Epistemic Uncertainty Is Not the Reducible Kind

2026-06-12 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

Robin Young's research from the University of Cambridge demonstrates a fundamental inconsistency in the standard taxonomy of predictive uncertainty, specifically regarding epistemic uncertainty. The paper proves that the common definition—epistemic uncertainty is reducible by more data—and its measure, the mutual-information term, are extensionally inconsistent. Through an explicit construction, the measure assigns all uncertainty to the epistemic class, yet no amount of training data reduces it. This leads to a proposed trichotomy: aleatoric, sample-reducible epistemic, and mechanism-reducible epistemic uncertainty, where the latter requires changing data acquisition methods. The work also shows that in-distribution data never reduces mechanism-irreducible uncertainty and can increase it, and that ensemble disagreement, a widely used epistemic estimate, tracks training procedures rather than the true epistemic term, collapsing to zero or reflecting initialization noise. Experiments confirm these theoretical findings with high statistical significance.

Key takeaway

For AI Scientists designing or deploying uncertainty quantification systems, you must critically re-evaluate assumptions about epistemic uncertainty. Your active learning pipelines, which often route high-epistemic inputs to additional in-distribution sampling, may be ineffective or even counterproductive for mechanism-reducible uncertainty. Furthermore, relying on ensemble disagreement as a proxy for epistemic uncertainty can be misleading, as it tracks training dynamics rather than the true underlying uncertainty. Consider diversifying data acquisition strategies beyond i.i.d. sampling to effectively reduce all forms of uncertainty.

Key insights

Standard epistemic uncertainty definitions and measures are inconsistent; reducibility depends on data acquisition methods.

Principles

• Uncertainty reducibility is a property of (uncertainty, acquisition class).
• In-distribution data can increase mechanism-irreducible epistemic uncertainty.
• Ensemble disagreement tracks optimization, not true epistemic uncertainty.

Method

A level-$\alpha$ finite-sample falsification test can reject mechanism-irreducibility by detecting a statistically significant decrease in the epistemic term under exact unidentifiability.

In practice

• Re-evaluate active learning strategies that assume i.i.d. data reduces all epistemic uncertainty.
• Do not rely on ensemble disagreement as a direct measure of epistemic uncertainty.
• Consider off-support data acquisition for truly reducing certain uncertainties.

Topics

• Uncertainty Quantification
• Epistemic Uncertainty
• Aleatoric Uncertainty
• Active Learning
• Ensemble Methods
• Data Acquisition
• Model Calibration

Best for: Research Scientist, AI Scientist

Related on AIssential

• Do Deep Ensembles Actually Capture Uncertainty in Graph Neural Networks? — Deep ensembles for GNNs suffer "epistemic collapse," limiting their uncertainty quantification benefits due to lack of diversity.
• Introduction to Deep Evidential Regression for Uncertainty Quantification — Evidential Deep Learning quantifies both epistemic and aleatoric uncertainty in a single inference pass.
• A consensus–but it’s a consensus of uncertainty. — Consensus in statistical analysis can appropriately include agreement on uncertainty or "not knowing."
• Structure-Aware Epistemic Uncertainty Quantification for Neural Operator PDE Surrogates — Restricting uncertainty sampling to the lifting module improves neural operator uncertainty quantification efficiency and accuracy.
• A Mutual Information Lower Bound for Multimodal Regression Active Learning — A Two-Index framework and MI-LB acquisition function improve active learning for multimodal continuous regression.
• Multi-Agent Reasoning with Consistency Verification Improves Uncertainty Calibration in Medical MCQA — Multi-agent reasoning with consistency verification significantly improves medical AI confidence calibration and discrimination.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.