Research map / Probabilistic and scientific modeling / Variational and generative sampling
Generalization Gap in Amortized Inference
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 |
|---|---|---|---|
| Train amortized inference with reverse half-asleep objective instead of standard VAE inference | VAE · Binary MNIST test | BPD: 0.200 -> 0.187 | Helped |
| Train amortized inference with reverse half-asleep objective instead of standard VAE inference | VAE · Grey MNIST test | BPD: 1.543 -> 1.513 | Helped |
| Train amortized inference with reverse half-asleep objective instead of standard VAE inference | VAE · CIFAR10 test | BPD: 4.592 -> 4.572 | Helped |
| Train amortized inference with reverse half-asleep objective | VAE · CIFAR100 test | BPD: 5.288 to 5.275 | Helped |
| Add a denoising regularizer with σ=0.1, σ=0.2, or σ=0.4 | VAE amortized inference · Binary MNIST and Grey MNIST test | BPD: Binary MNIST: 0.200 -> 0.195, 0.192, 0.191; Grey MNIST: 1.543 -> 1.527, 1.519, 1.515 | Helped |
| Increase denoising noise to σ=0.8 or σ=1.0 | VAE amortized inference · Binary MNIST and Grey MNIST test | BPD: Binary MNIST: 0.200 -> 0.196, 0.201; Grey MNIST: 1.543 -> 1.545, 1.550 | Hurt |
| Add a denoising regularizer with σ=0.1 | VAE amortized inference · CIFAR10 test | BPD: 4.592 -> 4.591 | Helped |
| Add a denoising regularizer with σ=0.2, σ=0.4, σ=0.8, or σ=1.0 | VAE amortized inference · CIFAR10 test | BPD: 4.592 -> 4.598, 4.614, 4.651, 4.667 | Hurt |
| Use reverse sleep rather than reverse half-asleep (α=0.5) (ablation) | VAE amortized inference · Binary MNIST test | BPD | Hurt |
| Use standard amortized inference rather than reverse half-asleep (α=0.5) (ablation) | VAE amortized inference · Binary MNIST test | BPD | Hurt |
| Use reverse half-asleep training from the beginning rather than post-hoc (ablation) | VAE · Binary MNIST test | ELBO | Helped |
| Use reverse half-asleep training from the beginning rather than standard VAE training (ablation) | VAE · Grey MNIST test | ELBO | Hurt |
Closest papers by meaning
- Amortized Inference Regularization (NeurIPS 2018 · similarity 0.94)
- Meta-Amortized Variational Inference and Learning (AAAI 2020 · similarity 0.91)
- Inference Suboptimality in Variational Autoencoders (ICML 2018 · similarity 0.91)
- Iterative Amortized Inference (ICML 2018 · similarity 0.90)
- Information-theoretic Generalization Analysis for VQ-VAEs: A Role of Latent Variables (NeurIPS 2025 · similarity 0.88)
- Generalization in VAE and Diffusion Models: A Unified Information-Theoretic Analysis (ICLR 2025 · similarity 0.87)
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.