深度學習一線實戰暑期研討基礎_第1頁
深度學習一線實戰暑期研討基礎_第2頁
深度學習一線實戰暑期研討基礎_第3頁
深度學習一線實戰暑期研討基礎_第4頁
深度學習一線實戰暑期研討基礎_第5頁
已閱讀5頁,還剩88頁未讀 繼續免費閱讀

下載本文檔

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

234

DiscreteValueContinuousValue圖像表示:Gabor圖像表示:Gabor圖像集表示:Manifold,GMM,LMNN,NCA…..詞典學習&CreditedtoProf.XiaogangWangandProf.Shiguang CreditedtoProf.Songchun DiscreteValueContinuousValue端到端學習(EndtoCreditedtoProf.Shiguang CreditedtoDr.Naiyan CreditedtoDr.Naiyan 多多PascalVOC目標檢測CreditedtoProf.ShiguangShanwith

Fastdescriptorcoding(LLCJinjunWang’sCVPR10&Super+SVMLin,Y’sCompressedFisherVectorsPerronnin,F’sECCV10,Sanchez,J’sDCNNAlexNet8層網絡KrizhevskyA’sDCNN:基于網絡可視化技術Zeiler,M.D’sECCV14在AlexNetDCNN:NetChristianSzegedy’sCVPR15,22層+Inception結DCNN:VGGNetKarenSimonyan’sarXiv14,19DCNN:ResidualNetKaimingHe’sarXiv15,152SelectiveSearchJ.R.R.Uijlings’sIJCV13+efficientencodingdenselysampledcolordescriptorsvandeSande’sTPAMI10,CVPR14RCNNRossGirshick’sCVPR14+NetworkInNetworkMinLin’sFasterRCNNShaoqingRen’sNIPS15DeepResidualNetworkKaimingHe’sMeanMean

CombinedMultishot

VectorVector

gh-dimintgh-dim

TLTL

Unrestricted Outside Unrestricted+LabeledOutside 網絡變大變深(VGGFace16FaceNet22層數據量不斷增大(DeepFace400萬,FaceNet2億3DeepID[Sun’sDeepID2[Sun’sDeepID2+[Sun’s-1FaceNet[Schroff’s1ModifiedfromProf.ModifiedfromProf.LeCunandProf. 來自:BoleiZhouetalObjectDetectorsEmergeinDeepSceneCNNsICLRConv,Conv,Pooling,ReLU,connected,BN,NIN, N,SiameseNet,機制LossFunc,e.g.SoftmaxLoss,SigmoidCross-entropy,…BP,SGD,AdaDelta,Path-Dropout,Fine-tune,CreditedtoProf.MeinaKan,PHDStudentXinLiuandShuzhe McCulloch,Warren;WalterPitts(1943)."ALogicalCalculusofIdeasImmanentinNervousActivity".BulletinofMathematicalBiophysics5(4):115–133F.Rosenblatt.Theperceptron:Aprobabilisticmodelforinformationstorageandorganizationinthebrain. PsychologicalReview,65:386-408,Minsky&Papert的專著Perceptron(1969) Novikoff,A.B.J.(1962).Onconvergenceproofsonperceptrons.ProceedingsoftheSymposiumontheMathematicalTheoryofAutomata(pp.615–622)IEEEFrankRosenblatt2014-GeoffreyE.Hinton2012-VladimirN.Vapnik2009-JohnJ.Hopfield多層感知機卷積網絡

DecisionSparseGraphRumelhart,DavidE.;Hinton,GeoffreyE.;Williams,RonaldJ.(8October1986)."Learningrepresentationsbyback-propagatingerrors".Nature323(6088):533–536.CreditedtoProf.ShiguangHinton,G.E.,Osindero,S.andTeh,Y.,Afastlearningalgorithmfordeepbeliefnets.NeuralComputation18:1527-1554,2006Hinton,G.E.andSalakhutdinov,R.R.Reducingthedimensionalityofdatawithneuralnetworks.Science,Vol.313.no.5786,pp.504-507,28July2006YoshuaBengio,PascalLamblin,DanPopoviciandHugoLaroce,GreedyLayer-WiseTrainingofDeepNetworks,NIPS2006CreditedtoProf.Eric deeplearningDCNNAlexKrizhevsky’s止過擬合,LocalResponseNormalization增強泛化能力評價評價 評價RNNLanguage評價CreditedtoProf.MeinaKan,PHDStudentXinLiuandShuzhe ReLU/

MoonMoonStart...:... ,其中??∈?????1,??∈?????1,??∈??:...Backward運算:????=???????,????=?? 連續卷積???,

=?????(??,??)

????,?????????,?????離散卷積????,Caffe實現:????

=?????(??,??) ??????,?????????,???=?????(??,??) ??????+??,??+??????,多個FeatureMap: 接

CreditedtoDr.ShaoxinCreditedtoDr.ShaoxinPleasealsoreferto“NotesonConvolutionalNeural …Image:??×??×…

FeatureMatrixX:(????????×????????)×(C×K×…Kernel:????????×??×??×…bias:????????×1×1×

??????+??:????????×(??×??× …………??:????????×(????????×……OutputMatrix:????????×(????????×

Kernelsize:2x2Kernelsize:2x24346378445555686823

S(x)=1/(1+e-xy=xifx>

a(e-1)ifx£max(w1x1+b1,w2x2+b2 LeCun組2010年的文章Whatisthebestmulti-stagearchitectureforobjectrecognition?嘗試了各種非線性激活函Hinton組的“RectifiedLinearUnitsImproveRestrictedBoltzmannMachines”將NReLU用于RBM訓練 48引入Bernoulli隨機數????代表dropout測試階段:DoG.E.Hinton,N.Srivastava,A.Krizhevsky,I.Sutskever,andR.R.Salakhutdinov.Improvingneuralnetworksbypreventingco-adaptationoffeaturedetectors.arXivpreprintarXiv:1207.0580,2012. Batch UntersuchungenzudynamischenneuronalenNetzen(1991LearningLong-TermDependencieswithGradientDescentisDifficult(1994引入三個Gate結構:ForgetGate,InputGate和Output 狀態與#1.狀態“遺忘”控制門forget狀態與#1.狀態“遺忘”控制門forget#2.輸入控制門input狀態與#1.狀態“遺忘”控制門forget#2.輸入控制門input狀態與#1.狀態“遺忘”控制門forget#2.輸入控制門input#3:#4.輸出控制門outputPeepholeGRU(inputgate和forgetgate合并為updateSoftmax+CrossEntropyE=-1N

log( ),l?[0,1,...,K-

n??nkn

CrossEntropyLossH(p,q)=-p(x)logxlog(x)'=x

當xSigmoidCrossEntropy

?npnlo1?Tips1:目標輸出需要歸一化到[0,1]XinLiu*,ShaoxinLi*etal.AgeNet:DeeplyLearnedRegressorandClassifierforRobustApparentAgeEstimation.InternationalConferenceonComputerVisionChaLearnLaPWorkshop(ICCVW),2015.SoftmaxSigmoidCrossEntropy

E=- E=-

log(?nl),ln?[0,1,...,K-n?n?nSigmoidCrossEntropyLossGroud_truth:Pn=(pn,1-pnSigmoid??,1?

?n)

log?

plo1?

L[ log o1? SoftmaxLoss

Euclidean損失函數:E=

?n Tips1:默認的,Data層和ImageData Tips2:歐式損失前可以增加Sigmoid操 ContrastiveLoss(Siamese

1

[(y)d2+(1-y)max(margin-d,0)2,d

a-

nTips2:Caffe中的例子:TripletBasic適用場景:Learningtorank,人臉識別F.Schroff,D.Kalenichenko,andJ.Philbin,FaceNet:AUnifiedEmbeddingforFaceRecognition MoonEthanM.Ruddetal,MOON:AMixedObjectiveOptimizationNetworkfortheRecognitionofFacialAttributesCreditedtoProf.Back Rumelhart,DavidE.;Hinton,GeoffreyE.;Williams,RonaldJ.(8October1986)."Learningrepresentationsbyback-propagatingerrors".Nature323(6088):533–536.GradientDescentanditswt+1=wt w

StochasticGradientDescendwt+1=wt wL(w,xi,yi

: =w-1h

L(w,x,y

Mini-batchMini-batchfori=1: WeightL(w)=w)+L(w)= L(wtwt+1=wt+Without WithSGDvs.給定等高線下不同優化算法的比 鞍點條件下不同優化算法的比Inmathematics,asaddlepointisapointinthe ofafunctionwheretheslopes(derivatives)oforthogonalfunctioncomponentsdefiningthesurfaceezero(astationarypoint)butarenotaloc9以為例,點(0,0)是鞍點,CreditedtoPHDStudentXinLiuandXiaoyi LearningCreditedtoProf.LearningRate在每隔stepsize輪迭代后減少base_lr: :base_lr: :“poly”power:0.5

base_lr:base_lr: :base_lr: :“step” :0.1stepsize:

LR(t)=base_lr·

t)Tbase_lr:0.01 :“inv”base_lr:0.01 :“inv” :power:LR(t)=base_lr·(1+g*iter)-LearningtolearnwithrecurrentneuralMa

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. 未經權益所有人同意不得將文件中的內容挪作商業或盈利用途。
  • 5. 人人文庫網僅提供信息存儲空間,僅對用戶上傳內容的表現方式做保護處理,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。

評論

0/150

提交評論