Abstract
What mechanisms underlie confidence ratings in perceptual tasks? In a two-choice discrimination task, the observer can potentially acquire two relevant sorts of information: evidence for, and evidence against, making a particular perceptual decision. For instance, deciding that a grating is tilting left rather than right can be driven either by strong evidence for a left tilt or weak evidence for a right tilt. Optimal performance requires perceptual and confidence decisions to take both sources of information into account. However, Zylberberg et al. (2012) found that confidence decisions, but not perceptual decisions, are insensitive to evidence against a selected perceptual choice. We develop a signal detection theory (SDT) model of this insensitivity to negative evidence and show that the model predicts a novel and counterintuitive dissociation. Typically, increases in perceptual task performance are accompanied by increases in metacognitive performance (Galvin et al., 2003; Maniscalco & Lau, 2012). However, if this model is correct, then there should be cases where increasing task performance is accompanied by decreasing metacognitive performance. Empirical tests confirm that this dissociation can be observed in human subjects, providing further support for the hypothesis that confidence ratings are insensitive to negative evidence. This model helps explain why human subjects are often metacognitively suboptimal. It also allowed us to develop new methods for experimentally manipulating subjective confidence without changing perceptual sensitivity. However, these dissociations and suboptimalities in metacognitive performance can be partially remediated in some (but not all) observers if they (1) evaluate task performance by wagering points rather than by rating confidence, and (2) are provided consistent feedback about perceptual and metacognitive performance. This suggests that the blindness of metacognitive mechanisms to negative evidence may not completely reflect a fundamental processing limitation, but rather may partially reflect a poor but correctable usage of decision strategies or cognitive resources.
Meeting abstract presented at VSS 2013