Explainable
Artificial Intelligence
A structured program on making machine learning models interpretable. Covers the theory behind explanation methods and the practical tools used to apply them on real datasets.
From foundational interpretability concepts to advanced attribution methods across model types.
Each notebook targets a specific tool or dataset type — tabular, image, or natural language.
SHAP, LIME, integrated gradients, and counterfactual methods — applied in context, not in isolation.
No fixed schedule. Work through modules at your own pace with sequential content unlocking.
What the program actually covers
Eight modules, structured around the questions practitioners genuinely face.
Most XAI tutorials show you how to run SHAP on a toy dataset. This program starts there and then moves into the harder questions — when does a local explanation mislead, how do you choose between global and instance-level methods, and what does an explanation actually need to communicate to a non-technical auditor.
The later modules address model-specific vs. model-agnostic approaches, with explicit coverage of where each breaks down. Counterfactual explanations get their own module because they surface a different class of problem than attribution methods do.
What each module type covers
Concept lecture
Defines the method, its assumptions, and where it applies. No code.
Code notebook
Annotated Python implementation on a real or realistic dataset.
Critique brief
Short written analysis of where the method produces unreliable output.
Comparison exercise
Side-by-side output from two methods on the same model — spot the differences.
Practical questions, direct answers
Things worth knowing before you start — not sales copy, just specifics.
Developers, data scientists, and ML engineers who already build models and need to explain their outputs — to auditors, product teams, or regulators. It is not an introduction to machine learning.
Working knowledge of Python and familiarity with how supervised learning models are trained and evaluated. You do not need prior exposure to XAI tools — that is what the program covers.
SHAP and LIME get the most coverage because they are the most widely used in production. The program also covers integrated gradients for neural networks, counterfactual explanations, and partial dependence plots. Each is applied to at least two dataset types.
Sequential video lectures, annotated notebooks, and short written briefs — all on-demand. Estimated time to complete all eight modules is 30 to 40 hours depending on how deeply you work through the code exercises.
One module is dedicated to explanation requirements under EU AI Act provisions and GDPR Article 22. It is framed practically — what an explanation needs to contain, not a legal analysis of the legislation itself.