报告题目：Unified distributed inference for a class of problems with non-smooth loss

主 讲 人：朱仲义

报告时间：2021年12月13日15:00-16:00

报告地点：腾讯会议 会议 ID：432-683-280

点击链接入会： https://meeting.tencent.com/dm/obb2lyp5ZA8E

报告摘要：

This article provides a unified distributed inference method for a class of statistical problems. The loss function in the class is not smooth. Typical examples include quantile regresion, support-vector machine, and Huber’s robust regression. The nonsmoothness of loss function poses a big challenge to distributed optimization. To overcome the diffculty caused by nonsmoothness, we propose to smooth the loss function via convolution in a unified way. Then based on the smoothed loss function, we develop two unified distributed algorithms. This first algorithm iteratively solves a sequence of smoothed loss functions with decreasing bandwidths. And the second algorthm is the communication-effient implementation of the first algorihtm, which avoids transferring matrices. In theory, the proposed algorithms can achieve the performance of full sample estimator in several iterations. The resulted estimator then can be used for statistical inference with little loss. Empirical studies also show good performance of the proposed algorithms.

主讲人介绍：

复旦大学统计系教授，博士研究生导师；曾任中国概率统计学会第八、九届副理事长,国际著名杂志”Statistica Sinica”副主编; “应用概率统计”, “数理统计与管理”杂志编委，中国统计教材编审委员会委员；现为 Elected Member of the ISI（国际数理统计学会）；”中国科学：数学”杂志编委。专业研究方向为：保险精算；纵向数据(面板数据)模型；分位数回归模型等。主持完成国家自然科学基金四项、国家社会科学基金一项，作为子项目负责人完成国家自然科学基金重点项目一项。已经培养毕业博士研究生13名。目前主持国家自然科学基金重大项目子项目一项，重点项目子项目一项，面上项目一项。近几年发表论文100多篇(其中包括在国际四大统计顶级刊物等SCI论文六十多篇)。获得教育部自然科学二等奖一次。

邀请人：林路

欢迎各位老师同学积极参加！