We propose a Sobolev-regularized classification framework for score difference estimation in diffusion models, resolving gradient divergence issues in standard transfer learning. The method improves ECG generation from PTB-XL to ICBEB2018 and achieves stronger robustness in small-sample regimes.
May 1, 2026
This AAAI 2024 paper studies stochastic nonconvex optimization beyond the standard globally Lipschitz-gradient assumption. It analyzes trust-region methods under a generalized smoothness setting motivated by deep learning and distributionally robust optimization, showing how first-order trust-region methods unify normalized and clipped stochastic-gradient updates and how second-order variants can obtain stronger stationarity guarantees with variance reduction.
Mar 1, 2024
Mar 1, 2024