Research map / Probabilistic and scientific modeling / Variational and generative sampling
Inference Suboptimality in 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 q_AF instead of q_FFG | VAE · Fashion-MNIST | inference gap (nats): 7.28 -> 6.35 | Helped |
| use q_AF instead of q_FFG | VAE · 3-BIT CIFAR | inference gap (nats): 52.22 -> 43.72 | Helped |
| use q_AF instead of q_FFG | VAE · MNIST | inference gap (nats): 2.77 -> 2.85 | Hurt |
| use q_AF instead of q_FFG | VAE · Fashion-MNIST and 3-BIT CIFAR | approximation gap (nats): 3.06 -> 1.69; 14.75 -> 1.60 | Helped |
| use q_AF instead of q_FFG | VAE · MNIST and 3-BIT CIFAR | amortization gap (nats): 1.34 -> 1.41; 37.47 -> 42.12 | Mixed |
| increase encoder capacity (ablation) | VAE with q_FFG or q_AF · MNIST and Fashion-MNIST | amortization gap: MNIST q_FFG: 1.49 -> 1.11; q_AF: 1.30 -> 0.75; Fashion-MNIST q_FFG: 3.05 -> 3.09; q_AF: 2.29 -> 3.76 | Helped |
| replace q_FFG with q_AF (ablation) | small encoder with fixed VAE decoder · MNIST | amortization gap (nats): 43.22 -> 13.10 | Helped |
| replace q_FFG with q_Flow (ablation) | small encoder with fixed VAE decoder · MNIST | amortization gap (nats): 43.22 -> 12.86 | Helped |
| increase generator capacity from 0 to 4 hidden layers (ablation) | VAE with q_FFG · MNIST | approximation gap (nats): 3.90 -> 1.63 | Helped |
| remove entropy annealing (ablation) | VAE with q_FFG and q_AF · MNIST and Fashion-MNIST | difference between optimal FFG and AF bounds: MNIST q_AF: -108.71 vs -92.25; Fashion-MNIST q_AF: -129.70 vs -105.80 | Hurt |
| use a larger encoder instead of the flow model (ablation) | VAE inference on held-out Fashion-MNIST · Fashion-MNIST validation set | amortized ELBO | Hurt |
| increase decoder capacity (ablation) | VAE · Fashion-MNIST training and validation | amortization gap | Mixed |
Closest papers by meaning
- Amortized Inference Regularization (NeurIPS 2018 · similarity 0.92)
- Iterative Amortized Inference (ICML 2018 · similarity 0.92)
- Meta-Amortized Variational Inference and Learning (AAAI 2020 · similarity 0.92)
- Gaussian Process Modeling of Approximate Inference Errors for Variational Autoencoders (CVPR 2022 · similarity 0.92)
- InfoVAE: Balancing Learning and Inference in Variational Autoencoders (AAAI 2019 · similarity 0.91)
- Semi-Amortized Variational Autoencoders (ICML 2018 · similarity 0.91)
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.