Diverse Visual Experience Promotes Integrated and Human-Aligned Face Representations in Deep Neural Networks

Abstract

Humans are experts at recognizing faces, yet this expertise is not uniform: people perceive faces from familiar facial groups more accurately than those from unfamiliar ones--a phenomenon known as the Other-Race Effect (ORE). Diverse facial exposure mitigates this bias, but how it reorganizes face representations to support cross-group recognition remains unclear. To address this question, we combined controlled training variation in deep convolutional neural networks (CNNs) with analyses of representational geometry, lesioning, and human behavioral data. CNNs trained exclusively on either Asian or White faces reproduced ORE-like biases, whereas a CNN trained on both groups showed reduced bias and balanced recognition performance. Critically, representational similarity analyses and lesioning revealed largely overlapping feature spaces across groups, indicating an integrated representational organization. Trial-by-trial comparisons with human choices further showed that the dual-trained CNN best captured human face-matching behavior across both Asian and White participants, outperforming single-trained networks on cross-group trials. Together, these findings show that diverse visual experience promotes an integrated representational geometry that supports cross-group generalization, providing a computational account of how exposure diversity may shape human face representations.