Multiple associative structures created by reinforcement and incidental statistical learning mechanisms.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Reinforcement learning (RL) explains how prediction errors update beliefs, but less is known about the resulting knowledge structures.
• Learning about the world's structure can be driven by explicit reward or incidental experience.

Purpose of the Study:

• To contrast associative structures formed through reinforcement versus statistical experience.
• To investigate the neural correlates of different learning processes and resulting knowledge representations.

Main Methods:

• Functional magnetic resonance imaging (fMRI) using Blood-Oxygen-Level-Dependent (BOLD) contrast in human volunteers.
• Analysis of neural representations associated with reinforcement learning and statistical learning of task structures.

Main Results:

• Rigid representations of rewarded sequences were found in the temporal pole and posterior orbito-frontal cortex, constructed backward from reward.
• Medial prefrontal cortex and a hippocampal-amygdala border region showed flexible statistical knowledge alongside reward-related knowledge.
• The ventral striatum encoded prediction error responses but not the complete learned task knowledge.

Conclusions:

• Task knowledge is derived from multiple learning processes operating at different timescales.
• Partially overlapping and specialized anatomical regions support these distinct learning processes.
• Neural representations reflect the distinct mechanisms of reinforcement and statistical learning.