Estimating and integrating the uncertainty of naturalistic stimuli

Poster Presentation: Saturday, May 18, 2024, 8:30 am – 12:30 pm, Pavilion
Session: Decision Making: Perceptual decision making 1

Corey Plate1,2 (), Zoe Boundy-Singer3, Corey Ziemba1,2,3; 1Laboratory of Neuropsychology, Division of Intramural Research Program, National Institute of Mental Health, 2Laboratory of Brain and Cognition, Division of Intramural Research Program, National Institute of Mental Health, 3Center for Perceptual Systems, University of Texas at Austin

Perceptual decisions are accompanied by a sense of confidence. To usefully guide behavior, confidence should integrate all sources of decision-relevant uncertainty into a single decision reliability estimate. However, this metacognitive ability is not perfect, and can vary across individuals and contexts. Confidence is often studied in experiments where decision reliability is only varied through changing stimulus strength (e.g. only varying orientation in an orientation discrimination task), or not varied at all. Here, we sought to test the limits of the ability to appropriately integrate the uncertainty of complex, naturalistic stimuli into their perceptual confidence reports. We created a set of synthetic texture stimuli matched to the statistics of natural images and asked participants to judge in which direction their dominant orientation was rotated from vertical. Subjects reported their decision and binary confidence level with a single button press. Accuracy and confidence reports were incentivized by rewarding a large number of points for high confidence correct responses, but a substantial loss of points for high confidence errors. To control decision reliability, we varied both the mean and variance of orientation energy in the stimulus. We additionally adjusted the presence of higher-order, naturalistic pixel correlations, whose strength determines neural signatures of uncertainty in visual cortical regions downstream of V1 without affecting the orientation content. We fit the responses of subjects with the CASANDRE model for confidence, allowing us to assess whether different sources of uncertainty were appropriately integrated into a single confidence value. Our results indicate that subjects generally demonstrate the ability to represent and integrate naturalistic uncertainty into their perceptual confidence reports, but with some subjects exhibiting striking and idiosyncratic failures.

Acknowledgements: Grant: ZIA MH002928, Grant: NSF GRFP, Grant: NIH Grant EY032102