You can find our paper at: https://t.co/3oRQxAgywI This paper would not have been possible without the support and supervision by @PascalMettes! 🙌

Our method leads to highly faithful embeddings that can easily be used in any downstream machine learning application! If you would like to know more or if you want to chat about hyperbolic geometry in general, come check out the poster on Wednesday at 16:30!

To avoid reliance on GPU-incompatible arbitrary-precision arithmetic, we switch to floating-point expansion arithmetic and implement this framework in a new high precision tensor library, involving several newly proposed routines for hyperbolic functions.

Our method follows the constructive Delaunay tree embedding approach, but we improve over existing methods by proposing a new way to uniformly distribute points on a hypersphere, leading to stronger separation and better embeddings.

Existing methods often heavily distort the distances between nodes and make use of arbitrary precision arithmetic, which is incompatible with GPUs, rendering the embeddings useless for downstream tasks. In our work, we propose a solution to both of these problems.

Hyperbolic embeddings of trees are incredibly powerful tools across many areas of machine learning, including network analysis, hierarchical learning and out-of-distribution detection. However, finding such embeddings has proven to be no easy task.

Excited to be in Vancouver for #ICML2025 this week! I’m here to talk about our latest work “Low-distortion and GPU-compatible tree embeddings in hyperbolic space”. If you're interested in graph embeddings and hyperbolic geometry, come and check it out! More details below 👇

(1/6)🥳 Excited to share my latest research done as part of my MSc AI thesis! We introduced Hyperbolic Compositional CLIP (HyCoCLIP)—a novel framework that leverages the hierarchical nature of hyperbolic space for learning vision-language representations using scene compositions.

Tomorrow I'll be presenting our hyperbolic learning library HypLL at the open-source software competition of #ACMMM2023! If you are around, come check it out!

Curious to check it out or want to learn more? Check out the library at https://t.co/ivL1al4vNp and https://t.co/R9IOaa6Tft. Or read the paper at https://t.co/R9DFgh3P4J. 3/3

Following recent successes hyperbolic geometry is rapidly gaining traction in the field of deep learning. Our library allows for easy integration of these new hyperbolic techniques into any PyTorch model, while also allowing for easy and transparent debugging. 2/3

Excited to share that our previously announced Hyperbolic Learning Library and the corresponding paper "HypLL: The Hyperbolic Learning Library" have been accepted for the open-source software competition at #ACMMM2023! w/ @phippli and @PascalMettes 1/3

Initial experiments show that our Poincaré ResNet is competitive and complementary to its Euclidean counterpart, while being more robust to adversarial attacks and out-of-distribution samples. Want to know more? Check out our paper at: https://t.co/IByznEK3x8 5/5

Hierarchies appear everywhere in the visual world, making hyperbolic learning an obvious choice for computer vision as well. Our Poincaré ResNet is a first step in this direction, extracting fully hyperbolic features from visual data for any downstream task you want. 4/5

In contrast, fully hyperbolic learning has already gained traction in other fields such as NLP and graph learning. This success is mostly attributed to the ability of hyperbolic space to embed hierarchical structures and the presence of these structures in many datasets. 3/5

The popularity of hyperbolic learning in computer vision is growing quickly. However, the currently available methods still use fully Euclidean networks for feature extraction and only apply hyperbolic geometry afterwards during classification. 2/5

Happy to share that our paper "Poincaré ResNet" has been accepted for #ICCV2023! w/ Erwin Berkhout and @PascalMettes 1/5

Curious to explore the benefits of hyperbolic geometry? Give it a shot! Check out the docs at https://t.co/ivL1al4vNp 4/4

With HypLL you can expect to - Seamlessly integrate hyperbolic geometry into any PyTorch model. - Enjoy easy and transparent debugging, even with multiple manifolds involved! 3/4

As the popularity of hyperbolic geometry in deep learning grows, the number of new techniques increases rapidly. However, using any of these techniques yourself has been difficult due to a lack of easy-to-use libraries for hyperbolic deep learning, until now! 2/4