Managing AI Risk: Beyond Explainability
summary
Argues that high‑risk banking needs inherently interpretable models rather than black boxes with post‑hoc explainers, showing LIME and SHAP mislead on local decisions. Demonstrates deep ReLU nets partition input space into many linear regions — one example has 530 regions, 85 single‑sample regions — and that constrained architectures (additive GAMs with interactions, EBM / RuleFit / GAMI‑Net) plus monotonicity and L1 regularization yield only a handful of interpretable components with almost unchanged AUC (~0.77) and support adverse‑action reason codes.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:48AI ML in Banking
- 4:25Model Risk Management
- 7:36Explainable AI
- 11:21In inherently interpretable model
- 12:22How to design interpretable model
- 15:23Examples of interpretable models
related talks
24mToward Trustworthy AI: Explainability and Robustness
16mBuilding Trust through Explainable AI
01h 29mNavigating ML Explainability: From Theory to Implementation
46mAI for High-Stakes Decision Making
35mWhy Do You Need an Explainable Ai?
15mIntroduction to Explainable AI
24mWhy and How to Build Explainability into your ML Workflow
26mStop Explaining Black Box Models, Use Interpretable Models Instead