A developmentally inspired computational model of face recognition that learns continuously through generative memory replay

Abstract

Deep convolutional neural networks (DCNNs) have emerged as powerful models for human face recognition, capturing several hallmark behavioral phenomena such as the face inversion effect and the other-race effect. Yet, key differences remain between how DCNNs and humans, particularly infants, learn to recognize faces. In this study, we present a developmentally-inspired model that addresses three critical gaps between humans and standard DCNNs: (1) DCNNs are trained via one-time batch learning, whereas infants encounter faces gradually over time; (2) DCNNs are trained on thousands of face images with high variability of static images for each identity, while infants are initially exposed to a small number of identities, often limited to close caregivers, through a continuous, dynamic stream of visual input with lower within-identity variability; (3) The goal of DCNNs is to recognize untrained (unfamiliar) faces, while the goal of human face system is to recognize familiar faces. To better approximate developmental face learning, we introduce a model comprising three interacting modules: an Embedder - a DCNN trained to encode face identity continually on up to 10 identities, based on images taken from videos of a TV series; an Autoencoder used for generative memory-replay, to construct images of missing identities in each training phase; and a Memory module - stores identity-specific latent codes. Results show that this model achieves recognition performance comparable to batch-trained DCNNs. We propose this new model as a framework for studying the developmental and mature mechanisms of human face recognition.