Extended Data Fig. 10: Representation of familiar stimuli in face patch ML and additional analysis of repetition suppression-related signals.

a, Responses of cells to screening stimuli from six stimulus categories (familiar human faces, unfamiliar human faces, familiar monkey faces, unfamiliar monkey faces, familiar objects, and unfamiliar objects), recorded the face patch ML. Left, responses were averaged between 50 to 300 ms after stimulus onset (“full” response window). Right, same for a “short” window 50 to 125 ms. b, Similarity matrix of population responses for full response window (left) and short response window (right). c, Left: Average response time course across the ML population to each of the screening stimuli. Right: Response time course averaged across cells and category exemplars. Shaded area, SEM. Earlier arrow indicates the mean time when visual responses to faces became significantly higher than baseline (77.5 ms). Later arrow indicates the mean time when responses to familiar versus unfamiliar faces became significantly different (175 ms and 145 ms for human and monkey faces, respectively). Responses also diverged briefly at very short latency (95 ms and 105 ms for human and monkey faces, respectively). d, Population analysis comparing preferred axes for familiar versus unfamiliar faces. Same conventions as Fig. 2b. e-i, Same analyses for the ML population (n = 154 cells) as in Fig. 2c, d, Fig. 3b–d. j, Time course of mean pairwise neural distance (Euclidean distance) between familiar or unfamiliar faces computed using a 50 ms sliding time window, step size 10 ms, normalized by mean baseline (0–50 ms) distance between unfamiliar faces. Distances were computed using a subset of familiar and unfamiliar feature-matched faces (see Extended Data Fig. 8). k, Time course of face identity decoding accuracy for 30 familiar (blue) or unfamiliar (orange) feature-matched faces, computed using a 50 ms sliding time window, step size 10 ms. Shaded area, SEM. Half the trials were used to train a linear classifier and decoding performance was tested on the remaining half of trials; chance performance was 1/30.