Individual Gaps in Prediction Generalization
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
Shows that DenseNet and ERM-trained image models encode sex, age, race and that greater encoding correlates with larger group performance gaps (over 20% gap reduced to ~5% using group DRO with similar AUROC). Builds an OODSelect algorithm to pick out-of-distribution subsets and finds up to 70% of MIMIC-CXR examples exhibit accuracy-on-the-inverse-line so model rankings can flip. For clinical QA, human-annotated relevance consistently improves LLM and clinician accuracy; GPT-4o approaches human performance.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:11AI model evaluation limits
- 3:23Spurious correlations in imaging
- 7:13Testing generalization across hospitals
- 10:00Inverse line phenomenon
- 16:22LLMs for diagnostic accuracy
- 19:50Human vs model relevance in records
- 23:10Summary and future work
- 25:02Q&A
related talks
38mRepresentation via Representations: Improving Generalization to Unseen Populations
41mDistribution shift in ecological data: generalization vs. specialization,
37mAI in medicine: where do we stand?
56mGlobal Bias in AI: When Western Data Shapes the World
53mAre We Measuring the Right Thing? Distribution Shift Lessons for Federated Learning
49mPanel Discussion
01h 01mINTERPRETABLE AND DIFFERENTIALLY PRIVATE MACHINE LEARNING
33mLessons Learned Applying Large Language Models in Healthcare