The paper introduces the Relation Network (RN), a flexible and general framework for few-shot learning, which aims to recognize new classes with only a few examples. The RN is trained end-to-end from scratch, learning a deep distance metric to compare a small number of images within episodes, each designed to simulate the few-shot setting. Once trained, the RN can classify images of new classes by computing relation scores between query images and the few examples of each new class without further updating the network. The framework is also easily extended to zero-shot learning. Extensive experiments on five benchmarks demonstrate that the RN provides improved performance on both few-shot and zero-shot learning tasks, outperforming existing methods with simpler and faster approaches. The RN's effectiveness is attributed to its ability to learn a deep non-linear metric for comparing images, which is jointly tuned with the embedding network, allowing for better identification of matching and mismatching pairs.The paper introduces the Relation Network (RN), a flexible and general framework for few-shot learning, which aims to recognize new classes with only a few examples. The RN is trained end-to-end from scratch, learning a deep distance metric to compare a small number of images within episodes, each designed to simulate the few-shot setting. Once trained, the RN can classify images of new classes by computing relation scores between query images and the few examples of each new class without further updating the network. The framework is also easily extended to zero-shot learning. Extensive experiments on five benchmarks demonstrate that the RN provides improved performance on both few-shot and zero-shot learning tasks, outperforming existing methods with simpler and faster approaches. The RN's effectiveness is attributed to its ability to learn a deep non-linear metric for comparing images, which is jointly tuned with the embedding network, allowing for better identification of matching and mismatching pairs.