时 间：2022年12月22日 10:00-11:00
地 点：腾讯会议ID:977-799-495
报告人：林伟 研究员
主持人：王小舟 助理教授
摘 要：
Federated learning trains models across multiple decentralized devices without sharing local data, and becomes increasingly popular in distributed machine learning. When the data are heterogeneous, however, the more data one aggregates, the more heterogeneity one may introduce. Communication heterogeneity adds further complexity to the scenario. To reconcile the aggregation–heterogeneity trade-off, we consider parameter estimation in federated learning with data and communication heterogeneity, by exploiting a communication graph among the local devices. We propose a general M-estimation framework with fused Lasso regularization to pursuit parameter homogeneity over the graph. We provide statistical guarantees for our estimator and show that it attains the optimal rate under a certain graph fidelity condition, as if we could aggregate all samples sharing the same distribution. We also suggest an edge selection procedure that is robust to graph misspecification. To implement our estimator, we develop FedADMM, a decentralized stochastic version of ADMM with convergence guarantees, and extend it to the case where devices are randomly inaccessible. The practical performance of our method is illustrated on simulations and 2020 US presidential election data.
报告人简介：
林伟，北京大学数学科学学院概率统计系、统计科学中心长聘副教授、研究员，统计学教研室主任。2011年获南加州大学应用数学博士学位，2011至2014年在宾夕法尼亚大学做博士后研究，2014年加入北京大学。主要研究方向为高维统计、统计机器学习、成分数据分析、因果推断、生存分析等，以及在基因组学、宏碁因组学和环境科学等领域的应用，代表性成果发表在JASA、Biometrika、Biometrics、IEEE TIT、Operations Research、Environmental Science & Technology、《中国科学：数学》等统计学及相关领域顶级期刊上。2015年入选国家高层次人才计划青年项目，主持国家重点研发计划课题、北京市自然科学基金重点研究专题项目、国家自然科学基金面上项目等。