Research map / Label-efficient and robust learning / Recognition and long-tail
Learning Augmentation Network via Influence Functions
Check 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 augmentation by maximizing the influence-based estimate of validation-loss reduction | ResNet-56 on CIFAR-10 and CIFAR-100 · CIFAR-10 and CIFAR-100, 100% labels | test error rate (%): 5.6 -> 5.2; 31.6 -> 29.6 | Helped |
| learn augmentation with influence functions rather than the Ratner augmentation method | ResNet-56 · CIFAR-10, 10% labels | test error rate (%): 18.5 -> 17.7 | Helped |
| use influence-learned augmentation instead of AutoAugment | Wide-ResNet-28-10, Shake-Shake, and PyramidNet+ShakeDrop · CIFAR-10 and CIFAR-100 | test error rate (%): 2.6 -> 2.8; 2.0 -> 2.1; 1.5 -> 1.7; 17.1 -> 17.3; 14.3 -> 14.6; 10.7 -> 11.9 | Hurt |
| use influence-learned augmentation instead of AutoAugment | ResNet-50 and ResNet-200 · ImageNet val | Top-1 accuracy (%): 77.6 -> 77.1; 80.0 -> 79.0 | Hurt |
| use influence-learned augmentation instead of AutoAugment | ResNet-50 and ResNet-200 · ImageNet val | Top-5 accuracy (%): 93.8 -> 93.4; 95.0 -> 94.6 | Hurt |
| add learned spatial and appearance transformations together in parallel (ablation) | ResNet-56 · CIFAR-10 | test error rate (%): 6.4% -> 5.2% | Helped |
| add appearance transformation to spatial transformation (ablation) | ResNet-56 · CIFAR-10 | test error rate (%): 6.1% -> 5.4% | Helped |
| add spatial transformation to appearance transformation (ablation) | ResNet-56 · CIFAR-10 | test error rate (%): 5.8% -> 5.4% | Helped |
| increase τ dimension from 32 to 64 (ablation) | Proposed augmentation model · CIFAR-10, ResNet-56 | test error rate (%): 5.8% -> 5.4% | Helped |
| increase τ dimension from 64 to 128 (ablation) | Proposed augmentation model · CIFAR-10, ResNet-56 | test error rate (%): 5.4% -> 5.2% | Helped |
| increase τ dimension from 128 to 256 (ablation) | Proposed augmentation model · CIFAR-10, ResNet-56 | test error rate (%): 5.2% -> 5.2% | No clear effect |
| omit GAN pretraining of G (ablation) | Proposed augmentation model · classification accuracy | accuracy | No clear effect |
Closest papers by meaning
- Influence Functions in Deep Learning Are Fragile (ICLR 2021 · similarity 0.86)
- Data-Efficient Augmentation for Training Neural Networks (NeurIPS 2022 · similarity 0.86)
- A Group-Theoretic Framework for Data Augmentation (NeurIPS 2020 · similarity 0.85)
- Characterizing the Generalization Error of Random Feature Regression with Arbitrary Data-Augmentation (NeurIPS 2026 · similarity 0.85)
- Adversarial AutoAugment (ICLR 2020 · similarity 0.85)
- Towards Robust Influence Functions with Flat Validation Minima (ICML 2025 · similarity 0.83)
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.