Research map / Reinforcement learning / Visual and representation RL
SLACK: Stable Learning of Augmentations With Cold-Start and KL Regularization
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 |
|---|---|---|---|
| Add KL anchoring regularization to the policy updates (ablation) | SLACK · CIFAR10 and CIFAR100 | accuracy: 96.27 ± .05 -> 96.29 ± .08; 97.06 ± .11 -> 97.46 ± .06; 79.61 ± .13 -> 79.87 ± .11; 83.79 ± .19 -> 84.08 ± .16 | Helped |
| Search with WRN-28-10 rather than WRN-40-2 (ablation) | SLACK · CIFAR10 and CIFAR100, evaluated with WRN-28-10 | accuracy: 97.43 ± .04 -> 97.46 ± .06; 83.94 ± .20 -> 84.08 ± .16 | Helped |
| Learn both transformation probabilities π and magnitude bounds μ rather than only π or only μ (ablation) | SLACK · CIFAR100 | accuracy: 79.22 ± .17 -> 79.87 ± .11; 79.36 ± .11 -> 79.87 ± .11; 83.57 ± .18 -> 84.08 ± .16; 83.45 ± .15 -> 84.08 ± .16 | Helped |
| Learn both transformation probabilities π and magnitude bounds μ rather than only π or only μ (ablation) | SLACK · CIFAR10 | accuracy: 96.20 ± .08 -> 96.29 ± .08; 96.22 ± .09 -> 96.29 ± .08; 97.42 ± .05 -> 97.46 ± .06; 97.35 ± .04 -> 97.46 ± .06 | Mixed |
| Replace SLACK's learned policy with the initial Uniform policy | SLACK · ImageNet-100 | accuracy: 85.78 ± .32 -> 86.06 ± .11 | Helped |
| Use TA (Wide) rather than TA (RA) magnitude ranges | TrivialAugment · ImageNet-100 | accuracy: 85.87 ± .30 -> 86.39 ± .18 | Helped |
| Use TA (Wide) ImageNet rather than SLACK's learned policy | DomainNet · DomainNet average | accuracy: 62.80 ± .06 -> 63.03 ± .07 | Helped |
| Use uniform magnitude sampling instead of the learned magnitude model (ablation) | DeepAA policy · CIFAR10 | accuracy: 96.46 -> 96.55; 97.48 -> 97.47 | Helped |
| Use uniform magnitude sampling instead of the learned magnitude model (ablation) | DeepAA policy · CIFAR100 | accuracy: 79.62 -> 78.89; 83.85 -> 83.62 | Hurt |
| Use uniform magnitude sampling instead of the original magnitude model (ablation) | FastAA policy · CIFAR10 and CIFAR100 | accuracy: 96.4 -> 96.37; 97.22 -> 97.30; 79.11 -> 79.15; 82.82 -> 82.84 | Mixed |
| Use uniform magnitude sampling instead of the original magnitude model (ablation) | DADA policy · CIFAR10 and CIFAR100 | accuracy: 96.33 -> 96.37; 97.19 -> 97.35; 79.07 -> 78.97; 82.05 -> 82.57 | Mixed |
| Use warm-start instead of restarting each round from the pretrained model (ablation) | SLACK · Policy search | policy quality | Hurt |
Closest papers by meaning
- Deep invariant networks with differentiable augmentation layers (NeurIPS 2022 · similarity 0.88)
- Tradeoffs in Data Augmentation: An Empirical Study (ICLR 2021 · similarity 0.87)
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.87)
- Population Based Augmentation: Efficient Learning of Augmentation Policy Schedules (ICML 2019 · similarity 0.86)
- SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation (ICLR 2024 · similarity 0.85)
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 2024 · 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.