LAUSR

Asking subjects to rate their confidence is one of the oldest procedures in psychophysics. Remarkably, quantitative models of confidence ratings have been scarce. The Bayesian confidence hypothesis (BCH) states that an observer’s confidence rating is monotonically related to the posterior probability of their choice. I will report tests of this hypothesis in two visual categorization tasks: one requiring rapid categorization of a single oriented stimulus, the other a deliberative judgment typically made by scientists, namely interpreting scatterplots. We find evidence against the Bayesian confidence hypothesis in both tasks. Model. Let s be the world state of interest and x a set of noisy visual observations that follow a distribution p(x|s). A Bayes-optimal observer would compute the posterior over s, denoted by p(s|x). We model the observer’s decision as a maximum-a-posteriori (MAP) estimate,   ˆ argmax | s s p s x  , and the