Research map / Probabilistic and scientific modeling / Variational and generative sampling
Implicit Reparameterization Gradients
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 implicit gradients instead of RSVI | LDA · RCV1 | test perplexity: 1505 ± 33 -> 896 ± 6 (RSVI B=1); 907 ± 13 -> 896 ± 6 (RSVI B=20) | Helped |
| Use implicit gradients instead of RSVI | LDA · 20 Newsgroups | test perplexity: 1066 ± 7 -> 876 ± 7 (RSVI B=1); 865 ± 11 -> 876 ± 7 (RSVI B=20) | Mixed |
| Increase RSVI shape augmentation from B=1 to B=20 (ablation) | LDA · 20 Newsgroups | test perplexity: 1066 ± 7 -> 865 ± 11 | Helped |
| Increase RSVI shape augmentation from B=1 to B=20 (ablation) | LDA · RCV1 | test perplexity: 1505 ± 33 -> 907 ± 13 | Helped |
| Use implicit gradients instead of RSVI | VAE with von Mises posterior and uniform prior · MNIST, D=40 | test negative log-likelihood: 92.8 ± 0.2 -> 91.5 ± 0.4 | Helped |
| Use implicit gradients instead of RSVI B=20 | VAE with Gamma posterior · MNIST, D=40, Gamma(0.3, 0.3) prior | test negative log-likelihood: 90.6 ± 0.1 -> 90.6 ± 0.2 | No clear effect |
| Use automatic differentiation rather than finite differences | Gamma CDF gradient computation · Gamma, float32 | mean absolute error: \num{1.9e-03} -> \num{2.3e-06} | Helped |
| Use automatic differentiation rather than finite differences | von Mises CDF gradient computation · von Mises, float32 | mean absolute error: \num{9.6e-05} -> \num{1.9e-07} | Helped |
| Use automatic differentiation rather than finite differences | Gamma CDF gradient computation · Gamma, float32 | time (s): \num{3.8e-08} -> \num{1.9e-08} | Helped |
| Use automatic differentiation rather than finite differences | von Mises CDF gradient computation · von Mises, float32 | time (s): \num{3.8e-08} -> \num{3.1e-08} | No clear effect |
| Add dropout and batch normalization to the LDA architecture (ablation) | LDA · 20 Newsgroups | test perplexity: 1059 -> 875 | No clear effect |
| Fix the prior distribution instead of learning prior parameters (ablation) | LDA with implicit gradients · 20 Newsgroups | test perplexity: +60 points | Hurt |
Closest papers by meaning
- The Generalized Reparameterization Gradient (NeurIPS 2016 · similarity 0.91)
- Reparameterization Gradient for Non-differentiable Models (NeurIPS 2018 · similarity 0.90)
- Categorical Reparameterization with Denoising Diffusion Models (ICML 2026 · similarity 0.89)
- Reducing Reparameterization Gradient Variance (NeurIPS 2017 · similarity 0.89)
- Rao-Blackwellised Reparameterisation Gradients (NeurIPS 2025 · similarity 0.87)
- Reparameterization invariance in approximate Bayesian inference (NeurIPS 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.