10
the Xight trajectory. b Changes in the sound pressure of pulse-echo
pairs. The values were normalized relative to the maximum pulse -20
amplitude of each Xight. c Echo attenuation during the landing Xight.
Data were pooled from all Xight sessions for all four bats
-30

bats received echoes that were regulated within a narrow

zone of frequencies (horizontal spread of red data-points in -40

Fig. 6b) and a narrow zone of amplitudes (vertical spread of

-50 Hr*\
red data-points in Fig. 6b). F$rF$
Echo
-60
-2000 -1500 -1000 -500 0 500
Discussion Normalized CF (Hz)

Fig. 6 Concurrent compensations for CF2 frequency (blue) and sound

Doppler-shift compensation
In the Rhinolophidae, Hipposideridae and Mormoopidae,
DSC behaviors have been well demonstrated through the
use of various techniques to induce DSC [e.g. recording
sounds with a distant Wxed microphone during landing
Xight in a laboratory (Schnitzler 1968; Gustafson and
Schnitzler 1979; Tian and Schnitzler 1997), presenting
electronically frequency-shifted playback sounds to a sta-
tionary bat (Schuller et al. 1974; Simmons 1974; Metzner
et al. 2002; Smotherman and Metzner 2003), and swinging
bats on a pendulum toward a large target wall to allow the
bat to experience quasi Xight (Henson et al. 1982; Gaioni
et al. 1990)]. Although quantitative analyses of the fre-
quency change of pulse and echo were conducted in these
DSC studies, in Xight experiments, measurements derived
intensity (red) by a Xying bat. a Changes in CF2 frequency and ampli-
tude of pulse-echo pairs. The green dashed line indicates the mean val-
ues of echo frequency and amplitude. b Distributions of the CF2
frequency and the amplitude from all Xight sessions of all four bats.
One typical Xight (corresponding to the data shown in a) is represented
by large solid circles and lines. The CF2 frequency of the echo is plot-
ted as normalized to the resting frequency of the bat
from the recorded sounds required several corrections to
eliminate Xight-induced errors due to Doppler shifts for fre-
quencies and due to the directional properties of the record-
ing microphone and the emitted pulse as well as
propagation losses in the air for amplitudes. To directly
record the pulses of Xying bats, Henson and his colleagues
employed a radio telemetry technique (Henson et al. 1987;
Lancaster et al. 1992). This innovative technique allowed
them to use a microphone carried on the Xying bat to moni-

123

122
J Comp Physiol A (2008) 194:841851
tor emitted pulses. Here, we have recorded not only the
pulses but also the returning echoes that the bats actually
listen to during Xight. By this means, we demonstrated that
the CF2 frequency of returning echoes was compensated to
within 83 Hz during DSC induced by natural Xight. Fre-
quency control of vocalization from pulse-to-pulse on aver-
age was 61.0 Hz for increments and 83.2 Hz for
decrement (Fig. 4d). Because the mean of the SDs of RF in
our four horseshoe bats was 90 Hz (Table 1), which
corresponds to the SDs of RF obtained previously from
R. ferrumequinum (e.g. 30120 Hz in Schuller et al. 1974),
the Xying bats compensated echo CF2 frequency with an
accuracy of regulation equivalent to bats at rest.
The Telemike recordings showed that the echo CF2 fre-
quency reaching to the bat was slightly but signiWcantly
higher than the RF itself (Fig. 4c). In previous DSC studies,
the estimated echo CF2 frequency also was found to be
maintained at a frequency that was slightly above the RF
(the frequency of compensation during active Xight has
been called the reference frequency). This diVerence
between resting and Xying bats indicates the occurrence of
systematic slight undercompensation by Xying CFFM bats
during DSC (Schnitzler 1968; Schuller et al. 1974). The
compensated frequency diVerence between resting and
Xying bats varies among bat species and studies [e.g.
50300 Hz for R. ferrumequinum at an RF of 8181.4 kHz
(Schuller et al. 1974); 110250 Hz for R. rouxi at an RF of
7279 kHz (Neumann and Schuller 1991) and 50250 for
Pteronotus parnellii parnellii at an RF of 61 kHz (Gaioni
et al. 1990)]. With regard to the frequency diVerence
between resting and Xying bats, physiological studies have
detected frequency diVerences between a preferred fre-
quency in the auditory system and the RF. For example, the
best frequency (BF) at the lowest threshold of cochlear
microphonic (CM) audiogram in mustached bats is 200 Hz
above the RF (Henson et al. 1982). The BFs of the common
type of audiovocal neurons in the paralemniscal tegmen-
tum, which are relevant to motor control for DSC, were not
found for frequency values between the RF and 150 Hz
above (Metzner 1993). Frequency diVerences found in bats
from diVerent geographic locations further indicate that
undercompensation in Xying bats is real. In R. ferrumequi-
num nippon, the RF is 6569 KHz, which is considerably
lower than the 80 kHz RF of European subspecies, and
the frequency of the lowest threshold of behavioral audio-
gram was reported to be 0.6 KHz higher than the RF
(Taniguchi 1985). The Telemike recordings obtained here
showed the gap between RF and the compensation value
for Xying bats to be only about 60 Hz. RFs of CFFM bats
change daily (Schuller et al. 1974; Suga et al. 1987; Hiryu
et al. 2006) and also instantaneously because of increase in
body temperature (HuVman and Henson 1993a, b) and
Xight activities (Henson et al. 1990). For this reason, we
849
employed the RF measured using the Telemike on the bat
just after each landing trial. The immediacy of the RF mea-
surements following Xights may result in the small fre-
quency gap observed. We suppose that plasticity in the
auditory system relevant to DSC function may enable such
Wne regulation of echo CF2 frequency because there are
both short- and long-term intra-individual variations of RF.
In a previous study using the same landing-Xight proce-
dure with the Telemike, we found patterns of alternation in
pulse CF2 frequencies between higher and lower values in
Taiwanese leaf-nosed bats (Hipposideros terasensis)
(Hiryu et al. 2005). These frequency diVerences were
approximately equivalent to DSC changes required for tar-
gets located in diVerent directionslower emitted frequen-
cies for the higher Doppler shifts from a target located to
the bats front, and higher emitted frequencies for the lower
Doppler shifts from a target oV to the side. The occurrence
of such large diVerences in the DSC response suggest that
the bats could shift their attention between echoes from the
front to the side and back again while DSC is taking place.
By rapidly focussing and shifting their DSC responses on
diVerent echoes (front and side target echoes), the bats may
segregate echoes from diVerent targets and process them as
multiple auditory streams. Such streaming may enabling
the bats to perceive several targets in real time. Some of the
recordings of R. ferrumequinum nippon in this study appear
to show alternating patterns in pulse CF2 frequency (low
and high) as well as duration and IPI (Fig. 2), but we do not
yet have strong evidence for periodic attention change as in
H. terasensis. One possible reason is that, during Xight, R.
ferrumequinum emit long pulses (from 30 to 10 ms). In
our measurements of the CF component, we recognized
only an intense front target echo for each pulse which con-
spicuously overlapped with a large portion of the corre-
sponding emitted pulse. In contrast, the pulse duration of H.
terasensis is shorter, ranging from only 5 to 10 ms. For
each pulse, two separate echoes are typically seen in the
Telemike data of H. terasensis while the bat is Xying
toward the target wall for landing; one is the echo from the
front target wall and the other is the echo from the side wall
(or the Xoor) of the Xight chamber. These separate echoes
will have very diVerent Doppler shifts and thus may attract
the bats attention at diVerent times, causing a shift in the
bats attention between echoes and corresponding shifts in
the DSC response itself. There also may be diVerences in
the echolocation strategy of these two CFFM bat species
in the context of DSC behavior.
Optimization of successive echoes in the task of target
range estimation by echo-intensity compensation
Echo-intensity compensation allows the bat to receive ech-
oes from targets (i.e. the target wall) with amplitudes that
123
123
850
are kept within a certain optimal sound pressure range
throughout the Xight. Receiving echoes with a stable ampli-
tude, and perhaps thus also signal-to-noise ratio, could help
the bats to sustain consistent analysis of successive echoes
without distortions of perception caused by amplitude
changes. The auditory system of CFFM bats is precisely
tuned to a narrow frequency band (Neuweiler 1970;
Ostwald 1984; Suga 1984; Riquimaroux et al. 1991). Bats
can detect and analyze small frequency diVerence induced
by the Xuttering of insect wings in this band (Schnitzler and
Henson 1980), and they adjust echoes to fall within this
band. It is widely understood that Doppler-shift frequency
compensation maintains a constant carrier frequency in the
echoes from Xuttering insects which facilitates detection of
insect prey by CFFM bats. The very narrow zone of fre-
quencies to which CF2 is adjusted form an acoustic fovea
where Xutter-detecting mechanisms are present. However,
the occurrence of DSC by bats approaching the landing
wall suggests that DSC may be important not only for
detection of Xuttering insects, but also for perceiving sur-
rounding objects especially for target range estimation by
the FM component attached to the CF component. Physio-
logical experiments have shown that the delay tuning of
FMFM neurons in mustached bats is aVected by echo
amplitude, suggesting that amplitude compensation acts to
stabilize range estimation (Edamatsu and Suga 1993).
Combined frequency and amplitude compensation may be
used for optimization of successive echoes in the task of
target range estimation to control approach and landing.
These combined compensations thus may serve more for
improving target range accuracy with the FM components
of echoes than for any involvement with velocity percep-
tion itself.
Adjustments in pulse amplitude occurs not only in
CFFM bats, but also in FM bats (Hiryu et al. 2007) and
echolocating dolphins (Au and Benoit-Bird 2003; Li et al.
2006). In these wideband echolocating species, perception
of objects is based on target range from ED, not perception
of Xutter from echo Doppler-shifts. We suggest that com-
pensation mechanisms may underlie the fundamental
design of biosonar systems. In contrast to the compensation
behavior of biosonar in animals, man-made sonar systems
generally are designed to transmit sonar sounds with a Wxed
frequency and intensity, and to obtain target information by
measuring deviations in frequency and amplitude in con-
junction with ED. CFFM bats adjust their call frequency
and amplitude together to maintain the echo frequency and
amplitude within a stabilized range. As a result, CF/FM
biosonar systems do not need to provide wide-frequency
and dynamic receiving ranges; instead they analyze the
echo frequency in a narrow range (the acoustic fovea) in
the auditory system, which may greatly reduce computa-
tional eVorts in the brain. Information gathered from Xying
123
J Comp Physiol A (2008) 194:841851
bats thus bring a novel adaptive perspective to biomimetic
sonar design.
Acknowledgments We thank Prof. James A. Simmons for careful
reading of the manuscript and valuable comments. We also thank T.
Hagino, M. Fukuda, E. Fujioka, M. Omura and Y. Osawa for analysis
and technical support during this experiment; N. Urano for assistance
in capturing bats in the Weld. The experiments complied with the Prin-
ciples of Animal Care, publication no. 8623, revised in 1985, of the
National Institutes of Health, and the procedures were approved by the
animal care committee of Doshisha University. This work was partly
supported by a grant to the Research Center for Advanced Science and
Technology (RCAST) at Doshisha University from the Ministry of
Education, Culture, Sports, Science, and Technology (MEXT) of Japan:
Special Research Grants for the Development of Characteristic Education
from the Promotion and Mutual Aid Corporation for Private Schools of
Japan and the Innovative Cluster Creation Project.
References
Au WWL, Benoit-Bird KJ (2003) Automatic gain control in the echo-
location system of dolphins. Nature 423:861863
Boonman A, Jones G (2002) Intensity control during target approach in
echolocating bats; stereotypical sensorymotor behaviour in Dau-
bentons bats, Myotis daubentonii. J Exp Biol 205:28652874
Edamatsu H, Suga N (1993) DiVerences in response properties of neu-
rons between two delay-tuned areas in the auditory cortex of the
mustached bat. J Neurophysiol 69:17001712
Gaioni SJ, Riquimaroux H, Suga N (1990) Biosonar behavior of
mustached bats swung on a pendulum prior to cortical ablation.
J Neurophysiol 64:18011817
GriYn DR (1958) Listening in the dark. Yale University Press,
New Haven
Gustafson Y, Schnitzler HU (1979) Echolocation and obstacle avoid-
ance in the hipposiderid bat Assellia tridens. J Comp Physiol A
131:161167
Hartley DJ, Campbell KA, Suthers RA (1989) The acoustic behavior
of the Wsh-catching bat, Noctilio leporinus, during pre-capture.
J Acoust Soc Am 86:827
Henson OW Jr, Pollak GD, Kobler JB, Henson MM, Goldman LJ
(1982) Cochlear microphonic potentials elicited by biosonar sig-
nals in Xying bats, Pteronotus p. parnellii. Hear Res 7:127147
Henson OW Jr, Bishop AL, Keating AW, Kobler JB, Henson MM,
Wilson BS, Hansen R (1987) Bisonar imaging of insects by Pter-
onotus p. parnellii, the mustached bat. Nat Geor Res 3:82101
Henson OW, Koplas PA, Keating AW, HuVman RF, Henson MM
(1990) Cochlear resonance in the mustached bat: behavioral adap-
tations. Hear Res 50:259273
Hiryu S, Katsura K, Lin LK, Riquimaroux H, Watanabe Y (2005)
Doppler-shift compensation in the Taiwanese leaf-nosed bat
(Hipposideros terasensis) recorded with a telemetry microphone
system during Xight. J Acoust Soc Am 118:39273933
Hiryu S, Katsura K, Nagato T, Yamazaki H, Lin LK, Watanabe Y,
Riquimaroux H (2006) Intra-individual variation in the vocalized
frequency of the Taiwanese leaf-nosed bat, Hipposideros terasen-
sis, inXuenced by conspeciWc colony members. J Comp Physiol A
192:807815
Hiryu S, Hagino T, Riquimaroux H, Watanabe Y (2007) Echo-inten-
sity compensation in echolocating bats (Pipistrellus abramus)
during Xight measured by a telemetry microphone. J Acoust Soc
Am 121:17491757
HuVman RF, Henson OW Jr (1993a) Labile cochlear tuning in the
mustached bat. I. Concomitant shifts in biosonar emission fre-
quency. J Comp Physiol A 171:725734
124
J Comp Physiol A (2008) 194:841851
HuVman RF, Henson OW Jr (1993b) Labile cochlear tuning in the
mustached bat. II. Concomitant shifts in neural tuning. J Comp
Physiol A 171:735748
Jen PH, Kamada T (1982) Analysis of orientation signals emitted by
the CFFM bat, Pteronotus p. parnellii and the FM bat, Eptesicus
fuscus during avoidance of moving and stationary obstacles.
J Comp Physiol A 148:389398
Kick SA, Simmons JA (1984) Automatic gain control in the bats sonar
receiver and the neuroethology of echolocation. J Neurosci
4:27252737
Kobler JB, Wilson BS, Henson OW Jr, Bishop AL (1985) Echo inten-
sity compensation by echolocating bats. Hear Res 20:99108
Lancaster WC, Keating AW, Henson OW Jr (1992) Ultrasonic vocal-
izations of Xying bats monitored by radiotelemetry. J Exp Biol
173:4358
Li S, Wang D, Wang K, Akamatsu T (2006) Sonar gain control in
echolocating Wnless porpoises (Neophocaena phocaenoides) in an
open water. J Acoust Soc Am 120(4):18031806
Metzner W (1993) An audio-vocal interface in echolocating horseshoe
bats. J Neurosci 13:18991915
Metzner W, Zhang S, Smotherman M (2002) Doppler-shift compensa-
tion behavior in horseshoe bats revisited: auditory feedback con-
trols both a decrease and an increase in call frequency. J Exp Biol
205:16071616
Moss CF, Surlykke A (2001) Auditory scene analysis by echolocation
in bats. J Acoust Soc Am 110:22072226
Neumann I, Schuller G (1991) Spectral and temporal gating mecha-
nisms enhance the clutter rejection in the echolocating bat,
Rhinolophus rouxi. J Comp Physiol A 169:109116
Neuweiler G (1970) Neurophysilogische Untersuchungen zum
Echoortungssystem der groben Hufeisennase Rhinolophus
ferrum equinum Schreber, 1774. Z Vergl Physiol 67:273306
Neuweiler G (2000) The biology of bats. Oxford University Press,
New York
Ostwald J (1984) Tonotopical organization and pure tone response
characteristics of single units in the auditory cortex of the greater
horseshoe bat. J Comp Physiol A 155:821834
Riquimaroux H, Watanabe Y (2000) Characteristics of bat sonar
sounds recorded by a telemetry system and a Wxed ground micro-
phone. Seventh Western PaciWc Regional Acoustics Conference
(WESTPRACVII):233238
851
Riquimaroux H, Gaioni SJ, Suga N (1991) Cortical computational
maps control auditory perception. Science 251:565568
Schnitzler HU (1968) Die Ultraschallortungslaute der Hufeisen-Fle-
dermuse (Chiroptera-Rhinolophidae) in verschiedenen Orient-
ierungssituationen [The ultrasonic sounds of horseshoe bats
(Chiroptera-Rhinolophidae) in diVerent orientation situations].
Z Vergl Physiol 57:376408
Schnitzler HU, Henson OW Jr (1980) Performance of airborne animal
sonar system, I. Microchiroptera. In: Busnel R-G, James FF (eds)
Animal sonar systems. Plenum Press, New York, pp 109181
Schuller G, Beuter K, Schnitzler HU (1974) Response to frequency
shifted artiWcial echoes in the bat Rhinolophus ferrumequinum.
J Comp Physiol A 89:275286
Simmons JA (1974) Response of the Doppler echolocation system
in the bat, Rhinolophus ferrumequinum. J Acoust Soc Am
56:672682
Smotherman M, Metzner W (2003) Fine control of call frequency by
horseshoe bats. J Comp Physiol A 189:435446
Suga N (1984) The extent to which biosonar information is represented
in the bat auditory cortex. In: Edelman GM, Gall WE, Cowan
WM (eds) Dynamic aspects of neocortical function. Wiley,
New York, pp 315373
Suga N, Jen PH (1976) Disproportionate tonotopic representation for
processing CFFM sonar signals in the mustache bat auditory
cortex. Science 194:542544
Suga N, Niwa H, Taniguchi I, Margoliash D (1987) The personalized
auditory cortex of the mustached bat: adaptation for echolocation.
J Neurophysiol 58:643654
Taniguchi I (1985) Echolocation sounds and hearing of the greater
Japanese horseshoe bat (Rhinolophus ferrumequinum nippon).
J Comp Physiol A 156:185188
Thomas JA, Moss CF, Vater M (2003) Echolocation in bats and
dolphins. University of Chicago Press, Chicago
Tian B, Schnitzler HU (1997) Echolocation signals of the greater
horseshoe bat (Rhinolophus ferrumequinum) in transfer Xight and
during landing. J Acoust Soc Am 101:23472364
Vogler B, Neuweiler G (1983) Echolocation in the noctule (Nyctalus
noctula) and horseshoe bat (Rhinolophus ferrumequinum).
J Comp Physiol A 152:421432
123
125


126
Echo-intensity compensation in echolocating bats (Pipistrellus
abramus) during flight measured by a telemetry microphone
Shizuko Hiryua
Research Center for Intelligent Information Science, Doshisha University, Kyotanabe 610-0321, Japan

Tomotaka Hagino
Department of Electrical Engineering, Doshisha University, Kyotanabe 610-0321, Japan

Hiroshi Riquimaroux
Department of Intelligent Information Engineering and Sciences, Doshisha University, Kyotanabe 610-0321,
Japan and Bio-navigation Research Center, Doshisha University, Kyotanabe 6100321, Japan

Yoshiaki Watanabe
Department of Electrical Engineering, Doshisha University, Kyotanabe 610-0321, Japan
and Bio-navigation Research Center, Doshisha University, Kyotanabe 610-0321, Japan

Received 3 October 2006; revised 20 November 2006; accepted 8 December 2006

An onboard microphone Telemike was developed to examine changes in the basic characteristics
of echolocation sounds of small frequency-modulated echolocating bats, Pipistrellus abramus.
Using a dual high-speed video camera system, spatiotemporal observations of echolocation
characteristics were conducted on bats during a landing flight task in the laboratory. The Telemike
allowed us to observe emitted pulses and returning echoes to which the flying bats listened during
flight, and the acoustic parameters could be precisely measured without traditional problems such as
the directional properties of the recording microphone and the emitted pulse, or traveling loss of the
sound in the air. Pulse intensity in bats intending to land exhibited a marked decrease by 30 dB
within 2 m of the target wall, and the reduction rate was approximately 6.5 dB per halving of
distance. The intensity of echoes returning from the target wall indicated a nearly constant intensity
42.6 5.5 dB weaker than the pulse emitted in search phase within a target distance of 2 m.
These findings provide direct evidence that bats adjust pulse intensity to compensate for changes in
echo intensity to maintain a constant intensity of the echo returned from the approaching target at
an optimal range. 2007 Acoustical Society of America. DOI: 10.1121/1.2431337
PACS numbers: 43.80.Ka JAS Pages: 17491757

I. INTRODUCTION
Bats Microchiroptera possess a highly developed sonar
system. As biosonar animals, they are capable of echoloca-
tion that recognizes the physical attributes of their environ-
ment with great accuracy by comparing emitted pulses to
returning echoes Griffin, 1958. Echolocation pulses can
differ markedly in structure, consisting of either only short
downward frequency-modulated FM pulses or long
constant-frequency CF pulses followed by a short FM com-
ponent. Such variation in the structure of echolocation pulses
is thought to reflect the adaptations of different species to the
constraints imposed by their foraging behavior Simmons
and Stein, 1980; Neuweiler, 1984.
Rhinolophids, Hipposiderids, and a Mormoopid the
mustached bat, Pteronotus parnellii all use a compound
CF-FM pulse. As a bat approaches a target, the frequency of
the returning echo is Doppler-shifted depending on the ve-
locity of the bat in relation to the target. These bats compen-
sate for the Doppler shifts by changing their pulse frequen-
cies so that the echo frequencies remain constant and are
precisely analyzed within the narrowly tuned frequency
a
Electronic mail: shiryu@mail.doshisha.ac.jp
range in which the bats can hear best. This behavior is
termed Doppler-shift compensation DSC and is an im-
portant behavioral adaptation for echolocation in CF-FM
bats Schnitzler, 1968; Schuller et al., 1974; Simmons, 1974;
Gustafson and Schnitzler, 1979; Trappe and Schnitzler, 1982;
Gaioni et al., 1990; Lancaster et al., 1992; Keating et al.,
1994.
During flight, bats decrease the intensity of their emitted
pulses when they approach a prey item or an obstacle e.g.,
Griffin, 1958; Jen and Kamada, 1982; Vogler and Neuweiler,
1983; Kick and Simmons, 1984; Hartley et al., 1989; Tian
and Schnitzler, 1997; Boonman and Jones, 2002. Kobler and
colleagues 1985 have shown that a bat attached to a pen-
dulum and swung toward a large fixed target decreases the
intensity of its emitted pulse during the forward swing and
increases the intensity during the backward swing. This be-
havior indicates the possibility of another compensation
mechanism, in which pulse intensity is adjusted in relation to
the distance to a target, resulting in maintenance of echo
intensity within the optimal sensitivity range echo-intensity
compensation. Subsequently, reduction of pulse intensity
has been quantitatively investigated in bats given a flight task
Hartley et al., 1989; Waters and Jones, 1995; Tian and
Schnitzler, 1997; Boonman and Jones, 2002 and using a

J. Acoust. Soc. Am. 121 3, March 2007 0001-4966/2007/1213/1749/9/$23.00 2007 Acoustical Society of America 1749
127
pendulum device or a moving-target apparatus to elicit and
measure changes in pulse intensity Gaioni et al., 1990; Hart-
ley, 1992b. Adjustment of pulse intensity by bats has been
discussed in a number of studies as being related to echo-
intensity compensation e.g., Kick and Simmons, 1984; Hart-
ley et al., 1989; Hartley, 1992a; b; Simmons et al., 1992;
Waters and Jones, 1995; Tian and Schnitzler, 1997; Boon-
man and Jones, 2002. Recently, Au and Benoit-Bird 2003
demonstrated that free-ranging dolphins in the wild changed
the amplitude of emitted echolocation signals depending on
the range of the target to compensate for propagating loss in
sound. Therefore, echo-intensity compensation may be effec-
tive in the biosonar systems of both bats and dolphins. How-
ever, accurate measurement of the intensity of directional
pulses emitted by biosonar animals is still considered diffi-
cult using the traditional microphone system. For quantita-
tive analysis of the pulse intensity which the flying bats ac-
tually emitted, recorded pulses required several corrections
for compensating the directional properties of the recording
microphone and the emitted pulse, and also traveling loss of
sound in the air depending on the distance between the bat
and the microphone e.g., Hartley et al., 1989; Tian and
Schnitzler, 1997; Boonman and Jones, 2002. Given these
constraints, the pulse intensity measured by the traditional
microphone potentially confounded interpretation of a bats
adjustment of pulse intensity in relation to its spatiotemporal
behavior during echolocation. Furthermore, measuring the
intensity of echoes as well as emitted pulses is necessary to
reveal any echo-intensity compensation behavior employed
by the animals. However, the relationship between the pulse
and returning echo intensities has not been experimentally
investigated in detail for bats during flight.
In this study, we examined echolocation behavior in
Japanese house bats Pipistrellus abramus, an FM bat given
a flight landing task in a laboratory. Echolocation sounds
were recorded by a telemetry microphone Telemike
mounted on the head of the bat Henson et al., 1987; Lan-
caster et al., 1992; Riquimaroux and Watanabe, 2000; Hiryu
et al., 2005. Since well-studied FM bat species are small
e.g., P. abramus weighs 5 8 g in relation to the weight of
telemetry microphones, sound recording using telemetry mi-
crophones has never been used to investigate echolocation in
flying FM bats. We developed a small Telemike that was
light enough to be carried by the animals, and combined the
Telemike with a high-speed video camera system in order to
examine the relationship between the echolocation sounds
and the spatiotemporal behavior of bats during flight.
Bats are assumed to adjust their call parameters by feed-
back control in response to the attended returning echoes.
Therefore, it is particularly important to investigate the be-
havioral response call parameters of bats associated with
their corresponding acoustic stimuli echoes under the spa-
tiotemporal conditions in which the pulse-echo pairs were
produced. Since P. abramus emit short downward FM pulses
with a mean duration of approximately 1 ms, the echoes ar-
riving from targets could be recorded by the microphone
above the head without overlapping with the outgoing
pulses. We analyzed the signal characteristics of pulse-echo
pairs to which the flying bats actually listened, and investi-
1750 J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007
gated adjustments of call parameters in the context of echo-
intensity compensation.
II. MATERIALS AND METHODS
A. Subjects
Three adult Japanese house bats P. abramus were used
in this study. The animals were captured from a large colony
roosting in bridge girders near the campus of Doshisha Uni-
versity, Japan. Pipistrellus abramus is a member of the Ves-
pertilionidae and is commonly found in Japan. Although
closely related pipistrelle bats such as P. pipistrellus and P.
kuhli have been studied extensively e.g., Schnitzler et al.,
1987; Kalko, 1995; Waters and Jones, 1995; Holderied and
Helversen, 2003, the echolocation characteristics of P. abra-
mus are not well known. The body mass of P. abramus
ranges from 5 to 8 g, and the wingspan measures approxi-
mately 10 cm. Echolocation pulses are emitted through the
mouth. The bats were kept in a rearing cage of 0.9 m L
0.9 m W 0.6 m H, and were allowed free access to
food mealworms and water. On alternate days, the bats
were allowed to fly freely for several hours in a large flight
room. The experiments complied with the Principles of Ani-
mal Care, publication no. 86-23, revised 1985, of the Na-
tional Institutes of Health, and with current Japanese laws.
B. General experimental procedure
All experiments were conducted in a flight chamber
measuring 8 m L 3 m W 2 m H under long wave-
length lighting with red filters 650 nm to avoid optical
effects Hope and Bhatnagar, 1979; Ghose and Moss, 2003.
The chamber was made of steel plates to minimize interfer-
ence from external electromagnetic waves. All internal walls
were painted in black. The bats were released at one end of
the flight chamber and allowed to fly freely to the opposite
end where a landing mesh 1 m W 0.7 m H was at-
tached to the wall 1.8 m above the floor. This wall is referred
to as the target wall during the landing flight task. Flight
behavior was recorded as a flying bat approached the target
wall for landing. Recording was conducted for three flight
sessions of each bat a total of nine flight sessions.
C. Telemike system
Echolocation sounds were recorded by a custom-made
telemetry microphone Telemike mounted on the back of the
bat Fig. 1. The Telemike consists of a 1/8-inch omnidirec-
tional condenser microphone Knowles, FG-3329, Illinois,
USA, an FM transmitter unit, a hearing aid battery of 1.5 V
Sony, SR421SW, Tokyo, Japan, and a transmitting antenna.
The Telemike was attached to the back of a bat with a piece
of double-sided glue tape, with the microphone positioned
approximately 1 cm above the mouth. Because the Telemike
weighed less than 0.6 g, including the battery, it was light
enough to be carried by bats weighing about 5 g. The bats
did not exhibit any fatigue during the experiments. Removal
of the Telemike from the back of a bat after each experiment
was facilitated by use of a parting agent to avoid skin irrita-
tion.
Hiryu et al.: Echo-intensity compensation in echolocating bats
128
FIG. 1. Japanese house bat P. abramus with an onboard microphone
Telemike mounted on its back. The Telemike consists of a 1/8-inch con-
denser microphone, FM transmitter, battery, and transmitting wire antenna
weight: 0.6 g, including the battery.

The Telemike transmitted signals with a carrier fre-

quency between 100 105 MHz to a wire antenna attached to
the ceiling of the flight chamber. Received signals were de-
modulated using a custom-made FM receiver. The signals
were then high-pass filtered at 20 kHz NF Corporation,
model 3625, Yokohama, Japan, digitized by a 16-bit,
384 kHz DAT recorder SONY, SIR-1000W, Tokyo, Japan,
and stored on a hard disk of a personal computer. Prior to
recording, the output of the Telemike system was calibrated
using a loudspeaker Pioneer, PT-R7, Tokyo, Japan and
Brel and Kajaer 1/8-in. microphone. The total frequency
response of the Telemike system was within 4 dB flat be-
tween 20 and 100 kHz. Since echolocation sounds were re-
corded directly by the microphone above the head of the
flying bat, acoustic parameters such as amplitude could be
measured precisely without interference from traveling loss
of sound in the air or directional properties of the emitted
pulse.

D. Three-dimensional reconstruction of
spatiotemporal echolocation behavior
The flight behavior was recorded using a dual digital
high-speed video camera system NIPPON ROPER Co.,
Ltd., CR Imager model 2000s, Chiba, Japan. Cameras were
placed at a corner of the flight chamber and did not interfere
with the bats flight path. The frame rate was 125 per second.
Three-dimensional coordinates of the flying bats were recon-
structed from these video images using a commercial motion
analysis software DITECT, Dipp-Motion 2D ver. 2.1. Prior
to recording bat flights, a reference frame with known coor-
dinates was positioned in the center of the flight chamber,
then recorded by two video cameras. The analysis software
calibrated the reconstruction system with the coordinate data
of the reference frame. Based on a direct linear transforma-
tion technique, the position of the flying bat or other object
was reconstructed from two-dimensional coordinate data in
the video images. The signal triggering the video cameras
was digitally stored using a DAT recorder so that flight co-
ordinates could be synchronized with sound data. Using
three-dimensional coordinate data, the flight trajectory of the
bat and the distance from the target wall target distance or
other objects could be determined. The echo delay t was
FIG. 2. Pulse-echo pairs recorded with the Telemike while a bat approached
the target wall. A Sonograms at a target distance of 4 m top and 1 m
bottom. Arrows indicate the echoes returning from the target wall see
text. B Changes in echo delay determined from sound data as a function
of target distance during the landing flight for two bats solid circles. Circle
size indicates relative variation in echo intensity where bigger circles des-
ignate sounds of greater intensity relative to smaller circles. The three lines
represent echo delays between the flying bat and three targetsthe target
wall, floor, and ceiling of the chambercalculated using three-dimensional
coordinate data of the bat.
calculated from the distance between the flying bat and an
object d using the formula t = 2d / c, where c is the sound
velocity in the air.
E. Sound analysis
The acoustical characteristics of echolocation sounds
were analyzed from the sonogram using a custom program of
Matlab on a personal computer. Figure 2A shows sono-
grams of typical pulse-echo pairs recorded by the Telemike

J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007 Hiryu et al.: Echo-intensity compensation in echolocating bats 1751
129
TABLE I. Signal characteristics of the echolocation pulse from P. abramus at rest. iFM: initial FM; tFM: terminal FM; BW: bandwidth.
Duration iFM tFM BW
N ms kHz kHz kHz
Bat A 100 0.90 0.14 79.3 4.71 42.7 2.73 36.6 5.72
Bat B 100 1.20 0.11 84.9 4.81 42.6 2.65 42.3 5.28
Bat C 100 1.57 0.30 83.6 4.60 46.0 2.86 37.6 5.87
when the flying bat approached the target wall. Near the
center of the chamber 4 m from the target wall, a number
of echoes referred to as an echo train were usually observed
with different echo delays the time difference between each
pulse-echo pair and intensities following an emitted pulse.
The second and higher harmonic components of the return-
ing echo were most often attenuated beyond 3 4 m from the
target wall. On the other hand, when the bat was within a
target distance of 1 and 2 m, the number of echoes reaching
the flying bat decreased. The echoes were clearly separated
from each other, and the second harmonic component ap-
peared in the sonogram. In this study, only the fundamental
component of the pulse-echo pair was analyzed. Each pulse
or echo was extracted from the sonogram, and the echo delay
was measured for all echoes observed by the Telemike. The
sound pressure level of the pulse was calculated from the
peak-to-peak amplitude voltage of the observed pulse in the
time domain. Simultaneously, the sonogram exhibited a peak
in energy at around 50 60 kHz in each sound of P. abramus
during flight. Therefore, the spectral energy at the peak en-
ergy portion of each pulse and echo was measured in the
displayed sonogram using the custom program of Matlab so
that changes in intensity of a weak echo could be quantita-
tively evaluated in relation to the pulse intensity. The maxi-
mum magnitude of the spectral energy in each sound is re-
ferred to as the sound intensity in this study.
Typical changes in the echo delay determined from
sound data are shown in Fig. 2B as a function of target
distance during landing as the bat approached the target wall.
The size of the solid circles indicates the relative variation in
echo intensity where bigger circles designate sounds of
greater intensity relative to smaller circles. In this flight ex-
periment, the Telemike allowed us to observe echoes with an
echo delay in the range of approximately 0.5 30 ms, corre-
sponding to targets 0.1 5 m away from the bat. The three
lines shown in Fig. 2B represent echo delays between the
flying bat and three different objectsthe target wall, floor,
and ceiling of the chambercalculated from the three-
dimensional coordinate data of the flying bat. By comparing
1752 J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007
Frequency at peak Min-Max SPL Mean SPL
energy kHz dB peak-to-peak dB peak-to-peak
54.5 5.91 97120 112
56.4 9.29 100117 112
57.4 8.57 99117 109
the echo delays shown in sound data solid circles, the ech-
oes from each surrounding wall of the flight chamber could
be identified in the observed sonogram. For example, the
arrows in Fig. 2A indicate echoes from the target wall.
Echoes from the floor and ceiling of the chamber consis-
tently showed intense sound pressure, as well as echoes from
the target wall directly in the flight path of the bat.
Prior to flight recording, pulses of the bat in a stationary
position were recorded using the Telemike for quantitative
analysis. The acoustical parameters of the echolocation
pulses were analyzed using the custom program of Matlab
described above. Signal bandwidth and duration were deter-
mined from the sonogram at 25 dB relative to the peak
intensity of the pulse.
III. RESULTS
A. General echolocation behavior of P. abramus
Pipistrellus abramus emits a short downward FM pulse
with maximum energy at the fundamental component. The
mean duration at rest is 1.22 0.34 ms n = 300, and the fun-
damental frequency was modulated from approximately 83
to 44 kHz by the three bats Table I. The sound pressure
level SPL of an emitted pulse at rest ranged from approxi-
mately 100 to 120 dB peak-to-peak re 20 Pa, with an
average of 111 dB at the microphone above the bats head.
The sonogram exhibited a peak in energy at 56.0 8.12 kHz,
which was 11 14 kHz higher than the terminal frequency.
For the landing experiment, a total of nine flight sessions
were recorded for three bats. In flight, the pulse duration was
elongated to 3 4 ms, and then decreased to 0.5 ms before
landing the target wall. Figure 3 shows the envelope and
frequency structure of a typical echolocation pulse emitted
by a P. abramus in flight, recorded by the Telemike at a
target distance of 3 m. The frequency of the fundamental
component was modulated exponentially from 100 to
40 kHz, and the bandwidth of the emitted pulse was ex-
tended when in flight. A prominent CF-like portion a shal-
low sweep portion at the end of the pulse; see Fig. 3 was
FIG. 3. Typical echolocation pulse emitted by P. abra-
mus during the landing flight, recorded by the Telemike.
Data were taken from sound recorded at a distance of
3 m from the target wall. Echolocation pulses usually
contained several harmonics, with the first being domi-
nant.
Hiryu et al.: Echo-intensity compensation in echolocating bats
130

observed beyond a target distance of approximately 2 m.
Flight trajectories for the three bats during landing ap-
proaches are illustrated in three-dimensional images in Fig.
4. The echolocation pulses recorded with the Telemike are
superimposed on the flight trajectory to indicate the spa-
tiotemporal characteristics of echolocation behavior. The
maximum flight speed for a direct approach to the target wall
was approximately 4 m / s at 3 4 m from the target wall.
Several common features were observed in the behavior of
the three bats, including a marked decrease in interpulse in-
terval and pulse amplitude as bats approached the target wall
approach phase, which usually started at a target distance
of between 1 and 2 m. The pulse emission rate increased
from 1020 pulses per second at 3 4 m from the target wall
to 130140 pulses per second immediately prior to landing.
The SPL of each pulse was calculated from the peak-to-peak
amplitude voltage of the observed pulse in the time domain
by the Telemike. The maximum SPL peak-to-peak at the
microphone above the bats head was approximately 130 dB
during flight, which was almost 20 dB higher than when the
bat was at rest. Before the bats started the approach phase,
the sound pressure level of emitted pulse was almost con-
stant at 130 dB peak-to-peak search phase.
B. Changes in sound intensity of the pulse-echo pair
Figure 5A shows changes in intensity in the peak en-
ergy portions of the pulse and echo as a function of target
distance for the three bats, normalized to the average of the
J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007
FIG. 4. Three-dimensional spatiotemporal reconstruc-
tions of echolocation behavior during landing flights for
three bats. Coordinate grids 1 m2 show the dimen-
sions of the flight chamber 8 3 2 m. Echolocation
pulses recorded by the Telemike are placed alongside
the flight trajectory. The arrow indicates the flight di-
rection of the bat. The amplitude of emitted pulses de-
creased as the bat approached the target wall.
pulse intensity during the search phase. Data were taken
from three flight sessions of each bat. Intensities of all ech-
oes that could be extracted from the displayed sonograms are
cumulatively plotted with open circles in Fig. 5A.
The pulse intensity in bats intending to land started to
decrease at a distance of between 1 and 2 m, and the reduc-
tion was approximately 30 dB before landing. On the other
hand, the observed echo intensities were mainly between 40
to 50 dB weaker than the pulse but remained almost constant
as the target distance decreased open circles in Fig. 5A.
The three bats decreased the intensity of their emitted pulses
logarithmically with target distance Fig. 5B. The reduc-
tion rate within 2 m of the target distance was 5.6 7.5 dB
per halving of the target distance for the three bats. On av-
erage, the reduction rate of pulse intensity was 6.5 dB
21.6 dB decrease per decade of the target distance.
Figure 5C shows the distribution of all observed echo
intensities relative to the average of the pulse intensity dur-
ing the search phase for nine total flight sessions of three
bats. The distribution showed a single peak, with a mean
of 42.6 5.5 dB, which corresponds to approximately
80 90 dB SPL peak-to-peak. Variation in observed echo
intensity may have been due to variation in the distance or
direction of the object from which the echo returned, and/or
the direction of the emitted pulse.
The change in intensity of a pulse-echo pair for one
flight session by a bat is shown in Fig. 6. The intensity of the
echo returning from the target wall to the bat asterisks in
Hiryu et al.: Echo-intensity compensation in echolocating bats 1753
131
FIG. 5. Changes in intensity in the peak energy portion the sonogram exhibited a peak in energy at about 50 60 kHz during flight, which was approximately
10 kHz higher than the terminal frequency of the pulse-echo pairs, normalized to the average of intensity of the pulse emitted during the search phase. A
Sound intensity of the pulse solid circles and echo open circles as a function of target distance for three bats. Data were taken from three flight sessions
of each of three bats. B Decrease in pulse intensity for a total of nine flights of three bats. The solid line indicates the correlation line. Pulse intensity
decreased by 6.5 dB on halving the target distance 21.6 dB decrease per decade of the target distance. C Distribution of observed echo intensity relative
to the average of the pulse intensity during the search phase during the landing approach for nine flight sessions of three bats.

Fig. 6 indicates a nearly constant intensity, while the bat

considerably decreased the pulse intensity as it approached
the target wall.

FIG. 6. Sound intensity of pulse-echo pairs as a function of target distance
for one flight session by Bat A. Asterisks indicate echoes from the target
wall, determined from sound and three-dimensional coordinate data see
Methods. The intensity of the echo returning from the target wall indicates
a nearly constant intensity, while the bat considerably decreased the pulse
intensity as it approached the target wall.
IV. DISCUSSION
A. Echo-intensity compensation during flight
Bats are thought to adjust their pulse intensity in re-
sponse to the intensity of the returning echoes that they at-
tend to during echolocation. In this experiment, bats were
expected to attend to the echo from the target wall. The echo
intensity from the target wall asterisks in Fig. 6 recorded
directly by the Telemike was maintained at a constant level
within a target distance of 2 m, whereas the pulse intensity
considerably decreased. This indicates that the bat adjusted
its pulse intensity depending on target distance such that the
intensity of the echo returning from its destination was con-
stant. In other words, this finding suggests that the echo from
the target destination could be a prominent stimulus to ad-
justment of pulse intensity by bats and that a bat could rec-

1754 J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007 Hiryu et al.: Echo-intensity compensation in echolocating bats
132
ognize an echo from its target the target wall among a
number of echoes reaching it e.g., Fig. 2A.
A number of studies have attempted to measure pulse
intensity for bats during flight e.g., Griffin, 1958; Jen and
Kamada, 1982; Vogler and Neuweiler, 1983; Kick and Sim-
mons, 1984; Hartley et al., 1989; Waters and Jones, 1995;
Tian and Schnitzler, 1997; Boonman and Jones, 2002; Hold-
eried and Helversen, 2003. These studies reported a de-
crease in pulse intensity as a bat approached its target, using
traditional fixed microphones that did not interfere with the
flight path of the bat. In several studies, the reduction rate of
pulse intensity as a function of distance to the target has been
quantitatively estimated e.g., Hartley et al., 1989; Waters
and Jones, 1995; Tian and Schnitzler, 1997; Boonman and
Jones, 2002. For example, Tian and Schnitzler 1997 esti-
mated the intensity of emitted pulses in Rhinolophus fer-
rumequinum during a landing approach, applying careful
corrections to directional sounds recorded by the fixed mi-
crophone. Similarly, Noctilio leporinus decreased its pulse
intensity while capturing a small target mealworm at a rate
of about 6 dB on halving the target distance Hartley, 1989.
Although the reduction of emitted pulses by flying bats ap-
proaching a target varied slightly among the studies e.g.,
Hartley et al., 1989; Waters and Jones, 1995; Boonman and
Jones, 2002, these studies suggest that the bats are supposed
to adjust their pulse intensity in response to the intensity of
returning echoes.
We found that P. abramus decreased their pulse intensity
at a rate of 6.5 dB on halving the target distance and exhib-
ited nearly constant echo intensity from the target wall,
which should be attended to by a bat on its landing flight.
This finding provides direct evidence of echo-intensity com-
pensation by echolocating bats during flight to maintain the
intensity of the resulting echo within the range necessary for
optimal signal processing. The leaf-nosed bat Hipposideros
terasensis, which emits a CF-FM pulse, decreased its pulse
intensity by approximately 6 dB on halving the target dis-
tance using the same landing flight task as used here Hiryu,
2005. Recently, echo-intensity compensation was demon-
strated for the echolocation system in dolphins Au and
Benoit-Bird, 2003. Free-ranging dolphins decreased the am-
plitude of their echolocation signals as they approached the
target by 6 dB per halving of distance. These results suggest
that adjusting pulse intensity may be a common strategy of
biosonar animals approaching a target.
The systems employed to compensate for variation in
intensity and frequency of echoes are unique to biosonar ani-
mals. Extant artificial sonar systems are generally designed
to emit a pulse with a fixed frequency to measure the Dop-
pler shift in the retuning echo in order to detect the velocity
of the target. In contrast, CF-FM bats adjust their call fre-
quency to maintain an echo frequency within the range they
can hear best Schnitzler, 1968; Suga, 1984. Consequently,
the auditory system in the mustached bats can detect a shift
in echo frequency with a high degree of accuracy, as small as
50 Hz which corresponds to less than 0.1% of the CF2 fre-
quency of the pulse emitted by the mustached bats
61 kHz Suga, 1984; Riquimaroux et al., 1991. Echo-
intensity compensation allows bats to receive echoes within
J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007
a certain optimal intensity range, which may facilitate con-
sistent and precise analysis of target information by the au-
ditory system as seen in DSC Kick and Simmons, 1984.
These compensation behaviors associated with adjustments
of pulse frequency and intensity may be a fundamental strat-
egy employed by the animals for streamlining their echolo-
cations, and will contribute to inspire the design of future
artificial sonar systems or echo-sensing devices.
B. Multiple echoes reaching a bat during flight
We observed that a number of echoes from surrounding
targets reached the flying bats with different echo delays and
different intensities after a pulse emission Fig. 2. When a
bat was flying, the frequency of echoes received was
Doppler-shifted depending on flight speed. A frequency
sweep was observed with a slight difference among echoes
from different targets Fig. 2A, which may be due to dif-
ferences in the extent of the Doppler shift according to the
direction or relative velocity of the targets in relation to the
flight path of the bat. We suggest that the difference in the
frequency slope between returning echoes may provide FM
bats in flight with direction and velocity information on mul-
tiple targets.
What specific behavioral strategies do echolocating bats
use for processing multiple echoes? Bats are supposed to
adjust their call parameters by feedback control in response
to information from previous echoes. Analysis of the emitted
pulse frequency and interpulse interval shows that bats may
change the focus of their attention between echoes during
flight Hiryu et al., 2005. By periodically focusing on dif-
ferent echoes, bats may process multiple auditory streams to
perceive several targets in real time. In this flight experiment,
the bats occasionally approached the wall attached with a
landing mesh referred to as the target wall in case of the
landing scenario and then returned to the starting point with-
out landing. Interestingly, the bats during the U-turn did not
decrease the intensity of emitted pulses as approached that
wall, as seen when they were landing Fig. 7. In addition,
the observed intensity of echoes from that wall did not ap-
pear constant, showing a decrease of approximately 20 dB
within a distance of 1 m before U-turn point marked with U
in Fig. 7B. This suggests that the bats making a U-turn did
not keep their attention on that wall as seen in case of land-
ing scenario. It is likely that a bat may divert its attention to
other targets depending on its flight direction in order to
avoid colliding with surrounding walls while making a
U-turn.
In this study, we focused on temporal changes in the
sound intensity of pulse-echo pairs in the context of echo-
intensity compensation. Further investigation of pulse-echo
pairs to which a bat actually listens may lead to other adjust-
ment mechanisms in echolocation systems for multiple tar-
gets in the immediate surroundings.
ACKNOWLEDGMENTS
This work was partly supported by a grant to the Re-
search Center for Advanced Science and Technology
RCAST at Doshisha University from the Ministry of Edu-
Hiryu et al.: Echo-intensity compensation in echolocating bats 1755
133

FIG. 7. Three-dimensional spatiotemporal reconstruction for a U-turn flight
A, and sound intensity of the pulse solid circles and echo open circles
as a function of the distance to the wall with a landing mesh referred to as
target wall in case of the landing flight while a bat making U-turn B. Kobler, J. B., Wilson, B. S., Henson, O. W., Jr., and Bishop, A. L. 1985.
Asterisk indicates the echo from that wall. The bat made a U-turn at a
distance of 0.2 m marked U from that wall. The bat during the U-turn did
not decrease the intensity of emitted pulses as approached that wall, as seen
when they were landing.
cation, Culture, Sports, Science, and Technology MEXT of
Japan: Special Research Grants for the Development of
Characteristic Education from the Promotion and Mutual Aid
Corporation for Private Schools of Japan and the Innovative
Cluster Creation Project.
Au, W. W. L., and Benoit-Bird, K. J. 2003. Automatic gain control in the
echolocation system of dolphins, Nature London 423, 861863.
Boonman, A., and Jones, G. 2002. Intensity control during target ap-
proach in echolocating bats; stereotypical sensori-motor behaviour in
Daubentons bats, Myotis daubentonii, J. Exp. Biol. 205, 28652874.
Gaioni, S. J., Riquimaroux, H., and Suga, N. 1990. Biosonar behavior of
mustached bats swung on a pendulum prior to cortical ablation, J. Neu-
rophysiol. 64, 18011817.
Ghose, K., and Moss, C. F. 2003. The sonar beam pattern of a flying bat
as it tracks tethered insects, J. Acoust. Soc. Am. 114, 11201131.
Griffin, D. R. 1958. Listening in the Dark Yale University, New Haven,
CT.
Gustafson, Y., and Schnitzler, H. U. 1979. Echolocation and obstacle
avoidance in the hipposiderid bat Assellia tridens, J. Comp. Physiol. A
131, 161167.
Hartley, D. J. 1992a. Stabilization of perceived echo amplitudes in
echolocating bats. I. Echo detection and automatic gain control in the big
brown bat, Eptesicus fuscus, and the fishing bat, Noctilio leporinus, J.
Acoust. Soc. Am. 91, 11201132.
Hartley, D. J. 1992b. Stabilization of perceived echo amplitudes in
1756 J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007
echolocating bats. II. The acoustic behavior of the big brown bat, Eptesi-
cus fuscus, when tracking moving prey, J. Acoust. Soc. Am. 91, 1133
1149.
Hartley, D. J., Campbell, K. A., and Suthers, R. A. 1989. The acoustic
behavior of the fish-catching bat, Noctilio leporinus, during pre capture,
J. Acoust. Soc. Am. 86, 827.
Henson, O. W., Jr., Bishop, A. L., Keating, A. W., Kobler, J. B., Henson, M.
M., Wilson, B. S., and Hansen, R. 1987. Bisonar imaging of insects by
Pteronotus p. parnellii, the mustached bat, Natl. Geograph. Res. 3, 82
101.
Hiryu, S. 2005. A study on biosonar engineering based on spatio-
temporal measurement of bats behavior, Ph.D dissertation, Doshisha
University, Kyotanabe, Japan.
Hiryu, S., Katsura, K., Lin, L. K., Riquimaroux, H., and Watanabe, Y.
2005. Doppler-shift compensation in the Taiwanese leaf-nosed bat
Hipposideros terasensis recorded with a telemetry microphone system
during flight, J. Acoust. Soc. Am. 118, 39273933.
Holderied, M. W., and Helversen, O. V. 2003. Echolocation range and
wingbeat period match in aerial-hawking bats, Proc. R. Soc. London, Ser.
B 270, 22932299.
Hope, G. M., and Bhatnagar, K. P. 1979. Electrical response of bat retina
to spectral stimulation: comparison of four microchiropteran species, Ex-
perientia 35, 11891191.
Jen, P. H., and Kamada, T. 1982. Analysis of orientation signals emitted
by the CF-FM bat, Pteronotus p. parnellii and the FM bat, Eptesicus
fuscus during avoidance of moving and stationary obstacles, J. Comp.
Physiol. A 148, 389398.
Kalko, E. 1995. Insect pursuit, prey capture and echolocation in pipist-
relle bats Microchiroptera, Anim. Behav. 50, 861880.
Keating, A. W., Henson, O. W., Jr., Henson, M. M., Lancaster, W. C., and
Xie, D. H. 1994. Doppler-shift compensation by the mustached bat:
quantitative data, J. Exp. Biol. 188, 115129.
Kick, S. A., and Simmons, J. A. 1984. Automatic gain control in the bats
sonar receiver and the neuroethology of echolocation, J. Neurosci. 4,
27252737.
Echo intensity compensation by echolocating bats, Hear. Res. 20, 99
108.
Lancaster, W. C., Keating, A. W., and Henson, O. W., Jr. 1992. Ultrasonic
vocalizations of flying bats monitored by radiotelemetry, J. Exp. Biol.
173, 4358.
Neuweiler, G. 1984. Foraging, echolocation and audition in bats, Natur-
wiss. 71, 446455.
Riquimaroux, H., Gaioni, S. J., and Suga, N. 1991. Cortical computa-
tional maps control auditory perception, Science 251, 565568.
Riquimaroux, H., and Watanabe, Y. 2000. Characteristics of bat sonar
sounds recorded by a telemetry system and a fixed ground microphone,
The seventh western pacific regional acoustics conference (WEST-
PRACVII), 233238.
Schnitzler, H. U. 1968. Die Ultraschallortungslaute der Hufeisen-
Fledermause Chiroptera-Rhinolophidae in verschiedenen Orientierungs-
situationen The Ultrasonic sounds of horseshoe bats Chiroptera-
Rhinolophidae in different orientation situations, Zeitschrift fur
Vergleichende Physiologie 57, 376408.
Schnitzler, H. U., Kalko, E., Miller, L., and Surlykke, A. 1987. The
echolocation and hunting behavior of the bat, Pipistrellus kuhli, J. Comp.
Physiol. A 161, 267274.
Schuller, G., Beuter, K., and Schnitzler, H. U. 1974. Response to fre-
quency shifted artificial echoes in the bat Rhinolophus ferrumequinum, J.
Comp. Physiol. A 89, 275286.
Simmons, J. A. 1974. Response of the Doppler echolocation system in
the bat, Rhinolophus ferrumequinum, J. Acoust. Soc. Am. 56, 672682.
Simmons, J. A., Moffat, A. J., and Masters, W. M. 1992. Sonar gain
control and echo detection thresholds in the echolocating bat, Eptesicus
fuscus, J. Acoust. Soc. Am. 91, 11501163.
Simmons, J. A., and Stein, R. A. 1980. Acoustic imaging in bat sonar:
echolocation signals and the evolution of echolocation, J. Comp. Physiol.
A 135, 6184.
Suga, N. 1984. The extent to which biosonar information is represented
in the bat auditory cortex, in Dynamic aspects of neocortical function,
Hiryu et al.: Echo-intensity compensation in echolocating bats
134
 edited by G. M. Edelman, W. E. Gall, and W. E. Cowan Wiley, New
York, pp. 315373.
Tian, B., and Schnitzler, H. U. 1997. Echolocation signals of the greater
horseshoe bat Rhinolophus ferrumequinum in transfer flight and during
landing, J. Acoust. Soc. Am. 101, 23472364.
Trappe, M., and Schnitzler, H. U. 1982. Doppler-shift compensation in
J. Acoust. Soc. Am., Vol. 121, No. 3, March 2007
insect-catching horseshoe bats, Naturwiss. 69, 193194.
Vogler, B., and Neuweiler, G. 1983. Echolocation in the noctule Nycta-
lus noctula and horseshoe bat Rhinolophus ferrumequinum, J. Comp.
Physiol. A, 152, 421432.
Waters, E., and Jones, G. 1995. Echolocation call structure and intensity
in five species of insectivorous bats, J. Exp. Biol. 198, 475489.
Hiryu et al.: Echo-intensity compensation in echolocating bats 1757
135


136
 Biosonar and Neural
Computation in Bats
Bats extract remarkably detailed information about their
surroundings from biosonar signals. Neurons in their auditory
systems are highly specialized for performing this task
I
t used to be a common misconcep
tion that bats' use of sound puls
es to navigate and locate prey is a
crude system, the acoustic equivalent
of feeling one's way in the dark with
a cane. But biosonar has since been
shown to be anything but crude: an
echolocating bat can pursue and cap
ture a fleeing moth with a facility and
success rate that would be the envy of
any military aerospace engineer.
In addition to providing information
about how far away a target is, bat
sonar can relay some remarkable de
tails. Doppler shifts-changes in the
frequency of the echo relative to the
original signal-convey information
not only about the relative velocity of
a flying insect but also about its wing
beat. The amplitude of the echo, com
bined with the delay, indicates the size
of the target. The amplitudes of the
component frequencies correspond to
the size of various features of the
target. Differences between the ears
in intensity and arrival time of sound
give the azimuth of the target, where
as the interference pattern of sound
waves reflected within the structure of
the outer ear gives the elevation.
The complex neural computations
needed to extract this information oc
cur within a brain the size of a large
pearl. For the past 27 years my col
leagues and I have been exploring the
neural mechanisms that underlie the
NORUO SUGA has been professor of
biology at Washington University in Saint
Louis, Mo. , since 1976. Suga was born in
Japan and attended the Tokyo Metropol
itan UniverSity, where he received his
B.A. in 1958 and his Ph.D. in biology in
1963. He then went to Harvard Universi
ty as a research associate, where he first
studied the auditory system of bats with
Donald R. Griffin.
60 SCIENTIFIC AMERICAN June 1990
by Nobuo Suga
echolocating abilities of bats. The
well-defined characteristics of a bat's
auditory world make the animal ideal
for elucidating the information proc
eSSing that goes on in its auditory
system. Similar mechanisms are un
doubtedly shared by other animals.
T
here are some 800 species of Mi
crochiropteran bats in the world
today, all of which are presumed
to echolocate. These species live in
diverse habitats and vary greatly in
behavior and physical characteristics.
Their biosonar pulses also differ, even
among species within the same ge
nus. Nevertheless, these pulses can
be classified into three types: constant
frequency (CF ), frequency modulated
(FM) and combined CF-FM. CF puls
es consist of a single frequency, or
tone. FM pulses sweep downward and
sound like chirps. Combined CF -FM
pulses consist of a long, constant tone
followed by a downward chirp, iiiiiiu.
In many bats the tones are not pure
but rather consist of a fundamental,
or first, harmonic and several higher
harmonics (multiples of the funda
mental frequency).
Most bat species emit only one type
of pulse. The little broWn bat, Myotis
lucifuqus, is an "FM" bat; it emits FM
pulses lasting between .5 and three
milliseconds and sweeping downward
by about one octave. The mustached
bat, Pteronotus parnellii, is a "CF-FM"
bat; it emits long CF pulses lasting
between five and 30 milliseconds fol
lowed by a short FM sweep lasting
between two and four milliseconds.
Several species change their pulses,
depending on the situation. The fish
catching bat, Noctilio leporinus, for ex
ample, emits CF and CF - FM pulses
while cruising in flight but emits FM
pulses while hunting prey.
A long CF pulse is excellent for de-
1990 SCIENTIFIC AMERICAN, INC
tecting targets larger than the wave
length of the signal, because the re
flected sound energy is highly con
centrated at a particular frequency. It
is also ideal for measuring Doppler
shifts. The CF pulse is not appropriate,
however, for locating a target precisely
or discerning its details. A larger num
ber of frequencies is needed to obtain
more information about target fea
tures. Bats broaden their frequency
bandwidth by producing harmonics
and by emitting FM bursts that sweep
over a wide frequency range. FM puls
es also contain more information
about time and so are used to com
pute echo delays and thereby deter
mine the distance to a target.
Certain bat species control the ener
gy in each harmonic depending on the
distance to a target. If the target is far
away, they amplify the lower harmon
ics, which are less attenuated by the
air. But if the target is nearby, they
enhance the higher harmonics to ob
tain finer details of the target. When
closing in on prey, Microchiropterans
shorten the duration of pulses and
increase the rate of pulse emission, up
to 200 per second in FM bats and up to
100 per second in CF-FM bats. This
adjustment occurs not only because
bats need to characterize the prey in
greater detail but also because when
the distance between a bat and its
prey is small, the angular position of
the prey changes more rapidly, and so
the bat needs to emit more signals to
track the prey accurately.
The hunting strategies and behavior
of a bat species are directly related to
the characteristics of its biosonar. The
MUSTACHED BAT sweeps in for a mid
flight drink from a pond. This species'
biosonar has been studied extensively.
137
key elements of the biosonar, in turn,
are reflected in the functional organi
zation of its auditory system. Since bat
biosonar was established by Donald R .
Griffin and Robert Galambos four dec
ades ago, neuroethologists have stud
ied the auditory system in several bat
species but most of all in the little
brown bat, the mustached bat and the
horseshoe bat, Rhinolophus ferrum
equinum. Each of these bats produces
distinctive biosonar pulses.
The auditory mechanisms of the
mustached bat have been the most
thoroughly examined. The biosonar
and peripheral auditory system of this
bat were first described in 1964 and
1972, respectively, by Alvin Novick
and his co-workers at Yale University.
In 1972 I began research on the pe
ripheral auditory system of the mus
tached bat with James A. Simmons,
now at Brown University, and on the
central auditory system with Philip
H.-S. Jen, now at the University of Mis
souri at Columbia. Subsequently, To
shiki Manabe and Kazuro Kujirai, now
at Yokohama City University, William
E . O'Neill, now at the University of
Rochester, and nearly two dozen oth
ers have joined in studying this bat's
central auditory system and its neural
mechanisms for processing biosonar
information. This article will mainly
describe findings for this species.
ying mustached bat detects the
relative velocity of objects by
the Doppler shift in the echoes.
When a bat flies toward a stationary
object, the pulses that strike and are
reflected by it become compressed, or
Doppler-shifted. The echo received by
the bat is therefore uniformly higher
in frequency than the emitted pulse.
When the animal flies toward a fly
ing insect, the insect's beating wings
introduce oscillating frequency shifts,
which are superposed on the overall
Doppler shift, like small surface rip
ples on an ocean wave.
Certain bats, such as the mustached
bat and horseshoe bat, can detect rip
ples from insect wings against the
echoes associated with stationary ob
jects, such as walls and vegetation.
How do they do it? Part of the answer
lies in a trick called Doppler-shift
compensation, first seen by Hans-Ul
rich Schnitzler of the University of
Tiibingen. A mustached bat at rest
emits a fundamental tone of around
30.5 kilohertz, along with three higher
harmonics; the "resting" frequency of
1990 SCIENTIFIC AMERICAN, INC
the second harmonic (CF2) is around
61 kilohertz. If the bat detects a Dop
pler-shifted echo at 63 kilohertz from
a stationary object, it reduces the fre
quency of emitted pulses by about 1.8
kilohertz, so that subsequent echoes
are stabilized at a "reference" fre
quency of around 61.2 kilohertz.
These bats turn out to be special
ized to analyze tiny differences in fre
quencies near the reference frequen
cy. Hence, Doppler-shift compensa
tion brings the echo CF2 into the range
at which the bat can most easily detect
ripples from beating insect wings.
This specialization begins in the
bat's ear. Of particular interest is the
inner ear, or cochlea, which contains
the basilar membrane, a thin, elongat
ed sheet curled up like a snail. When
sound waves vibrate the eardrum, this
vibration is conducted to the basilar
membrane, stimulating tiny hair cells
on the membrane. The excitation is
transmitted, via the spiral ganglion
cells, along the auditory nerve fibers
to the brain.
The neural signal produced at the
cochlea must contain all the infor
mation vital to the bat. The physical
properties of an acoustic signal-am
plitude, time and frequency-must in
138
 MUSTACHED BAT
N CF.
I-
c:r:
UJ
120 FM.\ \ \\
J: CF,
0 0
90 \
:;z
FM,\ \ \\
>-
u
z
UJ
=>
60
CF,
CF,
FM, \
\ \
0
UJ
30
I FM,\
c:r: PULSE ECHO
LL. 0 LL.
0 20 40
TIME (MILLISECONDS)
HUNTS IN VEGETATION
BIOSONAR PULSE of the mustached bat consists of a long,
constant-frequency ( CF ) component followed by a short, fre
quency-modulated (FM ) component. Each pulse contains four
harmonics (indicated by subscripts). When closing in on a
turn be translated into neural activity.
Amplitude is expressed by the rate at
which the auditory nerve fibers dis
charge impulses: the greater the am
plitude, the higher the discharge rate.
The duration of signals and the inter
vals between them are mimicked by
the pattern of the nerve impulses. The
frequency of the signal is expressed
by location on the basilar membrane:
high frequencies vibrate the portion
nearest the eardrum, whereas lower
ones stimulate portions farther in.
A certain portion of the mustached
bat's basilar membrane is unusually
thick. This thickness is related to ex
treme sensitivity to frequencies of be
tween 61.0 and 61.5 kilohertz (the CF2
of the Doppler shift-compensated
echoes) as well as insensitivity to fre
quencies of around 59.5 kilohertz (the
CF2 of the Doppler shift-compensat
ing pulses). In other words, the mem
brane is strongly stimulated by the
echoes but poorly stimulated by the
animal's own vocalizations.
The frequency selectivity of the spi
ral ganglion cells is extremely high
within the key range of 61.0 to 61.5
kilohertz. They are tuned to single
frequencies. That is, each neuron has a
"best" frequency (the frequency that
evokes the largest response), which
differs slightly from that of its neigh
bors. Indeed, these neurons are so
sharply tuned to their best frequen
cies that they can detect shifts as small
as .01 percent. Flying insects can eas
ily evoke frequency shifts an order of
magnitude greater. The auditory pe
riphery is also highly tuned to analyze
frequency shifts near CF1 (30-kilo-
62 SCIENTIFIC AMERICAN June 1990
lITILE BROWN BAT
N 120
,u'
\
UJ
J:
90
:;z
\\
\\,'
>-
u
z
UJ
=>
60
.\
ECHO
\\ 0
UJ
30
c:r:
0 I
60 80 100 0 20
target, the bat emits shorter pulses at a higher rate, while
keeping the same tones. The little brown bat emits only FM
chirps. When nearing a target, it emits shorter, lower chirps at
a faster rate. Each species emits pulses suited to its behavior.
hertz) and CF3 (92-kilohertz) signals.
In the mustached bat the great sen
sitivity and sharp tuning of the au
ditory periphery to the CF2 frequency
are combined with Doppler-shift com
pensation to proffer three advantages.
First, the auditory periphery is exqui
sitely sensitive to the CF2 echo (near
61 kilohertz) but is insensitive to the
bat's emitted CF2 pulse (near 59 kilo
hertz) during Doppler-shift compen
sation; hence, masking of the echo by
the emitted pulse is minimal. Second,
the sharply tuned neurons are well
able to detect the signal even if it is
embedded in background noise. Third,
the array of sharply tuned neurons
has a high likelihood of picking up the
echo from the beating wings of a fly
ing insect as the echo sweeps up and
down in frequency.
These advantages enable the mus
tached bat to hunt insects success
fully, even in dense vegetation. Dop
pler-shift compensation and frequen
cy tuning do not exist in FM bats, such
as the little brown bat, which hunt in
open air.
O
nce the auditory signal has
been coded into nerve Signals,
it must be further analyzed to
extract such information as the veloci
ty or distance of prey. This process
occurs in the central auditory system.
From the cochlea, signals are process
ed sequentially, beginning at the coch
lear nucleus and proceeding to the
lateral lemniscus, inferior colliculus,
medial geniculate body and finally to
the auditory cortex.
Using slender electrodes to record
1990 SCIENTIFIC AMERICAN, INC
\ ..\
,'. , '. ". \ , .. \',
i
40 60 80 100 120
TIME (MILLISECONDS)
" --,,,
X - ....
\
""tt!f? ',j :
nerve impulses from single neurons,
my colleagues and I studied the re
sponses of neurons in the mustached
bat's auditory system as we stimulat
ed the animal with biosonar signals.
What we discovered was an extraordi
narily developed system for process
ing that information. In particular, we
found that different processing tasks
are parceled out among several ana
tomically distinct areas of the audi
tory cortex. One region contains neu
rons that respond only to certain fre
quencies and amplitudes of echoes. A
second region responds only to fre
quency differences between pulses
and echoes. A third region is sensitive
to the time interval between pulses
and echoes.
By far the largest of the specialized
regions in the mustached bat's audi
tory cortex is the one that processes
Doppler-shifted CF2 signals. This re
gion, called the DSCF area, represents
only a narrow sliver of the frequen
cy range, between 60.6 and 62.3 kilo
hertz (when the bat's resting frequen
cy is 61.00 kilohertz). Yet it occupies
30 percent of the primary auditory
cortex. The exact frequencies overrep
resented differ among individual bats
according to their resting frequencies.
In other words, each bat's auditory
system is personalized.
Similar overrepresentation is found
in the brain wherever the signal being
processed is critical to an animal's
behavior. For example, in cats and
monkeys the visual cortex overrepre
sents the fovea, the area of the retina
where visual acuity is highest. the pri
mate somatosensory cortex overrep-
139
resents the tactile sense of the fingers.
Neurons in the DSCF area are sharp
ly tuned to particular frequencies,
even more so than neurons in the au
ditory periphery. They are also tuned DOPPLER SHIFT
to the amplitude of a signal. Hence, AND FLUTTER
each DSCF neuron has a particular
frequency and amplitude to which it
responds best. This sharpening of the
TIME DELAY
response is apparently the result of
lateral inhibition, a ubiquitous mecha ,,
nism in sensory systems by which ,
I

'
inhibitory signals from adjacent neu
:" SUBTENDED

,
- -
-
ANGLE PLUS
rons enhance the selectivity of a neu
DISTANCESI"
ron to a particular stimulus.

SUBTENDED ANGLE
I :

T
he auditory cortex of the mus
tached bat is about 900 microns, ,
or some 40 to 50 neurons, thick.
When we inserted a recording elec
trode into the cortex, we found that
all of the neurons perpendicular to
the surface are tuned to an identical
frequency and amplitude. Hence, the
DSCF area has a "columnar organiza
tion." Such columnar organization was
first discovered in 1959 in the somato
sensory cortex of monkeys by Vernon
B. Mountcastle of Johns Hopkins Uni
versity and later in the visual cortex of
cats by David H. Hubel and Torsten N.
Wiesel [see "Brain Mechanisms of Vi
sion," by David H. Hubel and Torsten
N. Wiesel; SCIENTIFIC AMERICAN, Sep
tember, 19791.
When we inserted an electrode tan
gentially to the cortical surface of the
DIVERSE INFORMATION can be extracted from biosonar signals. The time delay and
DSCF area, we found that the preferred
Doppler shift of the echo indicate the distance and relative speed of the target. Rap
frequency and amplitude gradually
id flutters betray the presence of beating insect wings. Echo amplitude depends on
change, indicating the existence of fre the relative size (subtended angle) and distance of the target. lnteraural time and
quency-versus-amplitude coordinates amplitude differences convey the azimuth of the target. Interference patterns of
along the surface of the DSCF area. sound waves reflected within the structure of the outer ear indicate the elevation.
One can (crudely) picture the area as a
bicycle wheel: as one moves outward
along a spoke, the best frequency of
the neurons increases; as one moves ----FOR-WARDSWING--> <--BACKWARDSWING----
circularly from one spoke to the next,
the best amplitude changes.
DOPPLER SHIFT BY
What is the function of the DSCF 62
./ PENDULUM SWING
area? Neurons in the area respond
purely to the amplitude and frequen
UJ
:I:
cy of the echo CF2 , regardless of the o
-' 6 1 :.
:: .::: .:: .....------------------------ - .-':;.- 1 : :.::.-:..:.: : .: !
.
- . .
.. . .....: :: :

52
. ..... . ..

frequency of the emitted pulse. DSCF e

. , -.
.. . .:- :. .
RESTING FREQUENCY
neurons, then, presumably are related
>
: :.
to the acuity of frequency and ampli u
z .. . .
/.::.\. ...
UJ ...
::>
tude discrimination, as well as to the .
.
;::
detection of changes in frequency and 8 60 -: ..... :.:.
c:r:
SOUNDS EMITTED

amplitude that would be evoked by u..
BY A BAT DURING ./ :6.;::
flying insects. FIVE SWINGS
According to recent experiments by
Stephen J. Gaioni, Hiroshi Rikimaru 59--------------------L----- --------------------
0 1 2 3
and me, if the DSCF area is destroyed,
TIME (SECONDS)
a bat can no longer discriminate tiny
differences in frequency-only large DOPPLER-SlDFT COMPENSATION is demonstrated by placing a mustached bat on a
ones. The animal requires twice as pendulum. During the forward swing the animal lowers the frequency of its emitted
much time to carry out Doppler-shift pulse (red) such that the echo stays at a "reference" frequency. The animal does not
compensation and performs the task compensate for Doppler shift during the backward swing. O'Dell W. Henson, Jr.,
only half as well. From this we spec- of the University of North Carolina at Chapel Hill first performed the experiment.

1990
SCIENTIFIC AMERICAN June 140 63
1990 SCIENTIFIC AMERICAN, INC
 C1
Vl
...J
UJ
co 400
HORSESHOE BAT id
' !
0
UJ
Cl ...J::J
0 MUSTACHED BAT
_UJ
...J
300 U
60
::::>
0:: Cl Vl
0 Vl
...J
UJ
I- 200 o
U
<{
I
Vl
u.. UJCl
O::Z 40
I::::>
100 1-0
Vl
BASILAR MEMBRANE
0
HIGH FREQUENCY -> LOW FREQUENCY 0 25 SO 75 100 67
FREQUENCY (KILOHERTZ)

BASIlAR MEMBRANE is vibrated at the outer part by high
frequencies and at the inner part by lower frequencies. The
vibrations excite hair cells, which in turn excite spiral ganglion
cells. The middle graph shows the high selectivity in the mus
tached bat of neurons tuned to the three higher CF harmon
ics, particularly at 6 1 kilohertz ( CF2) but also at 30 kilohertz
( CF 1) and 92 kilohertz ( CF 3)' High selectivity is also found
in horseshoe bat auditory neurons tuned to 83 kilohertz, the
CF2 harmonic of that species. Because the little brown bat
produces only an FM pulse, it has no such neurons. The graph
on the right shows response thresholds of 12 mustached bat
neurons; for each neuron the threshold plummets at a par
ticular frequency. The plunge is very steep for neurons that
respond to frequencies near 6 1 kilohertz, the CF2 harmonic.

N ECHO
I-
0::
UJ 90 FM, FM-FM AREA
I ---................

0
...J 60
:;;2 PULSE
>-
30 FM,
u
Z
NERVE
UJ 0
::::>
0 RESPONSE
UJ
0::
u..

0 10
20 30
TIME (MILLISECONDS)
18
L<,: 1
N
I-
0::
f&><s 4 6 14
'
90 8
(MILLISECO NDS)
10
ECHO
ECHO DELAY
UJ
I
0
...J 60
\ FM,
:;;2 PULSE
>-
30
FM,
u MEDULLA
NERVE
Z 0
UJ
RESPONSE
::::>
0
UJ
0::
u..
0 10
20 30
CF, FRE
QU ENCY
CF/CF AREA

(KILOHERrz
)
2
TIME (MILLISECONDS) "'-"7l :2'. 2 9.5
30.0
AUDITORY NERVE 30.5
8. 7
N
I-
0::
UJ
90 ECHO CF, \.

ACTUAL
I 60
0 SIZE -2
...J
:;;2 PULSE CF,
30
>-
u
Z
UJ
::::>
0
0
....K RVE RESPONSE
UJ
0::
u..
0 10 20 30
TIME (MILLISECONDS)

COMPUTATIONAL MAPS in the auditory cortex of mustached
bats represent echo delay (or distance) and Doppler shift (or
relative velocity). In the FM-FM area (green), neurons along
each black line respond to a specific echo delay. The top graph
(right) shows the delay-tuning curves of six FM-FM neurons;
each neuron responds to a specific echo delay and amplitude.
In the CF / CF area (tan), neurons along the blue lines respond
to a specific CF 1 combined with varying CF2' Neurons along the
black lines respond to Doppler shifts corresponding to a spe
cific relative target velocity. The bottom graph (right) shows

64 SCIENTIFIC AMERICAN June 1990 141

1990 SCIENTIFIC AMERICAN, INC
 ulate that the DSCF area is respon frequency relation between the con
sible for the precision of the Doppler stant-frequency harmonics. In particu
shift compensation but not for per lar, the CF,/CF2 and CFI/CF3 neurons
forming the actual compensation. We respond when a tone of between 28
do not know yet how the DSCF area and 30 kilohertz (CF,) is coupled with
is connected to other regions that are a tone around 61 kilohertz (CF 2) or 92
responsible for executing Doppler kilohertz (CF 3 ) respectively. (The CF2
'
shift compensation. and CF3 frequencies are not exactly
two and three times the CF I but are
O
ne important function that we slightly higher to account for Doppler
understand well is that of per shift in the echo.)
ceiving the relative velocity of When the pulse, echo, CF tone or FM
a target. To do this, the bat's brain sound is delivered alone, these neu
must compute the Doppler shift be rons respond very weakly. But when
tween the emitted pulse and the echo. the pulse and echo are combined, they
In other words, there must be neurons show a remarkably strong response,
that "examine" the frequency relation becoming sensitive to paired signals
between the two sounds. that are as much as 6,300 times weak
We have found such neurons in a er than the smallest unpaired signals
part of the auditory cortex that we call capable of evoking a response.
the CF /CF area. There are two types of The functional organization of the
CF /CF neurons: CF,/CF2 and CF,/CF3 , CF /CF area supports our conclusion
each of which forms a distinct region. that these neurons are primarily de
These neurons are sensitive to the voted to processing velocity infor
mation. Each column of neurons re
sponds best to a particular combina
tion of two frequencies, for example,
TARGET RANGE (CENTIMETERS) 29.60 and 61.20 kilohertz. Neurons at
50 100 150 200 250 a slightly different location respond to
90 '------Tr---
a different combination, say, 30.05
and 61.10 kilohertz.
:::; Further measurements showed that
u:j 70 the best combination of frequencies
COw
O....J varies in a regular way along the sur
Ww
0 0:: face of the cortex. The preferred CFI
::>
frequency increases along one axis,
g 50
o
::co..
and the CF2 and CF3 frequencies in
lflo crease along the axis at right angles to
o:: z it. In other words, the CF /CF region is
6 30
Vl organized by a frequency-versus-fre
quency coordinate system, in which a
specific location represents a particu
10 lar relative target velocity.
0 3 6 9 12 15 Because the harmonics are simply
ECHO DELAY (MILLISECONDS)
1oo r- ------ !\ ----
--r_ --_.
integer multiples of the fundamental
tone, CFI, a comparison between the
pulse CF, and the echo CF2 or CF3 will
yield the Doppler shift. For example, a
...J3 particular neuron may respond only if
80
O....J there is both a 30-kilohertz pulse CFI
C5 and a 61-kilohertz echo CF2 ; because
::>
the pulse CF2 frequency is double the
g
0 0::
tf] 60 pulse CFI frequency, or 60 kilohertz,
::co..
lflo
the Doppler shift in the echo is one
o:: z kilohertz. This shift would arise if the
::c::>
target were moving at a relative speed
I- Sl40
of 2.8 meters per second.
We have determined that within
both the CFI/CF2 and CFI/CF3 regions
20 -- -- --- 2 --- there is an axis representing veloci types of delay-sensitive neurons is
0 2 4 6 80
ties between minus two and plus nine
FREQUENCY (KILOHERTZ) meters per second. What is more, a
disproportionate number of neurons
the frequency-tuning curves of a CF Ii represent velocities from zero to four
CFz neuron. This neuron is most respon meters per second. These speeds arise
sive when stimulated by a CF, of 29.38 during activities critical to the animal,
kilohertz at 63 decibels combined with such as when a bat is about to land on
a CFz of 60.52 kilohertz at 45 decibels. its roost or capture an insect.
1990 SCIENTIFIC AMERICAN, INC
Having found the CF /CF area, we
wanted to know where along the audi
tory pathway the process of compar
ing frequencies first arises. We already
knew that it is absent in the peripher
al auditory system, and so it must oc
cur somewhere in the central auditory
system. Based on electrophysiological
measurements made along the audito
ry pathway, my graduate student John
F. Olsen, now at Stanford University,
found that neurons in certain regions
of the inferior colliculus are tuned to
single frequencies: CFI, CF2 or CF3
They send signals to a certain region
of the medial geniculate body, where
the signals are integrated, so that neu
rons there respond to the combina
tion of CFI frequencies with specific
CF2 or CF3 frequencies. These neurons
then project to the CF /CF area of the
s
auditory cortex.
o far I have described auditory
processing of only the CF com
ponents of biosonar. But the
mustached bat also produces an FM
sound at the end of the CF component.
What is the purpose of this FM sound?
The FM Signal provides the primary
cue for measuring the time interval
between a pulse and echo-and hence
the distance to a target. A one-milli
second echo delay corresponds to a
target distance of 17.3 centimeters (at
an air temperature of 25 degrees Cel
sius). Simmons found that several spe
cies of bats can detect a difference
in distance of between 12 and 17 mil
limeters, which means they can dis
criminate a difference in echo delay
of between 69 and 98 millionths of
a second!
As in the CF /CF area, the processed
information, in this case echo delay, is
represented by a "map" in a distinct
region of the auditory cortex, which
we call the FM-FM area. There neurons
respond poorly if a pulse, echo, CF
tone or FM sound is presented individ
ually, but they respond strongly if a
pulse is followed by an echo having a
particular delay time. In extreme cases
an FM-FM neuron can be 28,000 times
more sensitive to a pulse-echo pair
than it is to either signal alone.
FM-FM neurons compare the emit
ted pulse FMI with the delayed echo
FM2, FM3 or FM4 Each of the three
0
clustered in its own subdivision. Each
FM-FM neuron is tuned to a particu
lar echo delay, and most also prefer
a particular echo amplitude. As a re
sult, most FM -FM neurons respond to
a target that is located at a particular
distance and has a certain size.
The functional organization of the
SCIENTIFIC AMERICAN June 1990
142
65
FM -FM area supports our conclusion
that this area is primarily devoted to
processing distance information. Each
column of neurons responds to a par
ticular echo delay, and the columns
are arranged so that the preferred de
lay increases along one axis. This axis
represents delays from .4 to 18 milli
seconds, or target ranges of from sev
en to 310 centimeters. The resolving
power of this neuron array is presum
ably such that an animal can detect a
difference in target distance of about
10 millimeters. And indeed, studies of
bat behavior by Simmons bear out this
presumption.
gle group of neurons in the medial
geniculate to create neurons sensitive
to combinations of FM components.
My graduate student John A. Butman
has shown that this combination sen
sitivity is mediated by the receptor
for N-methyl-D-aspartate (NMDA) to a
large extent. The receptor's biophysi
cal properties cause the neuron's re
sponse to be amplified when neural
inputs coincide. Hence, the receptor
performs the logical AND operation
(as in "IF A AND B, TIIEN en).
How do these neurons become sen
sitive to echo delays? Olsen has found
that FM -FM neurons in the medial ge
niculate body do not respond to echo
at the same time as the undelayed
response to the echo. The coincidence
of these two responses causes that
neuron to generate its own signal.
This theory suggests a neural network
in which multiple delay lines form an
array of neurons that responds to a
range of echo delays.
Where are these delay lines? In the
medial geniculate body or closer to
the periphery? Because we first saw
the difference in response latency for
pulses and echoes in the signals is
suing from the inferior colliculus, we
concluded that some delay lines must
be within the inferior colliculus itself.
In the mustached bat the inferior col

H
ow do pathways in the audito
ry system give rise to neurons
that are sensitive to pulse-echo
delays? As with neurons that respond
to combinations of frequencies, those
that respond to pulse-echo delays are
first found in the medial geniculate
body. To find out how these neurons
are connected to the rest of the au
ditory pathway, Olsen injected them
with horseradish perOxidase, which
diffuses along nerve pathways. The
substance made its way to two distinct
groups of cells in the inferior collicu
Ius as well as to the FM-FM area of the
auditory cortex.
Clearly, the two groups of collicular
delays as strongly as do the FM -FM
neurons in the auditory cortex and
that they respond somewhat to un
paired pulses and echoes. Significant
ly, these neurons always take longer to
respond to the pulse FMj than they do
to the echo FM2 , FM3 or FM4 What is
more, the difference in response laten
cy is the same as the preferred pulse
echo delay for each neuron.
These observations indicate an in
teresting mechanism for the computa
tion of echo delays. When a bat hears
its own emitted pulse FMj, the neural
response is delayed as it travels to
ward the medial geniculate body. But
the response to the returning echo is
liculus is a huge nucleus protruding
between the cerebrum and cerebel
lum. Nerve fibers ascend from the
ventrolateral portion of the colliculus
to its dorsomedial portion. Impulses
travel a distance of about two millime
ters along these nerve fibers, exciting
100 or more neurons along the way,
and introduce delays of up to eight
milliseconds.
Recall, however, that delays in the
FM-FM area of the auditory cortex are
as long as 18 milliseconds. Hence, the
delay lines in the inferior colliculus do
not account for the full range of de
lays. The data suggest that additional
delays are created by inhibitory syn

neurons, one group tuned to the pulse
FMj and the other to higher harmon
ics in the echo FM, converge on a sin-
.J __
ECHO (FM" FM, OR FM.)
not delayed. The delayed response to
the pulse then arrives at a particular
neuron in the medial geniculate body
:>
s
apses in the medial geniculate body.
o far I have discussed how the
mustached bat extracts such in
formation as the velocity and
range of a target. But in doing so, I
have avoided the question of why the
bat emits a variety of harmonics. Why
is one harmonic not enough? Techni
cally it should be, but the animal has
to cope with another problem. Bats
live in colonies with hundreds of oth

i j'7"
er bats, and somehow they must be
able to use echolocation without get
0 31
.... 0 00000
rl;'r'VE
NEURONS
I .... ill! ting confused. Several mechanisms, in
cluding binaural hearing, help to cope
with this air-traffic controller's night

INTER-
mare. One important mechanism de
pends on the first harmonic

oro
NEURONS The first harmonic is the weakest
component of the emitted pulse, con

0----- A INHIBITORY NEURON A taining less than 1 percent of the total

energy in the pulse. lodeed, the pulse
PULSE (FM,) --A...- is so feeble that other bats can barely
hear it. About all a flying bat hears
from its roostmates are the higher
NEURAL NElWORK creates neurons that respond to specific echo delays. The net
harmonics. Combinations of the high
work delays the response to a pulse FMjo so that it arrives at a neuron at the same
time as the response to a higher harmonic of the echo FM, triggering a response from
er harmOnics, however, cannot excite
that neuron. The pulse response is slowed by a combination of axonal delays (be FM-FM or CF /CF neurons. When a bat
cause it takes time for a nerve impulse to travel along an axon) and synaptic de emits a pulse, however, it can hear its
lays. Even longer delays can be created by inhibition. Neuron A inhibits neurons own first harmonic, which is conduct
1 7 through 32; after a time, the inhibition wears off, beginning with neuron 1 7, and ed from its vocal cords to its ear
the neurons become briefly excited. (Paler colors indicate longer delays.) The re through the surrounding tissue. This
sponse to the echo, on the other hand, spreads to all of the neurons simultaneously. sound, in combination with higher

66 SCIENTIFIC AMERICAN June 1990

143
1990 SCIENTIFIC AMERICAN, INC
 RANGE

DELAY
LINES
I
CF SELECTIVITY FM SELECTIVITY

iii i iii

PARALLEL PATIIWAYS process different streams of biosonar
information (left). Various CF and FM harmonics excite differ
ent parts of the basilar membrane, and signals are sent to the
auditory cortex via several subcortical nuclei. Higher up in the
auditory pathway, the neurons become more narrowly selec
tive for frequency and amplitude. In the primary auditory cor
tex, frequencies from 10 to 100 kilohertz are tonotopically
organized (yellow); the large area (pink) at the center repre
sents 60.6 to 62.3 kilohertz. CF and FM signals are integrat
ed in the medial geniculate body, giving rise to neurons that
respond to specific combinations of CF or FM signals. These
combination-sensitive neurons project axons to the CF jCF or
FM-FM areas of the auditory cortex, creating maps that corre
spond to the relative target velocity (tan) or distance (green).

harmonics that are delayed or Dop
pler-shifted, can then stimulate FM
FM and CF /CF neurons. In this way, the
neural processing of biosonar signals
is shielded from the cacophony of
echoes generated by the colony.
The suppression of the first har
monic confers another important ad
vantage. To avoid being devoured by
bats, many species of moths have "bat
detectors," or auditory receptors that
are highly sensitive to sounds of be
tween 15 and 40 kilohertz but relative
ly insensitive to higher frequencies. By
suppressing the first harmOniC, which
Robert R. Capranica of Cornell Uni
versity and Albert S. Feng of the Uni
versity of Illinois at Urbana-Cham
paign have found that frogs process
complex sounds, such as mating calls,
by way of neurons that respond to
combinations of two essential signal
elements. Daniel Margoliash of the Uni
versity of Chicago has observed simi
lar neurons in songbirds. The process
ing of biologically crucial sounds by
combination-sensitive neurons, then,
appears to be an important neural
mechanism shared among diverse
types of animals.
sory cortex and visual cortex in mon
keys and cats. Our work on bats' audi
tory cortex, then, contributes not only
to the understanding of hearing but
also to the understanding of sensory
systems in general.
RJRlliER RFADING
PERFORMANCE OF AIRBORNE ANIMAL So
NAR SYSTEMS. H. -U. Schnitzler and O. W.
Henson, Jr. , in Animal Sonar Systems.
Edited by Rene-Guy Busnel and James
F. Fish. Plenum Press, 1980.
THE ExTENT TO WHICH BIOSONAR iNFOR
MATION Is REPRESENTED IN THE BAT Au

is between 24 and 31 kilohertz, a mus
tached bat can approach moths close
ly without alerting them.
It seems likely that the human audi
tory system also uses such neural ar
rays to process speech sounds. Even
though the processing of speech is
DITORY CORTEX. Nobuo Suga in Dynam
ic Aspects of Neocortical Function. Edit
ed by Gerald M. Edelman et al. John
Wiley & Sons, Inc., 1984.
iNHIBmoN AND LEvEL-TOLERANT FRE

H
earing is as critical to bats as unique and predominantly a function QUENCY TuNING IN THE AUDITORY COR
seeing is to visually oriented of the neocortex, the subcortical audi TEX OF THE MUSTACHED BAT. Nobuo

animals, and so it is not sur
pnsmg to find evidence of elegant
neural computations in bats' auditory
cortex. But sound is also biologically
important for many other animals. Do
they have similar neural mechanisms?
It turns out that many do.
tory regions, with their combination
sensitive neurons, may analyze the
sounds of speech to a greater extent
than has been generally thought. The
work pioneered by Mountcastle, Hubel
and Wiesel has shown that similar
mechanisms underlie the somatosen-
Suga and Koichi Tsuzuki in Journal of
Neurophysiology, Vol. 53, No. 4, pages
1109-1145; April, 1985.
AUDITORY FUNCTION: NEUROBIOLOGICAL
BASES OF HEARING . Edited by Gerald M.
Edelman, W. E. Gall and W. M. Cowan.
John Wiley & Sons, Inc. , 1988.

68 SCIENTIFIC AMERICAN June 1990

144
1990 SCIENTIFIC AMERICAN, INC
 Listening with Two Ears
Studies of barn owls oer insight into just how
the brain combines acoustic signals from two sides of
the head into a single spatial perception
W
hy do people have two ears?
We can, after all, make sense of
sounds quite well with a single
ear. One task, however, requires input
from both organs: pinpointing the ex-
act direction from which a sound, such
as the cry of a baby or the growl of a
dog, is emanating. In a process called
binaural fusion, the brain compares in-
formation received from each ear and
then translates the dierences into a
unied perception of a single sound is-
suing from a specic region of space.
Extensive research has shown that the
spatial cues extracted by the human
brain are dierences in the arrival time
and the intensity, or force, of sound
waves reaching the ears from a giv-
en spot. Dierences arise because of
the distance between the ears. When a
sound comes from a point directly in
front of us, the waves reach both ears
at the same time and exert equal force
on the receptive surfaces that relay in-
formation to the brain. But if a sound
emanates from, say, left of center, the
waves will reach the right ear slightly
after the left. They will also be some-
what less intense at the right because,
as they travel to the far ear, some frac-
tion of the waves will be absorbed or
deected by the head.
The brains use of disparities in tim-
ing and intensity becomes especially ob-
vious when tones are delivered sepa-
rately to each ear through a headset. In-
stead of perceiving two distinct signals,
MASAKAZU KONISHI has been Bing
Professor of Behavioral Biology at the
California Institute of Technology since
1980. He earned a doctorate in zoology
from the University of California, Berke-
ley, in 1963. Three years later he joined
the faculty of Princeton University, where
he studied hearing and vocalization in
songbirds as well as sound localization
in owls. Konishi moved to Caltech as a
professor in 1975. In his free time, he
enjoys hiking, skiing and visiting En-
gland to attend sheep-dog competitions.
66 SCIENTIFIC AMERICAN April 1993
by Masakazu Konishi
we hear one signala phantomorigi-
nating from somewhere inside or out-
side the head. If the stimuli fed to the
ears are equally intense (equally loud)
and are conveyed simultaneously, we
perceive one sound arising from the
middle of the head. If the volume is low-
ered in just one ear or if delivery to that
ear is delayed, the source seems to move
in the direction of the opposite ear.
This much has long been known.
What is less clear is how the brain man-
ages to detect variances in timing and
intensity and how it combines the re-
sulting information into a unied spa-
tial perception. My colleagues and I at
the California Institute of Technology
have been exploring this question for
more than 15 years by studying the be-
havior and brain of the barn owl (Tyto
alba). Recently we have uncovered al-
most every step of the computational
process in these animals. (The only oth-
er sensory system that is as completely
dened belongs to a sh.) We nd that
the owl brain combines aural signals
relating to location not all at once but
through an amazing series of steps. In-
formation about timing and intensity is
processed separately in parallel path-
ways that converge only late in those
pathways. It is highly probable that hu-
mans and other mammals achieve bin-
aural fusion in much the same manner.
I
rst thought of examining the neu-
ral basis of sound location in owls
in 1963, when I heard Roger S.
Payne, now at the Whale Conservation
Institute in Lincoln, Mass., report that
the barn owl can catch a mouse readily
in darkness, solely by relying on acous-
tic cues. I had recently earned a doc-
torate in zoology and wanted to know
more about how animals identify the
position of a sound source, but I had yet
to choose a species to study. Three years
later, at Princeton University, I observed
the exquisite aural abilities of barn owls
for myself after I obtained three of them
from a bird-watcher. When I watched
one of the owls through an infrared-sen-
sitive video camera in a totally dark
room, I was impressed by the speed and
accuracy with which it turned its head
toward a noise. I concluded that the
head-turning response might help un-
cover whether such animals use binau-
ral fusion in locating sound. If they did,
studies of their brain could help eluci-
date how such fusion is accomplished.
As I had anticipated, the head-turning
response did prove extremely useful to
me and my postdoctoral fellows, partic-
ularly after I established a laboratory at
Caltech in 1975. In some of our earliest
research there, Eric I. Knudsen, now at
Stanford University, and I obtained in-
direct evidence that barn owls, like hu-
mans, must merge information from the
two ears to locate a sound. When one
ear was plugged, the animals turned
the head in response to noise from a
loudspeaker, but they did not center
on the speaker [see The Hearing of the
Barn Owl, by Eric I. Knudsen; SCIENTIF-
IC AMERICAN, December 1981].
In the early 1980s Andrew Moise
and I additionally showed that the barn
owl extracts directional information
from disparities in the timing and the
intensity of signals reaching the two
earstechnically called interaural time
dierences and interaural intensity dif-
ferences. As part of that eort, we mea-
sured the dierences that arose as we
moved a speaker across the surface of
an imaginary globe around an owls
head. Microphones we had placed in the
ears relayed the signals reaching each
ear to a device that measured arrival
time and volume. When we eased the
speaker from the midline of the face
(zero angle) 90 degrees to the left or
BARN OWL PINPOINTS PREY in the dark
by listening. It determines the appro-
priate trajectory in which to y by com-
paring dierences in the timing and the
intensity of sounds reaching its two ears.
An infrared strobe ashing ve times
per second caught this barn owl in ac-
tion in the authors laboratory.
Copyright 1993 Scientific American, Inc.
145
right, the dierence in arrival time at
the two ears increased systematically.
Those results resembled the ndings
of human studies.
In contrast to human ndings, the
dierence in intensity did not vary ap-
preciably as the speaker was moved
Copyright 1993 Scientific American, Inc.
horizontally. But it did increase as the
speaker was moved up or down from
eye levelat least when the sound in-
cluded waves of frequencies higher than
three kilohertz, or 3,000 cycles per sec-
ond. Payne, who had seen the same in-
tensity changes in earlier studies, has at-
tributed them, apparently correctly, to
an asymmetry in the placement of the
owls ears. The left ear is higher than
eye level but points downward, where-
as the right ear is lower but points
upward. The net result is that the left
ear is more sensitive to sounds coming
146
from below, and the right is more sen- We intended to put a standard headset
sitive to sounds from above. on tame animals and to convey a noise
Satised that arrival time and inten- separately to each ear, varying the dif-
sity often dier for the two ears, we ference in delivery time or volume, or
could go on to determine whether the both. We would then see whether par-
owl actually uses specic combinations ticular combinations of time and inten-
of disparities in locating sound sources. sity dierences caused the animals to
a b TIME (MICROSECONDS)
126
84
42
0 few years earlier. We t two small coils
42
SOUND
SOURCE
20
DEGREES
c d INTENSITY (DECIBELS)
16
12
8
4
0
4 the owl turned its head mostly in the
8
12
20 16
20
DEGREES
e sounds so that both the delivery time
DIFFERENCES IN TIMING AND INTENSITY at which a sound reaches an owls two
ears vary as the source of the sound moves along the surface of an imaginary
globe around the owls head. Dierences in timing locate the sound in the horizon-
tal plane (a); the dierence increases 42 microseconds every 20 degrees a sound
source moves (b). Dierences in intensity locate the sound vertically (c ). Sound
from above eye level is more intense in the right ear, by the decibel levels shown
(d ); from below eye level, it is more intense in the left ear. Dierences vary with
frequency; they were measured for six kilohertz in this case. Combining the two
graphs (e ) denes each location in space. When an owl is exposed to a particular
pair of dierences, it quickly turns its head in a predictable direction (photograph).
68 SCIENTIFIC AMERICAN April 1993
turn the head reliably in specic direc-
tions. Unfortunately, we did not receive
cooperation from our subjects. When
we tried to ax the earphones, each
owl we approached shook its head and
backed o. We managed to proceed
only after we acquired tiny earphones
that could be inserted into the owls
ear canal.
We also had to devise a way to mea-
sure the direction of head turning, de-
termining both the horizontal and ver-
tical components of the response to
each set of stimuli. We solved the prob-
lem mainly by applying the search-coil
technique that Gary G. Blasdel, now at
Harvard Medical School, had designed a
84 of copper wire, arranged perpendicu-
126
larly to each other, on an owls head.
We positioned the owl between two big
coils carrying electric current. As the
head moved, the large coils induced cur-
rents in the small ones. Variations in the
ow of current in the smaller coils re-
vealed both the horizontal and vertical
angles of the head turning.
Sure enough, the owl responded rap-
idly to signals from the earphones, just
as if it had heard noise arising from out-
side the head. When the sound in one
ear preceded that in the other ear, the
head turned in the direction of the lead-
ing ear. More precisely, if we held the
volume constant but issued the sound
to one ear slightly before the other ear,
horizontal direction. The longer we de-
layed delivering the sound to the sec-
ond ear, the further the head turned.
Similarly, if we varied intensity but
held timing constant, the owl tended to
move its head up or down. If we issued
and the intensity of signals to the left
ear diered from those of the right, the
owl moved its head horizontally and
vertically. Indeed, combinations of inter-
aural timing and intensity dierences
that mimicked the combinations gener-
ated from a speaker at particular sites
caused the animal to turn toward exact-
ly those same sites. We could therefore
be condent that the owl brain does
fuse timing and intensity data to deter-
mine the horizontal and vertical coor-
dinates of a sound source. The process
by which barn owls calculate distance
is less clear.
T
o learn how the brain carries out
binaural fusion, we had to exam-
ine the brain itself. Our research
plan built on work Knudsen and I had
completed several years earlier. We had
identied cells that are now known
to be critical to sound location. Called
space-specic neurons, they react only
Copyright 1993 Scientific American, Inc.
147
to acoustic stimuli originating from spe-
cic receptive elds, or restricted areas
in space [see illustration at right]. These
neurons reside in a region of the brain
called the external nucleus, which is
situated within the auditory area of the
midbrain (the equivalent of the mam-
malian inferior colliculus). Collective-
ly, the space-specic neurons in the
left external nucleus form a map of pri-
marily the right side of auditory space
(the broad region in space from which
sounds can be detected), and those of
the right external nucleus form a map
of primarily the left half of auditory
space, although there is some overlap.
We identied the space-specic cells
by resting a microelectrode, which re-
sembles a sewing needle, on single neu-
rons in the brain of an anesthetized an-
imal. As we held the electrode in place,
we maneuvered a speaker across the
surface of our imaginary globe around
the owls head. Certain neurons red
impulses only if the noise emanated
from a particular receptive eld. For
instance, in an owl facing forward, one
space-specic neuron might respond
only if a speaker were placed within a
receptive eld extending roughly 20
degrees to the left of the owls line of
sight and some 15 degrees above or
below it. A dierent neuron would re
when the speaker was transferred else-
where on the globe.
But how did these neurons obtain di-
rectional information? Did they process
the relevant cues themselves? Or were
the cues extracted and combined to
some extent at one or more lower way
stations (relay centers) in the brain [see
illustration on page 72], after which the
results were simply fed upward?
Moise and I intended to answer
these questions by carrying out experi-
ments in which we would deliver sounds
through earphones. But rst we had to
be certain that signals able to excite
particular space-specic neurons truly
mimicked the interaural time and in-
tensity dierences that caused the neu-
rons to re under more natural condi-
tionsnamely, when a sound emanated
from a spot in the neurons receptive
eld. A series of tests gave us the en-
couragement we needed. In these stud-
ies, we issued sounds through the ear-
phones and monitored the response of
individual neurons by again holding a
microelectrode on or near the cells. As
we hoped, we found that cells respond-
ed to specic combinations of signals.
Further, the sets of timing and intensity
dierences that triggered strong ring
by space-specic neurons corresponded
exactly to the combinations that caused
an owl to turn its head toward a spot in
the neurons receptive eld. This con-
Copyright 1993 Scientific American, Inc.
OWLS BRAIN uses space-
specic neurons in the exter-
nal nucleus of the midbrain
auditory area to map pre-
cise regions (bars)called
receptive eldsin audito-
ry space. In probing to see
how space-specic neurons
work, the author and his
colleagues uncovered the
step-by-step procedure in
the brain that leads to the
ring of these neurons.
EXTERNAL NUCLEUS
40
20
LEFT 10
MIDBRAIN
AUDITORY AREA 0
10
gruence armed that our proposed ap-
proach was sensible.
In our initial eorts to trace the steps
by which the brain circuitry accomplish-
es binaural fusion, Moise and I tried to
nd neurons sensitive to interaural tim-
ing or intensity dierences in the way
stations that relay signals from the au-
ditory nerve up to the midbrain. These
preliminary investigations, completed
in 1983, suggested that certain stations
are sensitive only to timing cues, where-
as others are sensitive solely to intensi-
ty cues. The brain, it seemed, functioned
like a parallel computer, processing in-
formation about timing and intensity
through separate circuits.
S
uch clues led us to seek further
evidence of parallel processing.
Joined by Terry T. Takahashi,
now at the University of Oregon, we be-
gan by examining the functioning of
the lowest way stations in the brain
the cochlear nuclei. Each cerebral hemi-
sphere has two: the magnocellular nu-
cleus and the angular nucleus. In owls,
as in other birds, each ber of the audi-
tory nervethat is, each signal-convey-
ing axon projecting from a neuron in
the eardivides into two branches af-
ter leaving the ear. One branch enters
the magnocellular nucleus; the other
enters the angular nucleus.
We wondered how the space-specic
neurons would behave if we prevented
nerve cells from ring in one of the two
cochlear nuclei. We therefore injected a
minute amount of a local anesthetic
into either the magnocellular or angu-
lar nucleus. The results were dramatic:
the drug in the magnocellular nucleus
altered the response of space-specic
neurons to interaural time dierences
60
30
0
60 30
30
0
30
without aecting the response to inten-
sity dierences. The converse occurred
when the angular nucleus received the
drug. Evidently, timing and intensity are
indeed processed separately, at least
at the lowest way stations of the brain;
the magnocellular neurons convey tim-
ing data, and the angular neurons con-
vey intensity data.
These exciting results spurred me to
ask Takahashi to map the trajectories
of the neurons that connect way sta-
tions in the auditory system. His work
eventually revealed that two separate
pathways extend from the cochlear nu-
clei to the midbrain. The anatomic evi-
dence, then, added further support to
the parallel-processing model.
While Takahashi was conducting his
mapping research, W. E. Sullivan and I
explored the ways magnocellular and
angular nuclei extract timing and inten-
sity information from signals arriving
from the auditory nerve. To understand
our discoveries, one needs to be aware
that most sounds in nature are made up
of several waves, each having a dierent
frequency. When the waves reach a re-
ceptive surface in the ear, known as the
basilar membrane, the membrane be-
gins to vibrate, but not uniformly. Dif-
ferent parts of the membrane vibrate
maximally in response to particular fre-
quencies. In turn, neurons that are con-
nected to the maximally vibrating areas
(and thus are tuned to specic fre-
quencies) become excited. These neu-
rons propagate impulses along the au-
ditory nerve to the brain.
We and others nd that the intensity
of a sound wave of a given frequency is
conveyed to the brain from the ear by
the ring rate of auditory neurons tuned
to that frequency. This much makes
SCIENTIFIC AMERICAN April 1993 69
148
 SOUND SOURCE intuitive sense. Our next result is less
obvious. Neurons of the auditory nerve
also exhibit what is called phase lock-
ing : they re at characteristic points,
or phase angles, along the sound wave
[see bottom illustration on this page].
That is, a neuron tuned to one frequen-
LEFT RIGHT cy will tend to re, for example, when
EAR EAR
the wave is at baseline (zero degrees),
RELAY
STATION although it does not necessarily re
DELAY LINE IN BRAIN every time the wave reaches that posi-
tion. A neuron tuned to a dierent fre-
quency will tend to re at a dierent
phase angle, such as when a wave is
cresting (at the point called 90 degrees,
COINCIDENCE which is a quarter of the way through a
DETECTOR full 360-degree wave cycle), or reaches
some other specic point. In both ears,
MODEL CIRCUIT for detection of interaural time dierences was suggested in 1948. impulses produced by neurons tuned
The coincidence detectors receive inputs from both ears. They re only when im- to the same frequency will lock to the
pulses from the two sides arrive simultaneously through bers that serve as delay same phase angle. But, depending on
lines. The detector that responds (darkly colored circle) changes as a sound source when the signals reach the ears, the
moves from directly in front of an individual (left) to the side (right). The owl brain
train of impulses generated in one ear
operates in much the way the model proposed.
may be delayed relative to the impulse
train generated in the opposite ear.
a PHASE LOCKING It turns out that cells of the magno-
ONE CYCLE cellular nucleus exhibit phase locking.
SOUND WAVE But they are insensitive to intensity;
AMPLITUDE

90 changes in the volume of a tone do not

0 180 aect the rate of ring. In contrast, few
360 angular neurons show phase locking,
TIME 270 but they respond distinctly to changes
in intensity. These and other results in-
IMPULSE AXON
dicate that the owl depends on trains
of phase-locked impulses relayed from
the magnocellular nucleus for measur-
NEURON PHASE-LOCKED TO 90 ing interaural time dierences, and the
animal relies on the rate of impulses
b COINCIDENCE COINCIDENCE DETECTOR red by the angular nucleus for gauging
(MAXIMAL FIRING) interaural intensity dierences. Overall,
then, our analyses of the lowest way
NEURON FROM LEFT stations of the brain established that
the cochlear nuclei serve as lters that
pass along information about timing or
NEURON FROM RIGHT intensity, but not both.

W
c NONCOINCIDENCE e then proceeded to explore
WEAK FIRING
higher regions, pursuing how
the brain handles timing data
in particular. Other studies, which will
be discussed, addressed intensity. We
learned that when phase-locked im-
pulses induced by sound waves of a
single frequency (a pure tone) leave the
d PHASE AMBIGUITY magnocellular nucleus on each side of
FIRST IMPULSE
the brain, they travel to a second way
station: the laminar nucleus. Impulses
FIRST IMPULSE from each ear are transmitted to the
nucleus on both the opposite and the
same side of the head. The laminar nu-
SOUND WAVE OF A SINGLE FREQUENCY causes neurons sensitive to it to re trains
cleus is, therefore, the rst place where
of impulses at a particular phase angle (a). Coincidence detectors in the owls brain the information from both ears comes
re most strongly when impulses generated at the same phase angle reach the de- together in one place.
tectors simultaneously ( far right in b). Detectors can also re, but more weakly, The general problem of how the
when impulse trains reaching them are slightly asynchronous (c). In what is called brain combines timing data has been a
phase ambiguity, peak ring can occur if a sound to one ear is delayed or advanced subject of speculation for decades. The
by a full cycle from another delivery time that yields coincidence (d ). late Lloyd A. Jeress put forth a rea-

70 SCIENTIFIC AMERICAN April 1993

Copyright 1993 Scientific American, Inc.
149
sonable model in 1948, while spending
a sabbatical leave at Caltech. Jeress
proposed that the nerve bers carry-
ing time-related signals from the ears
(called delay lines) vary in how rapidly
they deliver signals to way stations in
the brain. They ultimately converge at
neurons (known as coincidence detec-
tors) that re only when impulses from
the two sides arrive simultaneously.
Signals reaching the ears at dierent
times would attain coincidencearrive
at coincidence detectors in unisonif
the sum of a sound waves transit time
to an ear and the travel time of impuls-
es emanating from that ear to a coinci-
dence detector were equal for the two
sides of the head. Consider a sound
that reached the left ear ve microsec-
onds before it reached the right ear. Im-
pulses from the two ears would meet
simultaneously at a coincidence detec-
tor in, say, the right hemisphere if the
delay lines from the left ear (the near
ear) prolonged the transit time of im-
pulses from that ear to a coincidence
detector by ve microseconds over the
time it would take impulses to traverse
bers from the right ear [see top illus-
tration on opposite page].
Since 1948, physiological studies ex-
amining neuronal ring in dogs and
cats and anatomic studies of chicken
brains have suggested that the brain
does in fact measure interaural time dif-
ferences by means of delay lines and
coincidence detection. In 1986 Cath-
erine E. Carr, now at the University of
Maryland, and I demonstrated in the
barn owl that nerve bers from magno-
cellular neurons serve as delay lines and
neurons of the laminar nucleus serve as
coincidence detectors.
B
ut the owls detection circuit, like
those of mammals that have been
examined, diers somewhat from
the Jeress model. Neurons of the lam-
inar nucleus respond most strongly to
coincidence brought about by partic-
ular time dierences. Yet they also re-
spond, albeit less strongly, to signals
that miss perfect coincidence. The num-
ber of impulses declines gradually as
the interaural time dierence increas-
es or decreases from the value that
produces coincidencethat is, until the
waves reaching one ear are 180 degrees
(a full half cycle) out of phase from the
position that would bring about coin-
cidence. At that point, ring virtually
ceases. (The neurons also respond, at
an intermediate level, to signals deliv-
ered to just one ear.)
In a way, then, coincidence detec-
tors, by virtue of the delay lines feeding
them, can be said to be maximally sen-
Copyright 1993 Scientific American, Inc.
LEFT LAMINAR NUCLEUS
FIBERS FROM THE MAGNOCELLULAR NUCLEUS serve as delay lines, and neurons
in the laminar nucleus act as coincidence detectors in the owls brain. When im-
pulses traveling through the left (blue) and right (green ) bers reach laminar neu-
rons (black dots) simultaneously, the neurons re strongly.
sitive to specic time dierences. They
are not, however, totally selective as
to when they produce a peak response.
They can be induced to re with rising
strength as the phase dierence increas-
es beyond 180 degrees from the value
that produces coincidence. When the
displacement reaches a full 360 degrees,
the arrival time of sound waves at one
ear is delayed by the time it takes for a
sound wave to complete a full cycle. In
that situation, and at every 360-degree
dierence, coincidence detectors will re-
peatedly be hit by a series of synchron-
ous impulses and will re maximally.
Thus, the same cell can react to more
than one time dierence.
Fortunately for the owl, some mech-
anism resolves such phase ambigui-
ty at higher stages, thereby prevent-
ing confusion. How this resolution is
achieved remains obscure. Another mys-
tery engages us as well: the owl can de-
tect interaural time dierences as short
as 10 microseconds (10 millionths of a
second). Yet a single impulse persists
considerably longer than that, on the
order of 1,000 microseconds. We are
seeking an explanation for this appar-
ent paradox.
The rest of the pathway for time de-
tection is more straightforward. After
a coincidence detector in the laminar
nucleus on one side of the brain de-
termines the interaural time dierence
produced by a sound of a given frequen-
cy, it simply passes the result upward
to higher stations, including to the core
region of the midbrain auditory area
on the opposite side of the head. Con-
TO RIGHT LAMINAR
NUCLEUS
LEFT RIGHT
MAGNO- MAGNO-
CELLULAR
CELLULAR
NUCLEUS
NUCLEUS
TO RIGHT LAMINAR
NUCLEUS
sequently, the higher areas inherit from
the laminar nucleus not only selectiv-
ity for frequency and interaural time
differences but also phase ambiguity.
The information in the core, in turn, is
passed to a surrounding areaknown
as the shell of the midbrain auditory
areaon the reverse side of the brain,
where it is nally combined with infor-
mation about intensity.
M
y colleagues and I understand
less about the operation of the
intensity pathway that converg-
es with the time pathway in the shell.
But we have made good progress. Unlike
the magnocellular nucleus, which pro-
jects up only one stage, to the laminar
nucleus, the intensity-detecting angular
nucleus projects directly to many high-
er stations (except the external nucle-
us). Among them is the posterior later-
al lemniscal nucleus.
The posterior lemniscal nucleus on
one side of the head receives direct in-
put only from the angular nucleus on
the opposite side. It nonetheless man-
ages to discern intensity dierences be-
tween the two ears. Indeed, it is the low-
est station in the brain to do so. The
lemniscal area can detect such dier-
ences because its twin (in the opposite
cerebral hemisphere) sends it informa-
tion from the other angular nucleus. In
essence, neurons of the lemniscal nu-
cleus on one side of the head receive ex-
citatory signals from the ear on the op-
posite side, and they receive inhibitory
signals from the ear on the same side.
The balance between excitatory and in-
SCIENTIFIC AMERICAN April 1993 71
150
 The Auditory Circuit

P arallel pathways in the barn owls brain sep-

arately process the timing (blue) and the in-
tensity (red ) of sounds reaching the ears (di-
agram and flow chart). The simplified diagram de-
picts the pathways only for the left ear except where
input from the right ear joins that pathway; brain
structures are not drawn to scale. Processing begins
as the magnocellular nucleus separates out informa-
tion about time and as the angular nucleus extracts
information about intensity from signals delivered
by the auditory nerve. The time pathway goes to the
laminar nucleus, which receives input from both the
right and the left magnocellular nuclei. Neurons of
the laminar nucleus are connected to two higher
stations: the anterior lateral lemniscal nucleus and
the core of the midbrain auditory area. Meanwhile
information about intensity travels from the angular
nucleus to the posterior lateral lemniscal nucleus,
where information from the two ears comes togeth-
er. The time and intensity pathways finally join in the
lateral shell of the midbrain auditory area. They pro-
ject from there to the external nucleus, which hous-
es the space-specific neurons and is the final station
in processing the acoustic cues for locating a sound.
If viewed in terms of an algorithm ( far right), a set
of step-by-step procedures for solving a problem,
these neurons are at the top of the hierarchy: they
represent the final results of all computations that
take place in the network.

hibitory signals determines the rate at In the shell, most neurons respond veal nothing on their own re impulses
which the lemniscal neurons re. strongly to both interaural intensity and collectively in a particular pattern.
We have observed, too, that neurons interaural timing dierences generated

of the posterior lemniscal nucleus vary
systematically in the intensity dierenc-
es that cause them to re most strong-
ly. Georey A. Manley and Christine
Kppl of the Technical University of
Munich showed in my laboratory that
neurons at the bottom of the left nucle-
us respond maximally when sound is
much louder in the left ear and that
those at the top of the nucleus re
most strongly when sound is louder in
the right ear. Similarly, neurons at the
bottom of the right posterior nucleus
respond most strongly when sound is
much louder in the right ear, and those
at the top of the nucleus prefer louder
sound in the left ear. This arrangement
clearly enables space-specic neurons
to determine that a noise is coming
from above or below eye level. The pro-
cess by which space-specic neurons
convert signals from the posterior lem-
niscal nucleus into vertical coordinates
remains to be established, however.
The next higher station is the later-
al shell of the midbrain auditory area;
neurons from the posterior lemniscal
nucleus on each side of the brain send
signals to the shell in both hemispheres.
by sounds within a narrow range of fre-
quencies. This station does not provide
the owl with sufcient information to
ensure accurate sound location, howev-
er, because phase ambiguity persists.
The ambiguity disappears only at the
level of the external nucleus, home of
the space-specic neurons. These neu-
rons are broadly tuned to frequency, re-
ceiving timing and intensity data from
many frequency channels. This converg-
ence somehow supplies the input need-
ed for the brain to select the correct co-
ordinates of a sound source. The selec-
tivity of space-specic neurons, then,
results from the parallel processing of
time and intensity data and from the
combination of the results in the shell
and in the external nucleus itself.
We have not yet resolved the number
of space-specic neurons that must re
in order for an owl to turn its head to-
ward a sound source. Nevertheless, we
know that individual neurons can carry
the needed spatial data. This fact belies
the view of some researchers that single
neurons cannot represent such complex
information and that perceptions arise
only when whole groups of cells that re-
T
ogether our neurological explora-
tions have elucidated much of the
algorithm, or step-by-step proto-
col, by which the owl brain achieves
binaural fusion. Presumably, we humans
follow essentially the same algorithm
(although some of the processing sta-
tions might dier). Recall, for example,
that several lines of evidence suggest
mammals rely on delay lines and coin-
cidence detection in locating sounds.
We can extrapolate even further. The
only other neural algorithm for a sen-
sory task that has been deciphered in
equal detail is one followed by electrici-
ty-emitting sh of the genus Eigenman-
nia. Walter F. Heiligenberg of the Uni-
versity of California at San Diego and
his associates have worked out the rules
enabling members of this species to de-
termine whether their electric waves are
of higher or lower frequency than those
of other Eigenmannia in the immediate
vicinity. (In response, a sh might alter
the frequency of the wave it emits.) Ei-
genmannia rely on parallel pathways to
process separate sensory information.
Also, relevant information is processed
in steps; the parallel pathways converge

72 SCIENTIFIC AMERICAN April 1993

Copyright 1993 Scientific American, Inc.
151
 PATHWAYS IN THE BRAIN NEURAL ALGORITHM

EXTERNAL NUCLEUS Formation of a map of auditory space

Elimination of phase ambiguity

LATERAL SHELL Convergence of different frequency channels

Convergence of time and intensity pathways

CORE

ANTERIOR LATERAL POSTERIOR LATERAL Detection and relaying of interaural intensity differences
LEMNISCAL NUCLEUS LEMNISCAL NUCLEUS

LAMINAR NUCLEUS Detection and relaying of interaural time differences

TIME INTENSITY

MAGNOCELLULAR ANGULAR NUCLEUS Separation of time and intensity data

NUCLEUS

INNER EAR Translation of frequency, time and intensity

cues into nerve signals

at a high station; and neurons at the top
of the hierarchy respond selectively to
precise combinations of cues. The sh
algorithm is thus remarkably similar to
that of the barn owl, even though the
problems that are solved, the sensory
systems involved, the sites of processing
in the brain and the species are dier-
ent. The similarities suggest that brains
follow certain general rules for informa-
tion processing that are common to dif-
ferent sensory systems and species.
Carver A. Mead, here at Caltech, thinks
the owl algorithm may also teach some-
thing to designers of analog silicon
chips, otherwise known as VLSI ( Very
Large Scale Integrated) circuits. In 1988
he and John Lazzaro, then his graduate
student, constructed an owl chip that
reproduces the steps through which the
barn owl measures interaural time dif-
ferences. The model, about 73 square
millimeters in area, contains only 64
auditory nerve bers in each ear (many
fewer than truly exist) and some 21,000
delay lines. (It also has 200,000 tran-
sistors, mainly to regulate the delay
lines.) Even in its pared-down version,
the electronic nervous system takes up
much more space and energy than does
the biological system. Historically, engi-
neers have constructed chips according
to principles drawn from electronics,
physics and chemistry. The economy of
the biological circuit suggests that nat-
ural principles may help engineers build
analog chips that consume less energy
and take up less space than usual.
My laboratorys research into the owl
brain is by no means nished. Beyond
lling in some of the gaps in our knowl-
edge of binaural fusion, we hope to be-
gin addressing other problems. For ex-
ample, Alvin M. Liberman of Haskins
Laboratories in New Haven, Conn., has
proposed that the human brain pro-
cesses speech sounds separately from
FURTHER READING
A NEURAL MAP OF AUDITORY SPACE IN THE
OWL. E. I. Knudsen and M. Konishi in Sci-
ence , Vol. 200, pages 795797; May 19,
1978.
NEURONAL AND BEHAVIORAL SENSITIVITY
TO BINAURAL TIME DIFFERENCES IN THE
OWL. Andrew Moiseff and Masakazu Ko-
nishi in Journal of Neuroscience, Vol. 1,
No. 1, pages 4048; January 1981.
NEUROPHYSIOLOGICAL AND ANATOMICAL
SUBSTRATES OF SOUND LOCALIZATION IN
THE OWL. M. Konishi, T. T. Takahashi, H.
Wagner, W. E. Sullivan and C. E. Carr in
nonspeech sounds. By the same token,
we can ask whether the owl separately
processes signals for sound location
and other acoustic information. Some
brain stations that participate in spatial
orientation may also take part in other
sensory activities, such as making owls
selectively attuned to the calls of mates
and chicks. How does the owl, using one
set of neurons, sort out the algorithms
for dierent sensory tasks? By solving
such riddles for the owl, we should be-
gin to answer some of the big questions
that relate to more complex brains and,
perhaps, to all brains.
Auditory Function: Neurobiological Bases of
Hearing. Edited by G. M. Edelman, W. E.
Gall and W. M. Cowan. John Wiley & Sons,
1988.
A CIRCUIT FOR DETECTION OF INTERAURAL
TIME DIFFERENCES IN THE BRAIN STEM OF
THE BARN OWL. C. E. Carr and M. Konishi
in Journal of Neuroscience, Vol. 10, No.
10, pages 32273246; October 1990.
THE NEURAL ALGORITHM FOR SOUND LO-
CALIZATION IN THE OWL. M. Konishi in
The Harvey Lectures, Series 86, pages 47
64; 1992.

SCIENTIFIC AMERICAN April 1993 73

Copyright 1993 Scientific American, Inc.
152
Overt attention toward oriented objects
in free-viewing barn owls
Wolf Maximilian Harmeninga,1,2, Julius Orlowskia, Ohad Ben-Shaharb, and Hermann Wagnera
a
Department of Zoology and Animal Physiology, Rheinisch-Westfaelische Technische Hochschule Aachen University, 52074 Aachen, Germany; and
b
Department of Computer Science, and the Zlotowski Center for Neuroscience, Ben-Gurion University of the Negev, Beer Sheva 84105, Israel
Edited by Thomas D. Albright, Salk Institute for Biological Studies, La Jolla, CA, and approved April 7, 2011 (received for review February 23, 2011)
Visual saliency based on orientation contrast is a perceptual product
attributed to the functional organization of the mammalian brain.
We examined this visual phenomenon in barn owls by mounting
a wireless video microcamera on the owls heads and confronting
them with visual scenes that contained one differently oriented tar-
get among similarly oriented distracters. Without being conned by
any particular task, the owls looked signicantly longer, more often,
and earlier at the target, thus exhibiting visual search strategies so
far demonstrated in similar conditions only in primates. Given the
considerable differences in phylogeny and the structure of visual
pathways between owls and humans, these ndings suggest that
orientation saliency has computational optimality in a wide variety
of ecological contexts, and thus constitutes a universal building
block for efcient visual information processing in general.
feature search | gaze map | pop-out | visual behavior | avian vision
I n his now-classic experiments, Yarbus (1) showed that humans
who freely view visual scenes move their eyes between salient,
discretely spaced features (e.g., eyes and mouth in face scenes,
people and objects in indoor scenes). Similar behavior in other
environments has been observed in primates (2, 3) and even in
birds (47). Indeed, the visual systems of humans and other ani-
mals have mechanisms to overtly shift attention to salient parts of
visual stimuli, a selective process that helps allocate the brains
limited computational resources to potentially important sensory
information (for a review see refs. 8 and 9). The selective nature of
the visual system perhaps is best expressed in what has been
termed visual search (10, 11).
Although visual search is a visual behavior occurring with nat-
ural and synthetic stimuli, controlled scientic studies of visual
search typically make use of well-dened, simple objects. In such
experiments, the subjects task is to detect one outstanding object
(the target) embedded among many similar objects (the dis-
tracters) (12). Studies in which the target differs from the dis-
tracters in one visual feature are referred to as feature search (11).
A typical feature would be an early visual cue, such as contrast,
color, motion, orientation, or even shape. When feature search
exhibits reaction times that do not change much with the number
of distracters, the behavior is usually characterized as pop-out,
which is indicative of a parallel preattentive process that precedes
any subsequent serial attentive processing (12).
Although much is known about the properties and the neural
networks involved in visual search in humans and primates (11
15), knowledge of this process (especially its neural substrate) in
nonprimate animals is limited. This is in stark contrast to the
important ecological function of search strategies in all animals,
which should have been optimized according to the ecological and
survival needs of a species in the course of evolution. Notable
exceptions are visual search strategies reported in pigeons (4, 5)
and recent orientation-saliency behavior found in archer sh (16).
In the present work we studied visual search in free-viewing
barn owls, which, as we argue, may become a model animal with
several important advantages for exploring this visual attentive
behavior in animals. First, barn owls eye movements are either
absent or very small, allowing the study of overt attention with an
www.pnas.org/cgi/doi/10.1073/pnas.1101582108
external camera xed to owls heads (6). Second, barn owls are
known to make conspicuous peering movements (17) and can also
covertly shift attention toward interesting targets (18), meaning
that their vision is likely to incorporate attention mechanisms.
Third, much is known about barn owls visual system (19, 20), and
the neural circuits underlying visual perception in this species
have been studied in some detail (2127). With this in mind, we
asked whether barn owls exhibit similar visual search behavior as
humans. The data presented here indeed show such similarities.
Results
In our barn owl experiments, the setup and procedures were
chosen to resemble the classical visual search studies performed
with humans (1, 12), with a specic focus on saliency due to ori-
entation. Two barn owls (subjects HB and WH) were trained
to carry the OwlCam, a head-mounted wireless microcamera
(Fig. 1A). In a typical experimental trial (Fig. 1B), the owl was
placed on a perch in a large illuminated room and was confronted
with an extended open-eld stimulus that contained several visual
objects (oriented bars), one of which differed in its critical visual
feature (orientation). No specic task was given, and the owls
could freely view the scene in the room. We measured and ana-
lyzed the owls gaze that could be derived directly from the camera
view, given that eye movements are negligible.
Because barn owls lack a visible fovea (28, 29), their true gaze
direction cannot be resolved by optical and geometrical analysis,
and thus the OwlCam must be calibrated by other means after it
is mounted and xed to an owls head (Materials and Methods
and Fig. 1 B and C). This yields the functional xation point of
the owl in camera-frame coordinates, which serves as a reference
frame for all of our reported data. As discussed in Materials and
Methods, both of our owl subjects were found to have a similar
xation spot of 2.3 of visual angle and steep anks. In our
OwlCam image plane, this amounts to a disk of 25 pixels in di-
ameter, which was used in all of our subsequent analyses.
In the visual search experiment, the room contained 25 ori-
ented bars scattered on the oor in a 5 5 jittered conguration
(Fig. 2). Of these 25 bars, 24 barsthe distracters had a sim-
ilar orientation (up to a small jitter, to avoid possible confounds
from strict regularity), and one barthe targetwas placed at
a very different orientation (Materials and Methods). The owl was
allowed to freely view the scene without any prescribed task.
Once the owl oriented its head toward the scene (usually im-
mediately after stimulus onset), it clearly moved its gaze from
one bar to another, sometimes returning to a bar on which it had
Author contributions: W.M.H., J.O., O.B.-S., and H.W. designed research; W.M.H. built the
OwlCam; W.M.H. and J.O. performed research; W.M.H., J.O., O.B.-S., and H.W. analyzed
data; and W.M.H., O.B.-S., and H.W. wrote the paper.
The authors declare no conict of interest.
NEUROSCIENCE
This article is a PNAS Direct Submission.
1
Present address: School of Optometry, University of California, Berkeley, CA 94720.
2
To whom correspondence should be addressed. E-mail: wolf@bio2.rwth-aachen.de.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1101582108/-/DCSupplemental.
PNAS | May 17, 2011 | vol. 108 | no. 20 | 84618466
153
Fig. 1. (A) OwlCam (indicated by the white arrow) attached to the head of a barn owl. Only the antenna and the frontal part of the camera unit are visible.
The total weight of the setup, including the battery, is 5.5 g. (B) Fixation spot calibration procedure. Single xation images (Bottom) are binarized into target
and background regions (n = 1). When several binarized frames are accumulated into a single normalized map, a 2D probability density function of target
locations within the camera frame begins to form (n = 10). A marked xation spot emerges after several thousand xation images are processed (n = 10,000).
(C) Fixation map of one of our owls (WH) after calibration. Absolute occurrences of targets within each pixel (in camera frame coordinates) are color-coded
(compare Inset). In this example, 10,662 xations were accumulated to yield a peak probability at pixel 322/339 of the image (horizontal/vertical coordinates).
The 1D probability functions along both axes are given as well (bright lines).

already xated earlier. An example of a typical xation-saccade
behavior is shown in Movie S1. In a total of 97 experimental
trials, 120 min (217,309 frames) of OwlCam video material was
recorded and analyzed from the two owls. To present data cor-
responding to both owls in a simple and clear fashion, in what
follows we present all results using the notation HB/WH.
In 28/69 experimental trials, a total of 45/75 min (82,090/135,219
frames) of video material was recorded. The average length of
each trial was 97.7/65.3 s (2,931.8/1,959.7 frames). The experi-
mental video material was separated into segments of image mo-
tion and nonmotion (17), corresponding to the phases of head
movement (sacchades) and nonmovement (xations). Note that
by such a denition, an equal number of saccades and xations
must occur, because each saccade is bracketed by two xations and
vice versa. A total number of 985/1,236 xations were analyzed,
which lasted a total of 71,673/120,567 frames and averaged 72.76/
97.55 frames per xation. Thus, the owls spent 87%/92% of the
total recording time on xations, with an average xation time of
2.43/3.25 s. The average duration of a saccade was 0.35/0.28 s.
Because of the static nature of xation segments, only one
frame was used to represent the content of each xation. For our
analysis, we extracted the middle frame of each xation, on
which the xation spot of 25 pixels in diameter was marked for
further analysis. First, xations were classied into four classes
based on their visual content (Fig. 2):
i) Target xations, in which the target item appeared within
or immediately at the border of the xation spot mark
(with up to 1 pixel tolerance).
ii) Control xations, in which the control item appeared within
or immediately at the border of the xation spot mark (1
pixel tolerance). The position of the control item was de-
ned by mirroring the target position about the center of the
3 3 array, and thus was trial-dependent and changed its
position according to the target position in each trial.
iii) Frontal xations, directed toward the stimulus scene while
being neither target nor control xations. Frontal xations
include cases in which bars other than the target and the
control items were looked at, along with cases in which the
gaze was directed toward the stimulus but the animal x-
ated at no specic bar. In general, frontal xations were
the most frequent of all xations.
iv) Back xations, directed at locations where no bars were
visible within the entire camera eld of view. Such xations
were directed at the walls, ceiling, or door of the ex-
periment room.
For simplicity of presentation, we also designate all nonback
xations (i.e., the sum of target, control, and frontal xations) as
scene xations.
Out of all 985/1,236 xations recorded, 347/264 (35%/21%)
were back xations. Such xations occurred due to the owls
overall state of alertness (or lack thereof) and its natural visual
scanning behavior. Naturally, these xations did not show useful
visual content that could be used to examine xations on target or
distracter items, and thus they were excluded from further analysis.
The remaining scene xations were used for our main analy-
ses, as described next. Out of all scene xations, 7%/8% were
outside of the region covered by the 25 bar items.
For stimulus presentation, we divided the scene into two
compartments: the total region of 5 5 items and a central
subarray of 3 3 items. Targets were placed only within the
central 3 3 subarray of items, to avoid possible margin effects
due to the fact that items outside the central subarray did not
have neighbors on all sides. Interestingly, increased saliency for
the bordering items was not observed. Moreover, both owls had a
tendency to focus their overt attention to the central 3 3 sub-
array, where indeed 76%/62% of their scene xations were di-
rected. This bias toward the central subarray was conrmed in a
number of control trials in which no target bar was present and all
bars were oriented similarly, where 79%/60% of all scene xations
were directed to the central subarray. Thus, the expectations for
randomly hitting one of the central items would be 0.76/9 = 0.08
and 0.62/9 = 0.07. With this in mind, the proportion of target and
control xations out of the total number of scene xations was
calculated and compared for each experimental trial separately.
The mean proportion of target xations was 0.21/0.16, whereas
that of control xations was 0.07/0.08 (Fig. 3A). More quantita-
tively, the mean proportion of xations directed to the target was
more than twice as high as those directed to the control bar.
To demonstrate the difference in the number of xations di-
rected at the target and the control, we also counted the number
of xations for each category in each trial. For example, in the
trial shown in Fig. 2B, the owl looked at the target three times,
whereas the xation spot appeared at the control item two times.

8462 | www.pnas.org/cgi/doi/10.1073/pnas.1101582108 Harmening et al.

154
Fig. 2. (A) Classes of xations. The xation classes were determined by the
content of the frame in the xation spot (marked by the yellow circle) and
away from it. In target xation, the owl was looking toward the stimulus
array, and the target item (in this case, rotated by 45 compared with all
other items) appeared within or immediately at the border of the xation
spot. In control xation, the owl was looking toward the stimulus array, and
the control item appeared within (or immediately at the border of) the
xation spot. The position of the control item was dened by mirroring
the target position about the center of the 3 3 array. In frontal xation,
the owl was looking toward the stimulus array, but neither the target
nor the control items appeared within the xation spot. This was the most
frequent type of xation. In back xation, none of the foregoing held. In the
demonstration case, the owl simply looked at the door of the experiment
room. (B) An exemplary stimulus scene as reconstructed from many xation
frames (Materials and Methods). The walls of the experiment room are
visible in the far end and along the sides. The reconstructed scan path of the
owl during this experimental trial is denoted by circles (loci of xations)
connected by straight lines (saccades). Note how the owl repeatedly shifted
its gaze between stimulus objects and often returned to specic locations/
items. Regions in which target and control xations were registered are
highlighted. Back xations are not shown.
A Wilcoxon matched-pair signed-rank test including all trials
revealed a highly signicant difference in the target and control
xations per trial (P < 0.001 for both owls). To demonstrate this
perceptual advantage of the target over the control in another
way, we also plotted the data as cumulative probability dis-
tributions (Fig. 3B). These curves demonstrated a rightward shift
of the target distribution compared with the control distribution,
suggesting that on average, target xations were more numerous
per trial. For example, Fig. 3B shows that in the control item was
not xated in 36%/38% of the trials, whereas the target was not
xated in only 7%/13% of the trials. The same graphs also show
that the control item was never xated more than 6/5 times per
trial, whereas the target was xated much more frequently, as
Harmening et al.
Fig. 3. Target advantage over control in terms of number and total xation
time. (A) Mean proportion of xations on target (green, lled) and control
items (white, open) out of all frontal xations for both owl subjects (HB and
WH). The difference between target and control conditions was highly sig-
nicant in both cases (P < 0.0001, Wilcoxon matched-pair signed-rank test).
Error bars are SEM. (B) Normalized cumulative occurrences of xations di-
rected to the target (green line) and to the control (dashed line) plotted
against the number of xations. In both owls, the right shift of the target
graphs generally indicates more xations. (C) Mean proportion of time
spent on xating the target and control items out of all frontal xations/
differences between target and control conditions (color-coded as in A),
again highly signicant for both owls (P < 0.002). (D) Normalized cumulative
occurrences of frames directed to the target and to the control plotted
against xation time given in the frames for both owls. A right shift denotes
longer xations.
many as 16/10 times per trial. In summary, our analyses clearly
show that the target exhibited increased saliency for both owls.
We aimed to conrm the signicant difference of target se-
lection for oriented objects in the time domain by computing the
relative time spent on each of the frontal, target, and control
xations out of all scene xations. The duration of each frontal,
target, and control xation was determined directly from the
video recordings, accumulated by category, and divided by the
total time spent on all three categories. The mean proportion of
NEUROSCIENCE
time spent on target items was 0.40/0.34 and that spent on control
items was 0.20/0.22 (Fig. 3C), whereas the mean proportion of
frontal viewing without xation of either the target or the control
bar was 0.40/0.34. Notably, the difference between target xation
time and control xation time was highly signicant for the in-
PNAS | May 17, 2011 | vol. 108 | no. 20 | 8463
155
dividual trials as well. The average time per trial spent on target
and control items yielded 548/254.17 frames on target items and
148.25/137.88 frames on control items, a highly signicant dif-
ference (P < 0.01/< 0.001, Wilcoxon matched-pair signed-rank
test). The difference in viewing time between target and control
is also obvious in the cumulative probability plots, assembled in
a way analogous to those shown in Fig. 3B. The curves reecting
the xations at target are shifted to the right compared with the
control curve (Fig. 3D). In >95% of all trials, the owls looked at
the control for <500/400 frames, compared with 1,000/800 frames
for the target. Overall, the owls xated on the target bars more
often and longer than on the control bars.
Whether or not the target bar was more salient than other
items also may be reected in how fast it drew the owls view. We
counted the number of saccades from the onset of a stimulus
(lights switched on) until the target or control item was rst
looked at. For example, in the trial shown in Fig. 2B, the owl rst
hit the target with its fth saccade, whereas it hit the control with
its third saccade. When all data were averaged, the mean number
of saccades until the target was looked at was 3.92/3.17, and that
until the control was looked at was 15.5/5.93 (Fig. 4A). Although
both differences were statistically signicant (P = 0.0022/0.0129),
the distribution of the number of saccades until rst hit was
positively skewed for both target and control items; that is, there
were generally more observations below the arithmetical average.
Specically, in 67% of all cases, owl HB looked at the target item
after one (i.e., the rst xation was at the target; n = 7), two (n =
6), or three (n = 4) saccades, suggesting that the median numbers
of saccades might be more informative than the averages in this
case. Indeed, the median number of saccades until the target item
was rst looked at was 1.5/2, compared with 8.5/4 for the control
item. The difference between the target and control conditions
was highly signicant for owl HB and signicant for owl WH (P <
0.01; n = 17/P = 0.014; n = 38, Wilcoxon matched-pair signed-
Fig. 4. Target advantage over control in terms of the number of saccades
before rst hit. (A) Mean number of saccades until target (green, lled) and
control (white, open) were rst looked at. Black bars indicate the mean
number of saccades until food items were looked at in training. The dif-
ferences between target and control conditions were highly signicant for
owl HB (**P < 0.0001, Wilcoxon matched-pair signed-rank test), and signif-
icant for owl WH (*P = 0.014, Wilcoxon matched-pair signed-rank test). Error
bars are SEM. (B) Normalized cumulative occurrences of saccades until the
target (green line) and the control (dashed line) were rst looked at plotted
against the number of saccades. Both owls looked at the target much faster,
thereby causing a left shift in the cumulative plot. Note that in many trials,
the owls never looked at the control item, and thus its curve does not
converge to the 1.0 asymptote.
8464 | www.pnas.org/cgi/doi/10.1073/pnas.1101582108
rank test). The faster gazing toward the target than toward the
control also becomes obvious in the cumulative probability plots
for both owls shown in Fig. 4B. The curves reecting the number
of saccades to the target are shifted to the left compared with the
control curves. In other words, the target is reached with the rst
saccade in 25%/19% of the trials, compared with 0%/7% for the
control. Likewise, after 10 saccades, the target was reached in
86%/84% of the trials, whereas the control was reached in only
36%/49%. These last values indicate that the control was not
reached at all in a considerable number of trials. In summary, our
data show that the target was reached after a lower number of
saccades compared with the control bar.
During the training phase, both articial objects and food
items were scattered on the oor (Materials and Methods). Be-
cause food items should be the most salient visual objects pre-
sented in an experimental room, it is interesting to examine how
they drew the owls attention compared with our oriented targets
and distracters. On repeating the foregoing analysis, the mean
number of saccades to food was 2.5/3.17, compared with the
median number of saccades of 3/3 (Fig. 4A). No signicant dif-
ferences were found compared with the mean number of sac-
cades to our differently oriented target (P = 0.73/0.59). Thus, the
most salient visual items that we observed at all timesfood
itemswere looked at after the same number of saccades as the
target item that was dened by a different orientation.
Discussion
Our study of visual search in barn owls demonstrates that the
free-viewing animals looked longer, more often, and earlier at
differently oriented targets than at a control item, in a manner
resembling visual search in humans. The expression of orientation
saliency in visual search, demonstrated here in a bird species, raises
intriguing questions and has important implications regarding the
neural machinery that might be responsible for the observed be-
havior, the evolutionary relationship between birds and primates,
and the role of orientation-based saliency in efcient visual infor-
mation processing.
The predatory barn owl, with its specialization for hunting in
low-light conditions (30) needs to catch approximately two food
items (mainly mice) each day to survive and more than 20 a day
to feed its offspring. The selective pressure on these birds is es-
pecially high if weather conditions are unfavorable due to rain or
snow. Indeed, in central Europe, 60% of barn owl yearlings do
not survive their rst winter (31). Under such high selective
pressure, it would be to the animals advantage to exploit every
possible cue available to nd its prey. Indeed, barn owls are known
to be effective hunters (32), and thus exploiting even minute visual
cues is likely to be an intrinsic part of their visual behavior.
The evolution of different forms of saliency may be related to
the high selective pressure experienced by this bird. The orien-
tation saliency reported here could help the owl detect prey more
easily and more quickly. Pigeons are able to group bars of similar
orientation and discriminate the resulting gure from bars with a
different orientation (4). Pigeons also can detect odd objects
in a scene and even discriminate letters and faces (5, 7). Thus,
birds seem to have the neural machinery necessary for complex
scene analysis.
In primates, orientation-based saliency is facilitated by certain
neural circuitries, particularly those creating orientation selec-
tivity (33, 34). Long-range lateral connections found in the pri-
mary visual cortex (35, 36) have been shown to be important as
well (37). Orientation sensitivity in the barn owls visual Wulst is
very similar to that seen in the visual cortex (19, 20). The func-
tion of the horizontal long-range connections in mammals may
be accomplished in birds through the interconnectivity of many
telencephalic nuclei (see ref. 38 for a review). Moreover, within
the visual Wulst, organizational complexity increases as with
increasing latency of neuronal responses, indicating a hierarchy
Harmening et al.
156
of processing (39). Furthermore, attentional mechanisms were
found in the barn owl in cross-modal experiments that success-
fully used spatial attention to modulate sound localization (18).
Cross-modal competition in barn owls also was found to occur in
intermediate and deep layers of the optic tectum, a structure
known to be involved in gaze control and attention (21).
The fact that two such distant species as humans and barn owls,
whose brain structures are substantially different, exhibit similar
visual search characteristics has profound implications. In this
sense, our research is similar to a recent study on archer sh,
which have been shown to exhibit orientation-based saliency
similar to humans (16). Unlike Mokeichev et al. (16), however,
who explored orientation-based saliency using a rapid forced
choice procedure, our experiment was based on free-viewing vi-
sual search, reminiscent of the conditions under which this be-
havior is tested in humans. In both cases, bottom-up mechanisms
are likely to play the main role in the observed behavior (although
the effects of top-down inuence, and of some implicit unspecied
task, cannot be excluded), and in both cases the behavioral simi-
larities in the reported ndings suggest that visual processes, such
as orientation-based visual search, may not necessarily require the
elaborate cortical structures typically seen in humans. Unveiling
the neural mechanisms that facilitate these processes in animals
like the barn owl may provide important insight into saliency pro-
cessing in other organisms as well.
In humans, classical visual search experiments are also used to
discriminate between preattentive, pop-out, parallel processes
and serial attentive processes by measuring how target detection
time varies with the number of distracters. Pop-out also has been
demonstrated in pigeons (40). Owing to the slower response time
of the owls in our free-viewing paradigm, the difculty in spec-
ifying a specic task, and the indirect way in which the subjects
responses must be measured, exploring true pop-out in barn owls
is more challenging. However, comparing our visual search results
for the differently oriented targets and the food items reveals that
the both were looked at after approximately the same number of
head saccades. This comparable performance to the most de-
sirable target may indicate a perceptual popping-out of the dif-
ferently oriented target against the distracter array, which draws
the animals attention equally effectively and may serve as a rst
indication of the existence of the pop-out effect in barn owls. To
conrm this hypothesis and the equivalence of visual behavior
between barn owls and humans in general will require an exten-
sive and more challenging examination of visual search in barn
owls. This will be done by confronting the animals with search
stimuli presented on a monitor and by recording from brain
structures putatively involved in visual search (21). Such experi-
ments may shed light on whether theories hypothesized for human
visual search (10, 11, 13) can model this visual process in the barn
owl as well, or whether nature has found a different solution.
Regardless of how this question is resolved, however, the fact that
species as distant as humans and barn owls exhibit striking simi-
larities in a fundamental visual behavior like orientation-based
visual search suggests that orientation saliency has computational
optimality in a wide variety of contexts and provides a universal
building block for efcient visual information processing.
Materials and Methods
Animals. The experimental animals were two adult American barn owls (T.
alba pratincola; subjects WH and HB) that were taken from the breeding
stock of the Department of Zoology at Rheinisch-Westfaelische Technische
Hochschule Aachen. The birds were hand-raised and tame. The wingspan of
barn owls is 1.1 m (41). During the phase of experimentation, the owls
body weight was maintained at 90% of their free feeding weight (415 g
and 470 g). Water was given ad libitum, and food (dead chicken) was given
only in the experimental room or as a reward directly after an experiment.
Training and experiments were performed on ve or six days per week. For
each owl, a small aluminum headpost, to which the OwlCam was later at-
tached (see below), had been xed to the skull on the forehead under an-
Harmening et al.
esthesia at an earlier time. Care and treatment of the owls was carried out in
accordance with the guidelines for animal experimentation as approved by
local authorities (Landesprsidium fr Natur, Umwelt und Verbraucherschutz
Nordrhein Westfalen, Recklinghausen, Germany) and in compliance with
the National Institutes of Healths guideline for the use and care of labora-
tory animals.
Experimental Setup and Procedure. Experiments and training were performed
in a large room (4.2 m long 3.2 m wide 3.2 m high), in which the owls were
allowed to move and y freely. Moderate illumination was provided by
ceiling-mounted tungsten lights that were switchable from outside. To ach-
ieve sound attenuation, the walls, ceiling, and oor were covered with planar
and pyramidal foam. A wooden perch placed 1.75 m above the oor close to
one short wall of the room served as a resting post just before and in be-
tween experiments. A retractable curtain made from thick black cardboard
was placed in front of the perching position, such that the animals view to
the oor could be blocked until the experiment was started. The owls were
trained to y toward food items presented on the oor and to return to the
perch after a successful strike, with the captured prey as a reward. During
training, ights normally occurred after the experimenter left the room.
Training trials also were interleaved irregularly between experimental trials
at a ratio of about 1:5, to ensure high motivation and active viewing be-
havior of the owls in experiments, where no specic task was given.
Between experimental trials, the following procedure was performed. The
curtain was moved into place to block the owls view of the oor. The ex-
perimenter then entered the room and placed 24 distracter items on the
oor to cover the virtual intersections of a sparsely arranged and randomly
jittered 5 5 orthogonal grid (Fig. 2). The visual items were identical rectan-
gular bar-like shapes (150 50 mm) cut from thick yellow cardboard. One
additional item, dened as the target item, was differently oriented and
slanted by 45 relative to the dominant orientation of the distracters. The
target item was placed quasi-randomly and counterbalanced in one out of
nine possible positions at the area of a concentrically arranged 3 3 grid. In
this way, the target item never appeared at the immediate edge of the whole
stimulus array, and possible margin confounds were avoided. The experi-
menter left the room, lights were switched off, and the curtain was retracted
to allow a free view onto the stimulus array. The beginning of a trial was
dened as the time when the lights were switched back on and the owl started
to visually inspect the room. The owl was allowed to look around freely for
a maximum of 3 min, after which the trial ended. Usually, the owls would y
toward one of the visual items after a shorter period of inspection and, upon
entrance of the experimenter, retreat to the perching position. Approximately
515 consecutive trials were performed with one owl per day.
OwlCam and Video Analysis. During all experimental and training sessions, the
owls carried a head-mounted lightweight wireless camera device, the Owl-
Cam (SI Materials and Methods). The OwlCam consisted of a miniature
complementary metal-oxide semiconductor active-pixel sensor and optics
unit, a 900-MHz video broadcasting unit, a rechargeable lithium-polymer
battery, and a custom-built attachment unit (Fig. S1). While maintaining
high rigidity at a total weight of 5.5 g, the OwlCam delivered a black-and-
white video signal at 30 frames per second with an effective vertical reso-
lution of 380 scan lines. The video signal was digitalized online and stored
in a 640 480 pixel video format for further processing.
Using a custom-written algorithm, the raw video material was later di-
vided into frame segments of image motion and nonmotion (6). Segments of
consecutive frames in which no image motion occurred were dened as
xation intervals, and the middle frame of each interval was extracted and
used as the xation frame representing the whole interval. All subsequent
processing steps were based on these xation frames.
Each OwlCam was calibrated with respect to the relative geometric ar-
rangement of camera eld of view and the owls gaze, to localize the owls
functional xation point in camera frame coordinates (Results). In several
calibration trials, a xation map was constructed for each owl and OwlCam
pair separately. Note that xation maps are valid only for a specic owl
OwlCam pair, because of idiosyncratic differences of the owls head post
position, camera layout, and prealignment procedure. Based on this map,
a single xation spot relative to the camera frame coordinates was revealed,
as described in detail elsewhere (6). In brief, during calibration, interesting
NEUROSCIENCE
bright targets scattered on a dark oor were presented (Fig. 1B). The owl
typically scanned the environment by xating one target and then making
a saccade to another target, and so on. As more individual xation frames
were overlaid and averaged, a distinct circular-shaped xation spot emerged
for each owl (Fig. 1C). The resultant xation map reects the probability to
encounter a bright target in camera frame coordinates. The xation spot
PNAS | May 17, 2011 | vol. 108 | no. 20 | 8465
157
itself indicates the image coordinates at which the owl would observe
a bright target most often. Such a bright spot did indeed occur in the xation
map of each owl, at camera coordinates (335,322) for owl HB and (322,339)
for owl WH. The diameter of the xation spot was calculated as the mean
width of the probability function at half height and was 26.71 pixels for owl
HB and 22.41 pixels for WH (corresponding to 2.5 and 2.1 of visual angle,
respectively). In quantitative terms, calibration targets appeared within the
xation spot in 94% (n = 9,804) of all xations for owl HB and in 96% (n =
10,662) of all xations for owl WH. Once determined, the xation spot of
each owlOwlCam pair was used in the analysis of all video data. Note that
the size of the xation spot does not necessarily represent the actual size of
the animals retinal area of preferred xation, because it is linked to the size
of the calibration targets used. However, the calibration targets were set to
have similar size as the bar objects used in our main experiments. After cal-
ibration, the xation spot could be marked in each recorded xation frame
of the main experimental trials to serve as an estimate of where the owl was
looking relative to the camera coordinates. This conclusion is possible with
the barn owl, which virtually lacks eye movements (42, 43). Thus, the view of
a properly aligned and xed head camera is in register with the animals gaze
at all times.
Given the xed relationship between the OwlCam and the owls gaze, the
individual xation frames collected during our experimental trials were
1. Yarbus AL (1967) Eye Movements and Vision (Plenum, New York).
2. Einhuser W, Kruse W, Hoffmann K-P, Knig P (2006) Differences of monkey and
human overt attention under natural conditions. Vision Res 46:11941209.
3. Mazer JA, Gallant JL (2003) Goal-related activity in V4 during free viewing visual
search: Evidence for a ventral stream visual salience map. Neuron 40:12411250.
4. Cook RG, Cavoto KK, Cavoto BR (1996) Mechanisms of multidimensional grouping,
fusion, and search in avian texture discrimination. Anim Learn Behav 24:150167.
5. Blough DS (1977) Visual search in the pigeon: Hunt and peck method. Science 196:
10131014.
6. Ohayon S, Harmening WM, Wagner H, Rivlin E (2008) Through a barn owls eyes:
Interactions between scene content and visual attention. Biol Cybern 98:115132.
7. Dittrich L, Rose J, Buschmann J-UF, Bourdonnais M, Gntrkn O (2010) Peck tracking:
A method for localizing critical features within complex pictures for pigeons. Anim
Cogn 13:133143.
8. Krummenacher J, Mller HJ, Deubel H, Wolfe JM, Humphreys GW (2010) Editorial:
Visual search and selective attention. Vision Res 50:13011303.
9. Wolfe JM, Horowitz TS (2004) What attributes guide the deployment of visual
attention and how do they do it? Nat Rev Neurosci 5:495501.
10. Wolfe JM (1994) A revised model of visual search. Psychon Bull Rev 1:202238.
11. Hochstein S, Ahissar M (2002) View from the top: Hierarchies and reverse hierarchies
in the visual system. Neuron 36:791804.
12. Treisman AM, Gelade G (1980) A feature-integration theory of attention. Cognit
Psychol 12:97136.
13. Lee TS, Yang CF, Romero RD, Mumford D (2002) Neural activity in early visual cortex
reects behavioral experience and higher-order perceptual saliency. Nat Neurosci 5:
589597.
14. Hegd J, Felleman DJ (2003) How selective are V1 cells for pop-out stimuli? J Neurosci
23:99689980.
15. Nothdurft H-C, Pigarev IN, Kastner S (2009) Overt and covert visual search in primates:
Reaction times and gaze shift strategies. J Integr Neurosci 8:137174.
16. Mokeichev A, Segev R, Ben-Shahar O (2010) Orientation saliency without visual cortex
and target selection in archer sh. Proc Natl Acad Sci USA 107:1672616731.
17. Ohayon S, van der Willigen RF, Wagner H, Katsman I, Rivlin E (2006) On the barn owls
visual pre-attack behavior, I: Structure of head movements and motion patterns. J
Comp Physiol A Neuroethol Sens Neural Behav Physiol 192:927940.
18. Johnen A, Wagner H, Gaese BH (2001) Spatial attention modulates sound localization
in barn owls. J Neurophysiol 85:10091012.
19. Liu GB, Pettigrew JD (2003) Orientation mosaic in barn owls visual Wulst revealed by
optical imaging: Comparison with cat and monkey striate and extra-striate areas.
Brain Res 961:153158.
20. Pettigrew JD, Konishi M (1976) Neurons selective for orientation and binocular
disparity in the visual Wulst of the barn owl (Tyto alba). Science 193:675678.
21. Mysore SP, Asadollahi A, Knudsen EI (2010) Global inhibition and stimulus
competition in the owl optic tectum. J Neurosci 30:17271738.
22. Reches A, Gutfreund Y (2008) Stimulus-specic adaptations in the gaze control system
of the barn owl. J Neurosci 28:15231533.
8466 | www.pnas.org/cgi/doi/10.1073/pnas.1101582108
ego-centered and perspective-limited representations of the owls view at
a given time. To study the owls viewing behavior as a consequence of the
global visual stimulus, the complete visual scene had to be taken into ac-
count. However, to a good approximation, the xation frames could be
considered limited-view images of the same exterior setting observed from
a specic vantage point at different viewing angles, resulting from the owls
head movements while keeping its body relatively xed. Thus, a full-scene
reconstruction was achieved by spatial transformation and alignment of the
individual xation frames to build a panoramic view of the entire scene as
would be observed by a wide-angle observer at the vantage point (44). By
mapping the coordinates of the spatially transformed loci of xations in
each single xation frame to the corresponding coordinates in the full-scene
panoramic image, global scan paths were created and analyzed. The owls
viewing behavior was then studied with respect to three main criteria: (i)
relative and absolute gaze time spent at a specic location, (ii) relative and
absolute number of xations directed onto such locations, and (iii) number
of head saccades performed until the locations were rst looked at.
ACKNOWLEDGMENTS. O.B.-S. thanks the generous support of the Zlotowski
center, the Frankel Fund, and the Paul Ivanier Robotics Center at Ben-Gurion
University. This study was funded by Deutsche Forschungsgemeinschaft
Grant WA606/17-1.
23. Zahar Y, Reches A, Gutfreund Y (2009) Multisensory enhancement in the optic tectum
of the barn owl: Spike count and spike timing. J Neurophysiol 101:23802394.
24. Harmening WM, Gbbels K, Wagner H (2007) Vernier acuity in barn owls. Vision Res
47:10201026.
25. Harmening WM, Nikolay P, Orlowski J, Wagner H (2009) Spatial contrast sensitivity
and grating acuity of barn owls. J Vis 9:112.
26. Nieder A, Wagner H (1999) Perception and neuronal coding of subjective contours in
the owl. Nat Neurosci 2:660663.
27. Wagner H, Frost B (1993) Disparity-sensitive cells in the owl have a characteristic
disparity. Nature 364:796798.
28. Wathey JC, Pettigrew JD (1989) Quantitative analysis of the retinal ganglion cell layer
and optic nerve of the barn owl Tyto alba. Brain Behav Evol 33:279292.
29. Oehme H (1964) Vergleichende untersuchungen an greifvogelaugen. Z Morphol
Oekol Tiere 53:618635.
30. van der Willigen RF, Harmening WM, Vossen S, Wagner H (2010) Disparity sensitivity
in man and owl: Psychophysical evidence for equivalent perception of shape-from-
stereo. J Vis 10:111.
31. Mebs T, Scherzinger W (2000) Die Eulen Europas: Biologie, Kennzeichen, Bestnde
(Franckh-Kosmos, Stuttgart).
32. Ilany A, Eilam D (2007) Wait before running for your life: Defensive tactics of spiny
mice (Acomys cahirinus) in evading barn owl (Tyto alba) attack. Behav Ecol Sociobiol
62:923933.
33. Supr H, Spekreijse H, Lamme VA (2001) Two distinct modes of sensory processing
observed in monkey primary visual cortex (V1). Nat Neurosci 4:304310.
34. Kastner S, Nothdurft HC, Pigarev IN (1997) Neuronal correlates of pop-out in cat
striate cortex. Vision Res 37:371376.
35. Hubel DH, Wiesel TN (1977) Ferrier Lecture: Functional architecture of macaque
monkey visual cortex. Proc R Soc Lond B Biol Sci 198:159.
36. Rockland KS, Lund JS (1982) Widespread periodic intrinsic connections in the tree
shrew visual cortex. Science 215:15321534.
37. Ben-Shahar O, Zucker S (2004) Geometrical computations explain projection patterns
of long-range horizontal connections in visual cortex. Neural Comput 16:445476.
38. Jarvis ED, et al.; Avian Brain Nomenclature Consortium (2005) Avian brains and a new
understanding of vertebrate brain evolution. Nat Rev Neurosci 6:151159.
39. Nieder A, Wagner H (2001) Hierarchical processing of horizontal disparity information
in the visual forebrain of behaving owls. J Neurosci 21:45144522.
40. Blough D (1992) Cognitive Aspects of Stimulus Control, eds Honig WK, Fetterman G
(Psychology Press, Lawrence Erlbaum Associates, Inc., Hillsdale, NJ), p 456.
41. Bachmann T, et al. (2007) Morphometric characterisation of wing feathers of the barn
owl Tyto alba pratincola and the pigeon Columba livia. Front Zool 4:23.
42. Steinbach MJ, Money KE (1973) Eye movements of the owl. Vision Res 13:889891.
43. du Lac S, Knudsen EI (1990) Neural maps of head movement vector and speed in the
optic tectum of the barn owl. J Neurophysiol 63:131146.
44. DAngelo P (2010) Hugin. Available at http://hugin.sourceforge.net/. Accessed
March 16, 2010.
Harmening et al.
158
 The Neural Basis
of Visually Guided Behavior
Techniques from ethology and neurophysiology are combined
to show how an animal localizes a visual object, discriminates
its significance and then makes the appropriate motor response
A nimals
ft
see things and then act on the
basis of what they see. What
chain of events connects some key
stimulus with a specific fixed pattern of
responses? In recent years workers in
several laboratories have sought by many
different means to analyze the nerve
mechanisms by which animals interpret
sensory. signals and select the most ap
propriate response. The most effective
way to understand the neural basis of be
havior appears to be to apply a broad
spectrum of experimental techniques: to
combine ethological studies of an ani
mal's behavior with experiments involv
ing brain anatomy and brain-cell stimu
lation and the recording of individual
nerve-cell activity.
For the past six years, in my labora
tory first at the Technical University of
Darmstadt and then at the University of
Kassel, we have been taking this broad
approach to learn about two kinds of
visually controlled behavior in the toad:
orienting (prey-catching) behavior and
avoidance (escape) behavior. There are
several good reasons for working with
the toad. Amphibians are vertebrates,
so that what we learn at their relatively
low level of behavioral integration con
tributes to our understanding of more
complex vertebrate functioning. Toads
in particular have a limited and easily
surveyed behavioral repertory. In re
sponse to specific stimuli one can repeat
edly elicit predictable reactions, such as
snapping at prey, fleeing from an enemy,
clasping during courtship and making
particular wiping motions after tactile
stimulation. (The fickle European frog,
in contrast, undergoes short-term chang
es in motivation and is not suitable for
behavioral experiments.) Finally, the
toad is not easily conditioned, so that its
34
by Jorg-Peter Ewert
innate behavioral functions can be mea
sured in successive experiments for some
time without being significantly affected
by accumulated experience.
Ta
oads respond to small objects, such as
piece of white cardboard moved
over a black background, with a series of
prey-catching reactions. First there is
orientation toward the prey, then binoc
ular fixation, then snapping, gulping and
mouth-cleaning. Two basic processes are
required to produce the overall orienting
reaction: the identification of a stimulus
and the location of it in space. The iden
tification process determines the type of
behavior. I t is dependent on specific fea
tures of the stimulus such as its angular
size, the orientation of the boundaries be
tween light and dark, its angular veloc
ity, its contrast with the background and
so on. A detection process then localizes
the stimulus and, together with the result
of the identification process, determines
the motor response, which can be either
to turn toward the stimulus if it is iden
tified as prey or to avoid it if it appears
to be an enemy. In what follows I shall
attempt to analyze the neurophysiologi
cal basis of signal identification, localiza
tion and the triggering of the associated
instinctive actions.
To begin one must analyze quantita
tively the key stimuli for orienting and
avoidance behavior. This is done by
changing various characteristics of a vi
sual stimulus in an ordered way. The toad
is placed in a cylindrical glass compart
ment where it observes a small square of
black cardboard moving against a white
background at a constant angular veloc
ity, describing a circle around the animal
at a distance of seven centimeters. The
toad interprets such a stimulus as prey
1974 SCIENTIFIC AMERICAN, INC
and tries, through successive turning
movements, to keep the object fixated in
the center of its visual field. The degree
of orienting activity is measured by
counting the number of turning re
sponses per minute.
The angular size of the stimulus-the
angle it subtends-influences the orient
ing activity [see illustration on page 36].
Of a variety of square objects toads pre
fer those with an edge length of four to
eight degrees. (The absolute size of such
stimuli is five to 10 millimeters. Experi
ments where the distance between ani
mal and stimulus is varied show that it is
the absolute-not the angular-size that
counts; in prey-catching behavior toads
display "size constancy.") The toads turn
away from objects larger than 30 degrees
on a side, exhibiting the avoidance re
sponse. More particular information is
obtained by substituting bars of various
lengths for the square stimuli. As a two
by-two-degree stimulus is elongated
along the horizontal axis the orienting
activity increases until a saturation level
is reached; wormlike objects turn out to
be particularly attractive to toads. In
contrast, the response decreases as a
small stimulus is extended vertically, or
perpendicularly to the direction of move
ment.
Other experiments indicate that toads
discriminate prey from enemy objects
through analysis of the visual stimulus in
terms of point or edge configurations,
also taking into consideration the direc
tion of movement. A horizontal chain
consisting of several two-by-two-degree
units moving along the same path signi
fies prey. One such unit moving alone
constitutes a prey stimulus just above the
response threshold. When the horizontal
chain is supplied with a separate vertical
159
 d

b e

c f

BEHAVIORAL PATTERNS characteristic of the toad Bulo bulo
are illustrated. The actions are commonly elicited in the animal by
the sight of visual objects. These drawings, however, are hased on
photographs of toads whose brains were being stimulated electri
cally as part of the author's investigation of the neural bases of vi
sually guided behavior. The electrode on the toad's bead penetrates
to the brain. An electric current applied to the optic tectum, a visu.
al center in the brain, elicits a prey.catching sequence: orienting,
or turning (a), snapping (b) and mouth.cleaning (c). Electrical
stimulation, instead, of a site in the left or right thalamus brings a
"plantingdown" defensive posture (d, e) and stimulation of anoth
er part of the thalamus brings a crouching avoidance response (f).

35

160
1974 SCIENTIFIC AMERICAN, INC
 extension (making it in effect an L
shaped structure moving on its long
side), it loses efficiency as a prey-catching
stimulus. The inhibitory effect of the ver
tical extension depends on its distance
from the horizontal element. If a second
vertical extension is introduced, in effect
making the stimulus a shallow U-shaped
structure, the total configuration signifies
"enemy." The ethological interpretation
is that it symbolizes a "swarm," and in
the toad's brain inhibitory interactions
first restrain prey-catching behavior and
then induce escape behavior.
For constant form and angular veloc
ity the behavioral activity generally in

STIMULUS NORMAL TOAD TOAD WITHOUT THALAMUS

40
creases as the amount of contrast be
tween stimulus and background increas
es. White objects moving against a black

1'\

(TURNS TOWARD) ./
v background are normally more attractive
as prey than black objects on white; the
latter, on the other hand, are more ef

SQUARE SIDE OF SQUARE (DEGRE S) SIDE OF SQUARE (DEGRE S)

w

::J
o L-_-'-_-'-_-'----_--'-"--"--'---_--.J
2 4 8 16 32 2 4 8 16 32
fective in eliciting avoidance behavior.
vVhen the size and contrast are held con
stant, behavioral activity increases with
increasing angular velocity, reaching a
Z

maximum at between 20 and 30 degrees
a: 40 per second. Stationary objects usually
w
elicit no prey-catching or avoidance re

I
Il.
<Jl
W sponse. The common critical feature for

<Jl
20 key stimuli representing both prey and
Il.
enemy is movement, and the two kinds
VERBTAIRCAL HEIGHT (DEGRE S) HEIGHT (DEGRE S)
ill
a:

Z
a:
::J
0
1
l"-t'-
2 4 8 16 32 2 4 8 16 32
of stimulus are differentiated primarily
on the basis of their form: extension of
the object in the horizontal direction of
the movement generally means prey,
40 whereas extension perpendicular to the
V
f-- +-- direction of the movement signifies "not

V VV
t--""'"
prey" or "enemy."
20

HORBIZAORNTAL V ,V hat does the toad's eye tell the

toad's brain? This question was first
o formulated for the frog and dealt with

B
II
<{
LENGTH (DEGRE S) LENGTH (DEGRE S)
1 2 4 8 16 32 2 4 8 16 32 in the fascinating research of Jerome Y.
Lettvin and his colleagues at the Massa
chusetts Institute of Technology, and
was later investigated quantitatively by

a:
O.-J. Grusser and his co-workers at the

!

d Free University of Berlin. To ask the

DOBUAABRLE DISTANCE (DEGRE S)

w
<Jl
Z
0
Il.
<Jl
w
a:
0
",
2 4 8
d
./
,/'"
16 32
,......

2
DISTANCE (DEGRE S)
4 8

d
16 32
question is to open the "black box" of the
toad's brain, or at least to examine the
brain functions that participate in trans
forming input from visual stimuli into
relevant behavioral patterns. At this
point I shall describe neurophysiological
BEHA VIORAL RESPONSES of the toad to objects of various shapes and sizes were quan findings concerning whether it is in the
tified. Small black objects were moved across the visual field at seven centimeters' distance retina of the toad's eye that the key
and the orienting response was determined for normal toads (left) and those whose thala stimuli "prey" and "enemy" are encoded.
mus had been removed (right). Prey-catching responses (turning toward the object) were In the toad retina there are three types
elicited most effectively in normal toads by squares with sides subtending four to eight de
of ganglion cells that send their fibers by
grees; the toads turned away from larger squares. Vertical bars were ineffective as prey ob
way of the optic nerve to the structure
jects-and increasingly ineffective with increasing height. Horizontal (wormlike) bars were
called the optic tectum in the midbrain.
increasingly effective as prey ohjects with increasing length, up to a limit. Double bars (a
One can record the action potentials, or
horizontal bar plus a vertical extension) were less attractive, the effect varying with distance
between bars; the ratio of their effect to that of a single bar is shown (bottom). In toads nerve signals, from the ends of these fi
lacking the thalamus the orienting response becomes "disinhibited." The animal tends to bers by introducing a microelectrode into
orient toward a target without discrimination, even if the target normally signals "danger." the tectum. John E. Dowling, then at

36

161
1974 SCIENTIFIC AMERICAN, INC
a <E--RECEPTIVE FIELD----;::;' c 40 r---,----.---.---.,---.

r'vL:RINE GA"NGLION-CEL
SQUARE o 0 L-
Z

8w
1
l.-__l.-__-L-__-'-_-'-_-'
__

SIDE OF SQUARE (DEGRE S)

2 4 8 16 32

\!!CAENLGLION
40r---.---,----r-----.

L @ EFXIECLITDA(TEORFY) VERTICAL BAR "" '-

t
w

I-@
II-
b en
w

R E C E P T I V E FIELDS ?\

::::l
II
:;

w
O
20f----t..o''''--+----+--l---l

-' -'- --'- --

J H E I G H T ( D E G R E S )
o 2 4 8 16 32

o
II
en

CGEALNGALXlON I N H I B I
F ELD (I F)T O R Y w
II:
40

--II
--@
--.,.,...
20
-

HORIZONTAL BAR LENGTH (DEGRE-IS)V ... -- -- . ---

r MICRO
o
1 2 4 8 16 32

ELE T DE B

II

I
ADOUBLE@ dt II:
..:
e::.

BARIRF DISTANCE (DEGRE S)

w
en
z
0
II-
en
w 0
II: 2 4 8 16 32
d

NEURAL RESPONSES of the toad to the same objects were mea
sured with recording electrodes. Tbe electrodes recorded impulses
at the terminals in the optic tectum of fibers from individual gan
glion cells, the cells in the retina of the eye on which signals from
the receptor cells converge via intermediate cells (a). Each gan
glion cell has an excitatory receptive field surrounded by an in
hibitory receptive field. The diameter of the excitatory fields and
the strength of the inhibitory surrounds are different for each of
three classes of ganglion cells (b). For the square object and the
vertical bar (which the ganglion cell "confuses"), maximum activi
ty is elicited when the size of the object matches the excitatory-field
size of each type of ganglion cell (c)_ Horizontal length does not
much affect these cells' response. Vertical extension of a horizontal
bar has less effect on these cells than on behavior (opposite page)_

the Johns Hopkins University School of
Medicine, showed through electron mi
crography that in the frog (or toad) retina
each ganglion cell is connected to a num
ber of receptor cells by bipolar and ama
crine cells. Each ganglion cell is thus fed
information from a particular part of the
animal's visual field. Lateral connections
established by horizontal and amacrine
cells play a role in determining the prop
erties of this receptive field. In toads as
well as frogs the field consists of a cen
tral circular excitatory receptive field im
mediately surrounded by an inhibitory
receptive field. The movement of an ob
ject through the excitatory field elicits a
ganglion-cell discharge, which is inhibit
ed if another object is simultaneously
moving through the inhibitory field. The
three ganglion-cell types in the toad (as
in the frog) differ in several character-
istics, including in particular the diame
ter of their excitatory receptive fields:
about four degrees for the so-called Class
II ganglion cells, about eight degrees for
Class III cells and from 12 to 15 degrees
for Class IV cells. (Class I cells have been
identified in frogs but not in toads.)
With microelectrodes we measure the
rate of ganglion-cell discharge to see how
it changes when objects (corresponding
to those in the behavioral experiments
described above) are moved through the
receptive field of the cell. The impulse
frequency increases with the length of
the side of a square object until the
length about equals the diameter of the
excitatory field; then it decreases as the
object becomes large enough to stimulate
part of the surrounding inhibitory field.
In accordance with the different sizes of
the excitatory fields the maximum activa-
tion of each cell type is therefore elicited
by objects of different sizes [see illustra
tion above J. Extending a small square
horizontally (making it a "worm") does
not bring about any change in nerve-cell
activation; this is in sharp contrast to the
previously noted effect of extension on
the behavioral response. The depen
dence of neuronal activation on the size
of the stimulus is instead primarily a
function of extension perpendicular to
the direction of movement. Indeed, the
discharge frequency is almost exactly the
same in response to a narrow vertical bar
as it is to a square with the same height
as the bar. A retinal ganglion cell "con
fuses" the two stimuli-but the toad does
not: the square excites behavioral activ
ity and the bar inhibits it. When the ob
ject size is held constant, however, the
dependence of the discharge rate on con-

37

162
1974 SCIENTIFIC AMERICAN, INC
trast between stimulus and background
and on angular velocity is the same as it
is in the behavioral experiments.
In summary, it is clear that the first
important operations on the visual input
from a prey stimulus or a threatening one
are performed by the toad retina. For
any particular prey or enemy stimulus
the behavioral response to velocity and
background contrast seems to depend on
information processing in the retina. The
size-dependent excitatory and inhibitory
processes, however, which were noted
the behavioral experiments and which
play an essential role in pattern discrimi
background contrast or from larger size?
The differentiation can be made only if
separate groups of cells receive different
inputs from different optic-nerve fibers.
In fact they do. The fibers of the optic
nerve pass from each eye through the
optic chiasm to the opposite side of the
brain, ending in various parts of the fore
brain and midbrain. Two of these desti
nations are of particular interest in our
work. One, to which most optic-nerve
fibers project, is in the surface layers of
the optic tectum in the midbrain. The
other is in the thalamus and the pretectal
region of the diencephalon.
27 degrees that are activated exclusively
by moving objects. These neurons prob
ably represent a localization system. This
supposition is reinforced by experiments
in which we stimulate the tectum of free
ly moving toads with trains of impulses
delivered by means of an implanted elec
trode. Stimulation of a given region of
the tectum always causes toads to turn
toward a particular part of the visual
field. Presumably the neurons we are
thus activating have a direct connection
with the animal's motor system, since (in
contrast to the natural orienting move
ments made in response to a prey object)

T he optic tectum constitutes a locali-

nation, cannot be traced to the influence
of the excitatory and inhibitory fields of
retinal ganglion cells. There are no reti
nal "worm-detectors" as distinct from
"enemy-detectors." The differential anal
ysis, and thus the behaviorally relevant
interpretation of the stimulus, must be
achieved in nerve-cell populations be
yond the retinal level.
Since different characteristics
coded by any one type of ganglion cell
the question becomes: Where is that cod
ing interpreted? What tells the central
nervous system whether an increased
rate of ganglion-cell firing stems, for ex
ample, from an increase in stimulus-
zation system. In the tectum there is
an exact topographical mapping of the
retina and hence of the entire visual field.
Movement of an object in a particular
part of the visual field excites a corre
sponding region of the tectum, where the
appropriate optic-nerve fibers terminate
are [see illustration on these two pages]. Re
cording from individual tectal neurons,
or nerve cells, tells one how the indi
vidual retinal ganglion cells that excite
them are reacting. In certain layers, for
example, there are tectal neurons with
excitatory receptive fields of about 10 to
the electrically induced orienting is not
disrupted by simultaneous presentation
of a threatening object.
If the recording electrode is driven
deeper into the tectum, it encounters
neurons with larger receptive fields.
Some of these cover the entire visual field
on the opposite side, some the entire
lower part of the field and some the en
tire field directly in front of the animal.
Interestingly enough, all three types in
clude the fixation point: the point of max
imum visual acuity near the center of the
visual field. The degree of activation of
these three types of neurons could pro
vide the toad with information about the

a VISUAL FIELD b VISUAL FIELD c VISUAL FIELD

CHIOPTIASMC MEDIAN
10 @
@@ @
-----'-----0'1...:

OPTIC OFSURFTECTUMACE
TECTUM
SPINAL CORD
BACK AVOIDANCE ORIENTIN(
__ H

I
NEUROPHYSIOLOGICAL EXPERIMENTS yield data on the
functions of different parts of the toad's visual system. Fibers of the
retinal ganglion cells project primarily to the optic tectum and to
the thalamus (a). The visual field of each eye is mapped (b), on a
one-to-one basis (numbers), on the dorsal surface of the opposite
side of the tectum_ By the same token experimental electrical stim-
ulation (c) of various parts of the tectum (letters) causes the toad
to turn to corresponding parts of the visual field. On the other
hand stimulation (black disks) of the thalamus, which partially un
derlies the tectum, causes the opposite action: avoidance, or turn
ing away. As a recording electrode penetrates below the surface of
the tectum it encounters successive populations of cells with differ-

38

163
1974 SCIENTIFIC AMERICAN, INC
location of a large object, since whenever
the three types are excited simultaneous
ly the object must be at the fixation
point.
In natural situations the behavior of
toads can be influenced by sensory mo
dalities other than vision. If, for exam
ple, a beetle crosses the field of vision,
the toad's orienting reaction can be ei
ther accelerated or retarded by simul
taneous vibratory and tactile stimuli.
Such results can be obtained in experi
ments if prey models are presented to
gether with acoustic or tactile stimuli.
The area for producing such changes in
behavioral activity seems to be in the
subtectal region, where multisensory in
tegration is achieved. In the area below
the third ventricle of the midbrain there
are large-field neurons with fields similar
to those of the large-field tectal cells.
These subtectal neurons receive addi
tional inputs from neurons excited by
tactile and vibratory stimuli. The "mech
anoreceptive" field of one of these bi
modal neurons is always localized on the
same side as the visual receptive field.
The additional inputs from nonvisual
neurons could serve to lower the thresh
old of a part of the visual field in which
a visual stimulus is anticipated and thus
VISUAL FIELDRECEPTIVE FIELDS
d
GANGL I O N CEL L
TECTAL
FIELD CELSMALLS L\ '
"'
\ ..:
TECTALL A RGE
F I E L D CEL L
II
II ; j=:
\ illIlIlIlIlI!l
ent receptive fields (d). There are small-field cells with fields a little larger than those of
ganglion cells and, lower down, three kinds of large-field cells, each with different coverage
(color, horizontal hatching and vertical hatching) _ A drawing based on a stained brain sec
tion indicates the layers at which each of these is found. The final drawing (e) relates the
various cell populations and shows another layer of large-field cells that receive inputs from
visual cells above them and also from cells that respond to tactile or vibratory stimuli.
in effect to raise the level of visual alert-
ness.
The optic tectum also comprises
neuronal system that processes behav
iorally relevant aspects of moving stimuli
[see illustrations on next page]. The cells
have excitatory receptive fields about 27
degrees in diameter. Those designated
Type I tectal neurons are activated main
ly if the stimulus surface of an object
moved through the receptive field is ex
tended in the direction of movement;
extension perpendicular to direction of
movement does not have the same effect.
Other cells, the Type II tectaI neurons,
differ from Type I neurons in that their
discharge rate actually diminishes with
surface extension perpendicular to direc
tion of movement. The response of these
neurons constitutes the key stimulus
"prey." That is, they can presumably be
considered the trigger system for the
prey-catching response.
T ond major destination of fibers from
he thalamic-pretectal region, the sec-
the retina, apparently provides what can
be called a "caution" system. I have re
cently identified four main types of visu
ally sensitive neurons in the toad's thala
mus by means of single-cell recordings.
R'C'''OR5
e Large objects extended perpendicularly
, , , , , BIPOLARS
2\!! AMACRINES
z
i=
\1\11 GANGLION CELLS
w neurons in turn inhibit the tectum and
a::
3 illustmtion on page 41].
SM''TECTAL
'-'''''' cms
---
"RG,"TECTAL,"," cm5
5
I I I SUBTECTAL LARGE
-I-r-g FIELD CELLS
- --'
VIBRATITACTINEURONS
ONLESENSIAND TIVE
7
8 ed by the stimulation of cells in the thala
11
1974 SCIENTIFIC AMERICAN, INC
They are activated respectively by four
distinct stimulus situations: (1) move
a ment of enemy objects extended perpen
dicularly to the direction of motion, ex
citatory receptive field of about 46 de
grees; (2) movement of an object toward
the toad, field about 90 degrees; (3) large
stationary objects, field about 45 de
grees; (4) stimulation of the balance sen
sors in the toad's ear by tilting. In gen
eral these thalamic neurons are activated
principally in situations that tend to call
for evasive movements-turning away
from an enemy, sidestepping or compen
sating for tilting of the body. Brain
stimulation experiments support our feel
ing that the thalamic-pretectal region is
one in which reactions can be assembled
that lead to protective movements. Elec
trical stimulation of various sites in the
region elicits the following reactions:
closing of the eyelids, ducking, turning
away, panicky springing away or tilting
of the body.
We constructed a working hypothesiS
involving connections between the optic
tectum and the thalamic-pretectal re
gion: Electrical triggers in the tectum
mainly elicit orienting, and triggers
in the thalamic-pretectal region elicit
avoidance. In a natural situation trigger
impulses in particular layers of the tec
tum are evoked by small wormlike prey.
to the direction of movement stimulate
particular neurons in the thalamic-pre
tectal region, both directly through ret
inal inputs and indirectly by way of the
optic tectum. These thalamic-pretectal
can also activate avoidance behavior [see
The existence of the postulated con
nections between the structures in the
midbrain and the diencephalon has been
demonstrated physiologically in two
ways. One way is by direct electrical
stimulation. Thalamic neurons that are
sensitive to movement can also be acti
vated by stimulation of points in the
optic tectum. When the stimulating and
recording electrodes are interchanged,
the response of Type II neurons in the
tectum to moving objects can be inhibit
mus. The other way is by surgical opera
tion: if the optic tectum is removed,
orienting movements are lost-and so are
avoidance reactions, which is evidence
for pathways from the tectum to the
thalamus. If the thalamic-pretectal re
gion is removed without damage to the
tectum, then avoidance behavior is lost
and the orienting response is dramatical
ly freed from inhibition even in the pres
ence of enemy objects; this may be
39
164
 T Y P E I T E C T A L NEURON THALAMIC NEURON

EJl "'II'I II 'I
I 1I\\\\\\Wlll,
8
-
I

8
l<E- :
I
,.

FEATURE DETECTION beyond the retinal level is accomplished

by cells in the tectum and the thalamus. Recordings from individ.
ual cells indicate that tectal Type I neurons (left) are most activated
I
8
if the object moving through the field is extended in the direction
of movement. The cells in the thalamic area (right) respond most
to an object extended perpendicularly to direction of movement.

evidence for the existence of inhibitory
pathways from the thalamus to the optic
tectum. In toads lacking the thalamic
pretectal region every moving stimulus
elicits the orienting movements; the cau
moved, the disinhibition extends to the
entire visual field on the opposite side;
small lesions in the thalamic-pretectal re
gion affect only local small parts of the
visual field. Quantitative experiments
II tectal cells to moving stimuli shows a
similar "disinhibition" effect after tha
lamic-pretectal removal [see illustrations
on page 36 and below].

tionary thalamic-pretectal system, which
ordinarily allows orientation toward the
stimulus only in behaviorally appropriate
situations, is missing. If one lateral half
of the thalamic-pretectal region is re-
with toads lacking the thalamic-pretectal
region make it clear that these animals
cannot discriminate between stimuli that
are behaviorally relevant and those that
are irrelevant. The response of the Type
T suggest
he findings I have described so far
the following sequence of
events: On the basis of retinal ganglion
cell input, the optic tectum tells the toad
where in the visual field a stimulus is

TYPE(NIOTREMCATLALTONAEDU)RON ( TOYAPDEWII THEOCTUALTNHEAULARMOUNS)

\ 1:11

TRIGGER UNITS for the entire preycatching response seem to be
the Type II tectal nenrons. In the normal toad (left) the cells are
most activated by wormlike objects (horizontal bar). They are less
activated {and the decrease is greater than in the case of Type I
- -8
1 -8
tectal neurons) by stimuli that in behavioral experiments are irrel
evant for prey.catching (vertical bar). After removal of the thala
mus, however (right), their response to those irrelevant stimuli is
greatly increased, suggesting that the thalamic signal is inhibitory.

40

165
1974 SCIENTIFIC AMERICAN, INC
situated, how large it is, how strongly it
contrasts with the background and how
fast it is moving. The connections from
the tectum to structures in the thalamic
pretectal region enable the toad to dis
cern the significance to its behavior of
the visual signals. The basic filtering
process for the prey-enemy differentia
tion can be conceived of as passage
through a series of "window discrimi
nators" [see illustration on next page],
each stage of which analyzes a particular
aspect of the object in question. Each
retinal ganglion cell acts as a vertical
window that codes extension perpendic
ular to the direction of movement. The
retinal analysis is repeated and amplified
these observations? One can speculate
that for toads, which are active at twi
light, biologically important prey stimuli
that appear in the lower half of the visual
field are paler then their background;
those in the upper part of the field, how
ever, are for the most part relatively
dark, or at least just as often dark as
pale. Each of these contrast relations
could be reflected in the sensitivity char
acteristics of the Class II retinal ganglion
cells. With the approach of winter and
the period of hibernation, toads stop
catching prey. What makes them stop?
One mechanism may be an inversion
of ganglion-cell response characteristics,
brought about by signals from the brain
to the retina, such that the stimulus
background contrast relation is out of
phase with the real world, making prey
objects less visible. For the toad, in other
words, identical objects appear to be dif-

OPTIC THALAMUS
in the thalamic-pretectal region, where a

TECTUM
neuron pool acts as another vertical win
dow, this one with a certain minimum
response threshold. Extension in a hori
zontal direction is coded primarily by
Type I tectaI cells, which constitute a

ORIENTING AVOIDANCE
horizontal window. Type II tectal cells
perform a summation, with signals arriv
ing from the thalamic-pretectal region

ELSTIECTRODE
MULUS
having an inhibitory effect and those

RECORDING
from the Type I cells having an excita
tory effect. The resultant signal acts as
the trigger stimulus for the orienting
movement. The triggering of avoidance

THALAMUS OPTIC
TECTUM
behavior is probably achieved through
the activation of still another pool of
thalamic-pretectal neurons, the activa
tion being proportional to an additive
function of inputs from two of the win
dow-discriminator pools.
One of the remarkable aspects of this
system is a degree of plasticity, or
changeability. During the summer
months white prey objects moved against

ELSTIECTRODE
MULUS
black backgrounds elicit orienting be
havior much more effectively than do
black objects against white. In fall and
winter the situation is reversed, and at

OPTIC
the same time the overall prey-catching
activity of the toads decreases. Recently
our recording electrodes revealed that
the activation of single Class II ganglion
cells in the retina exhibits this seasonal
TECTUM THALAMUS
I VI1ISIUALI11STI11MULUS
shift in white-black preference. In win
ter neurons with receptive fields in the
d c
lower part of the visual field are more

I :
I II
strongly activated by black objects than
by white ones; in the upper field the
situation is reversed. In summer, how
ever, the neurons whose receptive field
is in the upper half of the visual field are
CONNECTIONS between the optic tectum and the thalamus were indicated by preceding
activated primarily by black stimuli,
experiments: signals from the retina excite botb tectum and thalamus; subsequent impulses
whereas neurons receptive to the lower
from Type I tectal neurons further excite cells in the thalamus, whereas signals from the
half of the visual field become more thalamus inhibit activity in Type II tectal neurons. Two different kinds of motor activity
strongly activated by white stimuli and are thereupon initiated by the two structures (top). Confirmatory evidence was obtained by
remain so until in the fall black stimuli electrical stimulation. Stimulation of the tectum (middle) elicits impulses (a) from cells in
again become dominant. the thalamus that ordinarily respond to visual stimuli (b). Stimulation of the thalamus (bot
What is the biological significance of tom) inhibits (e) impulses normally elicited (d) in Type II neurons by moving objects.

41

166
1974 SCIENTIFIC AMERICAN, INC
ferent when they are seen in winter than
when they are seen in summer, a re
minder that an organism's picture of the
environment is a product of its brain.
In contrast to the plasticity of the
systems for the filtering and storage of
information and for pattern recognition,
brain mechanisms involved in instinctive
actions are quite inflexible and do not
adapt easily to changes in the stimulus
situation. For each instinctive action in
the behavioral repertory there is a pre
programmed "printed" neuronal circuit
that coordinates the appropriate motor
act-even if it becomes inappropriate! If
such a circuit is triggered (either natural
ly or by electrode stimulation), the innate
reaction proceeds automatically. Prey
catching behavior is a good example. On
the basis of brain-stimulation experi-
ments we believe the sequence of events
controlling a natural orientation response
is about as follows: A pattern is formed
by a natural stimulus on a portion of the
retina that is outside the fixation region;
the retinal locus has a corresponding pro
jection locus in the optic tectum. If the
filtering process described above has
identified the object as prey, then the
appropriate neuronal system is activated.
A value corresponding to the distance
between the prey's locus on the retina
and the fixation point is transferred to
the toad's motor system. The result is
orientation: a turning movement such
that the retinal representation of the prey
is brought to the fixation point. That trig
gers a locus in the optic tectum that cor
responds to the fixation point. As soon as
this triggering reaches a threshold value
the rest of the prey-catching sequence is
activated, quite independently of the re
sult, or even of the short-term benefit to
the animal, of such activation. For ex
ample, if an experimental prey object is
removed at the instant when it is fixated
by the toad, the entire normal prey
catching routine nevertheless proceeds.
The toad snaps, gulps and wipes its
mouth in spite of the "situational vac
uum." The sequence is similar in its in
eVitability to what happens when the
triggering region of the tectum is stimu
lated with an electrode.
As for avoidance behavior, the results
of thalamic stimulation indicate that it is
controlled by a single master program.
The response consists in a firm planting
of the extremities on one side of the
toad's body and a gathering together of
the limbs on the opposite side. With the

a
STIMULUS
t
b -
toad in this stationary, poised position
the additional behavior patterns for cor

>t<
recting tilting of the body or making the

Rn'NA various evasive movements can be read

ily incorporated.

4m T that
he evidence I have reviewed shows
in a lower vertebrate the neuro

e:)DACAM .ft-+- nal processes for localization and iden

tification of a visual signal and for re
leaSing the associated instinctive motor
responses are separated topographically

O
T( EYCPTEUIM) O but are intimately connected with one
another. In the course of evolution the
centers for two of these processes, visual
localization and instinctive action, have
apparently remained in about their orig
inal positions. They occupy the same
c d
I areas of the brain, the tectum and the
thalamus, in monkeys and cats as they do

. .
in toads. The organization of these parts
of the brain, to which both neurophysi
ological and ethological methods have
provided investigative access, shows re
markable constancy in all classes of ver
tebrates. That is not the case, however,

.} jV.
for stimulus identification. In toads this
process takes place primarily in the tha
lamic-pretectal region and also in the

. retina and the tectum. Mammals, how

O-w+- O-+-
ever, underwent further evolution, cor
responding to the importance of pattern
recognition in the evolution of their be
havior. A new substrate developed for
two associated but highly specialized
processes, filtering and storage of infor
IDENTIFICATION OF AN OBJECT AS PREY OR ENEMY is symbolized as a series of mation: the visual cortex. From the in
operations by "window discriminators" (a). A ganglion cell in the retina codes vertical ex vestigations of Gerald E. Schneider at
tension (perpendicular to direction of movement), in effect responding to as much of a vi
M.I.T. we learn that in this case ontog
sual object as appears in a vertical window; extension beyond the window has an inhibitory
eny reflects phylogeny. In newborn ham
effect. Cells in the thalamus do the same thing. In the tectum Type I cells code horizontal
sters subcortical pathways between the
extension (in the direction of movement). Type II tectal cells sum the excitatory signal
from Type I cells and the inhibitory signal from the thalamus, and the resultant signal trig.
tectum and the thalamus are implicated
gers an orienting movement. At each stage the cell discharge depends on the relation be in pattern discrimination. In adult ani
tween the object and the window that senses its extension either in or perpendicular to the mals, on the other hand, pattern discrim
direction of movement (b, c, d). (The celldischarge patterns shown here are schematic.) ination takes place in the cortex.

42

167
1974 SCIENTIFIC AMERICAN, INC

Glamourous, mysterious liquid crystals.
We make them.
Behind the mystery and glamour: Those
who could reasonably be expected to
phone Mr. Grau are probably interested
in making display devices ranging from
mass-marketed consumer goods to some
of the more ambitious reaches of the en
gineering imaginaHon.
Liquid-crystal technology has blazed up
after smoldering quietly for most of the
time since 1888. Organic chemists get to
collaborate with electronic engineers
through the medium of the marketplace,
if not personally. For now, Kodak has set
up its booth on the chemical side of the
street.
The present generation of electronic
engineers catch on fast to subjects they
didn't necessarily concentrate on in school.
In liquid-crystal work, one deals with the
different forms and degrees of orderedness
among molecules, ranging between the
randomness of an ordinary isotropic liq
uid and the periodic architecture when it
freezes to a crystal. Only certain com
pounds assume this mesomorphic state.
Molecules of such compounds have a gen
erally elongated shape.
The most highly ordered kind of liquid
crystal, called smectic, where the mole
cules tier up in layers and the layers can
only slide against each other, can be seen
when water sits in contact with a bar of
soap. In a nematic liquid crystal there are
no tiers, but the molecules stay parallel.
Possibilities exist for controlling their
alignment. And this alignment controls
what they do to light passing through. And
that's what the action is mostly about in
industry circles.
Field effect, one approach used to cre
ate a visible pattern in a thin layer of ne
matic material between patterned, trans
parent electrodes, amounts to electrically
tuned birefringence. (We hope you didn't
cut the lecture on birefringence in Physics
1.) Contrast between "on" and "off" is
attained through various arrangements of
light polarizers and quarter-wave retarda
tion plates. Where and when the field is off,
the molecules must line up parallel to each
other and to the cell walls. This is called
"homogeneous" alignment. Surface treat
ment to make it happen is protected by
patents or trade secrecy.
D y n a m i c s c a t t e r i n g is t h e o t h e r
approach. Here, i n the "off" state, the
alignment can be either homogeneous or
perpendicular to cell walls. The latter is
"homeotropic" in the lingo of the art.
Either way, light passes straight through,
and the layer looks clear. Where field is
applied, then regardless of previous orien-
1974 SCIENTIFIC AMERICAN, INC
News: Fo r fi e l d-e ffe c t de v i c e s,
EASTMAN 14080 is now ready to
ship. At 18 I'm spacing, a typical lot
shows a threshold of only 1.5 volts
rms for 60-Hz sine wave (or 2 volts
dc), an operating temperature range
of 0 to 75C, and a dielectric anisot
ropy of +11.9 at 25C. At 5 V rms,
tum-on time at room temperature is
about 100 ms; tum-off time, about
350 ms; contrast ratio, about 50: 1.
For dynamic scat tering de vices,
EASTMAN 14099 is equally ready,
with the same 0 to 75C operating
temperature range. It aligns itself ho
meotropically on cle a n electrodes
without additional magic-an impor
tant advantage. Typical threshold is
4 volts rms, 5 volts dc.
More details from George Grau at
Organic Chemical Sales, Kodak,
Rochester, N.Y. 14650 (716-325-
2000, ext. 57288).
tation, all the molecules strive for paral
lelism to the wall. But ions that have been
incorporated in the mixture migrate, col
liding with them to knock them every
which way. The resulting optical inhomo
geneity scatters light. These turbid areas
can be made to look either darker or
lighter than the clear areas, depending on
directions of illumination and viewing.
Appearance of the pattern is more sensi
tive to angle of view than with field effect.
Power drain is more, because of the tur
bulence to be maintained. But you don't
have to find (or license from somebody)
a way to align the molecules. You just use
EASTMAN 14099.
Say, if you've read this far you are
probably interested enough to ask Dept.
412-L, Kodak, Rochester, N.Y. 14650 for
Eastman Organic Chemical Bulletin, Vol.
45, No.2 (1973), where four scientists of
the Kodak Research Laboratories, who
write more like scientists than engineers,
will take you farther into this than you'd
want to go for just recreation. It has a
76-item bibliography. For our 3,281-item
Liquid Crystal Bibliography (Kodak Pub
lication 11-193) on microfiche, make that
Dept. 454 and send $25* (plus applicable
state and local taxes).
Price subject to change without notice.
43
168
 The Journal of Neuroscience, May 25, 2011 31(21):77537762 7753

Cellular/Molecular

Retinal Ganglion Cells with Distinct Directional Preferences

Differ in Molecular Identity, Structure, and Central
Projections
Jeremy N. Kay,* Irina De la Huerta,* In-Jung Kim,* Yifeng Zhang,* Masahito Yamagata, Monica W. Chu,
Markus Meister, and Joshua R. Sanes
Center for Brain Science and Department of Molecular and Cellular Biology, Harvard University, Cambridge, Massachusetts 02138

The retina contains ganglion cells (RGCs) that respond selectively to objects moving in particular directions. Individual members of a
group of ON-OFF direction-selective RGCs (ooDSGCs) detect stimuli moving in one of four directions: ventral, dorsal, nasal, or temporal.
Despite this physiological diversity, little is known about subtype-specific differences in structure, molecular identity, and projections. To
seek such differences, we characterized mouse transgenic lines that selectively mark ooDSGCs preferring ventral or nasal motion as well
as a line that marks both ventral- and dorsal-preferring subsets. We then used the lines to identify cell surface molecules, including
Cadherin 6, CollagenXXV1, and Matrix metalloprotease 17, that are selectively expressed by distinct subsets of ooDSGCs. We also
identify a neuropeptide, CART (cocaine- and amphetamine-regulated transcript), that distinguishes all ooDSGCs from other RGCs.
Together, this panel of endogenous and transgenic markers distinguishes the four ooDSGC subsets. Patterns of molecular diversification
occur before eye opening and are therefore experience independent. They may help to explain how the four subsets obtain distinct inputs.
We also demonstrate differences among subsets in their dendritic patterns within the retina and their axonal projections to the brain.
Differences in projections indicate that information about motion in different directions is sent to different destinations.

Introduction focus here are ON-OFF direction-selective retinal ganglion cells

The brain is remarkable not only for its staggering number of (ooDSGCs). ooDSGCs respond to both the onset and termina-
neurons, but also for its panoply of neuron types. To understand tion of a flashed spot of light. If the spot moves through the
how brain circuits work, it will be essential to define these classes, receptive field of an ooDSCG, it fires most strongly for one (pre-
recognize them in vivo, and match them with the functions they ferred) direction of motion and remains silent for the opposite
serve. Numerous molecular markers of neuronal subtypes have (null) direction (Barlow et al., 1964). The dendrites of ooDSGCs
been identified using immunohistochemistry, gene expression arborize in two sublaminae of the inner plexiform layer (IPL),
analysis, and transgenic technology (Lein et al., 2007; Wassle et al., where they receive synapses from bipolar and starburst amacrine
2009; Siegert et al., 2009). It remains unclear, however, whether ex- cells in circuits that produce the direction selectivity (Barlow and
pression of one or a few genes will suffice to distinguish closely re- Levick, 1965; Demb, 2007; Zhou and Lee, 2008; Lee et al., 2010).
lated neuronal subtypes. Directional preferences of ooDSGCs are not uniformly dis-
To address this issue, we sought differences among four tributed; instead, discrete groups of ooDSGCs prefer ventral,
closely related neuronal subtypes in mouse retina. The retina dorsal, temporal, or nasal motion on the retina (Oyster and Bar-
contains five main neuronal types, which are divided into 60 100 low, 1967; Elstrott et al., 2008). These distinct preferences suggest
subtypes (Masland, 2001; Sanes and Zipursky, 2010), a number that the four subtypes receive different inputs and might have
probably typical for any brain area. The subtypes on which we distinct synaptic targets in the brain. Furthermore, it is clear that
the four subtypes recognize themselves as distinct: gap junctions
selectively couple members of the same subtype (DeBoer and
Received Feb. 19, 2011; accepted March 25, 2011.
Author contributions: J.N.K., I.D.l.H., I.-J.K., Y.Z., M.Y., M.M., and J.R.S. designed research; J.N.K., I.D.l.H., I.-J.K., Vaney, 2005) and dendritic fields of a given subtype overlap min-
Y.Z., M.Y., and M.W.C. performed research; J.N.K., I.D.l.H., I.-J.K., Y.Z., M.Y., M.W.C., M.M., and J.R.S. analyzed data; imally, whereas dendrites of ooDSGCs with different preferred
J.N.K., I.D.l.H., I.-J.K., Y.Z., M.Y., M.M., and J.R.S. wrote the paper. directions overlap freely (Amthor and Oyster, 1995).
This work was supported by grants from the NIH to I.-J.K., M.M., and J.R.S., a grant from NIH to J.R.S., Collabor- Despite intensive study of ooDSCGs generally, and the
ative Innovation Award 43667 from HHMI, and fellowships from the Life Sciences Research Foundation to J.N.K. and
from the Charles A. King Trust to Y.Z. We thank Sara Haddad, Debbie Pelusi, and Laura Stoppel for assistance.
availability of mouse transgenes that label subsets of ooDSGCs
*J.N.K., I.D.l.H., I.-J.K., and Y.Z. contributed equally to this study. (Huberman et al., 2009; Kim et al., 2010), no differences in mor-
Correspondence should be addressed to Joshua R. Sanes, Center for Brain Science and Department of Molecular phology, endogenous gene expression, or targets have been de-
and Cellular Biology, Harvard University, 52 Oxford Street, Cambridge, MA 02138. E-mail: sanesj@mcb.harvard.edu. scribed among the four directionally distinct subtypes. Here, we
I.-J. Kims present address: Department of Ophthalmology and Visual Science, and Department of Neurobiology,
Yale University School of Medicine, New Haven, CT 06511.
used these previously described mice together with novel trans-
DOI:10.1523/JNEUROSCI.0907-11.2011 genic lines to seek distinctions among ooDSGCs. Using gene ex-
Copyright 2011 the authors 0270-6474/11/317753-10$15.00/0 pression profiling, we identify markers that, in combination,
169
7754 J. Neurosci., May 25, 2011 31(21):77537762 Kay et al. Diversity of Direction-Selective RGCs

Figure 1. Transgenic markers of ooDSGCs that prefer different directions. A, Responses of a BD-RGC to a white rectangle moving across the receptive field center in eight different directions at 575 m/s.
Average responses are displayed in a polar plot and surrounding traces show raster plots for seven repeats. Note ON and OFF phases of spiking in rasters. V, Ventral; D, dorsal; N, nasal; T, temporal. B, Preferred
directions of BD-RGCs recorded from 60% eccentricity. Arrow length indicates the extent of direction selectivity, calculated as in Kim et al. (2010). C, Preferred directions of W9-RGCs (black arrows) and
DRD4-RGCs(grayarrows).D,E,RetinawholemountshowingBD-RGCs(red)andDRD4-RGCs(green),intripletransgenicmouse(DRD4-GFPFSTL4-CreERROSA-CAGS-STOP-tdTomato).BlueboxinDshows
region enlarged in E. RGCs express either green or red fluorescent proteins, but not both, demonstrating that DRD4- and BD-RGCs are distinct populations.

label all four subtypes with 90% accuracy. Subtype identity is
specified molecularly before eye opening, suggesting that differ-
ences among ooDSGCs arise independent of visual experience.
We also document differences in dendritic arbors and axonal
projections that distinguish the ooDSGCs selective for ventral
motion from those selective for nasal motion. In addition to pro-
viding reagents for mechanistic studies of ooDSGC subtype di-
versification and synaptic specificity, our results provide support
for the idea that molecular markers can be used to distinguish
closely related neuronal subtypes.
Materials and Methods
Mice. FSTL4-CreER transgenic mice, used to label bistratified dendrite-
retinal ganglion cells (BD-RGCs), were described by Kim et al. (2010).
Briefly, the transgene was generated by insertion of a tamoxifen-
responsive cre recombinase (CreER) cDNA at the initiation codon of the
Fstl4 coding sequence in a bacterial artificial chromosome. These animals
were crossed to mice that express a fluorescent protein following Cre-
mediated excision of stop sequences. The following three reporter lines
were used: Thy1-STOP-YFP mice line 15 (Buffelli et al., 2003), and mice
that conditionally express tdTomato (Madisen et al., 2010) (obtained
from Jackson Laboratories, catalog #007909) or green fluorescent pro-
tein (GFP) (Kim et al., 2009) (kindly provided by S. Dymecki, Harvard
Medical School, Boston, MA) from the Rosa26 locus. Tamoxifen (100
g, Sigma) was injected intraperitoneally into double transgenics at
postnatal day (P) 0 1 to activate CreER and thereby initiate expres-
sion of reporter.
The W9 mouse line was generated in parallel with the W3 and W7 lines
described by Kim et al. (2010). In this transgene, Thy1 regulatory ele-
ments drive expression of YFP. YFP was expressed in distinct and non-
overlapping subsets of RGCs in the W3, W7, and W9 lines, presumably
due to effects of sequences near the site of transgene integration in the
genome (Feng et al., 2000).
To mark cdh6-expressing cells, CreER was inserted into the initiation
codon of the cdh6 gene by homologous recombination in embryonic stem
cells. The targeting vector was generated by lambda phage-mediated recom-
bineering (Chan et al., 2007). Chimeric mice with the targeted embryonic
stem cells were generated and mated to obtain germ line transmission. Mice
were mated to reporter lines as described above for FSTL4-CreER.
Thy1-YFP-H mice were generated and characterized by Feng et al.
(2000). Dopamine receptor D4-GFP (DRD4-GFP) mice were obtained
from Mutant Mouse Regional Resource Center-University of North Car-
olina (www.mmrrc.org).
Electrophysiology. Methods for electrophysiological analysis have been
described by Kim et al. (2008, 2010). Briefly, dark-adapted retinas were
isolated in Ringers solution and targeted for cell-attached recording with
patch microelectrodes. Light stimuli were delivered from a computer-
driven video projector through a custom-made substage lens. Receptive
field centers were determined with small flashing spots, then direction
selectivity was assessed with stimuli moving through the receptive field
center in eight different directions.
Histology. Methods for histological analysis of retina and brain are
described by Yamagata et al. (2006) and Kim et al. (2010). Antibodies
used were as follows: rabbit anti-GFP, goat anti-choline acetyltrans-
ferase, mouse anti-Brn3a (all from Millipore), goat anti-vesicular acetyl-
choline transporter (Promega), rabbit anti-CART (Phoenix), and rabbit
anti-MMP17 (Epitomics). Secondary antibodies were from Invitrogen
or Jackson ImmunoResearch.
Cell isolation and expression profiling. To isolate BD-RGCs, retinas
from P6 mice were dissociated using papain (Worthington). The cell
suspension was incubated with anti-Thy1 antibodies conjugated to mag-
170
Kay et al. Diversity of Direction-Selective RGCs J. Neurosci., May 25, 2011 31(21):77537762 7755

netic beads (Miltenyi Biotec), and passed over a magnetic column, ac-
cording to the manufacturers instructions. This RGC-enriched cell
fraction was then passed through a MoFlo cytometer (Dako) to select
YFP-positive cells, which were sorted directly into RNA-stabilizing lysis
buffer from the PicoPure RNA Isolation Kit (MDS).
To isolate starburst amacrines, we used a Thy1 transgenic line in which
all starbursts and a subset of RGCs are labeled with the Kusabira Orange
fluorescent protein (OFP; line Thy1-OFP3) (J. Livet and J. R. Sanes,
unpublished observation). Purification of these cells used the same pro-
cedure as above, except that the Thy1-negative fraction (cells not re-
tained on the magnetic column) was used for sorting. We confirmed that
all OFP-positive cells recovered in this manner were starbursts by plating
the sorted cells and staining for starburst and RGC markers.
For microarray hybridization, RNA was isolated using the PicoPure
kit. Two rounds of amplification were performed with the MessageAm-
pII system (Ambion/Applied Biosystems), the second resulting in biotin-
labeled samples. These were hybridized to Affymetrix Mouse 430 2.0
arrays according to the manufacturers instructions. We hybridized two
replicates per transgenic line, each representing a sample of 200 cells
collected from different litters. Data were analyzed, and cell-type-specific
genes identified using Resolver (Rosetta) and dChip software (Li and
Wong, 2001).

Results
Transgenic lines marking ooDSGCs that prefer ventral or
nasal motion
We recently generated a transgenic line in which a group of
ooDSGCs, which we call BD-RGCs, are labeled with a fluorescent
protein (Kim et al., 2010). To begin this study, we assessed the
directional preference of BD-RGCs in adult retina (eccentricity of
60% measured from the optic nerve head to the retinal periph-
ery). Each cell was stimulated with bars moving in eight direc-
tions. Over 90% of BD-RGCs responded best to bars moving
from dorsal to ventral on the retina (i.e., to ventral motion) (Fig.
1 A, B).
We also recorded from RGCs in a transgenic line called W9,
which was generated in parallel with the W3 and W7 lines de-
scribed previously (Kim et al., 2010). The ventral retina of the W9
line contained a nearly homogeneous population of labeled bis-
tratified RGCs. These cells were ooDSGCs that responded pref-
erentially to nasal motion (Fig. 1C). This preference was similar
to that reported previously for ooDSGCs labeled in a transgenic
line generated by the GENSAT project and called DRD4-GFP
(Huberman et al., 2009). We obtained this line and confirmed
that DRD4-RGCs prefer nasal motion (Fig. 1C).
These results imply that BD-RGCs are distinct from W9 and
DRD4-RGCs. To test this idea, we generated BDDRD4 double- Figure 2. Relationship between structure and function of ooDSGCs. A, B, Morphology of W9
RGCs does not correlate with their preferred direction. Confocal stack z-projection showing two
transgenic mice in which BD-RGCs were labeled with a red fluo-
W9 cells that were injected with Lucifer yellow following recording. Scale bar, 100 m. C,
rophore and could thus be distinguished from GFP-positive DRD-4-RGC filled with Lucifer yellow following recording. Arrow indicates preferred direction,
DRD4-RGCs. As expected, few fluorescent cells in these retinas which is distinct from the orientation of its dendritic asymmetry. D, Sketch of part of a whole
(1%) were both RFP and GFP positive (Fig. 1 D, E). We also mounted retina (see inset at bottom left) showing dendritic asymmetry of the BD-RGCs. Arrows
expect that W9-RGCs and DRD4-RGCs are overlapping subsets, originate from the somas and point in the direction of dendritic asymmetry. Length of arrow is
but because they are labeled with indistinguishable fluorophores proportional to degree of dendritic asymmetry. Dot, Optic disc. E, Micrograph of two BD-RGCs
(YFP and GFP), we could not test this idea directly. from the retina. Scale bar: (in E) C, E, 100 m. F, Polar plot summarizing the dendritic asym-
Based on these results, we used the BD, W9, and DRD4 lines to metry of BD-RGCs from a retina similar to that shown in D and E. G, Relationship between
compare the structural and functional properties of ooDSGCs dendritic asymmetry and direction selectivity of 22 BD-RGCs. The preferred direction is plotted
selective for ventral and nasal motion. We used both DRD4 and relative to the direction of the dendritic arbor (dot). Black lines indicate dorsal-preferring cells
from the dorsal margin of the retina.
W9 lines in most studies but, unless otherwise noted, present
results from DRD4 here.
varied among cells but in no case was as dramatic as that docu-
Different patterns of asymmetry in cells preferring ventral mented previously for OFF-DSGCs (Kim et al., 2008). Consistent
and nasal motion with previous reports (Oyster et al., 1993; Huberman et al., 2009),
BD-, DRD4-, and W9-RGCs all had bistratified dendritic arbors there was no obvious relationship between the structural asym-
that, when viewed en face, were generally displaced asymmetri- metry and directional preference of W9- or DRD4-RGCs. For
cally around the somata (Fig. 2 AC,E). The degree of asymmetry example, the W9-RGC in Figure 2 A and the DRD4-RGC shown
171
7756 J. Neurosci., May 25, 2011 31(21):77537762 Kay et al. Diversity of Direction-Selective RGCs

Figure 3. Molecular markers for subsets of ooDSGCs. A, B, In situ hybridization for Cdh6 RNA (red) combined with anti-GFP antibody staining (green) reveals expression of Cdh6 in BD- but not
DRD4-RGCs. Yellow arrows in AH indicate double-labeled cells. Red arrows in AH indicate marker-positive cells that do not express GFP. C, D, In situ hybridization for Col25a1 RNA (red) shows
expression in GFP-labeled (green) BD-RGCs but not DRD4-RGCs. E, F, In situ hybridization for Mmp17 RNA (red) shows expression in GFP-labeled (green) DRD4- but not BD-RGCs. G, H, CART is a marker
of both BD- and DRD4-RGCs as shown by immunostaining with anti-CART (red) and anti-GFP (green). Laminae marked in H (INL, inner nuclear layer; IPL, inner plexiform layer) apply to AH, and K.
Scale bars: 10 m. I, J, Quantification of the fraction of RGCs expressing Cdh6, Col25a1, or Mmp17 at P14 (I ) and at P7 (J ) (n 44 for BD-RGCs, 104 for DRD4-RGCs, 1198 for all RGCs). K, MMP17
immunoreactivity in section from P14 retina. Antibody labels the putative starburst layers of the IPL, suggesting immunoreactivity in the dendrites of starburst amacrines and/or ooDSGCs. A subset
of RGC cell bodies (arrows) as well as putative starburst amacrines (arrowheads) are labeled in the GCL and INL. Scale bar, 25 m.

in Figure 2C had dendrites displaced dorsonasally and dorso-
temporally from their somata, respectively, but preferred na-
sal motion.
In striking contrast, most of the asymmetric BD-RGCs had
dendrites that were displaced vertically from the soma. In most of
the retina, the predominant direction was ventral, but a small
number of BD-RGCs near the dorsal pole of the retina had arbors
directed dorsally (Fig. 2 DF ). Together with the physiological
results presented above, this finding suggested a correlation be-
tween the structure and function of BD-RGCs. To test this rela-
tionship stringently, we recorded from 13 BD-RGCs in central
retina and 3 BD-RGCs at the dorsal pole, then imaged their den-
drites following recording. In 15, including all 3 of the dorsal
cells, the preferred direction of the cell, determined physiologi-
cally, corresponded to the direction in which the dendritic arbor
was displaced from the soma (Fig. 2G); the 16th had a symmet-
rical dendritic arbor. Thus, not only the majority of BD-RGCs
that prefer ventral motion but also the minority in dorsal retina
that prefer dorsal motion exhibit an association of structural and
functional asymmetries.
Distinct molecular signatures of RGCs that prefer vertical and
nasal motion
To identify genes selectively expressed in BD-RGCs cells, we
isolated them based on their fluorescence and profiled gene ex-
pression using microarrays (see Experimental procedures). In
parallel, we isolated and profiled several other RGC and amacrine
subtypes (Kay, 2011), including starburst amacrines, which pro-
vide the main inhibitory input to ooDSGCs. We filtered the mi-
croarray data to identify genes expressed at several fold higher
levels in BD-RGCs and/or starburst amacrines compared with
other cell types in the dataset, then used in situ hybridization
(ISH) to ask they whether they were expressed by BD-, DRD4-, or
W9-RGCs.
Two genes, cadherin 6 (Cdh6 ) and collagen 25a1 (Col25a1),
were expressed by BD-RGCs but not by DRD4-RGCs (Fig. 3A
D,I ). Cdh6, a type II cadherin, was previously shown to be ex-
pressed in mouse retina (Honjo et al., 2000); Col25a1 encodes a
transmembrane collagen that is expressed in the nervous system
and accumulates in amyloid plaques in Alzheimers disease
(Hashimoto et al., 2002). Cdh6 and Col25a1 were also expressed
by starburst amacrines but not by other retinal cells.
We were also interested in identifying markers of ooDSGCs
that prefer nasal motion. The most obvious candidate was Drd4,
since regulatory elements from this gene drive expression of GFP
in the DRD4 transgenic line. In some cases, however, including
the line that labels BD-RGCs, transgenes are expressed in cells
that do not express the endogenous gene, because of omission of
critical regulatory elements from the transgene or to influences at
the genomic site of integration (Feng et al., 2000; Haverkamp et
al., 2009; Kim et al., 2010). Indeed, consistent with data from the
rat (Klitten et al., 2008), in situ hybridization showed that Drd4
172
Kay et al. Diversity of Direction-Selective RGCs
Figure 4. ISH for Drd4 RNA in P14 retina of wild-type or DRD4-GFP transgenic mice. A,
Wild-type mice show Drd4 expression in photoreceptors in the outer nuclear layer (ONL) and in
a subset of inner nuclear layer (INL) cells. Very low levels of transcript are detected in the GCL. B,
DRD4-GFP transgenic mice show the same pattern of transcripts in photoreceptors and INL as in
A but also exhibit Drd4 signals in DRD4-GFP-positive RGCs (arrowheads). The difference in
labeling between wild-type and transgenic mice suggests that Drd4 transcripts in DRD4-RGCs of
DRD4-GFP mice arise from the transgene, not the endogenous Drd4 gene. Scale bar, 25 m.
was expressed prominently by photoreceptors and at low, uni-
form levels in the ganglion cell layer (GCL) of wild-type mice
(Fig. 4 A). In contrast, the Drd4 probe strongly labeled GFP-
positive RGCs, along with photoreceptors, in the DRD4-GFP
J. Neurosci., May 25, 2011 31(21):77537762 7757
transgenic retina (Fig. 4 B, B). This result is not unexpected, in
that the genomic fragment used to generate the transgene con-
tains sequences that could be transcribed and recognized by the
ISH probe. Thus, Drd4 mRNA in DRD4-RGCs is derived from
the transgene, and it is unlikely that GFP-positive cells in the
DRD4 line express significant levels of endogenous Drd4.
As an alternative approach, we tested genes that we had found
by in situ hybridization to be expressed by starburst amacrines
and a small subset of RGCs but not by BD-RGCs. One of these,
matrix metalloprotease 17 (Mmp17 ), was expressed by 5% of
all RGCs and by few if any BD-RGCs (1%) or Col25a1-positive
RGCs (1/179) but by 90% of DRD4 RGCs (Fig. 3 E, F,I ).
MMP17 is a cell surface-associated enzyme (Sohail et al., 2008)
expressed by some neurons in brain (Rikimaru et al., 2007). Im-
munostaining confirmed the selective association of MMP17
with DRD4-RGCs. Immunoreactivity was localized in somata of
RGCs and starburst amacrines, and in the ON and OFF sublami-
nae of the IPL, where their processes arborize (Fig. 3K ). Approx-
imately 70% of MMP17-positive RGCs were DRD4-RGCs as
assessed by ISH and immunostaining. MMP17-positive, GFP-
negative RGCs might be ooDSGCs that prefer nasal motion but
are not labeled by the DRD4-GFP transgene or, as discussed be-
low, RGCs of other types.
Expression patterns of Cdh6, Col25a1, and Mmp17 at P7, be-
fore eye opening (P12), were similar to those at P14, after eye
opening (Fig. 3 I, J ). The correspondence of endogenous gene and
transgene expression at this early stage indicates that visual expe-
rience is not required for ooDSGC subtypes to acquire their dis-
tinct molecular identities.
CART selectively labels ooDSGCs
Another gene shown by microarray analysis to be expressed by
BD-RGCs was cocaine- and amphetamine-regulated transcript
(CART; gene symbol Cartpt). This gene encodes several neuro-
peptides, including CART (Rogge et al., 2008). Antibodies to
CART labeled 15% of all RGCs as well as a small group of
nonstarburst amacrines in the inner nuclear layer that we did not
characterize further. CART was present in nearly all BD-, W9-,
and DRD4-RGCs (Fig. 3GJ ) and in nearly all RGCs that
expressed Col25a1 or Mmp17 (data not shown). Moreover, the
fraction of RGCs that was CART positive corresponded approx-
imately to the fraction of RGCs that are ooDSGCs (Sun et al.,
2002; Badea and Nathans, 2004; Coombs et al., 2006; Volgyi et al.,
2009). We therefore considered the possibility that CART might
be expressed by all and only ooDSGCs.
To test this idea, we used a transgenic line, Thy1-YFP-H (Feng
et al., 2000). Approximately 200 RGCs are labeled per retina in
this line, so that RGCs are sufficiently well separated to permit
imaging of the complete dendritic arbor of a single cell. Labeled
cells include all known RGC subtypes (Coombs et al., 2006).
Whole mounts of retinas from Thy1-YFP-H mice were stained
with anti-CART and then imaged (Fig. 5 A, CE). All of the
CART-positive, YFP-positive RGCs imaged (22 in 3 retinas) were
morphologically similar and had bistratified dendritic arbors,
consistent with their being ooDSGCs (Fig. 5B). Seven of these
cells were reconstructed at high resolution; all had dendrites that
cofasciculated with the processes of starburst amacrines, con-
firming their identity as ooDSGCs (Fig. 5DF ). In contrast, none
of 140 CART-negative, YFP-positive RGCs imaged in these reti-
nas resembled ooDSGCs. Thus, CART is a specific marker of
ooDSGCs.
173
7758 J. Neurosci., May 25, 2011 31(21):77537762 Kay et al. Diversity of Direction-Selective RGCs

Cadherin 6 is selectively expressed by

ooDSGCs responsive to vertical motion
Of genes selectively expressed by BD-RGCs,
we focused on Cdh6 because previous work
suggested that cadherins are expressed by
RGC subsets (Yamagata et al., 2006) and in-
volved in retinotectal patterning (Inoue and
Sanes, 1997). We generated a targeted allele
in which tamoxifen-activated CreER was in-
serted in place of the first coding exon of the
Cdh6 gene. These mice were mated to mice
in which expression of YFP was Cre
dependent.
Tamoxifen was administered to neo-
nates, and retinas were examined at P21.
Two sets of retinal cells were YFP positive:
starburst amacrine cells (identified by dou-
ble labeling with antibodies to ChAT) and
RGCs (Fig. 6A,B). Most (99%) of the
Cdh6-positive RGCs had bistratified den-
drites that arborized with starburst ama-
crine processes (Fig. 6A), suggesting, along
with microarray results presented above,
that they were ooDSGCs. (Using other re-
porters described in Materials and Meth-
ods, only 95% of labeled RGCs were
ooDSGCs; most of the remainder had
dendrites in the outermost sublamina of
the IPL.) Physiological analysis confirmed
that the YFP-labeled Cdh6-RGCs were di-
rection selective. Over 90% (22/23) were
ooDSGCs; one was an ON-DSGC. Ap-
proximately 90% preferred vertical mo-
tion (10 dorsal and 11 ventral) (Fig. 6C).
Thus, Cdh6 is a selective marker of RGCs
that respond to vertical motion. Figure 5. CART antibody labels ooDSGCs. A, CART and anti-GFP immunostaining identify double-positive RGCs in a retinal
the morphology of a YFP CART immunoreactive RGC (arrow; z-projection of
Most BD-RGCs are selective for ven- whole mount from line YFP-H. GFP channel shows
confocal stack). CART channel (A) shows CART RGCs in a single confocal plane through the GCL. Arrow indicates the soma of the
tral motion, whereas similar numbers of YFP CART double-positive cell. B, Morphological analysis of RGCs in line YFP-H that were CART and CART (n 140). All
Cdh6-RGCs prefer dorsal and ventral mo- CART-immunoreactive RGCs (n 22) have the bistratified morphology of ooDSGCs. By contrast, none of the CART cells showed
tion. Thus, molecular differences between this morphology. C, D, Single confocal planes through the cell shown in A reveal morphological features that identify it as an
these two populations would provide clues ooDSGC. The cell dendrites (green) are bistratified, with both the ON (C) and OFF arbors (D) cofasciculating with the choline
to the molecular identity of DSGCs that acetyltransferase (ChAT)-positive processes of starburst amacrines (red). E, Rotation of a 3-D reconstruction of the dendrites of this
prefer dorsal motion. We used the panel cell show that it has bistratified projections to the OFF and ON starburst IPL sublaminae (arowheads). GFP (green) and ChAT (red)
of markers defined above to assess gene channels are shown. Scale bars: 20 m.
expression in Cdh6-RGCs. At P7, nearly
all Cdh6-RGCs (90%) were CART and Col25a1 positive, but BD-RGC axons (Kim et al., 2010). Small numbers of retinally
only 10% were Mmp17 positive (Fig. 6 D). Thus, expression of derived CART-positive fibers were present in the ventral LGN,
endogenous genes assessed to date does not distinguish ooDSGCs the medial temporal nucleus (MTN), and the nucleus of the optic
that prefer ventral and dorsal motion. It is possible that the tract (NOT). No retinally derived CART-positive fibers were
Mmp17-positive Cdh6-RGCs (Fig. 6 D) correspond to the small found in the suprachiasmatic nucleus, or in accessory optic or
fraction of Cdh6-RGCs that prefer temporal motion. pretectal nuclei other than MTN and NOT (data not shown).
Inputs from the contralateral and ipsilateral eyes are spatially
Central projections of ooDSGCs separated in both the dLGN and the superior colliculus (Gode-
To visualize the central projections of ooDSGCs, we stained cor- ment et al., 1984). In the dLGN, ipsilateral input is confined to a
onal brain sections with antibodies to CART. The full extent of all central core, surrounded by a contralaterally innervated shell. In
retinal projections was revealed by intraocular injection of a the superior colliculus, ipsilateral input is confined to a narrow
tracer, cholera toxin B, and the retinal origin of CART-labeled sublamina in the deepest portion of the retinorecipient zone.
fibers was assessed by their loss following monocular enucleation. Remarkably, in both structures few if any retinally derived
Retinally derived CART-positive fibers were abundant in the dor- CART-positive fibers were present in ipsilaterally innervated re-
sal lateral geniculate nucleus (dLGN) and superior colliculus gions (Fig. 7AC). Thus, central projections of ooDSGCs are pre-
(Fig. 7AC). Within the superior colliculus, CART-positive fibers dominantly if not entirely contralateral.
were concentrated in the superficial half of the retinorecipient Quina et al. (2005) have shown that RGCs expressing the tran-
zone (Fig. 7C), the region previously shown to be innervated by scription factor Brn3a are excluded from the ipsilateral pathway.
174
Kay et al. Diversity of Direction-Selective RGCs
Figure 6. Cadherin 6-positive ooDSGCs prefer vertical motion. A, Cdh6-RGCs and starburst
amacrine cells labeled with YFP (green) in retina sections from a cdh6 knock-in heterozygote
(Cdh6-CreER Thy1-stop-YFP). Choline acetyltransferase (ChAT) (red) labels starburst ama-
crine cell somas and dendrites. Cdh6-RGC dendrites project to starburst IPL layers. Blue, Fluo-
rescent Nissl stain. B, Retina whole mount from a cdh6 knock-in heterozygote. C, Responses of
Cdh6-RGCs, showing that the vast majority prefer dorsal or ventral motion (n 23 cells
from 7 retinas, 7 mice). D, Molecular markers expressed by Cdh6-RGCs. Scale bars: A, 50
m; B, 200 m.
Consistent with this result, we found that 95% of BD-RGCs,
DRD4-RGCs, Cdh6-RGCs, and CART-positive RGCs are Brn3a-
positive (data not shown).
Distinct projections of ooDSGCs that prefer vertical and
nasal motion
We used the transgenic lines described above to ask whether
ooDSGC subsets projected to distinct regions within their target
areas. BD-, DRD4-, and Cad6-RGCs all projected to the dLGN,
the ventral lateral geniculate nucleus (vLGN), and the superior
colliculus (Figs. 7 D, E, 8 A, B). In the dLGN, however, the laminar
position of their terminal arbors differed. DRD4-RGC arbors
occupied a narrow superficial lamina, directly beneath the optic
tract. In contrast, arbors of BD- and Cad6-RGCs extended deeper
(Fig. 7D). In the superior colliculus, DRD4-RGC arbors started
J. Neurosci., May 25, 2011 31(21):77537762 7759
branching at a more superficial point than those of BD- and
Cad6-RGCs, but the difference was not striking (Fig. 7E).
In contrast to these structures, which were innervated by BD-,
Cdh6-, and DRD4-RGCs, only the vertically preferring ooDSGCs
(BD- and Cdh6-RGC) projected to NOT (Fig. 8 E, F ). Likewise,
the MTN was innervated by BD- and Cad6-RGCs but not by
nasal-preferring DRD4-RGCs (Fig. 8C,D). Previous studies in
mice have shown that the MTN receives input from RGCs that
respond to vertical but not horizontal motion (Yonehara et al.,
2009). Our results are consistent with this pattern. Although
most RGCs that project to mouse MTN are ON-DSGCs rather
than ooDSGCs (Yonehara et al., 2009), rat MTN does receive
input from ooDSGCs (Dann and Buhl, 1987).
Discussion
In many respects, ooDSGCs appear to be a single major RGC
subtype: they are similar in dendritic morphology (the main cri-
terion by which RGCs are classified), in response properties and
in the biophysical mechanisms that underlie their direction selec-
tivity (Demb, 2007; Zhou and Lee, 2008). Yet in one respect they
are readily divisible: they can be grouped into four distinct sets,
each responsive to one of the four cardinal directions (Oyster and
Barlow, 1967). Thus, ooDSGCs are more properly viewed as
comprising four closely related subtypes that, despite many com-
mon features, are likely to be molecularly distinct and might send
their information to different central targets. However, despite
intensive study of ooDSGCs in the aggregate, differences among
them have not been explored until very recently (Huberman et
al., 2009). Here, we used a set of four transgenic lines to mark
ooDSGCs with preferences for ventral (BD), nasal (DRD4 and
W9), and vertical (dorsal and ventral; Cdh6) motion. By compar-
ing the lines, we identified molecular and structural differences
among ooDSGC subsets and documented differences in their
central targets.
Molecular markers for ooDSGCs with distinct
directional preferences
The fact that transgenic lines mark subsets of ooDSGCs with
distinct directional preferences implies that these subsets are mo-
lecularly distinct. Unfortunately, the three lines with which we
began our study (BD, W9, and DRD4) provided no information
on what these differences might be. This is because transgene
expression patterns reflect influences from the chromosomal re-
gion at which the transgene is integrated, rather than or in addi-
tion to regulatory sequences within the transgene itself. Such
ectopic expression is not uncommon. For example, Haverkamp
et al. (2009) recently characterized marked retinal subsets in four
BAC transgenic lines and showed that none of the marked cells
expressed the corresponding endogenous gene.
As an alternative, we isolated marked RGCs from several
transgenic lines, based on their fluorescence, and used expression
profiling to seek endogenous genes selectively expressed by one
line. This strategy led to identification of Cdh6 and Col25a1,
which are expressed selectively (though not exclusively) by BD-
RGCs. Further screening led to identification of Mmp17 as a gene
expressed by DRD4- and W9- but not BD-RGCs. Another gene
identified in this screen, Cartpt, is expressed by most if not all
ooDSGCs and by few if any other RGCs. Together, the panel of
markers we describe here provides a molecular signature for each
of the four ooDSGC subtypes (Fig. 9).
Of the many genes identified using our sorting and microarray
strategy, we chose these genes for investigation because their
products are cell surface or secreted proteins that might be in-
175
7760 J. Neurosci., May 25, 2011 31(21):77537762 Kay et al. Diversity of Direction-Selective RGCs

volved in intercellular interactions re-

quired for development or function of
ooDSGCs. Initial analysis of Cdh6-null
mutants has so far not revealed severe de-
fects in RGCs, but further studies of these
and other mutants are now in progress.
Our results also set the stage for com-
paring different types of RGCs that re-
spond to motion in the same direction.
Lines are now available that mark three
distinct sets of DSGCs selective for ventral
motion: OFF-DSGCs (J-RGCs) Kim et al.,
2008, 2010), ON-DSGCs (SPIG1-RGCs)
(Yonehara et al., 2008, 2009), and ooDS-
GCs (BD-RGCs, this article). Ongoing
electrophysiological studies indicate that
J-RGCs and BD-RGCs use radically dif-
ferent synaptic mechanisms to compute
image motion (Y. Zhang, unpublished
observation).

Correlated structural and functional

asymmetry in an ooDSGC subset
Oyster et al. (1993) recorded from a large
set of ooDSGCs in rabbit retina, then exam-
ined their dendritic morphologies. They
noted that some cells had asymmetric den-
dritic arbors but found no morphological
feature that is correlated with the cells pre-
ferred response directions. It was therefore
surprising to find that the asymmetry of
BD-RGC arbors was strongly associated
Figure 7. Lamina-specific projections of different ooDSGC functional subtypes. A, Cholera toxin subunit B (CTB)-labeled axons
with their directional preference. Remark- in the dLGN arising from the contralateral (red) and ipsilateral (blue) eyes. A, CART axons (green) overlap with CTB-labeled
ably, this association was not evident in axons from the contralateral eye (red). B, After monocular enucleation, the CART axon terminals (green) disappear from the
DRD4- and W9-RGCs, which otherwise dLGN contralateral to the eye removed, indicating that these terminals originate from RGCs. C, Projections of CART RGCs to
seemed structurally indistinguishable from superior colliculus (SC). After monocular enucleation CART axons (green) are present only in the SC contralateral to the remaining
BD-RGCs. eye. CTB-labeled axons in the contralateral (left) SC (red) overlap with CART axon terminals. No CART RGC terminals are seen
The correspondence of dendritic asym- in the CTB-labeled ipsilateral projection (right). D, BD- (red) and DRD4-RGC (green) axons in the dLGN. Overlay in D shows distinct
metry with preferred movement direction laminar targeting of the two RGC subsets. E, BD- (red) and DRD4-RGC (green) axons in the SC. Scale bars: AD, 200 m; E, 400 m.
in BD-RGCs resembles that in J-RGCs, a
far more strikingly asymmetric group of OFF-DSGCs that we We also found two differences in targets among ooDSGC sub-
described recently (Kim et al., 2008, 2010). We suspect, however, sets. First, BD-RGC and DRD4-RGC arbors were partially segre-
that the association differs in the two cases. Both J- and BD-RGCs gated within the dLGN. Second, BD- and Cdh6-RGCs project to
include some cells whose arbors appear symmetric. The symmet- the MTN and NOT, whereas DRD4-RGCs do not. These nuclei
ric J-RGCs are not direction selective, supporting the idea that are involved in generating optokinetic reflexes (OKRs) (Masseck
structure underlies function for these cells (Kim et al., 2008). In and Hoffmann, 2009). The finding that ooDSGCs provide verti-
contrast, structurally symmetric BD-RGCs are as direction selec- cal but not nasal motion information to these nuclei may have
tive as asymmetric ones, suggesting that for these cells structural implications for understanding OKR behavior. More generally,
asymmetry does not determine directional preference. We do not our results show that ooDSGC-derived directional information is
know why the association seen for BD-RGCs is not evident for segregated in distinct brain targets.
DRD4- and W9-RGCs.

Central targets of ooDSGCs References

We identified several differences between projections of ooDSGCs
and those of RGCs generally. First, ooDSGC axon arbors are
absent from many retinal targets, including the superchiasmatic
nucleus, the accessory optic nuclei LTN and DTN, and most
pretectal nuclei. Second, within the major retinal target, the su-
perior colliculus, ooDSGC arbors are confined to the upper por-
tion of the retinorecipient zone. Third, ooDSGC axons are nearly
if not entirely excluded from the ipsilateral projection, even
though this projection does include axons of multiple RGC sub-
types (Hong et al., 2011).
Amthor FR, Oyster CW (1995) Spatial organization of retinal information
about the direction of image motion. Proc Natl Acad Sci U S A 92:
4002 4005.
Badea TC, Nathans J (2004) Quantitative analysis of neuronal morpholo-
gies in the mouse retina visualized by using a genetically directed reporter.
J Comp Neurol 480:331351.
Barlow HB, Levick WR (1965) The mechanism of directionally selective
units in rabbits retina. J Physiol 178:477504.
Barlow HB, Hill RM, Levick WR (1964) Retinal ganglion cells responding
selectively to direction and speed of image motion in the rabbit. J Physiol
173:377 407.
Buffelli M, Burgess RW, Feng G, Lobe CG, Lichtman JW, Sanes JR (2003)

176
Kay et al. Diversity of Direction-Selective RGCs J. Neurosci., May 25, 2011 31(21):77537762 7761

Hashimoto T, Wakabayashi T, Watanabe A,

Kowa H, Hosoda R, Nakamura A, Kanazawa I,
Arai T, Takio K, Mann DM, Iwatsubo T
(2002) CLAC: a novel Alzheimer amyloid
plaque component derived from a transmem-
brane precursor, CLAC-P/collagen type XXV.
EMBO J 21:1524 1534.
Haverkamp S, Inta D, Monyer H, Wassle H
(2009) Expression analysis of green fluo-
rescent protein in retinal neurons of four
transgenic mouse lines. Neuroscience 160:
126 139.
Hong YK, Kim IJ, Sanes JR (2011) Stereotyped
axonal arbors of retinal ganglion cell subsets
in the mouse superior colliculus. J Comp Neu-
rol 519:16911711.
Honjo M, Tanihara H, Suzuki S, Tanaka T, Honda
Y, Takeichi M (2000) Differential expression
of cadherin adhesion receptors in neural ret-
ina of the postnatal mouse. Invest Ophthal-
mol Vis Sci 41:546 551.
Huberman AD, Wei W, Elstrott J, Stafford BK,
Feller MB, Barres BA (2009) Genetic iden-
tification of an On-Off direction-selective
retinal ganglion cell subtype reveals a layer-
specific subcortical map of posterior mo-
tion. Neuron 62:327334.
Inoue A, Sanes JR (1997) Lamina-specific connec-
tivity in the brain: regulation by N-cadherin,
neurotrophins, and glycoconjugates. Science
276:1428 1431.
Kay JN, Voinescu PE, Chu MW, Sanes JR (2011)
Neurod6 expression defines novel retinal
amacrine cell subtypes and regulates their fate.
Nat Neurosci, in press.
Kim IJ, Zhang Y, Yamagata M, Meister M, Sanes
JR (2008) Molecular identification of a reti-
nal cell type that responds to upward motion.
Nature 452:478 482.
Kim IJ, Zhang Y, Meister M, Sanes JR (2010)
Figure 8. ooDSGCs project to the vLGN and the pretectal system. A, BD-RGC axons (green) arborize in the vLGN. A, Merge of CTB Laminar restriction of retinal ganglion cell
(red) and BD-RGC (green) axons. B, DRD4-RGC axons (green) arborize in the vLGN. B, Merge of CTB (red) and DRD4-RGC (green) dendrites and axons: subtype-specific devel-
axons. C, BD-RGC axons (green) project to the MTN. Blue, fluorescent Nissl stain. D, DRD4-RGC axons (green) are not present in the opmental patterns revealed with transgenic
MTN. D, Merge of DRD4-RGC (green) and CTB (red) axons in the MTN. Blue, Fluorescent Nissl stain. E, BD-RGC axons (green) markers. J Neurosci 30:14521462.
arborize in the nucleus of the optic tract (NOT; arrowhead). CTB (red) labels the optic tract and the nucleus. F, DRD4 axons (green) Kim JC, Cook MN, Carey MR, Shen C, Regehr
are absent from the NOT (arrowhead), labeled with CTB (red). Scale bars: AE, 200 m. WG, Dymecki SM (2009) Linking geneti-
cally defined neurons to behavior through a
broadly applicable silencing allele. Neuron
Genetic evidence that relative synaptic efficacy biases the outcome of 63:305315.
synaptic competition. Nature 424:430 434. Klitten LL, Rath MF, Coon SL, Kim JS, Klein DC, Mller M (2008) Local-
Chan W, Costantino N, Li R, Lee SC, Su Q, Melvin D, Court DL, Liu P (2007) ization and regulation of dopamine receptor D4 expression in the adult
A recombineering based approach for high-throughput conditional and developing rat retina. Exp Eye Res 87:471 477.
knockout targeting vector construction. Nucleic Acids Res 35:e64. Lee S, Kim K, Zhou ZJ (2010) Role of ACh-GABA cotransmission in detect-
Coombs J, van der List D, Wang GY, Chalupa LM (2006) Morphological ing image motion and motion direction. Neuron 68:1159 1172.
properties of mouse retinal ganglion cells. Neuroscience 140:123136. Lein ES, Hawrylycz MJ, Ao N, Ayres M, Bensinger A, Bernard A, Boe AF,
Dann JF, Buhl EH (1987) Retinal ganglion cells projecting to the accessory Boguski MS, Brockway KS, Byrnes EJ, Chen L, Chen L, Chen TM, Chin
optic system in the rat. J Comp Neurol 262:141158. MC, Chong J, Crook BE, Czaplinska A, Dang CN, Datta S, Dee NR, et al.
DeBoer DJ, Vaney DI (2005) Gap-junction communication between sub-
(2007) Genome-wide atlas of gene expression in the adult mouse brain.
types of direction-selective ganglion cells in the developing retina. J Comp
Nature 445:168 176.
Neurol 482:8593.
Li C, Wong WH (2001) Model-based analysis of oligonucleotide arrays: ex-
Demb JB (2007) Cellular mechanisms for direction selectivity in the retina.
pression index computation and outlier detection. Proc Natl Acad Sci
Neuron 55:179 186.
Elstrott J, Anishchenko A, Greschner M, Sher A, Litke AM, Chichilnisky EJ, U S A 98:3136.
Feller MB (2008) Direction selectivity in the retina is established inde- Madisen L, Zwingman TA, Sunkin SM, Oh SW, Zariwala HA, Gu H, Ng LL,
pendent of visual experience and cholinergic retinal waves. Neuron Palmiter RD, Hawrylycz MJ, Jones AR, Lein ES, Zeng H (2010) A robust
58:499 506. and high-throughput Cre reporting and characterization system for the
Feng G, Mellor RH, Bernstein M, Keller-Peck C, Nguyen QT, Wallace M, whole mouse brain. Nat Neurosci 13:133140.
Nerbonne JM, Lichtman JW, Sanes JR (2000) Imaging neuronal subsets Masland RH (2001) The fundamental plan of the retina. Nat Neurosci
in transgenic mice expressing multiple spectral variants of GFP. Neuron 4:877 886.
28:4151. Masseck OA, Hoffmann KP (2009) Comparative neurobiology of the opto-
Godement P, Salaun J, Imbert M (1984) Prenatal and postnatal develop- kinetic reflex. Ann NY Acad Sci 1164:430 439.
ment of retinogeniculate and retinocollicular projections in the mouse. Oyster CW, Barlow HB (1967) Direction-selective units in rabbit retina:
J Comp Neurol 230:552575. distribution of preferred directions. Science 155:841 842.

177
7762 J. Neurosci., May 25, 2011 31(21):77537762
Figure 9. Markers that identify the four ooDSGC subsets and distinguish ooDSGCs from other
RGCs. The schematic includes genes identified in the microarray analysis (CART, Col25a1, Cdh6,
and Mmp17) as well as the transgenes, which do not correspond to endogenous genes(BD,
DRD4, and W9). *Some of the Cdh6-RGCs ooDSGCs that prefer temporal motion may be Mmp17
positive.
Oyster CW, Amthor FR, Takahashi ES (1993) Dendritic architecture of ON-
OFF direction-selective ganglion cells in the rabbit retina. Vision Res
33:579 608.
Quina LA, Pak W, Lanier J, Banwait P, Gratwick K, Liu Y, Velasquez T,
OLeary DD, Goulding M, Turner EE (2005) Brn3a-expressing retinal
Kay et al. Diversity of Direction-Selective RGCs
ganglion cells project specifically to thalamocortical and collicular visual
pathways. J Neurosci 25:1159511604.
Rikimaru A, Komori K, Sakamoto T, Ichise H, Yoshida N, Yana I, Seiki M
(2007) Establishment of an MT4-MMP-deficient mouse strain repre-
senting an efficient tracking system for MT4-MMP/MMP-17 expression
in vivo using beta-galactosidase. Genes Cells 12:10911100.
Rogge G, Jones D, Hubert GW, Lin Y, Kuhar MJ (2008) CART peptides:
regulators of body weight, reward and other functions. Nat Rev Neurosci
9:747758.
Sanes JR, Zipursky SL (2010) Design principles of insect and vertebrate vi-
sual systems. Neuron 66:1536.
Siegert S, Scherf BG, Del Punta K, Didkovsky N, Heintz N, Roska B (2009)
Genetic address book for retinal cell types. Nat Neurosci 12:11971204.
Sohail A, Sun Q, Zhao H, Bernardo MM, Cho JA, Fridman R (2008) MT4-
(MMP17) and MT6-MMP (MMP25), A unique set of membrane-
anchored matrix metalloproteinases: properties and expression in cancer.
Cancer Metastasis Rev 27:289 302.
Sun W, Li N, He S (2002) Large-scale morphological survey of mouse retinal
ganglion cells. J Comp Neurol 451:115126.
Volgyi B, Chheda S, Bloomfield SA (2009) Tracer coupling patterns of the
ganglion cell subtypes in the mouse retina. J Comp Neurol 512:664 687.
Wassle H, Puller C, Muller F, Haverkamp S (2009) Cone contacts, mosaics,
and territories of bipolar cells in the mouse retina. J Neurosci 29:106 117.
Yamagata M, Weiner JA, Dulac C, Roth KA, Sanes JR (2006) Labeled lines in
the retinotectal system: markers for retinorecipient sublaminae and the
retinal ganglion cell subsets that innervate them. Mol Cell Neurosci
33:296 310.
Yonehara K, Shintani T, Suzuki R, Sakuta H, Takeuchi Y, Nakamura-
Yonehara K, Noda M (2008) Expression of SPIG1 reveals development
of a retinal ganglion cell subtype projecting to the medial terminal nucleus
in the mouse. PLoS One 3:e1533.
Yonehara K, Ishikane H, Sakuta H, Shintani T, Nakamura-Yonehara K, Ka-
miji NL, Usui S, Noda M (2009) Identification of retinal ganglion cells
and their projections involved in central transmission of information
about upward and downward image motion. PLoS One 4:e4320.
Zhou ZJ, Lee S (2008) Synaptic physiology of direction selectivity in the
retina. J Physiol 586:4371 4376.
178
J. exp. Biol. 146, 255-275 (1989) 255
Printed in Great Britain The Company of Biologists Limited 1989

CODING AND PROCESSING OF ELECTROSENSORY

INFORMATION IN GYMNOTIFORM FISH

BY WALTER HEILIGENBERG
Scripps Institution of Oceanography, Neurobiology Unit, University of
California at San Diego, La Jolla, CA 92093, USA

Summary
Studies of the electrosensory system of gymnotiform fish have revealed
principles of neuronal coding and processing of information which also character-
ize more advanced systems, such as vision and audition in higher vertebrates.
1. Animals may have different classes of receptors adapted to code different
variables within a given modality, and the separation of their central projections
provides the basis for independent initial processing of these variables by higher-
order neurones.
2. These separate pathways, however, eventually converge at the level of still
higher-order neurones which are adapted to 'recognize' particular spatial and
temporal constellations, or patterns, of the stimulus variables conveyed by these
pathways.
3. As different stimulus patterns may control different forms of behavioural
responses, corresponding neuronal structures can be identified which are adapted
to recognize specific patterns. Neurones at an early level of pattern discrimination
may still show very general response properties, whereas neurones closer to the
ultimate control of a given behaviour show more specific response properties.
These latter are less sensitive to stimulus features which are irrelevant to the
control of the behaviour, and they code relevant features more purely and with
higher acuity than do lower-level neurones.
4. The acuity of stimulus discrimination displayed by some high-order neurones
may rival that observed at the behavioural level. This high sensitivity is achieved
through pooling and integration of information supplied by large populations of
less-sensitive receptors and lower-order neurones.

Introduction
Animals are naturally adapted to detect and to process very specific stimulus
patterns in their environment, and studies of their ecology and ethology help us to
identify such stimulus patterns as well as natural behavioural responses which they
evoke and control. These responses, in turn, provide crucial assays for the
experimental dissection of stimulus patterns and for the identification of their
relevant components. Moreover, experimental manipulation of these components
and their configuration within patterns enable us to pin down the computational

Key words: electrosensation, neuroethology, neuronal computations, neural networks, recog-

nition units.

179
256 W. HEILIGENBERG

rules which underlie the evaluation of patterns by the animal's brain. In the course
of this analysis, we will then gain confidence in our conceptual models of
perception to the extent that we are able to predict behavioural consequences that
are due to specific alterations in stimulus patterns.
After behavioural experiments have identified computational rules of percep-
tion, their specific neuronal implementations can be determined by physiological
and anatomical approaches. In view of the overwhelming structural and functional
complexities of brain mechanisms one would prefer to study neuronal mechanisms
in the relatively simple systems found in invertebrates and lower vertebrates, such
as fish and amphibians. The small size of their brains as well as the more modest
number of structural components and interconnections facilitate such an analysis
at the single-cell level.
The electrosensory system of certain orders of fish appears to have a relatively
simple structural and functional organization, but it also displays sophisticated
features, such as complex receptive fields, ordered central representations and
modulations of sensory processing by recurrent descending pathways, which are
known in more 'advanced' systems, such as mammalian vision and hearing. The
basic neuronal principles of sensory information processing appear to be the same
across different modalities and classes of organisms. The greater 'transparency' of
simple systems, however, makes the study of these principles more feasible, and
their successful exploration in simpler systems should ultimately enhance our
understanding of more complex systems.
What makes the electrosensory system particularly appealing are behavioural
responses, such as the jamming avoidance response (JAR), which remain intact in
physiological preparations. Studies at the single-cell level can thus be combined
with simultaneous behavioural assays to test the significance of stimulus patterns at
both levels as well as to monitor the overall intactness of the system.
The following review summarizes our knowledge of coding and central
processing in the electrosensory system of gymnotiform fish and points out
commonalities with other systems, particularly the auditory system of the barn owl
(Konishi et al. 1988). Various aspects of the anatomy and physiology of the
electrosensory system have recently been presented in great detail (Bullock &
Heiligenberg, 1986; Bastian, 1986a; Carr & Maler, 1986; Heiligenberg, 1986) so
that references can largely be limited to more recent work. Issues of information
processing have been addressed separately in later publications (Heiligenberg,
1987,1988). This presentation will start with a description of stimulus coding at the
periphery and then focus on central mechanisms of information processing, with
particular emphasis on the jamming avoidance response (JAR) of Eigenmannia.
The summary diagram in Fig. 1 shows the flow of information controlling this
behaviour and may serve as a reference throughout this presentation.

Coding at the periphery

Electric fish assess objects in their environment by monitoring distortions of the

180
 Processing of electrosensory information 257
LOCATION STRUCTURE FUNCTION
Electric organ (EO) discharges at frequency
Tail EO
of the PN.
\
Pacemaker nucleus (PN) drives each discharge
Hindbrain PN cycle of EO by single command pulse.

Diencephalon
t,
PPN
Neurones in the prepacemaker nucleus (PPN) are excited
by Df<0 and inhibited by Df >0. Their excitation
raises the frequency of the PN.
I i

Nucleus electrosensorius (NE) has two classes

Diencephalon NE of higher-order 'sign-selective' cells, excited
by Df < 0 and Df >0, respectively.

Neurones in lamina 6 of the torus semicircularis (TS)

Midbrain TS/TO compute differential phase. Differential phase
information and amplitude information converge
\ f in deeper layers of the TS and tectum opticum (TO),
yielding two classes of 'sign-selective' cells,
c excited by Df<0 and Df >0, respectively.
o
c
_c
- cc
5
=
4>
a3
" "

Local phase information is relayed to lamina 6

a of the TS.
a.
E
j =
D-
<

Electrosensory lateral line lobe (ELL) processes

Hindbrain ELL amplitude and phase information independently.

Body surface
tf
I> T
P- and T-type receptors code local amplitude and
phase, respectively.

Fig. 1. The flow of information controlling the jamming avoidance response (JAR) in
Eigenmannia. The fish generates a continuous, nearly sinusoidal electric organ
discharge (EOD) with a fundamental frequency of a few hundred hertz. The electric
signal caused by the interference of a neighbour's EODs with the animal's own EODs
is modulated in its instantaneous amplitude and phase at the difference frequency, Df,
of the two EODs. The animal minimizes the effect of jamming by a frequency too close
to its own by lowering its frequency in response to small positive Dfs (i.e. if the
neighbour's EOD frequency is slightly higher than its own) and by raising its frequency
in response to small negative Dfs. The necessary information about the sign of Df is
contained in the pattern of amplitude and phase modulations distributed over the
animal's body surface (see details in Heiligenberg, 1986, 1988).

181
258 W. HEILIGENBERG

H.-H

12

oJ i 0J
0
H.-H,
Fig. 2. Objects which differ electrically from the surrounding water distort the
animal's electric organ discharge (EOD) signal. (A) The nearly sinusoidal EOD of a
fish is recorded at two points, A and B, on its body surface. A small perturbation at A
could affect the local EOD amplitude, | S \A, as well as the EOD phase, or the timing of
the zerocrossing, HAHB, with reference to that of an unperturbed EOD at point B.
Sampling of amplitude and phase values at successive EOD cycles, t l t t2, ..., yields a
string of paired values which can be plotted in a two-dimensional amplitude-phase
plane to form a graph (B). This graph is coded by the activity of two classes of
'tuberous' electroreceptors (C), P-type receptors which fire intermittently, and T-type
receptors which fire one spike on each EOD cycle, phase-locked to the zerocrossing of
the signal. The probability, pA, of firing of P-type receptors at A codes the local
amplitude, ] S \A, whereas the difference, T^TB, in the timing of action potentials of
T-type receptors at A and B codes the differential phase, H^Hs. The EOD
distortions have been exaggerated for demonstration purposes. Realistic values are
shown in Figs 3 and 4.

electric field generated by their own electric organ discharges (EODs). Electrical
measurements in gymnotiform fish with continuous, nearly sinusoidal EODs show
that the amplitude as well as the phase (i.e. the timing of zerocrossings) of the
signal perceived by the animal are affected by the ohmic and capacitive loads of a
nearby object (Rose & Heiligenberg, 1986a). The spatial pattern of alterations of
amplitude and phase, monitored by large arrays of electroreceptors on the body
surface, represents the electric image of the object (Figs 2, 3).
Amplitude and phase of the signal at the animal's body surface are also
modulated when EODs of a neighbour interfere with the animal's own EODs. A
simultaneous evaluation of amplitude and phase modulations by the central^

182
 Processing of electrosensory information 259
nervous system enables the animal to determine the sign of the frequency
difference between the two interfering EODs and to shift its own EOD frequency
away from that of its neighbour. In this way, the animal avoids jamming of
electrosensory inputs by frequencies too close to its own (Fig. 4). This jamming
avoidance response (JAR) has become a powerful tool for analysing central
mechanisms for the processing of amplitude and phase information.
Amplitude and phase are coded by separate types of 'tuberous' electroreceptors
distributed over the body surface. P-type receptors fire intermittently and
modulate their rate of firing in accordance with modulations in the local amplitude
of the signal. T-type receptors fire one spike in each EOD cycle, at a fixed phase in
reference to the zerocrossing of the signal. Separate primary afferents relay
responses of P- and T-type receptors to different targets in the electrosensory
lateral line lobe (ELL) which contains three somatotopically organized represen-
tations of the animal's body surface. Each representation receives the same
primary afferent information via axon collaterals, and each has distinct strata of
neurones for separate processing of amplitude and phase information.
In addition to tuberous electroreceptors, which are tuned to the fundamental
frequency of the animal's own EODs, electric fish also have 'ampullary' electrore-
ceptors which are sensitive to low-frequency signals. Such signals originate from
various animate and inanimate sources which the fish passes in its environment
(Kalmijn, 1987). Ampullary receptors project to a separate, single map of the body
surface in the ELL.
Gymnotiform fish with wave-type EODs thus have separate classes of electrore-
ceptors coding different aspects of the electrical environment. This separation of
channels is maintained at the first-order processing station, the ELL of the
hindbrain. A similar separation is seen in gymnotiform fish with pulse-type
discharges, although the organization and functional significance of their ampli-
tude- and phase-coding systems have not yet been explored in detail.
Two classes of electroreceptors, sensitive to different frequency ranges, are also
found in the African mormyriform fish which evolved an electrosensory system
independently from the gymnotiforms. 'Small pores', or ampullary receptors,
respond to low-frequency signals, and two types of tuberous electroreceptors are
tuned to the higher frequency range of the animal's EODs. In contrast to the
situation in gymnotiforms, however, these two types of tuberous receptors cannot
be considered as phase and amplitude coders. Instead, one type, called Knollenor-
gans, is adapted to code EODs of neighbours (Hopkins, 1986), whereas the second
type, called Mormyromasts, is adapted to code alterations of the animal's own
EOD field. These two types project to different central targets, and their
information is gated differently by a corollary signal of the electric organ
pacemaker (see Bell, 1989).

Further filtering of information provided by primary afferents

Although different types of electroreceptors are adapted to code specific types
aspects of electric signals, their higher-order 'representations' in the central

183
260 W. HEILIGENBERG

A R closed

l 2-0-

<

-30 -10 0

R open R open
2-0- A 00
" no C
2-0-

E
>

D Eigenmannia
a
< 2-0- VA

1-9-

-20 - 1 0 6 0 10 20 -10 0
Phase (/)
Fig. 3

nervous system appear to be even more selective in their response properties. The
P- and T-type receptors of gymnotiform fish, for example, project to different
classes of first-order neurones in the electrosensory lateral line lobe (ELL) of the
hindbrain (Maler et al. 1981; Mathieson et al. 1987). P-receptors, whose firing rate
is modulated in accordance with variations of signal amplitude, contact 'basilar'
pyramidal cells (or E-units) directly, and 'non-basilar' pyramidal cells (or I-units)
indirectly via inhibitory interneurones. As a consequence of these connections and
additional network properties, such as descending recurrent inputs (Bastian,
1986a,b; Bastian & Bratton, 1988), E-units are excited by a rise in signa|

184
 Processing of electrosensory information 261
Fig. 3. Modulations in EOD amplitude and phase depend upon the electrical qualities
of the object. The inset in the upper left is a top view of a fish temporarily immobilized
by injection of a curare-like drug. Since this drug also silences the animal's EOD, a
substitute signal, S, of similar frequency (500Hz), amplitude and field geometry has
been provided by a signal generator and electrodes placed inside the mouth and at the
tip of the tail. The EOD-like signal is recorded by a pair of differential electrodes (+)
at the side of the animal's head. An object is moved sinusoidally at a rate of 0-5 Hz back
and forth between the locations A and B, approximately 1 cm away from and parallel
to the animal's body surface. In A-D the local peak-to-peak amplitude of 5 is plotted
on the ordinate, and its phase (i.e. its zerocrossing), measured in microseconds and in
reference to the cycle of the generator, is plotted on the abscissa. Since the frequency
of 5 is 500 Hz, a full cycle (i.e. a phase value of 2K) measures 2 ms. A signal amplitude
of 2mVcm"' and a phase defined as zero are recorded with the object in position B,
which is distant from the animal. In A-C, the object is a thin, 1-5 cm wide, vertically
suspended strip of metal foil, insulated on the side facing away from the animal. This
foil is connected via a capacitor, C, or an electrical short-circuit (R closed) to a distant
point in the water. In A dots mark every tenth cycle of the EOD-like signal, S, as the
object travels from A to B. Arrows indicate the direction of motion in the
amplitude-phase plane. Although the physical motion from A to B is, theoretically,
symmetrical to the motion from B to A, slight turbulence-related asymmetries in the
motion cause differences in the respective sections of the graph. Different natural
objects were chosen in D, a leaf of the water plant Vallisneria, approximately 1-5 cm
wide; a bulbous leaf stem of the water hyacinth, Eichhomia, approximately 2cm in
diameter; and a conspecific fish, with its EOD silenced by MS 222 anaesthesia. (From
Rose & Heiligenberg, 1986a.)

amplitude, whereas I-units are excited by a fall in signal amplitude. Both cell types
thus respond to changes (i.e. the temporal derivative) of signal amplitude and, in
this regard, resemble 'on' and 'off elements known in visual systems. Their
somatotopic organization within the ELL ensures that the animal can monitor the
spatial order of changes in signal amplitude over its entire body surface.
Although P-type receptor afferents fire action potentials within a preferred
phase range of the nearly sinusoidal wave form of the EOD and thus, at least
statistically, convey information about the phase of the signal, their representa-
tives in the ELL fire with almost constant probability at any phase of the EOD.
This is particularly the case for the I-units which, owing to the intervening
inhibitory interneurones, are synaptically even further removed from the recep-
tors. The E- and I-units in the ELL thus discard the small amount of phase
information that is left in the firing of their primary afferents and are totally
devoted to the coding of amplitude information.
T-receptor afferents code phase with some jitter as the timing of their action
potentials with reference to the EOD cycle shows a standard deviation of the order
of 30/is. Several T-receptor afferents from a given location on the body surface,
however, converge upon a single spherical cell in the ELL (Fig. 5). As a
consequence of their electrotonic synapses and the electrical properties of the
spherical cell and its spike initiation zone, the action potentials relayed by the
Spherical cell phase-lock to the EOD cycle with less jitter than do individual

185
262 W. HEILIGENBERG

Probability of P-unit firing

0 J ,-
0 2n
Beat cycle Beat cycle

Fig. 4. Modulations in EOD amplitude and phase are also induced by the interfering
EODs of a neighbour, and the temporal and spatial patterns of these modulations
reflect the sign of the difference between the interfering frequencies. Following the
notation introduced in Fig. 2, the diagrams in the top row show the modulation in
amplitude, j S |^ (left) and phase, H^HB, (right), at a point A on the body surface
which experiences a mixture of the animal's own signal and that of its neighbour. The
phase reference point, B, is assumed not to be exposed to the neighbour's signal, and
the amplitude at point A measured in the absence of the neighbour's signal is defined as
1. Amplitude and phase modulations are recorded over one full beat cycle, indicated
schematically underneath the abscissae. (Note that with the animal's EOD frequency
being 500 Hz and that of its neighbour being 501 Hz, the beat frequency would be 1 Hz,
i.e. the beat cycle would be 1 s long and would cover 500 EOD cycles.) A joint plot of
the amplitude and phase modulations in the amplitude-phase plane yields a circle, and
the sense of rotation of this circle reflects the sign of the frequency difference: the
counterclockwise sense indicated implies that the neighbour has a higher frequency.
The lower row of diagrams shows the neuronal representation of the events in the top
row. Amplitude is coded by the probability, p^, of local P-type receptor firing, whereas
phase is coded by the difference in the timing of action potentials, T^Tg, of T-type
receptors located at points A and B on the body surface. The values plotted for p^ are
the mean numbers of spikes recorded from a single primary afferent per successive
EODs in the beat cycle. Data from 25 beat cycles were averaged. Similarly, the values
plotted for T^TB are means from 25 beat cycles. (From Heiligenberg & Partridge,
1981.)

T-receptor afferents. Most importantly, a single action potential of a T-receptor

afferent not arriving in sufficient synchrony with those of its neighbours is
insufficient to trigger an action potential in the spherical cell. 'Outliers' are thus
discarded by this process of phase averaging. This non-linear mechanism of
averaging, originally proposed by Maler et al. (1981) on the basis of ultrastructurai

186
 Processing of electrosensory information 263
studies, has meanwhile been confirmed by intracellular recordings from spherical
cells in Eigenmannia (M. Kawasaki & W. Heiligenberg, unpublished obser-
vation). The same phenomenon has been observed in the giant cells of lamina 6 of
the torus which receive converging inputs from spherical cells (Carr et al. 1986a).
Although some T-afferents may respond to an increase in signal amplitude by
slightly advancing the timing of their action potential within the EOD cycle, the
timing of the action potentials of spherical cells is rather insensitive to modulations
of signal amplitude. Spherical cells thus provide phase information which is not
only less jittery but also purer than phase information provided by T-receptors.
A particular form of signal filtering is employed in the processing of primary
afferent information in the mormyrids. The Knollenorgans have lower thresholds
than the Mormyromasts and are, therefore, suitable for the detection of EODs of
distant neighbours. In addition, by firing only a single spike in response to an EOD
pulse, they only code the occurrence of this event but not its intensity. Afferent
information from Knollenorgans is gated by an inhibition generated by a corollary
discharge of the animal's electric organ pacemaker system, so that action
potentials caused by the animal's own EODs are not relayed to higher-order
neurones in the midbrain. Central information provided by Knollenorgans,
therefore, only represents the electrical presence of neighbours (for more details
see Bell, 1989).
The higher threshold of Mormyromasts makes this type of receptor less sensitive
to the EODs of neighbours and, by firing a burst of spikes which increases with the
amplitude of the signal, they code local intensity of the animal's own EOD.
Moreover, their input to higher-order neurones is facilitated by a corollary
discharge of the electric-organ pacemaker system so that the central nervous
system is informed preferentially about local intensities of the animal's own signal.
The Mormyromast system is thus dedicated to the analysis of electric images coded
by alterations of the animal's own EOD field (for more details see Bell, 1989).
Other sensory systems similarly show separate channels which are adapted to
code different aspects within their modality, and a further separation of these
aspects may be achieved by separate central pathways. Well-known examples are
the magno- and parvocellular systems in mammalian vision.
Although the availability of different classes of receptors facilitates separate
coding of different stimulus aspects, their presence is not always necessary to
achieve such a separation at a central level. The auditory system of the barn owl
presents an example. Primary auditory afferents are tuned to individual frequency
bands and respond to modulations in signal amplitude by modulating their rate of
firing accordingly. Although these afferents are unable to fire an action potential
on each cycle of a carrier signal of sufficiently high frequency, they are still able to
code its phase by firing their action potentials at a preferred moment within the
carrier cycle. Primary afferents send axon collaterals to two nuclei, the nucleus
angularis, which processes amplitude information, and the nucleus magnocellular-
is, which processes phase information. Neurones in the nucleus angularis
modulate their rate of firing in accordance with amplitude modulations of the

187
 264 W. HEILIGENBERG

TECTUM

CD
V \ Lamina 3

Q
t Lamina 5
TORUS
^- Lamina 6
c Lamina 7
t Lamina 8

PC Spherical cell Basilar Non-basilar

m pyramidal pyramidal *)
Q cell cell /T
Z
ELL (E-unit) (I-unit)

T-afferent P-afferent
(phase coder) 6 (amplitude coder)
RECEPTORS
o

Fig. 5

auditory signal within their frequency band, but the timing of their spikes shows
almost no correlation with the phase of the carrier. Neurones in the nucleus
magnocellularis, however, fire action potentials which appear to be even more
tightly locked to the phase of the carrier than are those of primary afferents, but
their rate of firing does not vary with the amplitude of the carrier. Differences in
dendritic and synaptic organization within these two nuclei must account for this
separation of amplitude information and phase information, which from then on
are processed separately up to the level of the midbrain, much as in the case of the
electrosensory system of gymnotiform fish (see Takahashi, 1989).

188
 Processing of electrosensory information 265
Fig. 5. The somatotopic representation of electrosensory information in the hindbrain
and midbrain. Four maps of the body surface are found on each side of the hindbrain in
the electrosensory lateral line lobe (ELL). Although the most medial map only
receives input from ampullary receptors, the remaining three maps receive identical
input through collaterals of primary afferents from tuberous electroreceptors,
P-afferents as well T-afferents. The organization of first-order neurones in the
tuberous pathway of the ELL is shown on the right. T-afferents from small receptive
fields on the body surface converge via electrotonic synapses upon spherical cells
which, in turn, project in somatotopic order to lamina 6 of the torus semicircularis of
the midbrain. P-afferents form excitatory synapses on the basilar dendrites of basilar-
pyramidal cells (E-units) and, via inhibitory interneurones, also contact non-basilar
pyramidal cells (I-units). As a consequence of these connections, E-units are excited by
a rise in stimulus amplitude, whereas I-units are excited by a fall. These amplitude-
coding cells project in somatotopic order to various laminae of the torus semicircularis
above and below lamina 6. Note that the torus as well as the next-order station, the
tectum opticum, both contain only a single map of the body surface. (From
Heiligenberg, 1988.)

Transformations of information and the coding of derived variables

The firing rate of P-type receptor afferents is determined by the level of the
signal amplitude as well as by its temporal changes. Units are strongly excited by a
sudden rise in amplitude and then partially adapt as the new amplitude level is
maintained. Conversely, a sudden drop in signal amplitude to a new level causes a
transient drop in firing rate. Owing to local network properties as well as
descending recurrent pathways (Bastian, 1986a,b; Bastian & Bratton, 1988;
Shumway & Maler, 1989), the next higher-order neurones in the ELL differ in
their response properties from primary afferents in that they almost exclusively
code changes in signal amplitude and largely ignore its level. These neurones of
the ELL are thus adapted to detect the motion of objects as well as temporal
modulations in interference patterns by monitoring changes in local signal
amplitude regardless of its mean level. An even more significant transformation of
information is seen in the phase-coding system.
Modulations in phase differences between signals at different sites of the body
surface are significant for the control of the JAR as well as for the detection of
moving objects. Such phase differences are detected within the somatotopically
organized lamina 6 of the torus semicircularis of the midbrain. This lamina
receives somatotopically ordered phase information from the spherical cells of the
ELL. Specific local connections within this lamina generate a network capable of
detecting phase differences between inputs from the area of body surface
represented at a given point on this map and inputs from a variety of other areas on
the body surface, which serve as phase references (Carr et al. 19866). The
spherical cells of the ELL as well as their relays, the giant cells in lamina 6, code
instantaneous local phase by firing one action potential on each EOD cycle at a
fixed latency with reference to the zerocrossing of the signal. The 'small' cells of
this lamina, however, compare the timing of spikes arriving from two different
sites on the body surface. They fire irregularly, and their probability of firing varies

189
266 W. HEILIGENBERG

with the difference in the timing of their two inputs. These small cells, therefore,
only code differential phase, and the timing of their spikes is, effectively, no longer
related to the individual timing of their inputs. The system discards information
about the absolute timing of signals at individual sites on the body surface and only
retains the more relevant information about differences in timing. This difference
is now coded by the rate of firing of neurones dedicated to the comparison of
inputs from two distinct sites on the body surface. A small neurone within the
location of lamina 6 representing a specific site, A, on the body surface, for
example, will raise its rate of firing if the signal in A experiences a small phase lead
with reference to the signal in some specific area, B, and the same neurone will
lower its rate of firing if area A experiences a phase lag instead. A nearby neurone
of this kind also represents area A, but it may have a different phase reference
area, C, on the body surface (Fig. 6).
Similar and, in some cases, more elaborate transformations of information
occur at still higher levels of the nervous system in connection with the analysis of
complex stimulus patterns. The differential phase information, for example,
provided by the small neurones in lamina 6 for a given area on the body surface is
combined with amplitude information from the same area in deeper laminae of the
torus and tectum.

Central computations and pattern analysis

Amplitude and phase information each have a specific temporal and spatial
order which requires a separate analysis. The ordered representation of infor-
mation in strata of higher-order neurones facilitates such computations.
As was just mentioned, the somatotopically organized lamina 6 of the torus
semicircularis of gymnotiform fish is dedicated to the computation of the
differential phase between signals at specific sites on the body surface. An auditory
structure with a function analogous to that of lamina 6 is the nucleus laminaris in
owls. This paired nucleus receives tonotopically ordered phase information from
the two ears and is adapted to detect specific interaural phase differences at each
carrier frequency (see details in Takahashi, 1989).
Neuroanatomical and ultrastructural studies suggest that a separate and ordered
representation of a stimulus variable, such as phase, facilitates local computations
of patterns within this variable by means of simple neuronal modules. Such
modules form a lattice of identical computational entities spread over the stratum
of neurones which constitute the neuronal map of the stimulus variable, and the
connection of these modules appears to follow fairly simple and statistical rules
(see Fig. 6). A rather limited set of genetic instructions, therefore, may suffice for
their ontogeny.
Modulations in the amplitude and phase of a periodic signal are two aspects
which require joint evaluation to identify the nature of the pattern. A specific
order of amplitude and phase modulations in the pattern formed by the
interference of two EODs, for example, reveals the sign of their frequency1

190
 Processing of electrosensory information 267

Giant
cells

Small
cells
I H.-Hr
Torus
Lamina 6

Ho

H fl

Spherical
cells I B(r c i D(t. ELL

Fig. 6. Lamina 6 of the torus semicircularis contains small cells which code differential
phase between signals arriving from different points, A,B,C,D, on the body surface.
Spherical cells of the ELL (see Fig. 5), coding signal phase, H^,... ,H D , in their
respective receptive fields, project in somatotopic order to the giant cells of lamina 6 as
well as to the dendrites of small cells in their vicinity. (Although several spherical cells
may converge on a single giant cell, only single-cell connections have been drawn here
for simplicity.) Giant cells, in turn, send axonal collaterals across wide areas of lamina
6 and contact the somata of small cells. The size of their synapses appears to preclude
that more than one giant cell can contact any given small cell. Small cells in the location
of lamina 6 representing point A on the body surface modulate their rate of firing in
accordance with the small temporal disparities, H^-H*, between action potentials
arriving from area A at their dendrites and from specific reference areas, B,C,D, on
their soma. (From Heiligenberg, 1987, and based upon data from Carr et al. 19866.)

difference and allows the fish to shift its own frequency in the appropriate
direction. If the neighbour's EOD frequency is lower than the animal's own, for
example, large areas of the body surface will predominantly experience rises in
local stimulus amplitude paired with phase leads and falls in local stimulus
amplitude paired with phase lags. Higher-order neurones in the deeper laminae of
the torus and tectum recognize one or the other of these combinations by
apparently gating amplitude inputs by phase inputs (Heiligenberg & Rose, 1986).
It is essential for this form of evaluation that amplitude and phase modulations be
compared in spatial register, and this ordered comparison is facilitated by the
laminar organization of the torus semicircularis of the midbrain. While lamina 6
processes phase information, laminae 5 and 7 receive somatotopically ordered

191
268 W. HEILIGENBERG

Phase discriminator
Spike rate

0
AM coder ( HB

Sign-selective neurone d
o o
Fig. 7. The gating of amplitude information by phase information. Sign-selective
neurones of the midbrain respond to particular combinations of changes in local
amplitude, | S \A, in a given area A on the body surface and values of differential phase,
HAHB, between this area and some reference area, B. Sign-selective neurones,
therefore, can discriminate the sense of rotation of graphs in the amplitude-phase
plane. The unit on the left is excited by a clockwise rotation centred at zero, whereas
the unit on the right is excited by a counterclockwise rotation. The AM coders
represent higher-order E- and I-type neurones which are excited by a rise and fall,
respectively, in local signal amplitude. Phase discriminators occur first in the form of
small cells in lamina 6 of the torus (see Fig. 6). Higher-order phase discriminators are
found in deeper laminae of the torus and in the tectum. The synaptic organization
shown here is inferred from responses of sign-selective neurones to independent
manipulations of amplitude and phase inputs. (From Heiligenberg & Rose, 1986.)

amplitude information. Since all laminae are stacked parallel to each other as well
as in somatotopic register, local perpendicular connections between laminae
provide connections necessary for the joint computation of local amplitude and
phase information. As a consequence of this ordered arrangement, very simple
rules for interlaminar connections should suffice to provide the necessary ordered
inputs of amplitude and phase information for higher-order neurones (Heiligen-
berg, 1987) (Fig. 7).

The emergence of 'recognition' neurones

A comparison of filter properties across neurones at different hierarchical levels
reveals that units closer to the motor control of a given behaviour respond more
selectively to the stimulus pattern eliciting this behaviour and are more insensitive
to irrelevant stimulus features. The JAR of Eigenmannia, for example, depends
upon the sign of the frequency difference, Df, between a neighbour's EOD and
the animal's own EOD, and absolute frequency differences of between 2 and 8 Hz
are most effective. The orientation of the neighbour's EOD field, however, is of
marginal significance in that it may only affect the strength of the response. This

192
 Processing of electrosensory information 269
behaviourally defined selectivity for a stimulus pattern gradually emerges at the
neuronal level through several levels of integration of sensory information.
Eigenmannia discriminates the sign of Df by simultaneous evaluation of
modulations in local signal amplitude and differential phase. Accordingly, the
lowest-order neurones sensitive to the sign of Df are found in the torus
semicircularis, where the amplitude- and phase-coding pathways converge. Since
individual 'sign-selective' neurones of the torus, however, receive information
only from limited receptive fields on the body surface, their responses depend
significantly upon the orientation of the interfering EOD field, and a change in the
orientation may even reverse a neurone's sign preference (Rose & Heiligenberg,
1986a). The activity of a single neurone, therefore, does not reflect the sign of Df
unambiguously. One can show, however, that the pooled activity of a large
population of such neurones, with receptive fields distributed over a wide area of
the body surface, should always yield reliable information about the sign of Df
(Heiligenberg & Rose, 1986). The population of sign-selective cells in the torus
can indeed be compared with a parliament in which members cast conflicting
votes, and a behavioural decision reflects the opinion of the majority. One can
even show that a minority of members, due to deficiencies in the transfer of phase
information, always casts the 'wrong' vote (Heiligenberg, 1987). The sign-selective
neurones of the torus differ widely with regard to the magnitude of Df at which
they respond best, although the majority prefers magnitudes of Df in the
behaviourally optimal range between 2 and 8 Hz.
The nucleus electrosensorius complex of the diencephalon receives direct and,
via the tectum opticum, indirect input from the torus. Sign-selective cells at this
level appear to obtain converging information from larger populations of sign-
selective cells of the torus and tectum and, as a likely consequence, discriminate
the sign of Df more reliably. Although their degree of sign discrimination may still
depend upon the orientation of the interfering EOD field, their sign preference
never reverses (Keller, 1988). Two additional features result from the assumed
convergence of sensory information from torus to nucleus electrosensorius.
Neurones in the latter structure are more sensitive to weak interfering EOD fields,
some of them almost reaching behavioural threshold, and the rate of their firing is
less strongly locked to the temporal modulation pattern of amplitude and phase
generated by the interfering EODs. The neurones of the nucleus electrosensorius,
however, do not yet fully reflect the dynamic properties of the JAR as at least
some of them prefer magnitudes of Df outside the behaviourally optimal range
2-8 Hz. Moreover, a neurone responding most vigorously to positive Dfs in the
range 2-8 Hz may not be inhibited evenly across all values of negative Dfs. It may
actually be more active for negative Dfs between 2 and 8 Hz than for positive
Dfs outside the range 2-8Hz. The firing rate of such a neurone, therefore, still
does not reveal the sign of Df unambiguously for all magnitudes of Df (Keller,
1988).
These minor differences between behavioural and neuronal response properties
vanish at a still higher level of integration, in the diencephalic prepacemaker

193
270 W. HEILIGENBERG

nucleus which innervates the final motor command structure, the electric-organ
pacemaker nucleus in the medulla of the hindbrain (Kawasaki et al. 1988a). Small
neurones of the prepacemaker raise their rate of firing in response to negative Dfs
(i.e. when the neighbour's EOD frequency is lower), and the animal raises its own
EOD frequency so that a larger separation of the interfering frequencies is
achieved. The same neurones lower their rate of firing in response to positive Dfs
which cause a lowering of the animal's EOD frequency. The firing rate of any
single neurone thus unambiguously reflects the sign of Df, and the strength of the
neuronal response is directly related to the strength of the behavioural response.
These sign-selective neurones are rather homogeneous in their response charac-
teristics, which closely resemble the behavioural response. In particular, they
show practically the same dependence upon the magnitude of Df as does the JAR,
many of them are as sensitive to weak interfering fields as is the behavioural
response of the intact animal, and the response of many neurones varies as little
with the orientation of the interfering EOD field as does the intact behaviour.
Finally, the firing rate of these neurones is no longer modulated at the difference
frequency, Df, with the exception of those very small magnitudes of Df which also
cause corresponding weak modulations in the animal's own EOD frequency (Rose
et al. 1988). The small cells of the prepacemaker nucleus, therefore, code a
behavioural effort, namely to change the frequency of the pacemaker, and their
activity is no longer affected by aspects of the stimulus regime which are irrelevant
to this particular behavioural response.
Although torus and tectum are structured somatotopically, no obvious somato-
topic order is found at the next higher level, the nucleus electrosensorius which,
instead, reveals a motor map. Whereas stimulation of one portion of this nucleus
raises the pacemaker frequency, stimulation of a more ventral and rostral portion
lowers the pacemaker frequency. Both structures are essential for the JAR, since
their selective, bilateral lesion abolishes the animal's ability either to raise or to
lower its frequency, respectively (Keller & Heiligenberg, 1989).

Multiple representations and sharing of information

The primary afferents from tuberous electroreceptors send collaterals to three
somatotopically ordered maps in the ELL which, at first glance, show very similar
anatomical and physiological properties. A recent analysis by Shumway (1989a,b),
however, has revealed significant functional differences between these three maps.
Owing to larger spatial convergence, pyramidal cells in the most lateral of the
three maps have larger receptive fields and are more sensitive to small modu-
lations in signal amplitude than pyramidal cells in the central-medial map. At the
same time, the pyramidal cells of the lateral map respond more rapidly to
amplitude modulations and thus provide better temporal resolution than the
pyramidal cells of the central-medial map. Neurones of the central-lateral map,
which lies between the lateral map and the central-medial map, show degrees of
spatial and temporal resolution intermediate to those of their neighbours. The

194
 Processing of electrosensory information 271
availability of multiple maps receiving identical information thus offers the
opportunity for processing this information under different aspects in separate,
specialized networks. It appears plausible that additional maps originated as
'accidental' reduplications of existing maps and were then progressively adapted
for specific tasks in sensory processing. An evolutionary scenario of this kind was
proposed by AUman et al. (1981) for multiple visual representations in mammals.
Multiple presentations are found at all levels of the nervous system, from
sensory processing to motor control. Neurones which respond very selectively to a
specific stimulus pattern and appear to be dedicated to the control of a particular
kind of behaviour emerge rather late within the neuronal hierarchy. Lower-order
neurones, even at a level as high as the tectum opticum, still have rather general
response characteristics. Electrosensory neurones of the tectum of gymnotiform
fish respond to certain combinations of phase and amplitude modulations which
may be generated by moving objects as well as by interfering EODs of neighbours
(Rose & Heiligenberg, 1986a; Heiligenberg & Rose, 1987). The firing of such a
neurone, therefore, tells us little about the world outside, unless we monitor the
activity of this neurone for a period much longer than the behavioural response
latency of the animal. Simultaneous recordings from many such neurones and
parallel evaluation of the spatial and temporal order of their activity, however,
would allow us to discriminate between moving objects and interfering EODs
almost immediately. These tectal neurones send axonal collaterals to various
targets in the diencephalon, midbrain and hindbrain so that their information can
be processed by different algorithms in separate higher-order structures. One of
these structures, the nucleus electrosensorius, receives additional inputs from the
torus semicircularis and participates in the control of the JAR (Keller &
Heiligenberg, 1989). Other structures in the reticular formation are apparently
adapted to monitor the motion of objects and to control movements of the body.
Various behavioural systems thus share rather general spatial and temporal
information provided by the tectum and are able to extract specific patterns for
their own control.

Neuronal convergence enhances the sensitivity of higher-order sensory neurones

The enhancement of sensitivity with convergence is illustrated by the accuracy
with which phase is coded during the JAR. The depth of phase and amplitude
modulations caused by the interference of a neighbour's EOD with the animal's
own EOD decreases monotonically with the relative amplitude of the interfering
EOD. The strength of the JAR, therefore, decreases as the distance to the
neighbour increases. Studies of the JAR of Eigenmannia have shown that these
animals can resolve extremely small modulations in the differential timing of
signals at different sites on their body surface. As long as a sufficiently large area of
the body is exposed to the interference pattern generated by the animal's own
EOD and the neighbour's EOD, the fish can detect phase differences smaller than
0-5/zs. Larger phase differences are required to drive the JAR if the interference

195
272 W. HEILIGENBERG

pattern is restricted to smaller regions of the body and thus to a smaller set of
electroreceptors (Rose & Heiligenberg, 1985). This extreme temporal resolution
is surprising in view of the considerable jitter observed in the phase-coding system.
As was mentioned previously, T-type afferents record the timing of the
zerocrossing of a signal by firing an action potential at a specific latency, with a
standard deviation of approximately 30/is. Through convergence at the level of
the ELL and within lamina 6 of the torus, this jitter is reduced by a factor of
approximately three at the level of lamina 6 (Carr et al. 1986a). Higher-order units
of the torus can resolve differential phase values as small as 10 pts,, and averaging of
their activity over periods much longer than the latency of the JAR reveals
thresholds for their phase discrimination close to that measured behaviourally
(Rose & Heiligenberg, 19866). Converging inputs from large numbers of such
neurones at the level of the nucleus electrosensorius and further convergence from
this nucleus to the prepacemaker nucleus should, therefore, yield single neurones
with a temporal resolution close to that of the intact behaviour. Neurones of such
sensitivity are indeed found in the prepacemaker nucleus, and, much as is true for
the intact behaviour, they require sensory inputs from sufficiently large areas of
the body surface to achieve this high sensitivity (Kawasaki et al. 19886).

Enhanced stimulus resolution in ordered maps of broadly tuned receivers

Biologically relevant stimulus variables, such as the location of a stimulus source
in space, the frequency of a tone or the time delay of an echo, are commonly coded
by broadly tuned neurones which form ordered neuronal maps. As one moves
along an axis of a map, neurones tuned to progressively higher values of the
stimulus variable are encountered in one particular direction. As a consequence of
this order, a given value of this variable is coded by the location of the centre of a
smooth distribution of excitation within the stratum of neurones. One can
demonstrate theoretically that a form of weighted averaging of inputs provided by
these neurones to higher-order elements can yield a degree of stimulus discrimina-
tion which far surpasses that of individual neurones in the map (Heiligenberg,
1987; Baldi & Heiligenberg, 1988). Although no clear example of this mechanism
can yet be demonstrated in the electrosensory system, its functional and
developmental simplicity suggest that nature has long made use of it somewhere.
Mittelstaedt & Eggert (1988) proposed such a mechanism of 'map weighting' for
the transformation of spatially ordered information into oriented motor com-
mands. J. P. Miller (personal communication) suggests that a set of broadly tuned
wind detectors on the cerci of crickets may yield more accurate spatial resolution
at a higher level by virtue of such weighted averaging.

The analysis and simulation of neural networks

Intracellular labelling of physiologically identified neurones is a common tool
for relating functional and morphological properties of neurones. To verify

196
 Processing of electrosensory information 273
specific types of synaptic connections and, thereby, to establish the rules for
connections within networks, rather time-consuming ultrastructural studies of
labelled neurones are necessary (Carr et al. 1986i>; Mathieson et al. 1987). In some
instances, in vivo experiments can be complemented by in vitro experiments on
structures isolated in a slice chamber. This approach, allowing for better control of
ionic conditions and the presence of agonists or antagonists of certain transmitters,
has been very fruitful in the study of the ELL (Mathieson & Maler, 1988) and the
medullary pacemaker (Dye, 1988) by revealing the functional significance of
certain transmitters. These findings could then be tested in vivo by applying
specific agonists or antagonists to particular locations within these structures
(Shumway & Maler, 1989; Dye etal. 1988). The structural and functional
organization of the ELL has now been identified in great detail, and the response
properties of its neurones can, at least qualitatively, be explained on the basis of
their specific connections and the action of known transmitters. A quantitative
theory of the operation of the ELL, however, will not only require more detailed
studies of its dynamic properties but also network simulations, preferably
performed on large computers structured in parallel.

Developmental and evolutionary considerations

Studies of neuronal systems always invite questions and speculations about their
developmental and evolutionary history. Neuronal computations postulated on
the basis of behavioural and physiological experiments appear plausible to the
extent that they can be implemented by simple and robust network structures, and
the development and self-organization of these structures should require a rather
limited amount of genetic information. The formulation of realistic theories of
brain function very much depends upon a thorough knowledge of developmental
constraints.
The analysis of brain mechanisms commonly reveals designs that human
engineers would not have chosen. Brains appear to be built rather sloppily and are
full of imperfections and patchwork which has accumulated in the course of their
evolution (Heiligenberg, 1987, 1988). That they are still able to function with such
perfection and that they are so tolerant of perturbations and trauma is all the more
remarkable.

I thank M. Konishi, M. Kawasaki, C. Keller and C. E. Carr for most helpful

comments on this manuscript. This work is supported by grants from NIMH,
NINCDS and NSF.

References
ALLMAN, J. M., BAKER, J. F., NEWSOME, W. T. & PETERSEN, S. E. (1981). Visual topography
and function. In Cortical Sensory Origination, vol. 2 (ed. C. N. Woolsey), pp. 171-185.
Qifton, New Jersey: Humana Press.
BALDI, P. & HEIUGENBERG, W. (1988). How sensory maps could enhance resolution through
ordered arrangements of broadly tuned receivers. Biol. Cybernetics 59, 313-318.

197
274 W. HEILIGENBERG

BASTIAN, J. (1986C). Electrolocation: behavior, anatomy, and physiology. In Electroreception

(ed. T. H. Bullock & W. Heiligenberg), pp. 577-612. New York: Wiley & Sons.
BASTIAN, J. (19866). Gain control in the electrosensory system: a role for the descending
projections to the electrosensory lateral line lobe. /. comp. Physiol. A 158, 505-515.
BASTIAN, J. & BRATTON, B. O. (1988). The nucleus praeeminentialis: Properties of neurons
providing descending inputs to the electrosensory lateral line lobe of weakly electric fish.
Neurosc. Abstr. 128.9, 18th Annual Meeting, Toronto.
BELL, C. C. (1989). Sensory coding and corollary discharge effects in momyrid electric fish.
J. exp. Biol. 146, 229-253.
BULLOCK, T. H. & HEILIGENBERG, W. (1986). Electroreception. New York: Wiley & Sons.
CARR, C. E., HEILIGENBERG, W. & ROSE, G. J. (1986a). A time-comparison circuit in the electric
fish midbrain. I. Behavior and Physiology. /. Neuroscience 6, 107-119.
CARR, C. E. & MALER, L. (1986). Electroreception in gymnotiform fish: Central anatomy and
physiology. In Electroreception (ed. T. H. Bullock & W. Heiligenberg), pp. 319-373. New
York: Wiley & Sons.
CARR, C. E., MALER, L. & TAYLOR, B. (19866). A time-comparison circuit in the electric fish
midbrain. II. Functional morphology. /. Neuroscience 6, 1372-1383.
DYE, J. (1988). An in vitro physiological preparation of a vertebrate communicatory behavior:
chirping in the weakly electric fish, Apteronotus. J. comp. Physiol. A 163, 445-458.
DYE, J., HEILIGENBERG, W., KELLER, C. & KAWASAKI, M. (1988). Different glutamate receptors
mediate distinct behaviors in a single brainstem nucleus in weakly electric fish. Neurosci.
Abstr. 86.8, 18th Annual Meeting, Toronto.
HEILIGENBERG, W. (1986). Jamming Avoidance Responses: Model systems for neuroethology.
In Electroreception (ed. T. H. Bullock & W. Heiligenberg), pp. 613-650. New York: Wiley &
Sons.
HEILIGENBERG, W. (1987). Central processing of sensory information in electric fish. /. comp.
Physiol. A 161, 621-631.
HEILIGENBERG, W. (1988). Neural mechanisms of perception and motor control in a weakly
electric fish. In Advances in the Study of Behavior, vol. 18 (ed. J. S. Rosenblatt, C. Beer,
M. C. Busnel & P. J. B. Slater), pp. 73-98. San Diego: Academic Press.
HEILIGENBERG, W. & PARTRIDGE, B. L. (1981). How electroreceptors encode JAR-eliciting
stimulus regimes: Reading trajectories in a phase-amplitude plane. J. comp. Physiol. 142,
295-308.
HEILIGENBERG, W. & ROSE, G. J. (1986). Gating of sensory information: Joint computations of
phase and amplitude data in the midbrain of the electric fish Eigenmannia. J. comp. Physiol.
A 159, 311-324.
HEILIGENBERG, W. & ROSE, G. J. (1987). The optic tectum of the gymnotiform electric fish,
Eigenmannia: Labeling of physiologically identified cells. Neuroscience 22, 331-340.
HOPKINS, C. D. (1986). Behavior of Mormyridae. In Electroreception (ed. T. H. Bullock &
W. Heiligenberg), pp. 527-576. New York: Wiley & Sons.
KALMIJN, A. (1987). Detection of weak elecric fields. In Sensory Biology of Aquatic Animals
(ed. J. Atema, R. R. Fay, A. N. Popper & W. N. Tavolga), pp. 151-186. New York: Springer
Verlag.
KAWASAKI, M., MALER, L., ROSE, G. J. & HEILIGENBERG, W. (1988a). Anatomical and
functional organization of the prepacemaker nucleus in gymnotiform electric fish: The
accommodation of two behaviors in one nucleus. J. comp. Neurol. 276, 113-131.
KAWASAKI, M., ROSE, G. J. & HEILIGENBERG, W. (19886). Temporal hyperacuity in single
neurons of electric fish. Nature, Lond. 336,173-176.
KELLER, C. (1988). Stimulus discrimination in the diencephalon of Eigenmannia: the emergence
and sharpening of a sensory filter. J. comp. Physiol. A 162, 747-757.
KELLER, C. & HEILIGENBERG, W. (1989). From distributed sensory processing to discrete motor
representations in the diencephalon of the electric fish, Eigenmannia. J. comp. Physiol. A
164, 565-576.
KONISHI, M., TAKAHASHI, T. T., WAGNER, H., SULLIVAN, W. E. & CARR, C. E. (1988).
Neurophysiological and anatomical substrates of sound localization in the owl. In Auditory
Function (ed. G. M. Edelman, W. E. Gall & W. M. Cowan), pp. 721-745. New York: John
Wiley & Sons.

198
 Processing of electrosensory information 275
MALER, L., SAS, E. & RODGERS, J. (1981). The cytology of the posterior lateral line lobe of high-
frequency weakly electric fish (Gymnotidae): Dendritic differentiation and synaptic
specificity in a simple cortex. J. comp. Neurol. 195, 87-140.
MATHIESON, W. B., HEIUGENBERG, W. & MALER, L. (1987). Ultrastructural studies of
physiologically identified electrosensory afferent synapses in the gymnotiform fish,
Eigenmannia. J. comp. Neurol. 255, 526-537.
MATHIESON, W. B. & MALER, L. (1988). Morphological and electrophysiological properties of a
novel in vitro preparation: the electrosensory lateral line lobe brain slice. /. comp. Physiol. A
163, 489-506.
MITTELSTAEDT, H. & EGGERT, T. (1988). How to transform topographically ordered spatial
information into motor commands? In Visuomotor Coordination: Amphibians, Comparisons,
Models and Robots (ed. M. A. Arbib & J. P. Ewert), pp. 569-585. New York: Plenum Press.
ROSE, G. J. & HEIUGENBERG, W. (1985). Temporal hyperacuity in the electric sense of fish.
Nature, Lond. 318, 178-180.
ROSE, G. J. & HEIUGENBERG, W. (1986a). Neural coding of difference frequencies in the
midbrain of the electric fish Eigenmannia: reading the sense of rotation in an amplitude-phase
plane. J. comp. Physiol. A 158, 613-624.
ROSE, G. J. & HEILIGENBERG, W. (1986b). Limits of phase and amplitude sensitivity in the torus
semicircularis of Eigenmannia. J. comp. Physiol. A 159, 813-822.
ROSE, G. J., KAWASAKI, M. & HEILIGENBERG, W. (1988). 'Recognition units' at the top of a
neuronal hierarchy? J. comp. Physiol. A 162, 759-772.
SHUMWAY, C. A. (1989a). Multiple electrosensory maps in the medulla of weakly electric
gymnotiform fish. I. Physiological differences. J. Neuroscience (in press).
SHUMWAY, C. A. (19896). Multiple electrosensory maps in the medulla of weakly electric
gymnotiform fish. II. Anatomical differences. /. Neuroscience (in press).
SHUMWAY, C. A. & MALER, L. (1989). GABAergic inhibition shapes temporal and spatial
response properties of pyramidal cells in the electrosensory lateral line lobe of gymnotoid fish.
/. comp. Physiol. A 164, 391-407.
TAKAHASHI, T. (1989). The neural coding of auditory space. J. exp. Biol. 146, 307-322.

199
200
The Journal of Experimental Biology 202, 13111318 (1999) 1311
Printed in Great Britain The Company of Biologists Limited 1999
JEB2088

CENTRAL MECHANISMS OF TEMPORAL ANALYSIS IN THE KNOLLENORGAN

PATHWAY OF MORMYRID ELECTRIC FISH
MATTHEW A. XU-FRIEDMAN1,* AND CARL D. HOPKINS2
1Department of Neurobiology, Harvard Medical School, 220 Longwood Avenue, Boston, MA 02115, USA and
2Section of Neurobiology and Behavior, Cornell University, Seeley Mudd Hall, Ithaca, NY 14853, USA

*e-mail: mfriedman@hms.harvard.edu

Accepted 1 February; published on WWW 21 April 1999

Summary
Mormyrid electric fish communicate using pulse-type appears to take place in the nucleus exterolateralis pars
electric organ discharges (EODs). The fine temporal anterior (ELa), where tightly phase-locked inputs from the
structure of the waveforms of EODs varies widely hindbrain drive a direct, excitatory input through a long
throughout the 200 or more species of mormyrids. These axonal delay line and also drive an indirect, inhibitory
signals carry information about the species, the sex and input with negligible delay through the ELa large cell.
even the individual identity of the signaller. Behavioral These two inputs converge on ELa small cells, where they
experiments have shown that some species of fish are are hypothesized to interact in a delay-line/blanking
capable of using this information. Of the four known types model. This initial temporal analysis is further refined in
of electroreceptors in mormyrid fish, the knollenorgan the nucleus exterolateralis pars posterior, where units
electroreceptor is the one most likely to be involved in the tuned to ranges of pulse durations have been identified
detection of conspecific EOD signals. Here, we review some physiologically.
recent advances in understanding how the central
knollenorgan pathway might be analyzing the temporal Key words: time-coding, knollenorgan, mormyrid, electric organ
structure of the EOD waveform. Fine temporal analysis discharge, waveform.

Introduction
Electric communication signals retain their fine temporal
structure as they are transmitted from signaller to receiver in
the aquatic environment, unlike auditory signals, which are
corrupted by echoes and reverberation (Hopkins, 1986b). This
enables weakly electric fish to communicate useful information
even with their brief electric organ discharges (EODs). EODs
from different species of African mormyrid electric fish vary
widely in waveform (14 phases), duration (60 s to 20 ms)
and polarity. The EODs of many species show sexual
differences in waveform (Kramer, 1997; Hopkins, 1999). It is
possible in some cases to recognize individuals on the basis of
their EODs (Crawford, 1992; Friedman and Hopkins, 1996).
Mormyrids use EOD waveforms to discriminate the species
and sex of the signaller (Hopkins, 1981; Hopkins and Bass,
1981; Graff and Kramer, 1992). One behavioral study
addressed the mechanism by which they perform this task.
Males of the species referred to as Brienomyrus brachyistius
(tp) (Hopkins and Bass, 1981) or Brienomyrus sp. 2 (Alves-
Gomes and Hopkins, 1997) exhibit a sex difference in the EOD
waveform. Males responded with courtship displays (rasps or
bursts of EODs delivered at a high rate) when presented with
square pulses of a duration that matched female EODs, but they
failed to respond to shorter or longer square pulses or to pulses
derived from a phase-shifted female EOD (Hopkins and Bass,
1981). These experiments suggested that the salient feature of
the EOD is its temporal waveform structure.
Mormyrid fish have four different kinds of electroreceptors
used for active and passive electrolocation and communication.
Among these, knollenorgan receptors appear the most likely to
detect and to respond to communication signals from other
fish. First, they have a high sensitivity and they are broadly
tuned to the spectrum of the species own EOD (Bass and
Hopkins, 1980, 1984). Furthermore, if parts of the
knollenorgan pathway in the central nervous system (CNS) are
lesioned, some communication behaviors are abolished
(Moller and Szabo, 1981). The motor areas that generate the
fishs own EOD send a strong inhibitory input to the primary
sensory areas in the hindbrain (Zipser and Bennett, 1976; Bell
and Grant, 1989). Thus, the primary afferents from
knollenorgan receptors respond to the fishs own discharge, but
this response never gets past the hindbrain electrosensory
lateral line lobe (ELL), making the fish deaf to its own
discharges. Together, these characteristics imply that the
knollenorgan pathway is specialized for detecting and
processing EOD signals generated by neighboring, conspecific
fish.
Knollenorgan electroreceptors are scattered over the body
surface. For a stimulus consisting of a brief pulse of electrical

201
1312 M. XU-FRIEDMAN AND C. D. HOPKINS
B Normal
A polarity
Reverse
polarity

EOD stimulus

0.5 ms

C D
OB

ELa ll
Tel ELp Ipsi
MV
Contra
OT
VA Valvula
ELa
IG SPE
ELp

1 Aff

ELL

nPLL Knollenorgan
1 mm NELL

Fig. 1. Summary of the knollenorgan pathway. (A) Drawing of a mormyrid fish, Brienomyrus brachyistius. (B) Response of a knollenorgan
(upper trace) to an electric organ discharge (EOD) stimulus (lower trace). The response is displayed as a post-stimulus time histogram. For the
upward histogram, the EOD stimulus shown was presented to the knollenorgan multiple times. For the downward histogram, the inverted EOD
stimulus was presented. The downward histogram is equivalent to the responses of knollenorgans on the opposite side of the body. Thus,
knollenorgans on one side of the body respond to the upward edge of the stimulus (with some delay) while knollenorgans on the opposite side
respond to the downward edge. Bar, 25 spikes per bin. (C) Overview of the brain of Brienomyrus brachyistius. On the right, the valvula
cerebelli has been removed. ELa, nucleus exterolateralis pars anterior; ELL, electrosensory lateral line lobe; ELp, nucleus exterolateralis pars
posterior; nPLL, posterior lateral line nerve; OB, olfactory bulb; OT, optic tectum; Tel, telencephalon; VA, valvula cerebelli. (D) Overview of
the known connections of the knollenorgan pathway. Contra, contralateral; IG, isthmic granule nucleus; Ipsi, ipsilateral; ll, lateral lemniscus;
MV, nucleus medialis ventralis; NELL, nucleus of the ELL; SPE, subpreminential nucleus; 1 Aff, primary afferent. Modified from Amagai
et al. (1998) and Friedman and Hopkins (1998).

current, either a square pulse or an EOD, knollenorgans on one
side of the body respond to the onset of current with a phase-
locked spike, and those on the opposite side of the body
respond to the offset (Fig. 1B; Hopkins and Bass, 1981;
Hopkins, 1986a). Therefore, fish could analyze the temporal
structure of the stimulus by comparing the temporal pattern of
spikes arriving from knollenorgan electroreceptors all over the
body. In the simplest case, they could measure the duration of
the pulse by measuring the time difference between the two
sides of the body (Hopkins and Bass, 1981). This comparison
must occur in the CNS, and here we describe some anatomical
and physiological characteristics of the knollenorgan pathway
that could provide the neural basis for this task.
Specializations for temporal analysis
Knollenorgan electroreceptors project roughly somato-
topically to the ipsilateral nucleus of the electrosensory lateral
line lobe (NELL) (Fig. 1D; Bell and Russell, 1978; Szabo et
al., 1983; Hopkins et al., 1993). The terminals onto the NELL
cells are large and electrotonic (Mugnaini and Maler, 1987b).
The NELL somata are large and adendritic (Szabo and
Ravaille, 1976) and are strongly immunoreactive for calretinin
(Friedman and Kawasaki, 1997). These anatomical and
neurochemical characteristics are commonly associated with
neural pathways specialized for the preservation of fine
temporal information, such as the auditory azimuthal
localization pathways in birds and mammals (Konishi, 1991)
and the electrosensory phase-coding systems in the electric fish
Eigenmannia and Gymnarchus (Carr, 1986; Kawasaki, 1997).
The NELL sends a thick axon bilaterally up the lateral
lemniscus to end in two nuclei in the torus semicircularis: the
nucleus medialis ventralis (MV) and the nucleus exterolateralis
pars anterior (ELa) (Fig. 1D; Szabo et al., 1983). The axon
branch to the MV is thin, and the terminals there are small en
passant boutons (Amagai et al., 1998; Friedman and Hopkins,

202

1998). In contrast, the axon branch to the ELa is thick and
heavily myelinated, and the terminals it makes are large and
electrotonic. Thus, the ELa shows anatomical specializations
typically associated with the preservation of fine temporal
information, but the MV does not. In addition, the midbrain is
the first stage in the knollenorgan pathway where information
from all parts of the body can be compared. Therefore, the ELa
is more likely to be the site where temporal analysis takes place
(Szabo et al., 1983), while the MV may perform some other
function, such as spatial analysis.
There are only two cell types in the ELa, large cells and
small cells (also called interstitial and granule cells by
Mugnaini and Maler, 1987a). They are easily differentiated by
size (approximately 10 m versus approximately 6 m in soma
diameter). The NELL axon terminates onto ELa large cells and
small cells, with mixed chemical and electrical synapses,
suggesting that synaptic delays are extremely short (Mugnaini
and Maler, 1987a). ELa large cells and small cells both appear
to be adendritic, according to intracellular dye-fills of ELa
large cells with biocytin and Lucifer Yellow and retrograde
labelling of small cells with biotinylated dextrans (Amagai et
al., 1998; Friedman and Hopkins, 1998), although Mugnaini
and Maler (1987a) found dendritic arborizations on ELa large
cells using Golgi staining. In any case, the terminals from
NELL axons all appear to be directly onto the large cell and
small cell somata, making it possible to identify completely
each postsynaptic cell contacted by an intracellularly labelled
NELL axon. Reconstructions show that NELL axons terminate
on 12 large cells, then wind extensively and terminate on
small cells throughout the ELa (Fig. 2A) (Friedman and
Hopkins, 1998).
ELa large cells project entirely within the ELa (Fig. 2B),
terminating with large calyceal terminals that envelop their
postsynaptic small cells. The large cells are probably inhibitory
because they are immunopositive for glutamic acid
decarboxylase, the enzyme that makes the neurotransmitter
-aminobutyric acid (GABA) (Mugnaini and Maler, 1987a).
The large cell axons project fairly directly across the nucleus,
ending on more restricted bands or patches of small cells
(Friedman and Hopkins, 1998).
Linear reconstructions (Fig. 2C) show that, when a NELL
axon enters the ELa, it first contacts a large cell and then travels
for approximately 1 mm, where it may contact a second large
cell; it then travels for 34 mm with few terminals, before
branching widely over a large number of small cells (Friedman
and Hopkins, 1998). The distance between the first large cell
terminal and the last small cell terminal may be as great as
7 mm, even though the ELa is only 1 mm in diameter. This
difference in axonal delay translates into a NELL cell
activating its first large cell 230460 s earlier than its last
small cell, using a first estimate of conduction velocity of
15 m s1 for the NELL axons (Enger et al., 1976). In contrast,
the large cell axon runs only approximately 1 mm from soma
to terminals (equivalent to 60 s), so it relays signals quickly
to its postsynaptic small cells. Thus, the small cell appears to
receive an indirect, inhibitory input from a large cell, and a
Knollenorgan pathway in mormyrids 1313
direct, but delayed, excitatory input from a NELL cell. Also,
the delays to different small cells may vary because NELL
cells show a range of axonal lengths.
Like knollenorgan electroreceptors, both NELL axons and
ELa large cells respond to electrosensory stimuli with a tightly
phase-locked spike (Amagai et al., 1998; Friedman and
Hopkins, 1998). NELL axons and ELa large cells are not
segregated according to their receptive field locations because
a single recording electrode advanced through the ELa runs
across cells at different depths that respond to either one edge
of the stimulus or the other (Amagai et al., 1998).
Mugnaini and Maler (1987a) first suggested that inhibition
could be used in a temporal blanking model. Refining this
model to include subsequent discoveries about physiological
responses and axonal arborizations, we show how the
anatomical specializations within the ELa could be responsible
for fine temporal analysis (Friedman and Hopkins, 1998). In
this model, a small cell receives one input from a large cell
with its receptive field on one side of the body (Fig. 3A). It
also receives a second input from a NELL cell with a receptive
field on the opposite side of the body, but with some delay
between the large cell and NELL inputs. For a sufficiently long
negative-going pulse, the excitatory NELL input will arrive
first, and the small cell will respond. However, if the pulse is
shorter than the axonal delay, then the inhibitory large cell
input will arrive first, and the NELL input will be suppressed.
In other words, a small cell responds if the pulse is longer than
some threshold duration, set by the length of the NELL axonal
delay. Fig. 3B illustrates the predicted response probability as
a function of square pulse duration and polarity. For a positive-
going pulse, the inhibitory large cell input would always arrive
first, suppressing responses to the excitatory input for the
duration of the inhibitory postsynaptic potential (IPSP), which
is presumably long enough to block all behaviorally relevant
stimuli. Thus, the ELa small cell should respond only to one
polarity of the stimulus. If different NELL axons have different
axonal delays, then different small cells would be tuned to
discriminate between stimuli of different duration. By
combining the inputs of many small cells, each specialized for
a different delay and for different patches of the body surface
and polarities of stimulation, there should be sufficient
information to make waveform duration discriminations.
This delay line-blanking model is supported by some
preliminary data. Recordings from the ELa, putatively from
small cells, have as predicted two different kinds of synaptic
activity, which can be tied to different edges of the stimulus
and with different latencies (Friedman and Hopkins, 1998).
These synaptic potentials are consistent with direct inhibitory
input from large cells and delayed excitation from NELL
axons. However, data are available from only a few small cells
because their small size and adendritic somata limit successful
penetrations.
Further processing of temporal information
ELa small cells project exclusively to the neighboring
203
1314 M. XU-FRIEDMAN AND C. D. HOPKINS

A B

1
2
3
4

100 m 5
6
7

100 m

0 1 2 3 4 5 6 7
Relative axon length (mm)
Fig. 2. Anatomy of cell types in the nucleus exterolateralis pars anterior (ELa). (A) Reconstruction of a nucleus of the electrosensory lateral
line lobe (NELL) cell axonal arborization within the ELa. Lateral is left, anterior is up. The cell was filled with biocytin, and the reconstruction
is made from 15 sections of 50 m thickness. Terminals onto small cells are green, and terminals onto large cells are red. The thickness of the
axon denotes its dorsal-to-ventral extent. The outline of the nucleus is taken from the largest section encompassing the entire reconstruction.
(B) Reconstruction of an ELa large cell, from seven sections of 50 m thickness. The soma is red, and the terminals onto small cells are green.
(C) Linear reconstructions of four NELL axonal arborizations and one ELa large cell. Modified from Friedman and Hopkins (1998).

midbrain nucleus exterolateralis pars posterior (ELp) They project straight across the nucleus, preserving their
(Hauged-Carr, 1979). The small cell axons within the ELp topographical organization (Friedman and Hopkins, 1998).
are thin, making small en passant boutons along their length. The ELp has two cell types, identified from extracellular

204

Fig. 3. Small cell model. (A) Small cell inputs. Knollenorgans
in one patch of the body surface (red) respond to the upward
edge of a square pulse, and contribute signals to a nucleus of
the electrosensory lateral line lobe (NELL) axon that
terminates on a nucleus exterolateralis pars anterior (ELa)
large cell, which in turn makes an inhibitory contact onto a
small cell. On the opposite side of the body (green patch), the
knollenorgans respond to the downward edge of the stimulus
and contribute signals to a NELL axon (green) that terminates
directly on the small cell, but with a relative delay tdelay. This
is only one example of possible receptive field organization.
(B) Proposed small cell responses. The abscissa represents the
duration and polarity of a square pulse stimulus, and the
ordinate is the response probability. For stimuli of one
polarity, small cells respond to stimuli longer than some
threshold duration. The threshold depends on the delay line,
which is different for different cells. Modified from Friedman
and Hopkins (1998).
recordings (Amagai, 1998). Type I cells respond at a shorter
latency (79 ms), with lower jitter and with higher response
probability, whereas type II cells respond at longer latency
(1220 ms), with higher jitter and with lower response
probability (Amagai, 1998). Most interestingly, type I cells
respond to square pulses, provided that the pulse is longer than
some threshold duration (long pass; Fig. 4A). The thresholds
of type I cells range from 0.02 to 0.2 ms. This response profile
is similar to that predicted for the ELa small cells, supporting
the delay line-blanking model described above. Type II cells
respond to square pulses, provided that the pulse duration is
within a restricted range of durations (band-pass; Fig. 4B).
The best response varied between 0.1 and 10 ms for different
cells. The responses of type II cells could therefore be used by
the fish to make decisions about the duration of a signallers
EOD. Important physiological issues left to resolve are the
complex effects of stimulus amplitude and geometry on these
responses (Amagai, 1998) and the mechanism that could
underlie discrimination of pulse durations as long as 10 ms,
since one based on axonal delay lines seems unlikely.
Of these two physiological types, only the anatomy of type
I cells has been reconstructed, as shown in Fig. 4C (M. A.
Friedman and C. D. Hopkins, unpublished results). They have
large (150 m), spiny dendritic arborizations that stretch
perpendicular to the thin, incoming small cell axons. This
morphology contrasts strongly with the anatomical
specializations described in the previous section for NELL and
ELa cells. This change in cellular structure is correlated with
the increase in response jitter in the ELp, suggesting that the
anatomical specializations evident in the NELL and the ELa
allow fine temporal discrimination to be made in the ELa, after
which spike times need not be so precise. The dendritic
arborizations in the ELp are more likely to function to integrate
Knollenorgan pathway in mormyrids 1315
A
NELL ELa large cell
t t
+
-
+ ELa small cell
tdelay
NELL
B
100% response probability
Pulse duration (t t )
the outputs from the ELa in a manner less dependent on precise
spike times, although more detailed physiological studies are
necessary.
Individual type I cells project widely throughout the
midbrain, with terminals in several areas: in two clusters in the
ELp, one near the soma and one several hundred micrometers
distant, in the medial ventral nucleus (MV), in the ipsilateral
and contralateral isthmic granule nucleus (IG) and in the
subpreminential nucleus (SPE) (Fig. 4C). The terminals
within the ELp imply the presence of complex local circuitry
and, indeed, intracellular recordings show several phases of
excitation and inhibition (M. A. Friedman and C. D. Hopkins,
unpublished results). The areas outside the ELp are unexplored
physiologically, so the role of the wide terminal field of
individual ELp cells is not known, but it implies that the
information that the type I cell carries is significant.
Anatomical studies have shown that the IG projects to the
valvula cerebelli (Finger et al., 1981) and that the MV projects
to the optic tectum (OT; Wulliman and Northcutt, 1990). The
OT (the homolog of the superior colliculus) plays a role in
spatial analysis of signals from many sensory modalities in
vertebrates (Bastian, 1982; Knudsen, 1982; Bartels et al., 1990;
Stein and Meredith, 1993), suggesting that the MV may be
involved in spatial analysis of knollenorgan information.
Comparative considerations
When we compare the known connections of the
knollenorgan pathway with other octavolateral pathways, such
as the ampullary/mormyromast system and the auditory
system, several parallels become evident (Fig. 5). In all three
systems, there is a direct pathway from the receptor, through
the hindbrain, to a distinct toral nucleus and finally to the optic
205
1316 M. XU-FRIEDMAN AND C. D. HOPKINS

Fig. 4. Physiology and anatomy of the nucleus exterolateralis pars posterior (ELp). (A) Normalized responses of type I cells, showing long-
pass tuning. (B) Normalized responses of type II cells, showing band-pass tuning. (C) Anatomical reconstruction of a type I cell and
terminals in other midbrain nuclei (outlined in bold). The cell was processed for biocytin staining and reconstructed from 26 sections of 50 m
thickness. Numbers next to bold outlines of other brain nuclei indicate their depths relative to the ELp soma. Terminals are found in the
subpreminential nucleus (SPE), the ipsi- and contralateral isthmic granule nucleus (IG) (enlarged in the lower right inset), the nucleus
medialis ventralis (MV) and the ELp, both close to the soma (enlarged in the lower left inset, in red) and in a distinct cluster 400 m ventral to
the soma. The cells have widely branching dendritic arborizations (lower left inset, in green). ELa, nucleus exterolateralis pars anterior; ll,
lateral lemniscus; mm, mesomesencephalic tract; OT, optic tectum; tp, toropreminential tract. A and B are modified from Amagai (1998).

206

OT OT OT
ELa MV L MD
ELp
IG SPE PE IG VPE
NELL ELL dzD
Knollenorgan Ampullary Auditory
Mormyromast
Fig. 5. Parallels in octavolateral pathways in mormyrid electric fish.
Outlined are the major connections of the knollenorgan (this review),
ampullary/mormyromast (simplified from Bell and Szabo, 1986) and
auditory (simplified from Bell, 1981a,b; Kozloski and Crawford,
1998) pathways. dzD, dorsomedial zone of the descending nucleus;
ELa, nucleus exterolateralis pars anterior; ELL, electrosensory lateral
line lobe; ELp, nucleus exterolateralis pars posterior; IG, isthmic
granule nucleus; L, nucleus lateralis; MD, nucleus medialis dorsalis;
MV, nucleus medialis ventralis; NELL, nucleus of the ELL; OT,
optic tectum; PE, preminential nucleus; SPE, subpreminential
nucleus; VPE, ventral preminential nucleus.
tectum. The toral nucleus may also feed back to a part of the
preminential nucleus, which may itself feed back to the
hindbrain. These parallels reinforce the hypothesis that
electrosensory pathways evolved by duplication of auditory/
mechanosensory pathways. The MV in the knollenorgan
pathway occupies a position analogous to that of the toral
nucleus lateralis (L) in the ampullary/mormyromast system
and medialis dorsalis (MD) in the auditory/mechanosensory
system, but the knollenorgan system uniquely adds a further
two toral nuclei, the ELa and the ELp. Interestingly, the sister
species to the mormyrids, Gymnarchus niloticus, lacks both
knollenorgans and a recognizable nucleus exterolateralis (Bass
and Hopkins, 1982). The ELa and ELp are presumably
necessitated by the special temporal analysis proposed to be
carried out in the knollenorgan pathway.
It is also instructive to compare the knollenorgan pathway
with pathways in other species that analyze fine temporal
information. In the avian auditory brainstem, time differences
between the two ears are compared to determine the azimuthal
location of a sound. Time comparison is performed by a delay-
line-coincidence detector, where signals from the left and right
nucleus magnocellularis project in opposite directions across
the nucleus laminaris (Young and Rubel, 1983; Carr and
Konishi, 1990). The magnocellular projections are highly
ordered compared with the axonal arborizations of the NELL
cells. This difference presumably reflects the more complicated
Knollenorgan pathway in mormyrids 1317
computational task that the ELa is performing. In the avian
brainstem, it is two ears being compared, whereas in the ELa,
all parts of the body surface are being compared. In this sense,
the phase-coding pathways of South American electric fish are
more similar to those of the mormyrid. In Eigenmannia, the
spherical cell afferent from the ELL ends on a giant cell and
nearby small cells in layer VI of the torus semicircularis (Carr
et al., 1986; Heiligenberg, 1991). In this case, the giant cell
distributes the signal widely throughout layer VI, analogously
to the NELL axon in mormyrids. The major departure of the
mormyrid system is that it appears to use an inhibitory
mechanism in measuring time differences, whereas the barn
owl and Eigenmannia use only excitatory synapses.
References
Alvez-Gomes, J. and Hopkins, C. D. (1997). Molecular insights into
the phylogeny of mormyriform fishes and the evolution of their
electric organs. Brain Behav. Evol. 49, 324351.
Amagai, S. (1998). Time-coding in the midbrain of mormyrid electric
fish. II. Stimulus selectivity in the nucleus exterolateralis pars
posterior. J. Comp. Physiol. A 182, 131143.
Amagai, S., Friedman, M. A. and Hopkins, C. D. (1998). Time-
coding in the midbrain of mormyrid electric fish. I. Physiology and
anatomy of cells in the nucleus exterolateralis pars anterior. J.
Comp. Physiol. A 182, 115130.
Bartels, M., Mnz, H. and Claas, B. (1990). Representation of
lateral line and electrosensory systems in the midbrain of the
axolotl, Ambystoma mexicanum. J. Comp. Physiol. A 167,
347356.
Bass, A. H. and Hopkins, C. D. (1980). Coding of species-specific
signals in mormyrid electric fish. I. Frequency characteristics. Soc.
Neurosci. Abstr. 6, 604.
Bass, A. H. and Hopkins, C. D. (1982). Comparative aspects of brain
organization of an African wave electric fish, Gymnarchus
niloticus. J. Morph. 174, 313334.
Bass, A. H. and Hopkins, C. D. (1984). Shifts in frequency tuning
of electroreceptors in androgen-treated mormyrid fish. J. Comp.
Physiol. A 155, 713724.
Bastian, J. (1982). Vision and electroreception: Integration of
sensory information in the optic tectum of the weakly electric fish
Apteronotus albifrons. J. Comp. Physiol. A 147, 287297.
Bell, C. C. (1981a). Central distribution of octavolateral afferents and
efferents in a teleost (Mormyridae). J. Comp. Neurol. 195,
391414.
Bell, C. C. (1981b). Some central connections of medullary
octavolateral centers in a mormyrid fish. In Hearing and Sound
Communication in Fishes (ed. W. N. Tavolga, A. N. Popper and R.
R. Fay), pp. 383391. New York: Springer-Verlag.
Bell, C. C. and Grant, K. (1989). Corollary discharge inhibition and
preservation of temporal information in a sensory nucleus of
mormyrid electric fish. J. Neurosci. 9, 10291044.
Bell, C. C. and Russell, C. J. (1978). Termination of electroreceptor
and mechanical lateral line afferents in the mormyrid
acousticolateral area. J. Comp. Neurol. 182, 367382.
Bell, C. C. and Szabo, T. (1986). Electroreception in mormyrid
fish: Central anatomy. In Electroreception (ed. T. H. Bullock
and W. Heiligenberg), pp. 375421. New York: John Wiley &
Sons.
207
1318 M. XU-FRIEDMAN AND C. D. HOPKINS
Carr, C. E. (1986). Time coding in electric fish and barn owls. Brain
Behav. Evol. 28, 122133.
Carr, C. E. and Konishi, M. (1990). A circuit for detection of
interaural time differences in the brain stem of the barn owl. J.
Neurosci. 10, 32273246.
Carr, C. E., Maler, L. and Taylor, B. (1986). A time-comparison
circuit in the electric fish midbrain. II. Functional morphology. J.
Neurosci. 6, 13721383.
Crawford, J. D. (1992). Individual and sex specificity in the electric
organ discharges of breeding mormyrid fish (Pollimyrus isidori). J.
Exp. Biol. 164, 79102.
Enger, P. S., Libouban, S. and Szabo, T. (1976). Fast conducting
electrosensory pathway in the mormyrid fish, Gnathonemus
petersii. Neurosci. Lett. 2, 133136.
Finger, T. E., Bell, C. C. and Russell, C. J. (1981). Electrosensory
pathways to the valvula cerebelli in mormyrid fish. Exp. Brain Res.
42, 2333.
Friedman, M. A. and Hopkins, C. D. (1996). Tracking mormyrid
electric fish in the field using individual differences in electric
organ discharges. Anim. Behav. 51, 391407.
Friedman, M. A. and Hopkins, C. D. (1998). Neural substrates for
species recognition in the time-coding electrosensory pathway of
mormyrid electric fish. J. Neurosci. 18, 11711185.
Friedman, M. A. and Kawasaki, M. (1997). Calretinin-like
immunoreactivity in mormyrid and gymnarchid electrosensory and
electromotor systems. J. Comp. Neurol. 387, 341357.
Graff, C. and Kramer, B. (1992). Trained weakly-electric fishes
Pollimyrus isidori and Gnathonemus petersii (Mormyridae,
Teleostei) discriminate between waveforms of electric pulse
discharges. Ethology 90, 279292.
Hauged-Carr, F. (1979). The mesencephalic exterolateral
posterior nucleus of the mormyrid fish Brienomyrus niger: Efferent
connections studied by the HRP method. Brain Res. 178, 179184.
Heiligenberg, W. (1991). Neural Nets in Electric Fish. Cambridge,
MA: MIT Press.
Hopkins, C. D. (1981). On the diversity of electric signals in a
community of mormyrid electric fish in West Africa. Am. Zool. 21,
211222.
Hopkins, C. D. (1986a). Behavior of Mormyridae. In
Electroreception (ed. T. H. Bullock and W. Heiligenberg), pp.
527576. New York: John Wiley & Sons.
Hopkins, C. D. (1986b). Temporal structure of non-propagated
electric communication signals. Brain Behav. Evol. 28, 4359.
Hopkins, C. D. (1999). Design features for electric communication.
J. Exp. Biol. 202, 12171228.
Hopkins, C. D. and Bass, A. H. (1981). Temporal coding of species
recognition signals in an electric fish. Science 212, 8587.
Hopkins, C. D., Harned, G. D. and Schmid, U. (1993). Anatomical
projections of the time coding pathway of the mormyrid electric
fish studied using fluorescent-labelled dextrans. Soc. Neurosci.
Abstr. 19, 376.
Kawasaki, M. (1997). Sensory hyperacuity in the jamming avoidance
response of weakly electric fish. Curr. Opin. Neurobiol. 7,
473479.
Knudsen, E. I. (1982). Auditory and visual maps of space in the optic
tectum of the owl. J. Neurosci. 2, 11771194.
Konishi, M. (1991). Deciphering the brains codes. Neural Computat.
3, 118.
Kozloski, J. and Crawford, J. D. (1998). Functional neuroanatomy
of auditory pathways in the sound producing fish Pollimyrus. J.
Comp. Neurol. (in press).
Kramer, B. (1997). Electric organ discharges and their relation to sex
in mormyrid fishes. Naturwissenschaften 84, 119121.
Moller, P. and Szabo, T. (1981). Lesions in the nucleus
mesencephali exterolateralis: Effects on electrocommunication in
the mormyrid fish Gnathonemus petersii (Mormyriformes). J.
Comp. Physiol. A 144, 327333.
Mugnaini, E. and Maler, L. (1987a). Cytology and
immunocytochemistry of the nucleus exterolateralis anterior of the
mormyrid brain: Possible role of GABAergic synapses in temporal
analysis. Anat. Embryol. 176, 313336.
Mugnaini, E. and Maler, L. (1987b). Cytology and
immunocytochemistry of the nucleus of the lateral line lobe in the
electric fish Gnathonemus petersii (Mormyridae): Evidence
suggesting that GABAergic synapses mediate an inhibitory
corollary discharge. Synapse 1, 3256.
Stein, B. E. and Meredith, M. A. (1993). The Merging of the Senses.
Cambridge, MA: MIT Press.
Szabo, T. and Ravaille, M. (1976). Synaptic structure of the
lateral line lobe nucleus in mormyrid fish. Neurosci. Lett. 2,
127131.
Szabo, T., Ravaille, M., Libouban, S. and Enger, P. S. (1983). The
mormyrid rhombencephalon. I. Light and EM investigations on the
structure and connections of the lateral line lobe nucleus with HRP
labelling. Brain Res. 266, 119.
Wulliman, M. F. and Northcutt, R. G. (1990). Visual and
electrosensory circuits of the diencephalon in mormyrids: An
evolutionary perspective. J. Comp. Neurol. 297, 537552.
Young, S. R. and Rubel, E. W. (1983). Frequency-specific
projections of individual neurons in chick brainstem auditory
nuclei. J. Neurosci. 3, 13731378.
Zipser, B. and Bennett, M. V. L. (1976). Interaction of
electrosensory and electromotor signals in lateral line lobe of a
mormyrid fish. J. Neurophysiol. 39, 713721.
208
The Journal of Experimental Biology 207, 1073-1084 1073
Published by The Company of Biologists 2004
doi:10.1242/jeb.00851

Central control of electric signaling behavior in the mormyrid Brienomyrus

brachyistius: segregation of behavior-specific inputs and the role of modifiable
recurrent inhibition
Bruce A. Carlson* and Carl D. Hopkins
Department of Neurobiology and Behavior, Cornell University, Ithaca, NY 14853, USA
*Author for correspondence (e-mail: bc6s@virginia.edu)

Accepted 23 December 2003

Summary
Like all mormyrid fish, Brienomyrus brachyistius of the excitatory neurotransmitter L-glutamate (L-Glu)
produces an electric organ discharge (EOD) with a into DP led to acceleration-like output patterns, while in
constant waveform and variable sequence of pulse PCN it led to scallop-like output patterns. Iontophoresis
intervals (SPI). Periodic bursts fall into two display of the inhibitory neurotransmitter -amino-butyric acid
categories termed scallops and accelerations, with a (GABA) into DP and PCN led to an elongation of
third category termed rasps that appears to combine the intervals, as did iontophoresis of L-Glu into VP.
two. The medullary EOD command nucleus (CN) receives Iontophoresis of the GABAA receptor blocker bicuculline
excitatory input from the midbrain precommand nucleus methiodide (BMI) into DP and PCN induced repetitive
(PCN) and the thalamic dorsal posterior nucleus (DP), bursting behavior and eliminated differences in the effects
both of which are regulated by a recurrent inhibitory of L-Glu iontophoresis in the two nuclei. These results
projection from the ventroposterior nucleus of the torus support our three hypotheses, suggesting that production
semicircularis (VP). We tested the following hypotheses: of different communication behaviors may be regulated
(1) PCN and DP are responsible for generating different by spatially distinct groups of neurons, and recurrent
burst types (scallops and accelerations, respectively), (2) inhibition and disinhibition may play an active role in
differences in the strength of recurrent inhibition are driving and shaping such behaviors.
related to physiological differences between PCN and
DP and (3) recurrent inhibition regulates the resting Key words: mormyrid, electric fish, Brienomyrus brachyistius,
electromotor rhythm, while disinhibition releases PCN electric organ discharge, electromotor, central pattern generator,
and DP, allowing them to generate bursts. Iontophoresis pacemaker, disinhibition, iontophoresis.

Introduction
The electric signaling behavior of mormyrid fish is
characterized by two components: the waveform of each
electric organ discharge (EOD) and the sequence of pulse
intervals (SPI). While the EOD is constant and signals the
senders identity (Carlson et al., 2000; Freedman et al., 1989;
Friedman and Hopkins, 1996; Hopkins, 1981), the SPI is
variable and appears to play a role in signaling motivation or
behavioral state (Carlson, 2002a; Hopkins, 1986; Kramer,
1993). The SPI is characterized by a relatively slow baseline
rhythm, with EOD intervals typically ranging from
approximately 100ms to 300ms (Carlson, 2002a; Teyssedre
and Boudinot, 1987). This baseline rhythm may be periodically
interrupted by a variety of bursts and cessations in the
discharge (Carlson, 2002a; Hopkins, 1986). Such displays
occur in specific behavioral contexts such as courtship and
aggression, suggesting that they play an important role in
social behavior (Bell et al., 1974; Bratton and Kramer, 1989;
Kramer, 1974, 1976; Kramer and Bauer, 1976; Moller et al.,
1989; Scheffel and Kramer, 1997, 2000).
A recent quantitative analysis of bursts in Brienomyrus
brachyistius has revealed three modal display categories based
on variation in the temporal patterning of EOD output (Fig.1B;
Carlson and Hopkins, unpublished observations). Scallops
are stereotyped pulse sequences in which intervals suddenly
drop to 1020ms and then immediately return to baseline
intervals of 100300ms. Accelerations are graded decreases
in interval, typically to values of 2060ms. Accelerations are
less stereotyped than scallops, and minimum intervals for
accelerations may be maintained over several EOD cycles with
a high degree of regularity. Subjectively, rasps appear to
combine an initial scallop-like onset with an acceleration-like
termination (Fig.1B), which is supported by the quantitative
characteristics of the three displays (Carlson and Hopkins,
unpublished observations). Thus, rasps in this species probably
result from a combination of two distinct displays.
Recent anatomical and physiological studies on the
electromotor system of mormyrids have suggested a closed-
loop circuit that may function as a relatively simple central

209
1074 B. A. Carlson and C. D. Hopkins
pattern generator (CPG) for regulating electromotor output Emde et al., 2000), apparently via a projection from the dorsal
(Carlson, 2002b, 2003; von der Emde et al., 2000). Fig.1A subdivision of the ventroposterior nucleus (VPd) in the torus
illustrates the functional connectivity of this network. Each semicircularis (Bell et al., 1983; Carlson, 2002b, 2003; Carlson
EOD is initiated in the medullary command nucleus (CN), and Hopkins, 2001). This recurrent inhibition probably provides
which activates spinal electromotor neurons (EMNs) indirectly a rate-limiting factor to the activity of DP and PCN neurons that
through a projection to the medullary relay nucleus (MRN; Bell may be responsible for producing rhythmic resting electromotor
et al., 1983; Grant et al., 1986). CN integrates excitatory input output (Carlson, 2003; von der Emde et al., 2000). A few large
from two distinct nuclei, the precommand nucleus (PCN) in the neurons at the ventral edge of VP also project to DP, PCN and
mesencephalon, and the dorsal posterior nucleus (DP) in the CN (Bell et al., 1983; Carlson, 2002b), although the functional
thalamus (Bell et al., 1983; Carlson, 2002b, 2003; von der Emde role of these neurons has not yet been explored.
et al., 2000). DP and PCN both receive inhibitory feedback DP and PCN neurons in B. brachyistius show a wide
from the electric organ corollary discharge pathway (von der diversity of firing patterns, and correlations between single unit
activity and burst production suggest that
distinct neuronal populations are responsible for
A C3
generating scallops and accelerations (Carlson,
2003). In distantly related gymnotiform electric
Val fish, the central posterior and prepacemaker
EGp
TM nuclei appear analogous to DP and PCN
EL (Carlson, 2002b), and the two are responsible
for generating distinct electrical behaviors
MCA Tel (Metzner, 1999). Based on these two lines of
ELL evidence, we hypothesized that DP and PCN are
DP
OB likewise responsible for driving different
VPd PCN
electrical behaviors in B. brachyistius.
BCA MRN Preliminary experiments using extracellular
CN electrical stimulation support this hypothesis,
IL suggesting that accelerations are generated
To 1 mm by DP, while scallops are generated by
EMN
PCN (Carlson and Hopkins, unpublished
observations). We tested this hypothesis using
B
Scallop Acceleration Rasp
iontophoresis of the excitatory neurotransmitter
250 250 250 L-glutamate (L-Glu) to stimulate DP and PCN
EOD interval (ms)

200 200 200 neurons and observe the effects on electromotor

150 150 150 output. It is known that L-Glu iontophoresis in
100 100 100
PCN drives decreases in EOD interval in
Gnathonemus petersii (von der Emde et al.,
50 50 50 2000), although these effects have not been
0 0 0 quantified in relation to natural signaling
0 0.5 1 0 0.5 1 0 0.5 1 behavior, and the effects of stimulating DP have
Time (s) not been assessed.
Fig.1. (A) Sagittal schematic showing the functional neuroanatomy of the mormyrid During scallop and acceleration production,
electromotor system, based on Bell et al. (1983), Carlson (2002b, 2003) and von der there is a decrease in the activity of VPd
Emde et al. (2000). Excitatory terminals are identified by flat lines, inhibitory neurons, suggesting that disinhibition may play
terminals by solid circles. Red denotes medullary electromotor nuclei, blue denotes a role in driving these displays by releasing DP
mesencephalic and diencephalic electromotor nuclei (topic of the current study), and and PCN neurons from negative feedback
green denotes corollary discharge nuclei. BCA, bulbar command-associated nucleus; control (Carlson, 2003). Conversely, increases
C3, third cerebellar lobule; CN, command nucleus; DP, dorsal posterior nucleus of in inhibition may be responsible for producing
the thalamus; EGp, eminentia granularis pars posterior; EL, exterolateral nucleus of cessations in the discharge. We tested these
the torus semicircularis; ELL, electrosensory lateral line lobe; EMN, electromotor hypotheses by several means. Preliminary
neurons; IL, inferior lobe of the hypothalamus; MCA, mesencephalic command-
immunohistochemical studies indicate that PCN
associated nucleus; MRN, medullary relay nucleus; OB, olfactory bulb; PCN,
precommand nucleus; Tel, telencephalon; TM, tectum mesencephali; Val, valvula of
is surrounded by terminals containing the
the cerebellum; VPd, dorsal subdivision of the ventroposterior nucleus of the torus inhibitory neurotransmitter -amino-butyric
semicircularis. (B) Examples of the three burst display types produced by freely acid (GABA; Niso et al., 1989). Thus, we used
behaving Brienomyrus brachyistius. Quantitative analysis indicates that they fall into iontophoresis of GABA in DP and PCN to test
distinct categories based on unique temporal patterns of EOD production (Carlson whether this causes increases in EOD interval
and Hopkins, unpublished observations). to verify that DP and PCN receive GABAergic

210
 Stereotyped temporal pattern generation in mormyrids 1075
inhibitory input. Second, we used iontophoresis of L-Glu in VP
to test whether this also causes increases in EOD interval.
Finally, we used iontophoresis of the GABAA receptor blocker
bicuculline methiodide (BMI) in DP and PCN to block
inhibitory input and determine whether eliminating recurrent
inhibition drives decreases in EOD interval.
Differences in the effects of DP and PCN on the SPI are
likely to be caused by differences in their physiology, which
may in turn relate to differences in the strength of recurrent
inhibition from VPd neurons (Carlson, 2003). To test this
hypothesis, we compared the effects of L-Glu iontophoresis
in DP and PCN before and after BMI iontophoresis. If the
observed differences resulting from stimulating the two
nuclei with L-Glu are due to variation in inhibitory feedback,
then blocking this inhibition should eliminate these
differences.
Materials and methods
Animals
We used a total of 27 Brienomyrus brachyistius (Gill 1862),
ranging in size from 8.0g to 51.0g in body mass and 7.9cm
to 18.2cm in total length. Fish were either wild-caught
or laboratory-bred. They were housed in 280-liter group
aquaria at a temperature of 2527C and conductivity of
150200Scm1 on a 12h:12h light:dark cycle and fed live differential AC amplifier (A-M Systems, Inc., Everett, WA,
black worms daily. All procedures were in accordance with the
guidelines established by the National Institutes of Health and
were approved by the Cornell University Institutional Animal
Care and Use Committee.
Surgery
Surgical procedures were identical to those described
previously (Carlson, 2002b, 2003). Animals were anesthetized
in a solution of 500mgl1 tricaine methanesulfonate (MS-222;
Sigma Chemical Co., St Louis, MO, USA) and then respirated
under a solution of 160mgl1 MS-222 during the surgery. Fish
were placed on a horizontal platform with lateral supports and
completely immersed in aquarium water except for the dorsal
surface of the head. A flap of skin was removed from the head
and the underlying tissue was scraped away to expose the dorsal
surface of the skull. Lidocaine (100200l of a 2% solution;
Radix Laboratories, Inc., Eau Claire, WI, USA) was used as a
local anesthetic. A metal post was affixed to the skull using
superglue, and a small rectangular portion of the skull and
meninges was removed to expose the dorsal surface of the
midbrain and caudal forebrain. A reference electrode was then
placed in the dorsal musculature at the posterior end of the skull.
The fish were then immobilized and electrically silenced with
an intramuscular injection of flaxedil (gallamine triethiodide;
100300l of a 3mgml1 solution; Sigma Chemical Co.), and
the respiration was switched to freshwater for recovery.
Experimental procedure
Triple-barrel electrodes were pulled using a Sutter Flaming
Brown Micropipette Puller model P-87 (Sutter Instrument Co.,
Novato, CA, USA) and broken to a composite diameter of
approximately 10m, resulting in individual barrel diameters
of ~23m. Each barrel was filled with one of the following
solutions: (1) 3moll1 NaCl for recording local field
potentials; (2) 2% alcian blue (Sigma Chemical Co.) in
Walpole acetate buffer (pH=4.0) for marking electrode
locations; (3) 0.1moll1 L-Glu (pH=8.0, adjusted with NaOH)
for excitatory iontophoresis; (4) 0.5moll1 GABA (pH=3.5,
adjusted with HCl) for inhibitory iontophoresis or (5)
20mmoll1 BMI in 165mmoll1 NaCl (pH=3.2) for blocking
GABAA receptors.
Electromotor output was monitored by placing a silver
wire against the caudal peduncle with a reference several
centimeters away. Although the electric organ is silenced by
flaxedil, the EOD command can be recorded as a three-spike
potential resulting from the synchronous activation of EMN
(Bennett et al., 1967). The first negative peak in the EMN
volley was defined as the reference time for EOD output (t0),
which in a natural situation precedes the EOD by 45ms. At
the start of each experiment, either DP, PCN or VP was
localized initially through landmarks on the dorsal surface of
the brain and then more precisely by recording characteristic
field potentials that were phase-locked to the EMN volley (see
Carlson, 2002b). Field potentials and EMN output were band-
pass filtered from 10Hz to 5000Hz, amplified 10000 on a
USA; model 1700) and monitored on a digital oscilloscope
(Tektronix, Inc., Beaverton, OR, USA; model 5223).
Iontophoretic currents were provided by a separate amplifier
(A-M Systems, Inc.; Neuroprobe model 1600).
After locating a given nucleus, the horizontal position and
depth of the electrode were adjusted using a microelectrode
drive (Burleigh Instruments, Inc., Fishers, NY, USA;
Inchworm 6000) that was held by a micromanipulator
(Newport Co., Fountain Valley, CA, USA; model 462-XY-M).
The position of the electrode was adjusted in 50m steps
in all three dimensions and the location where pulsed
iontophoresis of L-Glu (0.5A, 500ms pulses at 0.25Hz)
resulted in the strongest modulation in the SPI was used for all
subsequent iontophoretic injection experiments in that nucleus
(Fig.2).
For all experiments, the EMN signal was sent to a Schmitt
Trigger, which was output to an event timer that recorded the
time of t0 using a clock rate of 1MHz (Tucker-Davis
Technologies, Alachua, FL, USA; model ET1). Data on EOD
times of occurrence were saved using custom-made software.
For experiments involving L-Glu or GABA iontophoresis, 20s
of data were recorded before iontophoresis, followed by 20s
of iontophoresis (1.0A for L-Glu, +1.0A for GABA) and
an additional 20s of recording. In most cases, opposite polarity
current (+1.0A for L-Glu, 1.0A for GABA) was tested as
a control, and this resulted in no observable modulation in the
SPI. For experiments involving BMI iontophoresis, 1min of
data were recorded before iontophoresis, followed by 4min of
iontophoresis (+100nA), followed by 1min of recovery. The
longer duration of BMI iontophoresis compared with L-Glu
211
1076 B. A. Carlson and C. D. Hopkins
and GABA iontophoresis was chosen based on results from
previous studies in other systems in which the effects of BMI
iontophoresis occurred after relatively long latencies and
persisted for several minutes after termination (Fujita and
Konishi, 1991; Heiligenberg et al., 1996). The physiological
basis for this difference is unclear but is probably related to
differences in the pharmacological effects of a receptor blocker
compared with naturally occurring neurotransmitters. In some
experiments, the effects of L-Glu iontophoresis in DP and PCN
before and after BMI iontophoresis were determined. The

A PCN tc L tc L
+100

Dorsal
PGc
PGc +50

0 m
PGm PGm

Ventral
50
nLR IL nLR IL
Hyp Hyp 100

B pc pc 2s

v
v FR +100
FR

Dorsal
DP
+50

0 m

Ventral 50
CP
CP 100

2s

C ll
L L ll +100
MV
Dorsal

MV
VP +50

tt tt 0 m
Ventral

50
nLR
nLR 100
IL IL
2s
L-Glu

Fig.2. Localization of iontophoresis sites. (A) The precommand nucleus (PCN). (B) The dorsal posterior nucleus (DP). (C) The ventroposterior
nucleus (VP). The first column shows photomicrographs of transverse sections with retrogradely labeled neurons against a background of
cresyl violet counterstain, taken from the anatomical study by Carlson (2002b). In A and B, the label results from an injection of neurobiotin
into the command nucleus (CN), while in C it results from an injection of neurobiotin into PCN. The second column shows photomicrographs
of transverse sections with alcian blue staining against a background of neutral red counterstain, at the same approximate rostro-caudal
locations as the first column. The small blue dots provide a precise marker of electrode location, which was always accurately placed into one
of the three nuclei. In the third column, continuous voltage traces of electromotor neuron (EMN) activity are shown, with each spike
corresponding to a single EMN volley. Iontophoretic injections of L-glutamate (L-Glu; 500ms pulses of 500nA) occurred during the times
represented by horizontal green lines below each series of traces. Each example is taken from the site shown in the second column, with 0m
corresponding to the exact location of the alcian blue marker. The effects of L-Glu iontophoresis at 50m and 100m dorsal and ventral to
these sites are also shown. CP, central posterior nucleus of the thalamus; FR, fasciculus retroflexus; Hyp, hypothalamus; IL, inferior lobe of the
hypothalamus; L, lateral nucleus of the torus semicircularis; ll, lateral lemniscus; MV, medioventral nucleus of the torus semicircularis; nLR,
nucleus of the lateral recess; pc, posterior commissure; PGc, caudal subdivision of the preglomerular nucleus; PGm, medial subdivision of the
preglomerular nucleus; tc, tectocerebellar tract; tt, toro-praeeminential tract; v, ventricle. Scale bars, 200m in A, 50m in B and 200m in C.

212
 Stereotyped temporal pattern generation in mormyrids 1077
procedure for L-Glu iontophoresis in these cases was identical
to the normal L-Glu iontophoresis procedure, and L-Glu
iontophoresis was always performed within 2min after
terminating BMI iontophoresis. Statistica 6.1 (StatSoft, Inc.,
Tulsa, OK, USA) was used for all statistical analyses of data
on the SPI.
Histology
At the end of each experiment, we marked the location of
the electrode by iontophoretic injection of alcian blue, using a
500ms, 150V pulse (Grass Medical Instruments, Quincy,
MA, USA; model S88 stimulator). After completing the
experiments, fish were placed back under general anesthesia
(160mgl1 MS-222) and then perfused transcardially with
Hickmans ringer solution (6.48gl1 NaCl, 0.15gl1 KCl,
0.29gl1 CaCl2, 0.12gl1 MgSO4, 0.084gl1 NaHCO3,
0.06gl1 NaH2PO4) followed by ice-cold 4%
paraformaldehyde/1% glutaraldehyde in 0.1moll1 phosphate
buffer (PB; pH=7.2) for fixation. The brains were removed and
postfixed overnight and then transferred to 0.1moll1 PB for
storage. Brains were transferred to a solution of 30% sucrose
in 0.1moll1 PB on the night prior to sectioning. Transverse
sections were cut on a freezing microtome at 50m, mounted
on chrom-alum-subbed slides, counterstained with neutral red,
dehydrated in a graded alcohol series and coverslipped with
Permount (Sigma Chemical Co.).
Results
Glutamate iontophoresis
Brief pulses of L-Glu in DP (N=25 fish) and PCN (N=24 fish)
led to a shortening of EMN intervals, while in VP (N=20 fish)
it typically led to complete cessations (Fig.2). Alcian blue
marking of the locations where iontophoresis caused the
greatest response led to very restricted, bright blue dots in the
center of DP or PCN or in the region around VP (Fig.2),
verifying that the electrodes were located in the desired
A 500
DP
250
Median EMN interval (ms)
0
EMN interval (ms)
500
PCN
250
0
20 000
VP
10 000
0 (B) Doseresponse curves of the
0 50 100 150 200
Time (s)
regions. In general, L-Glu iontophoresis only led to changes in
EMN interval within a range of depths of 50150m from the
alcian blue mark (Fig.2).
In three different fish, doseresponse curves were
constructed by measuring the effects of L-Glu iontophoresis
on median EMN intervals with varying levels of current
magnitude (from 100nA to 900nA in steps of 200nA) in
all three nuclei (Fig.3A). In both DP and PCN, increasing
levels of current led to greater shortening of EMN intervals,
with the response beginning to saturate at approximately
500nA and showing complete saturation at 700nA to
900nA (Fig.3B). Similarly, in VP, increasing levels of
current led to a greater elongation of EMN intervals, with the
response saturating at 700nA to 900nA (Fig.3B). Thus, for
all experiments using L-Glu iontophoresis, we used current
magnitudes of 1.0A, which was well above the level of
saturation for all three nuclei and therefore provided maximal
stimulation of each nucleus.
20s injections of L-Glu into the three nuclei led to
characteristic modulations in the SPI (Fig.4). In both DP and
PCN, there was a marked, maintained decrease in EMN
interval that persisted throughout the duration of the stimulus,
while in VP there was a complete cessation of activity for the
whole period of stimulation and usually for many additional
seconds after terminating the current. There was a highly
significant decrease in EMN intervals during L-Glu
iontophoresis in DP and PCN and a highly significant increase
in EMN intervals during L-Glu iontophoresis in VP (Table1).
Although stimulation of both DP and PCN led to a
shortening of EMN intervals, the responses of the two
nuclei were typically quite different. Stimulation of DP
typically resulted in a smooth decrease in interval to values
of ~2050 ms that were maintained throughout the period of
stimulation with a high degree of regularity (Figs4,5).
Stimulation of PCN also led to a decrease in baseline
intervals, but this baseline was typically not as low or regular
as during DP stimulation (Table1) and was punctuated by the
B 500 Fig.3. Effects of varying L-
DP (N=3) glutamate (L-Glu) iontophoretic
400 PCN (N=3) current magnitudes on electromotor
300 neuron (EMN) intervals in the dorsal
posterior (DP), precommand (PCN)
200 and ventroposterior nuclei (VP).
100 (A) One example from each nucleus
0 in a single fish. EMN intervals are
20 000 plotted against time. The timing of
VP (N=3) L-Glu iontophoresis is indicated by
15 000 the horizontal bars beneath the
plots, with the current magnitude
10 000
increasing from 100nA to
5000 900nA in steps of 200nA.
0 effects of varying current magnitude
0 200 600 1000 on median EMN intervals. Values
Current (nA) shown are means S.E.M.
213
1078 B. A. Carlson and C. D. Hopkins
DP PCN VP
500
400
300
200
100
0
500
EMN interval (ms)

400
300
200
100
0
500
400
300
200
100
0
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60

Time (s)
Fig.4. Three representative examples of the effects of L-glutamate (L-Glu) iontophoresis on electromotor neuron (EMN) output in the dorsal
posterior (DP), precommand (PCN) and ventroposterior nuclei (VP). Each row corresponds to a single fish. The first column shows the effects
of L-Glu iontophoresis in DP, the second column shows the effects of L-Glu iontophoresis in PCN, and the third column shows the effects of L-
Glu iontophoresis in VP. In each case, iontophoretic currents consisted of 1A for 20s, which is indicated by the horizontal bar beneath each
column.

DP PCN repeated production of transient,

100 intense bursts reaching minimum
80 intervals of 1025 ms (Figs4,5). In
60 some cases, these transient bursts
40 appeared identical to scallops, while
20 in most cases, they simply appeared
0 as non-stereotyped scallop-like
100 bursts (Fig.5). Stimulation in PCN
EMN interval (ms)

80 led to a significantly greater

60 coefficient of variation (CV) in EMN
40 interval (Wilcoxon matched pairs
20 test: z22=3.912, P<0.0001), smaller
0
minimum EMN interval (z22=2.354,
100 P<0.02) and greater maximum EMN
80 interval (z22=3.360, P<0.001)
compared with stimulation in DP
60
(Fig.6), as expected from the burst-
40
like responses to PCN stimulation.
20
0 GABA iontophoresis
31 32 33 34 31 32 33 34
Iontophoresis of GABA in DP
Time (s)
(N=15 fish) and PCN (N=17 fish)
Fig.5. Differences in electromotor neuron (EMN) output patterns in response to L-glutamate typically led to an elongation of EMN
(L-Glu) iontophoresis in the dorsal posterior (DP) and precommand nuclei (PCN). Each row intervals that was maintained
corresponds to a single fish. The three representative examples for each nucleus show an throughout the duration of the
expanded view of the period from 31s to 34s of the data shown in Fig.4. stimulus (Fig.7), resulting in a

214
 Stereotyped temporal pattern generation in mormyrids 1079
Table 1. Changes in median EMN interval in response to iontophoresis of L-glutamate (L-Glu) for 20s using a current magnitude
of 1.0A in the dorsal posterior (DP), precommand (PCN) and ventroposterior (VP) nuclei
Median EMN interval (mean S.E.M.)
Nucleus (N) Pre-stimulation L-Glu stimulation Post-stimulation F P
DP (25) 381.1648.868 37.6602.6131 420.6762.789 29.617 <0.000001
PCN (24) 296.7032.242 67.2295.9719 339.8148.942 31.032 <0.000001
VP (20) 298.5636.302 213713744.8 275.4131.000 31.678 <0.000001

F-statistics and P-values from a repeated measures analysis of variance (ANOVA) for each nucleus are shown. N is number of fish.

significant increase in EMN intervals (Table2). By contrast, 1

A

CV in EMN interval (ms)

we observed no response to GABA iontophoresis in VP (N=14
fish; Fig.7) and there was no significant change in EMN 0.8
intervals (Table2). 0.6
BMI iontophoresis 0.4
Iontophoresis of BMI into DP (N=7 fish) and PCN (N=7
fish) resulted in repetitive bursting behavior that started after 0.2 *
a relatively long latency following stimulus onset (30150s)
0
and persisted for several minutes after stimulus termination 35
(Fig.8A). Comparing the last minute of BMI iontophoresis B

Min. EMN interval (ms)

with the 1min control period prior to BMI iontophoresis, there 28
was a significant shortening of median EMN interval in both *
DP (Wilcoxon matched pairs test; z7=2.3664; P<0.02) and 21
PCN (z7=2.2678; P<0.025). The bursts resulting from BMI
14
iontophoresis were qualitatively similar to those produced by
freely behaving animals, and included scallops, accelerations 7
and rasps (Fig.8B). Thus, rather than simply quantifying
overall activity with general descriptors, it was possible to 0
count the number of bursts produced. There were no significant 600
C
Max. EMN interval (ms)

differences in the numbers of scallops (MannWhitney U test:

z7,7=1.086; P>0.27), accelerations (z7,7=0.192; P>0.84) or
rasps (z7,7=0.639; P>0.52) produced by BMI iontophoresis in
DP compared with PCN (Fig.8C).
Comparing the effects of L-Glu iontophoresis before and
after BMI iontophoresis in both DP and PCN demonstrated
a significant increase in the CV in EMN interval
(F1,11=7.848; P<0.02), a significant decrease in the minimum
EMN interval (F1,11=34.95; P<0.02) and no significant
change in the maximum EMN interval (F1,11=3.809; P>0.07).
Before BMI iontophoresis, the CV was significantly greater
in PCN than in DP (Fig.9A; MannWhitney U test:
z7,6=2.571; P<0.02), although there was no significant
difference after BMI (z7,6=0.571; P>0.56). Similarly, the
minimum EMN interval was significantly smaller in PCN
than in DP before BMI iontophoresis (Fig.9B; z7,6=2.000;
P<0.05), although there was no significant difference
after BMI (z7,6=0.428; P>0.66), and the maximum EMN
interval was significantly greater in PCN than in DP before
BMI iontophoresis (Fig.9C; z7,6=2.000; P<0.05), though
there was no significant difference after (z7,6=0.428;
P>0.66). Thus, following BMI iontophoresis, the effects of
L-Glu iontophoresis in DP and PCN were statistically
identical.
400
200
*
0
DP (N=22) PCN (N=22)
Fig.6. The effects of L-glutamate (L-Glu) iontophoresis in the dorsal
posterior (DP) and precommand nuclei (PCN) on the coefficient of
variation in electromotor neuron (EMN) interval (A), the minimum
EMN interval (B) and the maximum EMN interval (C). Only those
fish with data from both nuclei are included. Values shown are
means S.E.M. Asterisks represent statistically significant differences
(Wilcoxon matched pairs test; P<0.05).
Discussion
We have shown that stimulation of DP and PCN neurons
using the excitatory neurotransmitter L-Glu leads to a
shortening of EOD intervals. This indicates that both nuclei
provide excitatory input to CN and are responsible for the
production of bursts, supporting previous studies that used L-

215
1080 B. A. Carlson and C. D. Hopkins
DP PCN VP
1000
800
600
400
200
0
1000
EMN interval (ms)

800
600
400
200
0
1000
800
600
400
200
0
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60

Time (s)
Fig.7. Three representative examples of the effects of -amino-butyric acid (GABA) iontophoresis on electromotor neuron (EMN) output in the
dorsal posterior (DP), precommand (PCN) and ventroposterior nuclei (VP). Each row corresponds to a single fish. In each case, iontophoretic
currents consisted of +1A for 20s, which is indicated by the horizontal bar beneath each column.

Glu stimulation in PCN (von der Emde et al., 2000) and
recordings of single unit activity in DP and PCN (Carlson,
2003; von der Emde et al., 2000). Furthermore, electromotor
output resulting from stimulating the two nuclei was
significantly different. DP stimulation resulted in smooth,
maintained accelerations to intervals of 2050ms. PCN
stimulation led to a more modest decrease in baseline intervals
that was interrupted by the repeated production of transient,
scallop-like bursts. This supports the hypothesis that these two
nuclei are responsible for generating distinct communication
displays. Recent studies of single unit activity in DP and PCN
have shown that units with low baseline firing rates experience
an increase in activity during accelerations, while units with
high baseline firing rates experience an increase in activity
during scallops (Carlson, 2003). In light of the findings of the
current study, this suggests that neurons within DP may
correspond to the former, while neurons within PCN may
correspond to the latter.
Neurons in PCN appear to be surrounded by GABAergic
inhibitory terminals (Niso et al., 1989), and single units in both
DP and PCN receive recurrent inhibitory feedback via the
corollary discharge pathway (Carlson, 2003; von der Emde et
al., 2000). Anatomically, DP and PCN receive a dense
projection from VPd, which in turn receives input from the
corollary discharge pathway (Carlson, 2002b). These three
lines of evidence suggest that VPd provides recurrent,
GABAergic inhibition to DP and PCN. The results of the
current study support this hypothesis: iontophoresis of GABA
into DP and PCN induces a significant elongation of
EOD intervals, as does stimulation of VPd using L-Glu
iontophoresis. In most cases, stimulation of VPd led to
complete cessations in the discharge, suggesting that cessations
may result from increases in the activity of VPd neurons and
therefore stronger recurrent inhibition.
We hypothesized that this recurrent inhibition is responsible
for regulating DP and PCN activity, and thereby maintaining

Table 2. Changes in median EMN interval in response to iontophoresis of -amino-butyric acid (GABA) for 20s using a current
magnitude of +1.0A in the dorsal posterior (DP), precommand (PCN) and ventroposterior (VP) nuclei
Median EMN interval (mean S.E.M.)
Nucleus (N) Pre-stimulation GABA stimulation Post-stimulation F P
DP (15) 260.2439.315 318.4242.941 223.0236.025 15.566 <0.00003
PCN (17) 284.6744.115 448.0471.982 255.1240.670 17.185 <0.00001
VP (14) 237.8235.102 243.6030.572 246.9436.056 0.8172 >0.45

F-statistics and P-values from a repeated measures analysis of variance (ANOVA) for each nucleus are shown. N is number of fish.

216
 Stereotyped temporal pattern generation in mormyrids 1081

A L-Glu iontophoresis BMI iontophoresis L-Glu iontophoresis

500
400 PCN
300
EMN interval (ms)

200
100
0
500
DP
400
300
200
100
0
0 10 20 30 40 50 60 0 50 100 150 200 250 300 350 0 10 20 30 40 50 60
Time (s)

B 250 250 250 C

80
DP (N=7)
PCN PCN (N=7)
60

Number of bursts
Scallop Scallop Scallop
0 0 0
0 3 0 3 0 3
250 250 250 40

20
DP

Accel. Accel. Rasp

0 0 0 0
0 3 0 3 0 3 Scallops Accelerations Rasps

Fig.8. (A) One example each of the effects of bicuculline methiodide (BMI) iontophoresis in the dorsal posterior (DP) and precommand nuclei
(PCN), and the effects of L-glutamate (L-Glu) iontophoresis immediately before and immediately after BMI iontophoresis. L-Glu iontophoretic
currents consisted of 1A for 20s, while BMI iontophoretic currents consisted of +100nA for 4min, each of which is shown as a horizontal
bar beneath each column. (B) Excerpts from A of examples of different burst types occurring during BMI iontophoresis in DP and PCN.
(C) Number of different burst types induced by BMI iontophoresis in DP and PCN. Values shown are means S.E.M. There were no significant
differences in the numbers of any of the three burst types.

the irregular, baseline rhythm of 100300ms EOD intervals.
During burst displays, the activity of VPd neurons decreases
(Carlson, 2003), so we hypothesized that disinhibition plays a
role in disrupting the baseline rhythm by releasing DP and
PCN from their normal rate-limiting factor, thereby allowing
them to drive burst displays. In support of this hypothesis,
application of the GABAA receptor blocker BMI to DP and
PCN caused the SPI to go from a resting rhythm to repetitive
bursting. It is unclear how the activity of VPd neurons is
modulated in a natural situation, although the tectum
mesencephali has a strong projection to VPd (Carlson, 2002b;
Wullimann and Northcutt, 1990) and retrograde labeling
suggests that VPd may also receive inputs from hypothalamic
and preoptic areas (Carlson, 2002b). Application of GABA to
VP did not elicit any changes in EMN activity, suggesting that
a reduction in VPd activity is mediated by a different
neurotransmitter or through neuromodulatory inputs.
One potential source of physiological differences between
DP and PCN may be variation in recurrent inhibition from VPd
neurons, which produce a stereotyped burst of action potentials
starting within a few milliseconds of EOD production
(Carlson, 2003; von der Emde et al., 2000). Across neurons,
there is wide variation in the duration of these bursts. If
different subsets of VPd neurons project to DP and PCN, this
variation could lead to differences in the baseline activity of
DP and PCN neurons and therefore different effects on
electromotor output when stimulated. In support of this
hypothesis, the effects of L-Glu iontophoresis in the two nuclei
were identical following BMI iontophoresis. Furthermore, if
recurrent inhibition to DP and PCN is separated into different
pathways, this may provide a means for differentially
activating the two nuclei and generating distinct behaviors.
Although we suggest that disinhibition plays a role in
driving burst displays in a natural situation, we were able to
elicit such displays solely through the application of L-Glu to
DP and PCN, which presumably did not affect recurrent
inhibition. Most likely, the increased excitatory input caused
by L-Glu application counterbalanced the ongoing recurrent

217
1082 B. A. Carlson and C. D. Hopkins
2 (Bullock et al., 1983). EOD production in both groups of fish
interval (ms) A DP (N=7) is controlled by a ventral midline nucleus in the medulla (CN
CV in EMN

1.5 PCN (N=7) in mormyrids and the pacemaker nucleus, or PN, in

1 gymnotiforms), which projects to larger, adjacent relay
0.5 neurons whose axons descend the spinal cord to innervate
* electromotor neurons (Dye and Meyer, 1986). A recent
0
anatomical study has suggested that DP and PCN are
35 analogous to the central posterior (CP) and prepacemaker
Minimum EMN

B
interval (ms)

28 nuclei (PPN) that provide input to PN in gymnotiforms

21 * (Carlson, 2002b). In both groups of fish, there is a rostral group
14 of cells located within a dorsal thalamic nucleus (DP and CP)
7 and a caudal group of cells that forms a ventrolateral extension
0 of the dorsal thalamus (PCN and PPN). While the caudal
groups of cells are relatively large with thick, extrinsic
1200 dendrites, the rostral groups of cells are small with thin,
Maximum EMN

C
interval (ms)

800
400
* chirps (transient, intense bursts; Metzner, 1999). This is
7 strikingly similar to electromotor output patterns induced by
Pre-BMI Post-BMI
Fig.9. The effects of L-glutamate (L-Glu) iontophoresis in the dorsal
posterior (DP) and precommand nuclei (PCN) before and after
bicuculline methiodide (BMI) iontophoresis on the coefficient of
variation (CV) in electromotor neuron (EMN) interval (A), the
minimum EMN interval (B), and the maximum EMN interval (C).
Values shown are means S.E.M. Asterisks represent statistically
significant differences (MannWhitney U test; P<0.05).
inhibition, resulting in increases in EMN activity. The different
effects of L-Glu stimulation in DP and PCN probably relate to
the different strengths of recurrent inhibition that counteract
the stimulatory effects of L-Glu to varying degrees. Application
of BMI to these nuclei, by contrast, leads to a maintained
blockage of recurrent inhibition but no excitatory input. Not
surprisingly, this also leads to increases in EMN activity.
Unlike stimulation with L-Glu, however, there were no
observable differences in the output patterns caused by BMI
stimulation in DP and PCN, which can be explained by the fact
that BMI application also removed the source of physiological
differences between the two nuclei, which was not the case
with L-Glu stimulation. In a natural situation, disinhibition is
typified by a modest, temporary reduction in inhibitory input
from the baseline level rather than a complete, maintained
removal (Carlson, 2003). Unlike BMI application, this would
not eliminate differences between DP and PCN but would
provide a transient excitatory effect similar to the effects of
L-Glu stimulation, causing the two nuclei to drive different
behaviors.
Convergence in the central control of electromotor behavior
Gymnotiform electric fish from South America have
electromotor and electrosensory systems that share many
striking similarities with those of mormyrids (Hopkins,
1995), despite overwhelming evidence that their electrogenic
and electrosensory capabilities have evolved independently
intrinsic dendrites.
Stimulating CP in gymnotiforms leads to rises (smooth,
graded increases in frequency), while stimulating PPN leads to
stimulation in DP and PCN, respectively, with the former
driving accelerations (smooth, graded increases in frequency)
and the latter driving scallops (transient, intense bursts). Thus,
convergence in anatomical substrates appears to be directly
linked to convergence in the behaviors they control. In
gymnotiforms, the different electrical behaviors resulting from
activation of CP and PPN are related to differences in the
location of synapses in PN and differences in the glutamate
receptor subtypes found at these synapses (Metzner, 1999). It
is possible that similar differences play a role in the differential
effects of DP and PCN on electromotor behavior in mormyrids,
but the findings of the current study suggest that these
differences may be due solely to the effects of variation in
recurrent inhibition.
Diversity in mormyrid electric signaling behavior
Every species of mormyrid that has been studied produces
acceleration-like displays that appear to play a role in
aggression, but there is wide diversity across species in the
other types of displays that may be produced (Carlson, 2002a).
In Gnathonemus petersii, agonistic encounters are often
accompanied by repetitive pulse pairs, with EOD intervals
alternating between 1516ms and 89ms (Bauer, 1972; Bell
et al., 1974). Such displays have never been observed in B.
brachyistius or any other species (Carlson, 2002a), although
the number of species studied is relatively small. By contrast,
scallops have never been described for G. petersii, although
they have been described for B. niger (Serrier and Moller,
1989). This suggests the hypothesis that pulse pairs may be
driven by PCN in species that do not produce scallops.
There is also wide diversity in the characteristics of certain
displays between species. For example, scallops in B. niger and
B. brachyistius are quite similar in their basic structure but
differ in minimum interval (Carlson, 2002a; Carlson and
Hopkins, unpublished observations; Serrier and Moller, 1989).
Such differences are possibly related to differences in the

218
 Stereotyped temporal pattern generation in mormyrids 1083
morphology and physiology of PCN neurons. Field recordings
from various Brienomyrus species in Gabon reveal the
production of rasps that differ dramatically from those in B.
brachyistius and do not appear to result from combining a
scallop and an acceleration (Hopkins, 1983; Hopkins and Bass,
1981). Scallops have not been described for these species, so
it is possible that PCN controls rasp production in this group.
The rich diversity of mormyrids and the signals they produce
provide a rare opportunity for studying the evolution of neural
circuits that govern communication behavior.
Motor networks and behavior
Much of our understanding of the mechanisms underlying
stereotyped motor output comes from relatively simple
networks involved in rhythmic behaviors such as locomotion,
digestion, respiration and heartbeat (Marder and Bucher,
2001). The stereotyped, rhythmic output of these central
pattern generators (CPGs) results from a combination of
several cellular and molecular specializations, many of which
are shared across different networks. The findings of the
current study, as well as other recent studies (Carlson, 2002b,
2003; von der Emde et al., 2000), reveal several similarities
between these networks and the mormyrid electromotor
system. For instance, recurrent inhibition plays an important
role in establishing rhythmic motor output in many different
networks (Friesen and Stent, 1978) as well as in regulating
electromotor output in mormyrids. Similarly, disinhibition
serves a permissive function in activating stereotyped motor
output in several motor systems (Faumont et al., 1998; Noga
et al., 1988; Wang and Bieger, 1991), which also seems to be
the case for generating burst displays in mormyrids.
Extracellular stimulation of restricted brain regions can
elicit the production of semi-natural, species-specific
communication signals in a variety of vertebrate species
(Apfelbach, 1972; Demski and Gerald, 1972; Fine and Perini,
1994; Fu and Brudzynski, 1994; Goodson and Bass, 2000;
Jrgens and Richter, 1986; Phillips and Youngren, 1973;
Schmidt, 1966; Schuller and Radtkeschuller, 1990; Seller and
Armitage, 1983; Valentine et al., 2002; Williams and Vicario,
1993). While it is clear that stereotyped signal production may
be controlled by spatially distinct groups of neurons, there is
often insufficient information about anatomical circuitry and
how it relates to patterns of neuronal activity to gain insight
into the network and cellular mechanisms involved in signal
generation. Part of the reason for this is that many vertebrate
communication signals are relatively complex, involving
several features that vary semi-independently over time. As a
result, the neural substrates underlying the generation of these
signals are similarly complex, making it difficult to formulate
hypotheses that directly link the activity patterns of individual
neurons to specific signal characteristics.
By contrast, electric signaling behavior in mormyrids is
relatively simple, consisting of two distinct components: a
stereotyped EOD waveform and a variable pattern of EOD
production (the SPI). The characteristics of the former are
controlled by the morphophysiological characteristics of the
electric organ (Bennett, 1971), while the latter is determined
by patterns of activity in CN (Grant et al., 1986). Thus,
unraveling the mechanisms involved in regulating the SPI
breaks down to a problem of understanding the generation of
spike times in CN. As this and other recent studies have shown
(Carlson, 2002b, 2003; von der Emde et al., 2000), this relative
simplicity makes the mormyrid electromotor network an
excellent model system for studying the mechanisms of
generating stereotyped temporal patterns in vertebrate
communication. The many similarities between this system
and the gymnotiform electromotor network, as well as with
CPGs in general, suggest that insights gained into the
functioning of this network are likely to be instructive towards
general issues in the motor control of behavior.
This research was supported in part by grants from the
National Institute of Mental Health (Grant MH37972) and
the National Science Foundation (0108372). B.A.C. was
supported by a National Science Foundation Predoctoral
Fellowship and a National Institute of Mental Health
Predoctoral Training Grant (MH15793). Thanks to M.
Kawasaki for advice on the alcian blue staining procedure.
Thanks to A. H. Bass and two anonymous reviewers for their
helpful comments on earlier versions of the manuscript.
References
Apfelbach, R. (1972). Electrically elicited vocalizations in the gibbon
Hylobates lar (Hylobatidae), and their behavioral significance. Zeit.
Tierpsychol. 30, 420-430.
Bauer, R. (1972). High electrical discharge frequency during aggressive
behavior in a mormyrid fish (Gnathonemus petersii). Experientia 28, 669-
670.
Bell, C. C., Libouban, S. and Szabo, T. (1983). Pathways of the electric organ
discharge command and its corollary discharges in mormyrid fish. J. Comp.
Neurol. 216, 327-338.
Bell, C. C., Myers, J. P. and Russell, C. J. (1974). Electric organ discharge
patterns during dominance-related behavioral displays in Gnathonemus
petersii (Mormyridae). J. Comp. Physiol. 92, 201-228.
Bennett, M. V. L. (1971). Electric organs. In Fish Physiology (ed. W. S. Hoar
and D. J. Randall), pp. 347-491. London: Academic Press.
Bennett, M. V. L., Pappas, G., Aljure, E. and Nakajima, Y. (1967).
Physiology and ultrastructure of electrotonic junctions. II. Spinal and
medullary electromotor nuclei in mormyrid fish. J. Neurophysiol. 30, 180-
208.
Bratton, B. O. and Kramer, B. (1989). Patterns of the electric organ
discharge during courtship and spawning in the mormyrid fish, Pollimyrus
isidori. Behav. Ecol. Sociobiol. 24, 349-368.
Bullock, T. H., Bodznick, D. A. and Northcutt, R. G. (1983). The
phylogenetic distribution of electroreception: evidence for convergent
evolution of a primitive vertebrate sense modality. Brain Res. Rev. 6, 25-46.
Carlson, B. A. (2002a). Electric signaling behavior and the mechanisms of
electric organ discharge production in mormyrid fish. J. Physiol. Paris 96,
405-419.
Carlson, B. A. (2002b). Neuroanatomy of the mormyrid electromotor control
system. J. Comp. Neurol. 454, 440-455.
Carlson, B. A. (2003). Single-unit activity patterns in nuclei that control the
electromotor command nucleus during spontaneous electric signal
production in the mormyrid Brienomyrus brachyistius. J. Neurosci. 23,
10128-10136.
Carlson, B. A., Hopkins, C. D. and Thomas, P. (2000). Androgen correlates
of socially induced changes in the electric organ discharge waveform of a
mormyrid fish. Horm. Behav. 38, 177-186.
Demski, L. and Gerald, J. (1972). Sound production evoked by electrical
stimulation of the brain in toadfish (Opsanus beta). Anim. Behav. 20, 507-
513.
219
1084 B. A. Carlson and C. D. Hopkins
Dye, J. C. and Meyer, J. H. (1986). Central control of the electric organ
discharge in weakly electric fish. In Electroreception (ed. T. H. Bullock and
W. Heiligenberg), pp. 71-102. New York: John Wiley & Sons.
Faumont, S., Simmers, J. and Meyrand, P. (1998). Activation of a lobster
motor rhythm-generating network by disinhibition of permissive
modulatory inputs. J. Neurophysiol. 80, 2776-2780.
Fine, M. L. and Perini, M. A. (1994). Sound production evoked by electrical
stimulation of the forebrain in the oyster toadfish. J. Comp. Physiol. A 174,
173-185.
Freedman, E. G., Olyarchuk, J., Marchaterre, M. A. and Bass, A. H. (1989).
A temporal analysis of testosterone-induced changes in electric organs and
electric organ discharges of mormyrid fishes. J. Neurobiol. 20, 619-634.
Friedman, M. A. and Hopkins, C. D. (1996). Tracking individual mormyrid
electric fish in the field using electric organ discharge waveforms. Anim.
Behav. 51, 391-407.
Friesen, W. and Stent, G. (1978). Neural circuits for generating rhythmic
movements. Annu. Rev. Biophys. Bioeng. 7, 37-61.
Fu, X. and Brudzynski, S. (1994). High-frequency ultrasonic vocalization
induced by intracerebral glutamate in rats. Pharmacol. Biochem. Behav. 49,
835-841.
Fujita, I. and Konishi, M. (1991). The role of GABAergic inhibition in
processing of interaural time difference in the owls auditory system. J.
Neurosci. 11, 722-739.
Goodson, J. L. and Bass, A. H. (2000). Rhythmic midbrain-evoked
vocalization is inhibited by vasoactive intestinal polypeptide in the teleost
Porichthys notatus. Brain Res. 865, 107-111.
Grant, K., Bell, C. C., Clausse, S. and Ravaille, M. (1986). Morphology and
physiology of the brainstem nuclei controlling the electric organ discharge
in mormyrid fish. J. Comp. Neurol. 245, 514-530.
Heiligenberg, W., Metzner, W., Wong, C. J. and Keller, C. H. (1996).
Motor control of the jamming avoidance response of Apteronotus
leptorhynchus: evolutionary changes of a behavior and its neuronal
substrate. J. Comp. Physiol. A 179, 653-674.
Hopkins, C. D. (1981). On the diversity of electric signals in a community of
mormyrid electric fish in west Africa. Am. Zool. 21, 211-222.
Hopkins, C. D. (1983). Neuroethology of species recognition in
electroreception. In Advances in Vertebrate Neuroethology (ed. J. P. Ewert,
R. R. Capranica and D. J. Ingle), pp. 871-881. New York: Plenum.
Hopkins, C. D. (1986). Behavior of Mormyridae. In Electroreception (ed. T.
H. Bullock and W. Heiligenberg), pp. 527-576. New York: John Wiley &
Sons.
Hopkins, C. D. (1995). Convergent designs for electrogenesis and
electroreception. Curr. Opin. Neurobiol. 5, 769-777.
Hopkins, C. D. and Bass, A. H. (1981). Temporal coding of species
recognition signals in an electric fish. Science 212, 85-87.
Jrgens, U. and Richter, K. (1986). Glutamate-induced vocalization in the
squirrel monkey. Brain Res. 373, 349-358.
Kramer, B. (1974). Electric organ discharge interaction during interspecific
agonistic behavior in freely swimming mormyrid fish: a method to evaluate
two or more simultaneous time series of events with a digital analyzer. J.
Comp. Physiol. A 93, 203-235.
Kramer, B. (1976). Electric signaling during aggressive behavior in
Mormyrus rume (Mormyridae Teleostei). Naturwissenschaften 63, 48-49.
Kramer, B. (1993). Electrocommunication in weakly electric fish: review of
signals sent and received. J. Comp. Physiol. A 173, 719-722.
Kramer, B. and Bauer, R. (1976). Agonistic behavior and electric signaling
in a mormyrid fish, Gnathonemus petersii. Behav. Ecol. Sociobiol. 1, 45-
61.
Marder, E. and Bucher, D. (2001). Central pattern generators and the control
of rhythmic movements. Curr. Biol. 11, R986-R996.
Metzner, W. (1999). Neural circuitry for communication and jamming
avoidance in gymnotiform electric fish. J. Exp. Biol. 202, 1365-1375.
Moller, P., Serrier, J. and Bowling, D. (1989). Electric organ dishcharge
displays during social encounter in the weakly electric fish Brienomyrus
niger L. (Mormyridae). Ethology 82, 177-191.
Niso, R., Serrier, J. and Grant, K. (1989). Mesencephalic control of the
bulbar electromotor network in the mormyrid Gnathonemus petersii. Eur.
J. Neurosci. Suppl. 2, 176.
Noga, B., Kettler, J. and Jordan, L. (1988). Locomotion produced in
mesencephalic cats by injections of putative transmitter substances and
antagonists into the medial reticular formation and the pontomedullary
locomotor strip. J. Neurosci. 8, 2074-2086.
Phillips, R. and Youngren, O. (1973). Electrical stimulation of brain as a tool
for study of animal communication behavior evoked in mallard ducks
(Anas platyrhynchos). Brain Behav. Evol. 8, 253-286.
Scheffel, A. and Kramer, B. (1997). Electrocommunication and social
behaviour in Marcusenius senegalensis (Mormyridae, Teleostei). Ethology
103, 404-420.
Scheffel, A. and Kramer, B. (2000). Electric signals in the social behavior
of sympatric elephantfish (Mormyridae, Teleostei) from the upper Zambezi
river. Naturwissenschaften 87, 142-147.
Schmidt, R. (1966). Central mechanisms of frog calling. Behaviour 26, 251-
285.
Schuller, G. and Radtkeschuller, S. (1990). Neural control of vocalization
in bats mapping of brainstem areas with electrical microstimulation
eliciting species-specific echolocation calls in the Rufous horseshoe bat.
Exp. Brain Res. 79, 192-206.
Seller, T. and Armitage, S. (1983). Diencephalic sites from which calling can
be evoked with small currents in Japanese quail. Behav. Brain Res. 9, 305-
314.
Serrier, J. and Moller, P. (1989). Patterns of electric organ discharge activity
in the weakly electric fish Brienomyrus niger L. (Mormyridae). J. Exp. Biol.
48, 235-244.
Teyssedre, C. and Boudinot, M. (1987). Rhythmicity as an intrinsic property
of the mormyrids electromotor command system. Physiol. Behav. 41, 201-
207.
Valentine, D., Sinha, S. and Moss, C. (2002). Orienting responses and
vocalizations produced by microstimulation in the superior colliculus of the
echolocating bat, Eptesicus fuscus. J. Comp. Physiol. A 188, 89-108.
von der Emde, G., Sena, L. G., Niso, R. and Grant, K. (2000). The midbrain
precommand nucleus of the mormyrid electromotor network. J. Neurosci.
20, 5483-5495.
Wang, Y. and Bieger, D. (1991). Role of solitarial GABAergic mechanisms
in control of swallowing. Am. J. Physiol. 261, R639-R646.
Williams, H. and Vicario, D. (1993). Temporal patterning of song
production: participation of nucleus uvaeformis of the thalamus. J.
Neurobiol. 24, 903-912.
Wullimann, M. F. and Northcutt, R. G. (1990). Visual and electrosensory
circuits of the diencephalon in mormyrids: an evolutionary perspective. J.
Comp. Neurol. 297, 537-552.
220
 Brain Evolution Triggers Increased Diversification of Electric Fishes
Bruce A. Carlson, et al.
Science 332, 583 (2011);
DOI: 10.1126/science.1201524

This copy is for your personal, non-commercial use only.

If you wish to distribute this article to others, you can order high-quality copies for your
colleagues, clients, or customers by clicking here.
Permission to republish or repurpose articles or portions of articles can be obtained by
following the guidelines here.

The following resources related to this article are available online at

www.sciencemag.org (this infomation is current as of June 23, 2011 ):

Downloaded from www.sciencemag.org on June 23, 2011

Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
http://www.sciencemag.org/content/332/6029/583.full.html
Supporting Online Material can be found at:
http://www.sciencemag.org/content/suppl/2011/04/27/332.6029.583.DC1.html
This article cites 61 articles, 22 of which can be accessed free:
http://www.sciencemag.org/content/332/6029/583.full.html#ref-list-1
This article appears in the following subject collections:
Evolution
http://www.sciencemag.org/cgi/collection/evolution

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
2011 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.
221
 REPORTS
Space Center. Support was provided by NSF grants and current records from 20042005 have SOM Text
OCE-0424953, OCE-0425361, and OCE-0647948; been deposited in the Marine Geoscience Data System: Figs. S1 to S6
Woods Hole Oceanographic Institution (WHOI) RIDGE 2000 Data Portal. The authors declare no Table S1
grants from the Deep Ocean Exploration Institute competing financial interests. References
and the Ocean Venture Fund; a National Defense Movies S1 and S2
Science and Engineering Graduate Fellowship to D.K.A.; Supporting Online Material
and the WHOI Jannasch Chair for Excellence in www.sciencemag.org/cgi/content/full/332/6029/580/DC1 30 November 2010; accepted 28 March 2011
Oceanography to L.S.M. Larval fluxes, chemical fluxes, Materials and Methods 10.1126/science.1201066

knollenorgans (18). Communication behavior is

Brain Evolution Triggers Increased mediated exclusively by knollenorgans (10). In a
region of the midbrain called the exterolateral
Diversification of Electric Fishes nucleus (EL; fig. S1), the timing of responses of
knollenorgans located on different parts of the
body is compared to extract information about
Bruce A. Carlson,1* Saad M. Hasan,1 Michael Hollmann,1 Derek B. Miller,1
electric signals (1013). Despite the importance
Luke J. Harmon,2 Matthew E. Arnegard3
of the EL for signal analysis, EL anatomy has
only been characterized in a few species (1012).
Communication can contribute to the evolution of biodiversity by promoting speciation and To investigate the role of brain evolution in mor-

Downloaded from www.sciencemag.org on June 23, 2011

reinforcing reproductive isolation between existing species. The evolution of species-specific myrid diversification, we performed a compara-
signals depends on the ability of individuals to detect signal variation, which in turn relies on tive analysis of EL anatomy. We obtained serial
the capability of the brain to process signal information. Here, we show that evolutionary change sections from the brains of 26 species (table S2).
in a region of the brain devoted to the analysis of communication signals in mormyrid electric After standard histological processing, we de-
fishes improved detection of subtle signal variation and resulted in enhanced rates of signal lineated the borders of the EL in each section
evolution and species diversification. These results show that neural innovations can drive the using established criteria (12). We then calculated
diversification of signals and promote speciation. total EL volume normalized to brain mass [see
supporting online material (SOM)].
lthough we assume that sensory process- tablished enhanced capacity for signal variation Previous studies identified distinct anterior

A ing is fundamentally important for the

detection of species-specific communi-
cations, or signals (1), we know relatively little
that is lacking in the Petrocephalinae, the only
other mormyrid subfamily (Fig. 1).
Mormyrids have three types of electrore-
and posterior subdivisions in the EL, referred to
as ELa and ELp, respectively (10). Sixteen of
the species that we studied clearly have separate
about how brain evolution might affect signal di- ceptors: ampullary organs, mormyromasts, and ELa/ELp subdivisions (Fig. 2A). However, the
vergence and speciation. African electric fishes with-
in the family Mormyridae provide an ideal model
system for relating brain evolution to diversifica- Paramormyrops sp. type 1
tion. The >200 described species in this family Paramormyrops sp. type 2
Paramormyrops sp. TEN
are phylogenetically and phenotypically diverse Paramormyrops curvifrons
0.06
(28); communicate using brief, species-specific, Paramormyrops longicaudatus
cytb nucleotide Paramormyrops kingsleyae
and easily quantified electric signals (9); and pro- substitutions per site Paramormyrops sp. SN9

Mormyrinae
cess these signals in a well-defined sensory path- Paramormyrops hopkinsi

clade A
Paramormyrops gabonensis
way devoted solely to the analysis of electric Paramormyrops sp. SZA
communication signals (1012) (fig. S1). origin of Marcusenius ntemensis
Mormyrids generate electric signals to com- ELa/ELp Boulengeromyrus knoepffleri
Ivindomyrus opdenboschi
municate and to actively sense their environment origin of electrocyte Ivindomyrus marchei

Mormyridae
(11). These signals have evolved more rapidly Stomatorhinus ivindoensis
stalk flexibility Marcusenius moorii
than body shape, size, and trophic ecology, sug- Isichthys henryi
gesting that electric communication behavior has Mormyrops zanclirostris
origin of Myomyrus macrops
played a key role in the radiation of mormyrids Petrocephalus simus
electrocyte
(3). Further, playback experiments in a few spe- Petrocephalus sullivani
stalks Petrocephalus valentini
cies suggest that these signals are critical for
Petrocephalinae

Petrocephalus sauvagii
species recognition during mate choice (1316). Petrocephalus pulsivertens
Electric signals are generated by an electric organ Petrocephalus christyi
Petrocephalus mbossou
in the tail, which consists of electrically excitable Petrocephalus grandoculis
cells called electrocytes (17). Electrocyte stalks Petrocephalus zakoni
origin of Petrocephalus binotatus
evolved with the origin of mormyrids, and devel- Petrocephalus balayi
opmental flexibility in stalk morphology arose ELa/ELp Petrocephalus odzalaensis
Petrocephalus microphthalmus (Gabon)
with the origin of the subfamily Mormyrinae (2). Petrocephalus microphthalmus (Congo)
This evolutionary change in the Mormyrinae es- Gymnarchus niloticus (Gymnarchidae)

1
Department of Biology, Washington University in St. Louis, St.
Louis, MO 63130, USA. 2Department of Biological Sciences,
University of Idaho, Moscow, ID 83844, USA. 3Human Biology
Division, Fred Hutchinson Cancer Research Center, Seattle, WA
98109, USA.
*To whom correspondence should be addressed. E-mail:
carlson.bruce@wustl.edu
Fig. 1. Inferred tree of phylogenetic relationships among mormyrid species and morphs. The phylogeny
was estimated by Bayesian analysis of cytb sequences (values at nodes are posterior probabilities). A
sequence from the closest outgroup to the Mormyridae (Gymnarchus niloticus) was used to root the tree.
Green branches represent a small exterolateral nucleus (EL) and magenta branches represent an enlarged
EL divided into anterior and posterior subdivisions (ELa/ELp); we reconstructed ancestral states using
parsimony (see text). Gray outline represents electric organs with electrocyte stalks, and black outline
represents electric organs with developmentally labile stalks, based on a previous study (2).

www.sciencemag.org SCIENCE VOL 332 29 APRIL 2011 222 583

REPORTS
remaining 10 species have a relatively small
EL without any apparent subdivisions (Fig. 2B).
Within the subfamily Mormyrinae, the genus
Myomyrus has a small EL, whereas all other
genera have an enlarged and subdivided ELa/ELp
(Fig. 2C and table S2). The monophyly of the
latter group is strongly supported by our cytb
phylogeny (Fig. 1), as well as several published
studies that used multiple molecular markers
(26). We therefore refer to this lineage as clade
A, and we conclude that all species in this
lineage have an ELa/ELp (Fig. 1). Within the sub-
family Petrocephalinae, we find that Petrocephalus
microphthalmus is the only species with an en-
larged ELa/ELp (Fig. 2C and table S2).
This pattern of EL anatomy suggests two
equally parsimonious scenarios: Either the EL
was ancestral and an enlarged ELa/ELp evolved
twice independently, or ELa/ELp was ancestral,
a broad distribution of receptors throughout the
body, or discrete clusters of receptors on the head
(20). We hypothesized that this difference might
relate to variation in EL anatomy. To test this hy-
pothesis, we mapped knollenorgan locations in
15 species (see SOM) and combined our findings
with published descriptions (6, 20, 21), yielding
data on 11 clade A species, 2 Myomyrus species,
and 13 petrocephaline species (table S2). All
clade A species and P. microphthalmus have
broad knollenorgan distributions (Fig. 3, A and B,
and fig. S2B). With one exception (Petrocephalus
zakoni), all other petrocephaline species have clus-
ters of knollenorgans located only on the head
(Fig. 3C and fig. S2C). Myomyrus has an in-
termediate phenotype, with a single cluster on the
head as well as a low-density, broad distribution
(Fig. 3D and fig. S2C). Thus, despite extensive in-
terspecific variation, ELa/ELp is universally as-
playback experiments in the field (see SOM).
Previous research identified two distinct behav-
ioral responses to electrosensory stimulation: in-
creases in electric discharge rate (22), or pauses
in electric output (23). We observed both re-
sponses. Among the six clade A species tested,
four responded to stimulation with rate increases,
and one responded with pauses. In the one re-
maining species (Brienomyrus brachyistius), two
individuals responded with rate increases, and
two responded with pauses. All four petrocepha-
line species responded with pauses.
In every species, repeated presentation of
the same stimulus led to a decrease in response
(i.e., habituation; fig. S4). We therefore used a
habituation-dishabituation paradigm to assess
signal discrimination ability in the field. Both
control and experimental stimuli consisted of 10
bursts of 10 pulses each. For controls, all 100

Downloaded from www.sciencemag.org on June 23, 2011

and a reduced EL evolved twice independently.
To distinguish between these two possibilities, we
related EL anatomy in mormyrids to a published
description of midbrain anatomy in the monotypic
Gymnarchidae, Gymnarchus niloticus (19), the
sister taxon to all mormyrids (Fig. 1). This study
identified an EL homolog without subdivisions,
leading us to conclude that the EL is the ancestral
mormyrid character state and that an enlarged
ELa/ELp evolved twice, once at the origin of
clade A and once in P. microphthalmus (Fig. 1).
Previous work has identified two distinct
patterns of knollenorgan receptor organization:
sociated with a broad distribution of knollenorgans pulses were an identical conspecific signal.
(table S2). Experimental stimuli were the same except that
Mormyrids analyze electric signals in the EL all 10 pulses in the ninth burst were a phase-
by comparing the response times of knollenor- shifted version of the same signal (fig. S3). Signal
gans located on different parts of the body (1013),
suggesting that species with broad receptor dis-
tributions should be better equipped for signal A Brienomyrus brachyistius
discrimination. Therefore, we hypothesized that
ELa/ELp evolved to facilitate the processing of
receptor responses for signal analysis, which pre-
dicts that species with an ELa/ELp should be
1 cm
better at detecting signal differences than species
with an EL. To test our hypothesis, we performed B Petrocephalus microphthalmus

A Brienomyrus Petrocephalus B Myomyrus Petrocephalus

brachyistius microphthalmus macrops soudanensis
tel
tel OT
OT 1 cm
OT OT
L L L
L
ll ll ll
ll ELa MD
ELa
ELp MD EL
C Petrocephalus soudanensis
MD ELp MD EL

val val val

val

1 cm

D Myomyrus macrops
C
Fig. 2. Anatomy of the exterolateral nucleus (EL). 5
EL volume/brain mass (mm3/g)

ELa/ELp
(A) Midbrain portion of 50-mm horizontal sections EL
from the clade A species B. brachyistius and the 4

petrocephaline P. microphthalmus. Dashed boxes in 1 cm

upper images delimit enlarged images below (scale
bars 1 mm and 500 mm, respectively). Both species
have an enlarged EL with distinct anterior and pos-
terior subdivisions (ELa/ELp). (B) The petrocephaline
P. soudanensis and mormyrine M. macrops both
have a small EL with no subdivisions. L, lateral
3
2 Fig. 3. An enlarged and subdivided exterolateral
nucleus (ELa/ELp) is universally associated with
1 broadly distributed knollenorgan electroreceptors.
Knollenorgan locations are indicated by red dots.
0 (A) B. brachyistius has a broad distribution of knol-
Campylomormyrus

P. microphthalmus
Paramormyrops

Stomatorhinus

Gnathonemus

P. soudanensis
P. pulsivertens

P. grandoculis
Marcusenius
Brienomyrus
Ivindomyrus

Mormyrops
Pollimyrus

P. valentini
P. sullivani
Myomyrus
Isichthys

P. zakoni

P. balayi
P. simus

nucleus; ll, lateral lemniscus; tel, telencephalon; OT, lenorgans, as found in all clade A species. (B) P.
optic tectum; MD, mediodorsal nucleus; val, valvula microphthalmus, the sole petrocephaline species
cerebellum. (C) Bar graph showing the median T with an ELa/ELp, also has a broad knollenorgan dis-
range of normalized EL volumes across all taxa tribution. (C) P. soudanensis has three knollenorgan
studied (table S2 gives sample sizes). clusters. (D) M. macrops has an intermediate pat-
clade A tern, with a single cluster and a low density of knol-
Mormyrinae Petrocephalinae lenorgans throughout the body.

584 29 APRIL 2011 VOL 332 SCIENCE www.sciencemag.org 223

 REPORTS
discrimination was assessed as the change in re- To test the importance of these traits on rates er compared to that of other mormyrids (s2A =
sponse from the eighth to ninth burst. In all clade of signal divergence, we analyzed the electric 0.000417, s2other = 0.000037, AICc = 13.9804,
A species, inserting a phase-shifted signal into signals of species collected in two locales: the P < 0.0002). For dimension 2, which reflects oth-
the stimulus train led to a partial recovery of re- Ivindo River of Gabon (3, 9) and Odzala Na- er signal features, we found a twofold higher rate
sponse (i.e., dishabituation) (Fig. 4A; discharge tional Park of the Republic of the Congo (5, 6), of signal divergence in clade A, but the difference
rate increases: n = 34, z = 4.57, P < 0.00001; homes to the largest known assemblages of Mor- was not significant (s2A = 0.000224, s2other =
pauses: n = 7, z = 2.37, P < 0.05). The three pet- myrinae and Petrocephalinae, respectively. The 0.000118, AICc = 1.1484; P = 0.25).
rocephaline species with an EL showed no evi- combined data set represents 18 species and To test our hypothesis that clade A has ex-
dence of discrimination (Fig. 4A; n = 12, z = 1.49, morphs within clade A and 13 outgroup species perienced higher rates of species diversification,
P = 0.13). However, in P. microphthalmus, the (fig. S5). Using cross-correlation (fig. S6) and we compared diversification rates in clade A to
sole petrocephaline species with an ELa/ELp, the multidimensional scaling (see SOM), we com- closely related outgroup lineages (25). There are
phase-shifted signal elicited dishabituation (Fig. puted coordinates of all 407 signals in this data at least 175 extant species in clade A (7). By con-
4A; n = 10, z = 2.80, P < 0.01). These results set within a two-dimensional space (Fig. 4B). We trast, there are only 3 known species of Myomyrus,
demonstrate that the independent evolution of then computed squared Mahalanobis distances and only 30 known species in the entire sub-
ELa/ELp established signal discrimination abil- (D2) between the centroids of each species or family Petrocephalinae (57). Every family of Os-
ities that are lacking in species with an EL. morph and plotted these values against ultra- teoglossomorpha outside the Mormyridae contains
Animal communication depends on both metric phylogenetic distances based on our phy- 10 or fewer extant species (7). Moreover, the on-
senders and receivers; in addition to the sensory logeny (Fig. 1). Assuming no particular model of ly two known osteoglossomorph species flocks
capacity for signal discrimination, traits per- evolution, the resulting plot reveals greater signal are restricted to clade A (3, 5, 8). We find sta-

Downloaded from www.sciencemag.org on June 23, 2011

mitting the evolution of signal variation are nec- variation and more rapid signal divergence in tistical support for net rates of diversification in
essary for a communication system that promotes clade A (Fig. 4C). We then formally compared clade A that are three to five times higher than
diversification. Therefore, we hypothesized that signal divergence rates using a Brownian motion those in closely related outgroup lineages (see
rapid signal divergence and species diversifica- framework (24). Considering dimension 1, which SOM). This result is robust across a range of pos-
tion should be restricted to clade A, the only reflects variation in temporal features of the sig- sible average extinction levels (table S3), support-
mormyrid lineage with both ELa/ELp and devel- nals (Fig. 4B), we found that the rate of signal ing our hypothesis that the evolution of ELa/ELp
opmentally flexible electrocytes (Fig. 1). divergence in clade A is more than 10 times fast- triggered explosive diversification in clade A
compared to that of other mormyroid lineages
(Fig. 1). However, this same evolutionary change
A in P. microphthalmus did not trigger rapid di-
clade A clade A other mormyrids
p<0.00001 8 p<0.05 versification. We propose that the lack of devel-
15
discharge rate (pulses/s)

opmentally flexible electrocyte stalks within the

Change in maximum

experimental
pause duration (s)

12
control 6 Petrocephalinae (2) impeded rapid diversification
through signal divergence in the P. microphthalmus
Change in

9 NS p<0.01
4
lineage.
6
2 Evolution of communication systems can have
3 profound effects on species radiation. Evolution-
0 0 ary change in the structure of the anuran inner ear
may have fostered increased rates of speciation
-3 -2
kingsleyae (N=26)

curvifrons (N=1)

moorii (N=2)

walkeri (N=2)

brachyistius (N=2)

marchei (N=5)

brachyistius (N=2)

balayi (N=1)

simus (N=7)

sullivani (N=4)

microphthalmus (N=10)
Paramormyrops

Paramormyrops

Marcusenius

Stomatorhinus

Brienomyrus

Ivindomyrus

Brienomyrus

Petrocephalus

Petrocephalus

Petrocephalus

Petrocephalus

through its effects on vocal communication (26).

Similarly, the adaptation of visual receptors to
different light environments has contributed to
cichlid speciation through its effects on mate
recognition (27, 28). The resulting radiations of
species can then lead to evolutionary divergence
B other mormyrids
N=13 species
C in brain structure as a secondary consequence of
clade A
0.5 N=18 species/morphs 500
adaptation to new ecological niches (29, 30). Our
results demonstrate the reverse relationship, in
Signal waveform distance

0.4 450
400 which brain evolution directly promotes diversi-
0.3
350 fication. We reveal that evolutionary change in
Dimension 2

0.2 300 the functional organization of sensory pathways

0.1 250 can establish new perceptual abilities that trigger
0 200 explosive diversification.
150
-0.1
100
-0.2 2 ms 50 References and Notes
-0.3 0 1. C. J. Hoskin, M. Higgie, Ecol. Lett. 13, 409
-0.4 -0.2 0 0.2 0.4 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 (2010).
Dimension 1 Phylogenetic distance 2. J. P. Sullivan, S. Lavou, C. D. Hopkins, J. Exp. Biol. 203,
665 (2000).
Fig. 4. Behavioral discrimination of signals and the evolution of signal diversity. (A) Dishabituation of 3. M. E. Arnegard et al., Am. Nat. 176, 335 (2010).
behavioral responses, shown as the change in response from the eighth to ninth stimulus (mean T SEM; 4. S. Lavou, J. P. Sullivan, C. D. Hopkins, Biol. J. Linn.
see text). Species in clade A are shown in red; all other species are shown in blue. Statistical significance Soc. Lond. 78, 273 (2003).
was assessed using Wilcoxons matched-pairs test. (B) Minimum polygons enclosing species or morphs in 5. S. Lavou, M. E. Arnegard, J. P. Sullivan, C. D. Hopkins,
J. Physiol. Paris 102, 322 (2008).
bivariate signal space (as obtained through multidimensional scaling of signal cross-correlations; see
6. S. Lavou, J. P. Sullivan, M. E. Arnegard, Zootaxa 2600,
SOM), including 11 example signals placed next to their location within this space. (C) Pairwise signal 1 (2010).
distances (Mahalanobis D2) between species or morphs plotted against pairwise phylogenetic distances 7. W. N. Eschmeyer, J. D. Fong, Catalog of Fishes (California
between cytb sequences. Academy of Sciences, San Francisco, 2010).

www.sciencemag.org SCIENCE VOL 332 29 APRIL 2011 224 585

REPORTS
8. P. G. D. Feulner, F. Kirschbaum, V. Mamonekene,
V. Ketmaier, R. Tiedemann, J. Evol. Biol. 20, 403 (2007).
9. C. D. Hopkins, Am. Zool. 21, 211 (1981).
10. M. A. Xu-Friedman, C. D. Hopkins, J. Exp. Biol. 202,
1311 (1999).
11. B. A. Carlson, in Communication in Fishes, F. Ladich,
S. P. Collin, P. Moller, B. G. Kapoor, Eds. (Science
Publishers, Enfield, NH, 2006), vol. 2, pp. 805848.
12. C. C. Bell, T. Szabo, in Electroreception, T. H. Bullock,
W. Heiligenberg, Eds. (John Wiley & Sons, New York,
1986), pp. 375-421.
13. C. D. Hopkins, A. H. Bass, Science 212, 85
(1981).
14. P. G. D. Feulner, M. Plath, J. Engelmann, F. Kirschbaum,
R. Tiedemann, Biol. Lett. 5, 225 (2009).
15. P. Machnik, B. Kramer, J. Exp. Biol. 211, 1969 (2008).
16. M. E. Arnegard, B. S. Jackson, C. D. Hopkins, J. Exp. Biol.
209, 2182 (2006).
17. A. H. Bass, in Electroreception, T. H. Bullock, W. Heiligenberg,
Eds. (Wiley, New York, 1986), pp. 1370.
18. H. H. Zakon, in Electroreception, T. H. Bullock,
W. Heiligenberg, Eds. (Wiley, New York, 1986),
pp. 103156.
19. A. H. Bass, C. D. Hopkins, J. Morphol. 174, 313
(1982).
20. W. Harder, Z. Vgl. Physiol. 59, 272 (1968).
21. S. Lavou, C. D. Hopkins, A. K. Toham, Zoosystema 26,
511 (2004).
22. N. Post, G. von der Emde, Physiol. Behav. 68, 115
(1999).
23. P. Moller, J. Serrier, D. Bowling, Ethology 82, 177
(1989).
24. B. C. OMeara, C. An, M. J. Sanderson, P. C. Wainwright,
Evolution 60, 922 (2006).
25. D. L. Rabosky, S. C. Donnellan, A. L. Talaba, I. J. Lovette,
Proc. Biol. Sci. 274, 2915 (2007).
26. M. J. Ryan, Proc. Natl. Acad. Sci. U.S.A. 83, 1379
(1986).
27. O. Seehausen et al., Nature 455, 620
(2008).
28. Y. Terai et al., PLoS Biol. 4, e433 (2006).
29. J. B. Sylvester et al., Proc. Natl. Acad. Sci. U.S.A. 107,
9718 (2010).
30. C. A. Shumway, Brain Behav. Evol. 72, 123 (2008).
Acknowledgments: We thank C. D. Hopkins and J. R. Gallant
for help with field work, J. P. Friel (Cornell University
Museum of Vertebrates) for providing specimens, and
S. Lavou and J. P. Sullivan for providing cytb sequences
(GenBank accession numbers provided in table S1).
Supported by NSF IOS-0818390 (B.A.C.); L.J.H. was
supported by NSF DEB-0919499.
Supporting Online Material
www.sciencemag.org/cgi/content/full/332/6029/583/DC1
Materials and Methods
Figs. S1 to S6
Tables S1 to S3
References
10 December 2010; accepted 21 March 2011
10.1126/science.1201524

take into account the state of neighboring auto-

Self-Organizing and Stochastic

Downloaded from www.sciencemag.org on June 23, 2011

mata. Automata in certain states can interact, gen-
erating complex, self-organizing patterns based
Behaviors During the Regeneration on a simple set of rules. Such patterning behavior
can be globally modulated by simple rule changes
of Hair Stem Cells in local automaton-to-automaton interactions (11).
To form regenerative patterns, activating sig-
nals among SCs should be able to spread and
Maksim V. Plikus,1 Ruth E. Baker,2 Chih-Chiang Chen,1,3 Clyde Fare,4 Damon de la Cruz,1 stop. This is possible when SCs can differentially
Thomas Andl,5 Philip K. Maini,2,6 Sarah E. Millar,7 Randall Widelitz,1 Cheng-Ming Chuong1,8* respond to the same signal at different times of
their regenerative cycle. We previously identified
Stem cells cycle through active and quiescent states. Large populations of stem cells in an organ four functional phases in the hair regenerative
may cycle randomly or in a coordinated manner. Although stem cell cycling within single hair cycle: signal-propagating (P) and nonpropagat-
follicles has been studied, less is known about regenerative behavior in a hair follicle population. ing phases (A), and phases refractory (R) and
By combining predictive mathematical modeling with in vivo studies in mice and rabbits, we competent (C) to such signals (12). Telogen HFs
show that a follicle progresses through cycling stages by continuous integration of inputs from in R phase cannot enter anagen because bone
intrinsic follicular and extrinsic environmental signals based on universal patterning principles. morphogenetic proteins (BMPs) in the surround-
Signaling from the WNT/bone morphogenetic protein activator/inhibitor pair is coopted to mediate ing skin macroenvironment keep hair SCs quies-
interactions among follicles in the population. This regenerative strategy is robust and versatile cent (12). Telogen HFs in C phase are devoid of
because relative activator/inhibitor strengths can be modulated easily, adapting the organism to these inhibitors and can enter anagen as long as
different physiological and evolutionary needs. the sum of intrinsic and extrinsic activators is
above the threshold. Intrinsic activators are pro-
ontinuous stem cell (SC) regeneration is ic structures and patterns by coordinating just a few duced as the result of hair SCs and dermal papilla

C essential for the maintenance of many

adult organs, for example, in the bone
marrow, skin, and gastrointestinal tract. Although
regenerative behavior within a single SC cluster
such as the hair bulge (1) or intestinal villi (2) has
been studied, it is largely unknown how the re-
generative behavior in populations of these SC
clusters is coordinated. During development,
thousands of cells can self-organize into anatom-
1
Department of Pathology, University of Southern California
(USC), Los Angeles, CA 90033, USA. 2Centre for Mathematical
Biology, Mathematical Institute, 24-29 St. Giles, Oxford OX1
3LB, UK. 3Institute of Clinical Medicine and Department of
Dermatology, National Yang-Ming University and Department of
Dermatology, Taipei Veterans General Hospital, Taipei, Taiwan.
4
Life Sciences Interface Doctoral Training Centre, Wolfson
Building, Parks Road, Oxford OX1 3QD, UK. 5Vanderbilt University
Medical Center, Nashville, TN 37232, USA. 6Oxford Centre for
Integrative Systems Biology, Department of Biochemistry, South
Parks Road, Oxford OX1 3QU, UK. 7Department of Dermatol-
ogy, University of Pennsylvania, Philadelphia, PA 19104, USA.
8
Research Center for Developmental Biology and Regenerative
Medicine, National Taiwan University, Taiwan.
*To whom correspondence should be addressed. E-mail:
cmchuong@usc.edu
morphogenetic signals (3), as seen in the periodic
patterning of skin appendages (4, 5). We hypoth-
esize that the regenerative cycling of adult organ
SCs can be similarly coordinated by diffusible sig-
nals and self-organize into spatiotemporal regen-
erative patterns.
Hair offers a suitable experimental model
because hair follicles (HFs) cycle through phases
of growth (anagen) and rest (telogen) (6). SCs are
clustered in hair bulges, making them easier to
study than SCs in other organs, where they are
usually scattered randomly (7) (fig. S1A). Grow-
ing hairs produce pigmentation patterns that allow
simultaneous monitoring of the regenerative
behavior of thousands of SCs (Fig. 1A) (8, 9).
Additionally, the skin is flat, restricting inter-
actions between HFs to two dimensions, further
simplifying the analysis.
We developed a cellular automaton (CA) model
consisting of a regular grid of automata, with one
automaton representing one HF (fig. S1B) (10).
The eight automata surrounding one automaton are
defined as its neighbors. With time, the state of
each automaton changes according to rules that
interactions. Extrinsic activators come from neigh-
boring P-phase anagen HFs and represent a form
of collective positive feedback. Thus, HFs can
enter anagen in two ways: autonomously, depend-
ing on the level of intrinsic activation, or non-
autonomously, when activators are delivered by
the surrounding macroenvironment. The probabil-
ity of anagen entry is based on the sum of these
fluctuating inputs.
We used mathematical simulations to test the
sufficiency and robustness of this model. We show
that the CA model encompassing PARC
cycling can reproduce the full spectrum of hair
regenerative patterns observed in mice: formation
of initiation centers, wave spreading, maintenance
of borders, and border instability (Fig. 1B, fig. S2,
and table S1).
For a model to be robust and capture con-
served patterning principles, it should be capable
of explaining the diverse regenerative patterns
seen in mutant mice and other animals. The
duration of each phase of the regenerative cycle
depends on the relative strengths of activators
and inhibitors (Fig. 1A). We suggest that dif-

586 29 APRIL 2011 VOL 332 SCIENCE www.sciencemag.org 225



226
 www.sciencemag.org/cgi/content/full/332/6029/583/DC1

Supporting Online Material for

Brain Evolution Triggers Increased Diversification of Electric Fishes

Bruce A. Carlson,* Saad M. Hasan, Michael Hollmann, Derek B. Miller, Luke J. Harmon,
Matthew E. Arnegard

*To whom correspondence should be addressed. E-mail: carlson.bruce@wustl.edu

Published 29 April 2011, Science 332, 583 (2011)

DOI: 10.1126/science.1201524

This PDF file includes:

Materials and Methods

Figs. S1 to S6
Tables S1 to S3
References

227
Materials and Methods
Phylogenetic Reconstruction
A sound phylogenetic framework is required to understand patterns of brain
evolution and how they relate to the diversification of signals and species. Mormyrids
have already been well studied phylogenetically (2-6, 8, 31-37). Our purpose in the
present analysis was not to extend this phylogenetic literature. Rather, our aim was to
estimate a single phylogeny, including all relevant species in the best-sampled
communities, for making a formal comparison of rates of signal divergence and species
diversification between mormyrid lineages (see below). Thus, we used published cytb
sequences from the mormyrid assemblage of the Ivindo River basin near Makokou,
Gabon (2, 3, 34) (aligned sequences provided by J. P. Sullivan) and from the
petrocephaline assemblage of Odzala National Park in the Lkoli River basin, Republic
of the Congo (5, 6) (aligned sequences provided by S. Lavou). In the case of the Odzala
petrocephaline assemblage, we only considered the most common cytb haplotype for
every species, each of which appears to be exclusively monophyletic with respect to all
other sympatric species (5, 6). To these focal species we added published cytb sequences
for the outgroup species, Gymnarchus niloticus (3) (aquarium specimen), and a specimen
of Myomyrus macrops collected by J. P. Sullivan from the Ubangi River, Central African
Republic (2). Table S1 lists the specimen numbers and GenBank accession numbers for
all specimens used to construct the phylogeny.
We conducted a Bayesian phylogenetic analysis on the matrix of aligned cytb
sequences using MrBayes ver. 3.1.2 (38, 39) and the GTR+ model of sequence
evolution: i.e., a general time reversible model with nucleotide substitution rate variation
assumed to follow a gamma distribution. The Metropolis-coupled, Markov chain Monte
Carlo (MCMCMC) process included four chains, three heated and one cold. Starting
from random trees, we simultaneously performed two independent Bayesian analyses
(runs) for 2,000,000 generations each. We were confident that stationarity had been
reached by this stopping point because the standard deviation of split frequencies had
dropped to 0.006. During the parallel runs, we sampled parameter values and trees every
100 generations. Log-likelihood values for the sampled trees stabilized by 100,000
generations into each run. Therefore, we only used the last 1,900,000 generations in both
runs to estimate the 50% majority-rule consensus tree and clade credibility values (i.e.,
posterior probabilities, PP). The consensus tree (Fig. 1) was computed using all 38,000
trees pooled across the two Bayesian runs. Although the monophyly of the
Petrocephalinae appears rather weakly supported in the consensus tree (PP = 0.52; see
Fig. 1), the monophyly of this group and its sister-clade relationship to the Mormyrinae
are strongly supported by morphological synapomorphies (21) and phylogenetic analysis
using whole mitogenome sequences (40). The phylogeny provides robust support for the
monophyly of clade A, the focus of the current study (PP = 1.00; see Fig. 1), in
agreement with several existing phylogenies based on multiple molecular markers (2-6,
32, 34, 37).
Next, we pruned the Ivindo River Petrocephalus microphthalmus specimen from the
tree, leaving the P. microphthalmus specimen collected in Odzala (this is the only species
with more than one representative in our phylogeny; see Fig. 1). Using the penalized
likelihood routine in the program r8s ver. 1.71 (41, 42), we converted the consensus

228
Bayesian tree to ultrametric form. We selected a smoothing parameter (log10 = 2.2)
using cross-validation and ran the penalized likelihood procedure, checking the stability
of the solution using the checkgradient command in r8s. All analyses of signal
divergence and species diversification (below) were conducted on this ultrametric tree.
When taxa were not included in an analysis, we simply pruned them from this ultrametric
tree, which served as the single phylogenetic framework for subsequent tests of signal
divergence and species diversification rates.

Brain Histology
We obtained fixed brains in one of three ways: (1) laboratory specimens after
anesthesia in 300 mg/l MS-222, each fish was perfused through the heart with Hickmans
Ringer, followed by 4% paraformaldehyde in 0.1 M phosphate buffer (e.g. Brienomyrus
brachyistius in Fig. 2A and Petrocephalus soudanensis in Fig. 2B); (2) freshly caught
field specimens after anesthesia in 300 mg/l MS-222, the skull was opened, followed by
immersion fixation in 4% paraformaldehyde in 0.1 M phosphate buffer (e.g.
Petrocephalus microphthalmus in Fig. 2A); (3) specimens donated by the Cornell
University Museum of Vertebrates brains were removed from curated specimens,
which had been fixed in 10% phosphate-buffered formalin for approximately two weeks
and subsequently stored in 70% ethanol (e.g. Myomyrus macrops in Fig. 2B). In each
case, we post-fixed brains in 4% paraformaldehyde in 0.1 M phosphate buffer for several
days, then embedded them in gelatin and post-fixed them overnight before slicing. We
obtained 50 m horizontal sections with a vibrating microtome. Sections were mounted
on chrom-alum subbed slides, stained with cresyl violet, dehydrated in a graded alcohol
series, cleared with xylene, and then coverslipped (43).
The exterolateral nucleus (EL) was clearly identified in each brain based on
topology (10, 12, 44-50), which is largely considered the single-most important criterion
for establishing the homology of brain regions (51-53). Previous anatomical studies of
EL in three clade A species describe separate anterior and posterior subdivisions that are
clearly divisible based on cytology; these are referred to as ELa and ELp, respectively
(10, 12, 44-50). As described in the main text, some of the species we studied fit this
previous description, while other species have just a single, small EL lacking any
subdivisions (Fig. 2); we refer to these two phenotypes as ELa/ELp and EL, respectively.
Evoked potential recordings from the ELa/ELp of P. microphthalmus reveal 2-3 ms
latency responses to electrosensory stimulation that are blocked at short delays following
the electric discharge motor command, exactly as described for species in clade A (48,
50, 54-56). This further satisfies the following additional criteria for establishing
homology (51, 57): physiological properties, connectivity (input from knollenorgans via
the nucleus of the electrosensory lateral line lobe, or nELL; see fig. S1), and function
(detection of electric signals generated by other fish). For each brain, we delineated the
borders of EL or ELa/ELp in each section based on cytology (12, 44-47). We then
determined the total volume of EL or ELa/ELp by summing cross-sectional areas,
multiplying by section thickness, and taking the mean of the left and right sides (no
lateral asymmetries were detected). We normalized this value by total brain mass,
determined after post-fixing and before embedding. No apparent differences in the
architecture or normalized volumes of these brain regions were observed between the

229
brains of closely-related specimens obtained through different fixation methods. A full
list of species for which we obtained EL anatomy data is provided in table S2.

Electroreceptor Histology
Curated specimens donated by the Cornell University Museum of Vertebrates had
been fixed in 10% phosphate-buffered formalin for approximately two weeks before
storage in 70% ethanol. Laboratory specimens were fixed in 70% ethanol. Mormyrids
have three distinct types of electroreceptors: ampullary organs used for passive
electrolocation, mormyromasts used for active electrolocation, and knollenorgans used
for electric communication (18, 20). There is a semi-transparent layer of skin that covers
the body where these receptors are located. In fixed specimens, this skin layer turns white
and is easily removed. We photographed specimens before removing this layer in pieces.
We then placed each piece in a solution of 0.05% toluidine blue for 10 minutes, followed
by three 5-minute washes in 70% ethanol. The three different types of electroreceptors
could then be visually identified and distinguished under a microscope (fig. S2A). We
mapped the location of knollenorgans in each piece of skin onto the original photograph,
and then outlined the specimen (Fig. 3; fig. S2B, C). A full list of species for which we
obtained data on knollenorgan distributions is provided in table S2.

Behavioral Playback Experiments

Playback experiments were performed on fish caught in Gabon near Lebamba
(2120 S, 11300 E) and Lambarn (04118 S, 101355 E) during July-August
2009. Methods for collecting mormyrids and recording electric signals have been
described in detail elsewhere (16, 34, 58, 59). For studying clade A, we focused our
efforts on Paramormyrops kingsleyae, a species that is widely distributed throughout
Gabon; however, we also performed experiments on small numbers of additional clade A
species (see Fig. 4A). For studying outgroup species, we focused on the four
petrocephaline species found in Gabon (21). For obtaining playback signals, we recorded
from individuals placed in water taken from the collection site (conductivity = 10-30
S/cm; temperature = 22-26 C). Using Ag/AgCl electrodes, electric signals were
amplified using a bandwidth of 0.0001-50 kHz (CWE, Inc. BMA-200), analog-to-digital
converted at 97.6 kHz (24-bit Sigma-Delta converter; Tucker-Davis Technologies RM1),
and then saved to disk using custom software written in Matlab 2007a (The MathWorks,
Inc.).
For playback experiments, each fish was placed in a rectangular PVC enclosure (3.5
x 3.5 x 20 cm) with Ag/AgCl stimulus electrodes spanning the length of both sides of the
middle of the inside of the enclosure, and Ag/AgCl recording electrodes at each end (fig.
S3A). Stimuli were digital-to-analog converted at 48.8 kHz (Tucker-Davis Technologies
RM1) and isolated from ground (A-M Systems, Inc. model 2200). The output of the fish
was amplified (CWE, Inc. BMA-200) and digitized at 48.8 kHz (Tucker-Davis
Technologies RM1). Custom software written in Matlab 2007a was used to deliver
stimuli and time-stamp the fishs electric discharges.
For stimuli, we used electric signal waveforms recorded from conspecifics, as well
as conspecific waveforms that were distorted by making a 90 phase-shift (fig. S3B). The
latter involves advancing the phase angles of the power spectrum by 90 for all positive
frequencies and retarding the phase angles by 90 for all negative frequencies, resulting in

230
a temporally distorted signal waveform with a frequency spectrum and total energy
equivalent to the original waveform (13, 60). Stimulus trains consisted of 10 bursts of 10
pulses each, with an intra-burst interval of 30 ms, inter-burst interval of 10 s, and peak-to-
peak intensity of 145 mV/cm. For control stimulation, all 100 pulses were an identical
conspecific waveform. Experimental stimuli were the same, except that all 10 pulses in
the 9th burst were a phase-shifted version of the same conspecific waveform (fig. S3C).
Each playback used a stimulus waveform recorded from a different conspecific (61).
In species that responded to stimulation with increases in discharge rate, we
determined the maximum discharge rates in response to each of the 10 bursts after
Gaussian smoothing (fig. S4A). In species that responded to stimulation by pausing, we
determined the duration of pauses in response to each of the 10 bursts (fig. S4B).
Discrimination was assessed as the change in maximum discharge rate or change in pause
duration from the 8th to 9th bursts (Fig. 4A).

Analysis of Signal Divergence Rates

To study broad patterns of signal evolution in the Mormyridae, we focused on the
most intensively studied assemblages of clade A species and petrocephaline species: i.e.,
the mormyrid assemblage of the Ivindo River basin (3) and the petrocephaline
assemblage of Odzala National Park (5, 6). The Ivindo assemblage is dominated by
numerous clade A species but also contains three petrocephaline species. In contrast, the
Odzala assemblage contains a high diversity of both clade A and petrocephaline species,
although the available molecular phylogenetic data for this assemblage only include the
eleven petrocephaline species known from Odzala. These two communities represent the
most diverse assemblages of clade A and Petrocephalinae, respectively, documented thus
far from any region of Africa (3, 5, 6, 9, 62). The numerous high quality electric signal
recordings and fine-scale biogeographical data collected for these assemblages are
unrivalled by any other mormyrid community that has been investigated previously. Our
comparison of signal divergence rates used previously published signal waveforms
recorded from these two assemblages (3, 5, 6) (fig. S5).
We performed cross-correlation between all 407 signal waveforms in the dataset
(63). We used the maximum of the absolute value of cross-correlation coefficients as a
pair-wise measure of waveform similarity (fig. S6). The absolute value was chosen since
waveform polarity is a variable feature that depends on the relative orientations of sender
and receiver (16, 64). This resulted in a matrix of pair-wise similarities ranging from 0
(no similarity) to 1 (identical waveforms). A phenotypic space describing signal variation
was then constructed by applying multidimensional scaling (MDS) to this similarity
matrix using the function mdscale with Kruskal's normalized stress1 criterion in Matlab
2007a (Fig. 4B). MDS is an ordination technique that projects similarities or distances in
some character or variable onto an N-dimensional coordinate space to maximally recreate
all pair-wise distances (65). The degree of correspondence between the pair-wise
distances in the input matrix relative to the MDS map is represented by stress, which can
range from 0 (perfect correspondence) to 1 (no correspondence). In our MDS analysis we
set N = 2, which resulted in a stress of 0.0635. Based on coordinate data for the two
resulting MDS axes, we computed the pair-wise Mahalanobis D2 between the centroids
of each species/morph distribution using the function mahalanobis in the program R
ver. 2.9.2 (66, 67). Since sample sizes were low for some species, we calculated D2 using

231
a pooled within-species variance-covariance matrix (3). We then plotted D2 against
patristic distances taken from the ultrametric tree for all clade A pair-wise comparisons
and all other pair-wise comparisons outside this clade. Patristic distances were scaled to a
maximum of 1.0. This plot allowed us to visualize the pattern of accumulation of signal
differences across distances in the phylogenetic tree without assuming any model of trait
evolution a priori (3). For clarity, we refer to D2 as signal waveform distance and
patristic distance as phylogenetic distance in Fig. 4C. We also computed D2 by
calculating separate pooled variance-covariance matrices for all clade A species and all
other species. This resulted in the same pattern of more rapid signal evolution in clade A
(data not shown).
We used Brownie ver. 2.1.1 to statistically compare rates of accumulation of signal
waveform disparity between clade A and all outgroup mormyrids in a Brownian motion
framework by means of maximum likelihood (24). The Brownian motion rate parameter
(2) describes the rate with which trait variance among species has increased over
evolutionary time. Using a censored rate test, we compared a model that fit a single 2
across the entire ultrametric tree of mormyrid relationships to a two-rate model allowing
clade A and all other mormyrids to differ in 2. This comparison was made separately for
each MDS-derived axis of signal variation. Given the moderate number of taxa under
consideration, we made each comparison using the small-sample-size-corrected version
of the Akaike Information Criterion (AICc) and a likelihood ratio test evaluated with a
parametric bootstrapping procedure (5,000 pseudoreplicates). These model evaluation
approaches avoid the inflated Type 1 errors that stem from using the 2 distribution with
small sample sizes (24, 68).

Analysis of Species Diversification Rates

We analyzed species diversification rates using the approach described by Rabosky
et al. (25) (table S3). This method is appropriate when one has an incompletely sampled
phylogenetic tree with branch lengths. We assigned missing species to lineages with
representatives in the tree using a combination of phylogenetic and taxonomic data, so
that we could account for the total diversity of each group. Net diversification rate (r) is
defined in the standard way as r = , where = rate of speciation and = rate of
extinction. We then compared a model with constant r across the tree to one where the
focal clade A was allowed to have a different net diversification rate (rA) from the rest of
the tree (rother). We compared the fit of these two nested models using a likelihood-ratio
test. Given our small dataset and the difficulty in estimating extinction rates from
comparative data (69), we set the relative extinction fraction to = / rather than
attempting to estimate it from the data (70). We repeated our analyses using a range of
values ( = 0, 0.5, 0.9, and 0.99) (25). Across this range of assumed , we found statistical
support for net diversification rates that are 3-5 times higher in clade A than in closely-
related out-group lineages (table S3). Patterns of species richness in sub-lineages
contained within clade A suggest that this increase in diversification immediately
followed the origin of clade A rather than arising more recently during its radiation. For
instance, the first extant lineage to split from the rest of clade A (Mormyrops spp.) is a
species-rich genus (21 extant species) with high signal waveform diversity (2, 4, 7, 58).

232
Fig. S1. Dorsal view of the brain of the clade A species Brienomyrus brachyistius,
highlighting the knollenorgan electrosensory pathway. Knollenorgan (KO) primary
afferent fibers project ipsilaterally to the nucleus of the electrosensory lateral line lobe
(nELL) in the hindbrain via the anterior (not shown) and posterior lateral line nerves
(nPLL). Neurons in the nELL project bilaterally to the anterior exterolateral nucleus
(ELa) in the midbrain, which in turn projects ipsilaterally to the adjacent posterior
exterolateral nucleus (ELp).

233
Fig. S2. Mapping of knollenorgan electroreceptor locations. (A) Portion of skin from the
clade A species Brienomyrus brachyistius stained with toluidine blue. Knollenorgans can
easily be distinguished from ampullary and mormyromast electroreceptors. (B) Maps of
knollenorgan locations from three species in clade A. (C) Maps of knollenorgan locations
from three species outside of clade A. Knollenorgan locations are indicated by red dots.
Scale bars = 1 cm.

234
Fig. S3. Behavioral discrimination experiments. (A) Set-up for field playback
experiments. Each fish was placed in a rectangular PVC enclosure with stimulus
electrodes spanning the length of both sides of the inside of the enclosure (red) and
recording electrodes at each end (blue). Stimuli were generated on a laptop and delivered
through the digital-to-analog converter (DAC) of a USB-connected portable processor.
Stimuli were isolated from ground prior to delivery (SIU). The output of the fish was
amplified (AMP), then digitized by the analog-to-digital converter (ADC) of the
processor. (B) Examples of stimulus waveforms used in experiments for three species.
Conspecific signals were recorded in the field. These waveforms were temporally
distorted by phase-shifting them by 90. (C) Two types of stimulus trains were used.
Control stimulus trains consisted of 10 bursts of 10 pulses each; each pulse was an
identical conspecific waveform. Experimental stimulus trains were the same except that
the 9th burst of pulses consisted of phase-shifted versions of the same conspecific
waveform.

235
Fig. S4. Habituation of behavioral responses to electrosensory stimulation. (A) In
Paramormyops kingsleyae, stimulation resulted in bursts of electrical output, quantified
as the maximum discharge rate after Gaussian smoothing. This response habituates
throughout the course of control stimulus trains. (B) In all four petrocephaline species,
Petrocephalus balayi (N = 1), P. microphthalmus (N = 10), P. simus (N = 7), and P.
sullivani (N = 4), stimulation elicited pauses in electrical output, quantified as pause
duration. This response also habituates throughout the course of control stimulus trains.
Values show the mean s.e.m. Statistical significance of the decrease in response with
repeated stimulus presentations was assessed using repeated-measures ANOVA.

236
Fig. S5. Electric signal waveforms used for analyzing rates of signal evolution. (A)
Waveforms from the Ivindo mormyrid assemblage of Gabon (3). (B) Waveforms from
the petrocephaline assemblage of Odzala National Park in the Republic of the Congo (5,
6). Species from clade A are shown in red, all other species are shown in blue. In each
case, waveforms are amplitude-normalized and plotted head-positive up. Multiple
waveforms from different individuals of the same species are superimposed and aligned
to the head-positive peak (except for Paramormyrops sp. TEN, for which waveforms
are aligned to the head-negative peak). The left and right columns show waveforms at
two different timescales (1 ms and 0.1 ms scale bars, respectively). The longest
waveforms are shown only in the left column.

237
Fig. S6. Cross-correlation method used to measure waveform similarity. (A) Example
showing the cross-correlation of two conspecific waveforms recorded from different
individuals of Paramormyrops curvifrons. The cross-correlation coefficients (bottom)
reveal the correlation between the two waveforms as a function of the relative delay (lag)
between them. The maximum of the absolute values of this function (C = 0.9587)
provides a scalar measure of waveform similarity. (B) Example showing the cross-
correlation of two heterospecific waveforms (P. curvifrons and Paramormyrops sp.
magnostipes type 1). Here the waveforms are less similar (C = 0.3996).

12

238
Table S1. Cytochrome b (cytb) sequences used for estimating mormyrid phylogeny,
showing corresponding specimen information. Museum catalog numbers preceded by
CU are specimens housed in the Cornell University Museum of Vertebrates, Ithaca,
NY. The single museum catalog number preceded by AMNH is a specimen housed in
the American Museum of Natural History, New York, NY.
Museum catalog GenBank
Taxon no., specimen no. Collection site accession no.
Paramormyrops sp. type 1 CU 78326, #2297 Loa Loa Rapids, Ivindo River, Gabon AF477452
Paramormyrops sp. type 2 CU 78344, #2295 Loa Loa Rapids, Ivindo River, Gabon AF477455
Paramormyrops sp. TEN CU 80809, #2011 Bial Stream, Ivindo R. basin, Gabon AF477453
Paramormyrops curvifrons CU 81661, #2050 Loa Loa Rapids, Ivindo River, Gabon AF477469
Paramormyrops longicaudatus CU 78355, #2289 Loa Loa Rapids, Ivindo River, Gabon AF201576
Paramormyrops kingsleyae CU 80816, #2116 Bial Stream, Ivindo R. basin, Gabon AF477466
Paramormyrops sp. SN9 CU 89360, #5552 Loa Loa Rapids, Ivindo River, Gabon FJ830628
Paramormyrops hopkinsi CU 78352, #2285 Loa Loa Rapids, Ivindo River, Gabon AF201575
Paramormyrops gabonensis CU 79702, #2048 Loa Loa Rapids, Ivindo River, Gabon AF201603
Paramormyrops sp. SZA CU 80848, #2008 Bial Stream, Ivindo R. basin, Gabon AF477475
Marcusenius ntemensis CU 79706, #2186 Loa Loa Rapids, Ivindo River, Gabon AF201593
Boulengeromyrus knoepffleri CU 79692, #2248 Loa Loa Rapids, Ivindo River, Gabon AF201573
Ivindomyrus opdenboschi CU 84642, #2106 Loa Loa Rapids, Ivindo River, Gabon DQ166689
Ivindomyrus marchei CU 81642, #2183 Loa Loa Rapids, Ivindo River, Gabon DQ166677
Stomatorhinus ivindoensis CU 70703, #2074 Bial Stream, Ivindo R. basin, Gabon AF201612
Marcusenius moorii CU 79697, #2013 Bal Creek, Ivindo R. basin, Gabon AF201595
Isichthys henryi CU 79705, #2179 Loa Loa Rapids, Ivindo River, Gabon AF201590
Makokou region, Ivindo River,
Mormyrops zanclirostris CU 79707, #2210 AF201599
Gabon
AMNH 228166, Ubangi River, Congo R. basin,
Myomyrus macrops AF201602
#2524 Central African Republic
Petrocephalus simus CU 79701, #2035 Bal Creek, Ivindo R. basin, Gabon AF201604
Petrocephalus sullivani CU 79700, #2038 Bal Creek, Ivindo R. basin, Gabon AF201606
Petrocephalus valentini CU 88058, #5175 Lkoli River, Congo R. basin, Congo EU770181
Petrocephalus sauvagii CU 87864, #5206 Lkoli River, Congo R. basin, Congo EU770160
Petrocephalus pulsivertens CU 88097, #5263 Lkoli River, Congo R. basin, Congo EU770174
Petrocephalus christyi CU 88095, #5261 Lkoli River, Congo R. basin, Congo EU770183
Petrocephalus mbossou CU 92389, #6183 Lkoli River, Congo R. basin, Congo EU770163
Petrocephalus grandoculis CU 92385, #6181 Lkoli River, Congo R. basin, Congo EU770155
Petrocephalus zakoni CU 92391, #6132 Lkoli River, Congo R. basin, Congo EU770170
Petrocephalus binotatus CU 88064, #5001 Lkoli River, Congo R. basin, Congo EU770164
Petrocephalus balayi CU 88111, #5314 Lkoli River, Congo R. basin, Congo EU770192
Petrocephalus odzalaensis CU 87852, #5148 Lkoli River, Congo R. basin, Congo EU770156
Petrocephalus microphthalmus CU 82208, #2199 Loa Loa Rapids, Ivindo River, Gabon EU770185
Petrocephalus microphthalmus CU 87940, #5092 Lkoli River, Congo R. basin, Congo EU770188
Gymnarchus niloticus CU 80334, no spec. # Aquarium import AF201586

239
Table S2. Specimens used for studying brain anatomy and knollenorgan distributions.
Taxonomic assignments are based on our cytochrome b (cytb) phylogeny (Fig. 1) as well
as several published studies (2-6, 32, 34, 37, 58).
Sub- EL anatomy Pattern of EL Knollenorgan
family Genus Species1 sample size organization2 distribution3
Brevimyrus niger B4
Brienomyrus brachyistius 5 ELa/ELp B
numenius 1 ELa/ELp
Campylomormyrus rhynchophorus B4
tamandua 1 ELa/ELp B
Gnathonemus petersii 1 ELa/ELp B
Isichthys henryi 2 ELa/ELp
Ivindomyrus marchei 2 ELa/ELp
2 ELa/ELp B4
Clade A

moorii
Mormyrinae

Marcusenius
senegalensis B4
nigricans 1 ELa/ELp B4
Mormyrops
zanclirostris 2 ELa/ELp B
Mormyrus caballus B4
gabonensis 2 ELa/ELp
longicaudatus 1 ELa/ELp
Paramormyrops
sp. BON 1 ELa/ELp
sp. VAD 1 ELa/ELp
Pollimyrus adspersus 2 ELa/ELp B
Stomatorhinus walkeri 3 ELa/ELp
macrops 1 EL C-B
Myomyrus
pharao 1 EL C-B
balayi 1 EL C4
binotatus C4
christyi C4
grandoculis 1 EL C
Petrocephalinae

microphthalmus 3 ELa/ELp B
odzalaensis C4
Petrocephalus pulsivertens 1 EL C
sauvagii C4
simus 4 EL C
soudanensis 4 EL C
sullivani 2 EL C
valentini 1 EL C
zakoni 3 EL B
1. Undescribed species are referred to by established 3-letter cheironyms (34, 59).
2. EL = relatively small exterolateral nucleus; ELa/ELp = relatively large exterolateral nucleus
divided into anterior and posterior subdivisions.
3. B = broad distribution of knollenorgan receptors throughout the head and trunk; C = distinct
clusters of knollenorgan receptors on the head; C-B = intermediate pattern with one cluster of
knollenorgan receptors on the head as well as a broad distribution of knollenorgan receptors at
relatively low density.
4. Knollenorgan distributions based on published studies (6, 20, 21). These previous studies used visual
examination of intact specimens rather than the staining method employed in the current study.

240
Table S3. Model comparisons testing for a shift in diversification rate in clade A relative
to other mormyrids. We set the relative extinction fractions () to different values, and
then compared a model with constant net diversification rate (r) across the tree to one
with different net diversification rates for clade A (rA) and the rest of the tree (rother).
Model Parameter estimates rA/rother lnL Delta p-value
=0
Single rate r=0.0079 -95.1
Two rate rother=0.0032, rA=0.0100 3.12 -84.0 22.2 <0.001
= 0.5
Single rate r=0.0062 -92.7
Two rate rother=0.0025, rA=0.0083 3.32 -84.1 17.1 <0.001
= 0.9
Single rate r=0.0030 -90.1
Two rate rother=0.0010, rA=0.0042 4.20 -85.7 8.8 0.003
= 0.99
Single rate r=0.0005 -90.5
Two rate rother=0.00014, rA=0.00073 5.21 -88.1 4.8 0.03

15

241
References and Notes
1. C. J. Hoskin, M. Higgie, Speciation via species interactions: The divergence of mating
traits within species. Ecol. Lett. 13, 409 (2010).
2. J. P. Sullivan, S. Lavou, C. D. Hopkins, Molecular systematics of the African electric
fishes (Mormyroidea: teleostei) and a model for the evolution of their electric
organs. J. Exp. Biol. 203, 665 (2000).
3. M. E. Arnegard et al., Sexual signal evolution outpaces ecological divergence during
electric fish species radiation. Am. Nat. 176, 335 (2010).
4. S. Lavou, J. P. Sullivan, C. D. Hopkins, Phylogenetic utility of the first two introns of
the S7 ribosomal protein gene in African electric fishes (Mormyroidea: Teleostei)
and congruence with other molecular markers. Biol. J. Linn. Soc. Lond. 78, 273
(2003).
5. S. Lavou, M. E. Arnegard, J. P. Sullivan, C. D. Hopkins, Petrocephalus of Odzala
offer insights into evolutionary patterns of signal diversification in the
Mormyridae, a family of weakly electrogenic fishes from Africa. J. Physiol. Paris
102, 322 (2008).
6. S. Lavou, J. P. Sullivan, African weakly electric fishes of the genus Petrocephalus
(Osteoglossomorpha: Mormyridae) of Odzala National Park, Republic of the
Congo (Lkoli River, Congo River basin) with description of five new species. M.
E. Arnegard, Zootaxa 2600, 1 (2010).
7. W. N. Eschmeyer, J. D. Fong, Catalog of Fishes (California Academy of Sciences, San
Francisco, 2010).
8. P. G. D. Feulner, F. Kirschbaum, V. Mamonekene, V. Ketmaier, R. Tiedemann,
Adaptive radiation in African weakly electric fish (Teleostei: Mormyridae:
Campylomormyrus): A combined molecular and morphological approach. J. Evol.
Biol. 20, 403 (2007).
9. C. D. Hopkins, On the diversity of electric signals in a community of mormyrid
electric fish in west Africa. Am. Zool. 21, 211 (1981).
10. M. A. Xu-Friedman, C. D. Hopkins, Central mechanisms of temporal analysis in the
knollenorgan pathway of mormyrid electric fish. J. Exp. Biol. 202, 1311 (1999).
11. B. A. Carlson, in Communication in Fishes, F. Ladich, S. P. Collin, P. Moller, B. G.
Kapoor, Eds. (Science Publishers, Enfield, NH, 2006), vol. 2, pp. 805-848.
12. C. C. Bell, T. Szabo, in Electroreception, T. H. Bullock, W. Heiligenberg, Eds. (John
Wiley & Sons, New York, 1986), pp. 375-421.
13. C. D. Hopkins, A. H. Bass, Temporal coding of species recognition signals in an
electric fish. Science 212, 85 (1981).

16

242
14. P. G. D. Feulner, M. Plath, J. Engelmann, F. Kirschbaum, R. Tiedemann, Electrifying
love: Electric fish use species-specific discharge for mate recognition. Biol. Lett.
5, 225 (2009).
15. P. Machnik, B. Kramer, Female choice by electric pulse duration: Attractiveness of
the males communication signal assessed by female bulldog fish, Marcusenius
pongolensis (Mormyridae, Teleostei). J. Exp. Biol. 211, 1969 (2008).
16. M. E. Arnegard, B. S. Jackson, C. D. Hopkins, Time-domain signal divergence and
discrimination without receptor modification in sympatric morphs of electric
fishes. J. Exp. Biol. 209, 2182 (2006).
17. A. H. Bass, in Electroreception, T. H. Bullock, W. Heiligenberg, Eds. (Wiley, New
York, 1986), pp. 1370.
18. H. H. Zakon, in Electroreception, T. H. Bullock, W. Heiligenberg, Eds. (Wiley, New
York, 1986), pp. 103156.
19. A. H. Bass, C. D. Hopkins, Comparative aspects of brain organization of an African
wave electric fish, Gymnarchus niloticus. J. Morphol. 174, 313 (1982).
20. W. Harder, Die Beziehungen zwischen Elektrorezeptoren, Elektrischem Organ,
Seitenlinienorganen und Nervensystem bei den Mormyridae (Teleostei, Pisces).
Z. Vgl. Physiol. 59, 272 (1968).
21. S. Lavou, C. D. Hopkins, A. K. Toham, The Petrocephalus (Pisces,
Osteoglossomorpha, Mormyridae) of Gabon, Central Africa, with the description
of a new species. Zoosystema 26, 511 (2004).
22. N. Post, G. von der Emde, The novelty response in an electric fish: Response
properties and habituation. Physiol. Behav. 68, 115 (1999).
23. P. Moller, J. Serrier, D. Bowling, electric organ discharge displays during social
encounter in the weakly electric fish Brienomyrus niger L. (Mormyridae).
Ethology 82, 177 (1989).
24. B. C. OMeara, C. An, M. J. Sanderson, P. C. Wainwright, Testing for different rates
of continuous trait evolution using likelihood. Evolution 60, 922 (2006).
25. D. L. Rabosky, S. C. Donnellan, A. L. Talaba, I. J. Lovette, Exceptional among-
lineage variation in diversification rates during the radiation of Australias most
diverse vertebrate clade. Proc. Biol. Sci. 274, 2915 (2007).
26. M. J. Ryan, Neuroanatomy influences speciation rates among anurans. Proc. Natl.
Acad. Sci. U.S.A. 83, 1379 (1986).
27. O. Seehausen et al., Speciation through sensory drive in cichlid fish. Nature 455, 620
(2008).
28. Y. Terai et al., Divergent selection on opsins drives incipient speciation in Lake
Victoria cichlids. PLoS Biol. 4, e433 (2006).

17

243
29. J. B. Sylvester et al., Brain diversity evolves via differences in patterning. Proc. Natl.
Acad. Sci. U.S.A. 107, 9718 (2010).
30. C. A. Shumway, Habitat complexity, brain, and behavior. Brain Behav. Evol. 72, 123
(2008).
31. J. Alves-Gomes, C. D. Hopkins, Molecular insights into the phylogeny of
mormyriform fishes and the evolution of their electric organs. Brain Behav. Evol.
49, 324 (1997).
32. S. Lavou, R. Bigorne, G. Lecointre, J.-F. Agnse, Phylogenetic relationships of
mormyrid electric fishes (Mormyridae; Teleostei) inferred from cytochrome b
sequences. Mol. Phylogenet. Evol. 14, 1 (2000).
33. S. Lavou, J. P. Sullivan, M. E. Arnegard, C. D. Hopkins, Differentiation of
morphology, genetics and electric signals in a region of sympatry between sister
species of African electric fish (Mormyridae). J. Evol. Biol. 21, 1030 (2008).
34. J. P. Sullivan, S. Lavou, C. D. Hopkins, Discovery and phylogenetic analysis of a
riverine species flock of African electric fishes (Mormyridae: Teleostei).
Evolution 56, 597 (2002).
35. J. P. Sullivan, S. Lavou, M. E. Arnegard, C. D. Hopkins, AFLPs resolve phylogeny
and reveal mitochondrial introgression within a species flock of African electric
fish (Mormyroidea: Teleostei). Evolution 58, 825 (2004).
36. P. G. D. Feulner, F. Kirschbaum, R. Tiedemann, Adaptive radiation in the Congo
River: An ecological speciation scenario for African weakly electric fish
(Teleostei; Mormyridae; Campylomormyrus). J. Physiol. Paris 102, 340 (2008).
37. M. E. Arnegard, D. J. Zwickl, Y. Lu, H. H. Zakon, Old gene duplication facilitates
origin and diversification of an innovative communication systemtwice. Proc.
Natl. Acad. Sci. U.S.A. 107, 22172 (2010).
38. J. P. Huelsenbeck, F. Ronquist, MRBAYES: Bayesian inference of phylogenetic
trees. Bioinformatics 17, 754 (2001).
39. F. Ronquist, J. P. Huelsenbeck, MrBayes 3: Bayesian phylogenetic inference under
mixed models. Bioinformatics 19, 1572 (2003).
40. S. Lavou et al., Remarkable morphological stasis in an extant vertebrate despite tens
of millions of years of divergence. Proc. Biol. Sci. 278, 1003 (2011).
41. M. J. Sanderson, Estimating absolute rates of molecular evolution and divergence
times: A penalized likelihood approach. Mol. Biol. Evol. 19, 101 (2002).
42. M. J. Sanderson, r8s: Inferring absolute rates of molecular evolution and divergence
times in the absence of a molecular clock. Bioinformatics 19, 301 (2003).
43. B. A. Carlson, Neuroanatomy of the mormyrid electromotor control system. J. Comp.
Neurol. 454, 440 (2002).

18

244
44. F. Hauged-Carr, The mesencephalic exterolateral posterior nucleus of the
mormyrid fish Brienomyrus niger: Efferent connections studied by the HRP
method. Brain Res. 178, 179 (1979).
45. E. Mugnaini, L. Maler, Cytology and immunocytochemistry of the nucleus
extrolateralis anterior of the mormyrid brain: Possible role of GABAergic
synapses in temporal analysis. Anat. Embryol. (Berl.) 176, 313 (1987).
46. M. A. Friedman, M. Kawasaki, Calretinin-like immunoreactivity in mormyrid and
gymnarchid electrosensory and electromotor systems. J. Comp. Neurol. 387, 341
(1997).
47. M. A. Friedman, C. D. Hopkins, Neural substrates for species recognition in the time-
coding electrosensory pathway of mormyrid electric fish. J. Neurosci. 18, 1171
(1998).
48. B. A. Carlson, Temporal-pattern recognition by single neurons in a sensory pathway
devoted to social communication behavior. J. Neurosci. 29, 9417 (2009).
49. S. Amagai, M. A. Friedman, C. D. Hopkins, Time coding in the midbrain of
mormyrid electric fish. I. Physiology and anatomy of cells in the nucleus
exterolateralis pars anterior. J. Comp. Physiol. A Neuroethol. Sens. Neural Behav.
Physiol. 182, 115 (1998).
50. S. Amagai, Time coding in the midbrain of mormyrid electric fish. II. Stimulus
selectivity in the nucleus exterolateralis pars posterior. J. Comp. Physiol. A
Neuroethol. Sens. Neural Behav. Physiol. 182, 131 (1998).
51. G. F. Striedter, Principles of Brain Evolution (Sinauer, Sunderland, MA, 2005).
52. R. Nieuwenhuys, Comparative neuroanatomy: Place, principles, practice and
programme. Eur. J. Morphol. 32, 142 (1994).
53. L. Puelles, L. Medina, Field homology as a way to reconcile genetic and
developmental variability with adult homology. Brain Res. Bull. 57, 243 (2002).
54. P. S. Enger, S. Libouban, T. Szabo, Fast conducting electrosensory pathway in the
mormyrid fish, Gnathonemus petersii. Neurosci. Lett. 2, 133 (1976).
55. C. J. Russell, C. C. Bell, Neuronal responses to electrosensory input in mormyrid
valvula cerebelli. J. Neurophysiol. 41, 1495 (1978).
56. T. Szabo, P. S. Enger, S. Libouban, Electrosensory systems in the mormyrid fish,
Gnathonemus petersii: Special emphasis on the fast conducting pathway. J.
Physiol. (Paris) 75, 409 (1979).
57. C. B. G. Campbell, W. Hodos, The concept of homology and the evolution of the
nervous system. Brain Behav. Evol. 3, 353 (1970).
58. C. D. Hopkins, S. Lavou, J. P. Sullivan, in The Fresh and Brackish Water Fishes of
Lower Guinea, West-Central Africa, M. L. J. Stiassny, G. G. Teugels, C. D.

19

245
 Hopkins, Eds. (IRD, Publications scientifiques du Musum, MRAC, Paris, 2007),
vol. 1.
59. M. E. Arnegard, S. M. Bogdanowicz, C. D. Hopkins, Multiple cases of striking
genetic similarity between alternate electric fish signal morphs in sympatry.
Evolution 59, 324 (2005).
60. W. Heiligenberg, R. A. Altes, Phase sensitivity in electroreception. Science 199, 1001
(1978).
61. D. Kroodsma, Suggested experimental designs for song playbacks. Anim. Behav. 37,
600 (1989).
62. C. D. Hopkins, Design features for electric communication. J. Exp. Biol. 202, 1217
(1999).
63. C. Chatfield, The Analysis of Time Series: An Introduction. (Chapman and Hall, New
York, 2004).
64. C. D. Hopkins, Temporal structure of non-propagated electric communication signals.
Brain Behav. Evol. 28, 43 (1986).
65. J. B. Kruskal, M. Wish, Multidimensional Scaling. (Sage, Beverly Hills, CA, 1978).
66. P. Mahalanobis, On the generalized distance in statistics. Proc. Natl. Inst. Sci. India
2, 49 (1936).
67. N. Eldredge et al., The dynamics of evolutionary stasis. Paleobiology 31, (suppl), 133
(2005).
68. S. A. Price et al., Functional innovations and morphological diversification in
parrotfish. Evolution 64, 3057 (2010).
69. D. L. Rabosky, Extinction rates should not be estimated from molecular phylogenies.
Evolution 64, 1816 (2010).
70. M. E. Alfaro et al., Nine exceptional radiations plus high turnover explain species
diversity in jawed vertebrates. Proc. Natl. Acad. Sci. U.S.A. 106, 13410 (2009).
Acknowledgments: We thank C. D. Hopkins and J. R. Gallant for help with field work,
J. P. Friel (Cornell University Museum of Vertebrates) for providing specimens,
and S. Lavou and J. P. Sullivan for providing cytb sequences (GenBank
accession numbers provided in table S1). Supported by NSF IOS-0818390
(B.A.C.); L.J.H. was supported by NSF DEB-0919499.

20

246
 Suthers, Roderick (2004) How birds sing and why it matters. In Nature's Music: the
science of birdsong. Marler, P. and Slabbekoorn, H. eds. Elsevier. New York. Chapter 9
pp. 272-295.
How birds sing and why it matters 273

Chapter 9 Tracheal syrinx ,

Variation in Syringeal Structure
Bronchial syrinx Tracheobronchial syrinx
I (Parrot) (Oilbird) (Canary)

I
How birds sing and why it matters
RODERICK A. SUTHERS 'I
1 I
! MT
!
I
INTRODUCTION briefly to indicate the variety of avian vocal I
systems, and to provide a perspective from which I

1 I
Songbirds have both an esthetic and a scientific to focus on the oscine songbirds. Readers 1

impact on our lives. Birdsong adds beauty and
vitality to our environment, The possibility of
its absence due to in c r eas ing levels of
environmental toxins, so eloquently described
by Rachel Carson (1962) in her goundbreaking
book, Silent Spring, was a potent factor in
mobilizing public opinion to become better
stewards of our environment. Scientifically, the
highly developed vocal communication of this
group, sharing as it does a number of parallels
with speech, has made songbirds especially suited
for the study of complex, learned vocal
communication at every level, from its molecular
biology to its evolution.
This chapter is about how birdsong is
produced. How do birds generate such a variety
of sounds and what vocal gymnastics are required
to produce them?The vocal system is the interface
between the bird's brain and his song. Knowing
how it functions, and coming to appreciate its
limitations as well as its capabilities, is important
in understanding vocal communication. Birdsong
is the product of the carefully coordinated activity
of many different muscles. Some of these are
associated with the vocal organ but many belong
to other muscle groups such as the respiratory
system and the vocal tract. Together they convert
nerve impulses from the brain into song.
Less is known about the vocal mechanisms of
non-songbirds, and they will be mentioned only
interested in additional information on songbirds
should consult reviews of this subject (Nowicki
& Marler 1988; Suthers 1997, 1999a, b; Gaunt
& Nowicki 1998; Doupe 81Kuhl 1999; Suthers
ef ale 1999; Goller & Larsen 2002)-
AN ORGAN FOR SINGING
The avian vocal organ, the syrinx, is located
deep in the chest in an air sac that is connected
to other air sacs and the lungs. It is present in all
birds except vultures, but its exact location and
structure varies considerably (King 1989). In
some species, including doves, pigeons and
parrots, the syrinx consists of modified rings of
cartilage near the lower end of trachea
(Fig. 9.1).
In other species, such as penguins, many owls,
nightjars, and some cuckoos, the syrinx
exists as two semisyrinxes located in the primary
bronchi (Fig. 9.1). In many birds, however,
including but by no means limited to the oscine
songbirds, the syrinx is at the junction where
the two primary bronchi join to form the trachea.
This tracheobronchial syrinx includes modified
cartilages from both the upper end of each
bronchus and the lower end of the trachea
(Fig. 9.1).
Figure 9.1 Examples of variation in syringeal anatomy. The tracheal parrot syrinx has two syringeal
muscles and a pair of lateral tympaniform membranes. The bronchial syrinx of the oilbird has one pair of
syringeal muscles and a pair of medial and lateral tympaniform membranes in each bronchus. Songbirds
have several pairs of syringeal muscles in their tracheobronchial syrinx. Tr, trachea; ST, sternotrachealis
muscle; SY SUP, superficial syringeal muscle; SY PROF, deep syringeal muscle; SY VALV, pneumatic
valve; LTM, lateral tympaniform me~brane;MTM, medial tympaniform membrane; BCI, first bronchial
cartilage; SY, syringeal muscle; ML, medial labium; LL, lateral labium; SYR, muscles of syrinx. (Oilbird
modified after Suthers & Hector 1985; parrot and canary modified after King 1989).
The Respiratory Rhythm direction through the syrinx and into the
The timing of vocalization and the tempo of esophagus. After each coo there is a short
song begins with the respiratory cycle (Vicario expiration followed by a brief inspiration (Gaunt
1991). Expiratory muscles compress the bellows- et al. 1982).
like air sacs in the thorax and abdomen, increasing
respiratory pressure so air flows out through the
syrinx and trachea (Hartley 1990). During song, The Source of Sound
a small inspiration called a 'mini-breath' (Calder It is now generally agreed that vocalizations are
1970) is usually taken after each syllable, except generated by airflow-induced oscillation of
at very high syllable repetition rates (Hartley & elements in the wall of the syrinx that convert
Suthers 1989). Inspiratory muscles expand the some of the air's kinetic energy into acoustic
thorax and abdomen, reducing pressure in the energy. These oscillations are presumably
air sacs and reversing the direction of airflow sustained by interaction between Bernoulli forces
through the syrinx (Wild et al. 1998). and the inertia of air in the vocal tract, in much
Vocalization occurs during expiratory airflow, the same way as are oscillations of human vocal
aside from rare exceptions when it occurs during folds (Titze 1994; Gardner et al. 200 1). An
inspiration, as in the 'wah' sounds of doves (Gaunt alternative mechanism for sound production
I et al. 1982), and certain syllables in the songs of based on the principle of an aerodynamic whistle
some zebra finches (Goller & Daley 2001). Most was put forth (Gaunt et al. 1982) to explain
birds sing with their beaks open, but doves and pure tone vocalizations in birds such as doves.
pigeons coo with their beaks and nares closed. However, this aerodynamic whistle hypothesis
Sound is produced during airflow in an expiratory is not supported by the observation of

247

Chapter 9
I (Parrot)
I
How birds sing and why it matters
RODERICK A. SUTHERS
! MT
INTRODUCTION
1 I
Songbirds have both an esthetic and a scientific
impact on our lives. Birdsong adds beauty and
vitality to our environment, The possibility of
its absence due to in c r eas ing levels of
environmental toxins, so eloquently described
by Rachel Carson (1962) in her goundbreaking
book, Silent Spring, was a potent factor in
mobilizing public opinion to become better
stewards of our environment. Scientifically, the
highly developed vocal communication of this
group, sharing as it does a number of parallels
with speech, has made songbirds especially suited
for the study of complex, learned vocal
communication at every level, from its molecular
biology to its evolution.
This chapter is about how birdsong is
produced. How do birds generate such a variety
of sounds and what vocal gymnastics are required
to produce them?The vocal system is the interface
between the bird's brain and his song. Knowing
how it functions, and coming to appreciate its
limitations as well as its capabilities, is important
in understanding vocal communication. Birdsong
is the product of the carefully coordinated activity
of many different muscles. Some of these are
associated with the vocal organ but many belong
to other muscle groups such as the respiratory
system and the vocal tract. Together they convert
nerve impulses from the brain into song.
Less is known about the vocal mechanisms of
non-songbirds, and they will be mentioned only
briefly to indicate the variety of avian vocal
systems, and to provide a perspective from which
to focus on the oscine songbirds. Readers
interested in additional information on songbirds
should consult reviews of this subject (Nowicki
& Marler 1988; Suthers 1997, 1999a, b; Gaunt
& Nowicki 1998; Doupe 81Kuhl 1999; Suthers
ef ale 1999; Goller & Larsen 2002)-
AN ORGAN FOR SINGING
The avian vocal organ, the syrinx, is located
deep in the chest in an air sac that is connected
to other air sacs and the lungs. It is present in all
birds except vultures, but its exact location and
structure varies considerably (King 1989). In
some species, including doves, pigeons and
parrots, the syrinx consists of modified rings of
cartilage near the lower end of trachea
(Fig. 9.1).
In other species, such as penguins, many owls,
nightjars, and some cuckoos, the syrinx
exists as two semisyrinxes located in the primary
bronchi (Fig. 9.1). In many birds, however,
including but by no means limited to the oscine
songbirds, the syrinx is at the junction where
the two primary bronchi join to form the trachea.
This tracheobronchial syrinx includes modified
cartilages from both the upper end of each
bronchus and the lower end of the trachea
(Fig. 9.1).
Suthers, Roderick (2004) How birds sing and why it matters. In Nature's Music: the
science of birdsong. Marler, P. and Slabbekoorn, H. eds. Elsevier. New York. Chapter 9
pp. 272-295.
How birds sing and why it matters 273
Tracheal syrinx ,
Variation in Syringeal Structure
Bronchial syrinx Tracheobronchial syrinx
(Oilbird) (Canary)
'I
1 I
!
I
I
I
1
Figure 9.1 Examples of variation in syringeal anatomy. The tracheal parrot syrinx has two syringeal
muscles and a pair of lateral tympaniform membranes. The bronchial syrinx of the oilbird has one pair of
syringeal muscles and a pair of medial and lateral tympaniform membranes in each bronchus. Songbirds
have several pairs of syringeal muscles in their tracheobronchial syrinx. Tr, trachea; ST, sternotrachealis
muscle; SY SUP, superficial syringeal muscle; SY PROF, deep syringeal muscle; SY VALV, pneumatic
valve; LTM, lateral tympaniform me~brane;MTM, medial tympaniform membrane; BCI, first bronchial
cartilage; SY, syringeal muscle; ML, medial labium; LL, lateral labium; SYR, muscles of syrinx. (Oilbird
modified after Suthers & Hector 1985; parrot and canary modified after King 1989).
The Respiratory Rhythm direction through the syrinx and into the
The timing of vocalization and the tempo of esophagus. After each coo there is a short
song begins with the respiratory cycle (Vicario expiration followed by a brief inspiration (Gaunt
1991). Expiratory muscles compress the bellows- et al. 1982).
like air sacs in the thorax and abdomen, increasing
respiratory pressure so air flows out through the
syrinx and trachea (Hartley 1990). During song, The Source of Sound
a small inspiration called a 'mini-breath' (Calder It is now generally agreed that vocalizations are
1970) is usually taken after each syllable, except generated by airflow-induced oscillation of
at very high syllable repetition rates (Hartley & elements in the wall of the syrinx that convert
Suthers 1989). Inspiratory muscles expand the some of the air's kinetic energy into acoustic
thorax and abdomen, reducing pressure in the energy. These oscillations are presumably
air sacs and reversing the direction of airflow sustained by interaction between Bernoulli forces
through the syrinx (Wild et al. 1998). and the inertia of air in the vocal tract, in much
Vocalization occurs during expiratory airflow, the same way as are oscillations of human vocal
aside from rare exceptions when it occurs during folds (Titze 1994; Gardner et al. 200 1). An
inspiration, as in the 'wah' sounds of doves (Gaunt alternative mechanism for sound production
I et al. 1982), and certain syllables in the songs of based on the principle of an aerodynamic whistle
some zebra finches (Goller & Daley 2001). Most was put forth (Gaunt et al. 1982) to explain
birds sing with their beaks open, but doves and pure tone vocalizations in birds such as doves.
pigeons coo with their beaks and nares closed. However, this aerodynamic whistle hypothesis
Sound is produced during airflow in an expiratory is not supported by the observation of
248
274 Nature's Music
tympaniform membrane vibration in pigeons
(Goller & Larsen 1997a), or by experiments with
collared doves vocalizing in light gas mixture
(Ballintijn & ten Cate 1998). In both collared
doves and ringdoves, there is evidence that the
sound generated in their syrinx contains
prominent harmonics, which must be filtered
out by the vocal tract to produce the tonal
properties of the coo (Beckers et al. 2003b).
The anatomical structures in the syrinx that
produce the sound differ with the syringeal
anatomy. Goller and Larsen used a fiberoptic
endoscope to make direct observations of
syringeal motion during vocalizations elicited
by stimulating vocal control regions in the brain
of anesthetized birds (Box 37, p. 277). In the
tracheal syrinx of pigeons they showed that the
lateral tympaniform membranes (LTMs), on each
side of the trachea bulge toward each other into
the tracheal lumen during vocalization where
they form a slit-like aperture and vibrate during
respiratory airflow (Goller & Larsen, 1997a, b;
Larsen & Goller 1999). The LTMs (Fig. 9.1)
also can be seen to fold into the tracheal lumen
and vibrate during vocalization in the tracheal
syrinx of the cockatiel (Larsen & Goller 1999,
2002).
In songbirds, the relatively well-developed
medial tympaniform membrane (MTM) at the
cranial end of each bronchus was long assumed
to be the source of sound ( ~ i s k i m e n1951;
Greenewalt 1968; Fletcher 1988; Gaunt 1988;
Figs. 9.1 & 9.2). However, endoscopic
observations of the songbird syrinx during
vocalization reveal that it is the connective tissue
forming the internal and external labia at the
anterior end of each bronchus that is adducted
into the syringeal lumen and vibrates during
vocalization (Larsen & Goller 1999). Surgical
removal of both MTMs has only a minor effect
on the vocalization, demonstrating that whatever
u
their function, they are not essential for song
production.
Syringeal Muscles Have Separate Functions
The muscular apparatus capable of acting on
the syrinx to control vocalization receives input
from.the brain by the hypoglossal nerve. In sohe
species, such as swiftlets, there are no intrinsic
syringeal muscles, meaning muscles having both
their origin and insertion on the syrinx (Box
36, p. 275). In these birds, syringeal function
during vocalization is controlled by muscles that
have at least one of their attachments outside
the syrinx. Other birds, such as parrots, have
one or more pairs of specialized syringeal muscles.
Doves have two pairs of extrinsic syringeal
muscles (see Fig. 10.5, p. 310). One, the
sternotrachealis, reduces the tension across the
syrinx and allows the lateral tympaniform
membranes to fold into the tracheal lumen where
airflow causes them to vibrate. Contraction of
the other muscle, the tracheolateralis, abducts
the tympaniform membranes out of the lumen
and terminates phonation (Goller & Larsen
1997a). The role of these muscles, if any, in
frequency modulation of coos is uncertain. Gaunt
(1988) reported that during a coo both muscles
are simultaneously active in ring doves, and
Beckers et al. (2003~)demonstrated that gradual
frequency modulation in the coo, but not abrupt
frequency jumps, closely parallels pressure changes
in the interclavicularair sac, suggesting expiratory
muscles may be important regulators of sound
frequency. Amplitude modulation of coos is
associated with interruptions of tracheal airflow
presumably due to the;losing and opening of a
valve, perhaps formed by the adducted
tympaniform membranes or the glottis (Beckers
et al. 2 0 0 3 ~ ) .
Songbirds have the most complex syiingeal
musculature of all. They have 5 pairs of muscles
that attach to the syrinx and a sikth pair that
acts on it indirectly via its insertion on the trachea
(King 1989). In general, more muscles provide
morgdegrees of freedom in configuring the syrinx
to produce different sounds (Gaunt 1983). The
complexity of song is roughly correlated with
the number of syringeal muscles, but this
relationship is not a simple one. Parrots, for
example, succeed in producing complex
vocalizations with only two pairs of syringeal
muscles (Fig. 9.1A) and the ~Lstralianlyreb&d,
249
How birds sing and why it matters 275
I
L
( BOX 36
I AVIAN SONAR: CLICKING IN THE DARK I
Two groups of birds, some Aerodramus Asian swiftlets (Medway 1967; Griffin & Suthers 1970; Griffin & Thompson 1982; Coles
et al. 1987) and the neotropical oilbird (Griffin 1953; Konishi & Knudsen 1979) make clicking sounds with their syrinx that enable
them to navigate in the dark by echolocation. Although swiftlets and oilbirds have quite different lifestyles, their sonar allows them
to nest in the relative safety of dark caves. Swiftlets are diurnal insectivores and spend the day on the wing catching insects which
they locate visually. Oilbirds are nocturnal frugivores with excellent night-time vision. They spend the daytime resting on ledges
in their nesting cave and use vision, supplemented by echolocation on dark nights, to feed on the fruits of tropical trees.
Clicks make good sonar signals because they contain energy over a wide range of frequencies and have a short duration,
making it easier to hear faint echoes reflected by surrounding objects (CD2 #64). Furthermore, clicks can be produced by an
anatomically simple syrinx. Although details of click production differ between oilbirds (Suthers & Hector 1985) and swiftlets
(Suthers & Hector 1982), in both it involves rapid, synchronous opening and closing of both sides of the syrinx. Each cycle
typically produces a pair of clicks (A). The first click occurs when the syringeal valve, located at the labia and tympaniform
membranes, briefly vibrates as it narrows the syringeal lumen just before closing it completely. After a short, silent interval while
the valves are closed, a second click occurs as the valve begins to open, again allowing airflow to vibrate the membranes until
they are withdrawn from the syringeal air stream.
Swiftlets have no intrinsic syringeal muscles so they operate their syringeal valves by using extrinsic muscles in sequence
to first close the valve by pulling the tracheobronchial syrinx towards'the bronchi (accomplishedby the sternotrachealis muscles)
and then opening the valve by stretching bronchi (through contraction of the tracheolateralis muscles; Suthers & Hector 1982).
The mechanism is similar in oilbirds with two important differences. Oilbirds have a bronchial syrinx (Muller 1841; Garrod 1873;
see p. 273; Fig. 9.1), in which each syringeal%alve (i.e. semisyrinx) is in the primary bronchus and the valve is opened by a
syringeal muscle that lies along the surface of each bronchus and attaches to a bronchial cartilage bordering the lateral tympanic
membrane (Suthers & Hector 1985). One of the beauties of this simple mechanism for producing sonar clicks is that two sonar
pulses are generated during each cycle of muscle contraction, thus doubling the number of echoes produced.
Why do some birds have a bronchial syrinx? Oilbirds may provide a clue. A puzzling feature of their syrinx is that the two
semisyrinxes are placed at different positions along the bronchus, with the distance from the trachea varying on the right and left
sides, and between individuals. By inserting a plug in either the left or right bronchus to seal it off from the trachea, we showed
that each bronchus had a quarter wave resonance, like that of a stopped tube, allowing frequencies with a wavelength 4 times
the length of the bronchial tube to pass while filtering out other frequencies (Suthers 1994; 6). The distribution of sound energy
as a function of frequency thus depends on the length of the section of rigid bronchus between the trachea and semisyrinx.
Because this length differs between the left and right bronchus and varies between individuals, the vocalizations of each oilbird
contain a pair of frequency bands of high intensity sound, or formants that are unique to that individual. Detectable in some
echolocating clicks (Suthers & Hector 1988), these formants are prominent in many of the bird's long duration broadband 'social'
vocalizations, and might provide a way for individuals to recognize each other in the dark (CD2 #65). Interestingly, most bronchial
syrinxes occur in nocturnal species (Suthers 1994).
Roderick A. Suthers
(A) Production of Echolocation Clicks by Swiftlets (B) Oilbird Vocal Tract Asymmetry gives.Individually
Unique Formants
i : TL
16
8
0
RB = right bronchus formant
4 Click
LB = left bronchus formant
T = tracheal formant '
3 Closed' 16
T
0.5 s
274 Nature's Music
tympaniform membrane vibration in pigeons
(Goller & Larsen 1997a), or by experiments with
collared doves vocalizing in light gas mixture
(Ballintijn & ten Cate 1998). In both collared
doves and ringdoves, there is evidence that the
sound generated in their syrinx contains
prominent harmonics, which must be filtered
out by the vocal tract to produce the tonal
properties of the coo (Beckers et al. 2003b).
The anatomical structures in the syrinx that
produce the sound differ with the syringeal
anatomy. Goller and Larsen used a fiberoptic
endoscope to make direct observations of
syringeal motion during vocalizations elicited
by stimulating vocal control regions in the brain
of anesthetized birds (Box 37, p. 277). In the
tracheal syrinx of pigeons they showed that the
lateral tympaniform membranes (LTMs), on each
side of the trachea bulge toward each other into
the tracheal lumen during vocalization where
they form a slit-like aperture and vibrate during
respiratory airflow (Goller & Larsen, 1997a, b;
Larsen & Goller 1999). The LTMs (Fig. 9.1)
also can be seen to fold into the tracheal lumen
and vibrate during vocalization in the tracheal
syrinx of the cockatiel (Larsen & Goller 1999,
2002).
In songbirds, the relatively well-developed
medial tympaniform membrane (MTM) at the
cranial end of each bronchus was long assumed
to be the source of sound ( ~ i s k i m e n1951;
Greenewalt 1968; Fletcher 1988; Gaunt 1988;
Figs. 9.1 & 9.2). However, endoscopic
observations of the songbird syrinx during
vocalization reveal that it is the connective tissue
forming the internal and external labia at the
anterior end of each bronchus that is adducted
into the syringeal lumen and vibrates during
vocalization (Larsen & Goller 1999). Surgical
removal of both MTMs has only a minor effect
on the vocalization, demonstrating that whatever
u
their function, they are not essential for song
production.
Syringeal Muscles Have Separate Functions
The muscular apparatus capable of acting on
the syrinx to control vocalization receives input
from.the brain by the hypoglossal nerve. In sohe
species, such as swiftlets, there are no intrinsic
syringeal muscles, meaning muscles having both
their origin and insertion on the syrinx (Box
36, p. 275). In these birds, syringeal function
during vocalization is controlled by muscles that
have at least one of their attachments outside
the syrinx. Other birds, such as parrots, have
one or more pairs of specialized syringeal muscles.
Doves have two pairs of extrinsic syringeal
muscles (see Fig. 10.5, p. 310). One, the
sternotrachealis, reduces the tension across the
syrinx and allows the lateral tympaniform
membranes to fold into the tracheal lumen where
airflow causes them to vibrate. Contraction of
the other muscle, the tracheolateralis, abducts
the tympaniform membranes out of the lumen
and terminates phonation (Goller & Larsen
1997a). The role of these muscles, if any, in
frequency modulation of coos is uncertain. Gaunt
(1988) reported that during a coo both muscles
are simultaneously active in ring doves, and
Beckers et al. (2003~)demonstrated that gradual
frequency modulation in the coo, but not abrupt
frequency jumps, closely parallels pressure changes
in the interclavicularair sac, suggesting expiratory
muscles may be important regulators of sound
frequency. Amplitude modulation of coos is
associated with interruptions of tracheal airflow
presumably due to the;losing and opening of a
valve, perhaps formed by the adducted
tympaniform membranes or the glottis (Beckers
et al. 2 0 0 3 ~ ) .
Songbirds have the most complex syiingeal
musculature of all. They have 5 pairs of muscles
that attach to the syrinx and a sikth pair that
acts on it indirectly via its insertion on the trachea
(King 1989). In general, more muscles provide
morgdegrees of freedom in configuring the syrinx
to produce different sounds (Gaunt 1983). The
complexity of song is roughly correlated with
the number of syringeal muscles, but this
relationship is not a simple one. Parrots, for
example, succeed in producing complex
vocalizations with only two pairs of syringeal
muscles (Fig. 9.1A) and the ~Lstralianlyreb&d,
How birds sing and why it matters 275
I
L
( BOX 36
I AVIAN SONAR: CLICKING IN THE DARK I
Two groups of birds, some Aerodramus Asian swiftlets (Medway 1967; Griffin & Suthers 1970; Griffin & Thompson 1982; Coles
et al. 1987) and the neotropical oilbird (Griffin 1953; Konishi & Knudsen 1979) make clicking sounds with their syrinx that enable
them to navigate in the dark by echolocation. Although swiftlets and oilbirds have quite different lifestyles, their sonar allows them
to nest in the relative safety of dark caves. Swiftlets are diurnal insectivores and spend the day on the wing catching insects which
they locate visually. Oilbirds are nocturnal frugivores with excellent night-time vision. They spend the daytime resting on ledges
in their nesting cave and use vision, supplemented by echolocation on dark nights, to feed on the fruits of tropical trees.
Clicks make good sonar signals because they contain energy over a wide range of frequencies and have a short duration,
making it easier to hear faint echoes reflected by surrounding objects (CD2 #64). Furthermore, clicks can be produced by an
anatomically simple syrinx. Although details of click production differ between oilbirds (Suthers & Hector 1985) and swiftlets
(Suthers & Hector 1982), in both it involves rapid, synchronous opening and closing of both sides of the syrinx. Each cycle
typically produces a pair of clicks (A). The first click occurs when the syringeal valve, located at the labia and tympaniform
membranes, briefly vibrates as it narrows the syringeal lumen just before closing it completely. After a short, silent interval while
the valves are closed, a second click occurs as the valve begins to open, again allowing airflow to vibrate the membranes until
they are withdrawn from the syringeal air stream.
Swiftlets have no intrinsic syringeal muscles so they operate their syringeal valves by using extrinsic muscles in sequence
to first close the valve by pulling the tracheobronchial syrinx towards'the bronchi (accomplishedby the sternotrachealis muscles)
and then opening the valve by stretching bronchi (through contraction of the tracheolateralis muscles; Suthers & Hector 1982).
The mechanism is similar in oilbirds with two important differences. Oilbirds have a bronchial syrinx (Muller 1841; Garrod 1873;
see p. 273; Fig. 9.1), in which each syringeal%alve (i.e. semisyrinx) is in the primary bronchus and the valve is opened by a
syringeal muscle that lies along the surface of each bronchus and attaches to a bronchial cartilage bordering the lateral tympanic
membrane (Suthers & Hector 1985). One of the beauties of this simple mechanism for producing sonar clicks is that two sonar
pulses are generated during each cycle of muscle contraction, thus doubling the number of echoes produced.
Why do some birds have a bronchial syrinx? Oilbirds may provide a clue. A puzzling feature of their syrinx is that the two
semisyrinxes are placed at different positions along the bronchus, with the distance from the trachea varying on the right and left
sides, and between individuals. By inserting a plug in either the left or right bronchus to seal it off from the trachea, we showed
that each bronchus had a quarter wave resonance, like that of a stopped tube, allowing frequencies with a wavelength 4 times
the length of the bronchial tube to pass while filtering out other frequencies (Suthers 1994; 6). The distribution of sound energy
as a function of frequency thus depends on the length of the section of rigid bronchus between the trachea and semisyrinx.
Because this length differs between the left and right bronchus and varies between individuals, the vocalizations of each oilbird
contain a pair of frequency bands of high intensity sound, or formants that are unique to that individual. Detectable in some
echolocating clicks (Suthers & Hector 1988), these formants are prominent in many of the bird's long duration broadband 'social'
vocalizations, and might provide a way for individuals to recognize each other in the dark (CD2 #65). Interestingly, most bronchial
syrinxes occur in nocturnal species (Suthers 1994).
Roderick A. Suthers
(A) Production of Echolocation Clicks by Swiftlets (B) Oilbird Vocal Tract Asymmetry gives.Individually
Unique Formants
i : TL
16
8
0
RB = right bronchus formant
4 Click
LB = left bronchus formant
T = tracheal formant '
3 Closed' 16
T
0.5 s
250
276 Nature's Music
known for its vocal versatility and now generally
considered to be a songbird (Higgins et al. 200 l),
has but 3 pairs. The remainder of this chapter
focuses on the oscine songbirds.
The function of syringeal muscles in songbirds
has been studied by recording their electrical
activity in the form of electromyograms (EMGs),
while measuring the rate of airflow through each
side of the syrinx, and monitoring respiratory
pressure as the bird moves about freely in its
cage and sings spontaneously in front of a
microphone (Suthers 1990, 1997; Suthers et al.
1994, 1996, 1999; Goller & Suthers 1996a, b;
Box 37, p. 277). Additional insights into syringed
function have been obtained by using an
endoscope to make direct observations on the
motion of syringeal structures in response to
electrical stimulation of individual muscles in
anesthetized birds (Larsen & Goller 2002).
Alth~ugh~different muscles may have more than
one role in sound production and interact in
complex ways that are still poorly understood,
two of the most basic aspects of song, its timing
and its frequency, are primarily controlled by
the dorsal and ventral syringeal muscles,
respectively, acting in conjunction with the
muscles that mechanically ventilate the respiratory
system during breathing.
The dorsal muscles, the syringealis dorsalis
and the tracheobronchialis dorsalis, play a key
role in operating a pneumatic valve at the upper
end of each bronchus that controls the timing
of phonation on each side of the syrinx (Goller
& Suthers 1995b, 199613). Each valve is formed
by the medial and lateral labia that are also the
structures generating sound (Fig. 9.2). When
these dorsal muscles are relaxed during quiet
respiration, and during mini-breaths between
syllables, the labia are abducted out of the air
stream. In this position they do not oscillate
and the syringeal lumen is open, presenting
minimal resistance to airflow. Contraction of
the dorsal muscles initiates sound production
by adducting the labia into the lumen where
they form a slit and oscillate, as air flows across
them (Goller & Suthers 1995b; Goller & Larsen
1997b; Larsen & Goller 1999). Stronger
contraction of these muscles, perhaps assisted
by other muscles (Larsen & Goller 2002),
terminates or prevents phonation on the same
side of the syrinx by moving the labia still closer
together and closing the slit, stopping airflow
I RECORDING FROM THE SYRINX OF A SINGING BIRD
and arresting labial oscillation.
The large ventral syringeal muscles, the
syringealis ventralis, appear to be involved in
regulating sound frequency (Fig. 9.2). The
amplitude of their EMG is correlated in a positive
way with the fundamental frequency of the
vocalization (Goller & Suthers 1995b, 1996a).
They presumably control fundamental frequency
by varying the tension or elasticity of the
oscillating labia. Upward-sweeping frequency
modulation (FM) is accompanied by a
corresponding increase in ventral muscle activity,
which is also high during high-pitched notes
with a high fundamental frequency but little
FM.
There is great variety in the frequency
structure, or 'syllable morphology' of birdsongs.
Producing such a variety of sounds would seem
to require very complicated patterns of syringeal
muscle activity to generate a different vocal
gesture for each syllable type, although some
theoretical models of song production suggest
the vocal gestures required may be much simpler
than previously assumed (Gardner et al. 2001;
Laje et al. 2002; Suthers & Margoliash 2002).
Sensory Feedback: Fine-tuning the Song
Motor Program i
To achieve or maintain the optimum respiratory
pressure and the appropriate configuration of
the syrinx during song, the bird needs to monitor
these variables as it sings. Using sensory feedback
to achieve this, it could then adjust the ongoing
motor program being sent from the brain to the
muscles to correct any deviations from the
intended vocal gesture.-Such deviations might
occur as a result of changes in posture or ongoing
physical activity. We know that auditory feedback
is essential during song learning (Konishi 1965a),
and continues to play a role in the long-term
maintenance of adult crystallized song (Nordeen
251
How birds sing and why it matters 277
Much can be learned about song production by monitoring respiratory dynamics and the activity of syringeal and
I
respiratory muscles during song. The bird is anesthetized and a tiny thermistor bead is placed in the lumen of
each primary bronchus, just below the syrinx (Suthers 1990; Suthers et al. 1994). The thermistors are connected
to a circuit that maintains them at a constant temperature of about 60C. Air flowing through the bronchus during
either inspiration or expiration will conduct heat away from the thermistor and the additional current required to
maintain its temperature provides a measure of the rate of airflow through each side of the syrinx.
The respiratory pressure driving airflow through the syrinx is measured with a miniature piezoresistive pressure
transducer mounted on the bird's back andAattachedto a small tube inserted into one of the air sacs. The bird's
respiratory system is different from that of mammals. Birds have several air sacs that act like bellows pumping
air through the lungs. Air sacs are connected to the lungs and to each other by complex passages and openings.
A midline interclavicular air sac connects the two sides of the respiratory system and contains the syrinx, so
respiratory pressure should be the same in both bronchi. By measuring the respiratory pressure, inspirations and
expirations can be distinguished. The movement and pokition of the labia on each side of the syrinx can be
monitored by using bronchial airflow and respiratory pressure to calculate changes in syringeal resistance. A
positive air sac pressure with no airflow indicates the labial valve on that side of the syrinx is closed, whereas a
high flow rate with low pressure indicates the labia are drawn back out of the air stream.
The function of various syringeal and respiratory muscles can be deduced by inserting a pair of very fine hair-
like stainless steel wires into them and recording the relationship of their electromyograms (EMG) while monitoring
syringeal airflow, respiratory pressure and the acoustic properties of the song at a microphone in front of the bird
(Goller & Suthers 1996b). All of these physiological signals are carried by fine wires to a velco tab on the bird's
back where they connect to other wires that travel up through the top of the cage to recording instruments. Birds
instrumented in this way are free to move around in their cages and sing spontaneously. At the end of an
experiment, which may last a week or more, the bird is again anesthetized, the apparatus is removed and he is
returned to the aviary.
Observing syringeal motion directly, Goller and Larsen did some elegant experiments using an angiofiberscope
to view the mechanical action of the syrinx in response to electrical stimulation of the song control system in the
brain of anesthetized birds (Goller & Larsen 1997b; Larsen & Goller 2002). A 1.4 mm diameter optical fiber was
inserted, either down the trachea to view labial movements, or into the air sac containing the syrinx to see the
muscles on its external surface. In some experiments they electrically stimulated individual muscles, and in
others (Larsen & Goller 1999) they used a laser optical system to detect the vibration of sound-producing
structures. Although the vocalizations produced by this method are not normal song, theyeprovide valuable
information about how sound is produced.
Roderick A. Suthers
Location of the transducers in the respiratory system Measuring Syringeal Function
used to study song production. FR, thermistor During Song
measuring airflow through the right side of the syrinx;
FL, thermistor measuring airflow through the left Inspiratory muscles
side of the syrinx; P, cannula in the air sac to measure
respiratory pressure. EMG electrodes may be placed
in syringeal, thoracic inspiratory andlor abdominal
expiratory muscles. lnspiratory muscles enlarge the
chest, reducing air sac pressure and drawing air
in. Expiratory muscles compress the chest and force
air out through the syrinx.
276 Nature's Music
known for its vocal versatility and now generally
considered to be a songbird (Higgins et al. 200 l),
has but 3 pairs. The remainder of this chapter
focuses on the oscine songbirds.
The function of syringeal muscles in songbirds
has been studied by recording their electrical
activity in the form of electromyograms (EMGs),
while measuring the rate of airflow through each
side of the syrinx, and monitoring respiratory
pressure as the bird moves about freely in its
cage and sings spontaneously in front of a
microphone (Suthers 1990, 1997; Suthers et al.
1994, 1996, 1999; Goller & Suthers 1996a, b;
Box 37, p. 277). Additional insights into syringed
function have been obtained by using an
endoscope to make direct observations on the
motion of syringeal structures in response to
electrical stimulation of individual muscles in
anesthetized birds (Larsen & Goller 2002).
Alth~ugh~different muscles may have more than
one role in sound production and interact in
complex ways that are still poorly understood,
two of the most basic aspects of song, its timing
and its frequency, are primarily controlled by
the dorsal and ventral syringeal muscles,
respectively, acting in conjunction with the
muscles that mechanically ventilate the respiratory
system during breathing.
The dorsal muscles, the syringealis dorsalis
and the tracheobronchialis dorsalis, play a key
role in operating a pneumatic valve at the upper
end of each bronchus that controls the timing
of phonation on each side of the syrinx (Goller
& Suthers 1995b, 199613). Each valve is formed
by the medial and lateral labia that are also the
structures generating sound (Fig. 9.2). When
these dorsal muscles are relaxed during quiet
respiration, and during mini-breaths between
syllables, the labia are abducted out of the air
stream. In this position they do not oscillate
and the syringeal lumen is open, presenting
minimal resistance to airflow. Contraction of
the dorsal muscles initiates sound production
by adducting the labia into the lumen where
they form a slit and oscillate, as air flows across
them (Goller & Suthers 1995b; Goller & Larsen
1997b; Larsen & Goller 1999). Stronger
contraction of these muscles, perhaps assisted
by other muscles (Larsen & Goller 2002),
terminates or prevents phonation on the same
side of the syrinx by moving the labia still closer
together and closing the slit, stopping airflow
I RECORDING FROM THE SYRINX OF A SINGING BIRD
and arresting labial oscillation.
The large ventral syringeal muscles, the
syringealis ventralis, appear to be involved in
regulating sound frequency (Fig. 9.2). The
amplitude of their EMG is correlated in a positive
way with the fundamental frequency of the
vocalization (Goller & Suthers 1995b, 1996a).
They presumably control fundamental frequency
by varying the tension or elasticity of the
oscillating labia. Upward-sweeping frequency
modulation (FM) is accompanied by a
corresponding increase in ventral muscle activity,
which is also high during high-pitched notes
with a high fundamental frequency but little
FM.
There is great variety in the frequency
structure, or 'syllable morphology' of birdsongs.
Producing such a variety of sounds would seem
to require very complicated patterns of syringeal
muscle activity to generate a different vocal
gesture for each syllable type, although some
theoretical models of song production suggest
the vocal gestures required may be much simpler
than previously assumed (Gardner et al. 2001;
Laje et al. 2002; Suthers & Margoliash 2002).
Sensory Feedback: Fine-tuning the Song
Motor Program i
To achieve or maintain the optimum respiratory
pressure and the appropriate configuration of
the syrinx during song, the bird needs to monitor
these variables as it sings. Using sensory feedback
to achieve this, it could then adjust the ongoing
motor program being sent from the brain to the
muscles to correct any deviations from the
intended vocal gesture.-Such deviations might
occur as a result of changes in posture or ongoing
physical activity. We know that auditory feedback
is essential during song learning (Konishi 1965a),
and continues to play a role in the long-term
maintenance of adult crystallized song (Nordeen
How birds sing and why it matters 277
Much can be learned about song production by monitoring respiratory dynamics and the activity of syringeal and
I
respiratory muscles during song. The bird is anesthetized and a tiny thermistor bead is placed in the lumen of
each primary bronchus, just below the syrinx (Suthers 1990; Suthers et al. 1994). The thermistors are connected
to a circuit that maintains them at a constant temperature of about 60C. Air flowing through the bronchus during
either inspiration or expiration will conduct heat away from the thermistor and the additional current required to
maintain its temperature provides a measure of the rate of airflow through each side of the syrinx.
The respiratory pressure driving airflow through the syrinx is measured with a miniature piezoresistive pressure
transducer mounted on the bird's back andAattachedto a small tube inserted into one of the air sacs. The bird's
respiratory system is different from that of mammals. Birds have several air sacs that act like bellows pumping
air through the lungs. Air sacs are connected to the lungs and to each other by complex passages and openings.
A midline interclavicular air sac connects the two sides of the respiratory system and contains the syrinx, so
respiratory pressure should be the same in both bronchi. By measuring the respiratory pressure, inspirations and
expirations can be distinguished. The movement and pokition of the labia on each side of the syrinx can be
monitored by using bronchial airflow and respiratory pressure to calculate changes in syringeal resistance. A
positive air sac pressure with no airflow indicates the labial valve on that side of the syrinx is closed, whereas a
high flow rate with low pressure indicates the labia are drawn back out of the air stream.
The function of various syringeal and respiratory muscles can be deduced by inserting a pair of very fine hair-
like stainless steel wires into them and recording the relationship of their electromyograms (EMG) while monitoring
syringeal airflow, respiratory pressure and the acoustic properties of the song at a microphone in front of the bird
(Goller & Suthers 1996b). All of these physiological signals are carried by fine wires to a velco tab on the bird's
back where they connect to other wires that travel up through the top of the cage to recording instruments. Birds
instrumented in this way are free to move around in their cages and sing spontaneously. At the end of an
experiment, which may last a week or more, the bird is again anesthetized, the apparatus is removed and he is
returned to the aviary.
Observing syringeal motion directly, Goller and Larsen did some elegant experiments using an angiofiberscope
to view the mechanical action of the syrinx in response to electrical stimulation of the song control system in the
brain of anesthetized birds (Goller & Larsen 1997b; Larsen & Goller 2002). A 1.4 mm diameter optical fiber was
inserted, either down the trachea to view labial movements, or into the air sac containing the syrinx to see the
muscles on its external surface. In some experiments they electrically stimulated individual muscles, and in
others (Larsen & Goller 1999) they used a laser optical system to detect the vibration of sound-producing
structures. Although the vocalizations produced by this method are not normal song, theyeprovide valuable
information about how sound is produced.
Roderick A. Suthers
Location of the transducers in the respiratory system Measuring Syringeal Function
used to study song production. FR, thermistor During Song
measuring airflow through the right side of the syrinx;
FL, thermistor measuring airflow through the left Inspiratory muscles
side of the syrinx; P, cannula in the air sac to measure
respiratory pressure. EMG electrodes may be placed
in syringeal, thoracic inspiratory andlor abdominal
expiratory muscles. lnspiratory muscles enlarge the
chest, reducing air sac pressure and drawing air
in. Expiratory muscles compress the chest and force
air out through the syrinx.
252
 II How birds sing and why it matters 279
I
Songbird Syrinx

Trachea
I SONG IS DISRUPTED BY DELAYED FEEDBACK AND THEN RECOVERS

Songbirds use auditory feedback to both leain and maintain their songs, and the dynamics of this process may
be investigated by manipulating the auditory feedback heard by singing birds. Deafening adult zebra finches
results in a complete loss of auditory feedback and, after roughly four weeks, a marked degradation in song
structure (Nordeen & Nordeen 1992). However, it is difficult, though not impossible (Woolley & Rubel 2002), to
1 restore auditory feedback to deafened birds. In contrast, by using a computer-controlled system to continuously
monitor a bird's vocalizations, specific types of artificial auditory feedback may be generated in real-time, played
back to the bird as he sings, and started and stopped at any moment (Leonardo & Konishi 1999). If the artificial
feedback signal is a 100 ms delay of the bird's song (ABCD) as he sings it, then when the bird sings syllable B
he will hear B+A, with the normal feedback for B superimposed with the delayed feedback for A. Two to four
weeks of exposure to delayed auditory feedback caused a dramatic loss of the spectral and temporal stereotypy
seen in crystallized song, consisting of stuttering, deletion, and distortion of song syllables and the creation of
new ones (CD2 #66). Stuttering sometimes increased the length of a song bout of 2-6 s to more than 60 s.
Restoration of normal feedback enabled the recovery of the original songs of each bird. All of the changes in song
structure gradually became more infrequent, and were eventually replaced by the temporal and spectral organization
characteristic of the original song. A complete recovery took 2-4 months. Thus adult zebra finches appear to
retain a great deal of potential song plasticity, even though they do not normally modify their songs after
I
Bronchus crystallization. Furthermore, despite destabilization of the behavior, a memory of the original song persists, and
can be used to correct the vocal output upon restoration of normal auditory feedback. This indicates that the song
Figure 9.2 A songbird syrinx showing its bipartite structure with 2 sound generators and multiple pairs of is maintained by an active control process which requires both a memorized song model and the auditory
muscles. On the left is a frontal section through a brown thrasher syrinx showing the location of microbead feedback generated during singing. Elucidating the neural mechanisms underlying this feedback control process
thermistors used to sense airflow through each side. The medial and lateral labia form pneumatic valves remains an area of vigorous research.
at the cranial end of each bronchus and oscillate to produce sound. On the right is a ventrolateral view of
the same syrinx depicting the syringeal muscles. vS, ventral syringeal muscle; vTB, ventral tracheobronchialis
muscle; dS, dorsal syringeal muscle; dTB dorsal tracheobronchialis muscle. For other abbreviations see
Figure 9.1. (Modified after Goller & Suthers 1996a).
Normal Song
& Nordeen 1992; Okanoya & Yamaguchi 1997; Similar recordings from syringed muscles indicate
Leonardo & Konishi 1999; Box 38, p. 279). that they also adjust their activity to compensate
Non-auditory feedback from mechano- for unexpected perturbations in subsyringeal
receptors or proprioceptors that may sense the pressure during phonation (Suthers & Wild
tension can provide a rapid response to correct 2000). The sensory receptors that provide the
a parameter that is too high or too low. Air feedback underlying these responses have not
pressure or the relative position of structures in been identified, but likely include
the respiratory and vocal system are also mechanoreceptors or proprioceptors in the
important, even in adult crystallized song. For
example, if the respiratory pressure is momentarily
muscles, air sacs or walls of the respiratory and
vocal systems.
Decrystallized Song with Dropped Syllables and Stuttering I
increased experimentally during a syllable by
injecting a small, randomly timed puff of air Lateral independence of Motor Control
into the air sac of a singing cardinal, it elicits a
compensatory partial relaxation of the abdominal Each side of the syrinx is capable of acting
expiratory muscles that returns the pressure to independently, since it is innervated separately
near its original level (Suthers et al. 2002). Since by the tracheosyringeal branch of the hypoglossal
this muscle response does not require extensive nerve on the same side, which in turn receives

processing at higher levels of the brain, it can
occur rapidly - often within the same syllable
during which the pressure perturbation occurred.
I I
0 Time (s) 4
most of its input from the same side of the brain
(Nottebohm & Nottebohm 1976; Nottebohm
1977; Vicario & Nottebohm 1988; Wild et al.

253
 II How birds sing and why it matters 279
I
Songbird Syrinx

Trachea
I SONG IS DISRUPTED BY DELAYED FEEDBACK AND THEN RECOVERS

Songbirds use auditory feedback to both leain and maintain their songs, and the dynamics of this process may
be investigated by manipulating the auditory feedback heard by singing birds. Deafening adult zebra finches
results in a complete loss of auditory feedback and, after roughly four weeks, a marked degradation in song
structure (Nordeen & Nordeen 1992). However, it is difficult, though not impossible (Woolley & Rubel 2002), to
1 restore auditory feedback to deafened birds. In contrast, by using a computer-controlled system to continuously
monitor a bird's vocalizations, specific types of artificial auditory feedback may be generated in real-time, played
back to the bird as he sings, and started and stopped at any moment (Leonardo & Konishi 1999). If the artificial
feedback signal is a 100 ms delay of the bird's song (ABCD) as he sings it, then when the bird sings syllable B
he will hear B+A, with the normal feedback for B superimposed with the delayed feedback for A. Two to four
weeks of exposure to delayed auditory feedback caused a dramatic loss of the spectral and temporal stereotypy
seen in crystallized song, consisting of stuttering, deletion, and distortion of song syllables and the creation of
new ones (CD2 #66). Stuttering sometimes increased the length of a song bout of 2-6 s to more than 60 s.
Restoration of normal feedback enabled the recovery of the original songs of each bird. All of the changes in song
structure gradually became more infrequent, and were eventually replaced by the temporal and spectral organization
characteristic of the original song. A complete recovery took 2-4 months. Thus adult zebra finches appear to
retain a great deal of potential song plasticity, even though they do not normally modify their songs after
I
Bronchus crystallization. Furthermore, despite destabilization of the behavior, a memory of the original song persists, and
can be used to correct the vocal output upon restoration of normal auditory feedback. This indicates that the song
Figure 9.2 A songbird syrinx showing its bipartite structure with 2 sound generators and multiple pairs of is maintained by an active control process which requires both a memorized song model and the auditory
muscles. On the left is a frontal section through a brown thrasher syrinx showing the location of microbead feedback generated during singing. Elucidating the neural mechanisms underlying this feedback control process
thermistors used to sense airflow through each side. The medial and lateral labia form pneumatic valves remains an area of vigorous research.
at the cranial end of each bronchus and oscillate to produce sound. On the right is a ventrolateral view of
the same syrinx depicting the syringeal muscles. vS, ventral syringeal muscle; vTB, ventral tracheobronchialis
muscle; dS, dorsal syringeal muscle; dTB dorsal tracheobronchialis muscle. For other abbreviations see
Figure 9.1. (Modified after Goller & Suthers 1996a).
Normal Song
& Nordeen 1992; Okanoya & Yamaguchi 1997; Similar recordings from syringed muscles indicate
Leonardo & Konishi 1999; Box 38, p. 279). that they also adjust their activity to compensate
Non-auditory feedback from mechano- for unexpected perturbations in subsyringeal
receptors or proprioceptors that may sense the pressure during phonation (Suthers & Wild
tension can provide a rapid response to correct 2000). The sensory receptors that provide the
a parameter that is too high or too low. Air feedback underlying these responses have not
pressure or the relative position of structures in been identified, but likely include
the respiratory and vocal system are also mechanoreceptors or proprioceptors in the
important, even in adult crystallized song. For
example, if the respiratory pressure is momentarily
muscles, air sacs or walls of the respiratory and
vocal systems.
Decrystallized Song with Dropped Syllables and Stuttering I
increased experimentally during a syllable by
injecting a small, randomly timed puff of air Lateral independence of Motor Control
into the air sac of a singing cardinal, it elicits a
compensatory partial relaxation of the abdominal Each side of the syrinx is capable of acting
expiratory muscles that returns the pressure to independently, since it is innervated separately
near its original level (Suthers et al. 2002). Since by the tracheosyringeal branch of the hypoglossal
this muscle response does not require extensive nerve on the same side, which in turn receives

processing at higher levels of the brain, it can
occur rapidly - often within the same syllable
during which the pressure perturbation occurred.
I I
0 Time (s) 4
most of its input from the same side of the brain
(Nottebohm & Nottebohm 1976; Nottebohm
1977; Vicario & Nottebohm 1988; Wild et al.

254
280 Nature's Music
2000). The implications of this independence
of right and left sides of the syrinx for vocal
communication are far reaching, for it gives
son g birds the possibilit y of two, virtually
independent sound sources. Although both right
and left sides are subjected to similar respiratory
pressures generated by the respiratory muscles
and both share the same supra-syringeal vocal
tract, the dorsal syringeal muscles can switch
sound production from one side of the syrinx to
the other very rapidly, silencing one side by closing
it while holding the labia on the opposite side
in a phonatory position. Likewise, the ventral
syringeal muscles can independently vary the
fundamental frequency generated on each side.
Songbirds have taken full advantage of this lateral
independence to greatly increase the variety and
complexity of their songs (Greenewalt 1968; Stein
1968; Suthers 1999a).
The Stereotypy of Adult Song
Most adult songbirds sing a stereotyped repertoire
in which reiterations of the same syllable type
vary little in their acoustic properties. This
acoustic stereotypy, together with the fact that
there is, at least sometimes, no immediate change
in the vocalization if auditory feedback is removed
by deafening, is part of the evidence suggesting
that the pattern of syringeal and respiratory
muscle activity required to produce a particular
syllable is stored in the brain as a 'motor program'
(konishi 1985; Vu et al. 1994). According to
this hypothesis, the syllables of crystallized song
might be produced by activating the appropriate
network of neurons in the brain; these are
presumably wired together in such a way that
they automatically generate the basic pattern of
coordinated muscle contractions, the vocal
gesture, needed to produce that syllable. Auditory
or somatosensory feedback may not be required,
at least in the short term, to produce the basic
vocal gesture, but it could adjust or modify it to
correct for errors that might otherwise occur if,
for example, conditions vary in the peripheral
vocal or respiratory system. Although central
pattern-generating circuits for different syllable
types have not yet been identified in songbirds,
they have an important role in controlling many
rhythmic behaviors in other animals (Getting
1989; Pearson 1993).
Whatever the neural mechanism underlying
syllable production, syllables rather than whole
songs, seem to represent functional units of song
production. Birds normally stop singing after
finishing a syllable, as opposed to in the middle
of a syllable. If one interrupts the song of a
zebra finch by flashing a strobe light, these stops
are almost always between syllables (Cynx 1990;
Franz & Goller 2002). The stereotypy of syllables
is reflected in the stereotypy of the pattern of
expiratory pressure and airflow that accompanies
each syllable (Allan & Suthers 1994; Suthers et
al. 1996; Franz & Goller 2002). These respiratory
parameters depend on both the syringeal muscle
activity influencing the aperture of the labial
valve in each side of the syrinx and on the activity
of the expiratory muscles that compress the air
sacs and generate the positive subsyringeal
pressure that drives airflow. Each syllable type
has a relatively invariant pattern of pressure and
airflow produced by its vocal gesture. When zebra
finches copy a syllable they typically also copy
the pattern of respiratory pressure that the tutor
used to produce that syllable (Franz & Goller
2002; Box 42, p. 322).
CARDINAL SONG: A CASE STUDY
Even relatively simple songs like that of the
northern cardinal require skillful coordination
of syringeal and respiratory muscles. An adult
male cardinal assembles its songs from a repertoire
of between 8 and 21 different syllables (Halkin
& Linville 1999). A song, a few seconds long,
typically consists of from one to several of these
syllables, each repeated several times to form a
series of phrases (CD1 #67).There are distinctive
patterns of airflow and pressure associated with
each syllable type. These patterns reflect the fact
that each is produced by a stereotyped motor
pattern coordinating the respiratory and syringeal
muscles to produce the particular vocal gesture
255
How birds sing and why it matters
for that syllable (Fig. 9.3). Many syllables include
extended upward or downward frequency sweeps
that sometimes exceed 2 octaves. Cardinals are
unusual in that females also sing, though their
song is less stereotyped and harmonics are more
prominent compared to the songs of males
(Lemon & Scott 1966; Dittus & Lemon 1969;
Yamaguchi 1998).
Except during a rapid trill, each syllable is
followed by a mini-breath that replenishes
respiratory air. Since both inspiration and
expiration are active processes, each syllable-
mini-breath cycle is accompanied in sequence
by contraction first of the abdominal expiratory
muscles (Hartley 1990), followed immediately
by the thoracic inspiratory muscles (Wild et al.
1998). During trills with high syllable repetition
rates there is not enough time for a mini-breath
between syllables. Instead, expiratory muscles
continue to contract and maintain a positive
pressure below the syrinx during the entire trill.
One side of the syrinx is kept closed and the
other side produces the trill by repetitively
opening and closing so that each puff of air
generates a syllable (Fig. 9.3).
To our ears, the long FM sweeps of a cardinal
sound smooth and continuous. It was therefore
surprising when closer examination revealed that
each side of the syrinx contributes to a different
part of the same syllable. Fundamental frequencies
below about 3.5 kHz are consistently sung on
the left side, and those above this frequency are
sung on the right side (Suthers & Goller 1996,
1997; Suthers 1997, 1999a; Suthers et al. 1999;
Fig. 9.3). The two sides of the cardinal syrinx
are thus specialized to operate over different
frequency ranges, facilitating the production of
extended FM sweeps that cross the boundary
between these registers, and thus contain
sequential contributions from each side. A syllable
that sweeps from 2to 7 kHz, for example, begins
in the left syrinx and is switched to the right
syrinx midway in the upward sweep. This
sequence of lateralized production is reversed in
syllables that sweep from high to low frequencies-
Most impressively the song generally contains
. no hint of this switch from one set of oscillating
28 1
labia to the other. To the human ear, and when
viewed as time-frequency sonograms, there is
usually no obvious interruption or discontinuity
in the frequency sweep at the moment of change
from one side to the other.
This is an extraordinary feat of virtuosity.
Consider the sequence of actions that a cardinal
must execute with precision to produce even an
acoustically simple note lasting less than half a
second. A single broadband FM sweep from high
to low begins with: (i) closure of the left syrinx;
(ii) expiratory muscle contraction; (3) opening
the right syrinx and configuring it to produce
the first portion of the sweep; (iv) closing the
?right syrinx while (v) opening the left syrinx
and configuring it to continue the sweep; (vi)
closing the left syrinx to terminate phonation;
(vii) relaxing expiratory muscles, and (viii)
contracting inspiratory muscles; (ix) opening both
sides of the syrinx for an inspiratory mini-breath
to replace the air used to produce the syllable.
Depending on the repetition rate of the syllable,
this entire sequence of events may take place in
scores of milliseconds and be repeated with
precision up to 16 times per second to produce
a phrase in the song!
LEARNING TO SING
Juvenile songbirds must meet many
developmental challenges on their way to
adulthood. Not the least of these is learning to
sing their species' song. Males must become expert
vocalists before they are a year old if they are to
have a chance of reproducing during their second
year. How does a young cardinal, of either sex,
go about learning to execute this 'checklist' of
actions that are necessary to produce even a simple
wideband frequency sweep (see Chapter 3)?
EcaZ Learning
Cardinals are age-limited, closed-ended learners,
and their adult song repertoire is learned during
their first year (Lemon & Scott 1966; Dittus &
Lemon 1969; Halkin & Linville 1999). Shortly
280 Nature's Music
2000). The implications of this independence
of right and left sides of the syrinx for vocal
communication are far reaching, for it gives
son g birds the possibilit y of two, virtually
independent sound sources. Although both right
and left sides are subjected to similar respiratory
pressures generated by the respiratory muscles
and both share the same supra-syringeal vocal
tract, the dorsal syringeal muscles can switch
sound production from one side of the syrinx to
the other very rapidly, silencing one side by closing
it while holding the labia on the opposite side
in a phonatory position. Likewise, the ventral
syringeal muscles can independently vary the
fundamental frequency generated on each side.
Songbirds have taken full advantage of this lateral
independence to greatly increase the variety and
complexity of their songs (Greenewalt 1968; Stein
1968; Suthers 1999a).
The Stereotypy of Adult Song
Most adult songbirds sing a stereotyped repertoire
in which reiterations of the same syllable type
vary little in their acoustic properties. This
acoustic stereotypy, together with the fact that
there is, at least sometimes, no immediate change
in the vocalization if auditory feedback is removed
by deafening, is part of the evidence suggesting
that the pattern of syringeal and respiratory
muscle activity required to produce a particular
syllable is stored in the brain as a 'motor program'
(konishi 1985; Vu et al. 1994). According to
this hypothesis, the syllables of crystallized song
might be produced by activating the appropriate
network of neurons in the brain; these are
presumably wired together in such a way that
they automatically generate the basic pattern of
coordinated muscle contractions, the vocal
gesture, needed to produce that syllable. Auditory
or somatosensory feedback may not be required,
at least in the short term, to produce the basic
vocal gesture, but it could adjust or modify it to
correct for errors that might otherwise occur if,
for example, conditions vary in the peripheral
vocal or respiratory system. Although central
pattern-generating circuits for different syllable
types have not yet been identified in songbirds,
they have an important role in controlling many
rhythmic behaviors in other animals (Getting
1989; Pearson 1993).
Whatever the neural mechanism underlying
syllable production, syllables rather than whole
songs, seem to represent functional units of song
production. Birds normally stop singing after
finishing a syllable, as opposed to in the middle
of a syllable. If one interrupts the song of a
zebra finch by flashing a strobe light, these stops
are almost always between syllables (Cynx 1990;
Franz & Goller 2002). The stereotypy of syllables
is reflected in the stereotypy of the pattern of
expiratory pressure and airflow that accompanies
each syllable (Allan & Suthers 1994; Suthers et
al. 1996; Franz & Goller 2002). These respiratory
parameters depend on both the syringeal muscle
activity influencing the aperture of the labial
valve in each side of the syrinx and on the activity
of the expiratory muscles that compress the air
sacs and generate the positive subsyringeal
pressure that drives airflow. Each syllable type
has a relatively invariant pattern of pressure and
airflow produced by its vocal gesture. When zebra
finches copy a syllable they typically also copy
the pattern of respiratory pressure that the tutor
used to produce that syllable (Franz & Goller
2002; Box 42, p. 322).
CARDINAL SONG: A CASE STUDY
Even relatively simple songs like that of the
northern cardinal require skillful coordination
of syringeal and respiratory muscles. An adult
male cardinal assembles its songs from a repertoire
of between 8 and 21 different syllables (Halkin
& Linville 1999). A song, a few seconds long,
typically consists of from one to several of these
syllables, each repeated several times to form a
series of phrases (CD1 #67).There are distinctive
patterns of airflow and pressure associated with
each syllable type. These patterns reflect the fact
that each is produced by a stereotyped motor
pattern coordinating the respiratory and syringeal
muscles to produce the particular vocal gesture
How birds sing and why it matters
for that syllable (Fig. 9.3). Many syllables include
extended upward or downward frequency sweeps
that sometimes exceed 2 octaves. Cardinals are
unusual in that females also sing, though their
song is less stereotyped and harmonics are more
prominent compared to the songs of males
(Lemon & Scott 1966; Dittus & Lemon 1969;
Yamaguchi 1998).
Except during a rapid trill, each syllable is
followed by a mini-breath that replenishes
respiratory air. Since both inspiration and
expiration are active processes, each syllable-
mini-breath cycle is accompanied in sequence
by contraction first of the abdominal expiratory
muscles (Hartley 1990), followed immediately
by the thoracic inspiratory muscles (Wild et al.
1998). During trills with high syllable repetition
rates there is not enough time for a mini-breath
between syllables. Instead, expiratory muscles
continue to contract and maintain a positive
pressure below the syrinx during the entire trill.
One side of the syrinx is kept closed and the
other side produces the trill by repetitively
opening and closing so that each puff of air
generates a syllable (Fig. 9.3).
To our ears, the long FM sweeps of a cardinal
sound smooth and continuous. It was therefore
surprising when closer examination revealed that
each side of the syrinx contributes to a different
part of the same syllable. Fundamental frequencies
below about 3.5 kHz are consistently sung on
the left side, and those above this frequency are
sung on the right side (Suthers & Goller 1996,
1997; Suthers 1997, 1999a; Suthers et al. 1999;
Fig. 9.3). The two sides of the cardinal syrinx
are thus specialized to operate over different
frequency ranges, facilitating the production of
extended FM sweeps that cross the boundary
between these registers, and thus contain
sequential contributions from each side. A syllable
that sweeps from 2to 7 kHz, for example, begins
in the left syrinx and is switched to the right
syrinx midway in the upward sweep. This
sequence of lateralized production is reversed in
syllables that sweep from high to low frequencies-
Most impressively the song generally contains
. no hint of this switch from one set of oscillating
28 1
labia to the other. To the human ear, and when
viewed as time-frequency sonograms, there is
usually no obvious interruption or discontinuity
in the frequency sweep at the moment of change
from one side to the other.
This is an extraordinary feat of virtuosity.
Consider the sequence of actions that a cardinal
must execute with precision to produce even an
acoustically simple note lasting less than half a
second. A single broadband FM sweep from high
to low begins with: (i) closure of the left syrinx;
(ii) expiratory muscle contraction; (3) opening
the right syrinx and configuring it to produce
the first portion of the sweep; (iv) closing the
?right syrinx while (v) opening the left syrinx
and configuring it to continue the sweep; (vi)
closing the left syrinx to terminate phonation;
(vii) relaxing expiratory muscles, and (viii)
contracting inspiratory muscles; (ix) opening both
sides of the syrinx for an inspiratory mini-breath
to replace the air used to produce the syllable.
Depending on the repetition rate of the syllable,
this entire sequence of events may take place in
scores of milliseconds and be repeated with
precision up to 16 times per second to produce
a phrase in the song!
LEARNING TO SING
Juvenile songbirds must meet many
developmental challenges on their way to
adulthood. Not the least of these is learning to
sing their species' song. Males must become expert
vocalists before they are a year old if they are to
have a chance of reproducing during their second
year. How does a young cardinal, of either sex,
go about learning to execute this 'checklist' of
actions that are necessary to produce even a simple
wideband frequency sweep (see Chapter 3)?
EcaZ Learning
Cardinals are age-limited, closed-ended learners,
and their adult song repertoire is learned during
their first year (Lemon & Scott 1966; Dittus &
Lemon 1969; Halkin & Linville 1999). Shortly
256
282 Nature's Music
How birds sing and why it matters 283

Song of an Adult Cardinal after fledging, juveniles begin to produce
sequences of low intensity, highly variable notes
having little resemblance to adult song. Gradually
during the autumn, and then again early in the
following spring, these vocal ramblings become
louder and some notes begin to recur in a
recognizably similar form. This early stage of
vocal development, 'subsong,' (Fig. 9.4) is slowly
transformed into the next stage, 'plastic song,'
as a recognizable, though initially poorly
controlled, syllable morphology develops
(CDI # 067). Eventually, when the bird is about
10 months old plastic song becomes
indistinguishable from adult song. Instead of
being plastic or malleable, the song becomes
muscle actions underlying the motor patterns
that are required to produce adult song based
on an auditory 'template' in the brain, which
may be either innate or shaped by experience
early in life (Marler 1997). Experiments on zebra
finches, in which this process of sensorimotor
integration was temporarily disrupted at different
stages of song development by using botulinum
toxin to reversibly block transmission of nerve
impulses to the syringeal muscles, suggest that
in addition to the sensitive period for song
memorization, there is another one for
sensorimotor integration during the latter part
of song development (Pytte & Suthers 2000).
+

fixed or 'crystallized.' Subsong and plastic song

are thought to involve trial and error motor
learning during which the juvenile relies on
0.5
(A) Time (s)
auditory feedback of his own vocalizations
(Konishi 1965a)' and presumably somatosensory
L
feedback (Suthers et al. 2002); he learns the
Milestones in Vocal Motor Control
Little is known about how the respiratory or
vocal motor skills of adult song develop. At what
stage in their development are young birds able
to coordinate respiratory and s~ringealmuscles,

Subsong of a Juvenile Cardinal

I
' *8 *L'@ 59

0.5
(B) Time (s)

Figure 9.3 Phrases from songs of an adult northern cardinal. (A) Two long, downward sweeping FM
syllables followed by faster upward FM syllables. In each syllable, frequencies below about 3.5 or 4.0 kHz
are sung on the left side (L) as indicated by airflow through the left syrinx; higher frequencies are produced
on the right side (R). Absence of flow through either side during a positive respiratory pressure indicates
that the labial valve is closed, preventing both airflow and sound production on that side. Note the mini-
breaths between each syllable and the stereotypical patterns of airflow and pressure that are characteristic 0.5
of each syllable type. (9)Another syllable with a low frequency fundamental produced entirely on the left Time (s)
side (L). This is followed by a trill using pulsatile expiration instead of mini-breaths. Expiratory pressure
remains positive during the entire phrase. FL and FR show the rate of airflow through the left and right Figure 9.4 Subsong from a 92-day-old juvenile cardinal. Arrows indicate periods of expiration- without
syrinx, respectively, P is the pressure in the cranial thoracic air sac. Horizontal lines indicate zero airflow sound production. Abbreviations as in Figure 9.3. In those marked B, both sides contribute.
and ambient air pressure. Shaded portions of the airflow are inspiratory mini-breaths.

257
282 Nature's Music
How birds sing and why it matters 283

Song of an Adult Cardinal after fledging, juveniles begin to produce
sequences of low intensity, highly variable notes
having little resemblance to adult song. Gradually
during the autumn, and then again early in the
following spring, these vocal ramblings become
louder and some notes begin to recur in a
recognizably similar form. This early stage of
vocal development, 'subsong,' (Fig. 9.4) is slowly
transformed into the next stage, 'plastic song,'
as a recognizable, though initially poorly
controlled, syllable morphology develops
(CDI # 067). Eventually, when the bird is about
10 months old plastic song becomes
indistinguishable from adult song. Instead of
being plastic or malleable, the song becomes
muscle actions underlying the motor patterns
that are required to produce adult song based
on an auditory 'template' in the brain, which
may be either innate or shaped by experience
early in life (Marler 1997). Experiments on zebra
finches, in which this process of sensorimotor
integration was temporarily disrupted at different
stages of song development by using botulinum
toxin to reversibly block transmission of nerve
impulses to the syringeal muscles, suggest that
in addition to the sensitive period for song
memorization, there is another one for
sensorimotor integration during the latter part
of song development (Pytte & Suthers 2000).
+

fixed or 'crystallized.' Subsong and plastic song

are thought to involve trial and error motor
learning during which the juvenile relies on
0.5
(A) Time (s)
auditory feedback of his own vocalizations
(Konishi 1965a)' and presumably somatosensory
L
feedback (Suthers et al. 2002); he learns the
Milestones in Vocal Motor Control
Little is known about how the respiratory or
vocal motor skills of adult song develop. At what
stage in their development are young birds able
to coordinate respiratory and s~ringealmuscles,

Subsong of a Juvenile Cardinal

I
' *8 *L'@ 59

0.5
(B) Time (s)

Figure 9.3 Phrases from songs of an adult northern cardinal. (A) Two long, downward sweeping FM
syllables followed by faster upward FM syllables. In each syllable, frequencies below about 3.5 or 4.0 kHz
are sung on the left side (L) as indicated by airflow through the left syrinx; higher frequencies are produced
on the right side (R). Absence of flow through either side during a positive respiratory pressure indicates
that the labial valve is closed, preventing both airflow and sound production on that side. Note the mini-
breaths between each syllable and the stereotypical patterns of airflow and pressure that are characteristic 0.5
of each syllable type. (9)Another syllable with a low frequency fundamental produced entirely on the left Time (s)
side (L). This is followed by a trill using pulsatile expiration instead of mini-breaths. Expiratory pressure
remains positive during the entire phrase. FL and FR show the rate of airflow through the left and right Figure 9.4 Subsong from a 92-day-old juvenile cardinal. Arrows indicate periods of expiration- without
syrinx, respectively, P is the pressure in the cranial thoracic air sac. Horizontal lines indicate zero airflow sound production. Abbreviations as in Figure 9.3. In those marked B, both sides contribute.
and ambient air pressure. Shaded portions of the airflow are inspiratory mini-breaths.

258
284 Nature's Music
or achieve independent control of their two
syringeal sound sources? Are there bottlenecks
in the process underlying motor development
and coordination that limit song acquisition?
Does the progression through successive stages
of juvenile song correspond to the mastery of
specific hurdles in vocal production? To better
understand the motor development of song, we
studied its production in juvenile male cardinals
at various stages of vocal learning (Suthers &
Goller 1998a, b).
Even young. nestlings may produce different
sounds simultaneously on each side of the syrinx.
Two-voice elements in the begging calls of young
chaffinches indicate that separate sounds are being
generated on each side of the syrinx in nestlings
only a few days old (Nottebohm 1971b;
Wilkinson 1980). It is not clear, however, if these
two-voice vocalizations are simply a passive
byproduct of poorly controlled airfl.ow through
both sides of the syrinx, or if they are under
active control of the syringeal muscles.
In cardinals, the ability to control which side
of the syrinx produces sound is present in
fledglings singing subsong when they are less
than 6 weeks old. Juveniles at this age are already
capable of unilateral sound production by closing
one or the other side of their syrinx, as do adults.
Sometimes both sides of the syrinx produce sound
at the same time, but often a note is generated
only on the left or the right side, the silent side
being closed to airflow (Fig. 9.4). Unlike adult
song, each note is quite variable and the frequency
structure of successive left and right notes is not
coordinated or linked (Suthers & Goller 1998a).
When adult cardinals sing, the timing of
expiratory airflow is carefully coordinated with
vocalization. Each syllable typically begins as soon
as the syringeal valve opens and ends when the
valve closes (Fig. 9.3). As a result, nearly all of
the exhaled air contributes to phonation and
very little is 'wasted' between syllables, even
du;ing the fast pulsatile trills. The respiratory-
syringeal coordination necessary for this efficient
;se of air during song is not present at the
beginning of juvenile song, and emerges only
gradually during the earl y stages of song
development. During early subsong and even
into plastic song, expiration sometimes continues
into or through the silent intervals between notes
(Fig. 9.4, arrows).
As subsong progresses, there is a tendency to
alternate sides and to produce a pair of notes,
one from each side, during a single expiration.
This foreshadows the adult motor pattern of
using the two sides in sequence to make a single
continuous frequency sweep. But in early juvenile
song the pair of notes within a single expiration
are not joined, nor is there any clear, consistent
coordinated sequence in their frequency patterns.
The linking of a high frequency sound from the
right with a lower frequency contribution from
the left side of the syrinx to form a coordinated,
seamless frequency sweep occurs in the latter
part of song development. J
PUSHING THE ACOUSTIC
ENVELOPE
In adult song, the respiratory and syringeal
muscles act together like experienced partners
in a dance choreographed by the brain, but the
dance steps differ between species. The possession
of two parallel sound sources, each with its own
pneumatic valve and neural control, gives oscine
songbirds an additional degree of freedom absent
in species that have a tracheal syrinx or that lack
independent control of the &o sides of the
bronchotracheal or bronchial syrinx. Studies of
how different species use the two sides of their
syrinx during song suggest that the distinctive
styles of singing among various songbird taxa
are achieved by using their duplex vocal organ
in different ways. These may optimize particular
acoustic effects at the cost of sacrificing virtuosity
in other aspects of sound production (Suthers
1997, 1999a, b; Suthers & Goller 1997).
Longer Song Bouts Versus Faster Tempos
Increasing Song Dlnration
Respiratory air is an essential, but potentially
limited, resource for vocalization. Sustained song
259
How birds sing and why it matters
or even the duration of individual notes might
be limited by the volume of air that can be
exhaled to produce them. This possible limitation
is particularly acute in small birds that sing big
songs. Canaries are small birds that can sing in
a seemingly continuous fashion for up to about
a minute. During a single song they may produce
scores or hundreds of syllables, yet their resting
respiratory tidal volume is only a fraction of a
milliliter! Calder (1970) showed that the depth
of the thorax increases slightly between most
notes in a canary song. From this he hypothesized
that the bird takes a mini-breath between syllables,
to replenish its air supply. Measurements of
respiratory pressure and tracheal airflow during
song showed that each mini-breath replaces about
the same volume of air that is exhaled to produce
the syllable (Hartley & Suthers 1989). As a result,
there is probably little or no net change in
respiratory volume during mini-breath song.
The mini-breath respiratory pattern is not
unique to songbirds. It appears to be used by
many species, including non-songbirds as diverse
as parrots and oilbirds (Suthers 2001). In our
own species, shortened inspirations, analogous
t o mini-breaths, are also present during
conversational speech (Hixon 1973). In all these
cases brief inspirations permit a continuity of
vocal communication that is free from periodic
interruptions by long inspirations.
Increasing Syllable Repetition Rate
Although mini-breaths allow a bird to sing longer
songs, they limit the song's tempo to rates of
syllable repetition having silent intervals long
enough to accommodate an inspiration. Some
birds get around this constraint by switching to
a different respiratory technique involving
pulsatile expiration, described above for cardinals,
which does not require the reversal of airflow
between syllables. Pulsatile expiration can almost
double the maximum syllable repetition rate that
is possible with mini-breaths. This increase in
syllable repetition rate is achieved at the cost of
depleting the volume of respiratory air available
for sound production. This limits the length of
phrases based on pulsatile expiration, but it is
285
unclear how often this limit is actually reached,
though this may sometimes occur (see Chapter
11).
The syllable repetition rate that requires a
bird to switch from mini-breaths to pulsatile
expiration is probably determined by the physical
properties of respiratory structures - particularly
the mass and elasticity of the thorax and abdomen
- that must oscillate at the syllable repetition
rate to drive ventilation. Small birds can sustain
a mini-breath respiratory pattern at higher syllable
repetition rates than can large birds. The 'cut-
off' repetition rate for mini-breaths for an 18 g
canary is about 30 syllables per second compared
. to about 16 syllables per second for a 35-40 g
cardinal. Pulsatile expiration is not subject to
this limit since it requires little movement of
the body wall, and the labia, whose valve-like
action gates airflow, have little mass.
Some zebra finches attempt to get the best of
both worlds by vocalizing during some mini-
breaths between 'normal' expiratory syllables
(Goller & Daley 2001). These inspiratory syllables
have a distinctive high fundamental frequency
and are copied by young zebra finches exposed
to them in tutor songs. It is not known if
inspiratory syllables have special perceptual
significance in communication or why more birds
don't use them. It is possible that they may be
energetically expensive to produce since the labia
must remain in the air stream during inspiration,
resulting in a higher resistance to airflow and
presumably requiring more effort from the
inspiratory muscles.
Increasing Spectral Diversity
Bandwidth: Left and Right Voral Registers
Songbirds use the two sides of their syrinx in
different ways to extend the bandwidth of their
song or to produce syllables that vary widely in
their tonal quality. The range of frequencies a
song encompasses, its bandwidth, is increased
by having each side of the syrinx specialized to
cover a different frequency band. In all species
studied, the fundamental bandwidth of the right
side is shifted upward compared to that of the
284 Nature's Music
or achieve independent control of their two
syringeal sound sources? Are there bottlenecks
in the process underlying motor development
and coordination that limit song acquisition?
Does the progression through successive stages
of juvenile song correspond to the mastery of
specific hurdles in vocal production? To better
understand the motor development of song, we
studied its production in juvenile male cardinals
at various stages of vocal learning (Suthers &
Goller 1998a, b).
Even young. nestlings may produce different
sounds simultaneously on each side of the syrinx.
Two-voice elements in the begging calls of young
chaffinches indicate that separate sounds are being
generated on each side of the syrinx in nestlings
only a few days old (Nottebohm 1971b;
Wilkinson 1980). It is not clear, however, if these
two-voice vocalizations are simply a passive
byproduct of poorly controlled airfl.ow through
both sides of the syrinx, or if they are under
active control of the syringeal muscles.
In cardinals, the ability to control which side
of the syrinx produces sound is present in
fledglings singing subsong when they are less
than 6 weeks old. Juveniles at this age are already
capable of unilateral sound production by closing
one or the other side of their syrinx, as do adults.
Sometimes both sides of the syrinx produce sound
at the same time, but often a note is generated
only on the left or the right side, the silent side
being closed to airflow (Fig. 9.4). Unlike adult
song, each note is quite variable and the frequency
structure of successive left and right notes is not
coordinated or linked (Suthers & Goller 1998a).
When adult cardinals sing, the timing of
expiratory airflow is carefully coordinated with
vocalization. Each syllable typically begins as soon
as the syringeal valve opens and ends when the
valve closes (Fig. 9.3). As a result, nearly all of
the exhaled air contributes to phonation and
very little is 'wasted' between syllables, even
du;ing the fast pulsatile trills. The respiratory-
syringeal coordination necessary for this efficient
;se of air during song is not present at the
beginning of juvenile song, and emerges only
gradually during the earl y stages of song
development. During early subsong and even
into plastic song, expiration sometimes continues
into or through the silent intervals between notes
(Fig. 9.4, arrows).
As subsong progresses, there is a tendency to
alternate sides and to produce a pair of notes,
one from each side, during a single expiration.
This foreshadows the adult motor pattern of
using the two sides in sequence to make a single
continuous frequency sweep. But in early juvenile
song the pair of notes within a single expiration
are not joined, nor is there any clear, consistent
coordinated sequence in their frequency patterns.
The linking of a high frequency sound from the
right with a lower frequency contribution from
the left side of the syrinx to form a coordinated,
seamless frequency sweep occurs in the latter
part of song development. J
PUSHING THE ACOUSTIC
ENVELOPE
In adult song, the respiratory and syringeal
muscles act together like experienced partners
in a dance choreographed by the brain, but the
dance steps differ between species. The possession
of two parallel sound sources, each with its own
pneumatic valve and neural control, gives oscine
songbirds an additional degree of freedom absent
in species that have a tracheal syrinx or that lack
independent control of the &o sides of the
bronchotracheal or bronchial syrinx. Studies of
how different species use the two sides of their
syrinx during song suggest that the distinctive
styles of singing among various songbird taxa
are achieved by using their duplex vocal organ
in different ways. These may optimize particular
acoustic effects at the cost of sacrificing virtuosity
in other aspects of sound production (Suthers
1997, 1999a, b; Suthers & Goller 1997).
Longer Song Bouts Versus Faster Tempos
Increasing Song Dlnration
Respiratory air is an essential, but potentially
limited, resource for vocalization. Sustained song
How birds sing and why it matters
or even the duration of individual notes might
be limited by the volume of air that can be
exhaled to produce them. This possible limitation
is particularly acute in small birds that sing big
songs. Canaries are small birds that can sing in
a seemingly continuous fashion for up to about
a minute. During a single song they may produce
scores or hundreds of syllables, yet their resting
respiratory tidal volume is only a fraction of a
milliliter! Calder (1970) showed that the depth
of the thorax increases slightly between most
notes in a canary song. From this he hypothesized
that the bird takes a mini-breath between syllables,
to replenish its air supply. Measurements of
respiratory pressure and tracheal airflow during
song showed that each mini-breath replaces about
the same volume of air that is exhaled to produce
the syllable (Hartley & Suthers 1989). As a result,
there is probably little or no net change in
respiratory volume during mini-breath song.
The mini-breath respiratory pattern is not
unique to songbirds. It appears to be used by
many species, including non-songbirds as diverse
as parrots and oilbirds (Suthers 2001). In our
own species, shortened inspirations, analogous
t o mini-breaths, are also present during
conversational speech (Hixon 1973). In all these
cases brief inspirations permit a continuity of
vocal communication that is free from periodic
interruptions by long inspirations.
Increasing Syllable Repetition Rate
Although mini-breaths allow a bird to sing longer
songs, they limit the song's tempo to rates of
syllable repetition having silent intervals long
enough to accommodate an inspiration. Some
birds get around this constraint by switching to
a different respiratory technique involving
pulsatile expiration, described above for cardinals,
which does not require the reversal of airflow
between syllables. Pulsatile expiration can almost
double the maximum syllable repetition rate that
is possible with mini-breaths. This increase in
syllable repetition rate is achieved at the cost of
depleting the volume of respiratory air available
for sound production. This limits the length of
phrases based on pulsatile expiration, but it is
285
unclear how often this limit is actually reached,
though this may sometimes occur (see Chapter
11).
The syllable repetition rate that requires a
bird to switch from mini-breaths to pulsatile
expiration is probably determined by the physical
properties of respiratory structures - particularly
the mass and elasticity of the thorax and abdomen
- that must oscillate at the syllable repetition
rate to drive ventilation. Small birds can sustain
a mini-breath respiratory pattern at higher syllable
repetition rates than can large birds. The 'cut-
off' repetition rate for mini-breaths for an 18 g
canary is about 30 syllables per second compared
. to about 16 syllables per second for a 35-40 g
cardinal. Pulsatile expiration is not subject to
this limit since it requires little movement of
the body wall, and the labia, whose valve-like
action gates airflow, have little mass.
Some zebra finches attempt to get the best of
both worlds by vocalizing during some mini-
breaths between 'normal' expiratory syllables
(Goller & Daley 2001). These inspiratory syllables
have a distinctive high fundamental frequency
and are copied by young zebra finches exposed
to them in tutor songs. It is not known if
inspiratory syllables have special perceptual
significance in communication or why more birds
don't use them. It is possible that they may be
energetically expensive to produce since the labia
must remain in the air stream during inspiration,
resulting in a higher resistance to airflow and
presumably requiring more effort from the
inspiratory muscles.
Increasing Spectral Diversity
Bandwidth: Left and Right Voral Registers
Songbirds use the two sides of their syrinx in
different ways to extend the bandwidth of their
song or to produce syllables that vary widely in
their tonal quality. The range of frequencies a
song encompasses, its bandwidth, is increased
by having each side of the syrinx specialized to
cover a different frequency band. In all species
studied, the fundamental bandwidth of the right
side is shifted upward compared to that of the
260
386 Nature's Music
left (Suthers 1999a), but there is often substantial
overlap in the frequencies produced by each side.
Cardinals are unique, among the small sample
of species studied, for their limited overlap
between frequencies produced on each side and
their habit of combining sound from the two
sides to produce a single FM note. The anatomical
or physiological basis for these separate left-right
vocal registers is not known. In many songbirds
there are anatomical asymmetries between the
two primary bronchi that might include subtle
differences in the labia or muscles on each side,
as yet unidentified. Whatever its mechanism,
the increased overall bandwidth available for
vocalization increases the possibilities for spectral
diversity in songs.
T h e influence on song of lateralized
specialization for particular frequency ranges is
evident in different strains of canaries. One of
these is the Waterschlager canary, which was
inbred for its distinctive low-pitched song.
Another strain is the outbred domestic canary.
Nottebohm and Nottebohm (1976) conducted
a series of experiments on Waterschlagers in which
either the left or right side of the syrinx had
been denervated by cutting the tracheosyringeal
nerve on one side. They demonstrated that this
strain produces about 90% of its syllable
repertoire in its left syrinx. As a result of the
strong lateral dominance of the left side, most
of this bird's song is below 3 kHz, except for an
occasional right side syllable that has a median
frequency range between 2.5 and 4.2 kHz
(Nottebohm & Nottebohm 1976). This left
dominance is in contrast to song of the outbred
domestic strain in which each side of the syrinx
contributes about an equal proportion of the
repertoire, and some syllables include notes from
both sides (Suthers et al. 200 1). In these domestic
birds, syllables frequently have fundamental
frequencies as high as 5 or 6 kHz, and a single
bilaterally generated syllable may have a
bandwidth of 6 or 7 kHz compared to 3 or 4
kHz in the case of unilaterally produced syllables.
Subsequent studies show that Waterschlager
canaries have an inherited auditory defect that
decreases their sensitivity to sounds above 2 kHz
How birds sing and why it matters 287
by as much as 40 dB (Okanoya & Dooling 1985; prominent in songs of some mimic thrushes such capacity of the two sides of the syrinx to function
Gleich et al. 1994a, b; Gleich & Klump 1994). as grey catbirds and brown thrashers (Suthers independentl y during sound production,
Their strong left syringeal dominance may well 1992; Suthers et al. 1994, 1996; Goller & Suthers inharmonic vocalizations can also result from
be the result of being partially deaf to most of 1995a, b, 1996a, by 1999). T h e separate, the nonlinear interaction between the sounds
the sounds from the right syrinx. simultaneous 'voices' that form the distinctive generated on each side. The 'dee' syllable of the
In the family FringilZidae, which includes inharmonic elements in these songs (Fig. 9.5; black-capped chickadee's song consists of a series
sparrows, cardinals and Darwin's finches, there CD I # 68) tend to differ less from each other in of overlapping frequency components that differ
is an inverse relationship between syllable their fundamental frequency than when each from those produced by either side of the syrinx
bandwidth and repetition rate - rapid trills have side sings alone, suggesting that simultaneous alone. It is not the sum of the sounds produced
narrower bandwidths (Podos 1996; see Chapter two-voice phonation comes with constraints on on each side afier unilateral syringed denervation.
11). It is suggested that syllable bandwidth might the degree of frequency separation between the It appears instead to be composed of the sum
be constrained at high repetition rates by the two sides that is possible when they operate and difference heterodyne frequencies generated
time required to change the shape and tuning of simultaneously. This limitation might be due to by nonlinear interactions between the sound
the mouth and throat to track the rapidly biomechanical interactions between the two produced by each side of the syrinx (Nowicki &
changing fundamental frequency (Nowicki et halves of the syrinx or to a need to keep both Capranica 1986a, b). The coupling between left
al. 1992; Podos 1997, 2001). In many cases fundamentals within the frequency pass band and right sound generators might be mediated
beak opening is correlated with sound frequency of the vocal tract, with which both sides connect. by air pressure fluctations associated with sound
and appears to be a factor in tuning resonances Whereas two-voice syllables depend on the or transmitted mechanically through the syringed
of the vocal tract by altering its effective length
and suppressing harmonics. Podos argues that Brown Thrasher Song
beak size and syllable repetition rate have evolved
together in Darwin's finches. He suggests that
the relatively low syllable repetition rate of finches
with large beaks is due to the difficulty of moving
larger mandibles rapidly to quickly change the
tuning of the vocal tract filter (Podos 2001).
The role of vocal tract resonance and beak
movements in birdsong is covered in Chapter
11 (Nowicki 1987; Westneat et al. 1993;
Moriyama & Okanoya 1996; Suthers & Goller
1997; Suthers 1999a; Suthers et al. 1999; Hoese
et al. 2000; Williams 2001).
Inharmonic Sounds
The songbird syrinx is well suited to achieve
dissonant sounds by singing two overlapping
notes that are different in pitch but not
harmonicall y related to each other. T h e
widespread occurrence of these 'two-voice' notes
and their likely origin from separate sides of the
syrinx was pointed out by Greenewalt (1968)
and Stein (1968). Two-voice syllables require
airflow through both sides of the syrinx, while
different motor programs to the muscles on each 0.2
side generate unrelated sounds that can be Time (s)
independently modulated in frequency and Figure 9.5 Brown thrasher song illustrating inharmonic, two-voice syllables (CDI #68). V is the time
amplitude. Two voice syllables are especially waveform of vocalization. Abbreviations as in Figure 9.3. Modified from Suthers et al. 1994.
261
386 Nature's Music
left (Suthers 1999a), but there is often substantial
overlap in the frequencies produced by each side.
Cardinals are unique, among the small sample
of species studied, for their limited overlap
between frequencies produced on each side and
their habit of combining sound from the two
sides to produce a single FM note. The anatomical
or physiological basis for these separate left-right
vocal registers is not known. In many songbirds
there are anatomical asymmetries between the
two primary bronchi that might include subtle
differences in the labia or muscles on each side,
as yet unidentified. Whatever its mechanism,
the increased overall bandwidth available for
vocalization increases the possibilities for spectral
diversity in songs.
T h e influence on song of lateralized
specialization for particular frequency ranges is
evident in different strains of canaries. One of
these is the Waterschlager canary, which was
inbred for its distinctive low-pitched song.
Another strain is the outbred domestic canary.
Nottebohm and Nottebohm (1976) conducted
a series of experiments on Waterschlagers in which
either the left or right side of the syrinx had
been denervated by cutting the tracheosyringeal
nerve on one side. They demonstrated that this
strain produces about 90% of its syllable
repertoire in its left syrinx. As a result of the
strong lateral dominance of the left side, most
of this bird's song is below 3 kHz, except for an
occasional right side syllable that has a median
frequency range between 2.5 and 4.2 kHz
(Nottebohm & Nottebohm 1976). This left
dominance is in contrast to song of the outbred
domestic strain in which each side of the syrinx
contributes about an equal proportion of the
repertoire, and some syllables include notes from
both sides (Suthers et al. 200 1). In these domestic
birds, syllables frequently have fundamental
frequencies as high as 5 or 6 kHz, and a single
bilaterally generated syllable may have a
bandwidth of 6 or 7 kHz compared to 3 or 4
kHz in the case of unilaterally produced syllables.
Subsequent studies show that Waterschlager
canaries have an inherited auditory defect that
decreases their sensitivity to sounds above 2 kHz
How birds sing and why it matters 287
by as much as 40 dB (Okanoya & Dooling 1985; prominent in songs of some mimic thrushes such capacity of the two sides of the syrinx to function
Gleich et al. 1994a, b; Gleich & Klump 1994). as grey catbirds and brown thrashers (Suthers independentl y during sound production,
Their strong left syringeal dominance may well 1992; Suthers et al. 1994, 1996; Goller & Suthers inharmonic vocalizations can also result from
be the result of being partially deaf to most of 1995a, b, 1996a, by 1999). T h e separate, the nonlinear interaction between the sounds
the sounds from the right syrinx. simultaneous 'voices' that form the distinctive generated on each side. The 'dee' syllable of the
In the family FringilZidae, which includes inharmonic elements in these songs (Fig. 9.5; black-capped chickadee's song consists of a series
sparrows, cardinals and Darwin's finches, there CD I # 68) tend to differ less from each other in of overlapping frequency components that differ
is an inverse relationship between syllable their fundamental frequency than when each from those produced by either side of the syrinx
bandwidth and repetition rate - rapid trills have side sings alone, suggesting that simultaneous alone. It is not the sum of the sounds produced
narrower bandwidths (Podos 1996; see Chapter two-voice phonation comes with constraints on on each side afier unilateral syringed denervation.
11). It is suggested that syllable bandwidth might the degree of frequency separation between the It appears instead to be composed of the sum
be constrained at high repetition rates by the two sides that is possible when they operate and difference heterodyne frequencies generated
time required to change the shape and tuning of simultaneously. This limitation might be due to by nonlinear interactions between the sound
the mouth and throat to track the rapidly biomechanical interactions between the two produced by each side of the syrinx (Nowicki &
changing fundamental frequency (Nowicki et halves of the syrinx or to a need to keep both Capranica 1986a, b). The coupling between left
al. 1992; Podos 1997, 2001). In many cases fundamentals within the frequency pass band and right sound generators might be mediated
beak opening is correlated with sound frequency of the vocal tract, with which both sides connect. by air pressure fluctations associated with sound
and appears to be a factor in tuning resonances Whereas two-voice syllables depend on the or transmitted mechanically through the syringed
of the vocal tract by altering its effective length
and suppressing harmonics. Podos argues that Brown Thrasher Song
beak size and syllable repetition rate have evolved
together in Darwin's finches. He suggests that
the relatively low syllable repetition rate of finches
with large beaks is due to the difficulty of moving
larger mandibles rapidly to quickly change the
tuning of the vocal tract filter (Podos 2001).
The role of vocal tract resonance and beak
movements in birdsong is covered in Chapter
11 (Nowicki 1987; Westneat et al. 1993;
Moriyama & Okanoya 1996; Suthers & Goller
1997; Suthers 1999a; Suthers et al. 1999; Hoese
et al. 2000; Williams 2001).
Inharmonic Sounds
The songbird syrinx is well suited to achieve
dissonant sounds by singing two overlapping
notes that are different in pitch but not
harmonicall y related to each other. T h e
widespread occurrence of these 'two-voice' notes
and their likely origin from separate sides of the
syrinx was pointed out by Greenewalt (1968)
and Stein (1968). Two-voice syllables require
airflow through both sides of the syrinx, while
different motor programs to the muscles on each 0.2
side generate unrelated sounds that can be Time (s)
independently modulated in frequency and Figure 9.5 Brown thrasher song illustrating inharmonic, two-voice syllables (CDI #68). V is the time
amplitude. Two voice syllables are especially waveform of vocalization. Abbreviations as in Figure 9.3. Modified from Suthers et al. 1994.
262

288 Nature's Music
tissue. Similar nonlinear bilateral coupling has
not yet been noted in other species, perhaps
because bilateral interaction is less prominent
in the larger syrinx of most other birds studied.
Other kinds of nonlinear phenomena may
also account for the abrupt transitions from
periodic to chaotic sounds, as well as period
doubling and other acoustic effects that are
present in some songs, such assthoseof the zebra
finch (Fee et al. 1998). The realization that non-
linear vocalizations may have a role in social
communication has led to an increased interest
in their production (Fitch et al. 2002).
Spectrdl Contvart
The duplicate sound sources in the oscine syrinx
facilitate the introduction into song of abrupt
changes in fundamental frequency. The song of
the brown-headed cowbird, for example, begins
with 2 or 3 expirations each containing a 'cluster'
of notes, rapidly produced using pulsatile
expiration (Allan & Suthers 1994; Fig. 9.6; CDl
# 69). Successive notes are produced on opposite
sides of the syrinx. In each case, the first note is
sung on the left side and is immediately followed,
or may partially overlap, the next note, which is
sung on the right side at a higher frequency.
The frequency of these alternating notes increases
in a staggered step-wise sequence until the end
of the expiration. Abrupt frequency steps between
successive notes that follow each other with little
or no silent interval between them are possible
if separate generators are involved. While one
side of the syrinx is producing a note, the muscles
of the other closed, silent side can adjust the
tension on its labia so that it starts the next note
on pitch without a slurred FM between the end
of one note and the beginning of the next,
something considered again below.
PERFORMANCE CONSTRAINTS:
INSIGHTS FROM A VOCAL MIMIC
Have these species-specific patterns of song
production evolved because they are the best
way to generate that species' style of song, perhaps
I
How birds sing and why it matters 289
i Brown-headed Cowbird Song
the only way, or are they simply independent
modes of singing, each with the potential to
generate a wide range of song types? To what
extent have the evolutionary forces shaping the
acoustic properties of a species' song influenced
the evolution of production mechanisms, and
vice versa? Here I approach from a different
viewpoint, some of the issues considered by ten
Cate (see Chapter 1O), and by Podos and Nowicki
(see Chapter 1 1).
Vocal mimics, such as mockingbirds, provide
an opportunity to investigate these questions.
When a mimic copies another species' song does
it also use the same mechanism employed by
that species to produce the song, or does it invent
a different way to produce the song?The northern
mockingbird is renowned for its ability to mimic
other species (Baylis 1982), but the vocal
mechanism it uses to accomplish this is unknown.
When a mockingbird imitates a cardinal's song,
for example, does it do so by joining sounds
produced sequentially on each side of the syrinx
to generate extended FM sweeps as the cardinal
does, or does it produce an unbroken sweep on
one or both sides together?
Zollinger and I (2004) tutored hand-reared
juvenile mockingbirds with recorded songs of
other species. We found that when a mockingbird
sings a cardinal song it usually mimics the vocal
mechanism of the cardinal, switching between
sides of the syrinx in mid-frequency sweep, but
the switch is seldom as seamless or the frequency
sweep as smooth as when the cardinal sings it. 0.2
Time (s)
When the mockingbird's motor pattern differed
from that of the cardinal, the mockingbird's Figure 9.6 Song of a brown-headed cowbird (CD1 #69). Each introductory note cluster is produced
vocalization also differed from that of the model. during a single expiratory phase by pulsatile expiration that alternates between sides, permitting abrupt
In a similar experiment, juvenile mockingbirds frequency changes between successive notes. The last expiration produces the final whistle and is always
were tutored with recorded cowbird songs. The on the right side. Abbreviations as in Figure 9.3.
cowbird's song covers an exceptionally wide
frequency range that often extends from a few sides, beginning with the lowest frequency note were tutored, not with natural songs, but with
hundred Hz for the lowest introductory notes on the left. computer synthesized cardinal-like FM sweeps
to 10 kHz or more in the final whistle. Although To test the possibility that the taped tutor or cowbird-like note clusters. When as an adult,
mockingbirds were unable to reproduce the songs of cardinals or cowbirds contained some the mockingbird mimicked the synthesized
highest or lowest frequencies of this tutor, they subtle cues such as brief pauses or discontinuities, cardinal-like FM sweeps, it sang the high
mimicked both the sound and the cowbird's vocal other than frequency, that might influence when frequency portion on its right side and the low
production mechanism for other notes in the the mockingbird switches from one side of the frequency portion on its left, switching sides at
note clusters, singing alternate notes on opposite syrinx to the other, some juvenile mockingbirds 3 or 4 kHz, during both upward and downward
263

288 Nature's Music
tissue. Similar nonlinear bilateral coupling has
not yet been noted in other species, perhaps
because bilateral interaction is less prominent
in the larger syrinx of most other birds studied.
Other kinds of nonlinear phenomena may
also account for the abrupt transitions from
periodic to chaotic sounds, as well as period
doubling and other acoustic effects that are
present in some songs, such assthoseof the zebra
finch (Fee et al. 1998). The realization that non-
linear vocalizations may have a role in social
communication has led to an increased interest
in their production (Fitch et al. 2002).
Spectrdl Contvart
The duplicate sound sources in the oscine syrinx
facilitate the introduction into song of abrupt
changes in fundamental frequency. The song of
the brown-headed cowbird, for example, begins
with 2 or 3 expirations each containing a 'cluster'
of notes, rapidly produced using pulsatile
expiration (Allan & Suthers 1994; Fig. 9.6; CDl
# 69). Successive notes are produced on opposite
sides of the syrinx. In each case, the first note is
sung on the left side and is immediately followed,
or may partially overlap, the next note, which is
sung on the right side at a higher frequency.
The frequency of these alternating notes increases
in a staggered step-wise sequence until the end
of the expiration. Abrupt frequency steps between
successive notes that follow each other with little
or no silent interval between them are possible
if separate generators are involved. While one
side of the syrinx is producing a note, the muscles
of the other closed, silent side can adjust the
tension on its labia so that it starts the next note
on pitch without a slurred FM between the end
of one note and the beginning of the next,
something considered again below.
PERFORMANCE CONSTRAINTS:
INSIGHTS FROM A VOCAL MIMIC
Have these species-specific patterns of song
production evolved because they are the best
way to generate that species' style of song, perhaps
I
How birds sing and why it matters 289
i Brown-headed Cowbird Song
the only way, or are they simply independent
modes of singing, each with the potential to
generate a wide range of song types? To what
extent have the evolutionary forces shaping the
acoustic properties of a species' song influenced
the evolution of production mechanisms, and
vice versa? Here I approach from a different
viewpoint, some of the issues considered by ten
Cate (see Chapter 1O), and by Podos and Nowicki
(see Chapter 1 1).
Vocal mimics, such as mockingbirds, provide
an opportunity to investigate these questions.
When a mimic copies another species' song does
it also use the same mechanism employed by
that species to produce the song, or does it invent
a different way to produce the song?The northern
mockingbird is renowned for its ability to mimic
other species (Baylis 1982), but the vocal
mechanism it uses to accomplish this is unknown.
When a mockingbird imitates a cardinal's song,
for example, does it do so by joining sounds
produced sequentially on each side of the syrinx
to generate extended FM sweeps as the cardinal
does, or does it produce an unbroken sweep on
one or both sides together?
Zollinger and I (2004) tutored hand-reared
juvenile mockingbirds with recorded songs of
other species. We found that when a mockingbird
sings a cardinal song it usually mimics the vocal
mechanism of the cardinal, switching between
sides of the syrinx in mid-frequency sweep, but
the switch is seldom as seamless or the frequency
sweep as smooth as when the cardinal sings it. 0.2
Time (s)
When the mockingbird's motor pattern differed
from that of the cardinal, the mockingbird's Figure 9.6 Song of a brown-headed cowbird (CD1 #69). Each introductory note cluster is produced
vocalization also differed from that of the model. during a single expiratory phase by pulsatile expiration that alternates between sides, permitting abrupt
In a similar experiment, juvenile mockingbirds frequency changes between successive notes. The last expiration produces the final whistle and is always
were tutored with recorded cowbird songs. The on the right side. Abbreviations as in Figure 9.3.
cowbird's song covers an exceptionally wide
frequency range that often extends from a few sides, beginning with the lowest frequency note were tutored, not with natural songs, but with
hundred Hz for the lowest introductory notes on the left. computer synthesized cardinal-like FM sweeps
to 10 kHz or more in the final whistle. Although To test the possibility that the taped tutor or cowbird-like note clusters. When as an adult,
mockingbirds were unable to reproduce the songs of cardinals or cowbirds contained some the mockingbird mimicked the synthesized
highest or lowest frequencies of this tutor, they subtle cues such as brief pauses or discontinuities, cardinal-like FM sweeps, it sang the high
mimicked both the sound and the cowbird's vocal other than frequency, that might influence when frequency portion on its right side and the low
production mechanism for other notes in the the mockingbird switches from one side of the frequency portion on its left, switching sides at
note clusters, singing alternate notes on opposite syrinx to the other, some juvenile mockingbirds 3 or 4 kHz, during both upward and downward
264

Nature's Music
290
sweeping sounds (Fig. 9.7; CDl #70). When
copying a sequence of frequency-stepped tones
similar to the cowbird's note cluster, the
mockingbird likewise switched back and forth
between sides of his syrinx producing each tone
on the same side as would a cowbird.
Perhaps the mockingbirds determine which
side of their syrinx to use simply on the basis of
the note's frequency. In mockingbirds, like other
songbirds, the frequency range of the right syrinx
is shifted upward compared to that of the left.
Although the absolute frequency of a note is a
factor in determining which side of the syrinx
will sing it, the context and other acoustic
properties of the song are also important and
can override frequency bias. This is shown by
an experiment in which mockingbirds were
tutored with two pairs of synthesized tones similar
to those often present in cowbird's note clusters.
In each pair, the second tone was at 2 kHz. In
one pair this sound was preceded by a lower
frequency tone whereas in the other pair, it was
preceded by a higher frequency tone, (Fig. 9.8).
In both cases, the frequency difference between
tones was 300 Hz and the second tone started
immediatel y after the first. W h e n the
mockingbird mimicked these tone pairs, the side
it used to produce the 2 kHz note differed,
depending on whether the immediately preceding
tone was higher or lower. If the first tone was
higher, it was sung on the right and the 2 kHz
tone was sung on the left. If, however, the first
tone was lower it was sung on the left and the 2
kHz tone was produced on the right (Fig. 9.8).
It is clear from this experiment that a 2 kHz
0 note can be produced on either side of the syrinx.
Why then did the mockingbird change sides
between the tones in each pair? Why not sing
the second tone on the same side that was used
to produce the first? The answer is apparent
from the occasional times a mockingbird failed
to change sides between tones. When both tones
I were sung on the same side of the syrinx, the
abrupt step-wise frequency change between tones
was lost, and the end of the first note became
slurred into the frequency of the second note so
that the two were connected with an FM sweep.
I
The ability to achieve abrupt step-wise frequency
I How birds sing and why it matters
Mockingbird Copying Synthesized Cardinal Song
291
changes without a significant silent period
81 Synthetic Tutor song
between notes may depend on exploitation of
the two voices of the bipartite syrinx. Perhaps it
is only by switching sound production from side
to side during the rapidly produced notes in
their note clusters that cowbirds are able to achieve
the clean frequency steps, or spectral contrast,
between notes. This use of the bipartite syrinx
circumvents the limited ability of a single sound
source to abruptly change the frequency at which
it oscillates. To achieve this acoustic effect without
slurrin g , the mimic must exploit the vocal
flexibility made possible by two sound sources, 0.5
as the cowbird does in its song. Time (sec)
Our findings on how mockingbirds copy other (A)
species support the view that the songs of different
species have been selected to maximize the Mockingbird Imitation
performance of certain species-specific acoustic
features, such as the abrupt frequency steps
between cowbird notes, or the smooth extended
FM sweeps of cardinals. Most songbirds have
become vocal specialists who push the
performance limits of singing in certain
directions, often at the expense of some other
kinds of acoustic skills. To be a successful vocal
mimic, the mockingbird must remain to some
extent a vocal generalist, a jack-of-all-trades but
master of none, who does a fairly good job of
mimicking other species, but usually with less
expertise than the rightful owner.
VOCAL GYMNASTICS: ARE
FEMALES IMPRESSED?
Two important functions of song are, a role in
competition between males for breeding
territories, and in attracting a mate. Exactly what
aspects of song are most important in fulfilling
these needs is poorly understood, and may differ
between species (Catchpole & Slater 1995; Gil
& Gahr 2002; see Chapter 2). Most studies of
0.5
Time (sec)
the relationship betwein singing and mating
(B)
success have focused on variables such as Figure 9.7 (A) A tutor 'song' consisting of a series of computer synthesized downward FM sweeps similar
repertoire size, or the frequency or duration of to syllables sung by cardinals. (B) A mockingbird's reproduction of these sweeps (CDI #70).The timing of
singing. These are all amenable to measurement, airflow through each side of the syrinx shows that the mockingbird begins each sweep on the right side and
changes to the left side for the portion of the sweep that is below about 3 kHz. Abbreviations as in Figure
265

Nature's Music
290
sweeping sounds (Fig. 9.7; CDl #70). When
copying a sequence of frequency-stepped tones
similar to the cowbird's note cluster, the
mockingbird likewise switched back and forth
between sides of his syrinx producing each tone
on the same side as would a cowbird.
Perhaps the mockingbirds determine which
side of their syrinx to use simply on the basis of
the note's frequency. In mockingbirds, like other
songbirds, the frequency range of the right syrinx
is shifted upward compared to that of the left.
Although the absolute frequency of a note is a
factor in determining which side of the syrinx
will sing it, the context and other acoustic
properties of the song are also important and
can override frequency bias. This is shown by
an experiment in which mockingbirds were
tutored with two pairs of synthesized tones similar
to those often present in cowbird's note clusters.
In each pair, the second tone was at 2 kHz. In
one pair this sound was preceded by a lower
frequency tone whereas in the other pair, it was
preceded by a higher frequency tone, (Fig. 9.8).
In both cases, the frequency difference between
tones was 300 Hz and the second tone started
immediatel y after the first. W h e n the
mockingbird mimicked these tone pairs, the side
it used to produce the 2 kHz note differed,
depending on whether the immediately preceding
tone was higher or lower. If the first tone was
higher, it was sung on the right and the 2 kHz
tone was sung on the left. If, however, the first
tone was lower it was sung on the left and the 2
kHz tone was produced on the right (Fig. 9.8).
It is clear from this experiment that a 2 kHz
0 note can be produced on either side of the syrinx.
Why then did the mockingbird change sides
between the tones in each pair? Why not sing
the second tone on the same side that was used
to produce the first? The answer is apparent
from the occasional times a mockingbird failed
to change sides between tones. When both tones
I were sung on the same side of the syrinx, the
abrupt step-wise frequency change between tones
was lost, and the end of the first note became
slurred into the frequency of the second note so
that the two were connected with an FM sweep.
I
The ability to achieve abrupt step-wise frequency
I How birds sing and why it matters
Mockingbird Copying Synthesized Cardinal Song
291
changes without a significant silent period
81 Synthetic Tutor song
between notes may depend on exploitation of
the two voices of the bipartite syrinx. Perhaps it
is only by switching sound production from side
to side during the rapidly produced notes in
their note clusters that cowbirds are able to achieve
the clean frequency steps, or spectral contrast,
between notes. This use of the bipartite syrinx
circumvents the limited ability of a single sound
source to abruptly change the frequency at which
it oscillates. To achieve this acoustic effect without
slurrin g , the mimic must exploit the vocal
flexibility made possible by two sound sources, 0.5
as the cowbird does in its song. Time (sec)
Our findings on how mockingbirds copy other (A)
species support the view that the songs of different
species have been selected to maximize the Mockingbird Imitation
performance of certain species-specific acoustic
features, such as the abrupt frequency steps
between cowbird notes, or the smooth extended
FM sweeps of cardinals. Most songbirds have
become vocal specialists who push the
performance limits of singing in certain
directions, often at the expense of some other
kinds of acoustic skills. To be a successful vocal
mimic, the mockingbird must remain to some
extent a vocal generalist, a jack-of-all-trades but
master of none, who does a fairly good job of
mimicking other species, but usually with less
expertise than the rightful owner.
VOCAL GYMNASTICS: ARE
FEMALES IMPRESSED?
Two important functions of song are, a role in
competition between males for breeding
territories, and in attracting a mate. Exactly what
aspects of song are most important in fulfilling
these needs is poorly understood, and may differ
between species (Catchpole & Slater 1995; Gil
& Gahr 2002; see Chapter 2). Most studies of
0.5
Time (sec)
the relationship betwein singing and mating
(B)
success have focused on variables such as Figure 9.7 (A) A tutor 'song' consisting of a series of computer synthesized downward FM sweeps similar
repertoire size, or the frequency or duration of to syllables sung by cardinals. (B) A mockingbird's reproduction of these sweeps (CDI #70).The timing of
singing. These are all amenable to measurement, airflow through each side of the syrinx shows that the mockingbird begins each sweep on the right side and
changes to the left side for the portion of the sweep that is below about 3 kHz. Abbreviations as in Figure
266

292 Nature's Music
Imitation of Stepped Tones
Reveals Motor Constraints
Mockingbird righttleft
2-tone imitations
f--- 2 kHz
Slurring with
righttright imitation
6l
0.5
Time (sec)
(B)
Figure 9.8 (A) A mockingbird's imitations of pairs of computer-synthesized tones showing that the context
as well as the frequency of a note influences the side on which it is produced. (A) When the 2 kHz tone
is preceded by a tone at a slightly higher frequency, the first tone is sung on the right and the 2 kHz tone
is sung on the left side. If the preceding tone is at a lower frequency, it is sung on the left, and the 2 kHz
tone is now sung on the right side. (B) Switching sides of production enhances the spectral contrast
between successive notes. When singing the second pair of synthesized tones, the mockingbird produced
both on the same side. As a result, the two tones are joined by an FM sweep. R and L indicate side of
production. Tutor tones and syringeal airflow and pressure are not shown. (A modified after Zollinger and
Suthers 2004).
in which the motor mechanisms of song
but are not necessarily correlated with the level
production have been studied, an aspect of singing
of motor skill, and are unlikely to reflect the
that seems likely to require special vocal motor
difficulty that is required to execute the vocal
skills is coordination between the two halves of
performance. A male's quality might be reflected
the syrinx. Examples are the acoustically seamless
not only in his stamina for sustained singing,
switch between right and left sides during-wide
but also in his ability to flawlessly and repeatedly
band frequency sweeps by cardinals or the
sing syllables or phrases that require special motor
concurrent production of independently
skills (Suthers & Goller 1997; Nowicki et al.
modulated sounds forming two-voice syllables
1998a; ten Cate et al. 2002). Among the species
267
How birds sing and why it matters 293
in brown thrashers. It is not yet known whether on its syllable type. Some syllables may be almost
females pay attention to these particular aspects a pure tone repeated a few times a second, others
of cardinal or thrasher song. However, in brown- include upward or downward frequency sweeps
headed cowbirds and domestic canaries, there is of varying complexity sung at various repetition
some evidence suggesting that parts of their songs rates which may exceed 30 per second.
requiring a high degree of motor coordination Researchers at the University of Paris (Vallet
may have a special perceptual significance to & Kreutzer 1995; Vallet et al. 1998) found that
prospective mates. certain kinds of 'type A' syllables are more effective
Male brown-headed cowbirds have a vocal than others in attracting the interest of a
repertoire that includes several different song prospective mate (Box 4, p. 58). They played
types. In the eastern subspecies, each song begins different phrases, each consisting of a single
with 2 or 3 note clusters followed by a whistle syllable type, to sexually receptive female canaries
(Fig. 9.6). West, King, and their colleagues and recorded their CSD responses. The female
showed that some songs are more effective than invites the male to mate with her by arching her
others in eliciting copulati,onsolicitation displays ,back, raising her tail, and fluttering her partially
(CSD) from female cowbirds (King &West 1977, spread wings. Vallet and his colleagues found
1983; West et al. 1979, 1981; see Chapter 14). that wide-band syllables that were more complex
A male's breeding success is correlated with the than simple frequency modulated sweeps,
potency of his songs. Experimental manipulation composed of two or more notes, and sung at a
of songs indicates that the note clusters are more repetition rate of 16 per second or higher evoked
effective than the final whistle in eliciting CSDs. significantly more female displays than other
A high-frequency note, designated the syllables. Similar type 'A' syllables recorded from
'interphrase unit' at the end of the last note wild canaries also elicit CSDs from domestic
cluster appears to be particularly important. Songs canaries when played back to them in the
of dominant males were more effective than those laboratory (Leitner et al. 200 I), indicating they
of other males. The inclusion of stereotyped note are not an artifact of domestication.
clusters in juvenile plastic song triples the Studies of how domestic canaries produce their
effectiveness of juvenile song in evoking a female songs suggest that 'A' syllables may be relatively
copulatory response (west & ~ i n 1988a). g difficult to sing (Suthers et al. 200 1). Domestic
Further experiments with computer-edited a canaries sing some syllables in their left syrinx,
songs are needed to identify the specific temporal others in the right, and still' other multi-note
or acoustic aspects of note clusters that are most syllables contain contributions from both sides
important in stimulating a female. But it already of the syrinx. The wide-band 'A' syllables of three
seems clear that a male cowbird's mating success birds studied, included contributions from both
is influenced by his bilateral motor skills in singing sides of the syrinx (Fig. 9.9; C D l #71). The
stereotyped patterns of notes that are rapidly syllables that are most effective in eliciting CSDs
produced, with alternating steps in frequency. from a female canary thus require the male to
The domestic canary provides another example produce stereotyped notes from each side of his
of a possible relationship between vocal motor syrinx at note repetition rates at least twice the
dexterity and female choice. Male domestic syllable rate. Furthermore, behavioral playback
canaries have a song repertoire of about 20-30 experiments (Vallet et al. 1998) suggest that the
different syllables. Individual songs contain a effectiveness of 'A' syllables in stimulating a female
subset of these syllables, repeated to form a is greater if the silent interval between the notes
sequence of phrases. Each syllable type is always within each syllable is short (between 16 and 23
sung in the same way and at the same repetition ms, as opposed to 23-30 ms). This is consistent
rate, but there are large differences in the spectral with the hypothesis that bilateral coordination
and temporal properties of a phrase, depending involving very rapid left-right switching between

292 Nature's Music
Imitation of Stepped Tones
Reveals Motor Constraints
Mockingbird righttleft
2-tone imitations
f--- 2 kHz
Slurring with
righttright imitation
6l
0.5
Time (sec)
(B)
Figure 9.8 (A) A mockingbird's imitations of pairs of computer-synthesized tones showing that the context
as well as the frequency of a note influences the side on which it is produced. (A) When the 2 kHz tone
is preceded by a tone at a slightly higher frequency, the first tone is sung on the right and the 2 kHz tone
is sung on the left side. If the preceding tone is at a lower frequency, it is sung on the left, and the 2 kHz
tone is now sung on the right side. (B) Switching sides of production enhances the spectral contrast
between successive notes. When singing the second pair of synthesized tones, the mockingbird produced
both on the same side. As a result, the two tones are joined by an FM sweep. R and L indicate side of
production. Tutor tones and syringeal airflow and pressure are not shown. (A modified after Zollinger and
Suthers 2004).
in which the motor mechanisms of song
but are not necessarily correlated with the level
production have been studied, an aspect of singing
of motor skill, and are unlikely to reflect the
that seems likely to require special vocal motor
difficulty that is required to execute the vocal
skills is coordination between the two halves of
performance. A male's quality might be reflected
the syrinx. Examples are the acoustically seamless
not only in his stamina for sustained singing,
switch between right and left sides during-wide
but also in his ability to flawlessly and repeatedly
band frequency sweeps by cardinals or the
sing syllables or phrases that require special motor
concurrent production of independently
skills (Suthers & Goller 1997; Nowicki et al.
modulated sounds forming two-voice syllables
1998a; ten Cate et al. 2002). Among the species
How birds sing and why it matters 293
in brown thrashers. It is not yet known whether on its syllable type. Some syllables may be almost
females pay attention to these particular aspects a pure tone repeated a few times a second, others
of cardinal or thrasher song. However, in brown- include upward or downward frequency sweeps
headed cowbirds and domestic canaries, there is of varying complexity sung at various repetition
some evidence suggesting that parts of their songs rates which may exceed 30 per second.
requiring a high degree of motor coordination Researchers at the University of Paris (Vallet
may have a special perceptual significance to & Kreutzer 1995; Vallet et al. 1998) found that
prospective mates. certain kinds of 'type A' syllables are more effective
Male brown-headed cowbirds have a vocal than others in attracting the interest of a
repertoire that includes several different song prospective mate (Box 4, p. 58). They played
types. In the eastern subspecies, each song begins different phrases, each consisting of a single
with 2 or 3 note clusters followed by a whistle syllable type, to sexually receptive female canaries
(Fig. 9.6). West, King, and their colleagues and recorded their CSD responses. The female
showed that some songs are more effective than invites the male to mate with her by arching her
others in eliciting copulati,onsolicitation displays ,back, raising her tail, and fluttering her partially
(CSD) from female cowbirds (King &West 1977, spread wings. Vallet and his colleagues found
1983; West et al. 1979, 1981; see Chapter 14). that wide-band syllables that were more complex
A male's breeding success is correlated with the than simple frequency modulated sweeps,
potency of his songs. Experimental manipulation composed of two or more notes, and sung at a
of songs indicates that the note clusters are more repetition rate of 16 per second or higher evoked
effective than the final whistle in eliciting CSDs. significantly more female displays than other
A high-frequency note, designated the syllables. Similar type 'A' syllables recorded from
'interphrase unit' at the end of the last note wild canaries also elicit CSDs from domestic
cluster appears to be particularly important. Songs canaries when played back to them in the
of dominant males were more effective than those laboratory (Leitner et al. 200 I), indicating they
of other males. The inclusion of stereotyped note are not an artifact of domestication.
clusters in juvenile plastic song triples the Studies of how domestic canaries produce their
effectiveness of juvenile song in evoking a female songs suggest that 'A' syllables may be relatively
copulatory response (west & ~ i n 1988a). g difficult to sing (Suthers et al. 200 1). Domestic
Further experiments with computer-edited a canaries sing some syllables in their left syrinx,
songs are needed to identify the specific temporal others in the right, and still' other multi-note
or acoustic aspects of note clusters that are most syllables contain contributions from both sides
important in stimulating a female. But it already of the syrinx. The wide-band 'A' syllables of three
seems clear that a male cowbird's mating success birds studied, included contributions from both
is influenced by his bilateral motor skills in singing sides of the syrinx (Fig. 9.9; C D l #71). The
stereotyped patterns of notes that are rapidly syllables that are most effective in eliciting CSDs
produced, with alternating steps in frequency. from a female canary thus require the male to
The domestic canary provides another example produce stereotyped notes from each side of his
of a possible relationship between vocal motor syrinx at note repetition rates at least twice the
dexterity and female choice. Male domestic syllable rate. Furthermore, behavioral playback
canaries have a song repertoire of about 20-30 experiments (Vallet et al. 1998) suggest that the
different syllables. Individual songs contain a effectiveness of 'A' syllables in stimulating a female
subset of these syllables, repeated to form a is greater if the silent interval between the notes
sequence of phrases. Each syllable type is always within each syllable is short (between 16 and 23
sung in the same way and at the same repetition ms, as opposed to 23-30 ms). This is consistent
rate, but there are large differences in the spectral with the hypothesis that bilateral coordination
and temporal properties of a phrase, depending involving very rapid left-right switching between
268
294 Nature's Music How birds sing and whv it matters 295

'Sexy' Syllables of a Domestic Canary Song vocal tract or cranial muscle groups involved in
161s singing. It may be that males who are not in top
1 R 181s
physical condition have difficulty producing these
phrases. If so, the 'A' phrase may be an honest
indicator of male fitness (see Chapter 2).
Another possibility is that the presence of 'A'
phrases in song may indicate a male's
,-h reproductive state - that he is ready to breed.
Field studies of wild canaries show that the
I syllable repetition rate of their repertoire, but
I not repertoire size, changes seasonally. X syllables
sung during the breeding season are replaced
during the winter with syllables at lower
repetition rates (Leitner et al. 2001). Seasonal
changes in steroid levels can affect syringeal
muscles and vocal behavior (DeVoogd 199 1;
! DeVoogd et al. 1991; Beani et al. 1995; Hartley
l et al. 1997; Tramontin et al. 2000). The presence
I
mate selection, and it is not clear if there is a
correlation between the presence of 'A' syllables
and reproductive success. Whatever the case, a
better understanding of how birds sing promises
to be a valuable tool in deciphering the
communicative functions and evolution of song.
CONCLUSIONS
Twenty years ago, in a review on avian sound
production, Brackenbury (1982) lamented that,
"many ideas about the functioning of the
passerine syrinx are based on guesswork." In the
.past decade new techniques for recording and
observing syringeal function during song have
yielded significant progress toward removing some
of this guesswork and advancing our

I of 'A' phrases in a male's song might indicate
his readiness to mate. Domestic canaries sing
I longer strings of 'A' phrases in the presence of
other male or female canaries than when they
are singing alone, even though the number and
duration of songs is similar in both situations
(Kreutzer et al. 1999).
These two possible functions of 'A' syllables
in intersexual communication are not mutually
exclusive. Other factors may also play a role in
understanding how birds sing, though the
number of species sampled is still very small.
There is every reason to believe the next decade
will be equally productive. As we learn more
about how the avian vocal system works, and its
acoustic possibilities and limitations, we will gain
new insights into the neural organization,
behavioral significance, and evolution of a
behavioral system that is unique in the animal
kingdom.

0.1
Time (s)

Figure 9.9 A segment of a domestic canary song containing two phrases of type 'A' syllables (CDI #74).
Each syllable is composed of two notes. The first note sweeps upward on the right side for about 20 ms
and is followed by an approximately 10 ms downward sweep at a lower frequency on the left side. To be
maximally effective in eliciting copulation displays from a receptive female, syllables must be complex and
sung at repetition rates greater than 16 per second. Abbreviations as in Figure 9.3.

sides of the syrinx may have special perceptual
importance in the vocal communication of birds.
The kind of information females get from 'A'
syllables remains to be determined. O n e
possibility is that they provide an indication of
the male's health or fitness. Of all the syllables
in a canary's repertoire, 'A' phrases may demand
the greatest precision in bilateral motor
coordination within the syrinx, and also between
the syrinx and the respiratory muscles and other

269
294 Nature's Music How birds sing and whv it matters 295

'Sexy' Syllables of a Domestic Canary Song vocal tract or cranial muscle groups involved in
161s singing. It may be that males who are not in top
1 R 181s
physical condition have difficulty producing these
phrases. If so, the 'A' phrase may be an honest
indicator of male fitness (see Chapter 2).
Another possibility is that the presence of 'A'
phrases in song may indicate a male's
,-h reproductive state - that he is ready to breed.
Field studies of wild canaries show that the
I syllable repetition rate of their repertoire, but
I not repertoire size, changes seasonally. X syllables
sung during the breeding season are replaced
during the winter with syllables at lower
repetition rates (Leitner et al. 2001). Seasonal
changes in steroid levels can affect syringeal
muscles and vocal behavior (DeVoogd 199 1;
! DeVoogd et al. 1991; Beani et al. 1995; Hartley
l et al. 1997; Tramontin et al. 2000). The presence
I
mate selection, and it is not clear if there is a
correlation between the presence of 'A' syllables
and reproductive success. Whatever the case, a
better understanding of how birds sing promises
to be a valuable tool in deciphering the
communicative functions and evolution of song.
CONCLUSIONS
Twenty years ago, in a review on avian sound
production, Brackenbury (1982) lamented that,
"many ideas about the functioning of the
passerine syrinx are based on guesswork." In the
.past decade new techniques for recording and
observing syringeal function during song have
yielded significant progress toward removing some
of this guesswork and advancing our

I of 'A' phrases in a male's song might indicate
his readiness to mate. Domestic canaries sing
I longer strings of 'A' phrases in the presence of
other male or female canaries than when they
are singing alone, even though the number and
duration of songs is similar in both situations
(Kreutzer et al. 1999).
These two possible functions of 'A' syllables
in intersexual communication are not mutually
exclusive. Other factors may also play a role in
understanding how birds sing, though the
number of species sampled is still very small.
There is every reason to believe the next decade
will be equally productive. As we learn more
about how the avian vocal system works, and its
acoustic possibilities and limitations, we will gain
new insights into the neural organization,
behavioral significance, and evolution of a
behavioral system that is unique in the animal
kingdom.

0.1
Time (s)

Figure 9.9 A segment of a domestic canary song containing two phrases of type 'A' syllables (CDI #74).
Each syllable is composed of two notes. The first note sweeps upward on the right side for about 20 ms
and is followed by an approximately 10 ms downward sweep at a lower frequency on the left side. To be
maximally effective in eliciting copulation displays from a receptive female, syllables must be complex and
sung at repetition rates greater than 16 per second. Abbreviations as in Figure 9.3.

sides of the syrinx may have special perceptual
importance in the vocal communication of birds.
The kind of information females get from 'A'
syllables remains to be determined. O n e
possibility is that they provide an indication of
the male's health or fitness. Of all the syllables
in a canary's repertoire, 'A' phrases may demand
the greatest precision in bilateral motor
coordination within the syrinx, and also between
the syrinx and the respiratory muscles and other

270
Fiete, I.R. and Seung, H.S. (2009) Birdsong Learning. In: Encyclopedia of Neuroscience.
Squire, L., Editor. pp. 227-239. Elsevier, New York.
Birdsong Learning
I R Fiete, University of California, Santa Barbara, and
Center for Neural Systems, Pasadena, CA, USA
H S Seung, Massachusetts Institute of Technology,
Boston MA, USA
2009 Published by Elsevier Ltd.
Introduction
Birdsong has been compared to human speech
because both are examples of animal vocalizations
that are at least partially learned. Indeed, some of
the neural pathways underlying these vocal behaviors
may be similar in songbirds and humans. More gen-
erally, birdsong can be compared to any motor behav-
ior that involves sequential control and learning and
is an ideal playground for experimental and theoreti-
cal explorations of the neural basis for motor learning
and control.
The neurons involved in song production and learn-
ing are clustered in a small number of discrete nuclei
with well-defined boundaries, fairly well-characterized
connectivity, and increasingly well-studied neuro-
physiology. The various species of oscine songbirds
display a vast spectrum of complexity in syllable
sequencing (syntax), improvisation, and vocal acqui-
sition. For example, song can be as simple as repeti-
tions of a set of syllables arranged in fixed order
(e.g., zebra finches), or it can be intricately impro-
vised, with syllable ordering governed by probabilistic
transitions between different syllables (e.g., domesti-
cated bengalese finches). Mockingbirds can acquire
novel songs and learn novel nonbird sounds through-
out life. Yet the basic neural song circuit is similar in
all oscines. This diversity of behavior subserved by an
underlying similarity of structure makes birdsong
an intriguing system for future comparative studies
of sophisticated motor sequencing and control.
Models have so far focused on the zebra finch, a
songbird with one of the simplest of oscine song reper-
toires, because most recordings of neural activity have
been obtained from these birds. The male zebra finch
learns one song motif with fixed syllable order and
sings repetitions of that motif throughout its life. Song
acquisition takes place in a sensory phase, when the
juvenile bird acquires a mental template of the song
of a possibly unrelated adult tutor, and an overlapp-
ing sensorimotor phase, when the juvenile begins
to vocalize and tries to match its vocalizations with
its mental template. In the laboratory the sensory
and sensorimotor phases can be separated: template
acquisition is fast and can be completed and the tutor
bird removed from the juveniles presence before the
Birdsong Learning 227
juvenile has begun to practice its song. Sensorimotor
learning is a slow process of self-generated trial and
error and can proceed successfully in complete audi-
tory and social isolation after template acquisition.
However, auditory feedback of the birds own voca-
lizations is crucial during sensorimotor learning
because a bird deafened at this stage cannot learn to
reproduce its mental template of the tutor song.
This article focuses on the motor and sensorimotor
aspects of song production, in particular the forma-
tion of neural sequences to drive song, the learning of
the motor map, and the reasons song may be encoded
the way it is by songbird premotor neurons. The aim
is to illustrate how theoretical work has contributed
to the understanding of the song system and highlight
the resulting predictions for experiment. Auditory
neural coding and dynamics and the physics (acous-
tics and mechanics) of songbird vocal production are
beyond the scope of the article.
Basic Functional Anatomy of
the Song Pathway
In the oscine vocal premotor circuit, high-level vocal
center nucleus (HVC) projects to nucleus robustus
archistriatalis (RA). RA drives the motor neurons
that innervate the syringeal and respiratory organs
for song production (Figure 1). Lesioning any of
these nuclei immediately disables song production,
while nuclei upstream of HVC are not necessary for
singing in adult zebra finches. Therefore, HVC origi-
nates the top-level premotor drive for adult zebra
finch song. Although the circuitry between the pre-
motor nuclei is feedforward, connectivity within
HVC and within RA is extensively recurrent. In
addition to the premotor circuit, the song system
contains the anterior forebrain pathway (AFP),
which connects HVC to RA through an indirect
path. Lesion studies reveal that the AFP is crucial
for the sensorimotor phase of song learning and
long-term song maintenance but not for song produc-
tion. The AFP is an avian homolog of the mammalian
thalamocorticalbasal ganglia pathway.
Based on anatomical, neurophysiological, and
lesion data, it is believed that HVC generates the
spatiotemporal premotor drive for sequential motor
activation in the form of sequential neural activity.
The HVC activity is abstract, in the sense that it
encodes only temporal ordering, not song features
(Figures 2(a) and 2(b)). The synapses from HVC to
RA and from RA to the motor neurons convert the
sequence of HVC activity patterns into the sequence
of motor commands that produces song (Figure 2).
271
228 Birdsong Learning
HVC
RA
UVA
Syrinx nXIIts
DLM
Premotor
AFP
a
Figure 1 The neural pathways underlying song learning and song production. (a) Song learning and production involves a discrete set
of nuclei. (b) The same areas, unfolded to better illustrate their connectivity. Lines terminating in arrows designate excitatory or putative
excitatory connectivity. Lines terminating in circles designate inhibitory synapses. Blue nuclei are part of the premotor pathway. Green
nuclei are part of the anterior forebrain pathway (AFP) and are important for sensorimotor song learning but not song production. HVC,
high-level vocal center nucleus; LMAN, lateral magnocellular nucleus of the anterior nidopallium; NIf, interfacial nucleus; RA, nucleus
robustus archistriatalis.
According to this perspective, there are two types of
learning. First, HVC acquires the capability of gener-
ating a long sequence of neural activity patterns
(Figures 2(f) and 2(g)). Second, the circuitry down-
stream from HVC acquires the capability of mapping
HVC activity patterns to the correct sound-producing
patterns of activity in RA. The HVCRA synapses
show extensive structural plasticity, with synaptic
growth followed by massive retraction, precisely
during the sensorimotor learning period. Electro-
physiological synaptic maturation accompanies such
structural remodeling. Therefore, motor map learning
from HVC to the vocal outputs is likely to involve
plasticity in the HVCRA synapses. Other possible
loci of plasticity for sensorimotor learning, not ruled
out by experiments, include the recurrent RARA
synapses or the RAmotor neuron connections.
Sequence Generation
The high-level premotor area HVC displays clear
sequential patterns of neural activation. In the awake
singing zebra finch, each recorded RA-projecting
HVC neuron (also called an HVCRA neuron) fires at
most one brief burst of spikes (lasting 6 ms) per
second-long song motif (Figure 2(a)). These neurons
are otherwise inactive during singing. Different RA-
projecting HVC neurons fire their single bursts at
different fixed times in the motif. Thus as a popula-
tion, the HVCRA neurons cover time during song.
In addition to RA-projecting neurons, HVC con-
tains inhibitory interneurons which ramify within
HVC and X-projecting neurons which send their out-
puts to the AFP. Inhibitory interneuron activity is very
HVC X Basal ganglia
NIf
LMAN DLM Thalamus
Cortex
RA LMAN
X
Motor
Excitatory b
Inhibitory
different from HVCRA activity: these neurons fire
tonically throughout song, with high rates and rela-
tively little modulation in firing frequency across song
(Figure 2(b)).
In the adult zebra finch, deafening causes no imme-
diate effect on song production, suggesting that audi-
tory feedback is not important for sequence generation.
Also, lesion studies indicate that input from the higher
interfacial nucleus to HVC is not necessary for singing
in zebra finches. What network connectivity and neural
dynamics underlie the sequence generation dynamics
observed in the top-level premotor area, and how
might that connectivity become established as hatchl-
ings mature?
The Synaptic Chain Model
The synaptic chain model is an old idea for how
networks of neurons might produce sequential pat-
terns of activity. (The idea was already old when it
was critiqued by Lashley in 1951 under the name
associative chaining model.) In the models simplest
version, the neurons in a network are divided into
groups, and the groups are ordered in a sequence.
From each group of neurons, there are excitatory
synaptic connections to the neurons in the next
group of the sequence. The groups and their synaptic
connections are like the links of a chain.
If the first group in the sequence is activated, it
sends synaptic drive to the second group, which
then becomes active. The second group in turn excites
the third group to become active, and so on down the
chain. (An alternative mechanism for the propagation
of sequential activity in songbird HVC is based on
the removal of lateral inhibition between groups of
272
 Birdsong Learning 229

neurons; however, this mechanism does not allow for

100 ms the self-propagation of sequential activity, instead
requiring an external periodic drive (clocked input)
HVCRA

to induce transitions between activity states.)

More-complex versions of the synaptic chain
model include inhibitory as well as excitatory
synapses or different types of synaptic currents with
a multiple time constants. In still other versions, pre-
cisely synchronous spiking of one group is required to
HVCI

activate the next group. These are known as synfire

(or synchronous firing) chain networks.
b

The role of intrinsic bursting Two recent models

have implemented the synaptic chain idea with excit-
atory conductance-based neurons to model sequence
RA

generation in the zebra finch HVC.

Recordings in the singing zebra finch show that
when an RA-projecting HVC neuron is active during
c song, it fires a high-frequency burst of four to six
spikes in a 610 ms interval (Figure 2(a)). Previously
ex

we discussed the question of the mechanism of

d sequential activation in HVC. Another basic question
concerns the mechanisms that give rise to the bursting
Song

behavior. One possibility is that the neurons burst

e
WHVC
1 2 3 4 HVC
WHVC-RA
WRA
a b c x y z RA
WRA-motor
m1 m2 Motor
f g Time
Figure 2 Summary of data and schematic model. (a)(e) Activ-
ity in the premotor pathway. Raster plot of single-unit activity in
nucleus robustus archistriatalis (RA)-projecting high-level vocal
center nucleus (HVC) neurons (HVCRA neurons; a), in HVC inter-
neurons (HVCI; b), and in RA (c). Consecutive lines of the same
color are multiple recordings of the same neuron. Ticks represent
single spikes. (d) Electromyogram of activity in the abdominal
expiratory muscle. (e) Spectrogram of resulting song. (a) and (b)
adapted from Hahnloser RH, Kozhevnikov AA, and Fee MS
(2002) An ultra-sparse code underlies the generation of neural
sequences in a songbird. Nature 419(6902): 6570; (c) from Leo-
nardo A and Fee MS (2005) Ensemble coding of vocal control in
birdsong. Journal of Neuroscience 25(3): 652661; (d) and (e)
from Suthers RA, Goller F, and Pytte C (1999) The neuromuscular
control of birdsong. Philosophical Transactions of the Royal Soci-
ety of London, Series B: Biological Sciences 354(1385): 927939;
(f,g) Schematic of loci of song production and plasticity. (f) Flow of
activity: HVC drives RA, which drives muscle contraction via
motor neurons in a feedforward path. In addition, connectivity
within HVC and RA is recurrent. Dashed lines represent weights
that are likely to be plastic; solid lines with arrows designate
weights assumed to be fixed. HVC, with its recurrent connectivity,
because they are driven by synaptic input that lasts
for the duration of the burst. Alternatively, the neu-
1 rons might burst as a result of the dynamics of their
2
3 intrinsic conductances.
4
a Li and Greenside have advanced a model that does
b
c not use intrinsic bursting. It is consistent with the fact
x that previous brain slice studies of HVC neurons have
y
z not revealed intrinsic bursting. The network model is
m1
able to produce sequential neural activity with multi-
m2 spike bursts (Figure 3(a)). However, the realism of
this model is compromised by the fact that the peak
firing rate within the propagating multispike burst
achieved by the model (15 ms interspike intervals)
is approximately 10 times lower than seen in the
actual HVC (1.5 ms interspike intervals).
In contrast, Jin et al. have advanced a model in which
HVC neurons are intrinsically bursting. The model
generates multispike bursts at peak firing rates consis-
tent with those seen in experiments (Figure 3(b)). For
comparison, Jin et al. also simulated a synaptic chain
model without intrinsic bursting. Unlike Greenside
and Li, they were able to achieve higher peak firing
produces the high-level sequential drive. Motor neurons, desig-
nated by motor neuron pools m1 and m2, sum the activity from RA
((g), middle eight traces) to generate drive for the muscles
((g), bottom two traces). Sequence learning involves plasticity in
the recurrent HVC connectivity. In this simple picture, sensori-
motor learning is dissociable from sequence learning and involves
plasticity in the HVCRA and perhaps RARA connections.

273
230 Birdsong Learning
100 ms 20 ms
HVC1
HVC2
HVC3
HVC4
HVC5
a Time b
1 2 3 4 5 6 7 during song, and this neuromodulator could be nec-
c
1 2 3 4 5 6 7 aptic chain models mentioned earlier, there are excit-
d While the synaptic chain model is intuitively plausi-
Figure 3 (a,b) Propagation of bursts in synaptic chain models of
high-level vocal center nucleus (HVC). (a) Neurons in the model
are not intrinsically bursting. Activity propagates due to synaptic
chain connectivity, and the burst of spikes in each neuron is due to
sustained synaptic drive from the network. Spike frequency during
the burst is 10 times lower than in nucleus robustus archistria-
talis (RA)-projecting HVC (HVCRA) neurons. Adapted from Li
M and Greenside H (2006) Stable propagation of a burst through
a one-dimensional homogeneous excitatory chain model of song-
bird nucleus HVC. Physical Review. E, Statistical, Nonlinear, and
Soft Matter Physics 74(1.1): 011918. (b) HVCRA neurons are
assumed to possess an intrinsic mechanism for burst generation.
A synaptic chain network can robustly produce stable sequences
of multispike neural activation. Adapted from Jin DZ, Ramazano-
glu F, and Seung HS (2007) Intrinsic bursting enhances the
robustness of a neural network model of sequence generation
by avian brain area HVC. Journal of Computational Neuroscience
23(3): 283299. (c,d) Unidirectional synaptic chain versus bidirec-
tional connectivity. (c) In a synaptic chain, connections between
different groups of neurons project in a predominantly forward
direction. Activity initialized at group 1 (red electrode) will drive
sequential activity up to the end of the chain, at group 7. Stimula-
tion of the last group of the chain (blue electrode) will produce no
propagating activity or sequence playback. Activation of a group
near the middle of the chain will generate playback of a fragment
of the original activity sequence. The resulting vocalization should
correspond to normal song, but beginning midsong or midsyllable.
(d) Bidirectional synaptic connectivity. Activity initialized at group 1
rates more consistent with experimental data, but this
required careful fine-tuning of parameters like synap-
tic strengths. In contrast, the model with intrinsic
bursting was quite robust to changes in parameters
like synaptic strengths.
Superficially, the Jin et al. model may seem incon-
sistent with previous experiments in vitro, in which
somatic current injection did not trigger intrinsic
bursting in RA-projecting HVC neurons. However,
one can imagine two scenarios in which this inconsis-
tency is resolved. First, somatic current injection
might be ineffective for revealing intrinsic bursting
that is caused by a dendritic calcium spike. This
depends on the strength of coupling between the
soma and dendrite and on the voltage threshold for
a dendritic spike. Second, the intrinsic properties of
RA-projecting HVC neurons appear to be different
in vitro and in the brain of a singing bird. For exam-
ple, the peak firing rates observed in vitro are never as
high as those observed during song. Possibly, HVC
neurons in vitro lack some neuromodulator secreted
essary for intrinsic bursting. Given the findings of Jin
et al. regarding robustness, it seems important to look
again for intrinsic bursting of HVC neurons.
Directionality of excitatory connectivity In the syn-
atory synapses from each group to the next group in
the chain. Unidirectionality is a fundamental feature.
ble, it is important to subject it to experimental test-
ing and to contrast the models predictions with those
of competing models. For example, one can imagine a
bidirectional synaptic chain in which there are excit-
atory synapses between adjacent groups in the chain,
but the synapses go in both directions, forward and
backward (Figures 3(c) and 3(d)).
In both the unidirectional and bidirectional mod-
els, it is assumed that song is initiated by activation
of the first group in the chain. This causes the
entire chain to be activated sequentially in the for-
ward direction (Figures 3(c) and 3(d), red electrode).
Both models can reproduce the sequential activity
patterns observed in HVC during singing. However,
(red electrode) will produce a unidirectional sequence of activity
propagating from beginning to end, as in the synaptic chain.
Stimulation at the end of the chain (blue electrode) will lead to
sequence playback in the reverse direction, from group 7 to 1. The
resulting vocalization should be the song or a syllable sung in
reverse. Stimulation of a central group in the chain will lead to
activity propagation in both forward and backward directions. The
resulting vocalization should be a superposition of the song (or a
syllable) sung forward and in reverse, starting in the middle.
274

the bidirectional model can also produce other kinds
of activity patterns that the unidirectional model can-
not (Figures 3(c) and 3(d)), which could be used as the
basis of experimental tests that distinguish between
the two models.
In these tests, activity in the HVC network would
be initiated by some means, and the subsequent
response to stimulation would be observed. If the last
group in the chain is stimulated (Figures 3(c) and 3(d),
blue electrode), there would be no subsequent activity
in the unidirectional model. In the bidirectional model,
the activity would propagate backward along the chain
and would result in a sequence of motor commands
that is reversed from normal song. Another experiment
would be to stimulate an intermediate group in the
chain (Figures 3(c) and 3(d), green electrode). Accord-
ing to the bidirectional model, subsequent activity
would simultaneously propagate both forward and
backward along the chain. (Note that the playback
will not continue from the forward to the backward
direction, or from the backward to the forward direc-
tion, because of the existence of a refractory period for
activation. This point is well known in the theory of
excitable media.) In the unidirectional model, activity
would propagate only forward.
While such experiments may sound straight-
forward, they are actually technically demanding. So
far no topographic organization of RA-projecting
HVC neurons has been detected. In other words, the
location of a neuron in HVC does not seem to be
related to the time at which it is activated during a
song motif. If coactive neurons were co-located in
HVC, it would be relatively straightforward to acti-
vate a hypothetical group in the synaptic chain models
by stimulation through a microelectrode. Instead, the
neurons of a hypothetical group appear to be scattered
throughout HVC. Therefore, a group would be best
identified through optical imaging of activity, and best
stimulated through optical means also.
So far, no one has attempted to precisely stimulate a
group of HVC neurons. However, cruder perturbation
experiments have been performed. For example, HVC
activity during song has been perturbed by stimulating
RA, which leads to retrograde effects on HVC. Stimu-
lation leads to skips in motif sequencing, but the
subsequent vocalization is a recognizable fragment of
the motif sung in the forward direction, beginning at a
different position within the motif. This is more consis-
tent with the unidirectional model than with the bidi-
rectional model. But more-conclusive tests will require
precisely controlled stimulation of activity in HVC.
A more direct way of testing the synaptic chain
models is to check their predictions concerning the
synaptic connectivity of HVC. Since the real HVC is
Birdsong Learning 231
no doubt messier than in these idealized models,
the predictions should be phrased probabilistically.
Suppose A and B are two RA-projecting HVC neurons
that are activated in succession. In the unidirectional
model, the probability of a connection from A to B is
expected to be higher than that of a connection from
B to A or of a connection between two neurons chosen
at random. (Alternatively, the predicted asymmetry of
coupling between A and B might be reflected in syn-
aptic strength rather than in the presence or absence of
connections.) In the bidirectional model, the probabil-
ity of a connection from A to B is equal to the proba-
bility of a connection from B to A.
Testing this kind of prediction could be done by
comparing physiological and neuroanatomical mea-
surements. Optical imaging could be used to find two
neurons A and B that are activated in succession. Then
optical stimulation could be used to see whether and
how they are connected to each other. Alternatively, a
completely neuroanatomical approach would be to
reconstruct the entire connectivity of HVC using serial
section electron microscopy, look for a chain within
this connectivity, and determine its directionality.
The role of inhibition The synaptic chain models
above were based on synaptic excitation, without
reference to the role of synaptic inhibition. Inhibition
is often included in such models to globally dampen
activity in a temporally nonspecific way and can
enhance the stability of sequential activity in the
excitatory neurons. To implement such a global inhi-
bition effect, each inhibitory interneuron receives
excitatory input from a large random set of projection
neurons, and each projection neuron in return
receives synapses from the interneurons. If the burst
times of the projection neuron population were
distributed across the entire song motif, then each
interneuron would receive roughly continuous drive
and would spike continuously.
Consistent with this picture, the temporal firing
patterns of inhibitory interneurons in HVC are
much less specific than those of the HVCRA neurons:
an inhibitory interneuron is typically active through-
out most of the motif. This suggests that the connec-
tivity of inhibitory neurons is also much less specific
than that of excitatory neurons.
Sequence Learning
Hand in glove with the question of what intrinsic
neural properties and network architectures could
underlie the stable propagation of activity in HVC is
the challenge of understanding how such architectures
could emerge from initially less-structured networks.
275
232 Birdsong Learning
Intuitively, associative (Hebbian) learning rules
such as spike-timing-dependent plasticity (STDP)
seem ideally suited to produce the kinds of asymmet-
ric connectivity characteristic of synaptic chains: if
neuron A fires before B, the connection from A to
B is strengthened, but that from B to A is weakened. If
a nave network is trained with this rule for synaptic
plasticity on a set of input patterns presented sequen-
tially, one would expect network weights to encode
the temporal correlations in the training patterns
and later to autonomously reproduce the same long,
stable chains of neural activity when stimulated.
But contrary to intuition, past modeling studies
show that associative learning rules are not sufficient
to configure networks to autonomously generate long
neural sequences. Networks trained in this way usu-
ally require finely tuned excitatory and inhibitory
connectivity that is continuously renormalized to bal-
ance network activity during learning, as well as
additional pacemaker inputs to propagate activity
from one time step to the next. The situation is even
worse if the network does not have access to training
inputs that are themselves sequential. In such cases,
the best response even with fine-tuning of parameters
is the formation of extremely short chains (fewer than
five steps) using thousands of neurons.
The reason for this is that associative learning
rules do not balance activity equitably across space
(neurons) and time; instead, they tend to amplify
inequities. For example, a neuron that by chance
receives strong synapses from its inputs will fire
often. Because of its frequent firing, it will tend to
rapidly form strong outgoing synaptic connections as
well. As a result, activity tends to accumulate around
a few overly active neural hubs, producing epileptic
or synchronous burstlike network states rather than
long, orderly chains. Empirically, these problems
remain unsolved with the imposition of local con-
straints on the maximum strengths of individual
synapses. For these reasons, the question of how
plasticity rules that are local in both space and time
can impose global constraints on activity balance
over all space and all time is nontrivial.
Two independent models provide a conceptual
framework for how associative learning rules must
be augmented to allow initially unstructured networks
to self-organize and generate stable, long sequences
of neural activation. Both models use STDP as the
primary activity-dependent mechanism for synaptic
change. But crucially, in addition to STDP, both mod-
els rely strongly on competitively enforced synaptic
bounds that are not computed on the level of single
synapses but are triggered by neuronwide constraints.
Although the rules governing synaptic remodeling are
not local to the individual synapses, they are local
to single neurons. Remarkably, these local neural
bounds nevertheless serve to uniformly and globally
distribute activity throughout the network in both
space and time.
The qualitative nature of these synaptic bounds,
and how they are triggered and enforced, is where
the two models diverge. The model of Jin and Jun
assumes that the formation of a finite number of
strong connections from a neuron triggers axon
remodeling and elimination of weak synapses. Fiete
et al. propose two separate schemes based on neural
resource constraints: one assumes that the production
rate of synapse-building resources is limited,
so potentiation in one synapse triggers a uniform
heterosynaptic weakening of all other synapses at
that neuron. The other scheme assumes a cap on the
total resources a neuron can devote to synaptic main-
tenance; therefore all synapses at a given neuron
undergo uniform weakening whenever STDP drives
their summed synaptic strength to exceed the cap.
Both models make predictions for the rules gov-
erning synaptic plasticity within HVC and for the
statistics of the resulting connectivity within HVC.
The Fiete et al. model produces several sequences
of different lengths, with a definite distribution of
chain lengths. Determining whether this distribution
matches syllable length distributions in birdsong could
indicate whether songbird syllables attain their specific
lengths largely due to bottom-up (low-level network)
properties or top-down (behavioral, evolutionary, or
respiratory) constraints.
Sensorimotor Learning of the Motor Map
The main goal-directed aspect of learning in zebra
finches takes place during the sensorimotor phase,
when the juvenile learns to match, through auditory
feedback and template comparison, its own string of
vocalizations with its memorized mental copy of the
tutor song. Because the goal-directed aspect of song
acquisition is so striking and because natural goal-
directed learning behaviors that are easy to quantify
are rare, several models have focused on this aspect of
song learning. Yet the microscopic synaptic plasticity
rules that underlie such learning remain experimen-
tally undiscovered and theoretically debated. Predic-
tions from existing theoretical or computational
models can serve to greatly narrow, in a principled
way, the dauntingly large number of possible plastic-
ity paradigms to investigate in experiment, and they
are now specific enough to be tested in detail in
slice preparations of the song system. (Experimental
testing of the proposed paradigms is contingent on
276
 Birdsong Learning 233

the identification of a reinforcement or critic signal, HVC

which is widely expected to exist but has not yet
been found.)
Existing models of goal-directed sensorimotor song
learning share the essential schema of the cartoon
model (Figure 2) for the roles assigned to the dif-
ferent premotor nuclei: HVC produces a high-level LMAN
sequence, and sensorimotor learning involves mapping
that abstract sequence into the correct vocal motor
sequence through plasticity in the HVCRA synapses. RA
The primary differences between models lie in the Spatially varying, temporally static perturbation and
roles assigned to the AFP (Figure 4), in the particular spatially uniform reinforcement
rules for how activity triggers plasticity in HVCRA a
synapses, and in their levels of biophysical realism
HVC
and detail. The plausibility of different assumed roles
for the AFP can already begin to be evaluated with
existing electrophysiological and lesion results.
Early experiments showed that lesioning the ter-
minal lateral magnocellular nucleus of the anterior
nidopallium (LMAN) of the AFP, which projects back
into the premotor pathway, abolishes the ability of LMAN
juvenile songbirds to learn song. Moreover, the imme-
diate and lingering effect of the LMAN lesion, besides
stoppage of song learning, is to dramatically reduce the RA
large trial-to-trial variability of juvenile song. These Spatially uniform reinforcement
results suggested that the role of LMAN in song b
learning is possibly to drive vocal experimentation HVC
through perturbation of the premotor song pathway.
An early, and in several respects prescient, com-
putational model is based in part on these observa-
tions. The model applied the specific algorithm of
weight perturbation (with momentum) from the rein-
forcement learning literature to suggest that LMAN
LMAN
directly generates small static weight changes in the
HVCRA connections (Figure 4(a)). According to
the model, these LMAN-driven static weight changes
RA
last for the duration of one song rendition, after
which they are selectively retained, based on correla- Spatially and temporally varying perturbation
tion with a song evaluation (reinforcement or critic) c
signal. The reinforcement (critic) signal is a function Figure 4 Proposed roles for lateral magnocellular nucleus of the
of the match between the actual and desired (tutor) anterior nidopallium (LMAN) in sensorimotor learning. The dashed
line indicates a plastic synapse between high-level vocal center
songs. Because the weight perturbation rule has been nucleus (HVC) and nucleus robustus archistriatalis (RA) neurons.
proven by others to perform gradient climbing on (a) The hypothesis of Doya and Sejnowski: LMAN must produce
the scalar reinforcement signal, it will improve song different, independent perturbations in each HVCRA weight ori-
performance and do so independent of most parameter ginating from different HVC neurons. These perturbations must be
values. Numerical simulation of a small network model static over the duration of one song rendition but vary across
renditions. LMAN must also broadcast a separate, spatially
of the birdsong network, consisting of rate-based units uniform reinforcement signal to the premotor pathway. (b) The
or neural assemblies with time coarse-grained into hypothesis of Troyer and Doupe. LMANs sole role is to provide
syllable-long chunks, has demonstrated how this may a spatially uniform reinforcement signal to the premotor pathway.
happen. Experimental studies have supported several Learning is associative (Hebbian). (c) The hypothesis of Fiete and
features of the model, including the high-level schema colleagues: LMAN must inject different perturbations to different
RA neurons. The perturbations are excitatory synaptic inputs to
of actorcritic reinforcement learning (Figure 5), the RA and vary in time over one song rendition. (A spatially uniform
role of perturbations in song learning, and the role of reinforcement signal is assumed to arrive at RA from elsewhere
the AFP in driving perturbations. in the network.)

277
234 Birdsong Learning
Critic Experimenter
Actor
a turbation in the HVCRA weights at the beginning of
HVC
Plastic?
RA LMAN
R
Empiric?
b 2050 inputs from LMAN. Thus, it is unlikely that
Figure 5 The empiric synapse hypothesis. (a) The empiric syn-
apse hypothesis for goal-directed learning consists of three parts:
The actor network, which is responsible for performing the desired
task and has plastic synapses, and two of its inputs, one an
experimenter, whose role is to drive perturbations in the actor
through empiric synaptic inputs, and the other a critic, which
provides through a reinforcement signal a scalar assessment of
the performance of the actor on the desired task. Actor neurons
are hypothesized to correlate experimenter-driven perturbations
with the reinforcement signal to derive a signal for synaptic
change, in the direction of the gradient of the reinforcement signal.
(b) In the song system, the actor network is hypothesized to be the
premotor pathway including high-level vocal center nucleus (HVC)
and nucleus robustus archistriatalis (RA) and plastic HVCRA
synapses. The experimenter is hypothesized to be the AFP.
A critic-generated reinforcement signal (R) is widely assumed to
exist in the song system, although its neural identity is unknown.
Adapted from Fiete IR, Fee MS, and Seung HS (2007) Model of
birdsong learning based on gradient estimation by dynamic per-
turbation of neural conductances. Journal of Neurophysiology 98:
20382057.
What are the biophysical assumptions and implica-
tions of this model? According to the model, both
the song evaluation (reinforcement) signal and the
perturbative drive are generated in the AFP, and deliv-
ered to the premotor pathway by LMAN. This means
that LMAN must carry and convey two independent
signals, for perturbation and reinforcement, to the
HVCRA synapses (Figure 4(a)).
The model is based on batch mode weight pertur-
bation, and it assumes that LMAN can somehow
directly perturb the weights of the HVCRA syn-
apses in a way that the perturbations vary from song
rendition to rendition but are static over the course
of a rendition. Moreover, for the weight perturbation
scheme to explore enough of the motor space for
song learning, HVCRA synapses that originate
from different HVC neurons must receive different,
and independent, perturbations from LMAN. There
are a few experimental inconsistencies with these
requirements in the birdsong system. First, single-
and multiunit activity in LMAN varies rapidly during
a song motif, without an unusually large transient at
the beginning. It is therefore difficult to imagine how
LMAN neurons could provoke an instantaneous per-
a song rendition, which is held constant for the rest of
the rendition. (It may be possible for weight perturba-
tions to be driven by slow neuromodulatory action of
LMAN onto the HVCRA synapses; within the
model, reinforcement would then have to be
mediated either by ordinary synaptic transmission
from LMAN to RA or through a second neuromodu-
lator released by LMAN neurons into RA.) Second,
each RA neuron receives an estimated 1000 inputs
from up to 200 HVC neurons but only approximately
synapses from different HVC neurons could receive
independent perturbations to fulfill the requirements
for weight perturbation.
Finally, Doya and Sejnowski found that learning
with purely random inputs from LMAN was too
slow, with poor convergence to the tutor song, even
though the simulated network was far smaller than
the songbird motor pathway and even though a song
motif consisted of very few time steps (each time
step represented an entire syllable). Therefore, they
assumed that LMAN activity itself evolves by rein-
forcement learning on the same global reinforcement
signal, so that its output is composed of a term pro-
portional to the true HVCRA weight gradient,
superimposed on a random component. With this
assumption, the AFP must perform at least three
functions: (1) generate and carry random perturba-
tion signals whose dimension might be as large as the
number of HVCRA neurons, (2) generate and carry a
true gradient estimate after correlating its random
activity with reinforcement, and (3) function as a
critic, generating and carrying the reinforcement sig-
nal. With these assumptions, the model indeed learns;
however, the AFP is assumed to possess these capabil-
ities de facto, without modeling how the necessary
computations are performed or how the resulting
signals are carried by LMAN and delivered to the
premotor pathway, leaving crucial questions unre-
solved. Whether the AFP actually possesses these cap-
abilities is an experimentally open question.
A model proposed by Troyer and Doupe in 2000
suggests that plasticity in HVCRA synapses and
RARA synapses may be largely associational or
Hebbian in nature. According to the model, plasticity
is gated by a reinforcement signal, which LMAN
278

delivers to the premotor network (Figure 4(b)). The
associational rule also includes a subtraction term for
synaptic weakening, bringing it closer to reinforce-
ment learning rules, but the specific form used is not
proven in the literature to guarantee improved perfor-
mance. As a result, there are no general assurances of
convergence or stability, even in networks of simplified
rate-based neurons. The model requires fine-tuning by
hand of several homeostatic mechanisms to maintain
stability during learning. The authors demonstrated
through numerical simulation in rate-based networks
that under particular parameter settings, with a small
set of RA assemblies, HVCRA connections, and a
few syllable-long time steps, the network can learn to
match a desired output.
A strength of this model is that, unlike others in
the literature, it stresses the importance of lateral
connections within RA, which, when subject to the
same plasticity rules as HVCRA synapses, form
assemblies that are capable of pattern completion
even with partially correct inputs, to generate good
output syllables. The AFPs only role in the model is
to generate and convey a spatially uniform reinforce-
ment signal to RA (Figure 4(b)). However, in the
zebra finch, LMAN activity is spatially inhomoge-
neous (neural activity is not strongly correlated across
the nucleus), and connections to RA are not highly
divergent, which implies that different RA neurons
are likely to receive fairly uncorrelated inputs from
LMAN rather than a uniform signal. As formula-
ted, the model cannot explain why, in experiments,
LMAN lesions reduce song variability. Experiments
fail to reveal qualitative changes in LMAN activity
when an adult birds auditory feedback is altered,
which suggests that LMAN may not primarily convey
an error or match signal. (If LMAN does generate
an error signal, but bases it on learned predictions
about auditory feedback, it is possible that prolonged
rather than transient exposure to distorted auditory
feedback is necessary to produce changes in LMAN
activity.)
Subsequent informal proposals, not grounded in
numerical simulation or modeling, have espoused
more information-rich primary roles for LMAN in
song learning. These include suggestions that LMAN
may provide detailed instructive or supervisory error
signals to RA. Unfortunately, these suggestions lack
mechanistic proposals or models of how the AFP
might correctly compute and generate these informa-
tion-rich, time-varying supervisory signals, leaving
the bulk of the computational burden of song learn-
ing unsolved and at the door of the AFP. In reality it
is possible that the AFP supplies such a supervisory
signal to RA, but no experimental or computational
studies exist yet to suggest that it does.
Birdsong Learning 235
A recent model of song learning by Fiete and colla-
borators returns to a reinforcement learning frame-
work and attempts to do two things: (1) propose that,
and illustrate how, spiking RA neurons might perform
reinforcement learning by using rapidly time-varying
LMAN synaptic inputs that perturb RA membrane
voltages (rather than using static synaptic pertur-
bations of the plastic HVCRA weights; Figure 4(c))
and (2) show as a proof-of-principle that although it is
widely thought to be slow, reinforcement learning
can be fast enough to explain song acquisition in
a full-scale model of the birdsong network with spik-
ing, conductance-based neurons, delayed feedback,
scalar reinforcement, and purely random independent
perturbations.
The model assumes that LMAN perturbs the pre-
motor pathway by injecting independent random
currents into different RA neurons (Figure 4(c))
through ordinary glutamatergic synaptic transmission
(Figures 5, 6(d), and (e)). This is consistent with recent
stimulation and pharmacological inactivation experi-
ments in LMAN, which show that LMAN helps drive
song variability through glutamatergic synaptic projec-
tions to RA. The model requires at most as many
independent LMAN inputs as there are RA neurons.
The LMAN inputs are assumed to be time varying
throughout song. Reinforcement is assumed to be com-
puted and delivered from elsewhere in the brain, with
learning following the high-level, actorcritic schema.
Under these conditions, the derivation of the rules for
how activity should modify synaptic weights, and the
rules themselves, are quite different from weight
perturbation-like algorithms (Figures 6(d)6(f)): each
HVCRA synapse must somehow deduce an instruc-
tion for change based on time-varying fluctuations
in the conductance of its postsynaptic RA neuron
rather than simply decide whether to retain or discard
direct static perturbations of its own weight through
correlation with the reinforcement signal. The resulting
rule is guaranteed to robustly (without parameter tun-
ing) improve song performance, even in recurrent net-
works of conductance-based spiking neurons and with
delayed reinforcement.
The model is able to produce learning in a biologi-
cally realistic network model of the song system,
with spiking neurons and an impoverished, delayed
reinforcement signal (Figures 6(a) and 6(b)). Learning
takes place online, and activity in the network evolves
on the millisecond timescale, driven by 56 ms bursts
in the HVC inputs. Within 2000 iterations, the net-
work output resembles the tutor song, which is a
recording of an actual zebra finch song (Figure 6(c)).
Fiete et al. have shown that learning time even with
random LMAN input should scale well in the full-
sized premotor network and may converge in far
279
236 Birdsong Learning

Tutor Before After

0

HVC
mV
60
48

mV
RA
56

arb units
Motor

5 kHz
10 30 50 70
a Time (ms) b 75 ms

400

300
Error

200

100

0
0 500 1000 1500 2000
c Iterations

HVC

Empiric LMAN
synapse
R
RA LMAN
d e

HVC

LMAN

R
f W > 0 W < 0 W = 0 W = 0
Figure 6 Empiric synapses: results and predictions. (a) Activity of two model high-level vocal center nucleus (HVC) neurons (top), two
representative nucleus robustus archistriatalis (RA) neurons (middle), and one of the motor output pools (bottom; black: tutor, blue:
network response) during song learning. (b) Spectrograms of the tutor song and the output of the model song network before and after
1200 iterations of learning, together with the respective sound pressure waves. (c) Mismatch between the network output and the tutor
song decreases relatively rapidly. Learning has reached an approximate asymptote by 1000 iterations. (d,f) Prediction for bidirectional
synaptic plasticity from the empiric synapse model. (d) Possible experimental design for testing the empiric synapse rule in the birdsong
pathway involves focused stimulation in HVC, intracellular recording in RA, and fiber bundle stimulation of the lateral magnocellular
nucleus of the anterior nidopallium (LMAN)RA axons, together with control of the reinforcement signal (R). (e) Magnified schematic of a
single RA neuron and its inputs. Each RA neuron receives a few empiric synapses from LMAN and several from different HVC neurons; in
addition, RA receives a reinforcement signal. Blue (gray): active (quiescent) premotor inputs; green: empiric input; red: reinforcement
signal. (f) Protocol for induction of bidirectional synaptic plasticity in active HVCRA synapses: Simultaneous stimulation of an HVC
neuron and the LMAN axon bundle, followed by administration of reinforcement, should lead to potentiation of all stimulated HVC
synapses.

fewer iterations than sung by the typical zebra finch
(100 000 renditions) during song acquisition.
The models central prediction is a specific paradigm
for the induction of bidirectional, heterosynaptic activ-
ity-dependent synaptic plasticity in the premotor vocal
pathway (Figure 6(f)) that should be possible to test
in slice experiments.
The model above is consistent with a large body
of experimental data in the song system and provides
concrete predictions for synaptic plasticity. How-
ever, LMANs role in the model is to provide purely
random exploration. This assumption does not con-
tradict the data but is a limited view of what the
AFP, a complex neural pathway, might be capable

280

of. More-sophisticated strategies for AFP-delivered
perturbation may speed up learning. This avenue is
wide open for discovery by theory and experiment.
More generally, models of song learning based
on the AFP as a source of random or directed pertur-
bations to the premotor system could be used as
comparative tools for studying the role of the basal
ganglia thalamocortical loop in mammalian sensori-
motor learning.
Premotor Codes and Representations
Up to 30 000 neurons are devoted to song production
and learning in the zebra finch. Together, these neu-
rons drive only eight to ten muscles that produce
song. What are some of the functional requirements
and costs that determine why so many neurons are
required for song production?
The ultrasparse or unary coding of song by HVCRA
neurons (Figures 2(a)) exacts a large energetic and
space cost in terms of the number of neurons required
to encode a given set of patterns. To produce a sequence
of 100 steps (L) with 200 neurons per time, unary
coding requires 20 000 neurons. At the other extreme,
a combinatorial code with a downstream readout capa-
ble of distinguishing between different combinatorial
patterns would require far fewer neurons, scaling as log
(L) instead of L; intermediate possibilities would
require intermediate numbers of neurons. What are
some of the properties of a unary code that might offset
its capacity costs and explain why it exists in HVC?
A possible explanation for ultrasparse HVC
sequences could be mechanistic: recurrent networks
cannot store many dense patterns because of increas-
ing interference between patterns. Studies show that
sparsely active attractor networks can store more
patterns. Nevertheless, in these studies, sparse is
far denser than unary, so the reason for ultrasparse
coding must lie elsewhere.
Because oscine birdsong is learned, a natural place
to investigate the role of unary HVC codes is in
learning. There are actually two distinct roles that
unary codes could play in song learning.
First, an analytical and numerical study by Fiete
et al. shows that unary coding within HVC enables
optimally fast learning of the feedforward motor
map from abstract sequences in HVC to sequences
of muscle activation. If each HVC neuron were to fire
twice instead of once, at random times in a song
motif of fixed length, the fastest possible speed with
which the inputoutput map can be learned is twice
as slow (Figures 7(a)7(d)). There is an intuitive
explanation of the underlying mathematical reason
for this effect. Suppose an HVC neuron and its
corresponding HVCRA synapses are active twice,
Birdsong Learning 237
when the tutor song has opposing characteristics
(e.g., pitch or amplitude), at those two times. The
synapses will be pulled in opposing directions to best
match the tutor song (Figure 7(c)) at each time.
Because the learning task is realizable, there is some
value of synaptic strength that achieves the desired
output, and in the end this value will be reached.
But the conflicting demands on synapse strength at
the two times in the motif make approach to the
solution with a hill-climbing rule slower, even if the
parameter controlling learning step size is optimized to
produce the fastest possible learning without network
instability.
Yet answers to why questions, however compel-
ling, risk being just-so stories unless they generate
specific testable predictions. The work described
above argued that unary codes may exist in HVC
because of their utility for fast song learning. How-
ever, single-unit HVCRA activity and unary codes
have been observed only in adults. Thus, the central
prediction which if false would invalidate the
hypothesis is that activity in HVCRA neurons,
even if nonsparse initially (early in the sensory
period), should be unary near the onset of senso-
rimotor learning in juvenile zebra finches instead of
emerging gradually after or toward the end of senso-
rimotor learning.
The analysis leads to an additional prediction. To
accommodate a larger song repertoire, HVC could
change in two a priori equally likely ways: (1) keep
HVC volume (number of neurons) fixed but increase
the number of bursts per neuron or (2) increase
HVC volume but maintain a unary code. Energetics
might favor the first possibility. But according to the
learning speed analysis, if zebra finches are under
pressure to learn their songs quickly, then unary cod-
ing should be conserved, even at the cost of a larger
HVC. This suggests that coding in HVC across zebra
finches and related songbirds with somewhat larger
repertoires (measured by number of unique syllables)
should remain unary while HVC size scales with rep-
ertoire size. Indeed, existing experimental findings
corroborate these predictions: HVC size scales with
repertoire size after correcting for overall brain size,
while RA, LMAN, or other song nuclei do not display
similarly consistent relationships.
Second, analysis and simulation by Fiete et al. has
shown that if the HVC code is unary and if, during
a song rendition, performance is continuously evalu-
ated and a reinforcement signal continuously delivered
(albeit delayed), then the time taken to learn a song
is independent of the length of the learned song.
Because the HVC code is unary in zebra finches, this
prediction can be used in experiments (where learning
time is monitored for birds trained on tutor songs of
281
238 Birdsong Learning

HVC1
HVC2

HVC3
a

b Time

HVC1 RA

HVC2 RA

HVC3 RA

Wj
Wi
Random nonunary representation
d Unary representation

HVC1

HVC2
e

HVC1 RA

HVC2 RA
g
Figure 7 Unary coding in high-level vocal center nucleus (HVC) helps speed the acquisition of the motor map. (a) The activity of
three hypothetical HVC neurons. The first two are each active only once in the song motif, but the third is active twice. (b) Hypothetical
pitch of the tutor song (black) and the pupil network (gray). (c) Direction in which synapses should change to reduce error between tutor
and pupil pitch. Synapses from HVC neuron 3 should be strengthened at the first and weakened at the second activity burst to improve the
tutorpupil match. Such conflicting demands on the synapses of neurons that are active at two or more random times in a motif cause a
slowdown in learning speed. (d) Contours of iso-error in the learning surface. Learning a feedforward map in a network with unary coding
in the top layer is like learning on an isotropic cost surface and can be fast. Denser coding in the input layer produces correlations and
makes the learning surface anisotropic. To keep the error from diverging along the steep directions, the learning rate must be kept low.
As a result, best-case learning is slower than in the isotropic case. HVC activity tuned to acoustic features may be helpful for generalizable
learning. (e) If HVC neurons fired multiple bursts at selected points when the acoustic features of tutor song are similar, rather than at
random times, there would be no interference in the learning update, (f). If HVC neurons acquired such a tuning to features in the tutor
song, it could be easy for the bird to quickly reproduce a specific heard sound by activating the requisite sound-tuned HVC neuron. RA,
nucleus robustus archistriatalis.

different lengths) as an indirect probe of whether the of song, then the time taken to learn a tutor song
elusive reinforcement signal in the song pathway is should scale linearly with its length. Similarly, if the
delivered online or in batch mode. If the reinforcement HVC code at each time consisted of a random pattern
signal is evaluated and delivered only once, at the end of activations, with half of all neurons active per time

282

(dense code), learning time should scale with song
length (even if reinforcement is delivered online).
Do other songbirds also display unary HVC repre-
sentations of song? At present, single-unit recordings
of RA-projecting HVC neurons are available only
from zebra finches. Learning with a pure time code
is nongeneralizable: like one piece of magnetic tape,
each pattern is used to produce song at only one
specific time. An entirely different set of HVC neu-
rons and HVCRA weights must be used and trained
to produce the same sound if it recurs elsewhere in the
motif. Therefore, songbirds that rapidly acquire and
imitate new songs (albeit after slowly acquiring a
first song) may use a different encoding strategy
(e.g., tuning codes based on acoustic features) from
a time-code (Figures 7(e)7(g)). In addition, song-
birds with vastly larger repertoires, regardless of
the speed of song acquisition, may display different
HVC codes because of capacity constraints.
Conclusions
Advances in the acquisition of behavioral and anato-
mical data and neurophysiological recordings from
awake, singing birds are rapidly turning the birdsong
circuit into an ideal system for unraveling the
mechanisms underlying motor control and learning.
Theoretical models, by quantitatively validating or
eliminating candidate explanations of network func-
tion, have helped lay a groundwork for understand-
ing the dynamical and coding principles that lead to
such functionality. Theoretical models also provide a
functionally motivated set of predictions about neural
activity, connectivity, coding, and plasticity that nar-
row the choices of future experiments and are pres-
ently ripe for experimental testing. Future experiments
and modeling in songbird species with more-flexible
learning behaviors should greatly enhance understand-
ing of generalizable motor learning and control.
Birdsong Learning 239
See also: Bird Song Systems: Evolution; Birdsong
Learning: Evolutionary, Behavioral, and Hormonal Issues;
Birdsong: The Neurobiology of Avian Vocal Learning;
Hebbian Plasticity; Learning, Action, Inference and
Neuromodulation; Motor Sequences; Signal Production
and Amplification in Birds; Spike-Timing-Dependent
Plasticity Models; Spiking Neuron Models; Synaptic
Mechanisms of Learning; Vocal Communication in Birds.
Further Reading
Brainard MS and Doupe AJ (2000) Interruption of a basal ganglia
forebrain circuit prevents plasticity of learned vocalizations.
Nature 404(6779): 762766.
Doya K and Sejnowski TH (1995) A novel reinforcement model of
birdsong vocalization learning. In: Tesauro G, Touretzky DS,
and Leen TK (eds.) Advances in Neural Information Processing
Systems, pp. 101108. Cambridge, MA: MIT Press.
Fiete IR, Burger L, Senn W, and Hahnloser RHR (2007) Enforcing
global coding constraints by synaptic competition a model for
the formation of long ultrasparse sequences in the songbird.
Submitted, 2007. (See also Society for Neuroscience Abstracts,
2005.)
Fiete IR, Fee MS, and Seung HS (2007) Model of birdsong learning
based on gradient estimation by dynamic perturbation of neural
conductances. Journal of Neurophysiology 98: 20382057.
Fiete IR, Hahnloser RH, Fee MS, and Seung HS (2004) Temporal
sparseness of the premotor drive is important for rapid learning
in a neural network model of birdsong. Journal of Neurophysi-
ology 92(4): 22742282.
Jin DZ, Ramazanoglu F, and Seung HS (2007) Intrinsic bursting
enhances the robustness of a neural network model of sequence
generation by avian brain area HVC. Journal of Computational
Neuroscience 23(3): 283299.
Jun JK and Jin DZ (2007) Development of neural circuitry for
precise temporal sequences through spontaneous activity, axon
remodeling, and synaptic plasticity. PLoS ONE 2(8): e723.
Li M and Greenside H (2006) Stable propagation of a burst
through a one-dimensional homogeneous excitatory chain
model of songbird nucleus HVC. Physical Review. E, Statistical,
Nonlinear, and Soft Matter Physics 74(1.1): 011918.
Troyer TW and Doupe AJ (2000) An associational model of
birdsong sensorimotor learning I. Efference copy and the
learning of song syllables. Journal of Neurophysiology 84(3):
12041223.
283


284
 Open access, freely available online

Primer

The Neural Basis of Birdsong

Fernando Nottebohm

T
here is a tradition in biology of using specic animal
models to study generalizable basic properties of a
system. For example, the giant axon of squid was used
for the pioneering work on nerve transmission; the fruit y
(Drosophila) has played a key role in researchers discovering
the role of homeobox genes in embryogenesis; the sea slug
(Aplysia) is used to study the molecular biology of learning;
and the round worm (Caenorhabditis elegans) is used to study
programmed cell death. Basic insights gained from these four
systems apply widely to other multicellular animals. Here, I
will review basic discoveries made by studying birdsong that
have helped answer more general questions in vertebrate
neuroscience.

Vocal Learning and the Song System

Oscine songbirds (e.g., zebra nches, canaries, and white-
crowned sparrows) learn their song by imitating those of
older members of their own species [1,2]. This is done
by modifying vocal output until the auditory feedback
it generates matches a memorized model [3]. In some
birds vocal learning gives rise to easily discernible song
dialects, which then act as local cultural traditions [4]. In
most songbirds mastery of a song model takes many weeks.
Song learning starts with a stage that has been likened to
human infant babbling called subsong, during which
highly variable, low-amplitude sounds are produced in a
non-communicatory context, often while the juvenile seems
to doze. The sounds of subsong provide the raw material
from which imitations emerge. As these imitations become
recognizable, they are referred to as plastic song. As the
imitations are perfected, song becomes less and less variable.
The stable song typical of adults is in place by the time the
sexually mature bird is ready to start to defend a territory
and woo a mate. Intriguingly, in birds as in human infants,
the path of vocal change that culminates with imitation of a
model can be very idiosyncratic, as if this were an exercise in
problem solving for which there is no single solution [5].
The acquisition and production of learned song is made
possible by a group of discrete brain nuclei and their
connecting pathways, referred to as the song system [6,7],
which has similarities in the three groups of birdssongbirds,
parrots, and hummingbirdsthat evolved learned song
[8,9]. This system, described in considerable detail in oscine
songbirds, has two main branches: the posterior descending
pathway (PDP), necessary for both the acquisition and
production of learned song, and the anterior forebrain
pathway (AFP), necessary for acquisition only (see Figure 1).
The high vocal center (HVC) is at the starting point of both
these pathways, but the HVC cells that project to the PDP and
AFP differ. In mammalian terms the PDP is homologous to a
motor pathway that starts in the cerebral cortex and descends
through the brain stem [6], while the AFP is homologous to
Primers provide a concise introduction into an important aspect of biology
highlighted by a current PLoS Biology research article.
DOI: 10.1371/journal.pbio.0030164.g001
Figure 1. The Song System of Songbirds
Nucleus HVC feeds information into two pathways that
ultimately lead to the neurons in the tracheosyringeal half of the
hypoglossal nucleus (nXIIts) that project to vocal muscles. HVC
projects to nucleus RA directly (PDP), and indirectly via Area X,
the dorsolateral anterior thalamic nucleus (DLM), and LMAN
(AFP) in a manner that shares similarities with the mammalian
pathway cortexbasal gangliathalamuscortex.
a cortical pathway through the basal ganglia and thalamus
[7,10,11].
Several of the telencephalic nuclei that participate in
the production and acquisition of learned song are small
in nestlings, before the onset of song development, and
their volume, cell number, cell size, and connections grow
during the subsequent weeks or months. As a result of these
changes, many of the components of the circuits for the
acquisition and production of learned song are formed and
connected during the very period when song rst develops
(reviewed in [12]). Another peculiarity of this system is that
Citation: Nottebohm F (2005) The neural basis of birdsong. PLoS Biol 3(5): e164.
Copyright: 2005 Fernando Nottebohm. This is an open-access article distributed
under the terms of the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the
original work is properly cited.
Abbreviations: AFP, anterior forebrain pathway; HVC, high vocal center; LMAN,
lateral magnocellular nucleus of the nidopallium; PDP, posterior descending
pathway; RA, robust nucleus of the arcopallium
Fernando Nottebohm is a professor at Rockefeller University in New York, New York,
United States of America, and head of the Rockefeller Universitys Laboratory of
Animal Behavior, and Director of the Rockefeller University Field Research Center for
Ethology and Ecology in Millbrook, New York. E-mail: nottebo@rockvax.rockefeller.
edu
DOI: 10.1371/journal.pbio.0030164

PLoS Biology | www.plosbiology.org 0759 May 2005 | Volume 3 | Issue 5 | e164

285
the right and left sides of the brain can operate, to some
extent, independently, each responsible for a different array
of sounds. In birds such as the canary, the chafnch, and the
white-crowned sparrow a majority of the sounds of song are
produced by the left syringeal half, under the control of left,
uncrossed pathways. This phenomenon has been referred to
as left hypoglossal or left hemispheric dominance [13].
Adult Variation, Neurogenesis, and Neuronal
Replacement
The song system of birds is sexually dimorphic: it is better
developed in males, which usually sing more and produce
a more complex repertoire than females. For example, the
nucleus HVC of canaries is three times larger in males than
in females; in zebra nches it is eight times larger [14]. In
seasonal singers such as the canary and song sparrow, song
system nuclei such as HVC are signicantly larger in the
spring than in late summer, after breeding stops [15,16].
Cells in many of the song control nuclei are androgen and
estrogen sensitive [17]. The nucleus HVC of adult female
canaries treated with physiological doses of testosterone
doubles in volume, and these birds start to sing in a male-like
manner [18]. Initially such changes were thought to result
solely from growth of dendritic trees and synapse formation
[19], but subsequently it was found that new neurons were
added, too. These new neurons, as in embryos, are born
in the wall of the forebrains lateral ventricle [20,21].
Interestingly, the addition of new neurons to nucleus HVC
occurred also in male and female adult canaries that had
received no hormonal treatment [22].
Earlier claims of neurogenesis in the adult mammalian
brain [23,24] had met resistance [25]. Nucleus HVC yielded
the rst unambiguous example of adult neurogenesis, in
that individual cells labeled with a cell birth marker provided
neurophysiological recordings that were unmistakably
neuronal [26]. We now know that the recruitment of new
HVC neurons is part of a process of constant replacement
[27,28]. This replacement is particularly active in canaries
during seasonal changes in the song repertoire [29]. Male
canaries develop a new song repertoire each year. New
neurons are constantly added, as well, to many regions of the
adult avian telencephalon, where they are probably involved
in a variety of brain functions. There is no evidence of
neuronal addition to other parts of the adult songbird brain
[22]. We now know that adult neurogenesis and neuronal
replacement are probably common to all vertebrates [30].
The song system of birds helped change the way in which we
think of brain circuits and their potential for rejuvenation
and repair. Just as important, the discovery of neuronal
replacement has raised basic questions about the brain
variables that set limits to learning [31].
Neurophysiology Offers Insights on the Mechanisms
for Vocal Learning
From early on, the song system drew the attention of
neurophysiologists. It was known that lesions of HVC and
the robust nucleus of the arcopallium (RA) affected the
organization of song differently, the former being more
devastating than the latter [6]. Likewise, stimulation of
HVC during song interrupted and reset the song program,
something that did not happen if the stimulating electrode
was in RA [32]. This hierarchical relation between HVC and
PLoS Biology | www.plosbiology.org
RA was conrmed by recording from HVC and RA while the
bird sang [33], but the manner in which the sounds of song
were represented in HVC remained unclear. This issue was
resolved by recording from individual HVC neurons that
projected to nucleus RA. These neurons, it was shown, red
very sparsely and at narrowly dened times, each neuron
ring always during the same six-millisecond window while
the bird produced its single learned song [34]. The inference
that these neuronsand the PDP of which they are part
carried the learned pattern of song seems inescapable. Since
these are the very HVC neurons that are replaced when birds
modify their song [31], it follows that the replacement cells
learn their score.
But what, then, is the role of the AFP in song learning? It
was known that the AFP was necessary for the acquisition but
not for the production of learned song [35]. It was known,
too, that the variable song typical of juvenile songbirds
became very stereotyped after bilateral lesions of the lateral
magnocellular nucleus of the nidopallium (LMAN) (part
of the AFP), from which it was inferred that LMAN played
a crucial role in fostering circuit plasticity necessary for
learning [36]. But the mechanism for this effect remained
unknown. Two recent independent studies now show how
this effect comes about [37,38]. In this issue of PLoS Biology,
lveczky and colleagues show that the LMAN neurons
that project to RA re in a quasi-random pattern when
variable song is produced in juvenile birds. Thus, while
the HVCRA projection carries the learned song, the
LMANRA projection carries the jitter that induces the
variability in motor output necessary for the imitation of a
model. This jitter, presumably, is imposed on the ring of the
same RA neurons that receive the more orderly output from
HVC. When the LMAN neurons are silent (or absent), the
HVCRA pathway produces a stereotyped pattern; when the
LMANRA neurons are ring, song is more variable. This is
a most elegant breakthrough. However, it is not clear exactly
how this pathway functions in song learning. One possibility
is that birds trying to imitate a model succeed by retaining,
from the variability generated, those patterns that more
closely approximate the model and discard the rest, thus,
over a period of time, achieving a perfect imitation. A second
possibility is that the variable mismatch between a model and
the attempted imitation drives output modication, so that
patterns that had not occurred before now rst appear. Both
mechanisms would depend on auditory feedback. It is the
rst time we are so close to a mechanism for vocal learning.
Open Questions
The stage is set for many more insights. An unsolved
question is the extent to which the very pathways that
produce learned song may also partake in the perception
of song. On a different front, why is it that HVCRA
neurons are periodically replaced? It was thought that
changes in dendritic conguration, dendritic spines, and
synaptic number and efciency provided all the plasticity
needed to change circuit conguration and explain how
new information was acquired and remembered. But if so,
why replace whole neurons? Is it possible that dendritic
and synaptic changes underlying learning are less easy to
achieve in older neurons? If so, are older neurons replaced
to reinstate a level of plasticity necessary for learning? Or
might it be that in some cases the whole neuron, rather than
0760 May 2005 | Volume 3 | Issue 5 | e164
286
the synapse, is the unit of learning? In this latter scenario,
changes associated with learning would be committed
as permanent gene expression changes akin to those
characterizing cellular differentiation. Such a change would
be a very stable way to encode learning, but it would have a
major drawback: the more learning that occurred, the fewer
neuronal pupils would remain. Thus, we are left to wonder
whether neuronal replacement takes place to make up for
the lost plasticity of aging neurons, or whether it takes place
as part of a normal recycling of memory space and of the
memories it holds.
During the early 1970s, before the song system was
discovered, it was widely believed that the learning of any
one skill had a wide representation in the vertebrate brain
[39]. The discovery of discrete brain regions devoted to song
learning and execution in the bird brain helped change
that view. It was also widely believed that the brains of male
and female vertebrates were virtually identical, with small
allowances for the levels of circulating hormones. The song
system changed that, too. And it was widely believed that
the anatomy of adult brains was set, but we now know that
the volume of brain structures can change seasonally and in
response to blood hormone levels. Most importantly, it was
widely held that though a straggling few neurons might still
be added after birth to late-developing parts of the brain,
brain cells, once lost, could not be replaced. Again, work
on the song system changed the prevailing view. It may well
be that our best understanding of how complex skills are
acquired and how broken circuits can be xed will come not
from humans, or other primates, but from the way birds learn
their song. 
References
1. Thorpe WH (1958) The learning of song patterns by birds, with special
reference to the song of the chafnch, Fringilla coelebs. Ibis 100: 535570.
2. Marler P (1970) A comparative approach to vocal learning: Song learning
in white-crowned sparrows. J Comp Physiol Psychol 71: 125.
3. Konishi M (1965) The role of auditory feedback in the control of
vocalization in the white-crowned sparrow. Z Tierpsychol 22: 770783.
4. Marler P, Tamura M (1964) Culturally transmitted patterns of vocal
behavior in sparrows. Science 146: 14831486.
5. Liu WC, Gardner TJ, Nottebohm F (2004) Juvenile zebra nches can use
multiple strategies to learn the same song. Proc Natl Acad Sci U S A 101:
1817718182.
6. Nottebohm F, Stokes TM, Leonard CM (1976) Central control of song in
the canary, Serinus canaria. J Comp Neurol 165: 457486.
7. Vates GE, Vicario DS, Nottebohm F (1997) Reafferent thalamo-cortical
loops in the song system of oscine songbirds. J Comp Neurol 380: 275290.
8. Paton JA, Manogue KR, Nottebohm F (1981) Bilateral organization of the
vocal control pathway in the budgerigar, Melopsittacus undulates. J Neurosci
1: 12791288.
9. Jarvis ED, Ribeiro S, da Silva ML, Ventura D, Vielliard, et al. (2000)
Behaviorally driven gene expression reveals song nuclei in hummingbird
brain. Nature 406: 628632.
10. Bottjer SW, Johnson F (1997) Circuits, hormones, and learning: Vocal
behavior in songbirds. J Neurobiol 33: 602618.
11. Luo M, Perkel DJ (1999) A GABAergic, strongly inhibitory projection to a
thalamic nucleus in the zebra nch song system. J Neurosci 19: 67006711.
PLoS Biology | www.plosbiology.org
12. Nottebohm F (1999) The anatomy and timing of vocal learning in birds.
In: Hauser MD, Konishi M, editors. The design of animal communication.
Cambridge (Massachusetts): MIT Press. pp. 63110.
13. Nottebohm F (1977) Asymmetries in neural control of vocalization in the
canary. In: Harnad S, editor. Lateralization in the nervous system. New
York: Academic Press. pp. 2344.
14. Nottebohm F, Arnold AP (1976) Sexual dimorphism in vocal control areas
of the songbird brain. Science 194: 211213.
15. Nottebohm F (1981) A brain for all seasons: Cyclical anatomical changes in
song control nuclei of the canary brain. Science 214: 13681370.
16. Brenowitz EA, Nalls B, Wingeld JC, Kroodsma DE (1991) Seasonal
changes in avian song nuclei without seasonal changes in song repertoire. J
Neurosci 11: 13671374.
17. Arnold AP, Nottebohm F, Pfaff DW (1976) Hormone concentrating cells
in vocal control and other areas of the brain of the zebra nch (Poephila
guttata). J Comp Neurol 165: 487512.
18. Nottebohm F (1980) Testosterone triggers growth of brain vocal control
nuclei in adult female canaries. Brain Res 192: 89107.
19. DeVoogd TJ, Nottebohm F (1981) Gonadal hormones induce dendritic
growth in the adult brain. Science 214: 202204.
20. Goldman SA, Nottebohm F (1983) Neuronal production, migration and
differentiation in a vocal control nucleus of the adult female canary brain.
Proc Natl Acad Sci U S A 80: 23902394.
21. Alvarez-Buylla A, Nottebohm F (1988) Migration of young neurons in adult
avian brain. Nature 335: 353354.
22. Nottebohm F (1985) Neuronal replacement in adulthood. Ann N Y Acad
Sci 457: 143161.
23. Altman J (1963) Autoradiographic investigation of cell proliferation in the
brains of rats and cats. Anat Rec 145: 573591.
24. Kaplan MS, Hinds JW (1977) Neurogenesis in the adult rat. Electron
microscopic analysis of light radioautographs. Science 197: 10921094.
25. Rakic P (1985) Limits of neurogenesis in primates. Science 227: 154156.
26. Paton JA, Nottebohm F (1984) Neurons generated in adult brain are
recruited into functional circuits. Science 225: 10461048.
27. Kirn JR, Nottebohm F (1993) Direct evidence for loss and replacement of
projection neurons in adult canary brain. J Neurosci 13: 16541663.
28. Scharff C, Kirn J, Grossman J, Macklis, Nottebohm F (2000) Targeted
neuronal death affects neuronal replacement and vocal behavior in adult
songbirds. Neuron 25: 481492.
29. Kirn JR, OLoughlin B, Kasparian S, Nottebohm F (1994) Cell death and
neuronal recruitment in the high vocal center of adult male canaries are
temporally related to changes in song. Proc Natl Acad Sci U S A 19: 7844
7848.
30. Gross CG (2000) Neurogenesis in the adult brain: Death of a dogma. Nat
Rev 1: 6772.
31. Nottebohm F (2002) Why are some neurons replaced in adult brain? J
Neurosci 22: 624628.
32. Vu ET, Mazurek MK, Kuo YC (1994) Identication of a forebrain motor
programming network for the learned song of zebra nches. J Neurosci 14:
69246934.
33. Yu AC, Margoliash D (1996) Temporal hierarchical control of singing in
birds. Science 273: 18711875.
34. Hahnloser RHR, Kozhevnikov AA, Fee MS (2002) An ultra-sparse code
underlies the generation of neural sequences in a songbird. Nature 419:
6570.
35. Bottjer SW, Miesner EA, Arnold AP (1984) Forebrain lesions disrupt
development but not maintenance of song in passerine birds. Science 224:
901903.
36. Scharff C, Nottebohm F (1991) A comparative study of the behavioral
decits following lesions of various parts of the zebra nch song system:
Implications for vocal learning. J Neurosci 11: 28962913.
37. Kao HM, Doupe AJ, Brainard MS (2005) Contributions of an avian basal
ganglia-forebrain circuit to real-time modulation of song. Nature 433:
638643.
38. lveczky BP, Andalman AS, Fee MS (2005) Vocal experimentation in the
juvenile songbird requires a basal ganglia circuit. PLoS Biol 3: e153. DOI:
10.1371/journal.pbio.0030153
39. Lashley KS (1950) In search of the engram. In: Physiological mechanisms in
animal behaviour, Symp Soc Exp Biol IV. Cambridge: Cambridge University
Press. pp. 454482.
0761 May 2005 | Volume 3 | Issue 5 | e164
287


288
letters to nature
Decrystallization of adult
birdsong by perturbation
of auditory feedback
Anthony Leonardo & Masakazu Konishi
Computation and Neural Systems Program and Division of Biology, MC 216-76,
California Institute of Technology, Pasadena, California 91125, USA
.........................................................................................................................
Young birds learn to sing by using auditory feedback to compare
their own vocalizations to a memorized or innate song pattern;
if they are deafened as juveniles, they will not develop normal
songs1,2. The completion of song development is called crystal-
lization. After this stage, song shows little variation in its tem-
poral or spectral properties. However, the mechanisms underlying
this stability are largely unknown. Here we present evidence that
auditory feedback is actively used in adulthood to maintain the
stability of song structure. We found that perturbing auditory
feedback during singing in adult zebra finches caused their song to
deteriorate slowly. This decrystallization consisted of a marked
loss of the spectral and temporal stereotypy seen in crystallized
song, including stuttering, creation, deletion and distortion
of song syllables. After normal feedback was restored, these
deviations gradually disappeared and the original song was
recovered. Thus, adult birds that do not learn new songs never-
theless retain a significant amount of plasticity in the brain.
The song of zebra finches consists of three levels of organization:
syllables, which are individual sound components of the song
separated by silent intervals; motifs, which are sequences of
syllables; and bouts, which are sequences of motifs3. The spectral
structure of the syllables is unstable in the juvenile bird. Similarly,
motifs and bouts are organized differently from song to song.
However, as the song gradually assumes its adult form, these
different levels of organization become highly stereotyped, the
variability of the spectral structure of the syllables becomes
extremely small, and the bird sings these syllables in a highly
predictable order. At this stage, the song is referred to as crystallized.
Some birds, like the zebra finch, maintain their crystallized song
throughout adulthood and are called age-limited learners. Open-
ended learners like canaries can, in contrast, learn new songs in
adulthood4.
The stability of song in age-limited learners was previously
thought to be maintained without auditory feedback1,2. Recent
reports, however, show that deafening these birds after crystal-
lization causes changes in song, suggesting that some auditory
feedback is important for song maintenance throughout life57.
Six to eight weeks after deafening, the song of adult zebra finches
deteriorates, and shows addition and deletion of syllables, abnormal
repetition of syllables (stuttering), and modified syllable sequences.
By opening the auditory feedback loop, deafening shows how well
the song pattern generator can maintain its original output without
this signal. However, to learn how the song control system works,
it is necessary to manipulate auditory feedback without disabling

Time (ms) 100 200 300 400 500 600 700

What the bird sings

Adaptive 100 200 300 400 500 600 700

Trigger Playback Trigger Playback Trigger Playback Trigger Playback

Detect singing

Generate feedback

What the bird hears

Syllable-triggered 100 200 300 400 500 600 700

Trig. Play Play Trig. Play Play

Detect target syllable

Generate feedback

What the bird hears

Figure 1 Protocols for constructing the feedback signals. In the adaptive protocol, monitored the birds song in 50-ms bins and detected the production of a single
the computer alternated between recording vocalizations and playing the last target syllable. In both protocols, the bird heard the superposition of his own
vocalization back to the bird. In the syllable-triggered protocol, the computer vocalizations and the computer-generated feedback.

466
1999 Macmillan Magazines Ltd 289
NATURE | VOL 399 | 3 JUNE 1999 | www.nature.com

either the auditory or vocal motor system. The ability to modify
feedback signals dynamically allows particular spectral and tem-
poral components of the song to be chosen for manipulation.
Furthermore, the effects of restoring normal auditory feedback
after exposure to abnormal feedback can be observed. We show
here how these manipulations produce dramatic changes in the
adult song.
We developed a computer-controlled system to perturb auditory
feedback and designed two feedback paradigms (Fig. 1). In both
methods, the computer detected singing and then played a feedback
signal to the bird. Because the playback signals were delivered
through an overhead speaker, the birds heard a superposition of
natural and artificial sounds. In the first paradigm, the adaptive
protocol (n 3 birds), a computer alternated between recording
letters to nature
vocalizations and playing back the last vocalization to the bird. The
position of the feedback varied, depending on the exact timing of
the birds song, and the feedback changed as the song changed (see
Methods). To investigate the effects of feedback on a more local
level, we designed a second syllable-triggered protocol (n 2 birds)
in which we perturbed the feedback of only a single target syllable.
The computer recognized and played back a stored copy of this
syllable each time the bird produced it. The timing of the feedback
signal was fairly constant across different song deliveries.
We recorded the songs of five adult male zebra finches in separate,
sound-attenuated chambers for several weeks before beginning the
experiment. The variability of their songs was well within that found
in crystallized songs. The birds were then placed in the feedback
system. After 14 months in this environment, four of the five birds

Motif 1 = Motif 2

i i A B C D E i A B C D E
9,000
Frequency (Hz)

0
0 2
Time (s)

Stuttering

A B B B B B B B B i A B
9,000
Frequency (Hz)

b A

0
0 3
Time (s)

Motif 1 = Motif 2 = Motif 3

i A B C D E F G i A B B C D E i A B H i A
9,000
Frequency (Hz)

0
0 4
Time (s)

Figure 2 Normal and decrystallized birdsong spectrograms. a, Two motifs of a
normal, crystallized zebra finch song. Each syllable has a stable and well defined
spectral structure. The ordering of the song syllables is identical in the two motifs.
The syllable sequence A B C D E represents the baseline motif for this bird. The
is represent introductory notes, which occur in variable numbers at the
beginning of motifs. b, The decrystallized song of the bird shown in a, with a
motif containing stuttering of the song syllable B. c, Another example of the
decrystallized song of the bird shown in a. The syllable sequences now vary from
motif to motif, indicating a loss of the fundamental temporal stereotypy which
characterizes crystallized zebra finch song. Motifs one and three contain new
song syllables (F, G, H) which were not present in the baseline motif shown in a.
Motif two contains a repetition of syllable B. After the feedback was removed, the
sequences shown in b and c gradually disappeared and only the baseline song
was produced.

NATURE | VOL 399 | 3 JUNE 1999 | www.nature.com

1999 Macmillan Magazines Ltd 290 467
letters to nature
showed dramatic changes in their songs. To study the progression of
these changes, we allowed some of the birds to sing without artificial
feedback on a randomly chosen 1015% of their song deliveries one
day per week. All of the analysed data consisted of recordings of the
bird singing by himself, with no artificial feedback. After the
feedback was permanently stopped, we tracked all the birds for
another 816 months to determine whether they could recover their
original songs.
Decrystallization of the song occurred in both global song
organization and local spectral structure, and consisted of the
emergence of new spectral and temporal properties and increased
occurrence of properties that were rare in the baseline song. We use
10,000
Frequency (Hz)
a sets of frequencies that are integer multiples of a common
0
0 120
Time (ms)
10,000
Frequency (Hz)
b the remainder of the feedback period. No changes were seen in the
0 initially tonal syllable was sometimes produced as a distorted
0 120
Time (ms)
10,000
Frequency (Hz)
c from the microphone into a series of syllable strings, such as A B C
0 one month of perturbed feedback, the probability of the baseline
0 120
Time (ms)
Figure 3 Decrystallization of a single syllable. a, Baseline (pre-feedback) version
of the syllable. b, After one week of syllable-triggered feedback. Harmonic
frequencies (white lines) have appeared around the single frequency in the early
portion of the syllable (at t ,30 ms). c, After one month of feedback, additional
harmonic frequencies now stretch throughout the duration of the syllable.
468
1999 Macmillan Magazines Ltd
the term decrystallization to refer collectively to all of the perturbation-
induced changes from the original quantitative and statistical
structure of the song; this does not necessarily imply a return to a
juvenile state of song structure. These changes include stuttering,
creation, deletion and distortion of song syllables. As
the decrystallization progressed, the proportion of normal songs
decreased and that of abnormal ones increased. However, baseline
songs were still produced with low probability even at the peak of
song degradation.
The changes seen in the three adaptive protocol birds were very
similar to those seen in deafened birds57. A series of spectrograms
representative of these results is shown in Fig. 2. Stuttering occurred
in all three adaptive-protocol birds and was the most dramatic
change in song organization. Both complex syllables and modified
introductory notes were stuttered. A secondary effect of stuttering
was a substantial increase in the maximum song length. Other
changes to song organization induced by the feedback were the
addition of new syllables to the song and, infrequently, the deletion
of old syllables from the song. Many zebra finch syllables contain
fundamental frequency; such sets are called harmonic stacks (Fig.
2a, syllable D). The same three birds also showed spectral distortion
in their song syllables, including wobbles in the harmonic structure
of simple notes and the production of two superimposed harmonic
stacks, indicating a loss of precise control over the vocal organ (the
syrinx)8,9. There was considerable variability in the magnitude and
time course of the changes between different birds, but significant
changes were generally seen within six weeks. Finally, all the changes
in song structure described above could occur within different
motifs in the same bout. This is significant because one of the
hallmarks of crystallized song is its robust temporal sterotypyan
identical syllable sequence is maintained within all the motifs of a
bout. Decrystallized song, in contrast, lacks this stereotypy (Fig. 2c).
Two birds received feedback in the syllable-triggered protocol. In
one of the birds, significant changes in the spectrum of the target
syllable appeared in less than a week and increased in magnitude for
spectrum or the ordering of any of the other syllables in the song.
The changes to the target syllable consisted of the appearance of
harmonic frequencies around previously single-frequency portions
of the syllables spectrum (Fig. 3). The presence of these additional
harmonics grew more frequent with time, until eventually this
harmonic stack. Humans can also alter the spectral structure of a
sound in response to altered auditory feedback10. The second bird
used in this protocol showed no changes in his song after 11 weeks
of feedback perturbation.
Decrystallization essentially consisted of a large increase in the
variability of the song. To quantify the changes associated with the
arrangement and variability of syllable sequences, we developed an
automatic method to sort the data acquired on a given day into
syllable types by using the different spectral and temporal features of
each syllable. This transformed the raw voltage waveforms recorded
D E, from which we calculated the probability of different syllable
sequences. We define the baseline motif as the most probable
sequence of syllables that the bird repeated in a bout before the
feedback period began. Figure 4a shows the probability of singing
the baseline motif as a function of time for one bird. After receiving
motif for this bird decreased by a factor of six.
The probabilistic sequencing of syllables by the bird on a
particular day can be fully characterized as a Markov chain11.
Thus, the likelihood of singing a particular syllable depends only
on the occurrence of the last syllable produced, and not on any prior
syllables. The birds song over the course of the experiment can then
be described as a Markov chain which evolves in time (Fig. 4a), and
291
NATURE | VOL 399 | 3 JUNE 1999 | www.nature.com

0.8
Probability (baseline motif)
a
0
0 540
Time (days of recovery)
Conditional entropy (bits)
1.5
0.5
0 540
Time (days of recovery)
Figure 4 Time course of decrystallization and recovery. a, The probability of one
bird singing his baseline motif as a function of time. The shaded area represents
feedback interval. Circles represent the probability calculated from the entire
sequence of syllables (Prob {A B C D E}); asterisks represent the probability
the conditional entropy of the Markov chain can be used as an
estimate of the variability of the song on a particular day (Fig. 4b)12.
The conditional entropy measures the uncertainty in observing
syllable B, given that syllable A was just produced. If each syllable is
followed only by a single other syllable (only one type of syllable
sequence is produced), then the conditional entropy will be low,
whereas if any syllable can follow any other syllable (many different
sequences are produced), the conditional entropy will be high.
We examined the data for all three adaptive protocol birds using
this method and found that birds in a decrystallized state had a
significantly higher conditional entropy than they did in their
baseline state. Taken together, the time course of the baseline
motif probability and the conditional entropy shows that the bird
in Fig. 4 went from being a low-variance singer, with essentially a
single motif, to being a high-variance singer who produced a large
number of different motifs.
After the removal of the artificial feedback, the temporal and
spectral variability in the songs of all three adaptive protocol birds
slowly decreased over the following weeks and months. The songs
eventually recrystallized and became stable again. Both the prob-
ability of the baseline motif and the conditional entropy of the song
returned to their baseline levels. Stuttering, abnormal sequencing
of syllables and modified spectral organization gradually became
infrequent and were replaced by the temporal and spectral organi-
zations characteristic of the original song. A complete recovery took
about 24 months. The syllable-triggered bird made a partial
recovery by 8 months after the cessation of feedback. The slow
progression of these changes suggests that auditory feedback does
not exert a great deal of instantaneous control over the production
of song, but instead has a cumulative effect on song maintenance.
The recrystallized songs appeared to remain stable indefinitely. We
tracked one bird for a year after his recovery and saw no departures
from the baseline song structure.
Our results demonstrate that zebra finches need auditory feed-
back to maintain their songs in adulthood. This species, which does
not modify its song or learn new songs after crystallization, appears
to retain a great deal of plasticity in its auditoryvocal control
system. This plasticity is sufficient to produce modifications in both
NATURE | VOL 399 | 3 JUNE 1999 | www.nature.com
1999 Macmillan Magazines Ltd
letters to nature
o P(ABCDE)
p Markov Chain
calculated from the syllable-to-syllable transition probabilities of the Markov
chain for the song (Prob{A}*Prob{B|A}*Prob{C|B}*Prob{D|C}*Prob{E|D}). The two
curves are identical to within 4% error. b, The conditional entropy of the song,
which is a measure of song variability, for the same bird.
the temporal pattern of song organization and the spectral structure
of individual syllables. This finding is not consistent with the
classical depiction of song development in which a dynamic learn-
ing period in youth ends in a static maintenance period in adult-
hood. Thus, the distinction between age-limited and open-ended
learners may not be as sharp as these names would indicate.
.........................................................................................................................
Methods
Birds were housed in custom-designed plexiglass cages and were paired with a
female who lived in a separate partition of the cage. The playback speakers
output was calibrated to be approximately the same as the sound level of the
birds vocalizations (8090 dB SPL in the birds ear). At the start of the baseline
song recordings, the birds ranged from 130 to 300 days old (mean age was 200
days; zebrafinches reach adulthood at 90 days). The singing rates of the
different birds ranged from tens to thousands of motifs per night. However,
there was no apparent correlation of age or singing volume with the magnitude
of the effects we observed. For the two birds who did not sing frequently, we
collected no baseline data during the feedback period.
Adaptive feedback perturbation. Microphone data were sampled at 40 kHz,
after being low-pass filtered (10 kHz cutoff, 7-pole anti-aliasing filter). As
is shown in Fig. 1, the computer continuously acquired data from the
microphone in segments of 100 ms. Each of these bins of data was passed
through a software-based infinite-impulse response (IIR) filter (the trigger
filter) that was used to detect song vocalizations while avoiding noise artifacts
(such as pecking, wing flaps and low amplitude calls). The playback of sound
occurred when the output of the trigger filter exceeded a root mean square
threshold. The artificial feedback was thus produced with a 100-ms delay, and
coincided with a silent interval in the birds song or with the following syllable,
depending on the timing of the song with respect to the bin borders created by
the computer. The alternation between recording vocalizations and playing
sounds prevented a positive feedback loop from developing. The playback
stimulus was constructed from the 100 ms of sound that had caused the trigger
event, and was narrowband-filtered before being played back to the bird. A
narrowband filter was chosen to make the playback sounds difficult to localize.
For one bird a wideband IIR filter was used. No significant differences in results
were observed between this bird and the two narrowband birds. Delayed
feedback was used instead of other interfering stimuli such as white noise so as
to replicate the structure of syllables that these birds normally hear. The effects
292 469
letters to nature
of white noise will be examined in future work.
The birds exposed to the adaptive protocol showed changes in their songs
similar to those of deafened birds. Because deafening can be caused by Neuronal correlates of
prolonged exposure to excessively loud sounds7, we used a behavioural test to
demonstrate that the conditions used in our experiment did not cause deafness. parametric working memory
Zebra finches respond to sounds by vocalizing. At the beginning and the end of
the feedback period, we examined differences in the calling probability of two in the prefrontal cortex
of the adaptive protocol birds to the playback of quiet sounds versus no sounds.
The sounds were a variety of natural stimuli including conspecific calls and Ranulfo Romo, Carlos D. Brody, Adrian Hernandez
sounds. Both birds produced significantly more calls during the presentation & Luis Lemus
of quiet sounds (P , 0:001, generalized likelihood-ratio test for Bernoulli Instituto de Fisiologa Celular, Universidad Nacional Autonoma de Mexico,
random variables), which indicates that the birds could hear the sounds. Mexico D.F. 04510, Mexico
Syllable-triggered perturbation. Recognition of the target syllable was .........................................................................................................................

achieved by using a series of IIR filters in conjunction with each other to
perform a logical operation (for example, power in band X and not in band Y).
The triggering was based on a small segment of the time-varying spectrum of
the target syllable which was unique to that syllable. An original copy of the
crystallized trigger syllable was used as the playback stimulus for the duration
of the experiment. Typical zebra finch syllables are 80150 ms in length. The
triggering resolution was 50 ms, which was short enough to ensure that the
feedback always overlapped the trigger syllable itself (and, partially, the
following syllable). The other syllables of the song received no feedback.
Spectral analysis. We calculated the time-frequency spectrogram for each
song with a sliding window (58 ms) in which each time point consisted of
the direct multitaper estimate of the power spectrum (with a time-bandwidth
product NW of 3 or 4) (ref. 13). The data shown in Fig. 2 were analysed in this
manner. A harmonic analysis was then used to determine the location and
magnitude of the jumps in the discrete spectrum of each syllable by calculating
the F-spectrum14 of each of the multitapered spectral estimates. This analysis
revealed additional statistically significant harmonic frequencies in the target
syllable after the feedback was presented to the bird (P , 0:01) (Fig. 3).
Syllable classification. For each syllable, we extracted the length and a
number of time-varying parameters (envelope, peak frequency, pitch, good-
ness-of-pitch and Wiener entropy13) based on the spectral analysis described
above. A modified K-means clustering algorithm15 was then used to partition
the syllables produced on a given day into subsets. These subsets were labelled
by the experimenter (syllable A, syllable B and so on; labelling was done blind to
the day on which the data were acquired). For each day of data, approximately
1,000 syllables were analysed. The standard deviations, which are shown as
Humans and monkeys have similar abilities to discriminate the
difference in frequency between two mechanical vibrations
applied sequentially to the fingertips13. A key component of
this sensory task is that the second stimulus is compared with
the trace left by the first (base) stimulus, which must involve
working memory. Where and how is this trace held in the brain?
This question was investigated by recording from single neurons
in the prefrontal cortex of monkeys while they performed the
somatosensory discrimination task. Here we describe neurons in
the inferior convexity of the prefrontal cortex whose discharge
rates varied, during the delay period between the two stimuli, as a
monotonic function of the base stimulus frequency. We describe
this as monotonic stimulus encoding, and we suggest that the
result may generalize: monotonic stimulus encoding may be the
a
500 ms
PD KD Base Comparison KU PB
b Set A c Set B
y

34 42
x=
Comparison (Hz)

Comparison (Hz)

error bars in Fig. 4, were obtained by bootstrapping the probability estimates 30 38

from the data16. 34
26 30
22 26
Received 23 February; accepted 12 April 1999. 22
1. Konishi, M. The role of auditory feedback in the control of vocalization in the White-crowned 18 18
sparrow. Z. Tierpsychol. 22, 770783 (1965). 14 14
2. Nottebohm, F. Auditory experience and song development in the chaffinch, Fringilla coelebs. Ibis 110,
% 10
549569 (1968). 10 6
3. Sossinka, R. & Bohner, J. Song types in the zebra finch. Z. Tierpsychol. 53, 123132 (1980). 10 14 18 22 26 30 34 10 14 18 22 26 30 34
4. Marler, P. & Peters, S. A sensitive period for song acquisition in the song sparrow, Melospize melodia
a case of age limited learning. Ethology 76, 89100 (1987). Base (Hz) Base (Hz)
5. Nordeen, K. & Nordeen, E. Auditory feedback is necessary for the maintenance of stereotyped song in
adult zebra finches. Behav. Neural Biol. 57, 5866 (1992).
d cs
6. Okanoya, K. & Yamaguchi, A. Adult bengalese finches require real-time auditory feedback to produce
normal song syntax. J. Neurobiol. 33, 343356 (1997). as
7. Wooley, S. & Rubel, E. Bengalese finches Lonchura-striata-domestica depend on auditory feedback for
the maintenance of song. J. Neurosci. 17, 63806390 (1997). ps
8. Suthers, R. Contributions to birdsong from the left and right sides of the intact syrinx. Nature 347,
473477 (1990).
9. Fee, M. S., Shraiman, B., Pesaran, B. & Mitra, P. P. The role of nonlinear dynamics of the syrinx in
birdsong production. Nature 395, 6771 (1998).
10. Houde, J. F. & Jordan, M. I. Sensorimotor adaptation in speech production. Science 279, 12131216
Figure 1 Discrimination task. a, Sequence of events during discrimination trials.
(1998).
11. Feller, W. An Introduction to Probability Theory and its Applications (Wiley, New York, 1968). The mechanical probe is lowered, indenting the glabrous skin of one digit of the
12. Cover, T. M. & Thomas, J. A. Elements of Information Theory (Wiley, New York, 1991). hand (PD); the monkey places his free hand on an immovable key (KD); the probe
13. Ho, C. E., Pesaran, B., Fee, M. S. & Mitra, P. P. Characterization of the structure and variability of zebra
finch song elements. Proc. 5th Joint Symp. on Neural Computation 7683 (1998). oscillates vertically, at the base frequency; after a delay, a second mechanical
14. Thomson, D. J. Spectrum estimation and harmonic analysis. Proc. IEEE 70, 10551096 (1982). vibration is delivered at the comparison frequency; the monkey releases the key
15. Selim, S. Z. & Ismail, M. A. K-means-type algorithms: a generalized convergence theorem and
(KU) and presses one of two push-buttons (PB) to indicate whether the
characterization of local optimality. IEEE Trans. Pattern Anal. Mach. Intell. 1, 8187 (1984).
16. Bradley, E. An Introduction to the Bootstrap (Chapman & Hall, New York, 1993). comparison frequency was higher or lower than the base. b, c, Stimulus sets
used during recordings. Each grey box indicates a base frequency/comparison
Acknowledgements. We thank B. Pesaran and M. Sahani for discussions on data analysis, and R. Egnor,
M. Fee, G. Laurent and M. Schmidt for comments on the manuscript. This work was supported by a grant frequency stimulus pair used; the number inside the box indicates overall per cent
from the NIMH. correct trials for that base/comparison pair. d, Location of recording sites that
gave somatosensory working-memory-related responses; CS, central sulcus;
Correspondence and requests for materials should be addressed to A.L. (e-mail: leonardo@cns.caltech.
edu). AS, arcuate sulcus; PS, principal sulcus.

470
1999 Macmillan Magazines Ltd 293
NATURE | VOL 399 | 3 JUNE 1999 | www.nature.com


294
letters to nature
................................................................. however, that alteration of auditory feedback by deafening or by the
Interruption of a basal ganglia playback of masking sounds can lead to gradual deterioration of
the songs of adult zebra nches1,5. This suggests the alternative
forebrain circuit prevents hypothesis that the neural mechanisms involved in the evaluation of
auditory feedback remain active in adult birds, but that normally
plasticity of learned vocalizations there is no impetus for change in song because it matches the stored
song target (Fig. 1b). According to this model, the changes in song
Michael S. Brainard & Allison J. Doupe following disruption of auditory feedback reect an active process,
in which the mismatch between feedback and the stored song target
Keck Center for Integrative Neuroscience, Departments of Physiology generates a large but aberrant instructive signal, which drives non-
and Psychiatry, University of California San Francisco, San Francisco, adaptive changes in song (Fig. 1c).
California 94143-0444, USA This interpretation raises the question of where in the brain
.............................................................................................................................................. auditory feedback of the bird's song is evaluated. One candidate is
Birdsong, like speech, is a learned vocal behaviour that relies
greatly on hearing; in both songbirds1 and humans2 the removal of a Learning
auditory feedback by deafening leads to a gradual deterioration of mismatch between evaluation
of auditory
adult vocal production. Here we investigate the neural mechan-
developing song
and stored target auditory feedback
feedback
isms that contribute to the processing of auditory feedback during
the maintenance of song in adult zebra nches. We show that the large
deleterious effects on song production that normally follow instructive signal

deafening can be prevented by a second insult to the nervous

systemthe lesion of a basal gangliaforebrain circuit. The adaptive modification song
results suggest that the removal of auditory feedback leads to of motor program motor
nuclei
song

the generation of an instructive signal that actively drives non-

adaptive changes in song; they also suggest that this instructive
signal is generated within (or conveyed through) the basal
b Normal adult
gangliaforebrain pathway. Our ndings provide evidence that evaluation
cortical-basal ganglia circuits may participate in the evaluation of good match between
mature song of auditory
sensory feedback during calibration of motor performance, and and stored target auditory
feedback
feedback

demonstrate that damage to such circuits can have little effect on

previously learned behaviour while conspicuously disrupting the small or stable
capacity to adaptively modify that behaviour. instructive signal

The learning of birdsong, an intricate, sequenced motor skill,

occurs in a two-stage process3,4. First, in a sensory learning period, stable song
motor song
song
young birds listen to and memorize the song of an adult tutor. Then, nuclei
in a sensorimotor learning period, they gradually match their own
developing vocalizations to the tutor song memory. This process
bears striking similarities to certain aspects of human vocal c Interruption of auditory feedback
learning4. In particular, both song and speech are strongly depen- mismatch between evaluation

dent on hearing: neither develops normally in deaf individuals, and,

(interrupted) feedback
and stored target
of
auditory
feedback
X auditory
feedback
in adults, hearing loss or experimental alteration of auditory feed-
back can lead to dramatic changes in vocal production1,2,5,6. Bird- large but aberrant
song thus provides a model system for investigating the neural instructive signal

processing whereby auditory feedback is used to calibrate and

maintain vocal production, as well as the more general question non-adaptive song
of how performance-based feedback is used to shape complex modification
of motor program
motor
nuclei
song

motor output. Here we show that a basal gangliaforebrain circuit

that is not required for normal song production in adult zebra
nches is nevertheless critical for plasticity in song following
d Interruption of error signal & feedback
interruption of auditory feedback. evaluation
The rationale for our experiment is illustrated by tracing the mismatch between
of auditory

hypothetical ow of auditory feedback through the brain as birds

(interrupted) feedback
and stored target auditory
feedback
X feedback

learn to produce and stabilize their own song (Fig. 1ad). During
normal sensorimotor learning, young birds use auditory feedback large
instructive signal X
to match their own vocalizations to the previously memorized tutor is interrupted

song7. To carry out this matching process, the nervous system must
rst encode auditory feedback from the bird's own song, and then stable
song
song
motor song
compare that feedback with the stored song target (tutor song). This nuclei
comparison is presumed to generate an instructive signal that
gradually drives adaptive changes in the motor pathway for song Figure 1 Hypothetical ow of auditory feedback through the song system. a, During
(Fig. 1a). At the end of sensorimotor learning, the bird's song sensorimotor learning, auditory feedback from song is compared with a previously
resembles that of the tutor and in many species is retained relatively memorized song target. Differences between the bird's own song and the target song
unchanged throughout adulthood. result in an instructive signal that drives adaptive modication of song. b, After song
The unchanging nature of adult birdsong might reect a learning, the bird's own song closely approximates the target song. Consequently, there is
stabilization of synapses in the motor pathway for song production, little drive for further changes. c, Removal of feedback leads to generation of an aberrant
such that they become independent of auditory experience once instructive signal and non-adaptive changes to song. d, Interruption of the instructive
sensorimotor learning is complete. Recent experiments have shown, signal removes impetus for changes to song, even when feedback is altered.

762 2000 Macmillan Magazines Ltd 295

NATURE | VOL 404 | 13 APRIL 2000 | www.nature.com

the anterior forebrain pathway (AFP; Fig. 2a), a basal ganglia
forebrain circuit8 belonging to a system of brain areas devoted to
vocal learning and production9. Lesions of this pathway in juvenile
birds prevent normal song learning1012. Moreover, the AFP is well
situated to guide changes in the connectivity of the motor pathway
for song; the pathway originates in the song motor control nucleus
HVc and passes through the basal ganglia (Area X)8, the thalamus
(DLM), and a dorsal forebrain nucleus analogous to frontal cortex
(LMAN: lateral magnocellular nucleus of the anterior neostriatum),
before returning to the motor pathway for song at the level of the
robust nucleus of the archistriatum (RA). Finally, neurons in the
AFP respond well to the sound of the bird's own song or the tutor
song, but poorly if at all to most other sounds, including the songs
of other individuals of the same species1315. This auditory selectivity
potentially endows the AFP with the capacity to `listen to' and
evaluate the bird's own song relative to the tutor song, and provide
instructive feedback to the motor pathway via its output to RA.
Although lesions of the AFP in young birds disrupt song learning,
lesions in adults have little apparent effect on song production912,16.
This lack of effect of adult lesions contrasts with the gradual
deterioration of song following deafening in adult birds1, and has
led to the suggestion that the AFP cannot participate in the
processing of auditory feedback during the maintenance of adult
song. However, although these results indicate that lesioning the
AFP is not equivalent to removing auditory feedback as it enters the
a HVc
RA Field L
LMAN
Area X
DLM
nXIIts
tracheo-
syringeal respiratory
nerve nuclei
respiratory motor pathway
muscles
syringeal anterior forebrain pathway
muscles
b Intact Lesion
LMAN lesion
mMAN mMAN centre
* cases were no greater than those observed in control birds that
Area X Area X
RA RA
dorsal
lateral
Figure 2 Song system nuclei in intact and lesioned birds. a, Schematic of connections
within the song system. The `motor pathway' (shaded) is required throughout life for song
production9. The `anterior forebrain pathway' (solid) is required for song learning, but not
for adult song production. Field L and related areas provide auditory input to the song
system. b, Transverse, CGRP-labelled29 sections through the right anterior forebrain (top)
and archistriatum (bottom). In an intact bird (left), LMAN was labelled along with its axonal
projections into RA and Area X. The medial nucleus of the anterior neostriatum (mMAN)
was also labelled. In a lesioned bird (right), nucleus LMAN was removed, with a
corresponding loss of label in RA and Area X. Arrows, 250 mm.
NATURE | VOL 404 | 13 APRIL 2000 | www.nature.com 2000 Macmillan Magazines Ltd
letters to nature
nervous system (Fig. 1c), they do not rule out the possibility that the
AFP is involved in the subsequent evaluation of auditory feedback.
If lesions of this pathway interrupt an instructive signal based on an
evaluation of auditory feedback, they might not have an effect on
normal adult song, which is well matched to its target (Fig. 1b).
Indeed, according to this hypothesis, lesions that interrupted such
an instructive signal might in fact occlude the effects of altering
auditory feedback (Fig. 1d). To test this hypothesis, we therefore
compared the effects of deafening in intact adult zebra nches with
the effects of deafening in birds that also received bilateral lesions of
LMAN, the output nucleus of the anterior forebrain (Fig. 2b).
In line with previous reports1, we found that songs of normal
adult zebra nches were stable over long periods of time, while
songs of zebra nches that had been deafened in adulthood
gradually deteriorated over a period of weeks to months. Examples
illustrating the range of song deterioration following deafening are
shown in Fig. 3A,C. The effects of deafening included alteration of
the structure of individual syllables, as well as changes to the overall
temporal organization of the song. In marked contrast, songs of
birds that received AFP lesions at the same time as they were
deafened remained essentially unchanged (for example, Fig. 3B,D),
and in most cases the subtle changes to syllable structure or syllable
sequencing were comparable to those observed in control birds.
Furthermore, lesions did not simply delay the effects of deafening;
rather, they prevented changes in song for as long as the birds were
followed (up to a year; see, for example, Fig. 3D, bottom panel).
The effects of experimental manipulations on song were character-
ized using two measures that focused separately on changes
to individual syllables and on changes to the overall organization
of song. These two aspects of song are thought to be controlled
separately by RA (syllables) and HVc (syllable sequencing/temporal
pattern)17,18, and thus might be subject to independent changes.
First, we used a subjective measure similar to that used in previous
studies of birdsong to score (blind to the treatment of individual
birds) how well syllables that were initially present in the bird's
repertoire were preserved in later songs (see Methods). This
measure was based on spectrographic representations of syllables
that were presented individually, and therefore emphasized changes
to the structure of syllables and discounted any changes to the
sequence in which syllables were sung or the overall temporal
pattern of song.
A summary of the changes in the structure of syllables for the
different experimental groups is shown in Fig. 4a. For deafened
birds, changes in syllables varied between individuals, and in some
retained normal hearing. Despite this variability, the songs of
deafened birds deteriorated signicantly relative to those of con-
trols. Indeed, on average, the degree of similarity to the birds'
original songs was not signicantly greater than the similarity
between songs randomly selected from unrelated birds. In contrast,
the songs of birds that received LMAN lesions before deafening did
not change more than those of control birds. Furthermore, in each
instance where there were recordings from pairs of brothers, which
controlled for some potential sources of variability, the syllables of
the birds lesioned before deafening were more stable than those of
their deafened but unlesioned brothers. Finally, `sham' lesions,
outside of the song system, did not prevent deafening induced
deterioration of song, indicating that the effects of lesions were not
due to non-specic damage to the brain. Thus, the absence of
LMAN prevented the signicant syllable deterioration that
normally results from deafening in adult zebra nches.
Because of the deterioration of syllables in deafened birds, it
generally was not possible to assess changes to the sequences in
which syllables were sung. We therefore used a cross-correlation
measure that was independent of the identication of syllables
to characterize changes in the overall temporal pattern of song
(see Methods). To illustrate this independence, we applied both
296 763
letters to nature
measures of song stability to two birds that were recorded imme-
diately before and after cutting the tracheosyringeal nerve (Fig. 2a),
which is essential for much of the structure of individual syllables,
but not for the timing of patterned expiration during song19. The
measures indicated a complete deterioration of syllable structure
(Fig. 4a, `nerve cut'), while the overall temporal pattern of song was
well preserved (Fig. 4b, `nerve cut'), demonstrating that changes to
these two aspects of song were dissociable.
Despite this dissociability, changes to the temporal pattern of
song for the experimental birds paralleled changes to syllable
structure; deafening caused a signicant deterioration of the tem-
poral pattern of song relative to controls, and this deterioration was
prevented by lesions of LMAN but not by sham lesions (Fig. 4b).
Thus, lesions of LMAN block deafening-induced vocal plasticity in
adult zebra nches, not only of syllable structure, but also of the
overall temporal pattern of song.
The results are consistent with the hypothesis outlined above
(Fig. 1c), that disruption of auditory feedback in adult birds leads to
the generation of an instructive signal that actively drives changes in
song, and that this signal either arises within or is conveyed through
the AFP. This possibility is supported both by the song-selective
properties of AFP neurons, and by the recent nding that this circuit
is active during singing, even in deaf birds15. It is critical to our
experiment that, in deafened birds, song production is initially

A Deafened B Lesioned & Deafened

i i a b c d e a b c d e a b c d e a b c d e
8kHz 8kHz

6 6
pre
4 4

2 2

0 0

a i i a a a a b c? d e a c? a b c d e a b c d e
8kHz 8kHz

6 6

4 182 d post 4

2 2

0 0

C Deafened D Lesioned & Deafened

i i a b c d e e f g i bc d e e f bc
8kHz 8kHz

6 6
pre
4 4

2 2

0 0

e? f? g? i bc d e f bc
8kHz 8kHz

6 6
36 d post
4 4

2 2

0 0

i bc d e f bc
8kHz 8kHz

6 6

321 d post
4 4

2 2

0 0
500ms

Figure 3 Changes to song following deafening in intact and lesioned birds. Shown are
spectrograms illustrating changes to song for 2 deafened birds (A and C) and their 2
brothers, which additionally received LMAN lesions (B and D). Each bird initially sang a
xed repertoire of syllables (labelled with letters) in a stereotyped sequence or `motif'
(boxes). One of the smaller effects of deafening is shown in A: gradual loss or deterioration
of some syllables (for example, c) and introduction of abnormal sequences (for example,
repeated a's). C, One of the larger effects of deafening: relatively rapid loss of identiable
syllables and eventual deterioration of song into a series of short, noisy notes. In contrast,
songs of the lesioned brothers (B and D) recognizably retained all original syllables and
sequencing. The largest change for these birds was loss of repetition of one syllable (e in
panel D). Songs for all groups of birds also speeded up, as previously observed in
controls30 and following LMAN lesions21. This is apparent from the time scale bars
(500 ms).

764 2000 Macmillan Magazines Ltd 297

NATURE | VOL 404 | 13 APRIL 2000 | www.nature.com

unchanged but song feedback to the nervous system is altered, that
is, that the deafening is a `pure' sensory manipulation. The data
therefore also provide insight into two previously reported effects of
adult lesions of LMAN: the prevention of incorporation of new
syllables into songs of birds experimentally manipulated to undergo
late learning20, and the prevention of gradual changes to the
abnormal songs of adult birds whose motor production has been
disrupted by denervation of the vocal musculature21; this motor
disruption also results in altered sensory feedback, and is thus a
`mixed', sensorimotor, manipulation. Our results with simple elim-
ination of auditory feedback suggest that in both of the previous,
more complex experimental situations, LMAN lesions may also
act by eliminating signals from the AFP to the motor pathway
about the mismatch between sensory feedback and the stored target
song, thereby eliminating any impetus for change in vocal output
(Fig. 1d).
An alternative (and not mutually exclusive) hypothesis about the
role of the AFP is that neural or trophic inputs from this circuit are
`permissive' for plasticity in the motor pathway. According to this
hypothesis, the instructive signal driving changes in song might
arise elsewhere, but without the integrity of the AFP would be
unable to bring about changes in song. This possibility is consistent
with LMAN's known trophic support of RA early in develop-
ment22,23. Assessing whether the AFP is instructive, permissive or
both will ultimately require characterization of the nature of signals
generated in the AFP by alteration of auditory feedback.
a feedback) where a large instructive signal to alter song is elicited.
'identical' 3 This `covert' contribution of the AFP to adult plasticity may have
Syllable similarity
letters to nature
Regardless of the means by which AFP lesions prevent song
plasticity, our study indicates that the changes in song that occur
following interruption of auditory feedback are the result of an
active process. This did not need to be the case. For example, the
slow changes in song that occur following hearing loss could have
reected an uncorrected `drift' in motor output that resulted from
passive processes such as cell death or neurogenesis in the motor
pathway24, or from changes in the efcacy of vocal musculature.
Since song changes are prevented by lesions of LMAN, however, this
seems unlikely. Presumably, in adult birds, the AFP normally
participates in the correction of any changes in song that result
from such passive processes or from external damage to the motor
pathway. Our results further demonstrate that in healthy adult birds
such changes to the motor pathway are remarkably small; although
removal of auditory feedback can lead to deterioration of song,
auditory feedback is not required (in birds with AFP lesions) for
song to remain stable.
Finally, the results are consistent with the AFP playing the same
role in adults as has previously been hypothesized for juveniles: the
auditory feedback-based evaluation and adaptive modication of
the bird's own song. Differences between the effects of lesions in
juvenile and adult zebra nches do not necessarily reect a changing
function of the AFP but rather a different state of motor com-
petence; in juveniles, lesions disrupt an instructive signal so that
normal song learning cannot progress, while in adults that have
completed song learning the presumptive signal for change is small
so that lesions have little effect. In the adult case, the effects of lesions
are only revealed under conditions (such as disruption of auditory
parallels to mammalian systems, where damage to similar reciprocal
cortical-basal ganglia circuits can impair procedural learning while

2.5
(average score)

having little effect on previously learned performance2527. In

2 particular, since speech, like song, is an apparently stable behaviour
that can nevertheless change in response to altered auditory
1.5
feedback2,6, our results suggest that it may be revealing to examine
the functioning of cortical-basal ganglia circuits during the mod-
Random
1
ication of speech; there is already evidence that such circuits are
similarity differentially active during production of speech in a second
0.5
language versus a native language28. Since the AFP is a discrete
0
basal gangliaforebrain circuit specialized for one well dened
'no similarity'
control deaf lesioned sham nerve cut behaviour, it may prove a particularly tractable system for elucidat-
& deaf & deaf ing the signals present in these structures and the role that they play
in the learning and modication of sequenced motor acts. M
b
identical 1
Methods
0.8 Animals
(xcorr-rand)/(1-rand)
Temporal similarity

Zebra nches (Taeniopygia guttata) were raised in individual breeding cages with their
0.6
parents and siblings until approximately 60 days of age when they were transferred to all-
male group cages. Preliminary experiments suggested that there were greater effects of
0.4
deafening on young adult birds (90140 days of age) than on older birds (.200 days of
age). For this study, baseline recordings and experimental manipulations were therefore
0.2
restricted to birds that were 98140 days of age, except for one pair of brothers (one
lesioned and one intact) that were deafened at 204 days of age. These two birds were
0
included in Fig. 4, but not in statistical analyses. Where possible, we studied pairs of
Random brothers that were housed together, experienced manipulations at the same ages, and had
-0.2
similarity similar songs by virtue of exposure to each other and the same adult tutor. There was a
signicant correlation between syllable stability (see below) for deafened birds and syllable
-0.4
stability for their lesioned brothers, indicating that the identity of brothers indeed
control deaf lesioned sham nerve cut accounted for some of the variability in song stability (see also Fig. 4, where data that
& deaf & deaf
derive from brothers are connected by lines).
Figure 4 Summary of song stability following deafening of intact or LMAN lesioned birds.
a, Syllable stability. Each point corresponds to an individual bird and shows the average Lesions
similarity (see Methods) between syllables from baseline songs and those from songs Birds were deafened by bilateral cochlear removal7. For some birds, LMAN was electro-
recorded at least six months later, except for nerve-cut birds, where song changes were lytically lesioned 15 days before deafening. Lesions were stereotaxically targeted at
assessed within 1 week. Lines connect points corresponding to brothers. Bars show LMAN, and evaluated in tissue labelled with an antibody to CGRP (Fig. 2)29. For all but
one of the nominally lesioned birds, the percentage of LMAN that was removed bilaterally
means and standard errors for each group. b, Temporal stability. Each point shows the ranged from 70% to 100%. The remaining bird had only a 37% lesion, but it included the
temporal similarity (see Methods) between baseline songs and songs recorded after the region where RA projecting axons leave LMAN, and almost all CGRP labelling in RA was
indicated manipulations. eliminated. `Sham' lesions were entirely anterior and dorsal to LMAN.

NATURE | VOL 404 | 13 APRIL 2000 | www.nature.com 2000 Macmillan Magazines Ltd 298 765
letters to nature
Analysis of changes to song
To characterize changes to the structure of individual syllables, we used a subjective
scoring procedure similar to that employed in previous studies of birdsong1620. Observers
who were blind to the experimental manipulation of each bird scored the similarity
between spectrographic representations of syllables from songs recorded before any
manipulation and those from later songs. For each baseline syllable that was initially
present in the repertoire of a bird (512 distinct syllables per bird), observers identied the
most similar syllable present in songs recorded at later dates. Observers additionally
judged the degree of similarity between each baseline syllable and its best match on a scale
of 0 (no similarity) to 3 (identical). As a measure of the maximum expected deterioration
of syllables, we determined the random similarity between syllables of songs from
unrelated birds (n = 12 pairs of songs). This random similarity averaged 0.9, indicating
some generic similarity between the syllables of unrelated zebra nches. Scores were
averaged across 3 observers (mean r2 for pairwise correlation of observers' scores, 0.81).
All songs were recorded from birds isolated in sound attenuating boxes and were
therefore `undirected'. Signicance of differences between groups was based on a criterion
of P , 0.05 using analysis of variance (ANOVA) and post-hoc Fischer's Protected Least
Signicant Difference tests.
To assess temporal stability, songs were represented by the timing of syllable production,
while the spectral structure of individual syllables was ignored. For each bird, the timing
pattern representing the most common motif before any experimental manipulation was
identied. We then searched songs from later recording sessions for the pattern of syllables
that provided the closest temporal match to the initial motif. At each offset (in increments
of 5 ms) between motif and song, we quantied the temporal overlap as the average of the
percentage overlap of syllables and the percentage overlap of intervals. For each bird, the
maximal overlap was calculated for 10 songs and then averaged to provide a measure
(`xcorr') of how well the temporal pattern of song was conserved following experimental
manipulation. Because the delivery of some songs speeded up `proportionately' over time
(that is, with no apparent change in the relative durations of notes and intervals), we
allowed for proportional changes in the temporal pattern of the song (ranging from 75%
to 110% of initial song duration) in searching for the maximal overlap. To control partially
for varying complexity of different birds' motifs, we calculated the maximal overlaps
between each bird's motif and randomly selected songs from unrelated birds (`rand'). The
temporal stability of song was then expressed as a normalized value ranging from 0
(indicating no more preservation of temporal pattern than random) to 1 (indicating
perfect preservation of temporal pattern). The baseline motif for one pair of brothers had a
very simple temporal pattern (3 syllables) and consequently showed a high degree of
temporal similarity to all songs (including those from unrelated birds). This pair of birds
was therefore excluded from further analysis of changes in temporal pattern.
Received 12 December 1999; accepted 1 February 2000.
1. Nordeen, K. W. & Nordeen, E. J. Auditory feedback is necessary for the maintenance of stereotyped
song in adult zebra nches. Behav. Neural Biol. 57, 5866 (1992).
2. Cowie, R. & Douglas-Cowie, E. Postlingually Acquired Deafness: Speech Deterioration and the Wider
Consequences 1304 (Mouton de Gruyter, Berlin, 1992).
3. Marler, P. A comparative approach to vocal learning: song development in white-crowned sparrows.
J. Comp. Physiol. Psychol. 71, 125 (1970).
4. Doupe, A. J. & Kuhl, P. K. Birdsong and human speech: common themes and mechanisms. Annu. Rev.
Neurosci. 22, 567631 (1999).
5. Leonardo, A. & Konishi, M. Decrystallization of adult birdsong by perturbation of auditory feedback.
Nature 399, 466470 (1999).
6. Houde, J. F. & Jordan, M. I. Sensorimotor adaptation in speech production. Science 279, 12131216
(1998).
7. Konishi, M. The role of auditory feedback in the control of vocalization in the white-crowned sparrow.
Z. Tierpsychol. 22, 770783 (1965).
8. Bottjer, S. W. & Johnson, F. Circuits, hormones, and learning: vocal behavior in songbirds. J. Neurobiol.
33, 602618 (1997).
9. Nottebohm, F., Stokes, T. M. & Leonard, C. M. Central control of song in the canary, Serinus canarius.
J. Comp. Neurol. 165, 457486 (1976).
10. Bottjer, S. W., Miesner, E. A. & Arnold, A. P. Forebrain lesions disrupt development but not
maintenance of song in passerine birds. Science 224, 901903 (1984).
11. Sohrabji, F., Nordeen, E. J. & Nordeen, K. W. Selective impairment of song learning following lesions
of a forebrain nucleus in the juvenile zebra nch. Behav. Neural Biol. 53, 5163 (1990).
12. Scharff, C. & Nottebohm, F. A comparative study of the behavioral decits following lesions of various
parts of the zebra nch song system: implications for vocal learning. J. Neurosci. 11, 28962913
(1991).
13. Doupe, A. J. Song- and order-selective neurons in the songbird anterior forebrain and their emergence
during vocal development. J. Neurosci. 17, 11471167 (1997).
14. Solis, M. M. & Doupe, A. J. Contributions of tutor and bird's own song experience to neural selectivity
in the songbird anterior forebrain. J. Neurosci. 19, 45594584 (1999).
15. Hessler, N. A. & Doupe, A. J. Singing-related neural activity in a dorsal forebrain-basal ganglia circuit
of adult zebra nches. J. Neurosci. 19, 1046110481 (1999).
16. Nordeen, K. W. & Nordeen, E. J. Long-term maintenance of song in adult zebra nches is not affected
by lesions of a forebrain region involved in song learning. Behav. Neural Biol. 59, 7982 (1993).
17. Vu, E. T., Mazurek, M. E. & Kuo, Y. C. Identication of a forebrain motor programming network for
the learned song of zebra nches. J. Neurosci. 14, 69246934 (1994).
18. Yu, A. C. & Margoliash, D. Temporal hierarchical control of singing in birds. Science 273, 18711875 (1996).
19. Vicario, D. S. Contributions of syringeal muscles to respiration and vocalization in the zebra nch.
J. Neurobiol. 22, 6373 (1991).
20. Morrison, R. G. & Nottebohm, F. Role of a telencephalic nucleus in the delayed song learning of
socially isolated zebra nches. J. Neurobiol. 24, 10451064 (1993).
21. Williams, H. & Mehta, N. Changes in adult zebra nch song require a forebrain nucleus that is not
necessary for song production. J. Neurobiol. 39, 1428 (1999).
766 2000 Macmillan Magazines Ltd
22. Johnson, F. & Bottjer, S. W. Afferent inuences on cell death and birth during development of a
cortical nucleus necessary for learned vocal behavior in zebra nches. Development 120, 1324 (1994).
23. Akutagawa, E. & Konishi, M. Two separate areas of the brain differentially guide the development of a
song control nucleus in the zebra nch. Proc. Natl Acad. Sci. USA 91, 1241312417 (1994).
24. Kirn, J. R. & Nottebohm, F. Direct evidence for loss and replacement of projection neurons in adult
canary brain. J. Neurosci. 13, 16541663 (1993).
25. Graybiel, A. M., Aosaki, T., Flaherty, A. W. & Kimura, M. The basal ganglia and adaptive motor
control. Science 265, 18261831 (1994).
26. Knowlton, B. J., Mangels, J. A. & Squire, L. R. A neostriatal habit learning system in humans. Science
273, 13991402 (1996).
27. Nakamura, K., Sakai, K. & Hikosaka, O. Effects of local inactivation of monkey medial frontal cortex
in learning of sequential procedures. J. Neurophys. 82, 10631068 (1999).
28. Klein, D., Zatorre, R. J., Milner, B., Meyer, E. & Evans, A. C. Left putaminal activation when speaking a
second language: evidence from PET. Neuroreport 5, 22952297 (1994).
29. Bottjer, S. W., Roselinsky, H. & Tran, N. B. Sex differences in neuropeptide staining of song-control
nuclei in zebra nch brains. Brain Behav. Evol. 50, 284303 (1997).
30. Arnold, A. P. The effects of castration on song development in zebra nches (Poephila guttata). J. Exp.
Zool. 191, 261278 (1975).
Acknowledgements
We thank M. Stryker, M. Churchland, T. Troyer and K. T. Moortgat for comments on the
manuscript, and A. Arteseros, G. Carrillo and A. Tam for technical assistance. This work
was supported by a Burroughs Wellcome Fund fellowship of the Life Sciences Research
Foundation (M.S.B.), and by the John Merck Fund, the EJLB Foundation and the National
Institutes of Health (A.J.D.).
Correspondence and requests for materials should be addressed to M.S.B. (e-mail:
msb@phy.ucsf.edu).
.................................................................
SHATTERPROOF MADS-box genes
control seed dispersal in Arabidopsis
Sarah J. Liljegren*, Gary S. Ditta*, Yuval Eshed, Beth Savidge*,
John L. Bowman & Martin F. Yanofsky*
* Section of Cell and Developmental Biology, University of California at San
Diego, La Jolla, California 92093-0116, USA
Section of Plant Biology, University of California at Davis, Davis,
California 95616, USA
..............................................................................................................................................
The fruit, which mediates the maturation and dispersal of seeds, is
a complex structure unique to owering plants. Seed dispersal in
plants such as Arabidopsis occurs by a process called fruit
dehiscence, or pod shatter. Few studies13 have focused on identi-
fying genes that regulate this process, in spite of the agronomic
value of controlling seed dispersal in crop plants such as canola4,5.
Here we show that the closely related SHATTERPROOF (SHP1)
and SHATTERPROOF2 (SHP2) MADS-box genes are required for
fruit dehiscence in Arabidopsis. Moreover, SHP1 and SHP2 are
functionally redundant, as neither single mutant displays a novel
phenotype. Our studies of shp1 shp2 fruit, and of plants con-
stitutively expressing SHP1 and SHP2, show that these two genes
control dehiscence zone differentiation and promote the lignica-
tion of adjacent cells. Our results indicate that further analysis of
the molecular events underlying fruit dehiscence may allow
genetic manipulation of pod shatter in crop plants.
The MADS-box gene family encodes transcriptional regulators
involved in diverse aspects of plant development, and phylogenetic
and functional studies show extensive redundancy between family
members69. SHP1 and SHP2 (previously known as AGL1 and
AGL5) were two strong candidates for functional redundancy, as
they share 87% identity at the amino-acid sequence level and show
almost identical expression patterns in developing Arabidopsis fruit
(refs 1012; and C. Ferrandiz, Y.E., J.L.B. and M.F.Y., unpublished
results).
Present address: Calgene, Davis, California 95616, USA.
299
NATURE | VOL 404 | 13 APRIL 2000 | www.nature.com


300
J Comp Physiol A (1986) 159: 827-840
Identified target-selective visual interneurons
descending from the dragonfly brain
Robert M. Olberg
Department of Biological Sciences, Union College, Schenectady, New York 12308, USA
Accepted August 1i, 1986
Summary. 1. Eight large interneurons descending
in the dragonfly (Aeshna umbrosa, Anax junius)
ventral nerve cord from the brain to the thoracic
ganglia were identified anatomically with intracel-
lular dye injection (Fig. 3). All eight were strictly
visual and responded only to movements of small
patterns, such as black squares, 'targets', moving
on a white background.
2. The target interneurons all projected from
the protocerebrum at least as far as the metatho-
racic ganglion. Within the protocerebrum they ar-
borized in the posterodorsal neuropil region, near
the base of the circumesophageal connectives
(Fig. 3).
3. The receptive fields of six of the cells were
large, including most of the forward hemisphere
of vision. For five of these, spiking responses were
often restricted to a much smaller region within
the receptive field, with stimulation of other areas
yielding only subthreshold responses (Figs. 4 and
5, Table 1).
4. The pattern of selectivity for target size var-
ied, with some neurons responding only to small
targets, some showing consistent responses over
a wide range of target sizes, and one preferring
larger targets (Fig. 6, Table 1).
5. Five of the interneurons were directionally
selective. Movement in the antipreferred direction
elicited hyperpolarizing responses in two of them.
Movements of large patterns, such as a checker-
board pattern covering the forward hemisphere,
elicited opposite directional responses, i.e., hyper-
polarizations in the preferred target direction and
subthreshold depolarizations in the antipreferred
direction (Fig. 7). A large pattern moving in any
Abbreviations: D I T Dorsal intermediate tract; D M T Dorsal
median tract; M D T Median dorsal tract; V N C Ventral nerve
cord; D C M D descending contralateral movement detector
Journal of
Comparative Sensory,
Neural,
and
P h y s i o l o g y A B....lo~,
Physiology
9 Springer-Verlag 1986
direction inhibited the response to target move-
ment (Fig. 8).
6. These neurons mediate, in part, the visual
control of flight orientation. I propose that they
convey turning signals to the wing motor in re-
sponse to objects moving relative to the animal.
Introduction
Descending interneurons in insects receive input
from the sensory structures of the head and pro-
vide input to thoracic motor centers to initiate or
guide behavior (Tanouye and Wyman 1980; Blon-
deau 1981; Pearson and Robertson 1981; Bacon
and M6hl 1983; Bicker and Pearson 1983; Rei-
chert and Rowell 1985). From studies of descend-
ing interneurons in a variety of insects, such as
grasshoppers (Catton and Chakraborty 1969;
Rowell 1971; Bacon and Tyrer 1978; Simmons
1980; Rowell and Pearson 1983; Kien and Altman
1984), flies (Strausfeld and Bacon 1983), and
moths (Olberg 1983), two generalizations emerge:
1) Many descending interneurons are multimodal
and therefore appear relatively nonspecific to the
experimenter (but see Olberg 198t b; Reichert et al.
1985) and 2) single descending interneurons usual-
ly do not elicit movements by themselves. The only
published exceptions to the second generalization
are the descending interneurons which elicit song
in the cricket (Bentley 1977) and the giant fibers
which initiate escape in flies (Tanouye and Wyman
1980; Tanouye and King 1983).
If descending interneurons appear relatively
nonspecific in their sensory reponses, how is the
precision which characterizes many insect behav-
iors (e.g. Collett and Land 1978) achieved? Kien
301
828
A Electrode~~ B interneurons (Olberg 1981 a) showed that their directional pref-
Projection L,/ I /
Screen ~/
Fig. 1 A, B. A, Visual stimulation arrangement. Dragonfly was
mounted ventral side up and tilted 15 ~ up from horizontal.
The head was centered between the back edges of a translucent,
wrap-around projection screen in line with the center of the
front 90 ~ x 90 ~ facet of the screen. B, View of screen from the
front. Square target 4 ~ visual angle at (a) has following dimen-
sions at indicated locations ( H o r x V e r t ) : b) 2.1~ ~ c)
1.4 ~ x 1.4 ~ d) 4 ~ x 2.8 ~ e) 2 ~ x 2 ~ f) 1.4 ~ x 1.4 ~ (Visual angle cal-
culated from the center of the dragonfly's head)
and Altman (1984) suggest that most behavior is
probably directed by many descending interneu-
rons acting in concert. Simultaneous transmission
by many parallel channels carries the potential for
sharpening the precision of the sensory informa-
tion.
A group of dragonfly descending interneurons
are of interest because they do not fit the above
generalizations. They are purely visual and re-
spond only to the movement of small patterns or
'targets' (Olberg 1981a). Most are also selective
for target direction. When stimulated intracellu-
larly some target interneurons can elicit move-
ments of from two to all four wings (Olberg 1978).
In the present study I have used intracellular
microelectrodes and Lucifer Yellow injection to
identify eight individual target interneurons. This
paper presents their anatomy, directional prefer-
ences, receptive fields, size selectivity and responses
to several standard stimuli.
Materials and methods
Research animals. The animals used in this study were adult
males and females of 2 dragonfly species, Aeshna umbrosa and
Anaxjunius, both of the family Aeshnidae. They were collected
as larvae and reared in the laboratory (23 ~ 16L/8D light
cycle) on a diet consisting mainly of mealworms (Tenebrio).
After adult eclosion they remained at 23 ~ for 2 days before
being moved to a 4 ~ cold room. I usually used the animals
within one week after emergence.
Preparation. Using a mixture of beeswax and rosin, I waxed
the animal's ventral surface behind the metathoracic legs to
a steel rod. I waxed the back of the head to the thorax dorsally
and on each side. I fixed the rod to a magnet stand, with the
animal ventral side up at a 15 ~ angle to horizontal, anterior
end up (Fig. 1). (Extracellular recordings from dragonfly target
R.M. Olberg: Identified visual target interneurons in the dragonfly
erences remained constant relative to the animal whether it
was dorsal or ventral side up.)
I cut off the legs and removed the ventral thoracic cuticle
anterior to the metathoracic leg sockets. To reduce movement
from flight muscles, I cut all of the nerve roots from the meso-
and metathoracic ganglia. To further reduce movement and
to expose the recording site, I removed the mouthparts and
gut from m o u t h to abdomen.
A manipulator on the magnet stand held a small, stainless
steel capillary tube. The tube (1) applied pressure downward
and backward on the prothoracic ganglion to stabilize the re-
cording site, (2) regulated the saline level around the recording
site, and (3) contained a Ag/AgC1 ground electrode.
Electrophysiological recording. I used glass micropipettes filled
with Lucifer Yellow CH (3% Lucifer in 1% LiC1 in the tip,
1 M LiC1 in the shaft) to penetrate axons in the cervical connec-
tive just posterior to the subesophageal ganglion. After filling
the electrodes, I etched their tips slightly by dipping them in
a 20 x dilution of hydrofluoric acid and washing them in dis-
tilled water. The two movements of a micromanipulator (Zeiss,
Jena) were used to hold the microelectrode and the stainless
steel spoon which supported the fused connective. This record-
ing situation was quite stable; penetrations often lasted for
more than an hour.
The signal from the electrode was amplified by a preampli-
fier (WPI M707). The signal to be recorded on magnetic tape
passed through a high pass filter ( t a u = l l 0 ms) to eliminate
the DC component. I used a 4 track cassette recorder (Fostex)
with F M adapters (Vetter) to store the signal for later analysis.
Signals were analysed with a digitizing waveform analyzer
(Data Precision D A T A 6000) and printed out with a digital
plotter (HP 7470A).
Visual stimuli. I rear-projected moving images on a translucent
plastic screen which covered nearly half of the spherical field
of vision of the dragonfly. The front face of the screen was
a square covering 90 ~ x 90 ~ visual angle. Dorsally and to both
sides, the screen angled back to cover the region of space be-
tween 45 ~ and 90 ~ back (Fig. 1).
The stimuli were black and white patterns photographed
on Kodak Ortho film projected from a 2 m distance with a
35 m m slide projector (Leitz Ortholux). The light intensity (sur-
face luminance) of the white areas on the screen, viewed from
the animal's side, was 170 cd/ft 2.
With this projection arrangement, both the size and the
shape of the projected pattern, measured in visual angle from
the surface of the eye, varied over the surface of the screen.
Figure 1 B shows examples of the size changes when measured
in visual angle at the dragonfly eye. These resulted from both
the variation in distance between the eye and screen and the
variation in projection and viewing angle. For example, a
square 4 ~ on a side when projected on the center of the 90 ~
square screen face was only 1.4 ~ visual angle when projected
on the corner. Here I will refer to each target by its size when
projected at the center of the screen. Variation was continuous
over each of the 4 faces of the screen, but there were discontinu-
ities at the borders between faces.
The beam reflected off a front-surface mirror before reach-
ing the screen. I produced pattern movements by tilting the
mirror horizontally or vertically, using the movement of two
loudspeaker cones to produce the deflections. These were driven
by a power amplified signal from a trapezoid generator. The
beam passed through a window in an opaque screen to restrict
its width to approximately the size of the screen, but some
projected movement was visible on the wall behind the animal.
302
R.M. Olberg: Identified visual target interneurons in the dragonfly
Fig. 2. Identification of target interneuron axons in cross sec-
tion of prothoracic ganglion. Top : Double exposure, using epi-
fluorescent illumination and phase contrast to show individual-
ly stained axon (MDT 1, arrow) with background detail. Bot-
tom: Enlargement of area outlined above showing the locations
of the target interneurons. Two arrows from DIT3 indicate
observed variation in the axon location. There was some varia-
tion in the order of MDT2, 3, and 4, but order shown here
was most typical. (Scale bar 50 gin)
Control experiments showed that these movements did not elicit
responses in the identified neurons except for the neuron
(DMTI) whose receptive field was in the backwards direction.
One series of experiments (Fig. 8) required simultaneous
movements of a small black target and a larger black and white
striped grating pattern. The grating pattern consisted of a series
of equally spaced black stripes, printed with a Kodak .copying
machine on a transparent acetate sheet. I used an XY recorder
(MFE) to move the grating back and forth about 5 cm in front
of the projection screen. The target was projected through the
acetate sheet onto the screen. The black target was visible even
on the dark stripes as a darker black square.
Anatomy. After recording from an axon, I injected Lucifer Yel-
low (5-10 nA) for 7 to 30 min. After allowing an additional
diffusion time of 3045 rain, I dissected out the head and thor-
acic ganglia. These I fixed in 4% formalin in a phosphate buffer
solution (10min) followed by 4% formalin in methanol
(45 min). I dehydrated the ganglia in propanol and cleared them
in methyl salicylate. Each filled neuron was drawn using a
Nikon drawing tube on a fluorescence microscope (Olympus
BHS) using a fluorescent pencil and UV light source to illumi-
nate the drawing. The ganglia were then embedded in soft Spurr
resin for sectioning. The cervical connectives and prothoracic
ganglion were sectioned transversely (20 gin) to locate the axon
829
within the cord and to determine its tract through the ganglion.
I photographed good sections within the connective and gangli-
on under phase contrast with Polaroid 667 Film and marked
the fluorescing axonal cross sections.
To determine the three dimensional arrangement of den-
dritic arborizations I sectioned the brain sagittally (30 gm) and
reconstructed the neurons from the sections with a drawing
tube.
In the adult dragonfly, the brain and optic lobes bend up-
ward to lie in a plane that is nearly at right angles to the
plane of the rest of the nervous system. Dorsal tracts in the
subesophageal and thoracic ganglia are therefore continuous
with and homologous to posterior tracts in the brain. For sim-
plicity I will refer here to the posterior surface of the brain
as its dorsal surface.
Results
M o s t l a r g e d i a m e t e r a x o n s in t h e a d u l t d r a g o n f l y
thoracic nerve cord were visual. The eight inter-
neurons described here have been grouped together
because they shared certain characteristics, both
a n a t o m i c a l a n d p h y s i o l o g i c a l . T h e y all d e s c e n d e d
from the brain, and their axons were among the
l a r g e s t in t h e v e n t r a l n e r v e c o r d ( V N C ) . T h e y all
responded selectively to small pattern (target) mo-
tion and not to the movement of large textured
backgrounds. Each was recorded and dye-injected
three or more times and showed consistent sensory
responses and anatomy.
O f t h e e i g h t t a r g e t i n t e r n e u r o n s , five s h o w e d
directional preferences. At least one target inter-
n e u r o n in e a c h c o n n e c t i v e w a s s e l e c t i v e f o r e a c h
d i r e c t i o n , left, r i g h t , u p a n d d o w n . ( D i r e c t i o n a l
p r e f e r e n c e s will b e p r e s e n t e d h e r e o n l y f o r t h e in-
terneurons whose axons extend down the right
c o n n e c t i v e . T h e i r s y m m e t r i c a l p a r t n e r s in t h e left
c o n n e c t i v e s h o w e d t h e e x p e c t e d r e v e r s a l in p r e f e r -
e n c e s t o left a n d r i g h t . ) S e v e n o f t h e i n t e r n e u r o n s
h a d r e c e p t i v e f i e l d c e n t e r s s o m e w h e r e in t h e f o r -
ward visual hemisphere, and the eighth (DMT1)
l o o k e d b a c k w a r d ( F i g . 3).
Anatomical identification
Injection of Lucifer Yellow into the axons within
the cervical connective between the subesophageal
a n d p r o t h o r a c i c g a n g l i a u s u a l l y e n a b l e d m e t o see
and draw the soma and major dendritic arboriza-
tions within the brain. The dye-filled axon profiles
w e r e a l w a y s v i s i b l e in c r o s s s e c t i o n s o f t h e c e r v i c a l
c o n n e c t i v e a n d t h e p r o t h o r a c i c g a n g l i o n ( F i g . 2,
top).
I have grouped and named the target interneu-
rons based on the longitudinal tract through which
t h e y p a s s in t h e p r o t h o r a c i c g a n g l i o n . T a r g e t i n t e r -
neuron axons always traveled through one of three
303
Fig. 3. Anatomical identification of the giant descending target interneurons. Camera lucida drawings (center) of Lucifer Yellow
stained neuron profiles in the brain, subesophageal ganglion and prothoracic ganglion viewed from dorsal aspect. Reconstruction
of serial sagittal sections (upper right) gives view of the same neuron looking from the right side of the brain (dorsal is left,
ventral right). Arrows in cross-sections of the cervical connectives (left top) at the anterior margin of the prothoracic ganglion
304
and cross-sections of the prothoracic ganglion (left bottom) show positions of the axons, retouched with white ink. (Scale bars
50 ~tm). Diagram of projection screen (lower right) indicates location and direction of greatest response to target movement.
(MDT3 and DIT3 were not directionally selective. DMT1 gave no responses to target movement on the screen - see text.)
In this and remaining figures, upward and rightward on screen mean upward and rightward with respect to the animal. (H,
Horizon; Z, line directly above the animal)

305
306
307
834 R.M. Olberg: Identified visual target interneurons in the dragonfly

Table 1. Anatomical and response characteristics of the target interneurons. Where there were discrepancies, the most commonly
observed responses are shown. Soma locations are with respect to axons recorded in the right connective

Cell MDT1 MDT2 MDT3 MDT4 DIT1 DIT2 DIT3 DMT1

Number of recordings 5 3 3 4 5 8 3 6
Soma location in brain LT RV LT RV LV RV LD LD
Left (L), Right (R),
Top (T), Dorsal (D)
Ventral (V)
Directional preference U L ND D R L ND -
Left (L), Right (R),
U p (U), Down (D)
Nondirectional (ND)
Spontaneous spike rate 0 0 0 0 0 0 0 0
Max. number of spikes to 7 10 4 20 20 8 17 8
target motion
Light on response E E E 0 E E E E
EPSP(E), No response (0)
Light off response E E E S E E S E
EPSP(E), Spike (S)
Air puff response 0 0 0 0 E 0 0 0
EPSP(E), No response (0)
Greater response to = = B B B = B -
white (W) or black (B) target?
Preferred target size A S L S S A A -
Small (S), Large (L), All (A)
Range in number of quadrants 8-10 7-10 7-11 4 7 10-12 10-12 12 -
where EPSPs were elicited
R a n g e i n number of quadrants 0- 7 1- 5 2- 5 4 1 11 1- 3 8 I0 -
where spikes were elicited

longitudinal tracts through the prothoracic gangli-
on, the median dorsal tract (MDT), dorsal inter-
mediate tract (DIT) or the dorsal median tract
(DMT). The tract chosen by a given interneuron
was consistent, and even the position of the axon
within the tract showed little variation (Fig. 2, bot-
tom).
The eight target interneurons descended from
the protocerebrum of the brain and projected at
least as far as the metathoracic ganglion. I could
discern very little branching below the prothoracic
ganglion, due to the distance from the recording
site. The soma location and branching pattern
within the brain and axon locations within the cer-
vical connective and prothoracic ganglion were
used to identify the cells anatomically (Fig. 3).
Within the brain the processes of the target
interneurons were restricted to the protocerebrum
(Fig. 3). The somata of three of them ( M D T I ,
MDT3, and DIT3) were located very near one an-
other in an antero-dorsal group of somata just me-
dial to the globuli cell cluster. All three of these
cells sent their axons down the connective contra-
lateral to their somata. The somata of DIT2,
M D T 2 and M D T 4 were located very near one an-
other, in a thin cortex of ventral cell bodies even
with or slightly medial to the alpha lobe of the
mushroom bodies. All of their axons projected
down the ipsilateral connective. The soma of DIT1
was also ventral, but lateral to those above. The
soma of DMT1 was located against the mid-dorsal
surface of the brain.
All of the target interneurons arborized in the
postero-dorsal region of neuropil very near the
point where their axons left the brain via the
circumesophageal connective (Fig. 3). None of
these interneurons made visible associations with
the mushroom bodies, central body or protocere-
bral bridge.
Within the protocerebrum the branching pat-
tern of the neurons was quite consistent from ani-
mal to animal, making each neuron easily recog-
nizable by its profile alone. Their large size, consis-
tent sensory responses, and consistent locations
within the longitudinal tracts of the prothoracic
ganglia provided further evidence that the neurons
were uniquely identified (Fig. 3).
There was no o b v i o u s correlation between the
profiles of the neurons within the brain and their
receptive field centers or directional preferences.
Figure 3 includes the target locations and move-
ment directions on the projection screen to which
each neuron responded maximally. Notice that
D M T I , whose soma and axon were not close to

308
R.M. Olberg: Identified visual target interneurons in the dragonfly 835

Fig. 4. Response of DIT2 to target movement on various screen locations. Outlines indicate the entire screen divided into 45 ~ 45 ~
quadrants. Target (4 ~ black square) moved right, left, down, and up with movements centered across each quadrant. Bars under
traces show 200 ms stimulus in indicated direction. Only one spike (s) was elicited by target movements (spike clipped to show
synaptic potentials). Note subthreshold EPSPs on most of screen and IPSPs to antipreferred (rightward) target movement. (Move-
ments 45~ ms; Scale bar 5 mV; Cell 6-28-1)

any of the others, was the only cell whose receptive
field was not in the forward direction.
Electrophysiological recordings
U p o n penetrating an interneuron, I presented a
preliminary set of sensory stimuli to determine
whether it was a target interneuron. I first moved
a white card (7.5 12.5 cm) in various locations
and directions around the animal ca. 3 cm from
the eye. The 3 mm black dot at its center was an
effective stimulus, producing spiking in all of the
target interneurons. A control card without a dot
never elicited spiking from the forward looking tar-
get interneurons (the backward looking DMT1
could not be tested in this way). A card covered
with black and white stripes also failed to elicit
spikes when moved at a variety of velocities.
To test the neuron for mechanosensory re-
sponses, I blew air onto the animal from the front
and sides, moved each of its wings up and down
and stroked its thorax and abdomen with a soft
brush. None of these stimuli elicited spiking in the
target interneurons. Table 1 summarizes some ana-
tomical and sensory properties of the target inter-
neurons.
In most target interneuron recordings com-
pound EPSPs and IPSPs were visible even though
the recording site in the axon was a long distance
(ca. 2 mm) from, and separated by a ganglion from
the probable site of spike initiation in the brain.
Such long length constants were due to the large
diameter of the axons. I did not see PSPs in other,
smaller axons descending from the brain recorded
during the study. Furthermore, when the largest
diameter fibers, M D T I and D M T I , were pene-
trated below the prothoracic ganglion, PSPs were
no longer visible.
Receptive fields
To determine the receptive field size and location,
I moved a 4 ~ or 16 ~ target up, down, left and right
in each of the twelve 45 ~ x 45 ~ quadrants compris-

309
836 R.M. Olberg: Identified visual target interneurons in the dragonfly

Cell 9 - 0 5 - a 2 Cell 11-18-bl

-Z

-H

9 9 t~
Preferred Preferred

-Z -Z

-H -H

9 4 9
Antipreferred Antipreferred
Fig. 5. Variation in responsiveness and apparent receptive field size. Recordings of DITI in two different animals (left and right).
Top screens show responses to rightward (preferred) target movement across entire 180 ~ screen along horizontal lines indicated.
Bottom screens show responses to three leftward target movements. In this and remaining figures, spikes are clipped to show
PSP activity. (Scale bars 200 ms, 5 mV)

ing the projection screen (Fig. 4). The receptive
field area in which this target motion evoked
EPSPs was much greater than the area in which
movements elicited spiking (Fig. 4). To quantify
this I counted the number of screen quadrants in
which I saw PSPs to target motion and the number
in which I saw spikes (Table 1). For all but MDT4
and DMT1 movements nearly anywhere on the
screen (10 to 12 quadrants) elicited EPSPs.
The receptive field area in which target move-
ment elicited spiking varied greatly across prepara-
tions for a given cell type. Figure 5 shows traces
recorded from DIT1 in two different preparations.
Rightward, preferred-direction movements elicited
a total of only 4 spikes within a very restricted
region in the first DITI, but the same stimuli eli-
cited 67 spikes in the second. The underlying pat-
tern of compound EPSPs in the two rcordings,
however, was almost identical. The peaks of depo-
larization of the left cell in Fig. 5 corresponded
to the highest spike densities in the right cell. What
appeared to vary was the overall level of excitabili-
ty. Even antipreferred motion (lower screens of
Fig. 5) elicited 9 spikes in the second DIT1, while
it elicited none from the first. Similar variability
occurred for M D T I and MDT2; the variability
was less extreme in MDT3, DIT2, and DIT3 (Ta-
ble 1). For all of these neurons, variability between
preparations was much greater than variability
over time within a given preparation. The spiking
receptive field area of M D T 4 was quite consistent
among preparations. It was always confined to a
ca. 30 ~ wide vertical strip along the midline above
the horizon.

310
R.M. Olberg: Identified visual target interneurons in the dragonfly 837

DIT1 DIT2

32 ~

Fig. 6. Target size selectivity in DIT1 and DIT2. Square black targets of indicated sizes moved in preferred direction (right
for DIT1, left for DIT2) during the period indicated by the stimulus marker below bottom traces. Checkerboard pattern (8 ~
black and 8~ white squares) covered entire screen. (All movements 45~ Scale bar 5 mV; DIT1 cell 9-5-2; DIT2 cell 9-2-1)

-=-=-=-=-_-.,,,,,,,~.~~ , ~ , ~ , r ~ ~ - the smaller targets, 2 ~ or 4 ~ on a side, over the

larger targets, 16 ~ or 32 ~ (Fig. 6). Other neurons
(DIT2, DIT3, M D T I ) responded about equally to
targets of all sizes (Fig. 6). One neuron (MDT3)
,;_-;_- - consistently preferred the larger targets, showing
little or no response to 2 ~ or 4 ~ squares. Target
interneurons rarely showed even a single spike to
movements of a checkerboard pattern, which cov-
ered the entire projection screen (Fig. 6, bottom
traces).
I

I Directional selectivity
Fig. 7. Opposing directional preference in responses of target
interneuron to target and checkerboard pattern movement. Pre-
ferred direction for target movement in this cell (MDT4) was
downward (trace 3). Downward movement of large pattern
showed hyperpolarizing response (trace 1) and upward pattern
movement showed depolarizing response (trace 2). (Movement
45~ ms; Scale bar 5 mV; Cell 7-12-a2)
Selectivity for target size
All of the target interneurons responded to object
movement but not to movement of a large textured
background. This visual discrimination occurred
over a range of velocities from 45~ to 225~
When moving black squares of various sizes
were presented to these interneurons, their re-
sponses fell into one of three types (Table 1). Some
cells (DIT1, MDT2, MDT4) consistently preferred
Five of the target interneurons showed clear direc-
tional selectivity (Fig. 3); movement in their pre-
ferred direction elicited spikes, and movement in
the antipreferred direction usually did not (Fig. 7).
For two neurons (MDT2, DIT2), compound
IPSPs indicated that target movement in the anti-
preferred direction was inhibitory (Fig. 4).
The directionally selective target interneurons
(MDT1, MDT2, MDT4, DIT1, DIT2) also
showed direction-dependent subthreshold re-
sponses to movement of the large checkerboard
pattern. For these cells, the preferred direction for
the large pattern movement was opposite to that
for the small target (Fig. 7, cf. 2nd and 3rd traces).
Checkerboard pattern movements very rarely eli-
cited any spikes from target interneurons, but these
invariably resulted from pattern movement in the

311
 838
-J,,,,...,,,,, , , ..... . . . .
-Illi II III . . . . . . . . . .
II1'tll
Fig. 8. Responses to target movement with and without a mov-
ing background. DITI produced spikes in response to right-
ward 4 ~ target movement (trace 3) and was depolarized (trace
2) by leftward movement of a striped pattern (8 ~ black and
white stripes covering entire screen), but leftward stripe move-
ment completely inhibited response to the rightward target
movement (trace 4). Bottom trace shows recovery of target
response in the absence of stripe movement. The five traces
are shown in the sequence they were obtained (i.s.i.=15 s).
(Movements 45~ ms; Scale bar 5 mV; Cell 9-5-2)
antipreferred direction for targets. Checkerboard
movement in the direction preferred for the target
elicited hyperpolarizing responses (Fig. 7, top).
Since the target interneurons preferred opposite
directions of target and background motion, per-
haps they would respond maximally to a combina-
tion of target and background movements in oppo-
site directions. Figure 8 shows the results of an
experiment designed to test this hypothesis. Al-
though this DIT1 interneuron showed spikes to
rightward target movement and a depolarization
to leftward background movement, the combina-
tion of the two produced no spikes at all. This
result was confirmed in two other preparations
(DIT1, DIT2).
Discussion
The most important properties of the dragonfly
target interneurons are: (1) their axons are very
large (and therefore easily recorded and recogniz-
able), (2) they are highly selective visual interneu-
rons, responding only to the movement of relative-
ly small objects and (3) the majority are direction-
ally selective.
Receptive field structure
An unexpected finding in this investigation was
that receptive field size for a given neuron may
vary greatly between preparations if spikes are
R.M. Olberg: Identified visual target interneurons in the dragonfly
used as a criterion (Fig. 5, Table 1). The visual area
in which there was some subthreshold response,
either depolarizing or hyperpolarizing, to target
motion was usually much greater than the limited
region in which targets elicited spiking responses.
When these subthreshold responses are taken into
account, the size and 'topography' of the receptive
field, its peaks and valleys of excitability, were
quite consistent for a given neuron from prepara-
tion to preparation.
It seems unlikely that the spike variability be-
tween preparations is an artifact of the recording
situation. It is possible that electrode injury to the
axon could cause the spread of depolarization to
the spike initiating zone, influencing spike produc-
tion. However, such injuries would also reduce
PSP amplitude at the recording site due to reduced
input resistance. Such systematic differences in
PSP amplitude were not seen (Fig. 5).
The variable receptive field area in which spik-
ing occurred was apparently a function of the over-
all excitability of the cell, that is, of the ability
of the visually evoked graded potentials to elicit
spiking. Three possible mechanisms may underlie
the differences in excitability from cells of the same
type: 1) postsynaptic responses, although qualita-
tively similar, may vary in their amplitudes, 2)
spike threshold may vary or 3) tonic excitation
or inhibition may bring the cell membrane closer
to or farther from threshold. The available data
do not exclude any of these possibilities.
Directional selectivity
Some target interneurons showed opposing direc-
tional preference to target and background motion
(Figs. 7 and 8). One explanation for such direction-
ally opposed responses is that the target inter-
neuron receives tonic inhibition from a wide field
movement detector with the identical directional
preference. A large textured pattern moving in the
preferred target direction would then hyperpolar-
ize the target interneuron. The depolarization seen
in response to the same pattern moving in the op-
posite direction would result from disinhibition,
due to the suppressed activity of the wide field
detector.
Wide field movement detectors with the appro-
priate directional preferences to mediate the cir-
cuitry proposed above are present in the dragonfly
CNS (Olberg 1981 b). However, the above explana-
tion does not account for the inhibition of target
response by the large textured pattern moving in
the reverse direction from the target (Fig. 8,
trace 4). T h e implication of this inhibition is that
312
R.M. Olberg: Identified visual target interneurons in the dragonfly
although background motion in the antipreferred
direction elicits a depolarizing response, this depo-
larization is in fact inhibitory. This suggests the
presence of at least two different ionic mechanisms,
one hyperpolarizing and one slightly depolarizing,
underlying inhibitory inputs to the target inter-
neurons. The nature of the inhibition underlying
directional and size selectivity will be considered
in more detail in a later paper.
Comparison with other descending
interneurons in insects
In its size, structure, and location within the central
nervous system, MDT1 bears a resemblance to the
orthopteran Descending Contralateral Movement
Detector (DCMD, for review see Rowell 1971).
The soma size and location, the profile within the
brain, and the axon size and location within the
nerve cord are all quite similar (cf. O'Shea et al.
1974). This suggests that MDT1 and DCMD may
be homologous giant neurons. Like MDTI,
DCMD responds maximally to small pattern
movement and is inhibited by large pattern move-
ment (Palka 1967; Pinter 1977, 1979).
There are differences, however, in the visual
properties of DCMD and MDT1. Whereas
DCMD's visual input is strictly from the contralat-
eral eye, MDT1 receives binocular input, strongly
weighted for input directly in front of the animal.
MDT1 is selective for upward pattern movement.
DCMD is not directionally selective.
In general, the large, descending interneurons
that have been described in insects have not been
as selective as the dragonfly target interneurons
in their visual properties. Although directionally
selective descending visual interneurons have been
found in grasshoppers (Kien 1974, 1975; Reichert
et al. 1985) and in the moth Manduca (Rind 1983),
these have responded to wide field movement. The
target interneurons described here represent an ad-
ditional, inhibitory step in the processing of the
visual signal. A target moving in the preferred di-
rection is only excitatory if it does not exceed a
certain dimension. Beyond that point, inhibition
outweighs the excitation produced by the target.
Role of the target interneurons in behavior
There are two reasons to assign an important be-
havioral role to the target interneurons. First is
the size of their axons. The eight axons represent
more than 10% of the total cross-sectional area
of each connective. Their large size implies a selec-
tive advantage to rapid transmission of the infor-
839
mation they carry down the cord. Second is their
direct influence on the flight system. High fre-
quency (200 Hz), intracellular stimulation of some
individual target interneurons produces move-
ments of from two to four wings, even when the
animal is not flying (Olberg 1978; Olberg in prepa-
ration). Their selective visual responses and their
influence on the flight musculature imply that the
target interneurons control, in part, oriented flight
responses to objects whose images move across the
dragonfly's retina.
The moving objects of greatest importance to
a flying dragonfly are flying insect prey, flying pre-
dators, and other dragonflies. Six of the target in-
terneurons are dominated by visual input from a
forward direction above the horizon. This is the
direction in which prey are viewed during the ap-
proach, the dragonfly sweeping up from below its
intended prey (Mayer 1957). Potential predators
(birds) also usually approach from above. Thus
both prey and predators would elicit a response from
various combinations of the target interneurons.
It seems likely that the activity of the target inter-
neurons governs the first turning response of the
flying dragonfly to the moving prey or predator.
Acknowledgements. I am indebted to David Wohlers for teach-
ing me embedding and sectioning techniques, and to Dawn
Tamarkin for her technical help. I thank Robert Pinter for
help with some of the experiments and with light level measure-
ments, and Andrea Worthington for her critical review of the
manuscript. This work was supported by National Science
Foundation Grant No. BNS-84-06254, by a Cottrell College
Science Grant from Research Corporation, and by grants from
the Union College Faculty Research Fund.
References
Bacon J, M6hl B (1983) The tritocerebral commissure giant
(TCG) wind-sensitive interneurone in the locust. I. Its activi-
ty in straight flight. J Comp Physiol 150:439-452
Bacon J, Tyrer M (1978) The tritocerebral commissure giant
(TCG): a bimodal interneurone in the locust, Schistocerca
gregaria. J Comp Physiol 126 : 317-325
Bentley D (1977) Control of cricket song patterns by descending
interneurons. J Comp Physiol 3[16:19-38
Bicker G, Pearson K G (1983) Initiation of flight by an identi-
fied wind sensitive neurone (TCG) in the locust. J Exp Biol
104:289-293
Blondeau J (1981) Electrically evoked course control in the
fly Calliphora erythrocephala. J Exp Biol 92:143-153
Catton WT, Chakraborty A (1969) Single neurone responses
to visual and mechanical stimuli in the thoracic nerve cord
of the locust. J Insect Physiol 15 : 245-258
Collett TS, Land M F (1978) How hoverflies compute intercep-
tion courses. J Comp Physiol 125:191-204
Kien J (1974) Sensory integration in the locust optomotor sys-
tem. II. Behavioral analysis. Vision Res 14:1255-1268
Kien J (1975) Neuronal mechanisms subserving directional se-
lectivity in the locust optomotor system. J Comp Physiol
102:337-355
313
840
Kien J, Altman JS (1984) Descending interneurones from the
brain and suboesophageal ganglia and their role in the con-
trol of locust behaviour. J Insect Physiol 30: 59-72
Mayer G (1957) Bewegungsweisen der Odonatengattung
Aeshna. Osterreich Arbeit Jb Stadt Linz 4: 211-219
Olberg RM (1978) Visual and multimodal interneurons in dra-
gonflies. PhD Dissertation, Univ. of Washington, Seattle
Olberg RM (1981a) Object- and self-movement detectors in
the ventral nerve cord of the dragonfly. J Comp Physiol
141 : 327-334
Olberg RM (1981b) Parallel encoding of direction of wind,
head, abdomen, and visual pattern movement by single in-
terneurons in the dragonfly. J Comp Physiol 142:27-41
Olberg RM (1983) Pheromone-triggered flip-flopping inter-
neurons in the ventral nerve cord of the silkworm moth,
Bombyx mori. J Comp Physiol 152:297-307
O'Shea M, Rowell CHF, Williams JLD (1974) The anatomy
of a locust visual interneurone; the descending contralateral
movement detector. J Exp Biol 80:191-216
Palka J (1967) An inhibitory process influencing visual re-
sponses in a fibre of the ventral nerve cord of locusts. J
Insect Physiol 13:235-248
Pearson KG, Robertson RM (1981) Interneurons coactivating
hindleg flexor and extensor motoneurons in the locust. J
Comp Physiol 144: 391-400
Pinter RB (1977) Visual discrimination between small objects
and large textured backgrounds. Nature 270:429-431
R.M. Olberg: Identified visual target interneurons in the dragonfly
Pinter RB (1979) Inhibition and excitation in the locust DCMD
receptive field: spatial frequency, temporal and spatial char-
acteristics. J Exp Biol 80:191-216
Reichert H, Rowell CHF (1985) Integration of nouphaselocked
exteroceptive information in the control of rhythmic flight
in the locust. J Neurophysiol 53:1201-1218
Reichert H, Rowell CHF, Griss C (1985) Course correction
circuitry translates feature detection into behavioural action
in locusts. Nature 315:142-144
Rind FC (1983) A directionally sensitive motion detecting neu-
rone in the brain of a moth. J Exp Biol 102:253-271
Rowell CHF (1971) The orthopteran descending movement de-
tector (DCMD) neurones: a characterisation and review.
Z Vergl Physiol 73:167-194
Rowell CHF, Pearson KG (1983) Ocellar input to the flight
motor system of the locust: structure and function. J Exp
Biol 103:265-288
Simmons P (1980) A locust wind and ocellar brain neuron.
J Exp Biol 85:281-294
Strausfeld N J, Bacon JP (1983) Multimodal convergence in the
central nervous system of insects. In: Horn E (ed) Multimo-
dal convergence in sensory systems. Fortschr Zool 28, Gus-
tav Fischer, Stuttgart, pp 47-76
Tanouye MA, Wyman RJ (1980) Motor outputs of giant nerve
fiber in Drosophila. J Neurophysiol 44 :405-421
Tanouye MA, King DG (1983) Giant fibre activation of direct
flight muscles in Drosophila. J Exp Biol 105:241-251
314
J Comp Physiol A
DOI 10.1007/s00359-007-0223-0
ORIGINAL PAPER
Eye movements and target Wxation during dragonXy
prey-interception Xights
R. M. Olberg R. C. Seaman M. I. Coats A. F. Henry
Received: 21 December 2006 / Revised: 5 March 2007 / Accepted: 23 March 2007
Springer-Verlag 2007
Abstract The capture of Xying insects by foraging drag-
onXies is a highly accurate, visually guided behavior.
Rather than simply aiming at the preys position, the drag-
onXy aims at a point in front of the prey, so that the prey is
intercepted with a relatively straight Xight trajectory. To
better understand the neural mechanisms underlying this
behavior, we used high-speed video to quantify the head
and body orientation of dragonXies (female Erythemis sim-
plicicollis Xying in an outdoor Xight cage) relative to an
artiWcial prey object before and during pursuit. The results
of our frame-by-frame analysis showed that during prey
pursuit, the dragonXy adjusts its head orientation to main-
tain the image of the prey centered on the crosshairs
formed by the visual midline and the dorsal fovea, a high
acuity streak that crosses midline at right angles about 60
above the horizon. The visual response latencies to drifting
of the prey image are remarkably short, ca. 25 ms for the
head and 30 ms for the wing responses. Our results imply
that the control of the prey-interception Xight must include
a neural pathway that takes head position into account.
Keywords DragonXy Eye movement Flight
Image Wxation Vision
Abbreviation
TSDNs Target selective descending neurons
R. M. Olberg (&) R. C. Seaman M. I. Coats A. F. Henry
Department of Biological Sciences,
Union College, Schenectady, NY 12308, USA
e-mail: olbergr@union.edu
Introduction
The interception of prey insects in Xight is an essential ele-
ment in the dragonXys behavioral repertoire. DragonXies
of the family Libellulidae spend much of their days
perched, only taking oV to capture small insects as they Xy
overhead. Remarkably, they do not Xy directly towards
their prey, but rather aim at a point in front of the prey, so
that they intercept the prey with a relatively straight Xight
trajectory (Olberg et al. 2000). This interception strategy is
highly successful; Baird and May (1997) reported a capture
rate of 76% in their study of foraging females (Pachydiplax
longipennis) and in our S-VHS video study of males of var-
ious Libellullid species foraging from ponds edge we saw
only a single miss in 38 Xights, a success rate of 97%
(Olberg et al. 2000). Prey capture Xights are usually quite
short, often less than 200 ms from takeoV to contact with
the prey item. These short Xights leave little time for eva-
sion, and in our Weld studies we have never observed obvi-
ous evasive maneuvers by the prey insect.
Prey interception is guided by the visual system. Of spe-
cial interest for this study is a distinct region of the com-
pound eye, with ommatidial facets much larger than those
in the rest of the eye, covering the dorsal surface of the
eyes. Dominated by blue and UV photoreceptors (Labhart
and Nilsson 1995), this region is well adapted for foraging
on insects against blue sky. The dorsal ommatidial axes are
nearly parallel, so that most of the dorsal surface of the eye
is directed at very narrow upward and forward crescent, a
fovea about 20 in height and intersecting visual midline
about 60 above the horizon (Fig. 1). Within the dorsal
fovea small inter-ommatidial angles (Sherk 1978) result in
resolution of ca. 15 min of arc (or even less for moving
objects, Rowe 1987)
123
315

Fig. 1 Method for estimating direction of view of the fovea. Shown is
a lateral view of head of Erythemis simplicicollis female, showing
direction viewed by the center of the dorsal fovea (black arrows) in
relation to the line of the back of the eye and the front of the head
(frons)
In earlier work (Olberg et al. 2000) we proposed the
following simple hypothesis to explain prey interception:
during pursuit the dragonXy steers so as to maintain the
prey at a constant bearing, resulting in a collision course
with the prey. This interception strategy, termed propor-
tional navigation, would allow the dragonXy to intercept
prey without any information about the preys distance. The
strategy would only require steering to minimize retinal slip
of the preys image. Target-selective descending neurons
(TSDNs) that could provide this control have been
described (Olberg 1986; Frye and Olberg 1995).
The simplest way in which proportional navigation
could be implemented by the Xying dragonXy would
require that the dragonXys head be rotationally stabilized
during pursuit. Such visual stabilization has been well doc-
umented in free Xying blowXies (van Hateren and Schilstra
1999). If the dragonXys head were stabilized relative to the
visual environment, small movements of the targets image
on the retina would indicate changes in the absolute angle
from the dragonXy to its prey. The TSDNs would respond
to such movements and send compensating instructions to
the Xight motor. By minimizing the drift of the target image
in this way, the dragonXy would intercept its prey.
123
J Comp Physiol A
The above scenario depends heavily on the rotational
stabilization of the head relative to the visual environment.
In this study, we used high resolution, high-speed video to
examine the orientation and movement of the head before
and during prey pursuit. Our goal was to determine whether
the head was rotationally stabilized during prey capture
Xights. Furthermore we sought to understand the role, if
any, of the dorsal fovea in monitoring prey movements.
Materials and methods
Flight cage
Adult dragonXies do not normally forage in captivity and
quickly starve even when their laboratory cage is loaded
with Xying insect prey. This problem was overcome with an
outdoor Xight cage whose mesh walls and ceiling admitted
full sunlight, including UV. The Xight cage was constructed
of 1.9 cm copper pipe (2.9 m square and 2.5 m high) cov-
ered with polyethylene/polypropylene aquatic netting (US
Netting, Erie, PA). The 0.5 cm gauge of the netting allowed
small insects, such as mosquitoes and midges to pass freely
through the cage walls providing natural prey for the cap-
tive dragonXies. The cage was located in a courtyard, pro-
viding shelter from wind but receiving full sunlight during
the videotaping period, generally from 9 a.m. to 12:30 p.m.
Perches constructed of wood or Styrofoam were dispersed
within the cage and a ca. 1-m diameter plastic pool and pot-
ted plants were added to provide water and shade.
Typically 68 female dragonXies Erythemis simplicicol-
lis (family Libellulidae) were netted in the wild and intro-
duced into the Xight cage. By the following day the
dragonXies showed behavior indistinguishable from that
seen in the Weld, perching and periodically taking oV after
small Xying insects. Male dragonXies were not used
because they did not forage in the Xight cage.
To elicit prey capture Xights, we moved 2 mm white
glass beads above the perched dragonXy. The white beads
were just as attractive as black beads and much easier to see
in the video. To restrict the behavior as much as possible to
a single plane, we moved the bead above the animal and in
the same plane as the animal, normal to the cameras Weld
of view. During the Wrst summer, the beads were aYxed
with 5 min epoxy to the end of a Wne (75
m) tungsten
wire, approximately 75 cm in length. This wire was Wne
enough to be nearly invisible to our eyes and stiV enough to
produce small jerky movements similar to those of small
Xying insects. The tungsten wire was attached to a stiV steel
wire (1.8 mm diameter) 50 cm in length with epoxy and
this steel wire was inserted into the hollow end of a 1.1 m
Wshing rod that was manipulated by hand. During the sec-
ond summer the bead was aYxed to the center of a Wne
316
J Comp Physiol A
nylon monoWlament Wsh line stretched between the ends of
a Y-shaped fork of bamboo. This more rigid mounting gave
us better control of the beads movements. We believe that
the Wsh line was not visible to the foraging dragonXies,
because they occasionally collided with it in Xight. We
found no diVerence in the response of the dragonXies to the
two diVerent methods of bead presentation.
We found that the animals became habituated to our
experimental conditions and after several capture attempts
stopped taking oV after our artiWcial prey object. In a few
cases a small insect was applied to the bead; this provided a
reward to the foraging dragonXy and reduced the behav-
ioral habituation to the bead stimulus. Because of this
habituation, we usually released the dragonXies after
23 days of videotaping and replaced them with fresh ani-
mals.
Video
High-speed video clips were captured with a Redlake
Motion Pro 2000 (Redlake MASD LLC, Tucson, AZ,
USA) camera with an Elicar 90 mm macro lens, which pro-
vided a 12 cm Weld at 1 m. We placed the camera from 0.5
to 1.5 m away from the perching animals. To aid in our
analysis of head position, we designed and oriented the
perches to encourage the dragonXies to perch either aligned
with and towards the camera or at right angles to the axis of
view. Because dragonXy foraging Xights are invariably
upward, we positioned the camera with the dragonXy at the
bottom of the cameras Weld of view. The camera was con-
nected to an Acme KB-108-1 Weld computer running
Microsoft Windows 2000 and Redlake MIDAS software
for capture at 500 frames per second and 1,280 1,024
pixel resolution.
Analysis
Video clips were digitized frame-by-frame by hand, using
the Redlake MIDAS analysis software. The coordinates
were then exported to Microsoft Excel spreadsheets for
analysis. For each clip we digitized the bead position, cen-
ter of the head and two other points on the head. In side
shots, we also digitized the point at the tip of the tail.
Twelve of the clips were digitized a second time by a diVer-
ent person with the results that were always within a few
pixels on one another.
To calibrate approximate distances we used average
dimensions of Erythemis females, 6 mm head width or
41 mm body length. Because we found about 3% variability
in the length of the animals we measured and because the
animal was not always oriented precisely head-on or at right
angles to the camera-viewing axis, we estimate that our dis-
tance calibrations are only accurate within about 5%.
Head orientation
One of our goals in this project was to determine whether
the dragonXy uses its dorsal fovea to Wxate the prey during
pursuit. When viewed from the side the front and back of
dragonXys head appear to be bounded by nearly straight
lines, and we used those lines to estimate the direction of
view of the center of the fovea. To do this we viewed iso-
lated Erythemis heads from directly above using a dissect-
ing microscope and axial illumination. We rotated the
head, mounted on a goniometer for angle measurements,
and observed the pseudopupil until the center of the fovea
was directed straight up. We identiWed the center of the
fovea as the region with largest pseudopupil; the entire
fovea was about 22 wide, extending about 11 forward
and backward from the center along the dorsal midline.
We then photographed the head from the side and mea-
sured the angles of the front and back of the head (Fig. 1).
In three animals we found that the center of the fovea was
approximately 34 (33.5, 34, 36) forward from the line
of the back of the head and 19 (17.3, 19.5, 20.3) for-
ward from the line of the front of the head. Given the 3
variability in our angle measurements and also our uncer-
tainty in assigning a precise line to the front or back of the
head, especially in images somewhat blurred by motion,
we estimate that our measurements are only accurate to
within about 5.
Results
In this study we obtained high-speed (500 frames/s) video
clips of 128 pursuit Xights by Erythemis simplicicollis
females. Forty-eight of these were views directly from the
front (n = 33) or directly from the side (n = 15), and we
analyzed these 48 clips for head position before and, if pos-
sible, during pursuit Xights. We estimate that these 48 clips
represent about 25 diVerent individual dragonXies. The 48
clips include data from both means of bead presentation;
the results from the two methods were indistinguishable
from one another. In this report we use the term prey to
refer to the stimulus, a white bead. The bead was usually
presented within about 15 cm of the animal and sometimes
as close as 5 cm, resulting in short pursuit-Xight durations,
typically less than 200 ms (mean 168 ms; range
62480 ms). These Xight durations were similar to the aver-
age duration of 184 ms measured in naturally foraging
dragonXies (Olberg et al. 2000).
In this high-speed video study we conWrmed the conclu-
sion of an earlier Weld study (Olberg et al. 2000) that drag-
onXies use an interception strategy, i.e., the dragonXys
Xight course is aimed at a point in front of its intended prey.
In almost all cases, when the prey maintained consistent
123
317
 J Comp Physiol A

Xight speeds and direction, the dragonXys Xight path was
nearly a straight line to the point of interception (Fig. 2ac).
Straight Xights such as those shown in Fig. 2ac do not
arise from a ballistic strategy in which the dragonXy Xies
along a predetermined interception Xight track. Flight
tracks in which the prey changed direction (Fig. 2df)
showed that the dragonXy reacts to changes in the preys
trajectory with changes in its own Xight path (Fig. 2df).
Behavioral response latencies
Circuitous Xight tracks such as those in Fig. 2df allowed
us to estimate the latency of the dragonXys reaction to
changes in the direction of its prey. In some Xights, we
could distinguish exaggerated wing strokes before prey
induced turns (Fig. 3a, lower inset). The latency measured
from prey deviation to wing correction ranged from 26 to
42 ms (mean = 29 ms, s = 6.4, n = 6). We measured an
average latency of 38 ms (s = 5.1, n = 16) to an observable
course correction after abrupt changes in prey direction.
The annotated Xight track shown in Fig. 3 illustrates a
wing-response latency of 28 ms resulting in a visible turn in
the Xight track 36 ms after the prey turn.
A second method conWrms and reWnes our behavioral
latency measurements. Rather than determining behavioral
latency by inspection of wing movements or by Xight path
deviation, which requires an accurate assessment of the cue
used by the dragonXy, we examined the x-coordinate of the
dragonXy and prey locations over time. For example, in
Fig. 3b a plot of the x component of the two paths illustrates
graphically the dragonXy position lagging behind the prey
position. The highest correlation between these two paths
(0.916) occurred with a latency of only 28 ms, very close to
the 29 ms average wing response latency measurement.
Head orientation
During pursuit, the dragonXys head was precisely oriented
with respect to the prey. In 18 cases we were able to quan-
tify the dragonXys head angle during the pursuit Xight.
Seven of these were viewed from head-on allowing us to
determine the direction of view of the visual midline
(Fig. 4a), and 11 of these were from the side, allowing us to
determine the direction viewed by the dorsal fovea
(Fig. 4bc). With only one exception, we found that the
dragonXy oriented its head during Xight to maintain its tar-
get on or very near the crosshairs formed by the intersec-
tion of the visual midline with the dorsal foveal band.
Two features of the top trace in Fig. 4a suggest that the
rotation of the head to Wxate the prey on midline includes a
predictive component to prey tracking. Firstly the head
rotation at takeoV continues along the same slope (i.e.,
angular velocity) that the prey moved about 26 ms earlier.
Thus the eye traces the path that the prey would have taken
had it not reversed direction. The change of direction of the
head rotation lags behind the change of direction of the
prey by about 26 ms. Secondly, as the prey swings back to
the left during the later part of the trace, the head angle fol-
lows the prey angle with no obvious time lag.
The histograms of Fig. 5 show that during most of the
pursuit the dragonXy is kept within about 10 of midline
and 10 of the foveal center. Examples with the prey mov-
ing forward (Fig. 5c) and backward (Fig. 5d) above the
dragonXy reveal no obvious bias in the foveal tracking of

Fig. 2 Six Xight tracks. Open

circles indicate prey location
with larger open circles every
10th frame (20 ms). Filled dia-
monds indicate center of dragon-
Xy head with large Wlled circle
every 20 ms. Large circles of
dragonXy and prey correspond
in time; the Wrst circle indicates
positions in Wrst frame in which
legs lose contact. Gray lines at
lower end of dragonXy trace
show position and length of
body at takeoV (animal viewed
head-on in c, d, f). Plus signs at
ac and f indicate head and prey
locations at capture. Calibration
bars = 2 cm. ac. DragonXy fol-
lows straight-line interception
course. (clips 6-29a, 7-8a, 7-8b).
df DragonXy turns in response
to changes in prey direction
(clips 8-20u, 6-24c, 7-29b)

123

318
J Comp Physiol A

Fig. 3 Flight track annotated to

indicate visual response latency.
a A reversal in the preys course
at 102 ms results in a very large
amplitude wing beat at 130 ms
(lower photo inset) compared to
normal wing beat (upper photo
inset). The large wing stroke
elicits a change in Xight track
visible by 138 ms. Symbols as in
Fig. 2. Calibration bar = 2 cm.
b A plot of the horizontal com-
ponent of the prey and predator
tracks shows that the dragonXy
oscillation trails the prey oscilla-
tion by 28 ms (clip 630f)

the prey. There may be a slight diVerence in tracking
between cases in which the dragonXy Xies forward and
cases when it Xies backward in pursuit. The median angle
between prey and fovea during forward Xight is 4.3,
indicating that the center of fovea is aimed behind the bead
relative to Xight direction. The median angle between prey
and fovea during backward Xight is 3.1, indicating that the
center of fovea is again aimed somewhat behind the bead
relative to Xight direction. These slight skews notwithstand-
ing in the accumulated histograms (Fig. 5a, b), as well as in
each of the histograms of diVerent pursuit scenarios
(Fig. 5cf), the bead is kept within 10 of the center of the
crosshairs at least 85% of the time.
Of the 18 clips analyzed for head orientation during
Xight, only one (Fig. 6) did not follow the pattern estab-
lished by all the others, suggesting that other visual pursuit
strategies are available to the foraging dragonXy. In this
instance the dragonXy was viewed from head-on as it Xew
precisely sideways in pursuit of the prey. For most of this
pursuit the dragonXy kept its head nearly in level, rather
than rotating the head to Wxate the prey on midline
(Fig. 6b). We could not ascertain whether the preys image
was maintained on the lateral region of the dorsal fovea, but
the head position as viewed from the front was consistent
with this possibility.
Discussion
Three principal points emerge from this high-speed video
study of dragonXy prey interception. (1) The dragonXy
steers an interception course, leading its prey, (2) during
pursuit the dragonXys head rotates freely with respect to
the rest of the body to Wxate the preys image on the cross-

123

319

Fig. 4 Fixation of the prey image on midline and dorsal fovea. Open
circles indicate angle from the dragonXy head to the prey. Filled dia-
monds indicate direction of view of midline (a) or dorsal fovea (b, c).
Dashed line indicates 0 ms, the time of takeoV. a Head orients in Xight
hairs formed by the intersection of dorsal foveal band with
visual midline, (3) the behavioral latency in response to the
movement of the prey image is remarkably fast, typically
less than 30 ms. In this report we have focused on relatively
short pursuit Xights (usually 515 cm), because we could
accurately measure head orientation only with close-up
camera views. However, short Xights such as these are not
uncommon in foraging Erythemis females. They are not
restricted to such short forays, however; we also observed
them taking oV after prey at a distance of 0.5 m or more.
Points (1) and (2) above imply that the dragonXy is not
usually Xying in the direction viewed by the midline dorsal
fovea. Instead, the dragonXys Xight course is directed at a
point in front of the preys movement, while the head rotates
Fig. 5 Histograms showing distribution of midline or fovea angles
relative to the angle to the prey. Visual angle is subtracted from prey
angle in each frame, and histogram of each clip is normalized to 100%.
Normalized histograms are averaged. a Angle to prey is centered on vi-
sual midline in head-on clips. b Angle to prey is centered on fovea in
the accumulated sideways clips. Positive values in bf indicate that the
123
J Comp Physiol A
to maintain the prey image near visual midline (clips 7-22i, 7-8b).
b, c Head orients to maintain the prey image near the center of the
fovea when the prey moves forward (a clips 7-29c, 6-29a) or backward
(b clips 7-29a, 7-22e) relative to the dragonXy.
to keep the target Wxated. These results then require us to
modify our earlier hypothesis (Olberg et al. 2000) for the
control of interception Xight. We had hypothesized that the
dragonXy would stabilize its head, and thus its eyes, on the
visual environment, responding with compensatory Xight
corrections to any drift of the preys image on the retina. In
this way the preys image would be stabilized on the eye and
the Xight would describe an interception course. In a land-
mark study of head and body positions and orientations of
blowXies in free Xight, van Hateren and Schilstra (1999)
showed that head is rotationally stabilized relative to the
visual environment with rapid head saccades accompanying
course changes. However, in the blowXy study Xies were
not actively orienting to an object, as were the dragonXies.
prey angle is behind the center of the fovea. c, d Subsets of the side-
ways clips with target moving forward (dotted arrow, c) or backward
(dotted arrow, d) relative to the animal show no obvious bias. e, f Sub-
sets of the sideways clips with the animal Xying forward (dotted arrow,
e) or backward (dotted arrow, f) show no obvious bias
320
J Comp Physiol A
Fig. 6 Head-on prey-pursuit Xight showing no Wxation of the prey im-
age on midline. a DragonXy Xies sideways on an intersection course.
Oscillations in Xight track reXect wing beats. Symbols as in Fig. 2.
Calibration bar = 2 cm. b Time course of angles to the prey and direc-
tion visual midline. Symbols as in Fig. 4a. c Histogram of diVerence
between prey angles and visual midline, as in Fig. 5a
It is clear from the results we report here that the
movement of the preys image elicits rotation of the head
as well as Xight steering and that the head rotation is
slightly faster than the wing response. A diagram of the
neural connections underlying interception (Fig. 7) must
therefore include two pathways, one controlling head
position and one eliciting Xight turns. While we have
much evidence showing that the TSDNs constitute the
major pathway for target-directed Xight steering (Olberg
1986; Frye and Olberg 1995), we have not identiWed the
neurons that control target-directed head rotation. One
possible scheme would have the TSDNs controlling both
head and wing movements. However, for the majority of
the TSDNs we think this is unlikely for the following
Target
image
moves
?? TSDNs
Head Wings steer
rotates to to stabilize
stabilize image
image SMDs?
Fig. 7 Diagram for the control of prey interception, as described in
text. SMDs refer to self movement detectors, a group of multimodal,
rotation-sensitive neurons known to respond to head position
reason. In order to stabilize the prey image on a speciWc
region of the eye, neuronal control of head movements
must include precise position information (Land and Col-
lett 1974). Most of the TSDNs show distinct directional
sensitivity to target movement, but their responses do not
vary systematically for target movements at diVerent
locations within their receptive Welds, making them
unlikely candidates for precisely Wxating the prey. Two
of the TSDNs, however, could reasonably have a role in
rotating the head to keep the target on the foveal midline.
These neurons (DIT1 and MDT4-Olberg 1986; Frye and
Olberg 1995) have small receptive Welds with their cen-
ters neatly centered on the dorsal foveal midline (Frye
and Olberg 1995). Whether stimulation of these two neu-
rons produces head rotation is still to be determined. It is
equally possible that centering of the prey image on the
dorsal foveal midline simply brings these neurons to bear
on the task of Xight control.
With both the head and wings responding to target-
image movement, the coordination between head and wings
is essential for orientation control. Thus a pathway must
exist which transmits information to the wings about head
rotation relative to the thorax (Fig. 7). A group of multi-
modal neurons that are directionally sensitive to head rota-
tion are known to provide input to the wings. In an earlier
study (Olberg 1981) we suggested that these neurons form
the pathway for rotational stabilization of the Xying drag-
onXy, and that one of their functions was to produce wing
torque tending to realign the thorax with the head. These
neurons, which were termed self movement detectors
(Olberg 1981), could, in principle, continue to provide
steering input to the wings while the image is well stabilized
on the fovea.
123
321

Eye specialization
Our results provide evidence for the role of the dorsal
foveal band in prey interception. It has been known for a
long time that the dorsal ommatidia of the dragonXy show a
remarkable degree of specialization, not only in their spec-
tral sensitivity (Labhart and Nilsson 1995), but also in their
large lens diameters and small inter-ommatidial angles
(Sherk 1978; Horridge 1978), properties that result in a
high acuity fovea (see Horridge 2005 for review). Various
authors (Sherk 1978; Sauseng et al. 2003) have suggested
that this region could be important in prey capture, since
dragonXies invariably attack their prey from below, placing
the prey on the upward-facing retina. Sherk (1978) also
suggested that the precise alignment of these facets and
their underlying rhabdomes could endow them with sensi-
tivity to the pattern of polarization in the sky above, which
could be useful in orientation and navigation. These sug-
gested functions are not mutually exclusive.
There are many examples of specialized eye regions
serving behavioral functions in insects. One of the best-
known examples is the male-speciWc acute zone of the Xy
eye. This region is located on visual midline and about 20
30 above the horizon (Land and Eckert 1985). Male
houseXies pursuing females keep the image of the female
on this acute zone at least 80% of the time (Wagner 1986).
Furthermore, male-speciWc neurons in the lobula of the
optic lobe are directionally selective for small target move-
ment, precisely the properties appropriate for controlling
tracking behavior (Gilbert and Strausfeld 1991).
Although there are some parallels between the tracking
Xight in male Xies mediated by the male-speciWc acute zone
and the prey pursuit Xight in dragonXies mediated by the
dorsal fovea, there are also fundamental diVerences.
Although tethered Xies rapidly rotate their heads in all
planes to follow moving gratings (Gilbert et al. 1995), dur-
ing tracking Xights the males body appears to remain
mostly in line with the head, which is directed towards its
quarry, another Xy. DragonXies show no tendency to align
their bodies towards their prey. The head is rotated to place
the preys image on the foveal midline, but the orientation
of the rest of the body is highly variable. The direction of
Xight relative to the body can as easily be backwards or
sideways as forward, but it is always in a direction that
leads the prey for interception (Olberg et al. 2000). In this
way the dragonXy pursuit is more similar to the interception
Xights of some hoverXies (Collett and Land 1978) than to
the tracking Xights of houseXies.
The dragonXys dorsal fovea is probably also used for
estimating the distance, and therefore the size of a potential
prey item (Olberg et al. 2005). Perched dragonXies typi-
cally bob their heads upward in response to an object pass-
ing above them (Kirmse and Lssig 1971; Miller 1995;
123
J Comp Physiol A
Sauseng et al. 2003). This movement usually results in the
objects image passing across the dorsal foveal band. The
rocking head motion includes a signiWcant translational
component (Olberg and Seaman, unpublished) that could
allow the dragonXy to extract distance information based
on motion parallax, a possibility discussed in detail else-
where (Olberg et al. 2005). The pre-Xight head movements
also include Wxation of the prey image on the foveal mid-
line, several examples of which can be seen in Fig. 4.
Response latencies
The higher spatial and temporal resolution in this study
allowed us to reWne our estimates of the latency between
detecting and responding to object image motion. In an
early study (Olberg et al. 2000), based exclusively on eval-
uating dragonXy and prey Xight tracks, we estimated this
latency to be between 33 and 50 ms, somewhat higher than
the average latency of 29 ms that we report in this study.
The remarkable speed of this visual reXex may account for
the impressive success rates of foraging dragonXies. Baird
and May (1997) measured a success rate of 76% in foraging
Pachydiplax longipennis, and we found an astonishing
combined capture rate of 97% in short Xights of Erythemis
simplicicollis and Leucorrhinia intacta (Olberg et al. 2000).
Perhaps not surprisingly, the visual reXexes of dragon-
Xies foraging in outdoor sunlight are much faster than
would be predicted from electrophysiological studies in the
laboratory. The shortest visual latency measured in
descending interneurons in the cervical connectives was
45 ms. The descending interneurons synapse onto thoracic
pre-motor or motor neurons, which must then excite mus-
cle, increasing appreciably the estimate for latency based
on lab measurements.
The role of prediction in prey interception
Whether the goal is to catch a ball, avoid a moving projectile,
or grab a Xying insect out of the air, it is important to be able
to predict where a moving object will be some time in the
future. This ability has been investigated in a variety of verte-
brate animals and especially in humans (see Regan et al. 1998
for review). The head movement plotted in the top trace of
Fig. 4a suggests that the dragonXy is able to predict the future
position of its prey. The Wrst head rotation in this trace shows
a simple tracking response to the preys movement with a
visual latency of about 26 ms. However, after the prey is again
Wxated in Xight (at about 60 ms) the image is held on midline
with no obvious response latency. This might suggest that the
head movement is now governed by the angular velocity,
rather than the position, of the moving image. Responding to
the velocity of the preys image would allow the dragonXy to
predict its future position (Adelman et al. 2003).
322
J Comp Physiol A
The predictive ability of the dragonXys visual response
is also seen in its choice of initial takeoV direction. The
Xight tracks in Fig. 2ac, show that the takeoV direction is
quite an accurate prediction of the line that the dragonXy
should take to intercept its prey. Like the prediction con-
trolling head orientation, determination of the best line for
interception is based on the angular velocity of the preys
image (Olberg and Henry, unpublished). The interplay
between target Wxation and prediction of future target posi-
tion underlies a far more sophisticated mechanism for prey
interception by dragonXies than we previously proposed
(Olberg et al. 2000).
Acknowledgments This study was supported in part by NSF RUI
0211467 grant to RMO, and by Union College summer research fel-
lowships to RCS and MIC. All experiments complied with the Princi-
ples of animal care, publication no. 8623, revised 1985, of the
National Institute of Health.
References
Adelman TL, Bialek W, Olberg RM (2003) The information content of
receptive Welds. Neuron 40:823-833-1500
Baird JM, May ML (1997) Foraging behavior of Pachydiplax longi-
pennis (Odonata: Libellulidae). J Insect Behavior 10: 655678
Collett TS, Land MF (1978) How hoverXies compute interception
courses. J Comp Physiol A 125: 191204
Frye MA, Olberg RM (1995) Visual receptive Weld properties of fea-
ture detecting neurons in the dragonXy. J Comp Physiol A 177:
569576
Gilbert C, Strausfeld NJ (1991) The functional organization of male-
speciWc visual neurons in Xies. J Comp Physiol 169:395411
Gilbert C, Gronenberg W, Strausfeld NJ (1995) Oculomotor control in
Calliphorid Xies: head movements during activation and inhibi-
tion of neck motor neurons corroborate neuroanatomical predic-
tions. J Comp Neurol 361:285297
Horridge GA (1978) The separation of visual axes in apposition com-
pound eyes. Phil Trans R Soc Lond B 285:159
Horridge A (2005) The spatial resolutions of the apposition compound
eye and its neuro-sensory feature detectors: observation versus
theory. J Insect Physiol 51:243266
Kirmse W, Lssig P (1971) Strukturanalogie zwischen dem system der
horizontalen blickbewegungen der augen beim menschen und
dem system der blickbewegungen des kopfes bei insekten mit Wx-
ationsreaktionen. Biol Zbl 90:175193
Labhart T, Nilsson D-E (1995) The dorsal eye of the dragonXy Sympe-
trum: specializations for prey detection against the blue sky. J
Comp Physiol A 176:437453
Land MF, Collett TS (1974) Chasing behavior of houseXies (Fannia
canicularis): a description and analysis. J Comp Physiol A 89:
331357
Land MF, Eckert H (1985) Maps of the acute zones of Xy eyes. J Comp
Physiol A 156:525538
Miller PL (1995) Visually controlled head movements in perched An-
isopteran dragonXies. Odonatologica 24:301310
Olberg RM (1981) Object- and self-movement detectors in the ventral
nerve cord of the dragonXy. J Comp Physiol A 141:327334
Olberg RM (1986) IdentiWed target-selective visual interneurons descend-
ing from the dragonXy brain. J Comp Physiol A 159:827840
Olberg RM, Worthington AH, Venator KR (2000) Prey pursuit and
interception in dragonXies. J Comp Physiol A 186:15562
Olberg RM, Worthington AH, Fox JL, Bessette CE, Loosemore MP
(2005) Prey size selection and distance estimation in foraging
adult dragonXies. J Comp Physiol A 191:791797
Regan D, Gray R, Portfors CV, Hamstra SJ, Vincent A, Hong XH,
Kohly R, Beverley K (1998) Catching, hitting, and collision
avoidance. In: Harris LR, Jenkin M (eds) Vision and action. Cam-
bridge University Press, Cambridge, pp 181214
Rowe RJ (1987) The dragonXies of New Zealand. Auckland Univ
Press, New Zealand
Sauseng M, Pabst M-A, Kral K (2003) The dragonXy Libellula quad-
rimaculata (Odonata: Libellulidae) makes optimal use of the dor-
sal fovea of the compound eyes during perching. Eur J Entomol
100: 475479
Sherk TE (1978) Development of the compound eyes of dragonXies
(Odonata) III. Adult compound eyes. J Exp Zool 203: 6180
van Hateren JH, Schilstra C (1999) BlowXy Xight and optic Xow II.
Head movements during Xight. J Exp Biol 202:149
Wagner H (1986) Flight performance and visual control of Xight of the
free-Xying houseXy (Musca domestica L.) II. Pursuit of targets.
Phil Trans R Soc Lond B 312:553579
123
323


324
Neuron, Vol. 17, 11451155, December, 1996, Copyright 1996 by Cell Press
Imaging the Functional Organization
of Zebrafish Hindbrain Segments
during Escape Behaviors
Donald M. OMalley, Yen-Hong Kao,
and Joseph R. Fetcho
Department of Neurobiology and Behavior
State University of New York at Stony Brook
Stony Brook, New York 11794-5230
Summary
Although vertebrate hindbrains are segmented struc-
tures, the functional significance of the segmentation
is unknown. In zebrafish, the hindbrain segments con-
tain serially repeated classes of individually identifi-
able neurons. We took advantage of the transparency
of larval zebrafish and used confocal calcium imaging
in the intact fish to study the activity of one set of
individually identified, serially homologous reticulospi-
nal cells (the Mauthner cell, MiD2cm, and MiD3cm)
during behavior. Behavioral studies predicted that dif-
ferential activity in this set of serially homologous neu-
rons might serve to control the directionality of the
escape behavior that fish use to avoid predators. We
found that the serially homologous cells are indeed
activated during escapes and that the combination
of cells activated depends upon the location of the
sensory stimulus used to elicit the escape. The pat-
terns of activation we observed were exactly those
predicted by behavioral studies. The data suggest that
duplication of ancestral hindbrain segments, and sub-
sequent functional diversification, resulted in sets of
related neurons whose activity patterns create behav-
ioral variability.
Introduction
Vertebrate brains are segmented structures (Lumsden
and Keynes, 1989; Rubenstein et al., 1994; Guthrie,
1995). The segmentation is most obvious in the hind-
brain where a series of repeated units can be defined
at all levels from the gross anatomical to the molecular
in both developing and adult brains (Metcalfe et al.,
1986; Fraser et al., 1990; Trevarrow et al., 1990; Clarke
and Lumsden, 1993). In animals as different as chicks
and fish, these segments contain repeated, serially ho-
mologous neurons (Trevarrow et al., 1990; Lee and Ea-
ton, 1991; Clarke and Lumsden, 1993). Although the
anatomical segmentation is evident even at the level of
single, identifiable neurons, the functional significance
of the segmentation is unknown. One obvious question
is whether similar neurons in successive segments act
together as a functional unit, or are functionally indepen-
dent. Because segments are thought to have arisen by
duplication of ancestral segments, subsequent diver-
gence of serially repeated neurons might constitute one
mechanism for the evolution of behavioral diversity.
Here, we examine the functional organization of one set
of serially repeated neurons in zebrafish larvae.
Zebrafish larvae are a particularly favorable prepara-
tion for study of hindbrain because many of their hind-
brain neurons are individually identifiable from fish to
fish (Kimmel et al., 1982; Metcalfe et al., 1986; Hanneman
et al., 1988). We have focused on one particular set of
hindbrain cells that includes the well-studied Mauthner
cell in hindbrain segment 4 and two other neurons,
MiD2cm and MiD3cm, in segments 5 and 6. All of these
cells are morphologically similar, each having two major
dendrites and an axon that crosses in the brain and
descends along the length of the contralateral spinal
cord. This set of homologs is of particular interest be-
cause one of the set, the Mauthner cell, is already known
to play an important role in the escape behaviors used
to avoid predators. There are, however, no physiological
data from the other cells to provide insight into the func-
tional organization of this serial set of neurons.
During escapes, a stimulus on one side of the fish
leads to a very fast and forceful C-shaped bend to the
opposite side (Foreman and Eaton, 1993). This is pro-
duced, at least in part, by the output of the Mauthner
cell that fires a single action potential during the escape,
thereby exciting motoneurons along the contralateral
side of spinal cord (Fetcho and Faber, 1988; Faber et
al., 1989). This is followed by a counter bend that propels
the fish through the water. The extent of the initial
C-bend and the subsequent direction of the escape vary
in a systematic and, apparently, adaptive manner (Eaton
and Emberley, 1991). Sensory stimuli from behind the
fish lead to a weak C-bend and an escape movement
directed forward, away from the location of the sensory
stimulus. Stimuli at the head lead to the strongest
C-bends, with the fish making up to a 1808 turn directed
away from the stimulus. Thus, the escape movement
varies in a way that assures that the fish moves away
from a potential threat.
Although there is strong evidence that the Mauthner
cell initiates the escape behavior, several observations
indicate that the Mauthner cell is not acting alone. The
variability of the escape with the location of the trig-
gering sensory stimulus suggests the involvement of
additional neurons, because the Mauthner cell only fires
a single spike and, by itself, produces a stereotyped
output (Nissanov et al., 1990). Also, lesion studies show
that after deletion of the Mauthner cell, goldfish can still
produce robust escapes, suggesting that other neurons
can mimic the function of the Mauthner cell (Eaton et
al., 1982). These observations led to the hypothesis that
there might be neurons other than the Mauthner cell
involved in the escape. Anatomical work revealed the
presence of MiD2cm and MiD3cm, and these are
thought, based upon their morphological similarities to
the Mauthner cell, to be the two most likely candidates
to act in concert with it. Foreman and Eaton (1993) have
proposed an explicit model of how the combined activity
in the Mauthner cell and its serial homologs might con-
tribute to the directional control of the escape behavior.
They propose that rostral stimuli might activate the en-
tire set of cells (Mauthner cell, MiD2cm, and MiD3cm)
leading to a very powerful C-bend and a large direction
change, whereas caudal stimuli would activate only the
Mauthner cell producing a smaller C-bend and the asso-
ciated weaker turn.
325
Neuron
1146

Thus, the prior work has generated two predictions

about the activity patterns in MiD2cm and MiD3cm. First,
if these are the neurons that can functionally substitute
for the Mauthner cell, then they should also be activated
during at least some escapes. Second, the Foreman and
Eaton model predicts that escapes elicited by rostral
stimuli should activate the entire set of homologs,
whereas caudal stimuli should activate just the Mauth-
ner cell. These predictions have not been tested previ-
ously because no one has been able to record the activ-
ity of MiD2cm and MiD3cm during escapes. We have
recently developed methods to label neurons with a
calcium indicator and observe their activity patterns dur-
ing behavior by using confocal calcium imaging in the
transparent, posthatching zebrafish larvae (Fetcho and
OMalley, 1995). Here, we show that this approach can
be used to label and monitor the activity of hindbrain
neurons in the live fish, which allowed us to evaluate
these predictions.
We asked first whether the Mauthner cell and its ho-
mologs were activated together during escapes, and
second, how their pattern of activation varied during
escapes elicited by sensory stimuli at different locations.
We examined both questions by directly observing
which members of this set of homologs showed fluores-
cence increases during escapes elicited by a pair of
piezoelectric tappers that were used to apply a sudden
light touch to either the head or the tail. We found that
the activity pattern is exactly that predicted by Foreman
and Eaton (1993). The observations are consistent with Figure 1. Identified Reticulospinal Neurons Observed in the Brain
of the Living Zebrafish
the idea that this set of similar neurons in successive
(A) Confocal projection of the array of reticulospinal neurons labeled
hindbrain segments forms a functional group and that
by a bilateral injection of calcium green dextran into caudal spinal
the activity pattern within the group contributes to vari- cord. Labeled neurons, shown in horizontal section, include the
ability in the form of the behavior (Morton and Chiel, bilateral Mauthner cells (arrows) as well as clusters of neurons in
1994). The organization into serial functional groups successive hindbrain segments.
spanning hindbrain segments may extend to other seri- (B) Confocal projection, from another fish, of the Mauthner cell and
ally repeated neurons present in the hindbrains of both its two homologs (MiD2cm and MiD3cm, left and right arrows, re-
spectively) labeled with calcium green dextran by a more selective
zebrafish and other vertebrates.
unilateral injection into the ventral portion of caudal spinal cord.
The axons and dendrites of the cells are visible and aid in their
identification when viewed either in optical sections or in three-
Results dimensional reconstructions.
Rostral is to the left in both panels. Scale bars represent 25 mm.
Imaging Identified Hindbrain Neurons In Vivo
Reticulospinal neurons were retrogradely labeled in lar-
val fish by injecting calcium green dextran into the cau- axons of the segmental homologs have similar trajecto-
dal spinal cord. Large injections labeled many reticulo- ries in caudal spinal cord.
spinal neurons bilaterally, as illustrated in Figure 1A,
which shows a projection from a stack of images ac-
quired from the hindbrain of a living fish. This array Calcium Responses in Mauthner Cells
contains several sets of serially homologous neurons. Because nerve cells have voltage-gated calcium chan-
We were specifically interested in the set that includes nels, we expected to observe fluorescence responses
the Mauthner cell and its segmental homologs (MiD2cm in calcium green dextranlabeled reticulospinal neurons
and MiD3cm). Small, unilateral injections of calcium when they fired action potentials during escapes. To
green dextran into ventral spinal cord more selectively confirm our ability to detect such increases, we initially
labeled this set on the contralateral side of the brain studied the Mauthner cell because of the compelling
(Figure 1B). The identity of the cells was confirmed in evidence that it fires in conjunction with escapes (Zottoli,
three-dimensional reconstructions of confocal data sets 1977; Eaton et al., 1982). Mauthner cells in unanesthe-
acquired from the living animal; the characteristic lateral tized fish (n 5 86) showed robust fluorescence increases
and ventral dendrites and decussating axons of these ranging from 12% to 110% in conjunction with escapes
cells were easily distinguished in these reconstructions. elicited by an abrupt, gentle tap to the ipsilateral side
The ability to label reproducibly this set by injections of the head (Figure 2A; escapes were monitored by visu-
restricted to ventral, postanal spinal cord indicates that alizing the large, rapid movement of the fish). These
not only is their dendritic morphology similar, but the fluorescence changes correspond to calcium increases

326
Functional Organization of Zebrafish Hindbrain
1147

Figure 2. Comparison of Calcium Green and Lucifer Yellow Dextran

(A) Response of a calcium green dextran labeled Mauthner cell during an escape elicited by a head tap. Panels show successive images (in
pseudocolor) taken at 400 ms intervals. The color scale on the left represents fluorescence intensity (blue, lowest; red, highest) for both (A)
and (B). The asterisk marks the frame in which the tap and escape occurred. A movement artifact from the escape is evident at the top of
this frame. Since these images are acquired line by line at 2 ms/line and the movement artifact lasts only about 30 lines (60 ms), this response
was the result of a brief movement, as occurs in escapes. The fluorescence of the cell increases quickly, but decays slowly over subsequent
frames. The size of the increase was about 50%, as illustrated in the inset plot of fluorescence versus time. The starting fluorescence level
was normalized to 100% in the plot.
(B) Fluorescence images of a Mauthner cell filled with the nonindicator, lucifer yellow dextran. An escape was elicited in the frame marked
by the asterisk. The inset shows that the fluorescence did not change. This control experiment demonstrates that movement artifact cannot
explain the fluorescence changes observed in cells labeled with calcium green dextran. The data point for the frame containing the escape
was omitted from the plot because movement of the cell during the escape precludes measurement of the fluorescence of the cell.
Scale bars represent 10 mm in (A) and (B).

327
Neuron
1148

ranging from 25 nM to 735 nM, assuming a resting cal-

cium of 100 nM. In experiments in which both Mauthner
cells were monitored, only the Mauthner cell ipsilateral
to the head stimulus responded to stimuli at or just
above threshold for escapes. Stimuli that were well
above threshold sometimes produced responses in
both Mauthner cells, possibly as a consequence of the
generation of multiple escapes.
In these experiments, we purposely did not paralyze
the fish so that we could visually monitor its behavior.
One consequence of this was a movement artifact dur-
ing the escape (e.g., frame marked by asterisk in Figure
2A). Although the agar returned the fish very close to
its initial position, there remained the concern that the
movement might induce some artifactual change in fluo-
rescence by changing, for example, the path length of
the emitted light. Since calcium responses have not
previously been imaged in single neurons in the brain
of an intact vertebrate, we were interested in both ruling
out movement related artifacts and quantitating the min-
imal fluorescence signal that could be reliably ascribed
to calcium dynamics. To this end, we backfilled Mauth-
ner cells with a calcium insensitive dye, lucifer yellow
dextran, and monitored the fluorescence of these cells
during escapes under conditions identical to those used
in the calcium imaging experiments. In these fish (n 5 Figure 3. Relationship between Action Potentials and Fluorescence
5), head taps again produced robust escape behaviors, Determined by Antidromic Stimulation of the Mauthner Cell
but only very minimal, if any, fluorescence changes, as (A) The fluorescence change in a Mauthner cell soma is plotted for a
illustrated in Figure 2B. Since lucifer yellow dextran does series of trials in which the axon of the Mauthner cell was stimulated
not respond to calcium increases, these changes pro- electrically in spinal cord 1, 2, 5, or 10 times to elicit a comparable
number of action potentials. Stimulus parameters were 10 mA cur-
vide a measure of the magnitude of movement related
rent strength, 0.2 ms pulse duration, and 50 ms interstimulus interval.
artifacts and indicate that after the escape attempt, the The stimulus strength was determined by gradually increasing the
fish returns quite closely to its original position, presum- strength to a level just sufficient to elicit a response. Given the
ably due to constraint by the agar. When fluorescence relatively large size of the Mauthner axon and the proximity of
increases did occur in lucifer experiments, they were the electrode to it, this is likely to result in a reasonably selective
typically less than 1% or 2%, and the largest ever ob- activation of the Mauthner axon. The increase in fluorescence is
roughly 10% for one stimulus and increases with additional stimuli,
served was 4%. These small increases might be due to
although the increase is not linearly related to the number of stimuli.
the focal plane not being precisely set at the maximally (B) shows a second set of stimuli for the same cell in (A) to illustrate
bright focal plane, or perhaps to some slight compres- the reproducibility of the relationship between stimuli and fluores-
sion of the fish during the escape. All of the increases cence. The baseline fluorescence is normalized to 100% to allow
seen with lucifer yellow dextran were considerably easy assessment of the percentage increase; however, the absolute
smaller than the calcium responses in zebrafish neurons baseline fluorescence was relatively stable, with little change from
trial to trial.
labeled with calcium green dextran in which the fluores-
cence increases to sensory stimulation were usually
much greater than 10%, well above the level that could we placed an extracellular metal stimulating electrode
be explained by shifts in the position of the fish. next to the Mauthner cell axon in caudal spinal cord and
stimulated it antidromically using an approach that is
standard in Mauthner cell studies. The Mauthner axon
Comparison of Antidromic and Sensory Responses is easily seen in bright field due to its large diameter,
The fluorescence increases observed in calcium indica- allowing the close apposition of the electrode to the
tor-loaded Mauthner cells during escapes support a link axon. Antidromic stimulation of the cell (in an anesthe-
between an action potential in the Mauthner cell and tized fish) produced small fluorescence increases (12%
a rise in intracellular calcium. However, we wanted to 19%) in the Mauthner cell soma in response to a single
confirm that we could detect the fluorescence increase stimulus, and larger increases in response to a rapid
associated with a single spike and to explore the rela- train of 2, 5, or 10 stimuli, as shown in Figure 3A. Figure
tionship between spikes and the calcium rise. The ideal 3B shows that the responses were similar when the
approach would be to record intracellularly from the cell same set of stimuli was repeated. The size of the calcium
and elicit one spike or a series of spikes. Unfortunately, response to a single antidromic spike, as well as the
all of the neurons in the larvae are small, with the 1214 accumulation and extrusion of calcium after trains of
mM diameter Mauthner cell being among the largest. action potentials, were similar to that observed pre-
While this small size is favorable for imaging (the larger viously in zebrafish motoneurons (Fetcho and OMalley,
surface to volume ratio results in larger calcium signals), 1995) and in cultured bullfrog sympathetic neurons
it makes intracellular recording more difficult. Instead, (OMalley, 1994).

328
Functional Organization of Zebrafish Hindbrain
1149
The small increase due to a single antidromic spike
was comparable to the smallest increases observed in
the Mauthner cell during escapes elicited by a tap. The
small tap-evoked increases occurred in early trials in a
series of escapes elicited at 2 min intervals. However,
the Mauthner cell responses were often larger, ranging
from 30% to 70% in most experiments. This was some-
what surprising because the evidence, including data
from larval and adult fish, indicates that the Mauthner
cell in teleost fishes fires only a single action potential
each time it triggers an escape behavior (Zottoli, 1977;
Featherstone et al., 1991). However, our data do not al-
low us to rule out the possibility that the larger calcium
responses are a consequence of multiple action poten-
tials.
Although the increase in the size of the responses in
the Mauthner cell with repetitive escapes is an interest-
ing phenomenon that may reflect plasticity in the size
of the calcium change evoked by a single spike or an
increase in the number of spikes fired (OMalley and
Fetcho, 1996, Soc. Neurosci., abstract), we were mainly
concerned with whether or not the cell was responding
during different sensory stimuli. Thus, the large size of
the responses typically seen was helpful in determining
whether or not the cell had been activated. There was
no evidence that these larger calcium responses were
a consequence of damage because they recovered rap-
idly and completely and could be observed reliably in
conjunction with escapes for up to 132 trials spread
over as long as 2 days. Thus, the cells remained healthy
and could be monitored physiologically for an extended
time, as long as the illumination was minimized.
Calcium Responses at Threshold
The experiments with antidromic stimulation indicate
that the smallest fluorescence increases in the Mauthner
cell during escapes could be explained by the calcium
influx associated with a single action potential. They do
not, however, rule out the possibility that some of the
fluorescence responses are a result of subthreshold
postsynaptic potentials. To determine whether we could
detect fluorescence increases during large, subthresh-
old input to the Mauthner cell, we varied the voltage
applied to the piezoelectric crystal driving the head tap-
per and examined the fluorescence changes in the
Mauthner cell at stimulus strengths near threshold for
eliciting an escape (i.e., a strength where the Mauthner
fires only 50% of the time). Figure 4 shows the responses
of a Mauthner cell to a series of stimuli of different
strength, including several near the threshold for es-
capes. The calcium response only occurred when an
escape was produced. In cases where the fish showed
no evidence of producing the escape, no calcium re-
sponse resulted. This included trials in which the stimu-
lus strength was the same for escape and nonescape
trials. Thus, in the absence of escapes, there is little if
any fluorescence change in the soma, even when one
would expect there to be a large postsynaptic potential
in the cell. This strengthens the link of the somatic cal-
cium responses to both the behavioral response and to
the firing of action potentials by the Mauthner cell. It
also supports other evidence that the calcium increases
in the soma seen with this approach are associated with
the firing of action potentials rather than subthreshold
postsynaptic potentials (ODonovan et al., 1993; McClel-
lan et al., 1994; Fetcho and OMalley, 1995).
Activation Pattern of Segmental Homologs
After using the well-studied Mauthner cell to establish
the reliability of the technique as an indicator of neuronal
firing, we went on to study the responses of MiD2cm
and MiD3cm, for which there were no prior physiological
Figure 4. Responses of the Mauthner Cell
near Threshold for Escapes
Plots of the fluorescence of the soma of a
calcium green dextran-labeled Mauthner cell
from a series of successive trials at 2 min
intervals in which the strength of a head tap
was varied around the threshold for escapes.
The voltage applied to the piezoelectric head
tapper in each trial is indicated, with larger
voltages associated with stronger taps. The
asterisks mark those trials in which an escape
occurred. Fluorescence increases occurred
only in conjunction with escapes, even at
stimulus strengths near threshold for es-
capes, as defined by 50% response probabil-
ity (about 13 V in this experiment). The fluores-
cence responses (30%40% increases) were
similar in all escapes independent of whether
near-threshold or suprathreshold stimuli
were used. Successive points are at 400 ms
intervals.
329
Neuron
1150

Figure 5. Differential Activation of the Mauthner Cell and Its Homologs during Escapes
Each row shows a trial consisting of a sequence of images during which an escape was elicited by an ipsilateral touch to the head or tail.
The color scale (applies to all figures) represents fluorescence intensity (blue, lowest; red, highest).
(A) Response of MiD3cm during an escape elicited by a head tap. MiD3cm is shown at relatively high magnification to illustrate clearly the
fluorescence increase. The tap was applied during the third frame. The escape event transiently causes the cell to move out of the frame,
which provides a convenient record of the behavioral event. This movement frame is omitted from subsequent rows.
(B) Simultaneous imaging of both the Mauthner cell (left) and MiD3cm (far right) during a head stimulus. Both cells respond.
(C) Same field as in (B), but with a tail stimulus. The Mauthner cell responds but not MiD3cm.
(D) Simultaneous imaging of the Mauthner cell (left) and MiD2cm (right) during a head stimulus. Both cells respond.
(E) Same field as in (D), but with a tail stimulus. Only the Mauthner cell responds. Arrows mark the first frame after the escape in (B)(E).
Images were acquired at 400 ms intervals.
Scale bars: 10 mm in (A), 15 mm in (B)(E). The size of the fluorescence increases in each trial were as follows: 41% (A); 62% in Mauthner
cell, 21% in MiD3cm (B); 53% in Mauthner cell, 2% in MiD3cm (not significant, see Figure 2) (C); 21% in Mauthner cell, 22% in MiD2cm (D);
67% in Mauthner cell, none in MiD2cm (E).

data. We compared the responses of the Mauthner cell
and its homologs during escapes elicited by one of the
two tappers positioned to stimulate either the head or
the tail. The Mauthner cell and its serial homologs
(MiD2cm and MiD3cm) all responded to a touch on the
ipsilateral side of the head at a strength above threshold
for eliciting an escape (Figures 5A, 5B, and 5D). These
responses consisted of fluorescence increases ranging
from 9% to 41% in the homologs and 21% to 110%
in the Mauthner cells. The fluorescence changes far
outlasted the movement artifact; they returned nearly
to baseline in 46 s, which is typical for calcium signals
in neuronal somata (Regehr et al., 1989; Lev-Ram et al.,
1992; Yuste et al., 1994). In marked contrast with the
response of this group during head taps, only the
Mauthner cell responded to ipsilateral tail stimuli even
when the stimuli were well above the threshold for es-
capes (Figures 5C and 5E).
If the homologs are indeed playing a role in determin-
ing the magnitude of the escape response, then their

330
Functional Organization of Zebrafish Hindbrain
1151
capes elicited by either head or tail taps. MiD2cm (middle) and MiD3cm (bottom) cells both respond during escapes elicited by head taps,
but not during escapes produced by tail taps.
trial-to-trial behavior should consistently show a differ-
ential response to head and tail stimulation. Addressing
this issue requires not only that the in vivo recordings
be relatively stable, but that successive trials be directly
observed without signal averaging. High efficiency op-
tics and the large fluorescence signals of calcium indica-
tors were important for this because they allow the use
of minimal excitation light, providing stable recording
conditions. This allowed the recording of neuronal re-
sponses over a series of successive trials in which head
and tail stimuli were alternately applied to elicit escapes
(Figure 6). Throughout these trials, head stimulation acti-
vated all of the cells (Mauthner cell, MiD2cm, and
MiD3cm), whereas tail stimulation activated only the
Mauthner cell; such differences were observed in 17
fish, with no exceptions to the differential response of
MiD2cm and MiD3cm to head and tail stimuli that elicited
escapes.
Timing and Threshold of Homolog Responses
These results suggest that the homologs of the Mauth-
ner cell are involved in the production of the more vigor-
ous escapes produced by stimulation of the head. Be-
cause the escape behavior is a very fast event (the initial
C-bend is completed 2025 ms after the stimulus onset;
Eaton and Farley, 1975), the segmental homologs must
be activated close in time to the Mauthner cell if they
are to contribute to the initial escape bend. To examine
the time course of activation, the laser scanning micro-
scope was used in a one-dimensional imaging mode,
i.e., a single line across these cells was repetitively
scanned. This provided a much faster acquisition rate (2
ms per line) than the 400 ms needed for two-dimensional
images. These experiments showed that MiD3cm and
the Mauthner cell are activated within 30 ms of one
another (Figure 7A). While the movement artifact pre-
vented a more exact determination of the response la-
tency, both cells were synchronously activated over a
Figure 6. Quantification of Escape Re-
sponses of the Mauthner Cell and Its Homo-
logs over Successive Trials
Each row shows a series of successive trials
in which we alternately elicited escapes by
either head or tail stimuli. Each point repre-
sents the mean pixel intensity per frame in a
rectangular box placed over the cell (400 ms
between frames). The escape was elicited
during the sixth frame in each trial. In some
trials, the cell moved out of the field during the
frame in which the escape occurred. Because
the fluorescence of the cell could not be mea-
sured when it was not in the frame, there is
no data point for those movement frames,
leading to a discontinuity in the plot. A 2 min
rest was allowed between each trial. Each
row is from a different fish. The mean pixel
intensity was normalized so that the starting
fluorescence baseline is 100 in each trial. This
normalization corrects for any slow changes
in the baseline brightness of the cells (due to
bleaching for example) and allows an easy
comparison of successive trials. (Top) The
Mauthner cell responds repeatedly during es-
series of trials, within the limits of the detection system
(Figure 7B).
If MiD2cm and MiD3cm contribute to the escape along
with the Mauthner cell, then their threshold should be
similar to that of the Mauthner cell, which is known to
fire an action potential only in conjunction with escapes
(Zottoli, 1977). The relative thresholds for activating
these cells were determined by varying the strength of
the stimulus applied to the head or tail. When the head
stimulus was near or at threshold (50% probability of
response) for an escape, the Mauthner cell and its homo-
logs responded together whenever an escape occurred
(e.g., Figure 7). The Mauthner cell showed a very clear
all or none response with large fluorescence increases
during escapes and none in response to threshold stim-
uli that did not lead to an escape. In some animals, as
in Figure 7, the homologs of the Mauthner cell also had
a clear threshold, responding only in conjunction with
escapes. In other fish, the homologs showed a weak
response to head stimuli that did not elicit an escape,
but even in these cases, they produced a larger increase
in fluorescence in conjunction with escape events.
MiD2cm and MiD3cm did not respond to tail stimuli at
any strength, even when those stimuli were three times
threshold for eliciting an escape.
Discussion
A role for the Mauthner cell homologs in escapes would
help to explain both the normal adaptive variability of
escapes as well as the production of escapes in fish
without Mauthner cells, thus unifying a series of obser-
vations made over the past 15 years. Until now, it had
not been possible to obtain functional data from the
identified Mauthner cell homologs (Metcalfe et al., 1986;
Foreman and Eaton, 1993). This is a consequence of
the smaller size of MiD2cm and MiD3cm, which makes
it difficult both to find them and to record their electrical
331
Neuron
1152

Figure 7. Responses of the Mauthner Cell and MiD3cm near the Threshold for Escapes
(A) The fluorescence of a line (black line in top panel) that passed through both the Mauthner cell (left) and MiD3cm (right) was monitored
repeatedly as shown in the bottom panel. Consecutive lines were acquired at 2 ms intervals (vertical scale bar, 200 ms) and are displayed
from top to bottom. An escape elicited by a tap to the ipsilateral side of the head is indicated by the arrows. The trace was briefly disrupted
by the escape movement, but following the movement, the fluorescence of the vertical bands representing the two cells had increased,
indicating that both cells had responded rapidly, within 30 ms of one another.
(B) Quantification of responses from the cells in (A) over several trials in which the head was tapped at a strength near threshold for escape
(escapes occurred in 67% of the trials). The points show the average intensity of successive sets of 15 lines (i.e. 30 ms per data point) plotted
over time. Successive trials, 2 min apart are separated by dashed lines. The starting baseline in each trial was normalized to 100 as in Figure
3. Repeated stimuli at a constant stimulus strength near threshold led to escapes in trials 2, 4, 5 and 6, but not in 1 and 3. When an escape
occurred, both the Mauthner cell and MiD3cm responded, but neither responded when there was no escape.
Horizontal scale bar, 20 mm. Size of increases in (A): 70% in Mauthner cell; 34% in MiD3cm.

activity. Our data indicate that the Mauthner cell and its
serial homologs (MiD2cm and MiD3cm) in the hindbrain
do indeed function together during escapes. They re-
spond together during the large escape turns produced
by lateral stimulation on the head, but in response to
stimulation of the tail, only the Mauthner cell is activated.
This is exactly the pattern of activation predicted by
behavioral studies. It suggests a population code, with
the extent of the bend related to the extent of activation
of a population of serially homologous hindbrain cells
(Eaton et al., 1991). We examined only the extremes
of the escape behavior by using stimuli that elicit the
weakest and strongest escapes. Intermediate forms of
the escape might be produced not only by variations in
which cells are activated, but also by changes in the
degree of activation of individual cells.
The large size of the somatic calcium signals seen
during sensory activation of MiD2cm and MiD3cm sug-
gests that these cells are firing one or more action poten-
tials. These responses are unlikely to result from sub-
threshold excitation. In the case of the Mauthner cell,
there is usually no detectable calcium increase from
stimuli that are just below threshold (Figures 4 and 7),
even though such stimuli lead to a large synaptic input
(Faber et al., 1991). This is not unexpected, as action
potentials should produce considerably larger calcium
signals than subthreshold inputs because action poten-
tials will open high threshold calcium channels and
NMDA-gated channels (if the NMDA receptors and glu-
tamate are present). The somatic calcium signals ob-
served in MiD2cm and MiD3cm after sensory stimulation
(Figures 6 and 7) are typically larger than those produced
either by single antidromic spikes in the Mauthner cell
(Figure 3) or after direct firing of single action potentials
in cultured bullfrog neurons (OMalley, 1994). In addition,
we are aware of no imaging studies reporting detectable
somatic calcium signals due to subthreshold EPSPs.
All of these observations support the conclusion that
MiD2cm and MiD3cm are firing in response to head
stimuli that elicit escapes.
The differential activation we observed among the
Mauthner cell, MiD2cm, and MiD3cm most likely arises
from differences in the sensory input to this set of hind-
brain neurons. Based upon previous work, the most
likely possibility is that rostral stimulation leads to acti-
vation of the auditory system, whereas caudal stimula-
tion activates cutaneous and/or lateral line inputs (Eaton
et al., 1984). All of these are known to excite the Mauth-
ner cell in goldfish, consistent with its responses to both
head and tail stimulation (Faber et al., 1989). Previously,
there was no information about the inputs to MiD2cm
and MiD3cm. Our work shows that they respond only
to head stimulation, suggesting that they receive little
or no input from the caudal cutaneous or lateral line
system. The similar thresholds and conjoint activation
of MiD2cm and MiD3cm with the Mauthner cell in re-
sponse to head stimuli are consistent with their sharing
common rostral sensory inputs, most likely from the
auditory system. Such differences in sensory input to
escape neurons is not unprecedented, as they are well
documented in crayfish. The crayfish system has obvi-
ous parallels with zebrafish in that rostral and caudal
stimuli also elicit different forms of escape by differential
activation of giant neurons (Wine and Krasne, 1972).
Our observations that MiD2cm and MiD3cm are co-
activated with the Mauthner cell during escapes and

332
Functional Organization of Zebrafish Hindbrain
1153
that they share a similar axonal trajectory are consistent
with their playing a role in the generation of escapes.
Although this evidence confirms previous predictions,
it is still correlative, and nothing is known about the
output connections of MiD2cm and MiD3cm. A more
causal relationship could be established by lesioning
the cells in the live fish and observing the effects on
performance. Our data lead to the prediction that le-
sioning MiD2cm and MiD3cm should reduce the perfor-
mance of escapes elicited by head taps, but not those
from tail taps. The expectation is that such lesions
should make the C-bend to a head stimulus more like
that to a tail stimulus. We have preliminary data that
neurons in intact zebrafish can be deleted by using opti-
cal methods, so it should eventually be possible to eval-
uate this prediction.
The escape system is particularly attractive for studies
of how activity in neuronal populations determines ver-
tebrate behaviors because the behavior is produced by
a population of neurons, but the number of neurons and
synapses in the circuit is manageably small (Faber et
al., 1989). The ability to image the activity of any of
the array of hindbrain neurons and to delete them and
monitor behavioral changes offers the prospect of a
more complete understanding of how a population of
neurons in a vertebrate determines the form of a behav-
ior. These approaches should prove useful for future
studies of neural circuits not only in the normal animal,
but also in the many behavioral mutants that have been
generated by large scale mutagenesis of zebrafish (Mul-
lins and Nusslein-Volhard, 1993; Driever et al., 1995).
Our work has several implications concerning the
functional organization of hindbrain segments. Hind-
brain segments are thought to have arisen from the
duplication of an ancestral segment, with subsequent
evolutionary divergence of the segments (Metcalfe et
al., 1986). This would lead to the similar structural organi-
zation in successive segments observed in vertebrates.
Our observations indicate that another consequence of
this duplication is the production of a series of function-
ally related neurons in successive segments. The serially
homologous neurons studied retain a similar functional
role in escapes, but they are not complete functional
clones of one another. Instead, there is variability in the
sensory inputs that drive the cells and this is associated
with variability in the resulting escape behavior. The
duplication of segments may have permitted an expan-
sion of the behavioral repertoire of the animal by produc-
ing segmental groups of neurons involved in particular
behaviors. A subsequent evolutionary divergence of the
inputs and outputs of these cells would allow for behav-
ioral diversification, much as gene duplication and diver-
gence has led to diversification at the molecular level.
This organization of the hindbrain into serial sets of
functionally related neurons is likely to be a very general
one because the hindbrain has changed relatively little
during the evolution of vertebrates, as illustrated by the
very similar organization of escape circuits in fish and
startle circuits in mammals (Cruce and Newman, 1984;
Lingenhohl and Friauf, 1994; Butler and Hodos, 1996).
Experimental Procedures
A 50% solution of calcium green dextran (10,000 MW) in 10% Hanks
solution was pressure injected via a glass microelectrode into the
caudal (postanal) spinal cord of posthatching larval zebrafish (Danio
rerio, usually within the first 2 weeks after hatching) that were anes-
thetized with 0.02% 3-aminobenzoic acid ethyl ester (Fetcho and
OMalley, 1995). In most cases, this injection was into ventral cord
to selectively label the Mauthner cell and its homologs without dis-
rupting more dorsal sensory pathways. After injection, the fish were
allowed to recover and were maintained in 10% Hanks solution. We
only studied fish that exhibited no obvious disruptions of swimming
or escape behaviors following the injection. Twelve or more hours
later, the fish were briefly anesthetized, embedded on their backs
in soft agar on a cover glass in a petri dish (Eaton et al., 1984) and
then rinsed with 10% Hanks solution to allow recovery from the
anesthetic. Confocal images were obtained by looking into the head
of the intact fish using a Zeiss IM35 inverted microscope with a
503 Leitz 1.0 NA objective and a Biorad MRC 600 laser-scanning
confocal imaging system (Hernandez-Cruz et al., 1990; OMalley,
1994; Fetcho and OMalley, 1995). The transparency of the fish al-
lowed us to not only see neurons inside the living animal but also
allowed easy monitoring of the viability of the fish by observing the
heart beat and blood flow. To confirm the identity of the cells studied
physiologically, stacks of images showing the morphology in suc-
cessive confocal sections were acquired. Signal averaging, usually
seven frames, was used when acquiring this morphological data.
Maximum projections were made from stacks of these sections. The
image stacks were also reconstructed in three-dimensions using the
VolVis program (Sobierajski et al., 1995), allowing us to examine the
details of the dendritic morphology and axonal projections of each
cell.
Escapes were elicited by an abrupt touch produced by the dis-
placement of a small glass probe attached to a piezoelectric crystal.
The voltage applied to the crystal could be varied to change the
excursion of the glass probe. Two probes were used: one to stimu-
late the head and the other the tail. All stimuli were ipsilateral to
the reticulospinal neurons, in the region of the otolith for the head
stimulus and postanally for the tail stimulus. The tail stimulus was
located rostral to the calcium green injection site so as to activate
sensory neurons whose pathways to the brain were presumably not
damaged by the injection. The responses of the Mauthner cell to
tail stimuli indicated that ascending sensory pathways were indeed
intact. The stimuli produced an escape movement during which the
cell(s) moved briefly out of the plane of section, but then returned
rapidly because the agar controlled the resting position of the fish.
The escape attempt could also be observed visually (with protective
goggles to block the laser light) through a dissecting scope mounted
above the inverted microscope. Free swimming larval zebrafish gen-
erate a variety of movements subsequent to the escape (Eaton and
Farley, 1975). Although the initial escape event is very obvious in
the agar, we could not distinguish subtle differences in movement
following escapes. However, the responses we observed were
clearly linked to the initial escape movement and not subsequent
behavior, because they occurred whether or not the fish continued
to move after the initial escape event.
The fluorescence intensity of neurons during escapes was moni-
tored either by collecting a sequence of images of a cell or group
of cells (usually at 400 ms intervals) or by repeatedly scanning a
single line through the cells at 2 ms intervals. To assure that an
increase in the brightness of the cell was not the result of movement
to a brighter plane, we collected a series of optical sections spanning
the cell(s) (this was done for each block of trials). We then focused
at the brightest focal plane immediately prior to each trial. In experi-
ments with multiple cells in the field, we picked a focal plane where
the cells of interest were at or near their maximal fluorescence. A
cell was deemed to have responded only if its fluorescence response
exceeded its highest resting value at any plane in the reference
image stack. For physiological experiments, the confocal aperture
was usually set fully open to maximize light collection. This yielded
an optical section of 45 mm. These dynamic calcium images were
acquired without signal averaging or other processing. For illustra-
tion purposes, the images are linearly contrast enhanced and
smoothed with a 3 3 3 or 5 3 5 low pass filter to reduce pixel to
pixel noise. All quantitation, however, was performed on the raw
data. The percent fluorescence increases associated with each set
of images are given in the figure captions.
The fluorescence responses can be converted to absolute levels
333
Neuron
1154
of free calcium by knowing the resting calcium level, the K D of the
indicator, and the dynamic range. The latter two numbers were
unknown for calcium green dextran in living neurons. We determined
them by using cultured bullfrog sympathetic neurons whose acces-
sibility and large size relative to zebrafish neurons allowed the stable
whole cell patch recordings needed for such determinations. The
cultured cells were patch clamped and filled with free calcium set
at fixed levels, using 10 mM BAPTA. The cells were then flooded
with calcium via repetitive voltage pulses under voltage clamp and
the maximal fluorescence increase that could be obtained at the
differing resting calcium levels determined. This allowed the intra-
cellular K D of calcium green dextran to be determined. The average
value obtained, 250 nM, was quite close to the in vitro value (252
nM; Molecular Probes, Eugene, Oregon). The maximum dynamic
range obtainable (upon raising calcium from subnanomolar levels
to indicator-saturating levels) was only 8.7-fold, considerably less
than the range obtainable in intracellular solution in cuvettes (15-
fold). Based on these numbers, and by assuming a resting calcium
level (e.g., 100 nM), the fluorescence responses may be equated to
changes in free calcium. Since the actual resting calcium levels are
not known, the calcium responses in the paper are expressed in
terms of relative fluorescence increases. An estimate of the sizes
of the increases based upon the calibration data is provided in the
Results.
Acknowledgments
Correspondence should be addressed to J. R. F. We thank B. J.
Burbach for technical assistance and P. R. Adams for advice during
the course of the study. Supported by National Institute of Neurolog-
ical Disorders and Stroke grants NS26539 (to J. R. F.) and NS09113
(to D. M. O.), The Sloan Foundation (to J. R. F.), and the Howard
Hughes Medical Institute.
The costs of publication of this article were defrayed in part by
the payment of page charges. This article must therefore be hereby
marked advertisement in accordance with 18 USC Section 1734
solely to indicate this fact.
Received October 1, 1996; revised October 30, 1996.
References
Butler, A.B., and Hodos, W. (1996). Comparative Vertebrate Neuro-
anatomy: Evolution and Adaptation (New York: Wiley-Liss).
Clarke, J.D.W., and Lumsden, A. (1993). Segmental repetition of
neuronal phenotype sets in the chick embryo hindbrain. Develop-
ment 118, 151162.
Cruce, W.L.R., and Newman, D.B. (1984). Evolution of motor sys-
tems: the reticulospinal pathways. Am. Zool. 24, 733753.
Driever, W., Stemple, D., Schier, A., and Solnica-Krezel, L. (1995).
Zebrafish: genetic tools for studying vertebrate development.
Trends Genet. 10, 152159.
Eaton, R.C., and Emberley, D.S. (1991). How stimulus direction de-
termines the trajectory of the Mauthner initiated escape response.
J. Exp. Biol. 161, 469487.
Eaton, R.C., and Farley, R.D. (1975). Mauthner neuron field potential
in newly hatched larvae of the zebra fish. J. Neurophysiol. 38,
502512.
Eaton, R.C., Lavender, W.A., and Wieland, C.M. (1982). Alternative
neural pathways initiate fast start responses following lesions of the
Mauthner neuron in goldfish. J. Comp. Physiol. 145, 485496.
Eaton, R.C., Nissanov, J., and Wieland, C.M. (1984). Differential acti-
vation of Mauthner and non-Mauthner startle circuits in the zebra-
fish: implications for functional substitution. J. Comp. Physiol. 155,
813820.
Eaton, R.C., DiDomenico, R., and Nissanov, J. (1991). Role of the
Mauthner cell in sensorimotor integration by the brain stem escape
network. Brain Behav. Evol. 37, 272285.
Faber, D.S., Fetcho, J.R., and Korn, H. (1989). Neuronal networks
underlying the escape response in goldfish: general implications for
motor control. Ann. NY Acad. Sci. 563, 1133.
Faber, D.S., Korn, H., and Lin, J.-W. (1991). Role of medullary net-
works and postsynaptic membrane properties in regulating
Mauthner cell responsiveness to sensory excitation. Brain Behav.
Evol. 37, 286297.
Featherstone, D., Drewes, C.D., and Coats, J.R. (1991). Noninvasive
detection of electrical events during the startle response in larval
Medaka. J. Exp. Biol. 158, 583589.
Fetcho, J.R., and Faber, D.S. (1988). Identification of motoneurons
and interneurons in the spinal network for escapes initiated by the
Mauthner cell in goldfish. J. Neurosci. 8, 41924213.
Fetcho, J.R., and OMalley, D.M. (1995). Visualization of active neural
circuitry in the spinal cord of intact zebrafish. J. Neurophysiol. 73,
399406.
Foreman, M.B., and Eaton, R.C. (1993). The direction change con-
cept for reticulospinal control of goldfish escape. J. Neurosci. 13,
41014113.
Fraser, S., Keynes, R., and Lumsden, A. (1990). Segmentation in the
chick embryo hindbrain is defined by cell lineage restrictions. Nature
344, 431435.
Guthrie, S. (1995). The status of the neural segment. Trends Neu-
rosci. 18, 7479.
Hanneman, E., Trevarrow, B., Metcalfe, W.K., Kimmel, C.B., and
Westerfield, M. (1988). Segmental pattern of development of the
hindbrain and spinal cord of the zebrafish embryo. Development
103, 4958.
Hernandez-Cruz, A., Sala, F., and Adams, P.R. (1990). Subcellular
calcium transients visualized by confocal microscopy in a voltage
clamped vertebrate neuron. Science 247, 858862.
Kimmel, C.B., Powell, S.L., and Metcalfe, W.K. (1982). Brain neurons
which project to the spinal cord in young larvae of the zebrafish. J.
Comp. Neurol. 205, 112127.
Lee, R.K.K., and Eaton, R.C. (1991). Identifiable reticulospinal neu-
rons of the adult zebrafish, Brachydanio rerio. J. Comp. Neurol. 304,
3452.
Lev-Ram, V., Miyakawa, H., Lasser-Ross, N., and Ross, W.N. (1992).
Calcium transients in cerebellar Purkinje neurons evoked by intra-
cellular stimulation. J. Neurophysiol. 68, 11671177.
Lingenhohl, K., and Friauf, E. (1994). Giant neurons in the rat reticular
formation: a sensorimotor interface in the elementary acoustic star-
tle circuit. J. Neurosci. 14, 11761194.
Lumsden, A., and Keynes, R. (1989). Segmental patterns of neuronal
development in the chick hindbrain. Nature 337, 424428.
McClellan, A.D., McPherson, D., and ODonovan, M.J. (1994). Com-
bined retrograde labeling and calcium imaging in spinal cord and
brainstem neurons of the lamprey. Brain Res. 663, 6168.
Metcalfe, W.K., Mendelson, B., and Kimmel, C.B. (1986). Segmental
homologies among reticulospinal neurons in the hindbrain of the
zebrafish larva. J. Comp. Neurol. 251, 147159.
Morton, D.W., and Chiel, H.J. (1994). Neural architectures for adap-
tive behavior. Trends Neurosci. 17, 413420.
Mullins, M.C., and Nusslein-Volhard, C. (1993). Mutational ap-
proaches to studying embryonic pattern formation in the zebrafish.
Curr. Opin. Genet. Dev. 3, 648654.
Nissanov, J., Eaton, R.C., and DiDomenico, R. (1990). The motor
output of the Mauthner cell, a reticulospinal command neuron. Brain
Res. 517, 8898.
ODonovan, M.J., Ho, S., Sholomenko, G., and Yee, W. (1993). Real-
time imaging of neurons retrogradely and anterogradely labelled
with calcium sensitive dyes. J. Neurosci. Meth. 46, 91106.
OMalley, D.M. (1994). Calcium permeability of the neuronal nuclear
envelope: evaluation using confocal volumes and intracellular perfu-
sion. J. Neurosci. 14, 57415758.
Regehr, W.G., Connor, J.A., and Tank, D.W. (1989). Optical imaging
of calcium accumulation in hippocampal pyramidal cells during syn-
aptic activation. Nature 341, 533536.
Rubenstein, J.L., Martinez, S., Shimamura, K., and Puelles, L. (1994).
The embryonic vertebrate forebrain: the prosomeric model. Science
266, 578580.
334
Functional Organization of Zebrafish Hindbrain
1155

Sobierajski, L.M., Avila, R.S., OMalley, D.M., Wang, S., and Kauf-
mann, A.E. (1995). Visualization of calcium activity in nerve cells.
IEEE Comp. Graphics Appl. 15, 5561.
Trevarrow, B., Marks, D.L., and Kimmel, C.B. (1990). Organization
of hindbrain segments in the zebrafish embryo. Neuron 4, 669679.
Wine, J.J., and Krasne, F.B. (1972). The organization of escape
behavior in the crayfish. J. Exp. Biol. 56, 118.
Yuste, R., Gutnick, M.J., Saar, D., Delaney, K.R., and Tank, D.W.
(1994). Ca2 1 accumulations in dendrites of neocortical pyramidal
neurons: an apical band and evidence for two functional compart-
ments. Neuron 13, 2343.
Zottoli, S.J. (1977). Correlation of the startle reflex and Mauthner
cell auditory responses in unrestrained goldfish. J. Exp. Biol. 66,
243254.

335


336
Neuron, Vol. 23, 325335, June, 1999, Copyright 1999 by Cell Press
Laser Ablations Reveal Functional Relationships
of Segmental Hindbrain Neurons in Zebrafish
Katharine S. Liu and Joseph R. Fetcho*
Department of Neurobiology and Behavior
State University of New York at Stony Brook
Stony Brook, New York 11794
Summary
Segmentation of the vertebrate brain is most obvious
in the hindbrain, where successive segments con-
tain repeated neuronal types. One such set of three
repeated reticulospinal neuronsthe Mauthner cell,
MiD2cm, and MiD3cmis thought to produce different
forms of the escape response that fish use to avoid
predators. We used laser ablations in larval zebrafish
to test the hypothesis that these segmental hindbrain
cells form a functional group. Killing all three cells
eliminated short-latency, high-performance escape
responses to both head- and tail-directed stimuli. Kill-
ing just the Mauthner cell affected escapes from tail-
directed but not from head-directed stimuli. These
results reveal the contributions of one set of reticulo-
spinal neurons to behavior and support the idea that
serially repeated hindbrain neurons form functional
groups.
Introduction
A striking feature of the vertebrate brain is its division
into segments that are recognizable at all levels of orga-
nization, from the gross anatomical to the molecular.
This segmentation is most evident in the hindbrain,
which contains reticulospinal neurons whose axons
project into spinal cord to influence spinal circuits in-
volved in motor control. Successive hindbrain segments
contain morphologically similar reticulospinal neurons
that form serially repeated arrays of cells. Even though
serially homologous neurons are present in the hind-
brain of diverse vertebrates, little is known about their
contributions to behavior.
In larval zebrafish, where the small number of neurons
allows the identification of individual cells, one can rec-
ognize single, morphologically similar cells in succes-
sive hindbrain segments (Kimmel et al., 1982; Metcalfe
et al., 1986). One such set includes three bilateral pairs
of neurons in adjoining segmentsthe Mauthner cell in
hindbrain segment 4 and two serial homologs, MiD2cm
and MiD3cm, located in successive segments caudal
to the Mauthner cell. The neurons in this set are thought
to generate an escape behavior used by fish to avoid
predators. When presented with a threatening stimulus,
a fish responds with a characteristic fast turn away from
the stimulus that allows the animal to swim off in the
direction opposite the threat. This highly reproducible,
* To whom correspondence should be addressed (e-mail: jfetcho@
neurobio.sunysb.edu).
extremely fast escape response, also known as the fast-
start response, is essential to the survival of the animal.
As reproducible as the behavior is, it is also modifiable.
Responses are directional, with the magnitude of the
turn depending upon the location of the stimulus (Eaton
et al., 1984; Eaton and Emberley, 1991). Thus, the cir-
cuitry that mediates successful escape responses pro-
vides not only short latency and high speed but also
proper timing and direction of movement.
The Mauthner cells are thought to be involved in the
initiation and laterality of the escape response. There is
a one-to-one correspondence between the Mauthner
cell firing and subsequent expression of the fast-start
response (Zottoli, 1977; Eaton et al., 1981), supporting
the cells role in the initiation of escape. The laterality
of the initial turn of the escape response is thought to
be determined by which of the two Mauthner cells is
activated (Eaton and Kimmel, 1980). A large sensory
input on one side of the fish leads to activation of the
Mauthner cell on that side. Because the Mauthner axon
crosses the midline to excite axial motoneurons on the
opposite side, activation of the Mauthner cell causes a
turn away from the stimulus.
The existence of a bilateral pair of Mauthner cells
might explain how lateral directionality of escape is me-
diated but is not sufficient to account for the varied
directionality of observed turns. Direct electrical stimu-
lation of the Mauthner cell in unrestrained goldfish re-
sults in weaker and less variable responses than sensory-
evoked responses (Nissanov et al., 1990), suggesting
that the Mauthner cell cannot, by itself, produce all forms
of escape. Lesion experiments also support the involve-
ment of other cells in escapes. In goldfish, electrolytic
removal of the Mauthner cell yielded no effect on the
behavior, other than a small increase in latency to re-
sponse (Eaton et al., 1982). These observations led re-
searchers to postulate the existence of parallel pathways,
acting along with the Mauthner cell during escapes (Ea-
ton and DiDomenico, 1985; Metcalfe et al., 1986; Fore-
man and Eaton, 1993). The obvious candidates for these
pathways were the Mauthner-like segmental homologs
MiD2cm and MiD3cm. Foreman and Eaton (1993) sug-
gested that differential activity of the cells in the
Mauthner array (Mauthner, MiD2cm, and MiD3cm) con-
trols the magnitude of the escape. Activity in all three
cells would excite more inter- and motoneurons, leading
to a larger turn, whereas activity in only one or two cells
would excite fewer downstream neurons, leading to a
smaller turn.
The proposal of Foreman and Eaton (1993) was sup-
ported by our previous functional imaging studies in
zebrafish, which showed that sensory stimuli known to
elicit different forms of escape behavior also produce
different patterns of activation in the Mauthner array
(OMalley et al., 1996). These studies took advantage of
the transparency of zebrafish larvae and used a calcium
indicator to image the activity of the Mauthner cell,
MiD2cm, and MiD3cm during escape responses. As pre-
dicted in the Foreman and Eaton model, all three homo-
logs were active during escapes elicited by stimuli to
337
Neuron
326
the head, while only the Mauthner cell was active during
escapes elicited by stimuli to the tail. These correlations
between neural activity and behavior suggested that
cells in the Mauthner array might generate different
forms of escape movements but offered no causal links
between the neurons and the behavior.
The transparency of zebrafish larvae offered us the
opportunity to conclusively link neurons and behavior
by examining the consequences of laser ablation of neu-
rons in the Mauthner array. Toward this end, we devel-
oped a noninvasive approach for killing single cells that
takes advantage of the relative phototoxicity of the dyes
we use for calcium imaging. The targets we chose for
ablation were the Mauthner cell alone and the entire
array, as these sets represent the extremes of activity
patterns observed in the imaging studies. Based upon
the functional imaging work, our predictions were that
killing all three cells in the Mauthner array would remove
the ability of the fish to produce high-performance es-
cape turns to both head and tail stimuli, whereas ablat-
ing just the Mauthner cell would have a greater effect on
turns in response to tail-directed than to head-directed
stimuli. The results from killing all three cells in the array
matched our predictions. The Mauthner lesions, while
partially supporting our predictions, produced unex-
pected results that allow us to resolve paradoxical ob-
servations from previous studies. Our results provide
direct support for the idea that repeated hindbrain neu-
rons form functionally related groups.
Results
Specificity and Effectiveness of Lesions
The indicator dyes typically used for visualizing cell ac-
tivity have adverse affects on the cells when exposed to
intense light. We observed such effects in our previous
Figure 1. Examples of Ablations
(A and B) A Mauthner array before (A) and the
day after (B) laser killing of Mauthner,
MiD2cm, and MiD3cm. Images show the
hindbrain viewed in horizontal section, with
rostral at the top.
(A) Prior to the lesions, there is bilateral label-
ing of hindbrain neurons with the large, bright
somata of the two Mauthner cells on the left
and right sides at the top (the asterisk on the
top left marks the Mauthner cell on that side).
Below the Mauthner cell on the left side are
the MiD2cm and MiD3cm cells, also marked
by asterisks.
(B) After killing the cells marked by the aster-
isks in (A), only minor debris is evident where
their somata were prior to the lesion. The
swollen axon of the ablated Mauthner cell
remains and terminates abruptly (arrow) near
where it once crossed toward the ablated
soma.
(C and D) A cluster of motoneurons in spinal
cord before (C) and immediately after (D) laser
illumination of the cell marked by an asterisk.
The laser is very selective, with only the tar-
geted cell affected by the illumination; imme-
diately adjacent cells retain their fluores-
cence.
Scale bars, 10 mm.
calcium imaging of neurons in zebrafish (Fetcho et al.,
1998) and set out to determine if this phototoxicity could
be used as a reliable method of killing single cells in
zebrafish larvae. Targeted cells were labeled by retro-
grade uptake after injection of a fluorescent indicator,
calcium green dextran (CGD), and identified by position
and morphology (Kimmel et al., 1982; Metcalfe et al.,
1986) on a confocal microscope. The laser was then
focused and maintained at highest intensity on the cen-
ter of a labeled cell. This led to an immediately noticeable
decrease in fluorescence of the cell that was likely due
to bleaching of the indicator. A day later, the fluorescent
cell body was no longer evident in confocal images,
suggesting cell death.
In the current work, a major concern was to identify an
accessible and reliable indicator that a cell was indeed
successfully killed by this illumination. Because larval
zebrafish are transparent, we can view the effects of
the photoablation directly in confocal optical sections.
The most dramatic examples of the effectiveness of
the approach were seen in lesions of the Mauthner cell
(Figures 1A and 1B) because its large axon was easily
identifiable, even without the cell body. In these cases,
optical sections 24 hr after ablations revealed a swollen,
truncated axon stump of the Mauthner cell leading to-
ward the original position of the soma (Figure 1B). Sur-
vival of the Mauthner axon after somatic lesions has
been reported before (Eaton et al., 1982; DiDomenico et
al., 1988). We found that a swollen, abruptly terminating
Mauthner axon stump was always associated with the
loss of a fluorescent soma. In about 25% of the ablation
attempts, the Mauthner cell body was fluorescent after
24 hr, with an intensity that was higher than that immedi-
ately after the attempted ablation. In these cases, the
axons of the cells always looked normal and the abla-
tions were obviously unsuccessful. We suspect that the
338
Behavioral Roles of Hindbrain Neurons in Zebrafish
327

Figure 2. Behavioral Trials

(A) Design of the experiments. Prelesion trials were collected by stimuli in each quadrant on successive trials (left); then, the targeted cells
were lesioned (middle); finally, postlesion trials were collected (right).
(B) Examples of two successive escape responses in an array-lesioned fish following stimuli on the intact side (top) and lesioned side (bottom).
Frames start when the water pulse contacts the head of the fish. For each trial, a single asterisk marks the initiation of response. The initial
movement is difficult to detect in the side-by-side view of the frames shown here but is evident when observed in video clips and on plots
of angular velocity. A double asterisk marks the frame of the maximal bend. Note that the response initiated from the lesioned side does not
begin until the response from the intact side is complete. The time from the start of the bend to the maximal bending is longer on the lesioned
side. Images were collected at 1,000 frames/s, and every third frame is shown (3 ms between frames).

fluorescence intensity recovered because dye that was
bleached out of the soma during the attempt to ablate
the cell was replaced by dye from the intact axon. The
one-to-one correlation between swollen axon stumps
and loss of fluorescence in the soma supports the con-
clusion that the loss of fluorescence (after 24 hr) is due
to somatic death.
This conclusion is also supported by experiments in
which we viewed cells under differential interference
contrast (DIC) optics shortly after exposure to the laser.
Targeted cells showed clear indicators of necrosisthe
formation of large vacuoles and a more pronounced
outline as the cell membrane begins to pull away from
its neighbors. These effects were always associated
with a loss of fluorescence the next day, again support-
ing the conclusion that this loss is a reliable indicator
of cell death.
To demonstrate further that the loss of fluorescence
was not simply due to photobleaching, we simultane-
ously labeled spinal neurons with Texas red (lex 5 568)
and CGD (lex 5 488) so that cells could be lesioned by
excitation of the former and their status monitored with
the latter. Cells were targeted with the confocal laser
until Texas red fluorescence was reduced to back-
ground levels. While these cells were still visible with
CGD shortly following the exposure, they were no longer
visible with either dye after 24 hr and beyond. Cells not
subjected to intense illumination remain fluorescent for
weeks. These results suggest that the absence of fluo-
rescence is not due simply to photobleaching but to cell
death. Thus, morphological evidence of degeneration
in fluorescence and DIC as well as double-labeling ex-
periments all indicate that the loss of fluorescence after
24 hr is a reliable indicator of cell death.
The amount of illumination time required to kill a cell
varied with cell size and location. The relatively small
and exposed motoneurons in the spinal cord (Figures
1C and 1D) required less than 1 min, while the much
larger cells of the Mauthner array deep in the hindbrain
(Figures 1A and 1B) required 1012 min. In the spinal
cord, the success rate was close to 100%. In the hind-
brain, where the cells are much deeper and occasionally
occluded by pigment, the success rate dropped to about
75%. It is possible that some targeted neurons that were

339
Neuron
328

Figure 3. Kinematic Analysis of Escape Trials

Rostral midline plots of an array-lesioned fish
(same trials as in Figure 2) show the position
of the fish in space over time. Midlines are
shown for all frames collected, with 1 ms be-
tween lines. Both trials are shown as if the
fish were turning to the right to allow easy
comparison; they were, however, produced
in response to stimuli on opposite sides of
the head of the same fish. The stimulus on
the intact side leads to a short-latency, rapid
turn away. Following a stimulus on the le-
sioned side, the fish remains still (except for
a slight drifting to the right from the push of
the water pulse) for much longer than on the
intact side and then turns away more slowly,
as shown by the more closely spaced lines. The right side shows this quantitatively in plots of the angle turned and the angular velocity of
the turn on intact and lesioned sides. The frame in which the stimulus arrives is marked by an arrow on the angle plots.

still fluorescent after 24 hr were damaged, but we did
not use these animals in our studies.
Laser illuminating unlabeled fish for the same time
periods and in the same regions of the brain as labeled
fish did not produce the behavioral effects observed
after illuminating labeled cells. This indicates that the
lesions were not a direct effect of the laser light, but
depended on interactions of the light with the dye in the
labeled cells.
These photoablations are highly specific. This is best
demonstrated in clusters of spinal motoneurons. A sin-
gle cell can be targeted in the midst of near neighbors
(Figure 1C). Neighboring cells retain their fluorescence
both immediately and 1 day after lesioning (Figure 1D).
In addition, in calcium imaging studies similar to OMal-
ley et al. (1996), we found that neurons that respond
during escapes retain that ability after lesioning of adja-
cent cells (data not shown). Thus, this technique is effec-
tive, minimally invasive, and highly specific compared
to ablation procedures previously used for behavioral
studies in vertebrate systems.
Behavioral Experiments
In response to a pulse of water on one side, fish perform
a large, fast turn away from the stimulus, followed by
a counterbend and further high-speed swimming. To
expedite analysis of this escape response, we wrote a
program in Labview that automatically tracked the fish
and generated plots of several kinematic parameters for
each trial. We focused our analysis on the performance
of the initial turn in the escape because previous data
suggested that the Mauthner array is important for gen-
erating the high-performance movements during this
turn. Several features of the initial turn were studied: the
latency to its initiation (time from the contact of the pulse
of water to the beginning of movement), the maximum
angle of the turn, its peak angular velocity, and its dura-
tion. Two classes of responsesnonescape turns that

Table 1. Means and Standard Errors for Kinematic Parameters

Array Lesions
Head Stimulus Tail Stimulus
Prelesion Postlesion Prelesion Postlesion
Lesioned Intact Lesioned Intact Lesioned Intact Lesioned Intact
Latency (ms) 3.8 6 0.1 3.1 6 0.1 43.1 6 2.2 4.4 6 0.6 6.2 6 0.3 8.2 6 0.5 51.3 6 3.0 12.7 6 1.5
Duration (ms) 7.9 6 0.1 8.0 6 0.1 17.2 6 1.2 9.4 6 0.2 7.3 6 0.2 7.2 6 0.2 13.5 6 1.0 9.1 6 0.5
Velocity (8/ms) 23.7 6 0.3 21.3 6 0.6 14.2 6 0.7 19.6 6 0.7 21.5 6 0.4 21.3 6 0.5 13.0 6 0.7 17.4 6 0.8
Angle (8) 128.7 6 2.5 117.5 6 2.1 118.4 6 4.8 127.4 6 5.6 108.7 6 2.0 111.6 6 2.1 89.9 6 6.3 96.1 6 5.5
Mauthner Cell Lesions
Head Stimulus Tail Stimulus
Prelesion Postlesion Prelesion Postlesion
Lesioned Intact Lesioned Intact Lesioned Intact Lesioned Intact
Latency (ms) 3.9 6 0.2 2.7 6 0.3 4.6 6 0.3 4.5 6 0.5 5.5 6 0.2 8.1 6 0.8 37.9 6 6.2 8.5 6 0.8
Duration (ms) 8.4 6 0.1 8.4 6 0.2 8.9 6 0.3 9.0 6 0.5 7.6 6 0.2 7.3 6 0.1 9.8 6 0.8 8.1 6 0.4
Velocity (8/ms) 24.5 6 0.5 23.6 6 0.7 22.3 6 0.6 21.4 6 0.3 22.1 6 0.5 22.6 6 0.6 17.0 6 0.7 20.8 6 0.4
Angle (8) 137.1 6 4.2 128.5 6 3.8 132.8 6 5.8 135.5 6 6.6 116.8 6 1.9 112.9 6 2.7 99.9 6 4.1 112.4 6 3.4
The means and standard errors for latency, maximum turn angle, peak angular velocity, and initial turn duration are presented for array lesions
(top) and Mauthner cell lesions (bottom). As in Figures 4 and 6, Table 1 is arranged to allow for comparisons between different categories.
Note that the prelesion values for the (to be) lesioned and intact sides are generally symmetrical. Exceptions are discussed in the text. The
values for the array lesions are calculated from 18 trials (3 trials for each stimulus from each of 6 fish). Those for the Mauthner cell lesions
are from 15 trials (3 trials for each stimulus from each of 5 fish).

340
Behavioral Roles of Hindbrain Neurons in Zebrafish
329

we call scoots, and turns toward the stimuluswere not

included in the data set. A more detailed account of the
basis for excluding them is presented in the Experimen-
tal Procedures.
A brief account of the experimental design (Figure
2A) is presented here to aid in the interpretation of the
results. Larval fish were injected with CGD, and the cell
labeling was verified by confocal microscopy. The day
after the labeling injection, prelesion behavioral trials
were collected at 1000 frames/s for each fish for detailed
kinematic analysis (Figure 3). The stimulus was delivered
to one of four quadrants on successive trialsleft head,
right head, left tail, or right tail (Figure 2A). The next day,
the targeted neuronseither the Mauthner cell alone or
all three array cellswere lesioned on one side of the
fish. Postlesion behavioral trials were then collected the
day after the lesions in the same manner as described
for prelesion trials (Figure 2B). Lastly, the animals were
remounted and examined under the confocal, in order
to directly verify that the lesioned cells were missing in
optical sections of the hindbrain.
A couple of aspects of this approach might lead to
behavioral deficits not specific to the lesions. First, the
injection of dye to fill the cells for ablation could result
in changes in performance. It might, for example, intro-
duce small asymmetries between the two sides not
present in uninjected fish. Second, manipulations of the
fish, particularly the embedding in agar, were physically
demanding to the delicate larval zebrafish and might
also result in nonspecific changes in performance. There-
Figure 4. Effects of Lesions on Response Latencies
fore, we anticipated general nonspecific effects as well
(Top) Latency to response before and after lesions of the entire
as slight asymmetries prior to lesions. The experiments Mauthner array (Mauthner, MiD2cm, and MiD3cm). The left side
were designed to control for these concerns by allowing shows the histograms of mean latency for stimuli on the two sides
us to make two sets of comparisons. We could compare of the head (top) or tail (bottom) prior to the lesions. The histogram
the performance of the same animals before and after bars on the side to be lesioned are in black. Prior to lesioning, the
lesions. The expectation was that pre- versus postlesion latencies are short (4 ms for head stimuli, slightly longer for tail) and
nearly symmetrical on the two sides. The right side shows similar
comparisons should reveal large changes on the le-
histograms for the two sides after the lesion. The latency on the
sioned side but not on the intact side. Any changes of lesioned side increases dramatically after the lesion to over 8 times
the intact side could be attributable to nonspecific ef- its prelesion value, whereas the intact side changes very little.
fects on the fish. We could also compare the two sides (Bottom) Histograms for lesions of the Mauthner cell alone in the
of the fish after the lesions, with the expectation that same format as the top panel. In these, the response latencies
the lesion would introduce a large asymmetry between increased on the lesioned side for responses to tail stimuli, but did
not change for head stimuli. The histograms for array lesions show
the two sides. Any observed changes could be confi-
means 6 SEM for 18 trials (3 trials for each stimulus from each of
dently attributed to the lesion if we saw both a significant 6 fish). Those for Mauthner lesions are from 15 trials (3 trials for
change on the lesioned side relative to its performance each stimulus from each of 5 fish). The p values report the signifi-
prelesion and a coincident, significant asymmetry in per- cance levels for ANOVA contrasts between the two sides of the fish
formance between the two sides of the fish after the before and after lesions. Asterisks mark those cases with significant
lesion. (p , 0.05) asymmetries between responses to stimuli on opposite
sides.

Prelesion Differences in Escapes to Head

and Tail Stimuli
Prior to lesions, the escapes produced by water pulses
directed at the head differed from those in response to
tail stimuli. Head escapes were faster and larger, and
they had a shorter latency and a longer duration (Table
1) than tail escapes. The differences in angular velocity
and angle turned were statistically significant (p , 0.05),
and the duration and latency differences were nearly so.
In pilot experiments, more robust differences between
head and tail escapes were seen in response to taps
on the head versus tail from a piezoelectrically driven
glass probe. The water pulses are a weaker and probably
less directional stimulus than taps, but we chose to use
water as our stimulus because repeated taps on the
head of the fish were detrimental to the animals escape
performance.
Effects of Lesions on Response Latency
The lesions had the most dramatic effects on the latency
to respond to the stimulus. In the prelesion trials, the
average latency for head-directed stimuli was 3.4 ms;
that for tail-directed stimuli was 7.0 ms. In animals in
which all three homologs were killed, the latency to
respond to head-directed stimuli on the lesioned side
increased 10-fold, from an average of 34 ms before
lesion to over 40 ms afterward, with little change on the

341
Neuron
330

unlesioned side (Figure 4, top; Table 1). There was also

a very significant asymmetry between the two sides after
the lesion. A similarly dramatic increase and asymmetry
in latency was seen for tail responses after array lesions
(Figure 4, top; Table 1).
Figure 2B shows frames taken from two successive
behavioral trials of an array-lesioned fish. In the first
trial, the stimulus was delivered to the intact side; in the
second, to the lesioned side. Frames were collected at
1 ms increments, but only every third frame from the
collected sequence is shown (3 ms between frames).
An asterisk denotes the initiation of response. Note that
the response on the lesioned side did not begin until
frame 12 (33 ms). By this time on the intact side, the
fish was well into its fourth bend and far from the stimu-
lus site. The lesion effect on latency was also seen in
the quantitation of movement in these trials (Figure 3).
These array-lesioned fish can still turn away from the
squirt, but typically do not even begin to do so until
after a normal escape response would be complete.
Thus, both the short-latency initial turn and the counter
turn of the escape are missing in the lesioned fish. Since
the ablated cells are thought to generate the initial turn,
the absence of a counter turn suggests that its produc-
tion is contingent upon the preceding initial turn.
The results for lesions of the Mauthner cell alone were
somewhat more complicated. These ablations left laten-
cies to head-elicited responses unchanged but resulted
in a significant increase in latencies to tail-elicited re-
sponses (Figure 4, bottom). However, the effects were
not as robust as those obtained by killing the entire
array. Figure 5 shows histogram plots of response la- Figure 5. Histograms of the Latency Values from Individual Trials
tency for all of the postlesion trials, both array and Each pair of histograms (AD) compares latency on intact and le-
Mauthner cell lesions. For array lesions, the effects on sioned sides. (A) and (B) show the effects on latency to response
latency were so dramatic that the values for intact versus for head and tail stimuli after lesions of the array. (C) and (D) show
lesioned sides were completely nonoverlapping for head effects on latency for head and tail stimuli after Mauthner lesions.
Lesions of the array removed all high-performance turns, such that
stimuli (Figure 5A) and largely so for tail stimuli (Figure
the latencies for head-elicited responses on intact and lesioned
5B). While lesions of the array eliminated all short- sides are completely nonoverlapping (A) and the values for tail-
latency responses to tail-directed stimuli, lesions of the elicited responses are largely so (B). Latencies to head-elicited re-
Mauthner cell alone resulted in a broader spectrum of sponses after Mauthner lesions were unchanged (C). Lesions of
latencies (Figure 5D). Most responses had long latencies the Mauthner cell produced a broad distribution of tail-elicited turn
like those following array lesion (over 20 ms), but some latencies ([D], bottom), which included some of the shortest-latency
turns and many that were longer than those on the intact side ([D],
of the shortest-latency responses (under 10 ms) still
top). The histograms include latencies from all trials in all fish studied
occurred. (n 5 18 trials in each histogram for array-lesioned fish and 15 trials
in each histogram for Mauthner-lesioned fish).
Effects of Lesions on Maximum Turn Angle
The lesions had the least dramatic and consistent ef-
fects on the maximum angle of the initial turn (Figure 6, strong conclusions regarding the affects of array lesions
turn angle). Before the lesions, the average maximum on turn angle. Photoablation of the Mauthner cell alone
turn angles in response to head-directed stimuli were had no effect on head-elicited responses but led to a
1288; those to tail-directed stimuli were 1128. There was small but significant (p 5 0.0268) decrease in angle for
an asymmetry between the two sides before the lesion tail-elicited responses on the lesioned side as compared
in one group (head stimuli in array lesions) that was to the intact one. These results indicate that while the
likely a consequence of the labeling injection, since we Mauthner cell and its homologs may contribute slightly,
did not observe such asymmetries in uninjected fish. In the control of maximum turn angle can largely be deter-
animals in which all three homologs were killed, there mined independently of the Mauthner array.
was a general, small decrease in maximum turn angle
in escapes elicited from the lesioned side, but there Effects of Lesions on Angular Velocity
were not significant asymmetries between lesioned and The lesions had clear effects on the angular velocity of
unlesioned sides after killing cells (Figure 6, turn angle; the turn (Figure 6, angular velocity), with the pattern of
Table 1). Because the array lesions did not produce both effects very similar to that seen for the latency of the
a change on the lesioned side and the development of response. Before the lesions, the average peak angular
a significant postlesion asymmetry, we could make no velocity of turns in response to head-directed stimuli

342
Behavioral Roles of Hindbrain Neurons in Zebrafish
331
was 23.28/ms; that to tail-directed stimuli was 21.8 ms.
In the array-lesioned group, there was an asymmetry
between the two sides in response to head stimuli prior
to lesions (likely due to the labeling injection), but the
other prelesion results showed good symmetry between
sides before the lesion. Killing all three homologs led
to a large, significant decrease in angular velocity (to
138148/ms) on the lesioned side for both head- and
tail-elicited responses. There were smaller decreases
in performance on the intact side after lesion that we
attribute to nonspecific effects. However, the large
change on the lesioned side led to the introduction of
an obvious and significant asymmetry between the two
sides after the lesion (Figure 6, angular velocity; Table
1). Photoablation of the Mauthner cell alone resulted in
a significant (p 5 0.0175) decrease in angular velocity
for tail-elicited responses but had no effect on head-
elicited responses. Thus, killing all three cells dramati-
cally slowed the turn in response to both head and
tail stimuli, whereas killing the Mauthner cell alone only
slowed tail responses.
Effects of Lesions on Duration of Initial Turn
The previous sections reveal that after array lesions, the
fish turn more slowly, but to about the same extent.
Thus, one might expect that they should be taking a
significantly longer time to complete the turn. This is
indeed what we found (Figure 6, duration). Before le-
sions, there was excellent symmetry in the duration of
the turns to both sides. Head turns averaged 8.2 ms;
tail turns, 7.3 ms. Killing all three homologs doubled the
duration of turns on the lesioned side for both head- (17
Figure 6. Effects of Lesions on Turn Angle,
Angular Velocity, and Duration of the Turn
Histograms of mean values (6 SEM) for each
kinematic parameter in the same format as
those shown for latency in Figure 4. The val-
ues for array lesions are in boxes on the top;
those for Mauthner (M-cell) lesions are on the
bottom. Each box contains histograms for
opposite sides of the fish in response to head
(top) and tail (bottom) stimuli pre- (left) and
post- (right) lesions. The lesioned side is in
black in both pre- and postlesion histograms.
Bottom panels show the same for lesions of
the Mauthner cell alone. The p values for
ANOVA contrasts between measurements
from opposite sides of the fish are listed, with
those pairs showing a significant (p , 0.05)
difference marked with asterisks. Further de-
tails are in the text.
ms) and tail-elicited responses (13.5 ms), with only small
changes on the intact side. This produced a significant
asymmetry between the two sides after the lesion (Fig-
ure 6, duration; Table 1). Photoablation of the Mauthner
cell alone resulted in a smaller but significant (p 5
0.0048) increase in turn duration for tail-elicited re-
sponses but had no effect on head-elicited responses.
Again, the pattern of these changes was similar to those
seen for latency and angular velocity.
Discussion
We set out to evaluate the role of the Mauthner array in
generating escape responses by examining the behav-
ioral consequences of killing these cells. The Foreman
and Eaton model (1993) proposed that the array cells
contribute differentially to control the strength of the
motor output. The Mauthner cell alone mediates tail-
elicited responses, resulting in a shallower bend, while
all three homologs are involved in head-elicited re-
sponses, resulting in a deeper bend. Using calcium im-
aging in larval zebrafish, OMalley et al. (1996) reported
that the activity patterns of the Mauthner array match
this model. Our lesion results now provide a causal link
between these neurons and behavior. Lesions of the
Mauthner array resulted in the elimination of high-per-
formance escape responses to both head- and tail-
directed stimuli, as measured by latency to response,
angular velocity, and turn duration. Lesions of the
Mauthner cell alone had significant, similarly detrimental
effects on tail-elicited responses, while head-elicited
responses appeared unchanged. Since we have not
343
Neuron
332
done separate lesions of MiD2cm and MiD3cm, we do
not yet know their individual contributions to escapes.
However, the greater impact of the array lesions support
the idea that MiD2cm and/or MiD3cm, like the Mauthner
cell, contributes to escape behavior.
These results resolve an apparent paradox in earlier
Mauthner lesion studies. Substantial physiological evi-
dence links activity of the Mauthner cell to escape be-
havior (Zottoli, 1977; Eaton et al., 1981; Nissanov et al.,
1990). It was therefore surprising when lesioning studies
reported few and inconsistent behavioral changes after
killing the Mauthner cell (Eaton and Kimmel, 1980; Kim-
mel et al., 1980; Eaton et al., 1982). After irradiating
zebrafish early in development, Kimmel et al. (1980)
found that those larvae that lacked a Mauthner cell also
exhibited decreased fast-start performances on the le-
sioned side. However, irradiation likely led to the re-
moval of other cells, some of which might be involved
in escape behavior. In the most cell-specific lesions,
Eaton et al. (1982) electrolytically killed the Mauthner
cell in adult goldfish and found no effects except a small
but significant increase in latency. Since the nature of
their stimulus did not allow them to determine where the
fish would perceive the stimulus along the rostrocaudal
axis, head- and tail-elicited responses were pooled to-
gether in their analyses. We now know that lesioning
the Mauthner cell alone has no effect on head-elicited
escapes but does substantially impair tail-elicited re-
sponses. These impairments included large effects on
the latency to respond and smaller, though still signifi-
cant, effects on angular velocity and duration of the
initial turn. If head- and tail-elicited escapes were pooled
together following Mauthner cell lesions, the unchanged
head responses would mask the smaller deficits in angu-
lar velocity and duration and reduce the large increases
in latency. This may explain why, in their pooled data,
Eaton et al. (1982) found only changes in latency follow-
ing Mauthner cell lesions.
Our observations demonstrate that the Mauthner cell
does indeed play an important role in generating escape
responses, particularly those elicited by caudal stimuli.
While clarifying the role of the Mauthner cell in escape
behavior, our observations raise new and interesting
questions about the relative contributions of the three
homologs in the Mauthner array. Based on the Foreman
and Eaton model, we would predict that the removal of
any of the three homologs should result in a decrease
in head-elicited escape performance. However, removal
of the Mauthner cell had no discernible effect on head-
elicited responses. This is surprising, given that activa-
tion of the Mauthner cell is sufficient to produce a strong
bend (Nissanov et al., 1990). These observations high-
light the difficulties in making predictions about le-
sioning experiments in complex systems, where several
neurons potentially contribute to a behavior (Eaton and
DiDomenico, 1985).
In this particular case, a likely explanation for the ap-
parent lack of effects of Mauthner lesions on head-elic-
ited responses is the behavioral context in which the
escapes were elicited. The escapes were produced while
the animal was still, instead of while swimming. This
allowed us more easily to control stimulus direction and
strength but ignored some of the adaptive relevance of
the fast-start response. A key function of the Mauthner
cell is to insure a powerful bend by overriding other
motor programs (Svoboda and Fetcho, 1996). It is possi-
ble that MiD2cm and MiD3cm can, by themselves, pro-
duce a high-performance escape from rest, when axial
motor activity is minimal, but could not do so without
the Mauthner cell if there was substantial conflicting motor
activity. If this explanation is correct, then Mauthner cell
lesions should produce impairments in escapes elicited
by head stimuli during swimming.
The Foreman and Eaton model and our previous cal-
cium imaging studies, which showed that the Mauthner
cell alone is active in escapes to tail stimuli, led us to
predict that ablating the Mauthner cell should remove
all high-performance tail-elicited responses. However,
even in tail-elicited escapes, we did occasionally ob-
serve high-performance responses from the lesioned
side after killing the Mauthner cell. One possible expla-
nation for the remaining high-performance turns was
that they occurred when the stimulus was more rostral
than usual, thereby activating one or both of the re-
maining homologs. However, there was no correlation
between these high-performance responses and the
rostrocaudal position of the stimulus. Another possibil-
ity is that MiD2cm and/or MiD3cm takes over after
Mauthner lesions. The fact that these high-performance
responses are completely eliminated in animals where
the entire array has been lesioned supports a role for
MiD2cm and/or MiD3cm in the production of these re-
sponses after Mauthner lesions. Functional substitution
by MiD2cm and MiD3cm could be revealed by calcium
imaging if these cells, which do not respond to tail-
directed stimuli in intact fish, do respond to tail-directed
stimuli after Mauthner lesions.
Our observations indicate that zebrafish can produce
turns of a given amplitude via multiple pathways. While
all short-latency, high-performance escapes are re-
moved by lesioning the Mauthner array, the animals
are still capable of sensing the stimulus, calculating its
direction, and responding with an appropriate amplitude
turn, albeit much more slowly and at a very long latency.
One can compare the Mauthner array circuit to a reflex
pathway, allowing quick and decisive retreat from dan-
ger. After its removal, fish can still turn away from stimuli,
but they most likely do so via other slower, polysynaptic
pathways that require more processing and hence a
longer latency to respond. Although these fish can re-
spond with a turn of appropriate amplitude, these slow
turns do not even begin until after a normal escape
would be complete. Such fish would likely be doomed
in a predatory attack.
Our experiments show that the serially repeated retic-
ulospinal neurons in the Mauthner array form a function-
ally related group. These experiments were possible
because zebrafish have relatively few reticulospinal neu-
rons and they form sets of repeated cells with individu-
ally identifiable members. However, the overall segmen-
tal organization of hindbrain neurons, so obvious in
zebrafish, is evident in vertebrates generally (Clarke and
Lumsden, 1993). Many of these hindbrain neurons are
part of a reticulospinal system that is strikingly similar in
fish and mammals (Fetcho, 1992; Lingenhohl and Friauf,
1994). Thus, it seems very likely that segmental func-
tional groups similar to those in zebrafish also exist in
other vertebrates.
344
Behavioral Roles of Hindbrain Neurons in Zebrafish
333
Laser ablations of single cells, used first to study de-
velopment (Sulston and White, 1980; Eisen et al., 1989),
have proved fruitful for linking neurons to behavior in
invertebrate systems (Selverston and Miller, 1980; Chal-
fie et al., 1985; Bargmann and Horvitz, 1991). Our lesion
studies of the Mauthner array in zebrafish extend this
approach to the analysis of the behavioral contribution
of identified neurons in a vertebrate system. Zebrafish
have many other small groups of identifiable neurons in
spinal cord and brain whose behavioral roles could be
analyzed with this approach. Cell-specific ablations,
when combined with the functional imaging at single-
cell resolution and the genetic tools also available in
zebrafish (Granato et al., 1996; Fetcho and Liu, 1998),
should make zebrafish a powerful model for exploring
the neural basis of vertebrate behavior.
Experimental Procedures
Larval fish were obtained from laboratory breeding stock. Zygotes
were kept in 10% Hanks buffer at 28.58C until the fish had developed
to the point where they could swim in the water column (about 4
days). At this stage, they were moved to room temperature in order
to acclimate the fish to the temperature at which the experiments
were done (268C).
Retrograde Labeling
Four- to five-day-old larval zebrafish were anesthetized with 0.02%
3-aminobenzoic acid ethyl ester (MS222). The targeted cells were
retrogradely labeled by pressure injection via a glass microelectrode
of a 50% solution of CGD (10,000 MW; Molecular Probes, Eugene,
OR) in 10% Hanks into the spinal cord (Fetcho and OMalley, 1995).
Injections were targeted to the ventral cord to selectively label
Mauthner and its homologs without disrupting more dorsal sensory
pathways. Since any injection would disrupt local circuits, injections
were also targeted to the most caudal part of the tail and all test
stimuli were given rostral to the injection site. After injection, the
fish were allowed to recover in 10% Hanks solution containing food
(cultured paramecia). Nine to ten hours later, the cell labeling was
verified under the confocal microscope. Between four to six of the
best-labeled animals were retained in 10% Hanks containing para-
mecia for behavioral trials.
Confocal Microscopy
The fish were briefly anesthetized in MS222, placed on a cover glass
in a petri dish, embedded on their backs in a thin layer of 1.2% agar
(Eaton et al., 1984), and screened with confocal microscopy for the
desired labeling. Confocal images were obtained by looking into
the head of the intact fish using a Zeiss inverted microscope with
a 633 water objective and a Zeiss laser-scanning confocal imaging
system (LSM 510). Mauthner cells and their homologs were identi-
fied by their highly characteristic morphology and position (Kimmel
et al., 1982; Metcalfe et al., 1986). The potential for photo-induced
damage to the cells during screening was reduced by minimizing
the illumination used and increasing the sensitivity of the system
(by opening the confocal aperture and increasing gain in the photo-
multiplier).
High-Speed Recording of Behavior
Labeled fish were placed into individual petri dishes (3.5 cm) filled
with 10% Hanks buffer at a depth of 34 mm. Escape responses
were recorded with a high-speed camera that captures images digi-
tally at 1,000 frames/s (EG&G Reticon, Sunnyvale, CA). At the same
time, a slow-speed video camera recorded the entire procedure. In
this way, we could verify later which trials were retained digitally
and the history of each trial. Escape responses were elicited by a
measured pulse of water, delivered from a picospritzer (General
Valve, Fairfield, NJ) through a syringe and 27 gauge needle cut blunt
and bent to about 1008. In order to monitor the stimulus, fast green
dye was added to the water at a concentration of 0.05 mg/ml. The
needle tip was kept about 0.5 mm from the fish at the time the
stimulus was given. A successful escape trial was recorded from
the beginning of the trigger to the picospritzer, as indicated by a
stimulus indicatora deflection of a piezoelectric crystal that was
wired in parallel with the picospritzer. Typically, the water left the
needle tip about 1014 ms after the crystal deflection and hit the
fish about 2 ms after that.
The water pulses were generated by a pressure pulse of 17 psi
and 5 ms duration. The pressure was occasionally adjusted in order
to achieve a consistent stimulus (as monitored by the speed of exit
of the dye) but adjustments did not vary more than 61 psi. These
settings were determined to be the lowest level that can reliably
produce an escape response without the water stream disturbing
the animals movements.
Squirting the colored water in front of the fish, so that the animal
might see but not feel the stimulus, did not elicit a response. There-
fore, we believe the animals are not responding to the visual aspects
of the stimulus. The most likely sensory modalities mediating the
response are the somatosensory system, the lateral line, or the
ear. If the response is visual, our latency measurements should be
calculated from the time the dye exits the needle, instead of from
the time the water made contact with the fish. This would have
added only about 2 ms to our latency data. This small change would
not have altered our conclusions about the consequences of lesions,
which had large affects on latency.
Stimuli to the head were directed at the ear. Stimuli to the tail
were directed caudal to the anal pore but well rostral to the injection
site. The order of stimulus presentation was to the left side of the
head and then the right, followed by the left side of the tail and then
the right. All of the fish in the group were tested for the same
quadrant before moving on to the next trial/quadrant. Therefore,
each fish rested between trials for an average of around 20 min. Also,
each fish was not presented with a stimulus in the same quadrant for
well over 1 hr. Such delays should prevent habituation, fatigue, etc.
On occasion, a fish either failed to respond to the stimulus, the
stimulus missed the fish, the fish gave a premature response, or
the fish turned slightly on its side. Only responses that occurred
after the stimulus was deployed and in which the fish remained
upright throughout the initial bend were analyzed digitally. A mini-
mum of 20 trials were collected per fish (5 trials per quadrant).
In pilot experiments, we elicited escapes by using a glass probe
driven by a piezoelectric device to abruptly touch the fish on the
head or tail. This produced more robust differences in the form of
the escapes to head and tail stimuli than the squirts, but because
the taps were more traumatic to the fish we elected to use squirts
to obtain the many trials needed for the lesion experiments.
Photoablation of Cells
After prelesion behavioral testing and overnight recovery, the fish
were mounted in agar as described above and the target neurons
either Mauthner alone or all three array cellswere reidentified by
their position and morphology (Kimmel et al., 1982; Metcalfe et al.,
1986). To lesion a targeted cell, the confocal microscope was fo-
cused at the highest zoom (Zeiss 633 water 0.9 NA) on the brightest
point of the labeled cell, and the cell was exposed to the laser (lex 5
488) at maximum intensity. To kill the Mauthner cell, exposure for
10 min was usually necessary. The smaller homologs, probably
because they contain less dye, required longer exposures of up to
12 min. We checked the region of the illumination to make sure
there was no evidence of general tissue damage immediately after
the ablation and 2 days later. Exposure for longer than 20 min could
lead to tissue damage, so we limited the exposure to well below
that duration. Our attempts to use the minimal effective exposure
time might explain why not all of our lesion attempts were success-
ful. After exposure to the laser, the fish were transferred from agar
into a small petri dish of 10% Hanks buffer containing paramecia
and allowed to recover overnight. Following postlesion behavioral
trials, the animals were remounted for confocal microscopy in order
to verify the success of the lesions and check for peripheral damage.
Data Analysis
Several kinematic parameters of the initial turn of the escape were
selected for analysis: the latency to its initiation (time from the con-
tact of the pulse of water to the beginning of the bending movement),
345
Neuron
334
the maximum angle of the turn, its peak angular velocity, and its
duration (time from the beginning of movement to the maximum
angle). The movements of the fish were analyzed for these kinematic
parameters by using a specialized program written in Labview (Na-
tional Instruments, Austin TX). The analysis was automated; the
image of the fish was thresholded and the binary silhouette of
the fish was used to determine, via fitting routines, the location
of the rostral midline, which does not bend much during the turns.
The midline from each successive frame was plotted to give a repre-
sentation of the animals movements (Figure 3, top). The program
calculated the angle between the position of the midline in succes-
sive frames and its original position (Figure 3, middle) and also
provided other kinematic data such as the duration of the turn and
the angular velocity, which was obtained by differentiation of the
curves for turn angle (Figure 3, bottom).
These parameters were then statistically compared using analysis
of variance (SuperAnova, Abacus Concepts, Berkeley CA) with re-
peated measures. We used statistical contrasts in the ANOVA to
perform specific comparisons that we identified as the important
ones prior to the experiment. These included comparisons of prele-
sion versus postlesion, intact side versus lesioned side, head versus
tail, and lesions of the array versus lesions of the Mauthner cell
alone.
Not every trial collected was used in our statistical analysis. Those
trials that were omitted fell into two classes: nonescape or scoot
responses, and escape responses in which the fish turned toward
the stimulus. These classes and the justification for their omission
are discussed below. The data presented comprise three trials prele-
sion and three postlesion (out of five collected in each case) from
each quadrant of eleven fish: six fish with array lesions and five with
just Mauthner lesions.
In a small percentage (17%) of trials, the fish moves away from
the stimulus with a movement that is sufficiently different from an
escape response that it can be distinguished visually during the
behavior. These responses consist of a very weak turn away from
the stimulus, followed by a somewhat larger turn in the same direc-
tion, with no evidence of the initial turn/counter turn sequence typi-
cal for escapes. This appears visually as a weak scoot away from
the stimulus. We analyzed the kinematics of these responses to
verify that they were easily distinguishable from escapes. They con-
sisted of an initial very small (about 108) and slow (less than 48/ms)
turn away, followed after a long (more than 20 ms) pause by another
turn in the same direction. This is very different from an escape,
which consists of a large, rapid, continuous turn away from the
stimulus, followed by a robust counter turn. The scoot responses
were seen with the same frequency in prelesion trials and postlesion
trials and were, therefore, not a result of cell lesion. Since the scoots
were qualitatively different movements from escapes, we did not
include them in our analysis.
On occasion, an animal turned toward the direction of the stimu-
lus. This inappropriate response was almost never seen in intact
fish but the frequency increased in fish that had been injected (4%)
and increased even more in lesioned fish (11%). There was no differ-
ence in the frequency of these turns between fish in which the entire
array was lesioned and those in which only Mauthner was lesioned.
Turns toward the stimulus exhibited a short latency and high angular
velocity that is characteristic of responses of intact fish. After le-
sioning, these turns occurred only when the stimulus was presented
on the lesioned side and not when it was presented on the intact
side. For these reasons, we believe that these turns were the result
of the Mauthner array cells on the remaining, intact side being inap-
propriately activated. Increased activation of the intact side follow-
ing unilateral Mauthner cell lesion has been reported previously in
goldfish (DiDomenico et al., 1988). Because these inappropriate
responses were most likely mediated by the cells on the unlesioned
side, they were not included in our analysis.
Acknowledgments
The authors would like to thank Melina Hale for many helpful com-
ments on the manuscript. The work reported here was supported
by NIH grant NS26539 (J. R. F.) and a postdoctoral fellowship from
the Helen Hay Whitney Foundation (K. S. L.).
Received April 15, 1999; revised May 17, 1999.
References
Bargmann, C.I., and Horvitz, H.R. (1991). Chemosensory neurons
with overlapping functions direct chemotaxis to multiple chemicals
in C. elegans. Neuron 7, 792742.
Chalfie, M.J., Sulston, J.E., White, J.G., Southgate, E., Thomson,
J.N., and Brenner, S. (1985). The neural circuit for touch sensitivity
in Caenorhabditis elegans. J. Neurosci. 5, 956964.
Clarke, J.D., and Lumsden, A. (1993). Segmental repetition of neu-
ronal phenotype sets in the chick embryo hindbrain. Development
118, 151162.
DiDomenico, R., Nissanov, J., and Eaton, R.C. (1988). Lateralization
and adaptation of a continuously variable behavior following lesions
of a reticulospinal command neuron. Brain Res. 473, 1528.
Eaton, R.C., and DiDomenico, R. (1985). Command and the neural
causation of behavior: a theoretical analysis of the necessity and
sufficiency paradigm. Brain Behav. Evol. 27, 132164.
Eaton, R.C., and Emberley, D.S. (1991). How stimulus direction de-
termines the trajectory of the Mauthner-initiated escape response
in a teleost fish. J. Exp. Biol. 161, 469487.
Eaton, R.C., and Kimmel, C.B. (1980). Directional sensitivity of the
Mauthner cell system to vibrational stimulation in zebrafish larvae.
J. Comp. Physiol. [A] 140, 337342.
Eaton, R.C., Lavender, W.A., and Wieland, C.M. (1981). Identification
of Mauthner-initiated response patterns in goldfish: evidence from
simultaneous cinematography and electrophysiology. J. Comp.
Physiol. [A] 144, 521531.
Eaton, R.C., Lavender, W.A., and Wieland, C.M. (1982). Alternative
neural pathways initiate fast-start responses following lesions of
the Mauthner neuron in goldfish. J. Comp. Physiol. [A] 145, 485496.
Eaton, R.C., Nissanov, J., and Wieland, C.M. (1984). Differential acti-
vation of Mauthner and non-Mauthner startle circuits in the zebra-
fish: implications for functional substitution. J. Comp. Physiol. [A]
155, 813820.
Eisen, J.S., Pike, S.H., and Debu, B. (1989). The growth cones of
identified motoneurons in embryonic zebrafish select appropriate
pathways in the absence of specific cellular interactions. Neuron 2,
10971104.
Fetcho, J.R. (1992). The spinal motor system in early vertebrates
and some of its evolutionary changes. Brain Behav. Evol. 40, 8297.
Fetcho, J.R., and Liu, K.S. (1998). Zebrafish as a model system for
studying neuronal circuits and behavior. In Neural Mechanisms for
Generating Locomotor Activity, O. Kiehn, R.M. Harris-Warrick, L.M.
Jordan, H. Hultborn, and N. Kudo, eds. Ann. NY Acad. Sci. 860,
333345.
Fetcho, J.R., and OMalley, D.M. (1995). Visualization of active neural
circuitry in the spinal cord of intact zebrafish. J. Neurophysiol. 73,
399406.
Fetcho, J.R., Cox, K.J.A., and OMalley, D.M. (1998). Monitoring
activity in neuronal populations with single-cell resolution in a be-
having vertebrate. Histochem. J. 30, 153167.
Foreman, M.B., and Eaton, R.C. (1993). The direction change con-
cept for reticulospinal control of goldfish escape. J. Neurosci. 13,
41044113.
Granato, M., van Eeden, F.J., Schach, U., Trowe, T., Brand, M.,
Furutani-Seiki, M., Haffter, P., Hammerschmidt, M., Heisenberg,
C.P., Jiang, Y.J., et al. (1996). Genes controlling and mediating loco-
motion behavior of the zebrafish embryo and larva. Development
123, 399413.
Kimmel, C.B., Eaton, R.C., and Powell, S.L. (1980). Decreased fast-
start performance of zebrafish larvae lacking Mauthner neurons. J.
Comp. Physiol. [A] 140, 343350.
Kimmel, C.B., Powell, S.L., and Metcalfe, W.K. (1982). Brain neurons
which project to the spinal cord in young larvae of the zebrafish. J.
Comp. Neurol. 205, 112127.
Lingenhohl, K., and Friauf, E. (1994). Giant neurons in the rat reticular
formation: a sensorimotor interface in the elementary acoustic star-
tle circuit? J. Neurosci. 14, 11761194.
Metcalfe, W.K., Mendelson, B., and Kimmel, C.B. (1986). Segmental
homologies among reticulospinal neurons in the hindbrain of the
zebrafish larva. J. Comp. Neurol. 251, 147159.
346
Behavioral Roles of Hindbrain Neurons in Zebrafish
335

Nissanov, J., Eaton, R.C., and DiDomenico, R. (1990). The motor

output of the Mauthner cell, a reticulospinal command neuron. Brain
Res. 517, 8898.
OMalley, D.M., Kao, Y.-H., and Fetcho, J.R. (1996). Imaging the
functional organization of zebrafish hindbrain segments during es-
cape behavior. Neuron 17, 11451155.
Selverston, A.I., and Miller, J.P. (1980). Mechanisms underlying pat-
tern generation in lobster stomatogastric ganglion as determined
by selective inactivation of identified neurons. I. Pyloric system. J.
Neurophysiol. 44, 11021121.
Sulston, J.E., and White, J.G. (1980). Regulation and cell autonomy
during postembryonic development of Caenorhabditis elegans. Dev.
Biol. 78, 577597.
Svoboda, K.R., and Fetcho, J.R. (1996). Interactions between the
neural networks for escape and swimming in goldfish. J. Neurosci.
16, 843852.
Zottoli, S.J. (1977). Correlation of the startle reflex and Mauthner
cell auditory responses in unrestrained goldfish. J. Exp. Biol. 66,
243254.

347


348
A structural and functional ground plan for neurons in
the hindbrain of zebrash
Amina Kinkhabwalaa,b, Michael Rileya, Minoru Koyamaa, Joost Monena, Chie Satouc,d, Yukiko Kimurac,
Shin-ichi Higashijimac,d, and Joseph Fetchoa,1
a
Department of Neurobiology and Behavior, Cornell University, Ithaca, NY 14853; bPrinceton Neuroscience Institute, Princeton University, Princeton, NJ 08544;
c
Okazaki Institute for Integrative Bioscience, National Institute for Physiological Sciences, National Institutes of Natural Sciences, Okazaki, Aichi 444-8787,
Japan; and dDepartment of Physiological Sciences, Graduate University for Advanced Studies (SOKENDAI), Okazaki, Aichi 444-8585, Japan
Edited by Lynn T. Landmesser, Case Western Reserve University, Cleveland, OH, and approved December 9, 2010 (received for review August 16, 2010)
The vertebrate hindbrain contains various sensory-motor net-
works controlling movements of the eyes, jaw, head, and body.
Here we show that stripes of neurons with shared neurotransmit-
ter phenotype that extend throughout the hindbrain of young
zebrash reect a broad underlying structural and functional
patterning. The neurotransmitter stripes contain cell types with
shared gross morphologies and transcription factor markers.
Neurons within a stripe are stacked systematically by extent and
location of axonal projections, input resistance, and age, and are
recruited along the axis of the stripe during behavior. The
implication of this pattern is that the many networks in hindbrain
are constructed from a series of neuronal components organized
into stripes that are ordered from top to bottom according to
a neurons age, structural and functional properties, and behav-
ioral roles. This simple organization probably forms a foundation
for the construction of the networks underlying the many behav-
iors produced by the hindbrain.
interneuron | locomotion | recruitment | topography
T he hindbrain contains a diverse set of sensory-motor networks
that control movements required for vision, respiration, mas-
tication, and locomotion in all vertebrates (1, 2). Most often these
different networks are studied separately from one another, per-
haps because the behaviors are distinct, and the regional differ-
entiation of hindbrain suggests that its several networks might
have little in common. Thus, we have strong data for the hindbrain
control of eye movements, respiration, and locomotion (310), but
fewer unifying principles of structural and functional organization
that apply across the different networks.
Structurally, the hindbrain is divided into segments, called
rhombomeres, which differ in the expression of homeotic genes,
in the morphological differentiation of neurons, and in their
sensory inputs and motor outputs (2, 11). Though there are clear
distinctions among rhombomeres, there are indications from
previous developmental work using in situ staining for tran-
scription factors and backlling of hindbrain neurons that there
may be structural patterns that cross rhombomere boundaries
(1216). Prior work has also revealed parallels in the deve-
lopment of hindbrain and spinal cord, with the hindbrain sharing
features of the now-classic transcription factor code that directs
development in spinal cord (1722). Though these studies did
not explore function because they were performed during early
development, they raised the possibility of a broader structural-
functional patterning that spans rhomobomeres and may un-
derlie the organization of circuits for different behaviors.
Here we show that there is indeed a broad structural and
functional patterning of neurons in the hindbrain of young
zebrash. The work was initially prompted by a striking patterning
observed in earlier work in which we used in situ staining for
markers of neurotransmitter phenotype to reveal putative glyci-
nergic, GABAergic, and glutamatergic neurons in the hindbrain
(23). We found that neurons of the same transmitter phenotype
were clustered together into stripes when viewed in cross-sections,
and that these extended as columns throughout much of the
rostrocaudal axis of the hindbrain. The pattern is evident in sh
11641169 | PNAS | January 18, 2011 | vol. 108 | no. 3
when they are freely swimming and most of the major adult
behaviors driven by hindbraine.g., swimming, feeding, and eye
movementsare functional. This suggests that the diverse net-
works in hindbrain might share an orderly functional patterning of
neurons. We now show that the transmitter stripes reect an
underlying order in the hindbrain in which stripes contain cell
types, stacked in order by age as well as structural and functional
properties. This pattern is tied to behavior because neurons are
recruited along the axis of a stripe as the speed of a motor be-
havior increases. We conclude that the diverse networks in the
hindbrain have at their foundation a common structural and
functional plan.
Results
Hindbrain Transmitter Stripes and Transcription Factor Patterning.
We examined the transmitter stripe patterning in more than 20
larval, posthatching sh from each of two BAC transgenic lines
expressing green (GFP) and red (DsRed) uorescent proteins
driven by promoter regions of glyt2 and vglut2.1 (vglut2) to mark
glycinergic and glutamatergic neurons, respectively (24, 25). Stripes
were clearly evident in the hindbrain of dual transgenic lines (vglut:
DsRed glyt2:GFP), as shown in the example from a 4-d post-
fertilization (dpf) sh in Fig. 1A (n = 16). In cross-sections, these
stripes were organized in an interleaved manner from medial to
lateral on each side of the brain in all hindbrain segments (rhom-
bomeres), with a glutamatergic stripe located medially followed by
alternating glycinergic and glutamatergic stripes that extend pre-
dominately dorsoventrally and rostrocaudally. This organization
was present in all of the sh. Although the glycinergic and gluta-
matergic neurons were segregated from one another in a columnar
pattern, there were other unlabeled neurons, both scattered within
stripes and in large contiguous areas between stripes that are
probably neurons with other transmitter phenotypes (cholinergic
and GABAergic, e.g.) based upon prior in situ staining (23).
We next used transgenic lines and immunostaining to in-
vestigate the relationship between the transmitter stripes and the
transcription factors alx (called chx10 in mammals), dbx1b, en-
grailed-1, and barhl2. The transcription factors were expressed
in stripe-like patterns resembling the patterning of transmitter
stripes. Neurons expressing the alx (chx10) transcription factor
were clustered medially and overlapped the most medial gluta-
matergic stripe (Fig. 1B). Three-dimensional colocalization
revealed that most, if not all, medial glutamatergic stripe neu-
rons express the alx transcription factor (Fig. 1B, panel 3; n = 2
sh). Immunostaining for alx protein in the alx:GFP transgenic
line conrmed that this line reliably marked the alx positive
Author contributions: A.K. and J.F. designed research; A.K., M.R., M.K., and J.M. per-
formed research; C.S., Y.K., and S.-i.H. contributed new reagents/analytic tools; A.K.,
M.R., M.K., J.M., and J.F. analyzed data; and A.K. and J.F. wrote the paper.
The authors declare no conict of interest.
This article is a PNAS Direct Submission.
Freely available online through the PNAS open access option.
1
To whom correspondence should be addressed. E-mail: jrf49@cornell.edu.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1012185108/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1012185108
349

Fig. 1. Interleaved transmitter stripes of neurons in hindbrain and their
overlap with transcription factor expression patterns. (A) Cross-sections of
hindbrain in vivo in rhombomeres 4, 6, and 8 (A, panels 13, respectively)
from a glyt2:GFP vglut:DsRed four dpf sh, showing interleaved stripes of
glycinergic (green) and glutamatergic neurons (red). (BE) Transcription
factor stripes are present in hindbrain and overlap glutamatergic and gly-
cinergic stripes. (B) Cross-sections from rhombomere 8 of a dual transgenic
alx (chx10):GFP and vglut: DsRed three dpf sh shows that alx expression
colocalizes with the most medial glutamatergic stripe. (C) Engrailed-1
immunostaining in cross-sections from a glyt2 transgenic line from rhom-
bomere 7 in a 5-dpf sh showing that all medial glycinergic stripe neurons
express engrailed-1. (D) Cross-section from rhombomere 7 in a four dpf
vglut:DsRed dbx1b:GFP transgenic sh shows that dbx colocalizes with
both the middle glutamatergic stripe as well as a more lateral stripe in
a region known to contain glycinergic neurons. (E) Cross-sections from
rhombomere 8 in a 4-dpf vglut:DsRed barhl2:GFP sh indicate that barhl2
staining overlaps the lateral portion of the most lateral glutamatergic
populations. (F) Cross-sections of the transformation of transcription factor
stripes from hindbrain (F1) to spinal cord (F5) in a cross of barhl2GFP and alx:
DsRed lines. White regions in panel 4 in BE indicate the location of the
transcription factor expression relative to the neurotransmitter stripes. All
cross-sections are maximum-intensity projections of 20- to 50-m volumes
from a confocal image stack of the entire hindbrain. (Scale bars, 20 m.)
neurons in the hindbrain stripes (n = 9). Engrailed-1 immu-
nostaining was located in a stripe just lateral to the alx (medial
glutamatergic) stripe and overlapped the most medial glycinergic
stripe (Fig. 1C). Colocalization indicated that most, if not all, of
the neurons in this glycinergic stripe express engrailed-1 (Fig. 1C,
panel 3; n = 4). Dbx1b is a transcription factor expressed in
progenitor cells that is also observed in differentiated neurons at
early times. In Fig. 1D, panel 1, a cross-section of dbx1b:GFP
expression in rhombomere 8 showed that these cells formed
a more laterally positioned stripe than both engrailed-1 and alx
stripes. The dorsalmost cells were present within the ventricular
zone; however, the majority of the expression was located in the
parenchyma and overlapped two stripesa glutamatergic stripe
just lateral to the engrailed-1 stripe (cells in blue/white), and
more lateral neurons in a region containing a glycinergic stripe
(the lateral population of green cells; Fig. 1D, panels 3 and 4;
n = 2). The dbx1b area thus encompassed at least a medial
Kinkhabwala et al.
glutamatergic domain and, more lateral, most likely, glycinergic
one. Neurons expressing the barhl2 transcription factor were
located in a band at the lateral edge of the hindbrain (Fig. 1E;
n = 10), overlapping a portion of a lateral crescent-like stripe of
glutamatergic neurons throughout hindbrain (Fig. 1E, panel 3).
In summary, neurons expressing particular transcription factors
are clustered into bands that align with neurotransmitter stripes
and, depending on the transcription factor, can be coextensive
with a transmitter stripe (alx, engrailed-1), overlap multiple
transmitter stripes (dbx1b), or overlap only a restricted, spatially
segregated portion of a stripe (barhl2).
We examined how the distribution of neurons expressing these
transcription factors changes from spinal cord (where they are also
expressed) into the hindbrain by examining optical cross-sections
at different rostrocaudal locations from confocal image stacks
from live sh in which neurons expressing two transcription factors
(alx and barhl2) were labeled in different colors. A dorsoventral
segregation of the different transcription factors in spinal cord
gradually changed at its rostral end into a mediolateral segregation
in hindbrain, indicating a topological transformation in expression
domains between the two regions (Fig. 1F).
Morphology of Neurons in the Transmitter Stripes. We used sto-
chastic expression of membrane targeted proteins (mMCherry,
mGFP, or Brainbow-1.1) to randomly label neurons in one color
in 5- to 6-dpf sh in which the transmitter stripes were labeled in
a different color. We then reconstructed the neurons in 3D and
sorted them by stripe membership. Reconstructed neurons from
four stripesthe medial glutamatergic stripe and the three gly-
cinergic stripesare shown in dorsal view in Fig. 2.
The morphology of the neurons differed systematically among
stripes. All 34 labeled medial glutamatergic stripe (alx) neurons
(12 shown in Fig. 2) imaged and reconstructed from hindbrain
rhombomeres 48 had ipsilaterally projecting axons that always
had a descending portion that often extended into spinal cord
(Fig. 2A); there was variation in whether rostral axonal branches
were evident. The medial (engrailed-1) glycinergic stripe neu-
rons were predominately cells with ipsilateral and ascending
NEUROSCIENCE
Fig. 2. Morphology of neurons in the neurotransmitter stripes. All panels
show horizontal projections of single-neuron reconstructions from different
sh sorted by stripe location. (A) Neurons in the medial glutamatergic stripe.
Their descending axons are truncated at the dashed line. (B) Neurons in the
most medial glycinergic stripe. (C) Neurons in the middle glycinergic stripe.
(D) Neurons in the lateral glycinergic stripe. Reconstructions are shown in
a dorsal view, with rostral to the top. In C and D, dotted lines indicate the
approximate position of the midline of the sh or, in one instance, the
continuation of an axon. In each panel, the left column illustrates the stripe
containing the cells. (Scale bars, 20 m.)
PNAS | January 18, 2011 | vol. 108 | no. 3 | 1165
350
axonal projections (13 of 15 neurons in rhombomeres 38; Fig.
2B) that sometimes projected beyond rostral hindbrain. All eight
neurons labeled in the middle (putative dbx1b) glycinergic stripe
(rhombomeres 68) had contralaterally extending axons with
both ascending and descending branches within hindbrain, often
of similar length (Fig. 2C). All seven neurons reconstructed from
the third, lateral glycinergic stripe (Fig. 2D) had both ipsilateral
and contralateral axonal projections that distinguished them
from those belonging to other stripes; there was, however, vari-
ation in the extent of ascending or descending projections on the
two sides of the hindbrain.
Although these 63 reconstructed single cells from four of the
transmitter stripes represent only a small fraction of the neurons
in hindbrain, they reveal a general pattern consistent with the
idea that these hindbrain stripes contain neurons with different
morphological features. This conclusion is also supported by
a series of backlling experiments designed to more specically
examine projection patterns within these transmitter stripes (26).

Age-Related Patterning of Neurons in the Stripes. We next exam-

ined neurons within a single stripe to determine if there was any
organization within stripes. We focused on the medial gluta-
matergic stripe of neurons expressing the alx transcription factor,
for which we had the most morphological data, transgenic lines
marking the stripe, and evidence that some neurons in the stripe
might be involved in swimming.
We examined whether neurons differed by age in an orderly
manner along the axis of a stripe by photoconverting the color
change protein Kaede, in an alx-Kaede transgenic line, at dif-
ferent time points during development to generate sh with the
earliest differentiating neurons labeled in red, and the latest
differentiating cells in green (Fig. 3 AD).
We found a broad and consistent pattern of age-related order
in which the oldest neurons were ventrally positioned and the
youngest ones populated more-dorsal regions within the alx
stripe throughout hindbrain, as in Fig. 3C. The exceptions were
two bands in which young alx neurons were ventrally positioned
in rhombomere 6, and one group of alx neurons that was located
medially and outside of the stripe region in rhombomere 8. The
broad age order was present across multiple sh photoconverted
and imaged at different times (n = 10, photoconverted between
24 and 55 hpf and imaged between 33 hpf and 8 dpf). In an
additional 22 experiments we crossed a sh expressing Kaede
under a general neuronal promoter (Huc:Kaede) into transgenic
lines with neurotransmitter and transcription factor stripes la-
beled in green to determine the location of the young and old
neurons in the various stripes (glyt2, n = 4; vglut, n = 5; dbx1b,
n = 1; barhl2, n = 6; Huc:Kaede alone, n = 6). These experi-
ments supported the conclusion that the pattern in which the
oldest neurons are located ventrally, with the younger ones
stacked above them, is present broadly across multiple hind-
brain stripes.
The processes of the alx neurons in the neuropil appeared
ordered by age as well. In the hindbrain, processes of older
ventral neurons (red and yellow) are located dorsally in the
neuropil, which lies just below the stripe of somata, whereas
younger, dorsal neurons (green) have processes located more
ventrally in the neuropil (Fig. 3E, panels 1 and 2). In contrast, in
the spinal cord, which contains processes of alx neurons from
both hindbrain and spinal cord, the older neuropil lies medial
and dorsal to the younger neuropil (Fig. 3F, panels 1 and 2). We
quantied the intensities of red and green uorescence within
different regions of the neuropil following photoconversion/im-
aging at different times and found the age-related order with the
neuropil to be present in all four cases quantied in spinal cord
and in multiple hindbrain rhombomeres (photoconvert 2 dpf/
image 3 dpf; photoconvert 33 hpf/image 2 dpf in spinal cord;
photoconvert 28 hpf/image 5 dpf; and photoconvert 2 dpf/image
4 dpf for R48 in hindbrain; n = 1 for each; Fig. 3, E2 and F2).
This age-related patterning of projections raises the possibility
that there might be age-related connectivity.
Fig. 3. Age-related patterning within the medial glutamatergic, alx hind-
brain stripe. (A) Location of images. (B) Timing of experiments. (C) Lateral
view of alx:Kaede expression in a photoconverted sh shown from hindbrain
through to spinal cord. White dashed lines indicate the rostrocaudal loca-
tions of the numbered cross-sections from different rhombomeres (R3R8)
and into spinal cord. (D, panels 13) Lateral views of hindbrain/spinal cord
regions of alx:Kaede transgenic sh photoconverted/imaged at different
times (timing shown in B). In this example we assigned colors so that green
cells on the gure had no red in them at the imaging time, and neurons
colored red on the gure are those with any red staining (they might also
have had green but it was removed to make the youngest, pure green cells
obvious). The yellow color here thus does not represent colabeling, but in-
stead red and green neurons that overlap in the z direction. (E, panel 1)
Reconstructed confocal cross-section shows that neuropil for older neurons
(red) in hindbrain tends to be dorsal to that for younger neurons (green). (E,
panel 2) Quantication of relative red/green expression at different dorso-
ventral locations within a cross-section of neuropil in rhombomere 7 from
a different sh than E, panel 1; ventral is at the bottom. Red expression is
shifted dorsally relative to green. (F) Photoconversion of an alx:Kaede
transgenic sh imaged in spinal cord also shows age-related separation in
the neuropil. In F, panel 1, a reconstructed confocal cross-section shows that
neuropil for older neurons (red) tends to be medial and dorsal to that of
younger neurons (green). (F, panel 2) Quantication of the dorsoventral
distribution of red and green expression in the neuropil. (Scale bars, panel 2
in E and F, 10 m; all other images, 20 m.)
Structural and Functional Organization of Neurons Within the alx
Stripe in Hindbrain. We suspected that some alx-positive neurons
in hindbrain would be active during swimming, based upon evi-
dence for involvement of spinal alx neurons in swimming (27,
28). To examine this, we developed an approach to do targeted
patch recording from neurons deep in the brain in vivo in 4- to 6-
dpf sh to allow us to explore their activity patterns. When we
patched from alx-positive neurons in the caudal hindbrain in
rhombomere 7 while recording from ventral roots to monitor the
motor pattern, we found that some (but not all) of the neurons
were rhythmically active following light or electrical stimulation,
in a pattern that matched the swimming motor output recorded
from the ventral roots (Fig. 4A, panels 1 and 2).
We then looked at structural variation of alx cells along the
axis of their stripe. We examined the morphology of 19 alx
neurons labeled stochastically with Brainbow-1.1m at different
positions along the axis of the stripe in different hindbrain
rhombomeres (Fig. 4B). We quantied the total axonal branch-
ing for a subset of 10 neurons that were located in the same

1166 | www.pnas.org/cgi/doi/10.1073/pnas.1012185108 Kinkhabwala et al.

351

Fig. 4. Structural and functional patterning within the alx stripe. (A) Patch
recording of an alx:GFP+ neuron in hindbrain segment 7 of a 5-dpf sh,
while simultaneously recording motor activity in a ventral root. (A, panel 1)
Cross-section view of the patched cell (red) within the alx stripe (green). (A,
panel 2) Recording from the alx cell on the left is shown below the simul-
taneous ventral root recording. (B) Variation in axonal projections into spinal
cord of hindbrain alx neurons. (B, panel 1) An example of a reconstruction of
a single alx neuron (red) within the alx:GFP transgenic line (lateral view). (A,
panel 2) Lateral views of 3D reconstructions of neurons from different
dorsoventral locations in the medial glutamatergic stripe in rhombomere 7,
with the most dorsal one at the top (normalized dorsoventral positions: 0.51,
0.28, and 0.16 from ventral edge of stripe, rostral to the left). (C) A plot of
the total 3D axonal length of 10 labeled alx neurons (two of them overlap,
shown with white circle) vs. their dorsal ventral location in the stripe. More
ventral neurons have systematically longer axonal length (P < 0.0001). Red,
green, and blue points correspond with the three neurons in B. (D) Plot of
the input resistance of a neuron vs. its dorsoventral location in a stripe.
More-dorsal neurons have systematically higher input resistances in both
more- (black dots, P < 0.05) and less-exposed brain preparations (gray
squares, P < 0.01). See text for further details. (E) Examples of calcium im-
aging of alx neurons from different locations in the stripe during different
frequencies of swimming in 5-dpf sh. (E, panels 13) Cross-sections images
of alx:DsRed sh with locations of the neurons (dots) shown relative to the
stripes. (E, panels 46) Calcium responses of the neurons on the left in two
example trials at swimming frequencies near those when the neurons are
rst recruited. (F) Plot of the minimum swimming frequency at which
a neuron responds vs. the dorsoventral location of the neuron, including
both alx:DsRed+ alx:DsRed neurons in the region of the alx stripe (P < 0.05
for each correlation). Neurons from Fig. 4B are shown in color. (Scale bars,
200 m.)
region (rhombomeres 68) of 6-dpf sh by reconstructing them
in three dimensions and measuring the total length of their axon.
A plot of location vs. axonal length for these neurons showed
a signicant (P < 0.0001) correlation between their total axonal
length and their location (Fig. 4C), with the length increasing
systematically, and nearly linearly, from dorsal to ventral along
the stripe axis. Tracking the growth of axons from dorsal and
ventral neurons in the alx stripe as they descend into spinal cord
indicates that the axons of ventral neurons grow along cord faster
after their extension than those of more dorsal cells, suggesting
that the differences in axonal extent are not just a result of dorsal
neurons having less time to grow, and raising the possibility that
the differences might persist later in life.
We also explored whether physiological properties might vary
systematically along the axis of a stripe. Using hyperpolarizing
current steps while patch recording from single alx-positive
neurons in ve dpf alx:GFP transgenic sh, followed by labeling
and 3D reconstruction, we examined whether the input re-
sistance of alx neurons varied systematically with location in
a stripe. We used both an exposed-brain preparation as well as
a more intact preparation and sampled neurons from the hind-
Kinkhabwala et al.
brain in R78 across the dorsoventral extent of the alx stripe.
Neurons positioned very dorsally had small amplitude, relatively
long duration action potentials, a depolarized resting membrane
potential, and very high input resistances in both preparations,
suggesting that they were very young, in accord with the Kaede
photoconversion data (see above). Interneurons above 70% of
the way up the stripe showed persistent activity that was not
correlated with behaviors such as swimming. Consequently, we
focused our input resistance and functional analysis on the
ventral 70% of the stripe, which included older neurons that
were already clearly incorporated into networks based upon their
rhythmic activation during swimming (see below).
Neurons with the lowest input resistance values were consis-
tently located in the most ventral stripe regions at 5 dpf, both in
the preparations with massive exposure of the brain (black dots
on Fig. 4D) and those with reduced exposure (gray squares on
Fig. 4D). There was a signicant correlation between the position
of a neuron along the stripe axis and its input resistance within
both preparations with either massive or reduced exposure of the
brain (Pearson correlation: P < 0.01 for reduced exposure, P <
0.05 for massive exposure, P < 0.05 for both together). The input
resistance was lowest at the bottom of the stripe and increased in
more dorsal locations in both cases, although the measurements
from the more exposed preparation were more scattered, probably
because of damage during exposure.
The presence of an orderly pattern of structural and physio-
logical properties along the stripe axis by age raised the question
of whether the neurons were recruited during behavior in an
orderly way along a stripe. To explore the involvement of neu-
rons at different dorsoventral locations in the alx stripe at dif-
ferent frequencies/speeds of swimming in a minimally invasive
way, we used calcium imaging with Oregon Green BAPTA-1
dextran (Invitrogen) electroporated into the neurons in the alx:
DsRed transgenic line. We recorded from ventral roots to
monitor the frequency of swimming produced by light or elec-
trical stimulation and determined, based upon the calcium re-
sponse, the lowest swimming frequency at which a neuron was
activated (25). We subsequently collected confocal image stacks
through the hindbrain in the region of labeled neurons so that we
could measure the position of the imaged cells relative to the axis
of the stripe.
NEUROSCIENCE
We collected many trials (averaging about 100) over a range
of frequencies of swimming for each neuron to determine its
minimum recruitment frequency. Some examples of calcium
responses of neurons at different dorsoventral locations are
shown in Fig. 4E. In these examples, the top neuron exhibited
a uorescence increase in a bout with a peak frequency of 24 Hz,
but not in one at 22 Hz; the middle neuron at 28Hz, but not
26Hz; and the bottom cell at 31 but not 29 Hz. We obtained
recruitment patterns and positions for 29 neurons from 20 sh.
Eleven of the neurons expressed the alx transcription factor and
18 did not, although the 18 non-alx neurons were in the general
region of the alx stripe, so their position along the stripe could be
measured. A plot of recruitment frequency vs. position for both
alx and non-alx neurons, in Fig. 4F, showed a consistent pattern
in which more-dorsal, younger cells were recruited at the lowest
swimming frequencies and increasingly more-ventral neurons
were recruited at increasingly higher swimming frequencies (P <
0.05 for both). Thus, the neurons are recruited from dorsal to
ventral along the axis of the stripe as the frequency of swimming
increases, indicating a systematic relationship between position,
age, and recruitment that maps onto the axis of the alx stripe and
extends to other non-alx neurons in hindbrain as well.
Discussion
Our experiments reveal that the striking organization into trans-
mitter stripes in hindbrain reects a broad patterning of neu-
rons by cell type, morphology, age, projections, cellular properties,
and activity patterns (summarized in Fig. 5). The sh we studied
are freely swimming animals with functional motor networks for
swimming, escape, jaw, and eye movements, all located in a hind-
PNAS | January 18, 2011 | vol. 108 | no. 3 | 1167
352

Fig. 5. Summary of the structural and functional rules governing hindbrain
patterning. (A) Interleaved neurotransmitter stripes overlap with transcrip-
tion factor expression patterns. Different neuronal morphologies (Right) are
associated (by matching colors here) with the different stripes. (B) Within an
individual stripe (alx stripe, Left), the somata of neurons are ordered along
the axis of the stripe by age, the location of their processes in the neuropil,
and their total axonal length. This corresponds to systematic changes in in-
put resistance and recruitment along the axis of the stripe.
brain with a strikingly regular organization. The transmitter stripes
extend throughout the hindbrain and across the well-studied
rhombomeres (2, 11, 13, 15, 29), through regions containing many
different networks. The implication is that the many neural circuits
in hindbrain are built from a set of neuronal types, which might be
viewed as circuit components, with particular morphologies, neu-
rotransmitters, and transcription factor phenotypes. The stripes
represent these different neuronal types. Within a given stripe, the
neurons are arranged in an orderly way by structural and functional
properties (and, importantly, age) so that those with more exten-
sive projections and higher thresholds for activation, which are
involved in the fastest movements, are at the bottom of a stripe,
with the neurons involved in increasingly rened movements
stacked in order above them. We infer that hindbrain networks are
constructed in an orderly manner by drawing cells from organized
stripes, much like selecting components from a set of parts where
the parts are arranged in order according to their structural or
functional properties.
Though the larval sh has nearly all of the same motor
behaviors as an adult, the brain of the adult is considerably larger
and the neurons are clustered into nuclei located among the ber
tractsan arrangement much different from the simple pat-
terning in the young sh. This raises the question of the re-
lationship between the networks in the young sh and those in
the adult. A wholesale reorganization of the connectivity of the
networks with growth seems unlikely, because the patterns of
motor output in the young sh parallel those in adults (e.g., the
alternation and rostrocaudal delays characteristic of swimming
are present throughout life) (3032). A reconstruction of the
networks would mean dispensing with a motor circuitry that al-
ready worked. The more likely possibility is that the essentials of
the larval pattern of network organization and recruitment are
retained even as the neurons disperse to reorganize into nuclei.
This would imply that circuits are set up early when brain pat-
terning is simple, but that the somata then migrate with growth
to form nuclei, obscuring a simple underlying network pattern-
ing. If so, we predict that if we could track neurons as the brain
differentiated further, the recruitment order and the basic pat-
terns of connectivity established early would be retained later in
life, after migration, and be related to the age of the neuron,
which corresponded to its location in the younger sh. This
migration may occur to minimize the extent of axonal and den-
dritic arbors in a much larger brain, given the evidence for
placement of neurons to minimize wiring in other systems (33).
This early organization in zebrash probably extends to the
hindbrain of other vertebrates. Longitudinal bands of some of
the transcription factors we studied, and others, extend through
the hindbrain of frogs, chicks, and mice, and their locations
1168 | www.pnas.org/cgi/doi/10.1073/pnas.1012185108
relative to the midline parallel the patterns in zebrash (1214,
16, 34, 35). Morphological data from backlling studies in em-
bryonic chicks indicates that neurons with different morpholog-
ical features occupy systematic locations relative to the midline
(15), consistent with an early orderly disposition of cell types.
Though these studies provide structural indications of a pat-
terning like that in sh, there is no information about the func-
tional organization at early stages to allow comparison with the
zebrash patterning. We would predict, however, that networks
are set up when the patterning is simple, with the later func-
tion of the neurons mapping systematically onto their age as
in zebrash.
The organization of hindbrain has clear similarities to the
patterning in spinal cord, suggesting that the two regions giving
rise to all of the motor output from the nervous system have
elements of the same structural and functional network plan (17
22, 25, 27, 28). Though there are undoubtedly changes with
growth, nervous system plasticity, and divergence in the networks
for specic tasks, the basic features of structural and functional
organization that we describe may lie at the foundation of the
construction of the many sensory-motor circuits throughout the
hindbrain, which produces a much broader range of sensory-
motor outputs than spinal cord. The challenge now is to un-
derstand how many different networks in hindbrain are con-
structed from this basic ground plan and how the early pattern is
reorganized as the brain continues to grow into its adult form.
We explore one of these specialized hindbrain networks, and its
relationship to the patterning, in the study by Koyama et al. (26).
Experimental Procedures
Fish Care. All experiments were performed on zebrash (Danio rerio) be-
tween 1 and 8 dpf obtained from a laboratory stock of wild-type and
transgenic adults. All procedures conform to the National Institutes of
Health guidelines regarding animal experimentation and were approved by
Cornell Universitys Institutional Animal Care and Use Committee.
Immunostaining. Standard whole-mount antibody staining procedures were
used as described previously (36, 37).
Stochastic Labeling. Stochastic labeling was performed as described previously
(23).
Transgenic Lines. The transgenic lines we used included ones described in
prior studies (24, 25, 38), as well as two new ones with the promoters dbx1b
and barhl2. These new lines, Tg(dbx1b:GFP) and Tg(barhl2:GFP), were con-
structed with the BACs zK17G17 and zC15L16 using a previously described
method (38). The detailed approaches for generating these lines are de-
scribed elsewhere.
Confocal Imaging. Confocal imaging was performed as described previously
(28, 39).
Colocalization. Colocalization was assessed using the colocalization add-on in
the Imaris software package (Bitplane).
Neuronal Tracing and Location Measurements. Neurons were traced using the
lament reconstruction feature provided in Imaris.
Photoconversion and Analysis. Huc:Kaede and alx:Kaede transgenic embryos
(38, 40) were illuminated with UV light from a mercury bulb source for 10
40 s within their chorion. Immediately afterward, photoconversion was
conrmed by observation of the presence of red expression and an absence
of green expression. Photoconverted sh were then kept at 28.5 C in
a light-tight container until the day of imaging.
Targeted Whole-Cell Patching of alx Neurons in an Exposed Brain Preparation for
Input Resistance Measurements. Larvae were anesthetized in 3-aminobenzoic
acid ethyl ester (0.02% in HBSS) and then immobilized using -bungarotoxin
(Sigma-Aldrich; 0.1% in HBSS). To measure input resistance of alx neurons in
hindbrain, we initially used a dissection procedure similar to one described
previously (41). We later switched to a less-extensive exposure of the brain,
leaving more of the head, including the eyes, intact to minimize potential
Kinkhabwala et al.
353
damage to the neurons. After the electrophysiological measurements, the
preparation was xed with 4% formaldehyde with the pipette in place to
avoid the movement of the cell body during the retraction of pipette. Then a z
stack was acquired with a confocal microscope (LSM 510 META; Zeiss) for the
measurement of the soma position relative to the alx stripe.
In Vivo Whole-Cell Recordings in the Hindbrain. Whole-cell recordings were
done in current clamp mode in 5-dpf alx:GFP nacre transgenic larvae using
methods similar to those described previously (25, 30, 41).
Calcium Imaging. Transgenic 4-dpf alx:DsRed Casper sh (42) were rst
anesthetized using 3-aminobenzoic acid ethyl ester (MS-222, 0.02% in HBSS)
and embedded in low-melting-point agarose (1.6% in HBSS; Sigma). A patch
electrode (510 M resistance) was lled with 20% Oregon Green BAPTA-1
(10,000 Mr; Invitrogen/Molecular Probes), and the indicator was electro-
porated along the dorsoventral axis of the alx stripe in caudal hindbrain by
using a single-cell Axoporator (4 V, 20-ms duration square pulse; Molecular
Devices). Larvae were then removed from the agarose, placed in a Petri dish
containing HBSS, and stored in an incubator (28.5 C) overnight.
1. Garcia-Campmany L, Stam FJ, Goulding M (2010) From circuits to behaviour: Motor
networks in vertebrates. Curr Opin Neurobiol 20:116125.
2. Tmpel S, Wiedemann LM, Krumlauf R (2009) Hox genes and segmentation of the
vertebrate hindbrain. Curr Top Dev Biol 88:103137.
3. Sparks DL (2002) The brainstem control of saccadic eye movements. Nat Rev Neurosci
3:952964.
4. Smith JC, Abdala AP, Rybak IA, Paton JF (2009) Structural and functional architecture
of respiratory networks in the mammalian brainstem. Philos Trans R Soc Lond B Biol
Sci 364:25772587.
5. Grillner S, et al. (1995) Neural networks that co-ordinate locomotion and body
orientation in lamprey. Trends Neurosci 18:270279.
6. Smith JC, Ellenberger HH, Ballanyi K, Richter DW, Feldman JL (1991) Pre-Btzinger
complex: A brainstem region that may generate respiratory rhythm in mammals. Science
254:726729.
7. Roberts A, Li WC, Soffe SR, Wolf E (2008) Origin of excitatory drive to a spinal
locomotor network. Brain Res Brain Res Rev 57:2228.
8. Aksay E, et al. (2007) Functional dissection of circuitry in a neural integrator. Nat
Neurosci 10:494504.
9. Dubuc R, et al. (2008) Initiation of locomotion in lampreys. Brain Res Brain Res Rev 57:
172182.
10. Faber DS, Fetcho JR, Korn H (1989) Neuronal networks underlying the escape response
in goldsh. General implications for motor control. Ann N Y Acad Sci 563:1133.
11. Moens CB, Prince VE (2002) Constructing the hindbrain: Insights from the zebrash.
Dev Dyn 224:117.
12. Moreno N, Bachy I, Rtaux S, Gonzlez A (2005) LIM-homeodomain genes as territory
markers in the brainstem of adult and developing Xenopus laevis. J Comp Neurol 485:
240254.
13. Schubert FR, Dietrich S, Mootoosamy RC, Chapman SC, Lumsden A (2001) Lbx1 marks
a subset of interneurons in chick hindbrain and spinal cord. Mech Dev 101:181185.
14. Cepeda-Nieto AC, Pfaff SL, Varela-Echavarra A (2005) Homeodomain transcription
factors in the development of subsets of hindbrain reticulospinal neurons. Mol Cell
Neurosci 28:3041.
15. Clarke JD, Lumsden A (1993) Segmental repetition of neuronal phenotype sets in the
chick embryo hindbrain. Development 118:151162.
16. Gray PA (2008) Transcription factors and the genetic organization of brain stem
respiratory neurons. J Appl Physiol 104:15131521.
17. Briscoe J, Pierani A, Jessell TM, Ericson J (2000) A homeodomain protein code species
progenitor cell identity and neuronal fate in the ventral neural tube. Cell 101:
435445.
18. Goulding M (2009) Circuits controlling vertebrate locomotion: Moving in a new
direction. Nat Rev Neurosci 10:507518.
19. Ericson J, et al. (1997) Pax6 controls progenitor cell identity and neuronal fate in
response to graded Shh signaling. Cell 90:169180.
20. Osumi N, et al. (1997) Pax-6 is involved in the specication of hindbrain motor neuron
subtype. Development 124:29612972.
21. Placzek M, Yamada T, Tessier-Lavigne M, Jessell T, Dodd J (1991) Control of
dorsoventral pattern in vertebrate neural development: Induction and polarizing
properties of the oor plate. Development 2 (Suppl 2):105122.
22. Takahashi M, Osumi N (2002) Pax6 regulates specication of ventral neurone
subtypes in the hindbrain by establishing progenitor domains. Development 129:
13271338.
Kinkhabwala et al.
The following day, we imaged the calcium indicator in hindbrain on a Zeiss
LSM 510 inverted confocal microscope while recording from a ventral root
using methods similar to those applied previously (25, 27).
Analysis of calcium imaging and ventral root recordings was performed
using custom-written MATLAB software with an approach similar to McLean
et al. (25).
Image stacks were then reconstructed in 3D using Imaris software (Bit-
plane). Neurons that were active during the experiment were identied. The
length of the left and right stripes (at the position of the cell body) and the
location of the neuron were each measured three times and averaged to
obtain the location of the neurons in the stripe.
More details of methods are provided in SI Experimental Procedures.
ACKNOWLEDGMENTS. We thank David McLean for comments on the
manuscript and Lindsay Heller for maintaining the many lines of sh. This
work was supported by National Science Foundation Integrative Graduate
Education and Research Traineeship Grant 0333366 (to A.K.), National
Institutes of Health Training Grant T32 GM007469 (to A.K. and M.R), the
Ministry of Education, Science, Technology, Sports and Culture of Japan
(S.-i.H.), the Japan Society for the Promotion of Science (M.K.), and National
Institutes of Health Grant NS26539 (to J.F.).
23. Higashijima S, Schaefer M, Fetcho JR (2004) Neurotransmitter properties of spinal
interneurons in embryonic and larval zebrash. J Comp Neurol 480:118.
24. Bae YK, et al. (2009) Anatomy of zebrash cerebellum and screen for mutations
affecting its development. Dev Biol 330:406426.
25. McLean DL, Fan J, Higashijima S, Hale ME, Fetcho JR (2007) A topographic map of
recruitment in spinal cord. Nature 446:7175.
26. Koyama M, et al. (2010) Mapping a sensory-motor network onto a structural and
functional ground plan in the hindbrain. Proc Natl Acad Sci USA, in press.
27. McLean DL, Fetcho JR (2009) Spinal interneurons differentiate sequentially from
those driving the fastest swimming movements in larval zebrash to those driving the
slowest ones. J Neurosci 29:1356613577.
28. McLean DL, Masino MA, Koh IY, Lindquist WB, Fetcho JR (2008) Continuous shifts in
the active set of spinal interneurons during changes in locomotor speed. Nat Neurosci
11:14191429.
29. Heyman I, Kent A, Lumsden A (1993) Cellular morphology and extracellular space at
rhombomere boundaries in the chick embryo hindbrain. Dev Dyn 198:241253.
30. Masino MA, Fetcho JR (2005) Fictive swimming motor patterns in wild type and
mutant larval zebrash. J Neurophysiol 93:31773188.
31. Liu DW, Westereld M (1988) Function of identied motoneurones and co-ordination
of primary and secondary motor systems during zebra sh swimming. J Physiol 403:
7389.
32. Fetcho JR, Svoboda KR (1993) Fictive swimming elicited by electrical stimulation of the
midbrain in goldsh. J Neurophysiol 70:765780.
NEUROSCIENCE
33. Chen BL, Hall DH, Chklovskii DB (2006) Wiring optimization can relate neuronal
structure and function. Proc Natl Acad Sci USA 103:47234728.
34. Passini MA, Raymond PA, Schechter N (1998) Vsx-2, a gene encoding a paired-type
homeodomain, is expressed in the retina, hindbrain, and spinal cord during goldsh
embryogenesis. Brain Res Dev Brain Res 109:129135.
35. Storm R, et al. (2009) The bHLH transcription factor Olig3 marks the dorsal
neuroepithelium of the hindbrain and is essential for the development of brainstem
nuclei. Development 136:295305.
36. Higashijima S, Masino MA, Mandel G, Fetcho JR (2004) Engrailed-1 expression marks
a primitive class of inhibitory spinal interneuron. J Neurosci 24:58275839.
37. McLean DL, Fetcho JR (2004) Ontogeny and innervation patterns of dopaminergic,
noradrenergic, and serotonergic neurons in larval zebrash. J Comp Neurol 480:
3856.
38. Kimura Y, Okamura Y, Higashijima S (2006) alx, a zebrash homolog of Chx10, marks
ipsilateral descending excitatory interneurons that participate in the regulation of
spinal locomotor circuits. J Neurosci 26:56845697.
39. Hale ME, Ritter DA, Fetcho JR (2001) A confocal study of spinal interneurons in living
larval zebrash. J Comp Neurol 437:116.
40. Sato T, Takahoko M, Okamoto H (2006) HuC:Kaede, a useful tool to label neural
morphologies in networks in vivo. Genesis 44:136142.
41. Drapeau P, Ali DW, Buss RR, Saint-Amant L (1999) In vivo recording from identiable
neurons of the locomotor network in the developing zebrash. J Neurosci Methods
88:113.
42. White RM, et al. (2008) Transparent adult zebrash as a tool for in vivo transplantation
analysis. Cell Stem Cell 2:183189.
PNAS | January 18, 2011 | vol. 108 | no. 3 | 1169
354
Mapping a sensory-motor network onto a structural
and functional ground plan in the hindbrain
Minoru Koyamaa, Amina Kinkhabwalaa,b, Chie Satouc,d, Shin-ichi Higashijimac,d, and Joseph Fetchoa,1
a
Department of Neurobiology and Behavior, Cornell University, Ithaca, NY 14853; bPrinceton Neuroscience Institute, Princeton University, Princeton, NJ 08544;
c
Okazaki Institute for Integrative Bioscience, National Institute for Physiological Sciences, National Institutes of Natural Sciences, Okazaki, Aichi 444-8787,
Japan; and dDepartment of Physiological Sciences, Graduate University for Advanced Studies (SOKENDAI), Okazaki, Aichi 444-8585, Japan
Edited by Lynn T. Landmesser, Case Western Reserve University, Cleveland, OH, and approved December 9, 2010 (received for review August 17, 2010)
The hindbrain of larval zebrash contains a relatively simple
ground plan in which the neurons throughout it are arranged
into stripes that represent broad neuronal classes that differ in
transmitter identity, morphology, and transcription factor expres-
sion. Within the stripes, neurons are stacked continuously accord-
ing to age as well as structural and functional properties, such as
axonal extent, input resistance, and the speed at which they are
recruited during movements. Here we address the question of
how particular networks among the many different sensory-motor
networks in hindbrain arise from such an orderly plan. We use
a combination of transgenic lines and pairwise patch recording to
identify excitatory and inhibitory interneurons in the hindbrain
network for escape behaviors initiated by the Mauthner cell. We
map this network onto the ground plan to show that an individual
hindbrain network is built by drawing components in predictable
ways from the underlying broad patterning of cell types stacked
within stripes according to their age and structural and functional
properties. Many different specialized hindbrain networks may
arise similarly from a simple early patterning.
development | locomotion | neuronal circuit | Mauthner neuron
T he vertebrate hindbrain contains many different sensory-
motor networks that control movements of structures in the
body and the head. Our recent work shows that despite the di-
verse networks it contains, there is a remarkably orderly pat-
terning of neurons that extends throughout the hindbrain in
young zebrash, across the well-studied segments, or rhombo-
meres (14). This pattern consists of a series of neurotransmitter
stripes that represent broad neuronal classes that differ in
transmitter identity, morphology, and transcription factor ex-
pression. Within these stripes, neurons are stacked continuously
according to age as well as structural and functional character-
istics, such as axonal extent, input resistance, and the speed at
which they are recruited during movements.
The presence of such an orderly array of neurons suggests that
networks are built by drawing components from particular stripes
that contain cell types with the required structure and transmitter
phenotype, and from the particular position within a stripe oc-
cupied by neurons with the appropriate functional properties,
much like selecting parts from a catalog. Here we test this pos-
sibility through a study of the hindbrain network for the escape
behavior initiated by the Mauthner cell (M-cell) (58). We con-
clusively identify neurons in the hindbrain escape network of the
M-cell in zebrash and map each cell type onto the arrangement
into the stripes that we describe in the previous work. Our work
reveals that this network is built from neurons in predictable
locations that are consistent with the idea that the stripe pat-
terning in hindbrain represents an orderly array of neurons from
which network components are drawn. Many different hindbrain
networks probably arise in a similar way by drawing neurons from
a shared structurally and functionally ordered set of parts.
Results
Retrograde Labeling of Interneurons. We set out to identify the
locations of neuronal candidates for a role in the Mauthner es-
cape network in zebrash by backlling them by injections of
11701175 | PNAS | January 18, 2011 | vol. 108 | no. 3
uorescent dyes (Alexa 647 or Texas Red) targeted to regions of
the M-cell on the basis of the innervation patterns of neurons in
the network known from goldsh, which is diagrammed in Fig. 1
(5). These backlls were performed in a transgenic line, Tg(glyt2:
GFP), in which the neurons in glycinergic stripes can be visual-
ized by GFP expression (Fig. 1 A1A3). Here we organize these
backlling data from sensory processing to motor output com-
ponents of the escape circuit.
A balance between direct sensory excitation of the M-cell and
a feed-forward inhibition regulates Mauthner ring and the initi-
ation of the escape response (Fig. 1A4) (5, 9, 10). To localize po-
tential glycinergic feedforward inhibitory neurons, we electro-
porated a red dye onto the cell body of one of the two M-cells in the
line with GFP labeled glycinergic neurons [Tg(glyt2:GFP)] and
looked for contralateral neurons that were both red (backlled)
and green (glycinergic/GFP) (Fig. 1B1; n = 8 sh). Nearly all of the
colabeled neurons were located in the most lateral of the three
glycine stripes (Fig. 1 B2 and B3; 64 of 66 cells) in hindbrain seg-
ment (rhombomere) 4 (Fig. 1 B2B4), the segment that contains
the M-cell body. The colabeled neurons were largely located in the
ventral portion of the stripe (Fig. 1B3; 57 of 64 cells).
Another less-well-studied interneuron thought to play a role in
sensory processing in the network is the spiral ber neuron (Fig.
1 A4 and C) (11, 12). To identify potential spiral ber neurons in
zebrash, we electroporated Alexa 647 or Texas Red into the
axon cap of the M-cell in Tg(GlyT2:GFP) sh and screened for
contralateral GFP-negative neurons that were lled with the red
dye (Fig. 1C1; n = 4 sh). Almost all of these were located in
rhombomere 3 between the medial and middle of the three
glycinergic stripes, at their ventral ends (Fig. 1 C2 and C3; 67 of
68 cells). The middle of the three glutamatergic stripes is located
in this region between the medial and the middle glycinergic
stripes, suggesting that the spiral ber neurons are located in the
ventral part of the middle glutamatergic stripe.
The outputs of the M-cell in hindbrain are relayed to other
neurons via an interneuron called a cranial relay neuron that is
monosynaptically activated by the M-cell (1315). To locate
cranial relay (also called T-reticular) neurons in zebrash, we
backlled from the medial longitudinal fasciculus in rhombo-
mere 8, where the axons of these neurons run, and looked for
red, GFP-negative neurons caudal to rhombomere 6 on the
contralateral side in the Tg(glyt2:GFP) line (Fig. 1D1). Most of
the backlled neurons were located ventrally near the middle
and the lateral glycinergic stripes (Fig. 1 D2D8; 48 of 51 cells).
The M-cell res single action potentials in goldsh because
multiple ring is blocked by feedback inhibition (Fig. 1A4) (15).
We located potential feedback inhibitory neurons in zebrash
Author contributions: M.K., A.K., and J.F. designed research; M.K. and A.K. performed
research; C.S. and S.-i.H. contributed new reagents/analytic tools; M.K., A.K., and J.F.
analyzed data; and M.K. and J.F. wrote the paper.
The authors declare no conict of interest.
This article is a PNAS Direct Submission.
Freely available online through the PNAS open access option.
1
To whom correspondence should be addressed. E-mail: jrf49@cornell.edu.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1012189108/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1012189108
355

Fig. 1. Retrograde labeling of prospective interneurons in the Mauthner
circuit. (A) Location of the M-cell relative to glycinergic neurons in the hind-
brain. (A1) Location of images on right. (A2) Dorsal view of a Tg(glyt2:GFP) line
at 5 dpf with reticulospinal neurons backlled from the spinal cord with Texas
Red. (A3) Cross-section view of the M-cell and the three glycinergic stripes at
rhombomere 4. White arrowheads mark the glycinergic stripes. (A4) Circuit
diagram of the Mauthner circuit from goldsh. Excitatory neurons, white;
inhibitory neurons, black. (B) Retrograde labeling of potential feedforward
inhibitory neurons. (B1) Experimental diagram. (B2) Cross-section view of the
backlled neurons at rhombomere 4 in the Tg(glyt2:GFP) line at 5 dpf. Cola-
beled voxels are orange. Arrowheads mark glycinergic stripes. (B3) Cross-
section from registrations showing the glycinergic neurons from four example
sh, with different colored spheres for each. (B4) Top-down view of neurons in
B3. White dashed line marks the M-cell. (C) Retrograde labeling of prospective
spiral ber neurons. (C1) Experimental diagram. (C2) Cross-section at rhom-
bomere 3 in the Tg(glyt2:GFP) line at 5 dpf. Voxels that only contain the far red
dye used to backll are shown in red. (C3) Cross-section of the registered
reconstructions for three different sh, as in B3. (C4) Top-down view of neu-
rons in C3. M-cell marked by white dashed line. (D) Retrograde labeling of
potential cranial relay neurons. (D1) Experimental diagram. (D2D7) Back-
lled neurons from four different sh shown in registered cross sections in
rhombomere 7/8. Arrowheads mark medial and middle glycinergic stripes.
(D8) Top-down view of the neurons in D2D7. White lines mark corresponding
cross sections. (E) Retrograde labeling of prospective feedback inhibitory
neurons. (E1) Experimental diagram. (E2) Cross-section of glycinergic neurons
at rhombomere 5 in the Tg(glyt2:GFP) line at 5 dpf. Orange indicates colab-
eling with GFP and the far red dye. White arrowhead marks medial glycinergic
neurons. (E3) Cross-section of the registered reconstructions from four sh.
(E4) Top-down view of the same neurons. r/c, rostral and caudal; d/v, dorsal and
ventral; m/l, medial and lateral in this and later gures. (Scale bars, 20 m.)
Koyama et al.
larvae by electroporating red dye into the axon cap (initial seg-
ment) of the M-cell in the transgenic line containing GFP-
labeled glycinergic neurons and then searching for the ipsilateral
glycinergic neurons that were both red and green (Fig. 1E1; n =
4 sh). All of the colabeled glycinergic neurons ipsilateral to the
backll were located ventrally among the most medial glycinergic
neurons (Fig. 1 E2 and E3; 15 neurons) and were in rhombomere
5, caudal to the M-cell (Fig. 1E4).
This backlling served to identify candidate neurons in the M-
cell network for patch recording, so we could both ll them with
dye to examine their structure relative to neurons in the associated
stripes and explore their connectivity and pharmacology in pair-
wise patch recording experiments to verify their roles in the
network and map rmly identied neurons onto the hindbrain
patterning.
Feedforward Glycinergic Neurons. To study the structure and
connectivity of putative feedforward inhibitory neurons, we
conducted a series of pairwise patch recordings from the M-cell
and from GFP-positive interneurons in the Tg(glyt2:GFP) line
that were located in the region of the candidate feedforward
neurons identied by backlling (Fig. 2A1). We found 17 glyci-
nergic neurons that when activated produced inhibitory post-
synaptic potentials (IPSPs) in the ipsilateral M-cell but were not
activated in response to ring of the M-cell by current injection
(Fig. 2 A2 and A3). Activation of the M-cell did lead to weak
subthreshold PSPs in these neurons (Fig. 2A3; n = 17 pairs). The
pathway mediating these PSPs and the role of these connections
onto the putative feedforward neurons are unknown. Firing the
glycinergic neurons led to short-latency IPSPs (0.32 0.10 ms,
n = 17 pairs) in the M-cell, which were most visible when the
membrane potential of M-cell was depolarized or hyperpolarized
(Fig. 2A2), suggesting that the reversal potential was close to the
resting membrane potential of the M-cell (79.5 1.8 mV, n =
17, junction potential corrected). The IPSPs were blocked by
a glycinergic blocker (Fig. 2A2; strychnine, 1 M, n = 10 pairs).
The soma of the glycinergic neuron whose physiology is shown in
Fig. 2 A2A4 was located in the ventral part of the lateral gly-
cinergic stripe, close to the lateral dendrite of the M-cell (Fig. 2
A5 and A6). Consistent with the short latency of IPSPs observed
in the M-cell, the axon coming from this neuron had puncta
NEUROSCIENCE
directly apposed to the cell body of the M-cell. This neuron also
had both an ipsilateral descending projection as well as a pro-
jection to the contralateral hindbrain that approached the other
M-cell and then extended further rostrocaudally from there.
To represent the location of all of the feedforward glycinergic
neurons identied physiologically, we registered the confocal im-
age stacks from different sh onto a template image stack of the Tg
(glyt2:GFP) line. As shown in Fig. 2 B1 and B2, all of the func-
tionally identied feedforward inhibitory neurons were located in
the ventral part of the lateral glycinergic stripe at rhombomere 4.
To conrm that the glycinergic neurons in this region actually
receive sensory inputs, we used dye electroporations into auditory
and facial nuclei and found that they innervate the region con-
taining cell bodies of the identied glycinergic neurons, although
their afferent terminations were somewhat separate within this
region. We then tested whether the glycinergic neurons receive
excitatory input from the auditory nerve by stimulating the audi-
tory nerve while doing pairwise patch recordings (n = 6 pairs; Fig.
2A4) from the M-cell and the interneuron. The stimulus evoked an
action potential in these glycinergic neurons at strengths at which
the M-cell only showed subthreshold PSPs (4/6), indicating an
input pattern consistent with a feedforward identity.
We reconstructed all of the recorded neurons to examine their
ipsilateral and contralateral projections and then quantied the
extent of these projections in rostral and caudal directions (Fig.
2C). All of the neurons had ipsilateral descending and/or as-
cending projections and axons that crossed to the contralateral
hindbrain, where they formed relatively long ascending and/or
descending branches (n = 17). All of these neurons also had an
axon close to the contralateral M-cell, suggesting bilateral M-cell
PNAS | January 18, 2011 | vol. 108 | no. 3 | 1171
356
Fig. 2. Feedforward inhibitory neurons. (A1) Pairwise whole-cell recording
of a GFP-positive feedforward inhibitory neuron and the ipsilateral M-cell
from a Tg(glyt2:GFP) relaxed line at 4 dpf. (A2) Firing the feedforward
neuron led to PSPs (average of ve traces) in the M-cell that inverted below
threshold (black) and were blocked by strychnine (1 M, gray). (A3) Response
of the feedforward inhibitory neuron to single spikes in the M-cell. (A4)
Eighth nerve stimulation led to ring of the interneuron and an EPSP in the
M-cell. (A5) Top-down view of the feedforward inhibitory neuron (green)
and the M-cell (red) from A2 relative to the glycinergic neurons (blue). (A6)
Cross-section of the same neurons, in rhombomere 4. Arrowheads mark
glycinergic stripes. (B1 and B2) Location in top-down (B1) and cross-section
views (B2) of 12 registered feedforward inhibitory neurons (orange spheres)
identied by pairwise recordings. Arrowheads mark the glycinergic stripes.
(C) Morphology of feedforward inhibitory neurons (16 cells). One cell is
excluded owing to poor lling. (C1) Box-and-whisker plots showing the ex-
tent of axonal projections in four directions: ipsilateral ascending (ia), ipsi-
lateral descending (id), contralateral ascending (ca), and contralateral
descending (cd). The whiskers extend to the most extreme data point, which
is no more than 1.5 times the interquartile range from the box. Asterisk
marks the line for the neuron in C2. (C2) Example reconstruction; dotted line
marks the midline. (Scale bars, 20 m.)
connections as in goldsh. The inhibitory inputs to the contra-
lateral M-cell were conrmed by pairwise patch recordings from
an interneuron and a contralateral M-cell in separate experi-
ments (n = 3 pairs). We conclude that these are indeed feed-
forward glycinergic neurons; they are located ventrally among
the oldest cells in the lateral glycinergic stripe, and they have the
ipsilateral and contralateral projections characteristic of neurons
stochastically labeled from many different locations in this stripe
in our previous work (2).
Spiral Fiber Neurons. Paired patch recordings from the M-cell and
contralateral GFP-negative cells in the Tg(glyt2:GFP) in the re-
gion containing putative spiral ber neurons led to the identi-
cation of 10 GFP-negative neurons that had excitatory con-
nections with the contralateral M-cell, characteristic of spiral ber
neurons (Fig. 3A1) (11). The majority of these neurons produced
biphasic excitatory postsynaptic potentials (EPSPs) in the con-
tralateral M-cell (Fig. 3A2; n = 9 of 10 pairs), with a short latency
(peak to peak, 0.24 0.03 ms, n = 10 pairs). The later component
in the EPSP was blocked by glutamatergic blockers [Fig. 3A2;
10 M 2,3-dihydroxy-6-nitro-7-sulfamoyl-benzo[f]quinoxaline-2,3-
dione (NBQX) and 50 M D-()-2-Amino-5-phosphonopentanoic
acid (D-AP5), n = 4 pairs), suggesting this slower component was
glutamatergic. Injection of negative current pulses into one cell led
to a hyperpolarizing response in the other neuron (n = 8 pairs),
and the fastest early component of the PSP was resistant to high-
frequency ring over 100 Hz (n = 6 pairs). Both of these observa-
tions indicate that the fast component was probably an electrical
synapse mediated by gap junctions. An action potential in the M-
1172 | www.pnas.org/cgi/doi/10.1073/pnas.1012189108
Fig. 3. Spiral ber neurons. (A1) Pairwise recording of a GFP-negative spiral
ber neuron at rhombomere 3 and the contralateral M-cell in the Tg(glyt2:
GFP) relaxed line at 4 dpf (Left). (A2) Firing the spiral ber led to EPSPs (black)
in the M-cell. The slow component of the EPSPs was blocked by the gluta-
matergic blockers 10 M NBQX and 100 M D-AP5 (gray). (A3) PSP in the spiral
ber neuron in response to ring the M-cell. (A4 and A5) Top-down and cross-
section views of the spiral ber neuron (green) and the M-cell (red) from A2
and A3 relative to the glycinergic neurons (blue). Arrowheads mark the gly-
cinergic stripes. (B1 and B2) Summary of the locations of eight spiral ber
neurons (blue) in the glycinergic line in rhombomeres 4 and 5, laid out as in Fig.
2B. (C1) Cross-section view of the backlled spiral ber neurons (red) in the
transgenic glutamatergic line [Tg(vglut2.1:GFP)] at rhombomere 3. Arrow-
heads mark the glutamatergic stripes. (C2) Cross-section of the backlled spiral
ber neurons (white) in the Tg(vglut2.1:RG dbx1b:Cre) line. Arrowhead
indicates the dbx1b-positive middle glutamatergic stripe (green). (D1 and D2)
Projections of the spiral ber neurons shown as in Fig. 2C. Asterisk on D1 marks
a point on the line for the neuron in D2. (Scale bars, 20 m.)
cell only led to subthreshold PSPs in SFNs (Fig. 3A3; n = 10
pairs); the pathway mediating these PSPs and its functional role
are unknown.
The neuron shown in Fig. 3 A2 and A3, like all physiologically
identied spiral ber neurons, was located in rhombomere 3
between the medial and the middle glycinergic stripes, where the
middle glutamatergic stripe is located (Fig. 3 A4, A5, B1, and B2;
n = 8). To conrm their location in the middle glutamatergic
stripe, we backlled the neurons in transgenic sh [Tg(vglut2a:
GFP)] in which the glutamatergic neurons are labeled with GFP,
and found that the neurons were indeed in the middle gluta-
matergic stripe (Fig. 3C1). We also tested whether the neurons
were in the dbx1b transcription factor stripe (dbx1b marks the
middle glutamatergic and glycinergic stripes) by backlling the
spiral ber neurons in sh from a cross of a dbx1b:cre line with
a line in which the vesicular glutamate transporter drives ex-
pression of oxed red uorescent protein, with GFP down-
stream. This cross produces sh in which the glutamatergic
neurons that are also dbx1b positive are green and the other
glutamatergic neurons are red. The spiral ber neurons were in
the ventral part of the dbx1b-positive middle glutamatergic stripe
(Fig. 3C2; n = 4 sh, 75 of 77 cells). For reasons we do not
understand, although they were in the glutamatergic stripe, they
were not stained red in the glutamatergic transgenic line. The
pharmacology of their synaptic connections, however, supported
the conclusion that they are glutamatergic.
Koyama et al.
357
 Reconstructions of all of the patched neurons (n = 8) showed
that they had only contralateral and descending axonal projec-
tions running directly into the axon cap of the contralateral M-
cell, with a characteristic axonal elaboration inside the cap (Fig.
3 D1 and D2). In summary, the evidence shows that the spiral
ber neurons are located ventrally in the middle glutamatergic
stripes and have contralateral projections like those of neurons
in the adjacent middle glycinergic stripe.
Cranial Relay Neurons. We conducted pairwise patch recordings
between the M-cell and GFP-negative neurons in the Tg(glyt2:
GFP) line in the region of putative cranial relay neurons and found
10 GFP-negative neurons that red in response to the M-cells
activation and whose activation led to (polysynaptic) IPSPs in the
M-cell (Fig. 4 A1A3).These neurons red an action potential in
response to the activation of the M-cell (Fig. 4A2; n = 10 pairs)
with a short latency (onset of EPSP: 0.15 0.09 ms; peak of action
potential: 0.70 0.31 ms, n = 10 pairs). A single action potential in
a contralateral neuron in this class led to a longer latency, pre-
sumably polysynaptic, IPSP in the M-cell (1.60 0.28 ms, n = 7
pairs; Fig. 4A3). This IPSP was blocked by a cholinergic blocker
(Fig. 4A3, Right, mecamylamine, 100 M; n = 5 pairs), as well as
a glycinergic blocker (strychnine, 1 M; n = 3 pairs), suggesting
that this class may excite glycinergic neurons via cholinergic
Fig. 4. Cranial relay neurons. (A1) Pairwise recordings of a GFP-negative
cranial relay neuron at rhombomere 7/8 and the contralateral M-cell in the
Tg(glyt2:GFP) relaxed line at 4 dpf. (A2) Firing the M-cell led to an action
potential in the cranial relay neuron. (A3) Firing the cranial relay neuron
with current injection led to a long-latency IPSP in the M-cell (black) that was
blocked by application of the cholinergic blocker mecamylamine (100 M,
gray). (A4) Top-down view of the recorded cranial relay neuron (green) and
the M-cell (red) relative to glycinergic neurons (blue). (A5) Cross-section of
the cranial relay neuron at rhombomere 8. Arrowheads mark glycinergic
stripes. (A6) Close up of axonal terminations from the cranial relay neuron in
region of the white rectangle in A4 showing apposition (arrowheads) to
glycinergic neurons in the region. (B1 and B2) Locations of 10 cranial relay
neurons at rhombomeres 7/8 (red) identied with pairwise recordings, in
top-down (B1) and cross-section (B2) views. (C1 and C2) Summary of the
morphology of seven cranial relay neurons, laid out as in Fig. 2C. Xs mark
neurons whose whole axonal extent was not captured after recording. Box-
and-whisker plots are from seven electroporated cells marked with blue
diamonds whose whole axonal extent was captured. Asterisk marks a point
on the line shown for the neuron in C2. (Scale bars, 20 m.)
Koyama et al.
transmission and the glycinergic neurons in turn feedback to in-
hibit the M-cell. Consistent with this possibility, the recorded pu-
tative cranial relay neurons innervated the region where the
feedback glycinergic neurons were located (Fig. 4 A4A6; n = 10).
These cranial relay neurons also exhibited extensive inner-
vation of the contralateral hindbrain (Fig. 4A4), including the
regions where cranial motor nuclei (fth, seventh, and 10th
motor nuclei) and pectoral n motoneurons are known to be
located. We did several experiments to examine these connec-
tions. Paired recordings of an M-cell and a cranial relay neuron in
the Tg(islet1:GFP) line in which motoneurons are GFP labeled
(n = 3) showed that cranial relay neurons contributing to the
feedback inhibition of M-cell projected to the cranial motor nu-
clei. Projections to the region of the pectoral n motoneurons
were conrmed morphologically by single-cell electroporation of
cranial relay neurons in sh with n motoneurons backlled
(n = 3). Finally, in the most direct approach, an excitatory con-
nection to the fth motor nuclei was observed electrophysiolog-
ically by paired recordings of cranial relay neurons and moto-
neurons in the fth motor nuclei in the Tg(islet1:GFP) line (n =
7). The somata of the recorded neurons in this class were located
below the middle or lateral glycinergic stripes, which contain
neurons with commissural projections like the cranial relay neu-
rons (Fig. 4 B1 and B2). Compared with their extensive contra-
lateral innervation of motor pools, these relay neurons had only
short or no ipsilateral axonal/dendritic processes (Fig. 4 A4, C1,
and C2).
Feedback Glycinergic Neurons. We used pairwise recording to iden-
tify nine glycinergic, feedback-inhibitory neurons that were acti-
vated by ring the ipsilateral M-cell and had inhibitory projections
back to that M-cell (Fig. 5 A1A6). These neurons red an action
potential in response to the M-cells action potential (Fig. 5A2),
with a relatively long latency (3.87 2.60 ms, n = 9 pairs) and
produced IPSPs in the same M-cell (Fig. 5A3) with a short latency
(0.38 0.15 ms, n = 9 pairs). The IPSPs were blocked by bath
application of strychnine (1 M, n = 4 pairs), indicating they were
glycinergic (Fig. 5A3). The strychnine application completely
blocked the feedback inhibition in the M-cell, suggesting that the
feedback inhibition in the 4-dpf zebrash was solely mediated by
glycinergic neurons (n = 23). The IPSP amplitude in the M-cell
NEUROSCIENCE
produced by a single feedback-inhibitory neuron was smaller than
the IPSP produced by ring a cranial relay neuron (Figs. 5A3 vs.
4A3), consistent with the possibility that a single cranial relay
neuron activates multiple feedback-inhibitory cells.
All of the electrophysiologically identied feedback inhibitory
neurons were located among the most ventral and medial gly-
cinergic populations in rhombomere 5 (Fig. 5 B1 and B2; n = 9).
Because the stripe-like structure at rhombomere 5 in the glyci-
nergic line was not always very crisp, we also examined the stripe
location of the neurons by immunostaining for engrailed-1,
which clearly marks the medial glycinergic stripe in other
rhombomeres (2). The feedback-inhibitory neurons identied by
backlling were in the ventral part of the engrailed-1positive
stripe (Fig. 5C1; n = 14). The engrailed labeling of these cells
was dim (probably owing to down-regulation with age), but in
single-neuron lling experiments we found engrailed-positive
staining of this cell type (Fig. 5C2).
Six of nine of these neurons had only ipsilateral ascending
projections directly to the M-cells cell body, ending at its initial
segment. The other three cells had only ipsilateral and primarily
ascending projections contacting the proximal part of the ventral
dendrite of the M-cell (Fig. 5 D1 and D2).We also identied four
additional glycinergic neurons in the same region that had in-
hibitory input to the ipsilateral M-cell but did not reach ring
threshold in response to M cells activation and so were excluded
from the more detailed analysis. These also had only ipsilateral
ascending projections directly to the axon cap. These several lines
of morphological and electrophysiological evidence show that
feedback-inhibitory neurons are located ventrally among the
oldest cells in the medial glycinergic stripe and have the ipsilateral
PNAS | January 18, 2011 | vol. 108 | no. 3 | 1173
358

Fig. 5. Feedback-inhibitory neurons. (A1) Pairwise recordings from a
feedback-inhibitory neuron at rhombomere 5 and the ipsilateral M-cell in
the Tg(glyt2:GFP) relaxed line at 4 dpf. (A2) Firing the M-cell led to an
action potential in the feedback neuron. (A3) Firing the feedback inhibitory
neuron led to an IPSP in the M-cell that was blocked with 1 M strychnine
(gray line). (A5 and A6) Top-down (A5) and cross-section (A6, at rhombomere
5) views of the recorded feedback-inhibitory neuron (green) and the M-cell
(red) relative to the glycinergic neurons (blue). Arrowhead marks the most
medial glycinergic neurons. (B1 and B2) Summary of the position of nine
recorded feedback-inhibitory neurons (pink) in top-down (B1) and cross-
section (B2) views at rhombomere 5. Arrowhead marks the medial glycinergic
stripe. (C1) Hemicross-section of backlled feedback-inhibitory neurons (red)
in the Tg(glyt2:GFP) at 5 dpf, immunostained forengrailed-1 (En1, blue). (C2)
Single electroporated feedback neuron (FB) stained for engrailed (En1), and
the merged image. (D1 and D2) Morphology of feedback-inhibitory neurons
(n = 7 cells) formatted as in Fig. 2 C1 and C2. Black Xs mark two cells whose
whole axonal extent was not captured. Asterisk marks the line for the neuron
in D2. (Scale bars, 2 m in C2, 20 um in other panels.)
the neurons located in stripes that correspond to their morpho-
logical and neurotransmitter phenotypes, they all lie at the ventral
end of the stripes, exactly where we would predict the fastest motor
behaviors to map on the basis of the continuous map of swimming
speed from slow to fast as one moves from dorsal to ventral along
a stripe, as we described in our previous article (2). Here the be-
havior is not a rhythmic one but is the fastest motor behavior
produced by the sh, and the neurons participating in it occupy
the low input resistance, high threshold, fast end of the stripes,
which also contain the oldest neurons. The Mauthner network
thus draws neurons by stripe and position just as expected given
their particular functional roles.
The hindbrain of vertebrates contains a variety of motor net-
works controlling sensory-motor aspects of movements of the
eyes, jaws, gills, and diaphragm (5, 16, 17). If the other networks
follow the same general patterning as the Mauthner network, we
predict that those neurons contributing to faster (more coarse)
sensory-motor behaviors will be drawn from early differentiating
cells located ventrally in those stripes that give rise to neurons
with the appropriate morphological and transmitter phenotype.
For example, a behavior that involves a range of movement
speeds, such as eye movements, which include very fast, large
saccadic movements as well as slow pursuit movements, would be
expected to draw from a range of dorsoventral locations (and
ages) in the stripes (16). Those neurons driving the fast saccadic
eye movements would be expected to have ventral (older)
locations in stripes, with slower eye movements driven by more
dorsal (younger) cells.
We do not yet know how the several other hindbrain networks
in zebrash or in any other animals map onto the overall order
we describe because physiologically dening even one network

and primarily ascending projections characteristic of this stripe.

Their features, like those of the other neurons in the Mauthner
network we examined here, thus mapped exactly as we would have
predicted on the basis of the overall ground plan in hindbrain.
Discussion
Our previous work showed that there is a very orderly structural
and functional patterning extending rostrocaudally throughout
the hindbrain segments (rhombomeres) in zebrash at a young
age, when their motor behaviors are functional and essential for
survival (2). One important prediction from this observation is that
sensory-motor networks in the hindbrain are built by drawing
neurons from a simple plan in which different cell types are
arranged into columns that are ordered by age as well as structural
and functional properties. The organization of the Mauthner
network, which includes neurons located in different rhombo-
meres with outputs across rhombomeres, conrms our predictions
about the relationships between particular hindbrain networks Fig. 6. Summary of the construction of the Mauthner network from the
and the broad neuronal patterning in hindbrain. Each of the ground plan in hindbrain. Neuronal members of the Mauthner circuit map
neuronal types in the M-cell network occupies a location consis- onto the stripes as predicted by our previous study of overall hindbrain
tent with what one would have predicted on the basis of the overall patterning. The colors of the neurons correspond to those on the stripe di-
stripe patterning in the brain, as summarized in Fig. 6. Not only are agram and on the circuit diagram below.

1174 | www.pnas.org/cgi/doi/10.1073/pnas.1012189108 Koyama et al.

359
and its connectivity is a major challenge. Nonetheless, morpho-
logical and transcription factor expression studies suggest that
the ground plan in zebrash is likely to be shared across the
vertebrate subphylum (1724). The extent to which this plan
presages age-related networks in other vertebrates that are akin
to those in zebrash remains to be determined. In humans, faster
behaviors such as saccadic eye movements and startle responses
develop before slower ones, consistent with a patterning like that
in zebrash (25, 26). Large Mauthner-like hindbrain neurons in
the mammalian startle response differentiate early (27, 28),
suggesting there might be similarities in patterning in mammals
at the neuronal as well as behavioral level. We expect that other
vertebrate networks, like that of the M-cell, will draw compo-
nents in predictable ways from the orderly array of parts that
underlies the construction of networks in hindbrain.
Experimental Procedures
Fish Care. All experiments were performed on 4- to 5-dpf zebrash (Dan-
iorerio) obtained from a laboratory stock of wild-type, transgenic, and
mutant adults. All procedures conform to the US National Institutes of
Health guidelines regarding animal experimentation and were approved by
Cornell Universitys Institutional Animal Care and Use Committee.
Transgenic Lines and Mutants. The transgenic lines Tg(glyt2:GFP (29) and Tg
(vglut2a:GFP) (30) were described previously. The transgenic lines Tg(vglut2a:
loxPDsRed-GFP) and Tg(dbx1b:Cre) were constructed using bacterial articial
chromosomes (DKEY-145P24 and zK17G17, respectively) (31). Details will be
described elsewhere. The mutant sh relaxed (red ts25a) was obtained from
the Max Planck Institute in Tubingen, Germany (32).
Electrophysiology. Patch-clamp recordings were performed as described
previously, with minor modications. We initially used cholinergic blockers
-bungarotoxin or d-tubocurarine to paralyze sh by blocking transmission
at the neuromuscular junction (33), but this treatment also blocked the
1. Higashijima S, Mandel G, Fetcho JR (2004) Distribution of prospective glutamatergic,
glycinergic, and GABAergic neurons in embryonic and larval zebrash. J Comp Neurol
480:118.
2. Kinkhabwala A, et al. (2011) A structural and functional ground plan for neurons in
the hindbrain of zebrash. Proc Natl Acad Sci USA 108:11641169.
3. Hale ME, Kheirbek MA, Schriefer JE, Prince VE (2004) Hox gene misexpression and
cell-specic lesions reveal functionality of homeotically transformed neurons. J
Neurosci 24:30703076.
4. Moens CB, Cordes SP, Giorgianni MW, Barsh GS, Kimmel CB (1998) Equivalence in the
genetic control of hindbrain segmentation in sh and mouse. Development 125:381391.
5. Faber DS, Fetcho JR, Korn H (1989) Neuronal networks underlying the escape response
in goldsh. General implications for motor control. Ann N Y Acad Sci 563:1133.
6. Liu KS, Fetcho JR (1999) Laser ablations reveal functional relationships of segmental
hindbrain neurons in zebrash. Neuron 23:325335.
7. Eaton RC, Bombardieri RA, Meyer DL (1977) The Mauthner-initiated startle response
in teleost sh. J Exp Biol 66:6581.
8. Kimmel CB, Sessions SK, Kimmel RJ (1981) Morphogenesis and synaptogenesis of the
zebrash Mauthner neuron. J Comp Neurol 198:101120.
9. Zottoli SJ, Faber DS (1980) An identiable class of statoacoustic interneurons with
bilateral projections in the goldsh medulla. Neuroscience 5:12871302.
10. Nakayama H, Oda Y (2004) Common sensory inputs and differential excitability of
segmentally homologous reticulospinal neurons in the hindbrain. J Neurosci 24:31993209.
11. Scott JW, Zottoli SJ, Beatty NP, Korn H (1994) Origin and function of spiral bers
projecting to the goldsh Mauthner cell. J Comp Neurol 339:7690.
12. Lorent K, Liu KS, Fetcho JR, Granato M (2001) The zebrash space cadet gene controls
axonal pathnding of neurons that modulate fast turning movements. Development
128:21312142.
13. Kimmel CB, Metcalfe WK, Schabtach E (1985) T reticular interneurons: A class of
serially repeating cells in the zebrash hindbrain. J Comp Neurol 233:365376.
14. Hackett JT, Faber DS (1983) Mauthner axon networks mediating supraspinal
components of the startle response in the goldsh. Neuroscience 8:317331.
15. Hackett JT, Faber DS (1983) Relay neurons mediate collateral inhibition of the
goldsh Mauthner cell. Brain Res 264:302306.
16. Sparks DL (2002) The brainstem control of saccadic eye movements. Nat Rev Neurosci
3:952964.
17. Gray PA (2008) Transcription factors and the genetic organization of brain stem
respiratory neurons. J Appl Physiol 104:15131521.
18. Clarke JD, Lumsden A (1993) Segmental repetition of neuronal phenotype sets in the
chick embryo hindbrain. Development 118:151162.
19. Cepeda-Nieto AC, Pfaff SL, Varela-Echavarra A (2005) Homeodomain transcription
factors in the development of subsets of hindbrain reticulospinal neurons. Mol Cell
Neurosci 28:3041.
Koyama et al.
cholinergic transmission from the M-cell. Therefore, we primarily used the
relaxed mutants, in which a mutation in the skeletal muscle dihydropyridine
receptor b1 gene, cacnb1, eliminates muscle contraction. We also used
formamide treatment to paralyze sh in some experiments (34, 35). These
two preparations produced similar results, so we pooled the data.
Neuronal Labeling. Four-day-old larvae were anesthetized and embedded in
1.7% agarose. Then, 20% Alexa Fluor Dextran, 10,000 MW, anionic, lysine
xable (Molecular Probes) was applied in extracellular solution through a patch
micropipette by electroporating the dye into regions of interest with 20-s trains
of 20-V, 1-ms pulses at 50 Hz. Confocal images were acquired at 5 dpf.
Immunohistochemistry. Standard whole-mount antibody staining procedures
were used (1, 36). A rabbit anti-Enhb1 antibody (a gift from A. Joyner, Sloan-
Kettering Institute, New York, NY; 1:2501:500) was detected with goat anti-
rabbit antibody conjugated with HRP (Invitrogen; 1:200) and the Alexa Fluor
647-Tyramide system (Invitrogen).
Image Analysis. To summarize the location of the recorded or backlled
neurons from multiple experiments, we registered volumes to a template
image of the Tg(glyt2:GFP) line by using custom MATLAB (MathWorks) soft-
ware to integrate the functionalityof Imaris, including ImarisXT (Bitplane), with
the coregistration/normalization functionality of SPM8 (http://www.l.ion.ucl.
ac.uk/spm). The quality of the registration was carefully examined by checking
the alignment of stripes, and unsatisfactory registrations were discarded from
the nal visualization. Volumetric colocalization was done in Imaris.
Details of the above methods are provided in SI Experimental Procedures.
ACKNOWLEDGMENTS. This work was supported by Japan Society for Pro-
motion of Science Postdoctoral Fellowships for Research Abroad (to M.K.),
Uehara Memorial Foundation Postdoctoral Fellowships (to M.K.), National
Science Foundation Integrative Graduate Education and Research Trainee-
ship 0333366 (to A.K.), National Institutes of Health (NIH) Training Grant T32
GM007469 (to A.K.), the Ministry of Education, Science, Technology, Sports
and Culture of Japan (S.-i.H.), and NIH Grant NS26539 (to J.F.).
20. Moreno N, Bachy I, Rtaux S, Gonzlez A (2005) LIM-homeodomain genes as territory
markers in the brainstem of adult and developing Xenopus laevis. J Comp Neurol 485:
240254.
21. Sieber MA, et al. (2007) Lbx1 acts as a selector gene in the fate determination of somato-
sensory and viscerosensory relay neurons in the hindbrain. J Neurosci 27:49024909.
22. Ericson J, et al. (1997) Pax6 controls progenitor cell identity and neuronal fate in
response to graded Shh signaling. Cell 90:169180.
NEUROSCIENCE
23. Placzek M, Yamada T, Tessier-Lavigne M, Jessell T, Dodd J (1991) Control of
dorsoventral pattern in vertebrate neural development: induction and polarizing
properties of the oor plate. Development 2(Suppl 2):105122.
24. Takahashi M, Osumi N (2002) Pax6 regulates specication of ventral neurone subtypes
in the hindbrain by establishing progenitor domains. Development 129:13271338.
25. Luna B, Velanova K, Geier CF (2008) Development of eye-movement control. Brain
Cogn 68:293308.
26. de Vries JI, Visser GH, Prechtl HF (1982) The emergence of fetal behaviour.
I. Qualitative aspects. Early Hum Dev 7:301322.
27. Lingenhhl K, Friauf E (1994) Giant neurons in the rat reticular formation: A sensori-
motor interface in the elementary acoustic startle circuit? J Neurosci 14:11761194.
28. Altman J, Bayer SA (1980) Development of the brain stem in the rat. IV. Thymidine-
radiographic study of the time of origin of neurons in the pontine region. J Comp
Neurol 194:905929.
29. McLean DL, Fan J, Higashijima S, Hale ME, Fetcho JR (2007) A topographic map of
recruitment in spinal cord. Nature 446:7175.
30. Bae YK, et al. (2009) Anatomy of zebrash cerebellum and screen for mutations
affecting its development. Dev Biol 330:406426.
31. Kimura Y, Okamura Y, Higashijima S (2006) alx, a zebrash homolog of Chx10, marks
ipsilateral descending excitatory interneurons that participate in the regulation of
spinal locomotor circuits. J Neurosci 26:56845697.
32. Granato M, et al. (1996) Genes controlling and mediating locomotion behavior of the
zebrash embryo and larva. Development 123:399413.
33. Drapeau P, Ali DW, Buss RR, Saint-Amant L (1999) In vivo recording from identiable
neurons of the locomotor network in the developing zebrash. J Neurosci Methods 88:113.
34. Wang M, Wen H, Brehm P (2008) Function of neuromuscular synapses in the zebrash
choline-acetyltransferase mutant bajan. J Neurophysiol 100:19952004.
35. Ono F, Higashijima S, Shcherbatko A, Fetcho JR, Brehm P (2001) Paralytic zebrash
lacking acetylcholine receptors fail to localize rapsyn clusters to the synapse.
J Neurosci 21:54395448.
36. McLean DL, Fetcho JR (2004) Ontogeny and innervation patterns of dopaminergic,
noradrenergic, and serotonergic neurons in larval zebrash. J Comp Neurol 480:3856.
PNAS | January 18, 2011 | vol. 108 | no. 3 | 1175
360
 613
The Journal of Experimental Biology 213, 613-620
2010. Published by The Company of Biologists Ltd
doi:10.1242/jeb.037085

Distinct startle responses are associated with neuroanatomical differences in

pufferfishes
A. K. Greenwood1,2,*, C. L. Peichel1 and S. J. Zottoli2,3
1
Division of Human Biology, Fred Hutchinson Cancer Research Center, Seattle WA, USA 2Marine Biological Laboratory,
Woods Hole MA, USA and 3Department of Biology, Williams College, Williamstown MA, USA
*Author for correspondence (akg@stanfordalumni.org)

Accepted 16 November 2009

SUMMARY
Despite the key function of the Mauthner cells (M-cells) in initiating escape responses and thereby promoting survival, there are
multiple examples of M-cell loss across the teleost phylogeny. Only a few studies have directly considered the behavioral
consequences of naturally occurring M-cell variation across species. We chose to examine this issue in pufferfishes, as previous
research suggested that there might be variability in M-cell anatomy in this group of fish. We characterized the M-cell anatomy
and fast-start responses of two pufferfish species, Tetraodon nigroviridis and Diodon holocanthus. T. nigroviridis showed robust
fast-starts to both tactile and acoustic startling stimuli. These fast-starts occurred with a latency typical of M-cell initiation in other
fish, and retrograde labeling of spinal-projection neurons revealed that T. nigroviridis does have M-cells. By contrast, D.
holocanthus only rarely exhibited fast-start-like behavior, and these responses were at a substantially longer latency and were
much less extensive than those of T. nigroviridis. Using three complementary anatomical techniques we were unable to identify
obvious M-cell candidates in D. holocanthus. These results provide a clear correlation between M-cell presence or absence and
dramatic differences in fast-start behavior. The rich diversity within the pufferfish clade should allow future studies investigating
the factors that contribute to this correlated anatomical and behavioral variation.
Key words: evolution, fish, behavior, neuroanatomy, escape.

INTRODUCTION
The spectacular array of behaviors exhibited by animals has been
generated through the processes of evolution. Behavioral
modifications are ultimately produced by differences in neural
circuits, but what specific circuitry changes contribute to behavioral
evolution? What are the mechanisms by which evolution shapes the
structure of these circuits? What constraints dictate how circuits can
respond to specific selective pressures? Probing the causes and
consequences of nervous system evolution requires a tractable neural
circuit and a group of animals that exhibit variation in this circuit
(for review, see Carr et al., 2001; Katz and Harris-Warrick, 1999;
Nishikawa, 1997; Rose, 2004; Tierney, 1996; Wright, 2000). The
circuitry that generates escape behavior in teleost fishes presents
such a system.
The escape response, also known as the fast-start or startle
response, is initiated by a pair of large, identifiable neurons in the
hindbrain: the Mauthner or M-cells (reviewed by Eaton et al., 2001;
Korn and Faber, 2005; Zottoli and Faber, 2000). Fast-starts in most
teleost fish consist of a C-type fast-start (C-start), the first stage of
which is characterized by a rapid unilateral contraction of trunk
musculature leading to head and tail movement, which causes the
fish to bend into a C-shape. Stage 1 is typically followed by
subsequent movements, including a tail stroke that results in a
forward propulsion of the center of mass (stage 2), and either gliding
or a burst swim (stage 3) (Domenici and Blake, 1997; Eaton et al.,
1981; Foreman and Eaton, 1993). M-cell activity precedes the C-
start, and electrical stimulation of Mauthner axons can elicit a C-
start (Nissanov et al., 1990; Zottoli, 1977). However, M-cells are
not necessary for the production of a C-start. When M-cells in
goldfish or zebrafish are ablated, a C-start can be elicited, although
the onset of the response is delayed (DiDomenico et al., 1988; Eaton
et al., 1982; Liu and Fetcho, 1999; Zottoli et al., 1999).
Although considered a key example of a highly conserved
identifiable neuron, M-cells exhibit substantial diversity within
teleosts (Bierman et al., 2009; Stefanelli, 1980; Zottoli, 1978a). M-
cells can be identified in the majority of teleosts, but many species
across the teleost phylogeny lack obvious M-cells or possess M-
cells with altered anatomy (Stefanelli, 1980; Zottoli, 1978a).
Variable M-cell morphologies include a substantially smaller size,
smaller axon diameter or differences in the structure of the axon
cap, which is a collection of inhibitory and excitatory fibers that
surround the initial segment and axon hillock regions of the M-cell
(Bierman et al., 2009; Stefanelli, 1980; Zottoli, 1978a). In some
cases, the discovery of small or missing M-cells has been associated
with specific types of life history variation, including bottom
dwelling and the use of crypsis or camouflage as an anti-predator
strategy (e.g. toadfish, flounder, lumpfish), or the extreme
modification of caudal fin anatomy (e.g. seahorse, pipefish, ocean
sunfish) (Marshall, 1971; Stefanelli, 1980; Uchihashi et al., 1960;
Zottoli, 1978a).
Given the diverse anatomy of M-cells among fishes, an
outstanding question is how does this natural variation affect fast-
start behavior? Only a single published study has directly examined
the fast-start behavior of fish with missing M-cells (Hale, 2000).
Lumpfish larvae exhibit significantly delayed C-starts when
compared to larval zebrafish (Hale, 2000), which is consistent with
predictions from M-cell ablation studies (Eaton et al., 1982; Zottoli
et al., 1999). However, this study compared fish species across vast
phylogenetic scales, and there may be a host of other factors that
contribute to the observed behavioral differences.

THE JOURNAL OF EXPERIMENTAL BIOLOGY

361
614 A. K. Greenwood, C. L. Peichel and S. J. Zottoli
Pufferfishes, which comprise two sister families within the Order
Tetraodontiformes, offer a useful system for further exploring the
behavioral consequences of M-cell loss. Pufferfishes exhibit a suite
of adaptations in anti-predator defenses, including the capacity to
harbor the neurotoxin tetrodotoxin and inflation behavior. These
anti-predator innovations might be predicted to alter selection on
an ancestral anti-predator behavior, the fast-start. In support of this
hypothesis, the only two species of pufferfish that have been
previously examined do not perform a fast-start response to a tactile
stimulus (Brainerd and Patek, 1998). In addition, anatomical
evidence suggests that some pufferfish lack M-cells (Otsuka, 1964;
Zottoli, 1978b). However, in several other pufferfish species, M-
cells are evident but apparently reduced in size relative to other
teleosts (Stefanelli, 1980; Uchihashi et al., 1960).
Several questions arise from this prior work, which could provide
insight into M-cell and fast-start evolution. From these previous
studies it was unclear whether the lack of M-cells was associated
with the absence of fast-start behavior within a single species. In
addition, no pufferfish were previously shown to exhibit fast-starts;
therefore the function of the M-cells identified in some species was
uncertain. Here, we examined fast-start responses using high-speed
video and M-cell anatomy using both retrograde labeling and serial
plastic sections in two species from the two pufferfish families:
Tetraodon nigroviridis Marion de Proc (green spotted puffer,
Family Tetraodontidae) and Diodon holocanthus Linnaeus
(porcupine puffer or balloonfish, Family Diodontidae).
MATERIALS AND METHODS
Animals
Choice of species
Tetraodon nigroviridis is closely related to, and often confused with,
T. fluviatilis, which was previously found to have identifiable, but
small, M-cells (Stefanelli, 1980; Tagliani, 1905). Diodon
holocanthus does not exhibit a fast-start in response to tactile
stimulation (Brainerd and Patek, 1998), but its M-cell anatomy has
not been previously described.
Housing conditions
Fish were obtained through the aquarium trade (T. nigroviridis:
Aquariumfish.net; D. holocanthus: Sea Dwelling Creatures and
Saltwaterfish.com) and were housed at both the Marine Biological
Laboratory (MBL) and the Fred Hutchinson Cancer Research
Center (FHCRC). Tetraodon nigroviridis were housed in brackish
water at 15p.p.t. and D. holocanthus were housed in either filtered
seawater (MBL) or artificial sea water (FHCRC; Instant Ocean,
Aquarium Systems, Mentor, OH, USA). Fish were fed freeze-dried
or frozen krill every other day (Hikari, Hayward, CA, USA). All
animals were treated in accordance with the Institutional Animal
Care and Use Committees at the MBL (protocol #07-07E) and
FHCRC (protocol #1781).
Behavior
The startle responses of six T. nigroviridis and six D. holocanthus
individuals were recorded using high-speed videography at 500 frames
per second. Two different high-speed cameras were used over the
course of the experiment: Fastcam-PCI (Photron, San Diego, CA,
USA) and 1200hs (The Cooke Corporation, Romulus, MI, USA). An
acrylic aquarium (22cm14cm13cm) was positioned on top of a
platform, and a mirror that was placed above the tank at a 45 deg.
angle enabled the top view to be recorded with the camera positioned
horizontally. Fish were presented with four or five trials of tactile and
acoustic stimuli, which were randomly interleaved and presented at
THE JOURNAL OF EXPERIMENTAL BIOLOGY
a minimum of 3-min intervals. The tactile stimulus consisted of a
sudden touch to the body using an opaque plastic rod (20cm1cm).
The responses to tactile stimuli were not analyzed quantitatively
because it was difficult to determine when the tactile stimulus actually
made contact with the fish; however, in both species the behavior
evoked by this stimulus was qualitatively similar to that evoked by
the acoustic stimulus (see Results).
Each fish received two acoustic trials, with the exception of one
D. holocanthus individual that had three trials. The acoustic stimulus
was provided by the impact of a rubber mallet (340g) on the side
of the platform holding the tank. The mallet was suspended next to
the platform and was raised to the same height for each trial in order
to provide a consistent stimulus.
We quantified several parameters of fast-start behavior from
acoustic stimulus trials. We measured some components that are
known to be M-cell mediated, as defined by the fact that they are
consistently affected by experimental M-cell ablation (latency and
probability of the response) (Eaton et al., 1982; Liu and Fetcho,
1999; Zottoli et al., 1999). In addition, we measured components
that probably involve the contribution of additional hindbrain
neurons (stage 1 duration and angle and peak angular velocity)
(Foreman and Eaton, 1993). The probability of a fast-start was
determined for each species by dividing the total number of trials
evoking a fast-start by the total number of trials (T. nigroviridis, 12
trials: six fish with two trials each; D. holocanthus, 13 trials: five
fish with two trials each and one fish with three trials). For additional
quantification of behavior, imported video files were compiled into
image stacks in ImageJ Software (NIH, Bethesda, MD, USA).
Latency was determined by counting the number of 2ms frames
from the visible impact of the mallet with the apparatus to when
the fish commenced axial movement of the head or tail. The end
of stage 1 was delineated by the fish beginning to straighten its tail
and, in some cases, to turn its head in the opposite direction
(Domenici and Blake, 1997; Hale et al., 2002). Although the
transition from stage 1 was not as rapid for D. holocanthus (see
Results), it was nevertheless an obvious change in movement. To
quantify the amount and direction of movement during the C-start,
we measured the turning angle of each fish during the response
(Foreman and Eaton, 1993; Zottoli et al., 1999). Using ImageJ, a
line was extended along the midline of the anterior portion of the
fish and the change in angle of this line at the end of stage 1, and
10ms and 70ms after movement onset was measured [escape
trajectory angle (Foreman and Eaton, 1993)]. These values were
also used to calculate the peak angular velocity at 4ms intervals
during stage 1. The minimum straight-line distance between the
position of each fish before the onset of movement and after 70ms
was also measured (Zottoli et al., 1999). To do this, an identifiable
point on the fish (on the midline directly in between the pectoral
fins) was tracked rather than the exact center of mass, which was
not determined in this study.
Statistical analysis was performed using R statistical software
(http://www.r-project.org). Fishers exact test was used to compare
the proportion of trials in which individuals of each species exhibited
a C-start. Owing to the small number of D. holocanthus fast-starts
that were recorded (see Results: three responses from two animals),
we did not perform statistical tests on other behavioral measures.
Neuroanatomy
Retrograde tracing
The spinal cords of eleven T. nigroviridis (3.95cm standard length)
and eight D. holocanthus (5.88.1cm standard length) were cut and
back-filled with biotin dextran amine (BDA) to identify all cells
362

that project from the brain into the spinal cord, including the M-
cells. Fish were anesthetized in 0.03% MS-222 and were maintained
under anesthesia during the surgery. An incision was made in the
skin, and muscle and bone were removed in order to expose the
spinal cord. The spinal cord was cut using iridectomy scissors, and
a solution of biotin dextran amine (Invitrogen, Carlsbad, CA, USA)
that had resolidified onto a thin wire was applied onto the cut surface
of the cord. The spinal cord was cut 14mm behind the base of the
cerebellum (pufferfishes have extremely short spinal cords). There
was no difference in the extent of labeling in preparations that had
tracer placement varying within this range (data not shown).
Following tracer application, the incision was closed with sutures
and Vetbond (3M Animal Care Products, St Paul, MN, USA). After
23days, fish were anesthetized in 0.03% MS-222 and perfused
with fixative [4% formaldehyde, 1.25% glutaraldehyde in 0.1moll1
phosphate buffer, pH7.4 (PB)]. Brains were removed, post-fixed
for 4h, equilibrated in 30% sucrose in PB, embedded in OCT
(Sakura Finetek, Torrance, CA, USA), frozen, and sectioned at
50m using a cryostat. Brains were sectioned in horizontal, sagittal
or transverse planes. Free-floating sections were then processed to
detect BDA using a Vector ABC kit (Vector Laboratories,
Burlingame, CA, USA) and DAB as a chromagen with nickel
intensification. Sections were processed as follows: 10min in 3%
hydrogen peroxide/10% methanol in PB, 5min in PB, 2 5min in
PB with 0.2% Triton X-100 (PBT), 3h in ABC reagent in PBT, 2
5min in PBT, 5min in PB, 3 5min in 0.05moll1 Tris, pH8.4, Epon. Serial transverse sections (12m) were mounted on slides,
15min in DAB presoak (0.4% DAB, 0.4% nickel ammonium sulfate
in Tris), DAB reaction (DAB presoak with 0.015% H2O2), then 3
5min in cold 0.05moll1 Tris, pH8.4. Sections were then transferred
to 0.1moll1 PB, mounted on gelatin-subbed slides, counterstained
with Cresyl Violet, dehydrated and a coverslip was placed on top.
All chemicals were from Sigma (St Louis, MO, USA) unless
otherwise indicated.
Silver stain and plastic sections
Two T. nigroviridis (3.9 and 4.1cm standard length) brains and one
D. holocanthus (8.1cm standard length) brain were prepared for
silver staining. Fish were anesthetized in 0.03% MS-222. Brains
were removed and immersed in AFA fixative (9 parts 80% ethanol,
5 parts formalin, 5 parts glacial acetic acid), then dehydrated, cleared
in methyl salicylate, embedded in paraffin, sectioned at 15m, and
stained with Morses modification of Bodians silver technique.
One T. nigroviridis (3.2cm standard length) and one D.
holocanthus (8cm standard length) brain were embedded in plastic
for thin sectioning. Fish were initially anesthetized in 0.03% MS-
222 and were then maintained in 0.012% MS-222 during perfusion
with 2.5% glutaraldehyde and 1% paraformaldehyde in 0.15moll1
sodium cacodylate (pH7.2). The brains were removed and placed
in fresh fixative for 12h, washed in buffer, post-fixed in 2% OsO4
Table 1. Comparison of fast-start behavior between species
C-starts to acoustic stimuli (number of trials)
Latency of movement onset
Angle 10ms after onset
Duration of stage 1
Angle at end of stage 1
Angle 70ms after onset
Distance moved after 70ms
Peak angular velocity
Values mean s.d.
THE JOURNAL OF EXPERIMENTAL BIOLOGY
M-cells and fast-start in pufferfish 615
A B
C
*
D *
Fig.1. Comparison of fast-start behavior in T. nigroviridis and D.
holocanthus. Photographs of (A) T. nigroviridis (size range: 3.95cm) and
(B) D. holocanthus (size range: 5.88.1cm). Silhouettes of (C) T.
nigroviridis and (D) D. holocanthus traced from high-speed videos of
representative responses to startling acoustic stimulus presentation, as
viewed from above. Silhouettes are at 2ms intervals. The onset of the
stimulus is indicated by an arrowhead. The asterisk indicates the latency of
the response, determined by the first occurrence of lateral head and/or tail
movement. Scale bars, 2cm.
in 0.15moll1 sodium cacodylate for 4h, washed in buffer,
dehydrated in ethanol, cleared in propylene oxide and embedded in
stained with Toluidine Blue and observed with a conventional light
microscope.
RESULTS
Behavior
Six T. nigroviridis individuals were presented with abrupt tactile
and auditory stimuli in order to elicit startle responses. All fish
readily produced fast-starts to both types of stimuli. An example of
a fish exhibiting a fast-start to the auditory stimulus is shown in
Fig.1. These responses were C-type fast-starts (C-starts), in which
the fish assumed a C-shape during the first stage of the response.
C-starts had a rapid onset of approximately 11ms after the auditory
stimulus (Table1). Superficially similar fast-start behavior was
elicited following a tactile stimulus, although the latency of this
response could not be determined (see Materials and Methods).
Thus, T. nigroviridis exhibits robust fast-start behavior with a short
latency that is suggestive of M-cell initiation (Eaton et al., 1977).
Following stage 1, fish exhibited a propulsive tail stroke (stage 2)
followed by either gliding or a burst swim (stage 3). The first stage
consisted of a variable amount of turning (up to 180 deg.), evidenced
by the large variation in the angle traveled during stage 1 (Table1).
Diodon holocanthus, by contrast, did not show robust fast-start
responses. Six fish were tested with both tactile and auditory stimuli.
T. nigroviridis D. holocanthus
12/12 (100%) 3/13 (23%)
11.21.3ms 27.31.2ms
42.310deg. 7.84deg.
23.59ms 7717ms
97.155deg. 397deg.
79.471deg. 34.44deg.
4.051.3cm 0.480.16cm
6.12 deg.ms1 0.80.3deg.ms1
363
616 A. K. Greenwood, C. L. Peichel and S. J. Zottoli
In response to the touch of a rod to the head or tail, D. holocanthus
individuals did not exhibit a fast-start, but instead used their pectoral
fins to swim away from the stimulus or initiated a burst swim (data
not shown). In the majority of trials with an auditory stimulus, fish
did not make any axial bending movements (Table1). However, we
did record three weak responses from two fish that consisted of slight
axial contraction resulting in minimal head and tail movement (Fig.1,
Table1). Although the C form attained by D. holocanthus was much
more shallow than that of T. nigroviridis, we will refer to D.
holocanthus responses as C-starts for consistency. The probability
of performing a C-start was significantly different between species
(Table1, Fishers exact test: P<0.001). Diodon holocanthus C-starts
occurred at an average latency of 26ms, which was over twice as
long as T. nigroviridis, and the duration of stage 1 was substantially
longer than that of T. nigroviridis (Table1). The angle at 10ms after
onset, and the angles at the end of stage 1 and after 70ms were also
considerably smaller in D. holocanthus (Table1). Diodon
holocanthus individuals also achieved a substantially lower peak
angular velocity during their responses (Table1).
Diodon holocanthus was not observed to make a propulsive tail
stroke (stage 2) or a burst swim (stage 3) following stage 1. The
distance moved 70ms after the onset of the fast-start was
approximately one body length in T. nigroviridis, whereas this
distance was almost negligible in D. holocanthus (Table1). Thus,
the D. holocanthus fast-start consists only of a delayed, weak stage
1 with no successive stages of movement.
A c,d B 1
T.n. b 2
e,f
D.h.
C D
E F
THE JOURNAL OF EXPERIMENTAL BIOLOGY
Neuroanatomy
Of the multiple elements of escape circuit anatomy that might cause
differences in the fast-start response, we focused our efforts on
identifying the M-cell. Retrograde labeling of axonal projections,
silver-stained preparations and thin plastic sections were used to
locate M-cells.
Retrograde tracing of cells projecting into the spinal cord with
BDA labeled an array of reticulospinal neurons with a segmental
arrangement in both T. nigroviridis (Fig.2) and D. holocanthus (data
not shown) similar to that described in other fish (Lee and Eaton,
1991; Lee et al., 1993). Vestibulospinal cells, which were more
lateral than the reticulospinal cells, were also labeled in both species
(data not shown). The extent of BDA labeling was similar in T.
nigroviridis and D. holocanthus. In all brains from both species, we
were able to clearly identify at least six reticulospinal segments; the
seventh and eighth segment were often difficult to discern because
of higher background in caudal hindbrain regions. In addition, we
counted the number of labeled cells in four horizontally sectioned
brains from both species, and found an average of 171 and 165
labeled cells in T. nigroviridis and D. holocanthus, respectively.
We were able to identify obvious M-cells in all T. nigroviridis
brains (N11; Fig.2). These cells showed hallmarks of M-cells (Lee
and Eaton, 1991; Zottoli, 1978b). They were located at the caudal
extent of the fifth motor nucleus at the level of the eighth nerve
(Fig.2). They were found in the fourth segment of reticulospinal
cells, which probably corresponds to the fourth hindbrain segment
Fig.2. Comparison of M-cell anatomy in T. nigroviridis and D.
holocanthus. (A)Schematic of brains from each species
indicating angle of sections in subsequent panels. T.n., T.
nigroviridis; D.h., D. holocanthus. (B)Horizontal section from
BDA-labeled T. nigroviridis brain. The dashed line delineates
3 the midline. DAB-reacted BDA appears black and Nissl
counterstain is purple. Hindbrain segments (14), delineated
by reticulospinal cells, are indicated to the right of this image.
4 A green arrow points to the distinctive M-cell on the right
side. Only a small portion of cell body of the left M-cell can
be seen in this section. (C)Transverse section from T.
nigroviridis taken at the caudal extent of the fifth motor
nucleus approximately at the level of the eighth nerve.
(D)Enlarged view of boxed region in C; the M-cell is
indicated by a green arrow. (E)Transverse section from D.
holocanthus taken at approximately the same position as C.
(F)Enlarged view of boxed region in E. No M-cell is visible in
this or neighboring sections. Scale bars, 1mm (C,E) and
100m (B,D,F) .
364

of zebrafish and goldfish (Lee and Eaton, 1991; Lee et al., 1993).
The M-cells were substantially larger than all other reticulospinal
cells and exhibited a characteristic anatomy, including two large
principal dendrites and large axons that crossed the midline (Figs2,
3). M-cells were clearly distinguishable in silver-stained paraffin
sections (data not shown) and Toluidine-Blue-stained plastic sections
(Fig.3). The M-axons were readily identified in the spinal cord
A holocanthus from brains that were BDA labeled and Nissl stained
B do not project far enough down the spinal cord to be labeled by a
C cell ablations in the laboratory (Eaton et al., 1982; Liu and Fetcho,
Fig.3. M-cells and the crossing of the M-axons in T. nigroviridis. (AC)
1m cross sections stained with Toluidine Blue. (A)M-cells can be seen on
either side of the midline (arrows). Dorsal is up and the midline is in the
center of the photograph. (B)Higher power micrograph of the left M-cell.
The beginning portion of the lateral dendrite can be seen to the left. The
axon hillock and initial segment can be seen to the right. Dorsal is up and
the midline is toward the right. (C)The M-axons can be traced caudally
where they cross (decussation is designated by an arrow). Dorsal is up and
the midline of the spinal cord is in the center of the image. Scale bars,
50m.
THE JOURNAL OF EXPERIMENTAL BIOLOGY
M-cells and fast-start in pufferfish 617
(Fig.4) and could be traced after crossing the midline to the
Mauthner cell body on the opposite side of the brain in Toluidine-
Blue-stained plastic sections (Fig.3). The axon cap did not have the
features attributed to the composite axon cap of the goldfish and
appears more similar to a simple axon cap (Bierman et al., 2009).
By contrast, we did not identify any obvious M-cell candidates
in D. holocanthus. In eight brains, we were unable to distinguish
any large retrogradely labeled neurons in the fourth segment of
reticulospinal cells at the caudal portion of the fifth motor nucleus
(Fig.2). Careful examination of sections of the spinal cord of D.
(N8), silver stained (N1) or Toluidine-Blue stained (N1) did not
reveal any axons that were distinctly larger than other axons (Fig.4
and data not shown).
The classification of a putative M-cell based solely on anatomical
location or size would potentially miss a cell that was substantially
modified compared to other fish. Therefore, we expanded our search
to account for this possibility. If M-cells were reduced in size but
still present in a homologous position, we would nevertheless expect
to be able to identify them anatomically because M-cells are
characteristically more dorsal than other cells in the fourth
reticulospinal segment (see Fig.2). However, we did not see any
smaller, retrogradely labeled cells in this anatomical location that
fitted this description. We also evaluated whether large M-cells
might indeed be present in the homologous location but that they
spinal backfill. However, no large cell bodies were distinguishable
in this region in D. holocanthus in Nissl-stained sections, silver-
stained paraffin sections, or Toluidine-Blue-stained thin plastic
sections (data not shown).
DISCUSSION
Mauthner cells have a long history of study in comparative
neurobiology (Bierman et al., 2009; Stefanelli, 1980; Zottoli,
1978b), although little work has been devoted to studying the
behavioral consequences of naturally occurring M-cell loss (Hale,
2000). Our results provide the first correlation between the presence
or absence of M-cells and variable fast-start behavior in related
species. These findings support and expand classic studies of M-
1999) by examining the behavioral consequences of M-cell loss in
an evolutionary context. To our knowledge, this work represents
the only published demonstration that naturally occurring M-cell
loss in adults is correlated with dramatically altered fast-start
behavior.
Using both retrograde labeling of axonal projections and
conventional anatomical techniques, we identified distinctive M-
cells in T. nigroviridis but were unable to distinguish M-cells in D.
holocanthus. We cannot entirely exclude the possibility that D.
holocanthus may have atypical M-cells that were missed by our
analysis, although we were careful to consider liberal definitions of
M-cells in our search. The M-cell was thought to be absent in adult
Xenopus laevis (Stefanelli, 1951) and freshwater eels (Zottoli,
1978b), but subsequently the M-cell was identified in both (Meredith
and Roberts, 1987; Will, 1991). Future work, including
developmental analysis or the use of molecular markers (Flores et
al., 2008), will be required to definitively confirm that these fish
lack M-cells. Our current results suggest that D. holocanthus either
lacks M-cells or possesses highly atypical M-cells. Therefore, we
can confidently conclude that we have identified the presence of
M-cells in T. nigroviridis and the absence of conserved M-cells in
D. holocanthus.
365
618 A. K. Greenwood, C. L. Peichel and S. J. Zottoli

A B

*
*

Fig.4. Comparison of spinally projecting axons in T. nigroviridis and D. holocanthus at the level of the obex. (A,B)1m cross sections of the fasciculus
longitudinalis medialis at the level of the obex, stained with Toluidine Blue, and shown in grayscale. (A)T. nigroviridis spinal cord in which the M-axons are
easily identified (asterisks). (B)D. holocanthus spinal cord in which the M-axons could not be identified. Note that no pair of axons stands out on the basis of
size or amount of myelination. Dorsal is up and the midline of the spinal cord is in the center of both images. The position of the midline is indicated with a
dashed line. Scale bar, 50m.

Differences in M-cell neuroanatomy are correlated with
distinctive fast-start behavior in these species. Tetraodon nigroviridis
performed robust C-type fast-start responses that were similar to
M-cell mediated C-starts from other teleosts, including goldfish
(Foreman and Eaton, 1993). However, under identical stimulus
conditions, D. holocanthus only rarely exhibited weak responses at
a substantially longer latency. Diodon holocanthus may have
different stimulus requirements than T. nigroviridis or the
behavioral context of the animal before the stimulus is delivered
might have an influence on the responsiveness of this species (Zottoli
et al., 1995). However, in pilot experiments, these fish did not exhibit
any rapid trunk movement following presentation of a variety of
acoustic, tactile and visual stimuli other than the weak C-start
following an acoustic stimulus. Diodon holocanthus also did not
exhibit fast-starts to a tactile stimulus in a previous study (Brainerd
and Patek, 1998).
Several aspects of the behavior of D. holocanthus are reminiscent
of the fast-start responses of other species of fish with missing M-
cells. Compared with T. nigroviridis, D. holocanthus exhibited a
longer latency of movement onset and a reduced probability of
performing a C-start. Experimental M-cell ablations result in fast-
start responses that occur at a longer latency (Eaton et al., 1982;
Kohashi and Oda, 2008; Liu and Fetcho, 1999; Zottoli et al., 1999),
and yield a reduced probability of generating a fast-start (Zottoli
et al., 1999). In addition, the delayed latency of onset is similar to
another species that naturally lacks M-cells (larval lumpfish: Hale,
2000). Although consistent with missing M-cells, a delayed latency
alone is not sufficient evidence to indicate the lack of an M-cell.
The sea robin has small M-cells and small M-axons (which
although small can nevertheless be distinguished, unlike D.
holocanthus) and exhibits delayed, weak C-starts at a reduced
probability compared with goldfish (S.J.Z., unpublished) (Zottoli
et al., 1992).
Aside from a reduced probability and longer latency, the behavior
of D. holocanthus is substantially different from fast-start behavior
that typically results from experimental M-cell ablations. Diodon
holocanthus did not bend as deeply during the C-starts and had a
protracted stage 1 duration compared with T. nigroviridis, and the
peak angular velocity of these responses was dramatically reduced.
In addition, D. holocanthus essentially did not move away from its
initial starting position (i.e. no stage 2) and did not exhibit subsequent
stages of movement (i.e. no stage 3) following the weak C-start
other than slow relaxation of the C-shape. It is unclear that this
response would actually be adaptive for predator avoidance and
whether it can be considered a true fast-start.
These substantial kinematic differences suggest that additional
factors beyond the lack of identifiable M-cells probably contribute
to the distinct behaviors of D. holocanthus and T. nigroviridis. Besides
latency differences, larval lumpfish do not show kinematic differences
in their fast-start performance (Hale, 2000). Similarly, M-cell ablations
alone do not affect angular, velocity, or distance measures in adult
goldfish (Eaton et al., 1982; Zottoli et al., 1999). The altered
kinematics of D. holocanthus fast-starts could result from additional
neuroanatomical and/or morphological factors. The reduced angle and
protracted stage 1 duration in D. holocanthus suggests that there may
be reduced ipsilateral muscle activation (Domenici and Blake, 1997;
Foreman and Eaton, 1993). In addition, D. holocanthus may lack
contralateral muscular activity that typically facilitates a propulsive
tail stroke (Domenici and Blake, 1997). These differences imply that
other aspects of fast-start circuitry might differ between D.
holocanthus and T. nigroviridis. In larval zebrafish, slight kinematic
differences in behavior that are apparent after M-cell ablations are
enhanced following additional ablations of the M-cell homologues
in adjacent segments (Liu and Fetcho, 1999). It will be interesting to
determine if pufferfish possess the M-cell homologues that are found
in both goldfish and zebrafish (Kimmel, 1982; Kohashi and Oda, 2008;
Lee and Eaton, 1991; Lee et al., 1993), and whether these cells also
show anatomical differences between pufferfish species. In addition
to variation in neural circuitry, kinematic differences between D.
holocanthus and T. nigroviridis fast-starts could result from
morphological differences, including musculature, body shape and
body stiffness. The relative contribution of neural versus
morphological factors to altered fast-start responses can be clarified
in the future using electromyography.

THE JOURNAL OF EXPERIMENTAL BIOLOGY

366

Our findings raise interesting evolutionary questions
concerning both the loss of M-cells in D. holocanthus relative to
ancestral teleosts and the variation in M-cells and behavior
between the two pufferfish species. The absence of M-cells and
robust fast-start behavior in D. holocanthus relative to other
teleosts could result from a relaxation of selection for the
maintenance of fast-start behavior in the presence of a multitude
of pufferfish-specific anti-predator adaptations. In addition, the
morphological changes that are associated with the gain of
inflation include loss of hypaxial musculature (Brainerd, 1994),
which may preclude a robust fast-start, leading to a loss of M-
cells. The outcome of relaxed selection could vary among species,
which might explain the differences in M-cell presence between
T. nigroviridis and D. holocanthus.
Alternatively, there may be direct selection for the loss of a
fast-start response and a functional M-cell in D. holocanthus. There
are obvious tradeoffs between performing a fast-start and
generating inflation behavior. Both of these behaviors are used to
avoid predation, but a fish cannot perform both fast-start and
inflation simultaneously because these behaviors require opposing
states of trunk muscle activation (contraction in fast-start and
relaxation in inflation). At a neural level, there must be a decision
to generate one of the two behaviors in response to a perceived
threat. If inflation is consistently a better strategy for D.
holocanthus, the chance of making the wrong choice (the fast-
start) might be selected against. However, potential tradeoffs might
be alleviated to some extent if these behaviors are utilized in
somewhat different stimulus contexts, for example, initial predator
detection versus capture. Systematic field observations will be
required to assess the relative use of these behaviors in different
species.
If the loss of M-cells in D. holocanthus is due to selection, T.
nigroviridis must have experienced a distinct selective regime.
These species are from two different pufferfish families, thought
to have diverged at least 50 million years ago (Alfaro et al., 2007),
and they exhibit multiple differences in morphology, coloration,
life history and habitat. At this point, data are lacking from a
sufficient number of species to determine whether the presence
or absence of M-cells represents the ancestral pufferfish state, so
we will discuss the findings in regard to known differences between
the two species. As described above, morphological differences
between these species may impact fast-start performance, and
thereby selection. In addition, because fast-start performance is
important in predator escape (Eaton and Hackett, 1984; Walker
et al., 2005), an obvious difference to consider between these
species is predation pressure. Unfortunately, little is known about
the life histories of these species. Diodon holocanthus is a marine
species, inhabiting mostly inshore and reef habitats throughout the
tropics. T. nigroviridis is a smaller euryhaline species that resides
in estuaries and freshwater streams in the Indo-Pacific. Large bony
fish and sharks are known predators of D. holocanthus (Leis, 2001;
Randall, 1967). Predators of T. nigroviridis are not well
characterized, but could include both birds and fish, given their
size and habitat. Divergence in several characteristics supports the
idea that these species may experience different predation regimes.
Although both species are known to contain tetrodotoxin (Chen
and Chou, 1998; Mahmud et al., 2003), T. nigroviridis has a
potentially aposematic coloration pattern (Fig.2), whereas D.
holocanthus exhibits mottled, disruptive coloration, most likely
for camouflage. In addition, our anecdotal laboratory observations
suggest that these fish employ inflation behavior to different
extents. Tetraodon nigroviridis was rarely observed to perform
THE JOURNAL OF EXPERIMENTAL BIOLOGY
M-cells and fast-start in pufferfish 619
inflation behavior, whereas D. holocanthus did so frequently
during routine handling.
Although at this stage we cannot pinpoint the key factors that
contribute to the species differences, our findings provide a clear
foothold into variation in M-cells and correlated fast-start behavior
characters in the pufferfish phylogeny. In the future, we can take
advantage of the diversity within pufferfish to examine the
evolutionary history of and ultimately the mechanisms contributing
to the gain or loss of fast-start behavior and circuitry.
ACKNOWLEDGEMENTS
We are grateful to Janice Simmons, Scott Lindell and Shaun McCann for advice
and assistance with fish husbandry; to Gwyneth Card, Michael Dickinson and
Jason Ritt for assistance with high-speed cameras; and to Tina Wong, Nancy
Piatczyc and Farah Musani for technical assistance. This work was supported by
a Grass Fellowship and a Helen Hay Whitney Postdoctoral Fellowship to A.K.G.,
a Burroughs Wellcome Career Award in the Biomedical Sciences to C.L.P. and a
Williams College Howard Hughes Medical Institute and Essel Foundation grant to
S.J.Z.
REFERENCES
Alfaro, M. E., Santini, F. and Brock, C. D. (2007). Do reefs drive diversification in
marine teleosts? Evidence from pufferfishes and their allies (Order
Tetraodontiformes). Evolution 61, 2104-2126.
Bierman, H. S., Zottoli, S. J. and Hale, M. E. (2009). Evolution of the Mauthner axon
cap. Brain Behav. Evol. 73, 174-187.
Brainerd, E. L. (1994). Pufferfish inflation: functional morphology of postcranial
structures in Diodon holocanthus (Tetraodontiformes). J. Morphol. 220, 243-261.
Brainerd, E. L. and Patek, S. N. (1998). Vertebral column morphology, C-start
curvature, and the evolution of mechanical defenses in Tetraodontiform fishes.
Copeia 4, 971-984.
Carr, C. E., Soares, D., Parameshwaran, S. and Perney, T. (2001). Evolution and
development of time coding circuits. Curr. Opin. Neurobiol. 11, 727-733.
Chen, C.-Y. and Chou, H.-N. (1998). Detection of tetrodotoxin by high performance
liquid chromatography in lined-moon shell and puffer fish. Acta Zool. Taiwanica 9,
41-48.
DiDomenico, R., Nissanov, J. and Eaton, R. C. (1988). Lateralization and adaptation
of a continuously variable behavior following lesions of a retiuclospinal command
neuron. Brain Res. 473, 15-28.
Domenici, P. and Blake, R. (1997). The kinematics and performance of fish fast-start
swimming. J. Exp. Biol. 200, 1165-1178.
Eaton, R. C. and Hackett, J. T. (1984). The role of the Mauthner cell in fast-starts
involving escape in teleost fishes. In Neural Mechanisms of Startle Behavior (ed. R.
C. Eaton), pp. 213-266. New York: Plenum Press.
Eaton, R. C., Bombardieri, R. A. and Meyer, D. L. (1977). The Mauthner-initiated
startle response in teleost fish. J. Exp. Biol. 66, 65-81.
Eaton, R. C., Lavender, W. A. and Wieland, C. M. (1981). Identification of Mauthner
initiated response patterns in goldfish: evidence from simultaneous cinematography
and electrophysiology. J. Comp. Physiol. A 144, 521-531.
Eaton, R. C., Lavender, W. A. and Wieland, C. M. (1982). Alternative neural
pathways initiate fast-start responses following lesions of the Mauthner neuron in
goldfish. J. Comp. Physiol. A 145, 485-496.
Eaton, R. C., Lee, R. K. and Foreman, M. B. (2001). The Mauthner cell and other
identified neurons of the brainstem escape network of fish. Prog. Neurobiol. 63, 467-
485.
Flores, C. E., Ene, S. and Pereda, A. E. (2008). An immunohistochemical marker for
goldfish Mauthner cells. J. Neurosci. Methods 175, 64-69.
Foreman, M. B. and Eaton, R. C. (1993). The direction change concept for
reticulospinal control of goldfish escape. J. Neurosci. 13, 4101-4113.
Hale, M. E. (2000). Startle responses of fish without Mauthner neurons: Escape
behavior of the lumpfish (Cyclopterus lumpus). Biol. Bull. 199, 180-182.
Hale, M. E., Long, J. H., Jr, McHenry, M. J. and Westneat, M. W. (2002). Evolution
of behavior and neural control of the fast-start escape response. Evolution 56, 993-
1007.
Katz, P. S. and Harris-Warrick, R. M. (1999). The evolution of neuronal circuits
underlying species-specific behavior. Curr. Opin. Neurobiol. 9, 628-633.
Kimmel, C. B. (1982). Reticulospinal and vestibulospinal neurons in the young larva of
a teleost fish, Brachydanio rerio. In Progress in Brain Research: Anatomy of
Descending Pathways to the Spinal Cord, vol. 57 (ed. H. G. J. M. Kuypers and G. F.
Martin), pp. 1-24. New York: Elsevier Biomedical.
Kohashi, T. and Oda, Y. (2008). Initiation of Mauthner- or non-Mauthner-mediated fast
escape evoked by different modes of sensory input. J. Neurosci. 28, 10641-10653.
Korn, H. and Faber, D. S. (2005). The Mauthner cell half a century later: a
neurobiological model for decision-making? Neuron 47, 13-28.
Lee, R. K. and Eaton, R. C. (1991). Identifiable reticulospinal neurons of the adult
zebrafish, Brachydanio rerio. J. Comp. Neurol. 304, 34-52.
Lee, R. K., Eaton, R. C. and Zottoli, S. J. (1993). Segmental arrangement of
reticulospinal neurons in the goldfish hindbrain. J. Comp. Neurol. 329, 539-556.
Leis, J. M. (2001). Family Diodontidae. In FAO Species Identification Guide for Fishery
Purposes, vol. Bony Fishes part 4 (ed. K. E. Carpenter and V. H. Niem), pp. 3958-
3965.
Liu, K. S. and Fetcho, J. R. (1999). Laser ablations reveal functional relationships of
segmental hindbrain neurons in zebrafish. Neuron 23, 325-335.
367
620 A. K. Greenwood, C. L. Peichel and S. J. Zottoli
Mahmud, Y., Arakawa, O., Ichinose, A., Tanu, M., Takatani, T., Tsuruda, K.,
Kawatsu, K., Hamano, Y. and Noguchi, T. (2003). Intracellular visualization of
tetrodotoxin (TTX) in the skin of a puffer Tetraodon nigroviridis by immunoenzymatic
technique. Toxicon 41, 605-611.
Marshall, N. B. (1971). Explorations in the Life of Fishes. Cambridge, MA: Harvard
University Press.
Meredith, G. E. and Roberts, B. L. (1987). Distribution and morphological
characteristics of efferent neurons innervating end organs in the ear and lateral line
of the European eel. J. Comp. Neurol. 265, 494-506.
Nishikawa, K. C. (1997). Emergence of novel functions during brain evolution.
BioScience 47, 341-353.
Nissanov, J., Eaton, R. C. and DiDomenico, R. (1990). The motor output of the
Mauthner cell, a reticulospinal command neuron. Brain Res. 517, 88-98.
Otsuka, N. (1964). Weitere vergleichend-anatomische untersuchlungen an
Mauthnerschen zellen von fischen. Z. Zelleforsch. 62, 61-71.
Randall, J. E. (1967). Food habits of reef fishes of the West Indies. Stud. Trop.
Oceanogr. Miami 5, 665-847.
Rose, G. J. (2004). Insights into neural mechanisms and evolution of behaviour from
electric fish. Nat. Rev. Neurosci. 5, 943-951.
Stefanelli, A. (1951). The Mauthnerian apparatus in the Ichthyopsida; its nature and
function and correlated problems of neurohistogenesis. Q. Rev. Biol. 26, 17-34.
Stefanelli, A. (1980). I neuroni di Mauthner degli ittiopsidi valutazioni comparative
morfologiche e funzionali. Atti Della Accad. Naz. dei Lincei 16, 1-43.
Tagliani, G. (1905). Le fibre del Mauthner nel midollo spinale deVertebrati inferiori
(anamni). Arch. Zool. 2, 385-435.
Tierney, A. J. (1996). Evolutionary implications of neural circuit structure and function.
Behav. Process. 35, 173-182.
THE JOURNAL OF EXPERIMENTAL BIOLOGY
Uchihashi, K., Shimamura, H. and Honda, A. (1960). A study on the Mauthner cell in
teleosts in relation to the swimming behavior. Ann. Rept. Jap. Sea Reg. Fish Res.
Lab. 6, 203-216.
Walker, J. A., Ghalambor, C. K., Griset, O. L., McKenney, D. and Reznick, D. N.
(2005). Do faster starts increase the probability of evading predators? Funct. Ecol.
19, 808-815.
Will, U. (1991). Amphibian Mauthner cells. Brain Behav. Evol. 37, 317-332.
Wright, W. G. (2000). Neuronal and behavioral plasticity in evolution: experiments in a
model lineage. BioScience 50, 883-894.
Zottoli, S. J. (1977). Correlation of the startle reflex and Mauthner auditory responses
in unrestrained goldfish. J. Exp. Biol. 66, 243-254.
Zottoli, S. J. (1978a). Comparative morphology of the Mauthner cell in fish and
amphibians. In Neurobiology of the Mauthner Cell (eds. D. S. Faber and H. Korn).
New York: Raven Press.
Zottoli, S. J. (1978b). Comparison of Mauthner cell size in teleosts. J. Comp. Neurol.
178, 741-757.
Zottoli, S. J. and Faber, D. S. (2000). The Mauthner cell: what has it taught us? The
Neuroscientist 6, 25-37.
Zottoli, S. J., Davis, G. W. and Northen, S. C. (1992). Comparative studies of the
Mauthner cell in teleosts. Neurology Today 1992, 53-63.
Zottoli, S. J., Bentley, A. P., Prendergast, B. J. and Rieff, H. I. (1995). Comparative
studies on the Mauthner cell of teleost fish in relation to sensory input. Brain Behav.
Evol. 1995, 151-164.
Zottoli, S. J., Newman, B. C., Rieff, H. I. and Winters, D. C. (1999). Decrease in
occurrence of fast startle responses after selective Mauthner cell ablation in goldfish
(Carassius auratus). J. Comp. Physiol. A 184, 207-218.
368
 REVIEWS
Corollary discharge across
the animal kingdom
Trinity B. Crapse and Marc A. Sommer
Abstract | Our movements can hinder our ability to sense the world. Movements can induce
sensory input (for example, when you hit something) that is indistinguishable from the input
that is caused by external agents (for example, when something hits you). It is critical for
nervous systems to be able to differentiate between these two scenarios. A ubiquitous
strategy is to route copies of movement commands to sensory structures. These signals,
which are referred to as corollary discharge (CD), influence sensory processing in myriad
ways. Here we review the CD circuits that have been uncovered by neurophysiological
studies and suggest a functional taxonomic classification of CD across the animal kingdom.
This broad understanding of CD circuits lays the groundwork for more challenging studies
that combine neurophysiology and psychophysics to probe the role of CD in perception.

Receptor (or sensor)
A sensory end organ that
detects changes in the external
world or the internal viscera.
Effector
An organ that becomes active
in response to a nerve signal.
Afferent
A neuronal projection that
conveys information to a
structure. The term is often
used in reference to sensory
channels.
Department of Neuroscience,
Center for the Neural Basis
of Cognition, and Center
for Neuroscience at the
University of Pittsburgh,
University of Pittsburgh,
Pittsburgh, Pennsylvania
15260, USA.
Correspondence to T.B.C.
e-mail: tbc6@pitt.edu
doi:10.1038/nrn2457
An animal that never moved could possess a relatively
simple sensory system. It would need receptors to detect
environmental changes, and a nervous system capable
of interpreting the information. The animal would be
limited by two main factors: the range of energy changes
it could detect and the sophistication of the subsequent
analyses it could perform. Within these limits, its
sensory receptors should reliably detect near (for exam-
ple, involving physical contact) and distant (for example,
involving light) environmental occurrences.
Once an animal moves, however, the situation
changes drastically. Movements contribute to form-
ing a sense of space, as discussed by Poincar1, and to
improved, discriminative sensory perception. But move-
ments have a potential downside as well: they introduce
ambiguity about the source of sensory input. A newly
mobile animal would be unable to determine whether a
disturbance registered by its sensory receptors was the
result of a change in the environment or simply a conse-
quence of its own movement. As it walked, it might not
know whether sudden activation of its skin receptors was
due to a predators paw or to an inanimate obstacle in
its path, or whether detected changes in light and shade
were caused by movements of external objects or by
movements of its own photosensory organs.
The cause of sensory input: self or other?
The fundamental distinction concerning the origin
of sensory input was discussed by von Holtz and
Mittelstaedt2. They termed input that results from occur-
rences in the environment as exafference; this is the
only kind of input that a stationary animal would expe-
rience. Exafference contrasts with reafference, which
refers to those inputs that inevitably result from an ani-
mals own movements. Reafference is action-dependent
and thus would be absent in immobile animals.
Sensory receptors are indifferent to the cause of
their activation, so in principle they would convey both
exafference and reafference equally, and downstream
processing would proceed identically for both. However,
von Holst and Mittelstaedt2 pointed out that this would
lead to a manifold sensory and interpretative problem:
reafference would be confused for exafference. This
potential confusion would be particularly significant for
the special abilities of some animals. For example, bats
emit sounds that reflect back to them, but the reaffer-
ent echoes are mingled with exafferent noise from other
sources (including nearly identical biosonar probes from
other bats). Furthermore, informational ambiguity is
only one facet of the underlying problem. Motor action
can be quite violent to receptors, many of which are
located on or near effectors (for example, muscles), and
so movements could desensitize the primary afferents of
a particular sensory channel. The animal would thus be
unresponsive to independently occurring inputs that
followed jarring motor acts.
Because the sensory problems that are introduced
by movements are potentially devastating, any ancestral
species that failed to solve them would have faced an
evolutionary disadvantage. Those species that survived
seem to have overcome the problem with a remarkably
uniform mechanism: the animals nervous systems keep

nature reviews | neuroscience volume 9 | august 2008 | 587

369
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS

Sensory processing stream
The series of neuronal areas
that are involved in analysing
the information acquired by
sense organs.
Efferent
A neuronal projection that
conveys information away from
a structure. The term is often
used when referring to motor
commands.
Decussation
The point where an axon or a
pathway crosses another.
Phylogeny
The evolutionary development
or history of a group of
organisms, often depicted in
family trees.
a Efference copy
Higher
Higher
b Corollary discharge
Higher
track of their movement commands and inform the
sensory processing stream about movements that are immi-
nent. In the terminology of von Holtz and Mittelstaedt,
the animal sends an efference copy signal that is, a
copy of the efferent motor command issued to an effector
to the sensory pathway. Thus, the combination of the
input received from the sensory receptors and the motor
copy that indicates how a movement might influence
the input signal allows a moving animal to resolve the
confusion in sensory input.
FIGURE1a shows the basic circuit in which a sensor
and a series of steps along a sensory pathway convey the
exafferent and reafferent sensory signals to a motor path-
way that impinges on the muscles to produce the appro-
priate movement. A branch from the motor pathway to
the sensory pathway provides the efference copy. The
general strategy involves the coincident production of a
motor command destined for an effector and a motor-
command copy destined for a sensory structure. On
receipt of the copy signal, the sensory structure adjusts
appropriately to minimize, eliminate or compensate for
the sensory consequences of the movement.
The term efference copy implies an actual copy of the
motor command (the efference) that targets the muscles;
this term seemed appropriate for the questions that von
Motor output
Action
Lower Motor neurons Effector
Efference copy Reafference
Lower Sensory neurons Sensor
Exafference
Sensory channel
Motor output
Action
Lower Motor neurons Effector
Holtz and Mittelstaedt were addressing in invertebrates
and for the general analysis of sensory processing that
takes place close to the motor output. However, it has
become apparent that the decussation from motor to sen-
sory areas might occur at any number of levels of motor
control, some of which are remote from the final effector
stage (FIG.1b). In studies on fish, Sperry3 coined the term
corollary discharge (CD) to denote motor-related sig-
nals that influence sensory processing, but his concep-
tion was less specific as to where the branch from motor
to sensory pathways should emerge. In this Review we
compare motor-to-sensory circuits across different spe-
cies and different levels of the nervous system, and so we
use what seems to be the more general of the two terms:
corollary discharge. At a mechanistic level CD can adopt
one of multiple forms depending on how it is used: it can
facilitate, inhibit or otherwise modulate its target.
Our main goal in this Review is to compare corol-
lary systems across the animal kingdom. What is evi-
dent from experiments that have been performed over
the past few decades is that multiple types of CD have
evolved, and that each is particularly well suited to the
problems faced by the species. Two recent reviews have
summarized a few of them4,5. Here we provide a broader
overview of the varieties of CD and place emphasis on
two themes. First, we focus on critical functional dif-
ferences, illustrated by the impact of CD on recipient
sensory structures. Animals that occupy diverse niches
show differences in their use of CD that, we propose, can
be summarized with a functional taxonomy. Second, we
consider the general similarities in CD circuits among
animals. We find that CD circuits conform to a com-
mon neuronal plan, with only minor modifications
in motor source and sensory termination. By classi-
fying each circuit we attempt not only to summarize
recent findings, but also to distill from them the core
circuits for CD that they exemplify. To help illustrate
our points, throughout this article we use FIG.1b and
its colour-coded conventions as a template onto which
each animals particular pathway is represented. We
close by discussing some implications of our current
understanding of CD and considering several questions
that remain unanswered.

A taxonomy for CD
Reafference
As is made evident in this Review, CD lends itself to a
Corollary discharge taxonomic classification scheme (FIG.2). Overall, CD can
be dichotomized into lower- and higher-order functional
(rather than phylogenetic) categories based on its opera-
Higher Lower Sensory neurons Sensor tional impact on the nervous system. The operational
Exafference impact is the way in which the CD signal influences the
Sensory channel
recipient structure, with the aim of achieving sensori
Figure 1 | Efference copy versus corollary discharge. a | A schematic of a motor harmony. The lower- and higher-order categories
sensorimotor circuit composed of a sensory pathway (shownNature in orange) and| Neuroscience
Reviews a motor are subject to further subdivisions that represent the
pathway (shown in brown). Each pathway consists of a number of tiers that represent
more specific functions of the signal.
the complexity of the processing that has been performed and the distance from the
Lower-order CD signalling is used for functions such
periphery. A branch from the motor pathway to the sensory pathway (shown in blue)
provides the efference copy. b | Corollary discharge. Motor-to-sensory signals are not as reflex inhibition and sensory filtration, both of which
confined to exact copies that target early tiers of the sensory channel. They can arise are examples of the control of sensation by the CNS. This
from almost all levels of the motor pathway and can target any tier of the sensory type of CD serves a sentinel function by intervening at
processing stream. These signals are known as corollary discharges (shown as thick various points along a sensorimotor pathway to regulate
arrows). Schematic inspired by von Holst and Mittelstaedt2. the sensory information that enters the system (in the

588 | august 2008 | volume 9 www.nature.com/reviews/neuro

370
2008 Macmillan Publishers Limited. All rights reserved.

Corollary discharge
Lower-order categories Higher-order categories
Reflex Sensory Sensory Sensorimotor
inhibition filtration analysis/stability learning/planning
Central control of sensation Central control of action and perception
Figure 2 | A taxonomic classification of corollary discharge. Corollary discharge can
be classified globally into lower- and higher-order categories according to the function
and operational impact of the signal. Lower-order functions Nature
includeReviews
reflex inhibition and
| Neuroscience
sensory filtration, which are examples of the control of sensation by the CNS. Higher-
order functions include sensorimotor learning/planning and sensory analysis/stability, all
of which are examples of the control of action and perception by the CNS.
case of sensory filtration) and the motor responses that
the information elicits (in the case of reflex inhibition).
Such activity seems to be ubiquitous, as it is necessary for
any animal that is equipped with sensory and motor sys-
tems. Most of the CD circuits that have been identified
in invertebrates are of this variety.
Higher-order CD signalling participates in func-
tions such as sensory analysis and stability, and senso-
rimotor planning and learning. On the perceptual side
it facilitates the contextual interpretation of sensory
information (analysis) and the construction and main-
tenance of an internal representation of this informa-
tion (stability). On the sensorimotor side it facilitates
novel motor-pattern acquisition (learning) and the
execution of rapid movement sequences (planning).
Examples of higher-order CD have been identified only
in vertebrates thus far.
Keeping this taxonomic framework in mind, below
we discuss the guises that lower- and higher-order CD
signals assume in the nervous systems of various spe-
cies. We first consider lower-order CD and investigate its
role in functions related to the central control of sensa-
tion. Such functions are relatively peripheral operations
Mechanoreceptor that are carried out by almost every nervous system to
A receptor that senses physical control the initial inflow of sensory information.
displacement.
Vestibular signal Lower-order CD
A signal that conveys changes Reflex inhibition. The nematode Caenorhabditis elegans
in head orientation, which are (see FIG.3a) has only 302 neurons6 and a relatively sim-
produced by head movements ple behavioural repertoire, which nevertheless requires
or changes in the position of
the head with respect to
coordinative signalling by CD. Its avoidance response
gravity. consists of two main reflexes: forward and backward
progression7. These behaviours are antagonistic: when
Proprioceptive signal the nematode jolts forward owing to a tail stimulus,
A signal that conveys
reafference from the head receptors would send it back-
information about the position
and movement of body parts. wards (FIG.3a). However, inhibitory neurons quickly
silence the second reflex pathway and annul the effect of
Giant command neuron the reafference. Among sensory neurons, interneurons
A motor neuron that is and motor neurons there are reciprocal patterns of con-
common in invertebrate
species and that facilitates
nectivity that inhibit the inappropriate reflex whenever
behaviours such as the rapid- the opposite behaviour is activated8, a CDlike func-
escape response. tion. A comparable problem is faced by Xenopus laevis
nature reviews | neuroscience volume 9 | august 2008 | 589
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS
tadpoles during locomotion, and a CD regulatory cir-
cuitry remarkably similar to that of C. elegans modulates
their avoidance reflexes9.
Gastropods must regulate their reflexes too, but in
a different context from the above examples10 (FIG.3b).
Normally, gastropods such as Pleurobranchaea repel if
contact is made with the tactile mechanoreceptors of its
oral veil (its mouth; FIG.3b, schematic). However, this
withdrawal response does not occur while the animal is
feeding, despite there being comparable tactile stimula-
tion. A CD of feeding commands, originating from a
corollary discharge interneuron (CDI), was found to
intervene during feeding. This effectively silences the
appropriate withdrawal command neurons and inhibits
maladaptive withdrawal from sustenance. For all the
reflex-inhibition examples, these rudimentary CD
circuits can be depicted in our organizing template as
providing routes of contact between lower motor and
lower sensory areas. These examples are thus instances
of CD regulating sensation by reflex inhibition (FIG.3c).
We have considered some specific cases of reflex
coordination that involve the suppression, by CD, of
signals that would drive incompatible behaviours.
Comparable circuits are present in more advanced organ-
isms as well and have a prominent role in regulating the
vestibular and proprioceptive signals that accompany
voluntary movements (BOX1).
Sensory filtration. Many organisms detect the presence
of prey or the approach of predators by monitoring
the energy that is captured by their hair-cell mechano
receptors. These sensory systems detect mechanical
perturbations that arise from events in the surrounding
environment. One organism with a reliance on hair-cell
sensors is the crayfish (see FIG.4a). The crayfish responds
to sudden and unpredicted events by tail-flipping to
safety11. This escape response is triggered by informa-
tion that enters the system through arrays of hair cells
located on its tail and abdomen1214. Signals related to
water or air displacements are reported to a network of
interneurons that synapse on to giant command neurons,
which directly elicit the escape response. In order to
escape only when it is most appropriate to do so, the
crayfish must maintain afferent sensitivity for optimal
threat detection while ignoring false escape cues arising
from reafference. However, repeated reafference could
result in habituation of the reflex, leaving the crayfish
unresponsive to exafferent information.
CD mechanisms allow crayfish to move freely with-
out inadvertently triggering or habituating their reflexive
escape responses15. Activation of their hair cells normally
causes a tail flip, but these same sensors are activated
repeatedly during escape movements and thus must be
regulated during escape. The circuitry contains CDIs
that receive signals from the lateral giant command
neurons, which initiate the escape response (FIG.4a).
The CDIs synapse on to interneurons (called primary
afferent depolarizing interneurons) that presynapti-
cally inhibit primary afferents from the tail hair cells.
This arrangement prevents a dizzying feedback cycle of
movementescape that would otherwise ensue with each
371
REVIEWS

Teleception
Sensory reception that is
specialized for the detection of
distant external stimuli, such as
light, sound and smell.
movement. This example illustrates how CD signals that
originate from a lower motor-control area target sen-
sory neurons positioned at an early stage of the sensory
processing stream (FIG.4c). In terms of function, this
circuit is almost identical to one that has been identified
in the cockroach, another organism with a behaviourally
important hair-cell/escape-response circuit16. In both of
these animals, CD signals prevent reafferent saturation
and inappropriate escape responses.
Fish face similar problems to the preceding inverte-
brate examples. The dogfish, for example, swims by gen-
erating rhythmic sinuous movements of its torso. These
movements induce water turbulence that displaces the
mechanosensitive hair cells of the lateral line system, a
network of mechanoreceptors that line the cephalic and
trunk regions17,18 and detect disturbances in the water
column19. As with the cockroach and the crayfish, these
hairs are stimulated by locomotion. To avoid reafferent
saturation, CDs of swimming commands directly inhibit
the lateral-line hair cells20. In this case the CD signals act
directly on a sensor.
The preceding examples illustrate how CD of loco-
motion modulates pathways that are associated with
contact and teleceptive mechanosensation. Next we assess
the role of CD in regulating sensory pathways associated
with two other teleceptive senses: audition and vision.

a b
Withdrawal Withdrawal
muscles muscles
Oral veil
Tactile
afferents
Front Back

Sensory WMNs WMNs

neurons

Brain
WCNs WCNs

Interneurons

Feeding CPGs

Motor
neurons
Buccal
CDIs CDIs ganglion

Forward Backward
movement movement

Inhibitory Excitatory

c Motor output
Action
Higher Lower Motor neurons Effector

Corollary Reafference
discharge

Higher Lower Sensory neurons Sensor

Exafference
Sensory channel
Figure 3 | Corollary discharge for reflex inhibition. a | The nematode Caenorhabditis elegans and a highly simplified
schematic of the neural circuitry of its locomotor system. Activation of sensory neurons in one portion of C. elegans body
Nature
(either the front or the rear) results in the excitation of motor neurons that drive a movement away Reviews
from | Neuroscience
the stimulus
(backwards or forwards, respectively). Inhibitory connections silence the antagonistic pathway whenever one of these
movements is elicited. Interneurons in the two sensorimotor circuits carry out corollary discharge (CD)like functions
(shown in blue). b | A member of the sea slug family Pleurobranchaea and a schematic diagram of portions of this familys
nervous system, showing how CD during feeding suppresses withdrawal behaviour. During feeding, gastric information
reaches CD interneurons (CDIs; shown in blue) from the feeding central pattern generators (CPGs). The CDIs inhibit the
withdrawal command neurons (WCNs) and prevent tactile stimulation of the oral veil from triggering the retreat response.
c | A basic schematic of the animals CD circuitry is depicted on a common diagram. WMN, withdrawal motor neuron.

590 | august 2008 | volume 9 www.nature.com/reviews/neuro

372
2008 Macmillan Publishers Limited. All rights reserved.

Box 1 | Vestibular and proprioceptive signals
Corollary discharge (CD) signals have a role in regulating the sensation of position,
velocity and acceleration that is mediated in part by the vestibular system and
proprioception. The vestibular system, which is composed of the semicircular canals
and otolith organs of the inner ear, helps to detect head motion and plays a critical part
in maintaining balance. It also drives the vestibulocollic reflex. In response to a sudden
displacement of the head, this reflex quickly restabilizes the head through activation of
the neck musculature. There are times, however, when a vestibulocollic response is
maladaptive, such as when an animal makes a willful attempt to move its head. It has
been demonstrated in monkeys that CD of head-movement commands suppresses
neurons of the vestibular nucleus during voluntary head movement, thus preventing
the inappropriate reflex response80. These CDs permit differential processing of active
and passive head movements by the vestibular nuclei, another case of CD acting in the
service of reflex inhibition.
Proprioceptors are found throughout the bodies of invertebrate and vertebrate
organisms and provide continual information about body position and orientation.
These receptors operate in static or dynamic contexts, signalling when parts of the
body are at rest or are moving from one configuration to the next. It is in the latter
context that CD has a role. Inputs from proprioceptors must be regulated during active
movement81 for two main reasons. If the proprioceptors are too sensitive, they could be
triggered by self-movements and initiate inappropriate reflexes. If they are not
sensitive enough, they could fail to protect the limb (for example, if an unexpected
obstacle is struck). One type of proprioceptor, the muscle spindle, is regulated directly
through projections from gamma motor neurons that receive CD of the movement
commands sent to muscle-innervating alpha motor neurons82. More generally, CD
signals influence proprioceptive processing at multiple points in the neuraxis, ranging
from the spinal cord83 to the cortex84.
Sound-producing and -receiving organs allow
organisms to exploit the physical characteristics of the
surrounding environment by inducing and detecting
pressure fluctuations. Among the many purposes of
sound for animals is communication between conspe-
cifics21,22. A serious problem that is faced by organisms
that are equipped with organs of sound production,
however, is the substantial input that their sonic emis-
sions impart to their auditory sensors. Following a sonic
event, the auditory circuits could be overwhelmed; this
would render the animal temporarily deaf to indepen
dently incoming sounds. CD mechanisms devoted to
sensory filtration have evolved concurrently with the
motor and sensory systems that are involved in acoustic
communication. The solution for this problem seems
to be ubiquitous, as evidenced by the examples of the
cricket and the marmoset. Despite these animals occu-
pying very different niches, they both adopted the same
solution: tight coordination between their auditory and
vocal systems by phasic CD that minimizes auditory
reafference.
Crickets (see FIG.4b) chirp by rubbing their forewings
together, a process that is known as stridulation23. They
hear through a tympanate membrane on their forelegs
that is only millimeters from the site of stridulation24.
Tympanate membrane
A crickets auditory system must therefore deal with
A thin membrane that detects
sound (also known as the ear a steady barrage of acoustic signals generated by the
drum). chirping. To prevent desensitization and ensure maximal
attunement to environmental events, the cricket selec-
Gain tively screens the signals that enter the system. A CDI
An inputoutput ratio that
defines a neurons
that performs this sensory filtration operation has been
responsiveness to incoming identified25 (FIG.4b). Positioned at the anatomical inter-
signals. face of the auditory and motor systems (an arrangement
nature reviews | neuroscience volume 9 | august 2008 | 591
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS
that is optimal for lower-motor to sensory-neuron con-
tact), the cricket CDI is well suited to perform selective
filtration operations that are coupled to motor activity.
Intracellular recordings revealed that the CDI fires
nearly simultaneously with wing motor-neuron bursts
and exerts a phase-locked inhibitory influence on the
interneurons of the auditory system25. By rhythmically
imposing and lifting a blockade in phase with the song
cycle, the cricket can chirp and remain receptive to the
auditory environment.
Marmosets encounter the same problem as crickets:
in principle, the sounds that they make should affect their
hearing26. A protective mechanism is observable in the
marmoset primary auditory cortex, where many neurons
are suppressed during self-vocalizations27,28. Suppression
begins ~200 ms before vocalization and continues for
its duration. Both the predictive nature of the effect and
the dependence on self-vocalization implicate CD as the
cause of the suppression. In primates there are reciprocal
connections between the primary auditory cortex and
motor regions of the cerebral cortex2931, and activity
in the frontal cortex can precede vocalization by up to
1s (ref. 32). Because the suppression begins ~200 ms
before the vocalization, the effect is most likely medi-
ated by direct, rather than indirect, inhibition from one
of these areas. As such, this could be a case in which
CD interconnects motor and sensory areas that occupy
comparable tiers of a sensorimotor pathway.
As with the sense of hearing, the sense of vision could
theoretically be severely compromised by an animals
behaviour. However, audition and vision differ with
respect to the nature of the motor interference. In audi-
tion, problems arise from reafferential acoustic noise.
In vision, problems arise not from autogenic energy
(energy originating from the animal itself for example,
the sound that results from a vocalization) but from the
actions of the animal. If the eyes or the head move,
the retinas go along for the ride. The retinas contain
sheets of photoreceptors that are passively responsive
to patterns of impinging photons. The moving of the
retinas called a gaze shift enables an animal to
sample a new part of visual space but, as we shall see,
every gaze shift has its costs.
A gaze shift can be achieved with a saccadic (quick
and discrete) eye movement and/or with a saccadic
head movement. Saccadic gaze shifts cause, at least in
principle, two serious reafferential problems. The first
is that the speed of the movement could transiently blur
the image. The second is the general displacement of the
visual scene from before to after the movement, which
could cause the scene to seem like it is jumping from one
place to another.
The transient blur that occurs during saccades is
minimized by a process called saccadic suppression33.
In this process, a CD of the gaze shift lowers the gain of
target visual structures to reduce information transfer
between visual areas33. Saccadic suppression mecha-
nisms have been identified in a host of different spe-
cies, such as pigeons, chickens, locusts, monkeys and
cats3438. Although the suppressive systems of these
species differ with respect to the anatomical structures
373
REVIEWS

concerned, each circuit involves a CD operating as a Higher-order CD

gate that regulates signal traffic from the periphery
by definition a lower-order CD function. The second
problem caused by saccades, the sudden jump of the
visual scene, seems to be solved by other mechanisms
(discussed below).
In the examples described up to now, the CD arises
from a range of motor levels but generally targets periph-
eral sensory levels. In the cockroach, the crayfish and
the cricket, the CD circuits discussed connect the lower
motor tier to sensory neurons at or near the periphery.
Overall, it seems that lower-order CD has a common
functional impact: transient, protective inhibition of
sensory networks. Next we turn to higher-order CD, the
other broad taxonomic category of function.
Higher-order CD circuits are those that enable predictive
control for perceptual cohesion and action sequencing.
CD for perceptual cohesion allows an organism to
move and yet continue to experience the world as it is
(stable and continuous) rather than as it is sensed at the
receptor level (in a chaotic and piecemeal fashion). CD
permits specific brain structures to carry out appropriate
adjustments in anticipation of the sensory input, and
enables them to construct a cohesive representation of
the world. The adjustments involve changes to neuronal
processing, such as sudden shifts in sensory sensitivity
(an example is described below). In addition, animals
can string together movements with great speed,
such that future movements are planned before prior

a Sensory filtration b

MoGs ANs
Prothorax From
tympanum

CDIs
SGs LGs Mesothorax MNs
CPG
CDIs To wings
PADIs
SAs SIs

Metathorax

c
Motor output
Action
Higher Lower Motor neurons Effector

Corollary Reafference
discharge

Higher Lower Sensory neurons Sensor

Exafference
Sensory channel

Figure 4 | Corollary discharge used for sensory filtration. a | The illustration on the left is of a crayfish species,
Procambarus clarkii. The crayfish escapes from potential threats by producing a rapid tail-flipNature
response. Mechanoreceptors
Reviews | Neuroscience
lining its abdomen detect events in the water column and elicit the escape behaviour. Corollary discharge (CD) signals of
movement commands protect the afferents to the mechanosensory escape system from maladaptive activation and
desensitization. The schematic on the right depicts the circuitry that is involved in producing the crayfishs tail-flip
response. Mechanical information enters the system through sensory afferents (SAs) and reaches the lateral giants (LGs) by
way of sensory interneurons (SIs). The LGs communicate with segmental giants (SGs) and movement generators (MoGs)
that activate flexor muscles in the abdomen. CD interneurons (CDIs) are activated by the SGs and convey signals to
primary afferent depolarization interneurons (PADIs) that inhibit the SAs. This transient inhibition briefly silences the
mechanosensory pathway and prevents tail-flip-induced reafference from generating further tail flips. Part c shows a
circuit that represents an example of CD signals that originate from a lower motor cortical area and target sensory
neurons at an early stage of the processing stream. b | The illustration on the left is of a cricket species, Gryllus bimaculatus.
Crickets communicate with one another by chirping. Chirps are generated by rubbing their forewings together, a process
that is known as stridulation. CDs of the wingbeats phasically inhibit components of the auditory system and prevent
saturation and desensitization. The schematic of the cricket thorax on the right depicts how CD coordinates the song
and auditory systems. The activity of the central pattern generator (CPG) drives the wing motor neurons (MNs) that
produce song (musical notes). Signals are routed concurrently from the CPGs to the CDIs. Collaterals of the CDIs synapse
on to the axon terminals of auditory primary afferents and on to the cell bodies of auditory interneurons (ANs) in the
prothorax. CPG-induced CDI activity rhythmically imposes and lifts an inhibitory gate on the auditory systems and
restricts auditory traffic from the tympanum. Part c shows how, for the system in part b, the point of contact between the
motor and sensory systems is at the lower levels of the idealized sensorimotor circuit.

592 | august 2008 | volume 9 www.nature.com/reviews/neuro

374
2008 Macmillan Publishers Limited. All rights reserved.
 REVIEWS

Box 2 | Shifting receptive fields and corollary discharge

Many neurons of the primate visual system alter their a Normal
visual sensitivity at the time of a saccade (see figure, Fixation Just before saccade Saccade
part a). Visual responsiveness shifts before the saccade
from the neurons current visual receptive fields (RFs) to
the location (the future field; FF) where the RFs will rest
after the saccade. Because this process enables the FF
(in the figure, the pepper) to be sampled both before
and after the saccade, it is thought to contribute to a Future field
percept of visual stability despite the jerkiness of Receptive
saccadic eye movements. One brain area that contains field
neurons with this property is the frontal eye field (FEF),
a cortical structure that is involved in vision and eye- Receptive
movement control42. FEF neurons that shift their RFs field
emit their visual-evoked burst of spikes in the middle of
the saccade, on average. Perceptually this is
advantageous for two reasons. First, it ensures that b Macaque brain
there is no premature perception of the world shifting
Recording
before the saccade. Second, it ensures that following Inactivation
the saccade the system does not have to wait for visual
reafference to arrive, which can take tens of
milliseconds. A CD pathway that ascends from the FEF
superior colliculus (SC) to the FEF through the medial
dorsal nucleus of the thalamus (MD) provides
important saccadic information that triggers the shift
(see figure, part b). Inactivation of this pathway at its MD
point of relay in the thalamus reduces the ability of
these FEF neurons to shift their RFs (see figure, parts b SC
and c). In part c of the figure, therefore, instead of
sensitivity shifting from the apple to the pepper during
the saccade, sensitivity largely remains at the apple. It c During inactivation
is only after the saccade that the neuron is fully
responsive to the pepper, just like any other neuron with Fixation Just before saccade Saccade
a classical receptive field. It has been proposed that
this would have a considerable impact on perception
during saccadic eye movements40. During inactivation,
the visual scene would seem to jump with each saccade,
a consequence of CD impairment and shifting RF
reduction. With the pathway disrupted, monkeys trained Receptive
to detect visual motion during saccades should report field
that stationary visual stimuli move with each saccade.
This would be the first direct test of the CD/shifting-RF
Receptive Receptive
visual-stability hypothesis. In parts a and c of the field field
figure, the red dot denotes the fixation point.

Nature Reviews | Neuroscience

movements finish. Higher-order CD is crucial in this
process of action sequencing because it provides internal
feedback that escapes the afferent lags and bandwidth
constraints of sensory or external feedback loops. Next
we discuss how these principles are realized in various
animals, first examining higher-order CDs devoted to
sensory analysis and stability.
Sensory analysis and stability. As discussed above, in
visuomotor behaviour, mechanisms of saccadic sup-
pression that are mediated by lower-order CDs reduce
the blur across the retinas. But even if blur were elimi-
nated, a substantial problem would remain: the image
on the retina is fully displaced by each saccade. Were we
to perceive the world exactly as we sense it, the visual
scene would seem to leap from place to place dozens of
times per minute. It is crucial that the brain be able to
distinguish what aspects of the jumpy visual inflow are
artefactual (due to saccades) as opposed to real (due to
changes in the world). CD has long been thought to have
a role in this process by providing advance warning of an
imminent movement39. With this warning, the brain can
enact a compensatory procedure to cancel the percept of
visual-scene displacement that, in principle, should be
perceived by an animal each time it moves its eyes. This
is most problematic for species that frequently make
saccades (the primate is a classic example).
In addition to CD, there is another candidate mecha-
nism for perceptual stability called spatial updating.
Spatial updating involves pre-saccadic changes to a
visual receptive field (RF). Although a typical RF is
firmly retinotopic and samples a new part of the visual
field (the new RF) only after the eye moves, a shifting
RF is dynamic and starts sampling the new RF location
even before a saccade (BOX2). Because a saccade can have
any direction and any amplitude, a neuron with a shift-

nature reviews | neuroscience volume 9 | august 2008 | 593

375
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS
Whisking
The act of tactile exploration in
which a whisker is rhythmically
swept across an object.
Vibrissae
Specialized long hairs located
near the mouth of most
mammals that are used for
tactile exploration.
594 | august 2008 | volume 9 www.nature.com/reviews/neuro
ing RF requires the appropriate CD to calculate the new
RF. At least some of this CD arises from a pathway that
ascends from the midbrain40,41 to the frontal eye field, a
cortical structure that is involved in visual processing
and eye-movement control42. Interestingly, this same
CD pathway assists in motor planning (discussed in the
sensorimotor learning and planning section).
Similar perceptual issues hold for animals for which
vision is a low priority. Rats explore the world largely
by whisking tactile objects of interest43. How does the
whisking rat construct a stable representation of the
world from the variable and volatile inputs that flow
in from the vibrissae ? One clue to this puzzle was
the identification of neural activity in the rat barrel
cortex that was modulated by the act of whisking and
was correlated with whisking amplitude44. The activity
persisted during inactivation of the facial motor nerve
to rule out proprioception. Because the signal was of
central origin, a CD seems to have been at work. Such a
signal could provide motor context to the barrel cortex
and assist in the interpretation of vibrissal inputs, a step
towards constructing a stable representation. Motor
pathways are known to target the barrel cortex from the
motor cortex, the superior colliculus and other brain-
stem areas45,46. Overall, the vibrissal system comprises a
web of sensorimotor loops that span much of the neur
axis. Any of these loops could convey CD signals to
the barrel cortex and thus mediate the effects observed
in the above study44 and the assembly of coherent
information from whisking events.
It is particularly intriguing to consider active sensing
systems (those in which energy generated by an animal
is used to probe the environment). Animals that are
equipped with these exotic signalling modes exercise
considerable control over the properties of the emitted
signal (most of which are species- and modality-specific)
and rely on complex predictive mechanisms for carrying
out detailed analysis of the echo (the returned carrier
signal)47. Two common modes of active sensing are used
by fish and bats, and CD is essential in both.
We first consider the mormyrid, a weakly electric
fish (see FIG.5Aa). The mormyrid uses self-generated
electrical currents for communication and object iden-
tification. These currents flow from the mormyrids
electric organ into the surrounding water column and
are analysed on their return by arrays of electroreceptors.
A number of neural processes have evolved that allow
the mormyrid to both extract the maximum information
content from the return signal and avoid interference
arising from reafference48.
Mormyrids sense the electromagnetic environment
using electroreceptors (FIG.5Aa) . Primary afferents
from these electroreceptors converge on layers of a
medullary structure known as the electrosensory lat-
eral line lobe (ELL)49. Separate mechanisms involving
CD accommodate and protect the highly variegated
electroreceptor systems50,51. For the ampullary system,
an adaptable CD mechanism was found to be at work52.
Each time a motor command exits the electric-organ
command centre, a CD of the motor command also
exits, targeting the ELL. One study dissociated reafferent
2008 Macmillan Publishers Limited. All rights reserved.
responses of the ELL from effects of the electric-organ
CD by pharmacologically decoupling the electric-
organ motor command centre from the electric organ52.
This preparation effectively silenced the electric organ:
no electrical currents were produced. Nevertheless,
the motor command and CD were still produced by the
motor command centre. ELL neurons responded to
the CD in a manner that was dependent on recent expe-
rience. If no external signals (that is, electrical current
reafference) preceded the motor command, no response
was elicited by the CD; however, if external electrical
stimuli were paired with motor commands, responses
that were opposite in sign and equal in duration to the
stimulus interval gradually emerged. If external stimuli
were removed, these responses were extinguished. The
CD of the ampullary system is effectively a plastic photo
negative of the expected reafference, and it varies as a
function of recent sensory experience. A modifiable
CD is evolutionarily advantageous because mormyrids
navigate diverse aquatic environments that differ subtly
in conductivity and resistivity. Interestingly, the
mormyrid has two additional electroreceptor subtypes
(mormyromasts and knollenorgans) that are suscep-
tible to electric-organ discharge (EOD) feedback and,
accordingly, compensatory CD systems are in place
to accommodate them50,51. These systems differ from
the ampullary system in that they require CDs that
amplify (mormyromast) or inhibit (knollenorgan) the
respective electroreceptor during EOD. As the mormy-
romast is specialized for electrolocation and the knollen
organ is specialized for conspecific communication,
CDs for amplification and suppression, respectively, are
appropriate. As the command centre and the ELL
are located in the hindbrain, the CD is shown in FIG.5Ab
as linking lower motor and lower sensory levels.
Outside of aquatic environments, electrical probes
have little utility. However, air is an excellent carrier
of acoustic probes, a fact that is exploited by bats (see
FIG.5B). Bats explore the environment by emitting high-
intensity, high-frequency beams of sound and then com-
paring spatiotemporal aspects of the returning echo with
the emitted sound53,54. Like the mormyrid, this active
process of emission and comparison enables the bat to
assemble an image of the external world5557 and thus
successfully navigate and predate.
Neurons that are sensitive to specific emissionecho
intervals have been identified in the inferior colliculus
(FIG.5Ba) and, together with CD, have a key role in the
neural circuitry for auditory-scene analysis54,58. The
echo-sensitive neurons are thought to be primed by CDs
of vocalizations that open time windows of echo analy-
sis. If the echo returns within a specific time frame, then
neurons that have been primed for analysis within that
particular window will respond optimally and convey
information to perceptual structures for further analysis.
Such signals convey information that is related to the
distance of approaching obstacles and the dimensions
of environmental objects. These signals could arise from
cortical or subcortical areas, so in FIG. 5Bb we represent
the putative CD signals as emerging from both higher
and lower motor levels.
376
 REVIEWS

Aa Mormyrid brain
Cerebrum

ELL

Motor Tectum
EOCD
command
Electric
organ Command
nucleus

Electrosensory
input

Ab Motor output
Action
Higher Lower Motor neurons Effector

Corollary
discharge Reafference

Higher Lower Sensory neurons Sensor

Exafference
Sensory channel

Ba
Priming Higher-order
Vocalization
corollary perceptual
centre
discharge centre

Vocalization Emissionecho
motor Inferior interval-specific
command colliculus auditory signals

Echo-sensitive
neurons
Vocalization Echo

Bb Motor output
Action
Higher Lower Motor neurons Effector

Corollary
discharge Reafference

Higher Lower Sensory neurons Sensor

Exafference
Sensory channel

Figure 5 | Corollary discharge used for sensory analysis and stability. Aa | The illustration on the
Nature left is|of
Reviews the
Neuroscience
mormyrid species Gnathonemus petersii. The mormyrid generates electrical signals to probe the aquatic environment.
Multiple types of corollary discharge (CD) of electrolocation commands allow the fish to gate, amplify or predict the
return signal. The schematic of the mormyrid brain on the right illustrates pathways of the electrosensory system. Electric
organ CD (EOCD; shown in blue) from the electric-organ motor command centre (which generates the motor command,
shown in purple) reaches cellular networks of the electrosensory lateral line lobe (ELL), following which a host of
interactions with electrosensory input (shown in orange) occur. Ab | The CD circuit that connects the command centre
and the ELL is represented as a link between lower motor and lower sensory levels. Ba | The illustration on the left is of the
bat species Rhinolophus rouxi. During high-speed flight, this bat uses sound to hunt. It compares a CD of the sonic probe to
the measured echo to interpret the acoustic input. The schematic on the right shows how CD is used in this system. CD
(shown in blue) represents the efferent motor command and innervates the inferior colliculus, where it is compared with
the echo input. Differences between the CD and the input (shown as dashed lines) are analysed by higher-order centres
to estimate the size, location and speed of the object that caused the echo. Bb | The CD signals could arise from any
number of subcortical and cortical structures, corresponding to multiple pathways emerging from both higher and lower
motor levels.

nature reviews | neuroscience volume 9 | august 2008 | 595

377
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS
In summary, these examples illustrate the diversity
of sensory systems that are assisted by higher-order CD.
In all of these different species, CDs aid the construction
of an internal representation of the environment. We
now consider motor planning and learning, the other
category of higher-order CD, in the saccading primate
and vocalizing bird.
Sensorimotor learning and planning. Among the most
stereotyped movements generated by primates are sac-
cadic eye movements. As a primate looks around a scene,
it must continuously update its internal record of the
current saccade to facilitate planning the next. CD has
repeatedly been implicated in this process5964.
Electrophysiological recordings from a transiently
inactivated circuit in the brain have shed some light
on how CD is used for saccadic sequencing in rhesus
monkeys62,63,65 (FIG.6Aa). Neurons that connect the mid-
brain to the prefrontal cortex transmit bursts of action
potentials that encode saccadic parameters. When
this pathway was inactivated by muscimol, a GABAA
(-aminobutyric acid receptor type A) agonist, behav-
ioural deficits in the simplest saccadic sequencing task
were observed. In this two-step task, the monkey was
required to saccade to the locations of two rapidly flashed
sequential targets. This particular task was chosen
because it required the monkey to retrace the flash
sequence without the aid of sensory feedback. The monkey
had only the internal feedback provided by CD to inform
it of its current eye position. If CD of the first eye move-
ments were impaired by the inactivation, then the mon-
key would make incorrect second saccades owing to the
absence of the feedback signal. The monkey would be
oblivious to the execution of the first saccade and would
perform the second saccade as if it had never made the
first. The results conformed to this prediction, revealing
a loss of CD. This is an example in which CD emerges
from a lower-level motor structure and ascends the
neuraxis to impinge on a higher-level executive centre
(FIG.6Ab). As discussed above and in BOX2, this pathway
also has a sensory analysis and stability function.
Vocalizations in birds are another prime example
of action sequencing. Singing birds are thought to use
updated, internal records of current phonations to
generate subsequent phonations. Bilateral feedback
pathways have been identified in the singing finch that
might convey such vital information66. These pathways
ascend from midbrain and medullary vocal-control
nuclei up to the forebrain song system. Although their
functional role in CD has not been tested directly using
causal methods, they carry signals that are consistent
with CDs of vocalization motor commands.
CD is especially important while vocalization is being
learned. Male juvenile songbirds (see FIG.6Ba) exhibit a
stereotyped song that emerges gradually following sev-
eral developmental stages67. The first stage is a sensory
phase in which the juvenile bird listens to the songs of
other adult birds known as tutors. The second stage con-
sists of sensorimotor training in which the bird begins
to sing spontaneously with syllables of varying intensity
and motifs of high variability. The final phase is known
596 | august 2008 | volume 9 www.nature.com/reviews/neuro
2008 Macmillan Publishers Limited. All rights reserved.
as crystallization, and this is the phase in which mature
song emerges. CD seems to have a key role during the
second phase of vocal learning68,69. The bird is thought
to fine-tune its song during this phase by comparing
auditory feedback of its own vocalizations with copies
of tutors songs that are stored in memory. The errors
between the memory and the actual feedback are cor-
rected through an iterative cycle of vocal production,
followed by adjustment of the motor programme. The
bird has to keep track of its own motor commands in
order to properly adjust and calibrate its subsequent
commands. This is where CD plays a part. Intricate
feedback loops that course throughout the avian song
system are thought to convey the auditory, memory and
CD signals70,71 (FIG.6Ba). The feedback loops execute a
delay function by increasing the distance over which
CD that is issued from vocal centres travels before it
collides with auditory feedback72,73. Evidence for this
CD function was recently found in neurons in the high
vocal centre (HVC) that fired with the same latency
irrespective of whether the bird was actively singing or
passively listening74. As the HVC is a premotor structure,
the neurons would be expected to fire earlier for singing
than for listening. The comparable latencies that were
found for singing and passive listening imply that there is
a delay function performed by a CD. Because a major site
of comparison is the HVC, which is a higher brain struc-
ture of the auditory forebrain, in FIG. 6Bb we represent
the CD as a signal between higher levels.
In these examples, CD is used for the higher-level
functions of planning and learning actions. Accordingly,
on exiting the motor domain, the CD signal impinges on
higher-level structures that are highly sensory and/or
executive in nature. As a result, appropriate behaviours
can be prepared for the future (planning) and modified
based on the lessons of the past (learning).
Conclusions and future directions
We have reviewed the operational impact of movement
on sensation, and the ways in which nervous systems
use CD signals to deal with movement-induced sen-
sory problems. CD systems have evolved in tandem
with the sensory systems of organisms that operate in
diverse environments. The CD and the sensory systems
are both shaped by the problems that are encountered
by each organism, resulting in many different uses
for CD signals. We have summarized these manifest
differences with a functional taxonomy. After decades
of experimentation, we now know a considerable
amount about these CD circuits in both invertebrate
and vertebrate organisms. In general, CD seems to be
necessary for the proper function of nearly all sensory
systems, and it exists in several guises at multiple levels
in diverse species (perhaps all species). Some common
alities and subtle differences are apparent from the
reviewed examples.
At one level (the lower-order level) CD is a discrimi-
natory mechanism that prevents maladaptive responses
and sensory saturation by restricting or filtering infor-
mation. Functions such as reflex inhibition and sensory
filtration are tightly controlled by this mode of signalling,
378
 REVIEWS

Aa Macaque brain
Visual input
FEF

CD
MD

SC

To saccade-
generating
circuitry

Ab Motor output
Action
Higher Lower Motor neurons Effector

Corollary
discharge Reafference

Higher Lower Sensory neurons Sensor

Exafference
Sensory channel

Ba Finch brain HVC

CD
LMAN
L

RA X

nXIIts DLM Uva

To syrinx

Bb Motor output
Action
Higher Lower Motor neurons Effector

Corollary
discharge Reafference

Higher Lower Sensory neurons Sensor

Exafference
Sensory channel

Figure 6 | Corollary discharge used for sensorimotor planning and learning. Aa | The illustration on the left is of the
macaque monkey species Macaca mulatta. The macaque monkey visually explores its arboreal environment with rapid
sequences of saccades. Corollary discharges (CDs) permit it to plan such sequences in rapid succession and enable it to
predict the visual outcome for purposes of perceptual stability. The schematic of the macaqueNature Reviews
brain on | Neuroscience
the right illustrates
the course of the CD (shown in blue): it ascends from the superior colliculus (SC) to the frontal eye field (FEF) by way of the
medial dorsal nucleus of the thalamus (MD). Ab | This pathway is an example in which CD emerges from a lower-level
motor area and targets a higher-level sensory area. Ba | The illustration on the left is of the songbird Poephila guttata (a
finch species). The developing male finch progresses through a series of song-learning stages that conclude with the
appearance of a mature, fully formed song. The schematic of the finch brain on the right depicts the circuitry and nuclei of
the avian song-learning system. Intricate feedback loops that reside in the finch forebrain are involved in the song-
learning process. CD has been proposed to course through some of these pathways and enable sensorimotor comparisons
to occur within appropriate temporal windows. Bb | The major site of comparison is proposed to reside in the forebrain;
this CD pathway would correspond to contact between higher motor and sensory levels. DLM, medial nucleus of the
dorsolateral thalamus; HVC, high vocal centre; L, field L; LMAN, lateral magnocellular nucleus of the anterior neostriatum;
nXIIts, tracheo-syringeal portion of the hypoglossal nerve nucleus; RA, robust nucleus of the archistriatum; Uva, uvaeform
nucleus of the thalamus; X, area X.

nature reviews | neuroscience volume 9 | august 2008 | 597

379
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS
which is a type of access control. Computationally,
the information content of these CD types is probably
minimal, as they generally implement a gain mechanism
that modulates a reflex or gates the inflow of sensory
information at the periphery. As such, these signals seem
to be one-dimensional in nature: time is the most crucial
variable. This point is illustrated best in the nematode
and gastropod examples of reflex inhibition. Here, CD
intervenes at the precise time of the motor act to prevent
an antagonistic reflex response. Timing is important too
in the sensory-filtration examples. In the cricket, for
instance, CD is synchronized rhythmically with singing
to protect the auditory system. In short, each subtype of
lower-order signal provides information about when the
sensory input that is elicited from the motor act should
occur. In other words, it is not so much what the signal
says that matters, but when it is said.
We explained that lower-order CD circuits are illus-
trated best in invertebrates. The main characteristic of
these circuits is the presence of a single interneuron, the
CDI, which exerts inhibitory actions on various cellular
components and controls interactions between motor
and sensory circuitries through its suppressive influ-
ences. CDIs have been identified in Pleurobranchaea,
crayfish and crickets. Suppressive influences by an
inhibitory CD are also found in vertebrates, and it
is tempting to speculate that in some species these
operations could also be carried out by a type of CDI.
CD is also present in the animal kingdom as a higher-
order mechanism that mediates sophisticated predictive
computations. This type of CD underwrites complex
behavioural phenotypes and cognitive operations, such
as motor sequencing, sensorimotor learning and per-
ceptual stabilization. Higher-order CD operates in the
context of internal feedback circuitry, which involves
multiple components and spans various levels of the neur
axis. Unlike CD that is involved with the central control
of sensation, which is almost solely time-modulated,
CD that is involved with the central control of action
and perception carries information that represents a
number of variables. In the monkey, for example, CD
in the visuosaccadic system encodes spatial as well as
temporal information about imminent saccades. The
songbirds CD carries rich information too, particularly
about song structure. In general, higher-order CDs are
multidimensional and encode more parameters than just
time. As we have described, the exact content and func-
tional role of the signal (whether the CD is the prediction
or whether it enables the recipient substrate to generate
the prediction) varies from species to species and from
system to system. But, in the end, what is said is just as
important as when it is said. This architecture seems to
be exclusive to higher vertebrates and might represent a
later phylogenetic advancement that is particular to the
demands of vertebrate life.
In all cases the corollary systems of remarkably
diverse species could be easily accommodated into a
common template that illustrates their underlying simi-
larities. For the species that use reflex inhibition and sen-
sory filtration, which are both cases of lower-order CD,
the motor-to-sensory crossing points were schematically
598 | august 2008 | volume 9 www.nature.com/reviews/neuro
2008 Macmillan Publishers Limited. All rights reserved.
similar (that is, they were all lower-motor-area-to-
sensory-area neurons). Although the CD signals of the
mormyrid and the bat, two sensory analysis/stability
cases, emerged from differing motor tiers, their sen-
sory-termination level was comparable (lower sensory
areas). Similarly, the two illustrative sensorimotor spe-
cies (a bird and a monkey) had different sites of motor
emergence but common points of sensory termination
(both terminated at higher sensory areas). An implicit
assumption in this comparison is that the sensory target
occupies increasingly higher tiers as one ascends from
lower-order CD through the stages of higher-order CD
(compare the nematode with the songbird). This illus-
trates the important point that although Sperrys original
conception of CD matches the general flow of informa-
tion from motor systems to sensory systems throughout
the animal kingdom, it seems to be inappropriately sim-
plistic to use a single term to describe the signal. There
is no single type of CD rather, there are numerous
subtypes that correspond both to anatomical levels of
the source and the target and, as we have emphasized,
to functional utilities.
As a next step, one could break down the taxonomic
subtypes even further. One way to do this would be to
consider underlying mechanisms. Not all instantiations
of CD are equal in their actions: CD signals can excite,
inhibit or modulate their targets. Although finer-grained
elaborations of the taxonomy would be beneficial, the
point of the present Review was to synthesize all of
the available data into what we see as the most essential
functional taxonomy.
We are still at an early age of CD research. Future
studies should search for signs of CD in sensory domains
that have been neglected, such as olfaction and gusta-
tion. There is evidence that components of the olfactory
system are modulated by motor signals during the act of
sniffing75. Are these motor signals CDs? When animals
masticate, do CDs modulate components of the gusta-
tory system and thereby enhance taste discrimination?
We should also further test the idea that aspects of cogni-
tion, such as thinking and decision making, use a type of
CD for cohesion of self-identity76,77.
At the systems level, work should be devoted to
understanding inter-areal circuits that mediate CD in the
behaving animal. To date, only a few large-scale circuit-
level investigations of CD that combined behaviour with
physiology have been performed38,40. We know quite a bit
about the anatomical connections that probably convey
CD for example, a motor area A that projects to a
sensory area B. But, until we introduce electrodes and
record from these circuits during self-movements of the
subjects, we can only speculate about their role in CD
function.
At the more reductionist level, researchers should
attempt to isolate CD mechanisms at ever-finer reso-
lution, towards the cellular and molecular levels. The
recent sequencing of various animal genomes is par-
ticularly exciting, as CD circuits could be manipulated
genetically at various stages of an animals life cycle. This
could shed light on how CD circuits develop, an area that
at present is poorly understood.
380
 REVIEWS

Another outstanding question concerns the range
of functional utilities of a given CD signal. A single
CD circuit can mediate many different functions. One
example we reviewed was from the visuosaccadic sys-
tem of the primate40,62. In this system, a CD circuit has
a role in sensorimotor planning as well as in sensory
stability (compare FIG. 6A with BOX2). What is the con-
nection between these two differing processes and the
CD circuit? Is there more than one CD signal coursing
through the circuit, or is there a single, multi-purpose
CD signal? Does behavioural context play a part in
determining what computation is implemented by the
CD, and its functional interpretation by the recipient
sensory structure? Similar CD complexities are likely
to be present in other sensorimotor systems. Does the
whisking rat use the same CD signal for sequencing its
whisker movements as for constructing a stable tactile
image of the whisked object? Or does it use different
CD signals for each functional application? These
perplexing questions remain to be addressed by fur-
ther circuit-level interventions that attempt to disen-
tangle the multifarious uses nervous systems have for
CD signals.
Finally, the ultimate goal is to discover how CD
influences perception. Experiments thus far have shown
that inactivation of CD pathways can alter behaviour,
and covert perceptual changes might accompany these
behavioural changes. Eye movements in a two-step
task, for example, change if CD about previous move-
ments are impaired. Comparable experiments could be
performed in other animals, such as rats that have been
trained to whisk an object in a particular sequence. Do
such animals, with their CD circuits interrupted, expe-
rience a sensory world that fractionates from stable to
jumpy? How could we determine this? Although it is
challenging to assess what an animal is perceiving, in
rhesus monkeys, for example, it seems to be possible78,79.
Now that the basic layout of many CD circuits has been
established, it will be exciting to manipulate the circuits
while probing animals with careful psychophysical tests
to determine how perception is altered.
Whatever is uncovered by future studies of CD,
it seems safe to say that we can currently recognize at
least one fundamental principle. All animals, from the
humble nematode to the cognitively advanced primate,
require the type of signalling that is enabled by CD. In
addition to the usual flow of information from sensory
systems to motor systems, there is extensive signalling
in the opposite direction by motor systems reporting
their activities to sensory structures. It is this coordina-
tion between the two systems that makes it possible to
analyse the world while moving within it.

1. Poincar, H. Science et methode (Flammarion, Paris,
1897).
2. Holst, E.V. & Mittelstaedt, H. The reafference
principle. Naturwissenschaften 37, 464467 (1950).
3. Sperry, R. Neural basis of the spontaneous optokinetic
response produced by visual inversion. J. Comp.
Physiol. Psychol. 43, 482489 (1950).
References 2 and 3 are two groundbreaking papers
that were published independently and nearly
simultaneously. They were the first to propose in a
rigorous manner, and with supporting experimental
evidence, that motortosensory feedback has a
critical role in regulating animal behaviour.
4. Cullen, K.E. Sensory signals during active versus
passive movement. Curr. Opin. Neurobiol. 14,
698706 (2004).
5. Poulet, J.F. & Hedwig, B. New insights into corollary
discharges mediated by identified neural pathways.
Trends Neurosci. 30, 1421 (2007).
6. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S.
The structure of the nervous system of the nematode
Caenorhabditis elegans. Philos. Trans. R. Soc. Lond. B
Biol. Sci. 314, 1340 (1986).
7. Rankin, C.H. Interactions between two antagonistic
reflexes in the nematode Caenorhabditis elegans.
J. Comp. Physiol. A 169, 5967 (1991).
8. Chalfie, M. etal. The neural circuit for touch sensitivity
in Caenorhabditis elegans. J. Neurosci. 5, 956964
(1985).
9. Sillar, K.T. & Roberts, A. A neuronal mechanism for
sensory gating during locomotion in a vertebrate.
Nature 331, 262265 (1988).
10. Davis, W.J., Siegler, M.V.S. & Mpitsos, G.J.
Distributed neuronal oscillators and efference copy in
the feeding system of pleurobranchaea.
J. Neurophysiol. 36, 258274 (1973).
This was one of the first electrophysiological
studies to characterize CD signals at the cellular
level.
11. Eaton, R. Neural Mechanisms of Startle Behavior
(Plenum, New York, 1984).
12. Edwards, D., Heitler, W. & Krasne, F. Fifty years of a
command neuron: the neurobiology of escape
behavior in the crayfish. Trends Neurosci. 22,
153161 (1999).
13. Hatsopoulos, N., Gabbiani, F. & Laurent, G.
Elementary computation of object approach by a wide-
field visual neuron. Science 270, 10001003 (1995).
14. Levi, R. & Camhi, J. Wind direction coding in the
cockroach escape response: winner does not take all.
Neuroscience 20, 38143821 (2000).
15. Krasne, F.B. & Bryan, J.S. Habituation: regulation
through presynaptic inhibition. Science 182,
590592 (1973).
16. Delcomyn, F. Corollary discharge to cockroach giant
interneurons. Nature 269, 160162 (1977).
17. Kroese, A.B.A. & van Netten, S.M. in The
Mechanosensory Lateral Line: Neurobiology and
Evolution (eds Coombs, S., Gorner, P. & Munz, H.)
265284 (Springer, New York, 1989).
18. Coombs, S. & Montgomery, J.C. in Comparative
Hearing: Fish and Amphibians (eds Fay, F.R. &
Popper, A.N.) 319362 (Springer, New York, 1999).
19. Harris, G.G. & van Bergeijk, W.A. Evidence that the
lateral-line organ responds to near-field displacements
of sound sources in water. J.Acoust. Soc. Am. 34,
18311841 (1962).
20. Roberts, B.L. & Russell, I.J. The activity of lateral line
efferent neurones in stationary and swimming dogfish.
J. Exp. Biol. 57, 435448 (1972).
21. Michelsen, A. in The Evolutionary Biology of Hearing
(eds Webster, D.B., Fay, F.R. & Popper, A.N.) 6177
(Springer, New York, 1992).
22. Popper, A.N., Platt, P. & Edds, P. in The Evolutionary
Biology of Hearing (eds Webster, D.B., Fay, F.R. &
Popper, A.N.) 4957 (Springer, New York, 1992).
23. Hedwig, B. Pulses, patterns, and paths: neurobiology
of acoustic behavior in crickets. J. Comp. Physiol. A
192, 677689 (2006).
24. Hoy, R.R. & Robert, D. Tympanal hearing in insects.
Annu. Rev. Entomol. 41, 433450 (1996).
25. Poulet, J.F.A. & Hedwig, B. The cellular basis of a
corollary discharge. Science 311, 518522
(2006).
This report is one of a series of elegant studies
carried out by the authors in which they homed in
on, and both anatomically and physiologically
identified, a CDI in the cricket auditory system.
26. Agamaite, J. & Wang, X. Quantitative classification of
the vocal repertoire of the common marmoset
(Callithrix jacchus jacchus). Assoc. Res. Otolaryngol.
Abstr. 20, 573 (1997).
27. Eliades, S.J. & Wang, X. sensory-motor interaction in
the primate auditory cortex during self-initiated
vocalizations. J. Neurophysiol. 89, 21942207
(2003).
28. Eliades, S.J. & Wang, X. Dynamics of auditory-vocal
interaction in monkey auditory cortex. Cereb. Cortex
15, 15101523 (2005).
29. Alexander, G., Newman, J. & Symmes, D. Convergence
of prefrontal and acoustic inputs upon neurons in the
superior temporal gyrus of the awake squirrel monkey.
Brain Res. 116, 334338 (1976).
30. Hackett, T., Stepniewska, I. & Kaas, J. Prefrontal
connections of the parabelt auditory cortex in
macaque monkeys. Brain Res. 817, 4558 (1999).
31. Morel, A. & Kaas, J. Subdivisons and connections of
auditory cortex in owl monkeys. J. Comp. Neurol. 318,
2763 (1992).
32. Gemba, H., Miki, N. & Sasaki, K. Cortical field
potentials preceding vocalization and influences of
cerebellar hemispherectomy upon them in monkeys.
Brain Res. 697, 143151 (1995).
33. Ross, J., Morrone, M.C., Goldberg, M.E. & Burr,
D.C. Changes in visual perception at the time of
saccades. Trends Neurosci. 24, 113121 (2001).
34. Marin, G., Letelier, J.C. & Wallman, J. Saccade-related
responses of centrifugal neurons projecting to the
chicken retina. Exp. Brain Res. 82, 263270 (1990).
35. Zaretsky, M. & Rowell, C.H.F. Saccadic suppression
by corollary discharge in the locust. Nature 280,
583584 (1979).
36. Thiele, A., Henning, P., Kubischik, M. & Hoffmann,
K.P. Neural mechanisms of saccadic suppression.
Science 295, 24602462 (2002).
37. Lee, D. & Malpeli, J.G. Effects of saccades on the
activity of neurons in the cat lateral geniculate
nucleus. J. Neurophysiol. 79, 922936 (1998).
38. Yang, Y., Cao, P., Yang, Y. & Wang, S.R. Corollary
discharge circuits for saccadic modulation of the
pigeon visual system. Nature Neurosci. 11, 595602
(2008).
39. von Helmholtz, H. Helmholtzs Treatise on
Physiological Optics (Optical Society of America, New
York, 1925).
40. Sommer, M.A. & Wurtz, R.H. Influence of the
thalamus on spatial visual processing in frontal cortex.
Nature 444, 374377 (2006).
41. Sommer, M.A. & Wurtz, R.H. Brain circuits for the
internal monitoring of movements. Annu. Rev.
Neurosci. 31, 317338 (2008).
42. Schall, J.D. On the role of frontal eye field in guiding
attention and saccades. Vision Res. 44, 14531467
(2004).

nature reviews | neuroscience volume 9 | august 2008 | 599

381
2008 Macmillan Publishers Limited. All rights reserved.
REVIEWS
43. Kleinfeld, D., Ahissar, E. & Diamond, M.E. Active
sensation: insights from the rodent vibrissa
sensorimotor system. Curr. Opin. Neurobiol. 16,
435444 (2006).
44. Fee, M.S., Mitra, P.P. & Kleinfeld, D. Central versus
peripheral determinants of patterned spike activity in
rat vibrissa cortex during whisking. J. Neurophysiol.
1997, 11441149 (1997).
45. Ahissar, E. & Kleinfeld, D. Closed-loop neuronal
computations: focus on vibrissa somatosensation in
rat. Cereb. Cortex 13, 5362 (2003).
46. Kleinfeld, D., Berg, R.W. & OConner, S.M.
Anatomical loops and their electrical dynamics in
relation to whisking by rat. Somatosens. Mot. Res. 16,
6988 (1999).
47. Nelson, M.E. & MacIver, M.A. Sensory acquisition in
active sensing systems. J. Comp. Physiol. A 192,
573586 (2006).
48. Caputi, A.A. Contributions of electric fish to the
understanding of sensory processing by reafferent
systems. J. Physiol. (Paris) 98, 8197 (2004).
49. Meek, J., Grant, K. & Bell, C. Structural organization
of the mormyrid electrosensory lateral line lobe.
J. Exp. Biol. 202, 12911300 (1999).
50. Mohr, C., Roberts, P.D. & Bell, C.C. The
mormyromast region of the mormyrid electrosensory
lobe. I. Responses to corollary discharge and
electrosensory stimuli. J. Neurophysiol. 90,
11931210 (2003).
51. Bell, C.C. & Grant, K. Corollary discharge inhibition
and preservation of temporal information in a sensory
nucleus of mormyrid electric fish. J. Neurosci. 9,
10291044 (1989).
52. Bell, C.C. An efference copy which is modified by
reafferent input. Science 214, 450453 (1981).
This was a pioneering study that unveiled a plastic
CD in the mormyrid that was modifiable by recent
sensory experience.
53. Moss, C.F. & Sinha, S.R. Neurobiology of
echolocation in bats. Curr. Opin. Neurobiol. 13,
751758 (2003).
54. Neuweiler, G. Evolutionary aspects of bat
echolocation. J. Comp. Physiol. A 189, 245256
(2003).
55. Simmons, J.A. A view of the world through the bats
ear: the formation of acoustic images in echolocation.
Cognition 33, 155199 (1989).
56. Simmons, J.A. & Kick, S.A. Physiological mechanisms
for spatial filtering and image enhancement in the
sonar of bats. Annu. Rev. Physiol. 46, 599614
(1984).
57. Simmons, J.A. etal. in Hearing by Bats (eds Fay, F.R.
& Popper, A.N.) 146190 (Springer, New York,
1995).
600 | august 2008 | volume 9 www.nature.com/reviews/neuro
58. Schuller, G. Vocalization influences auditory processing
in collicular neurons of the CFFM bat, Rhinolophus
ferrumequinum. J. Comp. Physiol. A 132, 3946
(1979).
59. Bellebaum, C., Daum, I., Koch, B., Schwarz, M. &
Hoffmann, K.P. The role of the human thalamus in
processing corollary discharge. Brain 128,
11391154 (2005).
60. Bellebaum, C., Hoffmann, K.P., Koch, B., Schwarz, M.
& Daum, I. Altered processing of corollary discharge in
thalamic lesion patients. Eur. J. Neurosci. 24,
23752388 (2006).
61. Guthrie, B.L., Porter, J.D. & Sparks, D.L. Corollary
discharge provides accurate eye position information
to the oculomotor system. Science 221, 11931195
(1983).
62. Sommer, M.A. & Wurtz, R.H. A pathway in primate
brain for internal monitoring of movements. Science
296, 14801482 (2002).
This was the first study to identify a CD pathway in
the primate brain.
63. Sommer, M.A. & Wurtz, R.H. What the brain stem
tells the frontal cortex. II. Role of the SCMDFEF
pathway in corollary discharge. J. Neurophysiol. 91,
14031423 (2004).
64. Tanaka, M. Inactivation of the central thalamus delays
self-timed saccades. Nature Neurosci. 9, 2022 (2006).
65. Lynch, J., Hoover, J. & Strick, P. Input to the primate
frontal eye field from the substantia nigra, superior
colliculus, and dentate nucleus demonstrated by
transneuronal transport. Exp. Brain Res. 100,
181186 (1994).
66. Striedter, G.F. & Vu, E.T. Bilateral feedback projections
to the forebrain in the premotor network for singing in
zebra finches. J. Neurobiol. 34, 2740 (1998).
67. Catchpole, D.K. & Slater, P.J.B. Bird Song: Biological
Themes and Variations (Cambridge Univ. Press,
Cambridge, 1995).
68. Brainard, M. & Doupe, A.J. Auditory feedback in
learning and maintenance of vocal behaviour. Nature
Rev. Neurosci. 1, 3140 (2000).
69. Margoliash, D. Evaluating theories of bird song
learning: implications for future directions. J. Comp.
Physiol. A 188, 851866 (2002).
70. Margoliash, D. Functional organization of forebrain
pathways for song production and perception.
J. Neurobiol. 33, 671693 (1997).
71. Reiner, A., Yamamoto, K. & Karten, H. Organization
and evolution of the avian forebrain. Anat. Rec. A
Discov. Mol. Cell. Evol. Biol. 287, 10801102 (2005).
72. Troyer, T.W. & Doupe, A.J. An associational model of
birdsong sensorimotor learning i. efference copy and
the learning of song syllables. J. Neurophysiol. 84,
12041223 (2000).
2008 Macmillan Publishers Limited. All rights reserved.
73. Troyer, T.W. & Doupe, A.J. An associational model of
birdsong sensorimotor learning II. Temporal
hierarchies and the learning of song sequence.
J. Neurophysiol. 84, 12241239 (2000).
74. Prather, J.F., Peters, S., Nowicki, S. & Mooney, R.
Precise auditoryvocal mirroring in neurons for
learned vocal communication. Nature 451, 305310
(2008).
75. Mainland, J. & Sobel, N. The sniff is part of the
olfactory percept. Chem. Senses 31, 181196 (2006).
76. Feinberg, I. & Guazzelli, M. Schizophreniaa disorder
of the corollary discharge systems that integrate the
motor systems of thought with the sensory systems of
consciousness. Br. J. Psychiatry 174, 196204
(1999).
77. Ford, J.M. etal. Neurophysiological evidence of
corollary discharge dysfunction in schizophrenia. Am.
J. Psychiatry 158, 20692071 (2001).
78. Logothetis, N.K. Single units and conscious vision.
Philos. Trans. R. Soc. Lond. B Biol. Sci. 353,
18011818 (1998).
79. Parker, A.J. & Newsome, W.T. Sense and the single
neuron: probing the physiology of perception. Annu.
Rev. Neurosci. 21, 227277 (1998).
80. Roy, J.E. & Cullen, K.E. Dissociating self-generated
from passively applied head motion: neural
mechanisms in the vestibular nuclei. J. Neurosci. 24,
21022111 (2004).
81. Rossignol, S., Dubuc, R. & Gossard, J.P. Dynamic
sensorimotor interactions in locomotion. Physiol. Rev.
86, 89154 (2006).
82. Matthews, P.B.C. Where does sherringtons muscular
sense originate? Muscles, joints, corollary
discharges? Annu. Rev. Neurosci. 5, 189218 (1982).
83. Seki, K., Perlmutter, S.I. & Fetz, E.E. Sensory input to
primate spinal cord is presynaptically inhibited during
voluntary movement. Nature Neurosci. 6,
13091316 (2003).
84. Voss, M., Ingram, J.N., Haggard, P. & Wolpert, D.M.
Sensorimotor attenuation by central motor command
signals in the absence of movement. Nature Neurosci.
9, 2627 (2006).
Acknowledgements
We thank R. H. Wurtz for comments on an earlier version of
the manuscript. Supported by the Alfred P. Sloan foundation
and RO1-EY017592 to M.A.S.
FURTHER INFORMATION
Marc Sommers homepage: http://cnup.neurobio.pitt.edu/
people/peopleDetail.aspx?uid=381
All links are active in the online pdf
382
J Comp Physiol A (2005) 191: 979986
DOI 10.1007/s00359-005-0027-z
R EV IE W
J. F. A. Poulet
Corollary discharge inhibition and audition in the stridulating cricket
Received: 8 December 2004 / Revised: 7 June 2005 / Accepted: 12 June 2005 / Published online: 25 October 2005

Springer-Verlag 2005
Abstract The romantic notion of crickets singing on a
warm summers evening is quickly dispelled when one
comes ear to ear with a stridulating male. Remarkably,
stridulating male crickets are able to hear sounds from
the environment despite generating a 100 db song
(Heiligenberg 1969; Jones and Dambach 1973). This
review summarises recent work examining how they
achieve this feat of sensory processing. While the
responsiveness of the crickets peripheral auditory sys-
tem (tympanic membrane, tympanic nerve, state of the
acoustic spiracle) is maintained during sound produc-
tion, central auditory neurons are inhibited by a feed-
forward corollary discharge signal precisely timed to
coincide with the auditory neurons maximum response
to self-generated sound. In this way, the corollary dis-
charge inhibition prevents desensitisation of the crickets
auditory pathway during sound production.
Keywords Corollary discharge Eerence copy
Stridulation Cricket Auditory Presynaptic
inhibition
Feedforward signals
To gather sensory cues from the environment, animals
have developed an array of highly sensitive sensory
systems. These systems, however, not only respond to
environmental sensory information, but also to stimuli
generated as a by-product of the animals own behav-
iour. In principle self-generated, or reaerent (von Holst
and Mittelstaedt 1950), information could desensitise
the animals own sensory pathway and/or also be con-
J. F. A. Poulet
Laboratory of Sensory Processing, Brain Mind Institute,
Faculty of Life Science, Ecole Polytechnique Federale de
Lausanne, CH-1015 Lausanne, Switzerland
E-mail: james.poulet@ep.ch
Tel.: +41-21-6938352
Fax: +41-21-6938353
fused for external, or exaerent, sensory information of
the same modality, which could elicit an inappropriate
behavioural response. Thus, the discrimination between
self-generated and external sensory stimuli is a funda-
mental problem in perception and a central question of
sensory neuroscience.
Philosophers and scientists over the centuries have
proposed that responses to self-generated stimuli are
modulated by neural signals that feedforward from
motor to sensory networks during behaviour (Grusser
1986). In 1950 two papers furthered modern thinking
about these concepts and termed the feedforward signals
corollary discharges or eerence copies (Sperry
1950; von Holst and Mittelstaedt 1950). More recently,
these terms have been rened so that motor to sensory
discharges indicate corollary discharge signals; whereas
an eerence copy is compared with the actual sensory
feedback and represents an accurate negative image of
the reaerent information (McCloskey 1981; Bell 1984).
Neurophysiological evidence for these mechanisms re-
quires recording neural activity during motor behaviour.
Despite this technical diculty, a growing body of evi-
dence suggests that feedforward signals are widespread
in the CNS and have been recorded from a number of
sensory systems including visual (Zaretsky and Rowell
1979; Sommer and Wurtz 2002), vestibular (Roy and
Cullen 2004), electrosensory (Bell 1989), mechanosen-
sory (Murphey and Palka 1974; Delcomyn 1977; El
Manira et al. 1996; Blakemore et al. 1998; Li et al. 2002)
and proprioceptive systems (Sillar and Skorupski 1986;
Gossard et al. 1991; Wolf and Burrows 1995).
Auditory responsiveness during sound production
During sound production, the auditory system of the
caller must somehow discriminate self-generated from
external sounds and prevent desensitisation. A modu-
lation in the response of auditory neurons during sound
production has been recorded in vertebrateshuman:
(Creutzfeldt et al. 1989; Numminen et al. 1999) monkey:
383
980

(Muller-Preuss and Ploog 1981; Kirzinger and Jurgens

1991); bat: (Suga and Schlegel 1972; Suga and Shimoz-
awa 1974; Schuller 1979; Metzner 1989; Metzner 1993);
bird: (McCasland and Konishi 1981); grasshopper:
(Hedwig 1986; Wolf and Helversen 1986; Hedwig 1990;
Hedwig and Meyer 1994); and cicada: (Hennig et al.
1994). In the vertebrates the modulation is assumed to
be the result of a centrally generated neural signal
however, the exact properties and source of the inhibi-
tion have never been established. In the invertebrate
systems studied so far, the modulation is the result of
peripheral noise-making eects (Hedwig 1990; Hedwig
and Meyer 1994; Hennig et al. 1994).
One animal that generates high-intensity sound pul-
ses for communication and oers signicant experi-
mental advantages to analyse the problems above, is the
stridulating cricket. Crickets posses well characterised
and highly sensitive auditory systems to listen out for
other crickets or predators. Although cricket auditory
interneurons are desensitised by loud sounds (Pollack
1988; Samson and Pollack 2002; Poulet and Hedwig
2003a, b), behavioural experiments have shown that
singing crickets are able to hear exaerent sounds
(Heiligenberg 1969; Jones and Dambach 1973). To
understand how crickets maintain auditory sensitivity Fig. 1 Peripheral auditory responses during stridulation. a The
during sound production, we examined the responsive- sound pattern and associated wing movements during pharmaco-
ness of the crickets peripheral and central auditory logically elicited calling song. b The tympanic membrane responded
system to environmental and self-generated sound dur- with a rapid oscillation to each sound pulse. However in one
animal, and in animals with their acoustic spiracle waxed shut,
ing pharmacologically elicited stridulation. there was a decrease in the amplitude of the oscillations in
synchrony with ventilation. During sonorous stridulation, c the
tympanic membrane oscillated vigorously and d the tympanic nerve
Sound production in crickets responded with bursts of action potentials in response to each self-
generated sound pulse. Wing, movement of the wing; Sound, sound
recordings; TM: velocity (RMS), velocity of tympanic membrane
Male crickets stridulate to attract mates and warn o oscillations given as the root mean square; TM: displacement, slow
rivals. During stridulation, or singing, crickets rub their displacement of the tympanic membrane; Abdomen, movement of
forewings against each other to generate a series of the abdomen; Acoustic stimuli, externally generated sound pulses;
sound pulses, or syllables, that are generated during the Tympanic nerve, extracellular recording of tympanic nerve. Used
with permission from Poulet and Hedwig (2001, 2003b)
closing wing movements of the chirp (Fig. 1a). In
Gryllus bimaculatus a chirp is composed of 35 syllables,
which can reach up to 100 dB SPL re. 20 lPa. Each vertebrates and other invertebrates employ this mecha-
chirp is separated by a 250 ms pause or chirp interval. nism (Suga and Jen 1975; Borg and Counter 1989; Na-
Recently, pharmacological methods have been devel- rins 1992; Hennig et al. 1994). The rst stage of sound
oped that elicit robust stridulation (Otto 1978; Wenzel processing in the cricket is at its tympanic membranes
and Hedwig 1999). Microinjections of acetylcholine, and located on the tibiae of the forelegs (Larsen et al. 1989;
other cholinergic agonists, into the crickets anterior Michelsen 1998). An H-shaped trachea, that has open-
protocerebrum are thought to activate the stridulation ings at the acoustic spiracles on either side of the thorax,
command neurons that spur the thoracic central pattern acoustically couples the tympanic membranes. There-
generator into action (Hedwig 2000). Armed with this fore, sound is transmitted to the tympanic membranes
technique, it was possible to record from dierent stages from both the external and internal side (Hill and Boyan
of the crickets auditory system during stridulation and 1976; Larsen and Michelsen 1978; Michelsen 1994).
analyse how the animals cope with the self-generated Oscillations of the tympanic membranes are thought to
sound. deform auditory aerent dendrites and initiate action
potentials (Kleindienst et al. 1983; Larsen et al. 1989).
About 60 auditory aerent neurons project from the ear
Peripheral auditory responses during stridulation along prothoracic nerve 5, the tympanic nerve, into the
prothoracic ganglion (Ball et al. 1989). For a complete
One strategy to cope with reaerent acoustic stimulation description of the activity of the peripheral auditory
could be to modulate the sensitivity of the peripheral system during stridulation we measured the movements
auditory system during sound production. Many of the acoustic spiracle, the oscillations of the tympanic

384

membranes and the activity of the whole tympanic nerve
(Poulet and Hedwig 2001).
To examine the responsiveness of the tympanic
membrane, we presented acoustic stimuli to stridulating
crickets and measured the wing movements, sound
produced and any oscillations or displacement with a
laser vibrometer/interferometer. In a resting cricket, the
tympanic membrane would oscillate with constant
amplitude to 4.5 kHz, 90 dB SPL sound pulses. Occa-
sionally, during ventilation, the tympanic membrane
was slowly displaced in synchrony with the abdominal
pumping movements, as seen in other Orthoptera
(Meyer and Hedwig 1995; Meyer and Elsner 1995). The
displacements were accompanied by a decrease in the
amplitude of the sound-induced oscillations (Fig. 1b).
Waxing the spiracle shut could produce the same eect.
Closing the acoustic spiracles could, therefore, control
the amplitude of reaerent auditory input. However,
during stridulation, the spiracles remained open.
Therefore, the self-generated auditory signals had full
access to both sides of the tympanic membranes.
During stridulation the tympanic membrane oscillated
in response to each self-generated syllable (Fig. 1c).
Further analysis of the response of the tympanic
membrane to computer generated sound pulses,
showed that there was no change in its responsiveness
during the chirp or chirp interval. Whole nerve
recordings of the tympanic nerve in the foreleg of the
cricket revealed a similar result: bursts of activity were
recorded in response to the crickets own song (Fig. 1d),
and the response to externally generated acoustic
stimuli presented were not modulated throughout the
song. Thus, in contrast to most other acoustically
communicating animals, the cricket does not take
advantage of a relatively simple mechanism to reduce
the responsiveness of its peripheral auditory pathway
during sound production. How then does the crickets
central auditory system copes with the mass of self-
generated sensory input during stridulation?
Central auditory responses during stridulation
The auditory aerents terminate in the auditory neuropil
in the anterior ventral prothoracic ganglion. Here they
forward their information to the dendritic branches of
auditory interneurons (Wohlers and Huber 1982). We
went on to make intracellular recordings and stainings in
the auditory neuropil from the axonal branches of the
auditory aerents and from the dendrites of the auditory
interneurons during stridulation. We recorded from
three identied interneurons: (1) the omega neuron 1
(ON1) is an W-shaped, local auditory interneuron that is
broadly tuned but is most sensitive to the carrier fre-
quency of male calling song (4.5 kHz) (Casaday and
Hoy 1977; Popov et al. 1978; Wohlers and Huber 1978);
(2) ascending neuron 1 (AN1) is tightly tuned to 4.5 kHz
sound stimuli but possesses an ascending axon that
terminates in the brain (Boyan 1980; Wohlers and
981
Huber 1982; Stumpner et al. 1995). (3) Ascending neu-
ron 2 (AN2) also ascends to the brain, it responds best to
higher frequencies (>10 kHz) and is inhibited, or only
weakly excited, by 4.5 kHz sound (Wohlers and Huber
1978, 1982; Boyan 1980; Nolen and Hoy 1987), which
supports roles in courtship song recognition (Wohlers
and Huber 1982) and bat detection during ight (Nolen
and Hoy 1983).
As expected from the whole nerve recordings, single
auditory aerent neurons responded with bursts of
spikes in response to each syllable (Fig. 2a). Addi-
tionally prior to the start of some of the quieter chirps,
the axonal arborisations of the aerents were depolar-
ised (arrow, Fig. 2a). The interneurons ON1 (Fig. 2b)
and AN1 (Fig. 2c) responded with bursts of spikes to
the crickets own song. However, as in the aerents,
besides the excitation a second input was present in
ON1 consisting of small hyperpolarising potentials
(arrow, Fig. 2b). Unlike the other neurons, AN2 was
inhibited during the chirps and rarely generated a spike
(Fig. 2d).
To characterise the inhibitory inputs, we removed
one of the wings and elicited silent stridulation. In the
absence of any self-generated or external sounds in the
aerent axon we recorded phasic groups of depolaris-
ing synaptic inputs termed primary aerent depolari-
sations or PADs (Fig. 3a). Each PAD started during
the closing wing movements of the chirp when sound
would normally have been generated. Recordings from
ON1 during silent stridulation revealed clear groups of
IPSPs in phase with each chirp (Fig. 3b). As with the
PADs they started during the closing wing movements
of each syllable and therefore must be expressed syn-
chronous with ON1s spiking response during singing.
Recordings from AN1 and AN2 showed small ampli-
tude hyperpolarising potentials during the chirps
(Fig. 3c, d). The large amplitude IPSPs recorded in
AN2 during sonorous singing must therefore have been
a response to reaerent input and correspond to the
4.5 kHz inhibition.
To test what eect the PADs and IPSPs have on
sound processing, we presented silently singing crickets
with a continuous series of sound pulses. In this way, it
was possible to evoke an auditory response in these
neurons and examine any modication during the PAD
or IPSP. Auditory aerents clearly responded with a
train of spikes when presented with the sound pulses
(Fig. 4a). The frequency of spiking in the aerent cells
was consistent throughout the chirp and chirp intervals.
However on closer inspection of the spike amplitude,
those spikes coinciding with the PADs were reduced.
The reduction in spike height is a good indication that
the PADs have a presynaptic inhibitory function (Bur-
rows and Matheson 1994; Clarac and Cattaert 1996).
The PADs momentarily shunt the membrane resistance
and cause a reduction in spike height that results in a
smaller depolarisation of the axon terminal. Therefore,
less neurotransmitter is released and the response in the
postsynaptic cell is reduced. In a resting cricket and
385
982
Fig. 2 Central auditory neural responses during sonorous, two
winged, stridulation. a Auditory aerents responded with bursts of
spikes during the chirp. Arrow indicates a depolarisation of the
membrane prior to the rst burst of spikes. b ON1 and c AN1
responded with bursts of spikes to syllables generated during the
chirps. Arrow in b indicates small hyperpolarisations of ON1
superimposed on the excitatory response during the chirp. d AN2
was inhibited during the chirp, IPSPs were generated in response to
each syllable. Aerent, intracellular recording of auditory aerent;
ON1, intracellular recording of omega neuron 1; AN1, intracellular
recording of ascending neuron 1; AN2, intracellular recording of
ascending neuron 2. The cartoon crickets in this and the following
gures represent a sonorously (two winged), silently (one winged)
and ctively stridulating cricket. For further details, see Fig. 1.
Used with permission from Poulet and Hedwig (2002, 2003a,
2003b)
during the chirp intervals, ON1, AN1 and AN2 all re-
sponded to the sound pulses with a continuous series of
spikes to the exaerent stimuli, however they all failed to
respond during the chirp (Fig. 4b, c, d). This inhibition
of the auditory interneurons must therefore be a com-
bination of the presynaptic inhibition acting at the
aerent axon terminal and the direct postsynaptic inhi-
bition mediated by the IPSPs.
Fig. 3 Responses of central auditory neurons during silent, one
winged, stridulation. a Phasic groups of PADs were recorded in
auditory aerent neurons during silent chirps. b IPSPs were
recorded in ON1 during silent chirps and closing wing movement.
Low amplitude hyperpolarising potentials were recorded in c AN1
and d AN2 during silent chirps. Both the PADs and IPSPs began
just after the start of the wing closing and reached a maximum
during the consecutive wing opening movement. For further
details, see Figs. 1 and 2. Used with permission from Poulet and
Hedwig (2002, 2003a, 2003b)
Sensory feedback or feedforward signal?
Alongside the reaerent input, we had identied an
inhibitory input that occurred at the exact time of the
maximum response to the crickets own song. The
inhibitory inputs may have two sources: reaerent
feedback from non-auditory sense organs activated
during stridulation or from the network of neurons
generating the stridulatory motor pattern and expressed
as a feedforward neural signal. To distinguish between
the two sources, we removed the sensory feedback and
eerent output of the motor neurons by cutting the
thoracic and abdominal lateral nerves and recorded the
responses of the auditory neurons in ctively stridulating
crickets (Poulet and Hedwig 2002, 2003a, b).
Recordings during ctive singing revealed the same
pattern of inhibitory input in the auditory neurons as in
the silently singing crickets. In the aerents, phasic
groups of PADs were present during the chirps (Fig. 5a);
IPSPs were recorded in ON1 (Fig. 5b) and small
amplitude hyperpolarisations occurred in AN1 (Fig. 5c)
and AN2 (Fig. 5d). The inhibitory inputs during the
chirps had the same eect on sound processing as during
silent singing on presenting a continuous series of sound
386

Fig. 4 Responses of auditory neurons to acoustic stimuli presented
during silent singing. a Auditory aerents showed no modulation
in their ring rate to acoustic stimuli and responded with a constant
train of spikes during the chirp and chirp intervals. However, those
spikes coinciding with PADs during the chirp were reduced in
amplitude (stippled line). b ON1, c AN1, and d AN2 all responded
with spikes to the acoustic stimuli during the chirp intervals but
their responses were completely inhibited during the chirps. For
further details, see Figs. 1 and 2. Used with permission from Poulet
and Hedwig (2002, 2003a, 2003b)
pulses. Just as in the silently singing crickets, the aerent
neuron continued to spike throughout the chirp and
chirp interval but the spike amplitude was reduced when
it coincided with the PADs (Fig. 6a). The auditory in-
terneurons responded to the sound during the chirp
intervals but failed to respond during the chirp (Fig. 6b,
c, d). This conrmed that the inhibitory inputs during
the chirp were generated within the nervous system and
were not the product of sensory feedback. As the inhi-
bition did not completely cancel out responses to reaf-
ferent stimuli it appeared to be a corollary discharge
signal rather than an eerence copy.
Biological significance of corollary discharge inhibition
in the stridulating cricket
To examine what long-term eect the inhibitory corol-
lary discharge had on sensory processing, loud chirps
were presented, which mimicked the crickets own song,
983
Fig. 5 Inhibitory inputs are present during ctive singing with all
motor and sensory neurons cut. a PADs occur in auditory aerent
neurons, b IPSPs occur in ON1 and low amplitude hyperpolarising
potentials are present in c AN1 and d AN2 during the chirps. These
inputs were therefore generated within the nervous system and the
result of a corollary discharge signal. For further details, see Figs. 1
and 2. Used with permission from Poulet and Hedwig (2003a)
followed by quieter test pulses (Poulet and Hedwig 2002,
2003b). In a resting cricket, bursts of spikes in the
auditory interneurons induce a long lasting hyperpola-
rising potential that can suppress responses to sub-
sequent quieter sounds (Pollack 1988; Poulet and
Hedwig 2003a, b). Responses to test pulses were de-
creased in ON1 if presented after a loud chirp (Fig. 7a,
b). When the eect of the corollary discharge was
mimicked, by injecting hyperpolarising current into
ON1 during the loud chirp, the response the test stimuli
returned (Fig. 7c). Thus, a reduction in response to loud
reaerent sounds, mediated by the corollary discharge
inhibition, will prevent desensitisation of the auditory
pathway and help the cricket hear sounds from the
environment.
Conclusions and future work
Sensory neural pathways are able to respond to exaf-
ferent stimuli, even during behaviours that generate in-
tense reaerent stimulation. Neural recordings of
sensory pathways made during motor behaviour, in
vertebrates and invertebrates, point towards a similar
solution to this feat: the central sensory pathway is
modulated during the phase of behaviour that generates
the reaerent stimuli. In more and more cases, both in
vertebrates and invertebrates, centrally generated neural
signals are responsible for the modulation. The logical
387
984
Fig. 6 Eect of corollary discharge inhibition on sound processing
in ctively singing crickets. a Auditory aerents responded to the
acoustic stimuli with a constant train of spikes during the chirp and
chirp intervals. Those spikes occuring at the same time as the PADs
were reduced in amplitude. b ON1, c AN1, and d AN2 all
responded with spikes to the test stimuli at rest and during the chirp
intervals. But their response was inhibited during the chirps. For
further details, see Figs. 1 and 2. Used with permission from Poulet
and Hedwig (2002, 2003a, 2003b)
next step in the study of feedforward signals is to iden-
tify their source. Experiments are underway to cha-
racterise the physiological and anatomical basis of the
corollary discharge signal in the stridulating cricket.
Acknowledgements I thank Dr Berthold Hedwig for his helpful
comments on an earlier version of this manuscript. This work was
supported by the BBSRC.
References
Ball EE, Oldeld BP, Rudolph KM (1989) Auditory organ struc-
ture, development and function. In: Huber F et al (eds) Cricket
behaviour and neurobiology. Cornell University Press, Ithaca
London, pp 391422
Bell CC (1984) Eects of motor commands on sensory inow, with
examples from electric sh. In: Bolis L et al (eds) Comparative
physiology of sensory systems. Cambridge University Press,
Cambridge, pp 636647
Fig. 7 Inhibition during loud chirps prevents subsequent desensi-
tisation. a ON1 responds to a series of 4.5 kHz, 80 dB test pulses
with bursts of spikes. b If a 100 dB chirp is presented immediately
prior to the quieter test stimuli, ON1s response is suppressed. c
Hyperpolarising current injection into ON1 during the loud chirp
prevented desensitisation to the test stimuli. For further details, see
Figs. 1 and 2. Used with permission from Poulet and Hedwig
(2003b)
Bell CC (1989) Sensory coding and corollary discharge eects in
mormyrid electric sh. J Exp Biol 146:229253
Blakemore S-J, Wolpert DW, Frith CD (1998) Central cancellation
of self-produced tickle sensation. Nature Neurosci 1:635640
Borg E, Counter S (1989) The middle-ear muscles. Sci Am 261:62
68
Boyan GS (1980) Auditory neurones in the brain of the cricket
Gryllus bimaculatus (De Geer). J Comp Physiol A 140:8193
Burrows M, Matheson T (1994) A presynaptic gain control
mechanism among sensory neurons of a locust leg proprio-
ceptor. J Neurosci 14:272282
Casaday B, Hoy RR (1977) Auditory interneurons in the cricket
Teleogryllus oceanicus: physiological and anatomical properties.
J Comp Physiol A 156:789801
Clarac F, Cattaert D (1996) Invertebrate presynaptic inhibition and
motor control. Exp Brain Res 112:163180
Creutzfeldt O, Ojemann G, Lettich E (1989) Neuronal activity in
the human lateral temporal lobe II. Responses to the subjects
own voice. Exp Brain Res 77:476489
Delcomyn F (1977) Corollary discharge to cockroach giant inter-
neurons. Nature 269:160162
388

El Manira A, Tegner J, Grillner S (1996) Locomotor-related pre-
synaptic modulation of primary aerents in the lamprey. Eur J
Neurosci 9:696705
Gossard J-P, Cabelguen J-M, Rossignol S (1991) An intracellular
study of muscle primary aerents during ctive locomotion in
the cat. J Neurophysiol 65:914926
Grusser O-J (1986) Interaction of eerenct and aerent signals in
visual perception a history of ideas and experimental para-
digms. Acta Psychologica 63:321
Hedwig B (1986) On the role in stridulation of plurisegmental in-
terneurons of the acridid grasshopper Omocestus viridulus L. II.
Anatomy and physiology of ascending and T-shaped inter-
neurons. J Comp Physiol A 158:429444
Hedwig B (1990) Modulation of auditory responsiveness in strid-
ulating grasshoppers. J Comp Physiol A 167:847856
Hedwig B (2000) Control of cricket stridulation by a command
neuron: ecacy depends on the behavioural state. J Neuro-
physiol 83:712722
Hedwig B, Meyer J (1994) Auditory information processing in
stridulating grasshoppers: tympanic membrane vibrations and
neurophysiology. J Comp Physiol A 174:121131
Heiligenberg W (1969) The eect of stimulus chirps on a crickets
chirping. Z Vergl Physiol 65:7097
Hennig RM, Weber T, Huber F, Kleindienst H-U, Moore TE,
Popov AV (1994) Auditory threshold change in singing cicadas.
J Exp Biol 187:4555
Hill KG, Boyan GS (1976) Directional hearing in crickets. Nature
262:390391
von Holst E, Mittelstaedt H (1950) Das Reaerenzprinzip: Wech-
selwirkungen zwischen Zentralnervensystem und Peripherie.
Naturwissenschaften 37:464476
Jones MDR, Dambach M (1973) Response to sound in crickets
without tympanal organs (Gryllus campestris L.). J Comp
Physiol A 87:8998
Kirzinger A, Jurgens U (1991) Vocalization-correlated single-unit
activity in the brain stem of the monkey. Exp Brain Res 84:545
560
Kleindienst H-U, Wohlers DW, Larsen ON (1983) Tympanal
membrane motion is necessary for hearing in crickets. J Exp
Biol 151:397400
Larsen ON, Michelsen A (1978) Biophysics of the ensiferan ear III.
The cricket ear as a four input system. J Comp Physiol A
123:217227
Larsen ON, Kleindienst K-U, Michelsen A (1989) Biophysical as-
pects of sound reception. In: Huber F et al (eds) Cricket
behaviour and neurobiology. Cornell University Press, Ithaca
London, pp 364390
Li W-C, Soe SR, Roberts A (2002) Spinal inhibitory neurons that
modulate cutaneous sensory pathways during locomotion in a
simple vertebrate. J Neurosci 22:1092410934
McCasland JS, Konishi M (1981) Interaction between auditory and
motor activities in an avian song control nucleus. Proc Natl
Acad Sci USA 78:78157819
McCloskey DI (1981) Corollary discharges: motor commands and
perception. In: Brooks VB (ed) Handbook of physiology. The
nervous system. Am Physiol Soc, Bethesda, pp 14151447
Metzner W (1989) A possible neuronal basis for Doppler-shift com-
pensation in echolocating horseshoe bats. Nature 341:529532
Metzner W (1993) An audio-vocal interface in echolocating
horseshoe bats. J Neurosci 13:18991915
Meyer J, Elsner N (1995) How respiration aects auditory sensi-
tivity in the grasshopper Chorthippus biguttulus (L.). J Comp
Physiol A 176:563573
Meyer J, Hedwig B (1995) The inuence of tracheal pressure
changes on the responses of the tympanal membrane and
auditory receptors in the locust Locusta migratoria L. J Exp
Biol 198:13271339
Michelsen A (1994) Directional hearing in crickets and other small
animals. Fort der Zool 39:195207
Michelsen A (1998) Biophysics of sound localization in insects. In:
Hoy RR et al (eds) Comparative hearing: insects. Springer,
Berlin Heidelberg New York, pp 1862
985
Muller-Preuss P, Ploog D (1981) Inhibition of auditory cortical
neurons during phonation. Brain Res 215:6176
Murphey RK, Palka J (1974) Eerent control of cricket giant -
bres. Nature 248:249251
Narins PM (1992) Reduction of tympanic membrane displacement
during vocalization of the arboreal tree frog, Eleutherodactylus
coqui. J Acoust Soc Am 91:35513557
Nolen TG, Hoy RR (1983) Initiation of behaviour by single neu-
rons: the role of behavioural context. Science 226:992994
Nolen TG, Hoy RR (1987) Postsynaptic inhibition mediates high-
frequency selectivity in the cricket Teleogryllus oceanicus:
implications for ight pyhonotaxis behaviour. J Neurosci
7:20812096
Numminen J, Salmelin R, Hari R (1999) Subjects own speech re-
duces reactivity of the human auditory cortex. Neurosci Lett
265:119122
Otto D (1978) Anderungen von Gesangsparametern bei der Grille
(Gryllus campestris L.) nach Injektion von Pharmaka ins
Gehirn. Verh Dt Zool Ges 245
Pollack GS (1988) Selective attention in an insect auditory neuron.
J Neurosci 8:26352639
Popov AV, Markovich AM, Andjan AS (1978) Auditory inter-
neurons in the prothoracic ganglion of the cricket, Gryllus
bimaculatus deGeer I. The large segmental auditory neuron
(LSAN). J Comp Physiol A 126:183192
Poulet JFA, Hedwig B (2001) Tympanic membrane oscillations and
auditory receptor activity in the stridulating cricket Gryllus
bimaculatus. J Exp Biol 204:12811293
Poulet JFA, Hedwig B (2002) A corollary discharge maintains
auditory sensitivity during sound production. Nature 418:872
876
Poulet JFA, Hedwig B (2003a) Corollary discharge inhibition of
ascending auditory information in the stridulating cricket.
J Neurosci 23:47174725
Poulet JFA, Hedwig B (2003b) A corollary discharge mechanism
modulates central auditory processing in singing crickets.
J Neurophysiol 89:15281540
Roy JE, Cullen KE (2004) Dissociating self-generated from pas-
sively applied head motion: neural mechanisms in the vestibular
nuclei. J Neurosci 24:21022111
Samson A-H, Pollack GS (2002) Encoding of sound localization
cues by an identied auditory interneuron: eects of stimulus
temporal pattern. J Neurophysiol 88:23222328
Schuller G (1979) Vocalization inuences auditory processing in
collicular neurons of the CF-FM-bat, Rhinolophus ferrumequi-
num. J Comp Physiol A 132:3946
Sillar KT, Skorupski P (1986) Central input to primary aerent
neurons in craysh, Pacifastacus leniusculus, is correlated with
rhythmic motor output of thoracic ganglia. J Neurophysiol
55:678688
Sommer MA, Wurtz RH (2002) A pathway in primate brain for
internal monitoring of movements. Science 296:14801482
Sperry RW (1950) Neural basis of the spontaneous optokinetic
response produced by visual inversion. J Comp Physiol Psych
43:482489
Stumpner A, Atkins G, Stout JF (1995) Processing of unilateral
and bilateral auditory inputs by the ON1 and L1 interneurons
of the cricket Acheta domesticus and comparison to other
cricket species. J Comp Physiol A 177:379388
Suga N, Jen P (1975) Peripheral control of acoustic signals in the
auditory system of echolocating bats. J Exp Biol 62:277311
Suga N, Schlegel P (1972) Neural attenuation of responses to
emitted sounds in echolocating bats. Science 177:8284
Suga N, Shimozawa T (1974) Site of neural attenuation of re-
sponses to self-vocalized sounds in echolocating bats. Science
183:12111213
Wenzel B, Hedwig B (1999) Neurochemical control of cricket
stridulation revealed by pharmacological injections into the
brain. J Exp Biol 202:22032216
Wohlers DW, Huber F (1978) Intracellular recording and staining
of cricket auditory interneurons (Gryllus campestris L., Gryllus
bimaculatus DeGeer). J Comp Physiol A 127:1128
389
986
Wohlers DW, Huber F (1982) Processing of sound signals by six
types of neurons in the prothoracic ganglion of the cricket,
Gryllus campestris L. J Comp Physiol A 146:161173
Wolf H, Burrows M (1995) Proprioceptive sensory neurons of a
locust leg receive rhythmic presynaptic inhibition during
walking. J Neurosci 15:56235636
Wolf H, Helversen OV (1986) Switching-o of an auditory
interneuron during stridulation in the acridid grasshopper
Chorthippus biguttulus L. J Comp Physiol A 158:861871
Zaretsky M, Rowell CHF (1979) Saccadic suppression by corollary
discharge in the locust. Nature 280:583585
390
REPORTS
Fig. 4. Compressive strength of
porous HAP scaffolds. Results
from literature [blue stars (22),
red stars (23), inverted green
triangles (24), black triangle
(25), blue circle (26), inverted
blue triangle (27), diamonds
(28), cross (29), and red circles
(30)] versus IT porous HAP
scaffolds. The typical pore sizes
of conventional porous HAP
scaffolds are on the order of
100 to 800 mm in order to al-
low bone ingrowth. In the IT
materials exhibiting the greatest
strength, the pores are typical-
ly 20 by 200 mm wide and
several millimeters long; pre-
vious studies have indicated
that these dimensions are ad-
equate for bone tissue engi-
neering (20). For the IT porous
materials, compression is ap-
plied in the direction parallel to the ceramic layers. The presence of inorganic bridges between the
ceramic layers (a feature that parallels the microstructure of nacre) prevents Euler buckling of the
ceramic layers and contributes to the high strength. (Inset) Typical compression load-displacement
curves for materials with 56% porosity (three different samples shown here). The samples fail gradually,
and, because of the large degree of control of hierarchical architecture, the mechanical behavior is very
consistent from one sample to another.
suspensions, we processed IT highly porous
lamellar scaffolds that are four times stronger
in compression than conventional porous HAP
(Fig. 4). These IT scaffolds exhibit well-defined
pore connectivity along with directional and
completely open porosity of an adequate size
to allow bone ingrowth (20). Hence, most of the
current shortcomings (low strength, random or-
ganization, multiple pore size, and uncontrolled
pore connectivity) that plague bone substitutes
are eliminated by this innovative approach.
Current ceramic and metallic implant ma-
terials have serious shortcomings because of
the mismatch of physical properties with those
of bone. In bone, intrinsically weak materials,
such as calcium phosphates and collagen, are
combined into composites exhibiting intermedi-
ate modulus (10 to 20 GPa), fairly high strength
(30 to 200 MPa), and high work of fracture
(100 to 1000 J/m2) (21). The unique proper-
ties of bone arise from the controlled integra-
tion of the organic (collagen) and inorganic
(apatite) components (5) with a sophisticated
architecture from the nano- to mesolevels. Our
approach to the problem is to infiltrate the IT
porous HAP scaffolds with a second organic
phase with tailored biodegradability. Because
the biodegradation rates of the scaffold and the
infiltrated compound can be designed to be dif-
ferent, porosity can be created in situ to allow
bone ingrowth. By using this approach, we have
been able to fabricate HAP-based composites
with stiffness (10 GPa), strength (150 MPa),
and work of fracture (220 J/m2) that match
that of compact bone for an equivalent mineral/
organic content (around 60/40 vol %).
518 27 JANUARY 2006
11. H. Ishiguro, B. Rubinsky, Cryobiology 31, 483 (1994).
12. J. O. M. Karlsson, Science 296, 655 (2002).
13. G. Gay, M. A. Azouni, Cryst. Growth Des. 2, 135 (2002).
14. A. P. Jackson, J. F. V. Vincent, R. M. Turner, Proc. R. Soc.
Lond. Ser. B 234, 415 (1988).
15. J. Aizenberg et al., Science 309, 275 (2005).
16. R. Z. Wang, Z. Suo, A. G. Evans, N. Yao, I. A. Aksay,
J. Mater. Res. 16, 2485 (2001).
17. F. Song, A. K. Soh, Y. L. Bai, Biomaterials 24, 3623 (2003).
18. E. Saiz, R. M. Cannon, A. P. Tomsia, Acta Mater. 48, 4449
(2000).
19. L. L. Hench, J. M. Polak, Science 295, 1014 (2002).
20. T. Dutta Roy, J. L. Simon, J. L. Ricci, E. D. Rekow,
V. P. Thompson, J. R. Parsons, J. Biomed. Mater. Res. 66A,
283 (2003).
21. Y. H. An, in Mechanical Testing of Bone and the
Bone-Implant Interface (CRC Press, Boca Raton, FL,
2000), pp. 4163.
22. A. Almirall et al., Biomaterials 25, 3671 (2004).
23. R. P. del Real, J. G. C. Wolke, M. Vallet-Regi, J. A. Jansen,
Biomaterials 23, 3673 (2002).
24. Y. Ota, T. Kasuga, Y. Abe, J. Am. Ceram. Soc. 80, 225
(1997).
Downloaded from www.sciencemag.org on August 23, 2007
25. A. Bignon, thesis, National Institute of Applied Science,
Lyon, France (2002).
26. H. R. Ramay, M. Q. Zhang, Biomaterials 24, 3293
(2003).
27. M. Sous et al., Biomaterials 19, 2147 (1998).
28. D. M. Liu, Ceram. Int. 23, 135 (1997).
29. M. Milosevski, J. Bossert, D. Milosevski, N. Gruevska,
Ceram. Int. 25, 693 (1999).
30. M. Kawata et al., J. Mater. Sci. Mater. Med. 15, 817 (2004).
31. This work was supported by the NIH National Institute
of Dental and Craniofacial Research under grant 5R01
DE015633 (Complex nanocomposites for bone
regeneration) and by the Director, Office of Science,
References and Notes Office of Basic Energy Sciences, Division of Materials
1. G. Mayer, Science 310, 1144 (2005). Sciences and Engineering of the U.S. Department
2. C. Sanchez, H. Arribart, M. M. G. Guille, Nat. Mater. 4, of Energy under contract DE-AC03-76SF00098
277 (2005). (Metal/Ceramic Composites). The authors wish to
3. L. Addadi, S. Weiner, Nature 389, 912 (1997). thank R. O. Ritchie for useful discussions and J. Wu
4. B. L. Smith et al., Nature 399, 761 (1999). for help with the synthesis of the aluminum-infiltrated
5. G. E. Fantner et al., Nat. Mater. 4, 612 (2005). composites.
6. Z. Y. Tang, N. A. Kotov, S. Magonov, B. Ozturk, Nat. Mater.
2, 413 (2003).
Supporting Online Material
www.sciencemag.org/cgi/content/full/311/5760/515/DC1
7. A. Sellinger et al., Nature 394, 256 (1998).
8. W. J. Clegg, Science 286, 1097 (1999). Materials and Methods
Figs. S1 and S2
9. M. G. Worster, J. S. Wettlaufer, J. Phys. Chem. B 101,
6132 (1997). 4 October 2005; accepted 6 December 2005
10. S. W. Sofie, F. Dogan, J. Am. Ceram. Soc. 84, 1459 (2001). 10.1126/science.1120937
The Cellular Basis of a
Corollary Discharge
James F. A. Poulet1,2* and Berthold Hedwig2
How do animals discriminate self-generated from external stimuli during behavior and prevent
desensitization of their sensory pathways? A fundamental concept in neuroscience states that
neural signals, termed corollary discharges or efference copies, are forwarded from motor to
sensory areas. Neurons mediating these signals have proved difficult to identify. We show that
a single, multisegmental interneuron is responsible for the pre- and postsynaptic inhibition of
auditory neurons in singing crickets (Gryllus bimaculatus). Therefore, this neuron represents a
corollary discharge interneuron that provides a neuronal basis for the central control of sensory
responses.
n animal_s behavior generates a con- external stimuli. One solution to these problems
A
stant flow of sensory information that is to forward a signal, or corollary discharge,
can update or fine-tune ongoing motor from motor to sensory regions during behavior
activity (1) but can also desensitize the animal_s to counter the expected, self-generated sensory
own sensory pathways and/or be confused with feedback (2, 3). A role for corollary discharges
391
VOL 311 SCIENCE www.sciencemag.org
 REPORTS

Fig. 1. Morphology of CDI. (A) A whole-mount staining of CDI

in the CNS of an adult male G. bimaculatus in ventral view. The
soma and dendrites are located in the mesothoracic ganglion,
and two axons project throughout the whole CNS with ex-
tensive varicose arborizations that are bilateral in every ganglion
except the brain. Arrow in brain indicates anterior branch of CDI
stained in two of six stainings of its axon in the brain. (B) Axonal
arborizations in the prothoracic ganglion; arrows indicate
overlap with the auditory neuropils. (C) Lateral view of CDI in
mesothoracic ganglion. The soma is positioned medially near
the dorsolateral edge of the ganglion. From the soma the
primary neurite extends in a loop toward the middle of the
ganglion and gives off a widespread bilateral array of smooth
branches typical of insect dendrites. Two axons originate
centrally in the ganglion and extend both anteriorly and
posteriorly. (D) Ventral axonal arborizations in the mesothoracic
ganglion. (E) Dendritic (dorsal) and axonal (ventral) ar-
borizations of CDI in the mesothoracic ganglion. (F) Axonal

Downloaded from www.sciencemag.org on August 23, 2007

arborizations of CDI in the metathoracic ganglion have a similar
morphology to those in the mesothoracic ganglion. Abbrevia-
tions: SOG, suboesophageal ganglion; Pro, prothoracic ganglion;
Meso, mesothoracic ganglion; Meta, metathoracic ganglion; Ab1
to Ab4, abdominal ganglia 1 to 4; TAb, terminal abdominal
ganglion. Scale bars, 100 mm.

in modifying sensory processing during behav-

ior has been identified in many sensory systems
(411). Despite their ubiquity and importance,
however, very little is known about the neurons
mediating corollary discharges.
An ideal model system to analyze a cor-
ollary discharge is the singing male cricket
(6, 12). Cricket song is composed of a series of
100 dB SPL chirps, repeated every 300 to 500
ms, each containing three to five sound pulses
or syllables. Sound is generated during the
closing movements of the wing. The crickets_
ears are located on their forelegs and remain
fully sensitive during singing (13). Therefore
the cricket_s central nervous system (CNS) has
to deal with a massive influx of auditory,
proprioceptive, and mechanoreceptive informa-
tion during sound production. Crickets maintain
auditory sensitivity during singing by inhibiting
their central auditory pathway with a corollary
discharge in phase with sound production (14).
In this study, we identify the neuron that
mediates this corollary discharge.
The corollary discharge interneuron (CDI)
was physiologically identified by simultaneous
intracellular recordings of auditory neurons in
the prothoracic ganglion, where primary audito-
ry signals are processed, and systematic probing
of interneurons in the mesothoracic ganglion,
which houses part of the singing pattern gen-
erating network. Consecutive stainings (n 0 12
crickets) revealed its extensive branching pat-
tern throughout the CNS (Fig. 1) (15).
1
Laboratory of Sensory Processing, Brain Mind Institute,
Ecole Polytechnique Federale de Lausanne, CH-1015
Lausanne, Switzerland. 2Department of Zoology, University
of Cambridge, Cambridge CB2 3EJ, UK.
*To whom correspondence should be addressed. E-mail:
james.poulet@epfl.ch

392
www.sciencemag.org SCIENCE VOL 311 27 JANUARY 2006 519
REPORTS
The morphology of CDI highlights three
major structural properties that are crucial for
its function. First, its cell body and extensive
dendritic arborization is in the mesothoracic
ganglion; therefore, it can receive synaptic input
from the singing central pattern generator
(CPG). Second, the neuron has profuse axonal
arborizations that overlap with the auditory
neuropil in the prothoracic ganglion, a pre-
requisite to forming direct output synapses with
auditory neurons. Third, its axon targets wide-
spread areas of the CNS and could affect other
sensory pathways activated during singing.
Recordings from CDI were made from its
dendritic branches during pharmacologically
elicited fictive singing, with all thoracic sensory
and motor nerves cut except for the auditory
nerves (3, 15). CDI generates bursts of spikes in
and also with the timing of inhibitory inputs in
auditory neurons during singing (14). To test
whether CDI is part of the singing CPG, we
used 100-ms depolarizing current pulses to
elicit spikes and measured any effect on the
singing motor pattern (Fig. 2C). Elicited bursts
of spikes in CDI never had an effect on the
timing of the ongoing singing motor pattern
(n 0 10) (Fig. 2D). Singing also continued
normally when CDI was prevented from spiking
by injection of hyperpolarizing current (n 0
10). We therefore conclude that CDI is not part
of the singing CPG but instead is driven by it.
Because CDI receives excitatory input dur-
ing wing-closing movements, we tested whether
it might also be activated during flying. When
the crickets generated the flight motor pattern
(n 0 7), CDI was always inhibited by a barrage
ings from CDI_s dendrite and auditory neurons
revealed the synaptic connectivity. About 60
auditory afferent neurons extend from the
cricket_s ears in the forelegs into prothoracic
ganglion, where they terminate in a median-
ventral auditory neuropil (17). Here they for-
ward excitatory information onto a small
number of auditory interneurons such as the
local omega neuron 1 (ON1) (18). Paired re-
cordings from CDI and auditory afferent
axonal arborizations demonstrated that spikes
in CDI occurred in synchrony with primary
afferent depolarizations (PADs) during wing-
closing motor activity (Fig. 3A). PADs cause a
reduction in spike height in cricket auditory
afferents (14) and mediate presynaptic inhibi-
tion in a number of sensory systems (1921).
When depolarizing current was injected into

Downloaded from www.sciencemag.org on August 23, 2007

synchrony with the wing motoneuron activity
indicating the chirps (Fig. 2A). CDI spikes
reach frequencies of 131 T 16 Hz during the
first syllable of the chirp and a maximum of
178 T 9 Hz during the following syllables
(mean T SEM; n 0 18). Quantitative analysis
reveals that each burst of spikes occurred
during wing-closing motor activity (stippled
lines in Fig. 2B). This corresponds to the phase
of wing movements when sound is produced
of inhibitory postsynaptic potentials (IPSPs)
(Fig. 2, E and F), as were some of the neurons
thought to be part of the cricket singing CPG
(16). CDI did not respond to 85 dB SPL
acoustic stimulation at 4.8 kHz, the carrier
frequency of cricket song (n 0 10) (Fig. 2G).
CDI_s anatomy and physiology strongly
suggested that it could mediate pre- and post-
synaptic inhibition of the auditory pathway
during singing (3). Paired intracellular record-
CDI, each spike in CDI reliably elicited a PAD
of 2.0 T 0.2 mV in the auditory afferent after a
constant latency of 3.8 T 0.1 ms (n 0 7) (Fig. 3,
B and C). Paired recordings of CDI and ON1
(Fig. 3D) revealed that ON1 received IPSPs
when CDI spiked during a chirp. Injection of
current into CDI revealed that each CDI spike
elicited an IPSP of 1.6 T 0.2 mV in ON1 after a
delay of 3.5 T 0.2 ms (n 0 8) (Fig. 3, E and F).
CDI has bilateral arborizations in the pro-

Fig. 2. Recordings of CDI and wing motor nerve during fictive singing and
flight. (A) Spikes in CDI as recorded in the dendrite during fictive chirps.
(B) The PSTH and superimposed instantaneous spike rate (n 0 59 chirps,
759 spikes) with the averaged wing motor nerve activity demonstrate that
CDI fires bursts of spikes during the wing-closing motor activity of each
chirp, as indicated by the stippled lines. (C) Injection of depolarizing
current into CDI elicited bursts of spikes but did not change the ongoing
singing pattern. (D) Phase response curve. Q 0 (duration of ongoing chirp
period N)/(duration of chirp period N 1) at the phase of stimulation of
CDI with a current pulse of 100 ms. Analysis shows no modulation in the
duration of chirps by CDI stimulation. (E) CDI is rhythmically inhibited
during fictive flight. (F) Transition from fictive flight to fictive singing
demonstrates the change in activity. (G) CDI does not respond to acoustic
stimulation with 4.8 kHz, 85 dB SPL pulses 21 ms in duration. (H) Paired
recording from the dendrite in the mesothoracic ganglion and axon in the
prothoracic auditory neuropil of the same CDI. (I) Average of paired
recording from the same CDI shows, in this example, a delay of 2.4 ms
from spike recorded in the dendrite to spike recorded in the axon in the
prothoracic ganglion (n 0 181 spikes). CDI, intracellular CDI recording;
Wing Nerve, activity of mesothoracic nerve 3A. Vertical scale bars, 20 mV
[(A), (C), (F), (G), and dendrite in (H)], 10 mV [(E) and axon in (H)];
horizontal scale bars, 250 ms [(A), (G), (H)], 200 ms [(C) and (E)], 1 s (F).

393
520 27 JANUARY 2006 VOL 311 SCIENCE www.sciencemag.org

thoracic ganglion, and spikes in CDI elicited
IPSPs and PADs in both the left- and right-side
ON1s and afferents. Paired recordings of CDI
in its dendrite and prothoracic axon showed
that CDI spikes take 2.8 ms to reach the
auditory neuropil (n 0 4) (Fig. 2, H and I),
which leaves less than 1 ms from the spike in
the prothoracic ganglion to the release of IPSPs
or PADs. We never recorded a failure of an
IPSP or PAD in response to a CDI spike.
Therefore, both the anatomical evidence of
axonal arborizations of CDI in the auditory
neuropil and the physiological evidence sug-
gest that the connection from CDI to auditory
afferents and to ON1 is monosynaptic.
What effect does CDI have on auditory
processing? Paired recordings were obtained
from CDI and ON1 during continuous presen-
neurons, both with bilateral outputs, the per-
sisting inhibition was most likely due to spikes
in CDI_s contralateral partner cell. However,
we could not rule out the possibility of paral-
lel inhibitory neurons. To examine whether
the CDIs are the only neurons mediating
the inhibition of the auditory pathway, we
made a paired recording of CDI and ON1
during singing and then cut the contralateral
prothoracic-to-mesothoracic connective, which
contained the axon of the partner CDI (n 0 2).
When the animal_s singing recovered, ON1 was
inhibited by CDI during fictive singing and
failed to respond to the ongoing auditory
stimuli (Fig. 4, C and D). When CDI was hy-
perpolarized and prevented from spiking, ON1
responded to the sound pulses both in the chirp
interval and in the chirp (Fig. 4, E and F). This
REPORTS
two populations of neurons (25, 26). CDI is a
rare example of a functionally and anatomical-
ly identified neuron that extends throughout the
entire insect nervous system. Simultaneously,
CDI mediates the presynaptic inhibition of
auditory afferents with PADs and the post-
synaptic inhibition of an identified auditory
interneuron with IPSPs. This twofold inhibition
reduces the auditory response to self-generated
sounds and protects the cricket_s auditory
pathway from desensitization during sound
production, allowing it to remain sensitive to
environmental sounds (6, 12, 14). Thus, even
in the small nervous system of the cricket, self-
generated sensory signals are processed in a
similar way to more complex vertebrate ner-
vous systems (27, 28). During flight, CDI is
inhibited and prevented from firing; hence,

Downloaded from www.sciencemag.org on August 23, 2007

tation of computer-generated sound pulses at
the carrier frequency of cricket song (Fig. 4). In
resting crickets, when CDI was not spiking, the
sound pulses elicited a train of spikes in ON1.
When CDI was depolarized to make it spike,
the auditory response in ON1 was inhibited
(Fig. 4, A and B). During silent, fictive singing,
ON1_s response to the external sound stimuli
was suppressed during the chirps, even when
one CDI was prevented from spiking by in-
jecting hyperpolarizing current (n 0 8) (22).
Because CDI exists as a bilateral pair of
confirmed that CDI is both sufficient and
necessary to mediate the corollary discharge
inhibition during singing, and there is no evi-
dence for any parallel inhibitory pathways.
Over the past 50 years, investigation into
the cellular basis of corollary discharges and
efference copies has been hampered by a lack
of data on identified neurons mediating these
signals. There are now only a small number of
neurons thought to mediate corollary discharges
(5, 2326). The synaptic connectivity has been
explored with intracellular recordings in only
flying crickets_ hearing will not be impeded by
CDI and they can listen for signaling con-
specifics or echolocating calls from predating
bats (29). The singing cricket also generates
substantial nonauditory sensory feedback Ee.g.,
(30)^. The complex and widespread branches
of CDI indicate that it may also inhibit other
sensory pathways. CDI therefore provides an
opportunity to understand not only the role of
timing in corollary discharge signals but also
the computation by which motor and sensory
signals are integrated.

Fig. 3. Inhibitory inputs in auditory neurons are elicited by CDI spikes. (A)
Paired recording of CDI and an auditory afferent during singing. Spikes in CDI
coincide with PADs in the afferent. (B) Stimulation of CDI with intracellular
depolarizing current injection. Each CDI spike elicits a PAD in the afferent. (C)
Superposition of CDI and afferent recording triggered by spikes in CDI. In this
example, PADs were elicited after a constant delay of 3.4 ms from CDI spike.
(D) Paired recording of CDI and ON1 during singing. Spikes in CDI coincide
with IPSPs in ON1. (E) Every spike elicited by depolarizing current injection in
CDI elicits an IPSP in ON1. (F) Superimposed traces of CDI and ON1 show a
constant latency (in this example 3.4 ms) from CDI spike to the IPSP. Afferent:
intracellular auditory afferent recording; ON1, intracellular ON1 recording.
Vertical scale bars, CDI: 15 mV (A), 30 mV (B), 20 mV (D), 25 mV (E); ON1:
15 mV (D), 5 mV (E); afferent: 20 mV (A), 10 mV (B); current: 5 nA; horizontal
scale bars, 250 ms [(A) and (D)], 200 ms [(B) and (E)].

394
www.sciencemag.org SCIENCE VOL 311 27 JANUARY 2006 521
REPORTS
Fig. 4. The effect of CDI on References and Notes
sound processing. (A) ON1s 1. K. G. Pearson, Annu. Rev. Neurosci. 16, 265 (1993).
2. R. W. Sperry, J. Comp. Physiol. Psychol. 43, 482 (1950).
response to a continuous se-
3. E. von Holst, H. Mittelstaedt, Naturwissenschaften 37,
quence of 4.8 kHz, 75 dB SPL 464 (1950).
sound pulses, 8 ms in du- 4. M. Zaretsky, C. H. F. Rowell, Nature 280, 583 (1979).
ration with intervals of 7 ms, 5. M. A. Sommer, R. H. Wurtz, Science 296, 1480 (2002).
is completely inhibited dur- 6. J. F. A. Poulet, B. Hedwig, Nature 418, 872 (2002).
ing periodic current injec- 7. K. T. Sillar, P. Skorupski, J. Neurophysiol. 55, 678 (1986).
8. K. T. Sillar, A. Roberts, Nature 331, 262 (1988).
tion in CDI. (B) PSTH and 9. S. J. Blakemore, D. Wolpert, C. D. Frith, Nat. Neurosci. 1,
superimposed instantaneous 635 (1998).
spike frequency of ON1 10. C. C. Bell, Science 214, 450 (1981).
averaged over 36 trials dem- 11. J. E. Roy, K. E. Cullen, J. Neurosci. 24, 2102 (2004).
onstrate that ON1 activity is 12. J. F. A. Poulet, B. Hedwig, J. Neurosci. 23, 4717 (2003).
reduced during CDI stimula- 13. J. F. A. Poulet, B. Hedwig, J. Exp. Biol. 204, 1281 (2001).
14. J. F. A. Poulet, B. Hedwig, J. Neurophysiol. 89, 1528 (2003).
tion. (C) In an animal with the 15. See supporting material on Science Online.
contralateral prothoracic-to- 16. R. M. Hennig, J. Comp. Physiol. A 167, 629 (1990).
mesothoracic connective cut, 17. K. Michel, Z. Morph. Tiere 77, 285 (1974).
ON1 responds with a train of 18. K. Schildberger, D. W. Wohlers, F. Huber, in Cricket Behavior
spikes during the chirp inter- and Neurobiology, F. Huber, T. E. Moore, T. E. Loher, Eds.
(Cornell Univ. Press, Ithaca, NY, 1989), pp. 423458.

Downloaded from www.sciencemag.org on August 23, 2007

vals, but it fails to respond
19. M. D. Kirk, J. J. Wine, Science 225, 854 (1984).
during the chirp if CDI is 20. F. Clarac, D. Cattaert, Exp. Brain Res. 112, 163 (1996).
spiking. (D) The PSTH and in- 21. P. Rudomin, R. F. Schmidt, Exp. Brain Res. 129, 1 (1999).
stantaneous spike frequency 22. J. F. A. Poulet, B. Hedwig, data not shown.
of ON1 highlights the reduc- 23. M. G. Weeg, B. R. Land, A. H. Bass, J. Neurosci. 25, 5967
tion in ON1 response during (2005).
24. C. C. Bell, K. Dunn, C. Hall, A. Caputi, J. Comp. Physiol. A
the chirp. (E) When CDI is 177, 449 (1995).
prevented from spiking by 25. W. C. Li, S. R. Soffe, A. Roberts, J. Neurosci. 22, 10924
hyperpolarizing current injec- (2002).
tion, ON1 responds to sound 26. S. C. Rosen, M. W. Miller, E. C. Cropper, I. Kupfermann,
during the chirp and the J. Neurophysiol. 83, 1621 (2000).
chirp interval. (F) When CDI 27. K. E. Cullen, E. Curr. Op. Neurobiol. 14, 698 (2004).
28. C. C. Bell, K. Grant, J. Neurosci. 9, 1029 (1989).
spikes were suppressed by 29. T. G. Nolen, R. R. Hoy, Science 226, 992 (1983).
inhibitory current injection, 30. M. Dambach, H.-G. Rausche, G. Wendler, Naturwissenschaften
the PSTH and instantaneous 70, 417 (1983).
spike frequency show no 31. We thank S. Atkinson, M. Burrows, C. Petersen, and S. Rogers
reduction in the activity of for comments on the manuscript. Supported by the UK
Biotechnology and Biological Sciences Research Council, the
ON1 during chirps. Vertical
Royal Society, and the Human Frontier Science Program.
scale bar, current: 2.5 nA
(A); CDI: 20 mV [(A) and (E)], Supporting Online Material
10 mV (C); ON1: 20 mV [(A), www.sciencemag.org/cgi/content/full/311/5760/518/DC1
Materials and Methods
(C), and (E)]; horizontal scale
References
bars, 500 ms (A), 200 ms [(C)
and (E)]. 3 October 2005; accepted 20 December 2005
10.1126/science.1120847

Scaling of Connectivity in mary opportunity for dispersal. Although

larvae have the potential for long-distance
dispersal (1, 2), evidence is mounting that lar-
Marine Populations val dispersal may be limited (311). These
studies challenge assumptions about the domi-
R. K. Cowen,1* C. B. Paris,1 A. Srinivasan2 nant distance mode of dispersal for marine pop-
ulations (whether larvae typically travel a long
Defining the scale of connectivity, or exchange, among marine populations and determining or short distance) (12, 13). The rates, scale, and
the factors driving this exchange are pivotal to our understanding of the population dynamics, spatial structure of successful exchange, or
genetic structure, and biogeography of many coastal species. Using a high-resolution biophysical connectivity, among local populations of ma-
model for the Caribbean region, we report that typical larval dispersal distances of ecologically rine organisms drive population replenishment
relevant magnitudes are on the scale of only 10 to 100 kilometers for a variety of reef fish species. and, therefore, have profound implications for
We also show the importance of the early onset of active larval movement mediating the dispersal population dynamics and genetics of marine
potential. In addition to self-recruitment, larval import from outside the local area is required to organisms; spatially oriented resource manage-
sustain most populations, although these population subsidies are very limited in particular ment (e.g., marine protected areas); and the
systems. The results reveal distinct regions of population isolation based on larval dispersal that 1
Marine Biology and Fisheries, 2Meteorology and Physical
also correspond to genetic and morphological clines observed across a range of marine organisms. Oceanography, Rosenstiel School of Marine and Atmo-
spheric Science, University of Miami, 4600 Rickenbacker
dentifying the scale of marine larval ography. Most coastal marine species have

I
Causeway, Miami, FL 33149, USA.
dispersal remains one of the fundamental limited adult movement, so the relatively *To whom correspondence should be addressed. E-mail:
challenges to marine ecology and ocean- short, pelagic larval phase represents the pri- rcowen@rsmas.miami.edu

395
522 27 JANUARY 2006 VOL 311 SCIENCE www.sciencemag.org


396
Sensory signals during active versus passive movement
Kathleen E Cullen
Our sensory systems are simultaneously activated as the
result of our own actions and changes in the external world.
The ability to distinguish self-generated sensory events from
those that arise externally is thus essential for perceptual
stability and accurate motor control. Recently, progress has
been made towards understanding how this distinction is
made. It has been proposed that an internal prediction of the
consequences of our actions is compared to the actual sensory
input to cancel the resultant self-generated activation.
Evidence in support of this hypothesis has been obtained
for early stages of sensory processing in the vestibular,
visual and somatosensory systems. These findings have
implications for the sensorymotor transformations that are
needed to guide behavior.
Addresses
Department of physiology, 3655 Promenade Sir William Osler,
Montreal, Quebec H3G 1Y6, Canada
e-mail: kathleen.cullen@mcgill.ca
Current Opinion in Neurobiology 2004, 14:698706
This review comes from a themed issue on
Motor systems
Edited by Marjorie E Anderson and Ole Kiehn
Available online 30th October 2004
0959-4388/$ see front matter
# 2004 Elsevier Ltd. All rights reserved.
DOI 10.1016/j.conb.2004.10.002
Abbreviations
fMRI function magnetic resonance imaging
Introduction
Most of our sensory experiences are gained by active
exploration of the world, for example, by locomotion, eye
movement and touch. Normally, we can readily distin-
guish between sensory signals that register changes in the
external world (exafference) and those that result from
our own actions (reafference). The ability to make this
distinction is essential both for our perceptual stability
and for spatial orientation and the construction of neural
representations of the environment to guide behavior
accurately. When we make eye movements, for example,
the world moves across our retinas, but we do not perceive
the world as moving. How do we distinguish sensory
events resulting from our own actions from those that
arise externally?
So far, most analyses of sensory processing have been
done in an experimental framework that is designed to
Current Opinion in Neurobiology 2004, 14:698706
test passive rather than active sensation. Recent results
from several laboratories have, however, yielded major
insights into our understanding of how sensory signals are
processed during movement. In this review, I consider
recent advances in this field, focusing on experiments in
the vestibular system that have provided evidence for the
differential processing of reafference early in sensory
systems [1,2,3]. Parallels between findings in this sys-
tem and those emerging from recent studies of the visual
[4,5,6] and somatosensory [7,8,9] systems are also
considered. This body of work is discussed in relation
to prominent theories of motor control and addresses the
following questions. What information or cues are avail-
able to achieve perceptual stability? At what level in
sensory processing are active and passive sensory signals
differentially encoded? What mechanisms underlie dif-
ferential sensory processing? Finally, I consider the impli-
cations of these results in relation to how the processing of
sensory signals during motion is reflected in the sensory
motor transformations that are needed to guide behavior.
Conceptual frameworks and model systems
Historically, the idea that animal behavior is triggered by
stimuli and based on reflex elements that are linked
together in a chain of activations has provided an impor-
tant conceptual framework in neurophysiology. The
focus on relationships between afferences and efferences
follows logically from the influential theory of a reflex
chain first made popular by Sherrington in 1906 [10]. In
classical reflex theory, any given stimulus results in a
predictable motor response (i.e. a reflex), and complex
behavior can be explained as the combined effect of a
chain of reflexes. Although this conceptual framework
remains to be abandoned completely, it is clear that the
cause of every response cannot be solely attributed to a
sensory stimulus.
In 1950, two separately but simultaneously published
studies by Von Holst and Mittelstaedt [11] and Sperry
[12] re-evaluated this dominant model and demonstrated
a need for the traditional perspective to be reversed.
Classical reflex theory asks what is the relationship
between afference (sensory information) and efference
(motor output)? These investigators argued, however,
that it is necessary to ask what happens when efference
causes changes in the state of an organism that are then
reverberated back into the nervous system as reafference?
To avoid responding to sensory inputs that arise from self-
generated actions, the sensory system needs to know what
the motor system has done. On the basis of their observa-
tions, Von Holst and Mittelstaedt [11] proposed the
principle of reafference (Figure 1a), in which a copy
www.sciencedirect.com
397
 Sensory signals during active versus passive movement Cullen 699

Figure 1

(a)
First-stage
Higher-order
central
processing
processing

Efference copy Effector

Command signal

Sensory exafference
Sensory receptor

(b)
Electric organ

Negative image EO
E.O. discharge command

AR
EMF exafference Ampullary receptor
ELL

EO AR

(c)
Respiration muscles

Motor command RM
PR
EMF exafference Proprioceptive receptor
DON

AR AR RMPR
Ampullary receptor

Current Opinion in Neurobiology

(a) Simplified scheme of the reafference principle of von Holst and Mittelstaedt [11]. The motor command is sent to the effector muscle and
in turn sensory activation, resulting from the activation of sensory receptors by an effector, is returned. This reafference is then compared to an
efference copy of the original motor command. Here, reafference is arbitrarily marked + and the efference copy is marked . When the
reafference and efference copy signals are of equal magnitude they cancel, and no sensory information is transmitted to the next levels of
processing. By contrast, a difference between the reafference and the efference copy indicates an externally generated event (i.e. exafference)
that is considered behaviorally relevant and thus is processed further. (b) Primary afferent fibers from electroreceptors project to the cerebellum-like
electrosensory lobe (ELL) of mormyrid electric fish. Descending motor commands result in the generation of electric organ discharge, which
results in the activation of ampullary afferents (i.e. reafference). Corollary discharge signals associated with the motor command that elicits the
electric organ discharge are prominent in the mormyrid electrosensory lobe and are used to reduce the central effects of activity in ampullary
receptors evoked by the electric organ discharge (i.e. exafference) of the fish. Abbreviation: EMF, electromagnetic field. (c) Skate do not generate
electric fields but can detect them. The first-order neurons in this system are the principal cells of the dorsal octavolateral nucleus (DON). These
neurons receive proprioceptor inputs from the parallel fibers inputs of the dorsal granular ridge, which cancel electric magnetic field reafference
in the DON during respiration.

of the expected sensory results of a motor command, ized this original proposal by suggesting that an internal
which they termed the efference copy, is subtracted prediction of the sensory consequence of our actions,
from the sensory signal to eliminate reafferent informa- derived from the motor efference copy, is compared to
tion. More recent behavioral investigations have general- the actual sensory input [1315].

www.sciencedirect.com Current Opinion in Neurobiology 2004, 14:698706

398
700 Motor systems
In some model systems, including the electrosensory
systems of mormyrid fish [16,17] and elasmobranchii
(sharks, skates and rays [18]), the mechanosensory system
of the crayfish [19,20], and the auditory system of the
cricket [21], evidence indicates that sensory information
arising from self-generated behaviors is selectively sup-
pressed at the level of afferent fibers and/or the central
neurons to which they project. The mechanisms by which
suppression occurs can differ. Bell and co-workers [16,17]
have provided concrete support for Von Holst and
Mittelstaedts principle of reafference in their investiga-
tions of the electrosensory system in mormyrid fish. In
this system, a negative image of the predicted reafference
is generated and added to neurons at the first stage of
central processing (Figure 1b). As a result, the fish does
not respond to the discharge from its own electric organ.
In the skate, by contrast, information from proprioceptive
inputs alone can generate a negative image of self-motion
during respiration [18]. This negative image is then used
at the first central stage of electrosensory processing to
remove the modulation of its electroreceptors that results
from its own motion (Figure 1c).
Vestibular processing during active head
movements
The vestibular system provides information about head
motion relative to space that is necessary for maintaining
posture, computing spatial orientation and perceiving
self-motion. The differential processing of actively versus
passively generated vestibular stimuli is crucial, however,
for controlling eye, neck and body movements, as well as
for perceptual stability (reviewed in [1]). This point can
be easily appreciated by considering the simple example
of a vestibularly driven spinal reflex the vestibulo-
collic reflex which, in response to head motion, stabi-
lizes the head in space via activation of the neck muscu-
lature (Figure 2a). The compensatory head movements
produced by this reflex are clearly beneficial when the
behavioral goal is to stabilize head position in space.
When the behavioral goal is to make an active head
movement, however, the vestibular drive to the reflex
pathway would command an inappropriate head move-
ment to move the head in the direction opposite to the
intended goal.
At what level are the vestibular signals that arise from
passive versus active head movement first differentially
encoded? In the alert primate, the vestibular system does
not seem to distinguish between active and passive head
movements at the level of the vestibular afferents (Figure
2b; [2]), but the differential treatment of vestibular sig-
nals is evident at the next stage of processing. The head-
velocity-related modulation of one population of vestib-
ular nuclei neurons, which receive direct inputs from the
vestibular afferents, is markedly attenuated in response to
vestibular inputs that result from active head movements
(Figure 2c; [22,23]). Notably, these same neurons con-
Current Opinion in Neurobiology 2004, 14:698706
tinue to respond selectively to passive head motion when
a monkey generates voluntary head movements while
undergoing passive whole-body rotation.
There are at least three extravestibular cues that could
contribute to canceling reafference at the level of the
vestibular nuclei. First, inputs from neck proprioceptors;
second, knowledge of the self-generated motion; and
third, neck efference copy signals. Roy and I [3,23]
have explored each of these possibilities in the rhesus
monkey. First, we found that the activation of neck
proprioceptors is not sufficient for suppressing vestibular
reafference: neurons in the vestibular nuclei encode head
velocity similarly during passive rotations of the head
relative to the body and during passive rotations of the
head and body together.
Second, higher-order areas, such as parietoinsular vestib-
ular cortex, that are involved in the perception of self-
motion [24] are known to send substantial projections to
each of the vestibular nuclei [25]. We found, however,
that knowledge of self-generated head motion alone is not
sufficient to suppress vestibular reafference. Neurons
respond robustly to head velocity when monkeys drive
themselves through space by rotating a steering wheel
connected to the motor controller of a vestibular turntable
[23]. Third, a copy of the command to the neck muscles
(i.e. a motor efference signal) could be used to cancel
vestibular inputs during active head movements. We
found, however, that when head-restrained monkeys
attempted to move their heads, generating levels of neck
torque comparable to those issued during large active
head movements, neuronal responses were not modu-
lated [3].
Further exploration has provided evidence for Von Holst
and Mittelstaedts principle of reafference in the primate
vestibular system [3]. The activity of individual neurons
in the vestibular nuclei was recorded in monkeys making
active head movements and the correspondence between
intended and actual head movement was experimentally
controlled. We found that a cancellation signal was gen-
erated only when the activation of neck proprioceptors
matched the motor-generated expectation during active
head movements (Figure 3). This mechanism functions
to eliminate selectively self-movements from the subse-
quent computation of orientation and posture control.
These findings are the first to confirm Von Holst and
Mittelstaedts proposal at the level of single neurons in a
mammalian system.
Correlates for the differential processing of active and
passive vestibular inputs have been identified upstream
of the vestibular nuclei. Head direction cells in several
areas of the classic limbic circuit discharge preferentially
when a monkey or rat orients its head in a specific
preferred direction. These neurons, which are important
www.sciencedirect.com
399
 Sensory signals during active versus passive movement Cullen 701

Figure 2

(a) Vestibular nuclei

H
Vestibular Only

Horizontal Neck
canal motoneurons
Cerebellum
thalamus/cortex?

(b) Vestibular afferent (c) Vestibular nucleus neurons

(i) Passive head movement Passive head movement

100 H 100 H
100 spikes 100 spikes
FR FR
2s

(ii) Active head movement Active head movement

100 100
H H
100 spikes 100 spikes
FR FR
100 ms

(iii) Combined head movement Combined head movement

100 deg/s
100 H passive 100 H passive

H active H active
50 spikes
50 sp/s 50 spikes

H combined H combined

FR FR
2s

In the vestibular system, second-order neurons distinguish between sensory inputs that result from self-actions and sensory inputs that arise
externally. (a) A subclass of neurons in the vestibular nucleus, named vestibular-only neurons on the basis of their responses in head-restrained
models, receive direct projections from the semicircular canals and in turn project bilaterally to spinal motor neurons to activate the neck
musculature. These neurons probably also send projections to the cerebellum, thalamus and cortex. (b) Activity of an example horizontal canal
afferent during (i) passive head movements, (ii) active head movements and (iii) combined active and passive head-in-space motion. Afferents
reliably encode head-in-space motion in all conditions. (c) Activity of an example VO neuron in the vestibular nucleus during the head movements
described in (b). Neuronal responses to the active component of head-in-space motion are significantly attenuated; by contrast, the neurons
show no attenuation in response to the passive rotation component. Abbreviations: FR, firing rate;H, horizontal head velocity. Afferent responses are
based on data from [2], and (c) is modified from data reported in [23].

for spatial memory and navigation, respond more robustly
during active than during passive head rotations [26]. In
addition, neurons in the ventral interparietal area, one of
several vestibular areas identified in the parietotemporal
cortex, have been studied during active and passive head
movements. The results show that there are considerable
differences in neural responses during active versus pas-
sive head movements [27]. This differential processing of
vestibular inputs most probably reflects the integration of
information from an efference copy of motor commands
and from proprioceptive and vestibular sources required
for the perception of self-motion and for the representa-
tion of extrapersonal space. I take this point up again
further below.
Visual processing during voluntary eye
movements
How the visual world is perceived as stable despite
movements of the eyes, head and body is an issue that
concerned many eminent scientists of the last century
including von Helmholtz, Hering, Mach and Sherrington.
Although targets rapidly jump across the retina as we
move our eyes to make saccades, we never see the world
move over our retina. Helmholtz [28] made the salient,

www.sciencedirect.com Current Opinion in Neurobiology 2004, 14:698706

400
702 Motor systems

Figure 3

(c) Active head rotation

(a) Passive whole body
(b) Active head rotation cancelled by passive whole-
rotation
body-rotation

100 100
HS 100
HB active

Passive table

FR prediction 100 spikes 100 spikes Difference 100 spikes

prediction
FR
1s 200 ms

In the vestibular system, an internal model of the sensory consequences of active head motion is used to selectively suppress reafference
at the level of the vestibular nuclei. (a) Activity of an example VO neuron (gray-filled trace) during passive whole-body rotation. In this condition,
only vestibular inputs are available to the central nervous system and there is no motor efference copy signal because the monkey does
not actively move its head. (b) Activity of the same neuron during active head-on-body movements. In this condition, the monkey commands an active
head movement and thus an efference copy signal is theoretically available. In addition, the movement of the head activates vestibular and
proprioceptive afferents as a result of the head-in-space (HS ) and head-on-body (HB ) movement, respectively. A prediction of the activity of the
neuron, based on its response to passive head motion, is superimposed (bold trace). (c) The neuron is recorded as the monkey actively moves its
head; however, the head-in-space velocity (HS ) that is generated by the monkey (broken arrow) is experimentally cancelled by simultaneously
rotating the monkey in the opposite direction (unbroken arrow). Consequently, in this condition the head moves relative to the body (HB ),
but not to space (HS ); as a result, an efference copy signal is generated and the neck proprioceptors are activated, but vestibular afferent input is
greatly reduced. This approach reveals a cancellation signal that is sent to the vestibular nuclei in conditions in which neck proprioceptive inputs
match those expected based on the neck motor command, but not when these inputs are vastly different. Indeed, the marked inhibition in the
response of the neuron shows excellent correspondence to that from the difference in response during passive (a) versus active (b) head
movements. Abbreviation: FR, firing rate. Modified from data reported in [3].

and easily replicated, observation that tapping on the
canthus of the eye to displace the retinal image, for
example during a saccadic eye movement, results in an
illusionary shift of the visual world. How can continually
changing retinal inputs to the visual system thus result
in the perception of a stable visual world during eye
movements?
Our visual sensitivity is reduced during and just before a
saccadic eye movement a phenomenon that is referred
to as saccadic suppression. Psychophysical data have
shown that suppression is strongest for stimuli that would
preferentially activate the magnocellular dorsal visual
processing stream, which carries transient, motion-related
visual information (e.g. see [29]). Neural correlates for
saccadic suppression have been now identified at several
stages of visual processing. Significant saccade-related
responses can be observed early in visual processing at
the level of the dorsal lateral geniculate nucleus of the
thalamus [4,5]. Reppas et al. [5] found that the most
common effect was a biphasic modulation of response
strength (weak suppression followed by strong enhance-
ment), which was far more prominent for neurons in the
magnocellular than the parvocellular layer. Moreover,
correlates of saccadic suppression are most evident in
magnocellular recipient areas including medial temporal
and medial superior temporal cortex of rhesus monkey
[30].
Results of studies in humans are remarkably consistent
with these findings. For example, functional magnetic
resonance imaging (fMRI) studies show saccade-related
changes in areas that receive magnocellular input includ-
ing medial temporal cortex, V7 and V4 [31]. In addition,
transcranial stimulation studies in humans have provided
evidence that saccadic suppression occurs via extraretinal
mechanisms in the thalamus or primary visual cortex [6].
Note, however, that there might be differences in the
strategies adopted across species (e.g. see the work of
Olveczky et al. [32] in rabbit).
The mechanisms that underlie saccadic suppression
remain to be fully elucidated. Proprioceptive information
from the stretch receptors in the extraocular muscles
could be used for this purpose; however, the relative
timing of saccadic suppression and eye movement sug-
gests that inputs from saccadic planning centers (e.g. an
efference copy or corollary discharge signal) are crucial.

Current Opinion in Neurobiology 2004, 14:698706 www.sciencedirect.com

401
 Sensory signals during active versus passive movement Cullen 703
Somatosensory processing during active
touch
Touch is also clearly an active process. Fingers or whis-
kers, like eyes, move as they scan the external world.
These movements thus determine the nature of the
sensory input that will be encountered. As pointed out
by Gibson [33], the constancy of the tactual world is
characterized by the stability of an object. For example,
movement of a skin surface over a corner of an object does
not usually give rise to recognition of the tactile motion.
The whiskers (vibrissae) system in rats has proved to be
an excellent experimental model of active touch. During
whisking, the responses of first-order neurons in the
trigeminal ganglion of the vibrissae system cannot be
inferred from their responses to comparable passive whis-
ker deflection. Ahissar and co-workers [7] have recently
shown that, during intervals of whisking, these neurons
carry both a reference signal that encodes the current
position of the rats vibrissae and a more typical sensory
signal that encodes contact of the vibrissae with an object.
Timing differences between the reference and the con-
tact signals carried by neurons can be used to compute the
position of an object in space. In this system, the position
of vibrissae within a cycle could be encoded primarily
through direct sensory activation rather than via central
pathways. Nevertheless, there is some evidence for the
existence of a modulatory efference copy at the cortical
level that might influence the amplitude of whisking [34].
Central information of this kind can be used to tune large-
scale feedback loops in the vibrissae sensorimotor system
to optimize sensory processing [8].
In primates, correlates for the differential active proces-
sing of touch have been found in somatosensory cortex,
where neuronal responses are attenuated for self-
produced versus externally produced tactile stimuli
[35]. In humans, self-generated tactile stimulation does
not result in the same tickling sensation that arises when
the stimulation is externally produced [3638]. Blakemore
et al. [37] have shown, however, that when an artificial
delay or trajectory perturbation is introduced between a
movement and the resultant tactile stimulation, self-
generated stimulation is rated as more ticklish. Thus,
as in the vestibular system, when the sensory sensation
no longer corresponds to the motor command, the pre-
dicted image of the afferent signal (in this case a tickle
stimulus) does not fully cancel the distorted reafference.
The analogy between findings in these two systems
implies that the vestibular system is effectively ticklish.
Common strategies?
Studies of the primate vestibular system have shown that
a cancellation signal is used at the first level of central
processing to suppress vestibular activation resulting from
active head movements. An investigation by Roy and
myself [3] found that this cancellation signal is gener-
www.sciencedirect.com
ated only when the activation of neck proprioceptors
matches the motor-generated expectation, thereby elim-
inating vestibular signals that result from active head-on-
body movements. Such a mechanism has clear analogies
to that used by the electrosensory system of the mormyrid
fish to cancel reafference. In addition, recent behavioral
studies in humans have shown that a similar strategy is
used by the somatosensory system.
Where is the essential cancellation signal actually com-
puted for each system? Evidence from work in electric
fish indicates that the cerebellum-like electrosensory
lobes provide the signal that is used to cancel the sensory
response to self-generated stimulation [39]. Moreover,
fMRI studies have suggested that the cerebellum has a
similar role in the suppression of tactile stimulation dur-
ing self-produced tickle [3638]. More recently, however,
the results of a fMRI study of active hand movements has
led to the proposal that efference copy information is not
located in the cerebellum per se, but instead is imple-
mented via interactions between perceptual areas and
motor areas in a task-dependent way [40]. Identifying
the neural representations of efference copy information
in different tasks promises to be an interesting area of
investigation.
It remains to be determined whether a comparable frame-
work can be used to describe the processing that selec-
tively suppresses reafference across all or even most
modalities. It is now clear that other mechanisms can
contribute to the differential processing of reafference.
For example, peripheral mechanisms have an essential
role in processing touch during active whisking, as
described above. In addition, a recent study has shown
that during active wrist movements in primates, cutaneous
inputs are presynaptically inhibited at the level of the
spinal cord afferents [9]. The timing of the attenuation
suggests that descending motor commands, rather than
peripheral feedback from the movement, generate the
inhibition. This strategy is markedly different from that
used by the vestibular system: vestibular afferents do not
differentially encode active versus passive head move-
ments; instead, reafference is distinguished only at the
next stage of processing in the vestibular nuclei.
Implications: consequences for motor
control
The differentiation of sensory stimulation that arises from
passive versus active movement is not only crucial for
perceptual stability but is also required to produce accu-
rate neural representations of the environment to guide
behavior accurately. The implications of this are particu-
larly obvious for the vestibular system, where the second-
order sensory neurons function as both sensory and pre-
motor neurons. Second-order vestibular neurons, which
are differentially sensitive to active and passive head
movements, project to the spinal cord and mediate
Current Opinion in Neurobiology 2004, 14:698706
402
704 Motor systems
postural reflexes such as the vestibulo-collic reflex
(Figure 2a). Accordingly, their reduced sensitivity during
active head movements is consistent with their functional
role in head stabilization, because their modulation is
actually counterproductive during these self-generated
movements. The fact that these same neurons continue
to encode information faithfully about passive head rota-
tions that occur during the execution of voluntary move-
ments [22,23] is also important. As a person explores the
environment, these neurons will selectively respond to
adjust postural tone in response to any head movements
that the brain does not expect. This selectivity can be
crucial, for example, in recovering from tripping over an
obstacle while walking or running, which requires a
robust postural response.
Vestibular pathways that control postural responses must
combine vestibular inputs with information from other
sources, such as neck and body proprioceptive informa-
tion, to generate appropriate motor responses. This can be
easily appreciated by considering two subjects: one whose
head is pointing forward, and another whose head is
rotated 908 to one side. To ensure postural stability,
the same vestibular stimulus will need to activate differ-
ent muscle combinations in each subject. In this example,
combining neck proprioceptive information with head-
referenced vestibular inputs to the brain will provide
information about motion of the body in space and,
indeed, there is recent evidence for such a multimodal
integration of vestibular and proprioceptive inputs. Some
neurons in the rostral fastigial nucleus of alert primates
encode movement of the body rather than the head in
space during passively applied rotational [41] and transla-
tional [42] motion. Thus, during passive movements
these neurons encode movement in a body-referenced
coordinate system.
To date, coordinate transformations, which have been
observed in many brain areas, have been characterized in
studies designed to test passive rather than active sensa-
tion. What do these neurons encode during active move-
ments? The rostral fastigial nucleus is a particularly
interesting example to consider with regard to this ques-
tion. It is thought to be important in vestibulo-spinal
control, including the regulation of gait and postural
mechanisms. Thus, the same arguments made above
for the utility of selective gating of vestibular reafference
at the level of the vestibulo-spinal neurons could be
applied to the rostral fastigial nucleus. By extension, a
similar logic can be applied to areas that have been
implicated in higher-order perceptual functions. Spatial
perception is more accurate in response to active than to
passive rotations [4345]. The absence of a signal related
to motor efference copy, at the level of higher-order
vestibular areas such as parietal cortex might prevent
the complete updating required for accurate spatial loca-
lization [46], consistent with the differences in the pro-
Current Opinion in Neurobiology 2004, 14:698706
cessing of active and passive movement that can be found
at these higher levels of processing.
Conclusions
Over 50 years ago, Von Holst and Mittelstaedt [11]
proposed that the brain generates a sensory expectation
based on the motor command, compares it with the actual
sensory feedback, and subtracts the self-generated sen-
sation. In this way, the nervous system could theoreti-
cally differentiate sensory inputs that arise from external
sources from those that result from self-generated move-
ments. Recent work in several systems has provided
evidence in support of this hypothesis, as well as evi-
dence for other mechanisms that suppress reafference in
the early stages of sensory processing. It remains a
challenge to understand how the differential processing
of sensory inputs in the early stages is used by the
upstream networks that mediate perceptual stability
and guide behavior.
Acknowledgements
I thank J Roy, S Sadeghi, M McCluskey, A Andrei and M Van Horn
for helpful discussions on this topic; and A Andrei for valuable
contributions to the design and exposition of the illustrations. This
work was supported by a grant from the Canadian Institutes of Health
Research (CIHR). The author holds a McGill Dawson Chair and a
CIHR Investigator Award.
References and recommended reading
Papers of particular interest, published within the annual period of
review, have been highlighted as:
 of special interest
 of outstanding interest
1. Cullen KE, Roy JE: Signal processing in the vestibular system
 during active versus passive head movements.
J Neurophysiol 2004, 91:1919-1933.
This review contrasts the responses of two classes of neuron in the
vestibular nuclei during active movement: PVP neurons, which mediate
the vestibulo-ocular reflex (VOR); and VO neurons, which are thought to
mediate the vestibulo-collic reflex (VCR) and to shape vestibular informa-
tion for the computation of spatial orientation by upstream structures. The
head-velocity signals carried by VOR interneurons (PVP neurons) are
reduced when the goal is to redirect gaze in space, regardless of whether
head movement is active or passive. By contrast, the vestibular signals
carried by VCR interneurons (the neurons discussed in the current review)
are reduced in response to active head-on-body movements. The
mechanisms that underlie this differential processing of vestibular infor-
mation by PVP and VO neurons use efference copies of gaze- and neck-
movement commands, respectively.
2. Cullen KE, Minor LB: Semicircular canal afferents similarly
encode active and passive head rotation: implications for the
role of vestibular efference. J Neurosci 2002, 22:RC226:1-7.
3. Roy JE, Cullen KE: Dissociating self-generated from passively
 applied head motion: neural mechanisms in the vestibular
nuclei. J Neurosci 2004, 24:2102-2111.
This study shows that vestibular signals that arise from self-generated
head movements are inhibited by a mechanism that compares the
internal prediction of the sensory consequences by the brain to the actual
resultant sensory feedback. Self-generated vestibular inputs are selec-
tively cancelled early in processing at the level of second-order neurons.
4. Ramcharan EJ, Gnadt JW, Sherman SM: The effects of saccadic
eye movements on the activity of geniculate relay neurons in
the monkey. Vis Neurosci 2001, 18:253-258.
5. Reppas JB, Usrey WM, Reid RC: Saccadic eye movements
modulate visual responses in the lateral geniculate nucleus.
Neuron 2002, 35:961-974.
www.sciencedirect.com
403
 Sensory signals during active versus passive movement Cullen 705
6. Thilo KV, Santoro L, Walsh V, Blakemore C: The site of saccadic
 suppression. Nat Neurosci 2004, 7:13-14.
Transcranial magnetic stimulation of the human occipital cortex and
retina is used to induce illusory visual perceptions (phosphenes). The
authors show that the perception of phosphenes induced by stimulation
of occipital cortex is immune to saccadic suppression, whereas that of
phosphenes induced by retinal stimulation is not.
7. Szwed M, Bagdasarian K, Ahissar E: Encoding of vibrissal active
 touch. Neuron 2003, 40:621-630.
During active touch, the response of neurons in the trigeminal ganglion of
the vibrassae system can be divided into two categories: first, touch-
sensitive neurons that encode either the interval of whisker contact with
the object or the timing of object contact or detachment; and second,
whisking neurons, which are active during the motor act of whisking
regardless of whether the object touches an object. Accordingly, timing
differences between the discharges of touch-sensitive neurons and those
of whisking-sensitive neurons encode horizontal object position. The
authors propose that locked loops, such as those in the paralemniscal
system [8], could be used to detect timing differences between these cells
and to translate them into a spike count code.
8. Ahissar E, Kleinfeld D: Closed-loop neuronal computations:
focus on vibrissa somatosensation in rat. Cereb Cortex 2003,
13:53-62.
9. Seki K, Perlmutter SI, Fetz EE: Sensory input to primate spinal
 cord is presynaptically inhibited during voluntary movement.
Nat Neurosci 2003, 6:1309-1316.
This study provides evidence for differential encoding of somatosensory
inputs at the afferent level in primates. Cutaneous afferents to the spinal
cord are presynaptically inhibited during active wrist movements. This
inhibition occurs preferentially during active movements and precedes
electromyographic activity, suggesting that descending motor com-
mands, rather than peripheral feedback from the motion itself, have a
dominant role in generating the inhibition. In addition, the monosynaptic
responses in active first-order interneurons are reduced simultaneously
with increased excitability of the relevant afferent terminals.
10. Sherrington CS: The Integrative Action of the Nervous System.
New York: Charles Scribners Sons; 1906: 306-307.
11. von Holst E, Mittelstaedt H: Das reafferenzprinzip.
Naturwissenschaften 1950, 37:464-476. [Translation of title:
The principle of reafference].
12. Sperry R: Neural basis of spontaneous optokinetic response
produced by visual inversion. J Comp Physiol Psychol 1950,
43:482-489.
13. Decety J: Neural representation for action. Rev Neurosci 1996,
7:285-297.
14. Farrer C, Franck N, Paillard J, Jeannerod M: The role of
proprioception in action recognition. Conscious Cogn 2003,
12:609-619.
15. Wolpert DM, Ghahramani Z, Jordan M: An internal model for
sensorimotor integration. Science 1995, 269:1880-1882.
16. Bell C: An efference copy which is modified by reafferent input.
Science 1981, 214:450-453.
17. Mohr C, Roberts PD, Bell CC: The mormyromast region of the
mormyrid electrosensory lobe. I. Responses to corollary
discharge and electrosensory stimuli. J Neurophysiol 2003,
90:1193-1210.
18. Hjelmstad GO, Parks G, Bodznick D: Motor corollary discharge
activity and sensory responses related to ventilation in the
skate vestibulolateral cerebellum: implications for
electrosensory processing. J Exp Biol 1996, 199:673-681.
19. Krasne FB, Bryan JS: Habituation: regulation through
presynaptic inhibition. Science 1973, 182:590-592.
20. Edwards DH, Heitler WJ, Krasne FB: Fifty years of a command
neuron: the neurobiology of escape behavior in the crayfish.
Trends Neurosci 1999, 22:153-161.
21. Poulet JF, Hedwig B: Corollary discharge inhibition of
 ascending auditory neurons in the stridulating cricket.
J Neurosci 2003, 23:4717-4725.
This article examines how the auditory pathways in crickets process self-
generated auditory signals during singing. Modulation of neurons during
pharmacologically elicited sonorous, silent and fictive singing indicates
www.sciencedirect.com
that responses to self-generated sound are reduced by a centrally
generated corollary discharge.
22. McCrea RA, Gdowski GT, Boyle R, Belton T: Firing behavior of
vestibular neurons during active and passive head
movements: vestibulo-spinal and other non-eye-movement
related neurons. J Neurophysiol 1999, 82:416-428.
23. Roy JE, Cullen KE: Selective processing of vestibular
reafference during self-generated head motion.
J Neurosci 2001, 21:2131-2142.
24. Deutschlander A, Bense S, Stephan T, Schwaiger M, Brandt T,
Dieterich M: Sensory system interactions during simultaneous
vestibular and visual stimulation in PET. Hum Brain Mapp 2002,
16:92-103.
25. Akbarian S, Grusser OJ, Guldin WO: Corticofugal connections
between the cerebral cortex and brainstem vestibular nuclei in
the macaque monkey. J Comp Neurol 1994, 339:421-437.
26. Wiener SI, Berthoz A, Zugaro MB: Multisensory processing in
the elaboration of place and head direction responses by
limbic system neurons. Brain Res Cogn Brain Res 2002,
14:75-90.
27. Klam F, Graf W: Vestibular signals of posterior parietal cortex
neurons during active and passive head movements in
macaque monkeys. Ann N Y Acad Sci 2003, 1004:271-282.
28. von Helmholtz H: Handbuch der Physiologischen Optik, vol 3, edn
1. Hamburg: Voss; 1867. Treatise on Physiological Optics (English
Translation), vol 3. Translated by Southall JPC. Rochester: Optical
Society of America; 1925:44-51.
29. Burr DC, Morrone MC, Ross J: Selective suppression of the
magnocellular visual pathway during saccadic eye
movements. Nature 1994, 371:511-513.
30. Thiele A, Henning P, Kubischik M, Hoffmann KP: Neural
mechanisms of saccadic suppression. Science 2002,
295:2460-2462.
31. Kleiser R, Seitz RJ, Krekelberg B: Neural correlates of saccadic
suppression in humans. Curr Biol 2004, 14:386-390.
32. Olveczky BP, Baccus SA, Meister M: Segregation of object and
 background motion in the retina. Nature 2003, 423:401-408.
This study describes a subclass of center surround retinal ganglion cells
in rabbit retina that respond optimally when there are differences in center
versus surround motion. In other words, these neurons respond to motion
in the visual field but only if the surround and center move with different
trajectories. Thus, these neurons would not be stimulated by global retinal
motion arising from retinal drift during fixation or more rapid movements
during saccades.
33. Gibson JJ: Observations on active touch. Psychol Rev 1962,
69:477-491.
34. Fee MS, Mitra PP, Kleinfeld D: Central versus peripheral
determinants of patterned spike activity in rat vibrissa cortex
during whisking. J Neurophysiol 1997, 78:1144-1149.
35. Chapman CE: Active versus passive touch: factors influencing
the transmission of somatosensory signals to primary
somatosensory cortex. Can J Physiol Pharmacol 1994,
72:558-570.
36. Blakemore SJ, Wolpert DM, Frith CD: Central cancellation of
self-produced tickle sensation. Nat Neurosci 1998, 1:635-640.
37. Blakemore SJ, Wolpert DM, Frith CD: The cerebellum
contributes to somatosensory cortical activity during self-
produced tactile stimulation. NeuroImage 1999, 10:448-459.
38. Blakemore SJ, Wolpert DM, Frith CD: Spatio-temporal
prediction modulates the perception of self-produced stimuli.
J Cogn Neurosci 1999, 11:551-559.
39. Bell CC, Han VZ, Sugawara Y, Grant K: Synaptic plasticity in the
mormyrid electrosensory lobe. J Exp Biol 1999, 202:1339-1347.
40. Leube DT, Knoblich G, Erb M, Grodd W, Bartels M, Kircher TT:
 The neural correlates of perceiving ones own movements.
NeuroImage 2003, 20:2084-2090.
In this study, subjects opened and closed their hand and their movement
was filmed and projected online onto a screen. The temporal delay
Current Opinion in Neurobiology 2004, 14:698706
404
706 Motor systems
between movement and its visual feedback was varied, and subjects
were asked to report whether there was a delay. fMRI imaging during this
task shows that activity in the posterior superior temporal sulcus (pSTS) is
positively correlated with the magnitude of the temporal delay between a
self-generated movement and visual feedback. By contrast, the cere-
bellum seems to be involved in providing information related to the
conscious detection of visual event timing.
41. Kleine JF, Guan Y, Kipiani E, Glonti L, Hoshi M, Buttner U: Trunk
position influences vestibular responses of fastigial nucleus
neurons in the alert monkey. J Neurophysiol 2004,
91:2090-2100.
42. Shaikh AG, Meng H, Angelaki DE: Multiple reference frames
for motion in the primate cerebellum. J Neurosci 2004,
24:4491-4497.
Current Opinion in Neurobiology 2004, 14:698706
43. Blouin J, Gauthier GM, Vercher JL: Failure to update
the egocentric representation of the visual space
through labyrinthine signal. Brain Cogn 1995,
29:1-22.
44. Fery YA, Magnac R, Israel I: Commanding the direction of
passive whole-body rotations facilitates egocentric spatial
updating. Cognition 2004, 91:B1-B10.
45. Jurgens R, Boss T, Becker W: Estimation of self-turning in the
dark: comparison between active and passive rotation.
Exp Brain Res 1999, 128:491-504.
46. Duhamel JR, Colby CL, Goldberg ME: Ventral intraparietal area
of the macaque: congruent visual and somatic response
properties. J Neurophysiol 1998, 79:126-136.
www.sciencedirect.com
405


406

Memory-based expectations in electrosensory systems
Curtis C Bell
Adaptive processing of electrosensory information occurs in the
cerebellum-like structures of three distinct groups of fish.
Associations within each of these structures result in the
generation of negative images of predictable features of the
sensory inflow. Addition of these negative images to the actual
inflow removes the predictable features, allowing the
unpredictable, information-rich sensory signals to stand out.
Evidence from all three groups of fish indicates that the negative
images are mediated by plasticity at parallel fiber synapses.
Addresses
Neurological Sciences Institute, Oregon Health Sciences University,
505 NW 185th Avenue, Beaverton, OR 97006, USA;
e-mail: bellc@ohsu.edu
Current Opinion in Neurobiology 2001, 11:481487
0959-4388/01/$ see front matter
2001 Elsevier Science Ltd. All rights reserved.
Abbreviations
DON dorsal octaval nucleus
ELL electrosensory lobe
EOD electric organ discharge
epsp excitatory post-synaptic potential
NMDA N-methyl-D-aspartate
Introduction
Removal of expected or predicted sensory input is one of
the very general functions of sensory processing.
Organisms are interested in what is new, and predictable
sensory inputs convey no information. Thus, adaptation of
receptors or neurons to maintained stimuli removes
expected levels of stimulation over time, and lateral
inhibitory networks remove expected mean levels over
space. Simple and universal expectations such as these can
be removed by non-plastic cellular or network mecha-
nisms, but plasticity and memory-like mechanisms are
needed to take full advantage of all the predictability avail-
able to an organism. Associations between stimuli of the
same modality, between stimuli of different modalities, or
between motor commands and their feedback can all be
used to predict sensory patterns. But such associations
come and go, varying with the individual and the circum-
stances, and memory-based mechanisms are needed to
make use of them.
The generation and subtraction of memory-based expecta-
tions has been demonstrated in the electrosensory systems
of three different groups of fish, and some of the cellular
mechanisms responsible for these expectations have also
been identified.
The electrosense and cerebellum-like sensory
structures
All animals in the water create electric fields around them,
and a surprisingly large number of different groups of fish
481
possess electroreceptors in their skin that are exquisitely
sensitive to such fields [1,2]. Electric fish, not only have
electroreceptors for sensing current but also posses electric
organs for generating current in the form of electric organ
discharges (EODs [3]). Such fish detect nearby objects by
distortions in the pattern of self-generated current that flows
through their skin. Thus, electric fish can not only sense
external voltage sources but can also sense the impedances
of nearby objects. The sensing of external voltage sources is
known as passive electrolocation, whereas the sensing of
impedances by means EODs and electroreceptors is known
as active electrolocation. Passive electrolocation is mediated
by ampullary electroreceptors, which respond to low fre-
quency electrical signals by modulating their discharge
frequency up and down. Active electrolocation is mediated
by tuberous electroreceptors which are especially sensitive
to the high frequency signals contained in EODs.
Primary afferent fibers from electroreceptors terminate in
cerebellum-like sensory structures ([4]; Figure 1). Similar
structures in non-electroreceptive fish, amphibians and
mammals receive input from mechanical lateral line
receptors, auditory receptors and vestibular receptors.
The primary afferent fibers from all of these receptors ter-
minate in the deep layers of the cerebellum-like
structures where they form maps of the sensory surface.
Principal cells of these structures have basilar dendrites
that are affected directly, or indirectly via interneurons, by
the primary afferents. These cells also have apical den-
drites in the molecular layer that are affected by parallel
fiber input. The parallel fibers arise from granule cells
located in an external granule cell mass.
The cerebellum-like structures that receive afferent input
from electroreceptors and which will be the focus of this
review are the electrosensory lobes (ELL) of teleost elec-
tric fish and the dorsal octaval nucleus (DON) of
elasmobranchs (sharks and rays). Granule cells giving rise
to parallel fibres in the ELL and DON receive signals
which include: corollary discharge input coupled to motor
commands; proprioceptive cues conveying information
about movements of the body or the fins; and electrosen-
sory information that descends from higher stages of the
electrosensory systems [59]. These signals are referred to
as predictive in Figure 1 because they are likely to be
associated with changes in electrosensory input and can
therefore be used to predict such changes.
Subtraction of memory-based expectations of
sensory input
The generation and subtraction of memory-based expecta-
tions has been observed in the electrosensory systems of
electric teleosts of the Mormyridae and Gymnotidae fami-
lies, and elasmobranch fish [10]. Each of these groups is
407
482 Sensory systems
Figure 1
Predictive inputs
Corollary discharge signals
Higher levels of the same modality
Other sensory modalities
(e.g. proprioception)
? their apical dendrites in the molecular layer.
Input from a sensory surface
Current Opinion in Neurobiology
believed to have evolved their electrosensory system
independently [11].
The mormyrid ELL
Memory-based expectations were first observed in electric
fish of the family Mormyridae in the region of the ELL that
receives afferents from ampullary electroreceptors, and is
responsible for passive electrolocation [12,13]. Ampullary
afferents respond to the pulsatile EOD with a change in
discharge frequency; but such responses could interfere with
the sensory function of the ampullary system. Responses of
these afferents to the fishs own EOD vary with water
conductivity and the surrounding environment. Thus,
cancellation or minimizing of the EOD-induced interference
requires an adaptive memory-based mechanism.
The principal cells of the mormyrid ELL are not only
affected by afferent input from the periphery but also by a
centrally originating corollary discharge signal coupled to
the motor command that elicits the EOD. Corollary dis-
charge effects are studied in fish in which the synaptic
effect of electromotor neurons on the electric organ has
been blocked by the drug curare. The EOD is blocked in
these fish but the motor command that would normally
elicit an EOD continues to be emitted. Artificial electrical
stimuli are delivered to the receptive field of the cell at
fixed times with regard to the motor command. Responses
to the corollary discharge are minimal when no electrosen-
sory stimuli have been given for several minutes but
become prominent after pairing the command with an
effective sensory stimulus for a few minutes (Figure 2).
The responses to corollary discharge alone after pairing are
a highly specific negative image of the previously paired
response to a sensory stimulus [13]. The negative image is
specific to the timing of the sensory response after the
Schematic drawing showing major features of
cerebellum-like sensory structures. The
principal cells of these structures receive
sensory input from the periphery on their
Granule layer
basal dendrites in the sensory input layer, and
central inputs from other brain structures on
The principal cells integrate these two inputs
and the results are conveyed to higher levels
of the same sensory system by the axons of
Molecular layer principal cells. The central inputs to the
molecular layer are parallel fibers that convey
various types of predictive signals.
Association between the predictive signals
and sensory input from the periphery results in
Principal cell layer
changes in synaptic efficacy at the parallel
fiber synapses. As a result of these synaptic
changes, the predictive signals come to elicit
Sensory input layer negative images of sensory patterns that are
predicted on the basis of past associations.
Inhibitory cells of the molecular layer are
shown in black (reproduced with permission
from [10]).
motor command (out to about 120 ms), to the polarity of the
sensory response (increase or decrease in firing frequency;
Figure 2a), and to the amplitude of the sensory response.
The negative image is also specific to the spatial distribu-
tion of the sensory response across the surface of the ELL,
as the negative image is only formed if the sensory stimu-
lus is within the receptive field of a cell and evokes a
response. During sensory stimulation, the addition of the
negative image of past or predicted input minimizes the
central effect of predictable features in concurrent afferent
input (compare the effect of command plus stimulus at the
beginning and at the end of pairing in Figure 2b).
The responses to corollary discharge alone disappear
within a few minutes after turning off the sensory stimulus
(Figure 2b). This disappearance is not a passive decay,
however, but rather an active rematching of the changed
sensory input that now follows the motor command [14].
The negative image can last 30 minutes if rematching is
prevented by silencing the EOD motor command during
the intervening period.
Adaptive corollary discharge effects are also found in
regions of the mormyrid ELL that receive afferent input
from tuberous electroreceptors and which are responsible
for active electrolocation [15,16]. The subtraction of pre-
dicted sensory input in these regions probably enhances
sensitivity to small changes in EOD-evoked input.
The gymnotid ELL
The gymnotid studies have been done in regions of ELL
that are responsible for active electrolocation in species
with continuous, near sinusoidal EODs. The electric
organs of these species extend from near the head to the
tip of the tail. When the tail is bent, the electroreceptors on
408

Figure 2
Negative images of predicted sensory (a)
responses in the ELL of mormyrid fish. Pairing with excitatory
Recorded cells were in the ampullary region of sensory stimulus
the ELL. (a) At the left, three histograms show
responses to EOD motor command alone
before pairing (top), responses to command
paired with a brief (1ms) excitatory sensory
stimulus (middle, stimulus indicated by arrow
head), and responses to command alone after
pairing for 5 minutes (bottom). Right: same cell
6 minutes later when the responses to
command alone have returned to a baseline
level. Same sequence as in (a), except that the
pairing was with an inhibitory sensory stimulus.
(In ampullary electroreceptors, a simple
inversion of the stimulus from a positive to a
negative voltage inverts the response.) In both
cases, the responses to command alone after
pairing are opposite to the effects of the
previously paired stimuli. (b) Raster display of occurred at time 0. The electrosensory
another cell in the ampullary region showing stimulus was delivered immediately afterwards,
the effect of pairing the EOD command with a as indicated by the black line. The response to
brief sensory stimulus that evoked a pause- command plus stimulus is reduced after
burst in cell activity. The EOD command several minutes of pairing and this is as a
one side of the body are more strongly affected by the
EOD, whereas those on the other side are less strongly
affected. These changes in EOD-evoked afferent
responses caused by tail movement, however, convey no
information about external conductances and would in fact
interfere with their perception.
The potentially interfering effect of tail movement is
removed by the addition of a negative image of predicted
sensory input, as in the ampullary region of the mormyrid
ELL. Cells in the ELL of a discharging gymnotid fish
were recorded while passively bending the tail ([5,17,18],
Figure 3a). Initially, tail bend alone had little effect on the
cells. An effective electrosensory stimulus was then deliv-
ered to the receptive field of the cells at a fixed phase of
the tail bend. When the electrosensory stimulus was
turned off after several minutes of pairing, tail bend alone
now evoked vigorous responses from the cells that were
opposite to the effects of the stimulus. As in the mormyrid
ELL, this negative image was specific to the timing
(phase), polarity, amplitude and spatial location of the
previously paired stimulus. Similarly, the disappearance of
the negative image after turning off the sensory stimulus is
an active rematching of the change in sensory input and
not a passive decay [17].
Two different types of predictive signals are used to evoke
the negative images in the gymnotid ELL proprioceptive
signals indicating changes in body position, and whole body
electrosensory stimuli. The role of proprioceptive signals
was demonstrated by repeating the tail bending experi-
ments in a fish in which the electric organ was silenced. The
role of whole body electrosensory stimuli was demonstrated
by keeping the fish still while delivering an electrosensory
Memory-based expectations in electrosensory systems Bell 483
(b) EOD command
Pairing with inhibitory
sensory stimulus EOD
Command alone
before
EOD
EOD
Command alone
before (9 min of C+S pairing) Command
plus
stimulus
2min
EOD
Command plus
stimulus EOD
Command alone
after
EOD
Command alone
after
0 160 msec
50ms
Current Opinion in Neurobiology
result of the development of a response to
command alone that is a negative image
(a burst-pause) of the previously paired
sensory response. (Reproduced with
permission from [14].)
stimulus to the whole body that mimicked the voltage
changes caused by tail movement. Pairing a local elec-
trosensory stimulus with the whole body electrosensory
stimulus at a fixed phase resulted in a negative image of the
response to a local electrosensory stimulus that was evoked
by the whole body stimulus.
The elasmobranch DON
Studies of adaptive sensory processing in elasmobranchs
have been done in the passive electrolocation system of
skates [19,20,21]. Elasmobranch electroreceptors are
extremely sensitive to electric fields (thresholds
>0.01 V/cm) and respond strongly to the fields created by
the fishs own ventilation [22], but such responses convey no
useful information and could interfere with the passive elec-
trolocation of external voltage sources.
Principal cells of the DON do not respond to the fishs own
ventilation even though the afferent fibers respond strongly
[22]. The cancellation of responses to afferent input is
largely due, once again, to the addition of a negative image
of predicted sensory responses. This was demonstrated by
delivering electrosensory stimuli to the receptive fields of
DON cells at a fixed phase of the ventilatory rhythm
(Figure 3b). Responses to the extra electrosensory stimulus
decreased after a few minutes of pairing and were essentially
gone after 25 minutes. When the extra stimulus was turned
off, ventilation alone evoked strong responses (Figure 3b)
which had not been present before the pairing and which
were negative images of the paired sensory effects. The
negative images were specific to the timing (phase), polari-
ty, amplitude and spatial location of the paired sensory
stimuli, as in the other two systems. The elasmobranch sys-
tem used three distinct types of predictive signals to evoke
409
484 Sensory systems
Figure 3
(a) (b)
Ex In
Tailbend
alone
0 min
1.5 min
Tailbend
plus
3 min
stimulus
6 min
15 min
25 min
Tail bend
alone
2 min
4 min
20 0 +20 0 20
8 min
Tail displacement (degs) 0 1 sec
Current Opinion in Neurobiology
the negative images: corollary discharge signals associated
with ventilatory motor commands; proprioceptive signals
associated with ventilation; and whole-body electrosensory
stimulation associated with ventilation [20].
Disappearance of the negative image within a few minutes
after turning off the electrosensory stimulus is again, as in
mormyrids and gymnotids, an active rematching of the
change in sensory response. This was shown using passive
bending of the fin as the predictive stimulus and pairing
such bending with a local electrosensory stimulus. The neg-
ative image was still present three hours after turning off the
electrosensory stimulus, if rematching was prevented by not
moving the fin during the intervening three hour period.
Sites and mechanisms of the plastic change
responsible for the negative images
The negative images of predicted input are due, in large
part at least, to plastic changes within the cerebellum-like
structures themselves. This was shown by pairing the
predictive signals with intracellular current injection rather
than with sensory stimuli, reasoning that if the intracellular
current affects only the recorded cell and if the pairing
results in the generation of a negative image, then changes
must have taken place at synapses onto the recorded cell.
Depolarizing intracellular current pulses were paired with
the EOD corollary discharge in mormyrid fish [23] and
with ventilation in elasmobranch fish [20]. After the
pairing, the same predictive signals alone evoked
hyperpolarizing responses in both groups of fish. Thus,
plastic change appeared to have taken place at synapses
between fibers conveying predictive signals and principal
cells. Parallel fiber synapses were hypothesized to be the
Negative images of predicted sensory
responses in the ELL of gymnotid fish and the
dorsal octaval nucleus (DON) of elasmobranch
Ventilation fish. (a) Raster display of responses of a cell in
alone the ELL of a gymnotid fish to tail bend before,
during and after pairing the tail bend with an
excitatory electrosensory stimulus to the
receptive field of the cell. Pairing with a
sensory stimulus that evokes a burstpause
yields a pauseburst response after pairing.
Ventilation (Reproduced with permission from [5].)
plus (b) Histogram display of responses of a cell in
stimulus the DON of a ray (elasmobranch) to ventilatory
rhythm of the fish before, during and after
pairing ventilation with an excitatory
electrosensory stimulus to the receptive field of
20 the cell. The response to ventilation plus
stimulus decreases during 25 minutes of
0
pairing as a result of the development of a
response to ventilation alone that is a negative
Ventilation image (a pauseburst) of the previously paired
alone sensory response. (Reproduced with
permission from [20].)
site of plastic change because all of the effective predictive
signals are conveyed by parallel fibers and some, such as
proprioceptive signals, are conveyed only by parallel fibers.
This hypothesis was also suggested by the known plasticity
of parallel fiber synapses in the cerebellum itself [2426].
Plasticity with appropriate characteristics has now been
demonstrated at parallel fiber synapses in all three elec-
trosensory structures. Pairing parallel fiber input with
postsynaptic depolarization causes a reduction in synaptic
efficacy at this synapse [20,27,28], a form of plasticity
referred to as anti-Hebbian [29]. As a result of such plasticity,
the association of parallel fiber activity with principal cell
activation by electrosensory input will reduce synaptic
strength at the paired parallel fiber inputs (relative to other
parallel fiber inputs), resulting in a hyperpolarizing response
to the predictive signals conveyed by the paired inputs.
The synaptic mechanisms responsible for robust depolar-
izing responses to predictive signals following pairing with
hyperpolarizing sensory responses (Figure 2a, right side)
are less clear. Associative synaptic plasticity following pair-
ing with purely hyperpolarizing stimuli is rare [30], and
possible mechanisms for such plasticity are obscure. In
addition, although pairing parallel fiber input with postsy-
naptic hyperpolarization yielded an increase in synaptic
efficacy in gymnotid and elasmobranch fish, such pairing
had no effect in mormyrid fish (C Bell, unpublished data).
This difference could reflect the fact that the mormyrid
results were obtained in vitro whereas results from the
other two groups of fish were obtained in vivo from the
whole animal. Postsynaptic spiking is critical for synaptic
depression in the mormyrid ELL [31], but spiking,
410
 Memory-based expectations in electrosensory systems Bell 485

although prominent in vivo, is either low or absent in vitro. Figure 4

Thus, the effects of pairing with hyperpolarizing current in
gymnotids and elasmobranchs could be due to lack of (a)
spikes during current injection and a consequent absence 100

% Change in epsp size

or reversal of synaptic depression. Modeling results in the
mormyrid are consistent with such an explanation [32].

Synaptic plasticity has also been demonstrated at a second 0

site in the gymnotid ELL, at synapses made by descend-
ing fibers from a higher order electrosensory nucleus Dendritic spike Dendritic spike
[28,33,34]. This plasticity is very similar to that found at before epsp after epsp
parallel fiber synapses in the same fish and could be part of 100
the explanation for negative image formation in response 400 0 400 ms
to whole body electrosensory stimuli. Time between epsp onset and dendritic spike
(b)
The characteristics and mechanisms of plastic change at
C
parallel fiber synapses have been examined in both the
mormyrid and gymnotid ELL. In the mormyrid, synaptic
depression was found to require the occurrence of a postsy-
naptic dendritic spike and to depend on the precise timing
of the spike relative to epsp (excitatory post-synaptic poten-
tial) onset [27,31]. Depression developed when the
postsynaptic spike occurred within 50 ms of epsp onset, C+S
whereas other timing relations yielded potentiation or no
effect. The depression required activation of NMDA recep-
tors and changes in postsynaptic calcium. The requirement
for NMDA receptor activation explains the timing depen-
dence of the learning rule in that simultaneous occurrence
of glutamate binding and postsynaptic depolarization
(caused by the dendritic spike) are needed to open the
NMDA channel. Potentiation also occurred at parallel fiber C
synapses in the mormyrid, but this potentiation appeared to
be non-associative and to depend on simple repetition of the
parallel fiber stimuli at a sufficiently high rate [31]. The
potentiation could reverse the depression and vice versa,
with both potentiation and depression appearing to have a 0 ms 70 140
Current Opinion in Neurobiology
presynaptic locus of expression, as indicated by changes in
paired pulse facilitation. Synaptic depression in the gym-
notid ELL also depended on NMDA receptors (J Bastian, Modeling the generation of negative images in the mormyrid ELL.
personal communication) and on changes in postsynaptic (a) Learning rule. The dots show results from in vitro slice experiments
calcium [35]. In the gymnotid, however, the site of expres- in which parallel fiber-evoked epsps were paired with postsynaptic
sion appeared to be postsynaptic for the depression and dendritic spikes evoked by intracellular current pulses. Changes in epsp
size were measured as a function of the relative timing of the epsp and
presynaptic for the potentiation [28]. the spike during pairing. Epsp depression was present only if the
postsynaptic spike was evoked within ~50 ms following epsp onset.
The characteristics of plasticity at parallel fiber synapses Other timing delays usually yielded potentiation. The thin continuous line
are consistent with such plasticity being responsible for the drawn through the experimental points shows the idealized form of the
learning rule that was used in the modeling (from [27,32]). (b) Modeling
formation of negative images. The associative anti-
results for a principal neuron in the mormyrid ELL. The model neuron
Hebbian characteristic provides a simple and direct included non-plastic synapses simulating electrosensory input and
mechanism for the generation of predictive responses that plastic synapses conveying corollary discharge signals at various delays
oppose depolarizing electrosensory responses. The ready following the EOD motor command, thus simulating parallel fiber input.
reversibility of depression by potentiation and vice versa Principal cells of ELL have broad dendritic spikes as well as narrow
axonal spikes. Most of the spikes shown in the raster are axonal spikes.
explains the ready reversibility of the negative images Changes in parallel fiber epsps were governed by the idealized learning
(Figure 2a), and the narrow timing window within which rule shown in (a). The illustrated simulation shows the effect of
depression occurs (Figure 4a) explains the temporal preci- command alone before (C), during (C+S), and after (C) pairing with a
sion with which negative images are formed [32]. sensory stimulus that evokes a pauseburst. The effect of command
alone after pairing is a negative image (burstpause) of the response to
the paired stimulus. Compare with Figure 2b. (Reproduced with
Modeling studies also support the conclusion that anti- permission from [32].)
Hebbian plasticity at parallel fiber synapses is responsible

411
486 Sensory systems
for the generation of negative images [32,36]. A variety of
different learning rules for parallel fiber plasticity were
tested in a model of the mormyrid ELL [32], and the neg-
ative images resulting from the use of an idealized form of
the experimentally-determined, asymmetric learning rule
(thin continuous line in Figure 4a) were significantly more
accurate and more stable than the negative images result-
ing from any other learning rule.
Conclusions
The role of parallel fiber synapses in generating negative
images of predicted sensory input is indicated by the
predictive signals that they convey, by the presence of
appropriate plasticity at these synapses, and by the model-
ing studies. Thus, the connection between synaptic
plasticity at the cellular level and a systems level adaptive
function is relatively secure for these cerebellum-like
structures, as secure perhaps as in any vertebrate system.
However, much remains to be elucidated about the gener-
ation of negative images before the full value of these
cerebellum-like structures for understanding adaptive
processing is obtained. Future work must determine the
following: first, the cellular mechanisms of plastic change;
second, the mechanism by which pairing with hyperpolar-
izing sensory stimuli results in depolarizing predictive
responses; third, the possible role of plasticity at inhibitory
synapses; and fourth, how the adaptive functions are
integrated with other known functions of the circuitry such
as lateral inhibition [37,38], gating [39,40], gain control
[41], and attention-like processes [42]. Behavioral demon-
strations of the subtraction of predictable sensory features
are also needed. Finally, work on other systems is required
to determine whether knowledge of these electrosensory
structures can contribute to an understanding of other
cerebellum-like structures, such as the dorsal cochlear
nucleus or the cerebellum itself [10,43,44].
Acknowledgements
I thank Joe Bastian, David Bodznick and Kirsty Grant for helpful
discussions. My research was supported by grants from the National
Institutes of Mental Health (MH49792 and MH60996).
References and recommended reading
Papers of particular interest, published within the annual period of review,
have been highlighted as:
of special interest
of outstanding interest
1. Bullock TH, Heiligenberg W: Electroreception. New York: Wiley; 1986.
2. Turner RW, Maler L, Burrows M (Eds): Electroreception and
Electrocommunication. J Exp Biol 1999, 202:1167-1458.
This entire issue of the Journal of Experimental Biology is devoted to
electrosensory systems. The volume is comprehensive and well-edited, like
the book edited by Bullock and Heiligenberg [1] but is focused on recent
work. It is a good place for the reader to review the many ways in which
electrosensory systems can contribute to our understanding of general
issues in sensory biology.
3. Moller P: Electric Fishes: History and Behavior. London: Chapman
and Hall; 1995.
4. Montgomery JC, Coombs S, Conley RA, Bodznick D: Hindbrain
sensory processing in lateral line, electrosensory, and auditory
systems: a comparative overview of anatomical and functional
similarities. Auditory Neuroscience 1995, 1:207-231.
5. Bastian J: Pyramidal-cell plasticity in weakly electric fish: a
mechanism for attenuating responses to reafferent
electrosensory inputs. J Comp Physiol 1995, 176:63-78.
6. Bell CC, Grant K, Serrier J: Corollary discharge effects and sensory
processing in the mormyrid electrosensory lobe: I. Field
potentials and cellular activity in associated structures.
J Neurophysiol 1992, 68:843-858.
7. Hjelmstad GO, Parks G, Bodznick D: Motor corollary discharge
activity and sensory responses related to ventilation in the skate
vestibulolateral cerebellum: Implications for electrosensory
processing. J Exp Biol 1996, 199:673-681.
8. Conley RA, Bodznick D: The cerebellar dorsal granular ridge in an
elasmobranch has proprioceptive and electroreceptive
representations and projects homotopically to the medullary
electrosensory nucleus. J Comp Physiol 1994, 174:707-721.
9. Sas E, Maler L: The organization of afferent input to the caudal
lobe of the cerebellum of the gymnotid fish Apteronotus
leptorhynchus. Anat Embryol 1987, 177:55-79.
10. Bell C, Bodznick D, Montgomery J, Bastian J: The generation and
subtraction of sensory expectations within cerebellum-like
structures. Brain Behav Evol 1997, 50:17-31.
11. Bullock TH, Bodznick DA, Northcutt RG: The phylogenetic distribution
of electroreception: evidence for convergent evolution of a primitive
vertebrate sense modality. Brain Res Rev 1983, 6:25-46.
12. Bell CC: An efference copy modified by reafferent input. Science
1981, 214:450-453.
13. Bell CC: Properties of a modifiable efference copy in electric fish.
J Neurophysiol 1982, 47:1043-1056.
14. Bell CC: Duration of plastic change in a modifiable efference
copy. Brain Res 1986, 369:29-36.
15. Bell CC, Grant K: Corollary discharge effects and sensory
processing in the mormyromast regions of the mormyrid
electrosensory lobe: II. cell types and corollary discharge
plasticity. J Neurophysiol 1992, 68:859-875.
16. Bell CC, Caputi A, Grant K: Physiology and plasticity of
morphologically identified cells in the mormyrid electrosensory
lobe. J Neurosci 1997, 17:6409-6422.
17. Bastian J: Plasticity in an electrosensory system. I. General features
of dynamic sensory filter. J Neurophysiol 1996, 76:2483-2496.
18. Bastian J: Plasticity of feedback inputs in the apteronotid
electrosensory system. J Exp Biol 1999, 202:1327-1337.
This article provides a good summary of both systems-level and cellular-level
studies of negative image formation in the ELL of gymnotid fish. The article
also shows the similarities and differences between plasticity at parallel fiber
synapses in the dorsal molecular layer of the gymnotid ELL and plasticity at
synapses made by descending fibers from a higher order electrosensory
nucleus (nucleus preeminentialis) in the ventral molecular layer.
19. Bodznick D: The specificity of an adaptive filter that suppresses
unwanted reafference in electrosensory neurons of the skate
medulla. Biol Bull 1993, 185:312-314.
20. Bodznick D, Montgomery JC, Carey M: Adaptive mechanisms in the
elasmobranch hindbrain. J Exp Biol 1999, 202:1357-1364.
This article provides a good summary of both systems-level and cellular-level
studies of negative image formation in the DON of elasmobranch fish. The
article also shows that in the elasmobranch fish, corollary discharge signals
associated with motor commands, proprioceptive signals, and descending
electrosensory signals from higher centers can all serve as predictive signals
which can elicit negative images of predicted sensory patterns after suitable
periods of association.
21. Montgomery JC, Bodznick D: An adaptive filter that cancels
self-induced noise in the electrosensory and lateral line
mechanosensory systems of fish. Neurosci Lett 1994, 174:145-148.
22. Montgomery JC: Noise cancellation in the electrosensory system
of the thornback ray; common mode rejection of input produced
by the animals own ventilatory movement. J Comp Physiol 1984,
155:103-111.
23. Bell CC, Caputi A, Grant K, Serrier J: Storage of a sensory pattern
by anti-Hebbian synaptic plasticity in an electric fish. Proc Natl
Acad Sci USA 1993, 90:4650-4654.
412

24. Ito M, Sakurai M, Tongroach P: Climbing fibre induced depression
of both mossy fibre responsiveness and glutamate sensitivity of
cerebellar Purkinje cells. J Physiol 1982, 324:113-134.
25. Crepel F, Jaillard D: Pairing of pre- and postsynaptic activities in
cerebellar Purkinje cells induces long-term changes in synaptic
efficacy in vitro. J Physiol 1991, 432:123-141.
26. Linden DJ, Dickenson MH, Smeyne M, Connor JA: A long-term
depression of AMPA currents in cultured cerebellar Purkinje
neurons. Neuron 1991, 7:81-89.
27. Bell CC, Han VZ, Sugawara S, Grant K: Synaptic plasticity in a
cerebellum-like structure depends on temporal order. Nature
1997, 387:278-281.
28. Bastian J: Plasticity in an electrosensory system III: Contrasting
properties of spatially segregated dendritic inputs. J Neurophysiol
1998, 79:1839-1857.
29. Hebb DO: The Organization of Behavior. New York: Wiley; 1949.
30. Fregnac Y, Burke JP, Smith D, Friedlander ME: Temporal covariance
of pre- and postsynaptic activity regulates functional connectivity
in the visual cortex. J Neurophysiol 1994, 71:1403-1421.
31. Han VZ, Grant G, Bell CC: Reversible associative depression and
nonassociative potentiation at a parallel fiber synapse. Neuron
2000, 27:611-622.
We show associative synaptic depression at parallel fiber synapses in the
mormyrid ELL depends on activation of NMDA receptors and the occurrence
of a postsynaptic dendritic spike. This work also provides evidence that both
depression and potentiation have a presynaptic locus of expression.
32. Roberts PD, Bell CC: Computational consequences of temporally
asymmetric learning rules: II. Sensory image cancellation.
J Comput Neurosci 2000, 9:67-83.
33. Bastian J: Plasticity in an electrosensory system. II. Postsynaptic
events associated with a dynamic sensory filter. J Neurophysiol
1996, 76:2497-2507.
Memory-based expectations in electrosensory systems Bell 487
34. Wang D, Maler L: In vitro plasticity of the direct feedback pathway
in the electrosensory system of Apteronotus leptorhynchus.
J Neurophysiol 1997, 78:1882-1889.
35. Bastian J: Modulation of calcium-dependent postsynaptic
depression contributes to an adaptive sensory filter.
J Neurophysiol 1998, 80:3352-3355.
36. Nelson ME, Paulin MG: Neural simulations of adaptive reafference
suppression in the elasmobranch electrosensory system. J Comp
Physiol 1995, 177:723-736.
37. Bastian J, Court Right J, Crawford J: Commissural neurons of the
electrosensory lateral line lobe of Apteronotus leptorhynchus:
morphological and physiological characteristics. J Comp Physiol
1993, 173:257-274.
38. Meek J, Hafmans TM, Han VZ, Bell CC, Grant K: Myelinated
dendrites in the mormyrid electrosensory lobe. J Comp Neurol
2001, 431:255-275.
39. Bell CC: Mormyromast electroreceptor organs and their afferents
in mormyrid electric fish: II. Intra-axonal recordings show initial
stages of central processing. J Neurophysiol 1990, 63:303-318.
40. Meyer JC, Bell CC: Behavioral measurements of sensory gating by
a corollary discharge. J Comp Physiol 1983, 151:401-406.
41. Bastian J: Gain control in the electrosensory system mediated by
descending inputs to the electrosensory lateral line lobe.
J Neurosci 1986, 6:553-562.
42. Berman NJ, Maler L: Neural architecture of the electrosensory
lateral line lobe: adaptations for coincidence detection, a sensory
searchlight and frequency-dependent adaptive filtering. J Exp Biol
1999, 202:1243-1253.
43. Paulin MG: The role of the cerebellum in motor control and
perception. Brain Behav Evol 1993, 41:39-50.
44. Devor A: Is the cerebellum like cerebellum-like structures? Brain
Res Rev 2000, 34:149-156.
413


414
 Further
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
ANNUAL
REVIEWS
Click here for quick links to
Annual Reviews content online,
including:
by Cornell University on 10/30/11. For personal use only.
Other articles in this volume
Top cited articles
Top downloaded articles
Our comprehensive search
Annu. Rev. Neurosci. 2008. 31:2546
First published online as a Review in Advance on
February 14, 2008
The Annual Review of Neuroscience is online at
neuro.annualreviews.org
This articles doi:
10.1146/annurev.neuro.31.060407.125639
Copyright  c 2008 by Annual Reviews.
All rights reserved
0147-006X/08/0721-0025$20.00
Spike TimingDependent
Plasticity: A Hebbian
Learning Rule
Natalia Caporale and Yang Dan
Division of Neurobiology, Department of Molecular and Cell Biology, and Helen Wills
Neuroscience Institute, University of California, Berkeley, California 94720;
email: caporale@socrates.berkeley.edu, ydan@berkeley.edu
Key Words
long-term potentiation, long-term depression, synapse, memory,
backpropagating action potential
Abstract
Spike timingdependent plasticity (STDP) as a Hebbian synaptic learn-
ing rule has been demonstrated in various neural circuits over a wide
spectrum of species, from insects to humans. The dependence of synap-
tic modication on the order of pre- and postsynaptic spiking within
a critical window of tens of milliseconds has profound functional im-
plications. Over the past decade, signicant progress has been made in
understanding the cellular mechanisms of STDP at both excitatory and
inhibitory synapses and of the associated changes in neuronal excitability
and synaptic integration. Beyond the basic asymmetric window, recent
studies have also revealed several layers of complexity in STDP, includ-
ing its dependence on dendritic location, the nonlinear integration of
synaptic modication induced by complex spike trains, and the modu-
lation of STDP by inhibitory and neuromodulatory inputs. Finally, the
functional consequences of STDP have been examined directly in an
increasing number of neural circuits in vivo.
25
415

Contents
INTRODUCTION . . . . . . . . . . . . . . . . . .
CELLULAR MECHANISMS . . . . . . . .
tLTP Window . . . . . . . . . . . . . . . . . . . . .
tLTD Window. . . . . . . . . . . . . . . . . . . . .
STDP OF INHIBITION . . . . . . . . . . . . .
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
STDP of Excitatory Synapses
onto Inhibitory Neurons . . . . . . . .
STDP of GABAergic Synapses . . . . .
STDP WITH COMPLEX
by Cornell University on 10/30/11. For personal use only.
SPIKE PATTERNS . . . . . . . . . . . . . . .
DEPENDENCE ON DENDRITIC
LOCATION . . . . . . . . . . . . . . . . . . . . . .
MODULATION OF STDP
BY OTHER INPUTS . . . . . . . . . . . . .
PLASTICITY OF NEURONAL
EXCITABILITY AND
SYNAPTIC INTEGRATION . . . . .
STDP IN VIVO . . . . . . . . . . . . . . . . . . . . . .
Electrical Stimulation . . . . . . . . . . . . . .
Paired Sensory Stimulation . . . . . . . . .
Motion Stimuli . . . . . . . . . . . . . . . . . . . .
Sensory Deprivation . . . . . . . . . . . . . . .
FINAL REMARKS . . . . . . . . . . . . . . . . . . .
INTRODUCTION
Electrical activity plays crucial roles in the
structural and functional renement of neu-
ral circuits throughout an organisms lifetime
(Buonomano & Merzenich 1998, Gilbert 1998,
Karmarkar & Dan 2006, Katz & Shatz 1996).
Manipulations of sensory experience that dis-
rupt normal activity patterns can lead to large-
scale network remodeling and marked changes
in neural response properties. Learning and
memory are also likely to be mediated by
LTP: long-term
potentiation activity-dependent circuit modications. Un-
derstanding the cellular mechanisms underly-
HFS: high-frequency
stimulation ing such functional plasticity has been a long-
standing challenge in neuroscience (Martin
LFS: low-frequency
stimulation et al. 2000).
In his inuential postulate on the cellular ba-
LTD: long-term
depression sis for learning, Hebb stated that when an axon
of cell A is near enough to excite a cell B and
26 Caporale Dan
repeatedly or persistently takes part in ring it,
some growth process or metabolic change takes
place in one or both cells such that As efciency,
26
as one of the cells ring B, is increased (Hebb
27
1949). This postulate gained strong experimen-
28
tal support with the nding of long-term po-
28
tentiation (LTP) of synaptic transmission, ini-
29
tially discovered in the hippocampus (Bliss &
Gardner-Medwin 1973, Bliss & Lomo 1973)
30
and subsequently reported in a large num-
30
ber of neural circuits, including various neo-
cortical areas (Artola & Singer 1987, Iriki
31
et al. 1989, Hirsch et al. 1992), the amyg-
dala (Chapman et al. 1990, Clugnet &
32
LeDoux 1990), and the midbrain reward cir-
cuit (Liu et al. 2005, Pu et al. 2006). Tra-
34
ditionally, LTP is induced by high-frequency
stimulation (HFS) of the presynaptic affer-
ents or by pairing low-frequency stimulation
35
(LFS) with large postsynaptic depolarization
36
(>30 mV). In contrast, long-term depression
36
(LTD) is induced by LFS, either alone or
37
paired with a small postsynaptic depolarization
37
(Artola et al. 1990, Dudek & Bear 1993,
38
Kirkwood & Bear 1994, Linden & Connor
38
1995, Mulkey & Malenka 1992, Stanton &
Sejnowski 1989). Together, LTP and LTD al-
low activity-dependent bidirectional modi-
cation of synaptic strength, thus serving as
promising candidates for the synaptic basis of
learning and memory (Bliss & Collingridge
1993; Ito 2005; Siegelbaum & Kandel 1991).
To characterize the temporal requirements
for the induction of LTP and LTD, Levi &
Steward (1983) varied the relative timing of
a strong and a weak input from the entorhi-
nal cortex to the dental gyrus and found that
synaptic modication depended on the tempo-
ral order of the two inputs. Potentiation was
produced when the weak input preceded the
strong input by less than 20 ms, and reversing
the order led to depression. Subsequent stud-
ies further demonstrated the importance of the
temporal order of pre- and postsynaptic spik-
ing in synaptic modication and delineated the
critical window on the order of tens of mil-
liseconds (Bi & Poo 1998, Debanne et al. 1998,
Magee & Johnston 1997, Markram et al. 1997,
416
 Zhang et al. 1998) (Figure 1a, I). Such spike- a Excitatory to excitatory
timing-dependent plasticity (STDP) (Abbott &
Nelson 2000) has now been observed at excita- I II
tory synapses in a wide variety of neural cir-
cuits (Boettiger & Doupe 2001, Cassenaer & 50
Laurent 2007, Egger et al. 1999, Feldman 2000, 50
Froemke & Dan 2002, Sjostrom et al. 2001,
Tzounopoulos et al. 2004). Compared with the
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

correlational forms of synaptic plasticity, STDP

captures the importance of causality in deter-
mining the direction of synaptic modication, b Excitatory to inhibitory
by Cornell University on 10/30/11. For personal use only.

which is implied in Hebbs original postulate.

I II
Recent studies have further characterized
the mechanism and function of STDP in both
in vitro and in vivo preparations, addressing
the following questions: Which cellular mech- 60 20
anisms determine the STDP window, and how
similar are they to the mechanisms underlying
LTP and LTD induced by HFS and LFS, re-
spectively? Does the window depend on the c Inhibitory to excitatory
dendritic location of the input, and can it be
regulated by neuromodulatory inputs? Does a I
II
similar learning rule apply to the inhibitory cir-
cuits? Can we observe the consequences of the 50
asymmetric window in vivo, and can it account 200
for the synaptic modications induced by com-
plex, naturalistic spike trains? In this review we
summarize recent progress in these areas. III

CELLULAR MECHANISMS 25

For many glutamatergic synapses, the induc-

tions of LTP by HFS and LTD by LFS both
require the activation of NMDA (N-methyl-d-
aspartate) receptors and a rise in postsynaptic
Ca2+ level (Malenka & Bear 2004). The NMDA
receptor is thought to serve as the coincidence
detector: The presynaptic activation provides
glutamate and the postsynaptic depolarization
causes removal of the Mg2+ block (Mayer et al.
1984, Nowak et al. 1984), which together allow
Ca2+ inux though the NMDA receptors. The
level and time course of postsynaptic Ca2+ rise
depend on the induction protocol: HFS leads
to fast, large Ca2+ inux, whereas LFS leads to
prolonged, modest Ca2+ rise (Malenka & Bear
2004, Yang et al. 1999). In the Ca2+ hypothesis
(Artola & Singer 1993, Lisman 1989, Yang et al.
Figure 1
Diversity of temporal windows for STDP induction. a: Windows for excitatory
to excitatory connections. b: Windows for excitatory to inhibitory connections.
c: Windows for inhibitory to excitatory connections. Temporal axis is in
milliseconds.
1999), these two types of Ca2+ signals cause
the activation of separate molecular pathways.
Activation of Ca2+ /calmodulin-dependent pro-
tein kinase II (CaMKII) by large Ca2+ rise STDP: spike
timingdependent
is required for LTP, whereas recruitment
plasticity
of phosphatases such as protein phosphatase
N-methyl-d-
1 (PP1) and calcineurin by modest Ca2+ in-
aspartate (NMDA)
crease is necessary for LTD (Malenka & Bear receptor: subtype of
2004). Spike timingdependent LTP (tLTP) glutamate receptors
and LTD (tLTD) also depend on NMDA

www.annualreviews.org Spike TimingDependent Plasticity 27

417
 receptor activation and the rise in postsy-
naptic Ca2+ level (Bi & Poo 1998, Debanne
et al. 1998, Feldman 2000, Magee &
BAP: back-
propagating action Johnston 1997, Markram et al. 1997, Sjostrom
potential et al. 2001, Zhang et al. 1998). Can this simple
AP: action potential model for conventional LTP and LTD account
for STDP, in particular for its temporal window
VDCC: voltage-
dependent Ca2+ on the time scale of tens of milliseconds?
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
channel
tLTP Window
by Cornell University on 10/30/11. For personal use only.
Induction of tLTP requires activation of
the presynaptic input milliseconds before the
backpropagating action potential (BAP) in
the postsynaptic dendrite (pre post, posi-
tive intervals). The BAP can facilitate Mg2+ un-
blocking of NMDA receptors and thus allow
Ca2+ inux, leading to tLTP induction. How-
ever, the width of the tLTP window cannot be
explained solely by the time course of NMDA
receptor activation. The dissociation of gluta-
mate from the NMDA receptors occurs on the
order of hundreds of milliseconds (Lester et al.
1990), much longer than the observed tLTP
windows (Figure 1a). The short duration of the
window may be due to the kinetics of Mg2+ un-
blocking NMDA receptors (Kampa et al. 2004),
such that the BAPs arriving soon after the onset
of the excitatory postsynaptic potential (EPSP)
are better able to open the NMDA receptors.
In addition to the Mg2+ unblock of NMDA
receptors, the tLTP window could also be
shaped by other types of interactions be-
tween the EPSP and the BAP. For example,
the EPSP can cause changes in the dendritic
conductances that affect the action poten-
tial (AP) backpropagation into the dendrites.
In the hippocampus, the distal dendrites of
CA1 pyramidal neurons express a high den-
sity of A-type K+ channels, which regulate
the BAP amplitude (Hoffman et al. 1997). An
EPSP that depolarizes the dendrite and inac-
tivates these channels can boost the BAPs ar-
riving within tens of milliseconds (Magee &
Johnston 1997, Watanabe et al. 2002). This
boosting of the BAPs can in turn increase the
Ca2+ inux through voltage-dependent Ca2+
channels (VDCCs), which can modulate the
28 Caporale Dan
magnitude of tLTP (Bi & Poo 1998, Froemke
et al. 2006, Magee & Johnston 1997). In the
neocortex, a similar boosting of the BAP by the
preceding EPSP is achieved by voltage-gated
Na2+ channel activation in the distal dendrites
(Stuart & Hausser 2001). Such nonlinear in-
teractions between the EPSP and BAP at short
positive intervals could explain the supralinear
summation of Ca2+ inux to the active synapse
in both hippocampal (Magee & Johnston 1997)
and neocortical (Koester & Sakmann 1998)
(Nevian & Sakmann 2004) neurons.
tLTD Window
Models based on the Ca2+ hypothesis have also
been used to explain the tLTD window (post
pre, negative intervals) (Karmarkar & Buono-
mano 2002, Shouval et al. 2002). Assuming that
the BAP contains an afterdepolarization last-
ing for tens of milliseconds and that all rele-
vant Ca2+ enters the postsynaptic cell through
NMDA receptors, the tLTD window can be ex-
plained by the interaction between the EPSP
and the BAP. Unlike pairing of the BAP and the
EPSP at positive intervals, which causes large
Ca2+ inux through the NMDA receptors, the
EPSP coinciding with the afterdepolarization
leads to a moderate Ca2+ inux, resulting in
tLTD. It should be noted that this model pre-
dicts an additional tLTD window at positive in-
tervals outside the tLTP window (Figure 1a,
II), where the rise in postsynaptic Ca2+ falls
within the range for LTD induction. This addi-
tional tLTD window has indeed been observed
in hippocampal CA1 neurons (Nishiyama et al.
2000, Wittenberg & Wang 2006) but not at
other synapses. This suggests a distinct form
of STDP at hippocampal synapses, or it could
reect insufcient sampling of long positive in-
tervals in the experimental studies of STDP in
other circuits.
In another model for tLTD based on the
Ca2+ hypothesis (Froemke et al. 2005), a
BAP preceding an EPSP induces Ca2+ in-
ux through VDCCs, which inactivates the
NMDA receptors (Rosenmund et al. 1995,
Tong et al. 1995). The reduced Ca2+ inux
418
 through NMDA receptors in turn leads to
tLTD. This model is supported by the obser-
vations that tLTD induction requires activation
of VDCCs (Bender et al. 2006, Bi & Poo 1998,
Froemke et al. 2005, Nevian & Sakmann 2006)
and that pairing EPSPs and BAPs at negative
intervals leads to sublinear summation of Ca2+
inux (Koester & Sakmann 1998, Nevian &
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
through presynaptic CB1 endocannabinoid
receptors is also required for tLTD for several
excitatoryexcitatory (Bender et al. 2006;
mGluR:
Nevian & Sakmann 2006; Sjostrom et al. metabotropic
2003) and excitatoryinhibitory connections glutamate receptor
(Tzounopoulos et al. 2007), presumably by
inhibiting presynaptic transmitter release. In
Figure 2, we have outlined the major signaling

Sakmann 2004). Furthermore, in L2/3 pyra- pathways implicated in STDP.

midal neurons in visual cortical slices, BAP-
induced Ca2+ -dependent NMDA receptor STDP OF INHIBITION
by Cornell University on 10/30/11. For personal use only.

inactivation varied with dendritic location, mir- Balanced excitation and inhibition are crucial
roring the location dependence of the tLTD for normal brain functions (Shu et al. 2003) and
window at these synapses (Froemke et al. 2005).
In some other synapses, tLTD induction
does not depend on activation of postsynaptic
NMDA receptors (Bender et al. 2006, Egger
et al. 1999, Nevian & Sakmann 2006, Sjostrom tLTD tLTP
et al. 2003). These studies suggest a model
involving two coincidence detectors, with the CB1-R
NMDA receptor for tLTP and an additional co-
incidence detector for tLTD. In a two-detector
model proposed by Karmarkar & Buonomano
(2002), tLTD induction requires activation of
postsynaptic mGluRs (metabotropic glutamate
receptors) and Ca2+ inux through VDCCs, Glu
eCB
a premise supported by experimental ndings
NMDA-R
in the barrel cortex (Bender et al. 2006, Egger
mGlu-R
et al. 1999, Nevian & Sakmann 2006). Signal-
VDCC
ing through mGluRs can lead to phospholipase
C (PLC) activation, and Ca2+ inux through Ca 2+

VDCCs can facilitate mGluR-dependent-PLC PLC

activation (Hashimotodani et al. 2005, Maejima
et al. 2005). Thus, PLC can serve as a potential
coincidence detector for tLTD.
Downstream of coincidence detection,
PLC may generate inositol 1,4,5-triphosphate
(IP3 ), which in turn triggers release of Ca2+
from internal stores through IP3 receptors
(IP3 Rs) (Bender et al. 2006). Both PLC
activation and Ca2+ level elevation (due
to inux through VDCCs and/or NMDA
receptors, or release from internal stores)
can promote endocannabinoid synthesis and
release (Hashimotodani et al. 2007). Endo-
cannabinoids play important roles in both
short- and long-term depression of many
synapses (Chevaleyre et al. 2006). Signaling
IP3
eCB
IP3-R
ER
Figure 2
Schematic representation of signaling pathways involved in STDP induction.
In tLTP induction (right), the NMDA receptors act as coincidence detectors
for pre- and postsynaptic spiking. In tLTD induction (left) the coincidence
detector may vary across synapses. The diagram includes several pathways that
have been suggested to play a role in tLTD. Red oval indicates possible
coincidence detectors. Arrow indicates activation/potentiation. Blunt-ended
line indicates inhibition/suppression. Abbreviations: eCB, endocannabinoids;
ER, endoplasmic reticulum; Glu, glutamate; IP3 , inositol 1,4,5-triphosphate;
PLC, phospholipase C; VDCCs, voltage-dependent Ca2+ channels.

www.annualreviews.org Spike TimingDependent Plasticity 29

419

Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
by Cornell University on 10/30/11. For personal use only.
30
for regulating experience-dependent develop-
mental plasticity (Hensch 2005). Although the
strength of excitatory synapses can be modi-
ed through STDP, an important question is
whether and how correlated pre- and postsy-
naptic activity affects inhibitory circuits. Inhi-
bition in a network depends on both the excita-
tory synapses onto inhibitory neurons and the
inhibitory synapses themselves. Spike timing
dependent plasticity has been studied at both
of these synapses.
STDP of Excitatory Synapses
onto Inhibitory Neurons
In a cerebellum-like structure in the electric
sh, Bell and colleagues (1997) measured the
excitatory inputs to Purkinje-like GABAergic
neurons to study the dependence of synap-
tic modication on the temporal order of
pre- and postsynaptic spiking. Pre post
pairing within a 60-ms window induces LTD,
whereas post pre pairing leads to LTP
(Figure 1b, I). This asymmetrical window
is thus opposite in polarity to the STDP
window for the synapses between excitatory
neurons (Figure 1a, I). However, given the
difference in the postsynaptic neurons, the
functional consequences of the two learning
rules may be similar, and they could act coop-
eratively in activity-dependent network mod-
ications. Mechanistically, LTD induced by
pre post pairings required NMDA re-
ceptor activation and postsynaptic Ca2+ ele-
vation (Han et al. 2000), similarly to tLTD
for excitatory-excitatory connections. How-
ever, LTP of these synapses can be induced by
EPSPs alone without postsynaptic spiking, in-
dicating a nonassociative component of the
synaptic plasticity.
Another study of excitatory inputs to
inhibitory neurons was conducted in mouse
brain stem slices by pairing parallel ber
stimulation with cartwheel neuron spik-
ing (Tzounopoulos et al. 2004, 2007).
Pre post pairing within a narrow window
(<10 ms) induces LTD, whereas post pre
pairing causes no change in synaptic strength
Caporale Dan
(Figure 1b, II). For these synapses, LTD
depends on postsynaptic NMDA receptor
activation, Ca2+ inux, and endocannabinoid
signaling, similar to the ndings at excitatory
synapses onto pyramidal neurons (see previous
section). Interestingly, synapses from the same
presynaptic bers onto excitatory postsynaptic
neurons (fusiform principal neurons) exhibit
a STDP window similar to that of other
excitatory-excitatory connections (Figure 1a,
I) (Tzounopoulos et al. 2004, 2007). This
target specicity of the learning rule can
be attributed to the selective distribution of
presynaptic endocannabinoid CB1 receptors
in different axonal terminals.
STDP of GABAergic Synapses
Compared with the glutamatergic synapses, the
learning rules for GABAergic synapses appear
more variable. In a study of inhibitory inputs
to neocortical L2/3 pyramidal neurons, synap-
tic modication was induced by pairing single
presynaptic spikes with high-frequency post-
synaptic bursts. Overlapping pre- and post-
synaptic spiking induced LTD, and nonover-
lapping post pre spiking within hundreds
of milliseconds induced LTP (Holmgren &
Zilberter 2001) (Figure 1c, I). In the hippocam-
pus, GABAergic synapses onto CA1 pyrami-
dal neurons exhibit a symmetrical window, with
pairing of single pre- and postsynaptic spikes
at short intervals (within 20 ms) leading to
LTP, and pairing at long intervals leading to
LTD (Woodin et al. 2003) (Figure 1c, II). In
contrast, in the entorhinal cortex GABAer-
gic inputs to layer II excitatory stellate cells
exhibit an asymmetric window similar to the
STDP window for excitatory-excitatory con-
nections: LTP was found at positive inter-
vals and LTD at negative intervals (Haas
et al. 2006) (Figure 1c, III). Despite the dif-
ferences between these temporal windows for
GABAergic synapses, both the induction mech-
anism and the loci of expression have similari-
ties. In both hippocampal CA1 (Woodin et al.
2003) and the entorhinal cortex (Haas et al.
2006), the induction of synaptic modication
420
 depends on postsynaptic Ca2+ inux through
the l-type Ca2+ channels, and presynaptic ex-
pression was excluded because no change was
observed in the paired pulse ratio. In the hip-
pocampus (Woodin et al. 2003), the changes in
inhibitory postsynaptic current (IPSC) ampli-
tude are due to changes in the Cl reversal po-
tential mediated by modication of the KCC2
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
K+ -Cl cotransporter, further indicating that
the expression is postsynaptic.
by Cornell University on 10/30/11. For personal use only.
STDP WITH COMPLEX
SPIKE PATTERNS
To study synaptic plasticity, the induction
paradigms are often selected for their effective-
ness rather than for their physiological rele-
vance, thus providing limited information on
how circuits are modied by natural patterns
of activity. Although most induction protocols
for STDP consisted of repetitive pairing of
pre- and postsynaptic spikes at regular inter-
vals, neuronal activity in vivo is far from regu-
lar (Softky & Koch 1993), with periods of al-
most no activity intermingled with short bouts
of high-frequency spike bursts. During each
presynaptic burst, transmitter release is likely
to be affected by short-term plasticity (Zucker
& Regehr 2002), and in each postsynaptic burst
the efcacy of individual spike propagation may
depend on the spike pattern (Spruston et al.
1995; Williams & Stuart 2000). How well does
the STDP learning rule measured with simple
spike patterns account for the synaptic changes
induced by naturalistic spike trains? When mul-
tiple spike pairs fall within the STDP window,
how are the contributions of individual spikes
integrated?
One simple strategy to study the interaction
among multiple spikes is to add one spike
at a time to the existing pairing protocol. In
L2/3 of visual cortical slices (Froemke & Dan
2002) and in hippocampal cultures (Wang
et al. 2005), spike triplets (pre post to induce LTP. This is likely caused by the
pre or post pre post) and quadruplets small EPSPs evoked by activating a single
(pre post post pre or post pre presynaptic neuron in paired recordings be-
pre post) were used to induce synaptic
modications. In both studies, the interaction
between multiple spikes was nonlinear, but the
specic forms of nonlinearity were different. In
cortical L2/3, the nonlinear interactions could
be accounted for by a suppression model, in
which the efcacy of later spikes in each train for
synaptic modication is reduced by the preced-
ing spikes (Froemke & Dan 2002). This model
accurately predicted the synaptic changes
induced by natural spike trains recorded in vivo
in response to visual stimulation. In cultured
hippocampal neurons, the pre post pre
triplets induce no synaptic change, which sug-
gests that LTP and LTD cancel each other, but
the post pre post triplets induce LTP,
which suggests that LTP wins over LTD
under this condition. A third study using spike
triplets showed that in hippocampal slices,
different learning rules are revealed with differ-
ent numbers of spike pairings (Wittenberg &
Wang 2006). With 2030 pairings at 5 Hz, LTP
was induced regardless of the temporal order
of the spikes. With 70100 repeats, however,
LTP was observed at short positive intervals
(<30 ms), and LTD was found at both negative
intervals and at long positive intervals (>30 ms)
(Figure 1a, II). These results suggest that the
integration across multiple spike pairs depends
on the activity patterns over several minutes.
The effects of pre- and/or postsynaptic
spike bursts on synaptic modication have
also been examined. Paired recordings from
L5 pyramidal neurons in visual cortical slices
showed that the synaptic change depends on
both the spike frequency within each burst
and the interval between the pre- and post-
synaptic spikes (Sjostrom et al. 2001). At
high frequencies (50 Hz), LTP is induced
regardless of the pre/post interval, whereas
at intermediate frequencies (1040 Hz), the
pre/post interval determines the sign and mag-
nitude of synaptic modication as described
by the STDP window (Figure 1a, I). Pair-
ing at low frequencies (<1 Hz) notably fails
cause LTP can be rescued by adding extra-
cellular stimulation that provides additional
www.annualreviews.org Spike TimingDependent Plasticity 31
421

Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
by Cornell University on 10/30/11. For personal use only.
32
depolarization. The combined dependence of
synaptic modication on burst timing and fre-
quency can be accounted for by a model
in which LTP wins over LTD, and only
the interactions between neighboring spikes
contribute to synaptic modication (Sjostrom
et al. 2001). In another study in L2/3 neu-
rons in rat visual cortical slices (Froemke et al.
2006), pairing of pre- and postsynaptic bursts
at high frequencies also favored LTP regard-
less of the pre/post spike timing. However,
systematic examination of the dependence of
synaptic modication on both the number and
the timing of pre- and postsynaptic spikes led
to a modied suppression model (Froemke
et al. 2006), which incorporates short-term
depression of the presynaptic input (Zucker
& Regehr 2002) and frequency-dependent
attenuation of postsynaptic spikes (Spruston
et al. 1995). Note that in both models described
above, burst-induced synaptic modication is
accounted for by integrating the contributions
of individual spike pairs. However, in some
synapses the learning rule for bursts seems to
be completely different from that for individ-
ual spikes (Birtoli & Ulrich 2004, Kampa et al.
2006, Pike et al. 1999).
Although the above studies focused on
synaptic modications induced by short bursts
lasting for tens of milliseconds, in some cir-
cuits bursts can last for hundreds of milliseconds
to several seconds. In the developing retino-
geniculate synapse, bursts of retinal ganglion
cells lasting seconds are believed to be critical
for circuit renement (Butts & Rokhsar 2001).
Temporally overlapping pre- and postsynaptic
bursts (interval within a window of 1 s) re-
sult in synaptic potentiation, whereas nonover-
lapping bursts cause a slight depression (Butts
et al. 2007). The degree of potentiation can be
predicted by a model in which LTP depends on
the interval but not the order between the pre-
and postsynaptic bursts, and it increases linearly
with the number of spikes in the burst. This is
reminiscent of the classic correlation-based
learning rule for synaptic plasticity (Stent
1973). A strikingly similar window for burst
timing was found in the hippocampal CA3 re-
Caporale Dan
gion for correlated activation of the associa-
tional/commissural (A/C) bers and the mossy
bers (Kobayashi & Poo 2004), although no
depression was observed. In both studies, the
width of the temporal window seems to scale
with the duration of the spike bursts used in the
induction protocol, and the changes in synap-
tic strength depend on the interburst interval
rather than the precise timing of individual
spikes. Such burst timingdependent plastic-
ity rules may be functionally advantageous for
the circuits in which the information relevant
for synaptic renement is contained in the tim-
ing of the bursts rather than that of individual
spikes (Butts & Rokhsar 2001).
Together, the studies described above indi-
cate that the integration across multiple spike
pairs for the induction of synaptic modication
is highly nonlinear. The nature of the nonlin-
ear interaction is likely to depend on short-term
plasticity of the presynaptic neurons, on the
biophysical properties of the postsynaptic den-
drites, and on the downstream signaling path-
ways present in different cell types. Further
characterization of the diversity of integration
mechanisms for STDP will allow better under-
standing of circuit remodeling induced by nat-
ural patterns of neuronal activity.
DEPENDENCE ON DENDRITIC
LOCATION
In the central nervous system, each neu-
ron may receive thousands of synaptic in-
puts distributed throughout its dendritic tree.
The processing of each input depends on
the dendritic location (Hausser & Mel 2003)
owing to both the passive cable properties
(Rall 1967) and the nonuniform distribution
of active conductances (Migliore & Shepherd
2002). Such location-dependent processing and
integration of synaptic inputs are believed
to be essential aspects of neuronal computa-
tion. Since a hallmark of STDP is its depen-
dence on the BAPs, which are strongly atten-
uated along the dendrite (Stuart & Sakmann
1994, Stuart et al. 1997b, Waters et al. 2005),
synaptic modication is likely to vary with
422
 Visual cortex
L2/3 to L2/3
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
by Cornell University on 10/30/11. For personal use only.
dendritic location (Rao & Sejnowski 2001b).
Recent studies have examined the location de-
pendence of both tLTP and tLTD. In L2/3
of rat visual cortex, the magnitude of tLTP
induced by pre post pairing of single
spikes was smaller at intermediate-distal (100
150 m) than at proximal (<50 m) segments
of the apical dendrite (Froemke et al. 2005)
(Figure 3, left column). This reduction of
tLTP amplitude is likely due to distance-
dependent attenuation of the BAP. In experi-
ments with paired recordings from a L5 and a
L2/3 pyramidal neuron or from two L5 neu-
rons (Sjostrom & Hausser 2006), burst pairing
at positive intervals led to LTP at the proxi-
mal synapses but LTD at the distal synapses
(Figure 3, middle column). Similar location
dependence was also found among L2/3 to
L5 connections by pairing a single EPSP
with a postsynaptic burst at positive intervals
(Letzkus et al. 2006) (Figure 3, right column).
BAP boosting by subthreshold local dendritic
depolarization or extracellular stimulation re-
covered tLTP at distal synapses (Letzkus et al.
2006, Sjostrom & Hausser 2006), which sug-
Visual cortex Barrel cortex
L2/3 to L5 L2/3 to L5
L5 to L5
Figure 3
Dependence of STDP
on dendritic location.
gests that distal tLTP requires cooperativity
among inputs.
Two distinct effects have been reported for
post pre pairing. In L2/3 pyramidal neu-
rons, the width of the tLTD window mea-
sured with single spike pairing is broader for
intermediate-distal than for proximal inputs
(Froemke et al. 2005) (Figure 3, left col-
umn). This difference in width is correlated
with the window for AP-induced suppression
of NMDA receptor activation, which sug-
gests that the suppression plays an important
role in setting the tLTD window. In L2/3
L5 synapses in rat barrel cortex, pairing sin-
gle presynaptic spikes with postsynaptic bursts
at negative intervals leads to LTD at proxi-
mal locations but LTP of distal inputs (Letzkus
et al. 2006) (Figure 3, right column). This
distal LTP could be explained by the induc-
tion of dendritic Ca2+ spikes by the later
BAPs in the burst (Larkum et al. 1999a, Stuart
et al. 1997a), such that the EPSP coincides
with the peak postsynaptic depolarization. Lo-
cal dendritic spikes can also play a prominent
role in coincidence detection in the neocortex
www.annualreviews.org Spike TimingDependent Plasticity 33
423

Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
by Cornell University on 10/30/11. For personal use only.
34
(Larkum et al. 1999b) and in LTP induction in
hippocampal CA1 (Golding et al. 2002) and the
amygdala (Humeau & Luthi 2007).
Comparison across these studies suggests
that the degree of spatial variation of the learn-
ing rule depends on the dendritic morphology,
with quantitative changes over short distances
(e.g., dendrite of L2/3 neurons) and qualita-
tive differences along long dendrites (e.g., apical
dendrites of L5 pyramids). Although the den-
dritic variations of STDP summarized above
can be explained largely by differences in the
local active conductances, the backpropaga-
tion of APs, or the local generation of Ca2+
spikes, differential distribution of other pre-
and postsynaptic molecular machineries could
also contribute to the observed heterogene-
ity. Functionally, the spatial variation of the
STDP rule may lead to differential input se-
lection at distal and proximal dendrites. For
example, the relative paucity of LTP at dis-
tal dendrites after pre post pairing predicts
that proximal inputs should be stronger than
distal inputs (Sjostrom & Hausser 2006). The
involvement of locally generated Ca2+ spikes
in LTP induction (Golding et al. 2002, Kampa
et al. 2006) likely rewards cooperativity among
distal inputs because their synchronous ac-
tivation is known to evoke dendritic spikes.
Furthermore, the broader LTD window for
intermediate distal inputs to L2/3 neurons sug-
gests that the distal dendrites strongly favor
transient over prolonged inputs (Froemke et al.
2005).
MODULATION OF STDP
BY OTHER INPUTS
In addition to the spiking of the pre- and postsy-
naptic neurons, STDP is also regulated by other
inputs. In particular, neuromodulators and in-
hibitory activity in the network can affect both
the magnitude and the temporal window of
STDP.
Neuromodulators such as norepinephrine
and acetylcholine (ACh) play important roles
in experience-dependent neural plasticity (Bear
& Singer 1986, Kilgard & Merzenich 1998).
Caporale Dan
At the cellular level, neuromodulators can
inuence AP backpropagation by modulat-
ing the activation and inactivation of var-
ious active conductances ( Johnston et al.
1999). For example, agonists to muscarinic
ACh receptors can reduce spike attenua-
tion during high-frequency bursts, probably
through reduction of Na+ channel inactivation
( Johnston et al. 1999, Tsubokawa & Ross 1997).
Both -adrenergic and muscarinic ACh recep-
tor agonists can boost AP backpropagation by
downregulating transient K+ channels through
protein kinase A (PKA) and protein kinase
C (PKC) activation, respectively (Hoffman &
Johnston 1998, 1999). Dopamine also has a sim-
ilar effect on the BAP (Hoffman & Johnston
1999).
Such modulations of the BAPs are likely to
have profound effects on STDP, particularly at
distal dendritic locations. In the Schaffer col-
lateral pathway to hippocampal CA1, pairing a
weak and a strong input (which evokes post-
synaptic spiking) at positive intervals can in-
duce NMDA receptordependent tLTP within
a narrow window of 310 ms. Bath applica-
tion of isoproterenol, a -adrenergic recep-
tor agonist, broadens the window to 15 ms
without changing the magnitude of tLTP (Lin
et al. 2003), an effect that depends on PKA
and mitogen-activated protein kinase (MAPK)
signaling. In the amygdala, dopamine can gate
the induction of tLTP by suppressing feed-
forward inhibitory inputs to the postsynap-
tic cell (Bissiere et al. 2003). In L5 pyrami-
dal neurons of the prefrontal cortex, nicotine
application converted tLTP to tLTD by
reducing dendritic Ca2+ signals during spike
pairing (Couey et al. 2007), and this reduction
is mediated by an enhancement of GABAer-
gic synaptic transmission. In L2/3 pyramidal
neurons, activation of M1 muscarinic recep-
tors promotes tLTD induction through a PLC-
dependent pathway, whereas -adrenergic re-
ceptor activation promotes tLTP through the
adenylate cyclase cascade (Seol et al. 2007).
Thus, neuromodulators can regulate both
the magnitude and the polarity of synaptic
modications.
424
 The timing and location of inhibitory in-
puts can also affect STDP. Somatic inhibition
can prevent AP propagation through hyper-
polarization and shunting (Miles et al. 1996,
Tsubokawa & Ross 1996), which may pre-
clude STDP induction. In contrast, inhibitory
inputs to the dendrites have a variety of ef-
fects, from reducing dendritic depolarization
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
through shunting to facilitating depolarization
and even spike generation (Gulledge & Stu-
art 2003). An additional layer of complexity
by Cornell University on 10/30/11. For personal use only.
is added by the fact that the strength and
distribution of inhibition are developmentally
regulated (Hensch 2005), predicting that the
learning rule can vary considerably across de-
velopmental stages. In hippocampal CA1 pyra-
midal neurons, pairing single pre- and postsy-
naptic spikes at positive intervals leads to tLTP
in juvenile (p9p14) but not in young (p22
p28) rats (Meredith et al. 2003). However, in
young rats tLTP can be rescued by replacing
the single postsynaptic spike with a burst or by
adding GABAA antagonists, suggesting that the
change in tLTP threshold might be due to a de-
velopmental enhancement of inhibition in this
circuit.
PLASTICITY OF NEURONAL
EXCITABILITY AND
SYNAPTIC INTEGRATION
Information processing by neuronal net-
works depends not only on the connectivity
between neurons, but also on the intrinsic
conductances in each neuron that deter-
mine its excitability and synaptic integration.
Changes in neuronal excitability have been re-
ported in a variety of invertebrate and verte-
brate neural circuits during associative learn-
ing (Daoudal & Debanne 2003, Zhang &
Linden 2003). At the cellular level, LTP in-
duction by tetanic stimulation also leads to
increases in intrinsic excitability in both the
hippocampus and the cerebellum (Aizenman
& Linden 2000, Armano et al. 2000, Bliss &
Gardner-Medwin 1973, Bliss & Lomo 1973).
These activity-dependent changes in intrinsic
neuronal properties may interact synergistically
with synaptic plasticity to mediate learning and
memory.
Changes in neuronal excitability have also
been examined in the context of STDP. In hip-
pocampal cell cultures (Ganguly et al. 2000)
and neocortical slices (Li et al. 2004), repeated
pre post pairing of single spikes leads to
LTP and to an enhancement of excitability and
spike time reliability of the presynaptic neu-
rons. Pairings at negative intervals result in
LTD and a reduction in presynaptic excitability
(Li et al. 2004). Mechanistically, these presy-
naptic changes require NMDA receptor acti-
vation and Ca2+ inux to the postsynaptic neu-
ron, suggesting the involvement of retrograde
signaling. On the presynaptic side, PKC is nec-
essary for the increase in excitability (Ganguly
et al. 2000), and both PKC and PKA are re-
quired for the decrease (Li et al. 2004). Inter-
estingly, the changes in excitability can be dis-
sociated from the changes in synaptic strength
because presynaptic blockage of PKC and/or
PKA abolished the excitability changes with lit-
tle effect on the synaptic modications.
Activity-dependent changes in intrinsic
membrane properties can also affect synaptic
integration (Magee & Johnston 2005). A
recent study examined the changes in spatial
summation between two input pathways in
hippocampal CA1 neurons following STDP in-
duction (Wang et al. 2003). Induction of tLTP
in one pathway resulted in an increase in the
linearity of spatial summation of the two path-
ways, whereas induction of tLTD produced the
opposite effect. The observed changes depend
on NMDA receptor activation and may be me-
diated by modications of the Ih channels. In
another study in hippocampal CA1, LTP induc-
tion by paired theta bursts causes an increase in
the linearity of temporal summation between
the potentiated input and a neighboring input
(Xu et al. 2006); the temporal specicity of this
effect varied with dendritic location. For distal
inputs, the increase in linearity is limited to
EPSPs arriving within 5 ms of each other,
favoring summation of coincident inputs.
In contrast, for proximal inputs the increase
can be observed for EPSPs arriving within
www.annualreviews.org Spike TimingDependent Plasticity 35
425

Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
by Cornell University on 10/30/11. For personal use only.
36
a broader window of 20 ms. Such location-
dependent modulation of synaptic integration
may interact with the location dependence of
the STDP learning rule (see above) to further
enrich dendritic processing.
STDP IN VIVO
Whereas most of the early experiments on
STDP were conducted in slices and cell cul-
tures, an increasing number of studies have
begun to address the functional consequences
of STDP in intact nervous systems. Neu-
ral circuits in vivo exhibit both spontaneous
activity and sensory-evoked responses, mod-
ulated by the behavioral states of the ani-
mal. Backpropagation of the APs may also be
more variable in vivo, as the neurons receive
barrages of excitatory and inhibitory inputs
(Destexhe et al. 2003). These factors could sig-
nicantly complicate the rules for synaptic plas-
ticity. How well does the STDP learning rule
described in vitro apply to activity-dependent
synaptic modication in vivo?
Electrical Stimulation
The rst demonstration of STDP in vivo
came from a study at the retinotectal pro-
jection in the developing Xenopus (Zhang
et al. 1998). Repetitive electrical stimulation
of the retinal ganglion cells within 20 ms be-
fore tectal neuron spiking leads to LTP, whereas
pairings at negative intervals lead to LTD. Both
LTP and LTD are NMDA receptor depen-
dent, and the temporal window is similar to the
STDP windows measured in vitro (e.g., Bi &
Poo 1998, Froemke & Dan 2002, Tzounopou-
los et al. 2004). In addition to the strength of
the retinotectal connection, the amplitude of
the tectal visual response can also be modied
by pairing visual stimulation with postsynaptic
spiking (Mu & Poo 2006, Vislay-Meltzer et al.
2006).
Plasticity with similar asymmetric windows
has also been demonstrated in the mammalian
visual cortex. Optical imaging in the kitten vi-
sual cortex showed that pairing visual stimula-
Caporale Dan
tion at a given orientation with cortical electri-
cal stimulation leads to changes in the orienta-
tion map (Schuett et al. 2001). Electrical acti-
vation after the arrival of the visual input causes
expansion of the cortical representation of the
paired orientation, whereas the reverse order
causes a reduction. Whole-cell recordings in
juvenile rat visual cortex showed that pairing
visual stimulation with single neuron spiking
leads to potentiation or depression of the vi-
sual response, depending on the order between
the visual inputs and the postsynaptic spiking
(Meliza & Dan 2006).
STDP has also been described in other sen-
sory modalities in vivo. In the somatosensory
cortex of anesthetized rats, pairing subthresh-
old whisker deections with postsynaptic spik-
ing at negative intervals leads to LTD of the
paired whisker ( Jacob et al. 2007). In an olfac-
tory circuit of the locusts (-lobe in the mush-
room body), pairing odor-induced synaptic ac-
tivity with postsynaptic spiking results in robust
synaptic modications, with a temporal win-
dow similar to those for vertebrate excitatory
synapses (Figure 1a, I) (Cassenaer & Laurent
2007).
In the motor system, STDP has been
demonstrated in human subjects. Pairing elec-
trical stimulation of a somatosensory afferent
nerve with transcranial magnetic stimulation
(TMS) of the motor cortex leads to long-lasting
changes in the motor-evoked potentials (MEPs)
elicited by TMS (Wolters et al. 2003). The di-
rection and magnitude of the change depend on
the relative timing between the afferent stim-
ulation and the TMS within a window of tens
of milliseconds, comparable to the STDP win-
dows measured in vitro. The potentiation in-
duced by pairing at positive intervals can be
blocked by NMDA receptor antagonists (Stefan
et al. 2002), and the depression at negative in-
tervals is blocked by both NMDA receptor and
VDCC antagonists (Wolters et al. 2003), con-
sistent with the pharmacological properties of
STDP found in several studies (Bi & Poo 2001).
Wolters et al. (2005) also used a similar exper-
imental protocol to demonstrate STDP in hu-
man somatosensory cortex.
426
 Paired Sensory Stimulation rule. In the Xenopus tadpole, repeated presen-
tation of a moving bar in a given direction se-
Although electrical stimulation affords excel-
lectively potentiated the response to the con-
lent control of spike timing in the study of
ditioned direction, resulting in the emergence
STDP, an important question is whether the
of direction sensitivity in the tectal neurons
temporal requirements of this learning rule can
(Engert et al. 2002). Induction of direction se-
be satised under natural conditions, as spik-
lectivity through STDP has indeed been pre-
ing responses to sensory stimuli are known
dicted in a theoretical study (Rao & Sejnowski
to be highly variable (Shadlen & Newsome
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

2001a). A follow-up experiment using both se-

1994). Several studies on the functional role
quentially ashed bars and moving bars pro-
of STDP in vivo have been performed with
vided further support for the role of STDP in
pure sensory stimulation. In anesthetized adult
by Cornell University on 10/30/11. For personal use only.

the induction of direction selectivity (Mu & Poo

cats, repetitive presentation of gratings at a
2006). The selective enhancement at the con-
pair of orientations induced shifts in orientation
ditioned direction manifests as a potentiation
tuning of individual V1 neurons; the direc-
of the early phase and a reduction of the late
tion of the shift depended on the temporal or-
phase of the visual response, consistent with
der of the two orientations (Yao et al. 2004,
the prediction from STDP. Blocking the cellu-
Yao & Dan 2001). In a parallel set of exper-
lar signaling pathways underlying STDP abol-
iments in the space domain, repeated visual
ished the effect of unidirectional motion stimuli
stimulation in two adjacent retinal regions in-
in inducing direction selectivity.
duced shifts in V1 receptive elds (Fu et al.
In the visual cortex, the interaction between
2002), with a similar dependence on the stimu-
motion stimuli and STDP has been used to pre-
lus order. In both the orientation and space do-
dict two receptive eld properties and to explain
main, signicant changes in cortical response
two motion-position illusions. Model simula-
properties were observed at intervals within
tions predicted that the prevalence of motion
40 ms, similar to the STDP windows observed
stimuli in various directions during visual cor-
in vitro. For the shift in orientation tuning, the
tical development would lead to a spatial asym-
effect showed complete interocular transfer, in-
metry in the direction-selective inputs to each
dicating that the underlying neuronal modi-
cortical neuron (e.g., inputs preferring right-
cations occur largely in the cortex, after the
ward motion are biased toward the left side
inputs from the two eyes converge (Yao et al.
of the receptive eld) (Fu et al. 2004). This
2004). Psychophysical experiments in human
asymmetry in the mature cortex in turn pre-
subjects using analogous induction protocols
dicts that (a) receptive eld position depends on
showed perceptual changes consistent with the
the local motion signals within the test stimuli,
electrophysiological effects (Fu et al. 2002, Yao
and (b) motion adaptation causes the receptive
et al. 2004, Yao & Dan 2001), which suggests
eld position to shift. Both effects were con-
that the neuronal changes have direct conse-
rmed experimentally in anesthetized cat V1.
quences in visual perception.
Psychophysical measurement using matching
stimulus parameters showed that these physio-
Motion Stimuli logical effects could each explain a known visual
Compared with the repetitively ashed stimuli illusion involving the interaction between mo-
used in the above studies, moving stimuli are tion and perceived object position (De Valois
much more common in nature. Motion stim- & De Valois 1991, Nishida & Johnston 1999,
uli are intrinsically sequential (e.g., an object Ramachandran & Anstis 1990, Snowden 1998,
moving across the visual eld should sequen- Whitaker et al. 1999).
tially enter the neuronal receptive elds dis- In addition to the motion signals in sensory
tributed along its trajectory) and are thus ideally inputs, locomotion of the animal may also in-
suited for interacting with the STDP learning duce circuit modication through STDP. The

www.annualreviews.org Spike TimingDependent Plasticity 37

427

Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
by Cornell University on 10/30/11. For personal use only.
38
place elds of hippocampal neurons are known
to be dynamically modied as the animal navi-
gates in a novel environment. During repeated
running of a linear track, the place elds of
both CA1 and CA3 cells are initially symmet-
rical, but they experience a gradual asymmet-
ric expansion against the direction of loco-
motion (Lee et al. 2004; Mehta et al. 1997,
2000). Simulation with a simple feedforward
network model showed that this effect can be
explained by STDP (Blum & Abbott 1996,
Mehta et al. 2000). In the orientation domain,
Yu et al. (2006) recently reported a similar
shift in head-direction tuning curves in thala-
mic head-direction cells as the animal runs in a
circular track.
Sensory Deprivation
STDP may also play a role in other forms
of experience-dependent plasticity, even if the
sensory inputs do not explicitly involve tim-
ing on the order of tens of milliseconds. In
an experiment measuring the neural activity
during sensory deprivation, rats were chroni-
cally implanted with electrode arrays to mon-
itor the spiking activity in L4 and L2/3 of
the barrel cortex during free-moving behav-
iors (Celikel et al. 2004). Stimulus deprivation
induced by trimming a single whisker, a ma-
nipulation known to induce whisker map re-
organization, caused an immediate reversal of
the ring order and decreased correlation be-
tween L4 and L2/3 neurons. Both of these
changes are known to drive tLTD in barrel cor-
tical slices (Feldman 2000), thus providing a
plausible explanation for deprivation-induced
LTD of L4 to L2/3 connections (Allen et al.
2003). In addition to the somatosensory system,
sensory deprivation induces circuit reorganiza-
tion in the visual and auditory systems (Buono-
mano & Merzenich 1998, Gilbert 1998). It
would be interesting to test whether depriva-
tion in these modalities (e.g., monocular depri-
vation of visual input) also induces changes in
the relative spike timing among neurons that
could cause the observed circuit modications
through STDP.
Caporale Dan
FINAL REMARKS
Over the past decade, the STDP learning rule
has been demonstrated in a range of species
from insects to humans, and our understanding
of its cellular mechanisms and functional impli-
cations has progressed signicantly. However,
many questions remain unresolved.
Regarding the mechanism, it remains un-
clear whether a single model can explain STDP
at different synapses or whether different neu-
rons employ distinct molecular machineries
to achieve similar outcomes. Studies are only
beginning to examine whether and how STDP
depends on several signaling events that have
been strongly implicated in conventional LTP
and LTD, including secretion of brain-derived
neurotrophic factor (BDNF) and nitric oxide
(Mu & Poo 2006), activation of CaMKII
(Tzounopoulos et al. 2007) and phosphatases
(Froemke et al. 2005), and modication
and insertion/removal of AMPA receptors.
It would also be interesting to investigate
whether the type of NMDA receptor subunits
(NR2A/NR2B) and their synaptic location
play a role in STDP (Sjostrom et al. 2003), as
has been suggested for LTP/LTD induced by
HFS/LFS (Cull-Candy & Leszkiewicz 2004,
Liu et al. 2004). In addition, whereas several
molecules have been proposed as coincidence
detectors at excitatory synapses (Figure 2),
there is so far no candidate for inhibitory
synapses. Furthermore, although postsynaptic
Ca2+ signals are required for STDP in most
cell types, recent imaging experiments showed
that volume-averaged Ca2+ transients in the
dendritic spines are poorly correlated with the
direction of synaptic modication (Nevian &
Sakmann 2006). Perhaps new techniques that
allow measurement of Ca2+ signals at a more
microscopic scale (e.g., microdomains) will
shed new light on the cellular mechanisms of
STDP.
To understand the functional consequences
of STDP, an important factor to consider is the
high level of ongoing activity in vivo. Spon-
taneous activity can signicantly affect mem-
brane potential, conductance, and intracellular
428
 Ca2+ levels, and in some cases it can boost AP
backpropagation in vivo (Waters & Helmchen
2004). These effects will likely modulate the
rules for synaptic modication. Furthermore,
spontaneous postsynaptic spiking reduces the
persistence of synaptic potentiation and depres-
sion (Zhou et al. 2003). An important question
is how experience-dependent synaptic modi-
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org
tivity in early sensory circuits (Galan et al.
2006, Yao et al. 2007) or be replayed in the
hippocampus during sleep ( Ji & Wilson 2007,
Louie & Wilson 2001, Nadasdy et al. 1999,
Ribeiro et al. 2004, Wilson & McNaughton
1994). These reactivated patterns may serve to
consolidate the transient effects of sensory stim-
ulation into long-lasting circuit modications.

cations can persist in vivo in the face of the Characterization of neuronal plasticity at the
ongoing network activity. Recent studies have network level during natural behaviors is a cru-
suggested that sensory-evoked activity patterns cial step in understanding the neural basis for
by Cornell University on 10/30/11. For personal use only.

can reverberate in subsequent spontaneous ac- learning and memory.

DISCLOSURE STATEMENT
The authors are not aware of any biases that might be perceived as affecting the objectivity of this
review.

LITERATURE CITED
Abbott LF, Nelson SB. 2000. Synaptic plasticity: taming the beast. Nat. Neurosci. 3(Suppl):117883
Aizenman CD, Linden DJ. 2000. Rapid, synaptically driven increases in the intrinsic excitability
of cerebellar deep nuclear neurons. Nat. Neurosci. 3:10911
Allen CB, Celikel T, Feldman DE. 2003. Long-term depression induced by sensory deprivation
during cortical map plasticity in vivo. Nat. Neurosci. 6:29199
Armano S, Rossi P, Taglietti V, DAngelo E. 2000. Long-term potentiation of intrinsic excitability
at the mossy ber-granule cell synapse of rat cerebellum. J. Neurosci. 20:520816
Artola A, Brocher S, Singer W. 1990. Different voltage-dependent thresholds for inducing long-
term depression and long-term potentiation in slices of rat visual cortex. Nature 347:6972
Artola A, Singer W. 1987. Long-term potentiation and NMDA receptors in rat visual cortex.
Nature 330:64952
Artola A, Singer W. 1993. Long-term depression of excitatory synaptic transmission and its rela-
tionship to long-term potentiation. Trends Neurosci. 16:48087
Bear MF, Singer W. 1986. Modulation of visual cortical plasticity by acetylcholine and nora-
drenaline. Nature 320:17276
Bell CC, Han VZ, Sugawara Y, Grant K. 1997. Synaptic plasticity in a cerebellum-like structure
depends on temporal order. Nature 387:27881
Bender VA, Bender KJ, Brasier DJ, Feldman DE. 2006. Two coincidence detectors for spike
timing-dependent plasticity in somatosensory cortex. J. Neurosci. 26:416677
Bi G, Poo M. 2001. Synaptic modication by correlated activity: Hebbs postulate revisited. Annu.
Rev. Neurosci. 24:13966
Bi GQ, Poo MM. 1998. Synaptic modications in cultured hippocampal neurons: dependence on
spike timing, synaptic strength, and postsynaptic cell type. J. Neurosci. 18:1046472
Birtoli B, Ulrich D. 2004. Firing mode-dependent synaptic plasticity in rat neocortical pyramidal
neurons. J. Neurosci. 24:493540
Bissiere S, Humeau Y, Luthi A. 2003. Dopamine gates LTP induction in lateral amygdala by
suppressing feedforward inhibition. Nat. Neurosci. 6:58792

www.annualreviews.org Spike TimingDependent Plasticity 39

429
 Bliss TV, Collingridge GL. 1993. A synaptic model of memory: long-term potentiation in the
hippocampus. Nature 361:3139
Bliss TV, Gardner-Medwin AR. 1973. Long-lasting potentiation of synaptic transmission in the
dentate area of the unanaesthetized rabbit following stimulation of the perforant path. J.
Physiol. 232:35774
Bliss TV, Lomo T. 1973. Long-lasting potentiation of synaptic transmission in the dentate area
of the anaesthetized rabbit following stimulation of the perforant path. J. Physiol. 232:33156
Blum KI, Abbott LF. 1996. A model of spatial map formation in the hippocampus of the rat.
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

Neural. Comput. 8:8593

Boettiger CA, Doupe AJ. 2001. Developmentally restricted synaptic plasticity in a songbird nucleus
required for song learning. Neuron 31:80918
by Cornell University on 10/30/11. For personal use only.

Buonomano DV, Merzenich MM. 1998. Cortical plasticity: from synapses to maps. Annu. Rev.
Neurosci. 21:14986
Butts DA, Kanold PO, Shatz CJ. 2007. A burst-based Hebbian learning rule at retinogeniculate
synapses links retinal waves to activity-dependent renement. PLoS Biol. 5:e61
Butts DA, Rokhsar DS. 2001. The information content of spontaneous retinal waves.
J. Neurosci. 21:96173
Cassenaer S, Laurent G. 2007. Hebbian STDP in mushroom bodies facilitates the synchronous
ow of olfactory information in locusts. Nature 448:70913
Celikel T, Szostak VA, Feldman DE. 2004. Modulation of spike timing by sensory deprivation
during induction of cortical map plasticity. Nat. Neurosci. 7:53441
Chapman PF, Kairiss EW, Keenan CL, Brown TH. 1990. Long-term synaptic potentiation in
the amygdala. Synapse 6:27178
Chevaleyre V, Takahashi KA, Castillo PE. 2006. Endocannabinoid-mediated synaptic plasticity
in the CNS. Annu. Rev. Neurosci. 29:3776
Clugnet MC, LeDoux JE. 1990. Synaptic plasticity in fear conditioning circuits: induction of
LTP in the lateral nucleus of the amygdala by stimulation of the medial geniculate body. J.
Neurosci. 10:281824
Couey JJ, Meredith RM, Spijker S, Poorthuis RB, Smit AB, et al. 2007. Distributed network
actions by nicotine increase the threshold for spike-timing-dependent plasticity in prefrontal
cortex. Neuron. 54:7387
Cull-Candy SG, Leszkiewicz DN. 2004. Role of distinct NMDA receptor subtypes at central
synapses. Sci. STKE 2004:re16
Daoudal G, Debanne D. 2003. Long-term plasticity of intrinsic excitability: learning rules and
mechanisms. Learn. Mem. 10:45665
Debanne D, Gahwiler BH, Thompson SM. 1998. Long-term synaptic plasticity between
pairs of individual CA3 pyramidal cells in rat hippocampal slice cultures. J. Physiol.
507(Pt. 1):23747
Destexhe A, Rudolph M, Pare D. 2003. The high-conductance state of neocortical neurons in
vivo. Nat. Rev. Neurosci. 4:73951
De Valois RL, De Valois KK. 1991. Vernier acuity with stationary moving Gabors. Vision Res.
31:161926
Dudek SM, Bear MF. 1993. Bidirectional long-term modication of synaptic effectiveness in the
adult and immature hippocampus. J. Neurosci. 13:291018
Egger V, Feldmeyer D, Sakmann B. 1999. Coincidence detection and changes of synaptic efcacy
in spiny stellate neurons in rat barrel cortex. Nat. Neurosci. 2:1098105
Engert F, Tao HW, Zhang LI, Poo MM. 2002. Moving visual stimuli rapidly induce direction
sensitivity of developing tectal neurons. Nature 419:47075

40 Caporale Dan

430
 Feldman DE. 2000. Timing-based LTP and LTD at vertical inputs to layer II/III pyramidal cells
in rat barrel cortex. Neuron 27:4556
Froemke RC, Dan Y. 2002. Spike-timing-dependent synaptic modication induced by natural
spike trains. Nature 416:43338
Froemke RC, Poo MM, Dan Y. 2005. Spike-timing-dependent synaptic plasticity depends on
dendritic location. Nature 434:22125
Froemke RC, Tsay IA, Raad M, Long JD, Dan Y. 2006. Contribution of individual spikes in
burst-induced long-term synaptic modication. J. Neurophysiol. 95:162029
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

Fu YX, Djupsund K, Gao H, Hayden B, Shen K, Dan Y. 2002. Temporal specicity in the cortical
plasticity of visual space representation. Science 296:19992003
Fu YX, Shen Y, Gao H, Dan Y. 2004. Asymmetry in visual cortical circuits underlying motion-
by Cornell University on 10/30/11. For personal use only.

induced perceptual mislocalization. J. Neurosci. 24:216571

Galan RF, Weidert M, Menzel R, Herz AV, Galizia CG. 2006. Sensory memory for odors is
encoded in spontaneous correlated activity between olfactory glomeruli. Neural. Comput.
18:1025
Ganguly K, Kiss L, Poo M. 2000. Enhancement of presynaptic neuronal excitability by correlated
presynaptic and postsynaptic spiking. Nat. Neurosci. 3:101826
Gilbert CD. 1998. Adult cortical dynamics. Physiol. Rev. 78:46785
Golding NL, Staff NP, Spruston N. 2002. Dendritic spikes as a mechanism for cooperative long-
term potentiation. Nature 418:32631
Gulledge AT, Stuart GJ. 2003. Excitatory actions of GABA in the cortex. Neuron 37:299309
Haas JS, Nowotny T, Abarbanel HD. 2006. Spike-timing-dependent plasticity of inhibitory
synapses in the entorhinal cortex. J. Neurophysiol. 96:330513
Han VZ, Grant K, Bell CC. 2000. Reversible associative depression and nonassociative potentia-
tion at a parallel ber synapse. Neuron 27:61122
Hashimotodani Y, Ohno-Shosaku T, Tsubokawa H, Ogata H, Emoto K, et al. 2005. Phospholipase
Cbeta serves as a coincidence detector through its Ca2+ dependency for triggering retrograde
endocannabinoid signal. Neuron 45:25768
Hashimotodani Y, Ohno-Shosaku T, Watanabe M, Kano M. 2007. Roles of phospholipase
C{beta} and NMDA receptor in activity-dependent endocannabinoid release. J. Physiol.
584:37380
Hausser M, Mel B. 2003. Dendrites: bug or feature? Curr. Opin. Neurobiol. 13:37283
Hebb DO. 1949. The Organization of Behavior; A Neuropsychological Theory. New York: Wiley. xix,
335 pp.
Hensch TK. 2005. Critical period plasticity in local cortical circuits. Nat. Rev. Neurosci. 6:87788
Hirsch JC, Barrionuevo G, Crepel F. 1992. Homo- and heterosynaptic changes in efcacy are
expressed in prefrontal neurons: an in vitro study in the rat. Synapse 12:8285
Hoffman DA, Johnston D. 1998. Downregulation of transient K+ channels in dendrites of hip-
pocampal CA1 pyramidal neurons by activation of PKA and PKC. J. Neurosci. 18:352128
Hoffman DA, Johnston D. 1999. Neuromodulation of dendritic action potentials. J. Neurophysiol.
81:40811
Hoffman DA, Magee JC, Colbert CM, Johnston D. 1997. K+ channel regulation of signal prop-
agation in dendrites of hippocampal pyramidal neurons. Nature 387:86975
Holmgren CD, Zilberter Y. 2001. Coincident spiking activity induces long-term changes in in-
hibition of neocortical pyramidal cells. J. Neurosci. 21:827077
Humeau Y, Luthi A. 2007. Dendritic calcium spikes induce bi-directional synaptic plasticity in
the lateral amygdala. Neuropharmacology 52:23443
Iriki A, Pavlides C, Keller A, Asanuma H. 1989. Long-term potentiation in the motor cortex.
Science 245:138587

www.annualreviews.org Spike TimingDependent Plasticity 41

431
 Ito M. 2005. Bases and implications of learning in the cerebellumadaptive control and internal
model mechanism. Prog. Brain Res. 148:95109
Jacob V, Brasier DJ, Erchova I, Feldman D, Shulz DE. 2007. Spike timing-dependent synaptic
depression in the in vivo barrel cortex of the rat. J. Neurosci. 27:127184
Ji D, Wilson MA. 2007. Coordinated memory replay in the visual cortex and hippocampus during
sleep. Nat. Neurosci. 10:1007
Johnston D, Hoffman DA, Colbert CM, Magee JC. 1999. Regulation of back-propagating action
potentials in hippocampal neurons. Curr. Opin. Neurobiol. 9:28892
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

Kampa BM, Clements J, Jonas P, Stuart GJ. 2004. Kinetics of Mg2+ unblock of NMDA receptors:
implications for spike-timing dependent synaptic plasticity. J. Physiol. 556:33745
Kampa BM, Letzkus JJ, Stuart GJ. 2006. Requirement of dendritic calcium spikes for induction
by Cornell University on 10/30/11. For personal use only.

of spike-timing-dependent synaptic plasticity. J. Physiol. 574:28390

Karmarkar UR, Buonomano DV. 2002. A model of spike-timing dependent plasticity: one or two
coincidence detectors? J. Neurophysiol. 88:50713
Karmarkar UR, Dan Y. 2006. Experience-dependent plasticity in adult visual cortex. Neuron
52:57785
Katz LC, Shatz CJ. 1996. Synaptic activity and the construction of cortical circuits. Science
274:113338
Kilgard MP, Merzenich MM. 1998. Cortical map reorganization enabled by nucleus basalis ac-
tivity. Science 279:171418
Kirkwood A, Bear MF. 1994. Hebbian synapses in visual cortex. J. Neurosci. 14:163445
Kobayashi K, Poo MM. 2004. Spike train timing-dependent associative modication of hippocam-
pal CA3 recurrent synapses by mossy bers. Neuron 41:44554
Koester HJ, Sakmann B. 1998. Calcium dynamics in single spines during coincident pre- and
postsynaptic activity depend on relative timing of back-propagating action potentials and
subthreshold excitatory postsynaptic potentials. Proc. Natl. Acad. Sci. USA 95:9596601
Larkum ME, Kaiser KM, Sakmann B. 1999a. Calcium electrogenesis in distal apical dendrites
of layer 5 pyramidal cells at a critical frequency of back-propagating action potentials. Proc.
Natl. Acad. Sci. USA 96:146004
Larkum ME, Zhu JJ, Sakmann B. 1999b. A new cellular mechanism for coupling inputs arriving
at different cortical layers. Nature 398:33841
Lee I, Rao G, Knierim JJ. 2004. A double dissociation between hippocampal subelds: differential
time course of CA3 and CA1 place cells for processing changed environments. Neuron 42:803
15
Lester RA, Clements JD, Westbrook GL, Jahr CE. 1990. Channel kinetics determine the time
course of NMDA receptor-mediated synaptic currents. Nature 346:56567
Letzkus JJ, Kampa BM, Stuart GJ. 2006. Learning rules for spike timing-dependent plasticity
depend on dendritic synapse location. J. Neurosci. 26:1042029
Levy WB, Steward O. 1983. Temporal contiguity requirements for long-term associative poten-
tiation/depression in the hippocampus. Neuroscience 8:79197
Li CY, Lu JT, Wu CP, Duan SM, Poo MM. 2004. Bidirectional modication of presynaptic neu-
ronal excitability accompanying spike timing-dependent synaptic plasticity. Neuron 41:25768
Lin YW, Min MY, Chiu TH, Yang HW. 2003. Enhancement of associative long-term potentia-
tion by activation of beta-adrenergic receptors at CA1 synapses in rat hippocampal slices. J.
Neurosci. 23:417381
Linden DJ, Connor JA. 1995. Long-term synaptic depression. Annu. Rev. Neurosci. 18:31957
Lisman J. 1989. A mechanism for the Hebb and the anti-Hebb processes underlying learning and
memory. Proc. Natl. Acad. Sci. USA 86:957478

42 Caporale Dan

432
 Liu L, Wong TP, Pozza MF, Lingenhoehl K, Wang Y, et al. 2004. Role of NMDA receptor
subtypes in governing the direction of hippocampal synaptic plasticity. Science 304:102124
Liu QS, Pu L, Poo MM. 2005. Repeated cocaine exposure in vivo facilitates LTP induction in
midbrain dopamine neurons. Nature 437:102731
Louie K, Wilson MA. 2001. Temporally structured replay of awake hippocampal ensemble activity
during rapid eye movement sleep. Neuron 29:14556
Maejima T, Oka S, Hashimotodani Y, Ohno-Shosaku T, Aiba A, et al. 2005. Synaptically driven
endocannabinoid release requires Ca2+ -assisted metabotropic glutamate receptor subtype 1
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

to phospholipase Cbeta4 signaling cascade in the cerebellum. J. Neurosci. 25:682635

Magee JC, Johnston D. 1997. A synaptically controlled, associative signal for Hebbian plasticity
in hippocampal neurons. Science 275:20913
by Cornell University on 10/30/11. For personal use only.

Magee JC, Johnston D. 2005. Plasticity of dendritic function. Curr. Opin. Neurobiol. 15:33442
Malenka RC, Bear MF. 2004. LTP and LTD: an embarrassment of riches. Neuron 44:521
Markram H, Lubke J, Frotscher M, Sakmann B. 1997. Regulation of synaptic efcacy by coinci-
dence of postsynaptic APs and EPSPs. Science 275:21315
Martin SJ, Grimwood PD, Morris RG. 2000. Synaptic plasticity and memory: an evaluation of
the hypothesis. Annu. Rev. Neurosci. 23:649711
Mayer ML, Westbrook GL, Guthrie PB. 1984. Voltage-dependent block by Mg2+ of NMDA
responses in spinal cord neurones. Nature 309:26163
Mehta MR, Barnes CA, McNaughton BL. 1997. Experience-dependent, asymmetric expansion
of hippocampal place elds. Proc. Natl. Acad. Sci. USA 94:891821
Mehta MR, Quirk MC, Wilson MA. 2000. Experience-dependent asymmetric shape of hippocam-
pal receptive elds. Neuron 25:70715
Meliza CD, Dan Y. 2006. Receptive-eld modication in rat visual cortex induced by paired visual
stimulation and single-cell spiking. Neuron 49:18389
Meredith RM, Floyer-Lea AM, Paulsen O. 2003. Maturation of long-term potentiation induction
rules in rodent hippocampus: role of GABAergic inhibition. J. Neurosci. 23:1114246
Migliore M, Shepherd GM. 2002. Emerging rules for the distributions of active dendritic con-
ductances. Nat. Rev. Neurosci. 3:36270
Miles R, Toth K, Gulyas AI, Hajos N, Freund TF. 1996. Differences between somatic and dendritic
inhibition in the hippocampus. Neuron 16:81523
Mu Y, Poo MM. 2006. Spike timing-dependent LTP/LTD mediates visual experience-dependent
plasticity in a developing retinotectal system. Neuron 50:11525
Mulkey RM, Malenka RC. 1992. Mechanisms underlying induction of homosynaptic long-term
depression in area CA1 of the hippocampus. Neuron 9:96775
Nadasdy Z, Hirase H, Czurko A, Csicsvari J, Buzsaki G. 1999. Replay and time compression of
recurring spike sequences in the hippocampus. J. Neurosci. 19:9497507
Nevian T, Sakmann B. 2004. Single spine Ca2+ signals evoked by coincident EPSPs and back-
propagating action potentials in spiny stellate cells of layer 4 in the juvenile rat somatosensory
barrel cortex. J. Neurosci. 24:168999
Nevian T, Sakmann B. 2006. Spine Ca2+ signaling in spike-timing-dependent plasticity.
J. Neurosci. 26:1100113
Nishida S, Johnston A. 1999. Inuence of motion signals on the perceived position of spatial
pattern. Nature 397:61012
Nishiyama M, Hong K, Mikoshiba K, Poo MM, Kato K. 2000. Calcium stores regulate the polarity
and input specicity of synaptic modication. Nature 408:58488
Nowak L, Bregestovski P, Ascher P, Herbet A, Prochiantz A. 1984. Magnesium gates glutamate-
activated channels in mouse central neurones. Nature 307:46265

www.annualreviews.org Spike TimingDependent Plasticity 43

433
 Pike FG, Meredith RM, Olding AW, Paulsen O. 1999. Rapid report: postsynaptic bursting is
essential for Hebbian induction of associative long-term potentiation at excitatory synapses
in rat hippocampus. J. Physiol. 518(Pt. 2):57176
Pu L, Liu QS, Poo MM. 2006. BDNF-dependent synaptic sensitization in midbrain dopamine
neurons after cocaine withdrawal. Nat. Neurosci. 9:6057
Rall W. 1967. Distinguishing theoretical synaptic potentials computed for different soma-
dendritic distributions of synaptic input. J. Neurophysiol. 30:113868
Ramachandran VS, Anstis SM. 1990. Illusory displacement of equiluminous kinetic edges. Per-
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

ception 19:61116
Rao RP, Sejnowski TJ. 2001a. Predictive learning of temporal sequences in recurrent neocortical
circuits. Novartis Found Symp. 239:20829; discussion 2940
by Cornell University on 10/30/11. For personal use only.

Rao RP, Sejnowski TJ. 2001b. Spike-timing-dependent Hebbian plasticity as temporal difference
learning. Neural Comput. 13:222137
Ribeiro S, Gervasoni D, Soares ES, Zhou Y, Lin SC, et al. 2004. Long-lasting novelty-induced
neuronal reverberation during slow-wave sleep in multiple forebrain areas. PLoS Biol. 2:E24
Rosenmund C, Feltz A, Westbrook GL. 1995. Calcium-dependent inactivation of synaptic NMDA
receptors in hippocampal neurons. J. Neurophysiol. 73:42730
Schuett S, Bonhoeffer T, Hubener M. 2001. Pairing-induced changes of orientation maps in cat
visual cortex. Neuron 32:32537
Seol GH, Ziburkus J, Huang S, Song L, Kim IT, et al. 2007. Neuromodulators control the polarity
of spike-timing-dependent synaptic plasticity. Neuron 55:91929
Shadlen MN, Newsome WT. 1994. Noise, neural codes and cortical organization. Curr. Opin.
Neurobiol. 4:56979
Shouval HZ, Bear MF, Cooper LN. 2002. A unied model of NMDA receptor-dependent bidi-
rectional synaptic plasticity. Proc. Natl. Acad. Sci. USA 99:1083136
Shu Y, Hasenstaub A, Badoual M, Bal T, McCormick DA. 2003. Barrages of synaptic activity
control the gain and sensitivity of cortical neurons. J. Neurosci. 23:10388401
Siegelbaum SA, Kandel ER. 1991. Learning-related synaptic plasticity: LTP and LTD. Curr. Opin.
Neurobiol. 1:11320
Sjostrom PJ, Hausser M. 2006. A cooperative switch determines the sign of synaptic plasticity in
distal dendrites of neocortical pyramidal neurons. Neuron 51:22738
Sjostrom PJ, Turrigiano GG, Nelson SB. 2001. Rate, timing, and cooperativity jointly determine
cortical synaptic plasticity. Neuron 32:114964
Sjostrom PJ, Turrigiano GG, Nelson SB. 2003. Neocortical LTD via coincident activation of
presynaptic NMDA and cannabinoid receptors. Neuron 39:64154
Snowden RJ. 1998. Shifts in perceived position following adaptation to visual motion. Curr. Biol.
8:134345
Softky WR, Koch C. 1993. The highly irregular ring of cortical cells is inconsistent with temporal
integration of random EPSPs. J. Neurosci. 13:33450
Spruston N, Schiller Y, Stuart G, Sakmann B. 1995. Activity-dependent action potential invasion
and calcium inux into hippocampal CA1 dendrites. Science 268:297300
Stanton PK, Sejnowski TJ. 1989. Associative long-term depression in the hippocampus induced
by Hebbian covariance. Nature 339:21518
Stefan K, Kunesch E, Benecke R, Cohen LG, Classen J. 2002. Mechanisms of enhancement of
human motor cortex excitability induced by interventional paired associative stimulation. J.
Physiol. 543:699708
Stent GS. 1973. A physiological mechanism for Hebbs postulate of learning. Proc. Natl. Acad. Sci.
USA 70:9971001

44 Caporale Dan

434
 Stuart G, Schiller J, Sakmann B. 1997a. Action potential initiation and propagation in rat neocor-
tical pyramidal neurons. J. Physiol. 505(Pt. 3):61732
Stuart G, Spruston N, Sakmann B, Hausser M. 1997b. Action potential initiation and backprop-
agation in neurons of the mammalian CNS. Trends Neurosci 20:12531
Stuart GJ, Hausser M. 2001. Dendritic coincidence detection of EPSPs and action potentials. Nat.
Neurosci. 4:6371
Stuart GJ, Sakmann B. 1994. Active propagation of somatic action potentials into neocortical
pyramidal cell dendrites. Nature 367:6972
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

Tong G, Shepherd D, Jahr CE. 1995. Synaptic desensitization of NMDA receptors by calcineurin.
Science 267:151012
Tsubokawa H, Ross WN. 1996. IPSPs modulate spike backpropagation and associated [Ca2+ ]i
by Cornell University on 10/30/11. For personal use only.

changes in the dendrites of hippocampal CA1 pyramidal neurons. J. Neurophysiol. 76:2896

906
Tsubokawa H, Ross WN. 1997. Muscarinic modulation of spike backpropagation in the apical
dendrites of hippocampal CA1 pyramidal neurons. J. Neurosci. 17:578291
Tzounopoulos T, Kim Y, Oertel D, Trussell LO. 2004. Cell-specic, spike timing-dependent
plasticities in the dorsal cochlear nucleus. Nat. Neurosci. 7:71925
Tzounopoulos T, Rubio ME, Keen JE, Trussell LO. 2007. Coactivation of pre- and postsynap-
tic signaling mechanisms determines cell-specic spike-timing-dependent plasticity. Neuron
54:291301
Vislay-Meltzer RL, Kampff AR, Engert F. 2006. Spatiotemporal specicity of neuronal activity
directs the modication of receptive elds in the developing retinotectal system. Neuron
50:10114
Wang HX, Gerkin RC, Nauen DW, Bi GQ. 2005. Coactivation and timing-dependent integration
of synaptic potentiation and depression. Nat. Neurosci. 8:18793
Wang Z, Xu NL, Wu CP, Duan S, Poo MM. 2003. Bidirectional changes in spatial dendritic
integration accompanying long-term synaptic modications. Neuron 37:46372
Watanabe S, Hoffman DA, Migliore M, Johnston D. 2002. Dendritic K+ channels contribute
to spike-timing dependent long-term potentiation in hippocampal pyramidal neurons. Proc.
Natl. Acad. Sci. USA 99:836671
Waters J, Helmchen F. 2004. Boosting of action potential backpropagation by neocortical network
activity in vivo. J. Neurosci. 24:1112736
Waters J, Schaefer A, Sakmann B. 2005. Backpropagating action potentials in neurones: measure-
ment, mechanisms and potential functions. Prog. Biophys. Mol. Biol. 87:14570
Whitaker D, McGraw PV, Pearson S. 1999. Non-veridical size perception of expanding and
contracting objects. Vision Res. 39:29993009
Williams SR, Stuart GJ. 2000. Backpropagation of physiological spike trains in neocortical pyra-
midal neurons: implications for temporal coding in dendrites. J. Neurosci. 20:823846
Wilson MA, McNaughton BL. 1994. Reactivation of hippocampal ensemble memories during
sleep. Science 265:67679
Wittenberg GM, Wang SS. 2006. Malleability of spike-timing-dependent plasticity at the CA3-
CA1 synapse. J. Neurosci. 26:661017
Wolters A, Sandbrink F, Schlottmann A, Kunesch E, Stefan K, et al. 2003. A temporally asymmetric
Hebbian rule governing plasticity in the human motor cortex. J. Neurophysiol. 89:233945
Wolters A, Schmidt A, Schramm A, Zeller D, Naumann M, et al. 2005. Timing-dependent plas-
ticity in human primary somatosensory cortex. J. Physiol. 565:103952
Woodin MA, Ganguly K, Poo MM. 2003. Coincident pre- and postsynaptic activity modies
GABAergic synapses by postsynaptic changes in Cl- transporter activity. Neuron 39:80720

www.annualreviews.org Spike TimingDependent Plasticity 45

435
 Xu NL, Ye CQ, Poo MM, Zhang XH. 2006. Coincidence detection of synaptic inputs is facilitated
at the distal dendrites after long-term potentiation induction. J. Neurosci. 26:30029
Yang SN, Tang YG, Zucker RS. 1999. Selective induction of LTP and LTD by postsynaptic [Ca2+ ]i
elevation. J. Neurophysiol. 81:78187
Yao H, Dan Y. 2001. Stimulus timing-dependent plasticity in cortical processing of orientation.
Neuron 32:31523
Yao H, Shen Y, Dan Y. 2004. Intracortical mechanism of stimulus-timing-dependent plasticity in
visual cortical orientation tuning. Proc. Natl. Acad. Sci. USA 101:508186
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

Yao H, Shi L, Han F, Gao H, Dan Y. 2007. Rapid learning in cortical coding of visual scenes. Nat.
Neurosci. 10:77278
Yu X, Yoganarasimha D, Knierim JJ. 2006. Backward shift of head direction tuning curves of the
by Cornell University on 10/30/11. For personal use only.

anterior thalamus: comparison with CA1 place elds. Neuron 52:71729

Zhang LI, Tao HW, Holt CE, Harris WA, Poo M. 1998. A critical window for cooperation and
competition among developing retinotectal synapses. Nature 395:3744
Zhang W, Linden DJ. 2003. The other side of the engram: experience-driven changes in neuronal
intrinsic excitability. Nat. Rev. Neurosci. 4:885900
Zhou Q, Tao HW, Poo MM. 2003. Reversal and stabilization of synaptic modications in a
developing visual system. Science 300:195357
Zucker RS, Regehr WG. 2002. Short-term synaptic plasticity. Annu. Rev. Physiol. 64:355405

46 Caporale Dan

436
 Annual Review of
Neuroscience

Contents Volume 31, 2008

Cerebellum-Like Structures and Their Implications for Cerebellar

Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

Function
Curtis C. Bell, Victor Han, and Nathaniel B. Sawtell p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1
Spike TimingDependent Plasticity: A Hebbian Learning Rule
by Cornell University on 10/30/11. For personal use only.

Natalia Caporale and Yang Dan p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p25

Balancing Structure and Function at Hippocampal Dendritic Spines
Jennifer N. Bourne and Kristen M. Harris p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p47
Place Cells, Grid Cells, and the Brains Spatial Representation System
Edvard I. Moser, Emilio Kropff, and May-Britt Moser p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p69
Mitochondrial Disorders in the Nervous System
Salvatore DiMauro and Eric A. Schon p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p91
Vestibular System: The Many Facets of a Multimodal Sense
Dora E. Angelaki and Kathleen E. Cullen p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 125
Role of Axonal Transport in Neurodegenerative Diseases
Kurt J. De Vos, Andrew J. Grierson, Steven Ackerley, and Christopher C.J. Miller p p p 151
Active and Passive Immunotherapy for Neurodegenerative Disorders
David L. Brody and David M. Holtzman p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 175
Descending Pathways in Motor Control
Roger N. Lemon p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 195
Task Set and Prefrontal Cortex
Katsuyuki Sakai p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 219
Multiple Sclerosis: An Immune or Neurodegenerative Disorder?
Bruce D. Trapp and Klaus-Armin Nave p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 247
Multifunctional Pattern-Generating Circuits
K.L. Briggman and W.B. Kristan, Jr. p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 271
Retinal Axon Growth at the Optic Chiasm: To Cross or Not to Cross
Timothy J. Petros, Alexandra Rebsam, and Carol A. Mason p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 295

437
 Brain Circuits for the Internal Monitoring of Movements
Marc A. Sommer and Robert H. Wurtz p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 317
Wnt Signaling in Neural Circuit Assembly
Patricia C. Salinas and Yimin Zou p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 339
Habits, Rituals, and the Evaluative Brain
Ann M. Graybiel p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 359
Mechanisms of Self-Motion Perception
Kenneth H. Britten p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 389
Mechanisms of Face Perception
Doris Y. Tsao and Margaret S. Livingstone p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 411
Annu. Rev. Neurosci. 2008.31:25-46. Downloaded from www.annualreviews.org

The Prions Elusive Reason for Being

Adriano Aguzzi, Frank Baumann, and Juliane Bremer p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 439
by Cornell University on 10/30/11. For personal use only.

Mechanisms Underlying Development of Visual Maps and

Receptive Fields
Andrew D. Huberman, Marla B. Feller, and Barbara Chapman p p p p p p p p p p p p p p p p p p p p p p p p 479
Neural Substrates of Language Acquisition
Patricia K. Kuhl and Maritza Rivera-Gaxiola p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 511
Axon-Glial Signaling and the Glial Support of Axon Function
Klaus-Armin Nave and Bruce D. Trapp p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 535
Signaling Mechanisms Linking Neuronal Activity to Gene Expression
and Plasticity of the Nervous System
Steven W. Flavell and Michael E. Greenberg p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 563

Indexes

Cumulative Index of Contributing Authors, Volumes 2231 p p p p p p p p p p p p p p p p p p p p p p p p p p p 591

Cumulative Index of Chapter Titles, Volumes 2231 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 595

Errata

An online log of corrections to Annual Review of Neuroscience articles may be found at

http://neuro.annualreviews.org/

vi Contents

438
 Available online at www.sciencedirect.com
Neural mechanisms for filtering self-generated sensory signals in
cerebellum-like circuits
Tim Requarth and Nathaniel B Sawtell
This review focuses on recent progress in understanding
mechanisms for filtering self-generated sensory signals in
cerebellum-like circuits in fish and mammals. Recent in vitro
studies in weakly electric gymnotid fish have explored the
interplay among anti-Hebbian plasticity, synaptic dynamics,
and feedforward inhibition in canceling self-generated
electrosensory inputs. Studies of the mammalian dorsal
cochlear nucleus have revealed multimodal integration and
anti-Hebbian plasticity, suggesting that this circuit may
adaptively filter incoming auditory information. In vivo studies in
weakly electric mormryid fish suggest a key role for granule cell
coding in sensory filtering. The clear links between synaptic
plasticity and systems level sensory filtering in cerebellum-like
circuits may provide insights into hypothesized adaptive
filtering functions of the cerebellum itself.
Address
Department of Neuroscience and Kavli Institute for Brain Science,
Columbia University Medical Center, New York, NY 10032, United
States
Corresponding author: Sawtell, Nathaniel B (ns2635@columbia.edu)
Current Opinion in Neurobiology 2011, 21:602608
This review comes from a themed issue on
Sensory and Motor Systems
Edited by Sascha du Lac and Rachel Wilson
Available online 23rd June 2011
0959-4388/$ see front matter
# 2011 Elsevier Ltd. All rights reserved.
DOI 10.1016/j.conb.2011.05.031
Introduction
Sensory receptors do not discriminate between activation
due to an animals own movements and activation due to
changes in the external world. Yet animals rarely confuse
self-generated signals with external events. How do ner-
vous systems solve the fundamental problem of dis-
tinguishing self from other [1]? One longstanding
hypothesis is that copies of motor commands routed to
sensory areas, termed corollary discharge or efference
copy, are used to predict and cancel out patterns of
sensory input due to an animals own movements [2,3].
Another possibility is that sensory input itself, from either
the same or different modalities, could be used to resolve
the ambiguity. Progress in understanding the neural
mechanisms for filtering self-generated sensory signals
Current Opinion in Neurobiology 2011, 21:602608
has been made in a number of model systems, from
primates to crickets [4]. This review focuses on recent
studies that have contributed to understanding how
synaptic plasticity contributes to sensory filtering in sen-
sory structures with circuitry resembling the cerebellum
that are found at the first stages of electrosensory proces-
sing in fish, auditory and auditory processing in mammals.
A combination of in vitro, in vivo, and computational
studies conducted over three decades has led to a com-
pelling account of how anti-Hebbian synaptic plasticity in
cerebellum-like circuits in fish acts to predict and cancel
the electrosensory consequences of the animals own
behavior [57]. In what appears to be a remarkable
case of convergent evolution, very similar predictive
mechanisms have been revealed in cerebellum-like cir-
cuits in three distinct groups of fish in which electrosen-
sory systems likely evolved independently [8]. These
include the dorsal octavolateral nucleus (DON) in elas-
mobranch fish (the group that includes sharks, skates, and
rays), the electrosensory lobe (ELL) of weakly electric
mormyrid fish from Africa and the ELL of weakly electric
gymnotid fish from South America.
Principal cells in these circuits integrate input from
peripheral electroreceptors with a diverse array of sensory
and motor signals conveyed by a granule cellparallel
fiber system similar to that found in the cerebellum [9]
(Figure 1a, left). Many of the inputs to granule cells carry
information about the animals own behavior, including
proprioceptive information about the positions and move-
ments of the tail and fins and corollary discharge signals
related to motor commands for gill movements, tail move-
ments, and the generation of electrical fields, known as
electric organ discharges. Hence parallel fiber input to
principal cells could, in principle, be used to predict
changes in electroreceptor input due to the animals
own behavior.
In all three groups of fish, pairing electrosensory stimuli
with predictive inputs conveyed by parallel fibers results
in changes in the responses of principal cells that
resemble highly specific negative images of the neural
response to the predictable stimulus [1014] (Figure 1a,
right). Addition of such negative images to the actual
electrosensory input cancels predictable components due
to the animals own behavior. The cellular mechanism for
the formation of negative images is a form of long-term
plasticity at synapses between parallel fibers and principal
cells. In contrast to more familiar Hebbian plasticity
www.sciencedirect.com
439
 Neural mechanisms for filtering self-generated sensory signals Requarth and Sawtell 603

Figure 1

(a) negative images and sensory cancellation in cerebellum-like circuits

before plasticity
parallel fiber inputs: individual PF inputs

membrane potential
corollary discharge
proprioception
other sensory modalities
summed PF input

membrane potential
(PF input+ sensory input)
site of anti-Hebbian
sensory input
plasticity

after plasticity
individual PF inputs

membrane potential
sensory input
(external+self-generated signals) filtered output
summed PF input
(negative image)
sensory cancellation

principal cell sensory input

time after predictable stimulus

(b) mormyrid ELL (c) gymnotid ELL (d) dorsal cochlear nucleus

post before pre before post before pre before post before pre before post before pre before
pre post pre post pre post pre post
CWC
-100 ms 100 ms

-100 ms 100 ms -20 ms 20 ms -20 ms 20 ms

MG
+ cholinergic
activation
to higher stages

electroreceptor input electroreceptor input auditory nerve input

Current Opinion in Neurobiology

Sensory filtering and synaptic plasticity in cerebellum-like circuits. (a) Left, schematic of a generic cerebellum-like circuit. Principal cells integrate
peripheral sensory input (orange) with predictive signals conveyed by a granule cellparallel fiber system (blue). Right, schematic representation of the
formation of a negative image (blue line in bottom panel) of a predictable temporal pattern of sensory input (orange). Negative images are formed via bi-
directional changes in the strength of parallel fiber synapses (blue) according to an anti-Hebbian learning rule (see text). After plasticity, the addition of
negative images (blue) with sensory input (orange) results in the cancellation of predictable features of the sensory input (black). (b) Circuitry and anti-
Hebbian learning rule at parallel fiber synapses onto GABAergic medium ganglion (MG) cells in the mormyrid ELL. Anti-Hebbian plasticity has been
observed in efferent cells as well, though the learning rule has not been studied in detail. (c) Circuitry and anti-Hebbian learning rule at parallel fiber
synapses onto glutamatergic efferent cells in the gymnotid ELL. Long-term depression is induced by correlated presynaptic and postsynaptic bursts
irrespective of relative timing. (d) Circuitry and learning rules at parallel fiber synapses onto GABAergic cartwheel cells (CWC) and glutamatergic
efferent cells in the DCN. PF synapses onto CWC synapses exhibit timing-dependent anti-Hebbian plasticity. PF synapses onto efferent cells exhibit
timing-dependent Hebbian plasticity that can be converted to anti-Hebbian plasticity by activating cholinergic inputs.

mechanisms that strengthen synapses between neurons
that are active together, the plasticity found in cerebellum-
like circuits is anti-Hebbian synapses that are active
together are weakened. In the context of cerebellum-like
circuitry, this learning rule provides a simple and powerful
mechanism for filtering out self-generated signals.
Increases in principal cell firing that occur together with
(i.e., can be predicted by) parallel fiber input are opposed
and eventually canceled by weakening of parallel fiber
synapses. Conversely, predictable decreases in principal
cell firing are opposed by increases in parallel fiber synaptic
strength. The mechanism is indifferent to the nature of the
signals conveyed by parallel fibers as long as they reliably
predict changes in principal cell output. On the basis of
these results it was suggested that cerebellum-like circuitry
could operate as an adaptive filter by continually generat-
ing and updating sensory predictions based on associations
between sensory and motor signals conveyed by parallel

www.sciencedirect.com Current Opinion in Neurobiology 2011, 21:602608

440
604 Sensory and Motor Systems
fibers and current sensory inputs, then subtracting these
predictions from the neural response of principal cells.
Similar adaptive filtering mechanisms have been proposed
for the mammalian cerebellar cortex in the context of
motor control [15,16].
Cellular mechanisms for negative image
formation
Anti-Hebbian plasticity has been characterized in some
detail in a brain slice preparation of the mormryid ELL
[17,18] (Figure 1b). Notably, the sign and magnitude of
the changes in parallel fiber strength depend critically on
the timing of the excitatory postsynaptic potential (EPSP)
evoked by parallel fiber stimulation relative to a postsyn-
aptic dendritic spike. This was among the first demon-
strations of spike timing-dependent plasticity in the
vertebrate brain. Parallel fiber EPSPs that preceded den-
dritic spikes by 50 ms or less were depressed, while
EPSPs occurring at all other delays were potentiated.
Because potentiation also occurred when parallel fiber
EPSPs were evoked at sufficiently high rates (>0.5 Hz)
without a postsynaptic spike, the potentiation was termed
nonassociative. The associative synaptic depression
required activation of NMDA-type glutamate receptors
and changes in postsynaptic calcium. The nonassociative
potentiation reversed the depression and vice versa, with
both forms of plasticity appearing to share a presynaptic
locus of expression. Modeling studies demonstrated that
the learning rule observed in vitro could account for the
formation of negative images observed in vivo [19].
In vivo studies have revealed negative images with similar
properties in the mormryid ELL, the gymnotid ELL, and
the elasmobranch DON. However, until recently, less
was known about the cellular mechanisms for plasticity in
the latter two structures. A recent study combining in vivo
electrophysiology and immunohistochemistry provides
strong evidence for a role for NMDA receptors in nega-
tive image formation in the DON [20]. Pharmacological
blockade of NMDA receptors reversibly blocked nega-
tive image formation in response to pairing artificial
electrosensory stimuli with either electrical stimulation
of parallel fibers or natural patterns of parallel fiber
activation related to the fishs ventilatory cycle. Venti-
latory movements are a significant source of self-gener-
ated electrosensory input for elasmobranchs and parallel
fibers are known to convey proprioceptive and motor
corollary discharge signals related to ventilation.
A recent in vitro study by Harvey-Girard et al. has pro-
vided the first detailed report of synaptic plasticity mech-
anisms in the gymnotid ELL [21] (Figure 1c). As in the
mormyrid ELL and the DON, plasticity in the gymnotid
ELL is anti-Hebbian and depends on NMDA receptors.
Interestingly, plasticity in the gymnotid ELL does not
depend on the relative timing of single presynaptic and
postsynaptic spikes as in the mormryid ELL. Instead a
Current Opinion in Neurobiology 2011, 21:602608
long-lasting depression is induced when presynaptic
spike bursts are temporally correlated with postsynaptic
bursts (within 50 ms) irrespective of temporal order. At
present, the functional significance of timing-dependent
anti-Hebbian plasticity in the mormryid ELL versus
correlation-based anti-Hebbian plasticity in the gymnotid
ELL is unclear, though it may relate to precisely timed
motor corollary discharge signals which are present in the
mormyrid ELL but not in the gymnotid ELL.
The Harvey-Girard et al. study also revealed another
important difference in plasticity mechanisms in morm-
ryids and gymnotids. Associative depression in the morm-
ryid ELL is reversed by a long-lasting nonassociative
potentiation induced by parallel fiber stimulation alone.
No such mechanism for reversing burst-induced depres-
sion was found in the gymnotid ELL. The lack of a
potentiation mechanism is troubling because in vivo
negative images are characterized by both decreased
responses that oppose predictable increases in principal
cell output (presumably due to associative depression at
parallel fiber synapses) as well as increased responses that
oppose predictable decreases in principal cell output.
How are increased responses explained given the lack
of a mechanism for potentiation in the gymnotid ELL? As
the authors point out, a mundane but important possib-
ility is that potentiation mechanisms exist but were
simply not engaged under the conditions of the exper-
iments. As a case in point, it took many years to discover a
form of parallel fiber potentiation that reversed parallel
fiber long-term depression found in the mammalian
cerebellum [22].
The authors suggest another, more intriguing, possibility
for increased responses motivated by recent in vitro
studies in the gymnotid ELL by Lewis et al. [23,24].
Lewis et al. utilized a data-based model of parallel fiber
synaptic dynamics [23] and dynamic clamp to explore
how changes in the balance of excitatory and inhibitory
inputs to ELL principal cells contribute to negative
image formation [24]. Lewis et al. found that the effects
of increasing excitatory parallel fiber input rate on prin-
cipal cell firing were nonmonotonic, such that higher
input rates actually inhibited principal cell firing. The
shift from net excitation to inhibition was due to the
interplay between synaptic dynamics and feedforward
inhibition to principal cells via molecular layer inter-
neurons. Harvey-Girard et al. extended this work by
showing that a model incorporating parallel fiber synaptic
dynamics, feedforward inhibition, and associative synap-
tic depression could account for negative images observed
in vivo. Although experiments are needed to validate the
model, attempts to understand the complex interactions
among long-term synaptic plasticity, short-term synaptic
dynamics and inhibitory circuitry are of general import-
ance. Such interactions are likely to occur in many circuits
where plasticity is present but are, in general, poorly
www.sciencedirect.com
441
 Neural mechanisms for filtering self-generated sensory signals Requarth and Sawtell 605
understood. Cerebellum-like circuits in weakly electric
fish provide a system in which to explore the significance
of multiple sites and forms of synaptic plasticity in the
context of well-defined systems level sensory filtering
functions.
Cerebellum-like circuitry and anti-Hebbian
plasticity in the mammalian auditory system
Recent studies have also provided insights into functional
circuitry and synaptic plasticity in the dorsal cochlear
nucleus (DCN) a cerebellum-like circuit at the first
stage of auditory processing in mammals(Figure 1d).
DCN principal cells integrate auditory nerve input with
parallel fiber input conveying both auditory and nonau-
ditory information from a variety of brain regions [25].
These signals could be used to filter incoming auditory
inputs, as in cerebellum-like circuits in fish. One possib-
ility is that proprioceptive and/or motor corollary dis-
charge information about the position and movements
of the head and external ears (conveyed by parallel fibers)
may be needed to properly interpret incoming auditory
nerve input. The DCN is believed to play a role in
processing spectral cues for localizing sound sources in
elevation. Changes in spectral cues could reflect move-
ment of an external sound source, movement of the
animal itself, or both. Hence, in order to localize sounds,
auditory information must be combined with information
about the position and movements of the animal. Given
its distinctive cerebellum-like circuitry and plasticity (see
below), this computation may begin at the first stage of
auditory processing in the DCN. Consistent with this
possibility, principal cells in the DCN of cats are known
to respond strongly to somatosensory stimulation, in-
cluding movements of the external ear [26]. A recent
study demonstrated that brief electrical stimulation of the
somatosensory dorsal column nucleus (a major input to
DCN granule cells) elicits complex and long-lasting
alterations in the responses of DCN neurons to sound
[27].
Another related possibility is that parallel fibers convey
signals related to the animals own chewing, vocalization,
or respiration that could be used to predict and cancel
their auditory consequences. Consistent with this, tract
tracing, immunohistocehmical, and in vivo electrophysio-
logical studies of multimodal integration in the DCN of
rodents suggest that signals related to chewing, respir-
ation, and vocalization may indeed reach DCN principal
cells via the granule cellparallel fiber system [28].
Responses to nonauditory inputs in DCN appear to be
mediated in large part by cartwheel cells a network of
inhibitory interneurons that receive the large majority of
parallel fiber inputs. Cartwheel cells powerfully inhibit
fusiform cells, the output cells of the DCN. A number of
recent in vitro studies have shed light on the interactions
among cartwheel cells as well as their effects on fusiform
www.sciencedirect.com
cells [29,30,31]. Using dual recordings to examine inter-
actions among cartwheel cells, fusiform cells, and parallel
fibers, Roberts and Trussell [31] made the surprising
observation that parallel fiber connectivity was extremely
sparse. Minimal stimulation showed that neighboring
cells were rarely activated by the same parallel fiber
synapses, suggesting that feedforward inhibition may
not be the dominant mode of cartwheel cell action.
Although in vivo evidence for sensory filtering in the DCN
has not yet been obtained, in vitro studies have revealed
parallel fiber synaptic plasticity mechanisms remarkably
similar to those in the mormyrid ELL [3234] (Figure 1b
and d). In both structures plasticity is anti-Hebbian,
NMDA receptor-dependent, presynaptically expressed,
and apparently reversed by a timing-independent form
of parallel fiber synaptic potentiation. Anti-Hebbian
plasticity is also found at analogous sites in the two struc-
tures at parallel fiber synapses onto the spiny dendrites
of interneurons (cartwheel cells in the DCN and medium
ganglion cells in mormyrid ELL) that inhibit nearby
glutamatergic output neurons. Interestingly, parallel fiber
synapses onto fusiform cells exhibit a Hebbian form of
spike timing-dependent plasticity, i.e. presynaptic inputs
that precede a postsynaptic spike result in long-term
potentiation. The mechanisms underlying this site-specific
plasticity have been studied in some detail [3335] and
depend in part on differences in endocannabinoid sig-
naling at cartwheel versus fusiform cell synapses.
Hebbian plasticity at parallel fiberfusiform cell synapses
is puzzling from the standpoint of a sensory filtering
function for the DCN. Such plasticity would apparently
strengthen the effects of predictable signals, rather than
canceling them as occurs in cerebellum-like structures in
fish. A recent study provides a possible resolution, show-
ing that Hebbian long-term potentiation at parallel fiber
synapses onto fusiform cells can be converted to anti-
Hebbian long-term depression if cholinergic neuromodu-
latory inputs are activated [36]. The switch was shown
to depend on interactions between postsynaptic M1/M3
muscarinic acetylcholine receptor activation and endo-
cannabinoid signaling pathways. The authors speculate
that neuromodulation could switch the polarity of
plasticity in a context-dependent manner and that both
Hebbian and anti-Hebbian plasticity may play function-
ally important roles. The presence of anti-Hebbian
plasticity at parallel fiber synapses onto both cartwheel
cell and fusiform cell synapses strongly motivates in vivo
tests of adaptive sensory filtering in the DCN.
Significance of granule cell coding for sensory
filtering
In both cerebellum-like circuits and the cerebellum itself,
diverse sensory and motor signals conveyed from various
brain regions by mossy fibers are recoded in granule cells
before being relayed to principal cells or Purkinje cells. Yet
Current Opinion in Neurobiology 2011, 21:602608
442
606 Sensory and Motor Systems
due to the difficulty of obtaining in vivo recordings from
these small, densely packed cells, little is known about
their function. Granule cell circuitry is extremely similar in
cerebellum-like circuits and the cerebellum in terms of
the granule cells themselves and the inhibitory Golgi cells
and excitatory unipolar brush cells that shape granule cell
responses to mossy fiber inputs. Hence insights regarding
the importance of granule cell circuitry for sensory filtering
in fish may be relevant for understanding functions of
cerebellar granule cell circuitry more generally.
Influential computational theories of cerebellar function
developed independently by Marr and Albus in the 1970s
posit that the large number of cerebellar granule cells
(around 70 billion in humans, or approximately 75% of the
total number of neurons in the entire brain) serve to
sparsely recode mossy fiber inputs, such that only a
small fraction of the granule cell population is active
for a given pattern of mossy fiber input [37,38]. Sparse
coding is expected to increase the capacity of downstream
Purkinje cells to discriminate between different mossy
fiber input patterns through associative plasticity at paral-
lel fiber synapses and is hypothesized to be important for
cerebellum-dependent motor learning. One way in which
sparse coding could be accomplished is if different mossy
fiber inputs converged onto individual granule cells and if
simultaneous activation of multiple inputs was required
to evoke action potentials in granule cells. Granule cell
responses have only begun to be explored in the mam-
malian cerebellum in vivo, and the question of sparse
coding remains open [39].
A recent study in the mormryid ELL provided a clear
example of multimodal integration in individual granule
cells, suggestive of sparse coding. The study also linked
multimodal responses in granule cells to the ability of
ELL principal cells to generate negative images via anti-
Hebbian plasticity [40]. Using in vivo whole-cell record-
ings in awake, paralyzed fish, it was shown that a subset of
granule cells integrate excitatory input from two distinct
classes of mossy fibers, one conveying proprioceptive
information about body position and the other corollary
discharge signals related to the motor command to dis-
charge the electric organ. Rates of spontaneous action
potential firing in granule cells were extremely low, and in
most cells activation of corollary discharge signals alone
did not evoke action potentials. However, if propriocep-
tive and corollary discharge signals were activated
together some granule cells did fire, thereby encoding
specific information about the conjunction of a motor
command and a particular body configuration. These
results are suggestive of sparse coding, though sampling
from a much larger number of cellse.g., using calcium
imagingwould be required to show this directly.
The author went on to show that principal neurons in
ELL were capable of generating negative images specific
Current Opinion in Neurobiology 2011, 21:602608
to the combinations of signals encoded by individual
granule cells, i.e. responses that were specific to particular
tail positions and particular times relative to motor com-
mands. Hence, integration of sensory and motor signals in
granule cells appears to provide a neural substrate for
generating predictions about the sensory consequences of
motor commands that take proprioceptive context into
account. Such predictions are likely to be of general use.
Predicting the consequences of a reaching movement, for
example, requires both the issued motor commands and
information about the starting position of the hand and
the target.
Future directions
Results described here span multiple levels of analysis and
hence provide many avenues for future studies from
detailed descriptions of molecular plasticity mechanisms,
to the identification of behavioral correlates for sensory
filtering in cerebellum-like circuits in fish, to in vivo
studies of sensory filtering in the DCN. Among the most
intriguing possibilities is that the relatively detailed
insights into sensory filtering mechanisms available from
studies of cerebellum-like circuits could provide general
insights into neural mechanisms for predicting sensory
events. Prediction is critical for a wide range of sensory,
motor, and cognitive functions. For example, predictions
about the sensory consequences of motor commands could
be used not only to cancel self-generated sensory signals,
as described above, but also for guiding rapid movements
in the face of sensory feedback that is delayed and/or
noisy. Such predictive mechanisms, termed forward
models, are widely discussed in the context of motor
control [41]. Interestingly, the generation of such forward
models is widely believed to occur in the cerebellum [42
44]. This raises the additional exciting possibility that
understanding mechanisms for sensory filtering in cerebel-
lum-like circuits will be a direct source of insight into the
function of the cerebellum itself.
Acknowledgements
This work was supported by a grant from the National Science Foundation
(IOS-1025849), the Kavli Institute for Brain Science, and the Alfred P. Sloan
Foundation.
References and recommended reading
Papers of particular interest, published within the period of review,
have been highlighted as:
 of special interest
 of outstanding interest
1. Churchland PS: Self-representation in nervous systems.
Science 2002, 296:308-310.
2. Sperry RW: Neural basis of the spontaneous optokinetic
response produced by visual inversion. J Comp Physiol Psychol
1950, 43:482-489.
3. von Holst E, Mittelstaedt H: The reafference principle.
Naturwissenschaften 1950, 37:464-476.
4. Crapse TB, Sommer MA: Corollary discharge across the animal
kingdom. Nat Rev Neurosci 2008, 9:587-600.
www.sciencedirect.com
443
 Neural mechanisms for filtering self-generated sensory signals Requarth and Sawtell 607
5. Bell C, Bodznick D, Montgomery J, Bastian J: The generation and
subtraction of sensory expectations within cerebellum-like
structures. Brain Behav Evol 1997, 50:17-31.
6. Bell CC: Memory-based expectations in electrosensory
systems. Curr Opin Neurobiol 2001, 11:481-487.
7. Montgomery JC, Coombs S, Conley RA, Bodznick D: Hindbrain
sensory processing in lateral line, electrosensory, and
auditory systems: a comparative overview of anatomical and
functional similarities. Audit Neurosci 1995, 1:207-231.
8. Bell CC: Evolution of cerebellum-like structures. Brain Behav
Evol 2002, 59:312-326.
9. Bell CC, Han V, Sawtell NB: Cerebellum-like structures and their
implications for cerebellar function. Annu Rev Neurosci 2008,
31:1-24.
10. Bastian J: Plasticity in an electrosensory system. I. General
features of dynamic sensory filter. J Neurophysiol 1996,
76:2483-2496.
11. Bell CC: An efference copy modified by reafferent input.
Science 1981, 214:450-453.
12. Bell CC, Caputi A, Grant K, Serrier J: Storage of a sensory
pattern by anti-Hebbian synaptic plasticity in an electric fish.
Proc Natl Acad Sci 1993, 90:4650-4654.
13. Bodznick D, Montgomery JC, Carey M: Adaptive mechanisms in
the elasmobranch hindbrain. J Exp Biol 1999, 202:1357-1364.
14. Sawtell NB, Williams A: Transformations of electrosensory
encoding associated with an adaptive filter. J Neurosci 2008,
28:1598-1612.
15. Dean P, Porrill J, Stone JV: Decorrelation control by the
cerebellum achieves oculomotor plant compensation in
simulated vestibulo-ocular reflex. Proc Biol Sci 2002, 269:1895-
1904.
16. Fujita M: Adaptive filter model of the cerebellum. Biol Cybern
1982, 45:195-206.
17. Bell CC, Han VZ, Sugawara S, Grant K: Synaptic plasticity in a
cerebellum-like structure depends on temporal order. Nature
1997, 387:278-281.
18. Han VZ, Grant G, Bell CC: Reversible associative depression
and nonassociative potentiation at a parallel fiber synapse.
Neuron 2000, 27:611-622.
19. Roberts PD, Bell CC: Computational consequences of
temporally asymmetric learning rules: II. Sensory image
cancellation. J Comput Neurosci 2000, 9:67-83.
20. Zhang Z, Bodznick D: The importance of N-methyl-D-aspartate
(NMDA) receptors in subtraction of electrosensory
reafference in the dorsal nucleus of skates. J Exp Biol 2010,
213:2700-2709.
21. Harvey-Girard E, Lewis J, Maler L: Burst-induced anti-Hebbian
 depression acts through short-term synaptic dynamics to
cancel redundant sensory signals. J Neurosci 2010, 30:6152-
6169.
This in vitro study provides the first characterization of long-term mechan-
isms in the ELL of weakly electric gymnotid fish, complementing previous
studious in the mormyrid ELL and the mouse DCN. Plasticity is NMDA
receptor-dependent and anti-Hebbian as in other cerebellum-like cir-
cuits, but is not spike timing-dependent. Also in contrast to findings in
other cerebellum-like circuits no mechanism for long-lasting potentiation
of parallel fiber synapses was found. The authors develop a model that
shows how interactions among parallel fiber synaptic dynamics, feedfor-
ward inhibition, and long-term depression could account for bi-directional
negative images observed in vivo.
22. Jorntell H, Hansel C: Synaptic memories upside down:
bidirectional plasticity at cerebellar parallel fiber-Purkinje cell
synapses. Neuron 2006, 52:227-238.
23. Lewis JE, Maler L: Dynamics of electrosensory feedback:
short-term plasticity and inhibition in a parallel fiber pathway.
J Neurophysiol 2002, 88:1695-1706.
24. Lewis JE, Lindner B, Laliberte B, Groothuis S: Control of neuronal
 firing by dynamic parallel fiber feedback: implications for
www.sciencedirect.com
electrosensory reafference suppression. J Exp Biol 2007,
210:4437-4447.
This in vitro study of the gymnotid ELL uses dynamic clamp and a model
of parallel fiber synaptic dynamics to demonstrate that the net effect of
parallel fiber activation on principal cell firing shifts from excitation to
inhibition as parallel fiber input rate increases. The authors suggest that
this shift could contribute to the formation of negative images appropriate
to cancel both predictable increases and decreases in principal cell firing.
25. Oertel D, Young ED: Whats a cerebellar circuit doing in the
auditory system? Trends Neurosci 2004, 27:104-110.
26. Kanold PO, Young ED: Proprioceptive information from the
pinna provides somatosensory input to cat dorsal cochlear
nucleus. J Neurosci 2001, 21:7848-7858.
27. Kanold PO, Davis KA, Young ED: Somatosensory context alters
 auditory responses in the cochlear nucleus. J Neurophysiol
2010, 105:1063-1070.
This study uses in vivo extracellular recordings and electrical stimulation
of the dorsal column nucleus to explore the interactions between auditory
stimuli and somatosensory input in the DCN of cats. A single stimulus
pulse evoked large and long-lasting effects on the responses of DCN
neuron to sound, consisting of both increases and decreases in evoked
firing. The authors provide evidence that inhibition from cartwheel cells
may mediate some of the observed effects.
28. Shore SE, Zhou J: Somatosensory influence on the cochlear
nucleus and beyond. Hear Res 2006, 216217:90-99.
29. Mancilla JG, Manis PB: Two distinct types of inhibition
mediated by cartwheel cells in the dorsal cochlear nucleus. J
Neurophysiol 2009, 102:1287-1295.
30. Roberts MT, Bender KJ, Trussell LO: Fidelity of complex spike-
mediated synaptic transmission between inhibitory
interneurons. J Neurosci 2008, 28:9440-9450.
31. Roberts MT, Trussell LO: Molecular layer inhibitory
 interneurons provide feedforward and lateral inhibition in the
dorsal cochlear nucleus. J Neurophysiol 2010, 104:2462-2473.
This elegant in vitro study of the mouse DCN uses dual recordings from
cartwheel and fusiform cells and minimal stimulation of parallel fibers to
explore the detailed interactions between these circuit elements. A
surprising finding of the study is that parallel fiber connectivity is extre-
mely sparse such that neighboring neurons rarely are activated by the
same parallel fiber input. This finding raises questions about the pre-
valence of true feedforward inhibition in the DCN.
32. Fujino K, Oertel D: Bidirectional synaptic plasticity in the
cerebellum-like mammalian dorsal cochlear nucleus. Proc Natl
Acad Sci U S A 2003, 100:265-270.
33. Tzounopoulos T, Kim Y, Oertel D, Trussell LO: Cell-specific, spike
timing-dependent plasticities in the dorsal cochlear nucleus.
Nat Neurosci 2004, 7:719-725.
34. Tzounopoulos T, Rubio ME, Keen JE, Trussell LO: Coactivation of
pre- and postsynaptic signaling mechanisms determines cell-
specific spike-timing-dependent plasticity. Neuron 2007,
54:291-301.
35. Zhao Y, Rubio ME, Tzounopoulos T: Distinct functional and
anatomical architecture of the endocannabinoid system in the
auditory brainstem. J Neurophysiol 2009, 101:2434-2446.
36. Zhao Y, Tzounopoulos T: Physiological activation of cholinergic
 inputs controls associative synaptic plasticity via modulation
of endocannabinoid signaling. J Neurosci 2011, 31:3158-3168.
This in vitro study shows that previously described timing-dependent
Hebbian plasticity (i.e. long-term potentiation) at parallel fiber synapses
onto fusiform cells in the mouse dorsal cochlear nucleus can be con-
verted to timing-dependent anti-Hebbian plasticity (i.e. long-term depres-
sion) if cholinergic inputs are activated during pairing. The paper provides
a striking example of how neuromodulation can alter rules for plasticity
and also strengthens the case for a sensory filtering function for the DCN
similar to those described previously in cerebellum-like circuits in fish.
37. Albus JS: A theory of cerebellar function. Math Biosci 1971,
10:25-61.
38. Marr D: A theory of cerebellar cortex. J Physiol (Lond) 1969,
202:437-471.
39. Arenz A, Bracey EF, Margrie TW: Sensory representations in
cerebellar granule cells. Curr Opin Neurobiol 2009, 19:445-451.
Current Opinion in Neurobiology 2011, 21:602608
444
608 Sensory and Motor Systems
40. Sawtell NB: Multimodal integration in granule cells as a basis
 for associative plasticity and sensory prediction in a
cerebellum-like circuit. Neuron 2010, 66:573-584.
This study of the mormyrid ELL uses in vivo whole-cell recordings to show
that individual granule cells integrate proprioceptive and motor corollary
signals conveyed by separate classes of mossy fibers. The study also
provides evidence that this multimodal integration allows for the genera-
tion of more specific negative images in principal cells than would be
possible if proprioceptive and corollary discharge signals were conveyed
separately. These findings provide the first experimental evidence for
multimodal integration in granule cells and suggest a function for this
integration that is in-line with longstanding MarrAlbus theories of cer-
ebellar function.
Current Opinion in Neurobiology 2011, 21:602608
41. Miall RC, Wolpert DM: Forward models for physiological motor
control. Neural Netw 1996, 8:1-15.
42. Bastian AJ: Learning to predict the future: the cerebellum
adapts feedforward movement control. Curr Opin Neurobiol
2006, 16:645-649.
43. Ebner TJ, Pasalar S: Cerebellum predicts the future motor
state. Cerebellum 2008, 7:583-588.
44. Miall RC, Weir DJ, Wolpert DM, Stein JF: Is the cerebellum a
smith predictor. J Motor Behav 1993, 25:203-216.
www.sciencedirect.com
445


446
 General and Comparative Endocrinology 157 (2008) 259265

Contents lists available at ScienceDirect

General and Comparative Endocrinology

journal homepage: www.elsevier.com/locate/ygcen

Minireview

Seasonal-like growth and regression of the avian song control system: Neural
and behavioral plasticity in adult male Gambels white-crowned sparrows
John Meitzen *,1, Christopher K. Thompson 1
Graduate Program in Neurobiology and Behavior, University of Washington, Box 356515, Seattle, WA 98195-6515, United States

a r t i c l e i n f o a b s t r a c t

Article history: Birdsong is regulated by a series of discrete brain nuclei known as the song control system. In seasonally-
Received 29 January 2008 breeding male songbirds, seasonal changes in steroid sex hormones regulate the structure and electro-
Revised 12 March 2008 physiology of song control system neurons, resulting in dramatic changes in singing behavior. Male song-
Accepted 17 March 2008
birds can be brought into the laboratory, where circulating levels of steroid hormone and photoperiod
Available online 25 March 2008
can be abruptly manipulated, providing controlled conditions under which rapid seasonal-like changes
in behavior and morphology can be carefully studied. In this mini-review, we discuss the steroidal and
Keywords:
cellular mechanisms underlying seasonal-like growth and regression of the song control system in adult
Testosterone
Estrogen
male Gambels white-crowned sparrows (Zonotrichia leucophrys gambelii), and its impact on song behav-
Androgen ior. Specically, we discuss recent advances concerning: (1) the role of androgen and estrogen receptors
Seasonal plasticity in inducing seasonal-like growth of the song control system; (2) how photoperiod modulates the time
Regression course of testosterone-induced growth of the song control system; (3) how bilateral intracerebral infu-
Growth sion of androgen and estrogen receptor antagonists near the song control nucleus HVC prevents sea-
Song control system sonal-like increases in song stereotypy but not song rate; and (4) the steroidal and cellular
Photoperiod mechanisms that mediate rapid regression of the song control system. Throughout this mini-review
we compare data collected from white-crowned sparrows to that from other songbird species. We con-
clude by outlining avenues of future research.
2008 Elsevier Inc. All rights reserved.

1. Introduction
Plasticity in brain and behavior is fundamental for animals to
react to changing environmental demands. An organisms ability
to adapt to seasonal changes in the environment is critical to
reproductive success, and breeding typically happens during the
season with the highest probability of successfully rearing off-
spring. It is thus not surprising that seasonal plasticity in brain
and behavior has been found in every vertebrate taxon (Tramontin
and Brenowitz, 2000). These seasonal changes in the environment
are often signaled to the brain via changes in hormone levels,
which can trigger massive restructuring of the neural substrates
that regulate behavior. In this mini-review we survey recent ad-
vances in understanding the proximate steroidal and cellular
mechanisms underlying an example of adult seasonal plasticity:
the seasonal growth and regression of the song control system
(Fig. 1) and the resulting changes in song (a reproduction-related
vocal behavior) in the male Gambels white-crowned sparrow.
White-crowned sparrows are age-limited learners (Marler and
Tamura, 1964) that as adults exhibit substantial seasonal changes
* Corresponding author. Fax: +1 206 543 5152.
E-mail address: jmeitzen@u.washington.edu (J. Meitzen).
1
These authors contributed equally to this work.
in the song control system and song behavior. Gambels white-
crowned sparrows are long-distance migrants that winter along
the west coast and western interior of the contiguous United
States, and breed in Alaska and Canada, unlike some other white-
crowned sparrow subspecies such as the non-migratory Nuttalls
(Farner and Lewis, 1973; Chilton et al., 1995). In the spring, longer
day lengths induce an increase in plasma testosterone (T) level,
which in turn triggers anatomical and electrophysiological changes
in the nuclei of the song control system and in white-crowned
sparrow song behavior, most notably increases in song rate, song
stereotypy, and duration (Nottebohm, 1981; Smith et al., 1995;
Tramontin et al., 2000; Brenowitz, 2004; Park et al., 2005; Meitzen
et al., 2007a). The volumes of the song control nuclei HVC, Area X,
and RA all increase, and the cellular mechanisms underlying these
increases vary by nucleus. The increase in HVC volume is largely
driven by an increase in neuron number (Fig. 2A), whereas growth
of RA (Fig. 2B) and Area X (Thompson and Brenowitz, 2005) is dri-
ven by changes in neuron size and density. The morphological
changes that underlie growth of the song control system nuclei
have been recently reviewed in detail elsewhere (Brenowitz,
2004), so we will not discuss this issue further.
Changes in the song control system and singing behavior can be
induced in the laboratory using the appropriate environmental and
hormonal cues, allowing carefully controlled experiments. To mi-

0016-6480/$ - see front matter 2008 Elsevier Inc. All rights reserved.
doi:10.1016/j.ygcen.2008.03.014

447
260 J. Meitzen, C.K. Thompson / General and Comparative Endocrinology 157 (2008) 259265

mic the spring and summer breeding season, we expose birds to

long-day (LD) photoperiod (20 h light, 4 h dark; typical of their
Alaskan breeding grounds) and a systemic T implant. The T implant
insures that plasma T levels are within the physiological breeding
range. To mimic the transition to non-breeding conditions, typi-
cally seen in the fall and winter, we expose birds to short-day
(SD) photoperiod (8 h light, 16 h dark). In addition, we remove
the subcutaneous T pellet and castrate the birds, to insure rapid
and synchronous withdrawal of circulating sex steroids. This sea-
sonal-like plasticity provides controlled conditions under which
proximate mechanisms underlying changes in behavior and mor-
phology can be carefully studied, which then informs eld studies
of non-captive animals (Just one example: Park et al., 2005; Meit-
zen et al., 2007b).

2. Mechanisms of growth
Fig. 1. Simplied schematic of the avian song control system showing the distri-
bution of steroid receptors. The nuclei HVC and RA comprise the main descending
motor circuit. RA projects to several brainstem nuclei (including nXIIts) that contain
2.1. Photoperiod increases plasma T level
motoneurons that project to the muscles controlling respiration and the sound
production organ, the syrinx. Both brainstem motoneurons and the syrinx express Seasonal-like plasticity in the song control system is modulated
androgen receptors. HVC, X, DLM, and LMAN comprise the anterior forebrain pat- by changes in photoperiod and circulating steroid sex hormones,
hway (AFP), which is necessary for song learning. HVC, used as a proper name; DLM,
with important contributions coming from both of these factors
the medial portion of the dorsolateral nucleus of the anterior thalamus; LMAN, the
lateral magnocellular nucleus of anterior nidopallium; RA, the robust nucleus of the (Fig. 3). Very early in the breeding season, increasing day length
arcopallium; nXIIts, the tracheosyringeal portion of the hypoglossal nucleus; X, activates the hypothalamicpituitary gonadal axis, which stimu-
Area X, a subdivision of the medial striatum. Nomenclature used here follows Re- lates growth of the testes and elevates circulating levels of T. This
iner et al. (2004).
increase in circulating T primarily regulates song control system

Fig. 2. Schematic, normalized representation of seasonal-like changes in morphology of HVC and RA. These representations are based upon the means of data taken from
Tramontin et al. (2000) and Thompson et al. (2007). We made these data sets congruent using the one time point shared between both sets (long-day photoperiod and
systemic testosterone (LD + T) for 20 days) and then normalized them to the short-day photoperiod (SD) time point. Birds were exposed to at least 10 weeks of SD
photoperiod in both studies to insure photosensitivity. (A) HVC grows within 7 days after the transition to LD + T, driven largely by changes in neuron number. After the
transition from LD + T to SD photoperiod, HVC volume regresses within 12 h, initially driven by a sharp increase in neuron density followed by a slower decrease in neuron
number. (B) RA volume increases signicantly within 20 days after the transition to LD + T, largely driven by an increase neuronal soma area and a decrease in density. After
the transition from LD + T to SD photoperiod, RA volume regresses signicantly by 20 days, with signicant changes in density and soma area by 2 days by continuing to
signicantly change over the entire time course.

448
 J. Meitzen, C.K. Thompson / General and Comparative Endocrinology 157 (2008) 259265
Fig. 3. Working model illustrating how steroid hormones induce changes in the
song control system and song behavior. Increasing photoperiod stimulates an inc-
rease in plasma testosterone (T) concentration. In the brain, T can be metabolized
into other androgens and 17-b estradiol (E2), an estrogen. The enzyme aromatase
converts T into E2. T and E2 act on HVC to increase neuron number and nucleus
volume. Steroid receptor binding in HVC is necessary and sufcient to create a
trophic signal that is transported by HVCs afferents to RA and X. Androgen receptor
binding in RA may be permissive for morphological and electrophysiological change
to occur in response to the trophic signal released by HVC. The mechanism trigg-
ering growth of the syrinx and nXIIts is unclear, but one possibility is direct acti-
vation of the androgen receptors present in both.
growth; T induces growth under both SD and LD photoperiod con-
ditions (Smith et al., 1997; Bernard et al., 1997; Gulledge and Devi-
che, 1997; Meitzen et al., 2007a; further reviewed in Brenowitz,
2004; for an alternate view see Ball et al., 2004), and T levels below
breeding maxima can induce growth of the song system (Tramon-
tin et al., 2001). LD photoperiod and systemic T induces song con-
trol system growth even when the bird has been deafened and
sings at a very low rate (Brenowitz et al., 2007).
2.2. Androgens and estrogens are both necessary for song control
system growth
Both androgen and estrogens may be necessary for maximal
growth of the song control system. In the brain, T can be metabo-
lized into other androgens, such as 5a-dihydrotestosterone (DHT),
or estrogens, such as 17-b estradiol (E2). These metabolites have
been implicated in song control system growth. In white-crowned
sparrows, systemic exposure to either DHT (a non-aromatizable
androgen) or E2 alone for 21 days results in signicant growth of
song control nuclei (Tramontin et al., 2003), but not RA spontane-
ous ring rate or soma area (Park et al., 2005; Meitzen et al.,
2007a). Systemic exposure to both an androgen and an estrogen,
however, produce full growth of song nuclei volumes and maximal
261
song rate (Tramontin et al., 2003), RA soma area, and RA spontane-
ous ring rate (Park et al., 2005; Meitzen et al., 2007a). Exposure to
the aromatase inhibitor ATD blocks the increase in RA soma area
and spontaneous ring rate induced by LD photoperiod and sys-
temic T (Meitzen et al., 2007a). Similarly, in wild, free-ranging male
song sparrows, exposure to the aromatase inhibitor fadrozole dur-
ing the breeding season decreased HVC volume (Soma et al., 2004).
Additional evidence comes from adult male canaries, where E2
promotes the survival of new neurons in HVC (Hidalgo et al.,
1995). In castrated male canaries exposed to SD photoperiod for
only 11 days, systemic T, but not DHT or E2 alone, increases HVC
volume and song rate (Sartor et al., 2005).
2.3. Androgens and estrogens act on HVC to stimulate growth of RA
and X
Studies using systemic hormone implants implicate both andro-
gens and estrogens in regulating growth of the song control nuclei
but provide no information on where the hormones act. The next
step in building a useful model of how T and its metabolites induce
growth of the song control system is to determine which steroid
receptors in which song system nuclei are necessary and sufcient
for T-induced growth. Experiments have focused on HVC because it
is the rst song control nucleus to grow in response to elevated
systemic T (Tramontin et al., 2000), and because it expresses estro-
gen receptors (ER) Fig. 1; (Gahr et al., 1993; Bernard et al., 1999;
Metzdorf et al., 1999; Fusani et al., 2000). Androgen receptors
(AR) or their mRNA are expressed in every song control nucleus
and the syrinx, the avian vocal organ Fig 1; (Arnold et al., 1976;
Gahr et al., 1993; Nastiuk and Clayton, 1995; Smith et al., 1996;
Bernard et al., 1999; Metzdorf et al., 1999; Soma et al., 1999b;
Fusani et al., 2000; Kim et al., 2004). AR immunoreactivity is ele-
vated in the HVC of Gambels white-crowned sparrows captured
in the spring; AR immunoreactivity in other song control nuclei
was not quantied (Soma et al., 1999b). Androgens and estrogens
could be activating activate different processes in different HVC
cell neuron types. Studies using steroid autoradiography and retro-
grade labeling found that X-projecting neurons accumulated estro-
gens (Johnson and Bottjer, 1995), and RA-projecting neurons
accumulated androgens (Johnson and Bottjer, 1993) in the canary.
In the zebra nch, both RA-projecting and X-projecting neurons
accumulated androgens (Sohrabji et al., 1989). Determining which
neuron types in HVC express ER and AR in the white-crowned
sparrow is a critical next step.
Several experiments validate the hypothesis that T and its
metabolites initially act on HVC, which then stimulates growth of
the other song control nuclei. Unilateral lesions of HVC in LD pho-
toperiod and systemic T-treated white-crowned sparrows block
growth of RA and X on the ipsilateral, but not contralateral, hemi-
sphere (Brenowitz and Lent, 2001). Small T implants unilaterally
implanted near HVC in SD photoperiod-treated white-crowned
sparrows increase the volumes of HVC, RA and X in the ipsilateral,
but not contralateral, hemisphere, while T implants near RA have
no effect (Brenowitz and Lent, 2002). Similarly, unilateral infusion
of DHT and E2 near HVC in SD photoperiod exposed male white-
crowned sparrows increases soma size and spontaneous ring rate
in the ipsilateral, but not contralateral, RA, but DHT and E2 infused
near RA have no effect (Meitzen et al., 2007a). Unilateral infusion of
androgen and estrogen receptor antagonists near HVC block the LD
photoperiod and systemic-T induced increase in soma size and
spontaneous ring rate in the ipsilateral, but not contralateral,
RA (Meitzen et al., 2007a). Together, these studies provide strong
evidence that androgens and estrogens act on HVC to stimulate
not only its own growth, but also that of RA and X. For a working
model (Fig. 3), we thus hypothesize that in HVC, AR and ER activa-
tion enhances incorporation of new RA-projecting neurons into
449
262 J. Meitzen, C.K. Thompson / General and Comparative Endocrinology 157 (2008) 259265
HVC (new neurons in HVC are thought to be RA-projecting, Scotto-
Lamasse et al., 2007) and increases synthesis of a transsynaptic tro-
phic signal that drives changes in intrinsic activity and morphology
in RA and Area X. The exact nature of the trophic signal that HVC
provides to RA and Area X is unclear, but it could involve neurotro-
phin release and/or changes in activity (further reviewed in Breno-
witz, 2004). Given that increases in Area X and RA nucleus volume
occur after HVC grows (Tramontin et al., 2000), the trophic signal
produced by HVC may either take time to be produced and trans-
ported, and/or new RA-projecting neurons may need to be re-
cruited to HVC and establish synapses on the increased dendritic
arbors of RA neurons (Hill and DeVoogd, 1991). AR in RA may be
permissive for RA neurons to respond to or amplify the incoming
trophic signal from HVC, as exposing RA neurons to the AR antag-
onist utamide prevents the increase in ring rate and soma size
induced by LD photoperiod and systemic T in three of ve birds
(Meitzen et al., 2007a). Further studies of the hypothesized permis-
sive role of AR in RA are required.
2.4. Do androgens act directly on nXIIts and the syrinx?
The steroidal mechanism underlying growth of the brainstem
motoneurons and of the syrinx (the avian vocal organ) is unclear,
although it may not depend upon activation of steroid hormone
receptors in either HVC or RA. When small intercerebral T implants
are unilaterally placed near either HVC or RA, neither nXIIts nor the
syrinx increase in volume or mass, respectively (Brenowitz and
Lent, 2002). Lesions of HVC prevented the growth of nXIIts (Breno-
witz and Lent, 2001). Since both nXIIts and the syrinx contain AR
(Smith et al., 1996), one reasonable hypothesis is that androgens
act individually and directly on them to increase growth. Other
hypotheses include but are not limited to: (1) retrograde support
of nXIIts growth from the syrinx, (2) afferent support of nXIIts from
either or both RA, and the dorsomedial portion of the intercollicu-
lar nucleus (ICo), which both contain AR, and (3) afferent support
of syringeal growth from nXIIts. These hypotheses remain to be
tested.
2.5. Photoperiod may modulate the growth of the song control system
Photoperiod might also modulate song control system growth
independently of stimulating the production of T by the testes (re-
cently reviewed in Ball et al., 2004; Brenowitz, 2004). In castrated
male starlings, American tree sparrows, and dark-eyed juncos and
Gambels white-crowned sparrows, LD photoperiod is sufcient to
promote at least partial growth of some song control system nuclei
(Bernard et al., 1997; Smith et al., 1997; Bentley et al., 1999; Dlo-
niak and Deviche, 2001). In castrated male house nches, LD pho-
toperiod is sufcient to induce growth of RA but not HVC (Strand
and Deviche, 2007). More recent data testing this hypothesis
comes from Gambels white-crowned sparrows, where LD photo-
period accelerates the systemic T-induced increase in spontaneous
ring rate and soma size in the nucleus RA (Meitzen et al., 2007a).
This acceleration can be blocked by systemic exposure to 1-4-6-
androstatrien-3,17-dione (ATD), an aromatase inhibitor; aroma-
tase is the enzyme that converts T into its estrogenic metabolite
E2 (Meitzen et al., 2007a). Conversely, exposure to systemic E2
accelerates the increase of spontaneous ring rates and soma size
in SD photoperiod and systemic T or DHT treated-birds (Park et al.,
2005; Meitzen et al., 2007a). These results suggest that photope-
riod might regulate aromatase expression or activity. Regional aro-
matase activity can change seasonally in starlings and song
sparrows (Riters et al., 2001; Soma et al., 2003), but has only been
measured in Gambels white-crowned sparrows captured in the
winter (Schlinger et al., 1992). Aromatase is positioned to have a
critical inuence on T-induced growth of the song control system.
In song sparrows (Melospiza melodia), a species closely related to
white-crowned sparrows (Carson and Spicer 2003), the nidopalli-
um surrounding HVC expresses aromatase mRNA, but HVC itself
does not (Soma et al., 2003), similar to other songbird species (For-
lano et al., 2006). Aromatase has been observed in pre-synaptic ter-
minals within HVC of zebra nches (Peterson et al., 2005).
Aromatase could thus modulate local concentrations of neuroster-
oids (reviewed in Forlano et al., 2006).
2.6. Changes in song stereotypy and song production may be
differentially regulated
LD photoperiod and systemic T induce dramatic changes in
Gambels white-crowned sparrow song behavior. These changes
include: an increase in song rate (Smith et al., 1995; Meitzen
et al., 2007a), an increase in syllable stereotypy (Smith et al.,
1995; Tramontin et al., 2000; Meitzen et al., 2007a), and an in-
crease in overall song length and buzz length (Smith et al., 1995).
White-crowned sparrows are closed-ended learners that do not
learn new syllables seasonally (Marler and Tamura, 1964), unlike
some other species (Brenowitz and Beecher, 2005). While numer-
ous studies have shown strong correlations between changes in
song behavior and song control nuclei, few have tested the causal
relationship between them. Recent studies have begun to provide
evidence regarding which seasonal changes in the song control
system mediate particular changes in song behavior. When song
system growth is blocked by infusing AR and ER antagonists near
HVC in white-crowned sparrows exposed to LD photoperiod and
systemic T, song stereotypy differs between treatment groups,
but song rate does not (Meitzen et al., 2007a). This result suggests
that activation of sex steroid receptors in HVC is not necessary for
the systemic T and LD photoperiod-induced increase in song rate,
and instead mediates changes in song stereotypy, and furthermore,
that increases in song rate are not sufcient for seasonal changes in
the song control system. In support of this, deafened white-
crowned sparrows greatly reduce song production, yet their song
control systems still grow and song becomes more stereotyped in
response to LD photoperiod and systemic T (Brenowitz et al.,
2007).
3. Mechanisms of regression
3.1. Transition to non-breeding conditions is not typically driven by
decreasing photoperiod
As detailed above, an increase in day length stimulates male
white-crowned sparrows to enter breeding condition, leading to
an increase in circulating T and growth of the song control system.
In contrast, much less is known about the mechanisms regulating
the transition from breeding to non-breeding conditions. In the
wild, once birds mate, establish nests, and start brooding, T levels
decline to basal levels within a matter of weeks (Wingeld and Far-
ner, 1978). This decline happens prior to the summer solstice while
photoperiods are still increasing. Thus, the transition to non-breed-
ing conditions is not actively driven by decreasing photoperiod. In-
stead, male white-crowned sparrows, like other seasonally-
breeding songbird species undergo absolute photorefractoriness;
their reproductive axis becomes insensitive to the stimulating ef-
fects of long-day photoperiods. This is in contrast to relative photo-
refractoriness, which is dependent upon a decrease in photoperiod.
The mechanisms regulating the transition to photorefractoriness
come from studies of various songbird species including Gambels
white-crowned sparrows (Dawson and Goldsmith, 1983; Wilson,
1985; Saldanha et al., 1994; Meddle et al., 1999). Photorefractori-
ness is characterized by a decrease of GnRH expression and/or re-
450
 J. Meitzen, C.K. Thompson / General and Comparative Endocrinology 157 (2008) 259265
lease, followed by a decrease in circulating levels of luteinizing
hormone and follicle stimulating hormone, which ultimately lead
to regression of the testes. The onset of photorefractoriness is
determined weeks earlier when individuals enter breeding condi-
tion in response to the initial exposure to LD photoperiod, given
that the transition to photorefractoriness is relative to the onset
of LD and largely independent of subsequent changes in photope-
riod (Dawson and Goldsmith, 1984; Bentley et al., 1997; Dawson,
2001; Dawson et al., 2001; Dawson, 2004). Ultimately, the de-
crease in circulating T removes steroidal trophic support for an en-
larged song control system, which then regresses via
neurodegenerative mechanisms.
It is not yet known if changes in photoperiod are sufcient to
induce steroid-independent regression of the song control system
in photostimulated males with high levels of circulating T. Despite
the fact that the onset of photorefractoriness is not dependent
upon a transition to SD photoperiod, we typically expose birds to
SD photoperiod at the same time circulating T is withdrawn to
insure that any stimulatory signal LD photoperiod may contribute
to maintaining a breeding-state song control system is removed
(Thompson et al., 2007). This manipulation is not the same as a
transition to photorefractoriness, because male white-crowned
sparrows are probably still photostimulated after just 34 weeks
of LD and high levels of circulating T. Yet the manipulation abruptly
removes two factors known to promote enlarged song control
system nuclei: high levels of circulating T and LD photoperiod.
The purpose of the manipulation is to remove as much exogenous
trophic support for the song control system as possible so that it
resembles that which is seen in birds held non-breeding conditions.
Regardless, there are many unanswered questions regarding the
steroid-independent contributions of photoperiod to song control
system plasticity.
3.2. Steroidal mechanism of regression
Acute withdrawal of circulating T induces a signicant decrease
in HVC volume within 12 h (Thompson et al., 2007). This decrease
occurs before the birds experience a change in photoperiod, which
demonstrates that, at the very least, the initial regression of HVC
volume is driven largely by the withdrawal of circulating sex ste-
roids and not by physiological changes driven by photoperiod. This
suggests that regression, like growth, of the song control system in
male white-crowned sparrows is largely driven by changes in cir-
culating levels of sex steroids.
3.3. Changes in neuronal morphology that underlie song control
system regression
In general, long-term changes in HVC volume result from
changes in neuron number. The initial decrease in HVC volume
within 12 h of T withdrawal, however, results from an increase in
neuron density (Fig. 2A, Thompson et al., 2007). HVC essentially
collapses in on itself as the space between neurons decreases. Over
the next four days, neuron number and soma area (a measure of
neuron size) signicantly decrease (Thompson et al., 2007). At this
point, the rate of neuron loss offsets the decrease in neuron spac-
ing, and neuron density signicantly decreases. Thus, the increase
in neuron density lasts only a few days.
The downstream nuclei of HVC, RA and Area X, regress more
slowly than HVC, taking days to weeks instead of hours to days.
The volumes of Area X and RA are not signicantly regressed until
seven and 20 days, respectively, following the transition to non-
breeding conditions (Thompson et al., 2007). RA soma area and
neuron density signicantly regress by two days after the transi-
tion to non-breeding conditions and continue to regress for at least
20 days (Fig. 2B, Thompson et al., 2007). Thus the neuronal mech-
263
anism mediating regression is the same as that underlying the
growth of RA: changes in soma area and neuron density with no
change in neuron number.
3.4. Neurodegenerative mechanisms of HVC regression and steroidal
neuroprotection
In male white-crowned sparrows, the loss of HVC neurons dur-
ing song control system regression is mediated by caspase-depen-
dent cell death pathways. Apoptotic-like processes are to be
expected, given that HVC neuron number decreases by nearly
30% following a seasonal-like transition to non-breeding condi-
tions (Thompson et al., 2007). HVC cells positive for activated cas-
pase-3 are apparent three days after the transition to non-breeding
conditions (accepted for publication, C.K. Thompson). In addition,
infusion of a cocktail of caspase inhibitors near HVC in male Gam-
bels white-crowned sparrows transitioned to non-breeding condi-
tions seems to prevent a decrease in HVC volume, neuron number,
neuron spacing, and soma area (accepted for publication, C.K.
Thompson).
T and/or its metabolites act within HVC to both induce and
maintain its seasonal growth (discussed above). This suggests that
T and/or its metabolites could play a neuroprotective role in rescu-
ing HVC from the regression induced by the withdrawal of sys-
temic T and exposure to SD photoperiod, not unlike the
hormone-mediated neuroprotection that is seen in several in vivo
animal models of neurodegenerative insult (Ramsden et al. 2003;
Pike et al. 2006). Consistent with this, we found that direct intrace-
rebral infusion of T near HVC unilaterally in castrated male white-
crowned sparrows transferred to SD photoperiod and systemic T-
withdrawal seems to ameliorate neurodegeneration of the ipsilat-
eral HVC (unpublished data, C.K. Thompson). These results suggest
that T and/or its metabolites directly act on HVC neurons to protect
them from degenerative mechanisms induced by withdrawal of
circulating sex steroids and photoperiod shift. Seasonal-like rapid
regression of the song control system therefore may serve as an
excellent model to further elucidate the molecular mechanisms
that underlie hormone-mediated neuroprotection.
3.5. Behavior
Male white-crowned sparrows stop singing immediately fol-
lowing the withdrawal of circulating T (Meitzen et al., unpublished
data). This observation strongly suggests that T is rapidly cleared
from the blood stream. Though a time course for the clearance of
circulating T following castration is not known in male white-
crowned sparrows, intramuscular injections of highly concentrated
T propionate dissolved in peanut oil into male song sparrows is
cleared within 90 min (Soma et al., 1999). In male house nches,
removal of subcutaneous T pellets results in signicantly reduced
circulating levels of T 24 h later (Deviche et al., 2006). If a similar
time course for clearance of circulating T applies to white-crowned
sparrows, it would suggest that abrupt cessation of singing in
males is driven by the rapid withdrawal of T independent of
photoperiod.
4. Future directions
There are many unanswered questions about seasonal-like
plasticity of the song control system in male white-crowned spar-
rows, and many ways that the model presented here could be fur-
ther tested, rened, and expanded. In addition to the questions
raised above, outstanding questions include: (1) what are the
downstream molecular cascades that are turned on or off by
changes in circulating sex steroids? (2) Do the various neuron
451
264 J. Meitzen, C.K. Thompson / General and Comparative Endocrinology 157 (2008) 259265
types in HVC (RA-projecting, Area X-projecting, and interneurons)
and RA (glutamatergic projection neurons and GABAergic inter-
neurons) contribute to song control system growth and regres-
sion in different ways? (3) How does the absence of
metabolites of T contribute to regression of the song control sys-
tem? (4) How do rates of proliferation, migration, and incorpora-
tion of new neurons into HVC change during seasonal-like growth
and regression? (5) What other electrophysiological properties, in
addition to the ones described here, change during the growth
and regression of the song control nuclei? (6) What changes occur
in the syrinx during the transition to breeding conditions and to
non-breeding conditions? (7) How similar are the proximate
mechanisms in Gambels white-crowned sparrow seasonal plas-
ticity to other steroid-sensitive models? These and other ques-
tions illustrate that study of seasonal plasticity within the avian
song control system will continue to serve as a useful model for
hormone-mediated neural plasticity.
Acknowledgments
We thank Eliot A. Brenowitz and David J. Perkel for their men-
torship and support. We also thank Eliot A. Brenowitz and two
anonymous reviewers for comments that improved this manu-
script. J.M. and C.K.T. made equal contributions to this work. Grant
Sponsor: NIH: MH53032 (E.A.B), MH068530 (D.J.P.); 5 T32
GM07108 (training grant supporting J.M. and C.K.T.)
References
Arnold, A.P., Nottebohm, F., Pfaff, D.W., 1976. Hormone concentrating cells in vocal
control and other areas of the brain of the zebra nch (Poephila guttata). J.
Comp. Neurol. 165, 487511.
Ball, G.F., Auger, C.J., Bernard, D.J., Charlier, T.D., Sartor, J.J., Riters, L.V., Balthazart, J.,
2004. Seasonal plasticity in the song control system: multiple brain sites of
steroid hormone action and the importance of variation in song behavior. Ann.
NY Acad. Sci. 1016, 586610.
Bentley, G.E., Goldsmith, A.R., Dawson, A., Glennie, L.M., Talbot, R.T., Sharp, P.J.,
1997. Photorefractoriness in European starlings (Sturnus vulgaris) is not
dependent upon the long-day-induced rise in plasma thyroxine. Gen. Comp.
Endocrinol. 107, 428438.
Bentley, G.E., Vant Hof, T.J., Ball, G.F., 1999. Seasonal neuroplasticity in the songbird
telencephalon: a role for melatonin. Proc. Natl. Acad. Sci. USA 96, 46744679.
Bernard, D.J., Wilson, F.E., Ball, G.F., 1997. Testis-dependent and -independent
effects of photoperiod on volumes of song control nuclei in American tree
sparrows (Spizella arborea). Brain Res. 760, 163169.
Bernard, D.J., Bentley, G.E., Balthazart, J., Turek, F.W., Ball, G.F., 1999. Androgen
receptor, estrogen receptor alpha, and estrogen receptor beta show distinct
patterns of expression in forebrain song control nuclei of European starlings.
Endocrinology 140, 46334643.
Brenowitz, E.A., 2004. Plasticity of the adult avian song control system. Ann. NY
Acad. Sci. 1016, 560585.
Brenowitz, E.A., Beecher, M.D., 2005. Song learning in birds: diversity and plasticity,
opportunities and challenges. Trends Neurosci. 28, 127132.
Brenowitz, E.A., Lent, K., 2001. Afferent input is necessary for seasonal growth and
maintenance of adult avian song control circuits. J. Neurosci. 21, 23202329.
Brenowitz, E.A., Lent, K., 2002. Act locally and think globally: intracerebral
testosterone implants induce seasonal-like growth of adult avian song control
circuits. Proc. Natl. Acad. Sci. USA 99, 1242112426.
Brenowitz, E.A., Lent, K., Rubel, E.W., 2007. Auditory feedback and song production
do not regulate seasonal growth of song control circuits in adult white-crowned
sparrows. J. Neurosci. 27, 68106814.
Carson, R.J., Spicer, G.S., 2003. A phylogenetic analysis of the emberizid sparrows
based on three mitochondrial genes. Mol. Phylogenet. Evol. 29, 4357.
Chilton, G., Baker, M.C., Barrentine, C.D., Cunningham, M.A., 1995. White-crowned
sparrow (Zonotrichia leucophrys). In: Poole, A., Gill, F. (Eds.), The Birds of North
America, No. 183. The Academy of Natural Sciences, Philadelphia, PA, and The
American Ornithologists Union, Washington, DC.
Dawson, A., 2001. The effects of a single long photoperiod on induction and
dissipation of reproductive photorefractoriness in European starlings. Gen.
Comp. Endocrinol. 121, 316324.
Dawson, A., 2004. Evidence against a period of relative photorefractoriness during
the recovery of photosensitivity in common starlings. Gen. Comp. Endocrinol.
136, 117121.
Dawson, A., Goldsmith, A.R., 1983. Plasma prolactin and gonadotrophins during
gonadal development and the onset of photorefractoriness in male and female
starlings (Sturnus vulgaris) on articial photoperiods. J. Endocrinol. 97, 253
260.
Dawson, A., Goldsmith, A.R., 1984. Effects of gonadectomy on seasonal-changes in
plasma-LH and prolactin concentrations in male and female starlings (Sturnus
vulgaris). J. Endocrinol. 100, 213218.
Dawson, A., King, V.M., Bentley, G.E., Ball, G.F., 2001. Photoperiodic control of
seasonality in birds. J. Biol. Rhythms 16, 365380.
Deviche, P., Martin, R.K., Small, T., Sharp, P.J., 2006. Testosterone induces testicular
development but reduces GnRH-I ber density in the brain of the house nch,
Carpodacus mexicanus. Gen. Comp. Endocrinol. 147, 167174.
Dloniak, S.M., Deviche, P., 2001. Effects of testosterone and photoperiodic condition
on song production and vocal control region volumes in adult male dark-eyed
juncos (Junco hyemalis). Horm. Behav. 39, 95105.
Farner, D.S., Lewis, R.A., 1973. Annual cycles of white-crowned sparrows. J. Reprod.
Fert. Suppl. 19, 3550.
Forlano, P.M., Schlinger, B.A., Bass, A.H., 2006. Brain aromatase: new lessons from
non-mammalian model systems. Front. Neuroendocrinol. 27, 247274.
Fusani, L., Vant Hof, T., Hutchison, J.B., Gahr, M., 2000. Seasonal expression of
androgen receptors, estrogen receptors, and aromatase in the canary brain in
relation to circulating androgens and estrogens. J. Neurobiol. 43, 254268.
Gahr, M., Gttinger, H.R., Kroodsma, D.E., 1993. Estrogen receptors in the avian
brain: survey reveals general distribution and forebrain areas unique to
songbirds. J. Comp. Neurol. 327, 112122.
Gulledge, C.C., Deviche, P., 1997. Androgen control of vocal control region volumes
in a wild migratory songbird (Junco hyemalis) is region and possibly age
dependent. J. Neurobiol. 32, 391402.
Hidalgo, A., Barami, K., Iversen, K., Goldman, S.A., 1995. Estrogens and non-
estrogenic ovarian inuences combine to promote the recruitment and
decrease the turnover of new neurons in the adult female canary brain. J.
Neurobiol. 27, 470487.
Hill, K.M., DeVoogd, T.J., 1991. Altered daylength affects dendritic structure in a
song-related brain region in red-winged blackbirds. Behav. Neural. Biol. 56,
240250.
Johnson, F., Bottjer, S.W., 1993. Hormone-induced changes in identied cell
populations of the higher vocal center in male canaries. J. Neurobiol. 24, 400
418.
Johnson, F., Bottjer, S.W., 1995. Differential estrogen accumulation among
populations of projection neurons in the higher vocal center of male canaries.
J. Neurobiol. 26, 87108.
Kim, Y.H., Perlman, W.R., Arnold, A.P., 2004. Expression of androgen receptor mRNA
in zebra nch song system: developmental regulation by estrogen. J. Comp.
Neurol. 469, 535547.
Marler, P., Tamura, M., 1964. Culturally transmitted patterns of vocal behavior in
sparrows. Science 146, 14831486.
Meddle, S.L., Maney, D.L., Wingeld, J.C., 1999. Effects of N-methyl-D-aspartate on
luteinizing hormone release and fos-like immunoreactivity in the male white-
crowned sparrow (Zonotrichia leucophrys gambelii). Endocrinology 140, 5922
5928.
Meitzen, J., Moore, I.T., Lent, K., Brenowitz, E.A., Perkel, D.J., 2007a. Steroid hormones
act transsynaptically within the forebrain to regulate neuronal phenotype and
song stereotypy. J. Neurosci. 27, 1204512057.
Meitzen, J., Perkel, D.J., Brenowitz, E.A., 2007b. Seasonal changes in intrinsic
electrophysiological activity of song control neurons in wild song sparrows. J.
Comp. Physiol. A 193, 677683.
Metzdorf, R., Gahr, M., Fusani, L., 1999. Distribution of aromatase, estrogen receptor,
and androgen receptor mRNA in the forebrain of songbirds and nonsongbirds. J.
Comp. Neurol. 407, 115129.
Nastiuk, K.L., Clayton, D.F., 1995. The canary androgen receptor mRNA is localized in
the song control nuclei of the brain and is rapidly regulated by testosterone. J.
Neurobiol. 26, 213224.
Nottebohm, F., 1981. A brain for all seasons: cyclical anatomical changes in song
control nuclei of the canary brain. Science 214, 13681370.
Park, K.H., Meitzen, J., Moore, I.T., Brenowitz, E.A., Perkel, D.J., 2005. Seasonal-like
plasticity of spontaneous ring rate in a songbird pre-motor nucleus. J.
Neurobiol. 64, 181191.
Peterson, R.S., Yarram, L., Schlinger, B.A., Saldanha, C.J., 2005. Aromatase is pre-
synaptic and sexually dimorphic in the adult zebra nch brain. Proc. Biol. Sci.
272, 20892096.
Pike, C.J., Rosario, E.R., Nguyen, T.V.V., 2006. Androgens, aging, and Alzheimers
disease. Endocrine 29, 233241.
Ramsden, M., Shin, T.M., Pike, C.J., 2003. Androgens modulate neuronal vulnerability
to kainate lesion. Neuroscience 122, 573578.
Reiner, A., Perkel, D.J., Bruce, L.L., Butler, A.B., Csillag, A., Kuenzel, W., Medina, L.,
Paxinos, G., Shimizu, T., Striedter, G., Wild, M., Ball, G.F., Durand, S., Guturkun,
O., Lee, D.W., Mello, C.V., Powers, A., White, S.A., Hough, G., Kubikova, L.,
Smulders, T.V., Wada, K., Dugas-Ford, J., Husband, S., Yamamoto, K., Yu, J., Siang,
C., Jarvis, E.D., 2004. Revised nomenclature for avian telencephalon and some
related brainstem nuclei. J. Comp. Neurol. 473, 377414.
Riters, L.V., Baillien, M., Eens, M., Pinxten, R., Foidart, A., Ball, G.F., Balthazart, J.,
2001. Seasonal variation in androgen-metabolizing enzymes in the
diencephalon and telencephalon of the male European starling (Sturnus
vulgaris). J. Neuroendocrinol. 13, 985997.
Saldanha, C.J., Deviche, P.J., Silver, R., 1994. Increased VIP and decreased GnRH
expression in photorefractory dark-eyed juncos (Junco hyemalis). Gen. Comp.
Endocrinol. 93, 128136.
Sartor, J.J., Balthazart, J., Ball, G.F., 2005. Coordinated and dissociated effects of
testosterone on singing behavior and song control nuclei in canaries (Serinus
canaria). Horm. Behav. 47, 467476.
452
 J. Meitzen, C.K. Thompson / General and Comparative Endocrinology 157 (2008) 259265
Schlinger, B.A., Slotow, R.H., Arnold, A.P., 1992. Plasma estrogens and brain
aromatase in winter white-crowned sparrows. Ornis Scand. 23, 292297.
Scotto-Lomassese, S., Rochefort, C., Nshdejan, A., Scharff, C., 2007. HVC interneurons
are not renewed in adult male zebra nches. Eur. J. Neurosci. 25, 16631668.
Smith, G.T., Brenowitz, E.A., Wingeld, J.C., Baptista, L.F., 1995. Seasonal changes in
song nuclei and song behavior in Gambels white-crowned sparrows. J.
Neurobiol. 28, 114125.
Smith, G.T., Brenowitz, E.A., Prins, G.S., 1996. Use of PG-21 immunocytochemistry to
detect androgen receptors in the songbird brain. J. Histochem. Cytochem. 44,
10751080.
Smith, G.T., Brenowitz, E.A., Wingeld, J.C., 1997. Roles of photoperiod and
testosterone in seasonal plasticity of the avian song control system. J.
Neurobiol. 32, 426442.
Sohrabji, F., Nordeen, K.W., Nordeen, E.J., 1989. Projections of androgen-
accumulating neurons in a nucleus controlling avian song. Brain Res. 488,
253259.
Soma, K.K., Hartman, V.N., Wingeld, J.C., Brenowitz, E.A., 1999. Seasonal changes in
androgen receptor immunoreactivity in the song nucleus HVc of a wild bird. J.
Comp. Neurol. 409, 224236.
Soma, K.K., Schlinger, B.A., Wingeld, J.C., Saldanha, C.J., 2003. Brain aromatase, 5
alpha-reductase, and 5 beta-reductase change seasonally in wild male song
sparrows: relationship to aggressive and sexual behavior. J. Neurobiol. 2003 56,
209221.
Soma, K.K., Tramontin, A.D., Featherstone, J., Brenowitz, E.A., 2004. Estrogen
contributes to seasonal plasticity of the adult avian song control system. J.
Neurobiol. 58, 413422.
265
Strand, C.R., Deviche, P., 2007. Hormonal and environmental control of song control
region growth and new neuron addition in adult male house nches,
Carpodacus mexicanus. Dev. Neurobiol. 67, 827837.
Thompson, C.K., Brenowitz, E.A., 2005. Seasonal change in neuron size and spacing
but not neuronal recruitment in a basal ganglia nucleus in the avian song
control system. J. Comp. Neurol. 481, 276283.
Thompson, C.K., Bentley, G.E., Brenowitz, E.A., 2007. Rapid seasonal-like regression
of the adult avian song control system. Proc. Natl. Acad. Sci. USA 104, 15520
15525.
Tramontin, A.D., Brenowitz, E.A., 2000. Seasonal plasticity in the adult brain. Trends
Neurosci. 23, 251258.
Tramontin, A.D., Hartman, V.N., Brenowitz, E.A., 2000. Breeding conditions induce
rapid and sequential growth in adult avian song circuits: a model of seasonal
plasticity in the brain. J. Neurosci. 20, 854861.
Tramontin, A.D., Perto, N., Wingeld, J.C., Brenowitz, E.A., 2001. Seasonal growth of
song control nuclei precedes seasonal reproductive development in wild adult
song sparrows. Gen. Comp. Endocrinol. 122, 19.
Tramontin, A.D., Wingeld, J.C., Brenowitz, E.A., 2003. Androgens and estrogens
induce seasonal-like growth of song nuclei in the adult songbird brain. J.
Neurobiol. 57, 130140.
Wilson, F.E., 1985. An androgen-independent mechanism maintains
photorefractoriness in male tree sparrows (Spizella arborea). J. Endocrinol.
107, 137143.
Wingeld, J.C., Farner, D.S., 1978. Annual cycle of plasma IRLH and steroid-
hormones in feral populations of the white-crowned sparrow, Zonotrichia-
leucophrys-gambelii. Biol. Reprod. 19, 10461056.
453


454

Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
ANNUAL
REVIEWS Further
Click here for quick links to
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
Annual Reviews content online,
including:
Other articles in this volume
Top cited articles
Top downloaded articles
Our comprehensive search
Annu. Rev. Neurosci. 2008. 31:124
First published online as a Review in Advance on
February 14, 2008
The Annual Review of Neuroscience is online at
neuro.annualreviews.org
This articles doi:
10.1146/annurev.neuro.30.051606.094225
Copyright  c 2008 by Annual Reviews.
All rights reserved
0147-006X/08/0721-0001$20.00
Cerebellum-Like Structures
and Their Implications for
Cerebellar Function
Curtis C. Bell,1 Victor Han,2
and Nathaniel B. Sawtell1
1
Neurological Sciences Institute, Oregon Health and Science University,
Beaverton, Oregon 97006; email: bellc@ohsu.edu, sawtelln@ohsu.edu
2
Oregon Regional Primate Center, Oregon Health and Science University,
Beaverton, Oregon 97006; email: hanv@ohsu.edu
Key Words
forward model, synaptic plasticity, electric sh, cerebellum
Abstract
The nervous systems of most vertebrates include both the cerebellum
and structures that are architecturally similar to the cerebellum.
The cerebellum-like structures are sensory structures that receive
input from the periphery in their deep layers and parallel ber input
in their molecular layers. This review describes these cerebellum-
like structures and compares them with the cerebellum itself. The
cerebellum-like structures in three groups of sh act as adaptive
sensory processors in which the signals conveyed by parallel bers in
the molecular layer predict the patterns of sensory input to the deep
layers through a process of associative synaptic plasticity. Similarities
between the cerebellum-like structures and the cerebellum suggest
that the cerebellum may also generate predictions about expected
sensory inputs or states of the system, as suggested also by clinical,
experimental, and theoretical studies of the cerebellum. Understanding
the process of predicting sensory patterns in cerebellum-like structures
may therefore be a source of insight into cerebellar function.
1
455
 parallel bers together with the dendrites and
Contents cell bodies on which the bers terminate. The
parallel bers are numerous and closely packed.
LOCAL CIRCUITRY, GENE
The granule cells that give rise to the paral-
EXPRESSION, AND
lel bers in cerebellum-like structures are mor-
EVOLUTION OF
phologically similar to cerebellar granular cells
CEREBELLUM-LIKE
(Mugnaini et al. 1980a,b) but are usually located
STRUCTURES . . . . . . . . . . . . . . . . . . . 2
in an external granule cell mass rather than in
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

General Features. . . . . . . . . . . . . . . . . . . 2
a granule cell layer beneath the molecular layer
Local Circuitry of Different
as in the cerebellum. Unipolar brush cells and
Cerebellum-Like Structures . . . . . 3
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Golgi cells similar to those present in the gran-

Comparison of the Local Circuitries
of Cerebellum-Like Structures
and the Cerebellum . . . . . . . . . . . . .
Patterns of Gene Expression in
Cerebellum-Like Structures
and the Cerebellum . . . . . . . . . . . . .
Evolution of Cerebellum-Like
Structures and the Cerebellum . .
PREDICTIONS AND PLASTICITY
IN CEREBELLUM-LIKE
STRUCTURES AND THE
CEREBELLUM . . . . . . . . . . . . . . . . . .
Predictions and Plasticity in
Cerebellum-Like Structures . . . . .
Predictions in the Cerebellum . . . . . .
DIRECTIONS FOR FUTURE
RESEARCH . . . . . . . . . . . . . . . . . . . . . .
Activity Patterns in Granule Cells . . .
Adaptive Filtering in Electrosensory
Systems . . . . . . . . . . . . . . . . . . . . . . . .
Adaptive Filtering in the DCN
and Less-Studied
Cerebellum-Like Structures . . . . .
Purkinje-Like Cells . . . . . . . . . . . . . . . .
Primitive Cerebellums . . . . . . . . . . . . .
LOCAL CIRCUITRY, GENE
EXPRESSION, AND EVOLUTION
OF CEREBELLUM-LIKE
STRUCTURES
General Features
A distinctive molecular layer is a key identi-
fying feature of all cerebellum-like structures
(Figure 1). The molecular layer is composed of
ular layer of the cerebellum are also present
in some cerebellum-like structures (Campbell
9
et al. 2007, Mugnaini et al. 1997).
Functionally, the parallel bers convey a rich
variety of information from other central struc-
10
tures, which includes corollary discharge in-
formation associated with motor commands,
11
information from higher levels of the same sen-
sory modality represented in the deep layers,
and information from other sensory modalities.
In general, the types of signals conveyed by par-
11
allel bers are signals that are likely to be asso-
ciated with changes in the sensory input to the
11
deep layers and that can therefore serve to pre-
15
dict such sensory input (predictive inputs in
Figure 1).
16
The parallel bers terminate on the den-
16
dritic spines of principal cells and on the smooth
dendrites of inhibitory stellate cells in a man-
17
ner very similar to the termination of paral-
lel bers on Purkinje cells and molecular layer
interneurons of the cerebellum. We use the
17
term principal cells to refer to large cells with
17
spine-covered dendrites that extend through-
17
out the molecular layer. Some of these prin-
cipal cells are excitatory efferent cells that
project to higher levels of the sensory system,
whereas others are inhibitory neurons that ter-
minate locally on each other and on the effer-
ent cells. The latter are sometimes referred to
as Purkinje-like. The cell bodies of principal
cells are usually located in a separate layer be-
low the molecular layer, like the Purkinje cell
layer of the cerebellum.
Afferent input from the periphery termi-
nates in the deep layers of cerebellum-like
structures, on basilar dendrites of principal

2 Bell
Han Sawtell
456
 Predictive inputs
Corollary discharge signals
Higher levels of the same modality
Other sensory modalities Granule layer
(e.g. proprioception)
???
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Molecular layer
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Principal cell layer

Sensory input layer

Input from a sensory surface

Figure 1
Schematic drawing showing major features of cerebellum-like sensory structures. Inhibitory stellate cells of
the molecular layer are shown in black. Blue upward arrows indicate afferent input from the periphery
terminating in the sensory input layer. In some cerebellum-like structures the afferent input also terminates
on the smooth proximal portion of the apical dendrites as indicated by the small blue arrowheads.

cells, on proximal apical dendrites of principal
cells, or on interneurons that relay the informa-
tion from the periphery to the principal cells.
Some of the interneurons of the deep layers
are inhibitory, allowing for a change of sign,
whereby excitation in the periphery is con-
verted into inhibition of some principal cells.
The peripheral input to the deep layers forms a
map of a sensory surface, such as the skin sur-
face, the retina, or the cochlea.
This review describes major features of the dif-
ferent cerebellum-like structures of craniates
but is not exhaustive. Recent reviews (Bell 2002,
Bell & Maler 2005, Montgomery et al. 1995)
and the original papers on individual structures,
as provided below, should be consulted for more
complete descriptions. Some of the structures
are also much better known than others, which
is reected in the level of detail in the following
descriptions.

Medial octavolateral nucleus. The medial

Local Circuitry of Different
Cerebellum-Like Structures
The brains of all major groups of craniates
except reptiles and birds have cerebellum-like
structures (Figures 2 and 3). The similari-
ties among the different cerebellum-like struc-
tures are clear, but so are the differences. Dif-
ferent structures may have different types of
cells in addition to the principal cells, stel-
late cells, and granule cells that are present in
all cerebellum-like structures. Moreover, some
structures have additional inputs besides the in-
puts from the periphery and the parallel bers.
octavolateral nucleus (MON) processes pri-
mary afferent input from the mechanical lateral
line system and, in some sh, from eighth nerve
end organs (Bell 1981b, McCormick 1999). It
is present in all basal aquatic craniates with me-
chanical lateral line sensory systems (Figures 2,
3ad, 4a). Myxinoids (atlantic hagsh; C.B.
Braun, personal communication) and aquatic
amniotes (reptiles, birds, and mammals; Mont-
gomery et al. 1995) do not have lateral line sys-
tems and do not have an MON.
MON: medial
The efferent cells of the MON extend their
octavolateral nucleus
spiny apical dendrites up into a molecular

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 3

457
 CBM MON DON OTML ELL RLN DCN

Myxinoidea Myxinoids ?
Eptatretidae ?
Petromyzontiformes
Lampetra ?
Elasmobranchii
Chondrichthyes
Holocephalii
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Chondrosteii
Holosteii
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Neopterygii Osteoglossomorpha
Gnathostomata

Actinopterygii
Vertebrata
Craniata

Elopomorpha
Teleosteii Clupeomorpha
Euteleostei
Dipneustii
Crossopterygii
Sarcopterygii

Urodela
Tetrapoda Amphibia
Anura
Apoda ? ?
Reptilia
Aves
Mammalia

Figure 2
Distribution of cerebellum-like structures and the cerebellum in different craniate groups. A lled circle means the structure is present
in all or almost all the members of that group. A lled half circle means the structure is present only sporadically in that group. A
question mark means that presence of the structure in that group is controversial. CBM, cerebellum; DCN, dorsal cochlear nucleus;
DON, dorsal octavolateral nucleus; ELL, electrosensory lobe; MON, medial octavolateral nucleus; OTML, marginal layer of the optic
tectum; RLN, rostrolateral nucleus of thalamus.

layer known as the cerebellar crest (Figure 3a
d ). The parallel bers of the cerebellar crest
descend from an anterior granule cell mass
known as the lateral granular mass in elas-
mobranchs and the eminentia granularis in
other sh. The inputs to these granule cells in-
clude lateral line primary afferents (Bodznick
& Northcutt 1980), eighth nerve primary affer-
ents (Puzdrowski & Leonard 1993), input from
the spinal cord (Schmidt & Bodznick 1987), and
descending input from higher-order lateral line
and acoustic centers (Bell 1981c, McCormick
1997, Tong & Finger 1983). The basilar den-
DON: dorsal drites of MON efferent cells are affected by pri-
octavolateral nucleus
mary afferent input.
Dorsal octavolateral nucleus (DON). The
dorsal octavolateral nucleus (DON) processes
primary afferent input from electroreceptors
and is present in many basal vertebrates with
an electrosense (Figures 2, 3a) Electrorecep-
tion is a vertebrate sense that may have orig-
inated as early as the lateral line or vestibular
senses (Bullock et al. 1983). The Myxinoidea
do not have electroreceptors and do not have
a DON (Ronan 1986). Electroreception was
lost during the evolution of neopterygian bony
sh, and these sh do not have a DON. Elec-
troreception reappeared independently at least
twice during the evolution of the teleost ra-
diation: once during the evolution of the two

4 Bell
Han Sawtell
458
 a b c d
Many aquatic Mormyrid Gymnotid Most aquatic
vertebrates fish fish vertebrates
DON ELL ELL MON
MON MON MON

p CC
EG
DG R
CB CB
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

CC CB MON
CB
p
EG
DO
N

nAll
ELL
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

ELL
CC CC CC nAll
MON
MON
nVIII
MON

e f g h
Ray finned Some bony Most Most
fish fish mammals vertebrates
OTML RLN DCN Cerebellum
TL
Climbing
Tel CB fibers

N
RL DCN
CB

VCN
TS Opt tr

Molecular layer Granule cell mass Sensory input map

Figure 3
Cerebellum-like structures in different vertebrate groups. The molecular layer, granule cell mass, and sensory input map are shown in
different colors, as indicated at the bottom of the gure. The climbing ber input to the cerebellum is shown here as a sensory input (see
text). CB, cerebellum; CC, cerebellar crest; DCN, dorsal cochlear nucleus; DGR, dorsal granular ridge; DON, dorsal octavolateral
nucleus; EGp, eminentia granularis posterior; ELL, electrosensory lobe; gran, granular layer; MON, medial octavolateral nucleus; mol,
molecular layer; nAll, anterior lateral line nerve; nVIII, eighth nerve; Opt tr, optic tract; RLn, rostrolateral nucleus; Tel, telencephalon;
TL, torus longitudinalis; TS, torus semicircularis; VCN, ventral cochlear nucleus.

related groups, Mormyriformes and Xeno-
mystinae, and a second time during the evolu-
tion of the other two related groups, Gymno-
tiformes and Siluriformes (Bullock et al. 1983).
However, the more recently derived electrore-
ceptors and associated electrosensory central
structures of teleosts are quite different from
those of other aquatic vertebrates (see elec-
trosensory lobe below).
The DON is located just dorsal to the MON
and is similar to the MON in its structure
and connections. Primary afferent input from
electroreceptors terminates on the basilar den-
drites of efferent cells and inhibitory neurons
of the deep layers, as in the MON (Bodznick &
Northcutt 1980, Puzdrowski & Leonard 1993).
The spine-covered apical dendrites of efferent
cells extend up into the overlying cerebellar
crest.
Parallel bers of the DON cerebellar
crest arise from the dorsal granular ridge,
which receives proprioceptive input, recurrent

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 5

459
 a b c
MON Mormyrid ELL Gymnotid ELL

Parallel fibers Parallel fibers Parallel fibers

(LGR, EGp) (EGp) (EGp)

PE PE
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

To torus To PE, torus To PE, torus

by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Corollary
discharge

Lateral line Electroreceptor Electroreceptor

afferents afferents afferents

d e f
DCN Teleost cerebellum Mammalian cerebellum
Parallel fibers
(GCD) Parallel fibers Parallel fibers

To IC To CN

Climbing Climbing
fiber fiber
Auditory
afferents

Figure 4
Local circuits of some cerebellum-like structures, the teleost cerebellum, and the mammalian cerebellum. Granule cells and parallel
bers are in red, afferent input from the periphery is in blue, and the additional inputs to the mormyrid and gymnotid ELLs are in
green. The Purkinje-like cells of the mormyrid ELL and mammalian DCN as well as the Purkinje cells of the teleost and mammalian
cerebellums are black. Excitatory efferent cells are white. IC, inferior colliculus; CN, cerebellar nucleus.

electrosensory input, and corollary discharge
input associated with motor commands
(Bodznick & Boord 1986, Conley & Bodznick
1994, Hjelmstad et al. 1996). All three types
of input are active in relation to the shs
respiratory cycle. Electroreceptors in elasmo-
branchs are strongly affected by the shs own
respiration (Montgomery & Bodznick 1993).
The activity in parallel bers can therefore
be used to predict the effect of these cyclic
OTML: marginal changes on electroreceptive input to the deep
layer of the optic
layers of DON (see Adaptive Processing in
tectum
Cerebellum-Like Structures, below).
Marginal layer of the optic tectum. The op-
tic tectum of actinopterygian (ray-nned) shes
is distinctive in that its outer layers are cerebel-
lum like (Figures 2, 3e) (Meek 1983, Vanegas
et al. 1979). The external layer of the optic tec-
tum in these sh is a molecular layer known
as the optic tectum marginal layer (OTML).
The cell bodies of principal cells, the type I
neurons of Meek (1983), are located below
the marginal layer. The type I neurons ex-
tend their spine-covered apical dendrites up
into the marginal layer and input from the
retina maps onto their basilar dendrites and

6 Bell
Han Sawtell
460
 the smooth proximal portions of their apical
dendrites.
The parallel bers of the marginal layer
arise from a medially located granule cell mass
known as the torus longitudinalis. Granule cells
of the torus longitudinalis respond to corol-
lary discharge signals associated with the motor
commands that evoke eye movements and re-
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
spond to visual stimuli as well (Northmore et al.
1983). Parallel ber activity driven by corol-
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
lary discharge signals associated with eye move-
ments could predict changes in retinal input to
the deep layers, a possible interaction between
the two types of input similar to that described
above for the DON.
Electrosensory lobe (ELL). Electrorecep-
tion is present in four groups of teleosts:
Mormyriformes, an order of electric sh from
Africa; Gymnotiformes, a superorder of elec-
tric sh from South America; Siluriformes, the
order of catsh; and Xenomystinae, an African
subfamily of the family Notopteridae (Bullock
& Heiligenberg 1986). All these sh have a
cerebellum-like electrosensory lobe (ELL) that
receives primary afferent input from electrore-
ceptors (Figures 2, 3b,c) (Bell & Russell 1978,
Braford 1982, Finger & Tong 1984, Maler et al.
1981).
The Mormyriformes and Gymnotiformes
are electric sh with electric organs as well
as electroreceptors. The order Mormyri-
formes includes the family Mormyridae, all of
which have electric organ discharges (EODs)
that are brief and pulse like, and the single-
species family Gymnarchidae, which has a con-
tinuous wave-like EOD. The order Gymno-
tiformes includes some families with wave-like
EODs and other families with pulsatile EODs.
The ELLs of pulsatile mormyrids and wave
gymnotids have been studied most extensively,
although some work has been done on the ELLs
of wave mormyriforms (Kawasaki & Guo 1998,
Matsushita & Kawasaki 2005) and pulse gym-
notiforms (Caputi et al. 2002, Schlegel 1973).
The spine-covered apical dendrites of ELL
principal cells extend up into the overlying
molecular layer. Primary afferent bers from
www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function
electroreceptors in the skin map onto the deep
layers, terminating on the basilar dendrites of
principal cells or on interneurons (Bell & Maler
ELL: electrosensory
2005). The ELL efferent cells of mormyrid lobe
(Bell et al. 1997b), gymnotid (Saunders & Bas-
EOD: electric organ
tian 1984), and silurid sh (McCreery 1977) are discharge
of two main types: E-cells, which are excited by
EOCD: electric organ
an increase in peripheral stimulus strength in
corollary discharge
the center of their receptive elds, and I-cells,
which are inhibited by such an increase. These
two functionally distinct cell types are also mor-
phologically distinct; the E-cells have more ex-
tensive basilar dendrites.
Parallel bers of ELLs arise from gran-
ule cells of the eminentia granularis posterior
(EGp), which in mormyrids, at least, also con-
tains Golgi cells and unipolar brush cells similar
to the same cell types in the mammalian cere-
bellum (Campbell et al. 2007). The inputs to
EGp in mormyrid and gymnotid sh include
proprioceptive signals associated with bending
of the body or the ns, recurrent electrosen-
sory input from a higher levels of the system,
and in mormyrids only, a corollary discharge
signal associated with the motor command that
elicits the electric organ (corollary) discharge
(EOCD) (Bastian & Bratton 1990; Bell et al.
1992; Carr & Maler 1986; Szabo et al. 1979,
1990). These different inputs to EGp are re-
layed to ELL as parallel ber inputs, where they
can predict changes in electroreceptor input to
the deep layers associated with tail movements,
some other electrosensory input, or the EOD
(see Adaptive Processing in Cerebellum-Like
Structures).
The mormyrid (Bell et al. 1981), gym-
notid (Carr & Maler 1986), and silurid (Tong
1982) ELLs receive additional input aside
from the peripheral and parallel ber inputs.
They receive direct recurrent input from a
higher-order electrosensory nucleus just ros-
tral to ELL, the nucleus preeminentialis dor-
salis (PE) (Figures 4b,c). The deep layers of the
mormyrid ELL also receive EOCD input di-
rectly from an EOD motor commandrelated
nucleus (Bell & von der Emde 1995). This in-
put is in addition to the EOCD input conveyed
via parallel bers.
7
461
 The ELLs of mormyrid and gymnotid sh
have differences as well as similarities. Most im-
portant, the mormyrid ELL includes a prin-
RLN: rostrolateral
nucleus of the cipal cell that is not present in the gymnotid
thalamus ELL (Figures 4b,c), the medium ganglion cell
DCN: dorsal cochlear (Meek et al. 1996). These cells are referred to as
nucleus Purkinje-like because they are GABAergic with
extensive spine-covered dendrites in the overly-
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
ing molecular layer. However, they differ from
Purkinje cells because they have basilar den-
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
drites and do not receive climbing ber input.
The medium ganglion cells are interneurons
that inhibit both nearby efferent cells and each
other (Figure 4b). They are more numerous
than the efferent cells and have many more den-
drites and spines in the molecular layer (Meek
et al. 1996). They must therefore have a central
role in the integration of peripheral and paral-
lel ber inputs in the mormyrid ELL. These
and other differences between the mormyrid
and gymnotid ELLs are consistent with their
independent evolutionary origins.
Rostrolateral nucleus of the thalamus. The
rostrolateral nucleus (RLN) (Figures 2, 3f ) of
the thalamus is a small, cerebellum-like struc-
ture found in the thalamus of a few widely
scattered neopterygian sh (Figure 2) (Butler
& Saidel 1992). The principal cells of RLN
receive topographically organized direct in-
put from the retina on the smooth proximal
parts of their apical dendrites. The more dis-
tal apical dendrites are covered with spines
and receive parallel ber input from the torus
longitudinalis.
Dorsal cochlear nucleus. All mammals
possess a dorsal cochlear nucleus (DCN)
(Figures 2, 3g, 4d ). The DCN is laminated
and cerebellum-like in marsupials and euthe-
rian mammals but not in monotremes (Cant
1992, Nieuwenhuys et al. 1997). Fusiform
cells are the major efferent cell type of the
DCN. Their basilar dendrites are contacted
by primary afferent bers from the cochlea,
which form a topographic map of the cochlea
in the deeper layers below the molecular layer.
The fusiform cells extend their spine-covered
8 Bell
Han Sawtell
apical dendrites up into the molecular layer
where they are contacted by parallel bers.
The parallel bers arise from granule cells
located around the margins of the nucleus.
The parallel bers course at right angles to
the isofrequency bands in the deeper layers.
Thus, parallel bers cross through different
frequency-specic regions of DCN.
The cartwheel cell is a second type of princi-
pal cell in the DCN (Cant 1992, Nieuwenhuys
et al. 1997). These cells are Purkinje-like
because they are GABAergic, have extensive
spine-covered dendrites in the molecular layer,
and inhibit the efferent fusiform cells. The cell
bodies of cartwheel cells are in the molecular
layer, and their dendrites are restricted to the
molecular layer.
The local circuits of the DCN and the
mormyrid ELL are very similar to the local
circuit of the cerebellar cortex in actinopteryr-
ian sh where most Purkinje cells are in-
terneurons that terminate locally on efferent
cells (Figure 4e) (Finger 1978, Meek 1998).
The parallel bers of the DCN, the mormyrid
ELL, and the actinopterygian cerebellum pass
through and excite the dendrites of both
efferent cells and Purkinje or Purkinje-like
cells. In all three cases, the Purkinje cells or
Purkinje-like cells inhibit nearby efferent cells
(Figures 4b,d,e). The efferent neurons of the
actinopterygian cerebellum are equivalent to
the cerebellar nucleus neurons of mammals
(Figure 4f ).
The granule cells of the DCN receive var-
ious types of input: recurrent auditory input
from the inferior colliculus (Caicedo & Herbert
1993) and auditory cortex (Weedman & Ryugo
1996); primary vestibular afferent input (Burian
& Gstoettner 1988); input from the pontine nu-
clei (Ohlrogge et al. 2001); somatosensory in-
put from the dorsal column nuclei (Weinberg
& Rustioni 1987), the trigeminal nuclei (Zhou
& Shore 2004), and the somatosensory cortex
(Wolff & Kunzle 1997); and direct input from
the cochlea via ne unmyelinated Type II affer-
ents (Brown et al. 1988). DCN granule cells
also receive input from brainstem nuclei as-
sociated with vocalization and respiration that
462
 may convey corollary discharge signals (Shore
& Zhou 2006). Proprioceptive input from the
pinna has particularly strong effects on DCN
granule cells in the cat (Kanold & Young 2001).
Movements of the animals pinna, head, or body
have predictable effects on how the cochlea
responds to an external sound source, and an
animals own vocalization and respiration will
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
have predictable consequences on auditory in-
put. Thus the signals conveyed by the parallel
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
bers in the DCN molecular layer could gen-
erate predictions about changes in afferent ac-
tivity from the cochlea that arrive at the deep
layers, as in other cerebellum-like structures.
Comparison of the Local Circuitries
of Cerebellum-Like Structures
and the Cerebellum
Many similarities in cell types and local cir-
cuitry between the cerebellum and cerebellum-
like structures have been described in the pre-
ceding section. The similar cellular elements
include the granule cells, the Golgi cells, the
unipolar brush cells, the parallel bers, the stel-
late cells, and the spine-covered molecular layer
dendrites of principal cells.
The most crucial similarity is that between
the two inputs to cerebellum-like structures
and the two inputs to cerebellar Purkinje
cells. Cerebellum-like structures receive paral-
lel ber and peripheral input, whereas Purkinje
cells of the cerebellum receive parallel ber in-
put and climbing ber input. In both cases, one
input, the parallel bers, conveys a rich variety
of information to an entire set of principal cells
or Purkinje cells. In both cases, a second input
peripheral input for cerebellum-like structures
and climbing ber input for the cerebellum
conveys specic information that subdivides the
set of Purkinje cells that share the same parallel
ber input.
Olivary input to Purkinje cells is more spe-
cic than the peripheral input to the deep lay-
ers of cerebellum-like structures insofar as it is
conveyed by just a single climbing ber. Effer-
ent cells and Purkinje-like cells in cerebellum-
like structures do not have such single ber in-
www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function
puts. The cerebellums of different vertebrates
can vary markedly, but all the cerebellums that
have been closely examined have a specic in-
put from the inferior olive that terminates as
climbing bers. We suggest that the presence
of a climbing ber is the dening characteris-
tic of the cerebellum that distinguishes it from
cerebellum-like structures.
Climbing bers and the peripheral sensory
input to cerebellum-like structures are similar
in many respects. Climbing bers signal rather
specic sensory events in most of the cases
where the information they convey has been
identied. Such sensory signals include retinal
slip in a particular direction (Maekawa & Simp-
son 1972), somatosensory stimulation within a
small region of skin (Ekerot & Jorntell 2001,
Robertson 1985), and vestibular stimulation
with tilt in a particular direction (Barmack &
Shojaku 1992). Moreover, the climbing bers of
vertebrates other than mammals do not termi-
nate throughout the molecular layer as in mam-
mals. They terminate instead on smooth, prox-
imal dendrites at the base of the molecular layer
(Nieuwenhuys et al. 1997) in a manner similar
to that of retinal input onto the smooth, prox-
imal dendrites of principal cells in the OTML
and RLN. This is not to say that the inferior
olive is a simple sensory relay. It is not. But
clearly sensory stimuli have a strong inuence
on the inferior olive and on climbing bers, a
result consistent with the origin of the inferior
olive from the embryos alar or sensory plate.
Devor (2002) has in fact suggested that the in-
ferior olive has been interposed between pe-
ripheral sensory structures and the cerebellum
to gate sensory signals by motor commands and
by the inferior olives own intrinsic rhythmicity.
As noted in the previous section, the parallel
bers of cerebellum-like structures convey in-
formation that is associated with sensory input
changes to the deep layers and that can there-
fore predict such changes. The parallel bers
of the cerebellum similarly convey information
that can predict the occurrence of climbing
ber input. Climbing bers in the occu-
lonodular lobe of the mammalian cerebel-
lum, for example, signal retinal slip (Maekawa
9
463
 & Simpson 1972), and the parallel bers in
this region convey vestibular information about
head movement (Lisberger & Fuchs 1974),
LTD: long-term
depression corollary discharge information about eye
movement (Noda & Warabi 1982), and propri-
oceptive information from the neck (Matsushita
& Tanami 1987), all of which could be used to
predict movement of an image on the retina.
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
The presence of a climbing ber is perhaps
the critical difference between the cerebellum
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
and cerebellum-like structures. Other differ-
ences include the presence of basilar dendrites
on most principal cells of cerebellum-like struc-
tures but not on Purkinje cells; the presence of
planar dendritic trees in most Purkinje cells but
not in most principal cells; the presence of cell
types in cerebellum-like structures not present
in the cerebellum; and the presence of other in-
puts besides parallel bers and climbing bers
in cerebellum-like structures not present in the
cerebellum, such as the preeminential input in
electroreceptive teleosts (Figures 4b,c).
Patterns of Gene Expression
in Cerebellum-Like Structures
and the Cerebellum
Similarities and differences between the differ-
ent cerebellum-like structures and the cerebel-
lum itself are also revealed in gene expression
patterns. Some genes are expressed in many
different cerebellar and cerebellum-like struc-
tures, whereas others are expressed in only a
few of these structures (Bell 2002). Common
patterns of gene expression between cerebellar
Purkinje cells and cartwheel cells of the DCN
are particularly prominent, and many muta-
tions affect both cell types (Berrebi et al. 1990).
One gene, the GluRdelta2 gene, may be ex-
pressed in most if not all cerebellum-like struc-
tures and also in the cerebellum, but not in other
structures. This gene is structurally related to
the ionotropic glutamate receptors but does
not form ion channels (Yuzaki 2003). The gene
is necessary for long-term depression (LTD)
at the parallel ber to Purkinje cell synapse
(Yawata et al. 2006). In mammals, the GluR-
delta2 gene is expressed in Purkinje cells (Yuzaki
10 Bell
Han Sawtell
2003) and in the principal cells of the DCN
(Petralia et al. 1996). In zebrash, the GluR-
delta2 gene is expressed in the molecular layers
of the cerebellum, the MON, and the OTML,
but not elsewhere in the brain as shown for
both the gene and the protein (Mikami et al.
2004). Similarly, in the mormyrid brain, the
GluRdelta2 protein is present in the molecular
layers of the cerebellum, the ELL, the MON,
and the OTML, but not elsewhere in the brain
( J. Zhang & C. Bell, unpublished observa-
tions). Expression of the GluRdelta2 gene in still
other cerebellum-like structures remains to be
established.
Some genes are expressed in some of the
cerebellum-like structures or the cerebellum in
the adult but are expressed only in other such
structures during development. The zebrin II
gene, for example, is expressed only in Purkinje
cells in adult mammals, birds, and sh (Hawkes
& Herrup 1995, Lannoo et al. 1991) but is ex-
pressed transiently during development in the
MON and in part of the ELL of gymnotid
sh (Lannoo et al. 1992). Similarly, functional
N-methyl-D-aspartate (NMDA) receptors are
present on principal cells of the adult mormyrid
and gymnotid ELLs (Grant et al. 1998, Berman
et al. 2001), as well as principal cells of the
adult DCN (Manis & Molitor 1996), but are
present on cerebellar Purkinje only during de-
velopment (Dupont et al. 1987).
The common features in the local circuitry
and in the gene expression patterns suggest
the presence of a shared genetic-developmental
program in all craniates, a program that
once activated can generate a cerebellum or
cerebellum-like structure. Some ndings from
experimental embryology support this idea.
Thus, ectopic cerebellum-like structures de-
velop in the forebrain or midbrain of a chick
embryo if beads are coated with broblast
growth factor 8 and placed at those sites in
the embryo (Martinez et al. 1999). Similarly,
cerebellar tissue will develop ectopically in the
midbrain and forebrain of a mouse embryo
with a genome that is Otx1+/- and Otx2+/-
(Drosophila orthodenticle protein, a transcrip-
tion factor) (Acampora et al. 1997).
464
 Evolution of Cerebellum-Like
Structures and the Cerebellum
The similarities between all the different
cerebellum-like and cerebellar structures can-
not be explained solely by homology in the
sense of historical or phylogenetic homology
(Butler & Saidel 2000). In this usage of the term,
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
a feature is considered homologous across dif-
ferent taxa if the taxa have inherited the fea-
ture from a common ancestor that also had the
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
feature. However, some of the individual struc-
tures described here are homologous. Thus the
most parsimonious explanation for the pres-
ence of a cerebellum in all vertebrates is that
it was present in a common ancestor. A com-
mon ancestor is also the most parsimonious
explanation for the presence of an MON, a
DON, or a DCN in some groups of craniates.
However, we nd no evidence for an ancestral
cerebellum-like structure from which the cere-
bellum, MON, DON, marginal layer of the tec-
tum, ELL, RLN, and DCN all evolved. (See
Bell 2002 for a more complete analysis of the
evolution of cerebellum-like structures.)
How then can we explain the clear simi-
larities among the different cerebellums and
cerebellum-like structures? The best explana-
tion may be the presence of a developmental-
genetic program that can generate a cerebellum
or cerebellum-like structure, as described pre-
viously, together with evolutionary pressure for
the type of information processing that these
structures can perform.
Cerebellum-like structures may have
evolved before the cerebellum itself. An MON
is clearly present in some myxinoids, and both
an MON and a DON are clearly present in
lampreys, but the presence of a cerebellum is
not well established in either of these groups.
Some comparative anatomists afrm the
presence of a cerebellum in myxinoids (Larsell
1967), whereas others deny it (Nieuwenhuys
et al. 1997), and arguments have also been
made both for (Larsell 1967, Nieuwenhuys
et al. 1997) and against (Crosby 1969) the
presence of a cerebellum in lampreys. As
suggested previously, the identication of
www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function
climbing bers on putative Purkinje cells
could indicate the presence of a cerebellum,
but no efforts to identify climbing bers
have been made in myxinoids and lampreys.
Purkinje cellspecic markers that do not stain
cerebellum-like structures could also help
determine the presence of a cerebellum. Thus
the nding that the Zebrin II antibody does
not stain cells in what some consider to be the
lamprey cerebellum is of interest (Lannoo &
Hawkes 1997) but is not conclusive because the
Zebrin II antibody does not stain all Purkinje
cells.
PREDICTIONS AND PLASTICITY
IN CEREBELLUM-LIKE
STRUCTURES AND
THE CEREBELLUM
Predictions and Plasticity in
Cerebellum-Like Structures
Cerebellum-like structures process informa-
tion from peripheral sensory receptors in com-
bination with an array of central signals con-
veyed by parallel bers. If a common function
exists among all cerebellum-like structures, it
must involve the interaction between these two
types of inputs. Progress toward understanding
these interactions has been made in cerebellum-
like structures concerned with the processing
of electrosensory information in three distinct
groups of sh: elasmobranchs, gymnoti-
form teleosts, and mormyrid teleosts. The
cerebellum-like structures of these sh act as
adaptive lters, removing predictable features
of the sensory input (for reviews, see Bastian
& Zakon 2005, Bell 2001, Bell et al. 1997a).
In these systems, the animals own behav-
ior strongly affects electroreceptors and could
interfere with sensing weak electrosensory sig-
nals from the environment. In the passive elec-
trosensory system of elasmobranch sh, for ex-
ample, ventilatory movements modulate the
shs standing bioelectric eld and can drive
electroreceptor afferents through their en-
tire dynamic range (Montgomery & Bodznick
11
465

Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
12
1999). In the active electrosensory systems of
mormyrid and gymnotid sh, movements of the
electric organ (located in the tail) relative to
sensory surface cause large changes in EOD-
evoked electroreceptor input that could over-
whelm the small changes resulting from nearby
objects.
Parallel ber inputs to cerebellum-like
structures involved in electrolocation convey
proprioceptive, corollary discharge, and elec-
trosensory signals that could be used to pre-
dict the electrosensory consequences of the an-
imals own behavior. Direct evidence for the
generation of such predictions has been ob-
tained from in vivo recordings from princi-
pal cells in the mormyrid and gymnotid ELL
and elasmobranch DON (Bastian 1996a, Bell
1981a, Bell et al. 1997b, Bodznick et al. 1999).
In each case, pairing articial electrosensory
stimuli with central predictive signalsa corol-
lary discharge signal at a particular delay after
the EOD motor command in the case of the
mormyrid ELL (Figure 5a), a proprioceptive
signal at a particular tail angle in the case of the
gymnotid ELL (Figure 5b), and a propriocep-
tive or corollary discharge signal at a particular
phase of the ventilatory cycle in the case of the
elasmobranch DON (Figure 5c)results in a
change in the response to the predictive sig-
nals alone that resembles a negative image of
the response to the previously paired (and now
predicted) stimulus. The negative images de-
velop rapidly over the course of a few minutes
of pairing and are specic to the sign as well
as to the spatial and temporal patterns of activ-
ity evoked by the stimulus. On the basis of these
results investigators suggested that cerebellum-
like circuitry could operate as an adaptive lter
by continually generating and updating sensory
predictions on the basis of associations between
central signals and current sensory inputs and
subtracting these predictions from the neural
response. Adaptive ltering could thus allow
external electrosensory signals to be detected
more easily.
Several lines of evidence conrm that for-
mation of negative images is due, at least in
large part, to plastic changes occurring within
Bell
Han Sawtell
the cerebellum-like structures themselves (Bell
2001). Pairing predictive signals with intracel-
lular current injections in vivo results in the
formation of negative images in principal cells
in all three groups of sh, indicating that the
inputs to the recorded cell are plastic (Bastian
1996b, Bell et al. 1993, Bodznick et al. 1999).
Given the types of predictive signals involved
in negative image formation, synapses between
parallel bers and principal cells are the most
natural candidates for the site of plastic changes.
Negative image formation requires that the
plasticity be anti-Hebbian in character, i.e., cor-
relations between pre- and postsynaptic ac-
tivity should decrease synaptic strength, and
researchers have obtained evidence for anti-
Hebbian plasticity at parallel ber synapses with
principal cells in all three classes of sh. Anti-
Hebbian plasticity at parallel ber synapses has
also been shown recently in the DCN of mam-
mals (Fujino & Oertel 2003, Tzounopoulos
et al. 2004) but has not yet been connected to
systems-level adaptive ltering.
Modeling studies have helped to link the
properties of negative image formation with
mechanisms of synaptic plasticity (Nelson &
Paulin 1995, Roberts 1999, Roberts & Bell
2000). Temporal specicity is a key feature of
negative image formation. In the mormyrid
ELL, parallel bers convey corollary discharge
signals related to the motor command that
drives the EOD. Pairing with electrosensory
stimuli at various delays relative to the motor
command results in negative images that are
specic to the paired delay (Bell 1982). Re-
sults of modeling studies suggest that tempo-
rally specic negative images could be gener-
ated using an anti-Hebbian learning rule similar
to that observed experimentally (see below) to-
gether with an array of parallel ber inputs
active at different delays following the mo-
tor command (Roberts 1999, Roberts & Bell
2000). The mechanisms for generating tem-
porally specic negative images in this model
are quite similar to those proposed for some
forms of cerebellar learning, such as the learn-
ing of adaptively timed responses in classical
eye-blink conditioning (Medina et al. 2000) or
466
 a b c
Mormyrid Gymnotid Elasmobranch
ELL ELL DON
EOD Ventilation
Ex In
Command Tail bend alone
alone alone before
before before
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

EOD
0 min
(9 min of C+S pairing) Command
plus
stimulus 1.5 min
Tail bend
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

plus 3 min Ventilation

stimulus plus
6 min stimulus

15 min
EOD
Command 25 min
2 min

alone
after Tail bend
alone
after 2 min Ventilation
4 min alone
after
8 min
0 ms 160 20 0 +20 0 20
0 1
EOD
Command Tail displacement () s

Figure 5
Formation of negative images of predicted sensory responses in three different cerebellum-like structures. (a) Raster display of the
responses of a cell in the ampullary region of the mormyrid ELL. Each dot represents an action potential. The EOD motor command
occurs at time 0. The command alone initially has no effect on the cell. An electrosensory stimulus (vertical black line) that evokes a
pause-burst response is then paired with the command. After several minutes of pairing, the stimulus is turned off and a response to
command alone is revealed, which was not present before the pairing and which is a negative image of the previously paired sensory
response. From Bell 1986. (b) Raster display of responses of cell in the gymnotid ELL. The tail is moved back and forth passively. Each
row of dots shows response to one movement cycle. Initially the tail bend has no effect on the cell. An electrosensory stimulus that
evokes a burst-pause is then delivered in phase with the movement. The electrosensory stimulus is turned off after several minutes of
pairing, which reveals a response to tail bending alone that was not present before the pairing and which is opposite to the previously
paired sensory response. From Bastian 1995. (c) Histogram display of responses of a cell in the elasmobranch DON. Initially the cell
does not respond to the exhalation (Ex)inhalation (In) ventilatory cycle of the sh (top histogram). An electrosensory stimulus that
evokes a burst-pause is then delivered in phase with the ventilatory cycle. The response to ventilation plus the electrosensory stimulus
decreases during 25 min of pairing. Turning off the electrosensory stimulus after pairing reveals the presence of a response to ventilation
alone, which was not present before and which is a negative image of the previously paired sensory response. From Bodznick 1993.

of appropriate phase relations in the vestibular
ocular reex (Raymond & Lisberger 1998).
The cellular properties of anti-Hebbian
synaptic plasticity have been studied in some
detail at synapses between parallel bers and
Purkinje-like medium ganglion cells in an in
vitro preparation of the mormyrid ELL (Bell
et al. 1997c, Han et al. 2000). Synaptic depres-
sion requires a postsynaptic dendritic spike and
depends on the precise timing of the spike rela-
tive to the parallel ber evoked excitatory post-
synaptic potential (EPSP) onset. Depression
develops when a postsynaptic dendritic spike
occurs within 50 ms of EPSP onset, whereas
other timing relations yield potentiation or
no effect. Potentiation as measured in vitro is
nonassociative and likely depends on simple
repetition of the parallel ber stimuli at a suf-
ciently high rate, although in vivo experiments
suggest a spike timingdependent component

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 13

467

Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
14
to the potentiation (Bell et al. 1997b, Sawtell
et al. 2007). The depression requires activa-
tion of NMDA receptors and changes in post-
synaptic calcium. The potentiation can reverse
the depression and vice versa, with both po-
tentiation and depression having a presynaptic
locus of expression. Plasticity at parallel ber
synapses onto Purkinje-like cartwheel cells of
the DCN is also anti-Hebbian, spike timing
dependent, NMDA dependent, and presynap-
tically expressed (Tzounopoulos et al. 2004,
2007).
Investigators have observed both similar-
ities and differences between plasticity in
cerebellum-like structures and plasticity in the
cerebellum itself. The depression of responses
to signals conveyed by parallel bers following
the pairing of these signals with postsynaptic
excitation in cerebellum-like structures is sim-
ilar to the depression of responses to parallel
ber stimulation in the mammalian Purkinje
cells following pairing with climbing ber input
or with postsynaptic depolarization (Ito 2001).
Such depression has been linked to the forma-
tion of negative images of predicted sensory in-
put in cerebellum-like structures and to mo-
tor learning in the mammalian cerebellum (Ito
1984). It is of interest in this regard that the
timing of stimulus-driven parallel berevoked
simple spike activity is consistently close to the
inverse of climbing ber responses in almost all
the systems where this relation has been exam-
ined (Barmack & Shojaku 1992, Ebner et al.
2002, Graf et al. 1988, Kobayashi et al. 1998,
Stone & Lisberger 1990). Thus in many sys-
tems, simple spike activity is a kind of negative
image of predicted climbing ber activity. Plas-
ticity at parallel ber synapses may play a role in
generating the antiphase relation, but it is only
part of the explanation because the antiphase
relation is still present when parallel ber LTD
is blocked (Goossens et al. 2004).
The timing constraints on parallel ber plas-
ticity may be more restrictive in cerebellum-
like structures than in the cerebellum. LTD in
the cerebellum-like structures where timing re-
lations have been tested occurred only when
the postsynaptic spike followed the presynap-
Bell
Han Sawtell
tic spike by 50 ms or less (Bell et al. 1997c,
Tzounopoulos et al. 2004). In the cerebellum,
however, depression of the parallel ber synapse
is present after pairings with climbing ber in-
put in which delays varied between occurrence
of the climbing ber 50 ms before the paral-
lel ber stimulus and occurrence of the climb-
ing ber 200 ms after the parallel ber stimulus
(Safo & Regehr 2007, Wang et al. 2000).
The mechanisms of synaptic plasticity are
clearly not the same in the cerebellum and in
the cerebellum-like structures where it has been
studied. Plasticity at parallel ber synapses onto
efferent or Purkinje-like cells in the mormyrid
ELL (Han et al. 2000) and the mammalian
DCN (Tzounopoulos et al. 2004) depends on
activation of NMDA receptors, but synap-
tic plasticity at parallel ber synapses onto
Purkinje cells does not (Ito 2001). However,
some aspects of the plasticity mechanisms may
be shared as indicated by the presence of
the GluRdelta2 gene in the cerebellum and in
cerebellum-like structures, and by the involve-
ment of this gene in plasticity at Purkinje cell
synapses (Hirano et al. 1995).
Adaptive processes in the cerebellum appear
similar to those in cerebellum-like structures. In
cerebellum-like structures, the pairing of par-
allel ber signals with excitatory input from
the periphery results in such signals eliciting
a predictive reduction in principal cell activity.
In the cerebellum, the pairing of parallel ber
signals with climbing ber input likely leads to
such signals eliciting a reduction in the ring
of Purkinje cells (but see Steuber et al. 2007 for
a contrary view). If the climbing bers convey
some type of sensory signal, gated through the
inferior olive, then the parallel ber signals that
are paired with the climbing bers, and which
predict their occurrence, will reduce Purkinje
cell activity, as shown by Jirenhed et al. (2007)
during eye-blink conditioning.
This review focuses on sensory predictions
through mechanisms of associative synaptic
plasticity and with those features of cerebellum-
like structures that are particularly relevant to
cerebellar function. Cerebellum-like structures
are also excellent sites for addressing other
468
 important issues in neuroscience, which cannot
be discussed here because of space constraints.
These include the roles of recurrent feedback
from higher to lower levels of the same sen-
sory system (Chacron et al. 2003, 2005; Do-
iron et al. 2003), the effects of motor commands
on sensory processing (Bell & Grant 1992), the
preservation and analysis of temporal informa-
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
tion (Kawasaki 2005), and the neural process-
ing of spectral cues for sound localization in the
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
DCN (Young & Davis 2002).
Predictions in the Cerebellum
The many similarities between cerebellum-like
structures and the cerebellum suggest that the
cerebellum too may be involved in generating
predictions concerning expected sensory input
or states of the system (Bell et al. 1997a, Devor
2000), and a variety of experimental, clinical,
and theoretical studies of the cerebellum sup-
port this hypothesis (Diedrichsen et al. 2007,
Nixon 2001, Paulin 2005, Wolpert et al. 1998).
The probable involvement of the cere-
bellum in predictive or feedforward control
through learning is well recognized (Bastian
2006, Ito 1984, Miall et al. 1993, Ohyama et al.
2003, Wolpert et al. 1998). Predictive control
allows for prior knowledge to shape an action,
as in knowing if a cup is full or empty before
picking it up. Several studies indicate that pre-
dictive feedforward control is decient in cere-
bellar patients (Morton & Bastian 2006, Smith
& Shadmehr 2005). Such patients do not adapt
their responses to predictable perturbations, al-
though they respond quite well to sudden un-
predictable perturbations of a movement, indi-
cating that feedback control from the periphery
is functional.
Theoreticians have proposed that the cere-
bellum may act in an adaptive and predictive
manner through the generation of two types
of models: forward models and inverse mod-
els (Wolpert et al. 1998). In a forward model,
copies of a motor command are conveyed to the
cerebellum together with information about
the current state of the system such as posi-
tions and velocities of the limbs. The cerebel-
www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function
lum then generates a prediction about the sen-
sory consequences of the commanded motor act
in the current context. In an inverse model, the
Forward model:
desired goal of an action together with infor- predicts the future
mation about the current state are conveyed to state of the system on
the cerebellum, which then generates the pre- the basis of the current
cise motor commands that will yield the desired state and the motor
command
goal. Both types of models must be capable of
plastic change or learning to adapt to changes Inverse model:
generates an
in the task or in the system, such as changes in
appropriate motor
load or initial limb position. command that will
Forward models are particularly important cause a desired change
in generating fast, coordinated movement se- in the state of the
quences. Feedback from peripheral sensory re- system
ceptors is slow. An appropriate command for
one phase of a movement must often be issued
before peripheral feedback can arrive about the
consequences of a motor command that evoked
a previous phase of the movement. A forward
model that predicts the sensory consequences
of a motor command, accounting for all that is
known about the current state of the system, al-
lows the next motor command in a sequence to
be issued appropriately and in accord with the
expected consequences of previous commands.
Such a process allows for the chunking of sepa-
rate components of a motor sequence and their
automatization, as described by Nixon (2001).
Moreover, classic symptoms of cerebellar dam-
age such as decomposition of movement, slow-
ness, and tremor can all be understood as due
to the absence of predictive forward models and
reliance on peripheral feedback (Bastian 2006,
Nixon 2001).
What is required in such automatization of
a sequence of movements is the predicted ef-
fect of the motor command: the sensory con-
sequences or state that results from the action,
not simply the motor command itself. Recent
experiments by Pasalar et al. (2006) suggest
that the Purkinje cell output from large regions
of the cerebellar hemispheres is indeed more
tightly coupled with predictions about conse-
quences of the movement than with the mo-
tor commands themselves (but see Yamamoto
et al. 2007). Pasalar et al. (2006) recorded from
Purkinje cells over a wide area of the hemi-
sphere in monkeys that had been trained to
15
469

Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
by CORNELL UNIVERSITY on 10/26/09. For personal use only.
16
control a cursor on a screen with a manipulan-
dum and to make the cursor track a circularly
moving stimulus. They then altered the forces
required to move the manipulandum. The elec-
tromyograms in the arm muscles varied sys-
tematically with the changes in required forces,
but Purkinje cell simple spike activity was un-
affected by the changes in force. Purkinje cell
simple spikes depended only on the position, di-
rection, and velocity of the movement. Purkinje
cell activity was phase advanced, that is, predic-
tive of the movement parameters or state of the
arm (T. Ebner, personal communication).
Pasalar et al. (2006) took their results as an
argument against an inverse model in the cere-
bellum because Purkinje cell activity had little
relation to the motor commands to the mus-
cles. Although one could argue that the ac-
tivity reects a high-level motor command, in
movement rather than muscle coordinates, the
simpler explanation is that the Purkinje cell ac-
tivity reects a forward model of expected con-
sequences, as required for the automatization of
movement sequences. Their experiments sug-
gest that not all sensory consequences are pre-
dicted; only those critical for accomplishing the
task are predicted. Thus presumed changes in
touch or muscle receptors associated with force
changes were not predicted by Purkinje cell ac-
tivity; only velocity and position of the limb
were predicted.
Examples of what are, in effect, forward
models in the cerebellum-like structures of
mormyrid and elasmobranch sh are described
in the previous section, showing that forward
models can indeed be generated within struc-
tures such as the cerebellum. In these sys-
tems, corollary discharge signals come to elicit
a prediction about the sensory input pattern
that is expected to follow the motor com-
mand. The possibility of such corollary dis-
charge effects in the OTML and DCN was also
mentioned.
Cerebellum-like structures can generate
predictions on the basis of other sensory in-
puts (Bastian 1996a, Bodznick et al. 1999), not
just on the basis of motor commands, and the
cerebellum may do so also. For example, in
Bell
Han Sawtell
eye-blink conditioning, which is thought to in-
volve the cerebellum, the timing of one sen-
sory signal, an air puff to the cornea (signaled
by the climbing ber), is predicted from an-
other sensory signal, a tone (signaled by mossy
bers) (Kim & Thompson 1997). Similarly,
cerebellar modulation of the vestibular ocu-
lar reex involves the prediction of one sen-
sory stimulus, retinal slip (signaled by climbing
bers), by the occurrence of another sensory
stimulus, vestibular input (signaled by mossy
bers). More broadly, Paulin (1993, 2005) has
suggested that the cerebellum estimates future
states of the organism or environment using
a combination of sensory, motor, and possibly
other types of information.
In simpler systems, such as the vestibular
ocular reex, in which Purkinje cell output is
coupled quite directly with motor pathways, the
adaptive alteration in Purkinje cell activity after
pairing with the climbing ber can be viewed as
either a prediction about a sensory input or as
a motor command. In more complex systems,
where Purkinje cell output is less tightly cou-
pled with motor pathways, as in the tracking
task studied by Pasalar et al. (2006), the hy-
pothesis of Purkinje cell activity as a predic-
tor of consequences may provide a more useful
perspective.
DIRECTIONS FOR
FUTURE RESEARCH
Our understanding of adaptive processing in
cerebellum-like structures is far from complete,
and future work will be useful both for under-
standing the neural mechanisms of sensory pro-
cessing and for understanding the cerebellum.
Promising lines of research are outlined briey
below.
Activity Patterns in Granule Cells
How the different types of predictive inputs are
combined and represented in the granule cells
that are associated with cerebellum-like struc-
tures remains unclear, as is also the case for cere-
bellar granule cells.
470
 Adaptive Filtering in
Electrosensory Systems
Several aspects of adaptive ltering in
cerebellum-like structures require further
investigation, including (a) the behavioral
consequences of adaptive ltering; (b) the
effects of adaptive ltering on encoding natu-
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org
ber inputs in auditory processing. Similarly,
very little is known about adaptive ltering in
the MON or OTML.
Purkinje-Like Cells
The functional roles of Purkinje-like cells
remain unclear. Recent work has shown

ralistic stimuli in the presence of self-generated that dendritic spikes that drive anti-Hebbian
interference; (c) the mechanisms of plasticity plasticity in Purkinje-like MG cells of the
and the presence of plasticity at other sites, mormyrid ELL are strongly regulated by
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

such as inhibitory synapses; and (d ) the possible
generation of more complex expectations such
as those based on memories of entire scenes
or sequences. The possibility of more complex
expectations is suggested by the massive de-
scending inputs that cerebellum-like structures
receive from higher levels of the same sensory
systems.
central signals, suggesting a parallel to su-
pervised learning mediated by climbing ber
inputs to the cerebellum (Sawtell et al. 2007).
In addition, the mormyrid ELL, the DCN, and
the teleost cerebellum all provide excellent op-
portunities for examining interactions between
Purkinje or Purkinje-like cells and neighboring
efferent cells (analogous to deep cerebellar nu-
clear cells in the mammalian cerebellum).

Adaptive Filtering in the DCN

and Less-Studied
Cerebellum-Like Structures
Recent studies have found synaptic plasticity
at parallel ber synapses onto Purkinje-like
cartwheel cells and fusiform cells in the DCN
in vitro. Yet very little is known at the systems
level regarding the role of such plastic parallel
Primitive Cerebellums
As discussed previously, the earliest craniates
possess cerebellum-like structures, but it is not
clear if they possess a cerebellum. Identica-
tion of a structure similar to the inferior olive
in hagsh or lampreys would help to resolve this
issue.

DISCLOSURE STATEMENT
The authors are not aware of any biases that might be perceived as affecting the objectivity of this
review.

ACKNOWLEDGMENTS
We thank Drs. Neal Barmack, Timothy Ebner, and Johannes Meek for their critical reviews of
the manuscript. The work was supported by grants from the National Institutes of Health (MH
49792 to C.C.B. and NS44961 to V.H.) and the National Science Foundation (IOB 0618212 to
N.B.S.) and by a National Research Service Award (NS049728 to N.B.S.).

LITERATURE CITED
Acampora D, Avantaggiato V, Tuorto F, Simeone A. 1997. Genetic control of brain morphogenesis
through Otx gene dosage requirement. Development 124(18):363950
Barmack NH, Shojaku H. 1992. Vestibularly induced slow oscillations in climbing ber responses
of Purkinje cells in the cerebellar nodulus of the rabbit. Neuroscience 50:15

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 17

471
 Bastian AJ. 2006. Learning to predict the future: the cerebellum adapts feedforward movement
control. Curr. Opin. Neurobiol. 16(6):64549
Bastian J. 1995. Pyramidal-cell plasticity in weakly electric sh: a mechanism for attenuating
responses to reafferent electrosensory inputs. J. Comp. Physiol. 176:6378
Bastian J. 1996a. Plasticity in an electrosensory system. I. General features of dynamic sensory
lter. J. Neurophysiol. 76:248396
Bastian J. 1996b. Plasticity in an electrosensory system. II. Postsynaptic events associated with a
dynamic sensory lter. J. Neurophysiol. 76:2497507
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Bastian J, Bratton B. 1990. Descending control of electroreception. I. Properties of nu-

cleus praeeminentialis neurons projecting indirectly to the electrosensory lateral line lobe.
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

J. Neurosci. 10:122640
Bastian J, Zakon H. 2005. Plasticity of sense organs and brain. See Bullock et al. 2005,
pp. 195228
Bell CC. 1981a. An efference copy modied by reafferent input. Science 214:45053
Bell CC. 1981b. Central distribution of octavolateral afferents and efferents in a teleost (Mormyri-
dae). J. Comp. Neurol. 195:391414
Bell CC. 1981c. Some central connections of medullary octavolateral centers in a mormyrid sh.
In Hearing and Sound Communication in Fishes, ed. RR Fay, AN Popper, WN Tavolga, pp.
38392. Berlin: Heidelberg, Springer-Verlag
Bell CC. 1982. Properties of a modiable efference copy in electric sh. J. Neurophysiol. 47:104356
Bell CC. 1986. Duration of plastic change in a modiable efference copy. Brain Res. 369:29
36
Bell CC. 2001. Memory-based expectations in electrosensory systems. Curr. Opin. Neurobiol.
11:48187
Bell CC. 2002. Evolution of cerebellum-like structures. Brain Behav. Evol. 59:312
26
Bell CC, Bodznick D, Montgomery J, Bastian J. 1997a. The generation and subtraction
of sensory expectations within cerebellum-like structures. Brain Behav. Evol. 50:17
31
Bell CC, Caputi A, Grant K. 1997b. Physiology and plasticity of morphologically identied cells
in the mormyrid electrosensory lobe. J. Neurosci. 17:640922
Bell CC, Caputi A, Grant K, Serrier J. 1993. Storage of a sensory pattern by anti-Hebbian synaptic
plasticity in an electric sh. Proc. Natl. Acad. Sci. USA 90:465054
Bell CC, Finger TE, Russell CJ. 1981. Central connections of the posterior lateral line lobe in
mormyrid sh. Exp. Brain Res. 42:922
Bell CC, Grant K. 1992. Corollary discharge effects and sensory processing in the mormyromast
regions of the mormyrid electrosensory lobe: II. Cell types and corollary discharge plasticity.
J. Neurophysiol. 68:85975
Bell CC, Grant K, Serrier J. 1992. Corollary discharge effects and sensory processing in the
mormyrid electrosensory lobe: I. Field potentials and cellular activity in associated structures.
J. Neurophysiol. 68:84358
Bell CC, Han VZ, Sugawara S, Grant K. 1997c. Synaptic plasticity in a cerebellum-like structure
depends on temporal order. Nature 387:27881
Bell CC, Maler L. 2005. Central neuroanatomy of electrosensory systems in sh. See Bullock
et al. 2005, pp. 68111
Bell CC, Russell CJ. 1978. Termination of electroreceptor and mechanical lateral line afferents in
the mormyrid acousticolateral area. J. Comp. Neurol. 182:36782
Bell CC, von der Emde G. 1995. Electric organ corollary discharge pathways in mormyrid sh:
II. The medial juxtalobar nucleus. J. Comp. Physiol. A. 177:46379

18 Bell
Han Sawtell
472
 Berman N, Dunn RJ, Maler L. 2001. Function of NMDA receptors in a feedback pathway of the
electrosensory system. J. Neurophysiol. 86:161221
Berrebi AS, Morgan JI, Mugnaini E. 1990. The Purkinje cell class may extend beyond the cere-
bellum. J. Neurocytol. 19(5):64354
Bodznick D. 1993. The specicity of an adaptive lter that suppresses unwanted reafference in
electrosensory neurons of the skate medulla. Biol. Bull. 185:31214
Bodznick D, Boord RL. 1986. Electroreception in Chondrichthyes: central anatomy and physi-
ology. See Bullock & Heiligenberg 1986, pp. 22556
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Bodznick D, Montgomery JC, Carey M. 1999. Adaptive mechanisms in the elasmobranch hind-
brain. J. Exp. Biol. 202:135764
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Bodznick D, Northcutt RG. 1980. Segregation of electro- and mechanoreceptive inputs to the
elasmobranch medulla. Brain Res. 195:31321
Braford MR. 1982. African, but not Asian, notopterid shes are electroreceptive: evidence from
brain characters. Neurosci. Lett. 32:3539
Brown MC, Berglund AM, Kiang NY, Ryugo DK. 1988. Central trajectories of type II spiral
ganglion neurons. J. Comp. Neurol. 278(4):58190
Bullock TH, Bodznick DA, Northcutt RG. 1983. The phylogenetic distribution of electrorecep-
tion: evidence for convergent evolution of a primitive vertebrate sense modality. Brain Res.
Rev. 6:2546
Bullock TH, Heiligenberg W. 1986. Electroreception. New York: Wiley
Bullock TH, Hopkins CD, Popper AN, Fay RR, eds. 2005. Electroreception. New York: Springer
Burian M, Gstoettner W. 1988. Projection of primary vestibular afferent bres to the cochlear
nucleus in the guinea pig. Neurosci. Lett. 84(1):1317
Butler AB, Saidel WM. 1992. Tectal projection to an unusual nucleus in the diencephalon of a
teleost sh, Pantodon buchholzi. Neurosci. Lett. 145:19396
Butler AB, Saidel WM. 2000. Dening sameness: historical, biological, and generative homology.
BioEssays 22(9):84653
Caicedo A, Herbert H. 1993. Topography of descending projections from the inferior colliculus
to auditory brainstem nuclei in the rat. J. Comp. Neurol. 328(3):37792
Campbell HR, Meek J, Zhang J, Bell CC. 2007. Anatomy of the posterior caudal lobe of the cere-
bellum and the eminentia granularis posterior in a mormyrid sh. J. Comp. Neurol. 502(5):714
35
Cant NB. 1992. The cochlear nucleus: neuronal types and their synaptic organization. In
The Mammalian Auditory Pathway: Neuroanatomy, ed. DB Webster, AN Popper, RR Fay,
pp. 66116. New York: Springer
Caputi AA, Castello ME, Aguilera P, Trujillo-Cenoz O. 2002. Electrolocation and electrocommu-
nication in pulse gymnotids: signal carriers, prereceptor mechanisms and the electrosensory
mosaic. J. Physiol. Paris 96(56):493505
Carr CE, Maler L. 1986. Electroreception in gymnatiform sh: central anatomy and physiology.
See Bullock & Heiligenberg 1986, pp. 31974
Chacron MJ, Doiron B, Maler L, Longtin A, Bastian J. 2003. Non-classical receptive eld mediates
switch in a sensory neurons frequency tuning. Nature 423(6935):7781
Chacron MJ, Maler L, Bastian J. 2005. Feedback and feedforward control of frequency tuning to
naturalistic stimuli. J. Neurosci. 25(23):552132
Conley RA, Bodznick D. 1994. The cerebellar dorsal granular ridge in an elasmobranch has pro-
prioceptive and electroreceptive representations and projects homotopically to the medullary
electrosensory nucleus. J. Comp. Physiol. A 174:70721
Crosby EC. 1969. Comparative aspects of cerebellar morphology. In Neurobiology of Cerebellar
Evolution and Development, ed. R Llinas, pp. 1941. Chicago: Am. Med. Assoc.

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 19

473
 Devor A. 2000. Is the cerebellum like cerebellar-like structures? Brain Res. Rev. 34(3):14956
Devor A. 2002. The great gate: control of sensory information ow to the cerebellum.
Cerebellum 1(1):2734
Diedrichsen J, Criscimagna-Hemminger SE, Shadmehr R. 2007. Dissociating timing and coor-
dination as functions of the cerebellum. J. Neurosci. 27(23):6291301
Doiron B, Chacron MJ, Maler L, Longtin A, Bastian J. 2003. Inhibitory feedback required for
network oscillatory responses to communication but not prey stimuli. Nature 421(6922):539
43
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Dupont JL, Gardette R, Crepel F. 1987. Postnatal development of the chemosensitivity of rat
cerebellar Purkinje cells to excitatory amino acids. An in vitro study. Brain Res. 431(1):5968
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Ebner TJ, Johnson MT, Roitman A, Fu Q. 2002. What do complex spikes signal about limb
movements? Ann. N.Y. Acad. Sci. 978:20518
Ekerot CF, Jorntell H. 2001. Parallel bre receptive elds of Purkinje cells and interneurons are
climbing bre-specic. Eur. J. Neurosci. 13:130310
Finger TE. 1978. Efferent neurons of the teleost cerebellum. Brain Res. 153:60814
Finger TE, Tong SL. 1984. Central organization of eighth nerve and mechanosensory lateral line
systems in the brainstem of ictalurid catsh. J. Comp. Neurol. 229:12951
Fujino K, Oertel D. 2003. Bidirectional synaptic plasticity in the cerebellum-like mammalian
dorsal cochlear nucleus. Proc. Natl. Acad. Sci. USA 100(1):26570
Goossens HH, Hoebeek FE, van Alphen AM, Van Der SJ, Stahl JS, et al. 2004. Simple spike and
complex spike activity of occular Purkinje cells during the optokinetic reex in mice lacking
cerebellar long-term depression. Eur. J. Neurosci. 19(3):68797
Graf W, Simpson JI, Leonard CS. 1988. Spatial organization of visual messages of the rabbits
cerebellar occulus. II. Complex and simple spike responses of Purkinje cells. J. Neurophysiol.
60(6):2091121
Grant K, Sugawara S, Gomez L, Han VZ, Bell CC. 1998. The Mormyrid electrosensory lobe in
vitro: physiology and pharmacology of cells and circuits. J. Neurosci. 18:600925
Han VZ, Grant G, Bell CC. 2000. Reversible associative depression and nonassociative potenti-
ation at a parallel ber synapse. Neuron 27:61122
Hawkes R, Herrup K. 1995. Aldolase C/zebrin II and the regionalization of the cerebellum.
J. Mol. Neurosci. 6(3):14758
Hirano T, Kasono K, Araki K, Mishina M. 1995. Suppression of LTD in cultured Purkinje cells
decient in the glutamate receptor d2 subunit. NeuroReport 6:52426
Hjelmstad GO, Parks G, Bodznick D. 1996. Motor corollary discharge activity and sensory re-
sponses related to ventilation in the skate vestibulolateral cerebellum: implications for elec-
trosensory processing. J. Exp. Biol. 199:67381
Ito M. 1984. The Cerebellum and Neural Control. New York: Raven
Ito M. 2001. Cerebellar long-term depression: characterization, signal transduction, and functional
roles. Physiol. Rev. 81(3):114395
Jirenhed DA, Bengtsson F, Hesslow G. 2007. Acquisition, extinction, and reacquisition of a cere-
bellar cortical memory trace. J. Neurosci. 27(10):2493502
Kanold PO, Young ED. 2001. Proprioceptive information from the pinna provides somatosensory
input to cat dorsal cochlear nucleus. J. Neurosci. 21(19):784858
Kawasaki M. 2005. Physiology of tuberous electrosensory systems. See Bullock et al. 2005,
pp. 154194
Kawasaki M, Guo YX. 1998. Parallel projection of amplitude and phase information from the hind-
brain to the midbrain of the African electric sh Gymnarchus niloticus. J. Neurosci. 18(18):7599
611

20 Bell
Han Sawtell
474
 Kim JJ, Thompson RF. 1997. Cerebellar circuits and synaptic mechanisms involved in classical
eyeblink conditioning. Trends Neurosci. 20:17781
Kobayashi Y, Kawano K, Takemura A, Inoue Y, Kitama T, et al. 1998. Temporal ring patterns
of Purkinje cells in the cerebellar ventral paraocculus during ocular following responses in
monkeys II. Complex spikes. J. Neurophysiol. 80(2):83248
Lannoo M, Hawkes R. 1997. A search for primitive Purkinje cells: zebrin II expression in sea
lampreys (Petromyzon marinus). Neurosci. Lett. 237:5355
Lannoo MJ, Maler L, Hawkes R. 1992. Zebrin II distinguishes the ampullary organ receptive map
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

from the tuberous organ receptive maps during development in the teleost electrosensory
lateral line lobe. Brain Res. 586:17680
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Lannoo MJ, Ross L, Maler L, Hawkes R. 1991. Development of the cerebellum and its extracellular
Purkinje cell projection in teleost shes as determined by zebrin II immunocytochemistry.
Prog. Neurobiol. 37:32963
Larsell O. 1967. The Comparative Anatomy and Histology of the Cerebellum from Myxinoids through
Birds. Minneapolis: Univ. Minn. Press
Lisberger SG, Fuchs AF. 1974. Response of occulus Purkinje cells to adequate vestibular stim-
ulation in the alert monkey: xation vs compensatory eye movements. Brain Res. 69:34753
Maekawa K, Simpson JI. 1972. Climbing ber activation of Purkinje cells in the occulus by
impulses transferred through the visual pathway. Brain Res. 39:24551
Maler L, Sas EKB, Rogers J. 1981. The cytology of the posterior lateral line lobe of high frequency
weakly electric sh (Gymnotidae): dendritic differentiation and synaptic specicity in a simple
cortex. J. Comp. Neurol. 195:87140
Manis PB, Molitor SC. 1996. N-methyl-d-aspartate receptors at parallel ber synapses in the
dorsal cochlear nucleus. J. Neurophysiol. 76:163955
Martinez S, Crossley PH, Cobos I, Rubenstein JL, Martin GR. 1999. FGF8 induces formation
of an ectopic isthmic organizer and isthmocerebellar development via a repressive effect on
Otx2 expression. Development 126(6):1189200
Matsushita A, Kawasaki M. 2005. Neuronal sensitivity to microsecond time disparities in the
electrosensory system of Gymnarchus niloticus. J. Neurosci. 25(49):1142432
Matsushita M, Tanami T. 1987. Spinocerebellar projections from the central cervical nucleus in
the cat, as studied by anterograde transport of wheat germ agglutinin-horseradish peroxidase.
J. Comp. Neurol. 266(3):37697
McCormick CA. 1997. Organization and connections of octaval and lateral line centers in the
medulla of a clupeid, Dorosoma cepedianum. Hear Res. 110(12):3960
McCormick CA. 1999. Anatomy of the central auditory pathways of sh and amphibians. In
Comparative Hearing: Fish and Amphibians, ed. RR Fay, AN Popper, pp. 155217. New York:
Springer Verlag
McCreery DB. 1977. Two types of electroreceptive lateral lemniscal neurons of the lateral line
lobe of the catsh Ictalurus nebulosus; connections from the lateral line nerve and steady-state
frequency response characteristics. J. Comp. Physiol. 113:31739
Medina JF, Garcia KS, Nores WL, Taylor NM, Mauk MD. 2000. Timing mechanisms in the
cerebellum: testing predictions of a large-scale computer simulation. J. Neurosci. 20(14):5516
25
Meek J. 1983. Functional anatomy of the tectum mesencephali of the goldsh. An explorative
analysis of the functional implication of the laminar structural organization of the tectum.
Brain Res. Rev. 6:24797
Meek J. 1998. Holosteans and teleosts. In The Central Nervous System of Vertebrates, ed. R Nieuwen-
huys, HJ Ten Donkelaar, C Nicholson, Vol. 15, pp. 759937. Berlin: Springer

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 21

475
 Meek J, Grant K, Sugawara S, Hafmans TGM, Veron M, Denizot JP. 1996. Interneurons of the
ganglionic layer in the mormyrid electrosensory lateral line lobe: morphology, immunocyto-
chemistry, and synaptology. J. Comp. Neurol. 375:4365
Miall RC, Weir DJ, Wolpert DM, Stein JF. 1993. Is the Cerebellum a smith predictor. J. Motor
Behav. 25(3):20316
Mikami Y, Yoshida T, Matsuda N, Mishina M. 2004. Expression of zebrash glutamate receptor
delta2 in neurons with cerebellum-like wiring. Biochem. Biophys. Res. Commun. 322(1):16876
Montgomery JC, Bodznick D. 1993. Hindbrain circuitry mediating common mode suppres-
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

sion of ventilatory reafference in the electrosensory system of the little skate Raja erinacea.
J. Exp. Biol. 183:20315
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Montgomery JC, Bodznick D. 1999. Signals and noise in the elasmobranch electrosensory system.
J. Exp. Biol. 202(Pt. 10):134955
Montgomery JC, Coombs S, Conley RA, Bodznick D. 1995. Hindbrain sensory processing in
lateral line, electrosensory, and auditory systems: a comparative overview of anatomical and
functional similarities. Auditory Neurosc. 1:20731
Morton SM, Bastian AJ. 2006. Cerebellar contributions to locomotor adaptations during splitbelt
treadmill walking. J. Neurosci. 26(36):910716
Mugnaini E, Dino MR, Jaarsma D. 1997. The unipolar brush cells of the mammalian cerebellum
and cochlear nucleus: cytology and microcircuitry. Prog. Brain Res. 114:13150
Mugnaini E, Osen KK, Dahl AL, Friedrich VL Jr, Korte G. 1980a. Fine structure of granule cells
and related interneurons (termed Golgi cells) in the cochlear nuclear complex of cat, rat and
mouse. J. Neurocytol. 9(4):53770
Mugnaini E, Warr WB, Osen KK. 1980b. Distribution and light microscopic features of granule
cells in the cochlear nuclei of cat, rat, and mouse. J. Comp. Neurol. 191(4):581606
Nelson ME, Paulin MG. 1995. Neural simulations of adaptive reafference suppression in the
elasmobranch electrosensory system. J. Comp. Physiol. A 177:72336
Nieuwenhuys R, Ten Donkelaar HJ, Nicholson C. 1997. The Central Nervous System of Vertebrates.
Heidelberg: Springer
Nixon PD, Passingham RE. 2001. Predicting sensory events: the role of the cerebellum in motor
learning. Exp. Brain. Res. 138:25157
Noda H, Warabi T. 1982. Eye position signals in the occulus of the monkey during smooth-
pursuit eye movements. J. Physiol. 324:187202
Northmore DPM, Williams B, Vanegas H. 1983. The teleostean torus longitudinalis: responses
related to eye movements, visuotopic mapping, and functional relations with the optic tectum.
J. Comp. Physiol. A 150:3950
Ohlrogge M, Doucet JR, Ryugo DK. 2001. Projections of the pontine nuclei to the cochlear
nucleus in rats. J. Comp. Neurol. 436(3):290303
Ohyama T, Nores WL, Murphy M, Mauk MD. 2003. What the cerebellum computes. Trends
Neurosci. 26(4):22227
Pasalar S, Roitman AV, Durfee WK, Ebner TJ. 2006. Force eld effects on cerbellar Purkinje cell
discharge with implications for internal models. Nat. Neurosci. 9(11):140411
Paulin MG. 1993. The role of the cerebellum in motor control and perception. Brain Behav. Evol.
41:3950
Paulin MG. 2005. Evolution of the cerebellum as a neuronal machine for Bayesian state estimation.
J. Neural. Eng. 2(3):S21934
Petralia RS, Wang YX, Zhao HM, Wenthold RJ. 1996. Ionotropic and metabotropic glutamate
receptors show unique postsynaptic, presynaptic, and glial localizations in the dorsal cochlear
nucleus. J. Comp. Neurol. 372(3):35683

22 Bell
Han Sawtell
476
 Puzdrowski RL, Leonard RB. 1993. The octavolateral systems in the stingray, Dasyatis sabina. I.
Primary projections of the octaval and lateral line nerves. J. Comp. Neurol. 332(1):2137
Raymond JL, Lisberger SG. 1998. Neural learning rules for the vestibulo-ocular reex.
J. Neurosci. 18(21):911229 (Abstr.)
Roberts PD. 1999. Computational consequences of temporally asymmetric learning rules: I. dif-
ferential Hebbian learning. J. Comp. Neurosci. 7:23546
Roberts PD, Bell CC. 2000. Computational consequences of temporally asymmetric learning
rules: II. sensory image cancellation. J. Comput. Neurosci. 9:6783
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Robertson LT. 1985. Somatosensory representation of the climbing ber system in the rostral
intermediate cerebellum. Exp. Brain Res. 61(1):7386
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Ronan M 1986. Electroreception in cyclostomes. See Bullock & Heiligenberg 1986, pp. 20924
Safo P, Regehr WG. 2008. Timing dependence of the induction of cerebellar LTD. Neurophar-
macology. In press
Saunders J, Bastian J. 1984. The physiology and morphology of two types of electrosensory neurons
in the weakly electric sh Apteronotus leptorhynchus. J. Comp. Physiol. A 154:199209
Sawtell NB, Williams A, Bell CC. 2007. Central control of dendritic spikes shapes the responses of
Purkinje-like cells through spike timing-dependent synaptic plasticity. J. Neurosci. 27(7):1552
65
Schlegel P. 1973. Perception of objects in weakly electric sh Gymnotus carapo as studied in record-
ings from rhombencephalic neurons. Exp. Brain Res. 18:34054
Schmidt AW, Bodznick D. 1987. Afferent and efferent connections of the vestibulolateral cere-
bellum of the little skate, Raja erinacea. Brain Behav. Evol. 30:282302
Shore SE, Zhou J. 2006. Somatosensory inuence on the cochlear nucleus and beyond. Hear. Res.
21617:9099
Smith MA, Shadmehr R. 2005. Intact ability to learn internal models of arm dynamics in Hunt-
ingtons disease but not cerebellar degeneration. J. Neurophysiol. 93(5):280921
Steuber V, Mittman W, Hoebeek FE, Silver RA, De Zeeuw CI, et al. 2007. Cerebellar LTD and
pattern recognition by Purkinje cells. Neuron 54(1):12136
Stone LS, Lisberger SG. 1990. Visual responses of Purkinje cells in the cerebellar occulus during
smooth-pursuit eye movements in monkeys. II. Complex spikes. J. Neurophysiol. 63(5):1262
75
Szabo T, Libouban S, Haugede-Carre F. 1979. Convergence of common and specic sensory
afferents to the cerebellar auricle (auricula cerebelli) in the teleost sh Gnathonemus demon-
strated by HRP method. Brain Res. 168:61922
Szabo T, Libouban S, Denizot JP. 1990. A well dened spinocerebellar system in the weakly
electric teleost sh Gnathonemus petersii. Arch. Ital. Biol. 128:22947
Tong S. 1982. The nucleus praeeminentialis: an electro- and mechanoreceptive center in the
brainstem of the catsh. J. Comp. Physiol. A 145:299309
Tong S, Finger TE. 1983. Central organization of the electrosensory lateral line system in bullhead
carsh Ictalurus nebulosus. J. Comp. Neurol. 217:116
Tzounopoulos T, Kim Y, Oertel D, Trussell LO. 2004. Cell-specic, spike timing-dependent
plasticities in the dorsal cochlear nucleus. Nat. Neurosci. 7(7):71925
Tzounopoulos T, Rubio ME, Keen JE, Trussell LO. 2007. Coactivation of pre- and postsynap-
tic signaling mechanisms determines cell-specic spike-timing-dependent plasticity. Neuron
54(2):291301
Vanegas H, Williams B, Freeman JA. 1979. Responses to stimulation of marginal bers in the
teleostean optic tectum. Exp. Brain Res. 34(2):33549
Wang SS, Denk W, Hausser M. 2000. Coincidence detection in single dendritic spines mediated
by calcium release. Nat. Neurosci. 3:126673

www.annualreviews.org Cerebellum-Like Structures and Their Implications for Cerebellar Function 23

477
 Weedman DL, Ryugo DK. 1996. Projections from auditory cortex to the cochlear nucleus in rats:
synapses on granule cell dendrites. J. Comp. Neurol. 371:31124
Weinberg RJ, Rustioni A. 1987. A cuneocochlear pathway in the rat. Neuroscience 20(1):209
19
Wolff A, Kunzle H. 1997. Cortical and medullary somatosensory projections to the cochlear
nuclear complex in the hedgehog tenrec. Neurosci. Lett. 221(23):12528
Wolpert DM, Miall C, Kawato M. 1998. Internal models in the cerebellum. Trends Cogn. Sci.
2:33847
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Yamamoto K, Kawato M, Kotosaka S, Kitazawa S. 2007. Encoding coding of movements dynamics

by Purkinje cell simple spike activity during fast arm movements under resistive and assistive
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

force elds. J. Neurophysiol. 97:158899

Yawata S, Tsuchida H, Kengaku M, Hirano T. 2006. Membrane-proximal region of gluta-
mate receptor delta2 subunit is critical for long-term depression and interaction with pro-
tein interacting with C kinase 1 in a cerebellar Purkinje neuron. J. Neurosci. 26(14):3626
33
Young ED, Davis KA. 2002. Circuitry and function of the dorsal cochlear nucleus. In Integrative
Functions in the Mammalian Auditory Pathway, ed. D Oertel, AN Popper, RR Fay, pp. 160206.
New York: Springer-Verlag
Yuzaki M. 2003. The delta2 glutamate receptor: 10 years later. Neurosci. Res. 46(1):1122
Zhou J, Shore S. 2004. Projections from the trigeminal nuclear complex to the cochlear nuclei: a
retrograde and anterograde tracing study in the guinea pig. J. Neurosci. Res. 78(6):9017

24 Bell
Han Sawtell
478
 AR346-FM ARI 20 May 2008 15:1

Annual Review of
Neuroscience

Contents Volume 31, 2008

Cerebellum-Like Structures and Their Implications for Cerebellar

Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

Function
Curtis C. Bell, Victor Han, and Nathaniel B. Sawtell p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 1
by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Spike TimingDependent Plasticity: A Hebbian Learning Rule

Natalia Caporale and Yang Dan p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p25
Balancing Structure and Function at Hippocampal Dendritic Spines
Jennifer N. Bourne and Kristen M. Harris p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p47
Place Cells, Grid Cells, and the Brains Spatial Representation System
Edvard I. Moser, Emilio Kropff, and May-Britt Moser p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p69
Mitochondrial Disorders in the Nervous System
Salvatore DiMauro and Eric A. Schon p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p91
Vestibular System: The Many Facets of a Multimodal Sense
Dora E. Angelaki and Kathleen E. Cullen p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 125
Role of Axonal Transport in Neurodegenerative Diseases
Kurt J. De Vos, Andrew J. Grierson, Steven Ackerley, and Christopher C.J. Miller p p p 151
Active and Passive Immunotherapy for Neurodegenerative Disorders
David L. Brody and David M. Holtzman p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 175
Descending Pathways in Motor Control
Roger N. Lemon p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 195
Task Set and Prefrontal Cortex
Katsuyuki Sakai p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 219
Multiple Sclerosis: An Immune or Neurodegenerative Disorder?
Bruce D. Trapp and Klaus-Armin Nave p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 247
Multifunctional Pattern-Generating Circuits
K.L. Briggman and W.B. Kristan, Jr. p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 271
Retinal Axon Growth at the Optic Chiasm: To Cross or Not to Cross
Timothy J. Petros, Alexandra Rebsam, and Carol A. Mason p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 295

479
v
 AR346-FM ARI 20 May 2008 15:1

Brain Circuits for the Internal Monitoring of Movements

Marc A. Sommer and Robert H. Wurtz p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 317
Wnt Signaling in Neural Circuit Assembly
Patricia C. Salinas and Yimin Zou p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 339
Habits, Rituals, and the Evaluative Brain
Ann M. Graybiel p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 359
Mechanisms of Self-Motion Perception
Kenneth H. Britten p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 389
Mechanisms of Face Perception
Doris Y. Tsao and Margaret S. Livingstone p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 411
Annu. Rev. Neurosci. 2008.31:1-24. Downloaded from arjournals.annualreviews.org

The Prions Elusive Reason for Being

by CORNELL UNIVERSITY on 10/26/09. For personal use only.

Adriano Aguzzi, Frank Baumann, and Juliane Bremer p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 439

Mechanisms Underlying Development of Visual Maps and
Receptive Fields
Andrew D. Huberman, Marla B. Feller, and Barbara Chapman p p p p p p p p p p p p p p p p p p p p p p p p 479
Neural Substrates of Language Acquisition
Patricia K. Kuhl and Maritza Rivera-Gaxiola p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 511
Axon-Glial Signaling and the Glial Support of Axon Function
Klaus-Armin Nave and Bruce D. Trapp p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 535
Signaling Mechanisms Linking Neuronal Activity to Gene Expression
and Plasticity of the Nervous System
Steven W. Flavell and Michael E. Greenberg p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 563

Indexes

Cumulative Index of Contributing Authors, Volumes 2231 p p p p p p p p p p p p p p p p p p p p p p p p p p p 591

Cumulative Index of Chapter Titles, Volumes 2231 p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p p 595

Errata

An online log of corrections to Annual Review of Neuroscience articles may be found at

http://neuro.annualreviews.org/

480
vi Contents
J. comp. Physiol. 89, 331--357 (1974)
9 by Springer-Verlag 1974

Chasing Behaviour of Houseflies (Fannia canicularis)

A Description and Analysis
M. F . L a n d a n d T. S. Collett

School of Biological Sciences, University of Sussex

Received November 21, 1973

Summary. 1. Chases in which male flies (2'annia canicularis) pursue other flies
were studied by filming such encounters from directly below. Males will start to
chase whenever a second fly comes within 10-15 cm (Fig. 3).
2. Throughout these chases there was a continuous relationship between the
angle (0e) made by the leading fly and the direction of flight of the chasing fly, and
the angular velocity of the chasing fly (w]). This relation was approximately linear,
with a slope of 20 ~ s-1 per degree 0e (Figs. 4-7).
3. The maximum correlation between w/and 0e occurs after a lag of approxi-
mately 30 ms, which represents the total delay in the system (Fig, 8).
4. In the region close to the chasing fly's axis (0e less than about 35 ~ a second
mechanism exists in which the angular velocity of the chasing fly (o)/) is controlled
by the relative angular velocity of the leading fly (we), rather than its relative
position. The ratio of w/to o~e in this region is approximately 0.7.
5. Using the results in 2-4 above, and an empirically determined relation
between the angular and forward velocities of the chasing fly, it was possible to
simulate the flight path of the chasing fly, given that of the leading fly (Fig. 11).
Because these simulations predict correctly the manoeuvres and outcomes of quite
complicated chases, it is concluded that the control system actually used by the
fly is accurately described by conclusions 2-4.
6. The physiological implications of this behaviour, and the possible function
of chasing, are discussed.

Introduction
I t is quite c o m m o n t o see houseflies e n g a g e d in w h a t a p p e a r t o be
s h o r t fast chases, which involve m a n y r a p i d changes of course. I f these
are i n d e e d chases, iu t h e sense t h a t one fly is t r y i n g t o c a t c h t h e o t h e r
t h e n this implies t h a t t h e fly which is doing t h e chasing m u s t possess a n
a c c u r a t e a n d r a p i d s y s t e m for d e t e r m i n i n g t h e course of t h e l e a d i n g fly,
a n d for controlling its own flight p a t h accordingly. This p a p e r p r e s e n t s
a n a c c o u n t of filmed records of such chases, a n d i t is shown t h a t these
are cases of real pursuit, a n d n o t meaningless d i s p l a y s of aerobatics.
B y e x a m i n i n g t h e form of t h e chases it should be possible to d e t e r m i n e
t h e k i n d of control s y s t e m t h e chasing fly uses. There are two basic
possibihties. The s y s t e m m i g h t be " c o n t i n u o u s " in t h e sense t h a t some
22*

481
332 M.F. Land and T. S. Collett

f e a t u r e of t h e l e a d i n g f l y ' s course (e.g. its i n s t a n t a n e o u s p o s i t i o n or

a n g u l a r v e l o c i t y r e l a t i v e t o t h e chasing fly) c o n t i n u o u s l y controls some
a s p e c t of t h e chasing fly's course (e.g. its a n g u l a r velocity). A l t e r n a t i v e l y ,
t h e chasing f l y m i g h t e m p l o y a "discontinuous" k i n d of s y s t e m in which
each m a n o e u v r e of t h e l e a d i n g f l y is m a t c h e d b y a n a p p r o p r i a t e b u t
stereotyped reply-manoeuvre.
To a n t i c i p a t e t h e results s o m e w h a t , i t seems t h a t this is a continuous
system, a n d t h a t t h e b e h a v i o u r of t h e chasing fly can be d e s c r i b e d b y a
quite simple " k a p u t - o u t p u t " relationship. F u r t h e r m o r e , using this
r e l a t i o n s h i p i t is possible t o s i m u l a t e t h e chasing fly's flight p a t h , given
t h a t of t h e l e a d i n g fly. T h e r e s u l t s of such s i m u l a t i o n s p r e d i c t quite
a c c u r a t e l y t h e ~orms a n d o u t c o m e s of chases, a n d this implies t h a t t h e
" m o d e l f l y " d e r i v e d f r o m t h e o b s e r v e d chases p r o v i d e s a n a d e q u a t e
e x p l a n a t i o n of w h a t is seen to occur.
I f t h i s b e h a v i o u r is v i s u a l l y m e d i a t e d , a n d it is difficult to i m a g i n e
how else it could be controlled, t h e n t h e i n f o r m a t i o n d e r i v e d from these
flights can b e u s e d t o infer some of t h e p r o p e r t i e s to be e x p e c t e d of
n e u r o n e s responsible for t h e visual control of flight.

Methods
Filming
There are usually two species of fly found inside houses in England, the common
house fly (Musca domestica) and the lesser house fly (~annia canieularis). The
species are easily distinguished by the sharp upward bend of wing vein 4 iC in
M. domestica compared with the straight vein in F. canicularis (see e.g. Colyer and
Hs,mmond, 1951). The other difference which is important for this study is that
male iv. canicularis tend to congregate round prominent objects such as lampshades,
making horizontal "patrolling" flights near them for long periods. M. domestica on
the other hand do not seem to prefer to fly in any particular parts of a room. Males
of both species engage in chases, of females or more commonly of other males, but
because only _~. canicularis can be relied upon to perform these at a known location,
it has so far only been possible to film this species.
Films were made by positioning the camera (16 mm Bolex with a constant
speed motor operating at 50 frames per second) directly beneath a lampshade
frequented by F. canicularis, and pointing upwards. A reduced shutter sector was
used to give an exposure time of 3 ms, which minimised blur. Inspection of chasing
flights from the side and below shows that most manoeuvres take place in the
horizontal plane, and that vertical components of chases are rather small in com-
parison. Where there are vertical components to the chases these seem to be mainly
in the form of spiralling movements, i.e. ones having a nearly constant upward or
downward component, which will reduce the apparent velocity of the flies along
the course as seen from below, but will not greatly change their angular relations,
which are the important determining features of chasing behaviour. Examination
of the flies which had taken part in filmed chases showed that most were males,
and the chases analysed here are almost certainly interactions between males.
~ilms were taken on hot days (about 25~ in normal room daylight (backgrounds
about 50 cd. era-2).

482
 Control of Chasing in Flies 333

/
5Of

Fig. 1. Alternative ways of measuring the error angle 0~, and the angular velocity
of the chasing fly cot= ~Oj~t, at the instants measured on the film, and instants
halfway between them

Analysis
The films were analysed frame by frame using a Specto projector, and the
courses of the flies in the horizontal plane were plotted. Usually it was impossible
to determine the exact direction of the fly's axis, and in subsequent analyses it is
assumed that this is the same as the direction of flight. From these course plots
three kinds of information were extracted.
(i) The angle made by the line of flight of the chasing fly and a line joining the
chasing and leading fly: this is referred to as the error angle (0e, Fig. 1), and it is
assumed that this represents the location of the leading fly on the retina of the
chasing fly at a particular instant. The extent to which flies roll about their long
axes during turns is not known ("banking") so this angle cannot simply be inter-
preted as an angle in the chasing fly's longitudinal plane, but in most instances it
is likely that the major component is in this plane, and for the following analysis
it does not matter whether or not this is true. The error angular velocity and error
angular acceleration (coe and ~e) can be obtained from 0e as indicated in Table 1.
Two methods of obtaining 0e are shown in Fig. 1.
(ii) The angular velocity of the chasing fly (~ol). This was measured as the
change in course angle between two frames of film ((~01) divided by the sampling
time ((~t~ 20 ms). Fig. 1 gives two methods of measuring ~0 I. This angular velocity
is taken as being appropriate to an instant halfway between the two sampling
instants. Angular~acceleration (o~/) can be obtained approximately from the dif-
ference between successive values of a)f divided again by the sampling time.
(iii) The forward velocities of leading and chasing flies (vl and vl). These are
taken as the distances travelled along the flight paths between frames of film,
divided by the sampling time. The errors involved in aU these estimates increase
as the angular velocities of the flies increase, but this is not likely to cause serious
errors at angular velocities in the observed range. There does not appear to be any
"fine structure" to the records described that is not resolved at 50 frames per
second.

483
334 M. 1~. Land and T. S. Collett

Table 1. Symbols

0~ Error angle; the angle between a tangent to the flight path of the chasing
fly and a line joining the two flies at a particular instant.
Error angular velocity; rate of change of 0e, measured as ~Oe/~t i.e. the
change in 0e during the sampling time ~t.
~be Error angular acceleration; measured as ~coe/6t.
Angular velocity of the following (chasing) fly; measured as ~O//~s i.e.
the change in course angle during the sampling time.
c51 Angular acceleration of the chasing fly; measured as ~co//($t.
vl, v! Horizontal forward velocities of the leading and following fly; measured
as the distances travelled along the respective flight paths between
sampling instants, divided by the sampling interval.
~t Sampling intervM used in measurement and reconstruction of flight
paths, usually 20 ms.
d Delay or latency of the response; defined as difference in time between
the measurement of an "input" variable (e.g. 0e) and the time of maxi-
mum correlation between this and an "output" variable (e.g. co]).
sign. con- All anti-clockwise angles, angular velocities and angular accelerations
vention are taken as positive, and vice-versa.

Results
General Description o/the _Flight Pattern o/Fannia canicularis
Male Fannia s p e n d m u c h of t h e i r t i m e m a k i n g c h a r a c t e r i s t i c " p a t r o l -
l i n g " flights a r o u n d p r o m i n e n t objects, a n d i t is in t h e course of such
flights t h a t e n c o u n t e r s w i t h o t h e r flies t a k e place. Colyer a n d H a m m o n d
(1951) give a n e x c e l l e n t d e s c r i p t i o n of this b e h a v i o u r , which t h e y s a y is
c h a r a c t e r i s t i c of t h e genus.
" T h i s flight is m a d e on a series of i r r e g u l a r t r i a n g u l a r or q u a d r i l a t e r a l
courses, a n a l m o s t i m p e r c e p t i b l e h o v e r i n g t a k i n g place a t t h e corners
a n d t h e sides being c o v e r e d in a r a p i d d a r t . W h e n u n d i s t u r b e d a n d alone,
t h e s e flies m a i n t a i n a m o r e or less c o n s t a n t h e i g h t a n d r e g u l a r course,
b u t w h e n m o r e t h a n one decides t o p a t r o l t h e same ' b e a t , ' i t u s u a l l y
h a p p e n s t h a t one d a r t s r a p i d l y t o w a r d s t h e other, a s h a r p f l u r r y a n d
p r o m p t dispersal ensues a n d e v e n t u a l l y one of t h e m r e c o m m e n c e s t h e
patrolling."
Two e x a m p l e s of such p a t r o l l i n g flights are shown in Fig. 2. T h e
s t r a i g h t s e g m e n t s are a b o u t 20 cm long, a n d t h e a v e r a g e v e l o c i t y of t h e
flies is 65 cm s -1, a b o u t half t h e speed seen d u r i n g chases. S o m e t i m e s
t w o or e v e n t h r e e m a l e s m a n a g e t o co-exist in t h e s a m e p a r t of t h e room,
b u t w h e n t h i s occurs t h e flies are f o u n d to be s t r a t i f i e d v e r t i c a l l y , so t h a t

484
Fig. 2. Two examples of "patrolling" flights made beneath a lampshade, which is
indicated by the heavy line. Time interval between points: 20 ms

4 4 ~"',

Fig. 3. Patrolling flight ending in a chase. Chasing fly e, leading fly o. Points are
at 20 ms intervals. Corresponding instants in the two flight paths are numbered
at 100 ms intervals. Dotted line indicates relative positions of the flies at the
probable instant of visual contact

there is a separation of 10-30 cm b e t w e e n the planes of the various

patrolling flights.
F r o m u n a s s i s t e d observation, it seemed t h a t whenever one fly came
withi~ a distance of 10-15 cm of another, a chase followed a u t o m a t i c a l l y .
The b e g i n n i n g of one such e n c o u n t e r is shown i n Fig. 3. Here t h e distance

485
336 M.F. Land and T. S. Collett

between the flies when the chasing fly makes its first turn is 9 cm, which
is probably a slight under-estimate because there must be some latent
period. However, if 10 cm is taken as the "visual contact distance,"
then a 6 m m long fly will occupy a visual angle of 3.4 ~ b y 2 ~ which is
comparable with an interommatidial angle of about 2 ~ (in the closely
similar housefly Musca).
Females rarely strayed into the area occupied b y males, although
when t h e y did t h e y were chased and sometimes caught, the two flies
joining, and flying in tandem to a nearby surface to copulate. Un-
fortunately, only one instance of what appeared to be a maleflemale
chase was filmed, and this was so brief t h a t little could be learned from
it. B y far the majority of chases were between males.

.Flight Paths during a Chase

Fig. 4 shows the longest and most complete film of a chase t h a t was
obtained. The entire sequence lasts just over a second: encounters are
often much briefer t h a n this and rarely last more than 2 seconds.
Inspection of Fig. 4 shows t h a t one animal (open circles) is being
chased and is taking continuous evasive action, and t h a t the other
(closed circles) is trying to follow the first as closely as possible. I n the
chase the leading animal makes six quite distinct manoeuvres: a sharp
left turn after 1, a right turn after 2, a 180 ~ right turn between 3 and 4,
a left turn at 4 immediately followed b y a right turn, and finally an
extraordinary right hand loop which leaves the fly almost on its former
course. I t is apparently this final manoeuvre t h a t causes the pursuer to
lose visual contact. I n contrast to this almost random behaviour, the
pursuing fly seems to behave in a much more comprehensible manner:
it follows each manoeuvre of the leading fly quite accurately, even
during the final loop. I t gives the impression of trying to keep its
course pointing in the direction of the leading fly, as though it were
attempting to catch up with it.
I n a chase situation like this the leading animal has the advantage
in t h a t it makes the manoeuvres first, and the pursuer, because of the
limitations of whatever control system it is using, wastes a certain
amount of time in making appropriate course changes. This is particu-
larly obvious during the turns between 3 and 4, and 4 and 5, where, in
spite of the pursuer's faster forward velocity (1.3 times t h a t of the
leading fly, overall) the pursuer's inability to follow each turn instantane-
ously results in it lagging behind the leader b y about 9 cm b y 5, although
both flies were level at 3. Thus the leading fly can only escape from its
pursuer b y outmanoeuvring it, since on a straight course the pursuer
would, in this case at least, overtake its target.

486
 Control of Chasing in Flies 337

I !

"~ 5cm
Fig. 4. Flight paths of chasing (o) and leading (9 flies during the longest recorded
chase. Points at 20 ms intervals. Corresponding instants on the two paths numbered
at 200 ms intervals

The remainder of this paper is devoted to an a t t e m p t to unravel the

nature of the control system t h a t enables the pursuing fly to follow as
closely as it evidently does. I t is assumed throughout t h a t the chasing
fly is guided b y vision, although the validity of the analysis does not
depend on this.
Input-Output Relations o/the ChasingFly
In the introduction it was suggested that the chasing fly might either
possess a set of stereotyped responses to manoeuvres made b y the leading
fly, or else t h a t it operated on the basis of a continuously running control
system in which some source of information available to it (0e, we, or o)e,
see Fig. 1) controlled some feature of its behaviour (o~],~l and possibly
vl). The best procedure seems to be to determine whether or not there
are a n y continuous relationships between "input" and "output"
variables, if there are to test whether or not they can explain the flight

487
338 M.F. Land and T. S. Collett

+lod

Oe

-too

+ 2,50C
O.S-1

0"2 s - - ~
Lof

-2,50C

Fig. 5. Plots of error angle (0e) and the angular velocity of the chasing fly (to1)
during the chase shown in Fig. 4, showing the continuous correspondence between
the two. Numbers on the time axis refer to the numbered points in Fig. 4

path, and only if this procedure fails to start looking for more com-
plicated kinds of responses.
On comparing input and output variables it became clear t h a t there
was a relationship, apparently continuous, between error angle (0e) and
the chasing fly's angular velocity (~ol). This is shown in two forms in
Figs. 5 and 6a. Fig. 5 shows t h a t whenever the leading fly departed
from a position ahead of the chasing fly (0~ = 0) the chasing fly turned
towards it with an angular velocity roughly proportional to the deviation.
Comparing the two plots, there seems to be a delay of between 20 and
40 ms between the two, and this will be examined further below. The
scatter diagrams in Fig. 6 are all based on the assumption t h a t the input
variables maximally affect the output 30 ms later. From Fig. 6a it can
be seen t h a t there is a strong and apparently linear relation between
o)I and 0e. The correlation coefficient is 0.76 and is highly significant.
The relationship can be approximated b y the regression line, which has
a slope of 21 s-1 (the full units are degrees per second per degree 0e, or
simply s-l).

488
 Control of Chasing in Flies 339

a)
+I T5,ooo
o.s_ .
b)
.
T +5,000

.tuf . . f , % 4

(1 , ~ o~o~t o_ 9 ,
.

-180~ ""'~.~_F" " ' " 9 +5,000Os-1

t I

, ,:,i"iI!?'"
, l I
*. ; -:.~'..." 9 .. 9D |
ee (,oe
".:~""~-o
5,00

C)
O.Jf
I+2' 9
~ d)
T+2,500
Wf t z's'l "
Q 9 9

. =;
9 "9
, - , l
~'I
' ,
9 Oe
l
+40
:
~ . . . . . .

. . . ' .:l.i.. ~
. .
I "'=, ~ 1

Fig. 6a--d. Scatter diagrams showing the relationships between co! and Oe, and
% and col for all points measured on Fig. 4 and 3 shorter chases (a and b), and for
only those points for which the error angle (0e) lay between + 3 5 ~ and --35 ~
(c and d). In all cases the delay has been taken as 30 ms, i.e. the ordinate of each
point corresponds to an instant 30 ms later than that of its abscissa. The correlation
coefficients are a) 0.76 (p<0.001), b) 0.09 (n.s.), c) 0.38 (0.01< p < 0.02), d) 0.65
( p < 0.002)

Interestingly, there was no correlation between 0e and coI which

indicates t h a t ar.gular position error does n o t control angular acceleration
towards the leading fly, only angular velocity. There was, however, a
strong correlation between we and (51, which is not surprising since these

489
340 M.F. Land and T. S, Collett

-2,500
o.s4
oo"

, , ~
ee
Fig. 7. Relation between o~l and 0e under conditions that approximate to a steady
state. The points are the average values of o~! measured during periods of 60 to
100 ms, in which the values of 0e were approximately constant. Closed circles are
from Fig. 4, other symbols are taken from four other chases. The straight line has
a slope of 20 s-1

are the first derivatives of 0 e and ~oi, between which there is a relationship.
The question arises as to whether one of these pairs of relationships is
causal, and the other its consequence. I t seems hkely t h a t the Oe, o~l
relation is the real one, because there is no reason w h y the differentiated
relationship should be " z e r o position error seeking." I n other words, if
o~e was actually being used to control d)1 one would not expect the plot
of 0~ vs. r I to pass t h r o u g h zero, which it does.
There is a problem in accepting the value of 21 s -1 as the value t h a t
the fly actually uses in converting 0~ into angular velocity o~t because
of the u n c e r t a i n t y a b o u t the exact latency of the response. Ideally, one
would like to be able to measure this conversion factor u n d e r steady
state conditions, t h a t is, to measure the angular velocity of the chasing
fly when 0~ is k e p t constant. I n free flight this is n o t strictly possible
because the angular velocity of the chasing fly tends to reduce 0e v e r y
rapidly. H o w e v e r during certain manoeuvres, like the loop between
5 and 6 in Fig. 4, a situation does arise in which 0 e stays a p p r o x i m a t e l y
the same for 3-5 frames. F r o m such situations a reasonable approxi-
m a t i o n to the steady state conversion factor can be obtained b y plotting
the average values of w I during these semi-stable periods against the
average value of 0~ in the period t a k e n as starting 30 ms earlier. Such a
plot is shown in Fig. 7. I t is well fitted b y a straight line whose slope is
20 s -1 (i.e. 20 ~ s -1 per o 0~).
F r o m these considerations, the operation performed b y the pursuing
fly can be defined as the conversion of the position of the retinal image

490
 Control of Chasing in Flies 341

of the target (the leading fly) into the angular velocity of the pursuer's
flight path. Or :
to] = k 0e, where k--~ 20 s -1. (1)
This, however, is not the only mechanism involved. An examination
of certain of the turns in Fig. 4, notably the right turn following 2 and
the left turn after 4, shows t h a t the chasing fly begins to turn in the
direction of motion of the leading fly, even before the leading fly has
crossed the chasing fly's axis. For such turns the above expression can-
not be true, and one would expect instead that there would be a closer
relation between error angular velocity (toe) and the angular velocity of
the chasing fly (tot)" Fig. 6b shows t h a t in general there is no such
relation between tot and toe: the correlation coefficient is 0.09, and not
significant. However, on closer examination it can be seen t h a t there is
a cluster of points near the origin t h a t appear to be correlated, and b y
breaking the distribution down with respect to 0~ it turns out t h a t nearly
all these points are derived from situations in which the leading fly is
close to the axis of the chasing fly. There is a highly significant correlation
between tot and toe when 0e is close to =t=30~ and between these values,
but this correlation disappears when 0e is ~= 40 ~ or more. This is shown
in Fig. 6 c and d, which demonstrate the virtual absence of a relationship
between 0e and tot in this central region, and a strong correlation (0.65,
p ~0.001) between to1 and toe- The latter relation is possibly sigmoid,
with the response saturating at values of toe around 1500 ~ s-1, although
on the basis of the present data this can only be a guess. I n any event,
this relation is not badly fitted b y the regression line, which has a slope
of 0.7, and this value will be adopted.
This can be regarded as meaning t h a t when the leading fly is more or
less directly in front, its relative velocity rather than its position determine
the angular velocity, and hence the course of the chasing fly. This can
be expressed:
tot--~ k'r where k' ~--0.7, (2)
provided t h a t 0e lies between q-35 ~
The significance of this mechanism will be examined in more detail
in the discussion section, where an a t t e m p t is made to simulate a chase.

Response Time
Inspection of Fig. 5 indicates a delay (d) of 20-40 ms between the
appearance of the leading fly at a particular angular position 0e, and
the attainment of the appropriate angular velocity tot b y the following
fly. To examine this further the cross-correlation coefficients between
0r ~nd 0)1 were determined for a range of time differences between the
two measurements from - - 3 0 ms (to] leads 0e) to Jr 70 ms. These are

491
342 M.F. Land and T. S. Collett

~ 0

-20 0 +20 +40 +60ms

t (wf)-t(ee)
Fig. 8. Correlation (R) between to/and 0e for different time intervals between the
two sets of measurements. Positive time differences mean that the 0e measurements
lead the to! measurements, by the times shown. The points are derived from the
same measurements as those used in Fig. 6. The line has been fitted by eye. The
maximum correlation occurs at about + 30 ms

plotted in Fig. 8. The results confirm t h a t the m a x i m u m correlation

occurs after a delay of 20 ms and it remains high up to 40 ms, t h e n falls
steeply. The positive correlations t h r o u g h o u t the entire period shown
merely reflect the fact t h a t both flies t e n d to be engaged in the same
m a n o e u v r e for around 100 ms and a casual correlation between the two
variables is to be expected within this range.
The simplest interpretation of Fig. 8 is t h a t the fly has a fixed
response time of a b o u t 30 ms. F u r t h e r evidence from simulations (see
discussion) indicates t h a t the response time cannot be m u c h greater t h a n
this: if it is as long as 40 ms the course would become a series of tortuous
knots (Fig. 12).
A similar correlation examination for the ~oI vs. eoe relationship
[Eq. (2)] showed t h a t this too has a delay of about 30 ms.

Angular Velocity and Eorward Velocity

A problem in trying t o analyse and predict the course of the chasing
fly is t h a t angular velocity and forward velocity are not independent of
each other. There is an obvious slowing down during large angle turns
in Fig. 4, which is shown more clearly in Fig. 9. This is a plot of forward
velocity vs. angular velocity, both measured in the same 20 ms period,
and shows b o t h t h a t there is considerable variation in the horizontal

492
 Control of Chasing in Flies 343

2 ]
m.s-
9 99

~f ~o~m~o'~ 9
"" .

9 9 o o~

0_1 1 I I I I
0 5~O00~ "1
LOf
Fig. 9. The relation between the forward velocity v t and angular velocity eot of the
chasing fly measured during each interval between points in Fig. 4. The angled
line is an arbitrary attempt to fit the data, and this line is used to adjust course
lengths in the simulations in Figs. 11-13

component of forward speed (v/) at low angular velocities (less t h a n

1000 ~ s-l), and also t h a t at angular velocities above this the forward
speed drops rapidly, and b y extrapolation would reach zero at about
5000 ~ s-1. There m a y be several reasons for this, but one is the reduction
of net forward thrust during turning caused either by the decreased stroke
angle on the side turned to (Nachtigall and Wilson, 1967), or the braking
effect of retaining the upstroke attitude of the wing on the side of the
turn during the downstroke (Faust, 1952). Some of the variability of the
forward speed at low angular velocities m a y be the result of the time
t a k e n to accelerate after turning, and part m a y be due to departures of
the flight p a t h from the horizontal. There is, however, no way of distin-
guishing these factors from the flight path records, and for the purpose
of simulating chases the relation between vI and co/ has simply been
approximated b y the arbitrarily fitted line shown on Fig. 9.
I n addition to the dependence of v] on co/, v / i s also likely to affect
cot because the fly's forward m o m e n t u m will tend to limit the rate at
which large turns can be made. There is some indication from the films
t h a t the fly actively counteracts this on very sharp turns b y actually
turning round before its forward velocity has reached a minimum, so
t h a t for about 20 ms it is in fact moving sideways. However, the quality
of the films was not good enough to examine this in more detail.

493
344 M.F. Land and T. S. Collett

Other Evidence Supporting a Continuous Relation between O~and col

While it is most unlikely t h a t the relationship in Fig. 6a could have
arisen fortuitously, there is still the remote possibility that it represents
the consequence of the chasing fly following a set of pre-set rules, in
which each manoeuvre of the leading fly is matched b y one from an
equivalent repertoire of the chasing fly. I t is in fact the case t h a t all but
one manoeuvre b y the chasing fly in Fig. 4 (the exception being the turn
between 3 and 4) has the same general form as the manoeuvre of the
leading fly. Thus it is just possible that, provided the chasing fly can
somehow detect the kind of manoeuvre made b y the leading fly, it can
control its flight b y matching it.
An example which seems to disprove this conclusively is shown in
Fig. l0 in which the manoeuvres made b y the two flies are completely
different. The leading fly makes a right turn followed b y a complete left
hand loop, whereas the chasing fly makes a right turn followed b y a
second very sharp right turn, and finally a sharp left turn. For much of
the flight the two flies are actually turning in opposite directions. Never-
theless, when one compares co/and 0e (Fig. 10b) it can be seen t h a t there
is still a nearly perfect correspondence between the two, in spite of the
speed with which 0r is changing. Fig. 10c shows the excellent relation
between 0~ and co! 20 ms later. I t is not believable either t h a t the chasing
fly's manoeuvres are simply replies to those of the leading fly, nor t h a t
the relationship in Fig. 10c is accidental.

Discussion
a) A n Attempt to Simulate a Chase
A valuable test of whether or not the behaviour of the chasing fly is
adequately described b y the relationships given in Eqs.(1) and (2) in
the results is to use these expressions to a t t e m p t to simulate a chase. If
the manoeuvres predicted are similar to those t h a t occur, and the overall
result is the same, then it can be concluded t h a t Eqs. (1) and (2) provide
a sufficient explanation.
Adding the two equations, and incorporating the delay (d) gives:
~o/(t+~) ~ b 0e(t) ~ b' o)~(t) (85~ > 0r > --36 ~ (3)
and this, together with the relation between forward speed (v]) and
angular velocity (w]) shown in Fig. 9, is used to reproduce the flight p a t h
of the chasing fly, given t h a t of the leading fly. I n the simulation~ shown
in Fig. 11 the p a t h of the leading fly is t a k e n directly from the chase
shown in Fig. 4, so t h a t direct comparison can be made.
The technique used is to start with the first three points of the chasing fly's
track as the given initial conditions (t = t, t ~ 2 0 and t-I-40 ms) and then, by
measuring 0e at t ~- 10 ms, to calculate the course change that must be made at
t ~ 40 ms, if the value of the delay (d) is to be 30 ms.

494
 Control of Chasing in Flies 345

2 a,

b)
+1o6 +5,000- c)
:S-1

Oe Wf

-lod J
_1~0 o 1 O O !

+ 2,500 QO

0e
~.S-1

OJf
9 -5,000

-2,500

Fig. 10a---e. Instance of a chase in which the two flies execute quite different
manoeuvres, b u t in which the relation between co! and 0e is preserved, a) Record
of ~he chase, b) Plots of 0e and o 1. Marks on the time axis correspond to the num-
bered points on the tracks, and are 200 ms apart, e) Relation between w / a n d 0e
measured 20 ms earlier

Since k O v : r (dOf)/dt, the required change b0! is given b y kOe 9(~t, where
6t, the sampling interval, is in this case 20 ms. Similarly, where the value of 0e is
appropriate, an angle proportional to we can be included b y measuring the dif-
ference between the values of 0e a t t ~ 20 and t = 0, multiplying this b y /r and
adding this to course change already obtained. This is the required change because,
if eel= k' o~e, t h e n (dO/)/dt = k" (dOe)/dt , or (~Ol= k" 30 e. This now gives the direction
23 E. comp. Physiol., Vol. 89

495
346 M.F. Land and T. S. Collett

o2 the line between the points at t + 40, and t + 60 ms and this is then drawn with
a protractor. The length of this line will be vt- ~t and by assuming that (o] is equal
to the course change divided by ~t, the appropriate value of v] can be obtained
from Fig. 9. This then establishes the fourth point in the course (t + 60 ms). By
repeating these operations (which are simpler than they sound) for successive points,
the complete course can be built up. Different values of/c and/c' can be used, and
with slight modifications to the method as described the delay (d) and sampling
time (~t) can be altered. The relative speeds of the two flies can also be changed by
multiplying the values of vI obtained from Fig. 9 by a constant factor. In general
the smaller the sampling time (~t) the better, but for most purposes 20 ms was
found to be short enough not to introduce appreciable distortion.
I n the simulation shown in Fig. l l b , / ~ 2 0 s -1, the delay is 20 ms
and the velocity t e r m has not for the m o m e n t been included. I t can be
seen t h a t there is a reasonably good m a t c h to the actual flight p a t h
shown in Fig. 11 a. E a c h manoeuvre of the " r e a l " pursuing fly is dupli-
cated b y a corresponding manoeuvre of the simulated fly, even t h o u g h
these differ in detail. The principal difference between the real and
simulated fly is t h a t in the former the turns are all rather tighter, and as
a result the simulated pursuer appears to s t r a y rather further from the
leading fly's course t h a n it should. Nevertheless, i~ one compares the
relative positions of the flies at the m a r k e d instants 3, 4 and 5 it can be
seen t h a t the simulated fly is not m u c h further behind t h a n the real fly.
I n the final loop the leading fly does "escape," in b o t h cases, because
the distance separating the flies comes to exceed the visual limit of
10 era. I t is concluded from this t h a t the postulated " b a s i c " mechanism,
in which error angle is converted b y the fly to angular velocity, is able
to explain most features of the observed chase, although capable of
refinement.
Before considering the effects of adding in the error velocity (/c')
term, it is interesting to see w h a t the effects of changing/c and d are,
a n d i m p o r t a n t to show t h a t changing t h e m makes the simulation worse
in the sense t h a t it looks less like the actual course. Fig. 12 shows a
n u m b e r of simulations of the first two manoeuvres of the same course
as iu Figs. 4 and l l a . I n Fig. 12a it is seen t h a t reducing the delay to
zero makes the course intolerably good, so t h a t the trailing fly has caught
up with the leader several times after only 0.4 s, and there is virtually
no overshoot on the first bend. E q u a l l y dramatically, with a 40 ms delay
the fly overcorrects wildly on the first b e n d producing a figure of eight
course which wastes time and lets the leading fly get well in front. I n
Fig. 12b, halving the value of/c from 20 to 10 s -1 (20 ms delay) makes
the first b e n d m u c h too wide, while increasing /c b y a factor of 1.75
clearly makes the b e n d m u c h too tight and causes some overshoot. F r o m
these and similar results it is concluded t h a t realistic simulations of the
flight p a t h require t h a t b o t h the delay and the constant/c s t a y close to
the values a d o p t e d in deriving Fig. 11 b.

496
 Control of Chasing in Flies 347

I a) 1!

c)
I

Fig. 11 a--c. Attempts to simulate the chase shown in Fig. 4. The course of the
leading fly is taken from Fig. 4, and the course of the chasing fly is plotted by the
method given in the text. a) Actual chase; b) Simulation using only the 0e, eoI
relationship: k = 2 0 s -1, d = 2 0 m s , e) As b) but incorporating the relationship
between me and co/: k = 2 0 s-1, k ' = 0 . 7 , d = 3 0 ms. The sampling time used in the
simulations was 20 ms in b) and 10 ms in c) although for clarity only every second
point is shown
23*

497
348 M.F. Land and T. S. Collett

k= 2 5 s -1

d = 20ms

Fig. 12. Effects of changing the parameters I: and d on the form of the simulation
of the first two bends in the course shown in l~ig. 4. All courses start in the same
way, and the cross line indicates where they would have reached when the track
of the leading f i t ends. I t can be seen that reasonable approximations to the actual
flight path are only obtained when I~ and d are close to values estimated in the
results

T h e inclusion of t h e error v e l o c i t y t e r m (Fig. 1] c), while n o t m u c h

i m p r o v i n g t h e overall a p p e a r a n c e of t h e course, does h a v e t w o v e r y
i m p o r t a n t effects. F i r s t l y i t " s t a b i l i s e s " t h e s i m u l a t e d fly s o m e w h a t ,
e s p e c i a l l y on a b r u p t turns, a n d t h e i n d i r e c t effect of this is t h a t t h e d e l a y
a d o p t e d in t h e s i m u l a t i o n can be increased from 20 to 30 ms, w i t h o u t
t h i s p r o d u c i n g m u c h overshoot. A 40 m s d e l a y is, however, still un-
a c c e p t a b l e . This is a n i m p o r t a n t p o i n t because t h e d e l a y o b t a i n e d in t h e
r e s u l t s section (Fig. 8) seems to be n e a r e r 30 t h a n 20 ms, a n d also from
a physiological p o i n t of view 30 ms is a l r e a d y a v e r y s h o r t t o t a l response
l a t e n c y , a n d 20 ms p r o b a b l y unrealistic. The second feature, which is
b e t t e r i l l u s t r a t e d in Fig. 13, is t h a t t h e inclusion of some v e l o c i t y com-
p o n e n t m a k e s i t possible to s i m u l a t e some t u r n s which a r e n o t even
q u a l i t a t i v e l y correct w i t h o u t it. :In ~he real chase, d u r i n g t h e r i g h t t u r n
b e t w e e n p o i n t s 2 a n d 3, t h e chasing fly h a s a c t u a l l y b e g u n t o t u r n r i g h t

498
 Control of Chasing in Flies 349

~ fse

k'2~ i
d=3Oms, k=2Os4,k =0"7 '

Fig. 13. The effect of including the error velocity term (Ic'eOe)on simulation of the
second bend in the chase shown in Fig. 4. The middle plot shows the path without
this term, and the lower left the improvement resulting from its inclusion. The
overshoot is reduced and the turn "anticipated" as in the actual track (upper right)

while the leading fly is still on the left, though moving right. A similar
situation occurs in the left turn after 4, and the right turn into the loop
after 5. This clearly cannot be simulated with only the position error
system, but as shown in Fig. 13 the inclusion of the angular velocity
term makes this possible b y opposing the effect of the position term while
the leading fly is approaching the chasing fly's line of flight, and aug-
menting it once the line of flight has been crossed. Thus an effect t h a t
almost looks like anticipation is explicable.

b) Comments on the Nature o] the Control System

The Geometry of Chasing Behaviour
I t is clear t h a t b y turning towards the leading fly the chasing fly will
tend to follow it, and if fast enough to overtake it. However, this cannot
be viewed simply as a feedback system in which the leading fly provides
the input, and the following fly the correcting output. The reason is t h a t
changes in the course of the following fly have two consequences: one is
to reduce the error angle (0r) but the other is to alter the respective
positions of the two flies, which changes the angle of the li~e E L and
hence the angle r (Fig. 14) in very complicated ways. This second effect
is particularly important when the two flies are close together, as a
small displacement of either fly can then result in a very rapid change
in 0e, and several instances of this can be seen in Fig. 4. I n Fig. 14 an

499
350 M.F. Land and T. S. Collett

v, g -~_ Oe
L Of
xfyf I

...... l . X

Fig. 14. Left: the geometrical relations between the leading (L) and following
fly (F) at an instant during a chase (represented here as two dimensional). Right:
diagram showing the double effect on 0e of changes in the angular velocity of the
following fly (col). The heavy line indicates the straightforward feedback loop in
which changes in course angle of the following fly (01) reduce the error angle (0e).
The thin lines indicate that eoI also affects the relative positions of the two flies at
time t, and hence the angle ~ of the line joining them (LF). In this diagram the
relation between coe and w! and the delay have been ignored

a t t e m p t has been m a d e to illustrate the complete system for the simpli-

fied case of two dimensional flight. The essential point is t h a t the angle
r is the angle of the line between F and L at a n y instant, and the co-
ordinates of these points are the resolved vector sums of all the elements
(o~. dt, v . dt) t h a t went to make up the respective courses since the
beginning of the chase. These vector sums are represented b y the two
integrals Ln the outer loop in Fig. 14. The problem of working out the
effects of changes of co! on r the input to the " s t r a i g h t f o r w a r d " loop,
are b e y o n d the scope of this paper, and will be ignored. However, this
does m e a n t h a t the conclusions which follow are only strictly applicable
to conditions in which the two flies are well separated, and the effect of
o~f on the angle r is small.
Stability
I n feedback systems containing delays there is always the potential
problem of instability. I t is therefore worth inquiring whether there is a
special reason w h y the fly should use the particular value of k (20 s -1)
t h a t is observed. I n a system like this there will be a value of k above
which a n y a t t e m p t e d manoeuvre will result in continuous oscillations
of increasing amplitude. I t can be shown theoretically and graphically
t h a t this occurs when k > II/2d, or in this case (d = 3 0 ms) when k > 5 2 s -1.
F o r lower values of k the fly will oscillate after each course change, b u t
these oscillations will gradually die out. At lower values still there is no
oscillation, and the new course is smoothly, b u t relatively slowly ap-
proached. The ideal value for k would p r e s u m a b l y be t h a t at which there

500
 Control of Chasing in Flies 351

were just no oscillations; this is the "aperiodic limit" and is given b y

lc -= e - l / d , i.e. 0.368/0.03 ----12.3 s-1, if d ----30 ms. The value of/c used b y
the fly is slightly greater t h a n this, and indeed simulations like t h a t in
Fig. l l b using a value of d of 30 ms do show slight " h u n t i n g " around
each new course ; if d = 40 ms (see Fig. 12) these overshoots become very
prouounced indeed. However, for a step input graphical estimates show
t h a t the size of the overshoot when /~~ 20 s -1 and d ~-30 ms is only
about 7%, i.e. the sudden appearance of an object at i00 ~ from the
fly's axis would result in an initial turn of 107 ~ and this error is probably
not important. I n any case, the small amount of instability introduced
is effectively damped out by the error rate feedback [the /~'coe t e r n
in (3)] t h a t operates in the region close to the axis. From these con-
siderations it seems t h a t the operation of the chasing fly is critically and
elegantly adjusted to its task--turning as fast as possible without
becoming unstable.
An interesting sideline to this is t h a t one would expect d a n d / c to
v a r y inversely with temperature, if this critical relationship is to be
maintained. I t would be interesting to know whether cold flies can chase
effectively, or whether the breakdown of the critical conditions for
effective chasing abolishes this kind of behaviour.
Relations between Torque, Acceleration and Velocity
Interesting questions arise out of the way the chasing fly changes
from one angular velocity to another. Clearly it must accelerate or de-
celerate in order to do so, and yet acceleration does not seem to be under
the control of 0,, but rather angular velocity itself. This apparent contra-
diction is resolved if we accept t h a t the initial output of the fly is torque
(rotational force about a vertical axis) and consider what happens to
t h a t torque. As l~eichardt has pointed out (l~eiehardt, 1973 ; Poggio and
Reichardt, 1973) the fly's rotational dynamics can be approximated by:
T = I d~ + F co,
where T is the torque (dyne 9cm), I is the moment of inertia of the fly
(g 9cm 2) and F the aerodynamic friction opposing rotation (gm 9cm 2 9s-l).
I n other words, at any instant part of the torque is taken up in accelerat-
ing the fly and part in overcoming air friction. If I is very large compared
with F one would expect t h a t co would correlate well with T (which we
are assuming is linearly related to 0e). The reverse, however, seems to be
the case: 0e and hence T correlate well with co/and not cb! so F must be
large compared with I.
For a step change in torque, the above equation is solved by:
co : T / F (1 -- e-t/(z/F)).
This expression indicates t h a t when the fly produces torque, it ac-

501
352 M.F. Land and T. S. Collett

celerates to an asymptotic angular velocity (T[F) with a time constant

I[F (s), and this time constant specifies how long the fly will take to
reach its final angular velocity. I n general, if I/F is similar to or longer
than the periods of movements during a chase, the chasing fly will show
instabilities of tracking, and attempts to simulate its behaviour will be
poor. I t is easy, b y graphical construction, to determine how oJ1 changes
with time when 0e changes, for different values of I/F, and these con-
structions can be compared with the " r e a l " results shown in Figs. 5
and 10b. The results of this exercise show that constructed and actual
time courses of ~of only begin to become importantly different when I/F
is greater than about 20 ms, and that any value of I/F below this value
will give a good approximation to the courses shown in Figs. 4 and 10a.
I t is interesting that Poggio and Reichardt (1973) have estimated
indirectly, b y methods quite different from those given here, that the
value of I]F for female Musca lies between 4 and 8 ms. There is no
reason for not accepting their conclusions, since Fannia and Musca are
of similar size, and both I and F are related to linear dimensions, and
not to any property of the control system governing a particular form
of flight.
These times mean in practice that although the course of the fly is
similar to that of one with zero moment of inertia, there will be some
delay between the application of torque and the attainment of maximum
angular velocity. These delays will vary with the kind of manoeuvres
made, but from the graphical reconstructions of Figs. 5 and 10b they
appear to be in the range 0.5 to 1.0 times the time constant (I/F). Using
Poggio and Reichardt's estimates this suggests acceleration delays
between 3 and 6 ms, and they should not in any case be greater than
15 ms, using the maximal estimate of I/E of 20 ms.

c) PhysiologicalIn/erences
If male flies track each other in the way that has been described,
they must have neural mechanisms for converting certain specific kinds
of stimuli into appropriate output patterns. The response time of the
fly is very rapid, about 30 ms, which includes every process from seeing
the stimulus to accelerating to an appropriate velocity. Thus it seems
likely that number of neural operations involved will be small, and that
it is sensible to inquire how the systems might be organised with as few
neurones as possible. The two operations that need explanations are
(i) the conversion of stimulus position into angular velocity, and (ii) the
conversion, in the anterior regions of the eye, of stimulus velocity into
angular velocity. For the moment we shall assume that there exist
neurones passiag to the thorax whose discharge rate specifies the fly's

502
 Control of Chasing in Flies 353

velocity ~ ~ position
system o~ system

Fig. 15. Possible scheme by which position and velocity information might be
extracted. The semicircles represent the two eyes, and it is assumed that each
contains an array of small-field movement detectors (represented by arrows). On
the right these feed onto interneurones (a) whose effect on the output neurone (b)
increases in the numbered order shown, from anterior to posterior. On the left the
movement detectors feed onto two units which collect selectively from the clock-
wise and anti-clockwise "on"-direction sets. These units in turn connect to the
outpu~ neurones. For further explanation see text

u l t i m a t e a n g u l a r velocity, a n d c o n c e n t r a t e on h o w t h e i n p u t s to t h e s e
n e u r o n e s m i g h t be a r r a n g e d .
A suggestion as t o h o w t h e first o p e r a t i o n m i g h t be p e r f o r m e d is
given i~ Fig. 15. This assumes t h a t t h e r e are neurones which d e t e c t
m o v e m e n t a n d collect f r o m r e s t r i c t e d regions of t h e eye (a). These feed
o n t o a collector n e u r o n e (b) which could be t h e n e u r o n e t o t h e wing
muscles. T h e (a) neurones m u s t h a v e t h e following properties.
(i) T h e y m u s t r e s p o n d o p t i m a l l y t o quite small objects. D u r i n g t h e
f i l m e d chases t h e average s e p a r a t i o n of t h e t w o flies was 3.7 cm, a n d
t h e r e was no sign of loss of visual c o n t a c t a t distances less t h a n 10 cm.
T h e r e l e v a n t stimulus angles on t h e r e t i n a for a 6 m m long f l y a r e 9.3 ~
a n d 3.4 ~, for t h e longest dimension of t h e stimulus.
(ii) T h e y m u s t be c a p a b l e of r e s p o n d i n g to objects m o v i n g a t a n g u l a r
velocities u p t o a t l e a s t 2 500 ~ s -1. This is a c o n s e r v a t i v e e s t i m a t e b a s e d
on t h e d i s t r i b u t i o n of values of e% (Fig. 6) which a c t u a l l y e x t e n d to
twice t h i s figure. I n t h e case of a 5 ~ o b j e c t passing o m m a t i d i a s p a c e d a t
2 ~ this r e p r e s e n t s a t o t a l "exposure t i m e " of a b o u t 2 . 8 m s p e r ore-

503
354 M.F. Land and T. S. Collett

matidium, which is not incompatible with flicker fusion frequencies of

around 300 s -1 (Calliphora; Autrum, 1950).
(iii) These cells must be able to distinguish background movement
from target movement. During chases the chasing fly turns at angular
velocities that cover the same range as the target velocity (Fig. 6b), and
so the cells must be capable of ignoring movements of large patterns
over the same range of velocities as they respond to small objects.
(iv) The output of these cells should not be a ftmction of stimulus
velocity, or stimulus size. Ideally they should fire at maximum velocity
if there is a target in their field of view, and not at other times. Finally,
the outputs of each (a) cell must be " w e i g h t e d " according to its location,
so that the (b) cell fires in proportion their position. This could be
achieved either by a gradation of the firing rate of the (a) cells according
to their position, or by differential weighting of their synaptic effec-
tiveness on the (b) cells. I t would be premature to speculate at present
whether any of the visual interneurones so far recorded in dipteran visual
systems have properties compatible with the (a) cells postulated here,
not least because most workers have used females for electrophysiological
work, and except for the predatory Asilidae and Empidae (Richards,
1927), females do not track other flies. The small-field non-directional
interneurones described by Mimura (1971) in the medulla of flies have
some of the properties required, notably an indifference to the actual
velocity of target motion (iv above), although the angular velocity range
given (0.2-2 rad s-1) is much smaller than that required here (up to
45 tad s-l). McCann and Dill (1969) describe what are probably the
same fibres as those Mimura found (class In) but they say these have
minimum latencies in the range 20-30 ms, which is too great for use in
a system whose total latency is 30 ms. Visual interneurones with the
essential property of distinguishing small target motion from background
motion (iii) have been found in moths (Collett, 1971, 1972) but not so far
in flies. Perhaps this emphasises the need to look for electrophysiological
properties of neurones in stimulus situations that are behaviourally
realistic, and in this case it is likely that realistic conditions will be sex
specific.
The other operation, converting angular velocity to output, is a
familiar one, since it is similar in many ways to the optomotor response
studied by Fermi and Reichardt (1963), and for which there already
exist plausible neural candidates in the class I I and class IV units (the
latter descending to the thorax) described by Bishop et al. (1968) and
McCann and Foster (1971). These unidirectional motion detectors are
maximally sensitive to movements occurring about 20 ~ from the fly's
axis, which would account elegantly for the positional restriction on the
angular velocity response seen in chasing behaviour. However, these

504
 Control of Chasing in Flies 355

units, like the (a) units postulated above, must also meet the condition
t h a t t h e y distinguish small target motion from background motion. I t
could simply be t h a t a small target ahead of the animal is a more powerful
stimulus to these cells t h a n the rest of the field moving in the opposite
(or the same) direction, but this is not yet clear. I t is also possible that
the velocity system (Fig. 15) is distributed over the whole eye, but t h a t
it is only detectable in front, where the position system has its smallest
effect.
Finally, the total response delay raises difficulties. Using available
data, we have McCann and Dill's (1969) estimate of 20 ms as the mini-
m u m latency for visual interneurones. Mulloney (1969) has shown in
Calliphora t h a t the total conduction time from optic lobe stimulation to
the appearance of an action potential in the "take-off jump muscle"
(tergo-trochanteral muscle) is only 3.6 ms. This is agreeably short and
includes one central synapse and a neuro-muscular junction. Heide
(1968) found t h a t the direct (non-fibrillar) flight muscles likely to be
responsible for yaw give a fused tetanus when stimulated at 60 Hz,
which suggests t h a t contraction and relaxation times cannot be much
less t h a n 17 ms (Calliphora). Acceleration time has already been discussed
and could be as short as 2 ms. E v e n using these optimistic estimates, the
total delay should be at least 43 ms, compared with the 30 ms found
here. Either visual latencies are rather shorter than existing measure-
merits suggest, or direct flight muscles operate more rapidly, or both.

d) Function o/Chasing
Male flies of m a n y species will chase any moving object of about the
right velocity and dimensions. Houseflies (Musca and Fannia) will chase
both males and females of their own or other species: for example it is
quite common to see Musca chasing CaUiphora twice their size. I t is
easy also to lure Musca off a wall, or Fannia from their patrolling flights
b y throwing small objects past them (e.g. peas). Provided the missile
does not pass too close, the animal's response is to fly towards it, and to
follow its trajectory for a while. This is thus pursuit and not evasion.
Greenbottles (probably Lucilia sp.) are especially spectacular. They will
sit on a particular leaf, and respond to small (1 cm) stones thrown just
above them b y darting up to ir~tercept, and then flying with the stone
for several metres before returning to the same leaf. One fly repeated
this performance 27 times in succession. Interestingly, a tendency to
chase thrown or flying objects turns out to be an excellent indicator of
sex. We have never seen a female muscid or calliphorid fly chase any-
thing.
Chasing thus seems to be confined to males and the stimuli t h a t
elicit it are not very specific. There are really only two plausible rune-

505
356 M.F. Land and T. S. Collett

tions for this behaviour: s e x - - " marriage b y c a p t u r e " to use Richards,

(1927) p r e t t y expression--or some kind of territorial defence. Terri-
toriality implies either the defence of a breeding site, or a food-gathering
locality, and neither of these possibilities fit the fly's situation. Fannia
neither feed nor deposit eggs near lampshades. The function of male
pursuits m u s t t h e n be to catch females, and indeed this often happens.
And, since males of one species all have preferences for the same kind
of locality, m a l e - - m a l e chases are also inevitable. Although proof of
this is lacking, males p r e s u m a b l y m a n a g e to escape from each other, b u t
females do n o t - - e i t h e r because t h e y are slower or less good at evasive
tactics.
The mechanism of chasing, as described in this paper, raises the
interesting possibility t h a t not only sex selection b u t also species selection
m a y be in a sense " c o d e d " in the chase itself, l~or example, a small slow
fly will never catch a large fast fly, b u t equally the fast fly with con-
siderable m o m e n t u m will n o t be able to catch the slower one if the latter
always makes a sharp t u r n as the fast fly approaches. Appropriate
m a t i n g could thus arise if the evasion s t r a t e g y of the female was m a t c h e d
to the manoeuverability of the " c o r r e c t " male, and t h e r e b y provide an
effective species sorting system t h a t operates prior to actual contact.
Given the large n u m b e r of different fly species to be found in gardens
in summer, some mechanism of this kind seems an attractive possibility.

We are very grateful to Alan King, Peter Slater and John )/Iaynard Smith
for critically reading the manuscript. This work was supported by a grant from ~he
Science Research Council of the U.K.

References
Autrum, H. : :Die Belichtungspotentiale und das Sehen der Insekten (Untersuchun-
gen an CaUiphora und Dixippus). Z. vergl. Physiol. 32, 176-227 (1950)
Bishop, L. G., Keehn, D. G., h[cCann, G. D. : Studies of motion detection by inter-
neurones of the optic lobes and brain of the flies Calliphora phaenicia and Musca
domestica. J. Neurophysiol. 81, 509-525 (1968)
Collett, T.: Visual neurones for tracking moving targets. Nature (Lond.) 2112,
127-130 (1971)
Collett, T. : Visual neurones in the anterior optic tract of the privet hawk moth.
J. eomp. Physiol. 78, 396-433 (1972)
Colyer, C. N., Hammond, C. O. : Plies of the British Isles. London: Warne 1951
Faust, R.: Untersuehungen zum Halterenproblem. Zool. Jb., Allg. Zool. Physiol.
63, 325-366 (1952)
Fermi, G., Reiehardt, W. : Optomotorische Reaktionen der Fliege Musca domestica.
Kybernetik 2, 15-28 (1963)
Heide, G.: Flugsteuerung durch nicht-fibrill~ire Flugmuskeln bei der Schmeil3fliege
Calliphora. Z. vergl. Physiol. 59, 456-460 (1968)

506
 Control of Chasing in Flies 357

McCann, G. D., Dill, J. C. : Fundamental properties of intensity, form and motion

perception in the visual nervous systems of CaUiphora ~haenicia and Musca
domestica. J. gen. Physiol. 58, 385413 (1969)
McCann, G. D., Foster, S. F. : Binocular interactions of motion detection fibers in
the optic lobes of flies. Kyhernetik 8, 193-203 (1971)
Mulloney, B. : Interneurones in the central nervous system of flies and the start of
flight. Z, vergl. Physiol. 64, 243-253 (1969)
NachtigM1, W., Wilson, D. M. : Neuromuscular control of dipteran flight. J. exp.
Biol. 47, 77-97 (1967).
Poggio, T., Reichardt, W.: A theory of pattern induced flight orientation of the
fly Musca domestica. Kybernetik 12, 185-203 (1973)
Reichardt, W.: Musterinduzierte Flugorientierung der ~liege Musca domestica.
Naturwissenschaften 60, 122-138 (1973)
Richards, O.W.: Sexual selection and allied problems in insects. Biol. Rev. 2,
298-364 (1927)
M. F. Land
T. S. Collett
School of Biological Sciences
The University of Sussex
Brighton, Sussex BIN1 9 QG
England

507


508

LETTERS
Cryptochrome mediates light-dependent
magnetosensitivity in Drosophila
Robert J. Gegear1, Amy Casselman1, Scott Waddell1 & Steven M. Reppert1
Although many animals use the Earths magnetic field for orienta-
tion and navigation1,2, the precise biophysical mechanisms under-
lying magnetic sensing have been elusive. One theoretical model
proposes that geomagnetic fields are perceived by chemical reac-
tions involving specialized photoreceptors3. However, the specific
photoreceptor involved in such magnetoreception has not been
demonstrated conclusively in any animal. Here we show that the
ultraviolet-A/blue-light photoreceptor cryptochrome (Cry) is
necessary for light-dependent magnetosensitive responses in
Drosophila melanogaster. In a binary-choice behavioural assay
for magnetosensitivity, wild-type flies show significant naive and
trained responses to a magnetic field under full-spectrum light
( 300700 nm) but do not respond to the field when wavelengths
in the Cry-sensitive, ultraviolet-A/blue-light part of the spectrum
(,420 nm) are blocked. Notably, Cry-deficient cry0 and cryb flies
do not show either naive or trained responses to a magnetic field
under full-spectrum light. Moreover, Cry-dependent magnetosen-
sitivity does not require a functioning circadian clock. Our work
provides, to our knowledge, the first genetic evidence for a Cry-
based magnetosensitive system in any animal.
The ability of an animal to detect geomagnetic fields has substan-
tial biological relevance as it is used by many invertebrate and verte-
brate species for orientation and navigation purposes, including
homing, building activity and long-distance migration2,4. Three gen-
eral modes of magnetoreception have been proposed5. One mode is
electromagnetic induction by the Earths magnetic field, which may
occur in electrosensitive marine fish, although there is little evidence
to support such sensing. The two other modes, for which experi-
mental evidence does exist, are a magnetite-based process68 and
chemical-based reactions9,10 that are modulated by magnetic fields.
One chemical model of magnetoreception proposes that magnetic
information is transmitted to the nervous system through the light-
induced product of magnetically sensitive radical-pair reactions in
specialized photoreceptors3.
Cry proteins are flavoproteins that have been postulated to gen-
erate magnetosensitive radical pairs that could provide a photoin-
duced electron transfer reaction for the detection of magnetic fields3.
Cry proteins are best known for their roles in the regulation of cir-
cadian clocks11,12 and can be categorized into two groups on the basis
of current phylogenetic and functional relationships13,14. Drosophila-
like Cry proteins are sensitive to light in the ultraviolet-A/blue range15
and function primarily as photoreceptors that synchronize (entrain)
circadian clocks. Vertebrate-like Cry proteins, which have also been
found in every non-drosophilid insect so far examined14, do not seem
to be directly light-sensitive. Instead, vertebrate-like Cry proteins are
potent repressors of the Clock and Bmal1 (known as Cycle in insects)
transcription factors which, as heterodimers, drive the intracellular
transcriptional feedback loop of the circadian clock mechanism in all
animals studied.
1
Department of Neurobiology, University of Massachusetts Medical School, Worcester, Massachusetts 01605, USA.
1014
2008 Macmillan Publishers Limited. All rights reserved
Vol 454 | 21 August 2008 | doi:10.1038/nature07183
Although there is good behavioural evidence for the involvement
of short-wavelength photoreceptors in the detection of a geomag-
netic field5,1618, an essential link between Cry and magnetoreception
has not been established in any animal. Drosophila are ideally suited
to investigate a role for Cry as a magnetoreceptor, because they only
have the light-sensitive Cry14 in which the action spectrum peaks in
the ultraviolet-A range (350400 nm) with a plateau in the near blue
range (430450 nm)19,20. Notably, flies that lack Cry (cry0)21 or har-
bour the chemically induced missense cryb mutation22,23 can be used
to evaluate the role of Cry in magnetosensitive responses.
We initiated our studies by developing a behavioural assay for
magnetosensitivity in Drosophila (Fig. 1a). In this illuminated appar-
atus, flies experience a magnetic field generated by an electric coil
system and display their magnetosensitivity in a binary-choice
T-maze. The two-coil system is ideal for behavioural studies of mag-
netosensitivity, because it produces a magnetic field on one side of
the T-maze, while producing no field on the opposite side. This
design eliminates non-magnetic differences such as heat generated
by the electric coils between sides during test sessions24. Flies were
tested either for their response to the magnetic field in the naive state
(naive group) or after a training session pairing the field with sucrose
reward (trained group).
Wild-type Canton-S, white-eyed w;Canton-S, Oregon-R and
Berlin-K strains all developed a learned preference for a magnetic
field (Fig. 1b). The trained groups in the two Canton-S lines showed
the greatest response to the field (P 5 0.002, one-way analysis of
variance (ANOVA)) and were the only ones to show a naive avoid-
ance of the field (P , 0.0001, one-sample t-test). Thus, Drosophila
consistently show magnetosensitivity that varies in magnitude in a
strain-dependent manner. The similarity of behavioural responses
between red-eyed, wild-type Canton-S flies and white-eyed
w;Canton-S flies shows that eye colour does not substantially alter
behavioural responses to the magnetic field.
Because wild-type Canton-S flies showed the most robust trained
and naive responses of the strains tested, we used them to determine
whether the magnetic responses we observed were light-dependent.
We assayed naive and trained Canton-S flies under different long-
wavelength pass filters that transmitted wavelengths of light
at .500 nm, .420 nm or .400 nm (Fig. 2a). In contrast to flies
assayed under full-spectrum light (Fig. 1b and Fig. 2a), flies did not
show either naive or trained responses to the field when wave-
lengths ,420 nm were blocked (Fig. 2b). Because the filter that
blocked light ,420 nm also caused a 13% decrease in total irradiance
(Fig. 2c, red line), we examined whether the filter-induced lack of
behavioural responses to the magnetic field was secondary to the
decrement in irradiance. When Canton-S flies were studied under
full-spectrum light, with a total irradiance level lower than that
imposed by the filter (Fig. 2c, blue line), the flies still showed signifi-
cant naive (P 5 0.0005, one-sample t-test) and trained responses to
509
NATURE | Vol 454 | 21 August 2008 LETTERS

the magnetic field (Fig. 2d). Thus, the filter-induced loss of beha-
vioural responses to the magnetic field is due to the loss of short-
wavelength light. a
12 Full spectrum 12 >500
Behavioural responses to the magnet were partially restored when
400420 nm light was included (Fig. 2b), which is consistent with the
8 8
action spectrum of Drosophila Cry tailing into the near blue19, and, as

Photons s1 cm2 nm1 (1012)

expected, the trained response was weaker than that under full-spec- 4 4
trum light (full spectrum versus .400 nm, P , 0.001, Students
t-test). This wavelength-dependent effect of the magnetic field on 0 0
300 400 500 600 700 300 400 500 600 700
behaviour suggests that Drosophila has a photoreceptor-based mag-
12 >420 12 >400
netosensitive system. Moreover, because the response to the mag-
netic field requires ultraviolet-A/blue light (,420 nm; Fig. 2e), these 8 8

4 4
a
Train Light source 0 0
300 400 500 600 700 300 400 500 600 700
Wavelength (nm)
b ****
Black box Filter
0.2

Coil
0.1 *

Preference index
12

11 11 9
0
11 11 9
10
0.1
Elevator T-port Training tube

Test
0.2
Full spectrum >500 nm >420 nm >400 nm
c d
Photons s1 cm2 nm1 (1012)

12 0.2 ****

Preference index
0.1
8 11
0
11
4
0.1

b **** 0 0.2
0.2 300 400 500 600 700
Full spectrum
**** e Wavelength (nm) (low intensity)
0.1 *
Photons s1 cm2 nm1 (1011)
Preference index

12
** 20
10 11
0 11
11 11 15
10 9
0.1 10

0.2
Canton-S w;Canton-S Oregon-R
Figure 1 | Behavioural apparatus for magnetosensitivity and behavioural
responses in different Drosophila strains. a, Behavioural apparatus for
magnetosensitivity. The top diagram (Train) shows the frontal view of the
choice chamber apparatus positioned for training. The chamber apparatus
consisted of a training tube, an elevator to transfer flies, and a duel-choice
point (T-port). For training, the apparatus, with training tube only, was
placed upright in an illuminated black box containing a two-coil system. A
population of flies (dots) was loaded into the training tube with or without
sucrose reinforcement and a magnetic field. The bottom diagram (Test)
shows the frontal view of the choice chamber apparatus positioned for
testing. For testing, the apparatus, with tubes attached to the T-port (T-
maze), was rotated to the horizontal and flies were transferred from the
elevator section to the T-port. Wavelength dependence was examined using
long-wavelength pass filters. b, Drosophila strains vary in their behavioural
response to a magnetic field under full-spectrum light (Fig. 2a). Bars show
the preference index of the naive (white) or trained (black) groups. Numbers
represent the groups tested and values are mean 6 s.e.m. *, P , 0.05; **,
P , 0.01; ****, P , 0.0001.
2008 Macmillan Publishers Limited. All rights reserved
5
Berlin-K 0
300 340 380 420
Wavelength (nm)
Figure 2 | Short-wavelength light is required for magnetosensitivity in
Canton-S flies. a, Irradiance curves for different light conditions. Light
measurements were taken inside the training and test tube. Dashed lines
denote cutoff points of the blocking filters. b, Wavelength-dependence of the
magnetic response. Bars show the preference index of the naive (white) or
trained (black) groups. Full-spectrum data are from Fig. 1b. Numbers
represent the groups tested. *, P , 0.05; ****, P , 0.0001. c, Irradiance
curves depicting full-spectrum light (black line), light .420 nm (red line)
and full-spectrum light with reduced total irradiance (full spectrum, low
intensity; blue line). d, Canton-S flies still elicited significant responses to the
magnetic field under full-spectrum, low intensity light. ****, P , 0.0001.
e, Irradiance values from 300420 nm. Data are expanded scale from full-
spectrum pattern in a. The irradiance values in ultraviolet-A/blue light in
our studies (300420 nm) are in line with those reported for Drosophila Cry
function using other biological responses19,20; that is, a range of 1011 to 1012
photons s21 cm22 nm21. Values from b and d are mean 6 s.e.m.
1015
510
LETTERS NATURE | Vol 454 | 21 August 2008

data are consistent with the hypothesis that Cry can function as a
magnetoreceptor in Drosophila.
We next used Cry-deficient cry0 mutant flies to examine directly
whether Cry is required for magnetosensitive behaviour. We tested
two of the newly generated cry0 fly lines, because, in cry0 flies, the
entire cry coding sequence has been replaced with mini-white1 by
homologous recombination, ensuring that, unlike in the more com-
monly used point-mutant cryb flies, there is no possibility of residual
Cry activity21. In addition, the three cry0 fly lines (from cry01 to cry03)
were backcrossed independently into a w1118 background21. Thus, we
magnetic field after at least 5 days in constant light when the flies
were shown to express arrhythmic locomotor behaviour (Fig. 4a), to
have disrupted Period abundance rhythms (Fig. 4b), and to express
constantly low levels of Cry (Fig. 4c). Notably, these arrhythmic flies
continued to show significant naive (P 5 0.004, one-sample t-test)
a
0.20
****
0.15

Preference index
were able to use the appropriate w1118 control flies to test the contri- 0.10
bution of the cry gene in magnetosensitive behaviour.
15
Control w1118 flies showed a clear naive preference for, rather than 0.05
avoidance of, the magnetic field (Fig. 3a). The difference in the dir-
ection of the naive response to the magnetic field between Canton-S 12
0
12
flies and the w1118 line re-emphasizes the importance of controlling
for genetic background in studies of magnetosensitivity in flies. 0.05 w1118 w1118 w1118
Nonetheless, like Canton-S flies, the naive response of w1118 flies to Full spectrum >500 nm >420 nm
the magnetic field was light dependent; the naive preference for the b
magnetic field was abolished in the absence of ultraviolet-A/blue light 0.20
(,420 nm; Fig. 3a). **
0.15 ***
Homozygous cry02 flies lacking Cry did not show a naive response

Preference index
to the magnet under full-spectrum light, in contrast to the significant
0.10
naive responses manifested by both w1118 and heterozygous cry02/1
10
flies (Fig. 3b). Training control w1118 flies to prefer the magnetic field 0.05
11
under full-spectrum light significantly enhanced their naive pref-
erence for the field (Fig. 3c). In contrast, homozygous cry01 flies 0
10
did not show either a naive preference for the field (like cry02 flies)
or an enhanced preference for the field after training (Fig. 3c). The 0.05 w1118 cry02 cry02/+
loss of the response to the magnetic field in the Cry-deficient flies
c
resembled the behaviour when w1118 flies were deprived of ultra- 0.20 *
violet-A/blue light (Fig. 3a), which is consistent with Cry being the
relevant light sensor. These data using two cry null strains strongly 0.15
Preference index

suggest that both naive and trained responses to the magnetic field in
Drosophila require Cry function. 0.10 10
The Cry-defective cryb mutant flies are also unable to respond to
0.05 10
the magnetic field; the cryb mutation renders CryB essentially non-
functional22,23. Because the genetic background of cryb mutant flies is 0
10
not well defined, we compared behavioural responses to the magnetic 10
field between homozygous cryb flies and heterozygous cryb/Canton-S 0.05
flies. Whereas homozygous cryb flies did not show either naive or w1118 cry01
d 0.20
trained responses to the magnetic field under full-spectrum light,
heterozygous cryb/Canton-S flies showed significant naive
0.15
(P 5 0.0004, one-sample t-test) and trained responses (Fig. 3d); the
trained response in the heterozygotes was less than that of wild-type ****
0.10
Canton-S flies (Fig. 1b) and probably results from differences in
Preference index

genetic background. 0.05

10
To rule out non-cry mutations as the reason for the lack of mag- *
10 10 14
netic responses in cryb mutants, we showed that the cryb mutation 0 14 14
fails to complement the cry01 null mutation. Transheterozygous cryb/
cry01 flies did not show significant naive or trained responses to the 0.05 14 10
magnet, whereas heterozygous cry01/Canton-S and cryb/Canton-S
flies did (naive response, P 5 0.006, one-sample t-test; Fig. 3d). 0.10
Taken together, these data indicate that the cry locus is necessary
for light-dependent magnetosensitivity in Drosophila. Furthermore, 0.15 cryb cryb/C-S cryb/cry01 cry01/C-S
the lack of a trained response in both cry01 and cryb mutant flies is Figure 3 | Drosophila Cry mediates magnetosensitivity.
consistent with Cry being an essential component of the magneto- a, Magnetosensitivity in w1118 flies depends on ultraviolet-A/blue light. Bars
sensitive sensory input pathway and perhaps the magnetoreceptor show preference index values for naive responses under full-spectrum light
itself. and light .500 nm and .420 nm. Numbers represent the groups tested.
Because light-activated Cry interacts with the critical circadian ****, P , 0.0001. b, Naive response to a magnetic field is impaired in Cry-
clock protein Timeless to reset the circadian clock mechanism25, we deficient cry02 flies, but not in cry02/1 flies. Bars show preference index
values for naive responses. **, P , 0.01; ***, P , 0.001. c, Naive and trained
examined whether an intact circadian system is necessary for the Cry- responses to a magnetic field are impaired in Cry-deficient cry01 flies. Bars
dependent magnetosensitive responses in wild-type Canton-S flies. are preference index values for naive (white) and trained (black) groups. *,
Circadian arrhythmicity was induced by constant light, which dis- P , 0.05. d, Naive and trained responses to a magnetic field are impaired in
rupts circadian clock function in Cry-containing cells by causing the homozygous cryb and transheterozygous cryb/cry01 flies. Bars show
constant degradation of not only Cry but also Timeless and then preference index values for naive (white) or trained (black) groups. C-S,
Period25. We subsequently tested behavioural responses to the Canton-S. *, P , 0.05; ****, P , 0.0001. Values from ad are mean 6 s.e.m.
1016
511
2008 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 454 | 21 August 2008
a c
LD
1
Day
5 0.5
LL
1
Day
5
b 0
2.0 Per
Relative protein levels
1.5
1.0
0.5
0 LD
LL
1 5
Zeitgeber time/circadian time (h)
Figure 4 | Constant light disrupts circadian function but not Cry-mediated
magnetosensitivity in Canton-S flies. a, Mean activity records in 12 h
light:12 h dark lighting cycle (LD; top) or constant light (LL; bottom) in
double-plotted format (n 5 62 for each group). The lighting conditions were
identical to those used for housing flies tested for responses to magnet; light
irradiance, 1.5 3 1015 photons s21 cm22 nm21. For the light-dark cycle, 94%
expressed circadian rhythms when released in constant darkness (period,
24.6 6 0.03 h). All the constant-light flies were arrhythmic. b, Temporal
profiles of Period (Per) in heads. Protein abundance was rhythmic in the
and trained responses to the magnetic field (Fig. 4d). Thus, the con-
tinuous activation of Cry by light does not disrupt its ability to sense
the magnet, and an intact circadian system is not required for the
magnetoreception mechanism to operate.
There are two other published reports of magnetosensitivity in
adult Drosophila26,27. One describes behavioural evidence that male
wild-type Oregon-R flies show a light-dependent magnetic compass
response in a radial maze whereas female flies did not respond to the
magnet27. Additionally, male flies responded in opposite directions
when tested under either 365 nm or 500 nm light. In our studies, both
male and female flies showed a magnetic response. Regardless of
experimental differences, both the previous study27 and ours dem-
onstrate that fruitflies can respond to a magnetic field in a wave-
length-dependent manner.
Our results extend substantially the presence of a light-dependent
magnetic sense in Drosophila by showing the necessity of Cry. We
cannot distinguish unequivocally whether fly Cry functions as the
actual magnetoreceptor or whether it is an essential component
downstream of the receptor. Cry is necessary for both the naive
and trained responses to the magnetic field, consistent with the
notion that Cry is in the input pathway of magnetic sensing. In
addition, the continued behavioural responses to the magnet in con-
stant light, in which the known Cry signalling components are being
constantly degraded and the circadian clock is rendered non-func-
tional, is also consistent with an input function. The most compelling
evidence supporting a magnetoreceptor role for Cry is that the Cry-
dependent behavioural responses to the magnetic field require ultra-
violet-A/blue light, which matches the action spectrum of Drosophila
Cry19,20.
2008 Macmillan Publishers Limited. All rights reserved
LETTERS
Cry
2.0
Relative protein levels
1.5
1.0
0
LD LL LD LL cry01
C-S
d
0.2
****
0.1
Preference index
10
10
0.1
0.2
LD
LL
9 13 17 21
light-dark cycle (P , 0.01, one-way ANOVA), but not in constant light.
Head extracts were analysed by western blot30 and normalized against
a-tubulin. Values are mean 6 s.e.m. from three sets of heads. c, Cry
abundance is decreased in LL. Values are mean 6 s.e.m. from three sets of
heads collected over 24 h in the light-dark cycle or in constant light. Right,
western blot probed for Cry30 showing presence (arrow) in the light-dark
cycle or in constant light in Canton-S (C-S) heads, and absence in cry01
heads. d, Flies in constant light elicit behavioural responses to the magnetic
field. Values are mean 6 s.e.m. ****, P , 0.0001.
Our behavioural assay for magnetosensitivity does not at present
have a pure directional component, and therefore it is difficult to
relate our findings directly to the use of geomagnetic fields for animal
orientation and navigation. Nevertheless, it is probable that the res-
ponse we have identified is the prototype for the involvement of Cry
in chemical-based magnetic sensing. Thus, our findings open new
avenues of investigation into the cellular and molecular basis of
chemical-based magnetic sensing in animals. The powerful genetics
of Drosophila will facilitate an understanding of the precise mech-
anism of action of Cry in magnetosensitivity, such as the actual
involvement of magnetosensitive radical pairs produced by photo-
induced electron transfer reactions28. Our data further show that the
biological functions of Drosophila Cry extend beyond those in cir-
cadian clocks.
METHODS SUMMARY
Fly stocks were raised on standard cornmeal/agar medium at 25 uC and 60%
relative humidity under a 12 h light:12 h dark lighting cycle. The w1118;;cry0 flies
(from cry01 to cry03) were a gift from J. C. Hall and are described in ref. 21. The
w1118 stock used in our experiments was the same stock used to create w1118;;cry0
flies21. The cryb line was a gift from P. Emery22. Our choice apparatus was based
on the olfactory conditioning apparatus as described 29. Our two-coil system was
based on the double-wrapped coil system described previously24. We adjusted
the current flowing through the coils so that the magnetic field intensity was no
more than 5 G in any area along the tube. Coils were positioned 45u to the
horizontal for experiments involving Canton-S, w;Canton-S, Berlin-K and
Oregon-R flies. Coils were positioned parallel to the horizontal for all other
experiments, because it produced a more robust response and eliminated a polar
gradient; that is, there was no horizontal magnetic gradient, as the field was
perpendicular to the T-port tubes. To assess the magnetoresponse of flies, we
used a simple choice paradigm. Flies were placed in a glass vial containing
1017
512
LETTERS
moistened Whatman paper and starved for 22 h before training. All experiments
were performed between 8:00 and 12:00 EST. For each population of flies tested
(100150 flies per group), we calculated a preference index value on the basis of
the equation: (PM 2 0.5)/[(PM 1 0.5) 2 (2PM 3 0.5)], in which PM is the pro-
portion of flies on the magnetic field side of the T-port. To test whether flies
responded to the experimental magnetic field, we used either a Students t-test to
compare preference index values between trained and naive groups or a one-
sample t-test to compare preference index values to zero (that is, preference
index values expected with no response to the magnetic field).
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 25 February; accepted 19 June 2008.
Published online 20 July 2008.
1. Lohmann, K. J., Lohmann, C. M. F. & Putman, N. F. Magnetic maps in animals:
natures GPS. J. Exp. Biol. 210, 36973705 (2007).
2. Wiltschko, W. & Wiltschko, R. Magnetic orientation and magnetoreception in
birds and other animals. J. Comp. Physiol. A Neuroethol. Sens. Neural Behav. Physiol.
191, 675693 (2005).
3. Ritz, T., Adem, S. & Schulten, K. A model for photoreceptor-based
magnetoreception in birds. Biophys. J. 78, 707718 (2000).
4. Luschi, P. et al. Marine turtles use geomagnetic cues during open-sea homing.
Curr. Biol. 17, 126133 (2007).
5. Johnsen, S. & Lohmann, K. J. The physics and neurobiology of magnetoreception.
Nature Rev. Neurosci. 6, 703712 (2005).
6. Kirschvink, J. L. & Gould, J. L. Biogenic magnetite as a basis for magnetic-field
detection in animals. Biosystems 13, 181201 (1981).
7. Walker, M. M. A model for encoding of magnetic field intensity by magnetite-
based magnetoreceptor cells. J. Theor. Biol. 250, 8591 (2008).
8. Kirschvink, J. L., Walker, M. M. & Diebel, C. E. Magnetite-based
magnetoreception. Curr. Opin. Neurobiol. 11, 462467 (2001).
9. Leask, M. J. M. Physicochemical mechanism for magnetic-field detection by
migratory birds and homing pigeons. Nature 267, 144145 (1977).
10. Schulten, K., Swenberg, C. E. & Weller, A. Biomagnetic sensory mechanism based
on magnetic-field modulated coherent electron-spin motion. Z. Phy. Chem.
(Frankfurt) 111, 15 (1978).
11. Cashmore, A. R. Cryptochromes: Enabling plants and animals to determine
circadian time. Cell 114, 537543 (2003).
12. Partch, C. L. & Sancar, A. Photochemistry and photobiology of cryptochrome
blue-light photopigments: The search for a photocycle. Photochem. Photobiol. 81,
12911304 (2005).
13. Zhu, H. et al. The two CRYs of the butterfly. Curr. Biol. 15, R953R954 (2005).
14. Yuan, Q., Metterville, D., Briscoe, A. D. & Reppert, S. M. Insect cryptochromes:
Gene duplication and loss define diverse ways to construct insect circadian
clocks. Mol. Biol. Evol. 24, 948955 (2007).
15. Ozturk, N., Song, S. H., Selby, C. P. & Sancar, A. Animal type 1 cryptochromes
analysis of the redox state of the flavin cofactor by site-directed mutagenesis. J.
Biol. Chem. 283, 32563263 (2008).
16. Phillips, J. B. & Borland, S. C. Behavioral evidence for use of a light-dependent
magnetoreception mechanism by a vertebrate. Nature 359, 142144 (1992).
1018
2008 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 454 | 21 August 2008
17. Ritz, T., Dommer, D. H. & Phillips, J. B. Shedding light on vertebrate
magnetoreception. Neuron 34, 503506 (2002).
18. Wiltschko, R. & Wiltschko, W. Magnetoreception. Bioessays 28, 157168 (2006).
19. VanVickle-Chavez, S. J. & Van Gelder, R. N. Action spectrum of Drosophila
cryptochrome. J. Biol. Chem. 282, 1056110566 (2007).
20. Helfrich-Forster, C. et al. The extraretinal eyelet of Drosophila: Development,
ultrastructure, and putative circadian function. J. Neurosci. 22, 92559266
(2002).
21. Dolezelova, E., Dolezel, D. & Hall, J. C. Rhythm defects caused by newly
engineered null mutations in Drosophilas cryptochrome gene. Genetics 177,
329345 (2007).
22. Emery, P. et al. CRY, a Drosophila clock and light-regulated cryptochrome, is a
major contributor to circadian rhythm resetting and photosensitivity. Cell 95,
669679 (1998).
23. Stanewsky, R. et al. The cryb mutation identifies cryptochrome as a circadian
photoreceptor in Drosophila. Cell 95, 681692 (1998).
24. Kirschvink, J. L. Uniform magnetic-fields and double-wrapped coil systems
improved techniques for the design of bioelectromagnetic experiments.
Bioelectromagnetics 13, 401411 (1992).
25. Stanewsky, R. Genetic analysis of the circadian system in Drosophila melanogaster
and mammals. J. Neurobiol. 54, 111147 (2003).
26. Wehner, R. & Labhart, T. Perception of geomagnetic field in fly Drosophila
melanogaster. Experientia 26, 967968 (1970).
27. Phillips, J. B. & Sayeed, O. Wavelength-dependent effects of light on magnetic
compass orientation in Drosophila melanogaster. J. Comp. Physiol. A Sens. Neural
Behav. Physiol. 172, 303308 (1993).
28. Maeda, K. et al. Chemical compass model of avian magnetoreception. Nature 453,
387390 (2008).
29. Tully, T. & Quinn, W. G. Classical-conditioning and retention in normal and
mutant Drosophila melanogaster. J. Comp. Physiol. A Sens. Neural Behav. Physiol. 157,
263277 (1985).
30. Rush, B. L., Murad, A., Emery, P. & Giebultowicz, J. M. Ectopic CRYPTOCHROME
renders TIM light sensitive in the Drosophila ovary. J. Biol. Rhythms 21, 272278
(2006).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank H. Zhu for the protein work in Fig. 4b, c; L. Foley for
assistance; J. C. Hall for the cry0 flies; P. Emery for the Per and Cry antibodies; and
P. Emery, P. Perrat, B. Leung, S. DasGupta, M. Krashes and H. Zhu for discussions.
This work was supported by grants from the NIH.
Author Contributions S.M.R. and R.J.G. conceived the idea of using Drosophila to
study magnetosensitivity; S.W. and R.J.G. conceived the idea of using appetitive
conditioning to study magnetoresponses; R.J.G. designed the experimental
apparatus; R.J.G., S.W., A.C. and S.M.R. designed the experiments and analysed the
data; R.J.G. performed the experiments with help from A.C.; R.J.G., S.M.R., S.W. and
A.C wrote the paper.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. Correspondence and requests for materials should be
addressed to S.M.R. (steven.reppert@umassmed.edu).
513
doi:10.1038/nature07183

METHODS
Fly strains. Oregon-R, Berlin-K and w1118 stocks were provided by the
Bloomington Drosophila Stock Center (product number 6326, 4269 and 8522,
respectively).
Behavioural apparatus. The main chamber consisted of a training tube, a centre
elevator section to transfer flies, and a two-tube choice point or T-port for testing
the relative response of flies to a magnetic field (Fig. 1a). The training and T-port
tubes were round-bottom polystyrene tubes and could be removed from the
main section of the apparatus. The magnetic stimulus was delivered to the
training and T-port tubes by placing the choice chamber apparatus inside an
opaque housing box that contained the magnetic coil system. The box was
constructed such that the main chamber could be placed between the two coils
in either an upright position (as shown in Fig. 1a, top panel) or a horizontal
position (as shown in Fig. 1a, bottom panel). The upright position of the cham-
ber apparatus was used for training so that the tube could be placed in the centre
of one coil, and the horizontal position of the apparatus was used to suspend the
two tubes of the T-maze in the same area of the coil during test sessions. In this
way, flies were subjected to the same light conditions and intensity of magnetic
field during training and test sessions.
Each of the two coils was wrapped with two wires; the wires were wrapped in
one direction on one coil and in the opposite direction on the other. Current flow
through the coils produced a magnetic field in one coil (parallel current flow) but
not in the other (opposite current flow). A double-pole, double-throw switch
reversed the current flow in one wire loop but not the other, which allowed us to
move the magnetic field easily from the right to the left side of the T-maze. A DC
power supply with current and voltage controls was used so that we could change
the intensity of the magnetic field produced by the coils. Each coil was mounted
on a plastic track so that it could be positioned directly under the tubes (field
perpendicular to the tubes), at the end of the tubes at a 45u angle (as shown in
Fig. 1a), or at the end of the tubes (field parallel to the tubes). We used a magnetic
field intensity of 5 G for our experiments, because it gave the most consistent
naive response; decreasing to 1 G increased variability to the point that responses
were no longer significant.
The housing box for the test or choice chamber was open on the top so that the
chamber, regardless of position, could be illuminated by one ZooMed Reptisun
10.0 UVB fluorescent tube (F20T12) and one Agrobrite full spectrum fluorescent
grow tube. Wavelengths entering the box were restricted by covering the top of
the box with a long-wavelength filter that transmitted wavelengths of
light .500 nm (Edmund Optics) or .420 nm or .400 nm (E400 and E420 from
Gentex). Irradiance measurements (from an Ocean Optics USB 2000 fibre optic
spectrometer) were taken from inside one arm of the T-maze portion of the
choice chamber apparatus; thus, lighting conditions represent those experienced
by flies while being either trained or assayed for sensitivity to a magnetic field.
Experimental procedure. For the training group, a population of 100150 flies
was loaded into the elevator section of the choice apparatus with an empty
training tube facing one of the coils (Fig. 1a, top panel). Flies were transferred
to the training tube for 2 min and then transferred back to the elevator and held
for a 1-min rest period. The empty training tube was next replaced with a tube
containing sucrose reinforcement and flies were allowed to feed for 2 min in the
presence of a magnetic field. Flies were then transferred back to the elevator and
held for 1 min while the coil system was turned off. During this time, the training
tube was also removed, and two empty tubes were added to form the two arms of
the T-port. The choice chamber was then positioned horizontally in the box
(Fig. 1a, bottom panel). The coil system was turned on, and flies were transferred
to the T-port, in which they were allowed to choose between the sides with or
without a magnetic field. After 2 min, the two arms of the T-port were blocked
and flies from each side were collected into separate empty vials and counted.
For the naive group, a second population of 100150 flies was immediately
loaded into the elevator section of the horizontally placed choice chamber and
the coil system was turned on. After 1 min, flies were transferred directly to the
T-port for 2 min.
Trained and naive groups were tested consecutively and with the magnetic
field on the same side. This was done to control for the possibility that the choice
behaviour of flies reflected a preference for one arm of the T-port and not a
response to the magnetic field. As an extra control for side preferences inde-
pendent of magnetic stimuli, we alternated the side of the T-port containing the
field after each consecutive set of trained and naive flies (that is, trained and naive
with magnet on the left side and then trained and naive with magnet on the
right).

514
2008 Macmillan Publishers Limited. All rights reserved
 Neuron

Review

Seeing Things in Motion: Models,

Circuits, and Mechanisms
Alexander Borst1,* and Thomas Euler2,*
1Department of Systems and Computational Neurobiology, Max-Planck-Institute of Neurobiology, Am Klopferspitz 18,

D-82152 Martinsried, Germany

2CIN-Centre for Integrative Neuroscience, Institute for Ophthalmic Research, University of Tubingen, Rontgenweg 11,

D-72076 Tubingen, Germany

*Correspondence: borst@neuro.mpg.de (A.B.), thomas.euler@cin.uni-tuebingen.de (T.E.)
DOI 10.1016/j.neuron.2011.08.031

Motion vision provides essential cues for navigation and course control as well as for mate, prey, or predator
detection. Consequently, neurons responding to visual motion in a direction-selective way are found in
almost all species that see. However, directional information is not explicitly encoded at the level of a single
photoreceptor. Rather, it has to be computed from the spatio-temporal excitation level of at least two photo-
receptors. How this computation is done and how this computation is implemented in terms of neural
circuitry and membrane biophysics have remained the focus of intense research over many decades.
Here, we review recent progress made in this area with an emphasis on insects and the vertebrate retina.

Introduction Models of Direction Selectivity

Motion vision serves many different tasks; when moving through
the environment, the images of the environment as projected
onto the photoreceptor layer are constantly in motion. Since
the particular distribution of motion vectors on the retina, called
optic flow, depends on the specific movement of the animal,
whether it is moving forward or making a turn, the optic flow
represents a rich source of information that is widely used for
navigation and visual course control. Motion cues also occur
when the observing animal is standing still but another animal
is moving. Obviously, detecting such a potential mate, prey, or
predator and knowing which direction it is moving can be of
utmost importance for the survival of the observer. Thus, it is
not surprising that neurons responding to visual motion cues in
a direction-selective (DS) way are found in different parts of the
nervous system across the animal kingdom. However, despite
the high behavioral relevance of motion vision, the direction of
motion is not encoded explicitly by the signals of individual
photoreceptors: When moving a bar from left to right and back
again, the output signal of a photoreceptor will be the same
both times, no matter in which direction the bar has been
moving. However, a few synapses downstream into the nervous
system, cells are found that respond differently to the two direc-
tions. In between, some computation is happening, turning the
direction unselective response of the photoreceptor into a DS
response of the interneuron. This problem has become a classic
example for neural computation that has attracted researchers
from different fields over many decades (see also review by Clif-
ford and Ibbotson, 2002). Focusing on the insect optic lobe and
the vertebrate retina, we will provide an overview of what has
been learnt about the circuits and biophysical mechanisms
underlying the extraction of motion information from image
sequences in different animal species. As will become evident,
much progress has been made recently so that a solution seems
to be within reach.
Before discussing the neurons that respond specifically to the
direction of a moving stimulus, we will first take a look at the
problem from a computational point of view and discuss models
that have been proposed to account for this computation.
Defining the Computations
In physics, the velocity of a moving object is defined as the
objects spatial displacement over time. For the visual detection
of displacement, physical motion has to go along with changes in
the spatial brightness distribution on the retina. What character-
izes visual motion? Consider a smooth edge in an image moving
from left to right, passing in front of a single photoreceptor (Fig-
ure 1A). If the edge is moving slowly, the output signal will ramp
up slowly, too. If the same edge is moving at a high velocity, the
photoreceptor output signal will climb up steeply. Obviously, the
faster the object moves, the steeper the output signal. Now
consider two edges of different steepness passing by the
same photoreceptor at the exact same velocity (Figure 1B): If
the steep edge is moving, the output signal will again rise
steeply, if the shallow edge is moving, the output signal will
rise slowly. Obviously, the steeper the gradient, the steeper the
output signal. Therefore, neither the speed nor the direction of
the moving object can be deciphered from this output signal
alone. However, both of the above dependences are captured
by the following formula, relating the temporal signal change
dR/dt to the product of the spatial brightness gradient dI/dx
and the velocity dx/dt (Limb and Murphy, 1975; Fennema and
Thompson, 1979):
dR dI dx
= 
dt dx dt
The velocity dx/dt can, thus, be recovered by dividing the
temporal change dR/dt by the spatial gradient dI/dx.

974 Neuron 71, September 22, 2011 2011 Elsevier Inc.

515
Neuron

Review

A B

x x
t t

C D E

HP HP

+
~dI/dt
.. ~dI/dx
AND NOT

Figure 1. Models of Motion Detection

(A) A flat gradient is moving at two different velocities (spatial profile, top left; corresponding xt plots bottom left). The slow velocity results in a shallow temporal
gradient (top right), the fast velocity in a steep temporal gradient (bottom right).
(B) Two different spatial gradients are moving at the same velocity (spatial profile, top left, corresponding xt plots bottom left). The shallow gradient results in
a shallow temporal gradient (top right); the steep gradient results in a steep temporal gradient (bottom right).
(C) The gradient detector calculates the direction and magnitude of image velocity by dividing the temporal brightness gradient dI/dt by the spatial gradient dI/dx.
The spatial gradient is approximated by the difference between the brightness values at two adjacent image points (high-pass filter [HP]).
(D) The Reichardt detector calculates the direction of image motion by multiplying (M) the brightness values at two adjacent image points after one of them has
passed a low-pass filter with a time constant t. This is done in two mirror-symmetrical subunits, the outputs of which are subtracted from one another ( ).
(E) The Barlow-Levick detector calculates the direction of image velocity by processing the brightness values at two adjacent image points through a logical
AND-NOT gate after one of them is delayed by ms.

Motion-Detector Models
Several models have been proposed in the past that calculate
the direction of motion from the brightness changes as captured
by the photoreceptors.
The gradient model (Figure 1C) describes the most straightfor-
ward way to implement a motion detector with the above
mentioned mathematical relationship. Here, the spatial gradient
dI/dx is approximated by the brightness difference dI, of the
pattern, I, sampled at two neighboring image points separated
by a distance, dx. Both input signals become high-pass filtered,
approximating the temporal derivative, and then added together.
These two quantities are then divided by each other yielding an
estimate of the local image velocity (Srinivasan, 1990). This esti-
mate will only depend on the image velocity and not on the spatial
structure of the moving pattern because the local image contrast
is expressed in a steeper spatial, as well as in a steeper temporal
gradient: Dividing them leads to a cancellation of image contrast.
However, as attractive as the gradient model of motion detec-
tion might appear, most models that were proposed to account
for biological motion detectors actually do not calculate the
spatial and the temporal gradient of the moving image. They
rather correlate the brightness values measured at two adjacent
image points with each other after one of them has been filtered
in time (correlation model, Figure 1D). Consequently, their output
is not proportional to image motion but rather deviates from it in
a characteristic way. In fact, this deviation has been the crucial
hint for researchers in motion vision to propose exactly this
type of model. The first correlation detector was proposed on
the basis of experimental studies on the optomotor behavior of
insects (Hassenstein and Reichardt, 1956; Reichardt, 1961;
1987). This correlation detector is commonly referred to as the
Reichardt detector (van Santen and Sperling, 1985), and has
also been applied to explain motion detection in different verte-
brate species including man (for review, see Borst and Egelhaaf,
1989). Such a detector consists of two mirror-symmetrical
subunits. In each subunit, the signals derived from two neigh-
boring inputs are multiplied with each other after one of them
has been shifted in time by a temporal low-pass filter. The final

Neuron 71, September 22, 2011 2011 Elsevier Inc. 975

516

detector response is given by the difference of the output sig-
nals. Various elaborations of the basic Reichardt model have
been proposed to accommodate this motion detection scheme
to perform in a species-specific way.
Perhaps the simplest correlation-type movement detector has
been proposed by Barlow and Levick to explain their experi-
mental findings on DS ganglion cells in the rabbit retina (Barlow
and Levick, 1965). The Barlow-Levick model (Figure 1E) is
almost identical with respect to its layout but with only one
subunit of the basic Reichardt model. It consists of two input
lines carrying the brightness signals which are compared after
one of the signals has been delayed. In contrast to the Reichardt
model, this comparison is accomplished by a specific logical
operation, an AND-NOT or veto gate, suppressing the detectors
activity when the delay line is activated first and, consequently,
both signals arrive simultaneously at the AND-NOT gate. The
corresponding direction of motion, i.e., from left to right, is,
therefore, the detectors null direction. For motion in the detec-
tors preferred direction the veto signal arrives too late to have
an effect.
Another model which is often applied to human psycho-
physics and motion-sensitive neurons in the mammalian cortex
is the so-called motion energy model (Adelson and Bergen,
1985). Interestingly, if the Reichardt model is equipped with the
same spatial and temporal filters in its input channels, it assumes
the same specific functional characteristics as the energy model
and even is mathematically equivalent (van Santen and Sperling,
1985; Adelson and Bergen, 1985). This identity, however, only
holds for the final, fully opponent output signal of both detectors
and does not pertain to its internal structure.
Despite many differences in detail, all models of motion detec-
tion share the following commonalities: (1) they all have at least
two spatially separated input lines that read the brightness levels
of adjacent pixels in the image, (2) they all have some sort
of asymmetry with respect to the temporal filtering of the input
(a temporal derivative in case of the gradient detector, a low-
pass filter in one of the input channels of the Reichardt detector,
a delay line in the Barlow-Levick model), and (3) they all possess
an essential nonlinearity (division in the gradient detector, a multi-
plication in the Reichardt detector, and an AND-NOT gate in the
Barlow-Levick model).
They differ, however, in many other aspects that can be used
to discriminate between them experimentally. (1) As a character-
istic hallmark, the gradient detector delivers a signal that is
proportional to image velocity independent of the local image
contrast. (2) The output of the Reichardt detector grows quadrat-
ically with image contrast. Furthermore, it displays a maximum at
a certain image velocity. The optimum velocity is proportional to
the spatial pattern wavelength such that the maximum response
is always at the same temporal frequency (image velocity divided
by pattern wavelength). (3) The Barlow-Levick model is charac-
terized by a null-direction inhibition.
For an experimental analysis, it is also important to make the
distinction between the response properties of the individual
local motion detector, and those of a spatially integrated de-
tector array. When stimulated by a periodic grating moving at a
constant velocity, the local gradient detector will signal a con-
stant value as well. In contrast, the output signal of a local Reich-
976 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
ardt detector will consist of two parts: a constant DC shift that is
DS and, superimposed, a periodic modulation with the local
brightness of the pattern. Only when the summed output of an
array of Reichardt detectors is considered, these local modula-
tions will disappear since they are phase-shifted with respect
to each other. This also holds true for the Barlow-Levick model.
DS Cells
Neurons responding differently to visual stimuli moving in oppo-
site directions are called DS. Such neurons have long been
known to exist in the visual system and other parts of the verte-
brate and invertebrate nervous system.
Insects
In invertebrates, the first DS neurons were found in flies, located
in a brain structure called the lobula plate. The lobula plate is the
third of a stack of neuropiles of the flys optic lobe, each forming
a retinotopic representation of the image as initially formed by
the compound eye. Starting from the periphery, these are called
lamina, medulla, and lobula complex, the latter being divided into
an anterior lobula and a posterior lobula plate (Figure 2A). As
a consequence of the retinotopic structure, each neuropile is
built from repetitive columns containing an identical set of
neurons first described anatomically by Ramon y Cajal on the
basis of Golgi staining (Cajal and Sanchez, 1915). For the fruit
fly Drosophila melanogaster, a large set of columnar neurons
has been cataloged (Fischbach and Dittrich, 1989). More re-
cently, this set has been complemented by assigning transmitter
systems to various columnar neurons (e.g., Morante and Des-
plan, 2008; Raghu and Borst, 2011; Raghu et al., 2011). Each
columnar neuron, whether located in the lamina, medulla, or lob-
ula complex, has distinct arborizations in particular layers of its
neuropile and some neurons connecting the lamina with the
medulla or the medulla to the lobula plate. Furthermore, all these
cells restrict their arborizations to a small part of their respective
neuropile, mostly respecting the columnar borders. This is
different for the lobula plate, where dendrites of the so-called
lobula plate tangential cells span large parts of the neuropile,
apparently collecting signals from local neurons within hundreds
of columns. These tangential cells have been thoroughly ana-
lyzed, first in the blow fly Calliphora (Hausen, 1982a; 1982b;
Hengstenberg, 1982; Hengstenberg et al., 1982; Borst and
Haag, 1996; Haag et al., 1997; 1999) and, more recently, also
in the fruit fly Drosophila (Joesch et al., 2008; Schnell et al., 2010).
Although the exact number depends on the species, the
tangential cells comprise roughly 50 neurons, each of which
can be uniquely identified on the basis of its anatomy, receptive
field, and electrical response properties. All tangential cells
respond to visual motion in a DS way. Among them, the three
cells of the horizontal system, called HS cells, respond most
strongly to horizontal image motion: When the pattern moves
from the front to the back, the cells depolarize (Figure 2B). This
direction of image motion is their preferred direction. When the
pattern moves from the back to the front, they hyperpolarize.
This direction of image motion is their null direction. When
stimulated by a moving bar instead of a grating, their preferred
and null direction remains the same, no matter whether a white
bar is moving a black background or a black bar on a white
517
Neuron

Review

Figure 2. Motion-Sensitive Neurons in the Fly Visual System

(A) Schematic of the Drosophila optic lobe showing the three horizontal cells (horizontal system northern in red, horizontal system equatorial in blue, horizontal
system southern in green) in the lobula plate.
(B) DS response of HS cells to a moving grating. When the grating is moving in the preferred direction (rightward), the cell depolarizes; when the grating is moving
in the null direction (leftward), the cell hyperpolarizes.
(C) DS response of HS cells to a moving bar. Again, as in B, the cell depolarizes during preferred direction motion of the bar and hyperpolarizes during null
direction motion. Note that the response is the same; no matter whether a black bar is moving on a white background (upper panel) or a white bar is moving on
a black background (lower panel).
(D) Responses of HS cells to a moving grating as a function of grating contrast.
(E) Response of HS cells to moving gratings with two different spatial wavelengths as a function of grating velocity. The larger wavelength results in a higher
velocity optimum.
(F) Same data as presented in (E), but plotted as a function of temporal frequency. The temporal frequency is defined as the grating velocity divided by the spatial
wavelength. Both gratings produce the maximum response at the same temporal frequency.
All data are from Schnell et al. (2010).

Neuron 71, September 22, 2011 2011 Elsevier Inc. 977

518

background (Figure 2C). Another prominent group, the cells of
the vertical system, called VS cells, in general respond most
strongly to vertical image motion; downward is their preferred
direction and upward is their null direction. However, precise
mapping of the cells local preferred directions revealed a
spatially nonuniform receptive field; the different preferred direc-
tions in different parts of the flys visual field resemble an optic
flow pattern as might be elicited by the fly during certain flight
maneuvers (Krapp and Hengstenberg, 1996; Krapp et al.,
1998). These large and elaborate receptive fields could be shown
to result from a combination of direct feed-forward input the
tangential cells receive from columnar motion-sensitive ele-
ments and lateral synaptic interactions between the various
tangential cells within the lobula plate (Borst and Weber, 2011;
for review, see Borst et al., 2010).
As for the nature of their retinotopic input elements, the lobula
plate tangential cells have been subjected to numerous tests
investigating whether they conform to the Reichardt model in
blow flies, hover flies, and fruit flies. In these experiments,
tangential cells were stimulated by periodic gratings moving at
a constant velocity (Haag et al., 2004; Joesch et al., 2008;
Schnell et al., 2010) or with a dynamic velocity profile (Egelhaaf
and Reichardt, 1987; Egelhaaf and Borst, 1989; Borst et al.,
2003; Reisenman et al., 2003; Borst et al., 2005; Spavieri et al.,
2010). Some studies investigated the local motion response by
restricting the field of view to a small window through which
the pattern was shown to the fly (Egelhaaf et al., 1989) or by using
intracellular calcium concentration changes as a readout for
local activity in the dendrite (Single and Borst, 1998; Haag
et al., 2004). Tangential cells were also stimulated by natural
images (Dror et al., 2001) or by apparent motion stimuli con-
sisting of spatially displaced sequences of discrete brightness
steps (Egelhaaf and Borst, 1992). All these studies concluded
that the Reichardt detector accurately describes the behavior
of these input elements. As an example, the responses of
Drosophila HS cells have been measured as a function of pattern
contrast (Figure 2D): Although the response does not rise
quadratically as predicted by a perfect multiplication, it clearly
increases with increasing pattern contrast, thus ruling out a
division of temporal by spatial gradient as specified in the
gradient detector. When stimulated by a periodic grating drifting
at different velocities, the response of HS cells displays a velocity
optimum, as predicted by the Reichardt detector (Figure 2E,
black trace). Furthermore, when the test is repeated with a
grating of twice the spatial wavelength, the optimum velocity is
doubled (Figure 2E, gray trace). When the pattern velocity is
divided by the spatial wavelength of the pattern, both curves
coincide, revealing a peak at the same temporal frequency
of 1 Hz (Figure 2F), exactly as predicted by the Reichardt
detector.
Vertebrate Retina
The first reports of DS neurons in the vertebrate retina appeared
in the 1960s (for references see Wyatt and Daw, 1975). In partic-
ular, an elegant series of papers by Barlow, Levick, and co-
workers (e.g., Barlow and Hill, 1963; Barlow et al., 1964; Barlow
and Levick, 1965) on DS ganglion cells in the rabbit retina initi-
ated more than 40 years of research that established the retinal
978 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
DS circuitry as one of the most investigated and best understood
neuronal circuitries in the vertebrate brain.
Types of DS Ganglion Cells
The first type of retinal DS ganglion cells fires both at the leading
and the trailing edge of a stimulus moving along the preferred
direction through the receptive field (Barlow and Levick, 1965).
In other words, a bright spot on a dark background evoked
very similar DS responses as a dark spot on a bright background.
Due to this contrast independence, this cell type is referred to as
ON/OFF DS ganglion cell (for review, see Masland, 2004; Vaney
et al., 2001). They have a distinct morphology with loopy
dendrites (Figure 3A; Amthor et al., 1984; 1989) ramifying in
both the ON and the OFF sublamina of the inner plexiform layer
(IPL) (Figure 3D, red cell). The two arborizations can differ in size
and shape (Oyster et al., 1993; Vaney, 1994), suggesting that the
ON and the OFF DS circuits work independently. ON/OFF DS
ganglion cells are inhibited by synchronous motion outside their
receptive field center and are, thus, sensitive to motion contrast
(Chiao and Masland, 2003). As a result of their response proper-
ties, ON/OFF ganglion cells are considered to be local motion
detectors. They display a rather broad tuning in both the
temporal and spatial frequency domain (see e.g., Figure 2 in
Grzywacz and Amthor, 2007). Nevertheless, they seem to be
tuned to the temporal frequency of the stimulus rather than to
its velocity, speaking in favor of the Reichardt detector as an
appropriate description of the underlying mechanism. ON/OFF
DS cells can be clustered into four functional subtypes (Oyster
and Barlow, 1967), each of which preferring a different motion
direction roughly parallel to the dorsal-ventral (superior, inferior)
or nasal-temporal (anterior, posterior) axis (Figure 3D, bottom).
A second type of DS cell responds to only the leading edge of
a bright stimuli moving on a dark background and is, therefore,
referred to as an ON DS ganglion cell. They are monostratified
(Figure 3B), and their dendritic arborization ramifies in the inner
(ON) sublamina of the IPL (Figure 3D, blue cell) (Amthor et al.,
1989; Buhl and Peichl, 1986; He and Masland, 1998). In contrast
to ON/OFF DS cells, ON DS cells respond best to global motion
(Wyatt and Daw, 1975) and are tuned to lower temporal frequen-
cies (Grzywacz and Amthor, 2007). With respect to their
preferred direction, ON DS ganglion cells can be clustered into
three subtypes (Figure 3D, bottom). There is also recent
evidence for further functional subdivision into transient and sus-
tained types, each of which has distinct anatomical features
(Kanjhan and Sivyer, 2010).
Recently, a third type of DS cell was discovered in transgenic
mice expressing green fluorescent protein under the control of
the junctional adhesion molecule B (JAM-B) promoter exclu-
sively in a subset of ganglion cells (Kim et al., 2008). JAM-B posi-
tive cells have a peculiar morphology: Their asymmetrical
wedge-shaped dendritic arbors are aligned with the dorsal-
ventral axis of the retina and point ventrally (Figure 3C). They
respond best to centripetal motion, i.e., from the soma to the
dendritic tips, and thus, are directionally tuned to upward motion
(Figure 3C, bottom)taking into account that the lens inverts the
retinal image. With the exception of very large diameter spots,
they fire only at the offset of a light spot and have their dendrites
at the distal border of the IPL (Figure 3D, green cell). Neverthe-
less, they respond to preferred direction motion for both
519
Neuron

Review

A B C

Figure 3. DS Cells in the Mammalian Retina

(AC) ON/OFF DS ganglion cell (A) and ON DS ganglion cell (B, from Kanjhan and Sivyer, 2010) in the rabbit retina; JAM-B positive OFF DS ganglion cell in mouse
retina (C, from Kim et al., 2008). The scale bars represent 50 mm; arrowheads in A and C indicate axons (which are out of focus in B).
(D) Schematic vertical section of the retina with the dendritic stratification of the three type of DS cells from (AC). Below the cells, the preferred directions of
motion with respect to the visual field for each (sub)type are indicated. The following abbreviations are used: SAC, starburst amacrine cell; OPL, outer plexiform
layer; INL, inner nuclear layer; IPL, inner plexiform layer; GCL, ganglion cell layer.
(E) Flow charts illustrating the current knowledge on the main signal pathway from the different types of retinal DS ganglion cells to their target areas/nuclei in the
brain. The following abbreviations are used: d, dorsal; v, ventral; MTN, medial terminal nucleus; OT, optic tract; LGN, lateral geniculate nucleus; V1, primary visual
cortex; SC, superior colliculus; AOS, accessory optic system; NOT, nucleus of the optic tract. For clarity, most interconnections between areas and nuclei are
omitted. Colors indicate different target regions in the brain (see also F), thickness of lines connecting the areas represent relative contributions.
(F) Schematic overview of the important projection targets of retinal DS cells (sagittal cross section of a mammalian brain; for abbreviations see [E]).

contrasts (Kim et al., 2008). Interestingly, ganglion cells with
asymmetrical dendrites but orientation-selective responses,
reminiscent of mouse JAM-B cells, have been reported in the
rabbit (Amthor et al., 1989). Thus, OFF DS ganglion cells might
also exist in other species.
Projections of Retinal DS Ganglion Cells
Starburst cells represent a type of amacrine cell (Famiglietti,
1983; Masland and Mills, 1979) that had been suggested to be
critical for direction selectivity. When selectively ablated through
a nifty genetic manipulation, ON and ON/OFF DS ganglion cell

Neuron 71, September 22, 2011 2011 Elsevier Inc. 979

520
 Neuron

Review

Figure 4. Mechanisms of Direction Selectivity in the Fly Visual System

(A) Preferred and null direction response of a Drosophila VS cell during de- and hyperpolarizing current injection (from Joesch et al., 2008).
(B) Simulation of a Reichardt detector (left) with the two subunits controlling excitatory and inhibitory conductances on a single-compartment model neuron
(from Borst et al., 2010).
(C) Horizontal view of the optic lobe, including columnar T4 and T5 cells that represent potential presynaptic neurons of the lobula plate tangential cells
(from Fischbach and Dittrich, 1989).
(D) Schematic 3D view of the fly lobula plate indicating 2-deoxy-glucose uptake in four different layers depending on the direction of the moving grating
(from Buchner et al., 1984).
(E) Parallel L1- and L2-pathways as proposed by Fischbach et al. (1992) based on costratification of columnar neurons and motion-dependent activity labeling.
(F) EM-micrograph of a single lamina cartridge of Drosophila. The two central profiles represent lamina monopolar cells L1 and L2 that are surrounded by the
terminals of the six photoreceptors R1R6.

980 Neuron 71, September 22, 2011 2011 Elsevier Inc.

521
Neuron
Review
responses became indiscriminate to directional motion (Yoshida
et al., 2001). Moreover, this manipulation also resulted in a com-
plete loss of the optokinetic nystagmus (OKN) (Amthor et al.,
2002; Yoshida et al., 2001). This indicates that one or both of
these DS cell types provide signals essential for the control of
eye movement and gaze stabilization (reviewed in Berson,
2008; Vaney et al., 2001). It is likely that ON DS cells are the
main source of visual input for these tasks (Oyster et al., 1972),
because they prefer global motion, as caused by image slip-
page. Furthermore, their preferred directions correspond to
the three axes of the semicircular canals in the inner ear
(Figure 3D, bottom; see also Simpson et al., 1988b). Instead of
projecting to the superior colliculus (SC) and the lateral genicu-
late nucleus (LGN), like the majority of other ganglion cell types,
ON DS ganglion cells indeed project to the accessory optic
system (AOS), a collection of nuclei that controls eye movement
(Figures 3E and 3F, for review, see Berson, 2008). Using trans-
genic mice, researchers confirmed that the axonal projections
of ON DS cells with different preferred direction form discrete
clusters in the medial terminal nucleus, the primary nucleus of
the AOS (Yonehara et al., 2009), as proposed earlier (Simpson
et al., 1988a). The ON/OFF DS ganglion cells also provide
some input to the AOS and, therefore, contribute to the control
of eye movement, possibly for higher velocities. Consistent
with this is also the fact that their preferred directions are roughly
aligned with the four directions of apparent movement caused by
eye muscles contractions (Oyster and Barlow, 1967). ON/OFF
DS ganglion cells send collaterals to the SC and the LGN and,
therefore, may serve other visual functions as well, such as di-
recting attention to moving objects (reviewed in Berson, 2008).
No projections to the AOS were found for the JAM-B positive
OFF DS ganglion cells; they project to the SC and the dorsal
LGN (Kim et al., 2008), but the functional role of these inputs is
not yet understood. Altogether, with the exception of the contri-
bution to the optokinetic system, little is currently known about
the functional role of retinal direction selectivity for higher visual
processing.
Retinal Direction Selectivity across Vertebrate Species
Only recently, with the tremendous increase in transgenic mouse
diversity, research on DS mechanisms started to shift from
rabbits, on which most studies had focused, toward mice.
Despite a few minor differences, ON and ON/OFF DS ganglions
cells are functionally and morphologically very similar in mice
(Sun et al., 2006; Weng et al., 2005) and rabbits. There is
evidence for retinal direction selectivity in other mammals (for
review see Vaney et al., 2001), and therefore, it is conceivable
that this function is largely conserved among mammals. Interest-
ingly, in primates the existence of retinal direction selectivity has
not yet been convincingly shown. It is possible that this absence
reflects a sampling bias specific to primates: Compared to the
overwhelming number of, for example, midget ganglion cells,
which underlie high acuity vision, DS cells may be too infrequent.
Supporting the notion that these cells might have been missed in
(G) Blocking the synaptic output from either L1 or L2 results in selective loss of tangential cell responses to either moving ON-edges (L1-block, in green) or moving
OFF edges (L2-block, two different driver lines, in red). Asterisks indicate the significance level of the difference between the mean values: *p < 0.05, **p < 0.001.
NS is an abbreviation for not significant. (from Joesch et al., 2010).
(H) Splitting of brightness information into ON and OFF pathways, corresponding to L1 and L2, respectively.
physiological recordings, primate ganglion cells that are
morphologically equivalent to rabbit DS cells have been docu-
mented (Dacey, 2004; Yamada et al., 2005). Also starburst
amacrine cells, which are crucial to the DS circuitry, have been
found (Rodieck, 1989). Furthermore, retrograde tracing data on
the retinal projections to the AOS are consistent with the pres-
ence of ON DS ganglion cells in primates (Telkes et al., 2000).
Direction selectivity has also been studied in several nonmam-
malian vertebrates (Vaney et al., 2001; Wyatt and Daw, 1975).
For instance, DS ganglion cells in turtle (Marchiafava, 1979)
have functional properties very similar to those of mammals
(Borg-Graham, 2001). Birds also possess retinal DS cells (for
research on pigeons see Pearlman and Hughes, 1976), but little
is known about the underlying circuitry (e.g., Uchiyama et al.,
2000). Results from fish suggest how the refinement of retinal
DS circuitry might have progressed during evolution: In ancient
fish, such as dogfish, DS ganglion cells were found by recording
retinal input from pretectal neurons, but no clustering of
preferred directions is evident (Masseck and Hoffmann, 2008).
By contrast, recordings from the optic tectum of modern fish,
like carp, provide evidence for retinal ON and OFF DS cells,
each with three clusters of preferred directions (Damjanovic 
et al., 2009).
Network, Cellular, Subcellular, and Biophysical
Mechanisms
Having introduced the neurons found in various animal species
that respond to image motion in a DS way, we will now discuss
what cellular, subcellular, and biophysical mechanisms give
rise to this particular response property.
Insects
As outlined above, there is overwhelming evidence that the lob-
ula plate tangential cells of flies receive input from arrays of local
motion detectors of the Reichardt type. However, the small size
of the columnar elements in the optic lobe has made it difficult
to determine which of the many cells take part in the neural
circuitry implementing this algorithm. However, this situation
has changed recently, largely due to the application of electro-
physiological recording techniques to Drosophila (Wilson et al.,
2004; Joesch et al., 2008; Maimon et al., 2010), in combination
with the wide armory of genetic tools already available for this
organism (for review, see Borst, 2009).
First of all, it was demonstrated that Drosophila tangential cells
receive excitatory and inhibitory input from local motion sensitive
elements with opposite preferred direction (Joesch et al., 2008).
This was done by injecting depolarizing and hyperpolarizing
current into the tangential cell during motion stimulation in the
preferred and null direction (Figure 4A): Without current injection,
visual stimulation leads to depolarization of the cell during
preferred direction motion and hyperpolarization during null
direction motion (Figure 4A, middle trace). When depolarizing
current is injected, the preferred direction response becomes
Neuron 71, September 22, 2011 2011 Elsevier Inc. 981
522

smaller and the null direction response larger (top trace). The
opposite is observed during injection of hyperpolarizing current
(bottom trace). This can be reproduced by simulation of a single
electrical compartment model that receives two synaptic inputs
with reversal potentials above and below the resting potential of
the cell: The depolarizing current injection reduces the driving
force for the excitatory input while increasing it for the inhibitory
input, and hyperpolarizing current injection does the opposite
(Figure 4B). These results suggest that the subtraction stage in
the Reichardt detector is localized within the tangential cells
dendrites. Earlier experiments on blow fly tangential cells arrived
at similar conclusions (Borst and Egelhaaf, 1990; Borst et al.,
1995; Single et al., 1997). The chemical identity of the transmitter
systems involved in this push-pull input organization was clari-
fied by in vitro studies of blow fly lobula plate tangential cells.
These studies indicated that excitation is mediated by excitatory
nicotinic acetylcholine receptors (nAChRs) and inhibition by
g-aminobutyric acid (GABA) receptors (Egelhaaf et al., 1990;
Brotz and Borst, 1996; Brotz et al., 2001; Single et al., 1997).
Through the use of a Gal4-driver line that leads to expression
in lobula plate tangential cells of two types of labeled reporter
genes, excitatory and inhibitory transmitter receptors were found
to be colocalized on the fine dendritic branches of HS and VS
cells of Drosophila (Raghu et al., 2007; 2009). Thus, direction
selectivity in the tangential cells results from summation of two
inputs with opposite preferred directions.
But what neurons represent these excitatory and inhibitory
input elements to the lobula plate tangential cells? For a number
of reasons, bushy T cells are the prime candidates for providing
input to the lobula plate tangential cells. T4 cells exist in four
different subtypes per column, with dendrites ramifying in the
most proximal layer of the medulla. Each of the four T4-cell
subtypes projects into one out of four different strata of the lob-
ula plate (Figure 4C). In a similar way, four subtypes per column
are found for T5 cells as well, and they connect the posteriormost
layer of the lobula to one of the four strata of the lobula plate.
Following extended stimulation by moving gratings, Buchner
et al. (1984) found strong 2-deoxy-glucose labeling in one of
the four layers in the lobula plate depending on the particular
direction of the motion stimulus (Figure 4D). The direction of
motion which activates a specific stratum, as labeled using the
2-deoxy-glucose method, matches the preferred direction of
those tangential cells extending their dendrite in that stratum.
In addition to the lobula plate, 2-deoxy-glucose labeling was
highest in the most proximal layer of the medulla, where T4 cells
ramify, and in the posterior most layer of the lobula, where T5
cells extend their branches (Buchner et al., 1984). Finally, an
electron microscopy study in the blow fly has shown unequivo-
cally a chemical synapse between an HS-cell dendrite and a
columnar T4 cell (Strausfeld and Lee, 1991). Because of their
small size, however, the visual response properties of T4 and
T5 cells have proven very difficult to study. The few successful
recordings showed that T5 cells reveal a fully DS response,
whereas T4 cells are direction unselective (Douglass and Straus-
feld, 1995; 1996). As to the type of transmitter these cells use,
recent studies identified T4 cells as among the group of neurons
activating the ChAT-promoter, which controls the expression of
the enzyme choline-acetyl-transferase (ChAT) involved in the
982 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
synthesis of acetylcholine (ACh) (Raghu and Borst, 2011), while
T5 cells activate the promoter upstream of the gene encoding
the vesicular glutamate transporter (VGluT) (Raghu and Borst,
2011). However, a conclusive physiological proof that indeed
T4 cells are cholinergic and T5 cells are glutamatergic, and
whether they exert excitatory or inhibitory action on the lobula
plate tangential cells, is still missing. Furthermore, it is not known
which cells provide the above mentioned GABAergic input to the
tangential cells.
Based on the costratification of columnar cells in individual
layers of the medulla, lobula, and lobula plate as well as
2-deoxy-glucose labeling, T4 and T5 cells were proposed to
be the target cells of two separate pathways starting from the
photoreceptor terminals R1-6 in the lamina (Figure 4C; Fisch-
bach et al., 1992). Here, the photoreceptor terminals surround
the dendrites of two large lamina neurons, called L1 and L2,
which contact separate strata in the medulla (Figure 4E). There,
the signals are supposed to be picked up by specific intrinsic and
transmedullary neurons that terminate in the dendritic areas of
T4 and T5 cells, respectively. Keeping in mind the limited
evidence for the existence of these pathways to begin with,
one could only speculate how the signals in these two pathways
differ and how they might correspond to the Reichardt model.
This situation has changed due to a study where tangential cell
responses were recorded in Drosophila while the chemical trans-
mitter release from L1 or L2 cells (Figure 4F) was genetically
blocked in a cell-specific way (Joesch et al., 2010). While block-
ing the output from either L1 or L2 led to reduced but still signif-
icant responses to drifting gratings, blocking L1 completely and
selectively abolished the response to drifting ON-edges, and
blocking L2 erased the response to drifting OFF edges (Fig-
ure 4G). Using a behavioral readout instead of tangential cell
responses, another study obtained similar results (Clark et al.,
2011). These findings demonstrate that in fruit flies, the photore-
ceptor signal from R1-6 is split in the lamina into separate ON
and OFF pathways, represented by L1 and L2 cells, respectively.
This is analogous to the vertebrate retina where cone photore-
ceptors contact ON and OFF bipolar cells in parallel (reviewed
in Wassle, 2004). However, in the vertebrate retina the split is im-
plemented by different types of glutamate receptors in ON and
OFF bipolar cells (Nomura et al., 1994) so that light depolarizes
ON bipolar cells and hyperpolarizes OFF bipolar cells. In fruit
flies, however, the dendritic membrane response to light is iden-
tical in L1 and L2 and consists of a transient hyperpolarization at
the beginning and a rebound excitation at the end of a light pulse.
In L2 cells, the selectivity for light decrements seems to originate
in the axon terminal, as suggested by Ca2+ imaging (Reiff et al.,
2010): Whereas the intracellular calcium concentration is only
slightly reduced at the onset of light, a large and long lasting
calcium increase is elicited by light offset. Thus, L2-terminals
amplify predominantly the off-signal to postsynaptic neurons.
L1-terminals reveal calcium signals similar to the ones of L2-
terminals but with a stronger decrease of calcium concentration
at light onset (Clark et al., 2011). The easiest way to explain the
selectivity for light increments in the L1-pathway would be to
assume a signal inversion by inhibition with subsequent rectifica-
tion in the neurons postsynaptic to L1. This, however, remains
a pure speculation at present.
523
Neuron
Review
An interesting question following from the finding about the
splitting of the input into ON and OFF pathways concerns the
number of motion detector subtypes being at work in the fly brain:
Do four different detectors exist, one for each stimulus combina-
tion (ON-ON, OFF-OFF, ON-OFF, OFF-ON), or are there only two
detectors (ON-ON, OFF-OFF)? The most intuitive experiment to
investigate this question is the use of apparent motion stimuli in
which the brightness in two adjacent bars is stepped sequentially
from an intermediate level, that is also present in the surround, to
either a high (ON-Step) or to a low (OFF-Step) level. By applying
such stimuli to blow flies and fruit flies, various studies con-
sistently found positive responses to ON-ON and OFF-OFF
sequences and negative responses to ON-OFF and OFF-ON
sequences, either at the level of lobula plate tangential cells or
in a behavioral assay (Egelhaaf and Borst, 1992; Eichner et al.,
2011; Tuthill et al., 2011; Clark et al., 2011). While these findings
seem to clearly indicate the existence of four detector subtypes,
careful quantitative modeling, including the peripheral filter
stages, suggests that responses to mixed-brightness steps can
also be obtained from only two detectors (ON-ON and OFF-
OFF) if some residual information about the average brightness
level is preserved at the motion-detector input. Using a more
selective stimulus sequence consisting of brief brightness pulses
instead of steps led to responses to pulse sequences of the same
sign only, ruling out the existence of mixed-sign motion detectors
in blow flies and fruit flies (Eichner et al., 2011). This conclusion is
supported by an earlier study on house flies, Musca domestica,
that used sophisticated optics to sequentially stimulate individual
photoreceptors within one ommatidium projecting to neigh-
boring cartridges in the lamina (Franceschini et al., 1989). While
ON-ON and OFF-OFF sequences along the preferred direction
of the cell led to strong responses in the H1 tangential cell, no
responses were detected for mixed-sign sequences, i.e., ON-
OFF and OFF-ON.
In contrast to the two-detector model, Clark et al. (2011) advo-
cate for a model consisting of six detectors, with an asymmetric
distribution of the mixed detectors across the two pathways (L1:
ON-ON, OFF-OFF, OFF-ON; L2: ON-ON, OFF-OFF, ON-OFF),
which are nevertheless selective for ON and OFF edges, respec-
tively. However, to achieve this selectivity, the model requires
highly specific stimulus conditions as well as model parameters
that are hard to reconcile with previous work. The delay-filter
time constant of 10 s, necessary to reproduce edge selectivity
in Clarks model, is two to three orders of magnitude larger
than the value derived from all previous studies (e.g., Guo and
Reichardt, 1987; Egelhaaf and Reichardt, 1987; Dror et al.,
2001; Safran et al., 2007). This large time constant introduces
a long-lasting ringing in response to the onset of motion (A.B.,
unpublished data), incompatible with the observed cellular
responses (Egelhaaf and Borst, 1989; Reisenman et al., 2003).
A further conflict arises in this models prediction of negative
responses to ON-OFF and OFF-ON pulses, which are clearly
absent in the experimental data (Eichner et al., 2011). Neverthe-
less, while we feel that there is evidence arguing against the
six-detector model with mixed channels, definitive clarification
of the discrepancies will require further direct investigation.
Taken together with the evidence of the pathways leading
from L1 to T4 and from L2 to T5 cells, respectively, our current
view is that in the fruit fly, two separate motion detection systems
operate in parallel, one analyzing the movement of light incre-
ments and the other one the movement of light decrements
(Figure 4H). While the exact nature and role of the participating
neurons is still unclear, the splitting of the positive- and nega-
tive-going brightness signal into two channels, one for signals
of the positive, the other for signals of the negative sign, has
interesting consequences for the multiplication as postulated in
the Reichardt detector. Without splitting, the output of such
a putative multiplication neuron would need to increase in
a supralinear way when both input signals go positive as well
as when they go negative. Such a mechanism is difficult to
realize. However, splitting the input into separate channels leads
to positive signals only, and while a number of biophysically
plausible mechanisms have been proposed to do that (Torre
and Poggio, 1978; Srinivasan and Bernard, 1976; Gabbiani
et al., 2002; Hausselt et al., 2007; Enciso et al., 2010), the exact
mechanism active within these neurons presynaptic to the fly
LPTCs remains to be determined. In a similar way, the bio-
physics underlying the temporal filtering as postulated by the
Reichardt detector represents another challenge for future
research.
Vertebrate Retina
Since the majority of studies focused on ON/OFF DS cells and
their circuitry, we will concentrate on those, while only briefly
touching upon other types (see Mechanisms in Other Types of
Retinal DS Ganglion Cells). The original Barlow-Levick model
(Barlow et al., 1964; reviewed in Masland, 2004) proposed that
DS ganglion cells receive delayed and/or long lasting inhibition
preferentially from interneurons displaced to the null side of their
dendritic field. Note that the term null/preferred side refer to
the positions from which the null/preferred direction stimulus
enters the ganglion cells dendritic field. This inhibition would
be triggered by a stimulus moving in the null direction toward
the cells receptive field center and would cancel out any exci-
tation caused by the stimulus when it eventually enters the
center. Understandably, the original model does not fully
capture the multitiered organization of the retinal DS circuitry
as it is known today. The original model did, however, identify
the key properties of any DS circuit (see above). A hallmark of
retinal ON/OFF DS cells is their surprising robustness: The
retinal cells easily outperform their counterparts in primary visual
cortex (V1) in many respectsexcept maybe for directional
tuning width (45 in retinal ON/OFF cells versus R 15 in
V1, reviewed in Grzywacz and Amthor, 2007). The direction of
motion within the ON/OFF DS cells receptive field center is reli-
ably detected largely independent of contrast (Merwine et al.,
1998) and velocity (Grzywacz and Amthor, 2007; Oyster et al.,
1972; Wyatt and Daw, 1975), even for small movements of a
few micrometers (Grzywacz et al., 1994). Although their spiking
frequency peaks at velocities of 30 /s, direction discrimination
is constant over a velocity range of more than two orders of
magnitude (reviewed in Grzywacz and Amthor, 2007). In the light
of this robustness, it is very likely that the underlying circuitry
relies on multiple pathways and computational mechanisms
to generate and enhance DS signals, as we discuss in the
following.
Neuron 71, September 22, 2011 2011 Elsevier Inc. 983
524
 Neuron

Review

A
photoreceptor

bipolar cell

OFF sublamina
SACs
ON sublamina Glu
excitation
ACh
GABA inhibition
ON/OFF DS ganglion cell
(w/preferred direction to the right)
B1 B2 B3 motion
motion

+ -- + -
SAC

50% - +
-- +

F/F onto different subtypes

input output 500 ms of DS ganglion cells

C1 D1

C2 D2

E preferred null

+ + + + + + + -- + + -- +

Figure 5. Different Levels of DS Mechanisms in the Retinal ON/OFF DS Circuitry

(A) Schematic retinal cross-section showing the key elements of the circuitry of ON/OFF DS ganglion cells (blue, with motion to the right as its preferred direction)
in the mammalian retina, including bipolar cells (green) and starburst amacrine cells (SACs, red). While for simplicity, the illustration is focused on the OFF
sublamina of the inner plexiform layer (IPL), the circuit is assumed to be largely the same in both the ON and the OFF sublamina. Also not explicitly shown are:
(1) GABAergic interactions between neighboring SACs, and (2) other GABA- and glycinergic amacrine cells participating in the DS circuitry (see The Circuitry of
ON/OFF DS Ganglion Cells).
(B) Ca2+ signals in the distal dendrites of SACs are intrinsically DS. SAC with dendritic input and output zones indicated; circular wave stimulus overlaid (B1).
Centrifugal motion evokes larger Ca2+ signals in the distal dendrites than centripetal motion (B2). Since dendritic sectors of SACs are largely isolated from each
other, different motion directions activate different sectors, resulting in direction-dependent release of neurotransmitter (B3).
(C) Preferred direction responses in DS ganglion cells are facilitated by DS cholinergic input from SACs (adapted from Lee et al., 2010). Apparent motion stimulation
scheme with the position of the stimuli along the preferred null axis of an ON/OFF DS ganglion cell (C1). Responses to single flashes (C2, left two columns) and flash
sequences (right column) in preferred (top) and null direction (bottom) without and in the presence of acetylcholine (ACh) receptor blockers (HEX and CPP).

984 Neuron 71, September 22, 2011 2011 Elsevier Inc.

525
Neuron
Review
The Circuitry of ON/OFF DS Ganglion Cells
DS ON/OFF ganglion cells receive excitatory input from bipolar
cells but also from the previously mentioned starburst cells
(Figure 5A), which are also known as cholinergic amacrine cells
(Famiglietti, 1983; Masland and Mills, 1979). Besides ACh, star-
burst amacrine cells (SACs) also release GABA (Brecha et al.,
1988; Masland et al., 1984b; Vaney and Young, 1988) and
provide DS ganglion cells with inhibition as well (Figure 5A). In
addition, the DS ganglion cells receive both GABA and glyciner-
gic inhibition from other amacrine cell types (reviewed in Da-
cheux et al., 2003). The role of this additional inhibition in the
DS circuitry, however, is not yet well understood (see e.g.,
Neal and Cunningham, 1995).
Starburst amacrine cells (Figure 5B1) feature a characteristic
morphology (Famiglietti, 1983; Tauchi and Masland, 1984; Va-
ney, 1984) that is well conserved across vertebrate species:
Their dendritic arbor is composed of 46 sectors, each arising
from a primary dendrite that radiates from the soma before
dividing into smaller branches. SACs come in an ON and an
OFF variety, which appear to be functionally equivalent. They
costratify with the respective dendritic subtrees of ON/OFF DS
cells; ON SACs also costratify with the ON DS type (Figure 5A)
(Famiglietti, 1992). Each SAC dendrite is anatomically and phys-
iologically strongly polarized: Synaptic inputs from both bipolar
and amacrine cells cover the whole dendritic length, but outputs
are restricted to the distal part (Figure 5B1) (Famiglietti, 1983). In
addition, some channels and transporters are differentially
distributed along SAC dendrites, which, in combination with
the morphology, leads to electrical isolation of the sectors from
each other (Miller and Bloomfield, 1983; Velte and Miller,
1997). As a result, the individual dendritic sectors can be consid-
ered as largely independent processing units (Borg-Graham and
Grzywacz, 1992; Euler et al., 2002; Miller and Bloomfield, 1983;
Vaney, 1990). In contrast to most retinal neurons, SACs display
an extensive dendritic overlap (Tauchi and Masland, 1984),
which enables them to provide independent neuronal hardware
for the different functional subtypes of DS cells.
It has been proposed earlier that SACs are importantly
involved in the DS computation (Borg-Graham and Grzywacz,
1992; Masland et al., 1984a; Vaney et al., 1989), but experimental
proof for this notion came less than 10 years ago, when it was
shown that massive ablation of SACs results in a selective loss
of retinal direction selectivity (Amthor et al., 2002; Yoshida
et al., 2001; but see He and Masland, 1997). Optical measure-
ments of light-stimulus-evoked Ca2+ concentration changes
(Denk and Detwiler, 1999) in the dendrites of SACs demon-
strated that stimuli moving from the soma to the dendritic tips,
i.e., centrifugal motion, evoked larger Ca2+ responses in the
distal dendrites than motion in the opposite direction, i.e.,
centripetal motion (Figure 5B2) (Euler et al., 2002). Because
SAC output synapses are located in the distal dendrites (Fami-
(D) Asymmetric synaptic wiring between SACs and DS ganglion cells, determined from large-scale ultrastructural reconstruction of retina that was functionally
characterized with optical recordings (see Briggman et al., 2011, for details). Specific connectivity of synaptic SAC outputs (black skeleton), with varicosities
indicated by black dots (D1). Synapses are color-coded by the preferred direction of the postsynaptic DS ganglion cells (color-coded dashed ellipses indicate
their dendritic trees; the inset shows their preferred motion directions). Locations relative to the SAC somas of output synapses (n = 831) from all 24 reconstructed
SACs (D2). The scale bars represent 50 mm.
(E) Direction selectivity in ganglion cells (blue) result from multiple presynaptic mechanisms: (1) DS in SAC dendrites (see B), resulting in (2) null direction inhibition
via asymmetrical GABAergic input (see D) and (3) preferred direction excitation via symmetrical cholinergic input (see C and The Role of Excitation).
glietti, 1991) and transmitter release from SACs is Ca2+-depen-
dent (OMalley et al., 1992; Zheng et al., 2004), this indicated
that SACs are able to provide DS ganglion cells with directionally
tuned input. In fact, since the dendritic sectors are electrically
isolated from each other, each sector can be thought of as an
independent detector for centrifugal motion (Figure 5B3). Around
the same time, the related long-standing question as to whether
retinal direction selectivity is computed in the ganglion cells
themselves or presynaptically by interneurons (reviewed in Mas-
land, 2004) was successfully addressed. Patch-clamp studies
revealed that the synaptic input to ON/OFF DS cells is already
DS (Borg-Graham, 2001; Fried et al., 2002; Taylor and Vaney,
2002): Preferred direction motion elicits more excitation and
less inhibition in the ganglion cells, whereas null direction motion
elicits more inhibition and less excitation. This suggested (1) that
both inhibitory and excitatory inputs are DS, (2) that the Barlow-
Levick model does not fully capture retinal DS computations,
and (3) that the latter are indeed already performed presynapti-
cally by interneurons. Note that the latter point does not exclude
that postsynaptic, i.e., ganglion cell-intrinsic mechanisms con-
tribute to the overall direction selectivity observed in ganglion
cells (see Mechanisms at the Ganglion Cell Level).
The Role of Inhibition
It was already known for long that blocking GABAA receptors
abolishes DS responses in the ganglion cells but leaves their
responsiveness intact (Caldwell et al., 1978; Massey et al.,
1997). This indicated that GABAergic inhibition plays a crucial
role not only in directly cancelling of null direction motion
responses in the ganglion cells (Barlow and Levick, 1965), but
also in rendering the excitatory input to the ganglion cells DS
(see The Role of Excitation). Furthermore, in combination with
the SAC ablation experiments (Amthor et al., 2002; Yoshida
et al., 2001), it suggested that SACs are a major source for
this GABAergic inhibition. While a contribution from other
GABAergic amacrine cells cannot be excluded, it seems unlikely
that they fit the role of SACs because they typically are not
numerous enough to provide the different functional subtypes
of DS ganglion cells with adequate input (reviewed in Vaney
et al., 2001). This lack of cells may, however, be compensated
for by highly localized dendritic processing.
To distinguish between different mechanistic hypotheses of
direction selectivity, it is important to differentiate between direc-
tionally tuned (Figure 5B) and spatially offset inhibition (Figure 5D).
While both originate, at least in part, in SACs, they represent
different aspects of the DS computation. Spatially offset inhibition
as such can be sufficient to render ganglion cell responses DS
(Figures 5D and 5E; Borg-Graham and Grzywacz, 1992; Koch
et al., 1983; Barlow and Levick, 1965). Spatially offset inhibition
is a recurrent theme in the retinal DS circuit (Fried et al., 2005).
It acts not only at the level of the ganglion cell, but possibly also
at the bipolar cell terminals presynaptic to the DS ganglion cells
Neuron 71, September 22, 2011 2011 Elsevier Inc. 985
526

(Figure 5E) and possibly between SACs. The first evidence for
DS ganglion cells receiving spatially offset inhibition from SACs
as a consequence of asymmetrical wiring came from paired
recordings. These revealed that SACs located on the null side
of a ganglion cell provide significantly stronger inhibition than
SACs located on its preferred side (Fried et al., 2002; Lee
et al., 2010; Wei et al., 2011). The anatomical correlate for
such asymmetrical wiring has been elusive, since neither the
DS cells morphology (Amthor et al., 1984) nor the distribution
of synaptic inputs (Jeon et al., 2002) allowed researchers to
predict its preferred direction. While complex synaptic arrange-
ments between SACs and DS ganglion cells have been described
at the ultrastructural level (Dacheux et al., 2003), very recently,
a new EM technique for recording tissue volumes of sufficient
size and resolution (Briggman and Denk, 2006) allowed re-
searchers to statistically analyze the SAC-DS ganglion cell
connectivity for the first time (Figure 5D). In their elegant study,
Briggman et al. (2011) revealed two important circuit properties:
(1) The great majority of SAC synapses a DS ganglion cell receives
come from SAC dendrites that roughly point in the ganglion cells
null direction, meaning that the respective SAC somata are
located preferentially on the ganglion cells null side. (2) SACs
provide input to all DS ganglion cells within their dendritic field
but only via dendrites pointing in the null direction of the respec-
tive ganglion cell. Moreover, Briggman et al. (2011) found that the
probability for a synaptic contact is substantially higher when the
local angle of the SAC dendrite closely matches the ganglion
cells null direction, suggesting that the synapse forming mecha-
nism makes use the local dendritic geometry and/or activity.
The retinal DS circuitry does not solely rely on spatially offset
inhibition from SACs but ensures that this inhibition itself is DS
(Figure 5B). A number of models have been put forward to
explain the generation of dendritic direction selectivity in SACs
(reviewed in Euler and Hausselt, 2008). While these models differ
in the neuron types recruited or in the biophysical mechanisms
employed, they are not necessarily mutually exclusive. One class
of models focuses on intrinsic properties of SACs and makes use
of the fact that their dendrites are polarized computational
subunits (see The Circuitry of ON/OFF DS Ganglion Cells).
Passive models predict that SAC dendrites generate weak DS
signals simply by input summation along the dendrite (Tukker
et al., 2004). As a result, centrifugal motion (which extends
from the soma to the dendritic tips) evokes a larger signal in
the dendritic tips, whereas centripetal motion evokes a larger
signal in the soma (see also Borg-Graham and Grzywacz,
1992; Branco et al., 2010). As the SACs output synapses are
located in the dendritic tips, centrifugal motion would lead to
a larger output signal. In this scenario, the location of the output
synapses serves as spatial asymmetry and the threshold for
transmitter release as the essential nonlinearity. The direction
selectivity found in passive models seems, however, small and
too sensitive to stimulus parameters to explain the robust direc-
tion discrimination observed at the ganglion cell level. Also,
passive models predict a larger electrical signal at the soma
for centripetal motion which contradicts experimental observa-
tions (Euler et al., 2002; Peters and Masland, 1996).
To circumvent the deficits of passive models, voltage-gated
channels have been incorporated as nonlinearities. Voltage-
986 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
gated channels are suited to boost the SACs DS response if
differentially activated in a way that depends on the motion
direction (e.g., Borg-Graham and Grzywacz, 1992; Hausselt
et al., 2007; Tukker et al., 2004). SACs possess a variety of suit-
able voltage-gated Ca2+ (Cohen, 2001) as well as tetrodotoxin-
resistant Na+ channels (OBrien et al., 2008). Recently, the latter
shifted into the focus of interest since blocking these Na+ chan-
nels led to a reduction in dendritic direction selectivity in SACs
(Oesch and Taylor, 2010). Aside from the radially asymmetric
distribution of SAC output synapses, two kinds of gradients
along SAC dendrites have been proposed to serve as functional
asymmetry for DS detection: a voltage and a Cl concentration
gradient. Optical Ca2+ measurements in SACs suggest that the
distal dendrite is tonically depolarized relative to the soma
(Hausselt et al., 2007), possibly due to tonic glutamatergic input
(Taylor and Wassle, 1995; but see Oesch and Taylor, 2010).
Modeling suggests that even a small dendritic voltage gradient
in combination with voltage-gated channels could generate
a robust DS signal in SAC dendrites (Hausselt et al., 2007). In
another model it was proposed, that SACs generate a Cl
concentration gradient along their dendrites due to a differential
distribution of Cl intruders and extruders, and that this results in
GABAergic input causing depolarization at the proximal and
hyperpolarization at the distal dendrite, respectively (for details
see Enciso et al., 2010; Gavrikov et al., 2003; Gavrikov et al.,
2006). According to this model, the asymmetry in the effect of
GABAergic inputs leads to dendritic direction selectivity. Other
than the voltage gradient model, the Cl gradient model requires
GABAergic input and therefore does not account for the finding
that SAC responses remain DS in the presence of GABA
receptor blockers (see below).
Ultrastructural (Millar and Morgan, 1987) and functional data
(Zheng et al., 2004) indicate that mature SACs form reciprocal
GABAergic synapses, which have been implicated in the compu-
tation of DS signals (e.g., Munch and Werblin, 2006). If a SAC is
excited, it inhibits its neighborthis in turn reduces the neigh-
bors GABA release and in effect enhances the first SACs
response. Such interaction may sharpen the DS contrast in
neighboring SAC dendrites pointing in opposite directions (Lee
et al., 2010; Lee and Zhou, 2006). However, since GABA receptor
antagonists do not abolish dendritic direction selectivity in SACs
(Euler et al., 2002; Hausselt et al., 2007; Oesch and Taylor, 2010),
it is unlikely that these interactions are essential for the SACs
intrinsic DS mechanism.
The Role of Excitation
In addition to inhibition, DS ganglion cells receive DS excitatory
input from bipolar cells (Fried et al., 2005). This tuning could arise
from DS suppression of bipolar cell output by GABAergic ama-
crine cells (Figure 5E), which would explain why this excitatory
DS pathway is eliminated by GABA receptor blockers (see The
Role of Inhibition). Because ablating SACs abolishes ganglion
cell DS responses (Amthor et al., 2002; Yoshida et al., 2001), it
is likely that SACs are involved in tuning bipolar cell outputif
other amacrine cells were crucial, some residual direction selec-
tivity after ablation would be expected.
Besides glutamatergic excitation, DS ganglion cells also
receive excitatory cholinergic input from SACs (reviewed in Va-
ney et al., 2001). Blocking cholinergic receptors in the presence
527
Neuron

Review

Figure 6. Postsynaptic Amplification in Retinal ON/OFF DS Ganglion Cells

(A) Voltage response recorded from an ON/OFF DS ganglion cell soma to preferred (red) and null (black) direction motion. Peak depolarization is significantly
larger for null compared to preferred motion, yet spikes are generated much more reliably for preferred direction.
(B) Results from a biophysical model of an ON/OFF DS ganglion cell. For a passive model, that is without voltage-gated channels, and for weak synaptic input
(at the location marked by a red dot), dendrites are electrotonically well isolated (color-coded voltage distribution shown 50 ms after the onset of the stimulus).
(C) Directional tuning of a modeled DS ganglion cell (from C) for different conditions: only presynaptic DS mechanisms with weak (blue) or strong (red) input to the
ganglion cell included, only postsynaptic mechanisms included (green), both pre- and postsynaptic mechanisms together (yellow).
All panels from Schachter et al. (2010), for details see there.

of GABA receptor antagonists reduces the responses of DS
ganglion cells independent of motion direction (Chiao and
Masland, 2002), suggesting that cholinergic excitation provides
motion-sensitive but not DS excitation (He and Masland, 1997).
On the other hand, there is also evidence that this cholinergic
input is DS (Figure 5C, Fried et al., 2005; Lee et al., 2010).
In fact, paired recordings confirmed that SACs provide DS
ganglion cells with cholinergic input from all sides but also
showed that this input is modulated by GABA (Lee et al., 2010).
SACs on the preferred side of a DS cell release ACh unhindered
and thereby facilitate the motion response of the ganglion
cell, whereas the ACh release from SACs on the ganglion cells
null side is suppressed by inhibitory inputs (Figure 5E, Lee et al.,
2010). In other words, in the cholinergic pathway direction selec-
tivity results from DS modulation of otherwise symmetrical input
to the ganglion cells. This is in stark contrast to the GABAergic
pathway where the asymmetry is implemented as spatially-
biased synaptic connectivity. Three findings highlight that the
interactions between the cholinergic and the GABAergic path-
ways are still not fully understood: (1) The cholinergic pathway
appears to be more relevant for grating stimuli (Grzywacz et al.,
1998). (2) EM data suggest that SACs located on the preferred
side of a DS ganglion cell make only few synapses with this cell
(Briggman et al., 2011), which leaves one wondering how the
cholinergic signals are relayed. Paracrine ACh release is a possi-
bility, which is supported by the fact that while SACs are the sole
source of retinal ACh, even ganglion cells that do not costratify
with SACs possess ACh receptors. (3) In SACs, GABA and ACh
are differentially released in a Ca2+ level-dependent way, likely
from separate vesicle populations (Lee et al., 2010), adding
another level of complexity to the circuitry.
Recent modeling data (Poleg-Polsky and Diamond, 2011;
Schachter et al., 2010) suggest that the observed direction-
dependent difference in excitatory input can be alternatively ex-
plained by interactions between excitatory and inhibitory
conductances in the ganglion cells, without requiring DS excita-
tion. Such electrotonic interactions are to be expected if a cells
membrane potential cannot be spatially well controlled, which
is likely considering the highly branched morphology of DS
ganglion cells.
Mechanisms at the Ganglion Cell Level
By using dendritic Ca2+ imaging, it was shown that light stimuli
can locally initiate spikes in DS ganglion cell dendrites and that
these dendritic spikes are independent of the somatic spike
generator (Oesch et al., 2005). The role of these dendritic spikes
in ON/OFF DS cells was recently studied in a detailed biophys-
ical compartment model (Figure 6, Schachter et al., 2010). The
simulation results suggest that the dendritic arbor of DS ganglion
cells is partitioned into separate electrotonic regions (Figure 6B),
each of which sums locally inhibitory and excitatory inputs to
decide whether or not a dendritic spike is fired. The dendritic
spikes not only sharpen the directional tuning of the synaptic
input, but are also needed to relay the decision of the dendritic
regionindependently of the activity in other regionsto the
soma, where a somatic spike can then be triggered. The model
also suggests that synaptic inhibition can prevent local spike
initiation but is not strong enough to suppress spike propagation.
Schachter and colleagues (2010) provide a sensible explanation
for the so-called nondiscriminating zone, an area without DS
responses located at the preferred side of the ganglion cells
dendritic arbor (Barlow and Levick, 1965): They suggest that
the distal dendrites of DS ganglion cells are intrinsically DS
with their preferred directions oriented radially from the cells
center. Such intrinsic direction selectivity would interact with
the DS synaptic input, enhancing it at the null side while antago-
nizing it at the preferred side. Taken together, dendritic spike
initiation and electrotonic isolation of dendritic regions allow
the ganglion cells to respond reliably to different scales of motion
within their receptive field center and to sharpen the rather broad
directional tuning of the synaptic inputs (Figure 6C).
Mechanisms in Other Types of Retinal DS Ganglion Cells
Although much less is known about the DS mechanisms in the
other types of DS ganglion cells, it is likely that the circuitry of
the ON DS cells resembles that of ON/OFF DS cells in many
respects. For instance, conductance measurements revealed
that ON DS cells receive directionally tuned inhibitory and

Neuron 71, September 22, 2011 2011 Elsevier Inc. 987

528

excitatory input (Sivyer et al., 2010)similar to the ON/OFF cells.
There are some circuit differences to be expected because
unlike their ON/OFF counterparts, ON DS cells respond to global
motion. The recent finding that wide-field amacrine cells affect
ON DS cell responses via gap junctions (Ackert et al., 2009)
may be relevant in this context.
While JAM-B positive OFF DS cells (Kim et al., 2008) have not
yet been studied in detail, there are a number of findings indi-
cating that they employ rather different mechanisms than the
ON and ON/OFF types: (1) Because the dendrites of OFF DS
ganglion cells stratify in a different IPL level than the SACs
(Figures 4C and 4D, Kim et al., 2008), a substantial involvement
of SACs is unlikely. (2) The DS tuning strength of OFF DS cells is
positively correlated with the degree of dendritic asymmetry,
suggesting that the cells curious morphology plays a crucial
role for DS (Kim et al., 2008). This is in contrast to ON/OFF DS
ganglion cells, where the cells morphology does not predict its
preferred direction. (3) In the OFF DS cells, the responses to
dark moving spots are tuned to higher velocities than those to
bright moving spots, suggesting different ON and OFF DS mech-
anismsagain in contrast to ON/OFF DS ganglion cells.
Establishment of DS during Development
Another intriguing question is what mechanisms lead to the
selective asymmetrical wiring in neuronal circuits that compute
the direction of motion during development.
Insects
The stereotyped anatomy of the insect optic lobe, which has a
high degree of invariance of identified neurons between
different individuals, seems to indicate, right from the begin-
ning, that visual experience does not play a major role in
shaping the circuits in the visual system, including the ones
responsible for direction selectivity. However, in contrast to
these expectations, early visual experience was demonstrated
to have a rather strong effect on the size of the optic lobes in
Drosophila (Barth et al., 1997). Monocular deprivation (done
by painting over one eye) decreased the total volume of the
lamina, medulla, and lobula plate by up to 6%, and the lamina
showed a volume difference of up to 30%. The changes in
the lamina were largely attributable to changes in the terminals
of the photoreceptor cell axons. Another study in house flies
similarly concluded that flies reared in constant darkness have
about 20% fewer synapses than the control group (Rybak and
Meinertzhagen, 1997).
However, the question is still whether the motion detection
circuit function is impaired by the changes reported above.
This question has been addressed by investigating the anatomy
and physiology of the lobula plate tangential cells. Studying their
anatomy in blow flies, Cuntz and colleagues (2008) quantified the
dendrites of the same identified neurons from different individ-
uals. They measured the branching topology in terms of numbers
and distribution of branch points and branch angles, as well as
the area covered by the dendrite, their so-called spanning field,
within the lobula plate. The astonishing outcome was that
branching topology varies widely from individual to individual
while the dendritic spanning field is highly invariant and can be
used as a single parameter to distinguish between different lob-
988 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
ula plate cells. Thus, the dendritic spanning field seems develop-
mentally tightly controlled, while the branching topology is not.
Assuming that synapses are located mostly toward the endings
of the dendritic branch, this would suggest that branch topology
has no influence on the wiring of the particular cell as long as the
branches reach their presynaptic input at the proper location.
Another study directly investigated the effect of sensory experi-
ence on the anatomy of the lobula plate tangential cell VS1 in
Drosophila (Scott et al., 2003). The authors compared this cell
from individuals that were raised in a 12 hr/12 hr light-dark cycle
to the same cell in flies that were dark-reared. When the mor-
phology of the dendrite as well as the axon terminals was exam-
ined, no conspicuous differences could be observed between
the two groups. Thus, visual experience has no influence
on the anatomy of the tangential cells. A third study investigated
the influence of visual experience on the receptive field proper-
ties of tangential cells in blow flies (Karmeier et al., 2001). Again,
the authors compared neurons from two groups, one raised at
a 12 hr/12 hr light-dark cycle and the other reared in the dark.
The receptive field of the neurons from both groups turned out
to be indistinguishable. This clearly indicates that, despite the
quite dramatic changes seen in the peripheral parts of the visual
system, the circuits responsible for direction selectivity in the fly
do not require visual experience to properly develop.
Vertebrate Retina
How the retinal DS circuitry arises during development has not
yet been solved (Elstrott and Feller, 2009). This process very
likely does not require visual input: DS responses in ganglion
cells can be recorded already at eye opening (Masland, 1977)
and attempts at modifying retinal direction selectivity by various
manipulations, such as dark-rearing (Chan and Chiao, 2008;
Chen et al., 2009; Elstrott et al., 2008; Yonehara et al., 2009),
blocking of GABAergic inhibition (Wei et al., 2011), or raising
animals in environments with biased motion direction statistics
(Daw and Wyatt, 1974), were unsuccessful. SACs appear very
early during development (Feller et al., 1996) and form a network
that is dominated by excitatory cholinergic interactions and
involved in generating activity waves (reviewed in Masland,
2005), which have been proposed to participate in the DS
circuitry generation. Recent work, however, suggests that these
activity waves are not crucial for setting up the asymmetrical
wiring of the DS circuitry because ganglion cell DS responses
were not altered in knock-out mice lacking cholinergic waves
(Elstrott et al., 2008). Also recent recordings of DS ganglion cells
in the developing retina showed that while the cells responded to
passing waves, the direction of wave propagation did not modu-
late the cells responses (Elstrott and Feller, 2010).
The dendritic architecture of DS ganglion cells and SACs is
established before the retina becomes light responsive (Wong
and Collin, 1989; Wong, 1990; Stacy and Wong, 2003). Two
recent studies investigated the development of the synaptic
connectivity between SACs and DS ganglion cells by using
two different but equally elegant experimental approaches. Yo-
nehara et al. (2009) used optogenetics to examine the develop-
ment of the connectivity between SACs and ON DS ganglion
cells in transgenic mice, in which (1) upward motion preferring
ON DS ganglion cells were fluorescently labeled, and (2) SACs
529
Neuron
Review
expressed Cre-recombinase. Via viral transfection, the latter
allowed for targeted expression of channelrhodopsin (Ch2R),
a light-sensitive cation channel (Nagel et al., 2003). In this way,
SACs could be directly light-activated before the retina becomes
light sensitive, which in mice is around postnatal day (P) 10. By
activating SACs at different positions around an ON DS cell,
they probed the synaptic connections between the two cell
types. They revealed that within a 2 day time window (P6P8)
inhibitory connections change from symmetrical to asymmet-
rical, and there is an increase in inhibition from SACs on the
null side and a decrease in inhibition from SACs on the preferred
side. In the second study, Wei et al. (2011) performed paired
recordings from SACs and one subtype of ON/OFF DS ganglion
cells. For mice at P3P4, they found that the inhibitory SAC input
to the ganglion cells was spatially symmetrical. At P14 or later,
however, the inhibitory input the ganglion cell received from
SACs on the null side was strongly increased, whereas the
preferred side inhibition had remained constant. Despite minor
deviations, the results from the two studies are largely consis-
tent: Retinal DS circuitry maturation takes place rather quickly
before bipolar cells connect and the retina becomes light sensi-
tive. That the inhibitory events measured in ganglion cells do not
change in amplitude but become more frequent (Wei et al., 2011)
suggests a change in synapse number rather than strength,
pointing at selective synapse formation and/or elimination that
facilitates the development of asymmetric processing along
the dendrite. It is likely that molecular cues or marker gradients
along the different retinal axes are involved in this process (El-
strott and Feller, 2009); however, the identity of these markers
and the underlying guidance mechanisms are unknown.
Species Commonalities and Differences
Given the still unclear situation about the cellular and subcellular
mechanisms of direction selectivity, both in the insect optic lobe
and in the vertebrate retina, a comparison is rather challenging.
Further complications arise from the fact that it is not obvious
which DS cell types and circuitries are functionally equivalent
between insects and vertebrates. Nevertheless, as premature
the situation might be, we will draw some conclusions in the
following based on the data available at present.
ON-OFF Input Rectification
Independent of DS, the most conspicuous commonality
between the visual processing in insects and vertebrates relates
to the splitting of the photoreceptor input into ON and OFF chan-
nels (Werblin and Dowling, 1969; Franceschini et al., 1989;
Joesch et al., 2010). One obvious function of such a split is to
increase the dynamic range for coding light increments and
decrements. Also, it is an economic way of handling metabolic
costs: If mainly the changes of brightness levels are to be trans-
mitted to downstream circuits, then without such a split, a
second order neuron would have to maintain a high constant
level of activity in order to signal both light increments and decre-
ments. In addition, splitting the visual input signal is beneficial for
subsequent motion detection circuits: Given the requirement for
nonlinear processing from at least two displaced input, dealing
exclusively with positive signals alleviates the biophysical imple-
mentation significantly.
Optic Flow Processing Neurons
Neurons responding to optic flow stimuli are found in the insect
lobula plate (Krapp and Hengstenberg, 1996) as well as in area
medial superior temporal visual area (MST) of primates (Duffy
and Wurtz, 1997). In both cases, these neurons have extremely
large receptive fields and possess different preferred directions
in different locations of the receptive field. Furthermore, both
classes of neurons receive DS input from upstream cells that
have small receptive fields. Whether this input to MST cells
also has a push-pull organization with neurons of opposite
preferred directions making excitatory and inhibitory contacts
onto the dendrite of MST cells such as is realized in fly lobula
plate tangential cells is not known at present. Interestingly, the
receptive fields match an optic flow occurring during certain
types of ego-motion such as translation for various heading
directions in case of MST cells and rotation around various
body axes in case of fly lobula plate tangential cells (Krapp
et al., 1998). For lobula plate tangential cells, it is known that
the network connectivity between the various tangential cells is
responsible for the exact spatial lay-out of the receptive field
(Borst and Weber, 2011; for review see Borst et al., 2010). For
MST neurons, it is unclear whether the layout of their receptive
field is due to dendritic sampling of appropriately oriented local
motion-sensitive input elements or to a connectivity between
the MST neurons themselves.
Multistep and Multilevel Computation of DS
A push-pull type of input organization, however, has indeed been
found for ON/OFF DS ganglion cells of the retina. As outlined
above, the excitatory as well as inhibitory inputs to these gan-
glion cells are already, to some extent, DS as a result of complex
presynaptic interactions. Other circuit features such as the
spatially offset inhibition, together with particular dendritic pro-
cessing, seem to significantly enhance direction selectivity at
the level of the ganglion cell output, as compared to the input
signals driving them. In a similar way, direction selectivity is pro-
duced in the insect optic lobe as a multistep process (Borst and
Egelhaaf, 1990; Single et al., 1997; Joesch et al., 2008). In fly lob-
ula plate tangential cells, the spatial layout of the input does not
contribute to the direction selectivity: Each part of the receptive
field can be stimulated separately with moving gratings, and the
cell will respond the same way provided it has the same sensi-
tivity in both locations. Interestingly, DS ganglion cells behave
in a similar way: With the exception of the nondiscriminating
zone (see Mechanisms at the Ganglion Cell Level), motion
restricted to different subsections of the receptive field elicits
similar DS responses. As to direction selectivity of fly neurons
further upstream in the processing chain, mechanisms similar
to the ones in ganglion cells and starburst amacrine cells might
account for direction selectivity, for example in the dendrites of
T4 or T5 cells.
Development of DS Circuitries
In higher visual centers, such as the amphibian tectum (Engert
et al., 2002) or ferret area V1 (Li et al., 2008), visual experience
is not only necessary but also instructive for the development
of DS (reviewed in Elstrott and Feller, 2009). In contrast, at early
sensory stages like in the vertebrate retina or the insects lobula
Neuron 71, September 22, 2011 2011 Elsevier Inc. 989
530

plate, the development of DS circuitries appears to be indepen-
dent of visual experience and even activity independent to some
extent. This suggests that the development of DS circuitries is
largely genetically guided by processes that are still hardly
understoodan interesting and exciting commonality between
vertebrate retina and insect DS pathways and a common chal-
lenge for future comparative research.
CONCLUSION
The last decade has witnessed much progress in our under-
standing of the cellular and subcellular mechanisms underlying
direction selectivity. To a large extent, this is due to the applica-
tion of advanced optical as well as genetic methods to this
problem. Optical methods are indispensible whenever different
anatomical compartments of a neuron turn out to be electrically
separated, operating almost in isolation from the rest of the cell,
such as the different dendritic branches of a SAC in the verte-
brate retina (Euler et al., 2002) and the output terminals versus
the dendrite of lamina cells (Reiff et al., 2010) or the dendrite of
lobula plate tangential cells in the fly (Elyada et al., 2009). Look-
ing at the corresponding circuits at the ultrastructural level
reveals an intriguing complexity, both within the IPL of the retina
(Briggman et al., 2011) as well as in the columns of the insect
optic lobe (Takemura et al., 2008). The above examples demon-
strate that this complexity has to be taken into account when
modeling the corresponding circuits (e.g., Poleg-Polsky and Dia-
mond, 2011; Schachter et al., 2010; Hausselt et al., 2007).
Another amazing fact is how much effort over so many years
had to be invested in this one single problem of direction selec-
tivity in order to achieve the current level of understanding,
a problem that, in terms of computation and information pro-
cessing, seems quite modest (telling leftward from rightward),
compared to the complex intellectual capabilities of humans.
Our hope is that understanding this simple neural computation
of direction selectivity in full detail will provide an important step-
ping stone toward our understanding of more complex functions
of the nervous system.
ACKNOWLEDGMENTS
T.E. is supported by the DFG (EXC 307); A.B. is supported by the Max-Planck-
Society and by the DFG (SFB 870).
REFERENCES
Ackert, J.M., Farajian, R., Volgyi, B., and Bloomfield, S.A. (2009). GABA
blockade unmasks an OFF response in ON direction selective ganglion cells
in the mammalian retina. J. Physiol. 587, 44814495.
Adelson, E.H., and Bergen, J.R. (1985). Spatiotemporal energy models for the
perception of motion. J. Opt. Soc. Am. A 2, 284299.
Amthor, F.R., Oyster, C.W., and Takahashi, E.S. (1984). Morphology of on-off
direction-selective ganglion cells in the rabbit retina. Brain Res. 298, 187190.
Amthor, F.R., Takahashi, E.S., and Oyster, C.W. (1989). Morphologies of rabbit
retinal ganglion cells with complex receptive fields. J. Comp. Neurol. 280,
97121.
Amthor, F.R., Keyser, K.T., and Dmitrieva, N.A. (2002). Effects of the destruc-
tion of starburst-cholinergic amacrine cells by the toxin AF64A on rabbit retinal
directional selectivity. Vis. Neurosci. 19, 495509.
990 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
Barlow, H.B., and Hill, R.M. (1963). Selective sensitivity to direction of
movement in ganglion cells of the rabbit retina. Science 139, 412414.
Barlow, H.B., and Levick, W.R. (1965). The mechanism of directionally
selective units in rabbits retina. J. Physiol. 178, 477504.
Barlow, H.B., Hill, R.M., and Levick, W.R. (1964). Rabbit retinal ganglion cells
responding selectively to direction and speed of image motion in the rabbit.
J. Physiol. 173, 377407.
Barth, M., Hirsch, H.V.B., Meinertzhagen, I.A., and Heisenberg, M. (1997).
Experience-dependent developmental plasticity in the optic lobe of Drosophila
melanogaster. J. Neurosci. 17, 14931504.
Berson, D.M. (2008). Retinal Ganglion Cell Types and Their Central Projec-
tions. In Vision I, The Senses A Comprehensive Reference, R.H. Masland
and T.D. Albright, eds. (San Diego: Academic Press), pp. 491519.
Borg-Graham, L.J. (2001). The computation of directional selectivity in the
retina occurs presynaptic to the ganglion cell. Nat. Neurosci. 4, 176183.
Borg-Graham, L.J., and Grzywacz, N.M. (1992). A model of the directional
selectivity circuit in retina: Transformations by neurons singly and in concert.
In Single Neuron Computation, T. McKenna, S.F. Zornetzer, and J.L. Davis,
eds. (St Louis, MO: Academic Press), pp. 347375.
Borst, A. (2009). Drosophilas view on insect vision. Curr. Biol. 19, R36R47.
Borst, A., and Egelhaaf, M. (1989). Principles of visual motion detection.
Trends Neurosci. 12, 297306.
Borst, A., and Egelhaaf, M. (1990). Direction selectivity of blowfly motion-
sensitive neurons is computed in a two-stage process. Proc. Natl. Acad.
Sci. USA 87, 93639367.
Borst, A., and Haag, J. (1996). The intrinsic electrophysiological characteristics
of fly lobula plate tangential cells: I. Passive membrane properties. J. Comput.
Neurosci. 3, 313336.
Borst, A., and Weber, F. (2011). Neural action fields for optic flow based navi-
gation: A simulation study of the fly lobula plate network. PLoS ONE 6, e16303.
Borst, A., Egelhaaf, M., and Haag, J. (1995). Mechanisms of dendritic integra-
tion underlying gain control in fly motion-sensitive interneurons. J. Comput.
Neurosci. 2, 518.
Borst, A., Reisenman, C., and Haag, J. (2003). Adaptation of response tran-
sients in fly motion vision. II: Model studies. Vision Res. 43, 13091322.
Borst, A., Flanagin, V.L., and Sompolinsky, H. (2005). Adaptation without
parameter change: Dynamic gain control in motion detection. Proc. Natl.
Acad. Sci. USA 102, 61726176.
Borst, A., Haag, J., and Reiff, D.F. (2010). Fly motion vision. Annu. Rev. Neuro-
sci. 33, 4970.
Branco, T., Clark, B.A., and Hausser, M. (2010). Dendritic discrimination of
temporal input sequences in cortical neurons. Science 329, 16711675.
Brecha, N., Johnson, D., Peichl, L., and Wassle, H. (1988). Cholinergic
amacrine cells of the rabbit retina contain glutamate decarboxylase and
gamma-aminobutyrate immunoreactivity. Proc. Natl. Acad. Sci. USA 85,
61876191.
Briggman, K.L., and Denk, W. (2006). Towards neural circuit reconstruction
with volume electron microscopy techniques. Curr. Opin. Neurobiol. 16,
562570.
Briggman, K.L., Helmstaedter, M., and Denk, W. (2011). Wiring specificity in
the direction-selectivity circuit of the retina. Nature 471, 183188.
Brotz, T.M., and Borst, A. (1996). Cholinergic and GABAergic receptors on fly
tangential cells and their role in visual motion detection. J. Neurophysiol. 76,
17861799.
Brotz, T.M., Gundelfinger, E.D., and Borst, A. (2001). Cholinergic and
GABAergic pathways in fly motion vision. BMC Neurosci. 2, 1.
Buchner, E., Buchner, S., and Buelthoff, I. (1984). Deoxyglucose mapping of
nervous activity induced in Drosophila brain by visual movement. J. Comp.
Physiol. A. 155, 471483.
531
Neuron
Review
Buhl, E.H., and Peichl, L. (1986). Morphology of rabbit retinal ganglion cells
projecting to the medial terminal nucleus of the accessory optic system.
J. Comp. Neurol. 253, 163174.
Cajal, S.R., and Sanchez, D. (1915). Contribucion al Conocimiento de los
Centros Nerviosos de los Insectos (Madrid: Imprenta de Hijos de Nicholas
Moja).
Caldwell, J.H., Daw, N.W., and Wyatt, H.J. (1978). Effects of picrotoxin and
strychnine on rabbit retinal ganglion cells: Lateral interactions for cells with
more complex receptive fields. J. Physiol. 276, 277298.
Chan, Y.C., and Chiao, C.C. (2008). Effect of visual experience on the matura-
tion of ON-OFF direction selective ganglion cells in the rabbit retina. Vision
Res. 48, 24662475.
Chen, M., Weng, S., Deng, Q., Xu, Z., and He, S. (2009). Physiological proper-
ties of direction-selective ganglion cells in early postnatal and adult mouse
retina. J. Physiol. 587, 819828.
Chiao, C.C., and Masland, R.H. (2002). Starburst cells nondirectionally facili-
tate the responses of direction-selective retinal ganglion cells. J. Neurosci.
22, 1050910513.
Chiao, C.C., and Masland, R.H. (2003). Contextual tuning of direction-selective
retinal ganglion cells. Nat. Neurosci. 6, 12511252.
Clark, D.A., Bursztyn, L., Horowitz, M.A., Schnitzer, M.J., and Clandinin, T.R.
(2011). Defining the computational structure of the motion detector in
Drosophila. Neuron 70, 11651177.
Clifford, C.W., and Ibbotson, M.R. (2002). Fundamental mechanisms of visual
motion detection: Models, cells and functions. Prog. Neurobiol. 68, 409437.
Cohen, E.D. (2001). Voltage-gated calcium and sodium currents of starburst
amacrine cells in the rabbit retina. Vis. Neurosci. 18, 799809.
Cuntz, H., Forstner, F., Haag, J., and Borst, A. (2008). The morphological iden-
tity of insect dendrites. PLoS Comput. Biol. 4, e1000251.
Dacey, D.M. (2004). Origins of perception: Retinal ganglion cell diversity and
the creation of parallel visual pathways. In The Cognitive Neurosciences,
M.S. Gazzaniga, ed. (Cambridge, MA: MIT), pp. 281301.
Dacheux, R.F., Chimento, M.F., and Amthor, F.R. (2003). Synaptic input to the
on-off directionally selective ganglion cell in the rabbit retina. J. Comp. Neurol.
456, 267278.
Damjanovic , I., Maximova, E., and Maximov, V. (2009). Receptive field sizes of
direction-selective units in the fish tectum. J. Integr. Neurosci. 8, 7793.
Daw, N.W., and Wyatt, H.J. (1974). Raising rabbits in a moving visual environ-
ment: An attempt to modify directional sensitivity in the retina. J. Physiol. 240,
309330.
Denk, W., and Detwiler, P.B. (1999). Optical recording of light-evoked calcium
signals in the functionally intact retina. Proc. Natl. Acad. Sci. USA 96, 7035
7040.
Douglass, J.K., and Strausfeld, N.J. (1995). Visual motion detection circuits in
flies: Peripheral motion computation by identified small-field retinotopic
neurons. J. Neurosci. 15, 55965611.
Douglass, J.K., and Strausfeld, N.J. (1996). Visual motion-detection circuits in
flies: Parallel direction- and non-direction-sensitive pathways between the
medulla and lobula plate. J. Neurosci. 16, 45514562.
Dror, R.O., OCarroll, D.C., and Laughlin, S.B. (2001). Accuracy of velocity esti-
mation by Reichardt correlators. J. Opt. Soc. Am. A Opt. Image Sci. Vis. 18,
241252.
Duffy, C.J., and Wurtz, R.H. (1997). Planar directional contributions to optic
flow responses in MST neurons. J. Neurophysiol. 77, 782796.
Egelhaaf, M., and Borst, A. (1989). Transient and steady-state response
properties of movement detectors. J. Opt. Soc. Am. A 6, 116127.
Egelhaaf, M., and Borst, A. (1992). Are there separate ON and OFF channels in
fly motion vision? Vis. Neurosci. 8, 151164.
Egelhaaf, M., and Reichardt, W. (1987). Dynamic response properties of move-
ment detectors: Theoretical analysis and electrophysiological investigation in
the visual system of the fly. Biol. Cybern. 56, 6987.
Egelhaaf, M., Borst, A., and Reichardt, W. (1989). Computational structure of
a biological motion-detection system as revealed by local detector analysis
in the flys nervous system. J. Opt. Soc. Am. A 6, 10701087.
Egelhaaf, M., Borst, A., and Pilz, B. (1990). The role of GABA in detecting visual
motion. Brain Res. 509, 156160.
Eichner, H., Joesch, M., Schnell, B., Reiff, D.F., and Borst, A. (2011). Internal
structure of the fly elementary motion detector. Neuron 70, 11551164.
Elstrott, J., and Feller, M.B. (2009). Vision and the establishment of direction-
selectivity: A tale of two circuits. Curr. Opin. Neurobiol. 19, 293297.
Elstrott, J., and Feller, M.B. (2010). Direction-selective ganglion cells show
symmetric participation in retinal waves during development. J. Neurosci.
30, 1119711201.
Elstrott, J., Anishchenko, A., Greschner, M., Sher, A., Litke, A.M., Chichilnisky,
E.J., and Feller, M.B. (2008). Direction selectivity in the retina is established
independent of visual experience and cholinergic retinal waves. Neuron 58,
499506.
Elyada, Y.M., Haag, J., and Borst, A. (2009). Different receptive fields in axons
and dendrites underlie robust coding in motion-sensitive neurons. Nat. Neuro-
sci. 12, 327332.
Enciso, G.A., Rempe, M., Dmitriev, A.V., Gavrikov, K.E., Terman, D., and
Mangel, S.C. (2010). A model of direction selectivity in the starburst amacrine
cell network. J. Comput. Neurosci. 28, 567578.
Engert, F., Tao, H.W., Zhang, L.I., and Poo, M.M. (2002). Moving visual stimuli
rapidly induce direction sensitivity of developing tectal neurons. Nature 419,
470475.
Euler, T., and Hausselt, S.E. (2008). Direction Selective Cells. In Vision I, The
Senses A Comprehensive Reference, R.H. Masland and T.D. Albright, eds.
(San Diego: Academic Press), pp. 413422.
Euler, T., Detwiler, P.B., and Denk, W. (2002). Directionally selective calcium
signals in dendrites of starburst amacrine cells. Nature 418, 845852.
Famiglietti, E.V., Jr. (1983). Starburst amacrine cells and cholinergic neurons:
Mirror-symmetric on and off amacrine cells of rabbit retina. Brain Res. 261,
138144.
Famiglietti, E.V. (1991). Synaptic organization of starburst amacrine cells in
rabbit retina: Analysis of serial thin sections by electron microscopy and
graphic reconstruction. J. Comp. Neurol. 309, 4070.
Famiglietti, E.V. (1992). Dendritic co-stratification of ON and ON-OFF
directionally selective ganglion cells with starburst amacrine cells in rabbit
retina. J. Comp. Neurol. 324, 322335.
Feller, M.B., Wellis, D.P., Stellwagen, D., Werblin, F.S., and Shatz, C.J. (1996).
Requirement for cholinergic synaptic transmission in the propagation of spon-
taneous retinal waves. Science 272, 11821187.
Fennema, C.L., and Thompson, W.B. (1979). Velocity determination in scenes
containing several moving objects. Computer Graphics and Image Processing
9, 301315.
Fischbach, K.F., and Dittrich, A.P.M. (1989). The optic lobe of Drosophila
melanogaster. I. A Golgi analysis of wild-type structure. Cell Tissue Res.
258, 441475.
Fischbach, K.F., Ramos, R.G.P., and Bausenwein, B. (1992). The optic lobe of
Drosophila melanogaster: Experimental approaches to an analysis of struc-
ture, function and development. Verh Dtsch Zool Ges 85.2, 133148.
Franceschini, N., Riehle, A., and Le Nestour, A. (1989). Directionally selective
motion detection by insect neurons. In Facets of Vision, H. Stavenga, ed.
(Berlin, Heidelberg: Springer), pp. 360390.
Fried, S.I., Munch, T.A., and Werblin, F.S. (2002). Mechanisms and circuitry
underlying directional selectivity in the retina. Nature 420, 411414.
Neuron 71, September 22, 2011 2011 Elsevier Inc. 991
532

Fried, S.I., Munch, T.A., and Werblin, F.S. (2005). Directional selectivity is
formed at multiple levels by laterally offset inhibition in the rabbit retina. Neuron
46, 117127.
Gabbiani, F., Krapp, H.G., Koch, C., and Laurent, G. (2002). Multiplicative
computation in a visual neuron sensitive to looming. Nature 420, 320324.
Gavrikov, K.E., Dmitriev, A.V., Keyser, K.T., and Mangel, S.C. (2003). Cation
chloride cotransporters mediate neural computation in the retina. Proc. Natl.
Acad. Sci. USA 100, 1604716052.
Gavrikov, K.E., Nilson, J.E., Dmitriev, A.V., Zucker, C.L., and Mangel, S.C.
(2006). Dendritic compartmentalization of chloride cotransporters underlies
directional responses of starburst amacrine cells in retina. Proc. Natl. Acad.
Sci. USA 103, 1879318798.
Grzywacz, N.M., and Amthor, F.R. (2007). Robust directional computation in
on-off directionally selective ganglion cells of rabbit retina. Vis. Neurosci. 24,
647661.
Grzywacz, N.M., Amthor, F.R., and Merwine, D.K. (1994). Directional hyper-
acuity in ganglion cells of the rabbit retina. Vis. Neurosci. 11, 10191025.
Grzywacz, N.M., Amthor, F.R., and Merwine, D.K. (1998). Necessity of acetyl-
choline for retinal directionally selective responses to drifting gratings in rabbit.
J. Physiol. 512, 575581.
Guo, A., and Reichardt, W. (1987). An estimation of the time constant of move-
ment detectors. Naturwissenschaften 74, 9192.
Haag, J., Theunissen, F., and Borst, A. (1997). The intrinsic electrophysiolog-
ical characteristics of fly lobula plate tangential cells: II. Active membrane
properties. J. Comput. Neurosci. 4, 349369.
Haag, J., Vermeulen, A., and Borst, A. (1999). The intrinsic electrophysiological
characteristics of fly lobula plate tangential cells: III. Visual response proper-
ties. J. Comput. Neurosci. 7, 213234.
Haag, J., Denk, W., and Borst, A. (2004). Fly motion vision is based on Reich-
ardt detectors regardless of the signal-to-noise ratio. Proc. Natl. Acad. Sci.
USA 101, 1633316338.
Hassenstein, B., and Reichardt, W. (1956). Systemtheoretische Analyse der
Zeit-, Reihenfolgen- und Vorzeichenauswertung bei der Bewegungsperzep-
tion des Riisselkiifers Chlorophanus. Z Naturforsch IIb, 513524.
Hausen, K. (1982a). Motion sensitive interneurons in the optomotor system of
the fly. I. The horizontal cells: Structure and signals. Biol. Cybern. 45, 143156.
Hausen, K. (1982b). Motion sensitive interneurons in the optomotor system of
the fly. II. The horizontal cells: Receptive field organization and response char-
acteristics. Biol. Cybern. 46, 6779.
Hausselt, S.E., Euler, T., Detwiler, P.B., and Denk, W. (2007). A dendrite-auton-
omous mechanism for direction selectivity in retinal starburst amacrine cells.
PLoS Biol. 5, e185.
He, S., and Masland, R.H. (1997). Retinal direction selectivity after targeted
laser ablation of starburst amacrine cells. Nature 389, 378382.
He, S., and Masland, R.H. (1998). ON direction-selective ganglion cells in the
rabbit retina: Dendritic morphology and pattern of fasciculation. Vis. Neurosci.
15, 369375.
Hengstenberg, R. (1982). Common visual response properties of giant vertical
cells in the lobula plate of the blowfly Calliphora. J. Comp. Physiol. A Neuroe-
thol. Sens. Neural Behav. Physiol. 149, 179193.
Hengstenberg, R., Hausen, K., and Hengstenberg, B. (1982). The number and
structure of giant vertical cells (VS) in the lobula plate of the blowfly Calliphora
erytrocephala. J. Comp. Physiol. A Neuroethol. Sens. Neural Behav. Physiol.
149, 163177.
Jeon, C.J., Kong, J.H., Strettoi, E., Rockhill, R., Stasheff, S.F., and Masland,
R.H. (2002). Pattern of synaptic excitation and inhibition upon direction-selec-
tive retinal ganglion cells. J. Comp. Neurol. 449, 195205.
Joesch, M., Plett, J., Borst, A., and Reiff, D.F. (2008). Response properties
of motion-sensitive visual interneurons in the lobula plate of Drosophila
melanogaster. Curr. Biol. 18, 368374.
992 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
Joesch, M., Schnell, B., Raghu, S.V., Reiff, D.F., and Borst, A. (2010). ON and
OFF pathways in Drosophila motion vision. Nature 468, 300304.
Kanjhan, R., and Sivyer, B. (2010). Two types of ON direction-selective
ganglion cells in rabbit retina. Neurosci. Lett. 483, 105109.
Karmeier, K., Tabor, R., Egelhaaf, M., and Krapp, H.G. (2001). Early visual
experience and the receptive-field organization of optic flow processing
interneurons in the fly motion pathway. Vis. Neurosci. 18, 18.
Kim, I.J., Zhang, Y., Yamagata, M., Meister, M., and Sanes, J.R. (2008).
Molecular identification of a retinal cell type that responds to upward motion.
Nature 452, 478482.
Koch, C., Poggio, T., and Torre, V. (1983). Nonlinear interactions in a dendritic
tree: Localization, timing, and role in information processing. Proc. Natl. Acad.
Sci. USA 80, 27992802.
Krapp, H.G., and Hengstenberg, R. (1996). Estimation of self-motion by optic
flow processing in single visual interneurons. Nature 384, 463466.
Krapp, H.G., Hengstenberg, B., and Hengstenberg, R. (1998). Dendritic
structure and receptive-field organization of optic flow processing interneu-
rons in the fly. J. Neurophysiol. 79, 19021917.
Lee, S., and Zhou, Z.J. (2006). The synaptic mechanism of direction selectivity
in distal processes of starburst amacrine cells. Neuron 51, 787799.
Lee, S., Kim, K., and Zhou, Z.J. (2010). Role of ACh-GABA cotransmission in
detecting image motion and motion direction. Neuron 68, 11591172.
Li, Y., Van Hooser, S.D., Mazurek, M., White, L.E., and Fitzpatrick, D. (2008).
Experience with moving visual stimuli drives the early development of cortical
direction selectivity. Nature 456, 952956.
Limb, J.O., and Murphy, J.A. (1975). Estimating the velocity of moving images
in television signals. Computer Graphics and Image Processing 4, 311327.
Maimon, G., Straw, A.D., and Dickinson, M.H. (2010). Active flight increases
the gain of visual motion processing in Drosophila. Nat. Neurosci. 13, 393399.
Marchiafava, P.L. (1979). The responses of retinal ganglion cells to stationary
and moving visual stimuli. Vision Res. 19, 12031211.
Masland, R.H. (1977). Maturation of function in the developing rabbit retina.
J. Comp. Neurol. 175, 275286.
Masland, R.H. (2004). Direction Selectivity in Retinal Ganglion Cells. In The
Visual Neurosciences, L.M. Chalupa and J.S. Werner, eds. (Cambridge, MA:
The MIT Press), pp. 451462.
Masland, R.H. (2005). The many roles of starburst amacrine cells. Trends
Neurosci. 28, 395396.
Masland, R.H., and Mills, J.W. (1979). Autoradiographic identification of
acetylcholine in the rabbit retina. J. Cell Biol. 83, 159178.
Masland, R.H., Mills, J.W., and Cassidy, C. (1984a). The functions of acetyl-
choline in the rabbit retina. Proc. R. Soc. Lond. B Biol. Sci. 223, 121139.
Masland, R.H., Mills, J.W., and Hayden, S.A. (1984b). Acetylcholine-synthe-
sizing amacrine cells: Identification and selective staining by using radioauto-
graphy and fluorescent markers. Proc. R. Soc. Lond. B Biol. Sci. 223, 79100.
Masseck, O.A., and Hoffmann, K.P. (2008). Responses to moving visual stimuli
in pretectal neurons of the small-spotted dogfish (Scyliorhinus canicula).
J. Neurophysiol. 99, 200207.
Massey, S.C., Linn, D.M., Kittila, C.A., and Mirza, W. (1997). Contributions
of GABAA receptors and GABAC receptors to acetylcholine release and
directional selectivity in the rabbit retina. Vis. Neurosci. 14, 939948.
Merwine, D.K., Grzywacz, N.M., Tjepkes, D.S., and Amthor, F.R. (1998).
Non-monotonic contrast behavior in directionally selective ganglion cells
and evidence for its dependence on their GABAergic input. Vis. Neurosci.
15, 11291136.
Millar, T.J., and Morgan, I.G. (1987). Cholinergic amacrine cells in the rabbit
retina synapse onto other cholinergic amacrine cells. Neurosci. Lett. 74,
281285.
533
Neuron
Review
Miller, R.F., and Bloomfield, S.A. (1983). Electroanatomy of a unique amacrine
cell in the rabbit retina. Proc. Natl. Acad. Sci. USA 80, 30693073.
Morante, J., and Desplan, C. (2008). The color-vision circuit in the medulla of
Drosophila. Curr. Biol. 18, 553565.
Munch, T.A., and Werblin, F.S. (2006). Symmetric interactions within a homo-
geneous starburst cell network can lead to robust asymmetries in dendrites of
starburst amacrine cells. J. Neurophysiol. 96, 471477.
Nagel, G., Szellas, T., Huhn, W., Kateriya, S., Adeishvili, N., Berthold, P., Ollig,
D., Hegemann, P., and Bamberg, E. (2003). Channelrhodopsin-2, a directly
light-gated cation-selective membrane channel. Proc. Natl. Acad. Sci. USA
100, 1394013945.
Neal, M.J., and Cunningham, J.R. (1995). Baclofen enhancement of acetylcho-
line release from amacrine cells in the rabbit retina by reduction of glycinergic
inhibition. J. Physiol. 482, 363372.
Nomura, A., Shigemoto, R., Nakamura, Y., Okamoto, N., Mizuno, N., and
Nakanishi, S. (1994). Developmentally regulated postsynaptic localization of
a metabotropic glutamate receptor in rat rod bipolar cells. Cell 77, 361369.
OBrien, B.J., Caldwell, J.H., Ehring, G.R., Bumsted OBrien, K.M., Luo, S., and
Levinson, S.R. (2008). Tetrodotoxin-resistant voltage-gated sodium channels
Na(v)1.8 and Na(v)1.9 are expressed in the retina. J. Comp. Neurol. 508,
940951.
OMalley, D.M., Sandell, J.H., and Masland, R.H. (1992). Co-release of
acetylcholine and GABA by the starburst amacrine cells. J. Neurosci. 12,
13941408.
Oesch, N.W., and Taylor, W.R. (2010). Tetrodotoxin-resistant sodium channels
contribute to directional responses in starburst amacrine cells. PLoS ONE 5,
e12447.
Oesch, N., Euler, T., and Taylor, W.R. (2005). Direction-selective dendritic
action potentials in rabbit retina. Neuron 47, 739750.
Oyster, C.W., and Barlow, H.B. (1967). Direction-selective units in rabbit retina:
Distribution of preferred directions. Science 155, 841842.
Oyster, C.W., Takahashi, E., and Collewijn, H. (1972). Direction-selective
retinal ganglion cells and control of optokinetic nystagmus in the rabbit. Vision
Res. 12, 183193.
Oyster, C.W., Amthor, F.R., and Takahashi, E.S. (1993). Dendritic architecture
of ON-OFF direction-selective ganglion cells in the rabbit retina. Vision Res.
33, 579608.
Pearlman, A.L., and Hughes, C.P. (1976). Functional role of efferents to the
avian retina. I. Analysis of retinal ganglion cell receptive fields. J. Comp. Neu-
rol. 166, 111122.
Peters, B.N., and Masland, R.H. (1996). Responses to light of starburst
amacrine cells. J. Neurophysiol. 75, 469480.
Poleg-Polsky, A., and Diamond, J.S. (2011). Imperfect space clamp permits
electrotonic interactions between inhibitory and excitatory synaptic conduc-
tances, distorting voltage clamp recordings. PLoS ONE 6, e19463.
Raghu, S.V., and Borst, A. (2011). Candidate glutamatergic neurons in the
visual system of Drosophila. PLoS ONE 6, e19472.
Raghu, S.V., Joesch, M., Borst, A., and Reiff, D.F. (2007). Synaptic organiza-
tion of lobula plate tangential cells in Drosophila: GABA-receptors and chem-
ical release sites. J. Comp. Neurol. 502, 598610.
Raghu, S.V., Joesch, M., Sigrist, S.J., Borst, A., and Reiff, D.F. (2009). Synaptic
organization of lobula plate tangential cells in Drosophila: Dalpha7 cholinergic
receptors. J. Neurogenet. 23, 200209.
Raghu, S.V., Reiff, D.F., and Borst, A. (2011). Neurons with cholinergic pheno-
type in the visual system of Drosophila. J. Comp. Neurol. 519, 162176.
Reichardt, W. (1961). Autocorrelation, a principle for the evaluation of sensory
information by the central nervous system. In Sensory Communication, W.A.
Rosenblith, ed. (New York, London: MlT Press and Wiley), pp. 303317.
Reichardt, W. (1987). Evaluation of optical motion information by movement
detectors. J. Comp. Physiol. A Neuroethol. Sens. Neural Behav. Physiol.
161, 533547.
Reiff, D.F., Plett, J., Mank, M., Griesbeck, O., and Borst, A. (2010). Visualizing
retinotopic half-wave rectified input to the motion detection circuitry of
Drosophila. Nat. Neurosci. 13, 973978.
Reisenman, C., Haag, J., and Borst, A. (2003). Adaptation of response
transients in fly motion vision. I: Experiments. Vision Res. 43, 12911307.
Rodieck, R.W. (1989). Starburst amacrine cells of the primate retina. J. Comp.
Neurol. 285, 1837.
Rybak, J., and Meinertzhagen, I.A. (1997). The effects of light reversals on
photoreceptor synaptogenesis in the fly Musca domestica. Eur. J. Neurosci.
9, 319333.
Safran, M.N., Flanagin, V.L., Borst, A., and Sompolinsky, H. (2007). Adaptation
and information transmission in fly motion detection. J. Neurophysiol. 98,
33093320.
Schachter, M.J., Oesch, N., Smith, R.G., and Taylor, W.R. (2010). Dendritic
spikes amplify the synaptic signal to enhance detection of motion in a simula-
tion of the direction-selective ganglion cell. PLoS Comput. Biol. 6, e1000899.
Schnell, B., Joesch, M., Forstner, F., Raghu, S.V., Otsuna, H., Ito, K., Borst, A.,
and Reiff, D.F. (2010). Processing of horizontal optic flow in three visual inter-
neurons of the Drosophila brain. J. Neurophysiol. 103, 16461657.
Scott, E.K., Reuter, J.E., and Luo, L. (2003). Dendritic development of
Drosophila high order visual system neurons is independent of sensory
experience. BMC Neurosci. 4, 14.
Simpson, J.I., Leonard, C.S., and Soodak, R.E. (1988a). The accessory optic
system of rabbit. II. Spatial organization of direction selectivity. J. Neurophy-
siol. 60, 20552072.
Simpson, J.I., Leonard, C.S., and Soodak, R.E. (1988b). The accessory optic
system. Analyzer of self-motion. Ann. N Y Acad. Sci. 545, 170179.
Single, S., and Borst, A. (1998). Dendritic integration and its role in computing
image velocity. Science 281, 18481850.
Single, S., Haag, J., and Borst, A. (1997). Dendritic computation of direction
selectivity and gain control in visual interneurons. J. Neurosci. 17, 60236030.
Sivyer, B., van Wyk, M., Vaney, D.I., and Taylor, W.R. (2010). Synaptic inputs
and timing underlying the velocity tuning of direction-selective ganglion cells in
rabbit retina. J. Physiol. 588, 32433253.
Spavieri, D.L., Jr., Eichner, H., and Borst, A. (2010). Coding efficiency of fly
motion processing is set by firing rate, not firing precision. PLoS Comput.
Biol. 6, e1000860.
Srinivasan, M.V. (1990). Generalized gradient schemes for the measurement of
two-dimensional image motion. Biol. Cybern. 63, 421431.
Srinivasan, M.V., and Bernard, G.D. (1976). A proposed mechanism for
multiplication of neural signals. Biol. Cybern. 21, 227236.
Stacy, R.C., and Wong, R.O.L. (2003). Developmental relationship between
cholinergic amacrine cell processes and ganglion cell dendrites of the mouse
retina. J. Comp. Neurol. 456, 154166.
Strausfeld, N.J., and Lee, J.K. (1991). Neuronal basis for parallel visual pro-
cessing in the fly. Vis. Neurosci. 7, 1333.
Sun, W., Deng, Q., Levick, W.R., and He, S. (2006). ON direction-selective
ganglion cells in the mouse retina. J. Physiol. 576, 197202.
Takemura, S.Y., Lu, Z., and Meinertzhagen, I.A. (2008). Synaptic circuits of the
Drosophila optic lobe: The input terminals to the medulla. J. Comp. Neurol.
509, 493513.
Tauchi, M., and Masland, R.H. (1984). The shape and arrangement of the
cholinergic neurons in the rabbit retina. Proc. R. Soc. Lond. B Biol. Sci. 223,
101119.
Taylor, W.R., and Vaney, D.I. (2002). Diverse synaptic mechanisms generate
direction selectivity in the rabbit retina. J. Neurosci. 22, 77127720.
Taylor, W.R., and Wassle, H. (1995). Receptive field properties of starburst
cholinergic amacrine cells in the rabbit retina. Eur. J. Neurosci. 7, 23082321.
Neuron 71, September 22, 2011 2011 Elsevier Inc. 993
534

Telkes, I., Distler, C., and Hoffmann, K.P. (2000). Retinal ganglion cells projec-
ting to the nucleus of the optic tract and the dorsal terminal nucleus of the
accessory optic system in macaque monkeys. Eur. J. Neurosci. 12, 2367
2375.
Torre, V., and Poggio, T. (1978). A synaptic mechanism possibly underlying
directional selectivity to motion. Proc. R. Soc. Lond. B Biol. Sci. 202, 409416.
Tukker, J.J., Taylor, W.R., and Smith, R.G. (2004). Direction selectivity in
a model of the starburst amacrine cell. Vis. Neurosci. 21, 611625.
Tuthill, J.C., Chiappe, M.E., and Reiser, M.B. (2011). Neural correlates of
illusory motion perception in Drosophila. Proc. Natl. Acad. Sci. USA 108,
96859690.
Uchiyama, H., Kanaya, T., and Sonohata, S. (2000). Computation of motion
direction by quail retinal ganglion cells that have a nonconcentric receptive
field. Vis. Neurosci. 17, 263271.
van Santen, J.P.H., and Sperling, G. (1985). Elaborated Reichardt detectors.
J. Opt. Soc. Am. A 2, 300321.
Vaney, D.I. (1984). Coronate amacrine cells in the rabbit retina have the
starburst dendritic morphology. Proc. R. Soc. Lond. B Biol. Sci. 220,
501508.
Vaney, D.I. (1990). The mosaic of amacrine cells in the mammalian retina. Prog.
Retin. Eye Res. 9, 49100.
Vaney, D.I. (1994). Territorial organization of direction-selective ganglion cells
in rabbit retina. J. Neurosci. 14, 63016316.
Vaney, D.I., and Young, H.M. (1988). GABA-like immunoreactivity in cholin-
ergic amacrine cells of the rabbit retina. Brain Res. 438, 369373.
Vaney, D.I., Collins, S.P., and Young, H.M. (1989). Dendritic relationships
netween cholinergic amacrine cells and direction selective retinal ganglion
cells. In Neurobiology of the Inner Retina, R. Weiler and N.N. Osborn, eds. (Ber-
lin: Springer), pp. 157168.
Vaney, D.I., He, S., Taylor, W.R., and Levick, W.R. (2001). Direction-Selective
Ganglion Cells in the Retina. In Motion Vision - Computational, Neural, and
Ecological Constraints, J.M. Zanker and J. Zeil, eds. (Berlin, Heidelberg,
New York: Springer), pp. 1356.
994 Neuron 71, September 22, 2011 2011 Elsevier Inc.
Neuron
Review
Velte, T.J., and Miller, R.F. (1997). Spiking and nonspiking models of starburst
amacrine cells in the rabbit retina. Vis. Neurosci. 14, 10731088.
Wassle, H. (2004). Parallel processing in the mammalian retina. Nat. Rev. Neu-
rosci. 5, 747757.
Wei, W., Hamby, A.M., Zhou, K., and Feller, M.B. (2011). Development of
asymmetric inhibition underlying direction selectivity in the retina. Nature
469, 402406.
Weng, S., Sun, W., and He, S. (2005). Identification of ON-OFF direction-selec-
tive ganglion cells in the mouse retina. J. Physiol. 562, 915923.
Werblin, F.S., and Dowling, J.E. (1969). Organization of the retina of the mud-
puppy, Necturus maculosus. II. Intracellular recording. J. Neurophysiol. 32,
339355.
Wilson, R.I., Turner, G.C., and Laurent, G. (2004). Transformation of olfactory
representations in the Drosophila antennal lobe. Science 303, 366370.
Wong, R.O.L. (1990). Differential growth and remodelling of ganglion cell
dendrites in the postnatal rabbit retina. J. Comp. Neurol. 294, 109132.
Wong, R.O., and Collin, S.P. (1989). Dendritic maturation of displaced putative
cholinergic amacrine cells in the rabbit retina. J. Comp. Neurol. 287, 164178.
Wyatt, H.J., and Daw, N.W. (1975). Directionally sensitive ganglion cells in the
rabbit retina: Specificity for stimulus direction, size, and speed. J. Neurophy-
siol. 38, 613626.
Yamada, E.S., Bordt, A.S., and Marshak, D.W. (2005). Wide-field ganglion cells
in macaque retinas. Vis. Neurosci. 22, 383393.
Yonehara, K., Ishikane, H., Sakuta, H., Shintani, T., Nakamura-Yonehara, K.,
Kamiji, N.L., Usui, S., and Noda, M. (2009). Identification of retinal ganglion
cells and their projections involved in central transmission of information about
upward and downward image motion. PLoS ONE 4, e4320.
Yoshida, K., Watanabe, D., Ishikane, H., Tachibana, M., Pastan, I., and
Nakanishi, S. (2001). A key role of starburst amacrine cells in originating retinal
directional selectivity and optokinetic eye movement. Neuron 30, 771780.
Zheng, J.J., Lee, S., and Zhou, Z.J. (2004). A developmental switch in the
excitability and function of the starburst network in the mammalian retina.
Neuron 44, 851864.
535


536
 t e c h n ic a l r e p o r t s

Functional imaging of hippocampal place cells at

cellular resolution during virtual navigation
Daniel A Dombeck1, Christopher D Harvey1, Lin Tian2, Loren L Looger2 & David W Tank1
Spatial navigation is often used as a behavioral task in studies revealed no correlation between place field location and a cells ana-
of the neuronal circuits that underlie cognition, learning and tomical location8. By contrast, IEG studies have provided evidence for
memory in rodents. The combination of in vivo microscopy with micro-clustering of cells that were active in restricted places within
genetically encoded indicators has provided an important new tool a larger environment5.
for studying neuronal circuits, but has been technically difficult Optical imaging of neuronal activity at a high spatial resolution can
2010 Nature America, Inc. All rights reserved.

to apply during navigation. Here we describe methods for imaging
the activity of neurons in the CA1 region of the hippocampus with
subcellular resolution in behaving mice. Neurons that expressed
the genetically encoded calcium indicator GCaMP3 were imaged
through a chronic hippocampal window. Head-restrained mice
performed spatial behaviors in a setup combining a virtual
reality system and a custom-built two-photon microscope.
We optically identified populations of place cells and determined
the correlation between the location of their place fields in the
virtual environment and their anatomical location in the local
circuit. The combination of virtual reality and high-resolution
functional imaging should allow a new generation of studies to
investigate neuronal circuit dynamics during behavior.
The location-specific firing of hippocampal place cells1 during
navigation represents a salient neural correlate of spatial informa-
tion in the mammalian brain. More generally, place cells and other
hippocampal neurons are thought to be involved in the encoding of
episodic memories2, making the behaving animal the most appro-
priate setting in which to study their activity patterns. The firing
properties of hippocampal neurons in behaving rodents have been
characterized using extracellular electrode techniques in numerous
experimental paradigms3,4. However, electrophysiological methods
have limitations in characterizing precise spatio-temporal activity
patterns, targeting genetically labeled subpopulations and monitor-
ing subcellular compartments.
For example, it is unclear whether hippocampal neurons with
similar place fields are spatially organized within the hippocam-
pus. Although some studies have reported organization on spatial
scales ranging from tens of micrometers5 to millimeters6,7, a separate
study found no organization on either of these scales8. This question
has been difficult to address because extracellular electrodes can-
not report the precise anatomical location of the recorded cells69.
Indirect methods, such as immediate early gene (IEG) expression5,8,
are also problematic because the relationship between gene expression
and neuronal spiking is not well established. Tetrode recordings have
address these and other questions by providing a full spatio-temporal
view of neuronal population activity within a microcircuit, by record-
ing from genetically labeled neuronal subtypes, and by monitoring
subcellular compartments such as dendritic branches and spines.
However, it is not currently possible to perform high-resolution func-
tional imaging of neurons in the hippocampi of awake, navigating mice.
To achieve this goal, three obstacles must be overcome: cellular reso-
lution imaging in the brain of a mobile mouse; imaging more than a
millimeter beneath the cortical surface; and imaging that is compatible
with navigation behavior. Two-photon microscopy10 in combination
with calcium-sensitive dyes has been used to provide cellular resolution
functional imaging in awake mobile mice on a spherical treadmill11,12.
Chronic functional imaging in this preparation has been made possible
using the genetically encoded calcium indicator GCaMP313. In an acute
anesthetized mouse preparation, non-activity dependent fluorescent
probes have been used for subcellular imaging of the hippocampus14. In
addition, a virtual reality system has been developed that allows spatial
behaviors to be studied in head-restrained mice15.
Here, we combined and extended these methods to make possible
high-resolution functional imaging of the hippocampus in mice navi-
gating in a virtual reality environment. We developed a preparation for
chronic imaging of CA1 neuronal activity with cellular and subcellular
resolution in awake mice with their heads restrained while they ran on
a spherical treadmill. This was combined with a custom two-photon
microscope and light-blocking methods designed for imaging in the
presence of the visual display surrounding the treadmill that defined
a virtual environment. Using these methods, we optically recorded
from populations of ~80100 GCaMP3-expressing neurons in the
CA1 region of the hippocampus while the mouse was running along
a linear track. Subpopulations of the recorded neurons were identi-
fied as place cells. Because our methods could precisely determine the
physical location of the place cells in the hippocampus, we could show
that for cells separated by more than a few tens of micrometers, there
was no strong relationship in our experimental paradigm between
the location of place fields in the virtual reality environment and the
position of the corresponding place cells in area CA1. Nearby cells

1Department of Molecular Biology and Princeton Neuroscience Institute, Princeton University, Princeton, New Jersey, USA. 2Howard Hughes Medical Institute, Janelia
Farm Research Campus, Ashburn, Virginia, USA. Correspondence should be addressed to D.W.T. (dwtank@princeton.edu) or D.A.D. (ddombeck@princeton.edu).

Received 3 May; accepted 1 September; published online 3 October 2010; doi:10.1038/nn.2648

nature NEUROSCIENCE VOLUME 13 | NUMBER 11 | NOVEMBER 2010 1433

537
 t e c h n ic a l r e p o r t s

(less than <35 m separation) showed enhanced correlation, although
the possibility that this was produced by mixing of optical signals
from adjacent cells could not be excluded. In addition, we show that
it is possible to record place-related activity in putative hippocampal
interneurons and CA1 apical dendrites.
RESULTS
Combining two-photon microscopy and mouse virtual reality
Our apparatus (Fig. 1a,b) was designed around our previously
described virtual reality system15. The limbs of a head-restrained
mouse rested on a spherical treadmill12. A toroidal screen that sub-
tended a mouses visual field surrounded the treadmill and displayed
a computer-generated image of a virtual environment15,16. Ball move-
ments recorded with an optical computer mouse provided information
on running speed and direction and this was used by the computer
program that implemented the virtual environment to update posi-
tion and view angle. As reported previously15, head-restrained mice
were trained using operant conditioning to run back and forth along
a 180 cmlong virtual linear track (Fig. 1c). The mice received water
rewards at the ends of the track after successfully traversing the full
2010 Nature America, Inc. All rights reserved.
A window for chronic imaging of CA1 neurons in awake mice
The CA1 region of the hippocampus is more than a millimeter below
the cortical surface and cannot be directly imaged using two-photon
microscopy17. We carefully removed the overlying cortex by aspira-
tion (down to the external capsule) and replaced it with a metal can-
nula with a coverslip sealing one of the openings (see Online Methods,
Fig. 1e). This created a chronic hippocampal window that allowed
direct imaging of the hippocampus.
We used genetically encoded calcium indicators to optically record the
activity of CA1 neurons. Hippocampal injections of adeno-associated
virus AAV2/1-synapsin-1-GCaMP3 (a moderate, neuron-specific pro-
moter18 in a serotype with efficient gene delivery to neurons19) resulted
in expression of GCaMP3 in a large population of CA1 neurons. Two-
photon imaging of the GCaMP3-expressing area at a shallow depth
through the hippocampal window showed axons in the external capsule
and alveus that appeared as a dense plexus of fibers running parallel to
the hippocampal surface (Fig. 1f). Inspection of images from the stra-
tum pyramidale (~130160 m below the external capsule) revealed the
densely packed neurons of this cell body layer. We did not see voids of
fluorescence indicative of unlabeled cells, suggesting that local labeling

track length; consecutive rewards at the same end were not available. was nearly ubiquitous (Fig. 1f; note that GCaMP3 is excluded from the
We designed and constructed a two-photon microscope that could fit nucleus, which causes a donut-like appearance of the cell bodies13). All
within the geometric constraints of our virtual reality apparatus with- pyramidal neuron cell bodies in the stratum pyramidale have a promi-
out obstructing the mouses view of the display (Fig. 1a). The physical nent apical dendrite that projects ventrally into the stratum radiatum
dimensions of the design were most severely constrained by the small (deeper into the hippocampus). Imaging perpendicular to the long axis
distance (~13 cm) between the top of the headplate on the mouses head of these dendrites in the stratum radiatum showed their cross-sections
and the bottom of the reflecting mirror in the virtual reality projec- as small dots (Fig. 1f; see Supplementary Movie 1 for a z-series through
tion path. In addition, we designed the microscope to be completely the labeled CA1 region).
shielded from the bright light of the virtual reality projection display so
that the smaller number of photons from the fluorescent probe could be Optical identification of CA1 place cells
detected by the photomultiplier tubes (PMTs) without contamination. We developed a timeline for the sequence of steps necessary to image
We then implemented additional light-blocking measures at the laser hippocampal activity in mice trained to perform the spatial behavior
input port and the hole for the microscope
objective (see Online Methods; Fig. 1a,d,e) so
that the amount of detected background light
a Projector b
AAM
was less than ~5% of the baseline fluorescence L BP
RM FL
level from labeled cells. M PMT
CL M
DM
Axis of
Figure 1 Experimental setup. (a) The Z translation
rotation
experimental apparatus, consisting of a spherical Rubber TL
treadmill, a virtual reality apparatus (projector, tube
reflecting mirror (RM), angular amplification
Toroidal SL
c
mirror (AAM), toroidal screen and optical screen
computer mouse to record ball rotation) and Optical X-Y
computer
a custom two-photon microscope (titanium: mouse Ti:S
sapphire laser (Ti:S), long-pass filter (LP),
X-Y translation View from left end of
galvanometers (X-Y), scan lens (SL), mirror (M), Air LP virtual linear track
tube lens (TL), dichroic mirror (DM), collection
Sliding Top view of virtual linear track
lens (CL), biconcave lens (L), bandpass filter stage
Microscope
(BP), focusing lens (FL), photomultiplier tube
(PMT), sliding stage (used to move microscope d e Opaque
Kwik-Sil
Stainless
180 cm
Skull meta-bond
for treadmill access), X-Y translation (moves cannula
f Axons
in the
treadmill and mouse), Z-translation (objective Cortex Glass
Kwik-Sil external
focus control) and rubber tube (shown in cross- capsule/
section, for light shielding)). (b) Photograph alveus
of experimental setup. (c) Top, view from one
end of the virtual linear track. Bottom, top view Rubber
tube CA1 cell
of the linear track. (d) View of materials used bodies in
stratum
to block background light from entering the Outer metal
pyramidale
ring
microscope objective hole. Hippocampal imaging Inner metal
window can also be seen. (e) Detailed view of ring
CA1 apical
hippocampal imaging window (from boxed region dendrites in
in d). (f) In vivo two-photon images at different Headplate stratum
40 m radiatum
depths through the hippocampal window.

1434 VOLUME 13 | NUMBER 11 | NOVEMBER 2010 nature NEUROSCIENCE

538
 t e c h n ic a l r e p o r t s

Figure 2 Imaging CA1 place cells in the dorsal

hippocampus. (a) Two-photon image of neuron
a 2
cell bodies in stratum pyramidale of CA1 1
labeled with GCaMP3. The indicator is excluded 5
from the nucleus13. ROIs for example cells 10
8
are shown in red (right). (b) GCaMP3 baseline 9
7 6
subtracted F/F traces are shown in black for
a subset of the cells labeled in a (right). Red 4 3
traces indicate significant calcium transients
20 m
with <5% false positive error rates (see Online
Methods). The position of the mouse along the b c
virtual linear track and reward times are shown 9
at the bottom. (c) Expanded view of boxed
region in b. (d) Mean F/F versus linear track

Cell number
position for a subset of the cells labeled in a 6
(right). (e) A plot of mean F/F versus linear
track position for all of the cells labeled in a 3
(right). (f) Place cells are colored according 50%
to the location of their place fields along F/F
2
the virtual linear track. Only place cells with 180 cm
significant place fields during running in the Linear
track
positive direction are shown.
position
2010 Nature America, Inc. All rights reserved.

0 cm
Rewards
10 s 2s
task. Mice were first implanted with a metal
headplate to allow their heads to be restrained
d f 180 cm
9 2
while they were on the spherical treadmill. 1
They were placed on water scheduling for sev- 5
Cell number

6 Position of place
eral days and then trained in the virtual real- 30% mean 8
10 fields along linear
F/F track
ity apparatus. Once the mice were proficient 3 7 9
60 cm 6
at the task (~2 rewards per minute, ~710 d
3
of training) we injected the GCaMP3 virus, 2 4
0 cm
and the next day we implanted the hippo 0 cm 180 cm

campal window. The mice were returned e 10 1

to behavioral training for ~57 d until the

GCaMP3 expression produced an acceptable
Cell number

Normalized
signal-to-noise ratio for imaging calcium mean F/F
transients. Using the hippocampal imaging
window and viral delivery of GCaMP3, we
could image activity in CA1 neurons repeat- 1
edly over the course of about 34 weeks. 0 cm 180 cm
0

To optically identify and characterize place Position along linear track

cells, we collected time-series movies (~64 ms

per frame) of fields of view (~200 100 m) in the CA1 region of the
hippocampus containing ~80100 neurons (Fig. 2a) while the mice
navigated the virtual linear track. We acquired time series at a single
location in CA1 for ~913 min. In all, we analyzed 47 time series from
~10 different labeled regions in four mice. During the time-series
acquisitions, the mice obtained rewards at a rate of 3.2 0.7 rewards
per min with a mean distance run between rewards of 2.5 0.5 m,
similar to reward rates during our previous electrophysiology experi-
ments15. To extract the fluorescence versus time traces for the indi-
vidual neurons, we first corrected the movies for brain motion that
occurred during the acquisition using a two-dimensional cross corre-
lation algorithm (see Online Methods). It was difficult to draw regions
of interest (ROIs) manually around individual neurons because CA1
cell bodies are tightly packed together, GCaMP3 is found only in the
cytoplasm (near the cell edges), and the axial extent of the imaging
focal spot is probably more than ~4 m (ref. 20). Instead, we used an
automated cell identification method based on independent compo-
nent analysis (ICA) and principal component analysis (PCA)21,22.
Regions of interest for 10 neurons identified using this procedure
are shown in red for the example field of view shown in Figure 2a.
We extracted F/F traces (Fig. 2b) from the ICA/PCA-defined
ROIs. The F/F traces revealed a baseline that was periodically inter-
rupted by calcium transients that varied in amplitude (mean peak
F/F = 28 32%), consistent with a difference in the number of
underlying action potentials13,23,24, and varied in duration (mean
transient duration = 1.2 1.1 s), consistent with the summation of
multiple transients12,13,25. We identified significant transients with
<5% false positive error rates11,12 and used them in all subsequent
analyses (see Online Methods; Fig. 2b,c). We used these traces as a
surrogate measure of spiking activity and refer to them as the tempo-
ral activity patterns of the neurons.
In the 47 time-series datasets, we analyzed epochs in which the
mouse performed goal-directed behavior to receive a reward (high
reward rate epochs, see Online Methods). In nearly all of these data-
sets many individual neurons had location-specific activity. Figure 2b
shows the temporal activity patterns of four neurons from the exam-
ple field of view shown in Figure 2a, along with mouse position.
When the mouse ran in the positive direction (left to right as seen
in Fig. 1c), each of these four cells was active at a different location.
This can be seen in the individual traversal (Fig. 2c) and in the plot
of mean activity versus linear track position averaged over all 21 posi-
tive direction traversals (Fig. 2d,e). These plots reveal well-defined

nature NEUROSCIENCE VOLUME 13 | NUMBER 11 | NOVEMBER 2010 1435

539
 t e c h n ic a l r e p o r t s

We found the main results of this research

a Positive direction place cells Negative direction place cells
180 cm to be similar using either method of analysis
(Supplementary Fig. 2).

Position of place
13
9 11

fields along
linear track
14

17
15
c Across our 47 time-series datasets, we
16 2
6 350 identified 808 place cells and 881 place fields

Number of place fields

7
12 8
5 10 4
3 in those place cells. The mean total number
Positive direction place cells
20 m
Positive direction place cells
0 cm Mean = 0.83 0.25
of place cells per imaging field was 17 9
175
b 17 in positive direction
17
in negative direction and the mean total number of place fields per
1
imaging field was 19 10 (number of place
Cell number

cells with one place field was 735, number

Normalized
mean F/F
0
10 10 0 0.5 1.0
Directionality index with two place fields was 73; 18 6 complete
runs through each place field). The calcium
1 1 0 transients that defined these place fields had
0 cm 180 cm 0 cm 180 cm
a mean duration of 1.1 0.9 s (Fig. 4a) and a
Position along linear track Position along linear track
mean peak F/F of 35 29% (Fig. 4b). The
Figure 3 Place cells differ depending on the running direction in the linear track. (a) An example distribution of place fields as a function of
imaging field in which the place cells are colored according to the location of their place fields
position along the track was inhomogeneous
along the virtual linear track. Significant place fields during running in the positive (left panel)
and negative (right panel) directions are shown. Example place cells with different place fields (Fig. 4c), as seen previously31, with more cen-
or no place fields depending on the running direction are highlighted with closed arrowheads or tered between the reward sites and the towers
open arrowheads, respectively. (b) A plot of mean F/F versus linear track position for the positive (tall distal cues located ~60 cm from each end
2010 Nature America, Inc. All rights reserved.

direction place cells labeled in a (left) during running in the positive (left) and negative (right) of the track; see Fig. 1c) than in the middle
directions. (c) Histogram of directionality index for all place fields. of the track. The mean width of the optically
identified place fields was 50 19 cm (Fig. 4d),
and statistically significant fields of neuronal activity (P < 0.05 from which is ~20% larger than the mean width of place fields defined by
bootstrapping, see Online Methods). The spatially restricted activity firing rate in extracellular recordings under similar conditions15 (41
of these neurons and the well-defined shape of the activity fields were 14 cm). To investigate this broadening, we convolved spike train elec-
similar to extracellular recording measurements of place cells in real trophysiology recordings with model calcium transient waveforms
and virtual reality linear tracks15,26,27. Therefore, we concluded that (see Online Methods) and observed a similar increase in the resultant
these neurons are place cells. As well as identifying the place field place field width (~20%; see Supplementary Fig. 3). This result is
for each of the place cells, imaging also reveals their exact relative consistent with the idea that our place fields are defined by underlying
anatomical locations (Fig. 2f and Supplementary Movie 2). spiking activity that is convolved with the time-broadening effects of
Consistent with previous studies15,26, optically identified place cells intracellular calcium dynamics (Supplementary Fig. 3).
had different place fields in the same environment depending on the When optically identified place cells were active, the activity most
direction of running (Figs. 2b and 3). To analyze the directionality of frequently occurred in the place field; however, the place cells are not
activity during place field traversals, we divided the fluorescence time necessarily active during every traversal of the place field. This is similar
series into epochs defined by running direction. The mean fluorescence to previous tetrode study reports of excess activity variance28 in place
versus track position was calculated separately for each direction. A cell cells. Figure 5a shows the activity trace (temporal activity pattern) for
that had a place field during at least one direction of running was called each individual traversal of the mouse in the positive running direction
a place cell; each place cell could therefore have up to two place fields. for a subset of the place cells shown in Figure 2. Although place cells
Many place cells had different place fields depending on the running 2, 3, 6 and 9 are reliably active at the same location during nearly every
direction (Fig. 3a) or had a place field for only one running direction traversal (active during >75% of traversals through the place field and
(Fig. 3a). This high degree of directionality (Fig. 3b) was consistent active for >66% of the time spent in the place field), place cell 10 shows
across all datasets and was quantified for all place fields using a direc- more variable activity. This cell is reliably active at the same location, but
tionality index (0.83 0.25, Fig. 3c; 1 represents high directionality, not during each traversal (active during 38% of traversals through the
0 represents no directionality; see Online Methods for definition). place field and active for 35% of the time spent in the place field). For
Although we analyzed only high reward rate periods in the Results, all of the 881 place fields, the place cells were active during 65 18% of
we also analyzed place fields during different behavioral contexts2830 traversals through the place field (Fig. 5) and active for 53 17% of the
including the low reward rate periods (see Supplementary Fig. 1). time spent traversing through the place field (Fig. 5).

a b c d Minimum
3,000 3,000 100 70 width
Number of place fields
Number of place fields
Number of calcium

Number of calcium

50
transients

2,000 Mean = 50 19 cm
transients

Mean = 1.1 0.9 s Mean = 0.35 0.29

1,500 50
30
1,000
10
0 0 0 0
0 2 4 6 0 0.4 1.2 2.0 0 60 120 180 0 40 80 120
Place cell calcium transient Place cell calcium transient Linear track position (cm) Place field width (cm)
width (s) peak F/F

Figure 4 Characterization of place cell calcium transients and place fields. Histograms of place cell transient widths (a) and transient peak F/F (b) are
shown for periods of mouse movement along the virtual track. Histograms of place field position along the linear track (c) and place field widths (d) are
also shown.

1436 VOLUME 13 | NUMBER 11 | NOVEMBER 2010 nature NEUROSCIENCE

540
 t e c h n ic a l r e p o r t s

Figure 5 Place cell activity variability in place fields. (a) Temporal activity
pattern versus virtual linear track position traces for a subset of the cells
a Cell 2 Cell 3 Cell 6 Cell 9 Cell 10

shown in Figure 2a (right). Each of the 21 positive running direction

track traversals is shown for each of the cells. (b) Mean and s.d. of F/F
versus linear track position for the traces shown in a. (c) Histogram of the
probability that a place cell is active during traversals through the place
field. (d) Histogram of the percentage of place field traversal time for
which the cell had a significant calcium transient.

Comparison with electrophysiology

We compared the properties of place cells measured optically in mice
with hippocampal windows (cortical excavation and region infected
with AAV2/1-synapsin-1-GCaMP3) with the same properties measured
previously by electrophysiology in control mice (no cortical excavation
or virus infection)15. The place field width in the mice with hippo
campal windows (50 19 cm) was very similar to that of the control
mice (41 14 cm)15, and the effects of intracellular calcium dynamics in
the presence of GCaMP3 can explain the ~20% broadening of the fields
(Supplementary Fig. 3). The percentage of neurons that were place 50%
cells (~1520%) in mice with hippocampal windows was similar to, but
2010 Nature America, Inc. All rights reserved.

F/F

slightly less than, the percentage in our whole-cell patch clamp record- 0 cm 180 cm

ing15 (~2030%). We also directly compared electrophysiologically b

measured place cell properties and local field potentials (LFPs) in mice 30%
F/F
with hippocampal windows (cortical excavation and region infected
with AAV2/1-synapsin-1-GCaMP3) with the same measurements in c 100 d Minimum
occupancy
control mice (no cortical excavation or virus infection)15. The LFP- 120
Number of place fields

Number of place fields

80 Mean = 0.65 0.18 Mean = 0.53 0.17
recorded theta frequency in mice with hippocampal windows (7.3
60
0.3 Hz, n = 4 mice; Supplementary Fig. 4) was very similar (P = 0.32, 60
40
two-tailed t-test) to that of control mice (7.4 0.2 Hz, n = 11 mice). 20
The phase precession of spike times relative to LFP-recorded theta 0 0
activity observed in control mice (phase = 73 48, P < 0.01; corre 0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0
Probability of calcium transient Place field calcium transient
lation between phase and position of spikes: correlation coefficient = in place field occupancy
0.2 0.1, P < 0.01; n = 10 place fields, 3 mice15) was similar in mice
with hippocampal windows (phase = 104 87, P < 0.01; correlation with control rodents in both real and virtual environments15,26,27.
between phase and position of spikes: correlation coefficient = 0.4 Because our mice with hippocampal windows required cortical exca-
0.3, P < 0.01; n = 3 place fields, 2 mice; Supplementary Fig. 4). Finally, vation and infection with AAV2/1-synapsin-1-GCaMP3, these results
theta-modulated high-frequency (>50 Hz) bursts of action potentials also show that viral infection, GCaMP3 expression, and cortical exca-
with decreasing spike amplitude that were characteristic of place cells vation did not substantially alter the hippocampal CA1 dynamics.
in control mice were also found in mice with hippocampal windows
(see inter-spike intervals in Supplementary Fig. 4). The anatomical organization of CA1 place cells
These measurements show the similarity between place cells and We next examined the anatomical layout of place cells on the microm-
CA1 network dynamics in mice with hippocampal windows compared eter scale (see Fig. 6a for example fields of view). To quantify the

a 180 cm
b c
fields along linear

0.16
Position of place

180
75 0.045
track

Pairwise distance between

Pairwise neuron-neuron
correlation coefficient
place fields (cm)

120
0 cm 0.005
50
0 180 0.08 0 180
Place cells
60 All neurons

0 0
0 60 120 180 0 60 120 180
Pairwise distance between place cells (m) Pairwise neuron-neuron distance (m)

20 m

Figure 6 Spatial organization of place cells in dorsal CA1. (a) Example images from different fields of view in which the place cells are colored according to
the location of their place fields along the virtual linear track. Each image shows place cells with significant place fields during running in either the positive
or negative direction. (b) Plot of mean distance between place cells in the hippocampus versus the mean distance between their place fields along the track
averaged over all 47 time series. The error bars represent s.e.m. (c) Plot of mean distance between cells in the hippocampus versus the mean correlation
between their temporal activity patterns averaged over all 47 time series for all place cells (gray) and all neurons (black). The error bars represent s.e.m.

nature NEUROSCIENCE VOLUME 13 | NUMBER 11 | NOVEMBER 2010 1437

541
 t e c h n ic a l r e p o r t s

relationship between anatomical location and

place field location, we calculated all pair-
a b c 0 cm 180 cm
Axons
wise distances between place cells (distance 1
between each cells ROI center of mass) and S-O
S-P 3
pairwise distances between the peaks in their S-R 2
1
2
place fields separately for each running direc-
O-LM
tion for each imaging field. We combined the 20 m 3
25%
results from all 47 datasets and plotted the F/F
mean distance between place fields as a func- d e f 45 cm

tion of mean distance between the place cells Axons

(Fig. 6b). For cells spaced over a wide range S-O 1
of distances, but excluding near-neighbors S-P 30%
F/F
less than about 35 m apart, we found a sta- S-R
1 45 cm
2
tistically insignificant relationship between O-LM
the mean distance between place fields and 2
0 cm 180 cm
20 m
the mean distance between the place cells
(Spearmans rank correlation coefficient of Figure 7 Imaging place-related activity in dendrites and putative interneurons. (a,b) A two-photon
image (b) of a field of view ~75 m ventral to the stratum pyramidale cell body layer (dashed line
0.57, P = 0.2). Therefore, on average, the in a). Bright spots in b are a cross-section through the apical dendrites from the overlying CA1
distance between place cells in the physical neurons. O-LM, oriens lacunosum molecular; S-O, stratum oriens; S-P, stratum pyramidale;
space of the hippocampus was not related to S-R, stratum radiatum. (c) Mean F/F versus linear track position for the dendrites labeled in b.
2010 Nature America, Inc. All rights reserved.

the distance between their place fields along (d,e) A two-photon image (e) of a field of view ~50 m dorsal to the stratum pyramidale cell body
the track when the place cells were more layer (dashed line in d). Sparsely distributed cell bodies in e are assumed interneurons. (f) Mean
than about 35 m apart (although there F/F versus linear track position for the interneurons labeled in the image in e.
were some exceptions to this average rule;
Supplementary Fig. 1). The data point that corresponded to the most in all of the fields. Back-propagating action potentials invade the den-
closely neighboring place cells (less than about 35 m apart) was sig- dritic arbors of CA1 pyramidal neurons32,33. Therefore, it is likely
nificantly lower (P < 0.025, two-tailed t-test) than the points for all that the activity in the apical dendrites is due to opening of voltage-
other distances between cells. This could be interpreted as fine-scale gated calcium channels induced by back-propagating action potentials
spatial clustering of place cells with similar place fields, but residual rather than to the smaller calcium transients that are associated with
crosstalk, residual brain motion or a common neuropil signal between synaptic input.
neighboring ROIs could also explain this result (see Discussion). To show that our methods can be used to record from inter
We next studied whether the correlation between the temporal neurons, we adjusted the imaging plane to ~50 m dorsal to the
activity patterns of pairs of neurons varied as a function of the distance stratum pyramidale (Fig. 7d) where we found sparsely distributed
between them in the hippocampus. For all place cells in each imag- cell bodies (Fig. 7e). The neurons in this region of the stratum oriens
ing field, we calculated all pairwise temporal activity pattern correla- have been identified as interneurons, suggesting that these cells
tions and pairwise distances between place cells and then combined are probably basket, oriens lacunosum moleculare or axo-axonic
the results from all 47 datasets. When we plotted the mean temporal interneurons34. As in the stratum pyramidale, we found neurons with
correlation as a function of mean distance between the place cells spatially modulated activity patterns in the stratum oriens (Fig.7f);
(Fig.6c), we found a statistically insignificant relationship (Spearmans some of these had off fields (Fig. 7f), as has been observed with
rank correlation coefficient of 0.25, P = 0.5) for cells spaced more extracellular recording35,36. We imaged interneurons in four fields of
than about 35 m apart. Again, the data point that corresponded to view (~23 interneurons per field) in two mice and found spatially
the near-neighbor place cells was statistically different (P < 0.035, modulated interneuron activity patterns in all of the fields.
two-tailed t-test) from all but one of the other data points. When we
repeated this calculation for all neurons, including both place cells DISCUSSION
and non-place cells, the overall correlation was lower than for the Our method relies on several combined technologies. We developed
place cells alone and there was a statistically significant relationship surgical methods to implant a hippocampal window that allowed
between mean temporal correlation and distance between the neurons chronic subcellular resolution imaging in the CA1 region of the hip-
(Spearmans rank correlation coefficient of 0.95, P < 105; Fig. 6c), pocampus in behaving mice. The combination of this window with
regardless of whether the data point for the most closely neighboring a custom two-photon microscope and background light suppression
neurons was included. methods allowed imaging in mice that were interacting with a visual
virtual reality system. Using the genetically encoded calcium indicator
Imaging activity in dendrites and putative interneurons GCaMP3, our methods allowed us to study the spatially modulated
To test whether our imaging method had the required resolution activity patterns of pyramidal neurons in the stratum pyramidale,
and signal-to-noise ratio to record activity patterns from dendrites, putative interneurons in the stratum oriens and apical dendrites in the
we acquired time series ~75 m ventral to the stratum pyramidale stratum radiatum. By imaging the activity of populations of ~80100
(Fig.7). In this plane, we could identify apical dendrites from the neurons in the stratum pyramidale in mice that had been trained to
overlying pyramidal neurons (individual bright spots in Fig. 7b). When navigate along a virtual linear track, we identified place cells that had
we plotted the mean activity versus linear track position, we found characteristics very similar to those of spiking rate-defined place cells
well-defined place fields in many of the apical dendrites (Fig. 7c). in both real and virtual environments15,26,27,37.
We imaged dendrites in four fields of view (~50 dendrites per field) Tetrodes are the most common method used to record the acti
in three mice and saw spatially modulated dendritic activity patterns vity of CA1 place cells in behaving mice. Tetrode array recordings

1438 VOLUME 13 | NUMBER 11 | NOVEMBER 2010 nature NEUROSCIENCE

542

combined with spike-sorting procedures can be used to identify
around five single units per tetrode38,39. Although sub-millisecond
temporal resolution is a major advantage of this method compared to
imaging, tetrode methods are limited by low spatial resolution (>100s
of micrometers) and sparse sampling within the microcircuitry38,40.
Our imaging method has several advantages over tetrode methods.
For example, it is possible to report the precise anatomical position
of functionally identified neurons within the microcircuitry. Imaging
also allows researchers to make functional recording from subcel-
lular compartments, to identify all neurons (even silent cells), and to
take advantage of the growing number of available genetic tools41. It
should be possible to image subcellular dynamics such as signal trans-
duction42 and structural plasticity43 in the context of learning and
memory during behavioral paradigms. In addition, optical methods
might allow researchers to identify the functional properties of spe-
cific neurons in a large population, followed by either the subsequent
reconstruction of the underlying connections 44 or modification of
their activity45. Finally, imaging methods can allow the unambiguous
identification and recording of the same neurons over many weeks.
Although this was not the focus of this research, to address the techni-
2010 Nature America, Inc. All rights reserved.
cal feasibility of such studies we have imaged the same region of CA1
over multiple days (Supplementary Fig. 5).
Here we assumed that calcium transients could be used as a proxy
for spiking activity. Sodium action potentials generate calcium tran-
sients in GCaMP3-expressing hippocampal pyramidal neurons in
brain slices. Calcium transient amplitude increases linearly with the
number of evoked action potentials, until saturation at more than
about 10 action potentials13. There is a quantitatively similar rela-
tionship between spiking activity and calcium transient amplitude
for neocortical neurons in brain slices and in vivo in both anesthe-
tized and awake mice13. It is therefore reasonable to assume that the
spiking activity of the CA1 neurons in vivo generated calcium tran-
sients similar to those seen in slices13. Using these previous calcium
transient measurements13 and our extracellular place cell recordings,
we calculated the spike-induced calcium transients that would be
expected in our imaging experiments (Supplementary Fig. 3). The
characteristics of the optically defined place fields are consistent with
and can be fully explained by calcium transients induced by sodium
action potentials. Therefore, the contribution of other sources to our
calcium transients, such as calcium influx due to synaptic input or
dendritic calcium spikes, is probably small.
It is unlikely that our methods allow us to detect single action
potentials13, to determine firing rates or to reliably count the number
of spikes. Although these limitations probably do not pose a problem
for identifying place fields in place cells owing to the marked increase
in spiking rate in the place field and the ability to average over many
traversals, it is still possible that place cells with low activity levels
may not be detected.
Cortical excavation was used to expose the hippocampus for acute
electrophysiology experiments in anesthetized cats nearly 50 years
ago46. More recently, cortical excavation was combined with the use
of a polystyrene tube filled with agarose and sealed with a cover-
slip to facilitate two-photon imaging of dendritic spines in an acute,
anesthetized mouse preparation14. Here, we used these methods as a
starting point to develop a chronic hippocampal window designed for
imaging in behaving mice. The stainless steel cannula and coverslip
directly bonded to the external capsule surface with Kwik-Sil formed
a rigid support structure that minimized brain motion during animal
movements and allowed repeated imaging of the hippocampus for
weeks. The removal of the cortex overlying the hippocampus did
not detectably alter the mouses performance of the task, or the
nature NEUROSCIENCE VOLUME 13 | NUMBER 11 | NOVEMBER 2010 1439
t e c h n ic a l r e p o r t s
hippocampal dynamics and place cell properties that were measured
in our preparation compared to those of control mice.
Miniaturized head-mounted microscopes47,48 may allow hippo
campal imaging in freely moving animals. Such experiments would
benefit from the natural array of inputs, as opposed to the lack of
vestibular input and potentially altered gait of our mice. Our methods,
however, have the advantages of not requiring a miniaturized micro-
scope, easy combination with electrophysiology, and the potential to
manipulate specific environmental cues using virtual reality in ways
that would be difficult or impossible in real environments.
We used an ICA/PCA algorithm to identify individual neurons on
the basis of both their spatial and temporal characteristics21. However,
this method could not identify silent neurons and occasionally missed
active cells, making it hard to estimate the number of potentially active
neurons. Therefore, it was difficult to quantify the exact fraction of
neurons that were place cells. Based simply on the morphology of
neurons in the fields of view, we estimate that our fields contained
~80100 neurons, meaning that ~1520% of the neurons were place
cells, slightly less than previous estimates15. In addition to the uncer-
tain number of potentially active neurons, this slight difference could
also be due to differences in the recording methods and differences
in the definitions of place fields.
The ICA/PCA algorithm was successful in limiting the crosstalk
between neighboring ROIs. However, when studying the spatiotempo-
ral organization of neurons in CA1, we could not rule out the possibil-
ity of residual crosstalk, residual brain motion or a common neuropil
signal between the most closely neighboring cells (less than about
35 m). This is one possible explanation for the statistically outlying
data points in Figure 6b,c that correspond to the most closely neigh-
boring cells. An alternative explanation for these outlying data points
is that they represent small clusters of functionally similar neurons,
as recent studies have suggested5,9.
For place cells that were further separated (more than about
35 m), the distance between their place fields along the track was
statistically unrelated to the distance between their positions in the
hippocampus. A few previous electrode recording studies found
anatomically organized clusters of functionally similar hippocampal
neurons on the 0.61 mm spatial scale6,7, whereas a separate study
of CA1 place cell location concluded that there was a lack of ana-
tomical spatial organization 8. These previous methods provided
only an indirect measure of the spatial organization of place cells
in the CA1 microcircuitry. By contrast, using our methods, we were
able to directly measure the spatial organization. Although possible
crosstalk concerns meant that we could not unambiguously measure
the organization down to the finest scale, it was possible to conclude
that place cells are anatomically distributed down to a scale of at
least ~35 m.
For place cells separated by at least this scale, there was also no
relationship between the correlation between their temporal activity
patterns and the physical distance between them. For all neurons,
we found a statistically significant decrease in temporal correlation
between the neurons the further they were from each other in the
physical space of the hippocampus (Fig. 6). Although this decrease
is significant, both the overall correlation and the rate at which the
correlation decreases as a function of distance are nearly an order of
magnitude smaller than has been observed in the motor cortex of
behaving mice11.
Finally, we note that spatiotemporal organization can occur in many
forms, most of which were not examined here. The methods outlined
in this research should allow future studies to search more thoroughly
for micro-organization within the hippocampus.
543
 t e c h n ic a l r e p o r t s
Methods
Methods and any associated references are available in the online
version of the paper at http://www.nature.com/natureneuroscience/.
Note: Supplementary information is available on the Nature Neuroscience website.
Acknowledgments
This work was supported by Princeton University, the National Institutes of Health
(grant 5R01MH083686-03), The Howard Hughes Medical Institute, The Helen Hay
Whitney Foundation and The Patterson Trust. We thank F. Collman for the brain
motion correction algorithm.
AUTHOR CONTRIBUTIONS
D.A.D. and D.W.T. designed the research. D.A.D. performed the imaging
experiments and developed the chronic hippocampal window system and surgery/
training sequences. D.W.T. designed and implemented the combined two-photon
microscope and virtual reality instrumentation. D.A.D. and C.D.H. performed
extracellular recording and optimized virtual reality training procedures. L.T. and
L.L.L. provided AAV2/1-synapsin-1-GCaMP3. D.A.D. analyzed the data. D.A.D.
and D.W.T. wrote the paper.
COMPETING FINANCIAL INTERESTS
The authors declare no competing financial interests.
2010 Nature America, Inc. All rights reserved.
Published online at http://www.nature.com/natureneuroscience/. 29. Frank, L.M., Brown, E.N. & Wilson, M. Trajectory encoding in the hippocampus
Reprints and permissions information is available online at http://www.nature.com/
reprintsandpermissions/.
1. OKeefe, J. & Dostrovsky, J. The hippocampus as a spatial map. Preliminary evidence
from unit activity in the freely-moving rat. Brain Res. 34, 171175 (1971).
2. Ergorul, C. & Eichenbaum, H. The hippocampus and memory for what, where, 32. Regehr, W.G., Connor, J.A. & Tank, D.W. Optical imaging of calcium accumulation in
and when. Learn. Mem. 11, 397405 (2004).
3. OKeefe, J. Hippocampal neurophysiology in the behaving animal. in The
Hippocampus Book (ed. P. Andersen) 475548 (Oxford University Press, Oxford,
2007).
4. Leutgeb, S., Leutgeb, J.K., Moser, M.B. & Moser, E.I. Place cells, spatial maps and
the population code for memory. Curr. Opin. Neurobiol. 15, 738746 (2005).
5. Nakamura, N.H. et al. Hippocampal cells encode places by forming small anatomical
clusters. Neuroscience 166, 9941007 (2010).
6. Eichenbaum, H., Wiener, S.I., Shapiro, M.L. & Cohen, N.J. The organization of
spatial coding in the hippocampus: a study of neural ensemble activity. J. Neurosci.
9, 27642775 (1989).
7. Hampson, R.E., Simeral, J.D. & Deadwyler, S.A. Distribution of spatial and
nonspatial information in dorsal hippocampus. Nature 402, 610614 (1999).
8. Redish, A.D. et al. Independence of firing correlates of anatomically proximate
hippocampal pyramidal cells. J. Neurosci. 21, RC134 (2001).
9. Takahashi, S. & Sakurai, Y. Sub-millisecond firing synchrony of closely neighboring
pyramidal neurons in hippocampal CA1 of rats during delayed non-matching to
sample task. Front. Neural Circuits 3, 9 (2009).
10. Denk, W., Strickler, J.H. & Webb, W.W. Two-photon laser-scanning fluorescence
microscopy. Science 248, 7376 (1990).
11. Dombeck, D.A., Graziano, M.S. & Tank, D.W. Functional clustering of neurons in
motor cortex determined by cellular resolution imaging in awake behaving mice.
J. Neurosci. 29, 1375113760 (2009).
12. Dombeck, D.A., Khabbaz, A.N., Collman, F., Adelman, T.L. & Tank, D.W. Imaging
large-scale neural activity with cellular resolution in awake, mobile mice. Neuron
56, 4357 (2007).
13. Tian, L. et al. Imaging neural activity in worms, flies and mice with improved GCaMP
calcium indicators. Nat. Methods 6, 875881 (2009).
14. Mizrahi, A., Crowley, J.C., Shtoyerman, E. & Katz, L.C. High-resolution in vivo
imaging of hippocampal dendrites and spines. J. Neurosci. 24, 31473151
(2004).
15. Harvey, C.D., Collman, F., Dombeck, D.A. & Tank, D.W. Intracellular dynamics of
hippocampal place cells during virtual navigation. Nature 461, 941946 (2009).
16. Hlscher, C., Schnee, A., Dahmen, H., Setia, L. & Mallot, H.A. Rats are able to
navigate in virtual environments. J. Exp. Biol. 208, 561569 (2005).
17. Denk, W. & Svoboda, K. Photon upmanship: why multiphoton imaging is more than
a gimmick. Neuron 18, 351357 (1997).
1440 VOLUME 13 | NUMBER 11 | NOVEMBER 2010 nature NEUROSCIENCE
18. Thiel, G., Greengard, P. & Sudhof, T.C. Characterization of tissue-specific
transcription by the human synapsin I gene promoter. Proc. Natl. Acad. Sci. USA
88, 34313435 (1991).
19. Burger, C. et al. Recombinant AAV viral vectors pseudotyped with viral capsids from
serotypes 1, 2, and 5 display differential efficiency and cell tropism after delivery to
different regions of the central nervous system. Mol. Ther. 10, 302317 (2004).
20. Ji, N., Milkie, D.E. & Betzig, E. Adaptive optics via pupil segmentation for high-
resolution imaging in biological tissues. Nat. Methods 7, 141147 (2010).
21. Mukamel, E.A., Nimmerjahn, A. & Schnitzer, M.J. Automated analysis of cellular
signals from large-scale calcium imaging data. Neuron 63, 747760 (2009).
22. Reidl, J., Starke, J., Omer, D.B., Grinvald, A. & Spors, H. Independent component
analysis of high-resolution imaging data identifies distinct functional domains.
Neuroimage 34, 94108 (2007).
23. Kerr, J.N., Greenberg, D. & Helmchen, F. Imaging input and output of neocortical
networks in vivo. Proc. Natl. Acad. Sci. USA 102, 1406314068 (2005).
24. Sato, T.R., Gray, N.W., Mainen, Z.F. & Svoboda, K. The functional microarchitecture
of the mouse barrel cortex. PLoS Biol. 5, e189 (2007).
25. Greenberg, D.S., Houweling, A.R. & Kerr, J.N. Population imaging of ongoing neuronal
activity in the visual cortex of awake rats. Nat. Neurosci. 11, 749751 (2008).
26. McNaughton, B.L., Barnes, C.A. & OKeefe, J. The contributions of position,
direction, and velocity to single unit activity in the hippocampus of freely-moving
rats. Exp. Brain Res. 52, 4149 (1983).
27. Nakazawa, K. et al. Hippocampal CA3 NMDA receptors are crucial for memory
acquisition of one-time experience. Neuron 38, 305315 (2003).
28. Fenton, A.A. & Muller, R.U. Place cell discharge is extremely variable during
individual passes of the rat through the firing field. Proc. Natl. Acad. Sci. USA 95,
31823187 (1998).
and entorhinal cortex. Neuron 27, 169178 (2000).
30. Griffin, A.L., Eichenbaum, H. & Hasselmo, M.E. Spatial representations of
hippocampal CA1 neurons are modulated by behavioral context in a hippocampus-
dependent memory task. J. Neurosci. 27, 24162423 (2007).
31. Gupta, A.S., van der Meer, M.A., Touretzky, D.S. & Redish, A.D. Hippocampal replay
is not a simple function of experience. Neuron 65, 695705 (2010).
hippocampal pyramidal cells during synaptic activation. Nature 341, 533536 (1989).
33. Spruston, N., Schiller, Y., Stuart, G. & Sakmann, B. Activity-dependent action
potential invasion and calcium influx into hippocampal CA1 dendrites. Science
268, 297300 (1995).
34. Klausberger, T. et al. Brain state and cell typespecific firing of hippocampal
interneurons in vivo. Nature 421, 844848 (2003).
35. Ego-Stengel, V. & Wilson, M.A. Spatial selectivity and theta phase precession in
CA1 interneurons. Hippocampus 17, 161174 (2007).
36. Wilent, W.B. & Nitz, D.A. Discrete place fields of hippocampal formation
interneurons. J. Neurophysiol. 97, 41524161 (2007).
37. Schnee, A. Rats in Virtual Reality: The Development of an Advanced Method to
Study Animal Behaviour (Eberhard-Karls-University, Tbingen, 2008).
38. Buzski, G. Large-scale recording of neuronal ensembles. Nat. Neurosci. 7,
446451 (2004).
39. Wilson, M.A. & McNaughton, B.L. Dynamics of the hippocampal ensemble code
for space. Science 261, 10551058 (1993).
40. Henze, D.A. et al. Intracellular features predicted by extracellular recordings in the
hippocampus in vivo. J. Neurophysiol. 84, 390400 (2000).
41. Luo, L., Callaway, E.M. & Svoboda, K. Genetic dissection of neural circuits. Neuron
57, 634660 (2008).
42. Miyawaki, A. Visualization of the spatial and temporal dynamics of intracellular
signaling. Dev. Cell 4, 295305 (2003).
43. Holtmaat, A. & Svoboda, K. Experience-dependent structural synaptic plasticity in
the mammalian brain. Nat. Rev. Neurosci. 10, 647658 (2009).
44. Denk, W. & Horstmann, H. Serial block-face scanning electron microscopy to
reconstruct three-dimensional tissue nanostructure. PLoS Biol. 2, e329 (2004).
45. Boyden, E.S., Zhang, F., Bamberg, E., Nagel, G. & Deisseroth, K. Millisecond-
timescale, genetically targeted optical control of neural activity. Nat. Neurosci. 8,
12631268 (2005).
46. Kandel, E.R., Spencer, W.A. & Brinley, F.J. Jr. Electrophysiology of hippocampal
neurons. I. Sequential invasion and synaptic organization. J. Neurophysiol. 24,
225242 (1961).
47. Helmchen, F., Fee, M.S., Tank, D.W. & Denk, W. A miniature head-mounted two-
photon microscope. high-resolution brain imaging in freely moving animals. Neuron
31, 903912 (2001).
48. Sawinski, J. et al. Visually evoked activity in cortical cells imaged in freely moving
animals. Proc. Natl. Acad. Sci. USA 106, 1955719562 (2009).
544
 ONLINE METHODS
Two-photon microscope and data acquisition. Our microscope design is
shown in Figure 1a. Focusing control was achieved by mounting the micro-
scope objective on a single-axis piezo-motorized translation stage (Z-translation;
Agilis AG-LS25) within the microscope. X-Y control was achieved by mounting
the spherical treadmill and mouse head-restraint on a manual actuator-driven
translation stage.
The microscope was built out of custom-machined and black anodized alumi-
num parts and was designed to entirely enclose all of the internal optics except for
one hole for the microscope objective and one hole for the entry of the excitation
laser beam. The first step in mitigating stray light entry into the microscope was
to create precise fitting lips at every joint between any two pieces of the outer
shell of the microscope. This had the effect of eliminating nearly all light entry
through the outer body of the enclosed microscope. The first major point of
entry for stray light into the microscope was then the excitation laser input port.
A long-pass colored glass filter (Thorlabs, 780 long pass) covered this hole to
allow only the excitation laser light to pass, but not the shorter wavelength light
from the projection system. The second point of entry for stray light was the hole
for the microscope objective (Olympus 40, 0.8 NA). To combat light entry at this
point, one end of a loose tube of thin black rubber was attached to the microscope
around the objective (Fig. 1d). The other end of the tube was attached to a metal
ring that was machined to tightly fit around a separate ring that was cemented
2010 Nature America, Inc. All rights reserved.
to the top of the headplate and centered around the craniotomy. The flexibility
of the thin rubber allowed movement of the microscope objective with respect
to the craniotomy while maintaining the light-tight connections at both ends of
the rubber tube. It was also possible for stray light (especially longer wavelength
red light) to enter the microscope by traveling through the animal and subse-
quently through the exposed skull between the craniotomy and the inner edge
of the metal ring that was attached to the animals skull. This area was therefore
covered with opaque dental cement (Meta-bond, Parkell, made opaque by adding
India Ink at ~5% vol/vol) so that the only remaining source of stray light through
the microscope objective hole was through the hippocampal window itself
(Fig. 1e). The light through this window proved to be inconsequential in the
green imaging channel (<~5% of the baseline fluorescence level from labeled
cells). Though red fluorophores were not used in this research, we found that a
red light blocking filter (for example, 500/40, Chroma) in front of the projector
was needed to reduce the amount of stray light in the red imaging channel to
<~5% of the baseline fluorescence level from labeled cells.
The Ti:sapphire excitation laser (Chameleon Ultra II, Coherent) was oper-
ated at 920 nm (~3050 mW average power at the sample). Green GCaMP3
fluorescence was isolated using a bandpass filter (Semrock, 542/50) and detected
using a GaAsP PMT (1077P-40, Hamamatsu). ScanImage (v3.6)49 was used for
microscope control and acquisition. Images (256 64 pixels, ~200 100 m
field of view) were acquired at 15.6 Hz. During the ~913 min of time-series
acquisition at a single location, photo-bleaching was observed at a mean rate of
~12% per minute.
Our spherical treadmill and virtual reality system has been described12,15. We
used a Digidata (Axon Instruments, 1440A) data acquisition system to record
(Clampex 10.2) and synchronize position in the linear track, reward timing and
two-photon image frame timing (using the command signal to the slow galva-
nometer). Instead of the MX-1000 computer mouse used previously, here we used
an MX-518 (Logitech) mouse to record ball rotation. The ball movements mea
sured using the LED based MX-518 mouse were less sensitive to the exact distance
between the sensor and ball surface than the MX-1000. In addition, the MX-518
mouse can be adjusted between three gain settings, the lowest of which allowed
measurements of larger ball velocity without saturation than the MX-1000.
Mouse training, hippocampal window and AAV injections. All experiments
were performed in compliance with the Guide for the Care and Use of Laboratory
Animals (http://www.nap.edu/readingroom/books/labrats/). Specific protocols
were approved by the Princeton University Institutional Animal Care and Use
Committee. Imaging experiments were performed on five male C57BL/6 mice
(postnatal day ~50). First, mice were anesthetized (Isoflurane, ~1%), a headplate
was attached to the skull using opaque metabond and the stereotactic coordi-
nates overlying the hippocampus were marked (1.8 mm lateral, 2.0 mm caudal
to bregma). Any exposed skull was covered with a silicone elastomer (Kwik-Sil,
World Precision Instruments). Anesthesia was then removed and the mouse
doi:10.1038/nn.2648
recovered. Beginning the next day, the mice were put on a water restriction
regimen so that they received 1 ml of water per day. After ~57 days, final body
weight was ~80% of the before restriction weight. Training then began in the
virtual linear track (one ~45 min session per day) and continued until the mice
routinely ran back and forth along the linear track to achieve a high reward rate
(~2 rewards per minute); this required ~10 training sessions. The mice were
trained the same and the learning was the same as previously15.
The mice were then anesthetized and a small (~0.51.0 mm) craniotomy
was made at the previously marked spot on the skull overlying the dorsal hippo
campus. Three injections of a solution containing AAV2/1-synapsin-1-GCaMP3
were made (~30 nl per injection, ~5 min per injection) ~200 m apart at a depth
of ~1,300 m below the dural surface using a beveled glass micropipette (~2 M
after beveling). The craniotomy and any exposed skull were again covered with
Kwik-Sil. The mice then recovered from anesthesia. For 2 d prior to the injections,
the mice received 3 ml of water per day.
After 2448 h, the mice were re-anesthetized and a trephine drill (Fine Science
Tools) was used to cut a ~2.72.8-mm diameter craniotomy centered over the
previously made small craniotomy. The dura was then removed with forceps
and aspiration was used to slowly remove the cortex within the craniotomy. The
removal was accomplished very slowly with aspiration of ~50100 m of tissue
at a time followed by repeated irrigation with saline until any bleeding stopped.
These steps continued until the external capsule was exposed. The cortex and
top-most layers of the external capsule were peeled away so that the remaining
external capsule was never physically touched. To reduce animal movement-
induced brain motion that could interfere with time-series image acquisition,
the surface of the external capsule was allowed to dry until tacky, a small drop of
uncured Kwik-Sil was applied to the surface, and then the cannula was inserted
(with a few hundred micrometers left above the skull surface) and cemented
to the skull using opaque Meta-bond. When cured, the small drop of Kwik-Sil
had the effect of bonding the cannula to the external capsule and reducing in-
focal plane (X-Y) brain motion. In some cases, the cannula did not fit through
the craniotomy and a hand drill was used to slightly widen the hole. After this
hippocampal window implantation surgery, the mice typically woke up after
~10 min and were walking around the cage within ~15 min. The next day, train-
ing in the virtual linear track resumed as before the surgery. Approximately
7 days after the window surgery, imaging experiments began. The experiments
took place every 23 days for up to ~34 weeks in the same mouse (up to
~45 weeks after the window surgery). GCaMP3 expression reached a somewhat
steady state level ~14 days after injection. A small fraction of nuclear-filled
GCaMP3 neurons with altered physiology have been described in the cortex13.
A similar small fraction of nuclear-filled neurons were observed here in CA1
(starting ~23 weeks after injection). These neurons often generated large, long-
duration calcium transients that rarely defined a place field.
The cannula was composed of a 1.5-mm segment of a 2.77 mm outer diameter
and 2.31 mm inner diameter thin-walled stainless steel tube (Small Parts Inc.).
A 2.5-mm diameter round coverglass (Erie Scientific, Thermo Scientific) was
cemented to one end of the tubing using UV curable adhesive (Norland Products).
Does cortical excavation alter hippocampal dynamics and place cell
properties? To image the CA1 region of the hippocampus with cellular resolu-
tion, it was necessary to unilaterally remove part of the overlying cortex (includ-
ing parietal cortex and parts of visual and hindlimb sensory cortex). This raises
the concern that cellular and network properties within the hippocampus may
be altered either by direct mechanical trauma to the hippocampus or by altering
upstream inputs to the hippocampus due to the cortical lesion (although the
excavated cortical regions do not provide strong direct projections to the hippo
campal region). Regarding mechanical trauma, our surgical procedures allow the
removal of the overlying cortex without the need to physically touch the surface
of the hippocampus. The cortex and the top-most axons in the external capsule
were peeled away from the remaining external capsule without applying direct
pressure or inadvertent mechanical trauma to the hippocampus. Furthermore, the
external capsule itself provides a protective barrier to the underlying CA1 region
of the hippocampus. Therefore, overt signs of damage, such as vesiculated den-
drites or reduced or aberrant neuronal activity observed via calcium transients,
were not observed. Although overt signs of damage were easily avoided, more
subtle effects due to alterations of the upstream circuits projecting to the hippo
campus were still possible. However, the properties of CA1 place cells measured
nature NEUROSCIENCE
545
 electrophysiologically and optically in our hippocampal window mice (cortical
excavation and region infected with AAV2/1-synapsin-1-GCaMP3) were found to
be similar to the same properties measured previously with electrophysiology in
control mice (no cortical excavation or virus infection; see Results). Additionally,
no significant difference was observed in the task performance of the same mice
in the 4 days before the hippocampal window surgery compared to the 7 days
after the surgery (1.7 1.2 versus 1.8 0.6 rewards per min, respectively, P = 0.51,
two-tailed t-test, n = 5 mice, one ~45-min session per mouse per day).
Bolus loading of calcium indicators in CA1. We first attempted to use multi-cell
bolus loading50 of the exogenous indicator Oregon Green Bapta-1-AM. Although
we could label large populations of CA1 neurons and record their activity, this
method had the drawback of only labeling cells for a few hours. In addition, we
could only label cells immediately before the cannula was implanted, meaning
that the imaging experiments had to be performed on the same day as the cortical
excavation procedure and during the short time window after the mice awoke
from surgery but before the cell labeling disappeared. During these periods, the
mice rarely performed well at the task. A few place cells per field could be identi-
fied during periods of running, but the percentage of cells that were place cells
(<~5%) and the width of the place fields (~23 times greater than expected from
our electrophysiology measurements15) indicated altered hippocampal place cell
and/or network dynamics. Because of these limitations, and the advantages of
2010 Nature America, Inc. All rights reserved.
genetically encoded indicators, we used GCaMP3 (ref. 13).
Data analysis. Analysis was performed using ImageJ (1.40 g) and custom scripts
written in MatLab (version 7). All data in the text and figures are presented
as mean s.d., except in the plots shown in Figure 6b,c and Supplementary
Figure 2, where the error bars represent s.e.m.
Time-series datasets were motion corrected using whole-frame cross-
correlation. Although our previously published HMM line-by-line motion
correction algorithm12 would probably provide superior correction, our imaging
frame rate (~15 Hz) was fast enough that in-frame distortions were minimal using
whole-frame correction. In addition, it was more straightforward to apply the
ICA/PCA cell ROI selection algorithm to the whole-frame corrected time-series
than to the HMM-corrected time-series in which blank lines made automatic ROI
detection difficult. The mean frame-to-frame in-plane (X-Y) Euclidean distance
brain motion during all time periods (the mice ran almost continuously during
our time-series acquisitions) was 1.8 1.3 m; the motion during periods of
mouse movement along the virtual track used to define place fields (see below)
was 1.5 1.2 m. Out of plane (Z) motion (measured as described12) was
measured in untrained mice running on a spherical treadmill; during periods of
running the Z-motion was 0.7 0.2 m.
ROIs were defined as described previously21 (mu = 0.5, 150 principal com-
ponents, 100 independent components, s.d. threshold = 1.25, 60 pixels < area
of ROI < 400 pixels). Manual inspection revealed that the ROIs nearly always
defined single cell regions. F/F versus time traces were generated for each ROI.
Slow time-scale changes in the fluorescence traces were removed by examining
the distribution of fluorescence in a ~15-s interval around each sample time
point and subtracting the 8% percentile value. The baseline-subtracted neuron
fluorescence traces were then subjected to analysis of the ratio of positive- to
negative-going transients of various amplitudes and durations as described12. We
used this analysis to identify significant transients with <5% false positive error
rates and generated the significant transient-only traces (Figs. 2b,c and 5a) that
were used for all subsequent analysis11. Note that the significant transient only
traces are not discontinuous traces; the significant transients are left untouched,
but the time points between the significant transients are all set to 0.
Place fields were identified and defined as follows. Long running periods of
mouse movement along the virtual track in which the virtual velocity was >8.3
cm s1 and the run length was >53 cm (straight run without changing direction or
hitting the end of the track) were identified. These periods were first categorized
based on the direction of running (positive or negative direction). Positive and
negative long running direction periods were then further subdivided into two
categories based on the animals current performance of the task. Segments of
time between two rewards in which long running periods of only one direction
occurred were defined as high reward rate periods, all other long running periods
were defined as low reward rate periods. Timeseries datasets were only included
if a mean of at least 10 long running segments during high reward rate periods in
nature NEUROSCIENCE
the positive and negative directions were completed during time series acquisi-
tion (this occurred in 47 datasets). Only high reward rate periods were analyzed,
except in Supplementary Figures 1,2 and 5 where the low reward rate periods
were also included as a comparison (note that for the analysis in these figures
the requirement of a mean of at least 10 long running/high reward periods was
dropped, allowing the inclusion of additional datasets: 103 datasets total). For
each running direction (during high reward rate periods) for each cell, the mean
F/F was calculated as a function of virtual position for 80 position bins and this
mean fluorescence versus position plot was then smoothed (averaged) over three
adjacent points. Potential place fields were first identified as contiguous regions of
this plot in which all of the points were greater than 25% of the difference between
the peak F/F value (for all 80 bins) and the baseline value (mean of the lowest
20 out of 80 F/F values). These potential place field regions then had to satisfy
the following criteria. (i) The field must be >18 cm wide. (ii) The field must have
one value of at least 10% mean F/F. (iii) The mean in-field F/F value must be
more than three times the mean out-of-field F/F value. (iv) Significant calcium
transients must be present >30% of the time the mouse spent in the place field.
Potential place field regions that met these four criteria were then defined as
place fields if their P-value from boot strapping was <0.05. Bootstrapping was
performed by breaking the F/F trace for each cell into at least nine segments
(determined by the significant transients) that were randomly shuffled. The mean
shuffled F/F versus position plots were then subjected to the same criteria out-
lined above. This process was repeated 1,000 times and the P-value was defined
as the ratio of the number of times out of 1,000 that the random shuffled trace
generated a place field that met the above criteria.
A directionality index for mean F/F in a place field in positive and negative
running directions (F+, F) was defined as |F+ F|/(F+ + F). A directionality
index of 0 indicates identical activity in both directions, whereas an index of 1
indicates activity in one direction only.
Note that while very little movement along the width of the virtual linear track
was possible, this position as well as the view angle were not controlled for in
our analysis. These factors could contribute to increased variability of the place
fields. Also note that during time-series acquisition, the mice rarely stopped in
the middle (between reward zones) of the track (Supplementary Fig. 1): only
5 3% of the time per time series was the virtual velocity <0.05 cm s1 in the
middle of the track.
Correlation values were defined as the Pearsons correlation coefficient.
Correlation versus distance between neuron and place field distance versus
distance between neuron plots were generated from imaging fields that had at
least two place cells and any two cells that had overlapping ROIs were excluded
from the analysis. The position of a place field along the virtual track was defined
as the position within the field with the maximal F/F value. For the mean place
field width calculation, only fields in which neither field edge was at the end of
the track were included. This avoided the inclusion of fields that were artificially
narrow due to clipping at one end of the track.
Electrophysiology. Extracellular place cell recordings were performed in four mice
with hippocampal windows using methods similar to those described15. All pro-
cedures for these mice were identical to those used in imaging experiments (water
restriction, training, virus injection, surgeries, and so on) except that a small hole
(~0.5 mm) was drilled in the center of the window coverslip before implantation.
After implantation, this small hole allowed electrode access to the hippocampus
through the cannula. A tungsten metal electrode (3 M impedance at 1 kHz, FHC)
or glass microelectrode (resistance ~2.5 M, filled with 0.5 M NaCl) recorded
CA1 extracellular currents that were amplified (DAM80, WPI or AxoClamp
2B) and filtered (3 Hz10 kHz, Brown-Lee Model 440) before being digitized at
20 kHz with a Digidata (Axon Instruments, 1440A). Note that this preparation
was intended to mimic the conditions of the imaging experiments, but did not
allow long-duration extracellular recordings; the hole in the coverslip to allow elec-
trode access probably caused instability due to reduced stabilizing pressure. While
these recordings (~35 min duration per cell) allowed clear identification of place
cells and phase precession, the reduced number of passes through the place field
compared to longer recordings led to noisier place field maps and phase preces-
sion plots (Supplementary Fig. 4). To reduce low frequency (<~50Hz) electrical
noise observed during treadmill rotation, the ball was coated with unscented fabric
softener (Ultra Downy, Procter and Gamble) and allowed to dry overnight for
electrical recordings. In other experiments not reported here, treating the ball with
doi:10.1038/nn.2648
546
 Benz-all or anti-static sprays also reduced this noise. Further work will be required
to determine whether this noise is a result of charge buildup on the ball.
Spikes were sorted offline using a threshold analysis. At most one unit was
isolated from a single recording. Spike waveforms were overlayed and inspected
visually to make sure that they matched (Supplementary Fig. 4), and inter-
spike interval distributions were plotted to make sure no spikes fell within the
refractory period (12 ms; Supplementary Fig. 4). To create firing rate maps
(Supplementary Figs. 3 and 4), we divided the virtual linear track into 80 bins
and calculated the firing rate as the total number of spikes in a bin divided by the
total amount of time spent in a bin. To identify place fields, we found groups of
adjacent bins with firing rates greater than 0.25 times the rate in the peak bin. We
selected only those fields that were larger than 8 bins (18 cm) in length, had mean
in-field firing rates of greater than 1.5 Hz, and had mean in-field firing rates more
than three times larger than the mean out-of-field firing rate. Theta frequency
was defined using the lag to the first peak in the autocorrelation of the LFP trace
after band-pass filtering between 6 and 10 Hz using an FIR filter.
Phase precession was calculated as described15. Briefly, to assign a phase to a
spike occurring at time t, we identified, in the filtered trace, the times of the theta
peaks immediately preceding and following the spike (t1 and t2, respectively)
2010 Nature America, Inc. All rights reserved.
doi:10.1038/nn.2648
and calculated the phase as 360(tt1)/(t2t1). We circularly shifted the phase of
the spikes in 1 steps from 0 to 360, continuing across the 360 border, and fit a
linear regression line to the phase versus position plot at each rotation. We found
the rotation with the best fit, such that the sum of squared errors between the fit
line and data was minimized, and used the correlation coefficient between phase
and position at this rotation as a measure of phase precession.
Computed fluorescence traces (Supplementary Fig. 3) were created by con-
volving a spike train with model fluorescence transients (rise time: t1/2 = 52 ms,
decay time: t1/2 = 384 ms, peak amplitude taken from Figure 6c of ref. 13). All
values in these traces below 10% F/F were then set to 0 (Supplementary Fig. 3)
to mimic the significant-transient-only traces generated for the real fluorescence
time-series (Fig. 2b,c).
49. Pologruto, T.A., Sabatini, B.L. & Svoboda, K. ScanImage: flexible software for
operating laser scanning microscopes. Biomed. Eng. Online 2, 13 (2003).
50. Stosiek, C., Garaschuk, O., Holthoff, K. & Konnerth, A. In vivo two-photon calcium
imaging of neuronal networks. Proc. Natl. Acad. Sci. USA 100, 73197324
(2003).
nature NEUROSCIENCE
547


548
Vol 461 | 15 October 2009 | doi:10.1038/nature08499

ARTICLES
Intracellular dynamics of hippocampal
place cells during virtual navigation
Christopher D. Harvey1,2,3, Forrest Collman1,2,3, Daniel A. Dombeck1,2,3 & David W. Tank1,2,3

Hippocampal place cells encode spatial information in rate and temporal codes. To examine the mechanisms underlying
hippocampal coding, here we measured the intracellular dynamics of place cells by combining in vivo whole-cell recordings
with a virtual-reality system. Head-restrained mice, running on a spherical treadmill, interacted with a computer-generated
visual environment to perform spatial behaviours. Robust place-cell activity was present during movement along a virtual
linear track. From whole-cell recordings, we identified three subthreshold signatures of place fields: an asymmetric ramp-like
depolarization of the baseline membrane potential, an increase in the amplitude of intracellular theta oscillations, and a
phase precession of the intracellular theta oscillation relative to the extracellularly recorded theta rhythm. These
intracellular dynamics underlie the primary features of place-cell rate and temporal codes. The virtual-reality system
developed here will enable new experimental approaches to study the neural circuits underlying navigation.

Hippocampal place cells encode spatial information during navigation
using rate and temporal codes1,2. The rate code refers to a selective
increase in firing rate at a specific location in a local environment3,
and the temporal code includes the precise timing of spikes relative to
the hippocampal theta rhythm (phase precession)46. To explain the
origins of these codes, theoretical network and cellular models have
been proposed4,716. These models make differing predictions about the
subthreshold membrane potential dynamics of place cells (Fig. 1ae,
gi), reflecting differences in the proposed mechanisms underlying
hippocampal coding. The predicted intracellular dynamics include
steady oscillations at theta frequencies that reflect a global hippocampal
theta rhythm79,12 (Fig. 1e), modulation of the amplitude and frequency
of membrane potential theta oscillations4,10,11,1315 (Fig. 1bd), and
ramps of depolarization of the baseline membrane potential710,12,13
(Fig. 1de). To account for phase precession, these models also make
differing predictions about the relationship between intracellular theta
oscillations and the local field potential (LFP) theta rhythm. Models
predict either that intracellular theta is phase-locked to the LFP theta
rhythm with phase precession resulting from a ramp of depolari-
zation79,12, or that intracellular theta in the place field is at a higher
frequency than the LFP rhythm4,10,11,1315 (Fig. 1gi).
Because the models of hippocampal coding make differing predic-
tions of subthreshold membrane potential dynamics, it is possible to
distinguish between these models by intracellular measurements
from place cells during spatial behaviours. However, intracellular
recording methods require a level of mechanical stability that is dif-
ficult to obtain in freely moving animals17,18. Previously, head
restraint on a spherical treadmill has been used to facilitate optical
imaging at cellular resolution in awake, mobile mice19. Furthermore,
a previous study has provided evidence that body-tethered rats can
navigate through virtual environments20. We reasoned that these
could be combined to facilitate whole-cell recordings to show the
intracellular dynamics of place cells, and thus distinguish between
models of hippocampal coding.
Spatial behaviour in a virtual environment
The visual virtual-reality system for head-restrained mice we
developed is shown in Fig. 2a. A mouse runs on top of an air-supported
1
Princeton Neuroscience Institute, 2Lewis-Sigler Institute for Integrative Genomics, 3Department of Molecular Biology, Princeton University, Princeton, New Jersey 08544, USA.
2009 Macmillan Publishers Limited. All rights reserved
spherical treadmill with its head held fixed in space using a head plate.
The mouse is surrounded by a toroidal screen that covers a wide area to
accommodate a rodents large field of view20. An image is projected
onto the screen from a digital light processing projector by an angular
amplification mirror21 (Fig. 2a, Supplementary Fig. 1 and Methods).
To control the virtual-reality system, we developed custom software
using the open source Quake2 video game engine. The visual display
was updated on the basis of the movements of the animal, measured as
rotations of the spherical treadmill using an optical computer mouse
(Methods).
We addressed whether head-restrained mice can perform visually
guided spatial behaviours in a virtual environment. We trained
water-scheduled mice using operant conditioning to run along a
virtual linear track (180 cm long) that had proximal and distal walls
with varying patterns for location cues (Fig. 2b). Mice were able to
turn around at any position along the tracks length. Small water
rewards were given for running between reward zones located at
opposite ends of the track; consecutive rewards were not available
at a single reward site. After several training sessions, mice ran large
total distances with high peak running speeds (session 4: total dis-
tance 5 217 6 97 m per 40 min, peak speed 5 41 6 17 cm s21 over a
2 s period, mean 6 s.d.). Individual mice received rewards at increas-
ing rates over time (Fig. 2c, d). Also, the average distance travelled
between rewards decreased across sessions (Fig. 2e), consistent with
learning of the task. After ten training sessions, mice ran 281 6 53 cm
between rewards, on average, approaching an ideal performance of
180 cm per reward (that is, the distance between reward zones). These
data indicate that head-restrained mice can perform visually guided
spatial behaviours in a virtual-reality environment.
Place cells in a virtual environment
Although our behavioural results indicate that mice have a spatial
understanding of the virtual environment, the activation of naviga-
tion circuits during these behaviours would provide further evidence.
To assess the function of the hippocampal place-cell circuitry, we
performed acute extracellular recordings in the dorsal hippocampus
from CA1 pyramidal neurons (Fig. 3a). We recorded during beha-
viour along the virtual linear track from mice that had been trained
941
549
ARTICLES NATURE | Vol 461 | 15 October 2009

a Firing rate a Toroidal DLP projector c Session 4

screen
270 160
AAM

Position (cm)
Time
RM 80

Predicted subthreshold Vm Vext Above view

Vm 0
Water 0 150
b 60 Time (s)
tube 20
Session 10

Optical 160

Position (cm)
c mouse VR
computer 80
Air

d 0
0 150
b Time (s)
d
3

Rewards per
e

minute
2

1
View from left end
f Measured subthreshold Vm 0
2 6 10
e Session

between rewards (m)

18

Distance run
5 mV
10
0.2 s View from right end

Predicted phase precession 9 cm 2

g 2
LFP
h of the experimental set-up. A head-restrained mouse runs on an air-
Vm
i mirror (AAM). Movements of the treadmill are measured using an optical
Vm
Figure 1 | Predicted and measured subthreshold membrane potential
dynamics during a run through a cells place field. a, Schematic of a place-
cells firing rate. be, Schematics of predicted subthreshold membrane
potentials (aligned to a) from a dual oscillator interference model2,4 (b), a
modified dual oscillator model14 (c), a soma-dendritic interference
model10,11,13,15 (d), and network79 and experience-dependent12 models (e). In
b, two sets of theta-modulated inputs at different frequencies interfere to
create a beat-like pattern of membrane potential fluctuations. In c, two
oscillations are phase-locked outside the place field. In the place field, the
frequency of one oscillation increases, resulting in a modulation of the
summed oscillation. In d, the cell receives theta-modulated inhibitory and
excitatory inputs. In the place field, the excitatory drive increases, resulting
in a ramp-like depolarization and an increase in the amplitude of excitatory
theta oscillations. Depending on the conductances used, the summed
oscillation can have either increased (grey)15 or decreased (black)11
amplitude. In e, a ramp of excitatory drive interacts with theta-modulated
inhibitory inputs. Asymmetric ramps have also been proposed12. Schematics
in be only illustrate depolarizations and changes in theta amplitude.
f, Example of a subthreshold membrane potential (filtered from DC-10 Hz)
recorded intracellularly from a place cell. Note the simultaneous ramp of
depolarization and increase in theta oscillation amplitude. Scale bars refer to
the experimentally measured trace only. g, Schematic of the LFP theta
rhythm. Dashed lines denote peaks. hi, Schematics of predicted
relationships between intracellular and LFP theta to account for phase
precession. In h, intracellular and LFP theta are the same frequency. Phase
precession of spikes occurs relative to both intracellular and LFP theta owing
to a ramp of depolarization79,12. An asymmetric ramp is shown. In
i, intracellular theta in the place field is a higher frequency than LFP theta.
Spikes precess relative to LFP theta but not intracellular theta2,4,10,11,1315.
Schematics in h and i only illustrate the relationships between spike times,
intracellular theta and LFP theta.
942
2009 Macmillan Publishers Limited. All rights reserved
6 10
180 cm Reward site Session
Figure 2 | Spatial behaviours in a virtual-reality environment. a, Schematic
supported spherical treadmill. An image from a digital light processing
projector is displayed on a toroidal screen (22060u vertically, 270u
horizontally) by a reflecting mirror (RM) and an angular amplification
computer mouse. Water rewards are delivered through a lick tube by a
computer-controlled solenoid valve. See Methods and Supplementary Fig. 1
for details. b, The virtual linear track. Screenshots (without the fisheye
perspective, see Methods) from the right and left ends of the track are shown.
The track (180 3 9 cm) was divided into three regions with different textures
on the proximal walls (black dots, vertical stripes, white dots). Distal walls
(horizontal stripes, green with black crosses) were present at the boundaries
between regions. Water rewards were given at the ends of the track, with
available rewards alternating between reward sites. c, Example trajectories
for an individual mouse on training sessions 4 and 10. Position is the
animals location along the tracks long axis. Blue dots indicate rewards.
d, Rate of rewards for individual mice. e, Average distance travelled by the
mouse between consecutive rewards. In d and e, grey lines indicate
individual mice, and the black line is the mean; n 5 7 mice.
for at least 5 days. Recorded cells had spiking patterns characteristic
of hippocampal pyramidal neurons. Cells fired bursts of action
potentials at high frequencies (.50 Hz) with decreasing spike ampli-
tudes2224; bursts occurred at intervals of ,130 ms corresponding to
theta frequencies of ,610 Hz (Fig. 3a, b). In all mice tested, we
identified units with place-cell characteristics (23 cells from 8 mice;
Methods). Place cells had low overall firing rates with spatially modu-
lated firing patterns (overall firing rate 5 1.0 6 0.3 Hz; in-field firing
rate 5 4.7 6 2.6 Hz; out-of-field firing rate 5 0.6 6 0.2 Hz; Fig. 3c).
Place field size was, on average, 41 6 14 cm. Place-cell activity in
some cases had directionality, with different place fields and firing
rates depending on the direction of running25 (directionality
index 5 0.6 6 0.2; Methods and Fig. 3c). We also measured phase
precession of spike times relative to LFP theta oscillations during
runs through the place field4,5. Spike times shifted to earlier phases
during movement through the place field, and the phase and position
of spikes were negatively correlated (Dphase 5 272.6 6 47.7u
550
NATURE | Vol 461 | 15 October 2009 ARTICLES

a b ran on the spherical treadmill. Recordings lasted many minutes on

100
average (7.7 6 3.8 min, maximum 20.4 min; n 5 46 cells from 15

Number of spikes
mice), during which time mice ran long distances at high speeds in
the virtual environment (27 6 5 m per min; total distance range 125
50
458 m). We did not detect any major motion-induced artefacts in the
electrophysiology recordings. Recordings could be performed from
50 V
100 V 100 ms 0
the same animal across several days ($6 days per mouse; $3 days per
1 ms 103 101 101 hemisphere).
c ISI (s) A subset of our whole-cell recordings was made from place cells
7 Both directions 7 9 (overall firing rate 5 2.2 6 0.4 Hz; in-field firing rate 5 7.3 6 1.4 Hz;
Right runs
out-of-field firing rate 5 1.5 6 0.4 Hz; field size 5 43 6 13 cm; n 5 8
Firing rate (Hz)

Left runs
cells from 8 mice; Fig. 4a, b and Supplementary Figs 2a and 3).
Approximately 36% of the spontaneously active putative pyramidal
neurons had a place field along the virtual track, which is consistent
with estimates from extracellular recordings and immediate early
0 0 0

a
0 90 180
Position (cm)
d e
720
Phase (degrees)

300

65 mV 10 mV
360 200
2s
b 10 12 16
100
Firing rate (Hz)

0
ld th

ld th

0.2 mV 120 180

fie gh

fie gh

0.1 ms Position (cm)

of t ei

of ei
st
rs

La
Fi

Figure 3 | Extracellular recordings of CA1 place cells along the virtual linear 0 0 0
track. a, An example extracellular recording filtered between 500 Hz and
7.5 kHz. The inset shows overlaid spike waveforms from the recording. b, ISI c
60
distribution for the full duration of the recording shown in a. The time axis is 61
plotted on a log scale. c, Example firing rate maps for 3 place cells from 3 66
Vm (mV)

different mice. Top, firing rates at positions along the track are shown for 62
rightward runs (red), leftward runs (blue) and runs in either direction 68 63
(black). Bottom, the position on the track of each spike in the recording is
shown as a vertical line. A total of 23 place cells from 8 mice were recorded. 64 70 65
d, Phase precession of spike times relative to LFP theta. Top left, an example
0 90 180
extracellular recording, filtered between 2 Hz and 10 kHz, during a run Position (cm)
through the place field. Spikes and the LFP were recorded on the same d 60 e
electrode. Bottom left, the extracellular recording band-pass filtered between 12 4
6 and 10 Hz. Grey lines denote peaks (0u) in the filtered trace, black lines
Firing rate (Hz)

V (mV)

denote the times of spikes. Right, an example plot of phase (two cycles) 64
Vm (mV)

2
versus position on the virtual track for all spikes during complete runs 6
through the place field for a single cell. e, Phase values for spikes in the first 68
and last eighth of the place field. Connected points represent a single place 0
field. Horizontal lines indicate the means. n 5 10 place fields from 8 cells and 72 0
0 1
In-field

Out-of-field

3 mice (multiple place fields are owing to the directionality of firing rates).
between the first and last eighth of the field, P , 0.01; correlation
coefficient (C) 5 20.17 6 0.09 between phase and position,
P , 0.01; n 5 10 place fields, 8 cells, 3 mice; Fig. 3d, e). These firing
rate and phase precession characteristics along the virtual linear track
are highly similar to those measured in freely moving mice in real
environments2629. These data therefore demonstrate that hippocam-
pal place-cell circuitry is operational in head-restrained mice during
visually guided spatial behaviours in the virtual-reality system.
Intracellular dynamics of place cells
We next developed methods to measure the intracellular activity of
hippocampal neurons during behaviour along the virtual linear
track. Because the mouses head is stationary in the virtual-reality
set-up, we were able to perform whole-cell patch-clamp recordings
using a patch electrode with a long taper mounted on a standard
micromanipulator positioned outside the mouses field of view
(see Methods). We obtained recordings in awake mice17,18,30,31 as they
2009 Macmillan Publishers Limited. All rights reserved
0 1
Position (fraction Position (fraction
of place field) of place field)
Figure 4 | Ramp-like membrane potential depolarization inside place
fields. a, Example whole-cell recording during runs through the cells place
field. Grey boxes indicate the place field (middle example from b). b, Firing
rates along the virtual linear track for 3 place cells recorded intracellularly
from 3 different animals. The grey boxes indicate the primary place field
determined by firing rates (Methods). Bottom, vertical lines mark the
location along the track of every action potential in the recording. c, Average
baseline membrane potential, excluding action potentials, sorted by position
along the track for the 3 place cells from b. d, Average membrane potential
inside and outside the place field. Each pair of connected points is from a
single cell. Horizontal lines indicate the means. n 5 8 cells from 8 mice.
e, Average firing rates and changes in baseline membrane potential during
complete runs through the place field. To combine data from several cells,
position values in the place field were normalized. Black lines indicate the
mean. Grey lines indicate the mean 6 s.e.m. Data are averaged over 84
complete runs through the place field (8 cells).
943
551
ARTICLES NATURE | Vol 461 | 15 October 2009

gene studies in rats exploring real environments32,33 (see Methods).
Place cells recorded intracellularly fired in theta bursts (high fre-
quency bursts at .50 Hz occurring at theta frequencies of ,6
10 Hz), resulting in interspike interval (ISI) distributions with peaks
at ,10 ms and ,130 ms, consistent with our extracellular measure-
ments (compare Supplementary Fig. 4 and Fig. 3b). The individual
action potentials during a theta burst often occurred during the
ascending phase of an underlying theta oscillation (Fig. 5a, d).
Spike amplitudes decreased within a burst without a change in the
peak potential of the spike, suggesting that depolarizing intracellular
oscillations contribute to intra-burst decreases in spike amplitude34.
Whole-cell recordings from place cells can therefore be obtained in
head-restrained mice behaving in virtual environments.
From our whole-cell recordings of place-cell activity, we examined
two types of membrane potential dynamics proposed by the theoretical
models of place-cell function: ramps of depolarization, and modu-
lation of theta oscillations (Fig. 1be, hi). We first analysed changes
in the average baseline membrane potential during behavioural epochs
inside and outside the place field. As a mouse approached the recorded
cells place field, the average membrane potential, excluding action
potentials, increased in a ramp-like manner and remained increased
until the place field was passed (Fig. 4a). The ramp of depolarization
often began before action potential firing in the place field started, and
in some cases reached a steady depolarization as large as ,10 mV (peak
depolarization 5 5.7 6 2.9 mV; Fig. 4a and Supplementary Figs 5 and
6). These depolarization events occurred preferentially in place fields;
the average membrane potential excluding action potentials was higher
in place fields than outside of place fields (Vin-field 2 Vout-of-field 5
2.5 6 0.5 mV, P , 0.0001; Fig. 4c, d and Supplementary Fig. 2b).
Consistently, the baseline membrane potential and firing rate were
strongly correlated (C 5 0.55 6 0.10, P , 0.001). On complete runs
through the place field, ramps of depolarization were asymmetric, with
the peak depolarization shifted towards the end of the field (position of
the peak depolarization 5 72 6 24% of the distance through the field,
P , 0.05 versus 50%; slope before peak 5 3.0 mV per place field length,
slope after peak 5 5.4 mV per field length, P , 0.001; Fig. 4e). In con-
trast, firing rates were symmetric within the place field (position of the
peak firing rate 5 52 6 17% of the distance through the field, P . 0.6
versus 50%; slope before peak 5 13.3 Hz per place field length, slope
after peak 5 13.6 Hz per field length, P . 0.4; Fig. 4e), perhaps because
firing rates were highest on the ascending part of a ramp depolariza-
tion13 (Supplementary Fig. 7a). An asymmetric ramp-like depolariza-
tion of the baseline membrane potential is therefore a subthreshold
signature of place fields.
We next examined the modulation of the amplitude and phase of
membrane potential oscillations occurring at theta frequencies during
runs along the virtual track. We measured intracellular theta oscilla-
tions by band-pass filtering our membrane potential recordings from
610 Hz, after action potentials were removed (Fig. 5a and Methods).
When the mouse entered the recorded cells place field, the amplitude

a b c
4 3
Theta power (mV2)

2.2 1.6
2
Vm
2
1.6 1
1
68 mV 15 mV 1 0 0.4

In-field

Out-of-field
2s 0 90 180
Position (cm)
4 mV
Filtered Vm d e
2s
(610 Hz) 720 400
Phase (degrees)

360 340
15 mV 64 mV
0.1 s
0 280
First eighth

Last eighth
5 mV 40 120

of field
of field
0.1 s Position (cm)
20 mV
10 mV
f 0.1 s
Spike times
Intracellular theta peaks g Spikes Spikes
Intracellular theta
LFP theta peaks peaks
720
LFP phase (degrees)

720 720
Intracellular phase

Vm
(degrees)

69 mV 360 360 360

LFP 15 mV 8 mV
0.4 mV 0.3 mV 0 0
0.2 s 110 160 0 110 160
110 160
0.2 s
Position (cm)

Figure 5 | Membrane potential theta oscillations in place cells. a, Top,
example raw and filtered (610 Hz after spikes were removed) membrane
potential traces during runs through the cells place field (grey boxes).
Bottom, expanded portions of the raw and filtered traces. b, Power in the
theta-frequency band sorted by position along the track for the cells from
Fig. 4. Power was measured as the squared amplitude of the filtered trace.
Grey boxes denote the primary place fields. c, Theta power inside and
outside the place field for individual cells. Horizontal bars are means. n 5 8
cells from 8 mice. d, Spike times relative to intracellular theta. Left, example
raw and filtered (610 Hz) membrane potential traces in the place field. Grey
lines indicate peaks of theta oscillations (0u), black lines are times of spikes.
Right, example phase (two cycles) versus position plot during place field
traversals for a single cell. e, Intracellular phase values for spikes in the first
944
2009 Macmillan Publishers Limited. All rights reserved
and last eighth of the place field for individual place fields. Horizontal lines
are the means. n 5 12 place fields and 8 cells from 8 mice. f, Raw (left) and
filtered (610 Hz, right) membrane potential and LFP traces in the place field
from a simultaneous LFP and whole-cell recording. The times of LFP theta
peaks (grey lines), intracellular theta peaks (circles) and spikes (crosses) are
shown to illustrate the phase precession of spikes and intracellular theta
relative to LFP theta oscillations, and the absence of phase precession of
spikes relative to intracellular theta oscillations. g, Phase precession of spike
times relative to intracellular theta (left), relative to LFP theta (middle) and
phase precession of intracellular theta peak times relative to LFP theta
(right) from a simultaneous LFP and whole-cell recording. In d and f, the top
and bottom scale bar labels denote the top and bottom traces, respectively.
552
NATURE | Vol 461 | 15 October 2009
of intracellular theta oscillations increased (Fig. 5a). Theta-band
power in the membrane potential trace was higher in place fields than
outside of place fields (powerin-field 5 1.7 6 0.4 mV2, powerout-of-field 5
0.8 6 0.2 mV2, P , 0.01; Fig. 5b, c and Supplementary Figs 2c and 8).
Consistently, theta oscillation amplitude and firing rate were highly
correlated (C 5 0.61 6 0.09, P , 0.001). In contrast, the amplitude of
membrane potential theta oscillations was similar at all spatial loca-
tions for putative CA1 pyramidal neurons that did not have a place
field (Supplementary Fig. 9) and for the LFP theta rhythm (Sup-
plementary Fig. 10).
To examine the modulation of the phase of intracellular theta, we
compared intracellular theta fluctuations with LFP theta oscillations.
We began by looking for phase precession of spike times relative to
membrane potential theta oscillations. The intracellular phases of
spike times did not change during runs through the place field,
and the phase and position of spikes were not correlated
(Dphase 5 0.6 6 10.4u between the first and last eighth of the field,
P . 0.9; C 5 20.01 6 0.08 between phase and position, P . 0.6;
Fig. 5dg). Because spike times advanced relative to LFP theta oscil-
lations but not intracellular theta (compare Figs 3d, e and 5d, e), it is
predicted that a phase shift between LFP and intracellular theta
occurs during place field traversals. We therefore performed simu-
ltaneous whole-cell and LFP recordings to compare directly the
phases of intracellular and extracellularly recorded theta. During
runs through the place field, the phase difference between intracel-
lular and LFP theta shifted such that the intracellular theta oscillation
phase precessed relative to the LFP theta rhythm (C 5 20.26 6 0.12
between LFP phase and position for the times of intracellular theta
peaks; n 5 2 cells from 2 mice; Fig. 5f, g and Supplementary Fig. 11a).
Consistently, the frequency of intracellular theta oscillations in the
place field was higher than the frequency of LFP theta fluctua-
tions (measured as a ratio of periods of intracellular theta to
periods of LFP theta; ratioin-field 5 0.97 6 0.21, P , 0.01 versus 1;
Dfrequency 5 0.22 Hz given a mean LFP frequency of 7.4 Hz; see
Methods and Supplementary Fig. 11b). In contrast, the frequencies
of intracellular theta and LFP theta during epochs outside the place
field were similar (ratioout-of-field 5 1.01 6 0.22, P . 0.2 versus 1;
Supplementary Fig. 11b). Intracellular theta oscillations in place cells
were therefore not constant in amplitude or phase (relative to LFP
theta) throughout runs on the linear track; rather, membrane poten-
tial theta oscillations were dynamically modulated across positions in
virtual space.
Ramp-like depolarizations of the membrane potential and
increases in intracellular theta oscillation amplitude were present
simultaneously, which can be demonstrated directly by filtering the
membrane potential trace from DC-10 Hz (Fig. 1f). Consistently,
intracellular theta power and the baseline membrane potential were
highly correlated (C 5 0.52 6 0.07, P , 0.001; Figs 4c and 5b and
Supplementary Fig. 2b, c). To determine whether ramps of depolariza-
tion trigger increases in theta amplitude, we injected ramps of current
at the soma while the animal was running along the virtual track, and
measured changes in theta power. Theta power increased weakly with
higher levels of depolarization; however, the increase in power was
smaller than during runs through the place field (P , 0.01 at similar
DV values; compare Figs 4c, 5b and Supplementary Fig. 7b). Ramp-like
depolarizations of the membrane potential therefore were not suf-
ficient to cause the increases in theta oscillation amplitude observed
in place fields.
Our whole-cell recordings revealed two further subthreshold
phenomena. First, in a fraction of our recordings we observed
spikelets, brief small amplitude deflections of the membrane poten-
tial (2 out of 8 place cells; amplitude 5 7.4 6 1.3 mV, full-width at
half-maximum 51.6 6 0.4 ms; 0.06 6 0.05 spikelets per second; Sup-
plementary Fig. 12a)35. Second, bursts of action potentials were occa-
sionally followed by large (,1025 mV) depolarizations lasting up to
50100 ms23 (Supplementary Fig. 12b). These events sometimes con-
tained broadened spikes of reduced amplitude, consistent with Ca21
2009 Macmillan Publishers Limited. All rights reserved
ARTICLES
spikes recorded in slices36. Further analysis is required to assess the
prevalence, significance and mechanisms of these events.
Discussion
Here we have developed a visual virtual-reality system in which head-
restrained mice performed spatial behaviours along a virtual linear
track. Hippocampal place-cell activities were triggered during runs
along the track, with similar properties to those recorded in real environ-
ments. Because the mouse was head-restrained, we were able to obtain
whole-cell recordings lasting many minutes in awake mice using
standard patch-clamp techniques. Furthermore, this set-up can prob-
ably be combined with two-photon laser scanning microscopy, which
has previously been performed in mice running on the spherical tread-
mill19. Virtual reality also offers the ability to design highly custom
environments, and to manipulate these environments rapidly through
software. We therefore anticipate that our virtual-reality system will
make new types of experiments exploring spatial information encoding
possible.
We identified three subthreshold signatures of place fields: an
asymmetric ramp-like depolarization of the baseline membrane
potential, an increase in the amplitude of membrane potential theta
oscillations, and a phase precession of intracellular theta relative to
LFP theta, such that spike times advanced relative to LFP theta but
not intracellular theta (Figs 1f, 4 and 5). These findings seem incon-
sistent with the mechanisms underlying dual oscillator models2,4,14,
network models79 and an experience-dependent model12 of hippo-
campal rate and temporal coding because each model predicts mem-
brane potential dynamics that differ significantly from our observed
subthreshold signatures (Figs 1b, c, eh). Our data are most consist-
ent with a soma-dendritic interference model that proposes interac-
tions between a spatially independent inhibitory oscillatory input
near the soma, and a spatially dependent (increasing in the place
field) and temporally patterned (at theta) dendritic excitatory
input10,11,13,15. With an appropriate choice of conductances15, the
soma-dendritic interference model predicts, in the place field,
ramp-like depolarizations (Figs 1d and 4), an increase in theta ampli-
tude (Figs 1d and 5ac), and precession of intracellular theta relative
to extracellular theta (Figs 1i and 5f, g). We note, however, that we
observed an asymmetric ramp-like depolarization (Fig. 4e), which
may be important for phase precession across the entire place field12,
compared with a symmetric ramp proposed in this model. We do not
exclude the possibility that revised forms of other models could
potentially explain the intracellular dynamics we observed. Other
experiments will be necessary to define and quantify the cellular
and synaptic mechanisms underlying the intracellular dynamics we
measured, including in the context of the entorhinal cortex2,16,37, and
to establish their causal relationship to rate and temporal codes in the
hippocampus. The virtual-reality system developed here, combined
with electro- and opto-physiological methods, will probably facilitate
this analysis.
METHODS SUMMARY
A virtual-reality system was designed using an air-supported spherical treadmill for
head-restrained mice19, in combination with a projection-based visual display
system20, in which a toroidal screen presented an image from a projector by an
angular amplification mirror21. Custom software to control the virtual-reality
system was developed on the basis of the Quake2 game engine. Rotations of the
spherical treadmill, measured by an optical computer mouse, were used to update
the visual display. Water-scheduled C57BL/6J mice (812 weeks old) were trained
using operant conditioning to run from end-to-end of a virtual linear track
(180 cm long) to obtain water rewards. For electrophysiology measurements, a
small craniotomy (,0.5 mm diameter) was made centred over dorsal hippocam-
pus (2.2 mm caudal, 1.7 mm lateral to bregma). The craniotomy was sealed with
silicone grease and then covered with silicone elastomer to allow recordings across
several days. Extracellular recordings were made using a glass electrode (filled with
0.5 M NaCl, ,2.5 MV pipette resistance) mounted on a micromanipulator posi-
tioned behind the mouse. Whole-cell recordings were obtained using standard
blind patch methods. Patch pipettes were pulled with a long taper (,100 mm
945
553
ARTICLES
diameter at 1 mm from the tip) to minimize damage to the overlying cortical tissue,
and were mounted on a micromanipulator positioned outside the field of view.
Firing rate maps were calculated for 80 spatial bins along the virtual track as the
number of spikes in a bin divided by the time spent in that bin. Changes in baseline
membrane potential in the place field were measured from membrane potential
traces excluding spikes. Theta oscillations were analysed after band-pass filtering
(610 Hz) of the membrane potential recording using a linear phase finite impulse
response filter.
Full Methods and any associated references are available in the online version of
the paper at www.nature.com/nature.
Received 8 July; accepted 15 September 2009.
1. Moser, E. I., Kropff, E. & Moser, M. B. Place cells, grid cells, and the brains spatial
representation system. Annu. Rev. Neurosci. 31, 6989 (2008).
2. OKeefe, J. & Burgess, N. Dual phase and rate coding in hippocampal place cells:
theoretical significance and relationship to entorhinal grid cells. Hippocampus 15,
853866 (2005).
3. OKeefe, J. & Dostrovsky, J. The hippocampus as a spatial map. Preliminary
evidence from unit activity in the freely-moving rat. Brain Res. 34, 171175 (1971).
4. OKeefe, J. & Recce, M. L. Phase relationship between hippocampal place units
and the EEG theta rhythm. Hippocampus 3, 317330 (1993).
5. Skaggs, W. E., McNaughton, B. L., Wilson, M. A. & Barnes, C. A. Theta phase
precession in hippocampal neuronal populations and the compression of
temporal sequences. Hippocampus 6, 149172 (1996).
6. Buzsaki, G. Theta oscillations in the hippocampus. Neuron 33, 325340 (2002).
7. Tsodyks, M. V., Skaggs, W. E., Sejnowski, T. J. & McNaughton, B. L. Population
dynamics and theta rhythm phase precession of hippocampal place cell firing: a
spiking neuron model. Hippocampus 6, 271280 (1996).
8. Jensen, O. & Lisman, J. E. Hippocampal CA3 region predicts memory sequences:
accounting for the phase precession of place cells. Learn. Mem. 3, 279287
(1996).
9. Wallenstein, G. V. & Hasselmo, M. E. GABAergic modulation of hippocampal
population activity: sequence learning, place field development, and the phase
precession effect. J. Neurophysiol. 78, 393408 (1997).
10. Kamondi, A., Acsady, L., Wang, X. J. & Buzsaki, G. Theta oscillations in somata and
dendrites of hippocampal pyramidal cells in vivo: activity-dependent phase-
precession of action potentials. Hippocampus 8, 244261 (1998).
11. Magee, J. C. Dendritic mechanisms of phase precession in hippocampal CA1
pyramidal neurons. J. Neurophysiol. 86, 528532 (2001).
12. Mehta, M. R., Lee, A. K. & Wilson, M. A. Role of experience and oscillations in
transforming a rate code into a temporal code. Nature 417, 741746 (2002).
13. Harris, K. D. et al. Spike train dynamics predicts theta-related phase precession in
hippocampal pyramidal cells. Nature 417, 738741 (2002).
14. Lengyel, M., Szatmary, Z. & Erdi, P. Dynamically detuned oscillations account for
the coupled rate and temporal code of place cell firing. Hippocampus 13, 700714
(2003).
15. Gasparini, S. & Magee, J. C. State-dependent dendritic computation in
hippocampal CA1 pyramidal neurons. J. Neurosci. 26, 20882100 (2006).
16. Maurer, A. P. & McNaughton, B. L. Network and intrinsic cellular mechanisms
underlying theta phase precession of hippocampal neurons. Trends Neurosci. 30,
325333 (2007).
17. Lee, A. K., Manns, I. D., Sakmann, B. & Brecht, M. Whole-cell recordings in freely
moving rats. Neuron 51, 399407 (2006).
18. Lee, A. K., Epsztein, J. & Brecht, M. Head-anchored whole-cell recordings in freely
moving rats. Nature Protocols 4, 385392 (2009).
19. Dombeck, D. A., Khabbaz, A. N., Collman, F., Adelman, T. L. & Tank, D. W. Imaging
large-scale neural activity with cellular resolution in awake, mobile mice. Neuron
56, 4357 (2007).
20. Holscher, C., Schnee, A., Dahmen, H., Setia, L. & Mallot, H. A. Rats are able to
navigate in virtual environments. J. Exp. Biol. 208, 561569 (2005).
946
2009 Macmillan Publishers Limited. All rights reserved
NATURE | Vol 461 | 15 October 2009
21. Chahl, J. S. & Srinivasan, M. V. Reflective surfaces for panoramic imaging. Appl.
Opt. 36, 82758285 (1997).
22. Ranck, J. B. Jr. Studies on single neurons in dorsal hippocampal formation and
septum in unrestrained rats. I. Behavioral correlates and firing repertoires. Exp.
Neurol. 41, 461531 (1973).
23. Kandel, E. R. & Spencer, W. A. Electrophysiology of hippocampal neurons. II.
After-potentials and repetitive firing. J. Neurophysiol. 24, 243259 (1961).
24. Quirk, M. C. & Wilson, M. A. Interaction between spike waveform classification
and temporal sequence detection. J. Neurosci. Methods 94, 4152 (1999).
25. McNaughton, B. L., Barnes, C. A. & OKeefe, J. The contributions of position,
direction, and velocity to single unit activity in the hippocampus of freely-moving
rats. Exp. Brain Res. 52, 4149 (1983).
26. Nakazawa, K. et al. Hippocampal CA3 NMDA receptors are crucial for memory
acquisition of one-time experience. Neuron 38, 305315 (2003).
27. Kentros, C. G., Agnihotri, N. T., Streater, S., Hawkins, R. D. & Kandel, E. R.
Increased attention to spatial context increases both place field stability and
spatial memory. Neuron 42, 283295 (2004).
28. Cacucci, F., Wills, T. J., Lever, C., Giese, K. P. & OKeefe, J. Experience-dependent
increase in CA1 place cell spatial information, but not spatial reproducibility, is
dependent on the autophosphorylation of the alpha-isoform of the calcium/
calmodulin-dependent protein kinase II. J. Neurosci. 27, 78547859 (2007).
29. Sun, L. D. & Wilson, M. A. Impaired and Enhanced Spatial Representations of the
PSD-95 Knockout Mouse. PhD thesis, Massachusetts Institute of Technology
(2003).
30. Margrie, T. W., Brecht, M. & Sakmann, B. In vivo, low-resistance, whole-cell
recordings from neurons in the anaesthetized and awake mammalian brain.
Pflugers Arch. 444, 491498 (2002).
31. Crochet, S. & Petersen, C. C. Correlating whisker behavior with membrane
potential in barrel cortex of awake mice. Nature Neurosci. 9, 608610 (2006).
32. Wilson, M. A. & McNaughton, B. L. Dynamics of the hippocampal ensemble code
for space. Science 261, 10551058 (1993).
33. Guzowski, J. F., McNaughton, B. L., Barnes, C. A. & Worley, P. F. Environment-
specific expression of the immediate-early gene Arc in hippocampal neuronal
ensembles. Nature Neurosci. 2, 11201124 (1999).
34. Henze, D. A. et al. Intracellular features predicted by extracellular recordings in
the hippocampus in vivo. J. Neurophysiol. 84, 390400 (2000).
35. Kandel, E. R. & Spencer, W. A. Electrophysiology of hippocampal neurons. IV. Fast
prepotentials. J. Neurophysiol. 24, 272285 (1961).
36. Wong, R. K. & Prince, D. A. Participation of calcium spikes during intrinsic burst
firing in hippocampal neurons. Brain Res. 159, 385390 (1978).
37. Hafting, T., Fyhn, M., Bonnevie, T., Moser, M. B. & Moser, E. I. Hippocampus-
independent phase precession in entorhinal grid cells. Nature 453, 12481252
(2008).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank E. Chaffin for help with mouse behaviour,
J. Carmack and id Software for providing the Quake2 code, A. Shishlov for
programming advice, G. Buzsaki, J. Magee, H. Dahmen and D. Markowitz for
discussions, and C. Brody, M. Berry and E. Civillico for comments on the
manuscript. This work was supported by the NIH (1R01MH083686-01,
5R01MH060651-09), a Helen Hay Whitney Fellowship (to C.D.H.), and a
Patterson Trust Fellowship (to D.A.D.).
Author Contributions C.D.H. performed behaviour and intracellular recording
experiments with technical assistance from D.A.D. C.D.H. and D.A.D. performed
extracellular recording experiments. F.C., D.A.D. and D.W.T. designed, and C.D.H.,
F.C. and D.W.T. implemented, the virtual-reality instrumentation. F.C. performed
virtual-reality software development. C.D.H. analysed all data with strategy and
methods contributions from all authors. C.D.H. and D.W.T. wrote the paper.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. Correspondence and requests for materials should be
addressed to D.W.T. (dwtank@princeton.edu).
554
doi:10.1038/nature08499
METHODS
Virtual-reality set-up. The virtual-reality system (Fig. 2a and Supplementary
Fig. 1) was designed around an air-supported spherical treadmill (8-inch dia-
meter Styrofoam ball) for head-restrained mice that was previously used for in
vivo optical imaging at cellular resolution19. A projection-based visual display,
similar to the system used in a body-tethered rat virtual-reality system20, pre-
sented a computer rendered scene with a 270u horizontal field of view and a
vertical field from 160u to 220u relative to the mouses head. The toroidal screen
displayed the two-dimensional image from a Mitsubishi HC3000 digital light
processing projector reflected off of a 15 cm diameter round mirror and a 15 cm
diameter angular amplification mirror (AAM)21 with an angular amplification
factor of 12. The AAM was machined out of aluminium using a CNC machine
and then polished by hand. The screen was ,46 cm tall and ,80 cm in outer
diameter, and was constructed out of a semi-reflective nylon material (Rose
Brand fabrics) supported by a frame of stainless steel rods. Water rewards were
delivered by a solenoid valve (NResearch) attached to a water-feeding tube
(Popper and Sons) positioned directly in front of the animals mouth such that
the mouse could lick the meniscus. Rotations of the Styrofoam ball were
measured by an optical computer mouse (Logitech MX1000) that was posi-
tioned below the field of view in front of the animal at the point where the balls
equator intersected the animals rostral-caudal body axis. A computer running
Labview used USB-communicated signals from the optical mouse to compute
low-pass filtered ball rotational velocity around both the horizontal axis (per-
pendicular to the body axis) and vertical axis. These velocities were then output
as analogue control voltages using a D/A converter (National Instruments) and
used as input control signals to a separate computer running the virtual-reality
software.
The virtual-reality software we developed was on the basis of the open source
Quake2 game engine (id software), using code ported to Visual Studio 2008
(Microsoft). The rendering engine was modified, using cube map texturing in
OpenGL graphics38, to display a fisheye transform of the perspective of the
virtual player, such that the image reflected off the AAM and displayed on the
screen had the geometrically correct perspective for the mouse on the ball.
Software to control the water reward system, input information on ball rotation
velocity, and output voltages proportional to position and view angle within the
environment was also developed by adding A/D and D/A control to the game
engine, using National Instruments multifunction cards. Ball velocity around the
vertical axis was used as a control signal for changing view angle in the virtual
environment, whereas velocity around the horizontal axis was used as a control
signal for forward and backward movement. Other open source gaming software
(Quark, http://quark.planetquake.gamespy.com/) was used to build the virtual
linear track. To synchronize behaviour data with electrophysiology recordings,
an independent computer with a Digidata 1440A interface running Clampex
software (Molecular Devices) digitized and stored real time information about
the mouses location in the virtual environment together with solenoid (water
reward) control signals, ball rotational velocities, and electrophysiological data.
Behavioural training. Eight-totwelve-week-old male C57BL/6J mice (Jackson
Labs) were used for all experiments. C57BL/6J mice were selected because they
have good vision compared to other inbred strains39 and are a common back-
ground strain for transgenic mice. All experimental procedures were performed
in compliance with the Guide for the Care and Use of Laboratory Animals
(http://www.nap.edu/readingroom/books/labrats/) and were approved by
Princeton Universitys Institutional Animal Care and Use Committee. Before
behavioural training, titanium head plates with a large central opening (2.5 cm
wide 3 0.9 cm long 3 0.08 cm thick; central opening: 0.89 cm wide 3 0.61 cm
long) were implanted on mice and affixed to the skull using dental cement
(Metabond, Parkell). Sites for future craniotomies over the left and right dorsal
hippocampi were marked using stereotactic coordinates (2.2 mm caudal,
1.7 mm lateral to bregma). After head plate implantation, mice were placed on
a water schedule in which they received 1 ml of water per day. Body weights were
checked to ensure mice were ,80% of their pre-water-restriction weight40,41.
After at least 5 days of water scheduling, behaviour training began. In each
training session, mice were placed on the experimental apparatus with their head
fixed in place. The head was centred over the middle of the Styrofoam ball with
the headplate ,2.6 cm from the top of the ball. A lick tube to deliver water
rewards was positioned in front of the mouses mouth. The water rewards earned
during behaviour were included in the mouses daily water allotment such that a
mouse received exactly 1 ml of water per day. Each training session (one per day)
lasted 45 min with the virtual-reality system turned on for all sessions. The first
several sessions mostly involved acclimation to the apparatus and learning to run
on the spherical treadmill.
Mice were trained to perform behaviours along a virtual linear track. The
virtual track was 180 cm long and 9 cm wide, measured as the number of rotations
2009 Macmillan Publishers Limited. All rights reserved
of the Styrofoam ball to move from one end of the track to the other times the
circumference of the ball. The effective width of the track (that is, the distance the
mouse could actually move) was ,1 cm, rather than 9 cm, because in Quake2 the
player is surrounded by a bounding box that gives him a fixed width. The track
had short proximal walls with different textures for each third of the track
(060 cm: white with black dots, 61120 cm: vertical white and black stripes,
121180 cm: black with white dots). The proximal walls at the ends of the track
were green with black dots to mark the reward zones. Tall distal walls were
positioned at 60 cm (horizontal black and white stripes) and 120 cm (green with
black crosses). The floor and ceiling were black throughout the track. Green was
selected as the only non-black or white colour because electrophysiological and
behavioural data indicate that mice can detect wavelengths of ,500 nm42. Mice
received small water rewards (4 ml) for running between reward zones positioned
at opposite ends of the track (zone 1 postion: 09 cm; zone 2 position: 171
180 cm). After receiving a reward at one reward zone, the mouse then had to
run to the other reward zone to receive the next reward; two consecutive rewards
could not be obtained from a single reward site. Linear track behavioural data
(Fig. 2ce) were from mice trained only on the linear track. Some animals used for
electrophysiology recordings were trained in other virtual environments before
training in the linear track. During electrophysiology experiments, behavioural
performance was in some cases degraded due to satiation from water rewards and
because the visual display was turned off while patch pipettes were changed.
Electrophysiology. After at least five training sessions on the virtual linear track,
a small craniotomy (,0.5 mm diameter) was made over the left hippocampus
(2.2 mm caudal, 21.7 mm lateral to bregma). The dura was left intact for both
extracellular and intracellular recordings. Because behavioural performance was
degraded on the day of surgery and anaesthesia, electrophysiology recordings
were performed starting the next day. To preserve the craniotomy across days, it
was covered with silicone grease (Dow Corning) and then with a layer of silicone
elastomer (Kwik-Sil, World Precision Instruments) until the time of recording.
Extracellular recordings (Fig. 3 and Supplementary Fig. 10) were made using a
single glass electrode filled with 0.5 M NaCl (,2.5 MV pipette resistance). The
electrode was mounted vertically (perpendicular to the surface of the brain) on a
micromanipulator (Sutter MP285) that was positioned behind the mouse and
thus outside the mouses field of view. The reference electrode was positioned
outside the craniotomy in extracellular saline containing (in mM): 150 NaCl,
2.5 KCl, 10 HEPES, pH 7.4. Signals were electronically filtered between 500 Hz
and 7.5 kHz and digitized at 20 kHz. For measurements of the LFP and phase
precession, signals were electronically filtered between 2 Hz and 10 kHz; spikes
and the LFP were recorded on the same electrode. The position of the top of the
brain was noted as a large resistance increase when the electrode made contact
with the dura. As the electrode was advanced through cortex, spikes were present
in each layer. Upon entering the external capsule, a change in the recording was
noted, especially from changes in sound quality using an audio monitor. The
CA1 cell body layer was ,200 mm beneath the external capsule and was char-
acterized by strongly theta-modulated spiking. Putative pyramidal neurons were
found as units firing complex bursts separated at theta frequencies22. In all
animals CA1 recordings were made at a depth of ,1.1 mm; cells were found
reproducibly at this depth across several electrodes and days. To confirm that
this recording position was in the CA1 cell body layer, in a separate experiment
we electroporated Alexa 488 dextran (5% (w/v) in extracellular saline, 4 mA
pulses, 25 ms pulse duration, 600 pulses at 2 Hz) using a recording pipette and
found bright labelling of CA1 cell bodies in histology sections. At this depth, we
searched for well-isolated units of large amplitude and recorded their activity for
,1530 min, which was enough time for the mouse to sample thoroughly the
virtual linear track. Spikes were sorted offline using a threshold analysis. At most
one unit was isolated from a single recording. To check the quality of unit
isolation, we overlaid all spike waveforms to make sure that they matched by
visual inspection (Fig. 3a), and we plotted ISI distributions to make sure no
spikes fell within the refractory period (,12 ms; Fig. 3b). Extracellular record-
ings were made from a total of 8 mice. In some cases a second craniotomy was
made over the right hemisphere to extend the number of recording sessions from
a single animal. Recordings were made for up to 4 days from the same cranio-
tomy. Place cells were found in every mouse tested.
Whole-cell recordings were made using the same experimental set-up as for
extracellular recordings. Pipettes (,57 MV) were filled with internal solution
containing (in mM): 135 K-gluconate, 10 HEPES, 10 Na2-phosphocreatine,
4 KCl, 4 MgATP, 0.3 Na3GTP (pH 7.25 with KOH, 285 mOsm). We pulled pipettes
with a long taper (,100 mm diameter at 1 mm from the tip) to minimize damage to
the overlying cortical tissue and to reduce compression of the tissue while advancing
the pipette. The pipette was mounted vertically on a standard micromanipulator
(Sutter MP285) positioned outside the mouses field of view; special methods to
anchor the pipette17,18 were not necessary. We used standard blind patch methods to
obtain whole-cell recordings30. In brief, we applied ,250 mbar of positive pressure
555
doi:10.1038/nature08499
while moving the pipette through the cortex and reduced the pressure to ,15 mbar
while searching for cells in the hippocampus. The CA1 cell body layer was located
using a depth coordinate obtained from extracellular recordings in the same animal.
We attempted to form a seal when we observed large (,50%), reproducible
increases in pipette resistance. Recordings were made in current clamp mode with
no holding current. Membrane potential values were corrected for liquid junction
potentials. We obtained recordings lasting longer than 3 min from 46 cells from 15
mice, with a peak success rate of approximately one such recording in every five
attempts. Recordings were abandoned when we observed large increases in resting
potential or large decreases in spike amplitude. We obtained whole-cell recordings
for up to 3 days from the same craniotomy. Whole-cell recordings from place cells
lasted on average 7.1 6 2.8 min and had series resistances of 48 6 16 MV, input
resistances of 98 6 23 MV, and resting membrane potentials of 267 6 4 mV. Firing
rates from place cells recorded intracellularly tended to be higher than those
recorded extracellularly, both inside and outside the place field, potentially because
a fraction of spikes measured extracellularly were lost during spike sorting. All other
place-cell parameters (for example, field size, ISI distribution) were similar between
cells recorded extracellularly and intracellularly. Although we cannot exclude
possible effects of dialysis of the cell during whole-cell measurements, such effects
are probably small owing to short recording durations and relatively high series
resistances. In experiments to assess the connection between ramps of depolariza-
tion and changes in theta oscillation amplitude (Supplementary Fig. 7), ramps of
current (4 s duration, 11.5 nA at the peak) were injected at the soma after 11 s
without current injection. Only trials in which the animal was running at greater
than 10 cm s21 were analysed.
To perform whole-cell and extracellular recordings simultaneously (Fig. 5f, g
and Supplementary Fig. 11), we mounted two glass electrodes on separate micro-
manipulators, which were both positioned behind the mouse, and advanced the
electrodes independently through separate craniotomies on the same hemi-
sphere. Whole-cell recordings were made at the standard coordinates as
described earlier. The extracellular recording electrode (filled with 0.5 M
NaCl) was mounted at a 40u angle relative to the midline and a 45u angle with
respect to the vertical. The craniotomy for the extracellular electrode was posi-
tioned caudal and either lateral for the left hemisphere or medial for the right
hemisphere relative to the craniotomy for intracellular recordings. The tip of the
extracellular electrode was positioned at a depth of ,1.2 mm and ,250 mm
caudal to the centre of the craniotomy for whole-cell recordings. The extracel-
lular recordings were filtered between 1 Hz and 10 kHz.
From our 46 hippocampal whole-cell recordings, 41 cells were spontaneously
active (overall firing rate .0.05 Hz). We classified 7 of the spontaneously active
cells as interneurons on the basis of high overall firing rates (.5 Hz) and the
absence of complex spike bursts. The remaining 34 spontaneously active cells
were categorized as pyramidal neurons. To estimate the fraction of sponta-
neously active pyramidal neurons that had a place field along the virtual linear
track, we considered only those recordings during which the animal visited each
of the 80 spatial segments at least 3 times, with visits separated by at least 2 s in
time. We estimate that 36% of spontaneously active pyramidal neurons had
place fields in the virtual environment (8 out of 22 cells). During whole-cell
place-cell recordings, the place field was visited 20 6 8 times on average, with
visits separated by at least 2 s in time.
Data analysis. To create firing rate maps (Figs 3c, 4b and Supplementary Fig. 2a),
we divided the virtual linear track into 80 bins (2.25 cm each) and calculated the
firing rate as the total number of spikes in a bin divided by the total amount of
time spent in a bin. The maps were smoothed using a five point Gaussian window
with a standard deviation of one. Periods in which the mouse ran slower than
5 cm s21, averaged over a 2-s window, were removed from the analysis. To
identify place fields, we found groups of adjacent bins with firing rates greater
than 0.25 times the rate in the peak bin. We selected only those fields that were
larger than 8 bins (18 cm) in length, had mean in-field firing rates of greater than
1.5 Hz, and had mean in-field firing rates more than 3 times larger than the mean
out-of-field firing rate. To verify that place fields were a statistically significant
increase in firing rate, we performed a bootstrap shuffle test. We shuffled the
times of spikes in 10 s segments, calculated new firing rate maps using the
unshuffled bin dwell times, and then checked for a place field using the above-
stated criteria. This procedure was repeated 1,000 times per place-cell recording.
We called a cell a place cell only if a place field was found in the unshuffled data
and in fewer than 2% of the shuffled tests (P , 0.02). Across the place cells we
recorded, we found place fields at all positions along the linear track. Spikes were
separated into leftward and rightward runs based on the head direction of the
animal with regard to the virtual track (measured as the players view angle in
Quake2). A directionality index for firing based on firing rates in leftward and
rightward directions (FRleft, FRright) was defined as jFRleft 2 FRrightj/
(FRlef t1 FRright). A directionality index of 0 indicates identical firing in both
directions, whereas an index of 1 indicates firing in one direction only.
2009 Macmillan Publishers Limited. All rights reserved
To quantify the average baseline membrane potential as a function of position
along the virtual track (Fig. 4c and Supplementary Fig. 2b), we first removed the
contribution of action potentials by discarding all time points between 4 ms before
and 7 ms after an action potentials peak. We then created a map of membrane
potential values by grouping the membrane potentials into 80 spatial bins along the
track and calculating an average membrane potential for each bin. The membrane
potential map was smoothed using a five point Gaussian window with a standard
deviation of one. Large, long-lasting depolarizations of the baseline membrane
potential occurred infrequently outside the place field (2.3 6 1.3 events per record-
ing for depolarizations of greater than 2 mV lasting longer than 0.5 s).
To analyse ramps of depolarization during complete runs through the place
field (that is, the mouse did not turn around inside the place field), we calculated
changes in membrane potential (DV) after action potentials were removed (Fig. 4e
and Supplementary Figs 5 and 6b). We defined the baseline membrane potential
as the mean membrane potential from 500 ms to 1 s before the mouse entered the
place field and subtracted this baseline from all membrane potential values during
the run to get DV values. The peak DV during a run through the place field
(Supplementary Fig. 6b) was calculated after smoothing the DV values over a
sliding window of length 50 ms such that the peak value was representative of the
ramp-like depolarization rather than brief, high frequency depolarizations. To
analyse the symmetry of firing rates and ramps of depolarization in place fields
(Fig. 4e), we considered runs starting half the place fields width before the place
field and ending half the place fields width beyond the place field (total length was
twice the length of the place field). Because place fields varied in size between cells
and runs through the field differed in duration, we divided each run into 60
equally sized spatial bins and calculated the mean firing rates and DV for each
1/60th of the run (that is, 15 spatial bins before the place field, 30 bins in the field,
and 15 bins after the field). All runs through the field were plotted with increasing
track position values so that runs in opposite directions could be combined.
To analyse subthreshold theta oscillations (Fig. 5 and Supplementary Figs 2c, 7
and 9), spikes were removed over a window of 3 ms preceding and 5 ms after the
peak and were replaced using linear interpolation. The resulting trace was then
band-pass filtered between 610 Hz (a peak in the ISI distribution) using a linear
phase finite impulse response (FIR) filter with a Hamming window of width 1 s
(Matlab function fir1). To create a map of theta power along the virtual track,
power was calculated as the mean of the squared amplitude for a sliding window
of length 1 s for the entire recording, with the centre of the window providing the
position along the track. The power values were grouped into 80 spatial bins
along the track and smoothed using a five point Gaussian window with a stand-
ard deviation of one. Power spectra for epochs in and out of the place field were
obtained using multi-taper spectral analysis methods (Chronux toolbox, http://
chronux.org; Supplementary Fig. 8). Running speed, which can influence theta
oscillations43, was similar inside and outside the place field (in-field
speed 5 49 6 6 cm s21, out-of-field speed 5 47 6 4 cm s21, P . 0.7). LFP theta
oscillations were present when the animal was running (Fig. 3d and
Supplementary Fig. 10) and absent when the animal was resting.
To analyse phase precession (Figs 3d, e and 5dg), we considered complete runs
through the place field (that is, the animal did not turn around in the field) that had
at least 5 spikes and in which the animal ran faster than 10 cm s21. Only cells with
more than 40 total spikes in the place field were included. Extracellular or intra-
cellular (after removing spikes) voltage traces were band-pass filtered between
610 Hz using an FIR filter. To assign a phase to a spike occurring at time t, we
identified, in the filtered trace, the times of the peaks immediately preceding and
following the spike (t1 and t2, respectively) and calculated the phase as 360(t 2 t1)/
(t2 2 t1). We circularly shifted the phase of the spikes in 1u steps from 0u to 360u,
continuing across the 360u border, and fit a linear regression line to the phase
versus position plot at each rotation4,37. We found the rotation with the best fit,
such that the sum of squared errors between the fit line and data was minimized,
and used the correlation coefficient between phase and position at this rotation as a
measure of phase precession. All phase values are from the raw data in relation to
theta phase, with 0u indicating the peak. To measure the precession of intracellular
theta relative to extracellular theta (Fig. 5g), we first identified the times of the
peaks of the filtered (610 Hz) membrane potential trace during runs through the
place field. For each peak of intracellular theta, we found the corresponding phase
of the simultaneously recorded extracellular theta and the position of the mouse
along the virtual track and followed the same procedures that were used for the
analysis of the phase precession of spike times. To quantify the change in times
between intracellular theta peaks and LFP theta peaks during place field traversals
(Supplementary Fig. 11a), we found the time difference between the first LFP theta
peak and the first intracellular theta peak in the place field, the time difference
between the second LFP theta peak and the second intracellular theta peak, and so
on. To quantify frequency differences between intracellular and LFP theta
(Supplementary Fig. 11b), we calculated the ratio of the period of the first intra-
cellular oscillation to the period of the first LFP oscillation, the ratio of the period of
556
doi:10.1038/nature08499
the second intracellular oscillation to the period the second LFP oscillation, and so
on using peaks from the filtered (610 Hz) traces. We determined these ratios for
complete runs through the place field and 3 s long epochs outside the place field.
Data are presented as mean 6 s.d. unless noted otherwise. P values are from
two-tailed t-tests unless stated otherwise. Correlation coefficients (C) are from
Pearsons correlations.
38. Greene, N. Environment mapping and other applications of world projections. IEEE
Comput. Graph. Appl. 6, 2129 (1986).
2009 Macmillan Publishers Limited. All rights reserved
39. Wong, A. A. & Brown, R. E. Visual detection, pattern discrimination and visual
acuity in 14 strains of mice. Genes Brain Behav. 5, 389403 (2006).
40. Rinberg, D., Koulakov, A. & Gelperin, A. Sparse odor coding in awake behaving
mice. J. Neurosci. 26, 88578865 (2006).
41. Huber, D. et al. Sparse optical microstimulation in barrel cortex drives learned
behaviour in freely moving mice. Nature 451, 6164 (2008).
42. Jacobs, G. H., Neitz, J. & Deegan, J. F. II. Retinal receptors in rodents maximally
sensitive to ultraviolet light. Nature 353, 655656 (1991).
43. Buzsaki, G., Leung, L. W. & Vanderwolf, C. H. Cellular bases of hippocampal EEG in
the behaving rat. Brain Res. 287, 139171 (1983).
557


558
J. Physiol. (1965), 178, pp. 477-504 477
With 11 text-ftgureas
Printed in Great Britain

THE MECHANISM OF DIRECTIONALLY SELECTIVE UNITS

IN RABBIT'S RETINA
BY H. B. BARLOW* AND W. R. LEVICKt
From the Physiological Laboratory, University of Cambridge
and the Neurosensory Laboratory, School of Optometry,
University of California, Berkeley 4, U.S.A.
(Received 29 September 1964)
Directionally selective single units have recently been found in the
cerebral cortex of cats (Hubel, 1959; Hubel & Wiesel, 1959, 1962), the
optic tectum of frogs and pigeons (Lettvin, Maturana, McCulloch & Pitts,
1959; Maturana & Frenk, 1963), and the retinae of rabbits (Barlow & Hill,
1963; Barlow, Hill & Levick, 1964). The term 'directionally selective'
means that a unit gives a vigorous discharge of impulses when a stimulus
object is moved through its receptive field in one direction (called the
preferred direction), whereas motion in the reverse direction (called null)
evokes little or no response. The preferred direction differs in different
units, and the activity of a set of such units signals the direction of move-
ment of objects in the visual field.
In the rabbit the preferred and null directions cannot be predicted from
a map of the receptive field showing the regions yielding on or off-responses
to stationary spots. Furthermore, the preferred direction is unchanged by
changing the stimulus; in particular, reversing the contrast of a spot or a
black-white border does not reverse the preferred direction. Hubel &
Wiesel (1962) thought that the directional selectivity of the cat's cortical
neurons could be explained by the asymmetrical arrangement of on and off
zones in the receptive field, and the simple interaction of effects summated
over these zones, but the foregoing results rule out this explanation, at
least in the rabbit's retina (Barlow & Hill, 1963).
In the present paper we go on from this point to describe experiments
which show, first, that directional selectivity is not due to optical aberra-
tions of some kind and, secondly, that it is not a simple matter of the
latency of response varying systematically across the receptive field. After
these negative results we describe experiments upon the organization of
directional selectivity within the receptive field, and upon its mechanism.
These lead us to the conclusion that the ganglion cells responding to a
* Present address: Neurosensory Laboratory, School of Optometry, University of
California, Berkeley 4.
t C. J. Martin Travelling Fellow on leave of absence from the University of Sydney.

559
478 H. B. BARLOW AND W. R. LEVICK
particular direction of motion are fed by a subset of bipolar cells that
respond to the corresponding sequence of excitation of two neighbouring
retinal regions with which they connect. Furthermore there is evidence
that this sequence-discrimination is brought about by a laterally connect-
ing inhibitory element from one of these regions, and this seems a likely
function for the horizontal cells to perform. At this stage identification of
the elements concerned is obviously tentative, but we were pleasantly
surprised to find well known histological structures already at hand to fill
the roles that the functional organization seemed to require.
All the experiments described here were performed upon on-off
directionally selective units. We have reason to believe that the mechanism
is different for the rare, on-type, directionally selective units and also for
the centrifugal and centripetal motion sensitivity of the ordinary con-
centric type of units.
METHODS
Action potentials were recorded from the unopened eyes of rabbits. As described else-
where (Barlow et al. 1964), it proved most effective to use fine tungsten electrodes, decere-
brate or lightly anaesthetized (urethane and chloralose mixture) animals, and immobilization
by continuous infusion of gallamine triethiodide (Flaxedil). Periodically the animal was
allowed to recover from paralysis by using an infusion fluid without relaxant. One could
thus ensure that the level of anaesthesia was neither too deep nor too light at the rate of
anaesthetic infusion employed.
When a good on-off, directionally selective unit was isolated the first step was always to
map out its receptive field on a plotting board 57 cm from the eye. A stationary spot was
turned on and off; in most cases this was I' diameter at an intensity of 12 cd/M2 and was
superimposed on a background of 0-6 cd/iM2. Directional selectivity was tested for and null
and preferred directions determined. Various techniques were used to provide the temporal
and spatial patterns of light stimuli. For the two-spot experiment we initially used two
glow modulator tubes controlled by pulse generators, but we later found that black and
white cards moved behind apertures in grey paper provided a more flexible means of
delivering the required stimuli. This method also has disadvantages (see p. 486) which we
finally overcame by illuminating the apertures from behind with thin Perspex light pipes lit
by low-current torch bulbs turned on and off manually. The changes of lulminance occurring
within the receptive field of the unit were monitored by a photocell whose output was
recorded with the action potentials.
RESIULTS
The results are presented in four sections. These are: (1) controls and
negative results which rule out various preliminary hypotheses on the
mechanism; (2) experiments which lead to the conclusion that the direc-
tional selectivity of the ganglion cell results from the sequence-discrimina-
ting activity of subunits-probably bipolar cells; (3) observations and
experiments which show that sequence-discrimination is achieved by an
inhibitory mechanism that prevents responses to sequences in the null
direction; (4) observations showing that inhibition also occurs with stimu-
lation of the surround of the receptive field.

560
 MECHANISM OF DIRECTIONAL SELECTIVITY 479
Controls and negative results
Optical controls. It might be thought that the unequal responses to
motion in the null and preferred directions were the result of a peculiarity
of the light distribution in the retinal image caused by aberrations of the
optical system. Although none of the schemes suggested to us, and none
we could imagine, seemed at all promising, we considered this possibility,
but were forced to abandon it at an early stage. The most direct disproof is
given by the observation that two units recorded simultaneously, or
within a short period of time, in the same retinal region can have their
preferred axes opposite or at right angles to each other. An optical
explanation of the phenomenon requires that there be some asymmetry in
the light distribution to cause the asymmetry in the responses, and two
different asymmetrical light distributions cannot co-exist in the same
region.
A second disproof is provided by observing how little the phenomenon is
affected by deliberately introduced optical aberrations. Figure 1 shows that
clear-cut directional selectivity persists with spherical supplementary lenses
causing more than 10 dioptres ofrefractive error in either direction. This was
for the normal pupil diameter of our preparation-6 mm or more. Further-
more, reducing the aperture of the rabbit's optical system to 3 or 125 mm
with an artificial pupil must greatly improve retinal image quality (since
diffraction is most unlikely to be a limiting factor); yet we found that such
stopping down merely extended the range of refractive error over which
the directionally selective property was shown. It did not even improve
acuity as judged by the finest grating giving any response to movement.
The best spherical correction was determined in each preparation,
judging this opthalmoscopically and sometimes retinoscopically. In all the
preparations the cornea and media were free from clouding, except
occasionally at the very end of a 2-day experiment. Optical effects from
the shank of the electrode can be excluded, for directional selectivity is
often observed when recording from a nerve fibre. In such cases, no
portion of the electrode intercepts light reaching the retina from the
receptive field projection.
Finally, we should point out that for most of the tests used here the
human eye's performance is superior by a factor of 10. The demands made
of the rabbit's optical system are thus not at aLl severe, but of course the
blur of the retinal image should be taken into account when interpreting
quantitatively the results of our later experiments.
Latencies. The first idea about mechanism was that the latency of the
response might be shorter at successive positions along a path through the
receptive field traced in the preferred direction. It was thought that when

561
480 H. B. BARLOW AND W. R. LEVICK
the image of an object moved in this direction the excitation from succes-
sive positions would arrive synchronously at the ganglion cell, and thus
might be more effective than when movement was in the null direction and
it was dispersed in time. An alternative scheme can be devised in which
the temporally dispersed sequence is more effective because it avoids

Unit 4-40
o0 I 0

45' I
0\~~~~~~~~~

30' -//\ 4

0-- -----0- -0/

15'_

-12 -6 0 +6 +12 +18 +24 +30

Power of correcting lens (Dioptres)
Fig. 1. The effect of refractive error and pupil diameter upon the acuity for a
moving grating. Pupil dimensions were 6 x 8 mm (0) and 3 mm diameter (0);
luminance of the white bars was 2-5 cd/M2 for the large, 25 cd/M2 for the small
pupil. The points show the finest grating for which there was a distinctly greater
response to movement in the preferred direction. Paradoxical responses, greater
in the null direction, were not observed. A small discharge with movement was
detectable in most cases for the next finer grating (71 % of the period), but this was
not obviously different in the two directions. Notice that directional selectivity
persists over a large range of refractive error, and that the finest resolvable grating
is not affected by pupil diameter: optical aberration is not likely to be the cause of
directional selectivity, and probably does not limit the optimum acuity of this
preparation.
refractoriness. Thomson's (1953) work on the rabbit's retina, together with
our own findings on the different latencies of centre and surround in con-
centric units (Barlow et al. 1964) lent some plausibility to the suggestion
that the latencies might vary with position, but the result shown in Fig. 2
is entirely negative. In this and other experiments 'on' and 'off' latencies
showed no sign of changing systematically with position in the receptive
field.

562
 MECHANISM OF DIRECTIONAL SELECTIVITY 481
Unsuccessful two-spot experiments. One sees from the records of Fig. 2
that excitation of a point in the field causes a response at on and off: yet
if a spot of light is moved through the field in the null direction no response
occurs, even though each point crossed by the spot must have received
an ' on' followed by an 'off' stimulus as the spot passed over it. The obvious
next step in the analysis seemed to be to stimulate at two points and see
how the response changed when the order and temporal interval between
the stimuli was varied. We hoped to decide whether the response to
Background 7 cd/M2 Position 45S 150A in visual field
Spot 50 cd/m2 aII- I -1I

Preferred
s
+ 0
;* ~~~~~c

Nul
X 0s 0i I I

Unit 5-26
Iw-'-- ---|i 0-25 sec
Fig. 2. Latencies of response at different positions. The receptive field map is
shown at the left; positions yielding on and off responses to a stationary spot are
marked + and -; positions yielding both are marked T if off was greater, if
on was greater; no responses were obtained on or outside the ring of O 's. Records
a, c, e, are the responses to a light turned first off, then on, at three successive posi-
tions along a line through the receptive field in the preferred-null axis, as shown to
the left (negativity is upwards). Records b, d, f, are from a photomultiplier
observing the receptive field (decreased light is downward), and the numbers show
the latency of response in msec (one impulse was ignored in a). There is no signifi-
cant trend in latency as one moves across the receptive field.

motion in the preferred direction was greater than the sum of responses to
excitation of separate points along the path, or whether the unequal
responses to motion in the two directions resulted from inhibition occurring
when the motion was in the null direction.
The question seemed clear-cut, but the results were not. In the first
place it was not nearly as easy to obtain unequal responses for the two
sequences as it was to obtain unequal responses with real moving objects.
Secondly, when we did get evidence of sequence-dependent responses, the
result seemed highly variable and we were unable to decide whether
summation or inhibition or both were occurring. This failure forced us to
31 Physiol. 178

563
482 H. B. BARLOW AND W. B. LEVICK
realize that we did not know whereabouts in the receptive field we should
put the spots, nor how far apart they should be. There was in fact a prior
question to answer before the two-spot type of experiment could be
performed and interpreted. The question, put in a form that avoids
implications as to mechanism, is this: does the ganglion cell respond
selectively to one direction of motion over all parts of its receptive field, or
is there some critical zone or line which must be crossed? The following
observations show that there is no such line and the directionally selective
property is distributed over the receptive field.
Background 7 cd/Mr2 Position 450S SOP
Spot 40 cd/M2 T Spotsge in visual field
Sposi.zO

[ 10 0MIWIi _tin;t
M&Uui.
ii02

o _,,____________
50[>
r~~ ~~~~~~~~ -e

fLg.3.Rpo:to _0-5 sec

i ;SlUnit 421
Fig. 3. Responses to motion along three different paths through the receptive
field. The map in the centre shows the field and the paths through it; symbols as in
Fig. 2. The records of the responses to traverses in the null direction are to the left,
those for the preferred direction to the right. The lower trace of each pair is from a
potentiometer and shows the position of the spot as it moved through the field
(calibration at left). Top, middle and bottom parts of the receptive field all show
the same directional selectivity.

Sequence-discrimination by subunits
Distribution of directional selectivity. Figure 3 shows the responses
obtained when a spot of light was moved across the receptive field and
back along three parallel lines. These were separated by more than the
breadth of the spot, and therefore different receptors were covered by the
geometric image of the moving spot in each case. It will be seen that the
selectivity clearly exists along all these three pathways.
Figure 4 shows typical responses obtained when the spot was moved
several times from one position to the next and back, as marked. It is
clearly not necessary for the spot to cross any definite line in order to
obtain different responses for the two directions of motion. If the experi-
ment is repeated using a black spot the same result is obtained; direction

564
 c_h | m-_ ~ n
MECHANISM OP DIRECTIONAL SELECTIVITY 483
of motion, independent of contrast, can be picked out in a large number of
widely separated regions of the receptive field. However, there is an
interesting exception to the rule that aU regions of the receptive field have
the capacity to distinguish between null and preferred sequences of
excitation of the receptors they contain. There is a zone adjacent to the
edge of the field that is first crossed when motion is in the preferred

Preferred s 1} af Background 7 cd/M2

Null At (@t | ml) ~~~~Spot 40 cd/M2
Null 1.% j) Bakg
Spot size 0 Unit 4-21 O-5 sec

b ~b
____b fgf__gf k k

b b
c g h g m

c d c . mm

~ml , 11...__-n.I .1;

dS- e rs zin _ o -

Fig. 4. Back and forth motion in different parts of the receptive field. The edge of
the receptive field is mapped at the top, and the positions a, b, c, . . . o, within it are
indicated. The spot was moved back and forth several times between a and b, then
between b and c, and so on. The records are samples of these back and forth
motions. The lower trace of each pair shows the position of the spot in the field:
downward movement of the trace corresponds to movement of the spot in the
preferred direction. Marked asymmetry of response for the opposite directions
holds in most positions in the field. Its absence in the top row of records is
expected in the inhibitory scheme (see Fig. 7 and p. 490).

direction where this capacity is lacking: motion in either direction causes a

response. This is shown in the top line of records in Fig. 4, and a possible
explanation for the effect is given later (see p. 490).
Smallest region giving directional selectivity. Figure 4 shows clear
directional selectivity when a spot of light is moved to and fro through 1l
in a receptive field whose total diameter is about 3. What then is the
31-2

565
484 H. B. BARLOW AND W. R. LEVICK
smallest distance over which responses to motion in the two directions
differ?
To answer this question we moved a strip of white card with a section
painted black behind an aperture of variable width in a sheet of grey
paper. The aperture breadth was 20 mm (subtending 20) and the width
measured in the direction in which the card moved varied from 1 mm (6')
to 11 mm (10 06'). When the card was moved there was no change until the
border of the black section entered the aperture, it then moved through a
variable distance before it disappeared behind the other edge, after which
there was again no further change and the aperture was wholly black. The
sequence described above was called 'off' stimulation since with black
trailing the receptors were successively exposed to a reduction of illumina-
tion. On the other hand, when the border of the white portion appeared in
TABLE 1. Single slit experiment
A black-white border was moved through the receptive field, but the view of the motion
was restricted by a fixed rectangular slit in a grey card placed immediately in front of the
moving border. The width of this slit, measured in the direction of motion, was varied.
Movement of the border causing the slit to change from white to black was called 'off'
stimulation, movement causing it to fill with white was called 'on' stimulation. Each of
these stimuli was applied with the edge moving in both null and preferred directions. We
graded the unit's response from 0 to 6 after listening to several repetitions of these stimuli on
the loudspeaker. This unit could distinguish the preferred from the null direction with slit
widths down to about 17' for both 'on' and 'off' stimuli.
'Off' stimulus 'On' stimulus
Slit width ,A_ A __,
Preferred Null Preferred Null
1 36' 6 1 4 0
106' 6 2 4 1
48' 5 2 4 0
34' 4 2 3 1
24' 4 2 3 1
17' 3 2 3 2
12' 3 3 3 2
8' 2 2 2 2

the aperture and moved along it they were successively exposed to 'on'
stimulation. Clearly 'off' and 'on' stimulation could be applied in any
direction and at any velocity, but we confined our attention to the pre-
ferred and null directions, and moved the cards by hand at velocities
chosen to give optimum discrimination between the two directions-in
most cases about 5/sec. The variable studied was the width of the aper-
ture, and the feature of the response attended to was the existence of a
clear-cut difference between responses to stimulation in the preferred and
null directions.
The amplitude of the response was graded subjectively from 0 to 6, and
Table 1 gives a typical result. The subjective grading was a crude but con-

566
 MECHANISM OF DIRECTIONAL SELECTIVITY 485
venient way of quantification: records of typical responses are shown in
Fig. 5. Altogether 10 units have been studied in this way, and the threshold
aperture for directional discrimination varied from 6' to 24'.
The straightforward interpretation of this result is that the complete
mechanism for directional selectivity is contained within a subunit of the
receptive field extending not much more than I' in the preferred-null axis.
Since the result does not depend critically upon the position of the slit
within the receptive field, it looks, again taking the straightforward view, as
Position 450S 50P in visual field
Background
7 cd/r2 /
I1111III

I*Il 1 11
111 I I I I[110111 1I I .

17't ..

|11.111 v
Unit 4-21
0o5 sec
Fig. 5. Responses to motion of a black edge across slits of various widths
(breadth was 20 in all cases). Results are shown for four slit widths, from 10 06' to
8' measured in the preferred-null axis. The records to the left were obtained when
the black border advanced across the slit in the preferred direction, those to the
right when it moved in the null direction. The lower trace of each pair is the
response of a photomultiplier aimed at the field. Notice that the differences
between preferred and null motions are obvious in the top two records, when the
distances through which the edge could be seen moving were 10 06' and 34'; the
difference is only just detectable at 17' and has vanished at 8'. The directionally
selective mechanism appears to be contained within a retinal region subtending
I' to f, whereas the whole receptive field subtended 4j0.
if the sequence-discriminating mechanism must be reduplicated perhaps a
dozen or more times to cover the whole receptive field. Would there be any
escape from this conclusion if the image was of very poor quality so that,
even for the small slits, it was diffused over a large part of the receptive
field?
If the optics are poor, then the time course of light intensity changes at
the two edges of the receptive field will be slightly different for the two

567
486 H. B. BARLOW AND W. R. LEVICK
directions of motion, and one can set up somewhat elaborate schemes in
which these differences form the basis of directional selectivity. The
schemes have to be made even more complicated to account for the fact
that black or white edges advancing through the slit show the same
directional preference. Now it will be shown in the next section that the
interaction responsible for directional selectivity occurs better at small
separations of the stimuli than at large separations. We can think of no
way in which neighbouring retinal regions would show more interaction
than widely separated ones unless the subunit responsible for directional
selectivity is itself a small compact one. In this way the experiment about
to be described is a useful control confirming the conclusion reached from
the single slit experiments.
Discrimination of sequence. The fact that movements within a region of
the receptive field subtending less than jI' are sufficient to give directionally
selective responses suggests a possible explanation for our failure to get
clear-cut sequence-dependent results when we first attempted to excite with
a pair of static stimuli at various temporal intervals. In these experi-
ments spots closer together than about 10 had never been tried and we
therefore designed new apparatus so that two strips of light each sub-
tending 0.10 x 20 could be brought to within 001' of each other and turned
on and off in either sequence (see Methods).
At small separations the experiment gave definite evidence of sequence-
dependence, as shown in Fig. 6. The response is much greater in the
sequence corresponding to movement in the preferred direction than for
null sequences, and this is true for both 'on' and 'off' stimuli. However,
these differences become less when the separation of the slits is increased
to over 10 even though both slits remain inside the receptive field.
We have done similar two-slit experiments in which moving cards were
used to provide 'on' or 'off' excitation at the two slits. These also indicated
that sequence discrimination occurs at small separations but is reduced at
large separations as shown in Table 2. It was not easy to judge the res-
ponses accurately but the difference between null and preferred sequences
faded out for separations greater than i' and other units showed a
similar reduction for large separations. There is, however, a defect in
these experiments which was not always fully controlled. In order
to provide a vigorous stimulus with each slit it was made 0.10 wide.
This is below the threshold for directional selectivity in most prepara-
tions, hence in these cases each slit by itself was equally effective for
null and preferred sequences. However, this was not always true, and
some of our results lack the necessary controls. In addition it might
be held that there are effects of movement which are subthreshold for
each slit by itself but become suprathreshold with the pair, and that it

568
 MECHANISM OF DIRECTIONAL SELECTIVITY 487
is the summation of these subliminal effects that produces the asymmetry
for the two sequences. A control observation (made on a unit in which
directional selectivity occurred across the 0-10 slit) meets this criticism and
is worth reporting because it reinforces the conclusion that sequence as
such is effective.
Unit 3-41 Preferred Null
t I 1 I l I

A B
On =

I-! Off
I!I 1 A_
17' A B B A
0-5 sec
w .

r - I I- I I I
A B
On

w
IIw I I
-%.-
Off
1006' A B B A
Fig. 6. Responses to different temporal sequences of two static stimuli. On the
left the positions of the pair of stimuli are shown within the outline of the receptive
field. Records for the small separation are shown above, those for the large
separation below. Within each half, records for on are above those for off. The
lower trace of each pair is the photomultiplier output: increasing light moves the
trace upwards, and slit A was arranged to give the bigger step in every case, even
though it was not brighter. Preferred sequences are on the left, null on the right.
Notice that the preferred sequence yields more spikes than the null at the small
spatial separation, but this difference ceases to be clearly visible when the separa-
tion of the slits is increased.

In this experiment the card behind the pair of slits had white and black
regions arranged so that when the card was moved in one direction the
sequence of lightening or darkening at the two slits corresponded to
motion in the opposite direction. Under these conditions each slit by
itself gave a greater response for movement of the card in the preferred-
direction. However, when both slits were used there was a greater response
for movement of the card in the null direction: that is, the sequence of
activation of the slits was overriding the effects of motion in the opposite
direction within each slit.

569
488 H. B. BARLOW AND W. R. LEVICK
It must be pointed out that the time interval between the two stimuli is
an important variable that we have not yet studied systematically. In
timing these stimuli manually we have varied the interval over as wide a
range as possible, but in spite of this we never found as strong an inter-
action at large as at small spatial separations of the slits. Some further
results with this type of experiment are given in Table 3 (see later) and it is
hoped that a more systematic exploration will be presented in the future.
TABLE 2. Two-slit experiment
Same unit as in Table 1. The stimulus was again a black-white border moved through the
receptive field in the null or the preferred direction, but the view of the motion was now
restricted by a pair of narrow slits in a grey card placed immediately in front of the moving
border. Each slit subtended only 6', and responses to preferred and null directions were
indistinguishable for each slit by itself. However, when the pair of slits was darkened or
lightened in sequence, the strength of the response was found to depend upon the order in
which the slits changed. The separation of the two slits, measured in the direction of motion,
was varied and the unit's response graded as in Table 1. The effect of the order of stimula-
tion was greatest at small separations, but null and preferred sequences were still dis-
tinguishable up to 24' separation.
'Off' stimulus 'On' stimulus
Slit A,AK
separation Preferred Null Preferred Null
10 36' 2 2 2 2
1006' 3 3 3 3
48' 3 3 2 2
34' 3 3 2 2
24' 4 3 3 2
17' 3 2 3 1
12' 4 2 3 1
8' 3 1 3 1
6' 3 1 3 2

We think the results already given are sufficient to establish that direc-
tional selectivity may be based upon the discrimination of the sequence of
excitation of only a pair of regions. Even though the image of a moving
object falls on a long succession of receptors in a continuous succession of
time intervals it is unnecessary to postulate the interaction of more than
two regions to account for the directionally selective property.
Responses to gratings. The results so far reported suggest that sequence-
discrimination is performed by subunits of the receptive field. Figure 1
shows that a directionally selective unit can discriminate the direction of
motion of the bars of a grating subtending 15' (period 30'). It is hard to
see how this discrimination could be performed if the bars of the grating
were small compared to the size over which each subunit integrates or
averages the light, and in fact the resolvable grating size fits, to a first
approximation, the size of subunit suggested by the preceding tests.
Another result that may also fall into line is the shape of the curves found
when determining threshold as a function of area (Fig. 5 of Barlow et at.

570
 MECHANISM OF DIRECTIONAL SELECTIVITY 489
1964; see also Barlow, 1953). Complete summation (that is, threshold
oc 1/area) does not hold out to the full diameter of the receptive field in the
directionally selective units of the rabbit, and it is tempting to identify
the limit to which it does hold (approx. 20') as the integrating area of the
subunits.
One negative result in the grating tests is worth comment. Hassenstein
(1951) found paradoxical optokinetic movement responses in beetles:
when a grating of period slightly greater than the angular separation of the
axes of the ommatidia was moved in one direction, the beetles responded as
for the opposite direction of movement. Nothing of this sort was seen in
the present tests: when motion of a grating caused a response, this was
never greater in the null direction than in the preferred direction. This is
not too surprising, for the occurrence of paradoxical movement responses
in the beetle must depend upon the regular spacing of its ommatidia.
Mechanism of sequence-discrimination
The foregoing experiments show that the directional selectivity of
ganglion cells is based upon sequence-discrimination within subunits of
their receptive fields, but they tell us nothing about the mechanism where-
by a pair of stimuli causes a greater discharge in one sequence than in
reverse. Figure 7 shows two schemes in which the preferred sequence,
corresponding to motion in the preferred direction, elicits a greater
response than the null sequence. These are intended to exemplify two
broad alternatives, not to make exact specifications.
The left-hand scheme works by detecting a specific conjunction of
excitations: activity aroused by increase or decrease of illumination in
region A is delayed and arrives at the 'and' gate in the next layer
synchronously with activity aroused when the image moves on to region
B. Activity from B passes to the 'and' gate below it, and is also passed
laterally to interact with activity from C. The sequence ABC is the
preferred sequence, and the gates only respond when their respective
conjunctions 'both B and delayed A', or 'both C and delayed B' occur.
Instead of selecting the preferred stimulus by a logical conjunction, the
right hand scheme rejects the null stimulus by veto. Activity aroused at
' on) or ' off' in region B or C is again passed laterally and acts after a delay,
but in this case it inhibits the next unit. As before, CBA is the null
sequence, and when it occurs the inhibition from C prevents the response
that would have resulted from B alone, and inhibition from B likewise
vetoes A's response. On the other hand if the sequence is the preferred one,
ABC, then the inhibition from B does not arrive until the excitation from
A has already got through, and likewise C is unsuccessful in vetoing B. It
will be observed that this scheme only requires that inhibition persists

571
490 H. B. BARLOW AND W. R. LEVICK
longer than excitation; a definite delay when it is passed laterally is not
strictly necessary.
Some evidence favouring the right-hand, inhibitory, scheme has already
been given. (1) As shown in Fig. 2 a stationary spot turned on and off
elicits a response. If the excitatory conjunction scheme was modified to
account for this it would probably still predict a considerably lower
threshold for a moving than for a stationary spot. As shown in Fig. 5
of Barlow et al. (1964), the thresholds for spots of various sizes moving in
the preferred direction differ by small and inconstant amounts from those
for the same spot turned on or off. (2) The most striking feature of these
directional units is the absence of any impulses when movement is in the
null direction. This prompts one to look for a mechanism that inhibits
unwanted responses. (3) WVhen testing for directional selectivity in
Excitatory mechanism Inhibitory mechanism

A BC A BC

At At At A

'And' "' And not'

(conjunction) A|
B B'. C A.- B' B.- C (veto)
gates gates

Preferred'direction Null direction

Fig. 7. Two hypothetical methods for discriminating sequence. For both, the
preferred direction would be from left to right, null from right to left. In the excita-
tory scheme activity from the groups of receptors A and B is delayed before it is
passed laterally in the preferred direction to the 'and' (conjunction) gates. If
motion is in the preferred direction A' (delayed A) occurs synchronously with B,
B' occurs synchronously with C, and these conjunctions cause the units in the next
layer to fire. In the scheme on the right the activity spreads laterally, but in the null
direction, from the groups of receptors B and C, and it has an inhibitory action at
the units in the next layer; hence these act as 'and not' (veto) gates. The inhibition
prevents activity from A and B passing through these gates if motion is in the null
direction, but arrives too late to have an effect if motion is in the preferred direction.
Notice that a special delay unit is not really necessary, for this scheme works if
inhibition simply persists longer than excitation and can thus continue to be
effective after a lapse of time. The excitatory scheme works by picking out those
stimuli with the desired property, whereas the inhibitory scheme works by vetoing
responses to unwanted stimuli; the latter is the one favoured by the experimental
evidence.

572
 MECHANISM OF DIRECTIONAL SELECTIVITY 491
different parts of the receptive field we found that the results obtained at
one edge were anomalous in that movements in both null and preferred
directions gave responses. Such responses are illustrated in the top row of
records in Fig. 4. On the inhibitory scheme it may be possible to resolve
this anomaly along the following lines. Responses from the last points
crossed by a spot moving in the null direction are normally prevented by
inhibition coming from the penultimate regions that have just been
crossed. If the spot is moved to and fro solely in the rim, this penultimate
region is avoided, and consequently it never inhibits the responses coming
from the rim. Measurements of this 'inhibition-free' zone at the rim
suggest that it may extend for as much as 10 inwards from the extreme
edge of the receptive field.
These observations are not decisive, but they brought the inhibitory
scheme to the front of our minds, and we now give some much stronger
evidence favouring it.
Movements in null direction evoking responses. A spot of light moved
continuously through the field in the null direction will evoke no impulses,
but if such continuous motion is interrupted while the spot is in the
receptive field, a burst of impulses occurs just when the movement starts
up again. Evidently the inhibition that prevents the response when
motion is continuous decays while the spot is stationary, so that when the
spot moves on to new receptors the activity excited escapes inhibition and
gets through to the ganglion cell. This response to intermittent motion is
illustrated in Fig. 8.
If motion in the null direction is slow enough, a discharge can also be
elicited, and this is imustrated in Fig. 9. Presumably the rate of rise and
decay of the inhibitory process, together with the distance at which it
operates, governs the range of speeds over which directional selectivity
occurs.
Responses to slits singly and in sequence. What was thought to be a
crucial test of the inhibition hypothesis was devised. Two slits were placed
close to each other and the responses to each in isolation were recorded
several times at on and off. The slits were then presented in null or
preferred sequence, and several responses again recorded. The records were
analysed by counting the impulses that occurred within i sec of stimula-
tion, and the averages of 4 to 7 responses are presented in Table 3.
First compare the figures in the last two columns, and notice that the
result confirms what has already been said. Preferred sequences are more
effective stimuli than null sequences at all separations studied, but the
difference is most pronounced at small separations and decreases at the
separations greater than 17'. Now compare the figures in the 'Null'
column with those in the 'A + B' column. In every case the 'Null' has the

573
492 H. B. BARLOW AND W. R. LEVICK
lower figure, so that inhibition certainly occurs: when the sequence is in
the null direction fewer impulses occur than when each region is excited
separately. Finally, compare the figures in the 'Preferred' column with
those in the 'A +B' column. Here there is an excess in the 'Preferred'
column at small separations, but not at large separations.
Position 50S 100A in visual field
Spot size III
/a b
Null directioni

ac
b _ b c

. \ ~ ~ ~~c
d

UniI3-26
Background 10 cd/M2 d e

Spot 60 cd/M2-

Unit 3-26

_I
-e ;c ~ ;

3- -I- 0 5 sec
Fig. 8. Escape of impulses with intermittent movement in the null direction.
Five positions are marked in relation to the outline of the receptive field shown on
the left. The lowest two pairs of records show the effect of sweeping continuously
through these positions in the null (abede) and preferred (edoba) directions. In the
upper four pairs the spot was moved discontinuously, first from a to b, then from
b to c, then from c to d, then from d to e just outside the field. The lower trace of
each pair shows the position of the spot in the field. As an example of the escape
phenomenon notice that no impulses occur when movement from c to d is part of a
continuous sweep (5th pair of records), but they do occur when this movement is
made in isolation (3rd pair). The suggested interpretation is that 'on' or 'off'
stimulation at any point inhibits 'on' or 'off' excitation of the next point in the
null direction, but this inhibition decays with time. When the spot pauses at c, off
excitation from c, and on excitation of the next point, occur after inhibition has
decayed and impulses therefore escape.

574
 MECHANISM OF DIRECTIONAL SELECTIVITY 493
. . I - . I
2
I Null Degrees/sec
b 3 -s_o__-_-_i_Preferre5 5
Null Preferred

07
d Preferred

06
f Null

h
hh H-- - -----__ 0.0
Stationary
Background 10 cd/M2 0 5 sec
Spot to 60 cd/M2 Unit 3-26
Fig. 9. Paradoxical response to very slow motion in the null direction. As before
the lower trace of each pair shows the position of the spot of light. For movements
at about 50/sec a vigorous response was obtainable in the preferred direction, but
the top pair of records shows that in the null direction there is no increase over the
maintained firing rate with no stimulation (lowest pair of records). When motion
was at about 0.70/sec there was still a vigorous response in the preferred direction
(2nd pair), but in the null direction (3rd pair) there was also a distinct increase
compared with the maintained discharge (4th pair). If movement is slow enough,
the inhibition at a point in front of the advancing spot must have declined by the
time the spot reaches it to a level where extra impulses are allowed to pass.

TABLE 3. Inhibition and sequence-discrimination

Two narrow rectangular slits A and B were lit from behind, and were spaced various
distances apart along the preferred-null axis of the receptive field. Responses were recorded
when each slit was turned on and off, first, in isolation, and then in sequences corresponding
to the null (BA) and preferred (AB) directions. The figures are the average numbers of
spikes that occurred within j sec of stimulation (4-7 responses averaged). For the null
sequence there was always a deficit of spikes compared with the sum of the spikes produced
by the two slits separately.
Slit Null Preferred
separation Stimulus A B A + B BA AB
10 06' On 2-6 1-7 4.3 2-0 1*8
Off 9.3 4.7 14-0 5.9 8*0
34' On 5*2 3-2 8-4 3.4 6-3
Off 10-2 6-2 16-4 4-6 14-9
17' On 5-1 3-2 8-3 1-6 13-9
Off 9-1 4*1 13-2 3-2 19-8
8' On 5.0 4-0 9.0 2-0 13-0
Off 8-5 5.5 14-0 1*8 17*7

575
494 H. B. BARLOW AND W. R. LEVICK
Individual responses are highly variable and the experimental situation
needs systematic exploration with averaging techniques. At this stage we
can say that the experiment obviously supports the idea that null
sequences are ineffective stimuli because of inhibition. It also indicates,
however, that there is some degree of facilitation for preferred sequences,
though it is fair to add that this seems a less important effect than the
inhibition.
Receptive field alone j Receptive field and surround 1 Surround alone

t
r e er re>S/// /
'I\
Preferred Preferred

62 30 l!4

0o5 sec

NullI z Null Null

_l40
unit E-u
'I.
Fig. 10. Lateral inhibition and responses to movement. A black edge was moved
behind a mask of grey paper (cross-hatched) so that the advancing border crossed
a 40 hole exposing the receptive field alone with the surround masked off (left), the
surround alone with the receptive field masked off (right), or it crossed both to-
gether, with no mask (centre). The records show the responses; the lower trace of
each pair came from a photocell aimed at the receptive field. No impulses were
obtained when motion was in the null direction (lower pair of records). In the pre-
ferred direction some were obtained in each case, but the response was much
greater with the surround masked off than when it crossed surround and centre
together (62 instead of 30 impulses). Motion in the surround inhibits the response
to motion in the centre, just as light going on or off in the surround inhibits on or
off responses from the centre.

576
 MECHANISM OF DIRECTIONAL SELECTIVITY 495
Inhibition from outside the receptive field
The type of inhibition postulated to account for directional selectivity,
and shown up in the experiment of Table 3, comes from within the recep-
tive field-that is, from within the region where light can evoke impulses.
There is also an inhibitory mechanism acting from outside the receptive
field-that is, from the surrounding region where light stimuli evoke little
or no response. Figure 10 shows an example of the effect of this inhibitory
mechanism on the discharge evoked by a moving object. Figure 5 of
Barlow et al. (1964) shows the effect of this inhibitory mechanism on the
threshold.
DISCUSSION
Physiological function and anatomical structure
We think that the experiments described establish without need of
further discussion these four points about directional selectivity. First, it
is not caused by optical aberrations, nor by simple differences of latency
for discharges evoked from different parts of the receptive field. Secondly,
it is not necessary to cross any critical region or line in the receptive field:
the mechanism responsible for the property resides in small subdivisions
of the field and must be extensively replicated. Thirdly, these replicated
subunits distinguish between null and preferred sequences of excitation of
a pair of regions with which they connect; thus the directional selectivity
of the ganglion cell is built up from sequence-discriminating subunits.
Fourthly, inhibition plays an important part in this discrimination by
preventing responses to sequences corresponding to motion in the null
direction.
By themselves these results probably do not justify any further con-
clusions, but the complexity of function that they have revealed is
beginning to match up to the long-known complexity of neural structure
in the retina. It is a challenging problem to fit together the jig-saw puzzle
of anatomical elements in the hope of revealing the picture of physiological
function, and a tentative solution is shown in Fig. 11. It is certainly
incomplete, for it does not specify the connexions of the concentric type of
ganglion cell, nor of those selectively responsive to fast and slow movement
(Barlow et al. 1964). Furthermore, we assume that there is a duplicate set
of bipolar and horizontal cells that are activated at 'off'. We have some
evidence, to be presented elsewhere, that 'on' and 'off' systems do not
interact with each other at this level, and therefore for simplicity we have
omitted the 'off' system. Because of the diversity of types of bipolar and
horizontal cells (on and off for at least four different directions) one can
see why a very large number of bipolar cells are required to handle the
input from a group of receptors.

577
496 H. B. BARLOW AND W. R. LEVICK
At various points there are alternatives to our scheme that are not
excluded by the evidence at present available. On the other hand the roles
of the anatomical elements and their postulated connexions are not as
arbitrarily assigned as a naive reader is liable to suppose. For discussion,
take what is perhaps the most controversial and interesting feature of the
scheme-the assignment to horizontal cells of the role of inhibitory
elements that prevent bipolar cells responding to null sequences. There are
two'main questions to be answered: why place the inhibitory element in
the inner nuclear layer? And why postulate that the horizontal cell

Null direction

Horizontal ' '

cells inhibit
Bipolar cells . Tdv . - . .H. . .

* ~~~~G

- =~~~~~j
1 00je
Fig. 11. Suggested functional connexions of the retinal elements concerned with
directional selectivity. The elements are freely adapted from Cajal (1893), and are
assembled in accordance with the functional organization suggested in this paper.
The scale of the diagram is approximate and a posterior nodal distance of 1 1l5 mm
has been assumed. The pathway of excitation is from receptors (R), through bipolars
(B), to the ganglion cell (G), but activity in this direct pathway is modified by the
associational cells. The horizontal cells (H) pick up from receptors, conduct laterally
in the null direction through a teledendron (Td), and inhibit bipolars in the neigh-
bouring region. This prevents responses when an image moves in the null direction,
but has no effect when motion is in the preferred direction. Horizontal cells have
the function of the laterally conducting elements in the inhibitory scheme shown in
Fig. 7. The amacrine cells (A) are thought to pick up from bipolar endings in the
inner plexiform layer and to conduct activity throughout their axo-dendritic
ramifications; they are assumed to make synaptic connexion with the ganglion cells
and inhibit them, thus mediating lateral inhibition of the type illustrated in Fig. 5
of Barlow et al. (1964) and Fig. 10 of this paper. The off-responding mechanism is
not illustrated, but seems to require duplicate horizontal cells and bipolar cells.
Notice that the ganglion cell must connect selectively to those particular bipolars
which respond selectively to the sequences for one particular direction. Its
response is specific for this pattern of stimulation but is invariant with respect to
contrast and position in the receptive field. It may be said to achieve some degree
of 'stimulus generalization'.

578
 MECHANISM OF DIRECTIONAL SELECTIVITY 497
connects from receptors to bipolar cells, rather than, for instance, from
bipolars to bipolars?
The strength of the proposed scheme arises from the fact that a function
can naturally be assigned to the neural elements that are known to exist,
without making esoteric or revolutionary assumptions about how they
work. Sequence-discrimination is assigned to bipolar cells because the
ganglion cell appears to pick up from subunits that are replicated in
different parts of the receptive field, and bipolar cells are the replicated
anatomical elements that feed ganglion cells. There is physiological
evidence of inhibitory interaction acting from one side on these subunits.
This is not like the classical lateral inhibitory interaction which counteracts
the pooled excitatory influences reaching the ganglion cell: the evidence
points to inhibition that acts locally. Excitation aroused from a particular
region of the receptive field is inhibited by preceding excitation of the
region that a light image has just crossed when motion is in the null
direction. This same inhibitory region has no influence on excitation
aroused from the neighbouring region on the opposite side, for it fails to
block the excitation when motion is in the preferred direction. The
physiological evidence thus indicates that each excitatory region has its
own private inhibitory region on one side, and one can construct a
number of schemes to account for this. Inhibition might act on the
ganglion cells, but in such a way that it only blocks one particular branch
of the dendritic tree: or it might act presynaptically on the bipolar cell
endings. Another possibility one might consider is that the inhibition is
mediated by the receptor-to-receptor connexions described by Sj6strand
(1958) in the guinea-pig. The distance over which the inhibitory effects
have to be passed may be a difficulty here, and this notion shares the
difficulty described below for other forms of inhibition which act on the
receptors. Since the horizontal cells are known to have processes conduc-
ting laterally the natural starting hypothesis is that they are the cells
carrying this inhibition from one region to another.
If this is granted there is still scope for argument as to where this
inhibition is picked up from, and where it feeds to. Might it not inhibit
receptors rather than bipolar cells? Might it not even pick up from bipolar
cells and feed back to receptors? The key observation here is that a region
which has itself been inhibited from its own private inhibitory zone on one
side can none the less inhibit activity aroused in the neighbouring zone on
the other side. Motion through the receptive field in the null direction may
elicit no impulses whatever. Consider what is happening half way through
such a traverse: one sees that excitation of the receptors at the mid-point
prevents the discharge from the next group of receptors the spot is going to
cross, even though no activity is transmitted centrally from the receptors
32 Physiol. 178

579
498 H. B. BARLOW AND W. R. LEVICK
at the mid-point. This indicates that the inhibitory connexion runs from
an early point on the path from the inhibition-arousing zone to a later
point on the path from the zone that is inhibited. Presumably then it
runs from receptors to bipolar cells, and in that case the inhibition can act
in the ordinary way by stabilizing the membrane potential of the bipolar
cells. However there is clearly a point here open to histological'investiga-
tion. Do the horizontal cells make this pattern of connexion in the rabbit?
Polyak (1941) describes the horizontal cells of the monkey as making
receptor-to-receptor connexions.
No further information has come to light on the pathway mediating
ordinary lateral inhibition of the type shown in Fig. 10. This probably acts
on the ganglion cells, and amacrine cells remain the most plausible guess.
It is clear that our allocation of functions to particular structures must
be regarded as provisional, but we were pleased to find how well the
physiological organization seems to fit in with the anatomical structure.
Other proposed anatomical correlates in other species. Maturana, Lettvin,
McCulloch & Pitts (1960) and Lettvin, Maturana, Pitts & McCulloch (1961)
have also attempted to relate structure and function in the retina, in their
case in the frog. Their discussion has something in common with ours, but
they place greater emphasis on the concept that the ganglion cell's
properties are determined by the shape and size of its dendritic tree. They
believe that the different strata of the inner plexiform layer carry informa-
tion as to different properties of the pattern of light falling on the recep-
tors; the ganglion cell is then thought to pick up the appropriate
combination of these properties by ramifying in the various layers. This
may explain how a ganglion cell is able to make connexion with a specific
subset of bipolar cells, and their notion does not contradict ours. Where we
feel that our scheme goes further is in showing how the complex task of
signalling direction of movement can be broken down into simpler tasks
that can be performed by elements making simple excitatory and inhibi-
tory connexions.
Maturana & Frenk (1963) have described directionally selective units in
the pigeon's retina. These obviously have much in common with the units
in the rabbit, for they show the same directional selectivity independent of
the path through the field and the contrast of the moving object. Further-
more, they made an interesting observation which led them to the con-
clusion that an inhibitory mechanism is involved in directional selectivity.
They turned a spot of light on and off in one place in the receptive field,
eliciting responses in the usual way. While the light was off they moved it
to another position in the field displaced in the null direction from the first
position, and turned it on and off again. No responses were obtained,
whereas, if the spot had been displaced in the preferred direction, responses

580
 MECHANISM OF DIRECTIONAL SELECTIVITY 499
were obtained as usual. Clearly this is similar to the two-spot experiment
described here, but it seems from their brief description that inhibition
must persist for a long time in the pigeon. They do not attempt to make
detailed suggestions about which anatomical structures are responsible
for the specificity of the stimuli that generate responses in a particular unit,
but they give the impression that they believe it is achieved by the ganglion
cell. In our view the specificity originates with the bipolars, and the
ganglion cell generalizes for position and contrast by picking up only from
those bipolars that respond to sequences of on or off stimuli corresponding
to one particular direction of motion.
Griisser-Cornehls, Griisser & Bullock (1964) tested movement-sensitive
units in the frog with various stimuli, and came to the conclusion that
movement detection was really 'change-of-position' detection. Their
experiment suggests that they are distinguishing between continuous and
discontinuous change of position, and in that case our conclusions are not
too far apart: discontinuous change of position, as in the two-spot experi-
ment, can activate the directionally selective mechanism. However, they
were not dealing with units responding selectively to the direction of
motion, for unlike Maturana et al. (1960) they failed to find such units in
the frog, although they confirmed many of these authors' other findings.
Directional system in insects. Reichardt (1957, 1961a, b) has proposed a
mechanism capable of explaining the responses of insects to movement in
their visual field. This seems at first sight very different from the one we
have arrived at, for his scheme depends upon evaluating the cross-correla-
tion between the signal from an ommatidium and that from its neighbour
modified by passage through a low-pass filter. This is closer to the excita-
tory-conjunction scheme that we rejected than it is to the inhibitory
scheme. However, one should probably regard Reichardt's proposal as the
simplest physical system with a performance specification similar to the
beetle's eye, and one should not be too surprised if the realization of a
system in 'biological hardware' is different from what it would be in
physical hardware, even if the operation performed is very similar.
Pattern recognition, trigger features, and stimulus generalization
Maturana & Frenk (1963) suggest that an understanding of the type of
behaviour they describe in the ganglion cells of the pigeon retina clarifies
certain problems of pattern recognition. We think there are two aspects of
recent work on the visual pathway that are interesting in this respect. The
first is the specificity of the features that are effective in triggering the
activity of sensory neurones. Examples of this are provided by the 'fly
detectors' (Barlow, 1953) and 'convexity detectors' (Lettvin et al. 1959) of
the frog's retina, the Jinear elements of the cat's cortex (Hubel & Wiesel,
32.2

581
500 H. B. BARLOW AND W. B. LEVICK
1959, 1962), the 'horizontal edge detectors' of the pigeon retina (Maturana
& Frenk, 1963), and the directionally selective elements found in all these
preparations as well as in the rabbit's retina. Now in pattern recognition
by machines, Grimsdale, Sumner, Tunis & Kilburn (1959) broke the task
into two stages by first detecting the presence of certain key features of the
patterns to be discriminated and then looking for the characteristic
combinations of these features. Most of the successful systems for recogni-
zing printed or handwritten characters make use of a similar scheme
(Selfridge & Neisser, 1960; Uhr & Vossler, 1961; Frishkopf & Harmon,
1961; Kamentsky & Liu, 1963), and it is interesting to see why it is
necessary for the computer to view its text through these 'feature filters'.
It is because even the largest computor cannot recognize letter A by com-
paring the input with a complete list of all members of the class of A's.
Such an approach would require the separate representation of each of the
2n possible states of the n binary inputs and this becomes unmanageable
for values of n that are very small by biological standards. Presumably
the trigger features of the visual system likewise enable the input states
to be classified in an effective way without requiring a googolian number
of separate representations.
The second aspect we want to draw attention to is the detailed manner
in which the specific and general properties of these trigger features are
picked out. This discussion will be based upon our suggested mechanism
for directional selectivity, and we shall introduce certain simplifications
which, though not entirely justifiable, make it easier to compare the
neural process with artificial pattern recognition.
According to our analysis the operation of abstracting direction of
movement is done in two stages, each with the same two steps. The first
step in each case is the summation or pooling of selected excitatory
influences, and the second step is the inhibitory interaction of another
element that has, as it were, the power of veto. The first step loses informa-
tion, for the bipolar cell which pools inputs from a number of receptors does
not reflect in its output which particular ones were active. As pointed out by
Reichardt (1961a, b) the inhibitory step could in principle regain this lost
information, but in the case of bipolar cells it does not do this; instead it
makes the response more selective by bringing in new information.
Without this inhibitory interaction a bipolar cell would simply say, when
it became active, 'Light fell in this region'; with the inhibition it says
'Light fell in this region and was not preceded by light falling in that
region'. Compared with the receptors in the preceding layer, the bipolars
have lost some information about the exact position of the stimulus, but
they have extracted some information about the presence of a particular
sequential pattern in the stimulus.

582
 MECHANISM OF DIRECTIONAL SELECTIVITY 501
The same two steps are taken in the next stage, occurring in the next
layer. Here a ganglion cell does not pool from all the bipolars in the
receptive field, but it picks up selectively from all those which respond
when a stimulus moves in a particular direction, irrespective of the location
of the bipolar cell or whether it belongs to the 'on' class or the 'off' class.
It thus discards the information as to the contrast of the stimulus object
and whereabouts in the receptive field it was; it 'generalizes' by grouping
together activity resulting from movement in a particular direction,
regardless of contrast and exact position. This is followed by inhibitory
interaction which again makes the response more specific. Light going on
or off in the surround (Fig. 5 of Barlow et al. 1964), or movement in
the surround (Fig. 10, this paper) reduces or prevents the response, so
that when activity occurs it implies that changes were not occurring
in the surrounding retina at the time they occurred within the receptive
field.
Let us now express the logical pattern of these repeated operations
symbolically. The pooling or generalizing operation is equivalent in some
ways to the formation of a logical union (inclusive 'or', symbolized by v),
and the inhibitory or veto operation is equivalent to 'and not. ..'
(symbolized by . -). If B is the class of inputs to which a bipolar cell
responds, and Ra, Rb, etc., are the inputs causing activity in the receptors
a, b, etc., then
'' B~~B= (RaVRbVRc...). (RrVRsVRt..
Likewise the class a of inputs causing activity in a ganglion cell is ex-
pressed in terms of Ba, Bb, etc., the inputs which activate the selection of
bipolars it connects with; thus
U = (BaVBbVBc...). (BrVB8VBt,..)
If we symbolize by Elk the class of inputs which is effective in exciting a
particular element after 3b synapses, and by EOA+1 the class effective for an
element after one more synapse, then E0r+1 is given by
EO+'= (E&vENvEg...). (ErvEfvE*...).
Notice that only a small proportion of the possible logical functions of the
PI can be expressed in this form, and it is therefore not at all a trivial
restriction.
We are suggesting that the classification system at one level in the
nervous system is built out of the classification at the preceding level by a
combination of pooling or union, and inhibition or veto (and not ... ). Can
we regard the proposed mechanism for directionally selective units as a
paradigm of the neural mechanisms responsible for the classification
system imposed on our sensory input? Is pooling analogous to 'stimlulus

583
502 H. B. BARLOW AND W. R. LEVICK
generalization', and is greater specificity of response always achieved by
the veto of an associational neurone, an interposed inhibitory element?
These are intriguing questions.

SUMMARY
1. The mechanism of directional selectivity has been investigated in
retinal ganglion cells of decerebrate or lightly anaesthetized rabbits.
2. The property of responding to one direction of motion (preferred) but
not to the opposite (null) direction occurs in on-off units, but the responses
to movement cannot be predicted from the map of the receptive field
obtained with static stimuli; the property cannot be explained by optical
aberrations (see Fig. 1), nor by progressive changes of latency across the
field (see Fig. 2).
3. There is no critical line or region that must be crossed to produce
unequal responses to preferred and null motion (Fig. 4): small subsections
of the receptive field possess the property (Fig. 5).
4. The response to successive stimulation of two small regions depends
upon whether the order corresponds to motion in the preferred or null
direction (Fig. 8). This effect is strong when the two regions are within
about I' of each other, but declines at greater separations.
5. This is thought to indicate that directional selectivity results from
the discrimination of sequence. Normal movement excites many points in
a long succession, but the mechanism works by discriminating the sequence
of individual pairs of regions.
6. When two stimuli are presented in the null sequence the number of
impulses elicited is much less than the sum of the numbers elicited from
each stimulus in isolation (Table 3). There is a small excess of impulses over
this sum when the stimuli are presented in the preferred sequence.
7. From this and other findings it is concluded that sequence-discrimina-
tion results primarily from an inhibitory mechanism that vetoes the
response to null sequences, rather than from the detection of the con-
junction of excitation from two regions with an appropriate delay (see
Fig. 7).
8. If the image of a moving object spreads outside the receptive field on
to its surround there are fewer impulses than when it is confined to the
receptive field alone (Fig. 10). This must be the inhibitory mechanism that
elevates the threshold for large compared with small spots, and it is
presumably different from the inhibition responsible for sequence-
discrimination.
9. The functional organization is discussed in relation to the anatomical
organization (Fig. 11). It is suggested that horizontal cells conduct

584
 MECHANISM OF DIRECTIONAL SELECTIVITY 503
laterally and inhibit the bipolars on one side, thus preventing them from
responding to null sequences; the ganglion cells then pick up from the
bipolars responsive to like sequences and it is thought that the inhibition
from the surround may be mediated by amacrine cells.
10. The ability to abstract direction of motion irrespective of the
position in the receptive field and the contrast of the moving object has
elements in common with much more complex feats of pattern recognition.
The two steps-inhibition by associational neurones and selective pooling
-may also play a part in these more complex feats.
We wish to thank W. A. H. Rushton, P. A. Merton, P. E. K. Donaldson and G. West-
heimer for the loan of apparatus, and W. Hail, C. Hood, R. Rumble and P. Starling for
help in construction and photography. This work was supported in part by Grants NB 05215
and NB 03154 from the U.S. Public Health Service. The micromanipulator for the Cam-
bridge experiments was constructed in the Department of Physiology, University of Sydney
to the design of P. 0. Bishop and W. Kozak.

REFERENCES
BARLOW, H. B. (1953). Summation and inhibition in the frog's retina. J. Phy8iol. 119,
69-88.
BARLow, H. B. & HILL, R. M. (1963). Selective sensitivity to direction of motion in
ganglion cells of the rabbit's retina. Science, 139, 412-414.
BARLow, H. B., HILL, R. M. & LEVICK, W. R. (1964). Retinal ganglion cells responding
selectively to direction and speed of image motion in the rabbit. J. Phy8iol. 173, 377-407.
CAJAL, S. RAmoN y (1893). La r6tine des vert6bres. In La Cellue, 9, 119-257, ed. CARNoY,
J. B., GILSON, G. & DENYS, J. Lierre: J. van In and Co.; Louvain: A. Uystpruyst, Libraire.
FRIsHKOPF, L. S. & HARMON, L. D. (1961). Machine reading of cursive script. In Proc. 4th
Lond. Symp. on Information Theory, pp. 300-316, ed. C. CEERRY. London: Butterworth.
GRIMSDALE, R. L., SUMNER, F. H., TuNIs, C. J. & KILBURN, T. (1959). A system for the
automatic recognition of patterns. Proc. Inst. elect. Eng8, B, 106, 210-221.
GRUSSER-CORNEHLS, U., GRJssEra, 0. J. & BuLLOCE, T. H. (1964). Unit responses in the
frog's tectum to moving and non-moving stimuli. Science, 141, 820-822.
HASsENSTEIN, B. (1951). Ommatidienraster und afferente Bewegungsintegration. Z. vergl.
Physiol. 33, 301-326.
HUBEL, D. H. (1959). Single unit activity in striate cortex of unrestrained cats. J. Phy8iol.
147, 226-238.
HUBEL, D. H. & WIESEL, T. N. (1959). Receptive fields of single neurones in the cat's
striate cortex. J. Phy8iol. 148, 574-591.
HUBEL, D. H. & WIESEL, T. N. (1962). Receptive fields, binocular interaction and functional
architecture in the cat's visual cortex. J. Physiol. 160, 106-154.
KAMENTSKY, L. A. & Liu, C. N. (1963). Computer-automated design of multifont print
recognition logic. I.B.M. Journal of Research and Development, 7, 2-13.
LETTVIN, J. Y., MATURANA, H. R., MCCULLOCH, W. S. & PITTS, W. H. (1959). What the
frog's eye tells the frog's brain. Proc. Inst. Radio Engr8, N.Y., 47, 1940-1951.
LETTvIN, J. Y., MATuRANA, H. R., PITTs, W. H. & McCuLLOCH, W. S. (1961). Two remarks
on the visual system of the frog. In Sensory Communication, pp. 757-776, ed. W. ROSEN-
BLITH. M.I.T. Press; New York: John Wiley.
MATURANA, H. R., LETTVIN, J. Y., MCCULLOCH, W. S. & PITTS, W. H. (1960). Anatomy
and physiology of vision in the frog (Rana pipiens). J. gen. Physiol. 43, suppl. 2,
Mechanisms of Vision, 129-171.
MATURANA, H. R. & FRENK, S. (1963). Directional movement and horizontal edge detectors
in pigeon retina. Science, 142, 977-979.
POLYAX, S. L. (1941). The Retina. Chicago: University of Chicago Press.

585
504 H. B. BARLOW AND W. R. LEVICK
REICHARDT, W. (1957). Autokorrelationsauswertung als Funktionsprinzip des Zentral-
nervensystems. Z. Naturf. 12, 447-457.
REICHARDT, W. (1961 a). Autocorrelation, a principle for the evaluation of sensory informa-
tion by the central nervous system. In Sensory Communication, pp. 303-317, ed. W.
ROSENBLI. M.I.T. Press; New York: John Wiley.
REICHARDT, W. (1961b). Yber das optische Auflosungsvermogen der Facettenaugen von
Limulus. Kybernetik. 1, 57-69.
SELIRIDGE, 0. & NEIssER, U. (1960). Pattern recognition by machine. Sci. Amer. 203,
60-68.
S TPs.AxD, F. S. (1958). Ultrastructure of retinal rod synapses of the guinea pig eye as
revealed by three-dimensional reconstructions from serial sections. J. Ultra8tr. Re8. 2,
122-170.
THOMSON, L. C. (1953). The localisation of function in the rabbit's retina. J. Physiol. 119,
191-209.
UEm, L. & VossLER, C. (1961). A pattern-recognition program that generates, evaluates
and adjusts its own operators. Proceeding8 of the We8tern Joint Computer Conference, 19,
555-570.

586
Proc. Natl. Acad. Sci. USA
Vol. 94, pp. 1264912654, November 1997
Neurobiology
Deciphering a neural code for vision
CHRISTOPHER PASSAGLIA*, FREDERICK DODGE*, ERIK HERZOG*, SCOTT JACKSON,
*Marine Biological Laboratory, Woods Hole, MA 02543; Department of Bioengineering and Neuroscience, Syracuse University, Syracuse, NY 13244; and
Center for Vision Research, Department of Ophthalmology, 3258 Weiskotten Hall, State University of New York Health Science Center, 750 East
Adams Street, Syracuse, NY 13210
Edited by John E. Dowling, Harvard University, Cambridge, MA, and approved September 10, 1997 (received for review June 2, 1997)
ABSTRACT Deciphering the information that eyes, ears,
and other sensory organs transmit to the brain is important
for understanding the neural basis of behavior. Recordings
from single sensory nerve cells have yielded useful insights,
but single neurons generally do not mediate behavior; net-
works of neurons do. Monitoring the activity of all cells in a
neural network of a behaving animal, however, is not yet
possible. Taking an alternative approach, we used a realistic
cell-based model to compute the ensemble of neural activity
generated by one sensory organ, the lateral eye of the horse-
shoe crab, Limulus polyphemus. We studied how the neural
network of this eye encodes natural scenes by presenting to the
model movies recorded with a video camera mounted above
the eye of an animal that was exploring its underwater habitat.
Model predictions were confirmed by simultaneously record-
ing responses from single optic nerve fibers of the same
animal. We report here that the eye transmits to the brain
robust neural images of objects having the size, contrast,
and motion of potential mates. The neural code for such
objects is not found in ambiguous messages of individual optic
nerve fibers but rather in patterns of coherent activity that
extend over small ensembles of nerve fibers and are bound
together by stimulus motion. Integrative properties of neurons
in the first synaptic layer of the brain appear well suited to
detecting the patterns of coherent activity. Neural coding by
this relatively simple eye helps explain how horseshoe crabs
find mates and may lead to a better understanding of how
more complex sensory organs process information.
Living in a world rich with information, animals are highly
efficient at extracting what is essential for their survival. The first
stage of visual processing begins in the eye where the retina
transforms patterns of incident light intensity into trains of
impulses in optic nerve fibers (1). The retina can encode infor-
mation in a reliable and efficient manner (2, 3) but does not
encode all of it. Rather the retina emphasizes certain features in
the visual scene at the expense of others, presumably to help
animals find food, locate mates, and avoid predators (46).
Knowing what an animal can see is crucial for exploring the
neural code the eye transmits to the brain.
We report here a study of neural coding in the lateral eye of the
horseshoe crab, Limulus polyphemus. We selected Limulus as a
model system because its lateral eye contains the largest neural
network for which a quantitative cell-based model exists (79),
and its visually guided behavior is well known (10, 12, 13).
Field studies show that vision has an important role in
Limulus mating behavior (10). In the spring millions of crabs
migrate to the oceans edge from Maine to Mexico to build
nests and deposit eggs (11). Male crabs use their lateral eyes
to find mates, whereas females use them to avoid nesting sites
of other crabs (10, 12). The crabs can see other horseshoe crabs
The publication costs of this article were defrayed in part by page charge
payment. This article must therefore be hereby marked advertisement in
accordance with 18 U.S.C. 1734 solely to indicate this fact.
1997 by The National Academy of Sciences 0027-8424y97y9412649-6$2.00y0
PNAS is available online at http:yywww.pnas.org.
12649
AND ROBERT BARLOW*
and objects resembling them under a wide variety of illumi-
nation conditions despite large differences in the contrast of
their carapace (12, 13). Fig. 1b shows underwater photographs
of black and gray cylindrical objects that approximate the size
and range of contrast of the female carapace. Even though the
gray object is much less apparent, male crabs can detect it
almost as well as the black one (12, 13).
Such visual performance is remarkable in view of the rather
simple design of the animals lateral eyes. Each compound eye
samples the underwater world with '1,000 hexagonally packed
receptor units (ommatidia). Individual ommatidia collect light
from a relatively large region of space ('6), giving the animal a
nearly hemispheric field of view (14). A cluster of retinular cells
within each ommatidium transduces the light into electrical
signals that passively propagate into a single spike-firing eccentric
cell. By using neural mechanisms generally found in more com-
plex retinas, the eccentric cell converts the signals into a train of
action potentials that it transmits to nearby ommatidia and the
brain. Detailed studies of the integrative properties of these cells
have yielded a quantitative description of the response of an
ommatidium in terms of its excitation by light and inhibition from
neighboring ommatidia (7, 8, 15, 16, 18).
Our strategy for deciphering the retinal code underlying
visually guided behavior of Limulus was first to videotape the
lateral eyes view of its underwater world with a crab-mounted
camera. Second, we recorded simultaneously from the optic
nerve fiber of an ommatidium viewing the central region of the
videotaped scene. Third, we computed the ensemble of optic
nerve activities in response to the video by using a model of the
eye. Fourth, if the response computed for the appropriate
neuron in the model matched that recorded from the nerve
fiber, we examined the array of computed responses for
putative coding of behaviorally relevant objects.
MATERIALS AND METHODS
Underwater Video Recording. We recorded underwater
movies with a miniature video camera (CrabCam) attached
to the carapace of crabs moving in the shallow waters (,1 m
depth) of an estuary (Fig. 1a). At the same time, we monitored
the activity of an optic nerve fiber from one of their lateral
eyes. The optic axis of the CrabCam (72 3 54 field of view;
model V-1210, Marshall Electronics, Culver City, CA; model
SVS-1, Chinon American, Mountainside, NJ) was 4 cm above
the eye and aligned parallel with that of the recorded neuron.
Video signals were recorded simultaneously with spike dis-
charges of the nerve fiber on a VHS camcorder. Animals freely
explored their surroundings or were moved along an under-
water track at a speed of '15 cmys, which is the average
velocity of swimming horseshoe crabs (17). They moved past
stationary cylindrically shaped black and gray objects that
approximate the size and range of contrasts of adult females
This paper was submitted directly (Track II) to the Proceedings office.
Abbreviation: ips, impulses per second.
Present address: Center for Biological Rhythms, University of Vir-
ginia, Charlottesville, VA 22903.
To whom reprint requests should be addressed.
587
12650 Neurobiology: Passaglia et al. Proc. Natl. Acad. Sci. USA 94 (1997)

FIG. 1. (a) A horseshoe crab, Limulus polyphemus, at the waters edge mounted with a video camera, CrabCam, for recording underwater movies
and a microsuction electrode for recording responses from a single optic nerve fiber. The barrel of the electrode protrudes to the right from the
recording chamber, which is sealed with a white cap (2.5 cm diameter). Cables lead the video and optic nerve signals to recording electronics located
on an overhead skiff as the animal moves about underwater at depths of 0.5 to 1 m. (b) CrabCam images of a high-contrast black object (Left)
and a low-contrast gray object (Right). Black disks indicate the fields of view of the ommatidia whose optic nerve responses were recorded with
the microsuction electrode. (c) Light intensities (given as contrast) incident on the recorded ommatidia plotted as a function of time as the animal
moved past the black (Left) and gray (Right) objects. Contrast is the ratio of the intensity of a pixel to the average intensity of the scene. (d)
Recordings with the microsuction electrode of spike trains from single fibers in response to the 10-s sequences of light intensities shown in c. (e)
Recorded trains of spikes in d plotted as instantaneous frequencies, which are the reciprocals of the intervals between nerve impulses. ( f)
Instantaneous frequencies computed by the cell-based model in response to digitized video images of the black and gray objects. The objects passed
through the fields of view of the recorded and model neurons from 4 to 6 s after the start of the runs evoking reduced (black object) and
quasi-periodic (gray object) responses. Responses are representative of those recorded under similar conditions in other field experiments (n 5
53 trials).

588
 Neurobiology: Passaglia et al. Proc. Natl. Acad. Sci. USA 94 (1997) 12651

(30 cm diameter; 15 cm high). The objects were placed 75 cm
away because field studies show that about half of the male
crabs detect and turn toward them at this distance (13).
Video sequences (10-s duration) were digitized at 20
framesys (120 3 160 pixels; 8 bits resolution) by using com-
mercial software (Videofusion, Maumee, OH). We rescaled
the digitized movies by using a nonlinear mapping function
that inversely corrects for the gamma response characteristic
of the video camera and for output saturation (18). We
compensated for automatic gain controls of the camera,
camcorder, and computer by subtracting from all frames of a
movie the black level of a black and white test patch viewed by
the CrabCam. The contrast of the black object determined by
these procedures matched well that measured from photo-
graphs taken underwater with high-latitude TMAX-100 film
(Kodak). The turbid, shallow waters where Limulus mate
reduce the contrast of underwater objects exponentially with
distance (attenuation coefficient of '1.2 m21) such that
objects beyond 1.52 m are essentially invisible (18).
Optic Nerve Recording. We recorded spike trains of single
optic nerve fibers from behaving animals by using a water-tight
microsuction electrode (17). Fig. 1a shows the recording
chamber and electrode attached to the carapace anterior to the
right lateral eye. To access the optic nerve, we trephined a hole
in the carapace '2 cm anterior to the eye, slid the tongue of
the saline-filled recording chamber under the nerve trunk, and
fastened the chamber to the carapace. The sheath of the nerve
was slit open, and a single active fiber was teased free and
sucked into the 200 mm electrode tip. The pixel of digitized
movies that intercepted the optic axis of the recorded neuron
was identified by averaging frames coinciding with nerve
impulses in response to a small probe light. Optic nerve signals
were led via a cable (15 m length) to an AC amplifier
(Electronics Shop, The Rockefeller University, New York)
and the audio channel of the camcorder. The audio channel
output was digitized off-line at 11 kHz, and the spike firing
times (Fig. 1d) were converted to instantaneous firing rates
and plotted as a function of time (Fig. 1e).
Computing Optic Nerve Responses. We computed patterns
of spike activity for the ensemble of optic nerve fibers by
feeding the CrabCam movies to a realistic cell-based model of
the lateral eye. To determine the light intensity incident on
each ommatidium of the model, we convolved each frame of
a movie with a Gaussian point-spread function that matched
the acceptance angle of an ommatidium (6 at half-maximal
sensitivity, ref. 19) by using public-domain software (NIH
Image Version 1.6), and then we sampled the blurred frames
at points of intersection with the optic axes of the array of
ommatidia (14). The size of the model was limited to a 12 3
16 array of neurons by the field of view of the CrabCam. Each
frame of a movie thus yielded a 12 3 16 array of pixels
specifying the relative light intensities incident on each om-
matidium at an instant in time (see Fig. 1c), and the sequence
of frames (Fig. 2a) comprised an optically transformed movie
of the stimulus to the retina as the animal explored its
underwater world.
The movie then was fed to the model, which computed spike
trains for the ensemble of optic nerve fibers by using a set of
differential equations representing the excitatory properties of
each ommatidium and the inhibitory interactions between them.
Details of the model are presented elsewhere (18). In brief,
incident light intensities were converted into voltage fluctuations
(quantum bumps) according to an adapting bump model (20, 21).
To account for transduction noise, which contributes 610% to
the coefficient of variation of fluctuations in instantaneous firing
rates in our experiments, quantum bumps were randomly gen-
erated by a Poisson process at a rate dependent on light intensity
(20) and summed to form the receptor potential by using a
four-stage version of the FuortesHodgkin model (22). The
receptor potential then was attenuated by the passive cable

FIG. 2. (a) CrabCam images of the black and gray objects after optical sampling. The arrays of pixels show the light intensities incident on the
12 3 16 array of ommatidia viewing the videotaped scene. Rows 1 and 2 show images of the black object while the animal was moving and stationary,
respectively. Rows 3 and 4 show corresponding images of the gray object. Interval between displayed sequences is 0.25 s. (b) Computed neural images
of the optically transformed visual images in a. The arrays of pixels give the computed firing rates of optic nerve fibers mapped onto a gray scale
with black set to 0 ips and white set to twice the mean firing rate. Mean firing rates to the uniform background illumination in rows 1, 2, 3, and
4 were 12, 20, 13, and 18 ips, respectively. (c) Computed neural images of activity in the brain generated by the eyes output in b. The arrays of
pixels give the simulated responses of brain neurons having integration times of 400 ms mapped onto a gray scale with black set to 0 and white
set to twice the mean. Each image sums eight sequential neural images from the eye. Computed images are based on the temporal properties of
brain neurons and not their spatial interactions. (d) Simulated responses of brain neurons evoked by the activities of an equal number of ommatidia
located in a horizontal row of the model eye (dashed lines in c). Movies of the underwater video recordings and their computed neural images
can be viewed at Center for Vision Research at http:yywww.hscsyr.eduy;eye.

589
12652 Neurobiology: Passaglia et al.
properties of the eccentric cell and integrated with lateral and
self-inhibitory potentials to form the generator potential (23).
The lateral inhibitory inputs to an eccentric cell were computed
by using a dynamic version of the original HartlineRatliff
formulation (7, 24) with inhibitory strength weighted as a function
of retinal distance (25, 26). The self-inhibitory input to an
eccentric cell was calculated by integrating a decaying exponential
function triggered by each impulse the cell fired (27). The
generator potential was further decremented by a putative elec-
trogenic pump (28, 29) and converted into a train of nerve
impulses with a leaky integrate-and-fire encoder at a rate of 1
impulseymV above a threshold of 1 mV (30). Computed trains of
impulses were expressed as instantaneous firing rates for each
frame of the movie, converted to a linear gray scale, and mapped
back onto two-dimensional arrays of pixels yielding a time series
of neural images (31), which are snapshots of the eyes input to reveals an interesting difference in optic nerve responses to
the brain (Fig. 2b).
For all computations we set the 10 parameters of the model
to those derived from the average spatiotemporal transfer
function of the eye measured in the laboratory (18, 26, 34). For
each experiment we scaled two parameters to simulate illu-
mination levels in the field. We trimmed one parameter, the
mean bump rate, to match the estimated transduction noise
and scaled another, the sensitivity of the spike encoder, to
match the mean firing rate of the recorded neuron. We
extended these parameters to all ommatidia in the model
because previous studies have shown that cells in a given eye
have the same properties (15, 25).
The accuracy of the model was checked first by comparing
optic nerve responses computed for controlled stimuli with those
recorded from single optic nerve fibers in the laboratory, and
second by comparing responses computed for underwater movies
with those recorded in the field. Optic nerve responses to drifting
bars recorded in the laboratory were indistinguishable from those
computed with the model, that is, responses recorded to succes-
sive presentations of the same stimulus were as correlated with
those computed by the model as with each other (18). Correlation
coefficients were consistently .0.95 for controlled laboratory
experiments (n 5 5) but were lower (.0.75) for field experiments
(n 5 5) because of difficulties in determining precisely the
stimulus to the recorded neuron. All field experiments yielded
results similar to those reported here.
Measuring Transfer Functions. We measured the eyes
integrative properties by recording the responses of single
optic nerve fibers to sinusoidally modulated light of different
spatial and temporal frequencies presented to animals sub-
merged in a seawater tank in the laboratory (n 5 8). Visual
stimuli (VENUS, Neuroscientific, New York, NY) were dis-
played on a monitor (10 3 13 cm, Tektronics, model 608,
Beaverton, OR) placed in front of a glass window on the tank,
4 cm from the eye, and centered on the optic axis of the
recorded neuron. Spatial transfer functions were measured by
drifting sinewave gratings of different spatial frequencies
across the monitor. Temporal transfer functions were mea-
sured by modulating a 6 spot of light (field of view of an
ommatidium) with the sum of five temporal frequencies
chosen to prevent phase locking of spike discharges (32).
Response gains were calculated for each of the five frequency
components by first estimating the relative modulation of the
recorded firing rate by using the method of least squares and
then normalizing the result with the contrast of the stimulus.
RESULTS AND DISCUSSION
Responses of Single Optic Nerve Fibers. Fig. 1 e and f plot the
responses recorded from a single optic nerve fiber and computed
for the corresponding neuron in the model as an animal moved
past high- and low-contrast crab-size objects in the ocean. Com-
puted and recorded responses match qualitatively in amplitude,
shape, and noise properties, indicating that the essential features
Proc. Natl. Acad. Sci. USA 94 (1997)
of optic nerve responses are well represented by the model. Note
that the variability in firing rate when the black object is outside
the recorded receptors field of view (coefficient of variation
'0.27) is significantly greater than in the case of the gray one
(coefficient of variation '0.1). This results from the field of view
(black disk) of one receptor being oriented slightly lower and thus
seeing more of the bright flickering light reflected from the sandy
bottom. The relative light intensities incident on these receptors
can be seen in Fig. 1c. Because transduction noise produces
'10% of the variation in a spike train for steady illumination,
higher variability results from underwater lighting. Hence, both
neural and environmental noise limit the detectability of these
behaviorally relevant objects.
The good correspondence between theory and experiment
in Fig. 1 not only confirms the accuracy of the model but also
high- and low-contrast objects. Note that as the animal moved
past the black object the firing rate of the optic nerve fiber
initially decreased and then briefly increased (Fig. 1 e and f,
Left); whereas, the firing rate to the gray object increased in a
quasi-periodic fashion because of flickering light in the un-
derwater scene (Fig. 1 e and f, Right). Although single optic
nerve fibers do not respond in the same way to the high- and
low-contrast objects in the ocean, male horseshoe crabs detect
both objects about equally well.
Responses of Arrays of Optic Nerve Fibers. Clues about how
horseshoe crabs detect other horseshoe crabs and objects
resembling them (Fig. 1b) can be found in the large arrays of
optic nerve responses, or neural images, computed by the
model eye (Fig. 2b). Each pixel of a neural image corresponds
to an ommatidium in the eye, and its gray level gives the optic
nerve firing rate computed for that ommatidium at an instant
in time. Inspection of the neural images in rows 1 and 3 of Fig.
2b reveals small clusters of ommatidia responding with tran-
sient increases (white pixels) and decreases (black pixels) in
firing rate to images of crab-size objects moving across the
visual field. Note that the eye encodes both moving high-
contrast (row 1) and low-contrast (row 3) objects with a
characteristic spatial pattern of optic nerve activity, even
though the responses of single nerve fibers to the objects differ
(Fig. 1 d and e). Clusters of ommatidia responding together at
high rates (light regions) lie next to clusters responding at low
rates (dark regions). In laboratory simulations the light and
dark regions merge together for smaller objects and separate
for larger ones, reducing the modulation of optic nerve re-
sponses in both cases (18). In the lower part of the visual field,
flickering light reflected off the sand modulates optic nerve
responses partially obscuring those to the crab-size objects.
FIG. 3. Temporal integration of optic-nerve responses by neurons
in the first synaptic layer of the brain. The gray and black symbols plot
the signal-to-noise ratio (SNR) of simulated responses of brain
neurons to moving images of the gray and black objects in Fig. 2,
respectively, as a function of synaptic integration time. SNR is defined
as the peak-to-peak response modulation of neurons viewing the
objects relative to twice the standard deviation of response modula-
tions of the rest of the network.
590
 Neurobiology: Passaglia et al.
Processing Patterns of Optic Nerve Activity in the Brain. How
might neurons in the brain distinguish responses to moving
crab-size objects from those evoked by other dynamic stimuli? A
possible answer is suggested by a recent finding that central
neurons integrate optic nerve signals with synaptic time constants
on the order of 300500 ms (18). Integrating optic nerve re-
sponses in Fig. 2b with a time constant of 400 ms (average of eight
sequential neural images computed every 50 ms) yields neural
images of average brain activity in which responses to moving
crab-sized objects are enhanced relative to flicker in the fore-
ground (Fig. 2c). Profiles of simulated brain activity evoked by a
horizontal row of optic nerve fibers reveal a 132% modulation to
the moving black object [(18 impulsesys (ips)3 ips)y12 ips] and
an 87% modulation to the moving gray one [(19 ips8 ips)y13 ips]
(Fig. 2d). The greater response modulation to the black object
may correspond to the field observation that animals can see it
slightly better than the gray one (13). Integrating optic nerve
responses over periods longer or shorter than 400 ms yields less
robust responses to the moving crab-size objects (Fig. 3). An
integration time of 400 ms gives maximal signal-to-noise ratio as
defined by half of the peak-to-peak modulation of the response
to the object divided by the standard deviation of responses to
surrounding regions of the scene. Integrative mechanisms in the
brain thus appear well suited for sensing moving crab-size objects
in the neural images it receives from the eye.
The Importance of Motion. The neural images of stationary
objects dramatically illustrate the essential role of motion to
retinal coding. Whereas visual images of stationary objects
appear similar to those of moving objects (Fig. 2a), their neural
images do not. The objects are hardly recognizable in the
computed neural images of optic nerve activity (Fig. 2b) and
nearly invisible in the neural images of simulated brain activity
(Fig. 2c) when they are not moving in the animals visual field.
Indeed, response profiles in Fig. 2d reveal an amplitude of
modulation of just 27% [(20 ips14 ips)y20 ips)] for the black
object and 22% [(20 ips16 ips)y18 ips)] for the gray one.
Image motion thus produced over a 4-fold increase in gain for
the black (132%y27%) and gray (87%y22%) objects.
Seeing High-Contrast Objects. Strongly modulated responses
to moving images of high-contrast objects (Fig. 2b, row 1) are
understandable in terms of the eyes optical and neural proper-
ties. Coarse sampling by the ommatidia (14) and lateral inhibitory
interactions among them (7) reduce the eyes sensitivity to high
and low spatial frequencies, respectively, yielding a broadly tuned
spatial filter (26). The eye is maximally sensitive (Fig. 4a, shaded
FIG. 4. Spatiotemporal response properties of the lateral eye. (a)
Spatial transfer function. Gain of the response of a single optic nerve fiber
plotted as a function of the spatial frequency of a drifting sinusoidal
grating. Negative slope lines denote the frequency band corresponding to
crab-size objects at the range of distances which Limulus detect potential
mates in their natural habitat (0.251.4 m). (b) Temporal transfer
function. Response gain plotted as a function of the frequency of a
flickering spot that fills the field of view of the recorded ommatidium.
Positive slope lines denote the frequency band corresponding to animal
velocities over the range of distances of mate detection (0.56 Hz).
Negative slope lines give the frequency range of strobic lighting (26 Hz)
in the animals natural underwater environment.
Proc. Natl. Acad. Sci. USA 94 (1997) 12653
region) for spatial frequencies from 0.008 to 0.04 cyclesydegree,
which correspond to crab-size objects at distances of 0.25 to 1.4 m
where 98% of visual detection occurs (13). With regard to the
temporal response properties of the eye, the duration of photo-
receptor quantal responses to light (22, 33) and the dynamics of
self-inhibition (27) reduce its sensitivity to high and low temporal
frequencies, respectively, resulting in a tuned temporal filter (26,
34). The eye is highly sensitive to images of crab-size objects
moving within the animals visual range at about the speed of a
horseshoe crab (15 cmys) and up to twice these speeds for animals
passing one another in opposite directions (17). These speeds
roughly correspond to temporal frequencies from 0.5 to 6 cyclesys
for which response gains exceed one (Fig. 4b, region shaded with
negative slope lines), underscoring the eyes high sensitivity to
moving objects. The spatiotemporal properties of the eye can
readily account for the animals ability to see high-contrast
objects but not low-contrast ones.
Seeing Low-Contrast Objects. Natural fluctuations of under-
water lighting enhance the visibility of low-contrast objects (35).
Surface waves, acting as dynamic lenses, create beams of light that
strobe the underwater scene at frequencies of '26Hz for which
the lateral eye is maximally sensitive (Fig. 4b, region shaded with
positive slope lines; ref. 36). Such wave-induced flicker is a
prominent feature of the shallow waters where Limulus mate and
persists over a range of lighting conditions (37), which is consis-
tent with the animals ability to find mates day and night (1013).
Behavioral studies in mating areas show that strobic light in-
creases the probability that male Limulus turn toward and
approach the low-contrast gray object (38). Model computations
show that the strobic light reflects off low-contrast objects,
enhancing their neural images as the objects move across the
animals visual field (Fig. 2b, compare images in rows 3 and 4). We
note that strobic light also evokes coherent bursts of nerve
impulses from neighboring ommatidia in the absence of relative
motion. How then does relative motion of objects enhance their
neural image? Motion of an image across the visual field activates
receptors along its path (Fig. 2b, row 3, white pixels) leaving
FIG. 5. Comparison of the visual performance of horseshoe crabs
to that predicted from the computational model. Black symbols replot
from a behavioral study (13) the probability of crabs turning toward
a black crab-size object as a function of the distance from the object.
Gray symbols plot the visual performance of an ideal horseshoe crab
predicted from optic nerve responses computed by the model. We
assume that the ideal horseshoe crab moved past the object at an
average speed of 15 cmys and water turbidity decreased the objects
contrast with distance (see Materials and Methods). We use signal
detection theory to estimate the probability of detecting the computed
optic nerve responses in the presence of noise. A coefficient of
variation of 0.1 in firing rate noise and a threshold for detection
permitting a 1% false-alarm rate provided the best description of the
measured visual performance. The higher coefficient of variation
measured in the field (0.2) suggests that the decision maker in the crab
brain pools the activities of a small cluster of optic nerve fibers ('4),
which is consistent with the number of ommatidia viewing the object
at behavioral threshold (13, 14). Significantly different estimates of
visual performance require at least a 10-fold change in detection
threshold or a 2-fold change in noise level.
591
12654 Neurobiology: Passaglia et al.
adapted ones in its wake (Fig. 2b, row 3, dark pixels). Without
motion of images across its receptor surface, the eye provides no
cues for distinguishing flickering light reflected off the object
from that reflected off the sand, making these parts of the scene
appear continuous (Fig. 2, row 4). In short, no visual cues
distinguish object from foreground in the absence of relative
motion. Subsequent integration by the brain further enhances the
neural images of moving objects (Fig. 2c) because the integration
time of brain neurons matches the velocity range of horseshoe
crabs and not that of the fast-moving flicker. Flicker-induced
responses (26 Hz) are shorter than the integration time of 400
ms whereas object motion across receptors is longer ('2 s),
yielding a substantial response envelope (Fig. 2 c and d). Image
motion thus binds together the activities of neighboring clusters
of ommatidia in space and time, sending robust signals to the
brain about potential mates.
CONCLUSION
Adrian and Hartline were the first to study neural coding in
single sensory fibers. They reported that the mean firing rate
of nerve impulses encoded various elementary features of
stimuli to proprioreceptors (39) and visual receptors (40).
Later studies in the auditory systems showed that the precise
time of occurrence of nerve impulses transmits information
about sound frequencies and their location in space (41, 42).
Similar claims have been made recently for the visual systems
of salamanders and rabbits (43). Such a timing code appears
not to have an important role in Limulus vision because
neurons in the brain integrate over periods of 250500 ms
trains of optic nerve impulses with interspike intervals ranging
from 25 to 200 ms (18). As a consequence of strobic under-
water lighting, spatiotemporal properties of the eye, and
temporal properties of brain cells, objects having the size,
contrast, and motion of horseshoe crabs are not distinguished
by unique neural codes. Low-pass integrative properties of
brain neurons filter out rapid changes, yielding response
envelopes that represent coherent changes in the mean firing
rates among ommatidia viewing a moving object. Fig. 5 shows
that estimates from such a rate code can account reasonably
well for the visual performance of horseshoe crabs.
The Limulus eye processes visual information with excitatory
and inhibitory mechanisms similar to those found in the first
synaptic layer, or outer plexiform layer, of the vertebrate retina.
Both retinas integrate photoreceptor signals with lateral inhibi-
tory mechanisms, but the vertebrate retina adds a second synaptic
layer of processing, the inner plexiform layer, before transmitting
optic nerve signals to the brain (1). With two layers of neural
processing, the frogs eye, for example, can extract specific
features of its visual world such as small, moving dark objects
about the size of a fly and transmit information about them to the
brain (4, 44). The integrative mechanisms of the Limulus eye also
extract information about moving objects, encoding those resem-
bling a horseshoe crab with characteristic patterns of coherent
activity among clusters of optic nerve fibers. Neurons in the first
synaptic layer of the brain integrate optic nerve signals, further
enhancing the neural images of potential mates.
We thank A. Meigs, R. Mitchell, A. Wixson, S. Dodge, M.E.
Kelly-Manglapus, D. Porcello, L. Sher, E. Kaplan, D. Samber, and M.
Powers for their help. This work was supported by the National
Institutes of Health (MH 49741 and EY00667) and the National
Science Foundation (IBN9696208).
1. Dowling, J. E. (1987) The Retina: An Approachable Part of the
Brain (Harvard Univ. Press, Cambridge, MA).
2. Rieke, F., Warland, D., DeRuyter van Steveninck, R. R. &
Bialek, W. (1997) Spikes: Exploring the Neural Code (MIT Press,
Cambridge, MA).
3. Bialek, W. & Owen, W. G. (1990) Biophys. J. 58, 12271233.
Proc. Natl. Acad. Sci. USA 94 (1997)
4. Lettvin, Y. L., Maturana, H. R., McColluch, W. S. & Pitts, W. H.
(1959) Proc. Inst. Radio Eng. 47, 19401951.
5. Levine, J. S. & MacNichol, E. F. (1979) Sens. Processes 3, 95131.
6. Laughlin, S. B. (1996) Vision Res. 36, 15291541.
7. Hartline, H. K. & Ratliff, F. (1957) J. Gen. Physiol. 40, 357376.
8. Hartline, H. K. & Ratliff, F. (1958) J. Gen. Physiol. 41, 10491066.
9. Barlow, R. B., Prakash, R. & Solessio, E. (1993) Amer. Zool. 33,
6678.
10. Barlow, R. B., Ireland, L. C. & Kass, L. (1982) Nature (London)
296, 6566.
11. Barlow, R. B., Powers, M. K., Howard, H. & Kass, L. (1986) Biol.
Bull. Woods Hole, Mass. 171, 310329.
12. Powers, M. K., Barlow, R. B. & Kass, L. (1991) Visual Neurosci.
7, 179189.
13. Herzog, E. D., Powers, M. K. & Barlow, R. B. (1996) Visual
Neurosci. 13, 3141.
14. Herzog, E. D. & Barlow, R. B. (1992) Visual Neurosci. 9, 571580.
15. Ratliff, F., ed. (1974) Studies on Excitation and Inhibition in the
Retina (Rockefeller Univ. Press, New York).
16. Knight, B. W., Toyoda, J. & Dodge, F. A. (1970) J. Gen. Physiol.
56, 421437.
17. Herzog, E. D. (1994) Dissertation (Syracuse University, New
York).
18. Passaglia, C. L. (1997) Dissertation (Syracuse University, New
York).
19. Barlow, R. B., Chamberlain, S. C. & Levinson, J. Z. (1980)
Science 210, 10371039.
20. Dodge, F. A., Knight, B. W. & Toyoda, J. I. (1968) Science 232,
15431545.
21. Wong, F., Knight, B. W. & Dodge, F. A. (1980) J. Gen. Physiol.
76, 517537.
22. Fuortes, M. G. F. & Hodgkin, A. L. (1964) J. Physiol. 172,
239263.
23. Purple, R. L. & Dodge, F. A. (1965) Cold Spring Harbor Symp.
Quant. Biol. 30, 529537.
24. Ratliff, F., Hartline, H. K. & Miller, W. H. (1963) J. Opt. Soc. Am.
53, 110120.
25. Barlow, R. B. (1969) J. Gen. Physiol. 54, 383396.
26. Brodie, S. E., Knight, B. W. & Ratliff, F. (1978) J. Gen. Physiol.
72, 167202.
27. Stevens, C. F. (1964) Dissertation (The Rockefeller University,
New York).
28. Smith, T. G., Stell, W. K., Brown, J. E., Freeman, J. A. & Murray,
G. C. (1968) Science 162, 456458.
29. Fohlmeister, J. F., Poppele, R. E. & Purple, R. L. (1977) J. Gen.
Physiol. 69, 849877.
30. Barlow, R. B. & Kaplan, E. H. (1977) J. Gen. Physiol. 69, 203220.
31. Laughlin, S. B. (1981) in Comparative Physiology and Evolution of
Vision in Invertebrates: Invertebrate Visual Centers and Behavior,
ed. Autrum, H. (Springer, New York), pp. 133280.
32. Victor, J. D., Shapley, R. M. & Knight, B. W. (1977) Proc. Natl.
Acad. Sci. USA 74, 30683072.
33. Kaplan, E., Barlow, R. B., Renninger, G. & Purpura, K. (1990)
J. Gen. Physiol. 96, 665685.
34. Batra, R. & Barlow, R. B. (1990) J. Gen. Physiol. 95, 229244.
35. McFarland, W. N. & Loew, E. R. (1983) Environ. Biol. Fishes 8,
173183.
36. Passaglia, C. L., Dodge, F. A. & Barlow, R. B. (1995) Biol. Bull.
Woods Hole, Mass. 189, 213215.
37. Loew, E. R. & McFarland, W. N. (1990) in The Visual System of
Fish, eds. Douglas, R. H. & Djamgoz, M. B. A. (Chapman and
Hall, London), pp. 143.
38. Passaglia, C. L., McSweeney, M., Stewart, K., Kim, E., Mole, E.,
Powers, M. & Barlow, R. B. (1997) Biol. Bull. Woods Hole, Mass.,
in press.
39. Adrian, E. D. (1928) The Basis of Sensation: The Action of the
Sense Organs (Norton, New York).
40. Hartline, H. K. & Graham, C. H. (1932) J. Cell Comp. Physiol. 1,
227295.
41. Kiang, N. Y.-S., Watanabe, T., Thomas, E. C. & Clark, L. F.
(1965) Discharge Patterns of Single Fibers in the Cats Auditory
Nerve (MIT Press, Cambridge, MA).
42. Carr, C. E. & Konishi, M. (1990) J. Neurosci. 10, 32273246.
43. Berry, M., Warland, D. & Meister, M. (1997) Proc. Natl. Acad.
Sci. USA 94, 54115416.
44. Barlow, H. B. (1953) J. Physiol. 119, 6988.
592
Cell-Based Model of the Limulus Lateral Eye
Christopher L. Passaglia, Frederick A. Dodge and Robert B. Barlow
J Neurophysiol 80:1800-1815, 1998.

You might find this additional info useful...

This article cites 51 articles, 37 of which can be accessed free at:

http://jn.physiology.org/content/80/4/1800.full.html#ref-list-1

This article has been cited by 3 other HighWire hosted articles

Growth, Visual Field, and Resolution in the Juvenile Limulus Lateral Eye
S. Meadors, C. McGuiness, F. A. Dodge and R. B. Barlow
Biol Bull, October , 2001; 201 (2): 272-274.
[Full Text] [PDF]

Limulus Vision in the Marine Environment

Robert B. Barlow, James M. Hitt and Frederick A. Dodge
Biol Bull, April , 2001; 200 (2): 169-176.
[Abstract] [Full Text] [PDF]

Outputs of Radula Mechanoafferent Neurons inAplysia are Modulated by Motor Neurons,

Downloaded from jn.physiology.org on November 23, 2011

Interneurons, and Sensory Neurons
Steven C. Rosen, Mark W. Miller, Elizabeth C. Cropper and Irving Kupfermann
J Neurophysiol, March 1, 2000; 83 (3): 1621-1636.
[Abstract] [Full Text] [PDF]

Updated information and services including high resolution figures, can be found at:
http://jn.physiology.org/content/80/4/1800.full.html

Additional material and information about Journal of Neurophysiology can be found at:
http://www.the-aps.org/publications/jn

This infomation is current as of November 23, 2011.

Journal of Neurophysiology publishes original articles on the function of the nervous system. It is published 12 times a year (monthly)
by the American Physiological Society, 9650 Rockville Pike, Bethesda MD 20814-3991. Copyright 1998 by the American
Physiological Society. ISSN: 0022-3077, ESSN: 1522-1598. Visit our website at http://www.the-aps.org/.
593
Cell-Based Model of the Limulus Lateral Eye
CHRISTOPHER L. PASSAGLIA, FREDERICK A. DODGE, AND ROBERT B. BARLOW
Departments of Ophthalmology and Physiology, Center for Vision Research, State University of New York Health
Science Center, Syracuse, New York 13210; and Marine Biological Laboratory, Woods Hole, Massachusetts 02543
Passaglia, Christopher L., Frederick A. Dodge, and Robert B.
Barlow. Cell-based model of the Limulus lateral eye. J. Neurophys-
iol. 80: 18001815, 1998. We present a cell-based model of the
Limulus lateral eye that computes the eyes input to the brain in
response to any specified scene. Based on the results of extensive
physiological studies, the model simulates the optical sampling of
visual space by the array of retinal receptors (ommatidia), the
transduction of light into receptor potentials, the integration of
excitatory and inhibitory signals into generator potentials, and the
conversion of generator potentials into trains of optic nerve im-
pulses. By simulating these processes at the cellular level, model
ommatidia can reproduce response variability resulting from noise
inherent in the stimulus and the eye itself, and they can adapt to
changes in light intensity over a wide operating range. Programmed
with these realistic properties, the model eye computes the simulta-
neous activity of its ensemble of optic nerve fibers, allowing us to
explore the retinal code that mediates the visually guided behavior
of the animal in its natural habitat. We assess the accuracy of
model predictions by comparing the response recorded from a
single optic nerve fiber to that computed by the model for the
corresponding receptor. Correlation coefficients between recorded
and computed responses were typically 95% under laboratory
conditions. Parametric analyses of the model together with optic
nerve recordings show that animal-to-animal variation in the opti-
cal and neural properties of the eye do not alter significantly its
response to objects having the size and speed of horseshoe crabs.
The eye appears robustly designed for encoding behaviorally im-
portant visual stimuli. Simulations with the cell-based model pro-
vide insights about the design of the Limulus eye and its encoding
of the animals visual world.
INTRODUCTION
What information does the eye transmit to the brain when
an animal sees? Pioneering attempts to answer this question
by recording from single optic nerve fibers yielded important
discoveries about how the retina processes information. Per-
haps the most fundamental one is that a retinal ganglion cell
integrates visual signals over a limited region of space,
termed its receptive field (Barlow 1953; Hartline 1938; photon responses, termed quantum bumps, sum to produce
Kuffler 1953). Much has since been learned about receptive
fields and other integrative properties of retinal neurons by
studying their physiology, pharmacology, and synaptic con-
nectivity (review: Dowling 1987). It is difficult, however,
to infer from such single-cell studies the information trans-
mitted by arrays of optic nerve fibers to the brain about the
complex patterns of illumination animals encounter in their
natural habitat.
The direct approach for studying retinal coding of natural
scenes would be to record simultaneously from large num-
bers of optic nerve fibers. Multielectrode arrays (Meister et
al. 1994) and voltage-sensitive dyes (Wong et al. 1995) have
1800 0022-3077/98 $5.00 Copyright q 1998 The American Physiological Society
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
recently been used to study patterns of activity generated by
ensembles of retinal neurons, but neither technique has yet
proven practical for use with behaving animals. An alterna-
tive approach is to construct a realistic computational model
of the eye from anatomic and physiological data. Because
of their relative simplicity, the eyes of lower vertebrates and
invertebrates appear to offer the best opportunities (Teeters
et al. 1997; Werblin 1991). The lateral eye of the horseshoe
crab, Limulus polyphemus, is a particularly attractive model
Downloaded from jn.physiology.org on November 23, 2011
system because it processes visual information with rela-
tively few retinal receptors using integrative mechanisms
shared by many higher animals (Barlow 1969; Ratliff 1974),
and it extracts sufficient information from underwater scenes
to enable horseshoe crabs to find mates (Barlow et al. 1982).
Precise measurements of the animals visual performance
(Herzog et al. 1996; Powers et al. 1991) guide our studies
and provide important tests of model predictions.
Quantitative analyses of the Limulus eye began with the
pioneering studies of Hartline and Ratliff (1957, 1958).
Building on the discovery of lateral inhibition in this eye
(Hartline 1949), they formulated a linear model of the eyes
integrative mechanisms and confirmed its predictions with
recordings of steady-state responses from single optic nerve
fibers (Ratliff and Hartline 1959). Barlow and Quarles
(1975) modified the Hartline-Ratliff formulation to include
an essential nonlinearity (Barlow and Lange 1974) and
showed that the modified formulation accurately predicted
the eyes spatial Mach-band response patterns. Knight et al.
(1970) and Ratliff et al. (1974) further modified the Hartline-
Ratliff formulation to extend it into the temporal domain.
Coleman and Renninger (1974, 1978) then used the model
to investigate the oscillatory properties of the inhibitory net-
work. Models were also developed to simulate excitatory
processes of the retina. Fuortes and Hodgkin (1964) em-
ployed a cascade of linear filters to model the time course
of excitatory signals generated by photoreceptor cells. Dodge
et al. (1968) extended their model, proposing that discrete
the excitatory signals and that light adaptation acts by modu-
lating the amplitude of these bumps. Wong et al. (1980) and
Wong and Knight (1980) later formalized this idea into the
adapting bump model of phototransduction. Finally, Bro-
die et al. (1978b) developed a comprehensive model of the
eye based on its spatial and temporal transfer functions and
with this model predicted the responses of single optic nerve
fibers to moving stimuli (Brodie et al. 1978a). Each of these
studies yielded important insights about properties of the
lateral eye and its response to light under specific conditions.
Our goal is to understand how the eye encodes visual
information in the animals natural habitat. An appropriate
neupa LP-Neurophys 594
 CELL-BASED MODEL OF THE LIMULUS EYE 1801

model for this purpose must incorporate the excitatory and

inhibitory mechanisms of the eye, as existing models have
done, as well as simulate the optical properties of the eye,
the mechanisms of photoreceptor adaptation, and the noise
inherent in the eye and light stimulus. For these reasons, we
constructed a cell-based model of the Limulus lateral eye
that has all of these features. It treats the retina as an array
of noisy neurons that sample visual space at discrete loca-
tions and adapt to changes in ambient illumination. This
configuration of the model enables us to examine issues of
retinal coding such as the detection of signals in the presence
of neural noise, the signaling capacity of single and multiple
neurons, the robustness of putative codes under diverse light-
ing conditions, and the role of different retinal mechanisms
in visually guided behavior.
We develop here the cell-based model from its experimen-
tal foundations, compute with it the eyes response to crab-
like stimulus patterns, evaluate its accuracy by comparing
computed responses to those recorded from single optic
nerve fibers, and explore its sensitivity to variations in pa-

Downloaded from jn.physiology.org on November 23, 2011

rameters. Model simulations yield insights about the design
of the eye and its ability to detect behaviorally important
stimuli in the animals natural habitat.

CELL-BASED MODEL OF THE EYE

The Limulus lateral eye views the world with a hexagonally

packed array of 1,000 retinal receptors, or ommatidia (Fig. 1A).
As shown schematically in Fig. 1B, the cylindrical corneal lens (l)
of each ommatidium focuses incident light through an aperture (a)
formed by the processes of pigment cells (p) onto the photosensi-
tive rhabdomeres (b) of 1012 retinular cells (r). Each retinular
cell transduces the incident light into a photocurrent that flows
passively into the eccentric cell (e) through gap junctions in its
dendrite. The individual photocurrents of retinular cells sum to-
gether in the eccentric cell forming a depolarizing receptor poten-
tial that is readily recorded with a microelectrode impaling the
soma of the eccentric cell. The summed excitatory photocurrent
propagates passively through the soma and down the axon to the
spike-generation site (x). Self- and lateral-inhibitory currents (i)
triggered by the spikes of the eccentric cell itself and those of its
neighbors also propagate to the spike-generation site via a dense
neuropil of synaptic connections. There the excitatory and inhibi-
tory currents sum to produce a generator potential that the spike
generator encodes as a train of action potentials. The spike train
propagates backward to the soma and along axon collaterals to FIG. 1. A: Limulus lateral eye. Dark spots in photograph are individual
exert self- and lateral inhibition and forward along the optic nerve ommatidia. Width of eye 1 cm. B: schematic drawing of an ommatidium
to the brain to mediate behavior. in the lateral eye. l, lens; a, aperture; b, rhabdomere; r, retinular cell; p,
The block diagram of Fig. 1C outlines how optical and neural pigment cell; e, eccentric cell; i, synaptic sites of self and lateral inhibition;
mechanisms of an ommatidium combine in our cell-based model x, site of spike generation. Diameter of ommatidium 250 mm. C: func-
tional diagram of optical and neural mechanisms operating in the lateral
of the Limulus eye. The model consists of a network of such
eye that transform visual scenes into patterns of optic nerve activity, or
ommatidial elements linked together by mutually inhibitory con- neural images.
nections. Although the model ommatidia are individually program-
mable, we assigned identical properties to them all using values
derived from empiric measurements of intact eyes. Although these edge of both the field of view and direction of view of each omma-
properties change with time of day as a result of efferent activity tidium.
generated by a circadian clock located in the brain (review: Barlow The direction of view of an ommatidium is determined by the
et al. 1989), all model simulations presented here apply to lateral orientation of its optic axis. It is convenient to express the orienta-
eyes in their daytime state. tion of an optic axis by its angle of divergence from the optic axis
of an adjacent ommatidium. For horseshoe crabs indigenous to
Optical transformation of visual scenes North Atlantic waters, such interommatidial angles average 67
but vary systematically such that the anteroventral region of the
The multifaceted compound eye transforms visual scenes into eye has minimal values, i.e., maximal resolution (Herzog and Bar-
points of light incident on its array of ommatidia. An accurate low 1992; Weiner and Chamberlain 1993). Because our stimulus
description of how the array samples visual space requires knowl- display illuminated about one-quarter of the eyes visual field, we

/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11 neupa LP-Neurophys 595

1802 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW
limited the size of our model to a 16 1 16 array of ommatidia in
the central region of the retina. Anatomic and physiological studies
(Fig. 2A) show that interommatidial angles of adjacent columns of
this array are roughly constant, but those of adjacent rows vary
such that the eye samples visual space more densely below the
horizon (Barlow et al. 1984; Dodge and Kaplan 1977; Herzog and
Barlow 1992; Weiner and Chamberlain 1993). We fit the diver-
gence of optic axes with the following empiric equations that map
the array of optic axes from retinal coordinates {i,j} to angular
coordinates {ui ,fj}
ui 6i (1)
fj 3j / 0.15j2 / 0.01j3 (2)
where ui is the angle of azimuth (in degrees) of column i relative
to that of the middle column of the array and fj is the angle of
elevation (in degrees) of row j relative to that of the row viewing
the horizon. From these relationships, we generate the sampling
mosaic of ommatidia shown in Fig. 2B.
The field of view of an ommatidium can be conveniently ex-
pressed by its acceptance angle, Dr, which is the angular width
of its field at half-maximal sensitivity. Acceptance angles are typi-
cally 67 during the day (Barlow et al. 1980). At night, efferent
signals from the circadian clock in the brain change the structure
of ommatidia that result in a doubling of their acceptance angles.
We approximate the light-collecting ability of ommatidia with a
normalized Gaussian point-spread function A(r) of spatial scale sa
S D
1 r2
A(r) 2
exp 0 2 (3)
2psa 2sa
where r is the angle (in degrees) of incidence with respect to the
optic axis. Because no evidence exists for regional differences in
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
the eye, all model ommatidia are assigned the same acceptance
angle (Dr 2.35sa).
We specify the relative light intensities In(t) incident at time t
on all n ommatidia of the model by first convolving individual
frames of the video input with the spread function A(r). We then
extract from the blurred images the pixel values at points of inter-
section {ui ,fj} with the optic axes of corresponding ommatidia
{i,j} in the crab eye as shown in Fig. 2B. Normalizing the individual
pixel values of an image by their collective mean over all images
yields the relative light intensities In(t) incident on the ommatidial
array at discrete moments in time.
Retinal processing of visual scenes
The retina processes visual scenes using excitatory and inhibi-
tory mechanisms of phototransduction, self inhibition, and lateral
inhibition (Fig. 1C). We incorporate these mechanisms in an equiv-
alent circuit of an ommatidium, shown in Fig. 3. This is the simplest
circuit that explicitly accounts for the electrical separation between
the sites of phototransduction in retinular cells and spike generation
in eccentric cells (Purple and Dodge 1965). We base the two-
compartment model on extensive intracellular recordings from the
Downloaded from jn.physiology.org on November 23, 2011
soma of eccentric cells: the integrators of excitation and inhibition
in the retina.
We model phototransduction of incident light intensity by the
nth ommatidium with a time-dependent excitatory conductance
gEn(t) having an equilibrium potential VE that yields the receptor
potential Sn(t) recorded from the eccentric-cell soma. This light-
sensitive conductance acts in parallel with the passive electrical
properties of the soma represented by resistance RS and capacitance
CS . The somatic compartment is coupled to an axonal compartment
by resistance RC . The coupling resistance represents the short
length of passive membrane separating the soma from the synaptic
FIG. 2. A: angle of azimuth (left) and
elevation (right) of optic axes of adjacent
columns and rows of ommatidia in the cen-
ter of the eye. s, physiological measure-
ments of Dodge and Kaplan (1977); , ana-
tomic measurements by Herzog and Barlow
(1992). Solid lines give the empiric fits
used in the model. B: schematic illustration
of how the ommatidial array samples an
underwater scene based on these measure-
ments. j, show where ommatidial axes in-
tersect the scene. Horizontally and verti-
cally elongated ellipses denote a row and
column of ommatidia, respectively.
neupa LP-Neurophys 596
 CELL-BASED MODEL OF THE LIMULUS EYE
FIG. 3. Electrical equivalent circuit of an ommatidium. Our model of
the Limulus lateral eye links many such cell-based circuits together into a
neural network. gEn(t), light-modulated excitatory conductance of the nth
ommatidium; VE , equilibrium potential of excitatory conductance (60 mV);
RS and CS , resistance (20.2 MV) and capacitance (0.002 mF) of soma; vSn(t),
receptor potential of somatic compartment; RC , coupling resistance (5.2
MV); RA and CA , resistance (8.0 MV) and capacitance (0.001 mF) of axon
collaterals; gSIn(t), self-inhibitory conductance; gLIn(t), lateral-inhibitory con-
ductance; VI , equilibrium potential of inhibitory conductances (015 mV);
C, electrogenic pump of constant current (00.25 nA); An(t), generator
potential of axonal compartment; fn(t), train of optic nerve impulses. Voltage
terms are defined relative to an arbitrary resting potential.
neuropil region where the axon generates nerve impulses. The
passive electrical properties of the numerous axon collaterals form-
ing the synaptic neuropil are lumped into resistance RA and capaci-
tance CA that operate in parallel with two variable inhibitory con-
ductances, gSIn(t) and gLIn(t), having an equilibrium potential VI .
Self-inhibitory (SI) inputs from the eccentric cell and lateral inhibi-
tory (LI) inputs from its neighbors modulate gSIn(t) and gLIn(t),
respectively, to inhibit the eccentric cell response after it or its
neighbors fires a spike. A putative electrogenic pump near the
spike-generation site provides a source of constant current C that
depends on the mean firing rate (Fohlmeister et al. 1977; Smith et
al. 1968). The combined action of all these elements yields a gener-
ator potential An(t) in the axonal compartment. A leaky integrate-
and-fire encoder (Knight 1972) then converts the generator poten-
tial into a train of optic nerve impulses, fn(t), which we represent
as sequences of delta functions at their times of occurrence.
The two-compartment model of an ommatidium yields the fol-
lowing pair of differential equations for computing the receptor
potential in the soma Sn(t) and generator potential in the axon
An(t)
dSn An 0 Sn Sn
CS 0 0 gEn(t)r(Sn 0 VE)
dt RC RS
d An Sn 0 An An
CA 0 0 [gSIn(t) / gLIn(t)]r(An 0 VI) / C
dt RC RA
where the voltage terms are defined relative to an arbitrary resting
potential. We next describe the processes of phototransduction and
inhibition that define the three time-varying conductances gEn(t),
gSIn(t), and gLIn(t) necessary for solving Eqs. 4 and 5.
PHOTOTRANSDUCTION gEn(t). In phototransduction, a rhodopsin
molecule activated by photon absorption triggers events that cause
a momentary increase in excitatory conductance and a discrete
change in membrane potential termed a quantum bump (Adolph
1964; Fuortes and Yeandle 1964; Yeandle 1957). Clearly detect-
able at low light levels, quantum bumps are Poisson distributed in
time with an average rate of occurrence that is directly proportional
to the incident light intensity. They are also similar in shape but
vary in amplitude. At higher light levels individual bumps summate
to form the receptor potential (Dodge et al. 1968). For such a
shot noise process, the mean and variance of the excitatory
conductance of a cell under steady illumination can be related
to the properties of individual bumps by the following pair of
equations
gV E lU rTraV
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
1803
[gE(t) 0 gV E]2 lU rTraV 2 (7)
where gV E is the mean excitatory conductance (mS), lV is the mean
bump rate (bumps/s), T is the effective bump duration (s), and aV
is the mean bump amplitude (mS). Bars indicate that these are
steady-state values of time-dependent variables.
In their analysis of the frequency response of the receptor poten-
tial to flickering light, Dodge and colleagues (1968) found that
bump duration decreases by a factor of 4 as light intensity increases
by a factor of 105. Because the mean bump rate lV is proportional
to mean light intensity, we can express this weak dependence of
bump duration on light level by the following empiric equation
T QlU 00.12, lU x 0 (8)
where the parameter Q sets the bump duration for a given experi-
ment. Dodge and colleagues (1968) also found, as have others
(Barlow and Kaplan 1977; Fuortes 1959; Hartline et al. 1952;
MacNichol 1956), that under steady-state conditions the steady-
state plateau of the receptor potential increases with the logarithm
of light intensity over an equally wide range. We can express
this logarithmic relationship as a dependence of mean excitatory
conductance gV E on mean bump rate lV using the following empiric
Downloaded from jn.physiology.org on November 23, 2011
equation
S
gV E 0.021rlog10 1 /
lU
1.4 D (9)
In relating the frequency response of the receptor potential and its
noise characteristics, Dodge and colleagues (1968) concluded that
increasing light intensity decreases the amplitude of quantum
bumps. Combining Eqs. 6, 8, and 9, we can express this dependence
of mean bump amplitude aV on mean bump rate lV as
S
0.021rlog10 1 /
lU
1.4 D
aV (10)
QlU 0.88
The modulation of mean bump amplitude by light constitutes the
photoreceptor mechanism of light adaptation known as the adapt-
ing bump model.
We present here a dynamic version of the adapting bump model
to convert time-dependent changes in incident light intensity into
changes in excitatory conductance. We model the adaptation of
(4)
quantum bumps as a dynamic interaction of two competing pro-
cesses, shrinkage and growth, which modulate the number of avail-
(5) able clusters of ion channels in the photoreceptor membrane. In
this model, a rhodopsin molecule excited by photon absorption
triggers a biochemical cascade that opens a cluster of ion channels
in the retinular cell membrane, thereby producing a bump. Other
rhodopsin molecules excited at the same time activate different
clusters of channels to generate other bumps. We define the number
of clusters g(t) available for activation at time t as
g(t) amax/an(t) (11)
where amax is the maximum bump amplitude (in conductance units,
mS) generated by opening all of the channels in the membrane and
an(t) is the expected bump amplitude at time t. Following activa-
tion, the cluster of channels immediately dissociates into smaller
clusters reducing the average amplitude of bumps generated by
subsequent photon absorptions. This shrinkage of bump amplitude,
however, is temporary because clusters can associate to form larger
clusters. As a result of the competition between bump shrinkage
and regrowth, there is at any given moment an exponential distribu-
tion of cluster sizes (Knight 1973) and bump amplitudes (Barlow
and Kaplan 1977).
We represent the transduction of light into bumps in the nth
(6) ommatidium by a Poisson shot-noise process Pln(t), which gener-
neupa LP-Neurophys 597
1804 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW

ates bumps of random amplitude an at random times of photon
absorption tn with a rate ln(t) that is a fixed percentage of the
incident light intensity In(t). Note that this is a nonhomogeneous
process because the rate changes as In(t) changes (Shanmugan and
Breipohl 1988). We select the amplitude of each bump an that
occurs at time t from an exponential distribution having an instanta-
neous mean value an(t), which constantly changes in response to
the incident light intensity In(t). That is, an(t) shrinks over time in
proportion to the fraction of the available clusters being activated.
We model bump shrinkage as
The output of the last stage gives the light-modulated excitatory
conductance gEn(t) of the ommatidium.
LATERAL INHIBITION gLIn(t). Excitatory photoresponses drive an
inhibitory network of eccentric cells whose properties have been
well described in the classic work of Hartline, Ratliff, and col-
leagues. By measuring the interactions among individual omma-
tidia, Hartline and Ratliff (1957, 1958) found that the steady-state
firing rate rV n of the nth ommatidium is equal to its uninhibited
response eV n minus inhibitory influences of its N 0 1 neighbors.
This relationship can be expressed by the following set of piece-

Shrinkage 0 F G
Pln(t)
gn(t)
where Pln(t) aP nrd(t 0 tO n)
i
(12)
wise linear equations

F N
rV n eV n 0 knmr(rV m 0 r onm)/ G /
,
nxm

H
where hats () indicate random variables. If, for example, excited
rhodopsin molecules activate one-half of the available clusters, z if z0
z/ (20)
then the bump amplitude shrinks toward one-half its previous value 0 otherwise n 1, . . , N
in the next time step. However, to maintain constant bump ampli-
tude under steady illumination, an(t) must also grow as
where superscript / is an operator that ensures nonnegative firing
Growth G[an(t)] (13) rates. The amount of inhibition exerted by neighboring ommatid-
ium m on n depends on its steady-state firing rate rV m , its threshold
We assume that the growth of bump amplitude G[an(t)] is a func-

Downloaded from jn.physiology.org on November 23, 2011

tion of only the current bump amplitude an(t) rather than the bump
rate ln(t), and determine its value from a look-up table of steady-
state growth rates that exactly compensate for the shrinkage rate.
That is, we set the growth and shrinkage rates equal and compute
the table by solving Eq. 12 using a range of mean bump amplitudes
aV whose values are specified by Eq. 10. Combining the processes
of bump shrinkage and growth, we represent the dynamics of adap-
tation of bump amplitude an(t) to time-varying illumination with
the following differential equation
dan
dt
0
an
amaxS D
rPln(t) / G[an]
We have not included in this formulation slow mechanisms of light
adaptation that change the amplitude of bumps on time scales of
tens of seconds (Biederman-Thorson and Thorson 1971) because
the stimuli in our experiments are comparatively short in duration.
We have also not included the large regenerative bumps character-
istic of the dark-adapted state (Barlow and Kaplan 1977) and the
nighttime state of the eye (Kaplan et al. 1990) because our experi-
ments are carried out during the day under ambient levels of illumi-
nation.
r onm to inhibit n, and its coefficient of inhibitory action knm on n.
The matrix knm represents the configuration of the inhibitory field
of the nth ommatidium. Although the inhibitory field contains non-
uniformities, they are small and do not significantly affect optic
nerve firing rates rV n computed in response to stationary patterns of
illumination (Barlow and Quarles 1975). For spatially uniform
stimuli, the uninhibited, eV n , and inhibited, rV n , responses of all om-
matidia are essentially the same and related to one another by a
constant of proportionality K / 1, where K is the sum of coeffi-
cients in the inhibitory weight matrix knm . Barlow and Lange (1974)
(14)
later modified Eq. 20 to include a nonlinear dependence of K on
the level of excitation.
We implement a time-dependent version of the original Hartline-
Ratliff formulation (Ratliff et al. 1963). We approximate the com-
plex waveform of lateral-inhibitory signals with a three-stage
Fuortes-Hodgkin model with each stage having the time constant
tLI . This description accounts for the relatively slow rise to peak
of the lateral-inhibitory potential (Coleman and Renninger 1974;
Knight et al. 1970; Ratliff et al. 1969). The three stages of lateral-
inhibitory conductance gLIn(t) yield the following differential equa-
tions
dl1n l1n M

We sum the individual quantum bumps to yield the excitatory

0 / knmGLI fm(t) (21)
dt tLI nxm
conductance gEn(t) using a four-stage version of the Fuortes and
Hodgkin (1964) bump integrator with each stage having the same dl2n l1n 0 l2n
(22)
time constant tb . Note that the duration of individual bumps T is dt tLI
6.4tb for a four-stage integrator (Wong and Knight 1980). Mea-
sured dispersions in the latencies of bumps are small (Wong et al. dl3n l2n 0 l3n
(23)
1980) and not included in the model. We drive the four stages dt tLI
of the Fuortes-Hodgkin model with the Poisson bump generator
gLIn(t) l3n(t) (24)
yielding the following set of differential equations
db1n b1n where GLI is the conversion factor from inhibitory coefficients to
0 / Pln(t) (15) inhibitory conductances, and fm(t) are the times of firing of model
dt tb
ommatidium m. The unitary inhibitory coefficients knm in our model
db2n b1n 0 b2n comprise a 256 1 256 matrix that is computed based on the func-
(16) tional dependence of k on the distance between ommatidia, given
dt tb
by a broad Gaussian function of scale si and sj in the horizontal
db3n b2n 0 b3n and vertical directions, respectively, that has a narrow Gaussian
(17) crater (Barlow 1969). We assume the ommatidial units are arranged
dt tb
in a square 16 1 16 array and compute each knm with the following
db4n b3n 0 b4n equation

F S D G
(18)
dt tb i2 j2
knm } KLI exp 0 2
0 2 0 exp(0i2 0 j2) (25)
si sj
gEn(t) b4n(t) (19)

/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11 neupa LP-Neurophys 598

 CELL-BASED MODEL OF THE LIMULUS EYE
where the horizontal separation between ommatidia n and m is
measured as i columns, and the vertical separation is measured as
j rows. After computing the matrix of coefficients, we rescale their
values so that the total inhibition converging on each unit is KLI .
For the simulation reported here, we assigned unit amplitude and
unit width to the central crater in Eq. 25 (Brodie et al. 1978b) and
scaled the broad function equally in the vertical and horizontal
directions (i.e., si sj sLI 4 ommatidia).
SELF INHIBITION gSIn(t). Ommatidia in the lateral eye inhibit not
only their neighbors but themselves as well. Every action potential
fired by an eccentric cell feeds back an inhibitory postsynaptic poten-
tial onto itself that reduces its likelihood of firing another spike in
the immediate future (Purple and Dodge 1966; Stevens 1964). Self-
inhibition was not explicitly included in the original Hartline-Ratliff
equations (1958), resulting in their overestimate of the strength of
lateral inhibition KLI . We treat the self-inhibitory conductance gSIn(t)
as a decaying exponential function having a strength KSI and time
constant tSI yielding the following differential equation
dgSIn gSIn
0 / GSI fn(t)
dt tSI
where GSI is the strength of self inhibition in terms of an equivalent
conductance and fn(t) are the times of firing of ommatidium n.
Encoding retinal signals with trains of optic
nerve impulses
Initial studies of the Limulus eye showed that the spike-firing
mechanism linearly converts voltage at the soma to frequency of
nerve impulses (Fuortes 1959; MacNichol 1956). The conversion
factor over the physiological range is about 1 impulse/s per milli-
volt (mV) of depolarization of the soma above a firing threshold
of 1 mV (Barlow and Kaplan 1977). More recent studies report
that the conversion factor may increase to 27 impulses/s/mV at
high levels of depolarization (15 mV) (Renninger et al. 1988).
Because inhibitory signals in the axon hillock are attenuated when
they reach recording sites in the soma, the conversion factor S
and threshold Vo of the voltage-to-frequency converter in the axon
hillock are undoubtedly 1 impulse/s/mV and 1 mV, respectively.
We use a simple integrate-and-fire model (Knight 1972) to en-
code the generator potential An(t) into a train of nerve impulses
fn(t). The form of this model is given by the following differential
equation
dFn
dt
Sr[An(t) 0 Vo] 0 FoV, where V H 1,
0
if Fn Fo
otherwise
where Fn(t) is an internal variable of the spike encoder and Fo is
its internal firing threshold that is set to yield a conversion rate of
1 impulse/mV. We adjust the sensitivity S of the encoder once
kinetic parameters have been set so that the rate at which Fn(t)
crosses Fo matches the recorded firing rate. Times of threshold
crossings give fn(t).
Setting initial conditions
Because all animals are not equally sensitive to light and daytime
light levels can vary over several orders of magnitude, we must
specify for each simulation the mean bump rate lV that sets the
operating point, or state of adaptation of an eye. Direct measure-
ment of lV requires precise information about the transmission of
light through the corneal apparatus, the absorption of photons by
the rhabdomere of photoreceptors, and the quantum efficiency of
excitation for absorbed photons. Because it is not possible to mea-
sure these quantities with precision, we cannot determine lV directly
from measurements of the absolute light intensity. Instead we esti-
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
1805
mate lV by taking advantage of the relationship between the state
of bump adaptation and the characteristics of retinal noise ex-
pressed in Eq. 7. Specifically, we assume that the variability in the
firing rate of optic nerve impulses under steady illumination reflects
noise generated by random fluctuations in the number and ampli-
tude of bumps (Kaplan and Barlow 1975). We then adjust the mean
bump rate lV until the coefficient of variation of firing rate computed
by the model matches that recorded from single nerve fibers of an
experimental eye in a given lighting environment.
To simulate phototransduction noise, we must generate two ran-
dom numbers for each bump: one to set the interval between bumps
and the other to set the amplitude of a bump. In our experiments,
mean bump rates lV typically range from 104 bumps/s in the
laboratory to 106 bumps/s outdoors. Simulation of such high
rates is a tremendous computational burden that we avoid by first
replacing the bump generator term Pln(t) in Eqs. 14 and 15 with
ln(t)ran(t) to compute a noise-free response. We then create bumps
at a lower rate and correspondingly greater amplitude without
changing the kinetics of bump adaptation and add the noise to the
noise-free response. This allows the model to simulate the dynamic
(26)
response of the receptor potential to time-varying stimuli in an
efficient and accurate manner over a wide range of light levels.
Downloaded from jn.physiology.org on November 23, 2011
In solving the differential equations of the model, we use a brief
time step (dt 0.2 ms) to achieve high resolution in times of spike
firings. This step is sufficiently short that time-dependent variables
can be accurately computed using a simple Euler method, but we
used a two-stage modified Euler method for solving the equivalent
circuit (Eqs. 4 and 5). The duration of time steps is also relevant
to the specification of model input, which for us is often a video
movie digitized at 20 frames/s. For such movies, we linearly inter-
polate successive frames to avoid sudden changes in light intensity.
In summary, our cell-based model of the Limulus lateral eye
combines previous models of retinal mechanisms in an equivalent
electrical circuit that converts a temporal sequence of optically
transformed images into arrays of optic nerve activities, or neural
images (Laughlin 1981; Passaglia et al. 1997), over a wide range
of daytime illumination conditions. Table 1 lists the values of
the 10 model parameters, and the legend briefly describes their
functional significance.
EXPERIMENTAL PROCEDURES
Animals
(27)
We performed experiments during the summer at the Marine
Biological Laboratory (Woods Hole, MA) and the winter at Syra-
cuse University (Syracuse, NY) on male horseshoe crabs measuring
1520 cm across the carapace. Female crabs were avoided because
their abundant supply of eggs impairs surgical isolation of optic
nerve fibers. We acquired animals from the MBL all year and from
the Gulf Specimens Company (Panacea, FL) during the winter.
Our measurements revealed no differences in the two populations
of crabs. Animals were maintained under natural lighting condi-
tions in ocean water in Woods Hole and circulating artificial seawa-
ter (Instant Ocean, Aquarium Systems, Eastlake, OH) in Syracuse.
They were fed fresh clams biweekly and used within 2 mo of
arrival.
Optic nerve recording
Details of the recording technique are described by Herzog et
al. (1993). In brief, we secured the animal to a rotatable wooden
platform and exposed its optic nerve by cutting a 2-cm circular
hole in the carapace anterior to the eye. The tongue of a nylon
recording chamber was then slid under the uncut nerve, and cotton
was placed in open slots to prevent fluid leakage. We filled the
chamber with saline (pH 7.4) and carefully removed the sheath
neupa LP-Neurophys 599
1806 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW

TABLE 1. Parameters of the cell-based model of the Limulus lateral eye for simulating optic nerve responses of three experimental
eyes and the standard eye

Parameter Units Effect Eye I Eye II Eye III Standard Eye SIg SIb

lU bumps/s * 50,000 50,000 150,000 50,000 0.1 0

Dr degrees c 4.7 5.4 6.1 6.1 1 1
tb s g 0.024 0.018 0.010 0.016 1 0
amax mS d 0.75 0.50 1 0.75 0.6 0.9
KLI a 4 4.5 4 4 0.5 0.4
sLI b 4 4 4 4 1 0.6
tLI s f 0.07 0.10 0.080 0.07 0.7 0.1
KSI e 2 2 3 2 0.3 0.8
tSI s d 0.14 0.20 0.16 0.20 0.4 0.7
S impulses/s/mV * 8.3 9.8 12.8 9.2 0 0

Letters refer to features of the spatiotemporal transfer function in Fig. 6 most affected by the parameter; asterisks indicate parameters having nonspecific
affects on the transfer function. Sensitivity indexes for drifting gratings (SIg) were calculated as the percent change in the affected feature divided by
the percent change in the value of the modified parameter (50%). Sensitivity indexes for moving crab-size bars (SIb) were calculated by integrating the
difference between the average response computed with the standard value of a parameter and a 50% increase in its value, and then normalizing the
result by that of the parameter having the maximal effect. lU , mean bump rate; Dr, acceptance angle; tb, bump time constant; amax, maximum bump
amplitude; KLI, strength of lateral inhibition; sLI, space scale of inhibitory field; tLI, time constant of lateral inhibition; KSI, strength of self inhibition;
tSI, time constant of self inhibition; S, sensitivity of spike encoder.

Downloaded from jn.physiology.org on November 23, 2011

encapsulating the nerve. Critical to surgical success was the expedi-
ent removal of this sheath because blood quickly clots inside the
vessel walls, starving tissue of oxygen and other nutrients. Using
fine probes, we gently teased free a single active fiber from the
optic nerve and guided it into the polished tip of a microsuction
electrode housed in the chamber. With practice, even the delicate
efferent fibers in the nerve remained active. The chamber was
sealed with a water-tight nylon plug, and a small tube carrying a
continuous supply of fresh Ringer was inserted through a hole in
the plug to maintain the stability of the recording over a long
period of time. We located the recorded ommatidium in the array
of retinal receptors by moving a fiber-optic lightpipe (70 mm diam)
along the corneal surface. We selected for analysis only those
ommatidia having optic axes within 10 and 307 from the horizontal
and vertical meridians of the eye, respectively. The wired ani-
mal was submerged in a glass-sided saltwater tank housed in a
light-proof, shielded cage.
Stimulus presentation and data acquisition
A fully automated system created the visual stimuli and recorded
the times of impulses of optic nerve fibers (VENUS, model 1010,
Neuroscientific, Farmingdale, NY, and Electronics Shop, The
Rockefeller University). The animal viewed the stimuli on a display
monitor (Model 608, Tektronics, Beaverton, OR) through the glass
window in the tank. The monitor was placed at a fixed distance
(4.5 or 9 cm) from the eye and positioned so that the optic axis of
the recorded neuron was aligned perpendicular with the center of
its screen (10 cm high, 13 cm wide). At such distances the screen
filled 75% of the inhibitory field of the recorded ommatidium
(Barlow 1969). Optic nerve signals were amplified (ISR Electronics
Shop, Syracuse University) and discriminated electronically
(World Precision Instruments, New Haven, CT). We discontinued
trials if the mean firing rate decreased more than 10%. Before data
collection, we calibrated the system (Pin-10DP photodiode, United
Detectors Technology, Hawthorne, CA) so that the output of the
monitor scaled linearly between 5 and 90% of its maximum
value. The system was not driven outside this range.
An experiment consisted of multiple trials beginning with a
minute of exposure to the mean light intensity (45 cd/m2) of the
stimulus display followed by a 15- to 30-min presentation of time-
varying stimulus patterns.
DRIFTING GRATINGS AND FLICKER. We measured the integ-
rative properties of the eye by drifting sinusoidal gratings of fixed
contrast (10%) across the monitor located 4.5 cm from the eye.
We presented the gratings at different spatial (0, 1, 2, 4, 8, 16, and
32 cycles/screen) and temporal (0.125, 0.25, 0.5, 1, and 2 cycles/
s) frequencies for 60 s each. Allowing the flat screen to approximate
a curved surface, the spatial frequencies are equivalent to 0.005,
0.008, 0.012, 0.021, 0.040, 0.079, and 0.157 cycles/deg, respec-
tively. This approximation holds for the center of the visual field
but underestimates the spatial frequency in the periphery. The sen-
sitivity of the eye to gratings drifting at higher temporal frequencies
(2 Hz) was not possible to measure because the recorded neuron
tended to synchronize its discharges to particular phases of the
periodic stimulus. We circumvented this problem by flickering both
a small 67 spot and the entire screen for 240 s with a sum of
multiple sinusoids (Victor et al. 1977). The wave numbers (2, 5,
11, 19, and 31 cycles) of the sinusoids were carefully chosen so
that harmonic and combination frequencies were nonoverlapping.
The repeat period (2, 4, and 8 s) of the sinusoids was selected to
adequately sample the entire transfer function, and their amplitudes
were set to yield about the same response modulation at the stimu-
lation frequencies (Brodie et al. 1978b).
We analyzed optic nerve responses by demodulating recorded
spike trains with a second-order low-pass filter having a time con-
stant of 10 ms and then sampling the resulting waveforms at 128
Hz. We estimated response amplitude using the method of least
squares with a fitting function composed of the mean, fundamental,
and second harmonic of the stimulus frequencies. Data were dis-
carded if the second harmonic was 20% of the fundamental,
indicating the presence of phase locking. We computed the gain
of the eye by normalizing the percent modulation of firing rate
with the contrast of the stimulus. Because the small spot stimulus
did not completely fill the excitatory center of the recorded neuron,
measurements using small spots were scaled to match the peak
gain measured using sinusoidal gratings.
Spike trains recorded in response to gratings of low spatiotempo-
ral frequencies were also used for analyzing stochastic fluctuations
in firing rate. We demodulated these trains by computing the recip-
rocal of the interval between spikes and then sampling the resulting
waveforms at 128 Hz. The random component was extracted by
subtracting stimulus-driven components, and its power spectrum
was estimated using the method of overlapping periodograms
(Welch 1967).
MOVING CRAB-SIZE OBJECTS. In addition to gratings and flicker,
we recorded responses of optic nerve fibers to a horizontally elon-
gated bar of fixed contrast (020 or 040% with respect to back-

/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11 neupa LP-Neurophys 600

 CELL-BASED MODEL OF THE LIMULUS EYE 1807

FIG. 4. Stages of analysis of model computations. A: firing rate recorded from an optic nerve fiber in response to a drifting
sinusoidal grating having a spatial frequency of 0.016 cycles/deg and temporal frequency of 1 Hz. B: firing rate computed
for an optic nerve fiber in the model eye in response to the same drifting grating. C: firing rate of an optic nerve fiber
predicted by the model in response to a simulated crablike object moving across the visual field. D: firing rate of an optic
nerve fiber subsequently recorded in response to the simulated crablike object. Top and bottom traces give responses of an
optic nerve fiber to multiple presentations of the stimulus and its average response, respectively.

ground) moving repeatedly for 90 s across the monitor placed 9 spatiotemporal frequencies, whereas eye II was the most

Downloaded from jn.physiology.org on November 23, 2011

cm from the eye. The moving bar is an abstraction of an adult
horseshoe crab. Its height (1.35 or 2.7 cm) and width (2.7 or 4.5
cm) corresponds to a crab located 60 or 100 cm away, and its
velocity (1, 2, 4, and 8 cm/s) simulates a crab moving below (6.6
cm/s), at (13.3 and 26.6 cm/s), and above (53.3 cm/s) its normal
swimming speed (Herzog 1994). We demodulated the recorded
spike train by computing the reciprocal of the interval between
spikes and sampling the resulting waveform at 128 Hz. Several
presentations were averaged to estimate the noise-free response of
the eye to the crablike objects. We then cross-correlated the average
recorded response with individual responses of the optic nerve
fiber and with the average response computed by the model to
quantitatively assess the accuracy of model predictions.
Model computations
We carried out model computations in four sequential stages.
Figure 4 illustrates sample records from each of the stages. First,
we measured an eyes spatiotemporal transfer function by re-
cording the discharge of a single optic nerve fiber in response to
drifting gratings (Fig. 4A, top) and flickering light. The eyes dy-
namic response was estimated by averaging the single fiber re-
sponses to repeated presentations of the stimulus (Fig. 4A, bottom),
and the eyes noise characteristics were estimated by power spec-
trum analysis of the difference between the average response and
individual ones for select stimuli (not shown). Second, we deter-
mined values for model parameters by matching average and differ-
ence records computed by the model for gratings and flicker with
those recorded from the nerve fiber (Fig. 4B). Third, we used the
parameter values to predict responses of a 16 1 16 array of optic
nerve fibers to a moving bar having the size and speed of an
adult horseshoe crab (Fig. 4C). And fourth, we compared responses
recorded from the single nerve fiber with those computed for
the corresponding neuron in the model for the same visual scenes
(Fig. 4D).
RESULTS
Experiments were performed on 12 lateral eyes in as many
animals. We present here complete analyses of model com-
putations for three of the eyes (I, 022196; II, 022696; III,
031296) whose integrative properties spanned the observed
range of variation among the horseshoe crabs used in this
study. Eyes I and III differed the most in sensitivity to high
responsive to low spatiotemporal frequencies. These experi-
ments were conducted with animals in the laboratory primar-
ily for the purpose of evaluating the model. Computational
studies of animals moving in their natural underwater habitat
are presented elsewhere (Passaglia et al. 1997).
Response properties of the eye
SPATIAL TRANSFER FUNCTION. The lateral eye is not equally
sensitive to all spatial frequencies. Rather, it exhibits maxi-
mal gain for spatial frequencies in the range of 0.010.03
cycles/deg (Fig. 5A). It exhibits such spatial tuning in re-
sponse to sinusoidal gratings drifting at rates below 12 Hz
and loses its tuning properties at higher temporal frequen-
cies. Above 2 Hz, the spatial transfer function flattens, and
the eye exhibits properties of a low-pass filter with a cutoff
frequency of 0.10 cycles/deg. Because of phase locking
of optic nerve impulses with the stimulus, we could not
reliably measure spatial transfer functions above 4 Hz with
drifting gratings. We conclude that the Limulus eye is
broadly tuned for relatively large, slow-moving objects.
Optical and neural mechanisms of the eye determine the
shape of the spatial transfer function. At low spatial frequen-
cies (0.01 cycles/deg), individual cycles of the sinusoidal
gratings more than fill the acceptance angle of an ommatid-
ium (4.577) maximally exciting the recorded neuron. How-
ever, the wide gratings also evoke lateral-inhibitory signals
of nearly equal magnitude from neighboring ommatidia that
oppose the excitatory signal, reducing the eyes response to
slowly changing stimuli. The strength of lateral inhibition
thus affects the gain of the eye at low spatiotemporal fre-
quencies (Fig. 6, arrow a). As spatial frequency increases,
the amplitude of the inhibitory signal evoked by the drifting
grating decreases because multiple cycles of the stimulus
oscillate within the inhibitory fields of ommatidia, canceling
time-dependent inhibitory interactions in the retinal network.
As a consequence, the gain increases with spatial frequency
at a rate dependent on the spatial extent of the inhibitory
field (Fig. 6, arrow b). The configurations of inhibitory fields
in our experiments were well approximated by a cratered
Gaussian function having a width at half-maximum of 547,

/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11 neupa LP-Neurophys 601

1808 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW

Downloaded from jn.physiology.org on November 23, 2011

FIG. 5. A: spatial transfer functions measured from 3 experimental eyes (I, II, and III) in response to vertically oriented
sinusoidal gratings drifting at 0.5 (j), 1 (m), and 2 Hz (l). Gain is the ratio of the percent modulation of firing rate to that
of the incident light intensity. B: temporal transfer functions of the 3 eyes in response to flickering small-spot (s) and full-
field () stimulation. Eyes I and III differed most in gain at high temporal frequencies, and eye II had the greatest gain at
low temporal frequencies among the 12 animals in the study. Solid lines were computed by the cell-based model of each
eye. C: power spectra of spike-rate fluctuations recorded () from the 3 eyes. Solid lines gives the power spectra of spike
trains computed by the model. Power is expressed as variance normalized by the mean-squared firing rate. This analysis
defined a set of parameter values used in all future model simulations regarding a given experimental eye.

which is equivalent to 9 ommatidia because their optic axes
diverge by 67 (Barlow et al. 1984; Herzog and Barlow
1992). Such a configuration agrees with previous inhibitory
field maps (Barlow 1969; Brodie et al. 1978b). Beyond the
peak, the gain of the eye falls sharply because the 67
acceptance angle of the optical apparatus (Barlow et al.
1980) filters fine patterns of light, limiting the spatial acuity
of the animal (Fig. 6, arrow c).
The spatial band-pass tuning conferred by the optics and
lateral-inhibitory network diminishes with increasing tempo-
ral frequency because synaptic delays shift the phase of in-
hibitory signals relative to that of excitatory signals. The
shift in phase can disinhibit retinal neurons enhancing
the eyes response to temporal frequencies corresponding to
one-half the synaptic delay for inhibition (Ratliff et al. 1967).
Amplification by lateral inhibition is evident in the temporal
transfer functions presented next.
TEMPORAL TRANSFER FUNCTION. The lateral eye is not
equally sensitive to all temporal frequencies either. It is max-
imally responsive to both small-spot (s) and full-field ()
flicker of intermediate temporal frequencies (Fig. 5B). Re-
sponses to small-spot stimuli reflect only the excitatory prop-
erties of the receptor that the spot stimulates, whereas re-
sponses to full-field stimuli include those of the inhibitory
network as well (Knight et al. 1970; Ratliff et al. 1967,
1969). For both small-spot and full-field conditions, the gain
of the eye increases monotonically with temporal frequency
to a peak in the region of 36 Hz and then falls sharply.
Over a large range of temporal frequencies (0.516 Hz), the
gain is greater than one, indicating that the eye amplifies the
stimulus in its output. We typically measured peak gains of
up to 510, with the eye showing greater sensitivity to small
spots at low temporal frequencies and large spots at interme-
diate frequencies. We conclude that the Limulus eye is mark-
edly tuned for light oscillating at temporal frequencies in
the range of 46 Hz.
Dynamic interactions of phototransduction, self inhibition,
and lateral inhibition shape the temporal transfer function.
Bump adaptation and self inhibition (KSI of 23) combine
to reduce the amplitude of the eyes response to slowly
modulated spots of light (Fig. 6, arrow e). Lateral inhibition
(KLI of 4) further decreases its response to larger areas of
stimulation, i.e., low spatial frequency (Fig. 6, arrow a).
These relatively slow mechanisms become less effective at

/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11 neupa LP-Neurophys 602

 CELL-BASED MODEL OF THE LIMULUS EYE 1809

noise spectra should be compared because we equated their

FIG. 6. Schematic spatial and temporal transfer functions illustrate with
arrows the major features affected by changes in model parameters. Letters
refer to parameters listed in Table 1.
higher temporal frequencies, allowing the gain to reach its
maximum value. The rate of growth (Fig. 6, arrow d) and
frequency of maximum gain (Fig. 6, arrow f ) depend on
the time constants of bump adaptation, self inhibition (tSI
of 0.10.15 s), and lateral inhibition (tLI of 0.070.1
total variance in setting the mean bump rate lV .
We note that the spatial and temporal transfer functions
shown in Fig. 5 agree well with those reported previously
for intact eyes (Batra 1983; Batra and Barlow 1990; Brodie
et al. 1978b). These authors fit their transfer functions with
analytic models having the same transduction mechanisms
as our cell-based model but required an arbitrary scaling
factor M to match empiric measurements. The M-factor
compensated for frequency-independent mechanisms not in-
cluded in their models, such as changes in membrane resis-
tance with light level and thresholds for spike firing (Knight
et al. 1970). Our cell-based implementation of the Hartline-
Ratliff equations, however, does not allow for arbitrary scal-
ing of model computations because it incorporates all of the
major physiological, anatomic, and biophysical mechanisms
known to operate in the eye. To achieve the high gains that
have been measured from the eye, we must include in the
model an electrogenic ion pump that acts near the spike-
initiation site and generates a small hyperpolarizing current

Downloaded from jn.physiology.org on November 23, 2011

s). The finite duration of bumps (tb of 0.010.02 s) limits
the ability of ommatidia to follow rapid changes in light
intensity, thereby attenuating the gain at high temporal fre-
quencies (Fig. 6, arrow g).
NOISE SPECTRUM. The power spectrum of random fluctua-
tions in optic nerve firing rate is also shaped by the response
properties of the eye. In all three experimental eyes, noise
power increased slightly with temporal frequency and then
fell precipitously at 45 Hz (Fig. 5C). The frequency of
maximum noise power corresponded to that of maximal gain
for flickering small-spot stimuli (Fig. 5B). The random fluc-
tuations in firing rate result largely from the stochastic arrival
of photons that generate bumps, with a minor contribution
at low temporal frequencies from asynchronous inhibitory
input from neighboring ommatidia (Shapley 1971a,b). The
coefficient-of-variation of the noise, given by the standard
deviation of the instantaneous firing rate relative to the mean
rate, was 0.06, 0.11, and 0.07 for eyes IIII, respectively,
at the level of illumination in our laboratory experiments.
in response to nerve impulses the cell fires (Fohlmeister et
al. 1977; Smith et al. 1968). Because the spike-initiation site
is electrically remote from the soma, the current is not readily
recorded by a microelectrode. The ionic pump increases the
overall responsivity of the eye without affecting its spatio-
temporal properties.
Modest changes in the values of many model parameters
do not significantly influence the shape of transfer functions.
To quantify this, we varied each parameter individually by
50% of its standard value and computed its sensitivity
index for drifting gratings (Table 1). The sensitivity index
is defined as the fractional change in the specific feature of
the transfer function most affected by the modified parameter
(Fig. 6). Examination of Table 1 shows that the spatiotempo-
ral response properties of the eye are sensitive to changes
in acceptance angle, inhibitory field configuration, and bump
duration. The fractional change in the appropriate feature
was identical to the fractional change in value of these pa-
rameters, i.e., the sensitivity index was one. This is because

TABLE 2. Comparison of model computations and

Determination of parameter values for the model experimental recordings in Figs. 6 and 7 for responses
Table 1 lists the values of model parameters used in simu- of eyes I, II, and III to crab-size bars
lating optic nerve responses of experimental eyes IIII. The mV eV mmV eeV mV mV
spatial and temporal transfer functions computed by the
model using these values are given by solid lines in Fig. IA 0.967 0.895 { 0.011 0.948 { 0.019 0.975 (IIA)
5, A and B. The computed transfer functions match those IIA 0.956 0.922 { 0.008 0.923 { 0.035 0.972 (IIIA)
measured for all three eyes with the possible exception of IIIA 0.977 0.936 { 0.011 0.969 { 0.007 0.963 (IA)
IB 0.972 0.923 { 0.014 0.972 { 0.012 0.976 (IIB)
eye III. The model predicted lower gains at the peak of the IIB 0.976 0.939 { 0.011 0.956 { 0.014 0.982 (IIIB)
temporal transfer function of this eye. The discrepancy may IIIB 0.984 0.948 { 0.008 0.975 { 0.006 0.978 (IB)
result from overestimating the peak gain of the recorded IC 0.983 0.928 { 0.013 0.972 { 0.009 0.969 (IIC)
neuron because of weak phase locking in its spike train or IIC 0.990 0.949 { 0.015 0.974 { 0.007 0.974 (IIIC)
IIIC 0.970 0.969 { 0.008 0.981 { 0.006 0.989 (IC)
from an incompletely characterized mechanism in the lateral
eyes of some animals, such as a delay in self inhibition, Values in mmV and eeV are means { SD. Crab-size bars were moving
which would boost the gain at temporal frequencies recipro- at 4 cm/s (A), 8 cm/s (B), and 16 cm/s (C). m, model computations; e,
cal to about twice the delay (Ratliff et al. 1969). Power experimental recordings; mV eV , correlation coefficient of average computed
spectra computed by the model also match well those mea- and recorded responses for the same eye; mmV , mean correlation coefficient
of individual computed responses and their combined average for the same
sured from the experimental eyes (solid lines in Fig. 5C), eye; eeV , mean correlation coefficients of individual recorded responses
except for a slightly reduced amount of power at low tempo- with their combined average for the same eye; mV mV , correlation coefficient
ral frequencies. Only the shapes of computed and recorded of average computed responses of different eyes.

/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11 neupa LP-Neurophys 603

1810 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW
no other mechanisms affect these features of the transfer
function. The spatiotemporal response properties are, on the
other hand, less sensitive to comparable changes in mean
bump rate and the time constant and strength of self inhibi-
tion. The eye is about threefold less sensitive to these param-
eters. Insensitivity to halving of the mean bump rate is under-
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
standable because the steady-state response of the eye de-
pends logarithmically on light intensity (Eq. 9). Insensitivity
to halving of self-inhibitory parameters suggests that self
inhibition is not a finely tuned element of the eyes design.
Note also that changes in more than one parameter of the
model often offset one another because many parameters
affect overlapping regions of the transfer function.
Computed responses to simulated moving crab-size objects
Having determined parameter values for the experimental
eyes, we can now predict their optic nerve responses to
repeated presentations of a simplified underwater scene. The
scene consists of a horizontally elongated bar, having the
size and contrast of an adult horseshoe crab, moving at dif-
ferent velocities across a uniform background. Figure 7 plots
the individual responses and average response computed for
a single nerve fiber in each of the three eyes. The computed
firing rate first decreases as the dark bar crossed the neurons
field of view and then momentarily increases on leaving its
view. The relative modulation of firing rate predicted by the
Downloaded from jn.physiology.org on November 23, 2011
model was 31, 46, and 34% for the slow speed (A), 48, 75,
and 54% for the medium speed (B), and 67, 97, and 80%
for the fast speed (C). Increasing the velocity of the stimulus
selectively enhances the amplitude of responses to leading
and trailing edges of the simulated crab. This edge en-
hancement boosted the relative modulation of firing rate by
300% (2- to 6-fold) in moving from the slowest to fastest
velocity. Over this range of velocities, the duration of re-
sponses to the leading and trailing edges was halved from
0.6 to 0.3 s.
Responses computed by the model for the three eyes
to moving crab-size objects are surprisingly similar even
though the eyes did not have identical properties. Correla-
tion coefficients between average responses of different
eyes were consistently above 0.95 at all 3 speeds, whereas
those between individual responses of the same eye and
their combined average were consistently below 0.95 (Ta-
ble 2). Hence the intrinsic variability of individual re-
sponses of the same eye resulting from transduction noise
is larger than the small differences in average responses
of different model eyes.
Accuracy of model predictions
We evaluated the accuracy of model predictions by re-
cording from single optic nerve fibers while presenting the
same underwater scenes to experimental eyes. Figure 8 plots
the individual and average responses of a neuron in each of
the three eyes to the crab-size object moving at the same
three velocities as before. The recorded firing rate first de-
creased and then momentarily increased in response to the
FIG. 7. Computed responses of optic nerve fibers to simulated crab-size
objects. A, left: firing rate of a neuron in model eyes I (top traces), II (middle
traces), and III (bottom traces) in response to a horizontally elongated bar
(4.5 cm wide, 2.25 cm tall, 035% contrast) moving at 4 cm/s at a distance
of 9 cm from the eye. Three stimulus presentations are shown. Right:
average computed response after 10 presentations. B, left: firing rate of
single neurons in the 3 model eyes for a bar moving at 8 cm/s. Right:
average computed response after 20 stimulus presentations. C, left: firing
rate of single neurons in the 3 model eyes for a bar moving at 16 cm/s.
Right: Average computed response after 40 stimulus presentations.
neupa LP-Neurophys 604
 CELL-BASED MODEL OF THE LIMULUS EYE
moving dark bar, as predicted for the corresponding omma-
tidium in the model (compare with Fig. 7). The relative
modulation of firing rate recorded from the three experimen-
tal eyes was 31, 48, and 38% for the slowest speed, 47, 73,
and 50% for the medium speed, and 64, 99, and 75% for
the fastest speed in good agreement with model simulations.
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
1811
Indeed, not only was the overall shape of recorded responses
to moving bars predicted by the model, but also the random
fluctuations in recorded spike rate.
There are, nonetheless, detectable differences between av-
erage computed and recorded responses. For example, com-
puted transients are slightly more abrupt than those recorded
from the eye. To assess the accuracy of model predictions,
differences must be viewed in relation to the intrinsic noise
level in single optic nerve fibers, because averaging is not
an option under natural conditions. The mean correlation
coefficient between individually recorded responses of the
three eyes and their combined average was 0.95, 0.97, and
0.97 at the three stimulus velocities, whereas the correlation
coefficients between average recorded and computed re-
sponses were 0.97, 0.98, and 0.98, respectively (Table 2).
That is, average responses of the recorded neuron to dark
bars moving at different velocities are as correlated with
those computed by the model as with individually recorded
responses. Hence one cannot distinguish between responses
of the model and the eye without significant averaging to
Downloaded from jn.physiology.org on November 23, 2011
remove noise.
DISCUSSION
Our main objective was to develop a cell-based model of
the Limulus lateral eye that predicts with precision the eyes
input to the brain. We outlined a procedure for computing
patterns of activity in arrays of optic nerve fibers and for
quantitatively evaluating the accuracy of the computations.
Theory and experiment were in good agreement. Average
responses of single optic nerve fibers to simple visual scenes
matched those predicted by the model better than their re-
sponses to individual presentations of the scenes. In other
words, random fluctuations in their firing rate masked small
discrepancies in computed responses that became apparent
only after significant averaging. Because the main purpose
of the model is to simulate optic nerve responses of crabs
in their natural habitat, such precision under well-controlled
conditions in the laboratory is more than sufficient for highly
variable conditions in the field. The cell-based model pre-
sented here is thus an accurate representation of the Limulus
eye in its daytime light-adapted state.
Uniqueness of model predictions
Having confidence in the validity of the model, to what
extent do computed patterns of optic nerve activity differ
among animals in response to the same visual scenes? Recall
in RESULTS that model simulations were carried out for three
experimental eyes whose spatiotemporal transfer functions
differed most among the animals in this study. Yet, differ-
ences in their average responses to moving bars were small
relative to the intrinsic variability of their individual re-
FIG. 8. Recorded responses of optic nerve fibers to simulated crab-size
objects. A, left: firing rate of a recorded neuron in crab eye I (top traces),
II (middle traces), and III (bottom traces) in response to the same bar as
in Fig. 5A moving at 4 cm/s. Three stimulus presentations are shown. Right:
average recorded response after 10 presentations. B, left: firing rate of the
same neurons to the bar moving at 8 cm/s. Right: average recorded responses
after 20 stimulus presentations. C, left: firing rate of the same neurons to
the bar moving at 16 cm/s. Right: average recorded responses after 40
stimulus presentations.
neupa LP-Neurophys 605
1812 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW
sponses (Figs. 6 and 7), suggesting that the eyes encoded
crablike bars in essentially the same manner. To further
investigate this point, Fig. 9 displays all of the temporal
transfer functions measured in this study to small-spot flicker
together with those reported by Brodie et al. (1978b) and
Batra and Barlow (1990). Notice that the transfer functions
measured from intact eyes by previous investigators encom-
pass the observed range of variation in our animals. Most
of this variation is in the peak gain and high-frequency cut-
off, which are determined largely by the duration of quantum
bumps (Wong 1978). Some lateral eyes, for example, have
a peak gain of 4 at 4 Hz, as reported by Batra and Barlow,
whereas others have a peak gain of 8 at 6 Hz, as reported
by Brodie and colleagues. Below 3 Hz, the response char-
acteristics of all tested eyes were nearly identical. In this
range of temporal frequencies, the slope of the transfer func-
tion, defined as the mean ratio ({SD) of response gains at
2 and 0.5 Hz, was consistently around 2.63 { 0.08. Because
a crab-size object moving at crab speed within the visual
range of the animal generates temporal frequencies mostly
in this range of 0.52 Hz (Passaglia et al. 1997), all three
sets of parameter values in Table 1 should yield similar
predictions of the eyes response to such objects, as was
observed (see Figs. 6 and 7). Consequently, we use a stan-
dard model of the Limulus lateral eye to compute patterns
of optic nerve activity in cases where spatiotemporal transfer
functions are unavailable or unmeasurable. Parameter values
of the standard eye are listed in Table 1. Optic nerve re-
sponses computed with the standard model to crablike ob-
jects of different contrast are essentially indistinguishable
from those of an experimental eye whose response properties
had not been characterized (Fig. 11, A and B).
Because vision plays an important role in the animals
mating behavior, we examined the robustness of the eyes
design in relation to this task. We quantified robustness
by computing sensitivity indexes (see RESULTS) for moving
crab-shaped bars. Table 1 lists the results of the analysis.
Computed responses were most sensitive to changes in
scale of the acceptance function. A 50% increase in accep-
tance angle significantly dampened the leading and trailing
edges of the response (Fig. 10, top). This is because the
stimulus enters and leaves the field of view of an ommatid-
ium at a slower rate. Because slow-moving, crab-size bars
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
FIG. 9. Temporal transfer functions mea-
sured from every animal of this study (n 12) in
response to small-spot (left) and full-field (right)
flickering light. Note that the small-spot transfer
functions of the eyes are nearly identical below
3 Hz but differ considerably at higher temporal
frequencies. Solid and dashed lines replot the
transfer functions reported for intact eyes by
Brodie et al. (1978b) and Batra and Barlow
(1990), respectively. Previous measurements en-
compass the range of variation of transfer func-
tions in our experiments.
generate primarily low temporal frequencies (2 Hz), the
response of the eye was also fairly sensitive to the maxi-
Downloaded from jn.physiology.org on November 23, 2011
mum bump amplitude and the strength and time constant
of self inhibition. These slow-acting mechanisms together
with the acceptance angle are thus important design param-
eters in the detection of crablike objects. Consistent with
this notion, the low-frequency slope of the temporal trans-
fer functions in Fig. 9 appears highly conserved among
animals. Conversely, the response of the standard eye was
largely unaffected by comparable changes in the rate and
time constant of bumps and the time constant of lateral
inhibition (Fig. 10, bottom), even though the latter two had
high sensitivity indexes with gratings (Table 1). This is
because these mechanisms affect the gain of the eye at
frequencies not present in a crablike stimulus. They thus
appear less important for signaling potential mates, consis-
tent with the large variability among eyes in their respon-
sivity to high temporal frequencies (Fig. 9). The function
of these mechanisms may be to signal faster-moving stim-
uli in the horseshoe crabs visual environment (Passaglia
et al. 1995) or to filter high-frequency noise.
Computed neural images of simulated crablike objects
Arrays of optic nerve firing rates computed by the model
can be converted to a gray scale and mapped back to their
appropriate location in the retina to form neural images of
the eyes input to the brain. Figure 11, C and D, shows
computed neural images of optic nerve activity at a moment
in time in response to the moving high- and low-contrast
crab-size objects, respectively. The leading and trailing
edges of the moving dark bar appear black and white, respec-
tively, in the neural image. That is, the firing rate of a small
cluster of nearby neurons is reduced (black), whereas that
of an adjacent cluster is elevated (white). The pattern of
optic nerve activity evoked by the high-contrast bar is easily
seen against the background of maintained activity, which
appears neutral gray in the neural image. The signal-to-noise
ratio, defined as the mean response variance of neurons view-
ing the bar relative to those viewing the background, was
9.0 for this snapshot.
Unlike the high-contrast object, the moving low-contrast
one is barely visible in the neural image. The signal-to-noise
neupa LP-Neurophys 606
 CELL-BASED MODEL OF THE LIMULUS EYE
FIG. 10. Responses of the lateral eye to a crab-size bar moving at crab
speed are sensitive to some parameters and not others. A 50% increase in
the acceptance angle from its standard value caused the maximal change
in response (top), whereas the same increase of bump duration had a mini-
mal effect (bottom). Dashed lines are the original responses of the standard
model eye, and solid lines are the responses after modifying one of its
parameter values.
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
1813
ratio was over threefold less (2.7) for this snapshot. Random
fluctuations in firing rate obscure the weak pattern of neural
activity evoked by the light object. Object-independent fluc-
tuations in firing rate will be even larger in the animals
underwater environment (Passaglia et al. 1995). Yet, behav-
ioral studies show that male horseshoe crabs can see crablike
objects of similar size and contrast as those used in this
study almost equally well day and night (Herzog et al. 1996).
Passaglia et al. (1997) address how horseshoe crabs achieve
this seemingly remarkable visual performance.
In summary, retinal coding of visual behavior has been
under investigation since Hartline first isolated the spike
trains of individual nerve cells in the eye (Hartline and Gra-
ham 1932). For over 65 years, inferences have been made
about retinal processing as a whole from the vantage point
of single neurons. Multielectrode recordings from excised
retina promise to expand our knowledge about how ensem-
bles of retinal neurons transmit information (DeVries and
Baylor 1997; Meister 1996). Studies of retinal coding may
eventually progress to the point when the major portion of
Downloaded from jn.physiology.org on November 23, 2011
the eyes output can be recorded while animals perform a
behavioral task. In the meantime, we have developed a real-
istic computational model of the Limulus lateral eye that
predicts the eyes input to the brain in response to specified
scenes. To anticipate what future studies may learn about
retinal coding by ensembles of neurons in behaving animals,
it will be instructive to examine how the model responds
to the collection of scenes that horseshoe crabs typically
encounter in their natural habitat.
FIG. 11. A: responses recorded from an op-
tic nerve fiber to a moving bar (4.5 cm wide,
2.25 cm tall, 4.5 cm distance, 2 cm/s velocity)
of high (035%, top traces) and low contrast
(010%, bottom traces). B: responses computed
for the corresponding neuron in the model us-
ing the parameters of the standard eye (Table
1). C and D: snapshots of the model input and
computed responses of the entire array of optic
nerve fibers for the high- and low-contrast
moving crablike bar, respectively. Computa-
tions used the parameters of the standard eye.
Black squares in the neural images indicate
neurons having a zero firing rate; white ones
indicate those having a firing rate of 18 im-
pulses/s.
neupa LP-Neurophys 607
1814 C. L. PASSAGLIA, F. A. DODGE, AND R. B. BARLOW
Address for reprint requests: R. B. Barlow, Center for Vision Research,
Dept. of Ophthalmology, SUNY Health Science Center, Syracuse, NY
13210.
Received 24 October 1997; accepted in final form 8 June 1998.
REFERENCES
ADOLPH, A. R. Spontaneous slow potential fluctuations in the Limulus pho-
toreceptor. J. Gen. Physiol. 48: 297322, 1964.
BARLOW, H. B. Summation and inhibition in the frogs retina. J. Physiol.
(Lond.) 119: 6988, 1953.
BARLOW, R. B. Inhibitory fields in the Limulus lateral eye. J. Gen. Physiol.
54: 383396, 1969.
BARLOW, R. B., CHAMBERLAIN, S. C., AND KASS, L. Circadian rhythms in
retinal function. In: Molecular and Cellular Basis of Visual Acuity, edited
S. R. Hilfer and J. B. Sheffield. New York: Springer-Verlag, 1984, p.
3153.
BARLOW, R. B., CHAMBERLAIN, S. C., AND LEHMAN, H. K. Circadian
rhythms in the invertebrate retina. In: Facets of Vision, edited by D. G.
Stavenga and R. Hardie, Berlin: Springer-Verlag, 1989.
BARLOW, R. B., CHAMBERLAIN, S. C., AND LEVINSON, J. Z. The Limulus
brain modulates the structure and function of the lateral eye. Science
210: 10371039, 1980.
BARLOW, R. B., IRELAND, L. C., AND KASS, L. Vision has a role in the
Limulus mating behavior. Science 296: 6566, 1982.
BARLOW, R. B. AND KAPLAN, E. Properties of visual cells in the lateral eye
of Limulus in situ. Intracellular recordings. J. Gen. Physiol. 69: 203
220, 1977.
BARLOW, R. B. AND LANGE, G. D. A nonlinearity in the inhibitory interac-
tions in the lateral eye of. Limulus. J. Gen. Physiol. 63: 579589, 1974.
BARLOW, R. B. AND QUARLES, D. A. Mach bands in the lateral eye of
Limulus. Comparison of theory and experiment. J. Gen. Physiol. 65:
709730, 1975.
BATRA, R. Efferent Control of Visual Processing in the Lateral Eye of the
Horseshoe Crab (PhD thesis). Syracuse, NY: Syracuse Univ., 1983.
BATRA, R. AND BARLOW, R. B. Efferent control of temporal response prop-
erties of the Limulus lateral eye. J. Gen. Physiol. 95: 229244, 1990.
BIEDERMAN-THORSON, M. AND THORSON, J. Dynamics of excitation and
inhibition in the light-adapted Limulus eye in situ. J. Gen. Physiol. 58:
119, 1971.
BRODIE, S. E., KNIGHT, B. W., AND RATLIFF, F. The response of the Limulus
retina to moving stimuli: a prediction by Fourier synthesis. J. Gen. Phys-
iol. 72: 129166, 1978a.
BRODIE, S. E., KNIGHT, B. W., AND RATLIFF, F. The spatiotemporal transfer
function of the Limulus lateral eye. J. Gen. Physiol. 72: 167201, 1978b.
COLEMAN, B. D. AND RENNINGER, G. H. Theory of delayed inhibition in
the compound eye of Limulus. Proc. Natl. Acad. Sci. USA 71: 2887
2891, 1974.
COLEMAN, B. D. AND RENNINGER, G. H. Theory of spatially synchronized
oscillatory responses in the retina of. Limulus. J. Math. Biosci. 38: 123
140, 1978.
DEVRIES, S. H. AND BAYLOR, D. A. Mosaic arrangement of ganglion cell
receptive fields in rabbit retina. J. Neurophysiol. 78: 20482060, 1997.
DODGE, F. A. AND KAPLAN, E. Transmission of visual information by the
Limulus eye (Abstract). Biophys. J. 142: 11, 1977.
DODGE, F. A., KNIGHT, B. W., AND TOYODA, J. Voltage noise in Limulus
visual cells. Science 160: 8890, 1968.
DOWLING, J. E. The Retina: An Approachable Part of the Brain. Cambridge,
MA: Harvard Univ. Press, 1987.
FOHLMEISTER, J. F., POPPELE, R. E., AND PURPLE, R. L. Repetitive firing:
quantitative analysis of encoder behavior of slowly adapting stretch re-
ceptor of crayfish and eccentric cell of. Limulus. J. Gen. Physiol. 69:
849877, 1977.
FUORTES, M.G.F. Initiation of impulses in visual cells of. Limulus. J. Phys-
iol. (Lond.) 148: 1428, 1959.
FUORTES, M.G.F. AND HODGKIN, A. L. Changes in time scale and sensitivity
in the ommatidia of. Limulus. J. Physiol. (Lond.) 172: 239263, 1964.
FUORTES, M.G.F. AND YEANDLE, S. Probability of occurences of discrete
potential waves in the eye of. Limulus. J. Gen. Physiol. 47: 443463,
1964.
HARTLINE, H. K. The responses of single optic nerve fibers of the vertebrate
eye to illumination of the retina. Am. J. Physiol. 121: 400415, 1938.
HARTLINE, H. K. Inhibition of activity of visual receptors by illuminating
nearby retinal areas in the Limulus eye. Federation Proc. 8: 69, 1949.
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
HARTLINE, H. K. AND GRAHAM, C. H. Nerve impulses from single receptors
in the eye. J. Cell. Comp. Physiol. 1: 227295, 1932.
HARTLINE, H. K. AND RATLIFF, F. Inhibitory interactions of receptor units
in the eye of Limulus. J. Gen. Physiol. 40: 357376, 1957.
HARTLINE, H. K. AND RATLIFF, F. Spatial summation of inhibitory influ-
ences in the eye of Limulus, and the mutual interactions of receptor units.
J. Gen. Physiol. 41: 10491066, 1958.
HARTLINE, H. K., WAGNER, H. G., AND MACNICHOL, E. F. The peripheral
origin of nervous activity in the visual system. Cold Spring Harb. Symp.
Quant. Biol. 17: 125141, 1952.
HERZOG, E. H. Vision in Limulus: From Optics to Neurons to Behavior
(PhD thesis). Syracuse, NY: Syracuse Univ., 1994.
HERZOG, E. H. AND BARLOW, R. B. The Limulus-eye view of the world.
Vis. Neurosci. 9: 571580, 1992.
HERZOG, E. H., PASSAGLIA, C. L., DODGE, S. A., LEVINE, N. D., AND BAR-
LOW, R. B. Limulus vision in the ocean: comparing neural and behavioral
thresholds. Biol. Bull. 185: 307308, 1993.
HERZOG, E. H., POWERS, M. K., AND BARLOW, R. B. Limulus vision in the
ocean day and night: effects of image size and contrast. Vis. Neurosci.
13: 3141, 1996.
KAPLAN, E. AND BARLOW, R. B. Properties of visual cells in the lateral eye
of Limulus in situ. Extracellular recordings. J. Gen. Physiol. 66: 303
326, 1975.
KAPLAN, E., BARLOW, R. B., RENNINGER, G., AND PURPURA, K. Circadian
Downloaded from jn.physiology.org on November 23, 2011
rhythms in Limulus photoreceptors. II. Quantum bumps. J. Gen. Physiol.
96: 665685, 1990.
KNIGHT, B. W. Dynamics of encoding in a population of neurons. J. Gen.
Physiol. 59: 734766, 1972.
KNIGHT, B. W. A stochastic problem in visual neurophysiology. In: Ameri-
can Mathematical Society Symposium on Stochastic Differential Equa-
tions, edited by J. Keller and H. J. McKean. Providence, RI: Amer. Math.
Soc., 1973, p. 119.
KNIGHT, B. W., TOYODA, J., AND DODGE, F. A. Quantitative description of
the dynamics of excitation and inhibition in the eye of Limulus. J. Gen.
Physiol. 56: 421437, 1970.
KUFFLER, S. W. Discharge patterns and functional organization of mamma-
lian retina. J. Neurophysiol. 16: 3768, 1953.
LAUGHLIN, S. B. Neural principles in the peripheral visual systems of inver-
tebrates. In: Comparative Physiology and Evolution of Vision in Inverte-
brates. B: Invertebrate Visual Centers and Behavior I, edited by H.
Autrum. New York: Springer-Verlag, 1981.
MACNICHOL, E. F. Visual receptors as biological transducers. In: Molecular
Structure and Functional Activity of Nerve Cells, edited by R. G. Grenell
and L. J. Mullins. Washington, DC: Amer. Inst. Biol. Sci., 1956, p. 3462.
MEISTER, M. Multineuronal codes in retinal signaling. Proc. Natl. Acad.
Sci. USA 93: 609614, 1996.
MEISTER, M., PINE, J., AND BAYLOR, D. A. Multi-neuronal signals from the
retina: acquistion and analysis. J. Neurosci. Methods 51: 95106, 1994.
PASSAGLIA, C. L., DODGE, F. A., AND BARLOW, R. B. Limulus is tuned into
its visual environment. Biol. Bull. 189: 213215, 1995.
PASSAGLIA, C. L., DODGE, F. A., AND BARLOW, R. B. Deciphering a neural
code for vision. Proc. Natl. Acad. Sci. USA 94: 1264912654, 1997.
POWERS, M. K., BARLOW, R. B., AND KASS, L. Visual performance of horse-
shoe crabs day and night. Vis. Neurosci. 7: 179189, 1991.
PURPLE, R. L. AND DODGE, F. A. Mechanisms of excitation and inhibtion
in the Limulus eye. Cold Spring Harb. Symp. Quant. Biol. 30: 529537,
1965.
PURPLE, R. L. AND DODGE, F. A. Self-inhibition in the Limulus lateral eye.
In: Functional Organization of the Compound Eye, edited by C. G. Bern-
hard. New York: Pergamon, 1966, p. 451464.
RATLIFF, F. Studies of Excitation and Inhibition in the Retina. New York:
The Rockefeller Univ. Press, 1974.
RATLIFF, F. AND HARTLINE, H. K. The responses of Limulus optic nerve
fibers to patterns of illumination on the receptor mosaic. J. Gen. Physiol.
42: 12411255, 1959.
RATLIFF, F., HARTLINE, H. K., AND MILLER, W. H. Spatial and temporal
aspects of retinal inhibitory interaction. J. Opt. Soc. Am. 53: 110120,
1963.
RATLIFF, F., KNIGHT, B. W., DODGE, F. A., AND HARTLINE, H. K. Fourier
analysis of dynamics of excitation and inhibition in the eye of Limulus:
amplitude, phase and distance. Vision Res. 14: 11551168, 1974.
RATLIFF, F., KNIGHT, B. W., AND GRAHAM, N. On tuning and amplification
by lateral inhibition. Proc. Natl. Acad. Sci. USA 62: 525532, 1969.
neupa LP-Neurophys 608
 CELL-BASED MODEL OF THE LIMULUS EYE
RATLIFF, F., KNIGHT, B. W., TOYODA, J.-I., AND HARTLINE, H. K. Enhance-
ment of flicker by lateral inhibition. Science 158: 392393, 1967.
RENNINGER, G. H., KASS, L., PELLETIER, J. L., AND SCHIMMEL, R. The ec-
centric cell of the Limulus lateral eye: encoder of circadian changes in
visual responses. J. Comp. Physiol. [A] 163: 259270, 1988.
SHANMUGAN, K. S. AND BREIPOHL, A. M. Random Signals: Detection, Esti-
mation, and Data Analysis. New York: Wiley, 1988.
SHAPLEY, R. Fluctuations of the impulse rate in Limulus eccentric cells. J.
Gen. Physiol. 57: 539556, 1971a.
SHAPLEY, R. Effects of lateral inhibition on fluctuations of the impulse rate.
J. Gen. Physiol. 57: 557575, 1971b.
SMITH, T. G., STELL, W. K., BROWN, J. E., FREEMAN, J. A., AND MURRAY,
G. C. A role for the sodium pump in photoreception in Limulus. Science
162: 456458, 1968.
STEVENS, C. F. A Quantitative Theory of Neural Interactions: Theoretical
and Experimental Investigations (PhD thesis). New York: The Rockefel-
ler Univ., 1964.
TEETERS, J., JACOBS, A., AND WERBLIN, F. How neural interactions form neural
responses in the salamander retina. J. Comp. Neurosci. 4: 527, 1997.
VICTOR, J. D., SHAPLEY, R. M., AND KNIGHT, B. W. Nonlinear analysis of
cat retinal ganglion cells in the frequency domain. Proc. Natl. Acad. Sci.
USA 74: 30683072, 1977.
/ 9k2d$$oc12 J-876-7 09-10-98 07:57:11
1815
WEINER, W. W. AND CHAMBERLAIN, S. C. The visual fields of American
horseshoe crabs: two different eye shapes in Limulus polyphemus. Vis.
Neurosci. 11: 333346, 1993.
WELCH, P. D. The use of fast Fourier transform for the estimation of power
spectra: a method based on time averaging over short, modified periodo-
grams. IEEE Trans. Audio Electroacoustics AU-15: 70, 1967.
WERBLIN, F. S. Synaptic connections, receptive fields, and patterns of activ-
ity in the tiger salamander retina. Invest. Ophthalmol. Vis. Sci. 32: 459
483, 1991.
WONG, F. Nature of light-induced conductance changes in ventral photore-
ceptors of Limulus. Nature 276: 7679, 1978.
WONG, F. AND KNIGHT, B. W. Adapting-bump model for eccentric cells of.
Limulus. J. Gen. Physiol. 76: 539557, 1980.
WONG, F., KNIGHT, B. W., AND DODGE, F. A. Dispersion of latencies in
photoreceptors of Limulus and the adapting-bump model. J. Gen. Physiol.
76: 517537, 1980.
WONG, R. O., CHERNJAVSKY, A., SMITH, S. J., AND SHATZ, C. J. Early
functional networks in the developing retina. Nature 374: 716 718,
1995.
YEANDLE, S. Studies on the Slow Potential and the Effects of Cations
on the Electrical Responses of the Limulus Ommatidium (PhD thesis).
Baltimore, MD: The Johns Hopkins Univ., 1957.
Downloaded from jn.physiology.org on November 23, 2011
neupa LP-Neurophys 609


610