ICLR Paper: Physics Theorems Reveal How Gradient Noise Shapes AI Representations

burny_tech · x · 2026-08-08

Researcher Liu Ziyin shared their ICLR 2024 work exploring the training dynamics of AI models from a theoretical physics perspective.

The study shows that gradient noise is a key determinant in how models learn representations. Drawing on fluctuation-dissipation theorems from physics, the research reveals that gradient noise, representations, and weights become mutually aligned during training. This serves as a compelling example of how physics can offer surprising predictions for AI model phenomenology.

Original post →

More from Research

Research channel →