marlin-codes/Awesome-Hyperbolic-Representation-and-Deep-Learning

What it is

Awesome-Hyperbolic-Representation-and-Deep-Learning is a community-curated list (an “Awesome” list) of academic papers, organized so researchers can quickly find relevant work in a specific area of machine learning: using hyperbolic space instead of the usual “flat” space to represent data.

The core idea it tracks

Most machine learning works with Euclidean space — the flat geometry we learn in school. But many real-world datasets have a tree-like or power-law structure (think social networks, knowledge graphs, taxonomies): a few things are highly connected, and the rest branch out hierarchically. In Euclidean space, the “volume” of space grows according to a polynomial formula as you move outward. In hyperbolic space, it grows exponentially. That exponential growth gives hyperbolic geometry a natural advantage at representing hierarchies and tree-like data, because there is simply “more room” to spread out nested or branching structures.

The README frames this as: hyperbolic space’s exponential-growth property acts as a prior for data with hierarchical organization, while Euclidean space’s polynomial growth is a poor fit for such data.

How it is organized

The repository groups papers into three main areas, each with subcategories:

• Core Methods & Geometry (Section 2) — the foundational techniques: hyperbolic embeddings, neural networks, graph neural networks, transformers/attention, theory and stability analysis, hierarchy modeling, mixed-curvature approaches, generative models, classifiers, operations/utilities, and LLM/foundation-model connections.
• Domain Applications (Section 3) — where these methods are applied: language/text, computer vision/graphics/robotics, graphs/networks, recommender systems, knowledge graphs, biology/molecular learning, code representation, and education.
• Task-Oriented Settings (Section 4) — cross-domain problem setups: multi-label/hierarchical classification, metric learning/retrieval/clustering, geometry learning and adaptation, few-shot/low-data learning, open-vocabulary/zero-shot/category discovery, safety/robustness/privacy, environmental monitoring, and multi-criteria learning.

Each paper entry lists the paper title, the conference it appeared at, and its authors.

What else it includes

• A linked survey paper (“A Survey of Non-Euclidean Geometric Learning from Hyperbolic to Mixed-Curvature Spaces”) that covers the field broadly.
• A “Latest Update” section showing recent additions from major AI conferences (NeurIPS, ECCV, ACM MM, MICCAI, RecSys, AISTATS, WACV, CVPR, and others), with dates and paper counts.
• A community Slack channel and website for discussions and collaboration.
• A contact email for corrections, because the maintainers acknowledge the list will inevitably have oversights.

Who it’s for

Researchers, graduate students, and practitioners working at the intersection of geometry and deep learning — especially anyone studying hierarchical or graph-structured data who wants to see what hyperbolic methods exist, where they’ve been applied, and who is publishing in the area. It’s a discovery tool, not a tutorial: it lists and links papers but does not provide code or explanatory text itself.