Hyperbolic Representation Learning for Computer Vision

ECCV 2022 Tutorial

Goal of the tutorial

Learning in computer vision is all about deep networks and such networks operate on Euclidean manifolds by design. But is Euclidean geometry the best choice for deep learning or simply a practical option? Recent literature in machine learning and computer vision has shown that hyperbolic geometry provides a strong alternative, with an improved ability to embed hierarchies, graphs, text, images, and videos.

In light of recent advancements in hyperbolic representation learning for computer vision, this tutorial seeks to advocate hyperbolic geometry and its strong potential for computer vision to a broader audience. The tutorials provides a theoretical and practical starting point for the field. At the conference, we will provide an easy-going introduction to hyperbolic geometry for non-mathematicians, where we focus on intuition and high-level understanding. We then outline the current state of hyperbolic geometry for vision from supervised and unsupervised perspectives. At the end, we dive into open research problems and future potential for hyperbolic geometry and visual understanding.

Unique for this tutorial is that we do not stop at a theoretical foundation. The tutorial website will also host a series of notebook-style code snippets with foundational works on hyperbolic geometry, to get a better understanding of its workings and lower the barrier to start your dive into this exciting new research direction in computer vision.