Fast Training of Convolutional Neural Networks via Kernel Rescaling
This addresses the time-consuming training process for deep learning practitioners, though it is incremental as it builds on existing architectures.
The authors tackled the problem of long training times for deep Convolutional Neural Networks by proposing a method that uses lower-resolution kernels and images initially, then refines at full resolution, achieving a 20% reduction in training time without accuracy loss on ImageNet and ResNet.
Training deep Convolutional Neural Networks (CNN) is a time consuming task that may take weeks to complete. In this article we propose a novel, theoretically founded method for reducing CNN training time without incurring any loss in accuracy. The basic idea is to begin training with a pre-train network using lower-resolution kernels and input images, and then refine the results at the full resolution by exploiting the spatial scaling property of convolutions. We apply our method to the ImageNet winner OverFeat and to the more recent ResNet architecture and show a reduction in training time of nearly 20% while test set accuracy is preserved in both cases.