Research map / Probabilistic and scientific modeling / Variational and generative sampling
Invertible Gaussian Reparameterization: Revisiting the Gumbel-Softmax
Andres Potapczynski, Gabriel Loaiza-Ganem, John P. Cunningham
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 Gumbel-Softmax with IGR-SB, linear VAE | VAE · MNIST | test log-likelihood: -106.17 -> -96.74 | Improved |
| replace Gumbel-Softmax with IGR-I, linear VAE | VAE · Omniglot | test log-likelihood: -138.98 -> -128.14 | Improved |
| replace Gumbel-Softmax with IGR-Planar, non-linear VAE | VAE · MNIST | test log-likelihood: -98.06 -> -92.91 | Improved |
| replace Gumbel-Softmax with IGR, non-linear VAE | VAE · FMNIST | test log-likelihood: -29.72 -> -34.10 | Hurt |
| add sticking-the-landing estimator (no closed-form KL) to IGR-Planar, non-linear (ablation) | VAE · MNIST | test log-likelihood: -92.91 -> -88.81 | Improved |
| add sticking-the-landing estimator to Gumbel-Softmax, linear (ablation) | VAE · MNIST | test log-likelihood: -106.17 -> -103.80 | Improved |
| add sticking-the-landing estimator to IGR-I, linear (ablation) | VAE · MNIST | test log-likelihood: -94.65 -> -94.18 | Improved |
| add sticking-the-landing estimator to IGR-SB, non-linear (ablation) | VAE · Omniglot | test log-likelihood: -139.82 -> -135.82 | Improved |
| replace Gumbel-Softmax with IGR-SB in structured output prediction | binary stochastic network · MNIST | test log-likelihood: -59.31 -> -45.12 | Improved |
| replace Dirichlet process mixture VAE truncation K (DLDPMM) with IGR-SB-MM | VAE with DP mixture prior · MNIST | test log-likelihood: -81.90 -> -81.79 | No clear effect |
| replace DLDPMM with IGR-SB-MM | VAE with DP mixture prior · Omniglot | test log-likelihood: -128.72 -> -129.06 | No clear effect |
| use IGR-SB instead of GS with fixed K to approximate Poisson (lambda 50) | discrete approximation · Poisson distribution | approximation quality | Improved |
Closest papers by meaning
- Rao-Blackwellised Reparameterisation Gradients (NeurIPS 2025 · similarity 0.84)
- Rethinking Approximate Gaussian Inference in Classification (NeurIPS 2025 · similarity 0.84)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.83)
- How Accurately Can a Gaussian Approximate Stochastic Approximation Iterates? (NeurIPS 2026 · similarity 0.83)
- Stochastic variance-reduced Gaussian variational inference on the Bures-Wasserstein manifold (ICLR 2025 · similarity 0.83)
- Is Temperature Sample Efficient for Softmax Gaussian Mixture of Experts? (ICML 2024 · similarity 0.83)
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.
