Research map / Probabilistic and scientific modeling / Variational and generative sampling
Advances in Black-Box VI: Normalizing Flows, Importance Weighting, and Optimization
Abhinav Agrawal, Daniel R. Sheldon, Justin Domke
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 |
|---|---|---|---|
| replace ADVI step scheme with comprehensive step search | full-rank Gaussian BBVI · 30 Stan models | ELBO: 10 nats or more on a small fraction | Improved |
| use STL gradient instead of closed-form entropy | full-rank Gaussian BBVI · 30 Stan models | ELBO | Improved |
| add Laplace initialization | full-rank Gaussian BBVI · 30 Stan models | ELBO | Mixed |
| add importance-weighted sampling, M=10 | Gaussian VI · 30 Stan models | ELBO: 1 nat or more on 30% of models | Improved |
| add importance-weighted training with DReG | Gaussian VI · 30 Stan models | ELBO: 1 nat of improvement to almost 30% | Mixed |
| use real-NVP normalizing flow instead of Gaussian | BBVI · 30 Stan models | ELBO: 1 nat or more on around half of models | Improved |
| use STL estimator instead of middle estimator for flows | real-NVP flow · 30 Stan models | ELBO | Mixed |
| add importance-weighted sampling to flows | real-NVP flow · 30 Stan models | ELBO | Improved |
| add importance-weighted training to flows | real-NVP flow · 30 Stan models | ELBO: 1 nat or more for almost 10% | Mixed |
| use full-rank instead of diagonal Gaussian (ablation) | Gaussian VI · 30 Stan models | ELBO: 1 nats or more on at least half | Improved |
| use DReG instead of standard IW-ELBO gradient | IW-training · 30 Stan models | ELBO | Improved |
| full path of beneficial components versus ADVI | best method (flow + STL + IW sampling) · 30 Stan models | ELBO: 1 nat or more on at least 50% | Improved |
Closest papers by meaning
- Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization (NeurIPS 2026 · similarity 0.90)
- Posterior Inference with Diffusion Models for High-dimensional Black-box Optimization (ICML 2025 · similarity 0.89)
- Efficient Mixture Learning in Black-Box Variational Inference (ICML 2024 · similarity 0.88)
- Batch and match: black-box variational inference with a score-based divergence (ICML 2024 · similarity 0.87)
- Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors (AAAI 2025 · similarity 0.87)
- Generalized Variational Inference via Optimal Transport (AAAI 2024 · similarity 0.87)
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.
