xxAI - Beyond Explainable AI. International Workshop Held in Conjunction with ICML 2020 July 18, 2020, Vienna, Austria Revised and Extended Papers

Preface
Organization
Contents
Editorial
xxAI - Beyond Explainable Artificial Intelligence
	1 Introduction and Motivation for Explainable AI
	2 Explainable AI: Past and Present
	3 Book Structure
	References
Current Methods and Challenges
Explainable AI Methods - A Brief Overview
	1 Introduction
	2 Explainable AI Methods - Overview
		2.1 LIME (Local Interpretable Model Agnostic Explanations)
		2.2 Anchors
		2.3 GraphLIME
		2.4 Method: LRP (Layer-wise Relevance Propagation)
		2.5 Deep Taylor Decomposition (DTD)
		2.6 Prediction Difference Analysis (PDA)
		2.7 TCAV (Testing with Concept Activation Vectors)
		2.8 XGNN (Explainable Graph Neural Networks)
		2.9 SHAP (Shapley Values)
		2.10 Asymmetric Shapley Values (ASV)
		2.11 Break-Down
		2.12 Shapley Flow
		2.13 Textual Explanations of Visual Models
		2.14 Integrated Gradients
		2.15 Causal Models
		2.16 Meaningful Perturbations
		2.17 EXplainable Neural-Symbolic Learning (X-NeSyL)
	3 Conclusion and Future Outlook
	References
General Pitfalls of Model-Agnostic Interpretation Methods for Machine Learning Models
	1 Introduction
	2 Assuming One-Fits-All Interpretability
	3 Bad Model Generalization
	4 Unnecessary Use of Complex Models
	5 Ignoring Feature Dependence
		5.1 Interpretation with Extrapolation
		5.2 Confusing Linear Correlation with General Dependence
		5.3 Misunderstanding Conditional Interpretation
	6 Misleading Interpretations Due to Feature Interactions
		6.1 Misleading Feature Effects Due to Aggregation
		6.2 Failing to Separate Main from Interaction Effects
	7 Ignoring Model and Approximation Uncertainty
	8 Ignoring the Rashomon Effect
	9 Failure to Scale to High-Dimensional Settings
		9.1 Human-Intelligibility of High-Dimensional IML Output
		9.2 Computational Effort
		9.3 Ignoring Multiple Comparison Problem
	10 Unjustified Causal Interpretation
	11 Discussion
	References
CLEVR-X: A Visual Reasoning Dataset for Natural Language Explanations
	1 Introduction
	2 Related Work
	3 The CLEVR-X Dataset
		3.1 The CLEVR Dataset
		3.2 Dataset Generation
		3.3 Dataset Analysis
		3.4 User Study on Explanation Completeness and Relevance
	4 Experiments
		4.1 Experimental Setup
		4.2 Evaluating Explanations Generated by State-of-the-Art Methods
		4.3 Analyzing Results on CLEVR-X by Question and Answer Types
		4.4 Influence of Using Different Numbers of Ground-Truth Explanations
		4.5 Qualitative Explanation Generation Results
	5 Conclusion
	References
New Developments in Explainable AI
A Rate-Distortion Framework for Explaining Black-Box Model Decisions
	1 Introduction
	2 Related Works
	3 Rate-Distortion Explanation Framework
		3.1 General Formulation
		3.2 Implementation
	4 Experiments
		4.1 Images
		4.2 Audio
		4.3 Radio Maps
	5 Conclusion
	References
Explaining the Predictions of Unsupervised Learning Models
	1 Introduction
	2 A Brief Review of Explainable AI
		2.1 Approaches to Attribution
		2.2 Neuralization-Propagation
	3 Kernel Density Estimation
		3.1 Explaining Outlierness
		3.2 Explaining Inlierness: Direct Approach
		3.3 Explaining Inlierness: Random Features Approach
	4 K-Means Clustering
		4.1 Explaining Cluster Assignments
	5 Experiments
		5.1 Wholesale Customer Analysis
		5.2 Image Analysis
	6 Conclusion and Outlook
	A Attribution on CNN Activations
		A.1 Attributing Outlierness
		A.2 Attributing Inlierness
		A.3 Attributing Cluster Membership
	References
Towards Causal Algorithmic Recourse
	1 Introduction
		1.1 Motivating Examples
		1.2 Summary of Contributions and Structure of This Chapter
	2 Preliminaries
		2.1 XAI: Counterfactual Explanations and Algorithmic Recourse
		2.2 Causality: Structural Causal Models, Interventions, and Counterfactuals
	3 Causal Recourse Formulation
		3.1 Limitations of CFE-Based Recourse
		3.2 Recourse Through Minimal Interventions
		3.3 Negative Result: No Recourse Guarantees for Unknown Structural Equations
	4 Recourse Under Imperfect Causal Knowledge
		4.1 Probabilistic Individualised Recourse
		4.2 Probabilistic Subpopulation-Based Recourse
		4.3 Solving the Probabilistic Recourse Optimization Problem
	5 Experiments
		5.1 Compared Methods
		5.2 Metrics
		5.3 Synthetic 3-Variable SCMs Under Different Assumptions
		5.4 Semi-synthetic 7-Variable SCM for Loan-Approval
	6 Discussion
	7 Conclusion
	References
Interpreting Generative Adversarial Networks for Interactive Image Generation
	1 Introduction
	2 Supervised Approach
	3 Unsupervised Approach
	4 Embedding-Guided Approach
	5 Concluding Remarks
	References
XAI and Strategy Extraction via Reward Redistribution
	1 Introduction
	2 Background
		2.1 Explainability Methods
		2.2 Reinforcement Learning
		2.3 Credit Assignment in Reinforcement Learning
		2.4 Methods for Credit Assignment
		2.5 Explainability Methods for Credit Assignment
		2.6 Credit Assignment via Reward Redistribution
	3 Strategy Extraction via Reward Redistribution
		3.1 Strategy Extraction with Profile Models
		3.2 Explainable Agent Behavior via Strategy Extraction
	4 Experiments
		4.1 Gridworld
		4.2 Minecraft
	5 Limitations
	6 Conclusion
	References
Interpretable, Verifiable, and Robust Reinforcement Learning via Program Synthesis
	1 Introduction
	2 Background on Reinforcement Learning
	3 Programmatic Policies
		3.1 Traditional Interpretable Models
		3.2 State Machine Policies
		3.3 List Processing Programs
		3.4 Neurosymbolic Policies
	4 Synthesizing Programmatic Policies
		4.1 Imitation Learning
		4.2 Q-Guided Imitation Learning
		4.3 Updating the DNN Policy
		4.4 Program Synthesis for Supervised Learning
	5 Case Studies
		5.1 Interpretability
		5.2 Verification
		5.3 Robustness
	6 Conclusions and Future Work
	References
Interpreting and Improving Deep-Learning Models with Reality Checks
	1 Interpretability: For What and For Whom?
	2 Computing Interpretations for Feature Interactions and Transformations
		2.1 Contextual Decomposition (CD) Importance Scores for General DNNs
		2.2 Agglomerative Contextual Decomposition (ACD)
		2.3 Transformation Importance with Applications to Cosmology (TRIM)
	3 Using Attributions to Improve Models
		3.1 Penalizing Explanations to Align Neural Networks with Prior Knowledge (CDEP)
		3.2 Distilling Adaptive Wavelets from Neural Networks with Interpretations
	4 Real-Data Problems Showcasing Interpretations
		4.1 Molecular Partner Prediction
		4.2 Cosmological Parameter Prediction
		4.3 Improving Skin Cancer Classification via CDEP
	5 Discussion
		5.1 Building/Distilling Accurate and Interpretable Models
		5.2 Making Interpretations Useful
	References
Beyond the Visual Analysis of Deep Model Saliency
	1 Introduction
	2 Saliency-Based XAI in Vision
		2.1 White-Box Models
		2.2 Black-Box Models
	3 XAI for Improved Models: Excitation Dropout
	4 XAI for Improved Models: Domain Generalization
	5 XAI for Improved Models: Guided Zoom
	6 Conclusion
	References
ECQx: Explainability-Driven Quantization for Low-Bit and Sparse DNNs
	1 Introduction
	2 Related Work
	3 Neural Network Quantization
		3.1 Entropy-Constrained Quantization
	4 Explainability-Driven Quantization
		4.1 Layer-Wise Relevance Propagation
		4.2 eXplainability-Driven Entropy-Constrained Quantization
	5 Experiments
		5.1 Experimental Setup
		5.2 ECQx Results
	6 Conclusion
	References
A Whale's Tail - Finding the Right Whale in an Uncertain World
	1 Introduction
	2 Related Work
	3 Humpback Whale Data
		3.1 Image Data
		3.2 Expert Annotations
	4 Methods
		4.1 Landmark-Based Identification Framework
		4.2 Uncertainty and Sensitivity Analysis
	5 Experiments and Results
		5.1 Experimental Setup
		5.2 Uncertainty and Sensitivity Analysis of the Landmarks
		5.3 Heatmapping Results and Comparison with Whale Expert Knowledge
		5.4 Spatial Uncertainty of Individual Landmarks
	6 Conclusion and Outlook
	References
Explainable Artificial Intelligence in Meteorology and Climate Science: Model Fine-Tuning, Calibrating Trust and Learning New Science
	1 Introduction
	2 XAI Applications
		2.1 XAI in Remote Sensing and Weather Forecasting
		2.2 XAI in Climate Prediction
		2.3 XAI to Extract Forced Climate Change Signals and Anthropogenic Footprint
	3 Development of Attribution Benchmarks for Geosciences
		3.1 Synthetic Framework
		3.2 Assessment of XAI Methods
	4 Conclusions
	References
An Interdisciplinary Approach to Explainable AI
Varieties of AI Explanations Under the Law. From the GDPR to the AIA, and Beyond
	1 Introduction
		1.1 Functional Varieties of AI Explanations
		1.2 Technical Varieties of AI Explanations
		1.3 Roadmap of the Paper
	2 Explainable AI Under Current Law
		2.1 The GDPR: Rights-Enabling Transparency
		2.2 Contract and Tort Law: Technical and Protective Transparency
		2.3 Banking Law: More Technical and Protective Transparency
	3 Regulatory Proposals at the EU Level: The AIA
		3.1 AI with Limited Risk: Decision-Enabling Transparency (Art. 52 AIA)?
		3.2 AI with High Risk: Encompassing Transparency (Art. 13 AIA)?
		3.3 Limitations
	4 Beyond Explainability
		4.1 Actionable Explanations
		4.2 Connections to Algorithmic Fairness
		4.3 Quality Benchmarking
		4.4 Interventions and Co-design
	5 Conclusion
	References
Towards Explainability for AI Fairness
	1 Introduction
	2 Fairness
	3 AI Explanation
	4 Explanation for AI Fairness
		4.1 Explanation Guarantees Fairness
		4.2 Influence of Explanation on Perception of Fairness
		4.3 Fairness and Properties of Features
		4.4 Fairness and Counterfactuals
	5 Discussion
	6 Conclusion
	References
Logic and Pragmatics in AI Explanation
	1 Introduction
	2 The Logic of Explanations
	3 The Pragmatics of Explanations
		3.1 Case 1: Conversational Explanations
		3.2 Case 2: Explainable AI-Mediated Communication (XAI-MC)
	4 Usability, Explaniability and Causability
	References
Author Index