Prototypical networks are a simple and effective approach for few-shot learning, where a classifier must generalize to new classes not seen in the training set with only a few examples per class. The method learns a metric space where classification is performed by computing distances to prototype representations of each class. These prototypes are computed as the mean of the embedded support examples for each class. The model uses a softmax over distances to class prototypes to classify query points. Prototypical networks are also extended to zero-shot learning, where class meta-data is used instead of examples. The approach achieves state-of-the-art results on benchmark tasks, including the CU-Birds dataset. The method is simpler and more efficient than recent meta-learning approaches, and it performs well with Euclidean distance, which is more effective than cosine similarity. The paper compares prototypical networks to matching networks and shows that simple design choices can yield substantial improvements. The model is trained using episodic learning, where each episode consists of a subset of classes and examples. The approach is effective for both few-shot and zero-shot learning, and it generalizes well to different distance metrics and training configurations. The results show that prototypical networks achieve high accuracy on several benchmark datasets, including Omniglot and miniImageNet, and outperform other methods in zero-shot learning. The paper also discusses the theoretical foundations of the method, relating it to clustering and mixture density estimation. Overall, prototypical networks provide a simple and effective solution for few-shot and zero-shot learning.Prototypical networks are a simple and effective approach for few-shot learning, where a classifier must generalize to new classes not seen in the training set with only a few examples per class. The method learns a metric space where classification is performed by computing distances to prototype representations of each class. These prototypes are computed as the mean of the embedded support examples for each class. The model uses a softmax over distances to class prototypes to classify query points. Prototypical networks are also extended to zero-shot learning, where class meta-data is used instead of examples. The approach achieves state-of-the-art results on benchmark tasks, including the CU-Birds dataset. The method is simpler and more efficient than recent meta-learning approaches, and it performs well with Euclidean distance, which is more effective than cosine similarity. The paper compares prototypical networks to matching networks and shows that simple design choices can yield substantial improvements. The model is trained using episodic learning, where each episode consists of a subset of classes and examples. The approach is effective for both few-shot and zero-shot learning, and it generalizes well to different distance metrics and training configurations. The results show that prototypical networks achieve high accuracy on several benchmark datasets, including Omniglot and miniImageNet, and outperform other methods in zero-shot learning. The paper also discusses the theoretical foundations of the method, relating it to clustering and mixture density estimation. Overall, prototypical networks provide a simple and effective solution for few-shot and zero-shot learning.