Research map / Label-efficient and robust learning / Recognition and long-tail
Learning to Compose Domain-Specific Transformations for Data Augmentation
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 sequence generator to compose user-provided transformations instead of using random heuristic compositions | CIFAR-10, 10% labeled · CIFAR-10 | accuracy: 77.5 -> 81.5 | Helped |
| Learn a sequence generator to compose user-provided transformations instead of using random heuristic compositions | ACE Employment · ACE Employment | F1: 62.8 -> 64.2 | Helped |
| Learn to compose transformations rather than use random compositions | DDSM + domain-specific transformations · DDSM + DS | accuracy: 53.7 -> 62.7 | Helped |
| Add segmentation-based domain-specific transformations to the standard DDSM transformation set | Heuristic augmentation · DDSM | accuracy: 59.3 -> 53.7 | Hurt |
| Add segmentation-based domain-specific transformations to the standard DDSM transformation set | LSTM augmentation · DDSM | accuracy: 61.0 -> 62.7 | Helped |
| Use an LSTM sequence model instead of the independent mean-field model | CIFAR-10, 10% labeled · CIFAR-10 | accuracy: 79.8 -> 81.5 | Helped |
| Use an LSTM sequence model instead of the independent mean-field model | MNIST, 10% labeled · MNIST | accuracy: 99.2 -> 99.1 | Hurt |
| Add transformation regularization with coefficient 0.1 (ablation) | CIFAR-10 heuristic augmentation · CIFAR-10 validation | accuracy: 77.4 -> 77.5 | Helped |
| Add transformation regularization with coefficient 0.1 (ablation) | CIFAR-10 LSTM augmentation · CIFAR-10 validation | accuracy: 80.4 -> 81.6 | Helped |
| Use the learned LSTM augmentation model instead of heuristic random transformations (ablation) | CIFAR-10 validation with transformation regularization · CIFAR-10 validation | accuracy: 77.5 -> 81.6 | Helped |
| Use the LSTM augmentation model instead of heuristic augmentation | CIFAR-10, 10% labeled folds · CIFAR-10 | accuracy: 80.42 ± 0.58; 81.37 ± 2.32; 81.47 ± 0.46 | No clear effect |
Closest papers by meaning
- AdvST: Revisiting Data Augmentations for Single Domain Generalization (AAAI 2024 · similarity 0.86)
- Domain Gap Embeddings for Generative Dataset Augmentation (CVPR 2024 · similarity 0.85)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.85)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.85)
- First-Order Manifold Data Augmentation for Regression Learning (ICML 2024 · similarity 0.84)
- UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation (NeurIPS 2025 · 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.