基于自步学习的开放集领域自适应刘星宏1,2周毅1,2周涛3秦杰41（东南大学计算机科学与工程学院南京211189）2（新一代人工智能技术与交叉应用教育部重点实验室（东南大学）南京211189）3（南京理工大学计算机科学与工程学院南京210094）4（南京航空航天大学计算机科学与技术学院南京210016）（xhoml158@gmail.com）Self-PacedLearningforOpen-SetDomainAdaptationLiuXinghong1,2,ZhouYi1,2,ZhouTao3,andQinJie41（SchoolofComputerScienceandEngineering,SoutheastUniversity,Nanjing211189）2（KeyLaboratoryofNewGenerationArtificialIntelligenceTechnologyandItsInterdisciplinaryApplications(SoutheastUniversity),MinistryofEducation,Nanjing211189）3（SchoolofComputerScienceandEngineering,NanjingUniversityofScienceandTechnology,Nanjing210094）4（CollegeofComputerScienceandTechnology,NanjingUniversityofAeronauticsandAstronautics,Nanjing210016）AbstractDomainadaptationtacklesthechallengeofgeneralizingknowledgeacquiredfromasourcedomaintoatargetdomainwithdifferentdatadistributions.Traditionaldomainadaptationmethodspresumethattheclassesinthesourceandtargetdomainsareidentical,whichisnotalwaysthecaseinreal-worldscenarios.Open-setdomainadaptation(OSDA)addressesthislimitationbyallowingpreviouslyunseenclassesinthetargetdomain.OSDAaimstonotonlyrecognizetargetsamplesbelongingtoknownclassessharedbysourceandtargetdomainsbutalsoperceiveunknownclasssamples.Traditionaldomainadaptationmethodsaimtoaligntheentiretargetdomainwiththesourcedomaintominimizedomainshift,whichinevitablyleadstonegativetransferinopen-setdomainadaptationscenarios.Weproposeanovelframeworkbasedonself-pacedlearningtodistinguishknownandunknownclasssamplesprecisely,referredtoasSPL-OSDA(self-pacedlearningforopen-setdomainadaptation).Toutilizeunlabeledtargetsamplesforself-pacedlearning,wegeneratepseudolabelsanddesignacross-domainmixupmethodtailoredforOSDAscenarios.Thisstrategyminimizesthenoisefrompseudolabelsandensuresourmodelprogressivelytolearnknownclassfeaturesofthetargetdomain,beginningwithsimplerexamplesandadvancingtomorecomplexones.Toimprovethereliabilityo...
