Recent studies have revealed that Convolutional Neural Networks requiring vastly many sum-of-product operations with relatively small numbers of parameters tend to exhibit great model performances. Asynchronous Stochastic Gradient Descent provides a possibility of large-scale distributed computation for training such networks. However, asynchrony introduces stale gradients, which are considered to have negative effects on training speed. In this work, we propose a method to predict future parameters during the training to mitigate the drawback of staleness. We show that the proposed method gives good parameter prediction accuracies that can improve speed of asynchronous training. The experimental results on ImageNet demonstrates that the proposed asynchronous training method, compared to a synchronous training method, reduces the training time to reach a certain model accuracy by a factor of 1.9 with 256 GPUs used in parallel.
展开▼
