We propose prototypical networks for the problem of few-shot classification,where a classifier must generalize to new classes not seen in the training set,given only a small number of examples of each new class. Prototypical networkslearn a metric space in which classification can be performed by computingdistances to prototype representations of each class. Compared to recentapproaches for few-shot learning, they reflect a simpler inductive bias that isbeneficial in this limited-data regime, and achieve excellent results. Weprovide an analysis showing that some simple design decisions can yieldsubstantial improvements over recent approaches involving complicatedarchitectural choices and meta-learning. We further extend prototypicalnetworks to zero-shot learning and achieve state-of-the-art results on theCU-Birds dataset.