It would be interesting to see an analysis of the models inferred by this method, to see what kinds of approximations they're learning.