Research map / Architectures and efficiency / Neural architecture search
FreeAugment: Data Augmentation Search Across All Degrees of Freedom
Tom Bekor, Niv Nayman, Lihi Zelnik-Manor
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 |
|---|---|---|---|
| jointly learn all four degrees of freedom instead of freezing magnitudes to uniform (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.64 -> 80.04 | Improved |
| learn types and order instead of freezing them to uniform (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.54 -> 80.04 | Improved |
| learn policy depth instead of freezing it to uniform (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.61 -> 80.04 | Improved |
| use Gumbel-Sinkhorn permutation sampling instead of independent Gumbel-Softmax per layer (ablation) | FreeAugment · CIFAR10 WRN-40-2 | repeated transformation rate | Improved |
| increase Sinkhorn iterations L up to 20 (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%) | Improved |
| sample a distinct augmentation per image instead of per batch (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.43 -> 80.04 | Improved |
| add warm-up for policy parameters (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.53 -> 80.04 | Improved |
| use uniform instead of Gaussian magnitude distribution (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.9 -> 80.04 | No clear effect |
| learn depth with Gumbel-Softmax over depths instead of application probabilities (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.11 -> 80.04 | Improved |
| train evaluation classifier from scratch instead of from end-of-search weights (ablation) | FreeAugment · CIFAR100 | top-1 accuracy (%): 79.85 -> 80.04 | Improved |
| apply FreeAugment policy versus TA(Wide) on CIFAR100 WRN-28-10 | WRN-28-10 · CIFAR100 | top-1 accuracy (%): 84.33 -> 84.13 | Hurt |
| apply FreeAugment policy found on ImageNet-100 to full ImageNet | ResNet-50 · ImageNet | top-1 accuracy (%): 76.3 -> 77.63 | Improved |
Closest papers by meaning
- Faster AutoAugment: Learning Augmentation Strategies Using Backpropagation (ECCV 2020 · similarity 0.90)
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.88)
- TrivialAugment: Tuning-Free Yet State-of-the-Art Data Augmentation (ICCV 2021 · similarity 0.88)
- TeachAugment: Data Augmentation Optimization Using Teacher Knowledge (CVPR 2022 · similarity 0.87)
- OnlineAugment: Online Data Augmentation with Less Domain Knowledge (ECCV 2020 · similarity 0.87)
- 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.
