Research map / Probabilistic and scientific modeling / Variational and generative sampling
Undirected Graphical Models as Approximate Posteriors
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 the directed posterior with an RBM undirected posterior | DVAE · MNIST (VAE, Prior RBM 200+200) | negative log-likelihood: 83.22 -> 82.75 | Helped |
| replace the directed posterior with an RBM undirected posterior | DVAE · OMNIGLOT (VAE, Prior RBM 200+200) | negative log-likelihood: 99.10 -> 98.61 | Helped |
| replace the directed posterior with an RBM undirected posterior | DVAE · MNIST (IWAE, K=25, Prior RBM 100+100) | negative log-likelihood: 83.38 -> 83.44 | Mixed |
| replace the directed posterior with an RBM undirected posterior | DVAE · OMNIGLOT (IWAE, K=25, Prior RBM 100+100) | negative log-likelihood: 100.37 -> 100.02 | Helped |
| add pairwise interactions to the mean-field posterior (ablation) | DVAE## · MNIST and OMNIGLOT (VAE) | negative log-likelihood: ∼0.5-nat improvement | Helped |
| increase relaxed Gibbs steps from 1 to 4 (ablation) | DVAE## · MNIST (VAE, RBM 100+100, s=10) | negative log-likelihood: 84.09{ $\pm$0.03} -> 84.34{ $\pm$0.02} | Hurt |
| increase discrete Gibbs steps from 5 to 10 while keeping t=1 (ablation) | DVAE## · MNIST (VAE, RBM 100+100) | negative log-likelihood: 84.06{ $\pm$0.04} -> 84.09{ $\pm$0.03} | No clear effect |
| increase discrete Gibbs steps from 1 to 5 while keeping t=1 (ablation) | DVAE## · MNIST (VAE, RBM 100+100) | negative log-likelihood: 84.11{ $\pm$0.04} -> 84.06{ $\pm$0.04} | No clear effect |
| include the high-variance I term in the gradient estimator (ablation) | undirected posterior gradient estimator · two-variable Gaussian toy optimization | KL divergence optimization | Hurt |
| increase the number of MCMC updates t (ablation) | proposed gradient estimator · 4+4-bit RBM toy optimization | optimization performance | No clear effect |
| use a shared context feature for directed posterior conditionals (ablation) | directed posterior baseline · MNIST and OMNIGLOT | negative log-likelihood | Mixed |
| use MCMC samples from the true posterior instead of an amortized RBM approximate posterior | DVAE## · MNIST (16-bit latent space) | test negative log-likelihood: 110.9{ $\pm$0.3} -> 113.4{ $\pm$1.5} | Hurt |
Closest papers by meaning
- Lagging Inference Networks and Posterior Collapse in Variational Autoencoders (ICLR 2019 · similarity 0.89)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 2024 · similarity 0.89)
- Neural Variational Inference and Learning in Undirected Graphical Models (NeurIPS 2017 · similarity 0.89)
- VAE Approximation Error: ELBO and Exponential Families (ICLR 2022 · similarity 0.88)
- Variational Autoencoder with Implicit Optimal Priors (AAAI 2019 · similarity 0.88)
- Variational Learning of Fractional Posteriors (ICML 2025 · 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.