Research map / Label-efficient and robust learning / Semi-supervised learning
PixelGAN Autoencoders
Alireza Makhzani, Brendan J. Frey
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 PixelGAN Autoencoders instead of the strongest listed baseline | MNIST unsupervised clustering · MNIST (Unsupervised) | error-rate: 4.10\;(\pm 1.13) -> 5.27\;(\pm 1.81) | Hurt |
| Use PixelGAN Autoencoders instead of the strongest listed baseline | MNIST semi-supervised classification · MNIST (20 labels) | error-rate: 16.77\;(\pm 4.52) -> 12.08\;(\pm 5.50) | Improved |
| Use PixelGAN Autoencoders instead of the strongest listed baseline | MNIST semi-supervised classification · MNIST (50 labels) | error-rate: 2.21\;(\pm 1.36) -> 1.16\;(\pm 0.17) | Improved |
| Use PixelGAN Autoencoders instead of the strongest listed baseline | MNIST semi-supervised classification · MNIST (100 labels) | error-rate: 0.89\;(\pm 0.50) -> 1.08\;(\pm 0.15) | Hurt |
| Use PixelGAN Autoencoders instead of the strongest listed baseline | SVHN semi-supervised classification · SVHN (500 labels) | error-rate: 7.05\;(\pm 0.30) -> 10.47\;(\pm 1.80) | Hurt |
| Use PixelGAN Autoencoders instead of the strongest listed baseline | SVHN semi-supervised classification · SVHN (1000 labels) | error-rate: 5.43\;(\pm 0.25) -> 6.96\;(\pm 0.55) | Hurt |
| Use PixelGAN Autoencoders instead of the strongest listed baseline | NORB semi-supervised classification · NORB (1000 labels) | error-rate: 9.40\;(\pm 0.04) -> 8.90\;(\pm 1.0) | Improved |
| Impose a categorical prior through adversarial training (ablation) | PixelGAN Autoencoder · MNIST first three digits, unsupervised | test error-rate: 0.3\% | Improved |
| Add a categorical prior and adversarial matching to the latent code (ablation) | PixelGAN Autoencoder · MNIST style and content representation | representation disentanglement | Improved |
| Use a deeper PixelCNN decoder (ablation) | PixelGAN Autoencoder with a 2-dimensional code · MNIST latent representation | digit-cluster separation | Hurt |
| Use location-invariant rather than location-dependent conditioning bias (ablation) | PixelGAN Autoencoder · MNIST latent representation | class-label encoding | Improved |
| Use additive Gaussian input noise rather than a deterministic posterior (ablation) | PixelGAN Autoencoder semi-supervised experiments · MNIST, SVHN, and NORB | training stability and mode coverage | Improved |
Closest papers by meaning
- Adversarial Latent Autoencoders (CVPR 2020 · similarity 0.87)
- Deep Automodulators (NeurIPS 2020 · similarity 0.86)
- Regularizing Discriminative Capability of CGANs for Semi-Supervised Generative Learning (CVPR 2020 · similarity 0.86)
- A Primal-Dual link between GANs and Autoencoders (NeurIPS 2019 · similarity 0.86)
- Improving Generative Adversarial Networks With Local Coordinate Coding (TPAMI 2020 · similarity 0.85)
- MCGAN: Enhancing GAN Training with Regression-Based Generator Loss (AAAI 2025 · similarity 0.84)
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.
