"New Machine Learning Approaches to Modeling Dynamic Systems"

Byron Boots
University of Washington

Monday, March 3
10:30 a.m.
LWSN 3102 A/B

ABSTRACT:

A major challenge in machine learning is to reliably and automatically discover hidden structure in high-dimensional data. This is an especially formidable problem for sequential data: revealing the dynamical system that governs a complex time series is often not just difficult, but provably intractable. Popular maximum likelihood strategies for contending with this problem are slow in practice and often terminate at poor local optima, problems that were long thought to be unavoidable. However, recent work has shown that progress can be made by shifting the focus to realistic instances that rule out the intractable cases. In this talk, I will present a new family of computational approaches for learning a wide range of dynamical system models. The key insight is that low-order moments of observed data often possess structure that can be revealed by powerful spectral decomposition methods, and, from this structure, model parameters can be directly recovered. Based on this insight, we design highly effective algorithms for learning parametric models like Kalman Filters and Hidden Markov Models, as well as a new class of nonparametric models via reproducing kernels. Unlike maximum likelihood-based approaches, these learning algorithms are statistically consistent, computationally efficient, and easy to implement using established matrix-algebra techniques. The result is a powerful framework for learning dynamical system models directly from data with state-of-the-art performance on real-world problems in several domains.

SPEAKER BIO:

Byron Boots is a postdoctoral researcher working with Dieter Fox in the Robotics and State Estimation Lab at the University of Washington. He received his Ph.D. in Machine Learning from Carnegie Mellon University in 2012 where he was advised by Geoffrey Gordon. Byron’s work on learning models of dynamical systems received the 2010 Best Paper award at the International Conference on Machine Learning (ICML-2010). His research focuses on statistical machine learning, artificial intelligence, and robotics.
_______________________________________________
Cs-colloq mailing list
Cs-colloq@lists.purdue.edu
https://lists.purdue.edu/mailman/listinfo/cs-colloq

Purdue University College of Science, 150 N. University St, West Lafayette, IN 47907 • Phone: (765) 494-1729, Fax: (765) 494-1736

Student Advising Office: (765) 494-1771, Fax: (765) 496-3015 • Science IT, (765) 494-4488

© 2018 Purdue University | An equal access/equal opportunity university | Copyright Complaints

Trouble with this page? Disability-related accessibility issue? Please contact the College of Science Webmaster.