Resolving Ambiguities of MVPA Using Explicit Models of Representation

Abstract

We advocate a shift in emphasis within cognitive neuroscience from multivariate pattern analysis (MVPA) to the design and testing of explicit models of neural representation. With such models, it becomes possible to identify the specific representations encoded in patterns of brain activity and to map them across the brain.

Keywords: computational modeling; encoding model; fMRI; multivariate pattern analysis; representation.

Figure 1. Resolving the representational ambiguity of MVPA

(A) Example MVPA experiment. Responses to several movie clips are measured. It is demonstrated that a linear classifier can predict the genre of a movie clip based on the multivoxel activity pattern elicited by that movie clip. (B) Representational ambiguity. Movie clips from different genres may differ with respect to one or more features. For example, clips from action movies (gray) may have larger amounts of visual energy than clips from romantic comedies (black). Thus, the voxels under consideration might represent a feature that is correlated with, but distinct from, movie genre. (C) Building encoding models. To adjudicate between competing hypotheses about the features that are represented, each feature of interest is used to build an encoding model and the various models are fit to the data (see Box 1). By directly comparing the accuracy (variance explained) of different models, it is possible to determine the features encoded in the population activity.