Thu, Nov 30, 2017 @ 03:30 PM - 04:50 PM
Conferences, Lectures, & Seminars
Speaker: Zi Wang, MIT
Talk Title: Bayesian Optimization and How to Scale it Up
Abstract: This lecture satisfies requirements for CSCI 591: Research Colloquium.
In recent years, Bayesian optimization (BO) has become a popular and effective approach to optimize an expensive black-box function with assumptions usually expressed by a Gaussian process prior. Successful applications include tuning hyper-parameters for neural networks, optimizing control parameters for robots, and designing biological experiments. Despite these successes, BO has been limited to small-scale and low-dimensional problems due to computational challenges with Gaussian processes and statistical challenges in high-dimensional settings. In this talk, I will present our recent work on scaling up BO from several aspects. First, I will introduce Max-value Entropy Search, a new BO strategy that improves sample-efficiency and obtains the first regret bound for a variant of the entropy search methods. Building on the new acquisition function, we extend our approach to high dimensions by learning the additive structures of the kernel. And finally, we propose a scalable high-dimensional BO approach that gives previously impossible results of scaling up BO to tens of thousands of observations within minutes of computation. We also show some interesting new findings on how evolutionary algorithms and BO are related, and establish novel connections among several well-known BO methods including entropy search, GP-UCB, and probability of improvement.
Biography: Zi Wang is a Ph.D. student at the MIT Computer Science and Artificial Intelligence Laboratory, advised by Stefanie Jegelka, Leslie Kaelbling and Tomás Lozano-Pérez. She received her S.M. in Electrical Engineering and Computer Science from MIT in Feb. 2016, and B.Eng. in Computer Science and Technology from Tsinghua University in Jul. 2014. Her research interests lie broadly in machine learning and artificial intelligence, currently with applications to robotics problems.
Host: Fei Sha
Audiences: Everyone Is Invited
Contact: Computer Science Department