Extended Data Fig. 9: Control analyses confirming axis robustness.

a, Top, Row 1: population analysis of preferred axes for familiar versus unfamiliar faces; same conventions as in Fig. 2b except 30 familiar and 30 unfamiliar feature-matched faces were used (see Methods and Extended Data Fig. 8). Row 2: time course from the same analysis; same conventions as in Fig. 2c. Shaded area, SEM. Note that new feature-matched 36 familiar and 36 unfamiliar faces were used for TP, thus the result shown in Fig. 2c for TP is already perfectly feature matched, and is replicated here for comparison. Middle, same analysis as in Fig 2b, c except a subset of neurons showing significant axis tuning were used. Shaded area, SEM. Bottom, same analysis as Fig. 2b,c except the preferred axes were computed using linear regression rather than spike-triggered averaging (see Supplementary Methods). Shaded area, SEM. b, Top: scatter plot of 20 feature sensitivities (see Supplementary Methods) from 134 AM cells and 72 PR cells, for familiar (y-axis) and unfamiliar (x-axis) faces. The dots in the blue rectangles (corralling points for which sensitivity to the familiar feature goes to ~0) indicate loss of tuning for familiar faces in some cells, while the dots in the red rectangles indicate gain of tuning. Bottom: Distribution of feature sensitivity values for familiar and unfamiliar faces. This shows that on average, sensitivity for familiar faces was larger than that for unfamiliar faces. c, Top: explained variance for responses to 36 unfamiliar (y-axis) or 36 familiar (x-axis) faces using unfamiliar axis (fitted on 1000 - 36 faces) with linear output function (each dot is one cell, n = 134 cells for AM and n = 72 cells for PR). Middle: explained variance for responses to 36 familiar faces using unfamiliar axis with linear output function (y -axis) or a logistic output nonlinearity (x-axis); the latter values are only slightly higher. Bottom: explained variance for responses to 36 unfamiliar faces using unfamiliar axis with linear output function (y-axis) or 36 familiar faces using axis model with a logistic output nonlinearity (x-axis). The slight increase in explained variance obtained by applying a logistic output nonlinearity cannot undo the decrease caused by axis change (however, explained variance is similar using familiar axes for familiar responses and unfamiliar axes for unfamiliar responses, Extended Data Fig. 3g). d, Comparison of average response time courses in AM and PR to the exact same set of familiar and unfamiliar stimuli, presented in two different temporal contexts. Scatter plot: average over time window [100 300] ms (AM, N = 80 cells; PR, N = 70 cells). Top: Responses to 9 personally familiar and 8 unfamiliar monkey faces presented as part of screening stimulus (experiment 1). Bottom: responses to the same set of stimuli presented as part of thousand face stimulus (experiment 2). Shaded area, SEM. e, Correlation in rank order (Spearman correlation) of neuronal responses to personally familiar face stimuli at short or long latency between split halves of trials (y-axis, correlation values averaged across experiments 1 and 2) is plotted against correlation between rank order of the same faces between experiments 1 and 2; each dot represents one cell (AM, N = 80 cells; PR, N = 70 cells; TP, N = 197 cells).