Research map / Architectures and efficiency / RNNs and normalization
Representing Hyperbolic Space Accurately using Multi-Component Floats
Tao Yu, Christopher De Sa
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 |
|---|---|---|---|
| represent upper-half space coordinates with 2-component floats (2-xMC-Halfspace) instead of plain floats | Halfspace · Mammals 2D | MAP: 92.07% -> 93.85% | Improved |
| use Poincare upper-half space instead of Lorentz model | hyperbolic embedding · Mammals 2D | MAP: 85.75% -> 92.07% | Improved |
| 2-component floats vs plain upper-half space | Halfspace · WordNet Nouns 2D | MAP (%): 23.31 -> 24.42 | Improved |
| 2-component floats vs L-tiling | 2-xMC-Halfspace · WordNet Nouns 2D | MAP (%): 23.39 -> 24.42 | Improved |
| 2-component floats vs plain halfspace | Halfspace · WordNet Verbs 2D | MAP (%): 63.65 -> 64.88 | Improved |
| 2-component floats vs plain halfspace | Halfspace · Gr-QC 2D | MAP (%): 55.60 -> 56.35 | Improved |
| increase embedding dimension from 2 to 5 and 10 | 2-xMC-Halfspace · WordNet Nouns | MAP (%): 24.42 -> 75.05 | Improved |
| 2-component floats vs Lorentz at 10D | 2-xMC-Halfspace · WordNet Nouns 10D | MAP (%): 78.43 -> 78.46 | No clear effect |
| 2-component floats vs Lorentz at 10D | 2-xMC-Halfspace · WordNet Verbs 10D | MAP (%): 90.71 -> 90.71 | No clear effect |
| use MCF on GPU instead of tiling models on CPU | 2-xMC-Halfspace · WordNet Nouns batch 128 | training time: saves 30.27 hours over 1000 epochs | Improved |
| MCF (high precision) vs ordinary float halfspace | Halfspace · WordNet Nouns batch 128 2D | seconds per epoch: 17.4 -> 54 | Hurt |
| MCF with points near origin | 2-xMC-Halfspace · embeddings close to origin | MAP | No clear effect |
Closest papers by meaning
- The Numerical Stability of Hyperbolic Representation Learning (ICML 2023 · similarity 0.89)
- Hyperbolic Representation Learning: Revisiting and Advancing (ICML 2023 · similarity 0.88)
- Random Laplacian Features for Learning with Hyperbolic Space (ICLR 2023 · similarity 0.87)
- Hyperbolic Dataset Distillation (NeurIPS 2025 · similarity 0.85)
- Low-distortion and GPU-compatible Tree Embeddings in Hyperbolic Space (ICML 2025 · similarity 0.85)
- Hyperbolic Embeddings of Supervised Models (NeurIPS 2024 · similarity 0.84)
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.
