Patterns, Predictions, and Actions

Moritz Hardt is a director at the Max Planck Institute for Intelligent Systems. Benjamin Recht is professor of electrical engineering and computer sciences at the University of California, Berkeley.

"A thorough, very clearly written overview on the subject of machine learning for those with the prerequisite mathematical tools of calculus, linear algebra and probability."---Jonathan Shock, Mathemafrica

• List of FiguresList of TablesPrefaceAcknowledgments1 IntroductionAmbitions of the twentieth centuryPattern classificationPrediction and actionChapter notes2 Fundamentals of PredictionModeling knowledgePrediction via optimizationTypes of errors and successesThe Neyman-Pearson LemmaDecisions that discriminateChapter notes3 Supervised LearningSample versus populationSupervised learningA first learning algorithm: The perceptronConnection to empirical risk minimizationFormal guarantees for the perceptronChapter notes4 Representations and FeaturesMeasurementQuantizationTemplate matchingSummarization and histogramsNonlinear predictorsChapter notes5 OptimizationOptimization basicsGradient descentApplications to empirical risk minimizationInsights from quadratic functionsStochastic gradient descentAnalysis of the stochastic gradient methodImplicit convexityRegularizationSquared loss methods and other optimization toolsChapter notes6 GeneralizationGeneralization gapOverparameterization: Empirical phenomenaTheories of generalizationAlgorithmic stabilityModel complexity and uniform convergenceGeneralization from algorithmsLooking aheadChapter notes7 Deep LearningDeep models and feature representationOptimization of deep netsVanishing gradientsGeneralization in deep learningChapter notes8 DatasetsThe scientific basis of machine learning benchmarksA tour of datasets in different domainsLongevity of benchmarksHarms associated with dataToward better data practicesLimits of data and predictionChapter notes9 CausalityThe limitations of observationCausal modelsCausal graphsInterventions and causal effectsConfoundingExperimentation, randomization, potential outcomesCounterfactualsChapter notes10 Causal Inference in PracticeDesign and inferenceThe observational basics: Adjustment and controlsReductions to model fittingQuasi-experimentsLimitations of causal inference in practiceChapter notes11 Sequential Decision Making and Dynamic ProgrammingFrom predictions to actionsDynamical systemsOptimal sequential decision makingDynamic programmingComputationPartial observation and the separation heuristicChapter notes12 Reinforcement LearningExploration-exploitation trade-offs: Regret and PAC-errorUnknown models and approximate dynamic programmingCertainty equivalence is often optimalThe limits of learning in feedback loopsChapter notes13 EpilogueBeyond pattern classification?14 Mathematical BackgroundCommon notationMultivariable calculus and linear algebraProbabilityEstimationBibliographyIndex