Research map / Optimization / Optimizers
Adaptive Batch Sizes Using Non-Euclidean Gradient Noise Scales for Stochastic Sign and Spectral Descent
Hiroki Naganuma, Shagun Gupta, Youssef Briki, Ioannis Mitliagkas, Irina Rish, Parameswaran Raman, Hao-Jun Michael Shi
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 non-Euclidean (l1) GNS instead of Euclidean GNS for adaptive batch (ablation) | signSGD · C4 160M Llama 3 | validation loss: 3.8906 -> 3.8396 | Improved |
| use non-Euclidean GNS instead of Euclidean GNS (ablation) | specSGD · C4 160M Llama 3 | validation loss: 3.7436 -> 3.7285 | Improved |
| use non-Euclidean GNS instead of Euclidean GNS (ablation) | Muon · C4 160M Llama 3 | validation loss: 3.3181 -> 3.3061 | Improved |
| adaptive batch size vs constant B=64 | signSGD · C4 160M Llama 3 | validation loss: 4.1390 -> 3.8396 | Improved |
| adaptive batch size vs constant B=256 | specSGD · C4 160M Llama 3 | validation loss: 3.9404 -> 3.7285 | Improved |
| adaptive batch size starting at B=64 | Muon · C4 160M Llama 3 | steps to baseline loss: 66.77% fewer steps | Improved |
| adaptive batch size starting at B=64 | Signum · C4 160M Llama 3 | steps to baseline loss: 66.61% fewer steps | Improved |
| adaptive batch size starting at B=64 | signSGD · C4 1B Llama 3 | steps to baseline loss: 31.84% fewer steps | Improved |
| adaptive batch size starting at B=64 | AdamW · C4 1B Llama 3 | validation loss: 2.7701 -> 2.8023 | Hurt |
| adaptive batch size starting at B=64 | Signum · C4 1B Llama 3 | steps to baseline loss: 12.11% fewer steps | Improved |
| adaptive batch size with matched-norm GNS | SGD · ImageWoof | validation loss: 1.4829 -> 1.3179 | Improved |
| adaptive batch size with matched-norm GNS | NormalizedSGD · ImageWoof | validation loss: 1.1979 -> 1.3286 | Hurt |
Closest papers by meaning
- A Second look at Exponential and Cosine Step Sizes: Simplicity, Adaptivity, and Performance (ICML 2021 · similarity 0.88)
- Stochastic Sign Descent Methods: New Algorithms and Better Theory (ICML 2021 · similarity 0.88)
- Efficient Sign-Based Optimization: Accelerating Convergence via Variance Reduction (NeurIPS 2024 · similarity 0.87)
- Sign-Based Optimizers Are Effective Under Heavy-Tailed Noise (NeurIPS 2026 · similarity 0.87)
- The Global Convergence Time of Stochastic Gradient Descent in Non-Convex Landscapes: Sharp Estimates via Large Deviations (ICML 2025 · similarity 0.87)
- When and Why SignSGD Outperforms SGD: A Theoretical Study Based on $\ell_1$-norm Lower Bounds (NeurIPS 2026 · similarity 0.87)
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.
