Explainable AI (XAI) & AI Governance

Explainability vs. interpretability vs. transparency, Scope of explanations (local/global, model-agnostic/model-specific), Explanation types (counterfactuals, prototypes, saliency maps), Human-centered XAI design principles, Mental models and explanation satisfaction, Trade-offs between accuracy and explainability, Regulatory requirements driving XAI (GDPR right to explanation, AI Act).

LIME (Local Interpretable Model-agnostic Explanations) algorithm, Kernel density weighting and sparse linear models, SHAP (SHapley Additive exPlanations) - Shapley values from game theory, Kernel SHAP vs. Tree SHAP, Anchors for high-precision rule extraction, Partial Dependence Plots (PDP) and Individual Conditional Expectation (ICE), Accumulated Local Effects (ALE) plots.

Gradient-based attribution methods (Vanilla gradients, Integrated Gradients, Saliency maps), Layer-wise Relevance Propagation (LRP), DeepLIFT and Deep Taylor decomposition, Attention visualization and interpretability, Decision tree surrogate models, Prototypical networks and case-based explanations, Intrinsic interpretable models (rule lists, generalized additive models).

Bias sources (representation, historical, measurement bias), Fairness metrics (demographic parity, equal opportunity, equalized odds), Group fairness vs. individual fairness, Counterfactual fairness, Algorithmic discrimination quantification, Fairness interventions (pre/pre/post-processing), Intersectional fairness and multi-attribute discrimination, Responsible AI frameworks.

AI governance maturity models, AI risk management frameworks (NIST AI RMF, ISO/IEC 42001), Algorithmic accountability and auditability, Model cards and datasheets for datasets, AI ethics boards and review processes, Impact assessments (Data Protection Impact Assessment), International AI regulations (EU AI Act, US Executive Order), Deployment governance (human oversight, contestability rights).