LATERAL CONNECTIONS IN VISUAL HIERARCHY ENCODE STATISTICAL STRUCTURE TO SUPPORT PERCEPTUAL INFERENCE

Perception is thought to rely on probabilistic inference over a hierarchy of sensory representations. The function of lateral connections within this hierarchy remains unclear. We propose that lateral connections encode statistical coupling between stimulus features. When these couplings between features are strong, local recurrent interactions amplify support between compatible interpretations, producing reverberatory dynamics. We predict that this is the phenomenon that causes bistable perception, where two mutually exclusive percepts alternate despite an unchanging stimulus. We formalize this idea as approximate probabilistic inference in a generative model of the world, and illustrate it with the Necker cube. The resulting dynamics are those of a recurrent network, where coupling strength determines whether inference converges to a single percept or alternates between attractors, yielding unimodal or bimodal confidence reports respectively. We developed an experimental paradigm to test our predictions. Participants reported depth perception in a rotating cylinder stimulus where we independently manipulated feature coupling and sensory evidence. As predicted, strong and weak coupling induced bimodal and unimodal confidence reports, respectively. Furthermore, increasing coupling enhanced hysteresis, where percepts tended to persist even when evidence favored the opposite interpretation, a hallmark of bistable dynamics. Our model further explains increased interneuronal correlations and choice probabilities during bistable perception (as reported in macaque area MT by Wasmuht et al. 2019): strong lateral coupling generates shared variability among neurons encoding related features. Our results suggest that bistable perception is caused by the statistical structure of stimuli, and highlight the fundamental role of lateral connections in encoding such structure.