Visual Attention Through Uncertainty Minimization in Recurrent Generative Models

Allocating visual attention through saccadic eye movements is a key ability of intelligent agents. Attention is both influenced through bottom-up stimulus properties as well as top-down task demands. The interaction of these two attention mechanisms is not yet fully understood. A parsimonious reconciliation posits that both processes serve the minimization of predictive uncertainty. We propose a recurrent generative neural network model that predicts a visual scene based on foveated glimpses. The model shifts its attention in order to minimize the uncertainty in its predictions. We show that the proposed model produces naturalistic eye movements focusing on informative stimulus regions. Introducing additional tasks modulates the saccade patterns towards task-relevant stimulus regions. The model’s saccade characteristics correspond well with previous experimental data in humans, providing evidence that uncertainty minimization could be a fundamental mechanisms for the allocation of visual attention.