Stop Explaining Black Box Models, Use Interpretable Models Instead
summary
Argument: replace black boxes with interpretable models for high‑stakes decisions because explanations of black boxes are unreliable. Examples: COMPAS and BreezoMeter produce real harms; a sparse integer model called 2HELPS2B seizure risk score is learned via mixed‑integer optimization (RiskSLIM) from 70+ variables, validated multicenter, and yields a 63.6% reduction in EEG monitoring time per patient while maintaining accuracy. A certifiably optimal rule method (COMPAS vs CORALS) produces an extremely small decision tree that matches COMPAS accuracy, showing interpretability need not cost performance.
AI-written summary · about
chapters (26)
- 0:00Introduction
- 0:32What is a Black Box
- 0:46Black Box Predictive Models
- 1:04Prison Time
- 3:19Health Care
- 3:34Black Box Models
- 4:04Interpretable Models
- 5:07Seizure Prediction
- 6:35To Help Stupid Score
- 8:35Data Driven Decision
- 10:16Continuous EEG
- 11:43Square Didnt Lose Accuracy
- 12:09Compass Score
- 12:50Compass vs Corals
- 14:04Tenfold Crossvalidation
- 14:22Boosted Decision Trees
- 14:40Using Interpretable Models
- 15:06Interpretable vs Explainable
- 15:53Double Trouble
- 19:19Machine Bias
- 21:44Problems with Compass
- 22:06Low Compass
- 22:37Learning Performance vs Explanation Effectiveness
- 23:59Stop Explanation Paper
- 25:06Chat Session
- 25:34Risk Slim
related talks
28mManaging AI Risk: Beyond Explainability
25mWhy No Model Is A Black Box
46mExplaining a Machine Learning blackbox
42mInterpreting Black-Box Models with Applications to Healthcare
15mInterpretable Machine Learning with R
01h 06mPanel on Societal Dimensions of Explanation
38mTo explain or to predict?
28mExplaining Machine Learning Models with LLMs