Conference on Cognitive Computational Neuroscience

When we enter a room, perception is challenged to convert sensory data into complex mental representations such as scene geometry. This leaves our percepts both strikingly sparse but also structured so as to support the flexibility of cognition. Here, we apply a new formal model of attention, adaptive computation, to reveal patterns of spatial attention and geometric selectivity in the context of perception of indoor scenes. The model uses the goal of navigating to a visible exit to guide selective processing over a multi-granular scene geometry model, which can represent regions in the room at different levels of resolution. Together the components of goal-driven processing and multi-granular scene states enables the efficient investment of computational resources to resolve geometry relevant to navigational affordances. In the context of a change detection task, we show that both goal-directed attention and multi-granular geometry representations are critical to explaining human responses. Together, adaptive computation and multi-granular geometry representations form powerful computational tools for studying scene networks.