A Common Face Space? - High/Low Variability Priming [dataset]

Both humans and Deep Convolutional Neural Networks (DCNN) are highly accurate at recognizing (familiar) faces, even though the same face can look surprisingly different in varying conditions. DCNNs achieve this by representing faces in a high-dimensional space, clustering different images of the same identity together while separating them from images of different identities. However, it is unclear whether human face representations are structured by a similar face space, incorporating both within- and between-identity variability. Here, we combined human event-related brain potential (ERP) data in four repetition priming experiments with DCNN-based analyses to approach this question. We found that both humans and DCNNs show larger priming for more similar, “low variability” relative to less similar, “high variability” prime/target pairs using naturalistic, ambient images as primes (Experiment 1). Moreover, this effect was not based on low-level image properties (Experiment 2). Further experiments, using face inversion (Experiment 3) and cropping images to inner facial features (Experiment 4), confirmed that the low-variability advantage in DCNNs is driven by facial identity information. Crucially, a corresponding effect in human ERP data was not observed when prime images were restricted to the face itself, i.e., to inner features. These results question the assumption that human face representations are organized in a common space that incorporates both within- and between-identity variability. They further inform recent debates about the suitability of DCNNs as models of human visual cognition by demonstrating a critical difference between human and DCNN face representations.