Abstract
Introduction: Bayesian models of visual perception and action presuppose that the visual system has access to accurate estimates of the probability of future events (priors). We examined how human observers would acquire such priors in a simple visual task with binary outcomes and compared their performance to beta-binomial Bayesian updating, the normative solution for such a task. Observers' successive priors were modeled as beta distributions. Methods: Observers sampled with replacement from visual sources that emitted black or white balls. Each sample was independent of all others and a source could be characterized by its probability p[B] of emitting a black ball. A total of 18 observers observed 50 samples taken one by one from the unknown source. We varied p[B] across subjects to take on values 0.1, …, 0.9. Task: Before drawing each sample, observers estimated p[B] and rated confidence. The dependent measure of greatest importance was the change in the estimates of p[B] from sample n to sample n+1 denoted Δn. Analysis: We assumed that the estimate of p[B] was the mode of the current prior. Then the beta-binomial updating model provided parameter-free predictions of successive estimates of p[B] for each trial as well as Δn. Results: We analyzed each observer's data separately. We found that observers updated their estimates from trial to trial, roughly following the pattern as the beta-binomial updating model. However, the actual weights observers gave to new information were markedly larger (factor of 2-3) than that predicted by the beta-binomial model with large disagreements between ideal and actual.
Meeting abstract presented at VSS 2017