Epistemic vs Aleatoric Uncertainty
Aleatoric uncertainty is irreducible randomness in the system; epistemic uncertainty is reducible ignorance about the model, and the two demand different responses.
Two kinds of not-knowing
Aleatoric uncertainty arises from inherent variability, such as measurement noise or stochastic physics, and cannot be reduced by collecting more data of the same kind. Epistemic uncertainty arises from limited knowledge, such as sparse training data or an uncertain model form, and shrinks as data and understanding grow. Separating them tells you whether more data will help.
Why the distinction is actionable
- High epistemic uncertainty: gather more data, refine the model, or restrict predictions to well-sampled regions
- High aleatoric uncertainty: accept the variability, design for robustness, or improve measurement precision at the source
How models express each
In a deep ensemble the average of member variances is aleatoric and the variance of member means is epistemic. In a Gaussian process the nugget term is aleatoric while the posterior variance from data sparsity is epistemic. In a Bayesian neural network the output likelihood is aleatoric and the weight posterior spread is epistemic.
Behavior in extrapolation
Epistemic uncertainty should grow far from training data; a model whose uncertainty stays flat in extrapolation is failing to represent its own ignorance, a common defect of poorly calibrated networks. Aleatoric uncertainty, by contrast, reflects the data-generating process and may be roughly constant or vary with input.
Cautions
The split is model-dependent, not absolute: what looks aleatoric under a coarse model may be explainable, and therefore epistemic, under a richer one. Report both components and treat the decomposition as a diagnostic guide rather than a fixed property of the world. For safety-critical use, verify that epistemic uncertainty actually rises where the model has no support.