Research map / Architectures and efficiency / Neural architecture search
Faster AutoAugment: Learning Augmentation Strategies Using Backpropagation
Ryuichiro Hataya, Zdenek Jan, Kazuki Yoshizoe, Hideki Nakayama
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 AutoAugment search with differentiable adversarial search (Faster AA) | AutoAugment · CIFAR-10 search cost | GPU hours: 5,000 -> 0.23 | Improved |
| replace AutoAugment search with differentiable adversarial search (Faster AA) | AutoAugment · ImageNet search cost | GPU hours: 15,000 -> 2.3 | Improved |
| add Faster AA policy instead of no augmentation policy | ResNet-50 · ImageNet top-1 | test error (%): 24.1 -> 23.5 | Improved |
| add Faster AA policy instead of Cutout | WideResNet-28-10 · Reduced CIFAR-10 | test error (%): 22.5 -> 14.8 | Improved |
| train policy on target CIFAR-100 instead of CIFAR-10 (ablation) | Faster AA WideResNet-28-10 · CIFAR-100 | test error (%): 17.8 -> 17.3 | Improved |
| increase number of sub-policies L (ablation) | Faster AA WideResNet-40-2 · CIFAR-10 | test error (%) | Improved |
| increase operation count K per sub-policy from 1 to 4 (ablation) | Faster AA WideResNet-40-2 · CIFAR-10 | test error (%) | Improved |
| train policy on full 50,000 CIFAR-10 images instead of 4,000, with L=10 (ablation) | Faster AA WideResNet-40-2 · CIFAR-10 | test error (%): 3.7 -> 4.1 | Hurt |
| train policy with L=80 sub-policies on full CIFAR-10 instead of L=10 (ablation) | Faster AA WideResNet-40-2 · CIFAR-10 | test error (%): 4.1 -> 3.8 | Improved |
| add trained policy instead of randomly initialized policy (ablation) | Faster AA WideResNet-28-10 · CIFAR-10 | test error (%): 2.7 -> 2.6 | Improved |
Closest papers by meaning
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.93)
- Fast AutoAugment (NeurIPS 2019 · similarity 0.92)
- FreeAugment: Data Augmentation Search Across All Degrees of Freedom (ECCV 2024 · similarity 0.90)
- Adversarial AutoAugment (ICLR 2020 · similarity 0.90)
- Deep AutoAugment (ICLR 2022 · similarity 0.90)
- RandAugment: Practical Automated Data Augmentation with a Reduced Search Space (NeurIPS 2020 · 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.
