ML@GT Seminar by Le Song

Abstract: Many big data analytics problems are intrinsically complex and hard, making the design of effective and scalable algorithms very challenging. Domain experts need to perform extensive research, and experiment with many trial-and-errors, in order to craft approximation or heuristic schemes that meet the dual goals of effectiveness and scalability. Very often, restricted assumptions about the data, which are likely to be violated in real world, are made in order for the algorithms to work and obtain performance guarantees. Furthermore, previous algorithm design paradigms seldom systematically exploit a common trait of real-world problems: instances of the same type of problem are solved repeatedly on a regular basis, differing only in their data. Is there a better way to design effective and scalable algorithms for big data analytics?

I will present a framework for addressing this challenge based on the idea of embedding algorithm steps into nonlinear spaces, and learn these embedded algorithms from problem instances via either direct supervision or reinforcement learning. In contrast to traditional algorithm design where every step in an algorithm is prescribed by experts, the embedding design will delegate some difficult algorithm choices to nonlinear learning models so as to avoid either large memory requirement, restricted assumptions on the data, or limited design space exploration. I will illustrate the benefit of this new design framework using large scale real world data, including a materials discovery problem, a recommendation problem over dynamic information networks, and a problem of learning combinatorial algorithms over graphs. The learned algorithms can reduce memory usage and runtime by orders of magnitude, and sometimes result in drastic improvement in predictive performance.

Bio: Le Song is an Associate Professor in the Department of Computational Science and Engineering, College of Computing, and an Associate Director of the Center for Machine Learning, Georgia Institute of Technology. He received his Ph.D. in Machine Learning from University of Sydney and NICTA in 2008, and then conducted his post-doctoral research in the Department of Machine Learning, Carnegie Mellon University, between 2008 and 2011. Before he joined Georgia Institute of Technology in 2011, he was a research scientist at Google briefly. His principal research direction is machine learning, especially nonlinear models, such as kernel methods and deep learning, and probabilistic graphical models for large scale and complex problems, arising from artificial intelligence, network analysis, computational biology and other interdisciplinary domains. He is the recipient of the Recsys’16 Deep Learning Workshop Best Paper Award, AISTATS’16 Best Student Paper Award, IPDPS’15 Best Paper Award, NSF CAREER Award’14, NIPS’13 Outstanding Paper Award, and ICML’10 Best Paper Award. He has also served as the area chair for many leading machine learning and AI conferences such as ICML, NIPS, AISTATS, AAAI and IJCAI, and the action editor for JMLR.

Event Details


  • Wednesday, April 5, 2017
    1:00 pm - 2:00 pm