Abstract
Perceptual judgments are often associated with a feeling of confidence, which is typically measured by collecting a confidence judgment after each perceptual decision. However, this single decision confidence is potentially quite different from the more general feeling of confidence that one has when facing a task. To address this issue, we designed a new global confidence task and found that observers can make a “global” confidence judgment over multiple perceptual trials (Lee et al., 2014, VSS). In the present study, we ask whether the global confidence system is capable of integrating information and delineate some characteristics and limits of such global confidence judgments. Observers (n=9) completed two small sets of first-order perceptual judgments (orientation-discrimination task: clockwise vs counterclockwise), one set after another. Both sets contained an equal number of trials (m = {1, 2, 4, 8}). After completing these trials, they performed a second-order 2IFC task (Barthelmé & Mamassian, 2009, 2010; de Gardelle & Mamassian, 2014): they had to choose the set for which they thought they had given more correct perceptual responses. Expected first-order performance for each set was targeted at d’=1 and d’=2, respectively (order randomized). If observers estimated global confidence from single perceptual decisions, the second-order confidence choice proportions would remain constant for different set sizes. On the contrary, we found that, the larger the set size (m), the more frequently observers chose the set with higher target d’ (i.e., more accurate confidence judgment). This suggests that observers integrate confidence information over trials when making global confidence judgments. Furthermore, when compared to the benchmark of an “ideal integrator”, we found that integration efficiency was close to ideal when m=2, but dropped significant when m>2. Finally, from an analysis of response times, this inefficiency seems in part related to a recency effect of confidence integration.
Meeting abstract presented at VSS 2015