Functional compartmentalization and viewpoint generalization within the macaque face-processing system

Abstract

Primates can recognize faces across a range of viewing conditions. Representations of individual identity should thus exist that are invariant to accidental image transformations like view direction. We targeted the recently discovered face-processing network of the macaque monkey that consists of six interconnected face-selective regions and recorded from the two middle patches (ML, middle lateral, and MF, middle fundus) and two anterior patches (AL, anterior lateral, and AM, anterior medial). We found that the anatomical position of a face patch was associated with a unique functional identity: Face patches differed qualitatively in how they represented identity across head orientations. Neurons in ML and MF were view-specific; neurons in AL were tuned to identity mirror-symmetrically across views, thus achieving partial view invariance; and neurons in AM, the most anterior face patch, achieved almost full view invariance.

Fig. 1

Face selectivity in different parts of the macaque temporal lobe. (A) Inflated macaque left hemisphere (dark gray areas mark sulci, light gray-dark gray boundaries mark the middle of the bank within a sulcus) showing six regions in the temporal lobe of monkey M1 that responded significantly more to faces than to objects in fMRI experiments. Color scale indicates negative common logarithm of the P value. (B) Coronal (left) and sagittal (right) anatomical fMRI images showing the electrode descending into MF (located 3 mm anterior to the interaural line, AP (anterior-posterior) + 3 mm), AL (at AP + 12 mm), and AM (at AP + 19 mm), respectively, in monkey M1. Coregistered face-selective functional activation is overlaid on the fMRI images. (C to E) Face selectivity of neural population responses in ML/MF, AL, and AM, respectively. Shown are distributions of face selectivity indices (FSIs) (see SOM) for visually responsive cells in ML/MF, AL, and AM; dotted lines indicate FSI of ±0.33, corresponding to 1:2 and 2:1 response ratios to faces versus nonface objects. (F to H) Mean response time courses of three typical cells to the 128-image FOB set (top) and the 200-image FV set (middle) in ML/MF (F), AL (G), and AM (H), respectively. For clarity, responses are shown using a binary color scale. For the FV data, the first 25 rows are responses to 25 individuals looking to the left at full profile, the next 25 rows are responses to the same 25 individuals looking to the left at half profile, and so on; the eight different views of one example individual are shown on the right of (F). [Colored traces at the bottom of (F) to (H)]: Mean response levels to the 25 individuals at each head orientation, with the color corresponding to each view indicated in (F); s and v denote sparseness and view-invariant identity correlation coefficients, respectively (see SOM).

Fig. 2

Selectivity of neural populations in ML/MF, AL, and AM to faces varied in view and identity. (A to C) Population response matrices to the FOB image set (left) and to the FV set (right), for cells visually responsive (top) and nonresponsive (bottom) in ML/MF (A), AL (B), and AM (C), respectively. Responses are sorted from top to bottom by the first (C) or second (A and B) principal component of the FV responses. Data combined from monkeys M1 (recordings in left hemisphere) and M3 (recordings in right hemisphere) for ML/MF, and from monkeys M1 and M2 (recordings in right hemisphere) for AL and AM. The FOB matrix for ML/MF contains only 96 elements because only the first 96 images were presented during recordings in monkey M3.

Fig. 3

Tuning of AL cells to a head randomly rotated in three dimensions. (A) Illustration of stimulus head and three axes of rotation. (B) View tuning in four typical cells. Cells 1 and 2 here are the same as Cells 1 and 2 in Fig. 1G. Cell 4 did not respond to any of the FV stimuli. (Top) Tuning to updown angle versus left-right angle (responses averaged across pictureplane angle). (Middle) Tuning to up-down angle versus picture-plane angle (responses averaged across left-right angle). (Bottom) Tuning to picture-plane angle versus left-right angle (responses averaged across up-down angle). Marginal tuning curves are also shown (vertical lines indicate tuning peak positions).

Fig. 4

Population representations of face view and identity in ML/MF, AL, and AM. (A to C) Comparison of multidimensional scaling plots of responses to the FV image set in ML/MF, AL, and AM. Each plot shows the location of the 25 faces (indicated by numbers 1 to 25) at eight head orientations [indicated by eight colors, key in (D)] within the first two dimensions of the MDS space (corresponding Eigenvalues in fig. S3). (D to F) Population similarity matrices in the three face patches. A 200 by 200 matrix of correlation coefficients was computed between responses of all visually responsive cells to the 200 FV stimuli from ML/MF (N = 121 cells), AL (N = 189 cells), and AM (N = 158 cells). The correlation patterns do not change when only the first 121 cells of each patch are considered (fig. S13). (G) Sharpness of identity tuning in ML/MF, AL, and AM (top to bottom). Central tendencies of distributions of identity tuning half-widths (see SOM) were significantly different from each other (P << 0.001, MannWhitney U tests). (H) Distributions of head-orientation tuning depths (see SOM). Tuning depths close to 0 indicate broad tuning. These distributions are significantly different from each other (P << 0.001 for ML/MF versus AL and AL versus AM, Mann-Whitney U tests; P < 0.002 for ML/MF versus AM and AL versus AM, F test). (I) Evolution of view-invariant identity selectivity over time. View-invariant identity-selectivity index, computed over a 200-ms sliding response window beginning at the indicated time point, plotted for AM, AL, and ML/MF (solid curves). The dotted curves show the mean view-invariant identity-selectivity index over time computed from shuffled similarity matrices. The grayscale traces show the time course of the mean response to the FV stimuli across the population in each face patch. The substantial delay between the peak of the mean response to the FV stimuli and the peak of the view-invariant identity-selectivity index suggests that recurrent mechanisms are involved in the computation of viewinvariant identity.