Abstract
Perceptual decision-making involves placing an optimal criterion on the axis of encoded sensory evidence to maximize outcomes for choices. Optimal criterion setting becomes critical particularly when neural representations of sensory inputs are noisy and feedbacks for perceptual choices vary over time in an unpredictable manner. Here we monitored time courses of decision criteria that are adopted by human subjects while abruptly shifting the criterion of stochastic feedback to perceptual choices with certain amounts in an unpredictable direction and at an unpredictable point of time. Subjects viewed a brief (0.3 s), thin (.07 deg) annulus around the fixation and were forced to judge whether the annulus was smaller or larger than an unknown boundary. We estimated moment-to-moment criteria by fitting a cumulative Gaussian function to the data within a sliding window of trials that are locked to a shift in feedback criterion. Unpredictable shifts in feedback criterion successfully induced shifts in actual decision criterion towards an optimal criterion for many of subjects, but with time delay and amount of shifts varying across individual subjects. There were disproportionately more numbers of overshooters (reaching and then surpassing the optimal criterion required) than undershooters (subpar reach, with a significant anti-correlation with sensory sensitivity). To find a mechanism that generates these individual differences, we developed a dynamic criterion learning model by modifying a reinforcement learning model, which assumes that a criterion is adjusted every trial by a weighted discrepancy between actual and expected rewards.