Research map / Optimization / Implicit bias and generalization
Sharpness-Aware Minimization Can Hallucinate Minimizers
Chanwoong Park, Uijeong Jang, Ernest K. Ryu, Insoon Yang
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 |
|---|---|---|---|
| add 10% SGD warm-start before SAM | SAM · CIFAR-100 ResNet-18 (stochastic, rho=0.3 peak) | Test accuracy (%): 77.05 -> 77.49 | Improved |
| add short SGD warm-start before SAM | SAM · MNIST 2-layer Tanh (full-batch) | training loss plateau | Improved |
| add short SGD warm-start before SAM | SAM · CIFAR-100 ResNet-18 (stochastic) | stability across rho | Improved |
| increase perturbation radius rho from 1.0 | SAM · MNIST 2-layer Tanh (full-batch) | training loss / test accuracy | Hurt |
| increase rho from 0.0 to 0.5 (SAM instead of plain gradient descent) | SAM · CIFAR-10 ResNet-18 (full-batch) | final training loss f(x_T): 0.001 -> 0.081 | Hurt |
| increase rho from 0.0 to 2.0 | SAM · CIFAR-100 ResNet-18 (full-batch) | final training loss f(x_T): 0.009 -> 0.828 | Hurt |
| increase rho from 0.0 to 1.5 | SAM · CIFAR-10 ResNet-18 (full-batch) | gradient norm ||grad f(x_T)||: 0.006 -> 11.932 | Hurt |
| increase rho from 0.0 to 0.5 | SAM · CIFAR-100 ResNet-18 (full-batch) | shifted gradient norm ||grad f(x_T+)||: 0.058 -> 0.100 | Mixed |
Closest papers by meaning
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 2025 · similarity 0.92)
- Practical Sharpness-Aware Minimization Cannot Converge All the Way to Optima (NeurIPS 2023 · similarity 0.92)
- Improving Sharpness-Aware Minimization by Lookahead (ICML 2024 · similarity 0.91)
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.90)
- Sharpness-Aware Minimization Enhances Feature Quality via Balanced Learning (ICLR 2024 · similarity 0.90)
- Tilted Sharpness-Aware Minimization (ICML 2025 · 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.
