Research map / Optimization / Implicit bias and generalization
The Crucial Role of Normalization in Sharpness-Aware Minimization
Yan Dai, Kwangjun Ahn, Suvrit Sra
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| remove gradient normalization from SAM | SAM · CIFAR-10 early-stage training, CNN and FCN | training loss stability | Hurt |
| use normalized SAM updates instead of un-normalized updates | SAM and USAM · CIFAR-10 test, ResNet18 adversarial checkpoint | test accuracy | Improved |
| use small ρ in USAM rather than normalized SAM | USAM · CIFAR-10 test, ResNet18 adversarial checkpoint | test accuracy | Hurt |
| use larger ρ in USAM rather than normalized SAM | USAM · CIFAR-10 training, ResNet18 adversarial checkpoint | training accuracy stability | Hurt |
| remove gradient normalization from SAM | SAM · sparse coding problem | test accuracy | Mixed |
| vary ρ in USAM, using a tiny ρ=0.01 (ablation) | USAM · sparse coding problem | test accuracy | No clear effect |
| vary ρ in USAM, using ρ=0.1 or ρ=0.3 (ablation) | USAM · sparse coding problem | test accuracy | Improved |
| vary ρ in USAM, using ρ=0.5 (ablation) | USAM · sparse coding problem | training stability | Hurt |
| vary η while fixing ρ (ablation) | SAM · sparse coding problem | test accuracy | Improved |
| vary η while fixing ρ (ablation) | USAM · sparse coding problem | test accuracy | Mixed |
Closest papers by meaning
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.94)
- A Universal Class of Sharpness-Aware Minimization Algorithms (ICML 2024 · similarity 0.92)
- Friendly Sharpness-Aware Minimization (CVPR 2024 · similarity 0.92)
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.92)
- Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.91)
- Fundamental Convergence Analysis of Sharpness-Aware Minimization (NeurIPS 2024 · similarity 0.91)
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.
