Research map / Optimization / Implicit bias and generalization
Friendly Sharpness-Aware Minimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Friendly-SAM (F-SAM), which removes the EMA-estimated full-gradient component from SAM's perturbation and perturbs along the batch-specific stochastic gradient noise, improving generalization and robustness to perturbation radius.
Problem
Understanding which components of SAM's adversarial perturbation drive its generalization gains, and improving SAM's generalization and radius sensitivity.
Approach
Decompose the minibatch gradient into full-gradient and stochastic-noise components, find the full-gradient part hurts generalization, and perturb along the noise component only, estimating the full gradient with an EMA of historical minibatch gradients (with a constant sigma=1 projection).
Key results
- F-SAM improves ResNet-18 on CIFAR-100 to 81.29% vs 80.88% for SAM and 77.90% for SGD (Table 2).
- On ImageNet with ResNet-50, F-SAM reaches 77.35% vs 77.14% for SAM (Table 4).
- Under 80% label noise on CIFAR-10 with ResNet-18, F-SAM achieves 59.39% vs 31.73% for SAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SAM perturbation direction with the full-gradient component (SAM-full) (ablation) | SAM · CIFAR-100, ResNet-18 | test accuracy (%): 80.88 -> 78.02 | Hurt |
| replace SAM perturbation direction with stochastic gradient noise only (SAM-noise) (ablation) | SAM · CIFAR-100, ResNet-18 | test accuracy (%): 80.88 -> 81.21 | Helped |
| use a different random minibatch for perturbation and minimization (SAM-db) (ablation) | SAM · CIFAR-10, ResNet-18 | test accuracy | No clear effect |
| increase the perturbation-step minibatch size while keeping minimization batch size fixed (ablation) | SAM, CIFAR-100 ResNet-18 · CIFAR-100 | test accuracy | Hurt |
| remove the full-gradient component using the F-SAM perturbation | SAM · CIFAR-100, WRN-28-10 | test accuracy (%): 84.88 -> 85.16 | Helped |
| integrate the friendly perturbation with ASAM (ablation) | ASAM · CIFAR-100, WRN-28-10 | test accuracy (%): 84.99±0.22 -> 85.24±0.16 | Helped |
| increase perturbation radius to twice the optimal radius (ablation) | SAM · CIFAR-100, ResNet-18 | test accuracy (%): 80.88% -> 78.83% | Hurt |
| increase perturbation radius to twice the optimal radius (ablation) | F-SAM · CIFAR-100, ResNet-18 | test accuracy (%): 80.30% | No clear effect |
| increase training batch size (ablation) | F-SAM relative to SAM · CIFAR-100, ResNet-18 | test accuracy | Helped |
| set projection constant σ to 1 (ablation) | F-SAM · CIFAR-10 and CIFAR-100 | test accuracy | No clear effect |
Surprising findings
- Using only the full-gradient component as the SAM perturbation (SAM-full) substantially degrades generalization, often below the base optimizer.
- Perturbing with only the stochastic gradient noise (SAM-noise) outperforms vanilla SAM.
Limitations
- Sigma is fixed to a constant 1 for efficiency and stability rather than computed per iteration.
- Gains over SAM on CIFAR and ImageNet are small (often 0.1-0.4 points).
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet Stanford Cars OxfordIIITPet Flowers102
Benchmarks: CIFAR-10 CIFAR-100 ImageNet Stanford Cars OxfordIIITPet Flowers102
Metrics: test accuracy
Baselines: SGD AdamW SAM ASAM FisherSAM
Builds on: Sharpness-Aware Minimization ASAM FisherSAM Entropy-SGD
Closest papers by meaning
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.93)
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.93)
- A Single-Step, Sharpness-Aware Minimization is All You Need to Achieve Efficient and Accurate Sparse Training (NeurIPS 2024 · similarity 0.92)
- Changing the Training Data Distribution to Reduce Simplicity Bias Improves In-distribution Generalization (NeurIPS 2024 · similarity 0.91)
- Unveiling m-Sharpness Through the Structure of Stochastic Gradient Noise (NeurIPS 2025 · similarity 0.91)
- Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.90)
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.