Research map / Probabilistic and scientific modeling / Variational and generative sampling
Improved Variational Inference with Inverse Autoregressive Flow
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 diagonal Gaussian posterior with IAF (Depth 2, Width 320) (ablation) | convolutional VAE · dynamic MNIST | log p(x): -81.08 -> -79.77 | Helped |
| widen IAF from 320 to 1920 at depth 2 (ablation) | IAF VAE · dynamic MNIST | log p(x): -79.77 -> -79.30 | Helped |
| increase IAF depth from 2 to 4 at width 1920 (ablation) | IAF VAE · dynamic MNIST | log p(x): -79.30 -> -79.17 | Helped |
| increase IAF depth from 4 to 8 at width 1920 (ablation) | IAF VAE · dynamic MNIST | log p(x): -79.17 -> -79.10 | Helped |
| replace diagonal posterior with IAF (Depth 8, Width 1920) (ablation) | IAF VAE · dynamic MNIST | VLB: -84.08 -> -80.80 | Helped |
| more expressive IAF posterior shrinks the gap between VLB and log p(x) | IAF VAE · dynamic MNIST | VLB-to-log-likelihood gap: -84.08 -> -80.80 | Helped |
| use ResNet VAE with IAF instead of Convolutional DRAW | ResNet VAE · CIFAR-10 | bits/dim: 3.58 -> 3.11 | Helped |
| use ResNet VAE with IAF versus PixelRNN | ResNet VAE · CIFAR-10 | bits/dim: 3.00 -> 3.11 | Hurt |
| sample from ResNet VAE instead of PixelCNN | ResNet VAE with IAF · CIFAR-10 | sampling seconds/image: 52.0 -> 0.05 | Helped |
| use VAE with deep IAF on statically binarized MNIST versus PixelCNN | IAF VAE · static MNIST | log-likelihood: -79.88 vs -79.2 | Hurt |
Closest papers by meaning
- Composing Normalizing Flows for Inverse Problems (ICML 2021 · similarity 0.89)
- Model-Informed Flows for Bayesian Inference (NeurIPS 2025 · similarity 0.88)
- Automatic variational inference with cascading flows (ICML 2021 · similarity 0.88)
- Bidirectional Normalizing Flow: From Data to Noise and Back (CVPR 2026 · similarity 0.88)
- Multilayer Matrix Factorization via Dimension-Reducing Diffusion Variational Inference (ICML 2025 · similarity 0.87)
- Normalizing Flows are Capable Generative Models (ICML 2025 · 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.