Bringing together researchers to uncover the principles of Non-euclidean representations.

Within deep learning, Euclidean geometry is the default basis for deep neural networks. The naive assumption that such a grid-like perspective is optimal for all problems in computer vision does not hold. There is clear evidence that data and the representations we aim to learn can be better captured by learning in corresponding geometries that exhibit non-Euclidean structures (Nickel et al. 17, Sarkar. 11). Interest in non-Euclidean deep learning has grown dramatically in recent years, driven by advancing methodologies, libraries, and applications. The Beyond Euclidean workshop series are the frontier workshops that advocate to look beyond Euclidean geometry only for deep learning in computer vision. 

Hyperbolic deep learning is rapidly gaining traction in computer vision and beyond. Learning representations in hyperbolic space brings various new perspectives and has the potential to address open issues in canonical deep learning. Within computer vision, hyperbolic learning has shown the unique ability to learn hierarchical embeddings with minimal distortion (Spengler et al. 25), which has already shown to benefit classification/segmentation (Franco et al. 23, Guo et al. 22), self-supervised learning (Durrant et.al. 22, Flaborea et al. 23, Franco et al. 23), and more (Mettes et al. 24). Moreover, hyperbolic learning has shown to make neural networks more robust, e.g., robustness to noise (Mishra et al. 26) and out-of-distribution samples (Spengler et al. 23). Moreover, hyperbolic learning has shown to be a strong candidate as the natural embedding space for vision-language models (Pal et al. 25), due to the inherent asymmetric and hierarchical nature of vision and language.

The Beyond Euclidean workshop bring together computer vision researchers with a shared interest in exploring non-Euclidean geometry. The workshop invites researchers to submit their latest work, fostering engaging discussions through invited and contributed talks. We aim to overcome the limitations of Euclidean representations and unlock new possibilities. 

Call for papers: join us to challenge conventional perspectives and shape the future of computer vision!

Accepted Papers in Proceedings Track

Hyper3-CLIP: Hierarchy-Conditioned Hyperbolic Vision-Language Training
Matin Mahmood, Antonio Rueda-Toicen, Mohamed M. ElBassat, Seifeldin Elkerdany, Weixing Wang, Gerard de Melo

No Free Robustness: On Bolting Hyperbolic Heads onto Frozen Vision Foundation Models
Max Wilde, Mattia Sferrazza, Selman Gul, Swasti Shreya Mishra

Geometry-Aware Hyperbolic Residual-Quantized Variational Autoencoders
Alessio Colombo, Melika Ayough

Group-Equivariant Poincaré Convolutional Networks
Aiden Durrant, Rahul Baburajan, Georgios Leontidis

LoViT: Intrinsic Lorentz Vision Transformer
Mees Lindeman, Gowreesh Mago, Pritam Mishra, Pascal Mettes

VP-HyperTree: Visual-Prototype Hyperbolic VLMs for Fine-Grained Recognition and Branch-Level OOD Detection
Ivan Malashin, Vadim Tynchenko, Aleksei Borodulin, Vladislav Kukartsev, Igor Masich