Looking outside the searchlight

Abstract

Searchlight analysis (information mapping) with pattern classifiers is a popular method of multivariate fMRI analysis often interpreted as localizing informative voxel clusters. Applicability and utility of this method is limited, however, by its dependency on searchlight radius, the assumption that information is present at all spatial scales, and its susceptibility to overfitting. These problems are demonstrated in a dataset in which, contrary to common expectation, voxels identified as informative do not clearly contain more information than those not so identified. © 2012 Springer-Verlag.