book

Explainable AI for Practitioners

by Michael Munn, David Pitman

October 2022

Beginner to intermediate

276 pages

8h 32m

English

O'Reilly Media, Inc.

Book available

Read now

Unlock full access

Who Should Read This Book?What Is and What Is Not in This Book?Code SamplesNavigating This BookConventions Used in This BookO’Reilly Online LearningHow to Contact UsAcknowledgments
Why Explainable AIWhat Is Explainable AI?Who Needs Explainability?Challenges in ExplainabilityEvaluating ExplainabilityHow Has Explainability Been Used?How LinkedIn Uses Explainable AIPwC Uses Explainable AI for Auto Insurance ClaimsAccenture Labs Explains Loan DecisionsDARPA Uses Explainable AI to Build “Third-Wave AI”Summary
What Are Explanations?Interpretability and ExplainabilityExplainability ConsumersPractitioners—Data Scientists and ML EngineersObservers—Business Stakeholders and RegulatorsEnd Users—Domain Experts and Affected UsersTypes of ExplanationsPremodeling ExplainabilityIntrinsic Versus Post Hoc ExplainabilityLocal, Cohort, and Global ExplanationsAttributions, Counterfactual, and Example-Based ExplanationsThemes Throughout ExplainabilityFeature AttributionsSurrogate ModelsActivationPutting It All TogetherSummary
Permutation Feature ImportancePermutation Feature Importance from ScratchPermutation Feature Importance in scikit-learnShapley ValuesSHAP (SHapley Additive exPlanations)Visualizing Local Feature AttributionsVisualizing Global Feature AttributionsInterpreting Feature Attributions from Shapley ValuesManaged Shapley ValuesExplaining Tree-Based ModelsFrom Decision Trees to Tree EnsemblesSHAP’s TreeExplainerPartial Dependence Plots and Related PlotsPartial Dependence Plots (PDPs)Individual Conditional Expectation Plots (ICEs)Accumulated Local Effects (ALE)Summary
Integrated Gradients (IG)Choosing a BaselineAccumulating GradientsImprovements on Integrated GradientsXRAIHow XRAI WorksImplementing XRAIGrad-CAMHow Grad-CAM WorksImplementing Grad-CAMImproving Grad-CAMLIMEHow LIME WorksImplementing LIMEGuided Backpropagation and Guided Grad-CAMGuided Backprop and DeConvNetsGuided Grad-CAMSummary
Overview of Building Models with TextTokenizationWord Embeddings and Pretrained EmbeddingsLIMEHow LIME Works with TextGradient x InputIntuition from Linear ModelsFrom Linear to Nonlinear and Text ModelsGrad L2-normLayer Integrated GradientsA Variation on Integrated GradientsLayer-Wise Relevance Propagation (LRP)How LRP WorksDeriving Explanations from AttentionWhich Method to Use?Language Interpretability ToolSummary
Alternative Explainability TechniquesAlternate Input AttributionExplainability by DesignOther ModalitiesTime-Series DataMultimodal DataEvaluation of Explainability TechniquesA Theoretical ApproachEmpirical ApproachesSummary
Who Uses Explainability?How to Effectively Present ExplanationsClarify What, How, and Why the ML Performed the Way It DidAccurately Represent the ExplanationsBuild on the ML Consumer’s Existing UnderstandingCommon Pitfalls in Using ExplainabilityAssuming CausalityOverfitting Intent to a ModelOverreaching for Additional ExplanationsSummary
Building with Explainability in MindThe ML Life CycleAI Regulations and ExplainabilityWhat to Look Forward To in Explainable AINatural and Semantic ExplanationsInterrogative ExplanationsTargeted ExplanationsSummary

ML ConsumersTaxonomy of ExplainabilityXAI TechniquesTabular ModelsImage ModelsText ModelsAdvanced and Emerging TechniquesInteracting with ExplainabilityPutting It All TogetherFurther ReadingExplainable AIInteracting with ExplainabilityTechnical Accuracy of XAI techniquesBrittleness of XAI techniquesXAI for DNNs

Content preview from Explainable AI for Practitioners

Chapter 6. Advanced and Emerging Topics

with Sheeraz Ahmad

The focus of the book so far has been on well-established techniques, modalities, and use cases. However, Explainable AI (XAI) continues to be an active area of research, so that new techniques are continually being developed and existing techniques are improved and scrutinized further. Feature-based explanations such as Shapley values and Integrated Gradients introduced in the previous chapters can cover many use cases, especially as applied to text, tabular, and image data. However, there are several emerging techniques and topics that can be valuable in your explainability toolbox in specific situations.

In this chapter, we will discuss three broad, emerging topics. First, we will introduce alternative explanation techniques like attribution to inputs (as opposed to features) and making models explainable by design. Second, we will briefly cover how some of the previously introduced techniques can be more generally applied to data formats that are not text, tabular, or image, specifically focusing on time-series and multimodal data (text + image). Third, we will discuss how explainability techniques can be evaluated in a systematic way, as opposed to spot checks on a handful of data points.

Alternative Explainability Techniques

In this section, we will discuss two alternative explainability methods, namely, alternate input attribution, which is attribution to training data points or user-defined concepts, as well as ...