Research map / Probabilistic and scientific modeling / Variational and generative sampling
Rao-Blackwellizing the Straight-Through Gumbel-Softmax Gradient Estimator
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 |
|---|---|---|---|
| Rao-Blackwellize ST-GS by averaging conditional Gumbel-gradient samples | ST-GS · ListOps | test accuracy | Helped |
| Increase conditional Monte Carlo samples from GR-MC10 to GR-MC1000 (ablation) | Gumbel-Rao on ListOps · ListOps, especially low temperatures | test accuracy: 38.8; 66.5 | Helped |
| Increase conditional Monte Carlo samples from K=10 to K=100 or K=1000 (ablation) | Gumbel-Rao on ListOps · ListOps | test accuracy | Mixed |
| Rao-Blackwellize ST-GS | ST-GS on MNIST discrete VAE · MNIST discrete VAE, small batch sizes | test negative log-likelihood: 98.8; 96.8 | Helped |
| Increase conditional Monte Carlo samples K (ablation) | Gumbel-Rao on MNIST discrete VAE · MNIST discrete VAE | test negative log-likelihood: 100.7; 100.5 | Mixed |
| Use GR-MC1000 instead of the best straight-through estimator | Discrete VAE · MNIST, binary latent variables | test negative log-likelihood: 100.5; 98.2 | Hurt |
| Use GR-MC1000 instead of GS | Discrete VAE · MNIST, 16-ary latent variables | test negative log-likelihood: 96.8; 96.8 | Helped |
| Increase conditional sample count K (ablation) | Gumbel-Rao · Gradient estimation | mean squared error | Helped |
Closest papers by meaning
- Cold Analysis of Rao-Blackwellized Straight-Through Gumbel-Softmax Gradient Estimator (ICML 2023 · similarity 0.92)
- Gradient Estimation with Stochastic Softmax Tricks (NeurIPS 2020 · similarity 0.89)
- Estimating Gradients for Discrete Random Variables by Sampling without Replacement (ICLR 2020 · similarity 0.89)
- Rao-Blackwellised Reparameterisation Gradients (NeurIPS 2025 · similarity 0.88)
- Coupled Gradient Estimators for Discrete Latent Variables (NeurIPS 2021 · similarity 0.88)
- Categorical Reparameterization with Gumbel-Softmax (ICLR 2017 · 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.