Explainable Machine Learning

Chapter 1

Local Explanations

You will be able to explain one prediction with a local surrogate, anchor rule, Shapley attribution, SHAP approximation, or ICE curve, and identify the assumptions each method adds.

Chapter 2

Global Explanations

You will be able to decompose model behavior into main and interaction effects, measure model reliance, compare PDP and ALE plots, and recognize when averages or synthetic combinations become misleading.

Chapter 3

Example-Based Explanations

You will be able to explain predictions with representative cases, actionable counterfactuals, and influential training instances while distinguishing similarity, feasibility, and causal effect.

Chapter 4

Attribution-Based Explanations

You will be able to compute and interpret pixel or token relevance with gradients, Grad-CAM, integrated gradients, relevance propagation, and occlusion, and test whether a highlighted region supports a model claim.

Chapter 5

Concept-Based Explanations

You will be able to connect internal activations to human concepts with TCAV, concept bottlenecks, and network dissection, and evaluate whether a concept vocabulary matches the model and task.

Chapter 6

Representation-Based Explanations

You will be able to visualize learned features while recognizing the gap between an evocative visualization and a complete model explanation.

Chapter 7

Attention-Based Explanations

You will be able to read attention visualizations, explain the attention-as-explanation debate, and compare attention weights with attribution evidence.

Chapter 8

Large Language Model Explanations

You will be able to select explanation methods for fine-tuning, prompting, RAG, and agentic systems, and separate generated rationales from behavioral or trace-based evidence.

Chapter 9

Evaluating Explanations

You will be able to evaluate explanations for fidelity, faithfulness, completeness, stability, and human usefulness, and design tests that match the explanation's intended claim and audience.