Diverse Visual Experience Promotes Integrated Representations and Mitigates Bias in Deep Neural Networks for Face Perception
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). Although the ORE is well established, it remains unclear how exposure diversity reshapes the underlying representations that give rise to it. Here, we combine deep convolutional neural networks (CNNs) with human behavioral data to investigate how training diversity influences the emergence and mitigation of recognition biases. CNNs trained exclusively on either Asian or White faces reproduced ORE-like biases, showing reduced performance for faces from unfamiliar groups. In contrast, a CNN trained on both Asian and White faces showed minimal bias, balanced performance across groups, and an integrated representational geometry spanning both groups. Critically, this model best captured human face-matching behavior across both Asian and White participants, outperforming single-trained networks on cross-group trials. These findings show how diverse visual experience reorganizes representational geometry to support cross-group generalization in artificial systems and provide computational insight into how diverse experience may shape more integrated face representations in humans.