Common attentional mechanisms, but distinct computational dynamics, across Statistical and Reward Learning

Abstract

Past experience guides attention through statistical learning (SL) and reward learning (RL). Yet, whether they do so through similar neural and computational mechanisms remains unclear. In a first EEG visual-search experiment, we tested whether RL establishes latent spatial representations, previously demonstrated for SL. Afterwards, we directly compared RL and SL in a double-session experiment. Using multivariate decoding, Experiment 1 showed that high-reward locations could be decoded before target onset following a task-irrelevant, spatially-unspecific visual perturbation. Experiment 2 revealed that SL and RL established comparable neural representations, and similarly facilitated target selection and evidence accumulation (drift-rate). However, combining Rescorla-Wagner and Drift Diffusion Model revealed slower learning-rate for RL than SL. Pupillometry further tracked spatially selective expectation violations during SL, but reflected increased arousal during RL. Together, our results reveal how the brain routes distinct computational algorithms (probability estimation versus value caching) into a unified attentional priority map.