If freely wired neutral networks are DAGs, I wonder how cyclic graphs of neurons behave, or even if that model would be theoretically meaningful, or computationally feasible.
You generally run them until they settle down. You can train them, in a difficult process, by using the idea of thermal equilibrium.