Spontaneous emergence of face-selective units in untrained deep neural networks

Abstract

In the visual cortex of primates, face-selective neurons are observed and considered the basis of facial recognition (Tsao, 2006). The observation of this type of intriguing neuronal tuning in the brain has inspired neuroscientists, raising important questions about its developmental mechanism — whether face-selective neurons can arise innately in the brain or require visual experience, and whether neuronal tuning to faces is a “special” type of function distinctive from tunings to other visual objects. Recently, model studies have used biologically-inspired deep neural networks (DNN) to predict the response of the ventral visual stream and provide insight into the underlying mechanisms of brain functions. Using a DNN model of the ventral visual stream, AlexNet, we showed that face-selective units can arise spontaneously in the complete absence of training. We measured the response of the last convolutional layer to a stimulus set of face and 5 non-face classes. We found face-selective units in the untrained network, where the face-units are defined as a unit that shows a significantly higher response to face images than non-face images. We found that their face-selectivity indices are comparable to those observed with face-selective neurons in the brain. The emergence of face-selectivity is robust across various conditions of initial randomization of the network. Next, to characterize the feature-selective responses of these face-selective units qualitatively, we reconstructed preferred feature images of individual units using a reverse-correlation method and a generative adversarial network algorithm. The obtained preferred feature images show that face-selective units are selective for a face-like configuration, distinct from units with no selectivity. Furthermore, we also tested whether the selective responses of these face units could provide reliable information with which to distinguish between faces and non-face objects. We confirmed that face-selective units enable the network to perform face detection. Intriguingly, we found that units selective to various non-face objects can also arise innately in untrained neural networks, implying that face-selectivity may not be a special type of visual tuning and that selectivity to various object classes can arise innately in untrained deep neural networks, spontaneously from random feedforward wirings. Overall, our results imply a possible scenario in which the random feedforward connections that develop in early, untrained networks may be sufficient for initializing primitive face-selectivity as well as selectivity to other visual objects in general.