English

Decentralized SGD and Average-direction SAM are Asymptotically Equivalent

Machine Learning 2023-11-10 v5 Computers and Society Distributed, Parallel, and Cluster Computing Systems and Control Systems and Control Machine Learning

Abstract

Decentralized stochastic gradient descent (D-SGD) allows collaborative learning on massive devices simultaneously without the control of a central server. However, existing theories claim that decentralization invariably undermines generalization. In this paper, we challenge the conventional belief and present a completely new perspective for understanding decentralized learning. We prove that D-SGD implicitly minimizes the loss function of an average-direction Sharpness-aware minimization (SAM) algorithm under general non-convex non-β\beta-smooth settings. This surprising asymptotic equivalence reveals an intrinsic regularization-optimization trade-off and three advantages of decentralization: (1) there exists a free uncertainty evaluation mechanism in D-SGD to improve posterior estimation; (2) D-SGD exhibits a gradient smoothing effect; and (3) the sharpness regularization effect of D-SGD does not decrease as total batch size increases, which justifies the potential generalization benefit of D-SGD over centralized SGD (C-SGD) in large-batch scenarios. The code is available at https://github.com/Raiden-Zhu/ICML-2023-DSGD-and-SAM.

Keywords

Cite

@article{arxiv.2306.02913,
  title  = {Decentralized SGD and Average-direction SAM are Asymptotically Equivalent},
  author = {Tongtian Zhu and Fengxiang He and Kaixuan Chen and Mingli Song and Dacheng Tao},
  journal= {arXiv preprint arXiv:2306.02913},
  year   = {2023}
}

Comments

40th International Conference on Machine Learning (ICML 2023)

R2 v1 2026-06-28T10:56:40.141Z