Prototypical Networks for Few-shot Learning

Authors: Snell et al.

Date: 2017

paper

Few-shot learning: prediction setting where a classifier is asked make predictions on classes not seen or seen a few times in training.

Related work: Matching newtorks in whcih learned embeddings are matched using attention between labeled and unlabeled points. paper

This work proposes the inductive bias that there exists a single representation (prototype) for all the points of a given class, and that a new point can be classified by matching a new point to the closest prototype. In the few-shot setting, several points (support set) of a given class are known, and in a zero-shot no examples are known, but some ‘meta data’ about the class is used instead.

Embeddings are optimized to maximize the probability of the true class given the embeddings. This probability is computed as a softmax over the distances from a given point $x$ and the prototype vectors of each class. In turn, the prototype vectors are a simple aggregation of all the points of the same class.