Research map / Optimization / Implicit bias and generalization
Penalizing Gradient Norm for Efficiently Improving Generalization in Deep Learning
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 loss-gradient-norm penalty with tuned balance coefficient | WideResNet-28-10 · CIFAR-10, basic augmentation | testing error rate: 2.78 -> 2.52 | Helped |
| Add loss-gradient-norm penalty with tuned balance coefficient | WideResNet-28-10 · CIFAR-10, cutout | testing error rate: 2.43 -> 2.16 | Helped |
| Add loss-gradient-norm penalty with tuned balance coefficient | WideResNet-28-10 · CIFAR-100, basic augmentation | testing error rate: 16.53 -> 16.02 | Helped |
| Add loss-gradient-norm penalty with tuned balance coefficient | ViT models · CIFAR-10 and CIFAR-100 | testing error rate | Helped |
| Add loss-gradient-norm penalty with tuned balance coefficient | ImageNet models · ImageNet | testing error rate | Helped |
| Use SAM's tied setting, balance coefficient α = 1, instead of tuning α independently (ablation) | Gradient-norm penalty scheme · CIFAR-10 and CIFAR-100 | testing error rate | Mixed |
| Set balance coefficient α to values below zero (ablation) | Gradient-norm penalty scheme · CIFAR-10 and CIFAR-100 | convergence | Hurt |
| Set balance coefficient α above 1 (ablation) | Gradient-norm penalty scheme · CIFAR-10 | testing error rate | Hurt |
| Use cutout augmentation rather than heavy augmentation (ablation) | Vision Transformers · CIFAR-10 | testing error rate | No clear effect |
Closest papers by meaning
- Generalization Error Bounds of Gradient Descent for Learning Over-Parameterized Deep ReLU Networks (AAAI 2020 · similarity 0.90)
- Sharpness Minimization Algorithms Do Not Only Minimize Sharpness To Achieve Better Generalization (NeurIPS 2023 · similarity 0.90)
- Optimal Rates for Generalization of Gradient Descent for Deep ReLU Classification (NeurIPS 2025 · similarity 0.89)
- Friendly Sharpness-Aware Minimization (CVPR 2024 · similarity 0.89)
- A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.89)
- Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization (NeurIPS 2024 · similarity 0.89)
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.