Working Memory Load Strengthens Reward Prediction Errors.

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Reinforcement learning (RL) typically models instrumental learning as a single process using reward prediction errors (RPEs).
• This view overlooks the roles of distinct memory and decision-making systems in even simple learning tasks.
• Working memory (WM) is increasingly recognized as a crucial cognitive function in human learning.

Purpose of the Study:

• To investigate the interaction between working memory (WM) and incremental reinforcement learning (RL) in guiding human learning.
• To determine how WM load affects neural signals associated with RL.
• To explore the relationship between individual differences in WM and learning strategies.

Main Methods:

• An fMRI experiment was conducted with participants learning stimuli under varying WM load conditions.
• Behavioral data were collected and analyzed using computational modeling to distinguish between WM and RL processes.
• fMRI data were analyzed using model-based approaches to identify neural correlates of learning and their modulation by WM.

Main Results:

• Behavioral and computational models indicated that learning is a mixture of a fast, capacity-limited WM process and a slower RL process.
• fMRI results showed that RPE-sensitive signals in the striatum and lateral prefrontal cortex were reduced when the learning task was within WM capacity.
• The degree of neural interaction between WM and RL signals correlated with individual differences in using WM for behavioral learning.

Conclusions:

• Human learning is not a monolithic RL process but involves an interaction between WM and RL systems.
• WM capacity can modulate neural signals associated with RL, specifically RPEs in key brain regions.
• These findings highlight the integrated nature of cognitive systems in guiding adaptive behavior and learning.