[Almaviva]

We know that AlphaGo uses a Monte Carlo tree search, and presumably contains innovations and refinements to techniques applicable to turn based board games with perfect information and a clear binary win condition.

We also know it uses a Deep Learning algorithm to imitate play from the best humans. And presumably, being stronger than any human now, it could also recursively refine and train on its own games.

Do we know how important the Deep Learning component is? Would AlphaGo be just as strong or nearly so without that part?

I know that over the last few years chess engines have become dramatically more powerful, even on the same hardware, to the tune of hundreds of Elo points. They still stick to the classical techniques though (Alpha-Beta pruning, minimax is still the core). This is not to understate the refinements, but the general ideas are (I think?) only applicable to turn based games with finite moves and a binary mathematical criterion for victory.

Is it possible that the improvements AlphaGo has made are mostly of this type, and the Deep Learning parts are not that important to its actual play strength? Or can we totally rule this out? Is there an in depth discussion of this question somewhere?