Aspectos matemáticos de las Redes Generativas Antagónicas

Abstract

Machine learning, and neural networks in particular, have become a very useful resource for solving problems such as decision making, where a model is trained using the information from a database (images or individuals data, for example) to predict future decisions. Thanks to enormous progress in this field, new aspirations have arisen, in our case, the industry desire to design data. In fact, we do not want to know how to react to training information, our objective is to infer the intrinsic qualities in order to generate new observations that appear so real that we are not able to distinguish them from the originals. In this relatively new approach, there are several ways of dealing with this problem. A possible solution will be GANs, generative adversarial networks, the central object of our study. We will see how they work and their mathematical foundations. Finally, due to the convergence problems that may arise during their learning process, we will make a literature review of different solutions that have been proposed to solve it, and following this, we will perform several experiments to test them.