Hyperbolic Generalized Category Discovery: Hyperbolic visual learning and clustering for generalized category discovery
Translated title
Hyperbolic Generalized Category Discovery
Author
Dalal, Mohamad
Term
4. semester
Education
Publication year
2025
Submitted on
2025-06-04
Pages
56
Abstract
Afhandlingen undersøger, om repræsentationer i hyperbolsk geometri kan hjælpe en open-world klassifikationsmodel med både at genkende kendte klasser og opdage nye (Generalized Category Discovery, GCD). Vi tilpasser to GCD-metoder til at lære indlejninger (kompakte numeriske repræsentationer) i Lorentz-modellen for hyperbolsk geometri, og vi ændrer K-Means-algoritmen, så den kan klynge direkte i hyperbolsk rum (en ikke-parametrisk tilgang). Forsøgene viser, at den ikke-parametriske metode nyder godt af at lære i Lorentz-rum, mens den parametriske metode ikke gør og klarer sig bedst i det almindelige euklidiske rum. Vi ser også højere nøjagtighed, når K-Means udføres direkte i Lorentz-modellen; derimod er K-Means i Poincaré-modellen ustabil og kan mislykkes med at danne gyldige klynger afhængigt af initialiseringen af klyngeprototyperne.
This thesis explores whether representing data in hyperbolic geometry can help an open-world classifier both recognize known classes and discover new ones, a task known as Generalized Category Discovery (GCD). We adapt two GCD methods to learn embeddings (compact numerical representations) in the Lorentz model of hyperbolic geometry, and we modify the K-Means algorithm so it can cluster points directly in hyperbolic space (a non-parametric approach). Experiments show that the non-parametric method benefits from learning in Lorentz space, whereas the parametric method does not and performs best in standard Euclidean space. We also find higher accuracy when running K-Means directly in the Lorentz model; in contrast, K-Means in the Poincaré model is unstable and may fail to form valid clusters depending on how the cluster prototypes are initialized.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents