Title: Modernizing Tracking with Machine Learning
Speaker: Ryan Turner, Research Scientist, Northrop Grumman
Abstract: In this talk I will review the work of two NIPS papers that combine methods from machine learning and the older field of tracking. Most of the methodology is motivated by radar tracking problems, but is equally applicable to other areas such as computer vision. I will cover research that shows how to replace what is effectively a sliding window MAP estimate for data association in the traditional tracking world with a variational Bayes algorithm. The final algorithm contains both loopy belief propagation and standard variational Bayes. It offers both speed and accuracy advantages over the traditional MAP approach. The loopy belief propagation algorithm for data association is similar to problems involved in estimating a matrix permanent and the Hungarian algorithm.
Speaker Bio: Ryan Turner did his PhD in machine learning at the University of Cambridge, UK under Carl Rasmussen and Zoubin Ghahramani. Previously, he did his undergraduate in computer engineering at the University of California Santa Cruz. He gained an interest in time series after working on many time series problems at Google as an intern in 2008. His research at Cambridge involved the intersection of Gaussian processes applied to time series problems as well as change point detection and state space models. After completing his PhD in 2011 he worked as a high frequency trading researcher for a year in London. In 2012, he moved back to the USA and began working on radar tracking problems for Northrop Grumman Corporation. He developed and published new algorithms combining elements of his PhD work, variational Bayes, and tracking. Since the beginning of 2015 he has been working on applying machine learning to analyze data sets involved in network security applications. He also has an interest in the overlap between machine learning and good software practices and organized a NIPS workshop in 2014 on the topic.