Those silly Harvard and Berkeley AI people have stumbled upon a cool method in AI. Augmenting extremely limited data sets with inexhaustible generated data from simple rules enables much better prediction. This technique is broadly applicable to many areas of machine learning (ML) where we have insufficient labels. Variations of this concept are applied in computer vision and robotics. It's not clear if they will work with human behavior data but I shall endeavor to find out in my "day job."
No comments:
Post a Comment