Abstract
Many perceptual effects have been successfully characterized within a Bayesian framework. Previously Stocker and Simoncelli (2006) used this framework to model sensory adaptation as an increase in the reliability of the measurement around the adaptor, leading to an asymmetrical likelihood function and the commonly observed aftereffects of repulsion and enhanced discriminability. We explored a Bayesian model of normalization aftereffects, in which the adapting stimulus appears more neutral or unbiased. This renormalization has been found for many stimulus attributes. We assumed the attribute is encoded by a uniform distribution of selective, labeled channels, and derived the likelihood functions from the pre- and post- adaptation channel responses. The prior corresponded to an unbiased stimulus and thus equal responses across the channels. Adaptation to a biased stimulus (e.g. too weak or strong a response in a subset of channels) leads to a compensatory gain change that restores an unbiased response for the adaptor. This corresponds to a change in the mean (along the intensity axis) and width (along the attribute axis) of the likelihood function, and accounts for a wide range of adaptation effects including changes in sensitivity and contrast, renormalization, and both repulsive and attractive aftereffects when adapting to stimuli that are too strong or weak, respectively. The model also provides a principled prediction for the magnitude of the adaptation. By this account, adaptation updates the likelihood function in order to match a fixed prior, equivalent to the visual system trusting its assumptions and error-correcting its measurements.