多标签分类：汉明损失和子集精度真的相互冲突吗？.pdf

《多标签分类：汉明损失和子集精度真的相互冲突吗？.pdf》由会员分享，可在线阅读，更多相关《多标签分类：汉明损失和子集精度真的相互冲突吗？.pdf（11页珍藏版）》请在三个皮匠报告上搜索。

1、Multi-label classification:do Hammingloss andconflictwitheachother？subsetreallyaccuracyGuogiang WuDepartment of Computer Science and Technology， Tsinghua UniversityG.WuJ.Zhu. Multi-label classification: do Hamming loss and subset accuracyreally conflict with each other？ NeuriPs 2020号24GGuoqiang Wu (

2、Tsinghua University）1/11#page#BackgroundMulti-Label Classification (MLC) is a fundamental task whereeach instance is associated with multiple labels simultaneouslyIt has plenty of applications in reality such as text classificationimage annotation especially the recommendation system inE-commercial

3、platforms.The efficient training of big data models can benefit from theaccelerated calculation of GPU(S）.We use a high-performanceserver with 8 RTX 2080Ti GPUs for MLC datasets (e.g.PASCALVOC and NUS-WIDE）.240Guoqiang Wu (Tsinghua University）2/11#page#MotivationFor MLC,various measures have been de

4、veloped includingHamming Loss (HL）， Subset Accuracy (SA） and Ranking Loss（RL）.However there is a gap in theory.F An algorithm often empirically performs well on somemeasurels） while poorly on others， and a formaltheoreticalanalysis is lackings In small label space cases， the algorithms optimizing HL

5、 canperform well on the SA measure, while existing theoreticalresults show that SA and HL are conflicting measures.This paper tries to fill this gap. Question: Whats the generalization performance of analgo240Guoqiang Wu (Tsinghua University）3/11#page#ChallengesTo answer this question， it needs to a

6、nalyze the generalizationbounds of an algorithmF Interms ofthe measureit aims to optimizeF In terms of other measures - More challengingThis requires us to reveal the intrinsic relationships among themeasures.Then， we can get the learning guarantees between themandtake insights to explain the phenom