Narrowing of the Cone-of-Direct Gaze Through Reinforcement Learning

The Cone of Direct Gaze (CoD) is described as the range of eye gaze deviations over which an observer reports gaze as being directed towards them. The CoD has been found to narrow with age across childhood (Mareschal et al. 2016). We investigated whether reinforcement learning, so critical in shaping eye gaze responses in infancy, was able to account for the emergence of a CoD and its narrowing in childhood. To this end, we adapted Triesch et al.'s (2006) reinforcement learning model by (1) defining a topology over object locations, and (2) introducing opponent non-linear reward profiles for looking at objects and caregivers. In Simulation 1 we show that these modifications give rise to a functional CoD in which there is reduced eye gaze following and increased fixation on the caregiver for locations with a small caregiver eye gaze eccentricity. In Simulation 2 we show that the width of this effect reduces with learning, suggesting that developmental decreases in the CoD may be driven by reinforcement learning. In Simulation 3 we explore how changes in model parameters can explain the CoD in high anxiety populations. Finally, the model provides one way of unifying the developmental gaze-following and CoD literatures, until now considered largely independent.