组会PPT-一种可控的超网络长尾多专家生成.pptx

下载文档

1
0
约3.34千字
约 14页
2025-02-26 发布于江苏
举报
版权申诉
保障服务

组会PPT-一种可控的超网络长尾多专家生成.pptx

1、本文档共14页，可阅读全部内容。
2、原创力文档（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)。

NeurIPS2024BreakingLong-TailedLearningBottlenecks:AControllableParadigmwithHypernetwork-GeneratedDiverseExperts

BackgroundLong-tailedmethods:re-samplingandmodifyingthelossfunctionfocusingonimprovingtheperformanceofthetailclasses.Assumethatthedistributionsofthetrainingandtestdataremaininvariant.Cannotwellhandlethedistributionshiftbetweentrainingandtestinginreal-worldscenarios.

BackgroundSomemorerecentworksproposeusingmultipleexpertmodelstoobtainstrongerdistributionadaptability.Thatis,themodelslearnedbyvariouslossesareskilledinhandlingclassdistributionswithdifferentskewness.Theuniformexpert:simulatetheuniformclassdistributionThebackwardexpert(inversesoftmaxloss):simulatetheinverselylong-tailedclassdistribution:Theforwardexpert:directlysimulatestheoriginallong-tailedtrainingclassdistribution:

BackgroundRIDE:trainarouterthatdynamicallyassignsambiguoussamplestoadditionalexpertsonanas-neededbasis.Thedistributionofinstancesseenbyeachexpertshowsthatheadinstancesneedfewerexpertsandtheimbalancebetweenclassesgetsreducedforlaterexperts.Long-tailedRecognitionbyRoutingDiverseDistribution-AwareExperts.ICLR2021

BackgroundSADE(Test-timeself-supervised):Empiricalobservation:strongerexpertshavehigherpredictionsimilaritybetweendifferentviewsofsamplesfromtheirfavorableclasses.alleviatetheproblemofdistributionmismatchbetweentrainingandtestingtosomeextent.Self-supervisedaggregationofdiverseexpertsfortest-agnosticlong-tailedrecognition.NeuraIPS2022

IntroductionGoal:Maximizetheoverallperformance.Newrequirements:usersmayhavedifferentpreferencesandneedsfortherelativetrade-offbetweenheadandtailclasses.Examples:lungcancer,wemayalsobewillingtomoderatelyincreasethefalsepositiverateinexchangeforhighercoverageofthetailclasses,toensurethatnopatientsarem