Akash Raj Gunniya Prakash¹ (), Jongmin Moon¹, Robbe L. T. Goris¹; ¹The University of Texas at Austin

Perceptual interpretations of the state of the environment are not flawless. Humans know this and typically indicate to be more confident in accurate than in inaccurate interpretations. This introspective knowledge resembles metacognition of theoretically ideal decision-makers who leverage the normative framework of Bayesian inference to compute the probability that an uncertain inference is correct. Key to this strategy is that two distinct types of knowledge are used: one pertains to the noisy nature of sensory measurements, the other to structural regularities in the environment (i.e., prior knowledge). Whether human decision confidence is also based on both types of knowledge is a topic of debate. To shed light on this question, we conducted a perceptual confidence experiment using a task designed to distinguish between different confidence computation strategies. Each trial, subjects judged whether a noisy visual stimulus was rotated clockwise or counterclockwise relative to 45 degrees and reported their confidence in this decision (high vs low). We manipulated both sensory uncertainty (3 levels of orientation dispersion) and stimulus strength (10 different orientations). Critically, we ran the experiments in blocks of 120 trials and used this temporal structure to manipulate environmental regularities. Within a block, stimulus orientation was either sampled from a uniform distribution or from a bi-delta distribution. Subjects were cued about the current stimulus distribution and knew the task statistics well. Under these circumstances, prior knowledge has no influence on the perceptual decisions of a Bayesian decision-maker but drastically impacts the pattern of their confidence reports. Human perceptual decisions closely resembled the Bayesian prediction, but this was not true of their confidence reports. Specifically, perceptual confidence seemed blind to prior information. Together, these results suggest that perceptual confidence results from a computation that leverages knowledge about the noisy nature of perception, but ignores informative regularities in the environment.

Acknowledgements: NIH R01 EY032999