Sorry for the delay of this answer.

I'm still at the very beginning with this project and have no interesting results yet :(

There is a model relationship between artificial neural networks and sets of sentences in a formal language. One idea is to track sets of sentences during the training of a network and see if something interesting can be found.