资源描述：
《Computational Baby Learning》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、ComputationalBabyLearningXiaodanLiangy?SiLiuyYunchaoWeiyLuoqiLiuyLiangLin?ShuichengYanyyNationalUniversityofSingapore?SunYat-senUniversityBeijingJiaotonguniversityfxdliang328,fifthzombiesi,wychao1987,llq667g@gmail.comlinliang@ieee.orgeleyans@nus.edu
2、.sgAbstractDuringcontinuouslyexploringand/orinteractingwithdi-verseinstancesandscenesinreallife,thebabycanasso-Intuitiveobservationsshowthatababymayinherentlyciatetheinitialsimpleinstanceswithothervariantsbyus-possessthecapabilityofrecognizinganewvisu
3、alconceptingvariousinformationlinkages.Basedontheaccumu-(e.g.,chair,dog)bylearningfromonlyveryfewpositivelatedinstancesabouttheconcept,thebabycangraduallyinstancestaughtbyparent(s)orothers,andthisrecogni-improveitsrecognitioncapabilityandrecognizedive
4、rsein-tioncapabilitycanbegraduallyfurtherimprovedbyexplor-stanceshe/sheneversaw.ingand/orinteractingwiththerealinstancesinthephysi-Recentsuccessesincomputervision[27],however,calworld.Inspiredbytheseobservations,weproposealargelyrelyonthelargenumberof
5、labeledinstancesofvi-computationalmodelforslightly-supervisedobjectdetec-sualconcepts,whichmayrequireconsiderablehumanef-tion,basedonpriorknowledgemodelling,exemplarlearn-forts.Theconstructionofanappearance-basedobjectde-ingandlearningwithvideocontext
6、s.Thepriorknowledgetectoriscostlyanddifficultbecausethenumberoftrainingismodeledwithapre-trainedConvolutionalNeuralNet-examplesmustbelargeenoughtocapturedifferentvaria-work(CNN).Whenveryfewinstancesofanewconceptaretionsintheobjectappearance.Someresearc
7、hershavemadegiven,aninitialconceptdetectorisbuiltbyexemplarlearn-effortsonimprovingtheinitialmodelsbyusingveryfewingoverthedeepfeaturesfromthepre-trainedCNN.Simu-labeleddata,alongwiththedetection/searchresultsfromlatingthebaby’sinteractionwithphysical
8、world,thewell-webimages[4][7][5]orweaklyannotatedvideos[25][3].designedtrackingsolutionisthenusedtodiscovermoredi-Inthispaper,wemakethefirstattemptandbuildacom-verseinstancesfromthemassiveonlineunlabeledvideos.putationalmodelforslightly-supervi
