Restoring the generalizability of SVM based decoding in high dimensional neuroimage data

Abstract

Variance inflation is caused by a mismatch between linear projections of test and training data when projections are estimated on training sets smaller than the dimensionality of the feature space. We demonstrate that variance inflation can lead to an increased neuroimage decoding error rate for Support Vector Machines. However, good generalization may be recovered in part by a simple renormalization procedure. We show that with proper renormalization, cross-validation based parameter optimization leads to the acceptance of more non-linearity in neuroimage classifiers than would have been obtained without renormalization. © 2012 Springer-Verlag.