Hyperbolic Embedding

A repository to investigate how to use Hyperbolic Embeddings in deep learning, with a focus on music generation.

Summary

Euclidean space has zero curvature. Hyperbolic space has constant negative curvature and naturally captures hierarchical structure, making it well-suited for tree-like and hierarchical data in deep learning.

This repository contains two independent projects:

Folder	Framework	Description
torchExperiments/	PyTorch	Hyperbolic NN experiments on classification and regression tasks
HyperbolicTimbrenet/	TensorFlow	TimbreNet model extended with hyperbolic layers for music generation

References

• Hyperbolic Neural Networks
• Hyperbolic Neural Networks ++

TODO

• Adapt wrapped hyperbolic layers to the TimbreNet model
• Train the model with hyperbolic layers
• Incorporate Hyperbolic CNN and Hyperbolic VAE

---

Advisors: Denis Parra, Mircea Petrache, Rodrigo Cadiz