Adaptive Multi-prototype Relation Network

Xiaoxu Li, Tao Tian, Yuxin Liu, Hong Yu, Jie Cao, Zhanyu Ma

Abstract: Based on Relation Network, we propose a new network structure that can adaptively adjust the number of prototypes according to data distribution. Our method, called the Adaptive Multi-prototype Relation Network(AMRN), aims at extracting more reasonable prototype representation for different data distribution in few-shot learning case. Instead of representing each class as a single prototype in the relational network, we represent each class with one or more prototypes, and solve the problem of embedding network with the relational network connection, which can improving the classification accuracy in few-shot learning. Besides, our method can easily extend to other network structures, which is also a useful reference for other metric learning approaches.