Research map / Probabilistic and scientific modeling / Variational and generative sampling
Cold Analysis of Rao-Blackwellized 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 |
|---|---|---|---|
| Replace GR-MC with its zero-temperature limit ZGR (average of ST and DARN) | Gumbel-Rao estimator in nonlinear VAE · MNIST-B and Omniglot VAE training | negative ELBO | No clear effect |
| Use ST alone instead of GR/ZGR | nonlinear VAE · MNIST-B C2 V192 and C16 V48 | negative ELBO: ~17-24 nats worse than ZGR | Hurt |
| Use REINFORCE RF(M=4) instead of ZGR | nonlinear VAE · MNIST-B C64 V32 | negative ELBO: 87.2 vs 81.9 | Hurt |
| Use ZGR instead of GS-ST variants for quantized network | quantized FMNIST network · W 2/A 2 test error | test error %: 8.2 vs GS-ST(t=1) 8.0 and ST 7.9 | No clear effect |
| Use ST alone (lowest variance) instead of ZGR in quantized network | quantized FMNIST network · W 2/A 2 test error | test error %: 7.9 vs 8.2 | No clear effect |
| Use REINFORCE RF(M=4) instead of ZGR in quantized network | quantized FMNIST network · W 2/A 2 test error | test error %: 22.4 vs 8.2 | Hurt |
| Use ZGR instead of GR-MC(t=0.1) for hierarchical VAE | hierarchical VAE · MNIST-B C16 V48 | negative ELBO: 115.6 vs 117.8 | Helped |
| Interpolate ST and ZGR (alpha mix) instead of GR-MC | nonlinear VAE · bias-variance tradeoff, C16 V16 epoch 10 | bias and variance | No clear effect |
| Increase GR-MC MC samples from 10 to 100 at t=0.1 (ablation) | Gumbel-Rao in nonlinear VAE · MNIST-B C16 V48 | negative ELBO: 81.0 to 80.5 | Helped |
Closest papers by meaning
- Rao-Blackwellised Reparameterisation Gradients (NeurIPS 2025 · similarity 0.85)
- Density-Softmax: Efficient Test-time Model for Uncertainty Estimation and Robustness under Distribution Shifts (ICML 2024 · similarity 0.85)
- On the Optimal Construction of Unbiased Gradient Estimators for Zeroth-Order Optimization (NeurIPS 2025 · similarity 0.85)
- Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations (ICLR 2025 · similarity 0.84)
- Flat Minima in Linear Estimation and an Extended Gauss Markov Theorem (ICLR 2024 · similarity 0.84)
- Stochastic Zeroth-Order Optimization Under Heavy-Tailed Noise (NeurIPS 2026 · 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.