Multifaceted Integration: Memory for Faces Is Subserved by Widespread Connections between Visual, Memory, Auditory, and Social Networks

Abstract

Our ability to recognize others by their facial features is at the core of human social interaction, yet this ability varies widely within the general population, ranging from developmental prosopagnosia to "super-recognizers". Previous work has focused mainly on the contribution of neural activity within the well described face network to this variance. However, given the nature of face memory in everyday life, and the social context in which it takes place, we were interested in exploring how the collaboration between different networks outside the face network in humans (measured through resting state connectivity) affects face memory performance. Fifty participants (men and women) were scanned with fMRI. Our data revealed that although the nodes of the face-processing network were tightly coupled at rest, the strength of these connections did not predict face memory performance. Instead, face recognition memory was dependent on multiple connections between these face patches and regions of the medial temporal lobe memory system (including the hippocampus), and the social processing system. Moreover, this network was selective for memory for faces, and did not predict memory for other visual objects (cars). These findings suggest that in the general population, variability in face memory is dependent on how well the face processing system interacts with other processing networks, with interaction among the face patches themselves accounting for little of the variance in memory ability.SIGNIFICANCE STATEMENT Our ability to recognize and remember faces is one of the pillars of human social interaction. Face recognition however is a very complex skill, requiring specialized neural resources in visual cortex, as well as memory, identity, and social processing, all of which are inherent in our real-world experience of faces. Yet in the general population, people vary greatly in their face memory abilities. Here we show that in the neural domain this variability is underpinned by the integration of visual, memory and social circuits, with the strength of the connections between these circuits directly linked to face recognition ability.

Trial registration: ClinicalTrials.gov NCT01031407.

Keywords: fMRI; face; human; memory; social.

Copyright © 2019 the authors.

Figures

Figure 1.

Individual centers of the seven ventral face ROIs. Colored vertices represent the location of the center of mass of each of the face patches, defined through the localizer for each of the 50 participants. All eight face patches were individually defined for each participant in the volume, and are projected here on the surface. ATL, blue; FFA, right: red, left: orange; OFA, right: green, left: yellow. Right amygdala shown in purple, left amygdala in cyan. pSTS on the lateral surface not shown. For all ROI analyses described later, ROIs were individually defined for each participant by drawing 6 mm spheres in the volume around these individually localized centers (for group localizer data, see Fig. 6).

Figure 2.

Face selectivity predicts CFMT score. Face–scene β during the two localizer runs averaged across all voxels in the eight individually defined face patches (bilateral OFA, bilateral FFA, bilateral amygdala, right STS, and right ATL) shown on the x-axis per participant, with CFMT scores shown on the y-axis. r = 0.41, p = 0.004, N = 46, as four participants were excluded because they had been given a different version of the localizer task (see methods).

Figure 3.

FFA–OFA correlation to memory. Correlation between right FFA and right OFA for each participant, plotted against their score on the CFMT. Note the wide range of correlations between right FFA and right OFA, despite the high mean correlation of these two regions (r = 0.72).

Figure 4.

Significantly predictive voxels and group defined ROI locations. Conjunction map of the eight predictiveness measure seed maps, which shows the voxels that were significantly predictive of CFMT scores in at least three seed maps. Colors indicate number of seed maps for which this voxel was significantly predictive, across both rest scans. Overlaid are the locations of all 23 group defined ROIs represented schematically by circles. Light green, somatosensory; blue, insula and anterior insula; dark purple, STG/auditory cortex; magenta, mid-STS; brown, pSPL; yellow, medial parietal; orange, cuneus; white, LOC; black, thalamus; cyan, hippocampus; red, parahippocampus; dark green, parahippocampus2. Insets, Locations of the ROIs in the volume, including separately for left and right hippocampus.

Figure 5.

Global connectivity. Map shows voxels whose global connectivity, i.e., average connectivity with all the other voxels, during rest is significantly correlated to performance on the CFMT, after corrections for multiple comparisons (orange/red voxels). Overlaid in green is the conjunction map from the previous analysis which was displayed in Figure 4, indicating voxels whose connectivity to at least three of the face ROIs was significantly predictive of performance on the CFMT. Note the high degree of overlap between the two analyses.

Figure 6.

Overlap with the face network. Map shows the group level Faces > Scenes contrast, with face selective voxels shown in red, and scene selective voxels in blue. Map is thresholded at an FDR corrected value of q < 0.05. Overlaid in green is the conjunction map from the previous analysis which was shown in Figure 4, using the face ROIs as seeds. Note the lack of overlap between these face memory predictive voxels and the face network.

Figure 7.

ROI pair correlations with behavior. Correlations of the connectivity between each of the ROI pairs (consisting of the 8 individually localized ROIs from the localizer, and the 23 group-defined ROIs from the seed analysis) and CFMT. Values indicate how predictive the correlation between each ROI pair is of performance on the CFMT. Blue denotes a nonsignificant correlation (determined through a permutation test, which was used to correct for multiple comparisons). Note the lack of significant predictive value of the correlation between the different face ROIs (top left).