Generative Adversarial Networks as a Data Augmentation Tool for Handwritten Digits

Abstract

In the field of data processing, handwritten digit recognition (HDR) has proven to be of great use. However, due to the vast differences in how different people write, accurate recognition of such characters from images is a challenging job. The labelled samples necessary for supervised learning methods are not always easy to come by. For instance, a lot of labelled examples are needed to train a model in deep learning approaches, where all the feature extraction steps are learned within the artificial neural network. To get around this problem, data augmentation methods can be used to fill in the gaps using variations in an example's label that are already known. The Generative Adversarial Network (GAN) is able to generate random samples from the latent space that are statistically indistinguishable from the training set's actual examples. In this study, we leverage the powerful features of GAN to learn from the MNIST data set and produce digital images of handwriting.