Abstract
Although confidence is commonly believed to be an essential element in decision making, it remains unclear what gives rise to our sense of confidence. Recent probabilistic theories propose that one's confidence is computed, in part, from the degree of uncertainty in sensory information. If confidence indeed reflects the imprecision of perceptual evidence, then greater levels of confidence should predict 1) less variable behavior, and 2) smaller biases in perception, as both behavioral variability and perceptual biases are linked to uncertainty. Here, we test these predictions using a combination of psychophysics and computational modeling. Participants viewed a stimulus that consisted of an array of 36 gabor patches, and reported both the mean orientation of the array, and their confidence in this estimate. Patches were variable in orientation (drawn from a Gaussian distribution), and five noise levels were used to parametrically manipulate uncertainty (s.d. = 0.5, 2, 4, 8 and 16). Corroborating the first prediction, we found that for a given stimulus orientation, confidence reliably predicted behavioral variability. Specifically, orientation estimates were more precise with higher confidence, both across and within levels of orientation noise. Surprisingly, however, the results deviated from our predictions when comparing between stimulus orientations: although orientation judgments were more accurate for cardinal orientations (a phenomenon known as the oblique effect), confidence was higher for oblique orientations. In addition, we observed no reliable link between confidence and the magnitude of behavioral biases. Rather than being consistent with Bayesian decision theory, we argue that these results are better explained by the ability of observers to perceive the degree of orientation noise in the stimulus – a heuristic to confidence.
Meeting abstract presented at VSS 2018