Fig. 2: AM and PR cells use different axes to represent familiar versus unfamiliar faces.
From: Temporal multiplexing of perception and memory codes in IT cortex
a, Three example cells showing axis tuning. Top, mean response as a function of distance along the preferred axis. Green (yellow) dots denote responses to eight random unfamiliar (nine personally familiar) faces. Error bar, s.e.m. Bottom, responses to 1,000 unfamiliar faces, projected onto the cell’s preferred axis and principal (longest) orthogonal axis in the face feature space. Response magnitudes are colour coded. b, Population analysis comparing preferred axes for familiar versus unfamiliar faces. Distribution of cosine similarities between axes computed using 1,000 − 36 unfamiliar faces and 36 omitted unfamiliar faces (orange), and between axes computed using 1,000 − 36 unfamiliar faces and 36 familiar faces (blue) are shown. Preferred axes were computed using the top ten shape and top ten appearance features of presented faces. Inset, control experiment with 36 low-contrast faces in place of 36 familiar faces. c, Time course of cosine similarity between preferred axes for unfamiliar–unfamiliar (orange) and unfamiliar–familiar (blue) faces as in b. Arrowheads indicate when differences became significant (AM, 105 ms; PR, 175 ms; TP, 255 ms; one-tailed t-test, P < 0.001; AM, n = 134 cells; PR, n = 72 cells; TP, n = 197 cells; Supplementary Methods and Supplementary Table 1 give additional information about statistical tests). Shaded areas, s.e.m. d, Time course of linear decoding of facial features (Supplementary Methods). Shaded areas, s.e.m. Lighter colour, same analysis using stimulus identity-shuffled data (ten repeats). e, Example linearly reconstructed faces from short (120–170 ms) and long (220–270 ms) latency responses combining cells from both AM and PR. Reconstructions were performed using linear decoders trained on a large set of unfamiliar faces, similar to d (Supplementary Methods).
