Research map / Optimization / Optimizers
Scalable and Practical Natural Gradient for Large-Scale Deep Learning
Kazuki Osawa, Yohei Tsuji, Yuichiro Ueno, Akira Naruse, Chuan-Sheng Foo, Rio Yokota
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 stale statistics (adaptive update frequency) (ablation) | SP-NGD emp+unitBN, BS=4K · ImageNet ResNet-50 | speedup / top-1: 75.2 % ->74.8 %, x 1.33 speedup, 23.6 % communication | Mixed |
| add stale statistics (ablation) | SP-NGD, BS=16K · ImageNet ResNet-50 | speedup: x 1.68 speedup, 5.4 % communication, 75.3 % ->75.2 % | Improved |
| add stale statistics (ablation) | SP-NGD, BS=32K · ImageNet ResNet-50 | speedup: x 1.40 speedup, 7.8 % communication, 75.6 % ->75.4 % | Improved |
| use empirical Fisher instead of single Monte Carlo Fisher estimate (ablation) | SP-NGD · ImageNet ResNet-50 | time per step | Improved |
| use unit-wise NGD on BatchNorm instead of full BN Fisher (ablation) | SP-NGD · ImageNet ResNet-50 | time per step | Improved |
| increase batch size from 32K to 131K | SP-NGD · ImageNet ResNet-50 | top-1 accuracy: 75.4 % ->74.9 % | Mixed |
| use second-order NGD instead of SGD for large batch | SP-NGD vs SGD · ImageNet ResNet-50 BS=32K | epochs: 45 epochs vs 90 epochs | Improved |
Closest papers by meaning
- On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima (ICLR 2017 · similarity 0.90)
- A Layer-Wise Natural Gradient Optimizer for Training Deep Neural Networks (NeurIPS 2024 · similarity 0.89)
- Diving into the shallows: a computational perspective on large-scale shallow learning (NeurIPS 2017 · similarity 0.89)
- Don't Use Large Mini-batches, Use Local SGD (ICLR 2020 · similarity 0.89)
- Neumann Optimizer: A Practical Optimization Algorithm for Deep Neural Networks (ICLR 2018 · similarity 0.89)
- MARS: Unleashing the Power of Variance Reduction for Training Large Models (ICML 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.
