Impact of conflicts between long- and short-term priors on the weighted prior integration in visual perception

Abstract

The prior distribution of values for a specific feature can be categorized as long- or short-term priors based on their respective learning durations. Studies have demonstrated that the visual system can integrate both priors through weighted averaging and then utilize the integrated prior to efficiently encode stimuli. It is unclear what determines the two priors' relative weights. To address this question, we arranged the orientations according to three distributions: natural, anti-natural, and natural with increased-amplitude distributions. The natural distribution mirrors the distribution of orientations in the natural world, so it does not conflict with the long-term prior; according to the Kullback-Leibler divergence analysis, the natural distribution had a higher conflict with the anti-natural distribution than with the natural distribution with increased amplitude. It was found that the cardinal bias — the orientation estimates are biased away from the cardinal orientations — was strongest in the natural distribution with increased amplitude but weakest in the anti-natural distribution. These results were accurately predicted by an efficient Bayesian observer model in which the prior is the weighted integration of the long- and short-term priors. Importantly, the weight of the short-term prior in the new prior decreased as the level of the conflict between the long- and short-term priors increased. Therefore, this study reveals that the visual system integrates the long- and short-term priors through weighted averaging, with the conflicting level between the two priors determining their relative weights in the integration prior. The integrated prior was used by visual systems to efficiently encode stimuli.