搜索
深深深度度度学学学习习习在在在图图图像像像识识识别别别中中中的的的应应应用用用戚戚戚锦锦锦秀秀秀院（系）：航天学院控制科学与工程系专业：自动化学号：1110410427指导教师：杨旭东2015年6月29日毕毕毕业业业设设设计计计（（（论论论文文文）））题题题目目目：：：深深深度度度学学学习习习在在在图图图像像像识识识别别别中中中的的的应应应用用用专专专业业业自自自动动动化化化学学学号号号1110410427学学学生生生戚戚戚锦锦锦秀秀秀指指指导导导教教教师师师杨杨杨旭旭旭东东东答答答辩辩辩日日日期期期2015年年年6月月月29日日日哈尔滨工业大学本科毕业设计（论文）摘要深度学习是近年来人工智能研究领域的一个热点，这个机器学习的分支以全连接的深度置信网络和局部连接的卷积神经网络为代表，融合概率图模型、马尔可夫链蒙特卡罗方法等多个技术，其基于大规模数据而训练出的模型在声音识别、图像识别等众多领域中获得了突破性成果。本文主要研究了深度置信网络及卷积神经网络的工作原理，并在MNIST与CIFAR数据集上进行模型的训练，利用GPU与CPU的异构计算在MNIST数据集上取得了98.9%的识别正确率，在CIFAR-10数据集上取得了62%的识别正确率。在本文中，我们还讨论了包括降维在内的多个神经网络设计技巧以及实验中发现的一些现象。关键词：深度学习；受限玻尔兹曼机；深度置信网络；卷积神经网络；MCMC；GPU计算-I-哈尔滨工业大学本科毕业设计（论文）AbstractDeeplearningasabranchofmachinelearningwhichisrepresentedbyafullcon-nectedneuralnetworknamedDeepBeliefNetworksaswellasalocalconnectednetworknamedConvolutionalNeuralNetworkshasdrawnlotsofattentioninthefieldofartificialintelligence.MultipletechnologieslikeprobabilisticgraphicalmodelsandMarkovChainMonteCarlomethodshaveintegratedintodeeplearningnowadaysandtheyhelpdeeplearningmakeagreatbreakthroughinmanyAItaskssuchasspeechandimagerecogni-tionbytrainingmodelsbasedonlarge-scaledata.ThispaperfocusontheprinciplesofDeepBeliefNetworksandConvolutionNeuralNetworks,andwetrainedsomemodelsontheMNISTandCIFAR-10datasetusingDBNsorCNNs.Asaresult,weachieved98.9%recognitionaccuracyonMNISTdatasetand62%recognitionaccuracyonCIFAR-10datasetwiththehelpofGPU&CPU-basedheterogeneouscomputing.Anumberofphenomenafoundinourexperimentsandtricksofdesigningtheneuralnetworks,includ-ingdimensionalityreduction,willalsobediscussedinthispaper.Keywords:deeplearning,restrictedboltzmannmachines,deepbeliefnetworks,convo-lutionalneuralnetworks,MarkovchainMonteCarlo,GPUcomputation-II-哈尔滨工业大学本科毕业设计（论文）目录摘要.............................................................................................IABSTRACT......................................................................................II第1章绪论....................................................................................11.1课题来源及研究的目的和意义..........................................................11.2国内外在该方向的研究现状及分析....................................................31.2.1神经网络与深度学习的发展状况.................................................31.2.2感知器时期...............................................................................31.2.3反向传播时期...........................................................................51.2.4深度学习时期...........................................................................51.3深度学习在人工智能上的应用..........................................................61.3.1语音识别..................................................................................61.3.2图像识别..................................................................................61.3.3自然语言处理...........................................................................71.4本文组织结构安排...........................................................................7第2章控制论与机器学习..................................................................82.1白盒模型与经典控制论....................................................................82.2灰盒模型与系统辨识........................................................................92.3黑盒模型与统计机器学习.................................................................92.4本章小结........................................................................................11第3章受限玻尔兹曼机.....................................................................123.1伊辛模型........................................................................................123.2玻尔兹曼机.....................................................................................133.3受限玻尔兹曼机..............................................................................143.4本章小结........................................................................................19第4章马尔可夫链蒙特卡罗方法.........................................................204.1蒙塔卡罗方法核心思想....................................................................204.2舍弃采样........................................................................................214.3重要性采样.....................................................................................254.4马尔可夫链.....................................................................................27-III-哈尔滨工业大学本科毕业设计（论文）4.5Metropolis-Hastings算法...................................................................304.6Gibbs采样.......................................................................................334.7对比离差........................................................................................354.8本章小结........................................................................................36第5章深度置信网络........................................................................375.1神经网络组成及表达能力.................................................................375.1.1神经元.....................................................................................375.1.2逻辑表达..................................................................................395.2神经网络的前馈.....