Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

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. This slows down the trai