Chapter 21 – Explainable AI

21.1  Overview

21.2  Introduction

21.2.1  Why We Need Explainable AI

21.2.2  Is Explainable AI Possible?

21.3  An Example -- Query-by-Browsing

21.3.1  The Problem

21.3.2  A Solution

21.3.3  How It Works

21.4  Human Explanation -- Sufficient Reason

21.5  Local and Global Explanations

21.5.1  Decision Trees -- Easier Explanations

21.5.2  Black-box -- Sensitivity and Perturbations

21.6  Heuristics for Explanation

21.6.1  White-box Techniques

21.6.2  Black-box Techniques

21.6.3  Grey-box Techniques

21.7  Summary

accuracy, adversarial learning, autonomous car, backpropagation, base rate, bias, big data, black-box machine learning, Boolean network, boundary example, Cambridge Analytica scandal, central example, Chi squared, classification, clustering, common ground, contingency table, database, database query, decision tree, deep neural network, divide and conquer, entropy, expert system, explainable AI, extensional representation, fitness function, forward reasoning, generative adversarial network, genetic algorithm, gig economy, grey-box techniques, Grices's conversational maxims, heuristic evaluation function, hotspot analysis, ID3, image recognition, information visualisation, intentional representation, latent space, LIME, logic, machine learning, neural network, open source, overfitting, penumbra, perturbation techniques, pinch-point layer, probability, Query-by-Browsing, recommender systems, relevance feedback, search engine, self-organising map, sensitivity analysis, SHAP, sigmoid activation function, sigmoid function, similarity matrix, similarity measure, social media, social media bots, sub-symbolic systems, sufficient reason, symbolic systems, threshold, transparency, trust, visualisation, white-box model, white-box techniques, word2vec