AMSGrad is a popular adaptive gradient based optimization algorithm that is widely used in training deep neural networks. The new variant assumes that minibatch gradients in consecutive iterations have some underlying structure, which makes the gradients sequentially predictable. By exploiting the predictability and some ideas from Optimistic Online learning, the proposed algorithm can accelerate the convergence and also enjoys a tighter regret bound. We evaluate Optimistic-AMSGrad and AMSGrad in terms of various performance measures (i.e., training loss, testing loss, and classification accuracy on training/testing data), which demonstrate that Optimistic-AMSGrad improves AMSGrad.
View on arXiv