It would be super interesting to see if this knowledge generalizes with transfer learning. For example, after seeing 50,000 episodes of PacMan, would the GAN be able to recreate Space Invaders with just 5,000 extra episodes?