Feature Induction Using Boosting and Logistic Regression on fMRI Images

Early efforts in fMRI classification were limited in that individual voxels were used as features (e.g. [1]), yet voxels divide images into regions that do not directly correspond to underlying neural activity. A growing trend is to perform spatial smoothing that captures the correlation between nearby voxels. Unfortunately, the optimal spatial resolution for this smoothing is unknown and likely varies across brain regions and cognitive tasks. The present work describes two methods that induce features of varying size and shape and use them to produce additive models that offer the potential for easy interpretability.