Transfer Learning through Greedy Subset Selection

In this work we study the binary transfer learning problem involving $10^2$ - $10^3$ sources. We focus on how to select sources from the large pool and how to combine them to yield a good performance on a target task. In particular, we consider the transfer learning setting where one does not have direct access to the source data, but rather employs the source hypotheses trained from them. The main result of this work is an efficient greedy algorithm that selects relevant source hypotheses and feature dimensions simultaneously. We show theoretically that, under reasonable assumptions on the source hypotheses, the algorithm can learn effectively from few examples. This claim is verified by state-of-the-art results involving up to 1000 classes, totalling 1.2 million examples, with the only 11 to 20 training examples from the target domain.