Domain adaptation -- can quantity compensate for quality? (Q2248544)

In this paper, the authors try to respond to the question: ``under which circumstances large sizes of not-perfectly-representative training samples can guarantee that the learned classifier performs just as well as one learned from target generated samples?'' Their answer is a positive one, particularly for a Nearest Neighbor algorithm, under specific assumptions. Furthermore, they show that, when the output classifier has to come from a predefined class, any learner needs access to data generated from the target distribution.