The paper introduces Matching Networks (MN), a novel neural architecture designed for one-shot learning, which maps a small labeled support set and an unlabelled example to its label without fine-tuning. MN combines ideas from metric learning and external memory augmented neural networks. The model is trained to match test and training conditions, enabling rapid learning from few examples. Experiments on ImageNet and Omniglot datasets show significant improvements over competing approaches, achieving 93.2% accuracy on ImageNet and 93.8% on Omniglot. The paper also defines one-shot language modeling tasks on the Penn Treebank, demonstrating the model's versatility. The key contributions include a non-parametric approach that allows for rapid adaptation to new classes and a training strategy tailored for one-shot learning.The paper introduces Matching Networks (MN), a novel neural architecture designed for one-shot learning, which maps a small labeled support set and an unlabelled example to its label without fine-tuning. MN combines ideas from metric learning and external memory augmented neural networks. The model is trained to match test and training conditions, enabling rapid learning from few examples. Experiments on ImageNet and Omniglot datasets show significant improvements over competing approaches, achieving 93.2% accuracy on ImageNet and 93.8% on Omniglot. The paper also defines one-shot language modeling tasks on the Penn Treebank, demonstrating the model's versatility. The key contributions include a non-parametric approach that allows for rapid adaptation to new classes and a training strategy tailored for one-shot learning.