On the Convergence Properties of a K-step Averaging Stochastic Gradient Descent Algorithm for Nonconvex Optimization

We adopt and analyze a synchronous K-step averaging stochastic gradient descent algorithm which we call K-AVG for solving large scale machine learning problems. We establish the convergence results of K-AVG for nonconvex objectives. Our analysis of K-AVG applies to many existing variants of synchronous SGD. We explain why the K-step delay is necessary and leads to better performance than traditional parallel stochastic gradient descent which is equivalent to K-AVG with $K=1$. We also show that K-AVG scales better with the number of learners than asynchronous stochastic gradient descent (ASGD). Another advantage of K-AVG over ASGD is that it allows larger stepsizes and facilitates faster convergence. On a cluster of $128$ GPUs, K-AVG is faster than ASGD implementations and achieves better accuracies and faster convergence for training with the CIFAR-10 dataset.