Prototypical Networks for Few-shot Learning

摘要: We propose Prototypical Networks for the problem of few-shot classification, where a classifier must generalize to new classes not seen in training set, given only small number examples each class. learn metric space which classification can be performed by computing distances prototype representations Compared recent approaches learning, they reflect simpler inductive bias that is beneficial this limited-data regime, and achieve excellent results. provide an analysis showing some simple design decisions yield substantial improvements over involving complicated architectural choices meta-learning. further extend zero-shot learning state-of-the-art results on CU-Birds dataset.