深度神经网络IDeepNeuralNetworks中国科学院自动化研究所吴高巍gaowei.wu@ia.ac.cn2016-12-6内容深度神经网络发展历史、背景动机——WhyDeepLearning?深度学习常用模型NeuralnetworkBackpropagation1986•解决了一般性学习问题•与生物系统相联系Nature历史NeuralnetworkBackpropagation1986Nature历史x1x2x3w1w2w3NeuralnetworkBackpropagation1986•解决了一般性学习问题•与生物系统相联系Nature历史Butitisgivenup…•SVM•Boosting•Decisiontree•…2006NeuralnetworkBackpropagation1986Nature历史2006DeepbeliefnetScience……………………•Unsupervised&Layer-wisedpre-training•Betterdesignsformodelingandtraining(normalization,nonlinearity,dropout)•Newdevelopmentofcomputerarchitectures–GPU–Multi-corecomputersystems•LargescaledatabasesBigData!Neuralnetworksiscomingback!深度学习浪潮ITCompaniesareRacingintoDeepLearningNeuralnetworkBackpropagation1986•Solvegenerallearningproblems•TiedwithbiologicalsystemButitisgivenup…2006DeepbeliefnetSciencedeeplearningresultsSpeech2011NatureObjectrecognitionover1,000,000imagesand1,000categories(2GPU)NeuralnetworkBackpropagation19862006DeepbeliefnetScienceSpeech20112012NatureA.Krizhevsky,L.Sutskever,andG.E.Hinton,“ImageNetClassificationwithDeepConvolutionalNeuralNetworks,”NIPS,2012.RankNameErrorrateDescription1U.Toronto0.15315Deeplearning2U.Tokyo0.26172Hand-craftedfeaturesandlearningmodels.Bottleneck.3U.Oxford0.269794Xerox/INRIA0.27058NeuralnetworkBackpropagation19862006DeepbeliefnetScienceSpeech20112012ImageNet2013–imageclassificationchallengeRankNameErrorrateDescription1NYU0.11197Deeplearning2NUS0.12535Deeplearning3Oxford0.13555DeeplearningMSRA,IBM,Adobe,NEC,Clarifai,Berkley,U.Tokyo,UCLA,UIUC,Toronto….Top20groupsalluseddeeplearning•ImageNet2013–objectdetectionchallengeRankNameMeanAveragePrecisionDescription1UvA-Euvision0.22581Hand-craftedfeatures2NEC-MU0.20895Hand-craftedfeatures3NYU0.19400DeeplearningNeuralnetworkBackpropagation19862006DeepbeliefnetScienceSpeech20112012ImageNet2014–ImageclassificationchallengeRankNameErrorrateDescription1Google0.06656Deeplearning2Oxford0.07325Deeplearning3MSRA0.08062Deeplearning•ImageNet2014–objectdetectionchallengeRankNameMeanAveragePrecisionDescription1Google0.43933Deeplearning2CUHK0.40656Deeplearning3DeepInsight0.40452Deeplearning4UvA-Euvision0.35421Deeplearning5BerkleyVision0.34521DeeplearningNeuralnetworkBackpropagation19862006DeepbeliefnetScienceSpeech20112012•GoogleandBaiduannouncedtheirdeeplearningbasedvisualsearchengines(2013)–Google•“onourtestsetwesawdoubletheaverageprecisionwhencomparedtootherapproacheswehadtried.WeacquiredtherightstothetechnologyandwentfullspeedaheadadaptingittorunatlargescaleonGoogle’scomputers.Wetookcuttingedgeresearchstraightoutofanacademicresearchlabandlaunchedit,injustalittleoversixmonths.”–BaiduNeuralnetworkBackpropagation19862006DeepbeliefnetScienceSpeech20112012Facerecognition2014Deeplearningachieves99.53%faceverificationaccuracyonLabeledFacesintheWild(LFW),higherthanhumanperformanceY.Sun,X.Wang,andX.Tang.DeepLearningFaceRepresentationbyJointIdentification-Verification.NIPS,2014.Y.Sun,X.Wang,andX.Tang.Deeplylearnedfacerepresentationsaresparse,selective,androbust.CVPR,2015.深度学习浪潮DeepLearning深度学习浪潮时代背景-数据爆炸还存在很多没有良好解决的问题,例如图像识别、语音识别、自然语言理解、天气预测、基因表达、内容推荐等。深度学习浪潮时代背景-计算性能提升动机——WhyDeepLearning?深度学习WhatisDeepLearning?“Deeplearningisasetofalgorithmsinmachinelearningthatattempttolearninmultiplelevels,correspondingtodifferentlevelsofabstraction.Ittypicallyusesartificialneuralnetworks.Thelevelsintheselearnedstatisticalmodelscorrespondtodistinctlevelsofconcepts,wherehigher-levelconceptsaredefinedfromlower-levelones,andthesamelower-levelconceptscanhelptodefinemanyhigher-levelconcepts.”(Oct.2013.)“Deeplearningisasetofalgorithmsinmachinelearningthatattempttomodelhigh-levelabstractionsindatabyusingmodelarchitecturescomposedofmultiplenon-lineartransformations.”(Aug.2014)传统机器学习解决这些问题的思路良好的特征表达,对最终算法的准确性起了非常关键的作用,而且系统主要的计算和测试工作都耗在这一大部分。但实际中一般都是人工完成的。特征表达能不能自动地学习一些特征呢?能!DeepLearning生物学启示人脑视觉机理“视觉系统的信息处理”:可视皮层是分级的神经-中枢-大脑的工作过程,或许是一个不断迭代、不断抽象的过程。关键词:一个是抽象,一个是迭代。从原始信号,做低级抽象,逐渐向高级抽象迭代。人类的逻辑思维,经常使用高度抽象的概念。不同水平的抽象层次化表示脑的深层结构whygodeep?深层结构能够有效被表达对相同的函数需要更少的计算单元深层结构可产生层次化特征表达允许非局部扩展可解释性多层隐变量允许统计上的组合共享深层结构有效(vision,audio,NLP等)!ComputerVisionFeaturesAudioFeaturesDeepLearning基本思想自动地学习特征假设有一堆输入I(如图像或者文本),我们设计了一个系统S(有n层),通过调整系统中参数,使得它的输出仍然是输入I,那么我们就可以自动地获取得到输入I的一系列层次特征,即S1,…,Sn。对于深度学习来说,其思想就是堆叠多个层也就是说这一层的输出作为下一层的输入。通过这种方式,就可以实现对输入信息进行分级表达了。可以略微地放松“输出等于输入”的限制深层vs浅层神经网络多隐层的人工神经网络具有优异的特征学习能力,学习得到的特征对数据有更本质的刻画,从而有利于可视化或分类深层网络结构中,高层可以综合应用低层信息低层关注“局部”,高层关注“全局”、更具有语义化深度神经网络在训练上的难度,可以通过“逐层初始化”(layer-wisepre-training)来有效克服。为自适应地学习非线性处理过程提供了一种可能的简洁、普适的结构模型深层vs浅层神经网络“深度模型”是手段,“特征学习”是目的。强调了模型结构的深度,通常有5层、6层,甚至10多层的隐层节点;明确突出了特征学习的重要性,也就是说,通过逐层特征变换,将样本在原空间的特征表示变换到一个新特征空间,从而使分类或预测更加容易。与人工规则构造特征的方法相比,利用大数据来学习特征,更能够刻画数据的丰富内在信息。BP算法的问题需要带标签训练数据几乎所有的数据是无标签的人脑可以从无标签数据中学习局部极小对深层网络远离了最优解学习时间不适合大数据梯度消失Belowtopfewlayers,correctionsignalisminimalniiiiijijjkkjiiiijjidyyyyyxdyyyx1111克服BP的限制梯度方法对输入的结构建模建立产生输入的生成式模型,调整参数使得生成式模型的概率最大Learnp(image)notp(label|image)Whatkindofgenerativemodelshouldwelearn?Deeplearning训练自下向上的非监督学习(greedylaye