Anarchic Convex Federated Learning
Anarchic Convex Federated Learning
复制标题
DOI：
10.1109/infocomwkshps57453.2023.10225908
复制
发表时间：
2023-05
期刊：
影响因子：
--
通讯作者：
Dongsheng Li;Xiaowen Gong
中科院分区：
文献类型：
--
作者：
Dongsheng Li;Xiaowen Gong
联邦学习（FL）在过去几年中的快速发展最近激发了对这一新兴主题的大量研究。FL的现有工作通常假设客户以某种特定的模式（如平衡参与）和/或以同步方式和/或以相同数量的局部迭代参与学习过程，而这些假设在实践中可能很难保持。在本文中，我们提出了AFLC，凸学习问题的无政府联邦学习算法，它给客户端最大的自由。特别地，AFLC允许客户端1）参与任意轮; 2）异步参与; 3）参与任意数量的局部迭代。提出的AFLC算法使客户能够根据自己的需求高效、灵活地参与FL，例如，基于它们的异构和时变计算和通信能力。我们表征性能界限AFLC的学习损失作为客户端的本地模型延迟和本地迭代次数的函数。我们的研究结果表明，收敛误差可以任意小，通过选择适当的学习率，收敛速度与现有的基准。研究结果还描述了客户的各种参数对学习损失的影响，这提供了有用的见解。数值结果表明了该算法的有效性。
The rapid advances in federated learning (FL) in the past few years have recently inspired a great deal of research on this emerging topic. Existing work on FL often assume that clients participate in the learning process with some particular pattern (such as balanced participation), and/or in a synchronous manner, and/or with the same number of local iterations, while these assumptions can be hard to hold in practice. In this paper, we propose AFLC, an Anarchic Federated Learning algorithm for Convex learning problems, which gives maximum freedom to clients. In particular, AFLC allows clients to 1) participate in arbitrary rounds; 2) participate asynchronously; 3) participate with arbitrary numbers of local iterations. The proposed AFLC algorithm enables clients to participate in FL efficiently and flexibly according to their needs, e.g., based on their heterogeneous and time-varying computation and communication capabilities. We characterize performance bounds on the learning loss of AFLC as a function of clients' local model delays and local iteration numbers. Our results show that the convergence error can be made arbitrarily small by choosing appropriate learning rates, and the convergence rate matches that of existing benchmarks. The results also characterize the impacts of clients' various parameters on the learning loss, which provide useful insights. Numerical results demonstrate the efficiency of the proposed algorithm.