How to detect silent failures in ML models
summary
Defines two concrete failure modes: data drift (change in input distribution) and concept drift (change in true mapping), and shows concept drift almost always breaks models. Uses the loan example with one-year label delay to argue labels are often unavailable, so performance is estimated by calibrating model probabilities and converting confidence shifts into an expected confusion matrix and metrics. Presents an open-source implementation that matched real ROC AUC on a California housing test, compares univariate tests (KS, chi-square) to multivariate reconstruction error via PCA and mean Euclidean distance, and lists false positives and interpretability limits.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:28Loan default prediction
- 2:31Data vs concept drift
- 6:05Monitor performance directly
- 7:49Ground truth challenges
- 9:55Confidence in performance
- 12:41Linking drops to data
- 13:44Feature-level drift detection
- 15:33Limitations of univariate drift
- 15:54Multivariate drift methods
- 21:10Summary
related talks
22mHow to Detect Silent Failures in ML Models- 22mHow to Detect Silent Failures in ML Models
25mWhy Good Models Fail After Deployment
18mReliable Maintenance of Machine Learning Models
34mTraining & Monitoring AI - Drift Detection
36mThe ML Monitoring Flow for Models Deployed to Production
01h 33mUnderstanding and diagnosing your machine-learning models
30mDetecting drift: how to evaluate and explore data drift in machine learning systems