基于Huber损失的稳健学习  

Robust learning based on Huber loss

作　　者：刘国旺 刘杰 黄收友 LIU Guo-wang;LIU Jie;HUANG Shou-you(China Telecom Fufu Information Technology Co.,Ltd.Fuzhou 350001,China;School of Mathematics and Statistics,Hubei Normal University,Huangshi 435002,China)

机构地区：[1]中电福富信息科技有限公司,福建福州350001 [2]湖北师范大学数学与统计学院,湖北黄石435002

出　　处：《湖北师范大学学报（自然科学版）》2023年第2期10-14,共5页Journal of Hubei Normal University：Natural Science

基　　金：湖北省自然科学基金—黄石联合资助项目(2022CFD042)。

摘　　要：研究了具有Huber损失的稳健学习的性能。作为传统经验风险最小化形成的稳健学习算法,Huber回归在机器学习中得到了广泛的应用。建立了一个新的比较定理,刻画冗余泛化误差与预测误差之间的差距。最后,通过仿真验证证明了Huber损失的稳健性。This paper studies the performance of robust learning with Huber loss.As an alternative to traditional empirical risk minimization schemes,Huber regression has been extensively used in machine learning.With the establishment of a new comparison theorem which characterizes the gap between the excess generalization error and the prediction error,this paper also verifies the robustness of Huber loss through simulation.

关 键 词：Huber损失 稳健学习 条件矩 

分 类 号：O211.5[理学—概率论与数理统计]