This paper proposes a graph neural network (GNN) approach for few-shot learning, semi-supervised learning, and active learning. The key idea is to model the few-shot learning problem as a supervised message-passing task on a graph, where nodes represent images and edges represent learned similarity between them. The GNN architecture generalizes several recent few-shot learning models and is trained end-to-end to capture task-specific invariances, such as permutations within input collections. The model is shown to achieve state-of-the-art performance on the Omniglot and Mini-Imagenet datasets with significantly fewer parameters than existing methods. It is also easily extended to semi-supervised and active learning settings, where the model can leverage unlabeled data or actively request labels to improve performance. The GNN is trained using a combination of graph convolutional layers and learned adjacency matrices, allowing it to propagate label information from labeled samples to unlabeled query images. The model is evaluated on various tasks, including few-shot classification, semi-supervised learning, and active learning, demonstrating its effectiveness in handling relational tasks. The results show that the GNN outperforms existing methods in terms of accuracy and efficiency, while maintaining a simple and generalizable architecture. The paper also discusses related work in graph-based learning and provides a detailed description of the model's architecture and training process.This paper proposes a graph neural network (GNN) approach for few-shot learning, semi-supervised learning, and active learning. The key idea is to model the few-shot learning problem as a supervised message-passing task on a graph, where nodes represent images and edges represent learned similarity between them. The GNN architecture generalizes several recent few-shot learning models and is trained end-to-end to capture task-specific invariances, such as permutations within input collections. The model is shown to achieve state-of-the-art performance on the Omniglot and Mini-Imagenet datasets with significantly fewer parameters than existing methods. It is also easily extended to semi-supervised and active learning settings, where the model can leverage unlabeled data or actively request labels to improve performance. The GNN is trained using a combination of graph convolutional layers and learned adjacency matrices, allowing it to propagate label information from labeled samples to unlabeled query images. The model is evaluated on various tasks, including few-shot classification, semi-supervised learning, and active learning, demonstrating its effectiveness in handling relational tasks. The results show that the GNN outperforms existing methods in terms of accuracy and efficiency, while maintaining a simple and generalizable architecture. The paper also discusses related work in graph-based learning and provides a detailed description of the model's architecture and training process.