Adversarial Networks — A Technology for Image Augmentation

Abstract

A key application of data augmentation is to boost state-of-the-art machine learning for completion of missing values and to generate more data from a given dataset. In addition to transformations or patch extraction as augmentation methods, adversarial networks can be used to learn the probability density function of the original data. Generative adversarial networks (GANs) are an adversarial method to generate new data from noise by pitting a generator against a discriminator and training in a zero-sum game trying to find a Nash Equilibrium. This generator can then be used in order to convert noise into augmentations of the original data. This short paper shows the usage of GANs in order to generate fake face images as well as tips to overcome the notoriously hard training of GANs.