![構建深度學習的理論基礎 -一種實證方法_第1頁](http://file4.renrendoc.com/view/d126e26c78e91dbb41f97c9b6b467f5f/d126e26c78e91dbb41f97c9b6b467f5f1.gif)
![構建深度學習的理論基礎 -一種實證方法_第2頁](http://file4.renrendoc.com/view/d126e26c78e91dbb41f97c9b6b467f5f/d126e26c78e91dbb41f97c9b6b467f5f2.gif)
![構建深度學習的理論基礎 -一種實證方法_第3頁](http://file4.renrendoc.com/view/d126e26c78e91dbb41f97c9b6b467f5f/d126e26c78e91dbb41f97c9b6b467f5f3.gif)
![構建深度學習的理論基礎 -一種實證方法_第4頁](http://file4.renrendoc.com/view/d126e26c78e91dbb41f97c9b6b467f5f/d126e26c78e91dbb41f97c9b6b467f5f4.gif)
![構建深度學習的理論基礎 -一種實證方法_第5頁](http://file4.renrendoc.com/view/d126e26c78e91dbb41f97c9b6b467f5f/d126e26c78e91dbb41f97c9b6b467f5f5.gif)
版權說明:本文檔由用戶提供并上傳,收益歸屬內容提供方,若內容存在侵權,請進行舉報或認領
文檔簡介
HarvardJohnA.paulsonschoolofEngineeringandAppliedsciencesveexaminedadissertationentitled"BuildingtheTheoreticalFoundationsofDeepLearning:AnEmpiricalApproach"presentedby:YaminiBansalGraduateschoolofArtsandsciencesDIssERTATIoNACCEpTANCECERTIFICATEersignedappointedbythesignatureTypedname:professorB.Bara一signaturesignatureTypedname:professorD.CoxMarch022DeepLearning:AnEmpiricalApproachDroTHEscHooioFENciNEERiNcANDAppiiEDsciENcEsiNpARriAiFuiFiiiMENroFrHEREguiREMENrsDocroRoFpHiiosopHyiNrHEsuBJEcroFCoMpurERsciENcEiDcEMAssAcHusErrsAiiRicHrsREsERvED.ABsrRAcrContentsTiriEpAcEiCopyRicHriiABsrRAcriiiCoNrENrsivLisriNcoFFicuREsviiiAcKNowiEDcEMENrsxv1INrRoDucrioNring 1.2Deviationsfromclassicalstatistics:Thegeneralizationpuzzle 31.3ourapproach 61.4overviewofthethesis 8iblebigpicture IUnderstandingRepresentationsroANEMpiRicAiiNvEsricArioNoFDEEpREpREsENrArioNsr2.1Introduction 14summaryofResults 152.2Modelstitching 172.3stitchingvs.representationalsimilarity 192.4stitchingConnectivity 212.5AllRoadsLeadtoRome 222.6MoreisBetter 242.7ConclusionandFuturework 25vwirHsEiF.supERvisioN263.1Introduction 263.1.1Relatedwork 293.1.2organization 3o3.2Formalstatementofresults 3o3.2.1CompleⅩitymeasures 323.3proofofTheorem4 333.4Thethreegaps 353.5EmpiricalstudyoftheRRMbound 373.6positiverationalitygapleavesroomforimprovement 393.7Conclusionsandopenquestions 4opolatingClassifers4INwHArwAyDoiNrERpoiAriNcciAssiFiERscENERAiizE2434.1Introduction 434.1.1DistributionalGeneralization 454.1.2summaryofContributions 464.1.3Relatedworkandsignińcance 474.2preliminaries 494.2.1DistributionalCloseness 494.2.2FrameworkforIndistinguishability 5o4.3FeatureCalibration 514.3.1EⅩperiments 534.3.2Discussion 574.3.31.Nearest.NeighborsConnection 574.3.4pointwiseDensityEstimation 594.4Agreementproperty 59 4.4.2potentialMechanisms 614.5LimitationsandEnsembles 624.5.1Ensembles 624.6DistributionalGeneralization:BeyondInterpolatingMethods 634.7ConclusionandDiscussion 664.7.1openQuestions 67IIIUnderstandingsampleEfciency68DArAscAiiNciAwsiNNMT6o5.1Introduction 69 5.2Relatedworks 725.3DatascalingLaws 735.3.1BLEUscore 755.3.2out.of.DistributionGeneralization 755.4TheE#ectofArchitecture 765.5TheE#ectofNoise 775.5.1DataFiltering 785.5.2AddingNoise 795.6Conclusions 81 n A.1.2CIFAR.1o: 84A.1.3ImageNet: 85A.1.4stitcher 85A.2Additionalresults 86A.2.1Ablations 86A.2.2ComparisonwithCKA 86A.2.3Comparisonwithńne.tuning 88 B.2EⅩperimentaldetails 92 B.2.4AdditionalResults 95 B.3.1Largerobustnessgap 96B.3.2Largerationalitygap 98 B.4simplerobustnessbounds 98B.4.1Robustnessofleastsquaresclassińers 98B.4.2Robustnessofempiricalriskminimizer 99 AppENDixCINwHArwAyDoiNrERpoiAriNcciAssiFiERscENERAiizE2Io2CEperimentalDetails......................1o2C.1.1Datasets..................................1o2C.1.2Models..................................1o3C.2FeatureCalibration:AppendiⅩ...........................1o5C.2.1Aguidetoreadingtheplots........................1o5C.2.2EⅩperiment1...............................1o6C.2.3Constantpartition............................1o7C.2.4Classpartition..............................1o7C.2.5MultipleFeatures 1o8C.2.6Coarsepartition 1o9C.2.7pointwiseDensityEstimation 11oC.3Agreementproperty:AppendiⅩ 113C.3.1EⅩperimentalDetails 113C.3.2Additionalplots 113C.3.3AlternateMechanisms 113C.4NerpolatingClassińers:AppendiⅩ 116C.5Nearest.Neighborproofs 117C.5.1Agreementproperty 117AppENDixDDArAscAiiNcLAwsiNNMTI22D.1scalingLawFittingDetails 122D.1.1separateFitsandvariance 122D.1.2optimizingHyperparameters 123D.1.3variance.LimitedRegime 123D.2BLEUscoreBehavior 125D.3scalingLawsforDi#erentooDDatasets 126D.4DatascalingphaseTransition 126D.5scalingLawswithDi#erentArchitectures 127D.6ChangingLanguagepairs 127REFERENcEsI46y true 7 formoreinformation 28RobustnessRationalityandMemorizationforCIFAR.ro.Eachbluepointisa nessRationalityandMemorizationforImageNetEachpointrepresents Ⅹ 4.3FeatureCalibrationwithoriginalclassesonCIFAR.ro:wetrainawRN.28.1o 4.5FeatureCalibrationforDecisiontreesonUCI(molecularbiology).weaddlabelbyConjecture13 56ing 56curacy........................................6o 61 4.1oDistributionalGeneralizationforwideResNetonCIFAR.ro.weapplylabelnoise p respectively 73s ndCKA C.6MNIsTEnsemble 112 CLaplaceKernelonFashion.MNIsT 119 LsTMhybrids 128
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
- 4. 未經(jīng)權益所有人同意不得將文件中的內容挪作商業(yè)或盈利用途。
- 5. 人人文庫網(wǎng)僅提供信息存儲空間,僅對用戶上傳內容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內容本身不做任何修改或編輯,并不能對任何下載內容負責。
- 6. 下載文件中如有侵權或不適當內容,請與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。
最新文檔
- 八年級上冊歷史人教版同步聽課評課記錄第6課《戊戌變法》
- 新版湘教版秋八年級數(shù)學上冊第二章三角形課題三角形高線角平分線中線聽評課記錄
- 五年級上美術聽評課記錄
- 北師大版道德與法治七年級下冊3.1《情緒使生活更美》聽課評課記錄
- 人教版地理八年級下冊第九章第一節(jié)《自然特征與農(nóng)業(yè)》聽課評課記錄
- 人教部編版八年級道德與法治上冊:8.1《國家好 大家才會好》聽課評課記錄2
- 中考道德與法治一輪復習九年級上第4單元和諧與夢想 聽課評課記錄 人教版
- 小學二年級數(shù)學乘法口算測試題人教版
- 蘇教版小學數(shù)學五年級上冊口算試題全套
- 班組長個人工作計劃書
- 降水預報思路和方法
- 工程設計方案定案表
- 第一章-天氣圖基本分析方法課件
- 虛位移原理PPT
- 暖氣管道安裝施工計劃
- 初二物理彈力知識要點及練習
- QE工程師簡歷
- 輔音和輔音字母組合發(fā)音規(guī)則
- 2021年酒店餐飲傳菜員崗位職責與獎罰制度
- 最新船廠機艙綜合布置及生產(chǎn)設計指南
- 可降解塑料制品項目可行性研究報告-完整可修改版
評論
0/150
提交評論