Task-optimized models of sensory uncertainty reproduce human confidence judgments.

Sensory input is often ambiguous, leading to uncertain interpretations of the external world. Estimates of perceptual uncertainty might be useful in guiding behavior, but it remains unclear whether humans explicitly represent uncertainty in naturalistic settings, and whether any such representations are normatively correct. Progress has been hindered by the absence of stimulus-computable models that estimate uncertainty. We developed a class of task-optimized models that generate probability distributions over perceptual estimates. To assess whether human uncertainty representations align with the model's, we compared human confidence judgments, which might indirectly reflect uncertainty representations, to confidence judgments extracted from the model's uncertainty. In both sound localization and pitch perception, human confidence varied systematically, being lower for stimuli that produced more variable estimates across trials. Human confidence tracked model confidence across conditions, suggesting that human uncertainty representations accurately reflect the actual uncertainty of perceptual estimation. The modeling framework is extensible to other perceptual domains.