Evaluating Epistemic Uncertainty: Beyond OOD Detection and Active Learning
For researchers in uncertainty quantification, this paper provides a theoretically grounded evaluation framework that reveals flaws in current proxy-task evaluations.
The paper evaluates epistemic uncertainty using the epistemic reject-option framework, proving that the optimal selector for selective prediction is a thresholded convex combination of aleatoric and epistemic uncertainties. It shows that standard correlation metrics for uncertainty disentanglement do not predict operational utility, and that decision-theoretic rankings can invert proxy-task rankings.
Current evaluation of epistemic uncertainty relies on tasks such as out-ofdistribution detection and active learning. However, the Bayes-optimal decision strategies for these tasks do not coincide with the scores commonly used to quantify epistemic uncertainty. Building on the epistemic reject-option framework, we evaluate epistemic uncertainty using its ability to identify regret, the reducible error. Formulating selective prediction as a constrained optimization over coverage, expected risk, and regret, we prove the optimal selector is a thresholded convex combination of the ground-truth aleatoric and epistemic uncertainties. This theoretical unification exposes a weakness in recent uncertainty disentanglement literature: we demonstrate that standard correlation metrics between learned components do not necessarily predict their actual operational utility. We instead propose to evaluate the achievable risk, regret, coverage surface of the decomposition as a diagnostic for joint disentanglement and utility. Benchmarking standard methods on datasets with dense human annotations reveals that decision-theoretic rankings can disagree substantially with proxy-task rankings, including pairwise rank inversions between methods that are top-ranked on one criterion and bottom-ranked on other.