Research map / Probabilistic and scientific modeling / Variational and generative sampling
Gradient Estimation with Stochastic Softmax Tricks
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 a Spanning Tree SST instead of independent directed edges for the NRI encoder | NRI · NRI graph layout T=10 | test ELBO: -1370 -> -1080 | Helped |
| use Spanning Tree instead of independent directed edges | NRI · NRI graph layout T=10 | edge precision (%): 48 -> 91 | Helped |
| use Spanning Tree instead of independent edges at longer series | NRI · NRI graph layout T=20 | edge precision (%): 97 -> 99 | No clear effect |
| use tree-structured (Spanning Tree) relaxation instead of independent undirected edges | ListOps GNN · ListOps | task accuracy (%): 89.4 -> 91.2 | Helped |
| use Arborescence (Gumbel) instead of independent directed edges | ListOps GNN · ListOps | task accuracy (%): 90.1 -> 95.0 | Helped |
| use Gumbel or Gaussian utilities instead of negative exponential for Arborescence (ablation) | Arborescence · ListOps | task accuracy (%): 71.5 -> 95.0 | Helped |
| use Gumbel instead of Gaussian utilities for Arborescence (ablation) | Arborescence · ListOps | edge precision (%): 65.3 -> 75.5 | Helped |
| use Corr. Top k instead of L2X relaxation | L2X aroma simple model · BeerAdvocate aroma k=5 | MSE (x10^-2): 3.6 -> 2.9 | Helped |
| use Corr. Top k instead of L2X relaxation | L2X aroma simple model · BeerAdvocate aroma k=5 | subset precision (%): 28.3 -> 63.1 | Helped |
| add correlation structure to top-k subset distribution (Corr. Top k vs E.F. Ent. Top k) (ablation) | L2X complex model · BeerAdvocate aroma k=15 | subset precision (%): 44.9 -> 57.7 | Helped |
| vary relaxation regularizer f (Euclid, categorical, binary, E.F. entropy) (ablation) | Top k SST · BeerAdvocate aroma | MSE | No clear effect |
| use generic Top k SSTs instead of bespoke L2X/SoftSub relaxations | L2X aroma simple model · BeerAdvocate aroma k=10 | subset precision (%): 25.5 -> 32.9 | Helped |
Closest papers by meaning
- Leveraging Recursive Gumbel-Max Trick for Approximate Inference in Combinatorial Spaces (NeurIPS 2021 · similarity 0.90)
- Rao-Blackwellizing the Straight-Through Gumbel-Softmax Gradient Estimator (ICLR 2021 · similarity 0.89)
- Marginalized Stochastic Natural Gradients for Black-Box Variational Inference (ICML 2021 · similarity 0.88)
- Categorical Reparameterization with Gumbel-Softmax (ICLR 2017 · similarity 0.88)
- Cold Analysis of Rao-Blackwellized Straight-Through Gumbel-Softmax Gradient Estimator (ICML 2023 · similarity 0.88)
- Policy Gradient with Tree Expansion (ICML 2025 · similarity 0.85)
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.