Epistemic Uncertainty Is Not the Reducible Kind

2026-06-10 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new analysis reveals that the standard taxonomy of predictive uncertainty, which defines epistemic uncertainty as reducible by more data, is inconsistent with its common measure, a mutual-information term. The study demonstrates, through an explicit construction, that this measure can assign all uncertainty to the epistemic class even when no amount of training data reduces it. The authors propose a refined dichotomy, resolving uncertainty into three parts: aleatoric, sample-reducible epistemic, and mechanism-reducible epistemic uncertainty. An exact identity shows that in-distribution data never reduces mechanism-irreducible uncertainty and often increases it. Furthermore, ensemble disagreement, the deployed epistemic estimate, tracks the training procedure rather than the true epistemic term, collapsing to zero under consistent training or equaling hyperparameter-scaled initialization noise under interpolation. Finite-sample falsification tests and seed-swept experiments confirm these theoretical findings.

Key takeaway

For AI Scientists and Machine Learning Engineers designing models or evaluating uncertainty, you should critically re-examine your understanding and measurement of epistemic uncertainty. The findings indicate that current methods may misclassify uncertainty and lead to ineffective data acquisition strategies. You must differentiate between sample-reducible and mechanism-reducible epistemic uncertainty to build more robust and reliable models, especially when interpreting ensemble disagreement as an uncertainty estimate.

Key insights

The standard definition and measure of epistemic uncertainty are inconsistent, requiring a refined taxonomy.

Principles

• Epistemic uncertainty's reducibility depends on the (uncertainty, acquisition class) pair.
• In-distribution data does not reduce mechanism-irreducible uncertainty and can increase it.
• Ensemble disagreement tracks training procedures, not true epistemic uncertainty.

In practice

• Re-evaluate uncertainty quantification methods based on the refined taxonomy.
• Interpret ensemble disagreement with caution, understanding its training dependency.
• Consider data acquisition strategies that target specific uncertainty types.

Topics

• Epistemic Uncertainty
• Uncertainty Quantification
• Aleatoric Uncertainty
• Ensemble Disagreement
• Machine Learning Theory
• Data Acquisition

Best for: Research Scientist, AI Scientist

Related on AIssential

• Epistemic Uncertainty Is Not the Reducible Kind — Standard epistemic uncertainty definitions and measures are inconsistent; reducibility depends on data acquisition methods.
• A consensus–but it’s a consensus of uncertainty. — Consensus in statistical analysis can appropriately include agreement on uncertainty or "not knowing."
• Introduction to Deep Evidential Regression for Uncertainty Quantification — Evidential Deep Learning quantifies both epistemic and aleatoric uncertainty in a single inference pass.
• 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.
• CauTion: Knowing When to Trust LLMs for Ensemble Causal Discovery — CauTion integrates LLM domain knowledge with statistical ensembles using trust calibration to improve causal discovery accuracy.
• Recent discoveries on the acquisition of the highest levels of statistical fallacies — Selection bias can create spurious negative correlations among highly selected groups.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.