Tahereh Toosi¹ (), Kenneth D. Miller¹; ¹Columbia University

One of the most debilitating aspects of psychosis is the wide spectrum of perceptual aberrations, ranging from altered contrast sensitivity to impaired facial emotion recognition and vivid visual hallucinations. While these symptoms appear disparate, we demonstrate that they likely emerge from a single computational principle: excessive integration of learned priors with sensory evidence. By leveraging a novel framework that combines image-computable object recognition networks with a generative inference algorithm, we provide the first unified computational account that not only explains normal vision but also bridges the gap between normal perception and psychosis-like visual dysfunctions. Unlike previous approaches where separate models were developed for normal vision and psychosis, our framework uniquely shows how a model of normal vision can transition to psychosis-like states by modulating a single parameter. This unification highlights the continuity between healthy and disordered perceptual processes and provides a novel computational explanation for psychosis symptoms emerging from disruptions in mechanisms supporting robust perception. Our approach makes three key contributions. First, we establish a formal mathematical link between robust object recognition and generative scoring, demonstrating why recognition networks can exhibit generative properties. Second, we trained robust object recognition models on naturalistic stimuli and incorporated feedback mechanisms during inference to simulate the effects of excessive prior integration. Third, we show that a single parameter controlling the integration of learned priors can reproduce the full spectrum of documented visual aberrations in psychosis, including excessive contrast sensitivity, systematic impairments in facial emotion recognition, and the emergence of complex visual hallucinations from spontaneous activity. These findings suggest that perceptual aberrations in psychosis arise from an imbalance in the integration of sensory evidence and learned priors, where excessive reliance on priors disrupts perceptual processing. This computational framework bridges low-level and high-level perceptual deficits, advancing our understanding of visual dysfunctions in psychosis.

Acknowledgements: T.T. is supported by NIH 1K99EY035357-01. This work was also supported by the Kavli Foundation and Gatsby Charitable Foundation GAT3708.