Abstract
Visual confidence refers to the observer's ability to predict her performance level in a visual task. High confidence judgments can result from low stimulus uncertainty or high predicted performance. While uncertainty and performance are usually inversely related, we provide here a method to decouple this relationship, and determine whether confidence is better accounted for by our ability to estimate our own level of uncertainty or by our ability to correctly predict our perception. In order to decouple uncertainty and performance, we used the fact that internal uncertainty (represented by the spread of the posterior distribution in Bayesian models) is determined both by sensory uncertainty (the likelihood) and expected uncertainty (the prior). Therefore, different levels of internal uncertainties can be achieved by manipulating the prior expectation. Observers were placed in a dual task on motion direction perception (Barthelmé & Mamassian, 2010, PNAS). In the first task, they chose one out of two noisy random-dot kinematograms (the one moving upwards or the one moving downwards) as the one they were more confident that it would maximise their performance in the second task. In the second task, they discriminated whether the mean direction of the chosen kinematogram was slightly to the left or the right of the vertical. Unknown to the observers, the mean direction of the downward stimulus was sampled from a more focused distribution (high prior) than the upward stimulus. This manipulation led to less internal uncertainty for stimuli moving downwards, but at the same time worse expected performance for these stimuli (the reverse is true for stimuli moving upwards). Observers chose significantly more often the stimuli moving downwards in the choice task when the choice was between two stimuli with identical physical uncertainty. In conclusion, internal uncertainty, rather than expected performance, determined visual confidence.
French Agence Nationale de la Recherche.