Figure 1.

Overview of reconstruction method. First, principal component analysis (PCA) was applied to a set of 300 training faces to generate component eigenfaces. Second, component scores from the training faces were mapped to evoked patterns of neural activity using a partial least squares regression (PLSR) algorithm. Third, based on patterns of activity elicited during the viewing of a distinct set of 30 test faces, the PLSR algorithm predicted each component score for each test face. Fourth, predicted component scores were used to reconstruct the viewed face. For comparison, test faces were also directly reconstructed based on component scores extracted from the test face (a ‘non-neural reconstruction’; grey box; see also Fig. S1).