Multi-variate Pattern Analysis (MVPA)

Traditional neuroimaging analysis techniques are designed to detect regional “activation” in the brain, which is often at centimetre scale. However, increasing number of studies have shown that fine-grained information can also be obtained from the fMRI signal at a scale that is smaller than this. In achieving this, pattern classifiers are often used to relate distinct patterns of activation within the brain to corresponding mental states. For example, MVPA has been used to reveal sensitivity to orientation in V1 region in human brains. Such information of orientation was lost in traditional fMRI analyses.

Many different classifiers have been used to build the correspondence between activity pattern and mental state, including linear-discriminant analysis and support vector machines. In some cases, one can test regions of interestes if a priori reason can be established. In others, the pattern of the whole brain is classified. In the middle ground, a roaming “searchlight” can be used to examine the whole brain one voxel at a time, hence providing specificity in localisation while still using local pattern information.

Representational Similarity Analysis (RSA)

RSA is a particular versatile version of MVPA. It goes beyond testing for information in regional response patterns and enables us to handle condition-rich experiments without predefined stimulus categories, to test conceptual and computational models, and to relate representations between humans and monkeys (Kriegeskorte et al. 2008, 2009). RSA characterizes the representation in each brain region by a representational dissimilarity matrix (RDM; Figure 1). An RDM is a square symmetric matrix, each entry referring to the dissimilarity between the activity patterns associated with two stimuli (or experimental conditions).


Figure 1 Computation of a representational dissimilarity matrix. For each pair of experimental stimuli, the response patterns elicited in a brain region or model representation are compared to determine the stimuli’s representational dissimilarity. The dissimilarity between two patterns can be measured as 1 minus the correlation (0 for perfect correlation, 1 for no correlation, 2 for perfect anticorrelation). These dissimilarities for all pairs of stimuli are assembled in the representational dissimilarity matrix (RDM). Each cell of the matrix then compares the response patterns elicited by two images and the matrix is symmetric about a diagonal of zeros. To visualize the representation, we can arrange the stimuli according to response-pattern dissimilarity, such that stimuli are placed close together if they elicited similar response patterns and far apart if they elicited dissimilar response patterns. The color of each connection line here indicates whether the response-pattern difference was significant (red: p < 0.01; light gray: p ≥ 0.05).


MATLAB Toolbox for RSA

In a collaboration between MRC CBSU and the Neurolex Group in Department of Psychology at University of Cambridge, we have developed a Matlab toolbox for representational similarity analysis. This toolbox is modular and work-flow-based for maximum flexibility and reusability. There are a set of top-level “recipe” functions in the toolbox that allow automatic ROI analysis as well as whole-brain searchlight analysis. Each recipe is formed form a number of “modules”. Each module performs a single type of analysis, e.g. one of the modules displays the representational geometries in a specific region of interest. A module of the toolbox may contain several functions. We call the collection of all functions the “engine” of the toolbox.

The toolbox also allows the users to simulate arbitrary regional pattern ensembles with different levels of fMRI noise. The users can then apply the toolbox to the simulated data and compare the results with the predefined ground truth. This allows them to become familiar with the structure of the codes and also allow them to test different hypothesises.

Extensions of Toolbox to EEG and MEG

The toolbox is currently being extented to analysis data of several different modalities, e.g. human fMRI, EEG, MEG, monky fMRI and electrophysiology. In particular, we have developed Spatiotemporal Searchlight RSA for time resolved imaging data, such as the source estimation of combined EEG and MEG data (Su et al., 2012).

Download Toolbox

Please click HERE.


The toolbox development is partly funded by Medical Research Council and European Research Council.

Meetings and Mailing List

There is a bi-weekly Representational Similarity Analysis Interests Group (RSAIG) meeting to discuss method development and applications of MVPA. Topics covered in the past relate to Multivariate Pattern Analysis (MVPA) including Representational Similarity Analysis (RSA) and other pattern classification approaches to fMRI and E/MEG analysis. For more details of this meeting, please contact the group coordinator Dr. Li Su. To subscribe or unsubscribe from the mailing list, please email or The mailing list hosts more than 40 members of the Unit and researchers in the University. It publishes the topics of each meeting and circulates information on the RSA toolbox.


Relating population-code representations between man, monkey, and computational models 
Kriegeskorte N (2009) Frontiers in Neuroscience. doi:10.3389/neuro.01.035.2009.

Representational similarity analysis – connecting the branches of systems neuroscience 
Kriegeskorte N, Mur M and Bandettini PA (2008) Frontiers in Systems Neuroscience. doi:10.3389/neuro.06.004.2008.

Spatiotemporal Searchlight Representational Similarity Analysis in EMEG Source Space 
Su L., Fonteneau E., Marslen-Wilson W. and Kriegeskorte (2012) 2nd International Workshop on Pattern Recognition in NeuroImaging (PRNI 2012). doi:10.1109/PRNI.2012.26