[coolness]

Looks great. Does this need a bunch of unlabeled training data to begin with, and then suggests datapoints to label, or can it also suggest labels for which data is needed next?