欢迎来到天天文库
浏览记录
ID:40720067
大小:1.97 MB
页数:10页
时间:2019-08-06
《Learning Image Embeddings using Convolutional Neural Networks for Improved Multi-Modal Semantics 》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、LearningImageEmbeddingsusingConvolutionalNeuralNetworksforImprovedMulti-ModalSemanticsDouweKiela∗LeonBottou´UniversityofCambridgeMicrosoftResearchComputerLaboratoryNewYorkdouwe.kiela@cl.cam.ac.ukleon@bottou.orgAbstractlabeleddataset(Krizhevskyetal.,2012).Theconvolutionallayersarethenusedasmid-
2、levelWeconstructmulti-modalconceptrepre-featureextractorsonavarietyofcomputervi-sentationsbyconcatenatingaskip-gramsiontasks(Oquabetal.,2014;Girshicketal.,linguisticrepresentationvectorwithavi-2013;ZeilerandFergus,2013;Donahueetal.,sualconceptrepresentationvectorcom-2014).Althoughtransferringc
3、onvolutionalnet-putedusingthefeatureextractionlayersworkfeaturesisnotanewidea(DriancourtandofadeepconvolutionalneuralnetworkBottou,1990),thesimultaneousavailabilityof(CNN)trainedonalargelabeledobjectlargedatasetsandcheapGPUco-processorshasrecognitiondataset.Thistransferlearn-contributedtotheac
4、hievementofconsiderableingapproachbringsaclearperformanceperformancegainsonavarietycomputervisiongainoverfeaturesbasedonthetraditionalbenchmarks:“SIFTandHOGdescriptorspro-bag-of-visual-wordapproach.Experimen-ducedbigperformancegainsadecadeago,andtalresultsarereportedontheWordSim353nowdeepconvo
5、lutionalfeaturesareprovidingaandMENsemanticrelatednessevaluationsimilarbreakthrough”(Razavianetal.,2014).tasks.Weusevisualfeaturescomputedus-ThisworkreportsonresultsobtainedbyusingingeitherImageNetorESPGameimages.CNN-extractedfeaturesinmulti-modalsemanticrepresentationmodels.Theseresultsareint
6、erest-1Introductioninginseveralrespects.First,thesesuperiorfea-Recentworkshaveshownthatmulti-modalse-turesprovidetheopportunitytoincreasetheper-manticrepresentationmodelsoutperformuni-formancegapachievedbyaugmentinglinguisticmodallinguisticmodelsonavarietyoftasks,in-featureswithmulti-modalfeat
7、ures.Second,thiscludingmodelingsemanticrelatednessandpre-increasedperformanceconfirmsthatthemulti-dictingcompositionality(FengandLapata,2010;modalperformanceimprovementresultsfromtheLeongandMihalcea,2011;Brunietal.,2012;informationcontai
此文档下载收益归作者所有