Multi-Class Few Shot Learning Task and Controllable Environment

Dmitriy Serdyuk, Negar Rostamzadeh, Pedro Oliveira Pinheiro, Boris Oreshkin, Yoshua Bengio

Keywords: few-shot, few shot, meta-learning, metalearning

TL;DR: We introduce a diagnostic task which is a variation of few-shot learning and introduce a dataset for it.

Abstract: Deep learning approaches usually require a large amount of labeled data to generalize. However, humans can learn a new concept only by a few samples. One of the high cogntition human capablities is to learn several concepts at the same time. In this paper, we address the task of classifying multiple objects by seeing only a few samples from each category. To the best of authors' knowledge, there is no dataset specially designed for few-shot multiclass classification. We design a task of mutli-object few class classification and an environment for easy creating controllable datasets for this task. We demonstrate that the proposed dataset is sound using a method which is an extension of prototypical networks.