How can an "architectural motif" be extracted from a trained MLP?

Question

I am trying to reproduce the paper Synthetic Petri Dish: A novel surrogate model for Rapid Architecture Search. In the paper, the authors try to reduce the architecture of an MLP model trained on MNIST (2 layers - 100 neurons) by initializing a motif network from it, that is, 2 layers, 1 neuron each, and extracting the sigmoid function. I have been searching a lot, but I have not found the answer of how can someone extract an 'architectural motif' from a trained neural network.