Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learning Multi-layer Latent Variable Model via Variational Optimization of Short Run MCMC for Approximate Inference
Erik Nijkamp, Bo Pang, Tian Han, Linqi Zhou, Song-Chun Zhu, Ying Nian Wu
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 |
|---|---|---|---|
| add latent layers (L=1 to L=5) (ablation) | short run MCMC inference · CelebA (64x64), K=25 | FID: 61.03 -> 47.95 | Improved |
| add latent layers (L=1 to L=5) (ablation) | short run MCMC inference · CelebA (64x64), K=25 | reconstruction MSE: 0.020 -> 0.015 | Improved |
| increase inference steps from K=5 to K=25 (ablation) | short run MCMC inference · CelebA (32x32), L=5 | FID: 82.79 -> 36.84 | Improved |
| increase inference steps from K=25 to K=400 (ablation) | short run MCMC inference · CelebA (32x32), L=5 | FID: 36.84 -> 35.16 | No clear effect |
| increase inference steps from K=5 to K=25 (ablation) | short run MCMC inference · CelebA (32x32), L=5 | reconstruction MSE: 0.045 -> 0.011 | Improved |
| use short run MCMC inference instead of Ladder-VAE inference network (L=5) | multi-layer generator · CelebA (32x32) | FID: 53.40 -> 36.84 | Improved |
| use short run MCMC inference instead of Ladder-VAE inference network (L=5) | multi-layer generator · SVHN (32x32) | FID: 39.26 -> 35.23 | Improved |
| use short run MCMC inference instead of Ladder-VAE inference network (L=5) | multi-layer generator · SVHN (32x32) | reconstruction MSE: 0.014 -> 0.011 | Improved |
| use short run MCMC inference instead of Ladder-VAE (L=1) | single-layer generator · SVHN (32x32) | reconstruction MSE: 0.019 -> 0.018 | No clear effect |
| use short run MCMC inference instead of VAE (L=1) | single-layer generator · CelebA (32x32), L=1 | FID: 69.90 -> 45.74 | Improved |
| use short run MCMC inference instead of ABP (L=1) | single-layer generator · SVHN (32x32), L=1 | FID: 49.71 -> 44.86 | Improved |
| optimize step size by gradient descent vs grid search (ablation) | short run MCMC step size · SVHN | step size: 0.05 | No clear effect |
Closest papers by meaning
- Learning Deep Latent Variable Models by Short-Run MCMC Inference With Optimal Transport Correction (CVPR 2021 · similarity 0.95)
- Langevin Autoencoders for Learning Deep Latent Variable Models (NeurIPS 2022 · similarity 0.89)
- Deep Attentive Variational Inference (ICLR 2022 · similarity 0.88)
- Learning Energy-Based Model with Variational Auto-Encoder as Amortized Sampler (AAAI 2021 · similarity 0.87)
- Online Variational Sequential Monte Carlo (ICML 2024 · similarity 0.87)
- Bayesian Basis Function Approximation for Scalable Gaussian Process Priors in Deep Generative Models (ICML 2025 · similarity 0.86)
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.
