Вы находитесь на странице: 1из 12

BY: MARI O B.

PEREZ
Multi-voxel Pattern Analysis (MVPA):
Fundamentals & Case of Study
A bit of context
Traditional fMRI analysis carry out univariate
analysis upon individual voxels (mass-univariate)

Conventional analysis try to find voxels that respond
significantly to an experimental condition.

To increase sensitivity, this methods spatially
average across voxels that respond to that cognition
(other steps like spatial smoothing)

The beauties of MVPA
MVPA performs multivariate analysis that takes into
account the relationships between voxels.

Uses classifier algorithms to learn and then recognise
those activity patterns.

It takes into account fine-grained activity that is
normally discarded in standard fMRI analysis.

Provides an enhanced sensitivity over one-voxel methods:
(no spatial smoothing or averaging)

Effects of spatial smoothing: Loss offine-
grained activity
Spatial smoothing generally increases the chances of obtaining
significant activity, but blurs vital information

MVPA basic procedure
Four main steps

Split your data in training and testing sets.

Feature selection: This first step tries to delimitate a set of voxels that
will be used further using training data.

Classifier training: In these stage the training data is used so our
classifier learns our activity patterns. Once training has finished a
decision plane its generated

Classifier or generalization testing: The classifier is exposed to new
data (testing data set) that belongs to the same category. Based on where
they fall and the identity of the example, the classification was
successful or not.

MVPA basic procedure
Classifiers & Decision Plane: How we distinguish
between conditions?
Voxels can work as coordinates to define a point in a plane.
We would only need a line to separate our conditions.
(linear classifiers)
Chadwick et al., 2010




Goal was to distinguish between memory traces
(a.patterns)
Searchlight feature selection (Sphere/radius)
Three ROIs were defined (hippocampus, entorhinal cortex
and parahippocampal gyrus)
K-fold cross-validation (9 iterations one with each memory
as test)

Chadwick et al., 2010
Results







You have been released!
Overfitting (subtle versions)
When you use all
your examples
(stimuli) to select
your voxels...voil!

You have included
some voxels that
are compatible
between your
testing and your
training data.

Вам также может понравиться