Research map / Vision-language and multimodal / CLIP and zero-shot
PHyCLIP: $\ell_1$-Product of Hyperbolic Factors Unifies Hierarchy and Compositionality in Vision-Language Representation Learning
Daiki Yoshikawa, Takashi Matsubara
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use l1-product of hyperbolic factors (k=64, d=8) instead of a single hyperbolic space (ablation) | HyCoCLIP · ImageNet zero-shot | top-1 accuracy (%): 43.80 -> 44.31 | Improved |
| use l1-product of hyperbolic factors instead of a single hyperbolic space (ablation) | HyCoCLIP · Food-101 zero-shot | top-1 accuracy (%): 52.64 -> 57.26 | Improved |
| use l1-product of hyperbolic factors instead of a single hyperbolic space (ablation) | HyCoCLIP · ImageNet hierarchical classification | TIE: 3.319 -> 3.294 | Improved |
| use l1 metric instead of l2-product metric (ablation) | PHyCLIP · ImageNet zero-shot | top-1 accuracy (%): 43.32 -> 44.31 | Improved |
| use l1 metric instead of l-infinity-product metric (ablation) | PHyCLIP · ImageNet zero-shot | top-1 accuracy (%): 6.55 -> 44.31 | Improved |
| use l1-product of hyperbolic factors instead of mixed-curvature space (ablation) | PHyCLIP · ImageNet zero-shot | top-1 accuracy (%): 39.34 -> 44.31 | Improved |
| use l1-product metric instead of mixed-curvature model (ablation) | PHyCLIP · ImageNet hierarchical classification | TIE: 3.712 -> 3.294 | Improved |
| increase factors from 64 to 128 (d=4) (ablation) | PHyCLIP · Food-101 zero-shot | top-1 accuracy (%): 57.26 -> 53.96 | Hurt |
| increase factors from 8 to 64 (total dim fixed) (ablation) | PHyCLIP · COCO text-to-image retrieval | R@5: 57.80 -> 58.03 | Improved |
| increase factors from 64 to 128 (ablation) | PHyCLIP · COCO image-to-text retrieval | R@5: 70.94 -> 71.18 | Mixed |
| add box annotations to training | CLIP · ImageNet zero-shot | top-1 accuracy (%): 38.87 -> 38.81 | No clear effect |
| add box annotations to training | MERU · ImageNet zero-shot | top-1 accuracy (%): 37.96 -> 38.08 | No clear effect |
Closest papers by meaning
- Compositional Entailment Learning for Hyperbolic Vision-Language Models (ICLR 2025 · similarity 0.90)
- Not Only Text: Exploring Compositionality of Visual Representations in Vision-Language Models (CVPR 2025 · similarity 0.88)
- Iterated Learning Improves Compositionality in Large Vision-Language Models (CVPR 2024 · similarity 0.87)
- HyperET: Efficient Training in Hyperbolic Space for Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.86)
- Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning (NeurIPS 2026 · similarity 0.86)
- Causal Graphical Models for Vision-Language Compositional Understanding (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
