Inferring affect from fMRI data