We don't have a good reference for GaussianNB iirc. We should check the references in the fMRI literature and see if they use some smoothing or anything like that.