Towards a dual learning system in humans: sensory and category information cause opposed biases in behavior

Introduction: Statistical properties of stimuli can bias evidence integration in decision making, managing a delicate balance between recency and primacy. These are modulated by the ratio of sensory information (stimulus perceptual strength; SI) vs category information (frequency in which pieces of evidence favor a response option; CI). We propose learning integrates these operations mainly through adjustments of the integration timescale and decision urgency, in interaction with reward. Hence, our goal was to study how integration biases arise from the features of evidence and rewards. Methods: 105 healthy university students (86% females) performed two decision-making tasks: preference trials (not shown) and Clouds-of-Dots trials. Participants identified the dominant color of a sequence of 10 static dot clouds (red or blue) with a right-left bottom press, and then rated confidence (1–4 Likert scale). Virtual coins were given if correct and lost if too slow. We manipulated both SI (Nred - Nblue dots per trial) and CI (ratio of red vs blue frames per trial). Analyses included reverse correlation to compute integration kernels. We designed a new reinforcement learning drift diffusion model, still being fit, whose parameters vary with SI and CI via adjustments in integration timescale and decision urgency. Results: Recency increased with SI/CI ratio, total information (SI+CI) and performance. This suggests these biases depend on specific statistical traits of evidence and task difficulty. Discussion We successfully uncovered how people learn exploiting recency-primacy balance depending on sensory and category information. Our model predicts these biases are produced by the online adjustment of different adaptive processing mechanisms.