Batch Normalization refers to a neural network layer which normalizes (i.e. sets the mean to 0 and the standard deviation to 1) the values of the data it receives, across the batch and spatial dimensions (but not the channel dimension), then applies scaling and shift operations (both by learnable tensors). Example: in a convolutional neural network for 2D images of 3 channels (say red, green, blue), a batch norm (BN) layer placed immediately after the input layer would compute one pair of average and standard deviation values per channel, regardless of the size of the batch.