Using Deep Convolutional Neural Networks to Model Face Learning

How do we learn faces? Current prevailing theories, largely based on early computational models, suggest that various instances of a face we have encountered are incorporated into averaged representations (Burton et al., 2005). However, recent human data supports an alternative account of face learning: the cost-efficient encoding mechanism (Devue & de Sena, 2023). This theory explains a surprising differentiation in recognition performance for faces that are stable or variable in appearance. The former would be represented more coarsely than the latter, which would incorporate higher-resolution areas corresponding to intrinsic diagnostic features. Concurrently, since the introduction of the averaging theory, artificial facial recognition systems have improved immensely, largely due to the emergence of deep learning-based systems.
Here, we capitalized on deep convolutional neural networks to test predictions derived from the cost-efficient theory. Specifically, we examined whether these networks would also show the appearance-based differentiation in representations.
To do so, we produced a database of images of 38 actors classified as “stable” and “variable” based on independent ratings that we fed to the well-established “DeepFace” model (Taigman et al., 2014). We then compared the content of the facial representations built by the model for each category of actors. Preliminary principal component analyses suggest that variable faces are represented in a more complex fashion than stable faces, similar to the hypothesized way of representing faces in humans.
The present results suggest that an AI-based investigation of the cost-efficient framework holds merit and will serve as a baseline for future research.