Research map / Probabilistic and scientific modeling / Variational and generative sampling
DVAE#: Discrete Variational Autoencoders with Relaxed 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 power-function rather than exponential smoothing (ablation) | DVAE++ · MNIST, 2 groups nonlinear, K=25 | negative log-likelihood: 84.85{ $\pm$0.13} -> 83.58{ $\pm$0.15} | Helped |
| Use power-function rather than exponential smoothing (ablation) | DVAE++ · OMNIGLOT, 2 groups nonlinear, K=25 | negative log-likelihood: 100.31{ $\pm$0.15} -> 99.49{ $\pm$0.07} | Helped |
| Use uniform+exponential rather than exponential smoothing (ablation) | DVAE# · MNIST, 2 groups nonlinear, K=25 | negative log-likelihood: 83.76{ $\pm$0.08} -> 83.30{ $\pm$0.04} | Helped |
| Use uniform+exponential rather than exponential smoothing (ablation) | DVAE# · OMNIGLOT, 2 groups nonlinear, K=25 | negative log-likelihood: 100.10{ $\pm$0.28} -> 99.59{ $\pm$0.16} | Helped |
| Use power-function rather than exponential smoothing (ablation) | DVAE# · MNIST, 2 groups nonlinear, K=25 | negative log-likelihood: 83.76{ $\pm$0.08} -> 82.85{ $\pm$0.03} | Helped |
| Use power-function rather than exponential smoothing (ablation) | DVAE# · OMNIGLOT, 2 groups nonlinear, K=25 | negative log-likelihood: 100.10{ $\pm$0.28} -> 98.93{ $\pm$0.05} | Helped |
| Train with K=25 rather than K=1 importance samples (ablation) | DVAE# Power · MNIST, 2 groups nonlinear | negative log-likelihood: 83.37{ $\pm$0.02} at K=1; 82.85{ $\pm$0.03} at K=25 | Helped |
| Train with K=25 rather than K=1 importance samples (ablation) | DVAE# Power · OMNIGLOT, 2 groups nonlinear | negative log-likelihood: 99.75{ $\pm$0.05} at K=1; 98.93{ $\pm$0.05} at K=25 | Helped |
| Use Gaussian integral relaxation rather than power-function overlap relaxation (ablation) | DVAE# · MNIST, 2 groups nonlinear, K=25 | negative log-likelihood: 82.85{ $\pm$0.03} -> 83.39{ $\pm$0.04} | Hurt |
| Use Gaussian integral relaxation rather than power-function overlap relaxation (ablation) | DVAE# · OMNIGLOT, 2 groups nonlinear, K=25 | negative log-likelihood: 98.93{ $\pm$0.05} -> 100.20{ $\pm$0.02} | Hurt |
| Use population annealing rather than persistent contrastive divergence (ablation) | DVAE# Power · MNIST, 4 groups nonlinear, K=25 | negative log-likelihood: 83.18{ $\pm$0.05} -> 82.82{ $\pm$0.02} | Helped |
Closest papers by meaning
- DVAE++: Discrete Variational Autoencoders with Overlapping Transformations (ICML 2018 · similarity 0.94)
- GumBolt: Extending Gumbel trick to Boltzmann priors (NeurIPS 2018 · similarity 0.89)
- Direct Optimization through $\arg \max$ for Discrete Variational Auto-Encoder (NeurIPS 2019 · similarity 0.89)
- Discrete Variational Autoencoding via Policy Search (ICLR 2026 · similarity 0.89)
- Statistical Guarantees for Variational Autoencoders using PAC-Bayesian Theory (NeurIPS 2023 · similarity 0.88)
- VAE Learning via Stein Variational Gradient Descent (NeurIPS 2017 · similarity 0.88)
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.