Hyperalignment of multi-subject fMRI data by synchronized projections

Abstract

Group analysis of fMRI data via multivariate pattern methods requires accurate alignments between neuronal activities of different subjects in order to attain competitive inter-subject classification rates. Hyperalignment, a recent technique pioneered by Haxby and collaborators, aligns the activations of different subjects by mapping them into a common abstract high-dimensional space. While hyperalignment is very successful in terms of classification performance, its “anatomy free” nature excludes the use of potentially helpful information inherent in anatomy. In this paper, we present a novel approach to hyperalignment that allows incorporating anatomical information in a non-trivial way. Experiments demonstrate the effectiveness of our approach over the original hyperalignment and several other natural alternatives.