Explainable AI

Full course description

A key component of an artificially intelligent system is the ability to explain to a human agent the decisions, recommendations, predictions, or actions made by it and the process through which they are made. Such explainable artificial intelligence (XAI) can be required in a wide range of applications. For example, a regulator of waterways may use a decision support system to decide which boats to check for legal infringements, a concerned citizen might use a system to find reliable information about a new disease, or an employer might use an artificial advice-giver to choose between potential candidates fairly. For explanations from intelligent systems to be useful, they need to be able to justify the advice they give in a human-understandable way. This creates a necessity for techniques for automatic generation of satisfactory explanations that are intelligible for users interacting with the system. This interpretation goes beyond a literal explanation. Further, understanding is rarely an end-goal in itself. Pragmatically, it is more useful to operationalize the effectiveness of explanations in terms of a specific notion of usefulness or explanatory goals such as improved decision support or user trust. One aspect of intelligibility of an explainable system (often cited for domains such as health) is the ability

for users to accurately identify, or correct, an error made by the system. In that case it may be preferable to generate explanations that induce appropriate levels of reliance (in contrast to over- or under-reliance), supporting the user in discarding advice when the system is incorrect, but also accepting correct advice.

The following subjects will be discussed:
(1) Intrinsically interpretable models, e.g., decision trees, decision rules, linear regression.
(2) Identification of violations of assumptions; such as distribution of features, feature interaction, non-linear relationships between features; and what to do about them.
(3) Model agnostic explanations, e.g., LIME, scoped Rules (Anchors), SHAP (and Shapley values)
(4) Ethics for explanations, e.g., fairness and bias in data, models, and outputs.
(5) (Adaptive) User Interfaces for explainable AI
(6) Evaluation of explanation understandability

Maximum number of 60 students can follow this course.