No More Pesky Learning Rate Guessing Games.
Leslie N. Smith · arXiv (Cornell University) · 2015
It is known that the learning rate is the most important hyper-parameter to tune for training deep convolutional neural networks (i.e., a guessing game). This report describes a new method for setting the learning rate, named cyclical learning rates, that eliminates the need to experimentally find the best values and schedule for the learning rates. Instead of setting the learning rate to fixed values, this method lets the learning rate cyclically vary within boundary values. This report shows that training with cyclical learning rates achieves near optimal classification accuracy without tuning and often in many fewer iterations. This report also describes a simple way to estimate reasonable bounds - by linearly increasing the learning rate in one training run of the network for only a few epochs. In addition, cyclical learning rates are demonstrated on training with the CIFAR-10 dataset and the AlexNet and GoogLeNet architectures on the ImageNet dataset. These methods are practical tools for everyone who trains convolutional neural networks.