Edit: I don't have time right now, but a toy example I like to throw at these kinds of problems is mapping primes to their indices (e.g. 2->0, 3->1, 5->2, ...). General-purpose learning algorithms can usually make a little headway with it, but not much, and only with substantial resources thrown at the problem. I'd be shocked if that toy example were any faster with their solution than a b-tree.
