Prototypical Siamese Networks for Few-shot Learning

Abstract

We propose a novel architecture, called Prototypical Siamese Networks, for few-shot learning, where a classifier must generalize to new classes not seen in the training set, given only a few examples of each class. Prototypical Siamese Networks add a new module to siamese networks to learn a high quality prototypical representation of each class. Compared to recent methods for few-shot learning, our method achieves state-of-the-art performance on few-shot learning. Experiments on two benchmarks validate the effectiveness of the proposed method.