Research map / Optimization / Implicit bias and generalization
Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training
Diego Granziol, Stefan Zohren, Stephen Roberts
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 |
|---|---|---|---|
| scale initial learning rate linearly with batch size (alpha_0 = 0.01B/128) | VGG-16 SGD without BN or weight decay · CIFAR-100 | validation accuracy: small drop at B=1024, larger drop at B=2048, no training at B=4096 | Mixed |
| double the maximal allowed learning rate (ablation) | VGG-16 SGD · CIFAR-100 | training stability | Hurt |
| use a mini-batch Hessian (B=128) instead of the full Hessian (ablation) | VGG-16 Hessian spectrum · CIFAR-100 | extremal eigenvalue magnitude | Hurt |
| use multiplicative noise model instead of additive noise for predicting batch eigenvalues (ablation) | Perturbation theory prediction · VGG-16 CIFAR-100 | fit to batch Hessian max eigenvalue | Improved |
| use larger initial learning rate at fixed B=1024 (ablation) | VGG-16 SGD · CIFAR-100 | validation/test accuracy | Improved |
Closest papers by meaning
- On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima (ICLR 2017 · similarity 0.90)
- Deep Linear Network Training Dynamics from Random Initialization: Data, Width, Depth, and Hyperparameter Transfer (ICML 2025 · similarity 0.90)
- On the Generalization Benefit of Noise in Stochastic Gradient Descent (ICML 2020 · similarity 0.90)
- Hessian-based Analysis of Large Batch Training and Robustness to Adversaries (NeurIPS 2018 · similarity 0.90)
- Understanding the Generalization of Stochastic Gradient Adam in Learning Neural Networks (NeurIPS 2025 · similarity 0.89)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.88)
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.
