- TL;DR: Bregman divergence learning for few-shot learning.
- Abstract: In this work, we approach one-shot and few-shot learning problems as methods for finding good prototypes for each class, where these prototypes are generalizable to new data samples and classes. We propose a metric learner that learns a Bregman divergence by learning its underlying convex function. Bregman divergences are a good candidate for this framework given they are the only class of divergences with the property that the best representative of a set of points is given by its mean. We propose a flexible extension to prototypical networks to enable joint learning of the embedding and the divergence, while preserving computational efficiency. Our preliminary results are comparable with the prior work on the Omniglot and Mini-ImageNet datasets, two standard benchmarks for one-shot and few-shot learning. We argue that our model can be used for other tasks that involve metric learning or tasks that require approximate convexity such as structured prediction and data completion.