ECE Seminar with Zhenming Liu

Learning in a Distributed System With Zhenming Liu Postdoctoral Research Associate Princeton University Faculty Host: Clem Karl Refreshments will be served outside Room 339 at 3:45 p.m. A central challenge in big data analytics is to design distributed algorithms for making real-time decisions and predictions. Like in a centralized setting, an optimal algorithm in a distributed environment needs to be able to quickly adapt its decision strategy as it continuously receives data. In addition, distributed algorithms bear communication costs and hence are subject to efficiency versus efficacy trade-offs. We will present the first theoretically grounded results addressing such trade-offs under both stochastic and non-stochastic models (i.e., no distributional assumption is made as to the input). Specifically, we focus on the following two problems: (1) The online non-stochastic experts problem, where one of the computers in the system has to pick one expert from a given set at each time period and is supplied with payoff information. The goal is to minimize regret with respect to the optimal choice in hindsight, while simultaneously keeping communication at a minimum. (2) The count tracking problem (counter), which is an important special case of the stochastic gradient descent problem and also serves as a building block for other more complex algorithmic problems. Here the goal is to keep track of all partial sums of items seen thus far in the stream, and we design a new randomized algorithm that balances the tracking accuracy and communication cost optimally. About the Speaker: Zhenming Liu is a postdoctoral research associate at Princeton University, working with Jennifer Rexford, Mung Chiang, and Vincent Poor. He received a PhD in theory of computation at Harvard in 2012. His doctoral research lies in the intersections among applied probability, combinatorial optimization, and machine learning. More recently, he works with networking and data mining researchers to understand how these theoretical tools can help in building scalable systems for analyzing massive datasets. He has received several awards for his research, including the Best Student Paper Award at ECML/PKDD 2010.