arxivcs.LGstat.ME2026-07-03
Statistically Meaningful Geometry (SMG) Beyond the Euclidean Paradigm, with Application to Generative AI
Bing Cheng, Yi-Shuai Niu, Howell Tong, Shing-Tung Yau
Conventional uniform convergence bounds and empirical risk minimization break down in massive over-parameterized models, such as large language transformers and biological sequence networks. With near-infinite unconstrained internal degrees of freedom, their optimization landscap…