This paper introduces a novel semi-supervised few-shot learning framework that incorporates unlabeled examples within each episode. The authors propose extensions of Prototypical Networks (Snell et al., 2017) that can utilize unlabeled data to refine prototypes, improving classification performance. The framework is evaluated on modified versions of the Omniglot and miniImageNet benchmarks, as well as a new dataset called tieredImageNet, which includes a hierarchical structure of classes. The experiments show that the semi-supervised variants of Prototypical Networks outperform purely supervised models, especially when unlabeled examples are available. The authors also propose a new split of ImageNet for few-shot learning. The key contributions include the extension of Prototypical Networks to semi-supervised learning, the introduction of a new dataset, and the demonstration of improved performance through semi-supervised refinement. The paper highlights the importance of incorporating unlabeled data in few-shot learning to achieve better generalization and performance.This paper introduces a novel semi-supervised few-shot learning framework that incorporates unlabeled examples within each episode. The authors propose extensions of Prototypical Networks (Snell et al., 2017) that can utilize unlabeled data to refine prototypes, improving classification performance. The framework is evaluated on modified versions of the Omniglot and miniImageNet benchmarks, as well as a new dataset called tieredImageNet, which includes a hierarchical structure of classes. The experiments show that the semi-supervised variants of Prototypical Networks outperform purely supervised models, especially when unlabeled examples are available. The authors also propose a new split of ImageNet for few-shot learning. The key contributions include the extension of Prototypical Networks to semi-supervised learning, the introduction of a new dataset, and the demonstration of improved performance through semi-supervised refinement. The paper highlights the importance of incorporating unlabeled data in few-shot learning to achieve better generalization and performance.