In step 1, hidden layers are created using DenseLayer, which are preceded by convolution/subsampling layers.

In step 2, note that we didn't mention the number of input neurons in hidden layers, since it would be same as the preceding layer's (SubSamplingLayer) outgoing neurons.