深度学习模型训练优化方法综述：收敛性与泛化性的理论视角 A Survey of Optimization Methods for Training DL Models - Theoretical Perspective on Convergence and Generalization.docx

下载文档

0
0
约26.89万字
约 234页
2025-02-09 发布于山西
举报
版权申诉
保障服务

深度学习模型训练优化方法综述：收敛性与泛化性的理论视角 A Survey of Optimization Methods for Training DL Models - Theoretical Perspective on Convergence and Generalization.docx

1、本文档共234页，可阅读全部内容。
2、有哪些信誉好的足球投注网站（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击这里二次下载。
4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。

ASurveyofOptimizationMethodsforTrainingDLModels:TheoreticalPerspectiveonConvergenceandGeneralization

jw5665@nyu.

jw5665@nyu.edu

ComputerEngineeringDepartmentofElectricaland

ComputerEngineering

ac5455@nyu.

ac5455@nyu.edu

arXiv:2501.14458v1[cs.LG]24Jan2025ComputerEngineeringDepartment

arXiv:2501.14458v1[cs.LG]24Jan2025

ComputerEngineering

Abstract

Asdatasetsgrowinsizeandcomplexity,itisbecomingmoredifficulttopullusefulfeaturesfromthemusinghand-craftedfeatureextractors.Forthisreason,deeplearning(DL)frameworksarenowwidelypopular.DLframeworksprocessinputdatausingmulti-layernetworks.Importantly,DLapproaches,asopposedtotraditionalmachinelearning(ML)methods,automaticallyfindhigh-qualityrepresentationofcomplexdatausefulforaparticularlearningtask.TheHolyGrailofDLandoneofthemostmysteriouschallengesinallofmodernMListodevelopafundamentalunderstandingofDLoptimizationandgeneralization.Whilenumerousoptimizationtechniqueshavebeenintroducedintheliteraturetonavigatetheexplorationofthehighlynon-convexDLoptimizationlandscape,manysurveypapersreviewingthemprimarilyfocusonsummarizingthesemethodologies,oftenoverlookingthecriticaltheoreticalanalysesofthesemethods.Inthispaper,weprovideanextensivesummaryofthetheoreticalfoundationsofoptimizationmethodsinDL,includingpresentingvariousmethodologies,theirconvergenceanalyses,andgeneralizationabilities.Thispapernotonlyincludestheoreticalanalysisofpopulargenericgradient-basedfirst-orderandsecond-ordermethods,butitalsocoverstheanalysisoftheoptimizationtechniquesadaptingtothepropertiesoftheDLlosslandscapeandexplicitlyencouragingthediscoveryofwell-generalizingoptimalpoints.Additionally,weextendourdiscussiontodistributedoptimizationmethodsthatfacilitateparallelcomputations,includi