How to Detect Silent Failures in ML Models
summary
Claims a practical way to estimate model performance without labels by calibrating predicted probabilities, turning each prediction into a partial confusion matrix, and summing them to get an expected confusion matrix; implements this in the NannyML open-source Python library and shows it matches realized AUC on a binarized California Housing experiment. For root cause it uses multivariate drift via PCA reconstruction error to detect changes in feature relationships. Notes real constraints: labels often delayed years or available for ~1% spot checks.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:27Setting the Stage: Loan Default Prediction
- 6:02Concept drift
- 6:51Performance matters.
- 15:07Performance Estimation - California Housing dataset
- 15:53Data Drift Detection
- 18:28Data reconstruction
- 20:10PCA Reconstruction error - examples
related talks
22mHow to Detect Silent Failures in ML Models
23mHow to detect silent failures in ML models
35mML: Beyond confusion matrix, model performance analysis
01h 33mUnderstanding and diagnosing your machine-learning models
44mHandling predictive uncertainty in Machine Learning
11mLearning from your (model's) mistakes
12mOn the Use of ML for Blackbox System Performance Prediction
25mThe Uncanny Valley of ML