Abstract
Within a few days after birth, a human infant begins to exhibit sophisticated visual skills. Primary amongst these is the ability to preferentially orient towards faces. Precisely identifying the endogenous and exogenous factors which facilitate this unsupervised face learning remains an unsolved challenge. Here we investigate one specific factor: the acuity of the infant's visual system. It is known that a newborn's visual acuity is quite poor. We examine how acuity interacts with face learning, and whether reduced acuity might in fact be beneficial for acquiring a robust face concept. We collected several short sequences of simulated parental interactions from a baby's perspective. This was accomplished via a small video-camera mounted on a life-sized baby doll's forehead. We then created acuity variants of these sequences by convolving them with diffusing filters of a range of different sizes. These sequences were then provided as input to a simple computational learning system that discovers clusters of salient image fragments without supervision. To demonstrate recognition, the clusters are then used as templates for detecting regions in new sequences/images that correspond to the learned concepts. We find that this procedure is effective for basic concept learning. Interestingly, our computational simulations reveal that learning performance is impaired with high-resolution inputs relative to low-resolution. The ‘concepts’ learned at high-resolution are fragmentary and do not capture the gestalt of faces. It appears, therefore, that the poor acuity of infants might indeed be advantageous for early face concept learning. An interesting implication of these findings lies in the domain of autism. Children with autism have been reported to have difficulties in orienting to faces as infants. Interestingly, recent work has found a strong correlation between autism and markedly elevated acuity. Together with our computational results, these findings might provide a partial explanation for impairments in face-learning in autism.