A Definition of a batch normalization layer When applying batch normalization to convolutional layers, the inputs and outputs of normalization layers are 4-dimensional tensors, which we denote by I