Module 27:
Multi-voxel Pattern Analysis
Data Processing Pipeline
Experimental Design
Data Acquisition
Preprocessing Data Analysis
Slice-time Localizing
Correction Brain Activity
Reconstruction
Motion Correction,
Co-registration & Connectivity
Normalization
Spatial
Prediction
Smoothing
Classification and Prediction
• There is a growing interest in using fMRI data for
classification of mental disorders and predicting the
early onset of disease.
• In addition, there is interest in developing methods
for predicting stimuli directly from functional data.
• This opens the possibility of inferring information
about subjective human experience directly from
brain activation patterns.
Machine Learning
• Predicting brain states is challenging and
requires the application of novel statistical and
machine learning techniques.
• Various techniques have successfully been
applied to fMRI data in which a classifier is
trained to discriminate between different brain
states and then used to predict the brain states
in a new set of fMRI data.
Machine Learning
• When applied to fMRI data the result is often a
pattern of weights across brain regions that can
be applied prospectively to new brain activation
maps to quantify the degree to which the pattern
responds to a particular type of event.
wT x > 0 Group A
wT x < 0 Group B
x =(x1,….xV)
w =(w1,….wV)
MVPA
• The application of machine learning methods to
fMRI data is often referred to as multi-voxel
pattern analysis (MVPA)
• Instead of focusing on single voxels, MVPA uses
pattern-classification algorithms applied to
multiple voxels to decode the patterns of activity.
Multivariate Analysis
Multivariate Analysis
Voxel
2
Voxel
1
Multivariate Analysis
Voxel
2
Voxel
1
Multivariate Analysis
Voxel
2
Voxel
1
MVPA vs GLM
• In MVPA the goal is to determine the model
parameters that allow for the most accurate
prediction of new observations.
– Seek to create rules that can be used to categorize
new observations.
• In contrast, the GLM seeks to determine the
model parameters that best fit the data at hand.
Classifiers
• A classifier is a function f(.) that takes the values
of observed features (e.g., voxels) and predicts
to which class the observation belongs (e.g.,
disease state).
• Let us denote the set of features x=(x1,…xV) and
the class label y.
^
• Predicted class: y = f(x)
Training Data
• A classifier has a number of parameters w that
needed to be estimated, or learned.
• The learning is typically performed on a subset
of the observations called the training data.
• The learned classifier models the relationship
between the features and class labels in the
training data set.
Test Data
• Once trained, the classifier is evaluated using an
independent set of observations called the test data.
• If the classifier truly captures the relationship
between features and classes, it should be able to
predict the class label for data it hasn’t seen before.
• The accuracy of the classifier measures the fraction
of observations in the test data for which the correct
label was predicted.
Illustration
Features
(voxels)
Class
Labels
x =(x1,….xV)
y
Data
Observa7ons
Illustration
The full data set is split into two parts:
training and test data
Training
Data
Test
Data
Illustration
Training
Data
Test
Data
Classifier
Predicted
True
labels
labels
End of Module
@fMRIstats