Demystifying Batch Normalization: Analysis of Normalizing Layer Inputs in Neural Networks

Abstract

Batch normalization was introduced as a novel solution to help with training fully-connected feed-forward deep neural networks. It proposes to normalize each training-batch in order to alleviate the problem caused by internal covariate shift. The original method claimed that Batch Normalization must be performed before the ReLu activation in the training process for optimal results. However, a second method has since gained ground which stresses the importance of performing BN after the ReLu activation in order to maximize performance. In fact, in the source code of PyTorch, common architectures such as VGG16, ResNet and DenseNet have Batch Normalization layer after the ReLU activation layer. Our work is the first to demystify the aforementioned debate and offer a comprehensive answer as to the proper order for Batch Normalization in the neural network training process. We demonstrate that for convolutional neural networks (CNNs) without skip connections, it is optimal to do ReLu activation before Batch Normalization as a result of higher gradient flow. In Residual Networks with skip connections, the order does not affect the performance or the gradient flow between the layers.