Research map / Reinforcement learning / Visual and representation RL
MetaAugment: Sample-Aware Data Augmentation Policy Learning
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 |
|---|---|---|---|
| Use sample-aware reweighting of augmented examples instead of dataset-level augmentation policies | MetaAugment · CIFAR-10, CIFAR-100, Omniglot, and ImageNet | Top-1 accuracy | Mixed |
| Use multiple transformations per sample | MetaAugment on WRN-28-10 · CIFAR-10 | Top-1 accuracy: 98.10 -> 98.26$\pm$0.02 | Helped |
| Use multiple transformations per sample | MetaAugment on WRN-28-10 · CIFAR-100 | Top-1 accuracy: 84.51 -> 85.21$\pm$0.09 | Helped |
| Reject augmented samples whose weights fall below the mini-batch mean (ablation) | MetaAugment · CIFAR-100, WRN-28-10 and WRN-40-2 | Top-1 accuracy: 83.79 -> 82.57\%; 80.60 -> 79.01\% | Hurt |
| Include transformation embedding in the policy-network input (ablation) | MetaAugment on WRN-28-10 · CIFAR-10, CIFAR-100, and Omniglot | Top-1 accuracy: 97.58 -> 97.76; 83.49 -> 83.79; 89.29 -> 89.61 | Helped |
| Share the task-network feature extractor with the policy network (ablation) | MetaAugment on WRN-28-10 · CIFAR-10, CIFAR-100, and Omniglot | Top-1 accuracy: 97.59 -> 97.76; 83.68 -> 83.79; 89.29 -> 89.61 | Helped |
| Sample transformations from the estimated distribution with randomness (epsilon=0.1), rather than uniformly at random (epsilon=1.0) (ablation) | MetaAugment on WRN-28-10 · Omniglot | Top-1 accuracy | Helped |
| Transfer the learned policy to WRN-40-2 rather than jointly train it (ablation) | MetaAugment · CIFAR-10 | Top-1 accuracy: 96.79$\pm$0.06 -> 96.82$\pm$0.10 | No clear effect |
| Transfer the learned policy to WRN-40-2 rather than jointly train it (ablation) | MetaAugment · CIFAR-100 | Top-1 accuracy: 80.60$\pm$0.16 -> 80.15$\pm$0.09 | Hurt |
Closest papers by meaning
- Data Augmentation for Meta-Learning (ICML 2021 · similarity 0.90)
- SelectAugment: Hierarchical Deterministic Sample Selection for Data Augmentation (AAAI 2023 · similarity 0.90)
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.90)
- KeepAugment: A Simple Information-Preserving Data Augmentation Approach (CVPR 2021 · similarity 0.88)
- RandAugment: Practical Automated Data Augmentation with a Reduced Search Space (NeurIPS 2020 · similarity 0.88)
- A Data-Augmentation Is Worth A Thousand Samples: Analytical Moments And Sampling-Free Training (NeurIPS 2022 · similarity 0.87)
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.