TalkRL: The Reinforcement Learning Podcast Thomas Akam on Model-based RL in the Brain
Aug 4, 2025
Thomas Akam, a prominent neuroscientist from Oxford, leads the Cognitive Circuits Research Group and explores the fascinating interface of the brain and behavior. He discusses how the brain adapts actions through internal models, revealing insights into decision-making mechanisms. The conversation also highlights the metabolic costs of brain tissue, contrasting biological intelligence with AI. Additionally, Akam shares advancements in brain measurement technologies, paving the way for exciting strides in understanding cognitive processes.
AI Snips
Chapters
Transcript
Episode notes
Brain Models Drive Flexibility
- Thomas Akam focuses on how the brain generates flexible behaviors using rich internal models.
- He aims to understand how these models are learned and guide action selection for adaptive behavior.
Dopamine Signals Reward Prediction
- Dopamine neuron activity corresponds to a temporal difference reward prediction error in reinforcement learning.
- The cortex likely learns a hierarchical state representation of the world, which basal ganglia use to learn values and policies.
Model-Based and Model-Free Blur
- The model-based vs. model-free RL distinction maps to goal-directed versus habitual behavior.
- But even model-free systems operate on rich learned state representations, blurring the classic divide.