Functional Magnetic Resonance Imaging (fMRI) provides a unique opportunity to study brain functional architecture, while being minimally invasive. Reverse inference, a.k.a. decoding, is a recent statistical analysis approach that has been used with success for deciphering activity patterns that are thought to fit the neuroscientific concept of population coding. Decoding relies on the selection of brain regions in which the observed activity is predictive of certain cognitive tasks. The accuracy of such a procedure is quantified by the prediction of the behavioral variable of interest - the target. In this paper, we discuss the optimality of decoding methods in two different settings, namely intra- and inter-subject kind of decoding. While inter-subject prediction aims at finding predictive regions that are stable across subjects, it is plagued by the additional inter-subject variability (lack of voxel-to-voxel correspondence), so that the best suited prediction algorithms used in reverse inference may not be the same in both cases. We benchmark different prediction algorithms in both intra- and inter-subjects analysis, and we show that using spatial regularization improves reverse inference in the challenging context of inter-subject prediction. Moreover, we also study the different maps of weights, and show that methods with similar accuracy may yield maps with very different spatial layout of the predictive regions.