Reproduce “Weakly-Supervised Residual Evidential Learning for Multi-Instance Uncertainty Estimation” on OpenResearch

About this paper

Pei Liu, Luping Ji

Researchers at the University of Electronic Science and Technology of China developed Multi-Instance Residual Evidential Learning (MIREL), a scheme for estimating predictive uncertainty in weakly-supervised multi-instance learning (MIL) tasks. MIREL improves existing MIL networks' uncertainty quantification and OOD detection by up to 20.85% in instance-level AUROC on synthetic datasets and over 10% on real-world histopathology data.

Key insights

• A direct instance estimator derived from the Fundamental Theorem of Symmetric Functions is more effective for instance-level uncertainty estimation in MIL than traditional attention-based proxies.
• Learning instance-specific residuals through a dedicated neural network can effectively compensate for potential biases from bag-level aggregation, enhancing fine-grained uncertainty quantification.
• Aggregating instance evidences with estimated probabilities of positivity under weak supervision provides a theoretically justified and empirically effective method for optimizing instance-level uncertainty in positive bags.

Summary by alphaXiv.