Resource-Efficient Decentralized and Federated Learning
Distributed optimization has been a classic topic, yet is attracting significant interest recently in machine learning due to its numerous applications such as distributed training, multi-agent learning, federated optimization, and so on. Often, the scale of modern datasets has exceeded the capacity of a single machine, and privacy and communication constraints prevent information sharing in a centralized manner and necessitates distributed infrastructures. Broadly speaking, there are two types of distributed settings: a distributed/federated setting, where a parameter server aggregates and shares parameters across all agents; and a decentralized/network setting, where each agent only aggregates and shares parameters with its neighbors over a network topology. The canonical problem of empirical risk minimization in the distributed setting leads to intriguing trade-offs between computation and communication that are not well understood; moreover, data unbalancedness and heterogeneity across agents poses additional challenges in both algorithmic convergence and statistical efficacy, often exacerbated by additional bandwidth and privacy constraints.
Overview
Federated and Decentralized Learning with Communication Compression and Differential Privacy
Escaping Saddle Points in Heterogeneous Federated Learning via Distributed SGD with Communication Compression [Arxiv]
S. Chen, Z. Li, and Y. Chi, International Conference on Artificial Intelligence and Statistics (AISTATS), 2024.
Convergence and Privacy of Decentralized Nonconvex Optimization with Gradient Clipping and Communication Compression [Arxiv]
B. Li and Y. Chi, preprint.
SoteriaFL: A Unified Framework for Private Federated Learning with Communication Compression [Arxiv] [Code]
Z. Li, H. Zhao, B. Li, and Y. Chi, Conference on Neural Information Processing Systems (NeurIPS), 2022.
-
BEER: Fast O(1/T) Rate for Decentralized Nonconvex Optimization with Communication Compression [Arxiv] [Code]
H. Zhao, B. Li, Z. Li, P. Richtarik, and Y. Chi, Conference on Neural Information Processing Systems (NeurIPS), 2022.
Communication-Efficient Federated Learning and Optimization
Communication-Efficient Federated Optimization over Semi-Decentralized Networks [Arxiv]
H. Wang and Y. Chi, preprint. Short version at ICASSP 2024.
A Multi-Token Coordinate Descent Method for Semi-Decentralized Vertical Federated Learning [Arxiv]
P. Valdeira, Y. Chi, C. Soares, and J. Xavier, preprint.
-
DESTRESS: Computation-Optimal and Communication-Efficient Decentralized Nonconvex Finite-Sum Optimization [Arxiv] [Code]
B. Li, Z. Li, and Y. Chi, SIAM Journal on Mathematics of Data Science, vol. 4, no. 3, pp. 1031-1051, 2022. Short version at OPT 2021 as a spotlight presentation.
-
Communication-Efficient Distributed Optimization in Networks with Gradient Tracking and Variance Reduction [Arxiv] [Code]
B. Li, S. Cen, Y. Chen, and Y. Chi, Journal of Machine Learning Research, vol. 21, no. 180, pp. 1-51, 2020. Short version at AISTATS 2019.
-
Convergence of Distributed Stochastic Variance Reduced Methods without Sampling Extra Data [Arxiv]
S. Cen, H. Zhang, Y. Chi, W. Chen and T.-Y. Liu, IEEE Trans. on Signal Processing, vol. 68, pp. 3976-3989, 2020.
Federated Reinforcement Learning
Federated Offline Reinforcement Learning: Collaborative Single-Policy Coverage Suffices [Arxiv]
J. Woo, L. Shi, G. Joshi, and Y. Chi, International Conference on Machine Learning (ICML), 2024.
Federated Natural Policy Gradient Methods for Multi-task Reinforcement Learning [Arxiv]
T. Yang, S. Cen, Y. Wei, Y. Chen, and Y. Chi, preprint.
The Blessing of Heterogeneity in Federated Q-Learning: Linear Speedup and Beyond [Arxiv]
J. Woo, G. Joshi, and Y. Chi, preprint. Short version at ICML 2023.
|