Research map / Optimization / Optimizers
AdamP: Slowing Down the Slowdown for Momentum Optimizers on Scale-invariant Weights
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 |
|---|---|---|---|
| Project the radial component out of momentum updates | SGD/Adam on ImageNet classification · ImageNet-1K | accuracy: +0.09 to +1.24 | Helped |
| Project the radial component out of Adam updates | CenterNet and SSD · MS-COCO | AP: +0.28 to +0.87 | Helped |
| Project the radial component out of Adam updates | Harmonic CNN · Audio classification benchmarks | accuracy / AUC / F1: +0.18 to +0.80 | Helped |
| Project the radial component out of Adam updates | Transformer-XL without WN · WikiText-103 | perplexity: 23.38 -> 23.26 | Helped |
| Add weight normalization, introducing scale invariance | Transformer-XL with Adam · WikiText-103 | perplexity: 23.33 -> 23.90 | Hurt |
| Apply AdamP to Transformer-XL with weight normalization | Transformer-XL + WN · WikiText-103 | perplexity: 23.90 -> 22.73 | Helped |
| Project the radial component out of Adam updates | ResNet-50 retrieval model · CUB, Cars-196, In-Shop, SOP | Recall@1: +0.0 to +2.2 | Mixed |
| Project the radial component out of Adam updates | Wide-ResNet adversarial training · CIFAR-10 | attacked accuracy: +2.94 to +13.26 | Helped |
| Project the radial component out of Adam updates | ReBias models · Biased MNIST and 9-Class ImageNet | accuracy: +1.4 to +7.9 | Helped |
| Project the radial component out of Adam updates | ResNet50 · ImageNet large-batch training | accuracy: +1.34 to +3.1 | Helped |
| Vary the detection threshold delta (ablation) | AdamP on ResNet18 · ImageNet-1K | accuracy: 70.66 to 70.83 | No clear effect |
| Use weight decay to regularize norm growth instead of projection | momentum GD · 3D scale-invariant Rosenbrock | convergence: r=1.0 -> 2.53 | Mixed |
Closest papers by meaning
- Towards understanding how momentum improves generalization in deep learning (ICML 2022 · similarity 0.89)
- Scaling Exponents Across Parameterizations and Optimizers (ICML 2024 · similarity 0.87)
- Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size (NeurIPS 2025 · similarity 0.87)
- A geometric framework for momentum-based optimizers for low-rank training (NeurIPS 2025 · similarity 0.87)
- The Marginal Value of Momentum for Small Learning Rate SGD (ICLR 2024 · similarity 0.87)
- Surge Phenomenon in Optimal Learning Rate and Batch Size Scaling (NeurIPS 2024 · similarity 0.86)
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.