Research map / Probabilistic and scientific modeling / Variational and generative sampling
Semi-Discrete Normalizing Flows through Differentiable Tessellation
Ricky T. Q. Chen, Brandon Amos, Maximilian Nickel
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 learnable Voronoi dequantization instead of Simplex dequantization | Discrete UCI likelihood modeling · Connect4, Forests, Mushroom, Nursery, USCensus90 | negative log-likelihood (nats): 13.46{ $\pm$0.01} -> 12.92{ $\pm$0.07}; 16.58{ $\pm$0.01} -> 14.20{ $\pm$0.05}; 9.26{ $\pm$0.01} -> 9.06{ $\pm$0.05}; 9. | Improved |
| Use learnable Voronoi dequantization instead of Discrete Flow | Discrete UCI likelihood modeling · PokerHands | negative log-likelihood (nats): 19.82{ $\pm$0.03} -> 19.86{ $\pm$0.04} | Hurt |
| Use Voronoi rather than BinaryArgmax dequantization | Permutation-invariant itemset CNF · Retail | negative log-likelihood (nats): 10.47{ $\pm$0.42} -> 9.44{ $\pm$2.34} | Improved |
| Use Voronoi rather than BinaryArgmax dequantization | Permutation-invariant itemset CNF · Accidents | negative log-likelihood (nats): 6.72{ $\pm$0.23} -> 7.81{ $\pm$2.84} | Hurt |
| Increase Voronoi embedding dimension from D=2 to D=8 (ablation) | Character-level language modeling · text8 | bits per character: 1.39{ $\pm$0.01} -> 1.36{ $\pm$0.00} | Improved |
| Increase Voronoi embedding dimension from D=2 to D=8 (ablation) | Character-level language modeling · enwik8 | bits per character: 1.46{ $\pm$0.01} -> 1.39{ $\pm$0.01} | Improved |
| Use Voronoi dequantization instead of BinaryArgmax | Character-level language modeling · text8 and enwik8 | bits per character: 1.38 -> 1.36{ $\pm$0.00}; 1.42 -> 1.39{ $\pm$0.01} | Mixed |
| Add a Voronoi disjoint mixture layer to a coupling flow | Continuous UCI density modeling · POWER, GAS, MINIBOONE | negative log-likelihood (nats): -0.44{ $\pm$0.01} -> -0.52{ $\pm$0.01}; -11.75{ $\pm$0.02} -> -12.63{ $\pm$0.05}; 10.87{ $\pm$0.06} -> 10.24{ $\pm$0.14} | Improved |
| Add a Voronoi disjoint mixture layer to a coupling flow | Continuous UCI density modeling · HEPMASS | negative log-likelihood (nats): 16.78{ $\pm$0.08} -> 16.16{ $\pm$0.01} | Hurt |
| Add a Voronoi disjoint mixture layer to a coupling flow | Continuous UCI density modeling · BSDS300 | negative log-likelihood (nats): -155.14{ $\pm$0.04} -> -156.59{ $\pm$0.14} | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- FFJORD — FFJORD: Free-Form Continuous Dynamics for Scalable Reversible Generative Models (ICLR 2019) · better on 2 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Tessellations of Semi-Discrete Flow Matching (NeurIPS 2026 · similarity 0.86)
- Annealing Flow Generative Models Towards Sampling High-Dimensional and Multi-Modal Distributions (ICML 2025 · similarity 0.85)
- Discrete Modeling via Boundary Conditional Diffusion Processes (NeurIPS 2024 · similarity 0.84)
- Winner-takes-all learners are geometry-aware conditional density estimators (ICML 2024 · similarity 0.84)
- Spherical Flows for Sampling Categorical Data (NeurIPS 2026 · similarity 0.84)
- Rectifying Categorical Flows on Statistical Manifolds for One-Step Generation (NeurIPS 2026 · 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.
