Research map / Probabilistic and scientific modeling / Variational and generative sampling
Scalable Normalizing Flows for Permutation Invariant Densities
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 an equivariant CNF with exact trace instead of the strongest reported baseline | point-process density modeling · Check-ins NY | per-point negative log-likelihood: -2.03 $\pm$ 0.07 -> -2.30 $\pm$ 0.09 | Helped |
| use an equivariant CNF with exact trace instead of the strongest reported baseline | point-process density modeling · Check-ins Paris | per-point negative log-likelihood: -3.57 $\pm$ 0.02 -> -3.58 $\pm$ 0.03 | No clear effect |
| use an equivariant CNF with exact trace instead of the strongest reported baseline | point-process density modeling · Crimes | per-point negative log-likelihood: -2.35 $\pm$ 0.00 -> -2.34 $\pm$ 0.01 | No clear effect |
| use an equivariant CNF with exact trace instead of IHP | point-process density modeling · Matérn | per-point negative log-likelihood: -0.21 $\pm$ 0.00 -> -0.77 $\pm$ 0.08 | Helped |
| use an equivariant CNF with exact trace instead of the strongest reported baseline | point-process density modeling · Mixture | per-point negative log-likelihood: -2.06 $\pm$ 0.00 -> -2.07 $\pm$ 0.00 | No clear effect |
| use an equivariant CNF with exact trace instead of the strongest reported baseline | point-process density modeling · Thomas | per-point negative log-likelihood: -0.23 $\pm$ 0.00 -> -0.55 $\pm$ 0.00 | Helped |
| add IHP coupling layers to the zero-trace flow (ablation) | zero-trace volume-preserving flow · Matérn | per-point negative log-likelihood: -0.11 -> -0.45 | Helped |
| add IHP coupling layers to the zero-trace flow (ablation) | zero-trace volume-preserving flow · Thomas | per-point negative log-likelihood: -0.15 -> -0.22 | Helped |
| add IHP coupling layers to the zero-trace flow (ablation) | zero-trace volume-preserving flow · Mixture | per-point negative log-likelihood: -0.02 -> -2.04 | Mixed |
| replace self-attention with induced attention (m = 8) (ablation) | attention-based CNF · Matérn | per-point negative log-likelihood: -1.03 $\pm$ 0.01 -> -0.99 $\pm$ 0.00 | Hurt |
| replace self-attention with induced attention (m = 8) (ablation) | attention-based CNF · Thomas | per-point negative log-likelihood: -0.55 $\pm$ 0.00 -> -0.29 $\pm$ 0.09 | Hurt |
| replace self-attention with induced attention (m = 8) (ablation) | attention-based CNF · Mixture | per-point negative log-likelihood: -1.92 $\pm$ 0.06 -> -2.02 $\pm$ 0.03 | Helped |
Closest papers by meaning
- Set Norm and Equivariant Skip Connections: Putting the Deep in Deep Sets (ICML 2022 · similarity 0.87)
- Regularizing Towards Permutation Invariance In Recurrent Models (NeurIPS 2020 · similarity 0.86)
- Equivariant Flows: Exact Likelihood Generative Learning for Symmetric Densities (ICML 2020 · similarity 0.86)
- Permutation Equivariance of Transformers and Its Applications (CVPR 2024 · similarity 0.84)
- Not-So-Optimal Transport Flows for 3D Point Cloud Generation (ICLR 2025 · similarity 0.84)
- On the Relation between Rectified Flows and Optimal Transport (NeurIPS 2025 · similarity 0.83)
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.