Research map / Probabilistic and scientific modeling / Variational and generative sampling
Discrete Variational Autoencoders
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 a discrete latent RBM with continuous latent layers and smoothing | Discrete VAE · MNIST (dynamic binarization) | log-likelihood: -81.74 -> -80.15 | Helped |
| Use a discrete latent RBM with continuous latent layers and smoothing | Discrete VAE · MNIST (static binarization) | log-likelihood: -81.32 -> -81.01 | Helped |
| Use a discrete latent RBM with continuous latent layers and smoothing | Discrete VAE · Omniglot | log-likelihood: -100.45 -> -97.43 | Helped |
| Use a discrete latent RBM with continuous latent layers and smoothing | Discrete VAE · Caltech-101 Silhouettes | log-likelihood: -100.0 -> -97.6 | Helped |
| Remove the RBM (ablation) | Discrete VAE · MNIST (dynamic and static binarization), Omniglot, Caltech-101 Silhouettes | test log-likelihood: 0.09, 0.37, 0.69, and 0.66 | Hurt |
| Remove continuous latent layers (ablation) | Discrete VAE · MNIST (static binarization) | log-likelihood: -81.01 -> -86.9 (128 RBM units); -81.01 -> -85.2 (200 RBM units) | Hurt |
| Restrict the decoder to a linear transformation followed by a logistic nonlinearity (ablation) | Discrete VAE without continuous latent layers · MNIST (static binarization) | log-likelihood: -86.9 -> -92.7 (128 RBM units); -85.2 -> -88.8 (200 RBM units) | Hurt |
| Remove RBM lateral connections (ablation) | Discrete VAE · MNIST (static binarization) | log-likelihood: -81.01 -> -97.0 | Hurt |
| Replace the hierarchical approximating posterior with a factorial posterior (ablation) | Discrete VAE · MNIST (static binarization) | log-likelihood: -97.0 -> -102.9 | Hurt |
| Increase the number of block Gibbs iterations per minibatch (ablation) | Discrete VAE · MNIST (dynamic binarization) | log-likelihood | Helped |
| Increase the RBM size beyond 64 units (ablation) | Discrete VAE · MNIST (dynamic binarization) | log-likelihood | No clear effect |
| Increase the hierarchical approximating posterior from 4 to 8 layers (ablation) | Discrete VAE · MNIST (dynamic binarization) | log-likelihood | Hurt |
Closest papers by meaning
- DVAE++: Discrete Variational Autoencoders with Overlapping Transformations (ICML 2018 · similarity 0.89)
- Shape your Space: A Gaussian Mixture Regularization Approach to Deterministic Autoencoders (NeurIPS 2021 · similarity 0.89)
- Variational Autoencoder with Arbitrary Conditioning (ICLR 2019 · similarity 0.88)
- Variational Autoencoders with Jointly Optimized Latent Dependency Structure (ICLR 2019 · similarity 0.88)
- Continuity-Preserving Convolutional Autoencoders for Learning Continuous Latent Dynamical Models from Images (ICLR 2025 · similarity 0.87)
- A Statistical Analysis of Wasserstein Autoencoders for Intrinsically Low-dimensional Data (ICLR 2024 · 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.