IA explicable (XAI)
Définition
L'IA explicable vise à rendre le comportement des modèles compréhensible: which inputs or features drove a décision, or what the model "thinks" in human terms. This supports trust, debugging, and compliance.
Il supporte AI safety (auditing) and bias in AI (understanding unfair outcomes). Required or recommended in regulated domains (par ex. credit, hiring, healthcare). Trade-off: post-hoc explanations (SHAP, LIME) are flexible but can be approximate; inherently interpretable models are limited in expressiveness.
Comment ça fonctionne
Feature attribution (par ex. SHAP, LIME) attribue des scores d'importance aux caractéristiques d'entrée pour une prédiction donnée. Attention visualization shows which tokens or regions the model attended to. Natural language explanations (par ex. from an LLM or a dedicated module) describe the décision in words. Inherently interpretable models (par ex. linear models, décision trees, rule lists) are interpretable by conception. Choice depends on the 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 match the model?) and usefulness (do they help users or auditors?). Integrate with evaluation and bias audits where needed.
Cas d'utilisation
Explainability matters when users or regulators need to understand why a model made a given décision (compliance, trust, debugging).
- Explaining credit, hiring, or medical décisions for compliance and users
- Debugging and improving model behavior via attributions
- Building trust and transparency in high-stakes applications
Documentation externe
- Interpretable Machine Learning (Molnar) — Free online book
- SHAP documentation