内容简介

Google、微软和Facebook等公司正在积极发展内部的深度学习团队。对于我们而言，深度学习仍然是一门非常复杂和难以掌握的课题。如果你熟悉Python，并且具有微积分背景，以及对于机器学习的基本理解，《深度学习基础（影印版）》将帮助你开启深度学习之旅。

*检验机器学习和神经网络基础
*学习如何训练前馈神经网络
*使用TensorFlow实现你的第1个神经网络
*管理随着网络加深带来的各种问题
*建立神经网络用于分析复杂图像
*使用自动编码器实现有效的维度缩减
*深入了解从序列分析到语言检验
*掌握强化学习基础

作者简介

NikhilBuduma是Remedy的联合创始人和首席科学家，该公司位于美国旧金山，旨在建立数据驱动为主的健康管理新系统。16岁时，他在圣何塞州立大学管理过一个药物发现实验室，为资源受限的社区研发新颖而低成本的筛查方法。到了19岁，他是国际生物学奥林匹克竞赛的两枚金牌获得者。随后加入MIT，在那里他专注于开发大规模数据系统以影响健康服务、精神健康和医药研究。在MIT，他联合创立了LeanOnMe，一家全国性的非营利组织，提供匿名短信热线在大学校园内实现有效的一对一支持，并运用数据来积极影响身心健康。如今，Nikhil通过他的风投基金QVenturePartners投资硬科技和数据公司，还为MilwaukeeBrewers篮球队管理一支数据分析团队。

《深度学习基础（影印版）》内容贡献者NickLocascio是一位深度学习顾问、作家和研究者。Nick在MIT的ReginaBarzilay实验室获得了本科和工程硕士学位，专业从事NLP和计算机视觉研究。他曾工作于多个项目，从训练神经网络到编写自然语言提示，甚至与MGHRadiology部门合作将深度学习应用于乳腺X线摄影的医学辅助诊断。Nick的工作已被MITNews和CNBC报道。在其闲暇之余，Nick为财富500强企业提供私人的深度学习咨询服务。他还联合创立了标志性的MIT课程6.S191IntrotoDeepLearning，教过300余名学生，听众包括博士后和教授。

目录

Preface
1.TheNeuralNetwork
BuildingIntelligentMachines
TheLimitsofTraditionalComputerPrograms
TheMechanicsofMachineLearning
TheNeuron
ExpressingLinearPerceptronsasNeurons
Feed-ForwardNeuralNetworks
LinearNeuronsandTheirLimitations
Sigmoid，Tanh，andReLUNeurons
SoftmaxOutputLayers
LookingForward

2.TrainingFeed-ForwardNeuralNetworks
TheFast-FoodProblem
GradientDescent
TheDeltaRuleandLearningRates
GradientDescentwithSigmoidalNeurons
TheBackpropagationAlgorithm
StochasticandMinibatchGradientDescent
TestSets，ValidationSets，andOverfitting
PreventingOverfittinginDeepNeuralNetworks
Summary

3.ImplementingNeuralNetworksinTensorFIow
WhatIsTensorFlow?
HowDoesTensorFlowComparetoAlternatives?
InstallingTensorFlow
CreatingandManipulatingTensorFlowVariables
TensorFlowOperations
PlaceholderTensors
SessionsinTensorFlow
NavigatingVariableScopesandSharingVariables
ManagingModelsovertheCPUandGPU
SpecifyingtheLogisticRegressionModelinTensorFlow
LoggingandTrainingtheLogisticRegressionModel
LeveragingTensorBoardtoVisualizeComputationGraphsandLearning
BuildingaMultilayerModelforMNISTinTensorFlow
Summary

4.BeyondGradientDescent
TheChallengeswithGradientDescent
LocalMinimaintheErrorSurfacesofDeepNetworks
ModelIdentifiability
HowPeskyAreSpuriousLocalMinimainDeepNetworks?
FlatRegionsintheErrorSurface
WhentheGradientPointsintheWrongDirection
Momentum-BasedOptimization
ABriefViewofSecond-OrderMethods
LearningRateAdaptation
AdaGrad——AccumulatingHistoricalGradients
RMSProp——ExponentiallyWeightedMovingAverageofGradients
Adam——CombiningMomentumandRMSProp
ThePhilosophyBehindOptimizerSelection
Summary

5.ConvolutionalNeuralNetworks
NeuronsinHumanVision
TheShortcomingsofFeatureSelection
VanillaDeepNeuralNetworksDon'tScale
FiltersandFeatureMaps
FullDescriptionoftheConvolutionalLayer
MaxPooling
FullArchitecturalDescriptionofConvolutionNetworks
ClosingtheLooponMNISTwithConvolutionalNetworks
ImagePreprocessingPipelinesEnableMoreRobustModels
AcceleratingTrainingwithBatchNormalization
BuildingaConvolutionalNetworkforCIFAR-10
VisualizingLearninginConvolutionalNetworks
LeveragingConvolutionalFilterstoReplicateArtisticStyles
LearningConvolutionalFiltersforOtherProblemDomains
Summary

6.EmbeddingandRepresentationLearning
LearningLower-DimensionalRepresentations
PrincipalComponentAnalysis
MotivatingtheAutoencoderArchitecture
ImplementinganAutoencoderinTensorFlow
DenoisingtoForceRobustRepresentations
SparsityinAutoencoders
WhenContextIsMoreInformativethantheInputVector
TheWord2VecFramework
ImplementingtheSkip-GramArchitecture
Summary

7.ModelsforSequenceAnalysis
AnalyzingVariable-LengthInputs
Tacklingseq2seqwithNeuralN-Grams
ImplementingaPart-of-SpeechTagger
DependencyParsingandSyntaxNet
BeamSearchandGlobalNormalization
ACaseforStatefulDeepLearningModels
RecurrentNeuralNetworks
TheChallengeswithVanishingGradients
LongShort-TermMemory（LSTM）Units
TensorFlowPrimitivesforRNNModels
ImplementingaSentimentAnalysisModel
Solvingseq2seqTaskswithRecurrentNeuralNetworks
AugmentingRecurrentNetworkswithAttention
DissectingaNeuralTranslationNetwork
Summary

8.MemoryAugmentedNeuralNetworks
NeuralTuringMachines
Attention-BasedMemoryAccess
NTMMemoryAddressingMechanisms
DifferentiableNeuralComputers
Interference-FreeWritinginDNCs
DNCMemoryReuse
TemporalLinkingofDNCWrites
UnderstandingtheDNCReadHead
TheDNCControllerNetwork
VisualizingtheDNCinAction
ImplementingtheDNCinTensorFlow
TeachingaDNCtoReadandComprehend
Summary

9.DeepReinforcementLearning
DeepReinforcementLearningMastersAtariGames
WhatIsReinforcementLearning?
MarkovDecisionProcesses（MDP）
Policy
FutureReturn
DiscountedFutureReturn
ExploreVersusExploit
PolicyVersusValueLearning
PolicyLearningviaPolicyGradients
Pole-CartwithPolicyGradients
OpenAIGym
CreatinganAgent
BuildingtheModelandOptimizer
SamplingActions
KeepingTrackofHistory
PolicyGradientMainFunction
PGAgentPerformanceonPole-Cart
Q-LearningandDeepQ-Networks
TheBellmanEquation
IssueswithValueIteration
ApproximatingtheQ-Function
DeepQ-Network（DQN）
TrainingDQN
LearningStability
TargetQ-Network
ExperienceReplay
FromQ-FunctiontoPolicy
DQNandtheMarkovAssumption
DQN'sSolutiontotheMarkovAssumption
PlayingBreakoutwthDQN
BuildingOurArchitecture
StackingFrames
SettingUpTrainingOperations
UpdatingOurTargetQ-Network
ImplementingExperienceReplay
DQNMainLoop
DQNAgentResultsonBreakout
ImprovingandMovingBeyondDQN
DeepRecurrentQ-Networks（DRQN）
AsynchronousAdvantageActor-CriticAgent（A3C）
UNsupervisedREinforcementandAuxiliaryLearning（UNREAL）
Summary
Index

其他推荐