Learning enables organisms to adapt to a dynamic world by forming and updating internal representations of their environment. Statistical Learning (SL) and Reinforcement Learning (RL) offer complementary perspectives on this process. RL is fundamentally goal-directed, focused on maximizing rewards through Reward Prediction Error (RPE). SL extracts the statistical structure of the environment without explicit instruction or reinforcement. Model-Based RL additionally incorporates State Prediction Error (SPE) to refine an internal model of the world, overlapping with SL, which may use SPE or associative mechanisms devoid of error computations to extract structure. Neurobiologically, current research shows that RL is linked to midbrain dopaminergic signaling, whereas SL is supported by cortical and subcortical networks including early sensory areas and the hippocampus. This review compares RL and SL across their historical foundations, objectives, computational principles, and neural implementation, suggesting ways to better delineate the boundaries and interconnections between these two fundamental forms of learning.
Ferrari et al. (Mon,) studied this question.