What this book covers
Chapter 1, Understanding Deep Learning Anomaly Detection, describes types of anomalies and real-world use cases for anomaly detection. It provides two PyOD example walk-throughs to illustrate fundamental concepts, including challenges, opportunities, and considerations when using deep learning for anomaly detection.
Chapter 2, Understanding Explainable AI, covers an overview of XAI, including its evolution since the US Defense Advanced Research Project Agency (DARPA) initiative, its significance in the Right to Explanation and regulatory compliance context, and a holistic approach to the XAI life cycle.
Chapter 3, Natural Language Processing Anomaly Explainability, dives deep into finding anomalies within textual data. You will complete two NLP example walk-throughs to detect anomalies using AutoGluon and Cleanlab and explain the model’s output using SHapley Additive exPlanations (SHAP).
Chapter 4, Time Series Anomaly Explainability, introduces concepts and approaches to detecting anomalies within time series data. You will build a times series anomaly detector using Long Short-Term Memory (LSTM) and explain anomalies using OmniXAI’s SHAP explainer.
Chapter 5, Computer Vision Anomaly Explainability, integrates visual anomaly detection with XAI. This chapter covers various techniques for image-level and pixel-level anomaly detection. The example walk-through shows how to implement a visual anomaly detector and evaluate discriminative image regions identified by the model using a Class Activation Map (CAM) and Gradient-Weighted Class Activation Mapping (Grad-CAM).
Chapter 6, Differentiating Intrinsic versus Post Hoc Explainability, discusses intrinsic versus post hoc XAI methods at the local and global levels. The example walk-through further demonstrates the considerations when choosing either approach.
Chapter 7, Backpropagation versus Perturbation Explainability, reviews gradient-based backpropagation and perturbation-based XAI methods to determine feature importance for a model’s decision. This chapter has two example walk-throughs covering the saliency map and Local Interpretable Model-Agnostic Explanations (LIME).
Chapter 8, Model-Agnostic versus Model-Specific Explainability, evaluates how these two approaches work with example walk-throughs using Kernel SHAP and Guided Integrated Gradients (Guided IG). This chapter also outlines a working-backward methodology for choosing the model-agnostic versus the model-specific XAI method, starting with analyzing and understanding stakeholder and user needs.
Chapter 9, Explainability Evaluation Schemes, describes fundamental XAI principles recommended by the National Institute of Standards and Technology (NIST). This chapter reviews the existing XAI benchmarking landscape on how to quantify model explainability and assess the extent of interpretability.