How does the human brain handle complex situations, like navigating through traffic in busy areas? Psychologists and neuroscientists propose that the brain creates causal models of the world, running mental simulations to plan and execute actions. This idea aligns with the concept of Reinforcement Learning (RL), a system developed by computer scientists to understand human thinking and decision-making.

In a recent study published in Neuron, researchers delved deeper into RL’s neural architecture by employing functional magnetic resonance (fMRI) to compare their algorithmic theory with real-world brain imaging. The goal was to better understand how RL plays out in the brain and potentially improve RL algorithms in artificial intelligence.

By studying 32 volunteers playing Atari-style video games while connected to fMRI scanners, the researchers observed activity theory-based models in the prefrontal cortex, with theory updates occurring in the posterior cortex. Surprisingly, they found evidence of theory-based models in the inferior frontal gyrus, previously associated with learning causal rules. Additionally, the occipital cortex and ventral pathway, essential for visual processing, were involved in model updating.

Moreover, the directional flow of information in the brain during game play revealed intriguing insights. Initially, information seemed to flow top-down, originating from the model stored in the prefrontal cortex and flowing down to posterior visual regions. However, when there was a discrepancy or update, the pattern flipped, and information flowed bottom-up, from posterior regions to frontal regions.

This research has implications for understanding human decision-making and learning, which could lead to advancements in RL. By applying these concepts to fields like self-driving cars, researchers hope to improve the efficiency of handling complex environments and enhance human-like performance in various domains.

By Impact Lab