Meta-Learning based Prototype-relation Network for Few-shot Classification

Pattern recognition has made great progress under large amount of labeled data, while performs poorly on a very few examples obtained, named few-shot classification, where a classifier can identify new classes not encountered during training. In this paper, a simple framework named Prototype-Relation Network is presented for the few-shot classification. Moreover, a novel loss function compared with pro- totype networks is proposed which takes both inter-class and intra-class distance into account. During meta-learning, the model is optimized by end-to-end episodes, each of which is to imitate the test few- shot setting. The trained model is used to classify new classes by computing min distance between query images and the prototype of each class. Extensive experimental results demonstrate that our proposed meta-learning model is competitive and effective, which achieves the state-of-the-art performance on Omniglot and mini ImageNet datasets.

Recommended citation: X. Liu, F. Zhou, J. Liu, and L. Jiang. "Meta-Learning based Prototype-relation Network for Few-shot Classification." Neurocomputing. vol. 383, pp. 224-234, 2020.
Download Paper