Research map / Optimization / Optimizers
Extrapolation for Large-batch Training in Deep Learning
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 |
|---|---|---|---|
| Use past local mini-batch gradients for extrapolation instead of mini-batch SGD | ResNet-20 · CIFAR-10 | test top-1 accuracy: 91.36 ± 0.19 -> 91.72 ± 0.11 | Helped |
| Use past local mini-batch gradients for extrapolation instead of mini-batch SGD | VGG-11 · CIFAR-10 | test top-1 accuracy: 86.64 ± 0.10 -> 87.00 ± 0.26 | Helped |
| Add extrapolation to mini-batch SGD without LARS (ablation) | ResNet-20 · CIFAR-10 | test top-1 accuracy: 90.00 ± 0.48 -> 90.47 ± 0.16 | Helped |
| Add extrapolation to mini-batch SGD without LARS (ablation) | VGG-11 · CIFAR-10 | test top-1 accuracy: 73.09 ± 9.35 -> 76.79 ± 3.5 | Helped |
| Use uniform noise for extrapolation instead of past local gradients (ablation) | extrapolated SGD · CIFAR-10, ResNet-20 | test top-1 accuracy: 91.53 ± 0.25 -> 91.72 ± 0.11 | Hurt |
| Use uniform noise for extrapolation instead of past local gradients (ablation) | extrapolated SGD · CIFAR-10, VGG-11 | test top-1 accuracy: 87.00 ± 0.31 -> 87.00 ± 0.26 | No clear effect |
| Use stochastic noise for extrapolation instead of past local gradients (ablation) | extrapolated SGD · CIFAR-10, ResNet-20 | test top-1 accuracy: 91.66 ± 0.24 -> 91.72 ± 0.11 | No clear effect |
| Use stochastic noise for extrapolation instead of past local gradients (ablation) | extrapolated SGD · CIFAR-10, VGG-11 | test top-1 accuracy: 86.04 ± 0.43 -> 87.00 ± 0.26 | Helped |
| Apply extrap-SGD in the early phase, then switch to post-local SGD (ablation) | ResNet-20 · CIFAR-10, global mini-batch size 16,384 | test top-1 accuracy | Helped |
| Increase the number of workers while keeping the global mini-batch size fixed (ablation) | extrap-SGD with ResNet-20 · CIFAR-10, global mini-batch size 8,192 | test top-1 accuracy | Helped |
| Remove momentum while retaining extrapolation (ablation) | extrap-SGD · CIFAR-10, ResNet-20 | optimization performance | Helped |
| Tune the momentum factor (ablation) | mini-batch SGD · CIFAR-10, ResNet-20 | optimization performance | Helped |
Closest papers by meaning
- On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima (ICLR 2017 · similarity 0.93)
- On the Generalization Benefit of Noise in Stochastic Gradient Descent (ICML 2020 · similarity 0.90)
- Train longer, generalize better: closing the generalization gap in large batch training of neural networks (NeurIPS 2017 · similarity 0.90)
- Augment Your Batch: Improving Generalization Through Instance Repetition (CVPR 2020 · similarity 0.89)
- Exact, Tractable Gauss-Newton Optimization in Deep Reversible Architectures Reveal Poor Generalization (NeurIPS 2024 · similarity 0.87)
- Optimal Rates for Generalization of Gradient Descent for Deep ReLU Classification (NeurIPS 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.