噪声标签-长尾学习-组会PPT-When Noisy Labels Meet Long Tail Dilemmas A Representation Calibration Method.pptx

ICCV2023WhenNoisyLabelsMeetLongTailDilemmas:ARepresentationCalibrationMethod

IntroductionLearningwithnoisylabels:1.memorizationeffect.E.g.Co-teaching2.thenoisetransitionmatrix.Learningwithlong-taileddata:1.re-samplingandre-weightingtechniques.Learningwithnoisylabelonlong-taileddata:1.distinguishmislabeleddatafromthedataoftailclassesforfollow-upprocedures.E.g.RoLT2.reducetheside-effectsofmislabeleddataandlong-taileddatainaunifiedway,relyingonstrongassumptions.E.g.HAR-DRW

IntroductionStep1:contrastivelearningtoachieverepresentationsforalltraininginstances.Step2:tworepresentationcalibrationstrategiesareperformed:distributionalandindividualrepresentationcalibrations.

Method/EnhancingRepresentationsThroughContrastiveLearning

Method/DistributionalRepresentationCalibrationPurpose:recoverrepresentationdistributions.Assumption:multivariateGaussiandistribution.Step1:giventhelearnedrepresentationsz.Step2:employLOFandremoveoutliers.Step3:estimatethemultivariateGaussiandistribution

Method/DistributionalRepresentationCalibrationProblem:thedataoftailclassesmaynotbeenoughtoestimate.Motivation:’FreeLunchforFew-shotLearning:DistributionCalibration’(similarclasseshavingsimilarmeansandcovarianceonrepresentations)

Method/IndividualRepresentationCalibrationPurpose:restrictthedistance.

Method

Experiments/Simulatednoisyandclass-imbalanceddatasets.Methodsforlong-taileddataMethodsfornoisylabelsMethodsforboth

Experiments/ResultsonReal-worldNoisyandImbalancedDatasetscombinesemi-supervisedlearning

Experiments/AblationStudy

Experiments/AblationStudy

Experiments/AblationStudy2021

Thankyou