Research map / Probabilistic and scientific modeling / Variational and generative sampling
Path-Gradient Estimators for Continuous 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 |
|---|---|---|---|
| path-gradient estimator replacing total derivative for CNF training | Equivariant CNF for phi^4 lattice field theory · Lattice 12x12 to 32x32 | ESS: +1.65 (12x12), +6.1 (20x20), +22.3 (32x32) points | Helped |
| path-gradient estimator replacing total derivative for FFJORD VAE training | FFJORD VAE · MNIST, Omniglot, Caltech Silhouettes, Frey Faces | negative ELBO: lower on all four datasets (e.g. -2.10 nats on Caltech) | Helped |
| path-gradient estimator replacing total derivative for CNF training | Equivariant CNF for phi^4 lattice field theory · Lattice 12x12 to 32x32 | reverse KL divergence: lower at all sizes (e.g. .1833 to .0498 at 32x32) | Helped |
| path-gradient estimator per-iteration runtime overhead over total derivative | Equivariant CNF for phi^4 lattice field theory · Lattice 12x12, 20x20, 32x32 | time per iteration increase: +14%, +12%, +5% per iteration | Hurt |
| path-gradient estimator replacing total derivative for FFJORD VAE, test NLL | FFJORD VAE · MNIST, Omniglot, Caltech, Frey Faces | test negative log likelihood: MNIST 79.87 vs 80.10 (path better on MNIST, Omniglot, Caltech); Frey Faces 4.28 vs 4.36 | Mixed |
| path-gradient estimator in VAE training: wall-time per iteration | FFJORD VAE · MNIST, Omniglot | runtime relative to total estimator: up to 111% (Omniglot) and 98% (MNIST) of standard estimator runtime | Hurt |
| path-gradient estimator on number of function evaluations | FFJORD VAE · Frey Faces | NFE | No clear effect |
Closest papers by meaning
- Fast and unified path gradient estimators for normalizing flows (ICLR 2024 · similarity 0.92)
- Accelerating Continuous Normalizing Flow with Trajectory Polynomial Regularization (AAAI 2021 · similarity 0.87)
- Asymptotically exact variational flows via involutive MCMC kernels (NeurIPS 2025 · similarity 0.85)
- Stream-level Flow Matching with Gaussian Processes (ICML 2025 · similarity 0.85)
- An Error Analysis of Flow Matching for Deep Generative Modeling (ICML 2025 · similarity 0.84)
- Minimizing $f$-Divergences by Interpolating Velocity Fields (ICML 2024 · 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.