Peter Bühlmann

Título

Distributionally robust invariance learning for domain generalization and adaptation

Abstract

Statistical models and machine learning systems are increasingly deployed in populations and settings that differ from those on which they were trained, a challenge that is especially pronounced in digital health. We introduce Distributionally Robust Invariance Learning, a framework designed to identify and exploit stable structure across heterogeneous environments. We illustrate its use for domain generalization and domain adaptation using a large-scale international intensive care unit (ICU) database spanning multiple countries. We conclude with a brief discussion of the opportunities and limitations of emerging foundation models in this setting.

Keywords

• Causality
• Digital Health
• Distributional Robustness