DYNAMIC PRIOR LEARNING FACILITATES PERCEPTUAL DECISION-MAKING IN A HIGH-DIMENSIONAL WORLD

The world is rich with information, but it is rarely obvious how this information should be translated into actionable plans. Instead, we must learn a model of the world and use these prior beliefs to guide our inference. To study this problem, we developed a two-dimensional perceptual decision-making task where random dot motion stimuli vary in both a colour and a motion dimension, but where the currently relevant dimension is not signaled and switches after a variable number of trials. When colour is relevant (colour task), participants must decide whether there are overall more red or blue dots; and when motion is relevant (motion task), participants must decide whether the average direction of dot motion is upward or downward. Crucially, the two response buttons overlap between tasks (e.g., left button for “red” and “upward”), so participant must disambiguate response feedback (“correct” or “error) by combining their prior beliefs about which task is currently active with the available sensory evidence and in turn update their prior beliefs accordingly. Providing an additional window onto this process, participants are on each trial asked to report their belief about which task is active and their decision confidence within this task. Here, we present preliminary results indicating that participants’ prior beliefs facilitate perceptual dimension reduction and that they update their prior beliefs in proportion to the strength of the available sensory evidence. These dynamics, explained by a Bayesian reinforcement learning model, demonstrate the intertwined nature of prior learning and perceptual inference.