Soheil Feizi: Generative Adversarial Networks: Formulation, Design and Computation
0A talk by Soheil Feizi at the Quantum Machine Learning Workshop, hosted September 24-28, 2018 by the Joint Center for Quantum Information and Computer Science at the University of Maryland (QuICS).
Abstract: Learning a probability model from data is a fundamental problem in machine learning and statistics. A classical approach to this problem is to fit (approximately) an explicit probability model to the training data via a maximum likelihood estimation. Building off the success of deep learning, however, there has been another approach to this problem using Generative Adversarial Networks (GANs). GANs view this problem as a game between two sets of functions: a generator whose goal is to generate realistic fake samples and a discriminator whose goal is to distinguish between the real and fake samples. In this talk, I will explain challenges that we face in formulation, design and computation of GANs. Leveraging a connection between supervised and unsupervised learning, I will then elaborate how we can overcome these issues by proposing a model-based view to GANs.