Guided Generative Adversarial Neural Networks

Abstract: Generative Adversarial Networks is relatively new approach of building unsupervised models that are able to translate a latent distribution of one nature to the different one. Typical applications of GANs are image-related, for instance grayscale-to-color image reconstruction, image impainting, text-to-image translation, and zoom without a quality loos.

Vanilla GANs consists of Generator and Discriminator that play against each other. While Generator tries to create a fake data that looks like genuine data, Discriminator’s goal is to recognize a fool and penalize the Generator for the bad job. In theory after a certain number of iterations Generator learns how to generate the output that is indistinguishable from the genuine dataset.
However in business applications the nature of input and output is usually connected to each other and the final result should inherit visual properties of both. In this session we would discuss a new type of GAN - ""Guided Generative Adversarial Neural Network"" that based on modified multi-level loss function of Generator, talk about difficulties of that approach and how to overcome them to build an end-to-end application. During the session we will show the solution built for the fashion industry that translates a colorful sketch to the detailed design of bodywear in a guided and predictable way.

Bio: Alexey has a background in applied math and physics and has been working as software developer, producing business applications for various companies across the globe. Gradually his interests turned towards Data Science with a focus on building sustainable, applicable solutions with a particular interest in Deep Learning. At the moment Alexey is heading Data Science team in PVH Europe (known by fashion brands like Tommy Hilfiger and Calvin Klein).