Abstract
Humans have the tendency to perceive the world the way they expect it to be. These prior expectations usually reflect knowledge about the statistical structure of the world that, if learned correctly, can help to improve the accuracy of the percept. Recent results suggest that these prior expectations can also be self-generated by a preceding categorical or structural interpretation of the sensory evidence by the observer (Stocker & Simoncelli 2007; Ding et al. 2017; Luu & Stocker 2018). In this case, however, these self-generated expectations should decrease perceptual accuracy as sensory information is counted twice, leading to a form of confirmation bias. Why, then, do people use this self-consistent conditioning strategy? To answer this question, we characterized perceptual accuracy for different inference strategies. We considered a perceptual task where an observer estimates the value of a stimulus feature that belongs to one of two categories with overlapping distributions. Using model simulations, we demonstrated that although an observer that conditions feature estimation on the self-generated expectation of the feature category is generally suboptimal, the decrease in accuracy is not homogeneous across the entire feature range. In fact, compared to a full Bayesian observer, the accuracy is better for feature values where the uncertainty of a correct categorical assignment is relatively low. That is to say, if an observer is reasonably certain about the categorical assignment of the feature, estimation accuracy is improved by fully committing to that particular category instead of considering both categories according to their posterior probabilities. Furthermore, these benefits of conditioning are substantially increased if additional noise is deteriorating sensory information after the categorical commitment (e.g. during retention in working memory). Our results suggest that inference strategies that rely on self-generated expectations (i.e. confirmation bias) can be beneficial and maximize the accuracy.