文档介绍
ColdCase:theLostMNISTDigitsChhaviYadavLéonBottouNewYorkUniversityAIResearchNewYork,NYandNewYorkUniversitychhavi@nyu.eduNewYork,NYleon@bottou.orgAbstractAlthoughthepopularMNISTdataset[LeCunetal.,1994]isderivedfromtheNISTdatabase[GrotherandHanaoka,1995],thepreciseprocessingstepsforthisderivationhavebeenlosttotime.Weproposeareconstructionthatisac-curateenoughtoserveasareplacementfortheMNISTdataset,withinsignificantchangesinaccuracy.WetraceeachMNISTdigittoitsNISTsourceanditsrichmetadatasuchaswriteridentifier,partitionidentifier,etc.WealsoreconstructthecompleteMNISTtestsetwith60,000samplesinsteadoftheusual10,000.Sincethebalance50,000wereneverdistributed,theycanbeusedtoinvestigatetheimpactoftwenty-fiveyearsofMNISTexperimentsonthereportedtestingperformances.Ourlimitedresultsunambiguouslyconfirmthetren