[deepnet]

Togelius is pure mental (in a good OOTB way) his evo-devo HyperNEAT approach to problems intractable to random-walk SGD learning learning leaps over snares and pitfalls, bypasses cul-de-sacs and avoids oubliettes -traversing the whole search space by a sort of teleportation - a fast global search.

Backprop can then refine the most efficient net architectures. Curiously some evolve structures akin to LSTM units.