Research map / Optimization / Optimizers
VAE Learning via Stein Variational Gradient Descent
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 the parametric VAE encoder training with Stein VAE training | VAE · MNIST semi-supervised classification, FNN and CNN, 10–300 labeled images per class | classification error (%): 3.33 ± 0.14 -> 2.78 ± 0.24; 2.59 ±0.05 -> 2.13 ± 0.08; 2.40 ±0.02 -> 1.92 ± 0.05; 2.18 ±0.04 -> 1.77 ± 0.03; 2.44 ± 0.17 | Helped |
| Add importance-weighted training to Stein VAE | Stein VAE · MNIST NLL | NLL: 85.21 -> 83.01 | Helped |
| Add S-ELBO evaluation instead of ELBO evaluation | Stein VAE · MNIST NLL | NLL: 85.21 -> 84.98 | Helped |
| Add S-ELBO evaluation instead of ELBO evaluation | Stein VIWAE · MNIST NLL | NLL: 83.01 -> 82.88 | Helped |
| Use Stein VIWAE rather than VAE | Semi-supervised classification · ImageNet, 1% labeled | classification accuracy (%): 35.92± 1.91 -> 36.91 ± 0.98 | Helped |
| Use Stein VIWAE rather than VAE | Semi-supervised classification · ImageNet, 2%–40% labeled | classification accuracy (%): 40.15± 1.52 -> 42.57 ± 0.84; 44.27± 1.47 -> 46.20 ± 0.52; 46.92± 1.02 -> 48.67 ± 0.31; 50.43± 0.41 -> 51.77 ± 0.12; 53.2 | Helped |
| Increase the number of parameter samples for model averaging (ablation) | Stein VAE · MNIST | NLL | Helped |
| Use Stein VIWAE rather than Stein VAE | Semi-supervised classification · MNIST, FNN and CNN, 10–300 labeled images per class | classification error (%): 2.78 ± 0.24 -> 2.67 ± 0.09; 2.13 ± 0.08 -> 2.09 ± 0.03; 1.92 ± 0.05 -> 1.88 ± 0.01; 1.77 ± 0.03 -> 1.75 ± 0.01; 1.94 ± 0 | Helped |
| Use Stein VIWAE rather than Stein VAE | Density estimation · MNIST and 20News, NYT, Science, RCV2 | NLL or perplexity: 85.21 -> 83.01; 849 -> 837; 2402 -> 2315; 1499 -> 1453; 549 -> 523 | Helped |
| Use Stein VIWAE rather than Stein VAE | Density estimation · 20News, NYT, Science, RCV2 | perplexity: 845 -> 829; 2401 -> 2277; 1497 -> 1421; 544 -> 518 | Helped |
| Use Stein VIWAE rather than DGDN | Semi-supervised classification · ImageNet, 1%–5% labeled | classification accuracy (%): 43.98± 1.15 vs 36.91 ± 0.98; 46.92± 1.11 vs 42.57 ± 0.84; 47.36± 0.91 vs 46.20 ± 0.52 | Hurt |
| Use Stein VIWAE rather than DGDN | Semi-supervised classification · ImageNet, 10%–40% labeled | classification accuracy (%): 48.41± 0.76 -> 48.67 ± 0.31; 51.51± 0.28 -> 51.77 ± 0.12; 54.14± 0.12 -> 55.45 ± 0.11; 57.34± 0.18 -> 58.21 ± 0.12 | Helped |
Closest papers by meaning
- Consistency Regularization for Variational Auto-Encoders (NeurIPS 2021 · similarity 0.91)
- Variational autoencoders in the presence of low-dimensional data: landscape and implicit bias (ICLR 2022 · similarity 0.90)
- Adversarial Symmetric Variational Autoencoder (NeurIPS 2017 · similarity 0.90)
- Simple and Effective VAE Training with Calibrated Decoders (ICML 2021 · similarity 0.90)
- On the Value of Infinite Gradients in Variational Autoencoder Models (NeurIPS 2021 · similarity 0.89)
- MAE: Mutual Posterior-Divergence Regularization for Variational AutoEncoders (ICLR 2019 · similarity 0.89)
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.