Abstract

Artificial Intelligence, most notably through deep learning, has made tremendous advances in the last few years, achieving ever-higher performances on benchmark data sets and outperforming humans on complex tasks such as image classification on certain data sets. However, these performances heavily rely on large data sets for training: generalizing image classification from a few samples is a much harder task. We explore this problem through 2 challenges respectively consisting of learning on limited samples of CIFAR-10 (Krizhevsky and others 2009) without and with access to external data. We first review relevant literature on Few-shot Learning and more briefly review the structure and reasoning behind the VGG and Residual networks archi- tecture. We then propose a simple architecture for learning without external data which outperforms a comparably heavier model and finally 2 models for few-shot learning with external data.

Report

Thumbnail source: https://www.cs.toronto.edu/~kriz/cifar.html