上午三位ordinal zero-shot learning有序类零样本学习.pptx

下载文档

0
0
约4.86千字
约 29页
2025-02-24 发布于四川
举报
版权申诉
保障服务

上午三位ordinal zero-shot learning有序类零样本学习.pptx

1、本文档共29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)。

XinGeng(耿新)PALMlab,SoutheastUniversityOrdinalZero-ShotLearning[IJCAI’17]

Outline:BackgroundOrdinalZero-ShotLearningExperimentsConclusion

Zero-ShotLearningThegoalofZero-ShotLearningistoclassifyunseenclassinstancesTrainingSet?

Zero-ShotLearningThegoalofZero-ShotLearningistoclassifyunseenclassinstancesBlack: noWhite: yesBrown: noStripes: noWater: yes Eatfish: yesMostzero-shotlearningmethodsarebasedonsideinformation(suchasattributesortextcorpora)

ObtainingsideinformationisdifficultNeedshumanexpertsorahugelanguagecorporaTime-consumingprocessHardtoobtaindiscriminativecategory-levelattributesNoisysideinformationCanwesolvezero-shotlearningwithoutusingsideinformation?Zero-ShotLearning

Zero-ShotLearningTrainingSetSideInfo.Zero-ShotLearningDataProperty

OrdinalLabelOrdinallabelsarecommoninrealapplicationsMovieRating(1,2,…,10)StudentGrade(A,B,C,D)

Outline:BackgroundOrdinalZero-ShotLearningExperimentsConclusion

OrdinalZero-ShotLearningRegressionmethodsTransformordinallabelsintorealnumbersABCD123410203040432190806020????

OrdinalZero-ShotLearningClassificationmethodsExpandsupervisioninformationfromseenclassestounseenclassesOrdinalLabelGradualnessOncealabeldescribesaninstance,itsneighboringlabelsintheordersequencemayalsodescribethesameinstancetosomeextent

OrdinalZero-ShotLearningSupervisionIntensityDistribution(SID)OrdinalZero-ShotLearningLabelDistributionLearning

(Geng,TKDE’16)

UnseenLabelPatternsOrdinalZero-ShotLearning

SupervisionIntensityDistribution(SID)Twocriteria:Supervisionintensityiswithin[0,1]andsumsupto1ThegroundtruthlabelisassignedwiththehighestsupervisionintensityandthesupervisionintensitydecreasesastherelevancedecreasesLabelspaceSIDMahalanobisdistanceGaussiandistribution

SupervisionIntensityDistribution(SID)SupervisionintensityexpansionthroughSIDExpand

LabelDist