Research map / Optimization / Implicit bias and generalization
On the Generalization Benefit of Noise in Stochastic Gradient Descent
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 |
|---|---|---|---|
| Increase batch size from 2048 to 16384 | SGD with Momentum, 16-4 Wide-ResNet · CIFAR-10, constant step budget | test accuracy: 94.9 ± 0.1 -> 92.5 ± 0.6 | Hurt |
| Increase batch size from 2048 to 16384 | SGD with Momentum, 16-4 Wide-ResNet · CIFAR-10, constant step budget | final training loss: 0.058 ± 0.000 -> 0.019 ± 0.004 | Helped |
| Increase batch size from 2048 to 16384 | SGD with Momentum, 28-10 Wide-ResNet · CIFAR-100, constant step budget | test accuracy: 80.2 ± 0.2 -> 72.2 ± 0.2 | Hurt |
| Increase batch size from 256 to 4096 | Momentum, ResNet-50 · ImageNet | test accuracy: 77.0 -> 76.8 | Hurt |
| Use Momentum instead of vanilla SGD at batch size 4096 | ResNet-50 · ImageNet | test accuracy: 76.1 -> 76.8 | Helped |
| Increase batch size from 64 to 4096 | SGD with Momentum, fully connected autoencoder · MNIST | test MSE: 2.91 ± 0.01 -> 2.98 ± 0.01 | Hurt |
| Increase batch size from 16 to 256 | SGD with Momentum, word-level LSTM · Penn TreeBank | test perplexity: 88.06 ± 0.26 -> 92.19 ± 0.26 | Mixed |
| Maintain a high initial learning rate (temperature) and decay it later | SGD with Momentum, 16-4 Wide-ResNet · CIFAR-10, varying epoch budget | test accuracy | Helped |
Closest papers by meaning
- A Bayesian Perspective on Generalization and Stochastic Gradient Descent (ICLR 2018 · similarity 0.94)
- On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima (ICLR 2017 · similarity 0.93)
- Understanding the Generalization of Stochastic Gradient Adam in Learning Neural Networks (NeurIPS 2025 · similarity 0.91)
- Bias of Stochastic Gradient Descent or the Architecture: Disentangling the Effects of Overparameterization of Neural Networks (ICML 2024 · similarity 0.90)
- Subsampling is not Magic: Why Large Batch Sizes Work for Differentially Private Stochastic Optimisation (ICML 2024 · similarity 0.89)
- The Global Convergence Time of Stochastic Gradient Descent in Non-Convex Landscapes: Sharp Estimates via Large Deviations (ICML 2025 · similarity 0.89)
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.