Research map / Architectures and efficiency / Neural architecture search
Direct Differentiable Augmentation Search
Aoming Liu, Zehao Huang, Zhiwu Huang, Naiyan Wang
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 |
|---|---|---|---|
| learn p_tp adaptively instead of fixing it to 1.0 (ablation) | DDAS · CIFAR-100 (WRN-28-10) | test error (%): 24.7 -> 16.6 | Improved |
| learn p_tp adaptively instead of fixing it to 1.0 (ablation) | DDAS · ImageNet (ResNet-50) | test error (%): 23.1 -> 22.3 | Improved |
| learn p_tp adaptively instead of best fixed value (ablation) | DDAS · CIFAR-100 (WRN-28-10) | test error (%): 16.6 vs 16.6 | No clear effect |
| learn p_tp adaptively instead of fixed 0.4 (ablation) | DDAS · ImageNet (ResNet-50) | test error (%): 22.2 -> 22.3 | Hurt |
| use dynamic epoch-wise policy instead of fixing p_tp and p_o (ablation) | DDAS · CIFAR-100 (WRN-28-10) | test error (%): 16.9 -> 16.6 | Improved |
| use dynamic epoch-wise policy instead of fixing p_tp and p_o (ablation) | DDAS · ImageNet (ResNet-50) | test error (%): 22.4 -> 22.3 | Hurt |
| use dynamic policy instead of fixing only p_o (ablation) | DDAS · ImageNet (ResNet-50) | test error (%): 22.2 -> 22.3 | Hurt |
| increase sampled policies per step L-hat from 3 to 11 (ablation) | DDAS · CIFAR-100 (WRN-28-10) | test error (%): 16.6 -> 17.1 | Hurt |
| increase sampled policies per step L-hat from 3 to 7 (ablation) | DDAS · CIFAR-100 (WRN-28-10) | test error (%): 16.6 -> 16.8 | Hurt |
| DDAS instead of Faster AA / DADA on CIFAR-100 | DDAS · CIFAR-100 (WRN-28-10) | test error (%): 17.3 -> 16.6 (Faster AA), 17.5 -> 16.6 (DADA) | Improved |
| DDAS instead of AA search for detection | DDAS · COCO (RetinaNet R50) | mAP: 38.4 -> 38.1 mAP with 19800 -> 20 search hours | Mixed |
| DDAS augmentation vs baseline | RetinaNet R50 · COCO val | mAP: 36.9 -> 38.1 | Improved |
Closest papers by meaning
- Differentiable Image Data Augmentation and Its Applications: A Survey (TPAMI 2023 · similarity 0.90)
- Faster AutoAugment: Learning Augmentation Strategies Using Backpropagation (ECCV 2020 · similarity 0.88)
- RandAugment: Practical Automated Data Augmentation with a Reduced Search Space (NeurIPS 2020 · similarity 0.87)
- DADA: Differentiable Automatic Data Augmentation (ECCV 2020 · similarity 0.86)
- Sample-Aware RandAugment: Search-Free Automatic Data Augmentation for Effective Image Recognition (IJCV 2025 · similarity 0.86)
- Your Image is My Video: Reshaping the Receptive Field via Image-To-Video Differentiable AutoAugmentation and Fusion (CVPR 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.
