Research map / Reinforcement learning / Visual and representation RL
Population Based Augmentation: Efficient Learning of Augmentation Policy Schedules
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 |
|---|---|---|---|
| learn a nonstationary augmentation schedule rather than a fixed policy (ablation) | PBA · CIFAR-10, Wide-ResNet-28-10 | test error (%): 2.76 -> 2.576 $\pm$ 0.062 | Helped |
| learn a nonstationary augmentation schedule rather than a fixed policy (ablation) | PBA · CIFAR-10, PyramidNet+ShakeDrop | test error (%): 1.55 -> 1.46 $\pm$ 0.077 | Helped |
| shuffle schedule order while keeping policy durations (ablation) | PBA · CIFAR-10, Wide-ResNet-28-10 | test error (%): 2.576 $\pm$ 0.062 -> 2.66 | Hurt |
| replace the epoch schedule with a stationary distribution (ablation) | PBA · CIFAR-10, Wide-ResNet-28-10 | test error (%): 2.576 $\pm$ 0.062 -> 2.89 | Hurt |
| use the final schedule policy throughout training (ablation) | PBA · CIFAR-10, Shake-Shake (26 2x96d) | test error (%): 2.03 $\pm$ 0.11 -> 2.33 | Hurt |
| use a ResNet-20 population instead of WRN-40-2 (ablation) | PBA · Reduced CIFAR-10 | test error: 0.8484 -> 0.7484 | Mixed |
| increase PBT population size beyond 16 (ablation) | PBA · Reduced CIFAR-10 | test error: 0.8484 -> 0.8446 | No clear effect |
| reduce PBT search to 100 epochs (ablation) | PBA · CIFAR-10 | performance | Hurt |
| vary perturbation interval between 2 and 4 (ablation) | PBA explore function · CIFAR-10 | performance | No clear effect |
| use PBA instead of AutoAugment for policy search | CIFAR-10 policy search · CIFAR-10 | GPU hours: 5000 -> 5 | Helped |
| use PBA instead of AutoAugment for policy search | SVHN policy search · SVHN | GPU hours: 1000 -> 1 | Helped |
| use PBA instead of AutoAugment | PyramidNet+ShakeDrop · CIFAR-100 | test error (%): 10.67 -> 10.94 $\pm$ 0.094 | Hurt |
Closest papers by meaning
- Tradeoffs in Data Augmentation: An Empirical Study (ICLR 2021 · similarity 0.89)
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.88)
- AdaAug: Learning Class- and Instance-adaptive Data Augmentation Policies (ICLR 2022 · similarity 0.88)
- AutoAugment: Learning Augmentation Strategies From Data (CVPR 2019 · similarity 0.87)
- SLACK: Stable Learning of Augmentations With Cold-Start and KL Regularization (CVPR 2023 · similarity 0.86)
- Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation (NeurIPS 2026 · similarity 0.84)
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.