Weight Update Skipping: Reducing Training Time for Artificial Neural Networks

Abstract

Artificial Neural Networks (ANNs) are known as state-of-the-art techniques in Machine Learning (ML) and have achieved outstanding results in data-intensive applications, such as recognition, classification, and segmentation. These networks mostly use deep layers of convolution and/or fully connected layers with many filters in each layer, demanding a large amount of data and tunable hyperparameters to achieve competitive accuracy. As a result, storage, communication, and computational costs of training (in particular time spent for training) become limiting factors to scale them up. In this paper, we propose a new training methodology for ANNs that exploits the observation of improvement of accuracy shows temporal variations which allow us to skip updating weights when the variation is minuscule. During such time windows, we keep updating bias which ensures the network still trains and avoids overfitting; however, we selectively skip updating weights (and their time-consuming computations). This training approach virtually achieves the same accuracy with considerably less computational cost and reduces the time spent on training. We developed two variations of the proposed training method for selectively updating weights, and call them as i) Weight Update Skipping (WUS), and ii) Weight Update Skipping with Learning Rate Scheduler (WUS+LR). We evaluate these two approaches by analyzing state-of-the-art models, including AlexNet, VGG-11, VGG-16, ResNet-18 on CIFAR datasets. We also use ImageNet dataset for AlexNet, VGG-16, and Resnet-18. On average, WUS and WUS+LR reduced the training time (compared to the baseline) by 54%, and 50% on CPU and 22%, and 21% on GPU, respectively for CIFAR-10; and 43% and 35% on CPU and 22%, and 21% on GPU, respectively for CIFAR-100; and finally 30% and 27% for ImageNet, respectively.