The Local Learning Coefficient: A Singularity-Aware Complexity Measure

Abstract

This paper introduces the Local Learning Coefficient (LLC), a singularity-aware measure of model complexity grounded in singular learning theory. We provide scalable estimators using SGLD and demonstrate how LLC reveals inductive biases in deep learning. An earlier version of this work was released on arXiv under a different title and slightly different formulation emphasising the mathematics.

Publication
The 28th International Conference on Artificial Intelligence and Statistics (AISTATS 2025)