Extended Prototypical Network for Few-Shot Learning

This paper proposes a learning method based on extended distance for unsupervised meta-learning. Compared with previous algorithms, the network can also learn new classes and correctly classify them, and each class only requires a few shots for training. Among them, extension distance emerged as a new distance measurement method, which has obvious advantages over the previous Euclidean distance. This paper compares it with the prototypical network based on Euclidean distance and compares and analyzes the experimental results, the experiment that works best is the Miniimagenet dataset which improves accuracy by about 2.36%. The meta-learning method based on metric space is further explored, and experiments are carried out on three data sets and achieved good experimental results on MNIST, miniimagenet, and omniglot data sets.