(a) Axial and coronal slices of the averaged, normalized T1-weighted structural scan from all subjects showing anatomically defined regions of interest. Odor-evoked ensemble patterns across all voxels within a given ROI were used in a two-step multivariate classification analysis. First, we trained a linear SVM on a training data-set (b, left panel) to separate two odors belonging to different categories. Second, odor category coding was assessed in an independent test data-set (b, right panel), specifically by testing how well the SVM classified the other pair of odors from the corresponding categories; here, cross-decoding is only successful if similar patterns code different odors of the same category. (c) Category decoding from all subjects during the pre-drug session showed that classification accuracy in PPC, OFC, amygdala, and pHIP significantly exceeded chance (mean ± between-subject s.e.m., n = 32, *P < 0.05, one-tailed).