Research map / Optimization / Optimizers
Positive-Negative Momentum: Manipulating Stochastic Gradient Noise to Improve Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 conventional momentum with positive-negative momentum | SGD on CIFAR-10 and CIFAR-100 image-classification models · CIFAR-10 and CIFAR-100 | test error: 4.48_{0.09} vs 5.01_{0.03}; 6.26_{0.05} vs 6.42_{0.02}; 20.59_{0.29} vs 21.52_{0.37}; 19.76_{0.28} vs 19.81_{0.33}; 20.3 | Helped |
| Replace Adam with AdaPNM | Adaptive optimization on CIFAR-10 and CIFAR-100 · CIFAR-10 and CIFAR-100 | test error: 4.94_{0.05} vs 6.53_{0.03}; 5.99_{0.11} vs 7.31_{0.25}; 20.41_{0.18} vs 27.16_{0.55}; 20.68_{0.11} vs 25.11_{0.15}; 20.2 | Helped |
| Replace SGD with PNM | ResNet50 on ImageNet · ImageNet | Top-1 test error: 23.28 -> 23.21 | Helped |
| Replace SGD with PNM | ResNet50 on ImageNet · ImageNet | Top-5 test error: 6.75 -> 6.73 | Helped |
| Replace AdamW with AdaPNM | ResNet50 on ImageNet · ImageNet | Top-1 test error: 23.62 -> 23.12 | Helped |
| Replace AdamW with AdaPNM | ResNet50 on ImageNet · ImageNet | Top-5 test error: 7.09 -> 6.82 | Helped |
| Replace Adam with AdaPNM | LSTM language model · Penn TreeBank | test perplexity: 66.67 -> 64.25 | Helped |
| Use positive-negative pairs with beta_0 > 0 rather than beta_0 in [-1, 0] (ablation) | PNM on ResNet18 · CIFAR-10 | test error | Helped |
| Vary beta_0 within [-1, 0], which does not enhance SGN (ablation) | PNM on ResNet18 · CIFAR-10 | test error | No clear effect |
| Use decoupled weight decay instead of L2 regularization (ablation) | PNM · CIFAR-10 and CIFAR-100 | test error | Helped |
| Use cosine annealing or warm-restart learning-rate schedules (ablation) | PNM and AdaPNM · CIFAR-10 and CIFAR-100 | test error | Helped |
| Use PNM instead of SGD with momentum under label noise | ResNet34 · CIFAR-10 with 40% label noise | test error: more 20 points | Helped |
Closest papers by meaning
- The Marginal Value of Momentum for Small Learning Rate SGD (ICLR 2024 · similarity 0.90)
- Towards understanding how momentum improves generalization in deep learning (ICML 2022 · similarity 0.89)
- Perfect Parallelization in Mini-Batch SGD with Classical Momentum Acceleration (NeurIPS 2026 · similarity 0.88)
- Accelerated Convergence of Stochastic Heavy Ball Method under Anisotropic Gradient Noise (ICLR 2024 · similarity 0.88)
- Random Scaling and Momentum for Non-smooth Non-convex Optimization (ICML 2024 · similarity 0.88)
- Gradient correlation is a key ingredient to accelerate SGD with momentum (ICLR 2025 · 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.