Research map / Probabilistic and scientific modeling / Variational and generative sampling
Representational aspects of depth and conditioning in 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 |
|---|---|---|---|
| Use Gaussian padding instead of zero padding (ablation) | RealNVP · Synthetic datasets, including the 4-component Gaussian mixture, Swissroll, 2 Moons, and Checkerboard | distribution quality | Helped |
| Use zero padding instead of no padding (ablation) | RealNVP · Synthetic datasets | distribution quality | Hurt |
| Use zero padding instead of no padding (ablation) | RealNVP · Synthetic datasets | Jacobian conditioning | Hurt |
| Increase depth from 1 or 2 layers to 8 layers (ablation) | Linear affine coupling network · Random linear maps | regression loss | Helped |
| Increase depth from 4 layers to 8 layers (ablation) | Linear affine coupling network · Random linear maps | regression loss | Helped |
| Increase depth from 8 layers to 16 layers (ablation) | Linear affine coupling network · Random linear maps | regression loss | No clear effect |
| Fit random Toeplitz maps rather than random Gaussian matrices (ablation) | Linear affine coupling network · Synthetic linear-map regression | regression loss | Hurt |
| Use nonlinear RealNVP instead of a linear affine coupling network (ablation) | Affine coupling regression · Random linear maps | regression loss | Hurt |
| Use an MLP instead of an affine coupling architecture (ablation) | Elementwise tanh regression · Elementwise tanh | squared error | Helped |
| Use Gaussian padding instead of zero padding (ablation) | RealNVP · Synthetic datasets, including the 4-component Gaussian mixture, Swissroll, 2 Moons, and Checkerboard | Jacobian conditioning | Helped |
Closest papers by meaning
- On the Universality of Volume-Preserving and Coupling-Based Normalizing Flows (ICML 2024 · similarity 0.90)
- Universal Approximation Using Well-Conditioned Normalizing Flows (NeurIPS 2021 · similarity 0.89)
- Densely connected normalizing flows (NeurIPS 2021 · similarity 0.89)
- Normalizing Flows are Capable Generative Models (ICML 2025 · similarity 0.87)
- CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices (NeurIPS 2025 · similarity 0.86)
- Lifting Architectural Constraints of Injective Flows (ICLR 2024 · 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.