Generative Adversarial Neural Networks: A Pictorial Approach

Abstract

This current work explores the basis of the ambit of Generative Modeling approached through Adversarial Neural Networks. The project starts by introducing Discriminative and Generative Modelling, both explained and set apart from one another, so that the latest, which is the main focus of this work, can be properly understood. Then, examples of Deep Learning architectures modeled as means to approach this kind of modeling, through a setting inspired in Game Theory, are showcased. In the end, a proof of concept to showcase the capabilities of this kind of approaches, capable of learning to reproduce the work of pictorial artist, referred to as PictorialGAN, is presented. In the annex there are references to the theoretical ambit over which this work rises, a whole detailed explanation of Multilayer Perceptrons, as well as the practical bricks used during the development of the project, those being the computational environment and the datasets used to test the models.