Research map / Probabilistic and scientific modeling / Variational and generative sampling
A Wrapped Normal Distribution on Hyperbolic Space for Gradient-Based Learning
Yoshihiro Nagano, Shoichiro Yamaguchi, Yasuhiro Fujita, Masanori Koyama
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 hyperbolic rather than Euclidean latent geometry | VAE on the synthetic binary tree · Synthetic Binary Tree | Correlation: 0.644 {\scriptstyle \pm .011} -> 0.768 {\scriptstyle \pm .003} | Improved |
| Use hyperbolic rather than Euclidean latent geometry | VAE on the noisy synthetic binary tree · Synthetic Binary Tree with noise | Correlation: 0.550 {\scriptstyle \pm .012} -> 0.590 {\scriptstyle \pm .018} | Improved |
| Use hyperbolic rather than Euclidean latent geometry | VAE on binarized MNIST with 2 latent dimensions · MNIST | log-likelihood: -140.45 {\scriptstyle \pm .47} -> -138.61 {\scriptstyle \pm 0.45} | Improved |
| Use hyperbolic rather than Euclidean latent geometry | VAE on binarized MNIST with 5 latent dimensions · MNIST | log-likelihood: -105.78 {\scriptstyle \pm .51} -> -105.38 {\scriptstyle \pm 0.61} | Improved |
| Use hyperbolic rather than Euclidean latent geometry | VAE on binarized MNIST with 10 latent dimensions · MNIST | log-likelihood: -86.25 {\scriptstyle \pm .52} -> -86.40 {\scriptstyle \pm 0.28} | Hurt |
| Use hyperbolic rather than Euclidean latent geometry | VAE on binarized MNIST with 20 latent dimensions · MNIST | log-likelihood: -77.89 {\scriptstyle \pm .36} -> -79.23 {\scriptstyle \pm 0.20} | Hurt |
| Use hyperbolic rather than Euclidean latent geometry | VAE trained on Breakout trajectories · Atari 2600 Breakout | correlation with cumulative reward: 0.712 -> 0.8460 | Improved |
| Use hyperbolic rather than Euclidean geometry | Probabilistic word embedding with diagonal variance, 5 dimensions · WordNet nouns | MAP: 0.296 {\scriptstyle \pm .006} -> 0.506 {\scriptstyle \pm .017} | Improved |
| Use hyperbolic rather than Euclidean geometry | Probabilistic word embedding with diagonal variance, 10 dimensions · WordNet nouns | MAP: 0.778 {\scriptstyle \pm .007} -> 0.795 {\scriptstyle \pm .007} | Improved |
| Use hyperbolic rather than Euclidean geometry | Probabilistic word embedding with diagonal variance, 20 dimensions · WordNet nouns | MAP: 0.894 {\scriptstyle \pm .002} -> 0.897 {\scriptstyle \pm .005} | No clear effect |
| Use hyperbolic rather than Euclidean geometry | Probabilistic word embedding with diagonal variance, 50 dimensions · WordNet nouns | MAP: 0.942 {\scriptstyle \pm .003} -> 0.975 {\scriptstyle \pm .001} | Improved |
| Use hyperbolic rather than Euclidean geometry | Probabilistic word embedding with diagonal variance, 100 dimensions · WordNet nouns | MAP: 0.953 {\scriptstyle \pm .002} -> 0.978 {\scriptstyle \pm .002} | Improved |
Closest papers by meaning
- A Rotated Hyperbolic Wrapped Normal Distribution for Hierarchical Representation Learning (NeurIPS 2022 · similarity 0.91)
- Random Laplacian Features for Learning with Hyperbolic Space (ICLR 2023 · similarity 0.89)
- Representation Tradeoffs for Hyperbolic Embeddings (ICML 2018 · similarity 0.89)
- Robust large-margin learning in hyperbolic space (NeurIPS 2020 · similarity 0.89)
- Hyperbolic Embeddings of Supervised Models (NeurIPS 2024 · similarity 0.89)
- Hyperbolic Representation Learning: Revisiting and Advancing (ICML 2023 · similarity 0.89)
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.
