Research map / Adversarial robustness and security / Certified and empirical robustness
Adversarial AutoAugment
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 fixed standard augmentation with randomly generated dynamic policies (ablation) | ResNet-50 training · ImageNet | top-1 test error (%): 22.70 -> 21.68 | Helped |
| replace randomly generated policies with adversarial policies from the policy network (ablation) | ResNet-50 training · ImageNet | top-1 test error (%): 21.68 -> 20.60 | Helped |
| increase augmentation instances per batch and use fixed standard augmentation (ablation) | ResNet-50 training · ImageNet | top-1 test error (%): 23.69 -> 22.70 | Helped |
| use adversarial policy learning instead of direct transfer of ImageNet-learned policies (ablation) | Wide-ResNet-28-10 · CIFAR-10 | top-1 test error (%): 2.45±0.13 -> 1.90 | Helped |
| use adversarial policy learning instead of direct transfer of ImageNet-learned policies (ablation) | Wide-ResNet-28-10 · CIFAR-100 | top-1 test error (%): 16.48±0.15 -> 15.49 | Helped |
| use adversarial policy learning instead of direct transfer of ImageNet-learned policies (ablation) | ResNet-50-D · ImageNet | top-1 test error (%): 20.20±0.05 -> 20.00 | Helped |
| use adversarial policy learning instead of direct transfer of ImageNet-learned policies (ablation) | ResNet-200 · ImageNet | top-1 test error (%): 19.05±0.10 -> 18.68 | Helped |
| increase the number of sampled augmentation policies per batch beyond 8 (ablation) | Adversarial AutoAugment with Wide-ResNet-28-10 · CIFAR-10 | test accuracy | No clear effect |
| search policies online and reuse target-network computation rather than perform AutoAugment search and retraining | ResNet-50 training · ImageNet | GPU hours: 15160 -> 1280 | Helped |
| search policies online and reuse target-network computation rather than perform AutoAugment search and retraining | ResNet-50 training · ImageNet | time overhead (days): 11 -> 1 | Helped |
Closest papers by meaning
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.92)
- Efficient and Effective Augmentation Strategy for Adversarial Training (NeurIPS 2022 · similarity 0.90)
- AutoDO: Robust AutoAugment for Biased Data With Label Noise via Scalable Probabilistic Implicit Differentiation (CVPR 2021 · similarity 0.88)
- Adversarial Auto-Augment with Label Preservation: A Representation Learning Principle Guided Approach (NeurIPS 2022 · similarity 0.87)
- KeepAugment: A Simple Information-Preserving Data Augmentation Approach (CVPR 2021 · similarity 0.87)
- Adversarial AutoMixup (ICLR 2024 · 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.