Abstract

The distinction between aleatoric (irreducible) and epistemic (reducible) uncertainty is central to uncertainty-aware machine learning. Yet its conceptual meaningfulness and practical usefulness have recently been questioned. In this talk, I argue that the dichotomy remains both meaningful and indispensable. I revisit recent criticisms, highlight the importance of distinguishing the underlying concept from its mathematical frameworks and making modeling assumptions explicit. I then discuss decision-making problems in which optimal performance provably requires distinguishing between reducible and irreducible uncertainty. These considerations motivate improving existing methods rather than abandoning the distinction itself.

About the Speaker

Yusuf Sale is a PhD student in Eyke Hüllermeier’s group at LMU Munich. He is affiliated with the Munich Center for Machine Learning (MCML) and a member of the Konrad Zuse School of Excellence in Reliable AI (relAI). His research focuses on the foundations of uncertainty in machine learning, particularly on how different types of uncertainty should be understood, represented and quantified. In this context, he also studies theory and applications of distribution-free uncertainty quantification. His work has appeared at leading machine learning venues.