Reproduce “Brain in the Dark: Design Principles for Neuromimetic Inference under the Free Energy Principle” 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
Mehran H. Bazargani, Szymon Urbas, Karl Friston
The paper offers a roadmap for applying the Free Energy Principle to predictive-coding networks, with CPU-based PyTorch code. The method predicts sensory inputs from an internal model, then updates hidden-state estimates using prediction errors. In one predator–prey example, the trigonometric-flow model had lower free action than linear flow (423.21 versus 576.98), with similar MSEs (0.52 versus 0.53); this finding is limited to that setup.
Key insights
- Minimizing Variational Free Energy (VFE) enables neuromimetic AI to perform parameter, precision, and state estimation, balancing model accuracy with complexity (Occam's razor).
- Predictive coding provides a neurobiologically plausible mechanism for Bayesian filtering by continuously minimizing prediction errors across hierarchical layers.
- The choice of generative model dynamics significantly impacts inference quality, with models whose assumed dynamics align better with the underlying generative process yielding superior performance in terms of Free Action.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.