A neural basis for flexible perceptual inference

Abstract

What seems obvious in one context can take on an entirely different meaning if that context shifts. While context-dependent inference has been widely studied, a fundamental question remains: how does the brain simultaneously infer both the meaning of sensory input and the underlying context itself, especially when the context is changing? Here, we study flexible perceptual inference--the ability to adapt rapidly to implicit contextual shifts without trial and error. We introduce a novel change-detection task in dynamic environments that requires tracking latent state and context. We find that mice exhibit first-trial behavioral adaptation to latent context shifts driven by inference rather than reward feedback. By deriving the Bayes-optimal policy under a partially observable Markov decision process, we show that rapid adaptation emerges from sequential updates of an internal belief state. In addition, we show that artificial neural networks trained via reinforcement learning achieve near-optimal performance, implementing Bayesian inference-like mechanisms within their recurrent dynamics. These networks develop flexible internal representations that enable real-time adjustments to the inference model. Our findings establish flexible perceptual inference as a core principle of cognitive flexibility, offering computational and neural-mechanistic insights into adaptive behavior in uncertain environments.