Evaluating Model Performance Under Worst-case Subpopulations

Mike Li, Daksh Mittal, Hongseok Namkoong +1

The performance of ML models degrades when the training population is different from that seen under operation. Towards assessing distributional robustness, we study the worst-case…