Faster SGD training by minibatch persistency

It is well known that, for most datasets, the use of large-size minibatches for Stochastic Gradient Descent (SGD) typically leads to slow convergence and poor generalization. On the other hand, large minibatches are of great practical interest as they allow for a better exploitation of modern GPUs. Previous literature on the subject concentrated on how to adjust the main SGD parameters (in particular, the learning rate) when using large minibatches. In this work we introduce an additional feature, that we call minibatch persistency, that consists in reusing the same minibatch for K consecutive SGD iterations. The computational conjecture here is that a large minibatch contains a significant sample of the training set, so one can afford to slightly overfitting it without worsening generalization too much. The approach is intended to speedup SGD convergence, and also has the advantage of reducing the overhead related to data loading on the internal GPU memory. We present computational results on CIFAR-10 with an AlexNet architecture, showing that even small persistency values (K=2 or 5) already lead to a significantly faster convergence and to a comparable (or even better) generalization than the standard "disposable minibatch" approach (K=1), in particular when large minibatches are used. The lesson learned is that minibatch persistency can be a simple yet effective way to deal with large minibatches.