Research map / Architectures and efficiency / Pruning and efficient tuning
Efficient Augmentation via Data Subsampling
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 |
|---|---|---|---|
| Rank augmentation candidates by training loss or model influence instead of sampling uniformly | Data augmentation subsampling · MNIST, CIFAR10, and NORB; translation, rotation, and crop | test accuracy | Helped |
| Use loss- or influence-based sampling to augment only 10% of points rather than all points | Data augmentation · Translation on MNIST, CIFAR10, and NORB | test accuracy: MNIST: 99.8% vs. 99.6%; CIFAR10: 97.2% vs. 97.3%; NORB: 89.8% vs. 89.7% | No clear effect |
| Use loss- or influence-based sampling rather than random sampling for rotation augmentation | Data augmentation subsampling · MNIST, CIFAR10, and NORB; rotation | test accuracy: 5--10% vs. 90--100% | Helped |
| Use model influence scores rather than training loss scores (ablation) | Loss- and influence-based augmentation selection · MNIST, CIFAR10, and NORB; translation, rotation, and crop | test accuracy | No clear effect |
| Add sample reweighting when selected augmented points are added (ablation) | Influence-based rotation augmentation · MNIST, CIFAR10, and NORB | test accuracy | Mixed |
| Update augmentation scores as the model is retrained (ablation) | Influence-based augmentation selection · MNIST, CIFAR10, and NORB | test accuracy | No clear effect |
| Select augmentation points using an SVM margin score instead of loss or influence (ablation) | Data augmentation subsampling · MNIST, CIFAR10, and NORB; tested augmentations | test accuracy | Hurt |
| Use clustered stratified sampling instead of standard influence-based sampling (ablation) | Data augmentation subsampling · MNIST, CIFAR10, and NORB; tested augmentations | test accuracy | Hurt |
| Combine influence-proportional selection with clustering (ablation) | Data augmentation subsampling · MNIST, CIFAR10, and NORB; tested augmentations | test accuracy | Hurt |
Closest papers by meaning
- Automatic Data Augmentation via Invariance-Constrained Learning (ICML 2023 · similarity 0.90)
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.89)
- Learning Instance-Specific Augmentations by Capturing Local Invariances (ICML 2023 · similarity 0.88)
- How Much Data Are Augmentations Worth? An Investigation into Scaling Laws, Invariance, and Implicit Regularization (ICLR 2023 · similarity 0.88)
- A Data-Augmentation Is Worth A Thousand Samples: Analytical Moments And Sampling-Free Training (NeurIPS 2022 · similarity 0.88)
- Revisiting Data Augmentation in Deep Reinforcement Learning (ICLR 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.