Summary of Evaluating Model Performance Under Worst-case Subpopulations, by Mike Li et al.
Evaluating Model Performance Under Worst-case Subpopulations
by Mike Li, Hongseok Namkoong, Shangzhou Xia
First submitted to arxiv on: 1 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computers and Society (cs.CY); Machine Learning (stat.ML)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed paper investigates the degradation of machine learning (ML) models when trained on one population but deployed on another. The study aims to assess distributional robustness by evaluating the worst-case performance of a model over various subpopulations defined by core attributes Z. This notion of robustness can account for complex intersectionality in disadvantaged groups and considers arbitrary continuous attributes Z. A scalable two-stage estimation procedure is developed to evaluate the robustness of state-of-the-art models, ensuring finite-sample convergence guarantees. The method’s evaluation error depends on the dimension of Z only through the out-of-sample error in estimating performance conditional on Z. Experimental results demonstrate that the approach certifies model robustness and prevents unreliable deployments. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Machine learning models can perform poorly when trained on one group but used with another. Researchers want to know how well these models will work when they’re deployed in real-world situations. They came up with a new way to measure this “distributional robustness” by looking at how well the model does on different subgroups of people, defined by characteristics like age or location. This approach can handle complex interactions between these groups and works even if there are many different attributes being considered. The team developed a method to test how well these models do in real situations and proved that it’s accurate. They tested their approach on some real-world datasets and showed that it helps identify when a model isn’t reliable enough for deployment. |
Keywords
» Artificial intelligence » Machine learning
