View article

Here we introduce a simple actor-critic model of eye movements during category learning that we call RLAttn (Reinforcement Learning of Attention). RLAttn stores the rewards it receives for making decisions or performing actions, while attempting to associate stimuli with particular categories. Over multiple trials, RLAttn learns that a large reward is most likely when the values of the relevant stimulus features have been revealed by fixations to them. The model is able to approximate human learning curves in a common category structure while generating fixation patterns similar to those found in human eye tracking data. We additionally observed that the model reduces its fixation counts to irrelevant features over the course of learning. We conclude with a discussion on the effective role eye movements might play in bridging structural credit assignment and temporal credit assignment problems.