A nonmonotone learning rate strategy for SGD training of deep neural networks | IEEE Conference Publication | IEEE Xplore