In step 1, we create a fully connected layer that returns a two-dimensional tensor. The first dimension (?) represents any number of (input) images and the second dimension represents the number of output neurons (here, 10 class labels). In step 2, we provide the option for dropout primarily used during the training of the network.