Title: Fast RL
Abstract: Current RL algorithms, while strong in some ways, learn much more slowly than humans. This is caused, in part, by their lack of any experience whatsoever with the world. As a result, their early actions are completely random, and learning is a lot slower than desired.
In reaction to this problem, I will present the Retro Challenge, which is a collection of Sonic the Hedgehog levels. This is the first RL benchmark that comes with a training set and a secret test set, consisting of new levels specifically designed for this contest. It has the property that joint training works a lot better than training from scratch, but humans still learn far faster than jointly-trained systems.
I will present the details of the Retro Challenge and the results we have so far, and will conclude by presenting some of our still unpublished results on generative models.
Bio: Ilya Sutskever is a computer scientist working in machine learning and is the Cofounder and Research Director of OpenAI. Sutskever obtained his B.Sc, M.Sc, and Ph.D in Computer Science from University of Toronto’s Department of Computer Science under the supervision of Geoffrey Hinton. After graduation, Sutskever became a postdoc with Andrew Ng at Stanford University. Before joining Google Research’s Brain team, he was the co-founder of DNNresearch. Sutskever was named in MIT Technology Review’s 35 Innovators Under 35.