Testing the reward prediction error hypothesis with an axiomatic model.

Area of Science:

• Cognitive Neuroscience
• Neuroeconomics
• Computational Psychiatry

Background:

• Neuroimaging studies often link brain activity to predictions from complex models, such as reward prediction error (RPE) models in reinforcement learning.
• A key question is whether identified brain regions truly encode RPEs or merely show correlated activity.

Purpose of the Study:

• To formally test the entire class of RPE models against neural data using an axiomatic approach derived from economic theory.
• To determine if specific brain regions' activity is consistent with RPE models or if alternative explanations are necessary.

Main Methods:

• Applied an axiomatic framework rooted in economic theory to analyze neural activity data.
• Formally assessed whether the entire class of RPE models could account for observed neural signals across different brain regions.

Main Results:

• Neural activity in the striatum, medial prefrontal cortex, amygdala, and posterior cingulate cortex satisfied the necessary and sufficient conditions for RPE models.
• Activity in the anterior insula falsified the axiomatic model, indicating that no RPE model can explain the observed neural signals in this region.

Conclusions:

• The findings support the validity of RPE models for explaining neural activity in several key brain areas involved in reinforcement learning.
• The anterior insula’s activity is inconsistent with RPE models, suggesting it may encode outcome salience or other non-RPE related information.
• Formal, axiomatic approaches are crucial for rigorously testing entire classes of models in cognitive neuroscience as models and data proliferate.