Hyperbolic Hierarchical Representation Learning for Generalized Category Discovery.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Generalized Category Discovery (GCD) is a challenging semi-supervised learning task involving known and novel categories.
• Existing methods often map features to Euclidean space, failing to capture the inherent semantic hierarchy of data.
• This limitation hinders performance in discovering new categories and exploring rich semantic information.

Purpose of the Study:

• To propose a novel approach, Hyperbolic Hierarchical Representation Learning for GCD (HypGCD), to address limitations in current GCD methods.
• To leverage hyperbolic geometry for improved representation learning in GCD tasks.
• To enhance the discovery of novel categories by better preserving data's latent semantic structure.

Main Methods:

• HypGCD enhances data representations in hyperbolic space, complementing Euclidean space representations.
• It constructs hierarchical clusters at the instance-class level and models tree-like structures at the instance-instance level.
• The method jointly optimizes both Euclidean and hyperbolic spaces for refined feature extraction.

Main Results:

• HypGCD achieves state-of-the-art (SOTA) performance on multiple benchmark datasets.
• The approach demonstrates superior ability in generalized category discovery compared to existing methods.
• Enhanced representation in hyperbolic space proves effective for capturing semantic hierarchies.

Conclusions:

• HypGCD offers a significant advancement in generalized category discovery by effectively utilizing hyperbolic geometry.
• The proposed method provides a more robust way to learn data representations, preserving semantic hierarchies.
• This work opens new avenues for research in semi-supervised learning and representation learning.