Human Visual Search as a Deep Reinforcement Learning Solution to a POMDP.

Abstract

When people search for a target in a novel image they oftenmake use of eye movements to bring the relatively high acuityfovea to bear on areas of interest. The strategies that controlthese eye movements for visual search have been of substantialscientific interest. In the current article we report a new com-putational model that shows how strategies for visual searchare an emergent consequence of perceptual/motor constraintsand approximately optimal strategies. The model solves a Par-tially Observable Markov Decision Process (POMDP) usingdeep Q-learning to acquire strategies that optimise the trade-off between speed and accuracy. Results are reported for theDistractor-ratio task.