After training a batch-normalized network, the learned scale ($\boldsymbol{\gamma}$) and shift ($\boldsymbol{\beta}$) parameters of each batch normalization layer can be inspected to verify that they have diverged significantly from their initial values of $1$ and $0$, respectively. For example, after training a batch-normalized LeNet on Fashion-MNIST for $10$ epochs, the first batch normalization layer—which normalizes the $6$-channel output of the first convolutional layer—exhibits $\boldsymbol{\gamma}$ values in the approximate range of $1.4$ to $2.1$ and $\boldsymbol{\beta}$ values spanning from roughly $-1.4$ to $1.3$. These non-trivial learned values confirm that the network exploits the affine transformation to recover representational capacity beyond the unit-variance, zero-mean normalization, adapting the output distribution of each channel to values that are most effective for the downstream activation function and subsequent layers.