Research map / Label-efficient and robust learning / Recognition and long-tail
ASAM: Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks
Jungmin Kwon, Jeongseop Kim, Hyunseo Park, In Kwon Choi
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 |
|---|---|---|---|
| Replace SAM's fixed-radius perturbation with scale-adaptive perturbations (ASAM) | PyramidNet-272 · CIFAR-10 | test accuracy: 98.55_{\pm 0.05} -> 98.68_{\pm 0.08} | Improved |
| Replace SAM's fixed-radius perturbation with scale-adaptive perturbations (ASAM) | PyramidNet-272 · CIFAR-100 | test accuracy: 89.36_{\pm 0.20} -> 89.90_{\pm 0.13} | Improved |
| Replace SAM with ASAM | ResNet-50 · ImageNet | Top1 accuracy: 76.39_{\pm 0.03} -> 76.63_{\pm 0.18} | Improved |
| Replace Adam+SAM with Adam+ASAM | Transformer · IWSLT'14 DE-EN test | BLEU: 34.78_{\pm 0.01} -> 35.02_{\pm <0.01} | Improved |
| Replace SAM with ASAM | ResNet-32 · CIFAR-10 with 20%, 40%, and 60% label noise | test accuracy: 92.94_{\pm 0.12} -> 93.21_{\pm 0.10}; 90.62_{\pm 0.18} -> 90.89_{\pm 0.13}; 86.58_{\pm 0.30} -> 87.41_{\pm 0.16} | Improved |
| Replace SAM with ASAM | ResNet-32 · CIFAR-10 with 80% label noise | test accuracy: 69.92_{\pm 0.98} -> 67.69_{\pm 1.34} | Hurt |
| Use p=2 rather than p=∞ for the adaptive maximization region (ablation) | ASAM · CIFAR image classification | test accuracy | Improved |
| Add bias normalization (ablation) | ASAM with element-wise normalization and p=2 · CIFAR image classification | test accuracy | No clear effect |
| Use element-wise rather than filter-wise normalization (ablation) | ASAM with WideResNet-16-8 · CIFAR image classification | test accuracy | No clear effect |
| Use adaptive sharpness rather than ordinary sharpness (ablation) | Generalization-gap measure · WideResNet-28-2 on CIFAR-10 | Kendall rank correlation τ: 0.174 -> 0.636 (p=2); 0.257 -> 0.616 (p=∞) | Improved |
Later papers that compared with this method
3 later accepted papers report ASAM as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Inconsistency-Aware Minimization: Improving Generalization with Unlabeled Data (ICML 2026 · better on 3 of 4 shared results)
- Quantifying and Optimizing Simplicity via Polynomial Representations (ICML 2026 · better on 1 of 1 shared result)
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · better on 0 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.91)
- A Universal Class of Sharpness-Aware Minimization Algorithms (ICML 2024 · similarity 0.91)
- Friendly Sharpness-Aware Minimization (CVPR 2024 · similarity 0.89)
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.89)
- $\boldsymbol{\mu}\mathbf{P^2}$: Effective Sharpness Aware Minimization Requires Layerwise Perturbation Scaling (NeurIPS 2024 · similarity 0.89)
- Implicit Regularization of Sharpness-Aware Minimization for Scale-Invariant Problems (NeurIPS 2024 · similarity 0.88)
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.
