An Extension of Prototypical Networks

Meta-learning is playing an increasingly important role in the field of machine learning and artificial intelligence. Few-shot learning, which only exploits a few data to learn different tasks, has been evolved greatly using meta-learning approaches like metric learning, optimization learning and etc. In this paper, we develop Prototypical Networks in two aspects: first one is that we propose a novel feature extractor architecture which not only being trained to learn useful feature but also learn to reduce useless feature. Secondly, we propose a regularization method that tends to increase entropy of output, improving flexibility under the paradigm of few-shot learning. Our experiments show that our extension of prototypical networks outperforms the vanilla version.