Bayesian Causal Inference Accounts for Multisensory Filling-In at the Blind Spot

Abstract

We asked three questions about multisensory perception across the physiological blind spot: (1) Does audiovisual integration persist without bottom-up visual input? (2) Does the brain adjust its sensory uncertainties and priors accordingly? (3) Are the underlying causal-inference computations preserved? Participants judged flashes and beeps in an audiovisual illusion presented across the blind spot or a matched control location. Responses were fit with a Bayesian Causal Inference (BCI) model, estimating sensory noise, numerosity priors, and causal-inference priors under multiple decision strategies evaluated using BIC. Illusions were robust at both locations, indicating preserved integration. Model fits showed higher visual uncertainty and broader prior expectations at the blind spot, while auditory precision and the causal prior remained stable. Thus, the computational architecture of causal inference is maintained, but its parameters flexibly adapt to local sensory reliability. These findings demonstrate that perceptual inference remains intact even in regions without retinal input, achieved by adjusting internal uncertainty rather than altering core multisensory computations.