IEEE Congress on Evolutionary Computation tutorial Adversarial Deep Learning by Using Coevolutionary Computation
In recent years, machine learning with Generative Adversarial Networks (GANs) has been recognized as a powerful method for generative modeling. Generative modeling is the problem of estimating the underlying distribution of a set of samples. GANs accomplish this using unsupervised learning. They have also been extended to handle semi-supervised and fully supervised learning paradigms. GANs have been successfully applied to many domains. They can generate novel images (e.g., image colorization or super-resolution, photograph editing, and text-to-image translation), sound (e.g., voice translation and music generation), and video (e.g., video-to-video translation, deep fakes generation, and AI-assisted video calls), finding application in domains of multimedia information, engineering, science, design, art, and games.
GANs are an adversarial paradigm. Two NNs compete with each other using an antagonistic lost function to train the parameters with gradient descent. This connects them to evolution because evolution also exhibits adversarial engagements and competitive coevolution. In fact, the evolutionary computation community’s study of coevolutionary pathologies and its work on competitive and cooperative coevolutionary algorithms offers a means of solving convergence impasses often encountered in GAN training.
In this tutorial we will explain:
- The main concepts of generative modeling and adversarial learning.
- GAN gradient-based training and the main pathologies that prevent ideal convergence. Specifically, we will explain mode collapse, oscillation, and vanishing gradients.
- Coevolutionary algorithms and how they can be applied to train GANs. Specifically, we will explain how algorithm enhancements address non-ideal convergence
- To demonstrate we will draw upon the open-source Lipizzaner framework (url: http://lipizzaner.csail.mit.edu/). This framework is easy to use and extend. It sets up a spatial grid of communicating populations of GANs.
- Students will be given the opportunity to set up and use the Lipizzaner framework during the tutorial by means of a jupyter notebook expressly developed for teaching purposes.
Jamal Toutouh Massachusetts Institute of Technology, CSAIL, USA
- Email: firstname.lastname@example.org
- Bio: I am a Marie Skłodowska Curie Postdoctoral Fellow at Massachusetts Institute of Technology (MIT) in the USA, at the MIT CSAIL Lab. I obtained my Ph.D. in Computer Engineering at the University of Malaga (Spain). The dissertation, Natural Computing for Vehicular Networks, was awarded the 2018 Best Spanish Ph.D. Thesis in Smart Cities. My dissertation focused on the application of Machine Learning methods inspired by Nature to address Smart Mobility problems. My current research explores the combination of Nature-inspired gradient-free and gradient-based methods to address Adversarial Machine Learning. The main idea is to devise new algorithms to improve the efficiency and efficacy of the state-of-the-art methodology by mainly applying co-evolutionary approaches. Besides, I am working on the application of Machine Learning to address problems related to Smart Mobility, Smart Cities, and Climate Change.
Una-May O’Reilly Massachusetts Institute of Technology, CSAIL, USA
- Email: email@example.com
- Bio: Una-May O’Reilly is leader of the AnyScale Learning For All (ALFA) group at MIT CSAIL. ALFA focuses on evolutionary algorithms, machine learning and frameworks for large scale knowledge mining, prediction and analytics. The group has projects in cyber security using coevolutionary algorithms to explore adversarial dynamics in networks and malware detection. Una-May received the EvoStar Award for Outstanding Achievements in Evolutionary Computation in Europe in 2013. She is a Junior Fellow (elected before age 40) of the International Society of Genetic and Evolutionary Computation, which has evolved into ACM Sig-EVO. She now serves as Vice-Chair of ACM SigEVO. She served as chair of the largest international Evolutionary Computation Conference, GECCO, in 2005.
IEEE Congress on Evolutionary Computation conference: 28 June 2021 - 1 July 2021