Robin Young's arXiv paper shatters ML uncertainty dogma

The standard taxonomy in machine learning divides predictive uncertainty into aleatoric (irreducible inherent noise) and epistemic (reducible by more data), with the standard measure using mutual information between model parameters and predictions. In a bombshell new paper, Robin Young proves these two definitions are extensionally inconsistent. Using an explicit construction, the author shows the mutual-information measure assigns all uncertainty to the epistemic class, yet no amount of in-distribution training data reduces it. This reveals a critical flaw: reducibility is not a property of the uncertainty itself but of the pair (uncertainty, acquisition class). The paper further demonstrates that in-distribution data never reduces what they call “mechanism-irreducible” uncertainty and can actually increase it.

Young proposes a refined taxonomy: aleatoric, sample-reducible epistemic, and mechanism-reducible epistemic—the latter requiring changes to the learning mechanism (e.g., model architecture or training procedure). The paper also takes aim at ensemble disagreement, the most common deployed estimate of epistemic uncertainty. Under consistent training, ensemble disagreement collapses to zero even when positive truth remains; under interpolation it equals hyperparameter-scaled initialization noise. This means it tracks training procedure, not genuine uncertainty. A finite-sample falsification test and seed-swept experiments confirm the theory. For ML practitioners, this is a wake-up call: current uncertainty quantification methods may be systematically flawed, demanding a fundamentally new conceptual foundation.