Reproduce “SEAL: Semantic-Aware Hierarchical Learning 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
Zhenqi He, Yuanpei Liu, Kai Han
Researchers from The University of Hong Kong's Visual AI Lab developed SEAL, a framework for Generalized Category Discovery that explicitly integrates natural semantic hierarchies to refine contrastive learning and ensure cross-granularity consistency, outperforming prior methods on fine-grained datasets by up to 1.4%.
Key insights
- Incorporating semantic hierarchical labels provides a strictly tighter upper bound on mutual information, offering a theoretical justification for the benefits of multi-level supervision in representation learning.
- Refining contrastive learning with semantically aware 'soft negatives' significantly enhances discriminative power by accurately capturing nuanced relationships between classes.
- Explicitly enforcing cross-granularity consistency through self-distillation is crucial for mitigating contradictory predictions and ensuring a coherent understanding of categories across a hierarchy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.