Large Learning Rates Improve Generalization: But How Large Are We Talking About?
This work provides practical guidance for tuning learning rates in neural network training, though it is incremental in nature.
The study investigated the optimal initial learning rate ranges for neural network training to improve generalization, finding that these ranges are significantly narrower than previously assumed.
Inspired by recent research that recommends starting neural networks training with large learning rates (LRs) to achieve the best generalization, we explore this hypothesis in detail. Our study clarifies the initial LR ranges that provide optimal results for subsequent training with a small LR or weight averaging. We find that these ranges are in fact significantly narrower than generally assumed. We conduct our main experiments in a simplified setup that allows precise control of the learning rate hyperparameter and validate our key findings in a more practical setting.