内容简介

  通过具体的例子、很少的理论以及两款成熟的Python框架:Scikit-Learn和TensorFlow,作者AurélienGéron会帮助你掌握构建智能系统所需要的概念和工具。你将会学习到各种技术,从简单的线性回归及发展到深度神经网络。每章的练习有助于你运用所学到的知识,你只需要有一些编程经验就行了。
  探索机器学习,尤其是神经网络
  使用Scikit-Learn全程跟踪一个机器学习项目的例子
  探索各种训练模型,包括:支持向量机、决策树、随机森林以及集成方法
  使用TensorFlow库构建和训练神经网络
  深入神经网络架构,包括卷积神经网络、循环神经网络和深度强化学习
  学习可用于训练和缩放深度神经网络的技术
  运用实际的代码示例,无需了解过多的机器学习理论或算法细节

作者简介

  AurélienGéron,是一名机器学习顾问。作为一名前Google职员,在2013至2016年间,他领导了YouTube视频分类团队。在2002至2012年间,他身为法国主要的无线ISPWifirst的创始人和CTO,在2001年他还是Polyconseil的创始人和CTO,这家公司现在管理着电动汽车共享服务Autolib'。

精彩书评

  “《Scikit-Learn与TensorFlow机器学习实用指南(影印版)》很好地介绍了利用神经网络解决问题的相关理论与实践。它涵盖了构建高效应用涉及的关键点以及理解新技术所需的背景知识。我向有兴趣学习实用机器学习的读者推荐这《Scikit-Learn与TensorFlow机器学习实用指南(影印版)》。”
  ——PeteWarden
  TensorFlow移动部门主管

目录

Preface

PartⅠ.TheFundamentalsofMachineLearning
1.TheMachineLearningLandscape
WhatIsMachineLearning?
WhyUseMachineLearning?
TypesofMachineLearningSystems
Supervised/UnsupervisedLearning
BatchandOnlineLearning
Instance-BasedVersusModel-BasedLearning
MainChallengesofMachineLearning
InsufficientQuantityofTrainingData
NonrepresentativeTrainingData
Poor-QualityData
IrrelevantFeatures
OverfittingtheTrainingData
UnderfittingtheTrainingDatateppingBack
TestingandValidating
Exercises
2.End-to-EndMachineLearningProject
WorkingwithRealData
LookattheBigPicture
FrametheProblem
SelectaPerformanceMeasure
ChecktheAssumptions
GettheData
CreatetheWorkspace
DownloadtheData
TakeaQuickLookattheDataStructure
CreateaTestSet
DiscoverandVisualizetheDatatoGainInsights
VisualizingGeographicalData
LookingforCorrelations
ExperimentingwithAttributeCombinations
PreparetheDataforMachineLearningAlgorithms
DataCleaning
HandlingTextandCategoricalAttributes
CustomTransformers
FeatureScaling
TransformationPipelines
SelectandTrainaModel
TrainingandEvaluatingontheTrainingSet
BetterEvaluationUsingCross-Validation
Fine-TuneYourModel
GridSearch
RandomizedSearch
EnsembleMethods
AnalyzetheBestModelsandTheirErrors
EvaluateYourSystemontheTestSet
Launch,Monitor,andMaintainYourSystem
TryItOut!
Exercises
3.Classification
MNIST
TrainingaBinaryClassifier
PerformanceMeasures
MeasuringAccuracyUsingCross-Validation
ConfusionMatrix
PrecisionandRecall
Precision/RecallTradeoff
TheROCCurve
MulticlassClassification
ErrorAnalysis
MultilabelClassification
MultioutputClassification
……

PartⅡ.NeuralNetworksandDeepLearning
A.ExerciseSolutions
B.MachineLearningProjectChecklist
C.SVMDualProblem
D.Autodiff
E.OtherPopularANNArchitectures
Index

精彩书摘

  《Scikit-Learn与TensorFlow机器学习实用指南(影印版)》:
  3.Itisquitepossibletospeeduptrainingofabaggingensemblebydistributingitacrossmultipleservers,sinceeachpredictorintheensembleisindependentoftheothers.ThesamegoesforpastingensemblesandRandomForests,forthesamereason.However,eachpredictorinaboostingensembleisbuiltbasedonthepreviouspredictor,sotrainingisnecessarilysequential,andyouwillnotgainanythingbydistributingtrainingacrossmultipleservers.Regardingstackingensembles,allthepredictorsinagivenlayerareindependentofeachother,sotheycanbetrainedinparallelonmultipleservers.However,thepredictorsinonelayercanonlybetrainedafterthepredictorsinthepreviouslayerhaveallbeentrained.
  4.Without-of-bagevaluation,eachpredictorinabaggingensembleisevaluatedusinginstancesthatitwasnottrainedon(theywereheldout).Thismakesitpos-sibletohaveafairlyunbiasedevaluationoftheensemblewithouttheneedforanadditionalvalidationset.Thus,youhavemoreinstancesavailablefortraining,andyourensemblecanperformslightlybetter.
  5.WhenyouaregrowingatreeinaRandomForest,onlyarandomsubsetofthefeaturesisconsideredforsplittingateachnode.ThisistrueaswellforExtra-Trees,buttheygoonestepfurther:ratherthansearchingforthebestpossiblethresholds,likeregularDecisionTreesdo,theyuserandomthresholdsforeachfeature.Thisextrarandomnessactslikeaformofregularization:ifaRandomForestoverfitsthetrainingdata,Extra-Treesmightperformbetter.Moreover,sinceExtra-Treesdon'tsearchforthebestpossiblethresholds,theyaremuchfastertotrainthanRandomForests.However,theyareneitherfasternorslowerthanRandomForestswhenmakingpredictions.
  6.IfyourAdaBoostensembleunderfitsthetrainingdata,youcantryincreasingthenumberofestimatorsorreducingtheregularizationhyperparametersofthebaseestimator.Youmayalsotryslightlyincreasingthelearningrate.
  ……

其他推荐