In this study, we investigate the performance of various meta-learning algorithms in the context of
few-shot learning scenarios. Specifically, we evaluate Model-Agnostic Meta-Learning (MAML),
Prototypical Networks, and Matching Networks across three benchmark datasets: Mini-ImageNet,
Omniglot, and CIFAR-FS. The evaluation focuses on classification accuracy at 1-shot, 5-shot, and 10-
shot learning settings. Our results demonstrate that Prototypical Networks generally outperform both
MAML and Matching Networks, achieving the highest accuracy across most datasets and shot levels.
MAML shows strong adaptability with competitive performance but exhibits variability depending
on the dataset complexity. Matching Networks offer a balanced performance with effective memory
mechanisms and similarity functions. These findings underscore the strengths and limitations of each
algorithm and highlight the importance of choosing an appropriate meta-learning approach based on
task requirements and dataset characteristics.
