Abstract
Bayesian observer models have been successful in explaining a large range of perceptual behavior. Their specification, however, is challenging because formal constraints on the likelihood function and the prior belief are typically missing. We propose a new model formulation based on the general assumption that an observer is optimally adapted to the statistical regularities of the perceptual environment. Specifically, we assume that Bayesian decoding operates on an efficient sensory representation, resulting in a likelihood function and prior belief of the observer that are jointly constrained by the stimulus distribution. A Bayesian observer model based on these assumptions makes two novel and counter-intuitive predictions. First, it predicts that perception is in many instances biased away from the peaks of the prior distribution, a prediction that is clearly at odds with the traditional Bayesian perspective. Second, it predicts that stimulus uncertainty differentially affects perceptual bias depending on whether the uncertainty is induced by internal (sensory) or external (stimulus) noise. We show that these predictions well match reported psychophysical measurements of perceived visual orientation and spatial frequency, two stimulus variables for which the natural stimulus distributions are known. Our work augments the standard Bayesian approach in perception by combining it with the idea of Efficient coding. This leads to a better constrained Bayesian observer model that is, at the same time, also more powerful in reconciling seemingly contradicting experimental findings.