Constraint Satisfaction Problems can be modeled as Hopfield-type neural networks, such as Boltzmann machines (RBMs of deep learning are a variant) which can be

Constraint Satisfaction Problems can be modeled as Hopfield-type neural networks, such as Boltzmann machines (RBMs of deep learning are a variant) which can be tuned using simulated annealing. There was plenty of research on this around 1985-1995 (the "second coming" of neural networks), but it died out with the hype because it was never an actually practical way to solve CSPs. Given the recent innovations in deep learning, it should eventually pick up again, to enable deep learners to solve constraint satisfaction subproblems within the same integrated architecture.