Humans often look toward the eyes when viewing another face. While this behavior is known to carry significant social value in many cultures, its functional role in basic perceptual tasks remains unclear. Here, three studies explore a sensory optimization hypothesis of face recognition eye movement behavior: Observed human eye movements during face recognition reflect the brain's ability to learn optimized fixation strategies for the task-specific statistical distribution of information across the human face combined with the foveated nature of the visual system.

In the first study, eye movements were recorded during three common face-related tasks: identification, emotion recognition, and gender discrimination. On average, humans looked toward the middle of the face and slightly displaced downward from the eyes. Next, a novel forced-fixation paradigm was used to show that the choice of fixation has functional consequences. A model is then developed, the foveated ideal observer, that simulates the decreasing acuity and sensitivity of the visual system in the periphery. Simulations show that humans' preferred points of fixation are consistent with a performance-maximizing eye movement strategy.

The second study explores inter-individual variability in eye movement behavior. It is shown that individuals who tend to look away from the group-optimal point of fixation do not experience performance detriments. Instead, each individual enacts an idiosyncratic eye movement strategy that leads to maximal performance for their own visual system.

The third study investigates one possible mechanism by which humans might learn to optimize their eye movement strategies for face recognition. A major component of formulating an optimal strategy is knowledge of the statistical distribution of discriminating information. Here, novel faces with odd spatial distributions of information were learned by humans. This learning was fast, and revealed an initial bias towards features (e.g., the eyes) that carry high information content in natural faces.

Taken together, these studies describe how two major neural networks, the face recognition and eye movement systems, work in concert to learn and enact strategies optimized for the computationally challenging visual tasks associated with faces.