Reproduce “PHyCLIP: $\ell_1$-Product of Hyperbolic Factors Unifies Hierarchy and Compositionality in Vision-Language Representation Learning” 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
Daiki Yoshikawa, Takashi Matsubara
PHyCLIP, a vision-language model, integrates hierarchical and compositional semantic structures by embedding instances into an ℓ1-product metric space of hyperbolic factors. This approach, supported by theoretical correspondences, enhances performance in zero-shot classification, retrieval, hierarchical classification, and compositional understanding tasks.
Key insights
- The ℓ1-product metric applied to a Cartesian product of hyperbolic factors provides a theoretically grounded space that effectively unifies hierarchical and compositional semantic representations.
- Individual hyperbolic factors within the product space specialize in encoding specific concept families and their internal taxonomies, leading to disentangled representations.
- Compositional concepts activate corresponding specialized factors, with the overall structure resembling a Boolean algebra where the ℓ1-product distance penalizes mismatches across these factors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.