Abstract:

Bayesian decision theory explains perception in terms of encodings, priors, and losses, but it remains unclear how much of this structure can be inferred from behavior. We develop a unifying theory of perceptual biases, showing they decompose into three components: attraction to priors, repulsion from regions of high encoding precision, and regression away from boundaries. We derive a method to fit priors and encoding models to behavioral data, and we establish general conditions under which they are uniquely identifiable. These results also specify which experimental design features are necessary for identifiability, providing practical guidance for behavioral studies. Applying our framework to vision and decision making—including orientation, direction, color, and probability—we reveal common principles underlying diverse perceptual biases.