The author of the winning bot (tscmoo) is currently experimenting with training NNs on data acquired by parsing replays of pro-level games. He's been mostly focusing on automatic learning of build orders (BOs) in specific match ups.