Reproduce “Jaynes Machine: The universal microstructure of deep neural networks” 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
Venkat Venkatasubramanian, N. Sanjeevrajan, Manasi Khandekar
Jaynes Machine proposes that the magnitudes of connections in sufficiently wide neural-network layers arise from an equilibrium between utility, maintenance cost, and competition. A statistical-teleodynamics model derives a lognormal distribution for these magnitudes, and analysis of 798 layers across six convolutional and transformer models finds strong fits in many highly connected layers, with average R² values from 0.93 to 0.98 and similar scale parameters near 0.65 in large language models.
Key insights
- The theory frames weight distributions as a macroscopic consequence of many local connections reaching an arbitrage-like equilibrium, rather than requiring each weight to have an individually interpretable role.
- Lognormal behavior is most evident in highly connected layers; small layers suffer from sampling noise, while some extremely large layers may not have fully equilibrated during training.
- Large language models show especially consistent scale parameters, but the evidence supports an approximate regularity rather than a universal law for every architecture or training process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.