Epistemic Uncertainty

Epistemic uncertainty is uncertainty about what is true — uncertainty that arises from incomplete knowledge, limited data, or imperfect models, and that could in principle be reduced by gathering more evidence. It is distinguished from aleatoric uncertainty, which is uncertainty inherent in the system itself — the irreducible randomness of quantum measurement, thermal fluctuations, or genuinely stochastic processes. The distinction, rooted in decision theory and the philosophy of probability, is not merely academic: in safety-critical systems, conflating the two leads to dangerous overconfidence. A model that mistakes epistemic uncertainty for aleatoric randomness will declare itself calibrated when it is merely ignorant, while a model that treats aleatoric randomness as eliminable will waste resources chasing noise.

The proper quantification of epistemic uncertainty is one of the central challenges in modern machine learning. Bayesian neural networks attempt to capture it by placing distributions over weights, but the resulting posteriors are often computationally intractable and may underrepresent true epistemic uncertainty when the model class is misspecified. The deeper problem is that epistemic uncertainty is not a property of the data alone but of the relationship between the data, the model, and the space of hypotheses under consideration — a triadic structure that standard statistical formalism struggles to express.