Meta-Cal:Well-controlledPost-hocCalibrationbyRankingXingchenMa1MatthewB.Blaschko1AbstractToachievethisobjective,onecandesignmodelswithin-trinsicallylowCalibrationerrors.Thesemodels(WilsonInmanyappl...
Distribution-FreeCalibrationGuaranteesforHistogramBinningwithoutSampleSplittingChiragGupta1AadityaRamdas1Abstractthegoalistorecalibrateexisting(‘pre-learnt’)classifiersthatarepowerfuland(statisti...
AccuratePostTrainingQuantizationWithSmallCalibrationSetsItayHubara12YuryNahshan1YairHanani1RonBanner1DanielSoudry2Abstract1.IntroductionLately,post-trainingquantizationmethodshaveThepursuitofadvanc...
UnlabelledDataImprovesBayesianUncertaintyCalibrationunderCovariateShiftAlexJ.Chan1AhmedM.Alaa2ZhaozhiQian1MihaelavanderSchaar123Abstracttityoflabelledi.i.ddataareavailable(Krizhevskyetal.,2012).How...
IndividualCalibrationwithRandomizedForecastingShengjiaZhao1TengyuMa1StefanoErmon1Abstractbrated,itispossiblethatthetrueoutcomeisalwaysbelowthepredictedmedianonasubgroup(e.g.,men),andalwaysMachinele...
MaximumLikelihoodwithBias-CorrectedCalibrationisHard-To-BeatatLabelShiftAdaptationAmrM.Alexandari1AnshulKundaje12AvantiShrikumar1Abstract1.IntroductionLabelshiftreferstothephenomenonwheretheImagine...
DistributionCalibrationforRegressionHaoSong1TomDiethe2MeelisKull3PeterFlach14Abstractmaking,suchascost-sensitiveclassification[33].However,fromsimplemodels(e.g.Na¨ıveBayes)tocomplexonesWeareconce...
TrainableCalibrationMeasuresForNeuralNetworksFromKernelMeanEmbeddingsAviralKumar1SunitaSarawagi1UjjwalJain1Abstractlearnedmodelsfeedintodecisionsystemsorarehumanin-terpreted.AlsoCalibrationisuseful...
MultiCalibration:Calibrationforthe(Computationally-Identifiable)MassesU´rsulaHe´bert-Johnson1MichaelP.Kim1OmerReingold1GuyN.Rothblum2Abstractanumberofreasons.First,thetrainingdatamaycontainbiases...
OnCalibrationofModernNeuralNetworksChuanGuo1GeoffPleiss1YuSun1KilianQ.Weinberger1AbstractLeNet(1998)ResNet(2016)1.0CIFAR-100CIFAR-100ConfidenceCalibration–theproblemofpredict-ingprobabilityestimat...