Recurrent networks (e.g. Hopfield networks and Boltzmann machines) can have arbitrary topologies.
Can model systems with internal state based on node activation levels as well as weights.
Training is more difficult .
Performance can be problematic:
