Where learning paths meet: Convergence and divergence of statistical and reinforcement learning

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Biology

Background:

• Organisms adapt to dynamic environments through learning, forming internal representations.
• Statistical Learning (SL) and Reinforcement Learning (RL) are complementary learning paradigms.
• RL focuses on goal-directed behavior and reward maximization via Reward Prediction Error (RPE).

Purpose of the Study:

• To compare RL and SL across historical foundations, objectives, computational principles, and neural implementation.
• To delineate the boundaries and interconnections between RL and SL.
• To integrate perspectives on adaptive learning mechanisms.

Main Methods:

• Comparative analysis of existing literature on RL and SL.
• Review of computational principles, including error signals like RPE and State Prediction Error (SPE).
• Examination of neurobiological underpinnings and neural network associations.

Main Results:

• SL extracts environmental structure without explicit rewards, potentially using SPE or associative learning.
• RL is driven by RPE to maximize rewards; Model-Based RL uses SPE to refine world models, overlapping with SL.
• RL is primarily associated with midbrain dopaminergic signaling.
• SL involves cortical and subcortical networks, including sensory areas and the hippocampus.

Conclusions:

• RL and SL represent distinct yet interconnected adaptive learning systems.
• Understanding their differences and overlaps provides insights into brain function.
• Further research can refine the delineation and integration of these learning processes.