Synthetic Data Reveals Generalization Gaps in Correlated Multiple Instance Learning

This paper evaluates how well current multiple instance learning (MIL) models capture critical inter-instance correlations in medical imaging. By generating our own synthetic dataset, we construct a Bayes estimator that serves as an optimal upper bound on model performance. We benchmark both correlated MIL methods (such as attention-based and transformer architectures) and non-correlated approaches, revealing that there is a generalization gap in correlated and non-correlated MIL architectures and that narrowing that gap takes much more labeled data than is normally available in the medical domain.

Recommended citation: Ethan Harvey, Dennis Johan Loevlie, and Michael C. Hughes. (2025). "Synthetic Data Reveals Generalization Gaps in Correlated Multiple Instance Learning." ML4H 2025 Symposium, Findings Track.
Download Paper