Hyperbolic Generalized Category Discovery: Hyperbolic visual learning and clustering for generalized category discovery
Translated title
Hyperbolic Generalized Category Discovery
Author
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-04
Pages
56
Abstract
This thesis investigates the use of hyperbolic visual learning in Generalized Category Discovery (GCD), an open-world classification problem in which a model is tasked with classifying both known and novel classes. To that end, two GCD methods are adapted to learn embeddings in the Lorentz model of hyperbolic geometry. Furthermore, the K-Means algorithm is modified to perform non-parametric clustering on hyperbolic embeddings. The experiments showcase that the non-parametric method benefits from learning in Lorentz space, while the parametric method does not, with its best results in Euclidean geometry. Furthermore, experiments with the hyperbolic K-Means demonstrate increased accuracy when clustering embeddings directly in Lorentz space. However, K-Means in the Poincaré model of hyperbolic geometry suffers from instability, failing to generate valid clusters depending on the initialization of the cluster prototypes.
Keywords
Documents