Reproduce “Lessons from Neuroscience for AI: How integrating Actions, Compositional Structure and Episodic Memory could enable Safe, Interpretable and Human-Like AI” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rajesh P. N. Rao, Vishwas Sathish, Linxing Preston Jiang, Matthew Bryan, Prashant Rangarajan

The authors propose a neuroscience-informed architecture for foundation models combining action-conditioned prediction, hierarchical representations, and selective episodic memory. A world model predicts action outcomes, while separate policies choose actions and sensory feedback updates predictions. In a mouse visual choice task, cortical activity predicted upcoming movements before they occurred. This supports action-related signals, not the proposed architecture's effectiveness.

Key insights

  • While LLMs utilize predictive coding similar to the brain, they fundamentally lack the tight integration of actions, hierarchical compositional structure, and episodic memory critical for biological intelligence.
  • Intelligence in biological systems is inherently tied to the ability to act and interact with the environment, deriving meaning from sensory-motor contingencies, a crucial missing component in current AI.
  • The brain's energy-efficient and robust cognition stems from its hierarchical organization for compositional processing and dedicated memory systems for contextual understanding and consolidation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.