Functional Magnetic Resonance Imaging (FMRI) provides an unprecedented window into the complex functioning of the human brain, typically detailing the activity of thousands of voxels for hundreds of time points. The interpretation of FMRI is complicated, however, because of the unknown connection between the hemodynamic response and neural activity, and the unknown spatiotemporal characteristics of the cognitive patterns themselves. Recent work has exploited techniques from machine learning to find patterns of voxel activity related to brain processes (see e.g., [1]). Many of these techniques involve decoding, inferring the value or category class of a stimulus !S given a pattern of voxel activations !V . Decoding can generally be split into two approaches, discriminative and generative [2]. With a discriminative model one learns the conditional distribution P(!S |!V ) directly by minimizing a loss such as minimum classification error. Alternatively, the generative approach obtains this conditional probability through Bayes rule; one posits and fits models for P(!S ) and P(!V |S) instead. Both approaches can reliably establish the existence of sufficient decoding information.