Wikileaks' release of classified data detailing the activity of the US military in Afghanistan in 2004-09 caused enormous sensation in the political and diplomatic world, yet it was unclear whether the leaked data contained any useful information which would help understand the conflict dynamics. We took a data-driven approach and tried to fit a simple spatio-temporal modelling approach to the data, the spatio-temporal log-Gaussian Cox process. Surprisingly, the resulting parametrised model could capture many non-trivial statistics of the conflict, and proved extremely accurate in forward forecasting. In this seminar, I will explain the modelling framework and the machine learning techniques we used. I will also discuss some novel approximate inference techniques which make the methodology scale to higher dimensional data, enabling us to capture finer resolution dynamics of the process.
Ref: Zammit-Mangion et al, PNAS 109(31), 2012; Cseke et al, J. Am. Stat. Ass. in press 2017.
Guido Sanguinetti is a Reader in Machine Learning at the School of Informatics, University of Edinburgh. After a degree in Physics (Genova) and a PhD in Mathematics (Oxford), his interests shifted towards machine learning and computational biology, with a particular focus on devising statistical models for the dynamics of gene regulation. He has published over 60 articles in leading specialist and interdisciplinary journals, including Science, Proceedings of the National Academy of Sciences (PNAS), Nature Communications, Bioinformatics, and JASA. He holds an ERC starting grant and is the recipient of the 2012 PNAS Cozzarelli Prize in Engineering and Applied Science, the 2013 QEST Best Paper Award and the 2016 ECCB Best Paper Award.
This talk is organized by the Cyber-Physical Systems Group at the Institute of Computer Engineering.