Research map / Probabilistic and scientific modeling / Variational and generative sampling
Langevin Autoencoders for Learning Deep Latent Variable Models
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 datapoint-wise Langevin sampling iterations with encoder-parameter Langevin updates (LAE) | Deep latent variable models · MNIST test | negative ELBO per data dimension: 1.189 -> 1.177 | Helped |
| Replace datapoint-wise Langevin sampling iterations with encoder-parameter Langevin updates (LAE) | Deep latent variable models · SVHN test | negative ELBO per data dimension: 4.442 -> 4.412 | Helped |
| Replace datapoint-wise Langevin sampling iterations with encoder-parameter Langevin updates (LAE) | Deep latent variable models · CIFAR-10 test | negative ELBO per data dimension: 4.820 -> 4.773 | Helped |
| Replace datapoint-wise Langevin sampling iterations with encoder-parameter Langevin updates (LAE) | Deep latent variable models · CelebA test | negative ELBO per data dimension: 4.671 -> 4.636 | Helped |
| Add the Metropolis-Hastings rejection step (ablation) | LAE · MNIST test | negative ELBO learning-curve stability | Helped |
| Increase the number of ALD iterations to at least 2 (ablation) | LAE · MNIST test | negative ELBO | No clear effect |
| Increase encoder final-layer input dimensionality to at least the number of data points (ablation) | ALD · Bivariate Gaussian posterior toy example | posterior sample quality | Helped |
| Use Gaussian variational inference instead of ALD | Posterior inference · Neural likelihood toy example | posterior approximation quality | Hurt |
| Increase encoder-initialized LD iterations from T=0 to T=10 (ablation) | Encoder-initialized LD · MNIST test | negative ELBO per data dimension: 1.189 \pm 0.002 -> 1.188 \pm 0.001 | No clear effect |
| Increase encoder-initialized LD iterations from T=0 to T=10 (ablation) | Encoder-initialized LD · SVHN test | negative ELBO per data dimension: 4.442 \pm 0.003 -> 4.437 \pm 0.009 | Helped |
| Increase encoder-initialized LD iterations from T=0 to T=10 (ablation) | Encoder-initialized LD · CIFAR-10 test | negative ELBO per data dimension: 4.820 \pm 0.005 -> 4.832 \pm 0.006 | Hurt |
| Increase encoder-initialized LD iterations from T=0 to T=10 (ablation) | Encoder-initialized LD · CelebA test | negative ELBO per data dimension: 4.671 \pm 0.001 -> 4.664 \pm 0.004 | Helped |
Closest papers by meaning
- Learning Deep Latent Variable Models by Short-Run MCMC Inference With Optimal Transport Correction (CVPR 2021 · similarity 0.91)
- Learning Energy-Based Model with Variational Auto-Encoder as Amortized Sampler (AAAI 2021 · similarity 0.90)
- Fully Bayesian Autoencoders with Latent Sparse Gaussian Processes (ICML 2023 · similarity 0.89)
- Sample as you Infer: Predictive Coding with Langevin Dynamics (ICML 2024 · similarity 0.89)
- Microcanonical Langevin Ensembles: Advancing the Sampling of Bayesian Neural Networks (ICLR 2025 · similarity 0.88)
- The Poisson Midpoint Method for Langevin Dynamics: Provably Efficient Discretization for Diffusion Models (NeurIPS 2024 · 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.