This paper introduces a Relation Network (RN) for few-shot learning, which is a simple and effective framework for learning to recognize new classes with only a few examples. The RN is trained end-to-end and can classify new classes by computing relation scores between query images and few examples without further network updates. It is also easily extended to zero-shot learning. The framework is evaluated on five benchmarks and shows improved performance on both few-shot and zero-shot learning tasks. The RN consists of two modules: an embedding module that generates image representations and a relation module that computes similarity scores between images. The model is trained using an episode-based strategy, where each episode consists of a set of classes with a few examples each. The relation module learns to compare query images with few examples to determine if they belong to the same class. The RN outperforms previous approaches in terms of simplicity and speed, and is able to generalize to zero-shot learning by modifying the sample branch to input category descriptions instead of images. The model is implemented using convolutional neural networks and is trained end-to-end. The results show that the RN achieves state-of-the-art performance on both few-shot and zero-shot learning tasks.This paper introduces a Relation Network (RN) for few-shot learning, which is a simple and effective framework for learning to recognize new classes with only a few examples. The RN is trained end-to-end and can classify new classes by computing relation scores between query images and few examples without further network updates. It is also easily extended to zero-shot learning. The framework is evaluated on five benchmarks and shows improved performance on both few-shot and zero-shot learning tasks. The RN consists of two modules: an embedding module that generates image representations and a relation module that computes similarity scores between images. The model is trained using an episode-based strategy, where each episode consists of a set of classes with a few examples each. The relation module learns to compare query images with few examples to determine if they belong to the same class. The RN outperforms previous approaches in terms of simplicity and speed, and is able to generalize to zero-shot learning by modifying the sample branch to input category descriptions instead of images. The model is implemented using convolutional neural networks and is trained end-to-end. The results show that the RN achieves state-of-the-art performance on both few-shot and zero-shot learning tasks.