Research map / Probabilistic and scientific modeling / Variational and generative sampling
Ladder 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 |
|---|---|---|---|
| replace bottom-up VAE inference with ladder (top-down precision-weighted) inference, same generative model | 5-layer latent hierarchy · MNIST test | log-likelihood: best VAE −82.74 vs LVAE −82.12 at 5000 IS | Helped |
| add batch-normalization to VAE models (ablation) | VAE · MNIST test | log-likelihood: Figure 3 comparison; BN improved all models | Helped |
| add deterministic warm-up (linearly increasing KL weight over 200 epochs) (ablation) | VAE with BN · OMNIGLOT test | log-likelihood: e.g. 64-32-16 VAE+BN −106.09 to +WU −102.52 | Helped |
| add deterministic warm-up to VAE without BN | plain VAE · OMNIGLOT test | log-likelihood: plain VAE 64-32-16-8-4 −110.46 vs VAE +BN −105.45 (warm-up not isolated) | No clear effect |
| ladder inference on 5-layer model vs 2-layer VAE with normalizing flows / VGP | MNIST · MNIST test (1 IS) | log-likelihood: LVAE −82.12 vs VGP −81.90 and IWAE IW=50 −82.90 | Mixed |
| ladder inference on NORB Gaussian observation model | NORB · NORB test | log-likelihood: LVAE 3455 vs VAE+BN+WU 3422 (5 layers) | Helped |
| increasing stochastic depth beyond three layers with ladder inference on NORB | LVAE on NORB · NORB test | log-likelihood: 64-32-16 3519 vs 64-32-16-8-4 3455 | No clear effect |
| ladder inference with 5 layers vs more latent variables in IWAE | OMNIGLOT · OMNIGLOT test | log-likelihood: −102.11 vs IWAE best −103.38 (with more latents) | Helped |
| increase depth from 1 to 5 stochastic layers in LVAE (ablation) | LVAE on MNIST · MNIST test (1 IS) | log-likelihood: LVAE reaches −85.23 with five layers; best VAE −87.49 at three layers | Helped |
Closest papers by meaning
- Variational Autoencoders with Jointly Optimized Latent Dependency Structure (ICLR 2019 · similarity 0.91)
- Deep Attentive Variational Inference (ICLR 2022 · similarity 0.91)
- InfoVAE: Balancing Learning and Inference in Variational Autoencoders (AAAI 2019 · similarity 0.91)
- Variational Laplace Autoencoders (ICML 2019 · similarity 0.91)
- Inference Suboptimality in Variational Autoencoders (ICML 2018 · similarity 0.90)
- Beyond Vanilla Variational Autoencoders: Detecting Posterior Collapse in Conditional and Hierarchical Variational Autoencoders (ICLR 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.