Why No Model Is A Black Box
summary
Claim: any model can be explained using model-agnostic methods rather than declaring neural nets “black boxes.” The session demonstrates permutation importance, partial dependence plots, SHAP and DataRobot’s faster ZEMP as concrete tools, cites McKinsey adoption stats (78% embed AI; 22% report >5% revenue), shows a British Airways driver-analysis and an A/B tested call-centre deployment that raises outcomes when explanations accompany scores, and quantifies SHAP’s combinatorial cost (20 variables → ~1M model evaluations) motivating ZEMP.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker background
- 2:05AI’s growing business impact
- 6:20Why AI projects succeed or fail
- 9:49The black box debate explained
- 13:21Model agnostic explainability methods
- 14:25Permutation importance for business
- 16:31Partial dependence for insights
- 17:54SHAP values and their power
- 20:02Zemp for real-time explanations
- 21:51Rethinking black box models
- 23:59Key takeaways and closing
related talks
46mExplaining a Machine Learning blackbox
26mStop Explaining Black Box Models, Use Interpretable Models Instead
01h 36mOpen the Black Box: an Introduction to Model Interpretability with LIME and SHAP
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
28mManaging AI Risk: Beyond Explainability
35mTaking a peek under the hood - Interpreting black box model
31mBeyond the Black Box: Interpreting ML Models with SHAP
47mBreaking Black-Box AI