Abstract
Animals (including humans) are able to assess the quality of incoming sensory information and act accordingly while taking decisions. The computations underlying such ability are unclear. If neuronal activity encodes probability distributions over sensory variables, then uncertainty – hence confidence – about their value is explicitly represented and, at least in principle, readily accessible. On the other hand, if neuronal activity encodes point-estimates, then confidence must be obtained by comparing the level of the evoked response to fixed (possibly learned) criteria. To address this issue we developed a novel task allowing the behavioral read-out of confidence on a trial-by-trial basis. Each trial consisted of two consecutive decisions on whether a given signal was above or below some reference value, call it zero. The first decision was to be made on a signal uniformly drawn from an interval centered at zero. Correct/incorrect responses resulted into signals uniformly drawn from the positive/negative sub-intervals to be judged when making the second decision, and subjects were told so. The task reliably elicited confidence assessments as demonstrated by the finding that second decisions were more frequently correct than first decisions. We compared the ability of Bayesian and non-Bayesian observers to predict the empirically observed pattern of both first and second decisions. The non-Bayesian observer was designed to have discrete confidence levels instantiated by one or more second-decision criteria representing different levels of the evoked response. Different confidence levels resulted into different second-decision criteria. The non-Bayesian observer with two-three confidence levels systematically (over 9 subjects) outperformed the Bayesian observer in predicting the actual behavior. Hence, contrary to previous claims, confidence appears to be a discrete rather than continuous quantity. Simple heuristics are sufficient to account for confidence assessment by humans making perceptual decisions.
Meeting abstract presented at VSS 2016