Research map / Label-efficient and robust learning / Semi-supervised learning
Adversarial Transformations for Semi-Supervised Learning
Teppei Suzuki, Ikuro Sato
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 |
|---|---|---|---|
| replace additive adversarial noise with adversarial, class-preserving transformations | VAT-style consistency regularization · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 11.70$\pm$0.32\% | Improved |
| replace additive adversarial noise with adversarial, class-preserving transformations | VAT-style consistency regularization · SVHN, 1,000 labels | test error rate: 5.32$\pm$0.25\% -> 3.10$\pm$0.12\% | Improved |
| add ε-rampup (ablation) | VAT · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 13.26$\pm$0.20\% | Improved |
| add ε-rampup (ablation) | RAT with affine and noise transformations · CIFAR-10, 4,000 labels | test error rate: 12.68$\pm$0.29\% -> 11.26$\pm$0.34\% | Improved |
| add affine transformation to noise injection (ablation) | RAT · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 11.70$\pm$0.32\% | Improved |
| add affine transformation to noise injection (ablation) | RAT · SVHN, 1,000 labels | test error rate: 5.32$\pm$0.25\% -> 3.10$\pm$0.12\% | Improved |
| add channel-wise weighting to noise injection (ablation) | RAT · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 14.33$\pm$0.44\% | Hurt |
| add flow field to noise injection (ablation) | RAT · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 14.27$\pm$0.41\% | Hurt |
| add thin plate spline to noise injection (ablation) | RAT · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 13.06$\pm$0.44\% | Improved |
| add channel-wise weighting alongside affine transformation (ablation) | RAT · CIFAR-10, 4,000 labels | test error rate: 11.70$\pm$0.32\% -> 11.32$\pm$0.44 | No clear effect |
| increase VAT maximum ε from 6 to 10 without rampup (ablation) | VAT · CIFAR-10, 4,000 labels | test error rate: 13.68$\pm$0.25\% -> 14.31$\pm$0.33\% | Hurt |
| use RAT with ε-rampup rather than random transformations | RAT · SVHN, 1,000 labels | test error rate: 6.06$\pm$0.50\% -> 2.86$\pm$0.07\% | Improved |
Closest papers by meaning
- Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning (TPAMI 2018 · similarity 0.90)
- Tangent-Normal Adversarial Regularization for Semi-Supervised Learning (CVPR 2019 · similarity 0.89)
- Enhancing Adversarial Robustness in Low-Label Regime via Adaptively Weighted Regularization and Knowledge Distillation (ICCV 2023 · similarity 0.88)
- Regularizing Discriminative Capability of CGANs for Semi-Supervised Generative Learning (CVPR 2020 · similarity 0.88)
- Kernel Learning with Adversarial Features: Numerical Efficiency and Adaptive Regularization (NeurIPS 2025 · similarity 0.85)
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.85)
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.
