My understanding is that the neural network is used to predict the "magnitude" of each eigenvector/axe (extracted by using PCA) in order to reconstruct an approximation of the original behavior.