Current evaluation of epistemic uncertainty primarily relies on tasks such as out-of-distribution (OOD) detection and active learning. However, the Bayes-optimal decision strategies for these tasks do not align with the scores commonly used to quantify epistemic uncertainty.
Building on the epistemic reject-option framework, we evaluate epistemic uncertainty by focusing on its ability to identify regret, the reducible error. We formulate selective prediction as a constrained optimization problem over coverage, expected risk, and regret, proving that the optimal selector is a thresholded convex combination of the ground-truth aleatoric and epistemic uncertainties.
This theoretical unification reveals a weakness in recent uncertainty disentanglement literature: standard correlation metrics do not necessarily predict operational utility. Instead, we propose evaluating the achievable risk, regret, and coverage surface of the decomposition as a diagnostic for joint disentanglement and utility.
Benchmarking standard methods on datasets with dense human annotations shows that decision-theoretic rankings can significantly disagree with proxy-task rankings, including pairwise rank inversions between methods that are top-ranked on one criterion and bottom-ranked on another.
Blogger's Review: This paper introduces a novel method for evaluating epistemic uncertainty, highlighting the gap between theory and practical application, especially in uncertainty disentanglement. By optimizing selection strategies, the authors provide a more reliable decision framework for future machine learning models, which is worth noting.