Research map / Probabilistic and scientific modeling / Variational and generative sampling
Variational Russian Roulette for Deep Bayesian Nonparametrics
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 fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · SYNTH | IWAE testing: 44.93 vs 43.28 (MF-IBP) and 42.47 (S-IBP) | Helped |
| Replace fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · MNIST | activated features K~: 14 vs 50 (fixed truncation) | Helped |
| Replace fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · MNIST | IWAE testing: -104.54 vs -104.28 (S-IBP) | No clear effect |
| Replace fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · Fashion-MNIST | IWAE testing: -240.34 vs -241.52 (S-IBP) and -244.15 (MF-IBP) | Helped |
| Replace fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · Fashion-MNIST | activated features K~: 9 vs 50 | Helped |
| Replace fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · MNIST running time | time per iteration (s): about 1.35x speed-up vs fixed truncation of 50 | Helped |
| Replace fixed truncation with Russian roulette adaptive truncation (RRS-IBP) | infinite VAE with IBP prior · SYNTH linear-Gaussian | expected features per image M (true value 2.275): 3.464 vs 5.647 (MF-IBP) and 6.021 (S-IBP) | Helped |
| Use fixed truncation with structured inference (S-IBP) rather than mean-field (MF-IBP) (ablation) | infinite VAE with IBP prior · MNIST and Fashion-MNIST | IWAE testing: -104.28 vs -111.45 (MNIST); -241.52 vs -244.15 (Fashion-MNIST) | Helped |
| Increase fixed truncation level of S-IBP from 14-ish inferred size to 50 (dummy features) (ablation) | S-IBP with deep decoder · MNIST | test IWAE with first k features: IWAE maxes out right after non-collapsed features; extra features add nothing | No clear effect |
Closest papers by meaning
- Variational Inference for Sequential Data with Future Likelihood Estimates (ICML 2020 · similarity 0.88)
- Efficient Online Variational Estimation via Monte Carlo Sampling (ICML 2026 · similarity 0.88)
- Bayesian Online Natural Gradient (BONG) (NeurIPS 2024 · similarity 0.88)
- Variational Inference with Mixtures of Isotropic Gaussians (NeurIPS 2025 · similarity 0.87)
- EigenVI: score-based variational inference with orthogonal function expansions (NeurIPS 2024 · similarity 0.86)
- Theoretical Guarantees for Variational Inference with Fixed-Variance Mixture of Gaussians (ICML 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.