Learning Algorithms for Active Learning

Philip Bachman, Alessandro Sordoni, Adam Trischler

Abstract: We present a model that learns active learning algorithms via metalearning. For each metatask, our model jointly learns: a data representation, an item selection heuristic, and a one-shot classifier. Our model uses the item selection heuristic to construct a labeled support set for the one-shot classifier. Using metatasks based on the Omniglot and MovieLens datasets, we show that our model performs well in synthetic and practical settings.

TL;DR: We present a model and experiments for meta active learning.

Keywords: Deep learning, Supervised Learning

Conflicts: maluuba.com, microsoft.com