Learning to Compare: Relation Network for Few-Shot Learning

This paper presents the Relation Network (RN), described as a conceptually simple, flexible, and general framework for few-shot learning, where a classifier must recognize new classes given only a few examples from each. The network is trained end-to-end from scratch. Through meta-learning, it learns to learn a deep distance metric that compares a small number of images within episodes, each designed to simulate the few-shot setting.

Once trained, a Relation Network classifies images of new classes by computing relation scores between query images and the few examples of each new class, without any further updating of the network. Besides improving few-shot performance, the framework extends easily to zero-shot learning, and extensive experiments on five benchmarks show that this single, simple approach is unified and effective for both tasks.