Erklärbare KI (XAI)
Definition
Erklärbare KI zielt darauf ab, das Modellverhalten verständlich zu machen: which inputs or features drove a Entscheidung, or what tdas Modell in menschlichen Begriffen "denkt". This supports trust, debugging, and compliance.
Es unterstützt AI safety (auditing) and bias in AI (understanding unfair outcomes). Required or recommended in regulated domains (z. B. credit, hiring, healthcare). Trade-off: post-hoc explanations (SHAP, LIME) are flexible but can be approximate; inherently interpretable models are limited in expressiveness.
Funktionsweise
Feature attribution (z. B. SHAP, LIME) weist Eingabemerkmalen Wichtigkeitswerte für eine gegebene Vorhersage zu. Attention visualization shows which tokens or regions the model attended to. Natural language explanations (z. B. from an LLM or a dedicated module) describe the Entscheidung in words. Inherently interpretable models (z. B. linear models, Entscheidung trees, rule lists) are interpretable by Entwurf. Die Wahl hängt ab vauf dem model type and use case: post-hoc methods work with black boxes but may not reflect the true mechanism; interpretable models are more faithful but less flexible. Evaluate explanations for fidelity (do they nachzuahmen the model?) and usefulness (do they help users or auditors?). Integrate with evaluation and bias audits where needed.
Anwendungsfälle
Explainability ist wichtig, wenn users or regulators need to understand why a model made a given Entscheidung (compliance, trust, debugging).
- Explaining credit, hiring, or medical Entscheidungs for compliance and users
- Debugging and improving model behavior via attributions
- Building trust and transparency in high-stakes applications
Externe Dokumentation
- Interpretable Machine Learning (Molnar) — Free online book
- SHAP documentation