Beyond Euclidean Deep Learning: Aiden Durrant (University of Aberdeen)

Описание: Aiden Durrant, Assitant Professor at the University of Aberdeen, gave a tutorial on "Beyond Euclidean Deep Learning" at the NLDL 2026 Winter School in Tromsø, Norway.

For more information about the Northern Lights Deep Learning Conference, visit www.nldl.org.

Abstract:
Undoubtedly, the vast majority of deep learning advancements have implicitly assumed that data resides in a flat, Euclidean space. However, many real-world datasets, from phylogenetic trees to natural language, possess intricate hierarchical structures that are not naturally Euclidean. This tutorial challenges the commonn flat-world assumption and introduces the fundamental concepts of non-Euclidean deep learning. We will explore the motivation behind moving to curved spaces, with a keen focus on hyperbolic geometry, which provides a more natural and efficient embedding space for tree-like data. This shift in perspective has the potential to revolutionise deep learning by enabling models that better capture the inherent structure of complex data.

This session is designed to equip you with both the theoretical foundations and the practical skills to apply these techniques. We will begin by building an intuition for non-Euclidean geometries before diving into the core principles of hyperbolic deep learning and showcasing its successful application in various domains. The tutorial will feature a hands-on practical component, where you gain exposure to specialised libraries to implement, train, and visualise non-Euclidean representations. By the end of this session, you will not only understand the rationale behind non-Euclidean methods but will also be prepared to explore them for your own research problems.