[D] Is anyone interested in the RL ↔ Neuroscience “spiral”? Thinking of writing a deep dive series

A thought-provoking Reddit post is gaining traction by reframing the historical relationship between artificial intelligence and brain science. User Kooky_Ad2771 argues against the standard narrative of parallel development between reinforcement learning (RL) and neuroscience, proposing instead a dynamic 'spiral.' In this model, foundational ideas originate in one domain, are formalized into computational models in the other, and then loop back to inspire new experiments and theories, creating a continuous feedback cycle that has accelerated progress in both fields.

The proposed series, tentatively titled 'The RL Spiral,' aims to trace this intricate exchange. It would start with early 20th-century concepts like Edward Thorndike's 'law of effect'—a precursor to reward learning—and move through pivotal moments such as the formalization of Temporal Difference (TD) learning by Richard Sutton and Andrew Barto, which was directly inspired by neuroscientific discoveries about dopamine signaling. The series would also explore how modern RL architectures, like actor-critic models, find striking parallels in the brain's basal ganglia, and question what current frontiers like deep RL and world models can still learn from biological intelligence.

This conceptual shift from 'parallel' to 'spiral' underscores a more collaborative and interdependent history. It highlights that breakthroughs like understanding dopamine as a reward prediction error signal were not merely coincidences but results of active cross-pollination. The poster is crowdsourcing interest and suggestions, indicating a strong community desire to better understand the biological roots of AI and how future AI development might be guided by principles of neural computation.