From Perception to Algorithm: Quantifying Facial Distinctiveness with a Deep Convolutional Neural Network

Background: Face distinctiveness is a pivotal concept in face recognition, due to long-standing findings indicating that the rated typicality of a face is inversely related to its recognizability (e.g., Light et al., 1979). In traditional face space models (Valentine et al., 2016), distinctive faces are perceptually distant from the average or prototype face. Objectively measurable face spaces are available from deep convolutional neural network (DCNN), which are highly accurate at face recognition. Here we test whether DCNN models capture human-rated perceptions of face typicality, as well as other human-ratings (memorability, attractiveness, etc.). Method: We utilized FaceNet (Schroff et al., 2015, 2018), a pre-trained DCNN, to derive a distinctiveness index based on the average distance of a face to other faces in the DCNN space. First, we quantified distinctiveness for 418 male and 631 female faces using FaceNet's 512-dimensional embedding space and the cosine as a measure of distances. Second, we computed correlations between ach face’s distinctiveness and various human ratings of the same faces (Bainbridge et al., 2013). Results: Male faces classified as more distinctive are rated as less common (r = 0.27) and atypical (r = 0.21). Female faces classified as more distinctive are rated as less common (r = -0.35), more attractive (r = 0.37), more egotistical (r = 0.22), and more confident (r = 0.25) (all p < 0.001). Conclusion: DCNN provides a model of memory-related traits that relate to face distinctiveness across both genders. For female faces, distinctiveness also strongly correlates with personality and social attributes. This underscores the utility of DCNN models for understanding distinctiveness and face recognition. Current efforts focus on extending these findings using an alternate face distinctiveness model and another stimulus set to bolster the generalizability of our results.