Training Deep Neural Networks is complicated by the fact that the distribution of each layer’s inputs changes during training, as the parameters of the previous layers change. Sergey Ioffe, et al., "Batch Normalization: Accelerating Deep N…