A Relativistic Perspective of Reliability in Machine Learning

Machine learning systems are expected to know that they don't know, and a large literature operationalizes this by categorizing uncertainty as aleatoric (irreducible) or epistemic (reducible with more data). We argue that this categorization implies two consequences, the goal of learning and how to define and measure epistemic uncertainty, that are not faithful to the practice of machine learning. We take this as an argument to question the epistemic-aleatoric categorization, and instead propose a relativistic alternative grounded in subjective and game-theoretic probability: internal coherence is the absence of a Dutch book on the forecaster's prices, external coherence is non-exploitability by a Sceptic with bounded information, and calibration is always relative to a class of tests, in the sense of outcome indistinguishability and multi-calibration. Reliability, on this reading, refers not to a metaphysical ground truth but to the hypothesis class, the Sceptic, and the realized data.