Active learning using meta-learned priors

Abstract

Deep learning models have found enormous success across a variety of displines, but training the parameters within these models generally requires huge amounts of labelled data. One technique to reduce the burden of this data requirement is meta-learning, which involves extracting knowledge from previously experienced tasks so we can train new tasks with less data. Another such technique is active learning, where we selectively query labels for examples to train on, which potentially allows us to train a new task with fewer labelled data points. We investigate the potential to use active learning in the meta-learning setting, combining the advantages of the two techniques and further reducing the requirements for labelled data. In particular, we consider several techniques for actively querying labels within the modular meta-learning framework. We apply these techniques to several empirical settings, finding significant advantages over a baseline of random queries.