Research map / Probabilistic and scientific modeling / Variational and generative sampling
Self Normalizing Flows
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 |
|---|---|---|---|
| replace exact log-determinant gradient updates with learned approximate inverses | FC 2-Layer flow · MNIST test | negative log-likelihood (nats): 947.6 ± 0.2 -> 947.1 ± 0.2 | No clear effect |
| replace exact log-determinant gradient updates with learned approximate inverses | Conv. 9-Layer flow · MNIST test | negative log-likelihood (nats): 637.4 ± 0.2 -> 638.6 ± 0.9 | Hurt |
| replace exact log-determinant gradient updates with learned approximate inverses | Glow 2L-16K · MNIST test | negative log-likelihood (nats): 575.7 ± 0.8 -> 575.4 ± 1.4 | No clear effect |
| replace exact log-determinant gradient updates with learned approximate inverses | Glow · CIFAR-10 | bits per dimension: 3.36 ± 0.002 -> 3.37 ± 0.004 | Hurt |
| replace exact log-determinant gradient updates with learned approximate inverses | Glow · ImageNet 32x32 | bits per dimension: 4.12 ± 0.002 -> 4.14 ± 0.007 | Hurt |
| use learned approximate inverses instead of exact gradients | FC 2-Layer flow · MNIST | time / batch (ms): 44.9 ± 4.4 -> 18.7 ± 0.8 | Helped |
| use learned approximate inverses instead of exact gradients | Conv. 9-Layer flow · MNIST | time / batch (ms): 372.2 ± 24.5 -> 212.5 ± 37.3 | Helped |
| use learned approximate inverses instead of exact gradients | Glow 2L-16K · MNIST | time / batch (ms): 583.4 ± 21.2 -> 476.4 ± 16.7 | Helped |
| set reconstruction weight λ to ≤ 0.1 (ablation) | SNF FC 2-Layer · MNIST | training stability | Hurt |
| increase reconstruction weight λ from 1.0 to 10.0, 100.0, or 1000.0 (ablation) | SNF FC 2-Layer · MNIST | negative log-likelihood (nats): 946.3 -> 954.6; 946.3 -> 1007.4; 946.3 -> 1039.8 | Hurt |
| increase reconstruction weight λ from 1.0 to 10.0, 100.0, or 1000.0 (ablation) | SNF Conv. 9-Layer · MNIST | negative log-likelihood (nats): 639.6 -> 642.9; 639.6 -> 646.3; 639.6 -> 660.6 | Hurt |
| use reconstruction weight λ below 100.0 (ablation) | SNF Glow 2L-16K · MNIST | training stability | Hurt |
Closest papers by meaning
- OT-Flow: Fast and Accurate Continuous Normalizing Flows via Optimal Transport (AAAI 2021 · similarity 0.89)
- Implicit Normalizing Flows (ICLR 2021 · similarity 0.88)
- Fast and unified path gradient estimators for normalizing flows (ICLR 2024 · similarity 0.88)
- On the Universality of Volume-Preserving and Coupling-Based Normalizing Flows (ICML 2024 · similarity 0.87)
- Kernelised Normalising Flows (ICLR 2024 · similarity 0.87)
- Annealing Flow Generative Models Towards Sampling High-Dimensional and Multi-Modal Distributions (ICML 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.