A deep reinforcement learning account of competition and integration of visual and goal vector signals for spatial navigation

Abstract

Abstract Integrating different sources of information is essential to successful spatial navigation. For instance, animals often rely on a combination of visual impressions, self-motion, olfaction, and other signals to navigate towards a goal. This is especially important when navigating in uncertain environments, where switching from one source of information to another or integrating multiple sources of information may be required to make navigation decisions. We propose a deep reinforcement learning model of the interaction of visual and goal-vector signals based on reinforcement learning and use it to study behavior and spatial representations. We show that the nature and degree of signal noise strongly influences how the signals drive behavior and spatial representations. Our model also demonstrates that the ability to navigate using each information source independently, in addition to integrating them, is crucial to successfully navigating in uncertain environments. Counterintuitively, our model shows that when one of the signals is removed, navigation may be improved if the remaining signal is reliable and sufficient to navigate. However, this improvement comes at the expense of robustness. Our modeling results demonstrate that combining redundant sources of information in biological systems is far more complex than suggested by sensor fusion in the engineering context.