Research map / Probabilistic and scientific modeling / Variational and generative sampling
GumBolt: Extending Gumbel trick to Boltzmann priors
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 GumBolt instead of dVAE++ with k=1 training samples | VAE with BM prior · MNIST, linear | test-set log-likelihood: 90.40 -> 88.88 | Helped |
| Use GumBolt instead of dVAE++ with k=1 training samples | VAE with BM prior · OMNIGLOT, nonlinear | test-set log-likelihood: 101.97 -> 101.61 | Helped |
| Increase training importance samples from k=1 to k=20 | GumBolt · MNIST, linear | test-set log-likelihood: 88.88 -> 87.45 | Helped |
| Increase training importance samples from k=1 to k=20 | GumBolt · OMNIGLOT, two nonlinear hierarchies | test-set log-likelihood: 99.82 -> 98.81 | Helped |
| Add BM coupling weights rather than setting W to zero (ablation) | GumBolt · MNIST, linear | test-set log-likelihood: 99.94 -> 87.45 | Helped |
| Add BM coupling weights rather than setting W to zero (ablation) | GumBolt · OMNIGLOT, nonlinear | test-set log-likelihood: 107.221 -> 100.68 | Helped |
| Increase decoder network width and depth (ablation) | GumBolt-nW · MNIST | test-set log-likelihood | No clear effect |
| Use 2 PCD steps instead of the 200-step setup (ablation) | GumBolt · MNIST | test-set log-likelihood: 0.35 nats | Hurt |
| Use 20 PCD steps instead of the 200-step setup (ablation) | GumBolt · MNIST | test-set log-likelihood: 0.12 nats | Hurt |
| Vary GumBolt temperature among 1/9, 1/8, and 1/7 (ablation) | GumBolt · MNIST and OMNIGLOT | test-set log-likelihood | No clear effect |
Closest papers by meaning
- DVAE#: Discrete Variational Autoencoders with Relaxed Boltzmann Priors (NeurIPS 2018 · similarity 0.89)
- DVAE++: Discrete Variational Autoencoders with Overlapping Transformations (ICML 2018 · similarity 0.87)
- Undirected Graphical Models as Approximate Posteriors (ICML 2020 · similarity 0.86)
- Shape your Space: A Gaussian Mixture Regularization Approach to Deterministic Autoencoders (NeurIPS 2021 · similarity 0.86)
- Generative Marginalization Models (ICML 2024 · similarity 0.85)
- Learning with Restricted Boltzmann Machines: Asymptotics of AMP and GD in High Dimensions (NeurIPS 2025 · 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.