Research map / Optimization / Nonconvex and smooth optimization
Practical Sharpness-Aware Minimization Cannot Converge All the Way to Optima
Dongkuk Si, Chulhee Yun
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 |
|---|---|---|---|
| use constant perturbation size rho with gradient normalization instead of decaying rho, for deterministic SAM | deterministic SAM · smooth strongly convex functions | convergence rate of best-iterate suboptimality | Mixed |
| switch from deterministic to stochastic SAM (constant rho, normalized) | SAM · smooth strongly convex functions | suboptimality at convergence | Hurt |
| keep constant rho in deterministic SAM on smooth nonconvex functions | deterministic SAM · smooth nonconvex functions | average squared gradient norm | Hurt |
| use gradient normalization in the ascent step versus no normalization (ablation) | SAM on f(x,y)=(xy-1)^2 · trajectory illustration | limit point | Mixed |
| use m-SAM (same sample for ascent and descent) instead of n-SAM (independent samples) | stochastic SAM · smooth convex functions | convergence to global minima | Hurt |
| stochastic m-SAM with high gradient noise (sigma > beta rho) | m-SAM · smooth strongly convex functions | suboptimality gap | Hurt |
| use small gradient noise (sigma <= beta rho) in n-SAM bound | stochastic n-SAM · smooth strongly convex and convex functions | convergence rate modulo additive term | Improved |
| deterministic SAM on nonsmooth Lipschitz convex functions with constant rho | deterministic SAM · nonsmooth Lipschitz convex functions | distance to global minimum | Hurt |
| step size eta of order 1/beta or larger | deterministic SAM · strongly convex quadratic | iterate stability | Hurt |
Closest papers by meaning
- Fundamental Convergence Analysis of Sharpness-Aware Minimization (NeurIPS 2024 · similarity 0.91)
- Sharpness-Aware Minimization: General Analysis and Improved Rates (ICLR 2025 · similarity 0.91)
- Unveiling m-Sharpness Through the Structure of Stochastic Gradient Noise (NeurIPS 2025 · similarity 0.91)
- Explicit Eigenvalue Regularization Improves Sharpness-Aware Minimization (NeurIPS 2024 · similarity 0.91)
- A Universal Class of Sharpness-Aware Minimization Algorithms (ICML 2024 · similarity 0.91)
- Sharpness-Aware Minimization Efficiently Selects Flatter Minima Late In Training (ICLR 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.
