(A) Classification of single trial FDMs using support vector machines. Before training, single FDMs of CS+ and CS- trials were reduced in their dimensionality (pixels) by a principal component analysis (PCA), resulting in a new representation in this feature space (red and cyan dots). A SVM was then trained to decode CS+ from CS- trials by finding a multidimensional hyperplane (dotted lines) separating CS+ from CS- patterns. The same was repeated for stimuli from the unspecific dimension (not shown). (B) Accuracy of SVM classification along the specific and unspecific dimension, trained within- subject for baseline (white) and generalization phase (gray) (M ± SEM, ***: t(73) = 4.8, p<0.001, *: t(73) = 2.1, p<0.05 paired t-test) (see Source code 1) (C) Activation patterns derived from hyperplanes of classification of CS+ vs CS- trials as shown in (A). Activation patterns were z-scored and averaged across subjects with the same underlying physical CS+ facei (ni = [10, 12, 8, 10, 8, 10, 8, 8]), then superimposed on the respective face stimulus (see Source code 1).