الفهرس | Only 14 pages are availabe for public view |
Abstract Most modernsystemsutilizeartificialintelligenceandmachine learning, especiallydeeplearningmodelswhichhavegainedan increasedpopularityinthelastdecade.Theutilizationofthese models hasbeenmoreplausibleduetothecontinuousprogress in hardwareinnovations. This canbeseeninthewidespreadoftouchscreendevices that facilitatedthehumancomputerinteractionandincreased the useofhandgesturesanddrawings.Asaresult,image- relatedapplicationshavegainedaremarkableattention.Fur- thermore,theemploymentofconvolutionalneuralnetworksin image-relatedapplicationshaveshownanexceptionalperfor- mance. Alongside theprevalenceofmachinelearningmodels,con- cerns regardingthosemodelshavebeenontherise.Theneedto understand howsomemodels,likedeeplearningmodels,oper- ate hasincreased.Hence,aresearchfielddedicatedtotheinter- pretationofsuchmodelshasemerged.Inthiskindofresearch, the aimistoexplaintheunderlyingworkflowofanexistingma- chine learningmodel,ratherthantodevelopanewone. v In thisthesis,thesketchrecognitionproblemisusedasa testbed forinvestigatingsomeofthelearningbehaviorsofcon- volutional neuralnetworks.Ontheonehand,thenatureof sketches ischallengingowingtotheirabstractandinconsistent nature.Theexplanationofimage-relatedmodelsismorein- formative andappealingduetotherichnatureofimagesover sketches incolors,textures,backgroundscenery,etc.Onthe other hand,thehumanfactorinsketch-relatedapplicationscan be usedtoimprovetheresultsobtainedusingasketchrecogni- tion model.Someofthehumantraitsinthesketchrecognition task isexperimentedandusedasareferenceinexplainingthe learning behaviorofconvolutionalneuralnetworks.Oneofthe main propertiesaddressedinthisthesisistheabilityofanartifi- cial networktodetectvisualanalogies(i.e.,similarities)among differentobjects. This thesisisorganizedintosevenchapters.Inwhatfollows, the contentsofeachisbrieflyoutlined. Chapter 1 presentsanintroductiontothethesisincluding the motivationandthemaingoalsofthethesis.Thischapter also givesadetaileddescriptionoftheorganizationofthethesis’ chapters. vi In Chapters 2 and 3, themainaspectsofartificialneuralnet- works areintroduced.Architecturalvariationslikeconvolutional and recurrentneuralnetworksarealsodiscussed. Chapter 4 introducestheconceptofinterpretablemachine learning models.Itshedslightonsomeoftherelatedworkin the literature. Chapter 5 outlines thesketchrecognitionproblem.Thischap- ter highlightssomeoftherelatedrepresentationandrecognition approaches. Chapter 6 presentsanexperimentalapproachtostudythe learning behaviorofconvolutionalneuralnetworksusingthe sketch recognitionproblem.Theresultsobtainedfromthecon- ducted experimentsalongwiththeiranalysesarediscussed.The methodology andresultspresentedinthischapterispublished in [6]. Finally inChapter 7, themainideaspresentedinthethesis aresummarized.Thischaptergivestheconclusionandmain resultsof[6]. Moreover,thesuggestionsforresearchpointsto be conductedandattempingtosolveinthefutureareprovided. Appendix A providespartsoftheprogramthatwasimple- mented usingPython. |