Explainable
AI systems make decisions that affect real people — hiring, lending, diagnosis. Understanding those decisions is not optional. This program examines the methods, tools, and limitations of making machine learning interpretable.
Three areas the program addresses directly
Interpretability is not a single technique — it is a cluster of distinct problems that require different approaches depending on the model type, the audience, and the decision context.
Lectures move from foundational concepts to applied methods, with attention to where each tool works and where it breaks down. No single framework is treated as a universal solution.
- Distinction between interpretable models and post-hoc explanation
- Global vs. local explanations — scope and trade-offs
- Model complexity and the accuracy-interpretability tension
- When black-box models are and are not acceptable
- LIME — local surrogate model construction
- SHAP values — Shapley-based feature attribution
- Attention mechanisms in transformer architectures
- Counterfactual explanations and recourse
- Regulatory requirements — EU AI Act, GDPR Article 22
- Communicating explanations to non-technical stakeholders
- Auditing model behaviour in production
- Failure modes of explanation tools
What makes a model explanation actually useful?
An explanation that satisfies a regulator may confuse the person it is meant to help. An explanation that feels intuitive may be technically misleading. The gap between these is where most XAI work fails.
Lectures draw on documented cases — loan rejections, medical triage tools, recidivism scoring — where explanation quality had measurable consequences. The goal is to develop judgment about when an explanation is sufficient, not just present.
How the program is structured
Sequential delivery — each stage builds directly on the previous one
Conceptual Ground
Definitions, problem framing, and the vocabulary used across the field. Establishes shared language before any method is introduced.
FoundationMethod Survey
LIME, SHAP, integrated gradients, and surrogate models examined side by side — with explicit attention to their assumptions and limits.
MethodsApplied Cases
Real deployment scenarios from healthcare, finance, and public sector. Analysis focuses on what went wrong and what a better approach would look like.
PracticeCritical Evaluation
How to assess whether an explanation is faithful, stable, and appropriate for its intended audience — including regulators and end users.
Assessment