[1502.03167] Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

http://arxiv.org/abs/1502.03167

Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch.