Fig. 3.

The impact of low- and high-level manipulations on the DCNN to neural face-space match. a Manipulated face exemplars included in set 1, from top to bottom: Or - original face images, Lum - matched luminance, BG - removed background, Gr - gray scale conversion, Ap - different appearance of the same identities which differ in pictorial aspects such as facial expression and haircut, 45° and 90° - different images of the same identities, rotated at ~45° and ~90° view-points, respectively. b The impact of feeding the manipulated images presented in panel a to the DCNN on its correlation to neural distances in layers that originally showed a significant correlation in set 1 (layers conv5–3 and pool5). Error bars denote bootstrap s.e.m. on image pairs. Whereas low-level manipulations did not impact the correlation to neural data, higher level yet identity-preserving manipulations significantly reduced the correlation to chance. P values were defined as the proportion of correlations delta for shuffled image labels (1000 permutations) that exceeded the delta obtained from correctly labeled data. *p < 0.05, **p < 0.01. A 95% image bootstrap confidence interval test around the original correlation value yielded the same significance profile across manipulations. The differential sensitivity to image manipulations points to a high-level pictorial rather than personal-identity function of the iEEG recorded face neuronal groups