Renormalization-Group Theory Quantifies Grokking Transitions in Neural Networks

This research applies renormalization-group (RG) theory to describe the transition from memorization to generalization in deep networks. By using weight norms as order parameters and constructing a coupling matrix via loss expansion, the authors provide a physical framework for understanding grokking dynamics.