Research map / Label-efficient and robust learning / Recognition and long-tail
TeachAugment: Data Augmentation Optimization Using Teacher Knowledge
Check 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 |
|---|---|---|---|
| Add teacher-loss guidance so augmentations remain recognizable to the teacher | TeachAugment · CIFAR-10 and CIFAR-100, WRN-28-10 | error rate (%): 3.1 -> 2.5; 18.4 -> 16.8 | Helped |
| Use the non-saturating loss when updating the augmentation model (ablation) | TeachAugment · CIFAR-100, WRN-28-10 | error rate (%): 18.7% to 17.4% | Helped |
| Use an EMA teacher rather than a pretrained WRN-40-2 teacher (ablation) | TeachAugment · CIFAR-10, WRN-40-2 | error rate (%): 4.4 -> 3.7 | Helped |
| Use an EMA teacher rather than a pretrained WRN-40-2 teacher (ablation) | TeachAugment · CIFAR-100, WRN-40-2 | error rate (%): 21.3 -> 20.3 | Helped |
| Use a stronger pretrained WRN-28-10 teacher instead of the WRN-40-2 teacher (ablation) | TeachAugment · CIFAR-100, WRN-40-2 | error rate (%): 21.3 -> 21.6 | Hurt |
| Add color regularization (ablation) | TeachAugment · CIFAR-10, WRN-28-10 | error rate (%): 3.0 -> 2.5 | Helped |
| Add color regularization (ablation) | TeachAugment · CIFAR-100, WRN-28-10 | error rate (%): 16.5 -> 16.8 | Hurt |
| Add experience replay (ablation) | TeachAugment · CIFAR-10 and CIFAR-100, WRN-28-10 | error rate (%): 2.7 -> 2.5; 17.3 -> 16.8 | Helped |
| Add label smoothing (ablation) | TeachAugment · CIFAR-10 and CIFAR-100, WRN-28-10 | error rate (%): 3.0 -> 2.5; 17.3 -> 16.8 | Helped |
| Replace the TeachAugment objective with the Adv. AA objective (ablation) | TeachAugment framework · CIFAR-10 and CIFAR-100, WRN-28-10 | error rate (%): 2.5 -> 3.6; 16.8 -> 19.4 | Hurt |
| Replace the TeachAugment objective with the PointAugment objective (ablation) | TeachAugment framework · CIFAR-10 and CIFAR-100, WRN-28-10 | error rate (%): 2.5 -> 2.9; 16.8 -> 17.5 | Hurt |
| Replace the proposed neural augmentation model with DDA (ablation) | TeachAugment · CIFAR-10 and CIFAR-100, WRN-28-10 | error rate (%): 2.5 -> 2.8; 16.8 -> 18.0 | Hurt |
Closest papers by meaning
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.88)
- Learning to Compose Domain-Specific Transformations for Data Augmentation (NeurIPS 2017 · similarity 0.88)
- On the Generalization Effects of Linear Transformations in Data Augmentation (ICML 2020 · similarity 0.87)
- Understanding the detrimental class-level effects of data augmentation (NeurIPS 2023 · similarity 0.87)
- Adversarial AutoAugment (ICLR 2020 · similarity 0.87)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.86)
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.