Few Shot Learning with Generative Adversarial Networks

Abstract

GANs (generative adversarial networks) are a recent machine learning method for generative modeling, which has seen tremendous success in rich datasets such as images.
A common problem in ML is not having enough data for a specific task to perform well. This is particularly a problem for data-hungry tasks such as GAN training and generative modeling in general.
A common way to improve the performance of the data-deficient task is to leverage data from othersimilar tasks; this is a technique known as transfer learning or few-shot learning. We are investigating such techniques for GANs using a new deep learning architecture known as DeepSets.