Reproduce “Comparing Euclidean and Hyperbolic K-Means for Generalized Category Discovery” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mohamad Dalal, Thomas B. Moeslund, Joakim Bruslund Haurum
Mohamad Dalal and colleagues developed Hyperbolic Clustered GCD (HC-GCD), a method that performs both representation learning and clustering directly within hyperbolic space using a novel hyperbolic K-Means algorithm for the Lorentz Hyperboloid model. This approach achieved comparable or superior accuracy to state-of-the-art hyperbolic GCD methods that use Euclidean clustering, demonstrating improved cluster consistency across varying label granularities.
Key insights
- Clustering directly in hyperbolic space, following hyperbolic representation learning, consistently leads to improved accuracy in Generalized Category Discovery compared to performing Euclidean clustering.
- Hyperbolic K-Means within the Lorentz Hyperboloid model results in more consistent and semantically meaningful cluster assignments across different levels of ground truth label granularity.
- The Lorentz Hyperboloid model provides improved training stability and more consistent accuracy across random seeds when compared to the Poincar4e model in this context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.