Trustworthy AI - opening up the black box
summary
AI decisions power credit, hiring and medical care but hide errors: adversarial noise flips labels and a wolf vs husky model reports 99.9% accuracy fooled by background; a DARPA camouflage tank case shows clouded scenes drive classification. XAI tools include SHAP (Shapley) feature importance, pixel‑level saliency and model‑simplification (decision trees), applied during training or post‑hoc, explaining whole models or local decisions. A proposed real‑time explain‑and‑retrain loop surfaces where data or features are missing and whether a deployed model truly generalizes.
AI-written summary · about
chapters (4)
related talks
47mML: Explainable Artificial Intelligence explained
39mThe Search for Transparency and Accountability in the Age of...
46mCan you trust your AI?
15mIntroduction to Explainable AI- 39mSponsored Session: The Search for Transparency and Accountability in the Age of AI: RAI & XAI as Essential Tools
37mTrusted and Responsible AI (Explainability, Adversarial, Bias, Fairness)
39mTrust Cards For AI
48mExplainable Artificial Intelligence (XAI): Why, When, and How?