Rethinking Aleatoric and Epistemic Uncertainty

TL;DR

This paper critiques the traditional aleatoric and epistemic uncertainty framework in machine learning, proposing a decision-theoretic perspective for clearer understanding and better measurement of uncertainty and data quality.

Contribution

It introduces a decision-theoretic approach to better distinguish and measure different types of uncertainty in machine learning predictions.

Findings

Existing information-theoretic measures can poorly estimate true uncertainties.

The proposed framework clarifies the relationship between uncertainty, performance, and data variability.

Insights into how information-theoretic quantities can guide data acquisition.

Abstract

The ideas of aleatoric and epistemic uncertainty are widely used to reason about the probabilistic predictions of machine-learning models. We identify incoherence in existing discussions of these ideas and suggest this stems from the aleatoric-epistemic view being insufficiently expressive to capture all the distinct quantities that researchers are interested in. To address this we present a decision-theoretic perspective that relates rigorous notions of uncertainty, predictive performance and statistical dispersion in data. This serves to support clearer thinking as the field moves forward. Additionally we provide insights into popular information-theoretic quantities, showing they can be poor estimators of what they are often purported to measure, while also explaining how they can still be useful in guiding data acquisition.