UAI 2016 - Invited Speakers
UAI 2016 is pleased to announce the following invited speakers:
Statistics, Machine Learning & the Detection of Gravitational Waves
Gravitational waves are extremely weak ripples of spacetime that propagate at the speed of light. They were first detected on September 14, 2015 by Advanced LIGO (the Laser Interferrometer Gravitational wave Observatory).The origin of these gravitational waves has been attributed to the merger of two black holes, more than 1.3 billion years ago, with masses around 29 and 36 solar masses. This resulted in the formation of a single more massive black hole. About 3 solar masses were converted into gravitational waves in the last fraction of the second of the merger which was then detected by LIGO. This is also the first time that the collision of two black holes has been observed. This discovery open a new window on our universe as gravitational waves carry unique information and studying them is expected to provide important insights into the evolution of stars, supernovae, neutron stars, gamma ray bursts and black holes. I will describe the ideas behind gravitational waves and the analyses that first detected them. I will also discuss how Bayesian inference techniques and machine learning methods are currently being used in the analysis of LIGO data and what role they are expected to play in future discoveries.
Dr Farhan Feroz is a researcher at the Department of Physics, University of Cambridge. He completed his PhD in Astrophysics at Cambridge in 2008 for which he was awarded the Salje medal for the best Science PhD by Clare Hall, Cambridge. Since then he has held Junior Research Fellowship at Trinity Hall Cambridge and Leverhulme & Newton Trust Research Fellowship in Astrophysics at the Department of Physics, University of Cambridge. Dr Feroz's research has focused on the development of novel Bayesian analysis and machine-learning methods and their application to a wide variety of problems in cosmology, particle physics, gravitational wave astronomy and beyond. He has published over 70 peer reviewed journal articles in these areas. Statistical method developed by Dr Feroz have been instrumental in the detection of several extra-solar planets, galaxies and clusters of galaxies. In recognition of this work, Dr Feroz was awarded the UK Scopus Young Researcher award in physical science by Elsevier and US/UK in 2011.
Online optimization of power networks
We are at the cusp of a historical transformation of our power systems into a more sustainable, dynamic, intelligent, and distributed form with hundreds of millions of intelligent endpoints. The optimization and control of such a network requires solving power flow equations which is well-known to be hard. The grid, however, solves them in real-time at scale, and we propose to exploit it explicitly to carry out part of our optimization algorithm. This approach not only improves scalability, but also naturally adapts to evolving network conditions. In this talk, we present two examples.
The first example presents an online algorithm to solve an optimal power flow problem at a slow timescale on a radial network where the controllable devices continuously interact with the network that implicitly computes a power flow solution given a control action. Collectively these devices and the network implement a gradient projection algorithm in real time. We prove that the proposed algorithm converges to a set of local optima and provide sufficient conditions under which it converges to a global optimum. We derive an upper bound on the suboptimality gap of any local optimum. This bound suggests that any local optimum is almost as good as any strictly feasible point.
In the second example, the online algorithm integrates primary frequency regulation, secondary frequency regulation, and congestion management at a fast timescale. The algorithm is distributed in that it requires communication only between neighboring buses. Collectively, the controllable devices and the swing dynamics of the network implement a primal-dual algorithm to rebalance power, restore nominal frequency and inter-area flows, and enforce line limits at a minimum control cost. We prove sufficient conditions under which the algorithm converges to a global optimum.
Steven H. Low is a Professor of the Department of Computing & Mathematical Sciences and the Department of Electrical Engineering at Caltech. Before that, he was with AT\&T Bell Laboratories, Murray Hill, NJ, and the University of Melbourne, Australia. He is a Senior Editor of the IEEE Transactions on Control of Network Systems and the IEEE Transactions on Network Science & Engineering, is on the editorial boards of NOW Foundations and Trends in Networking, and in Electric Energy Systems, as well as Journal on Sustainable Energy, Grids and Networks. He is an IEEE Fellow and received his B.S. from Cornell and PhD from Berkeley, both in EE. His research interests are in the control and optimization of power systems and Internet.
University of Massachusetts Amherst
Structured Prediction and Deep Learning
Deep neural networks have recently revolutionized speech recognition, computer vision, natural language processing and other areas. Dramatically improved empirical results have been enabled by their ability to learn rich representations of their inputs, where setting up the learning in layers is made easier by automatic differentiation. In structured prediction we capture dependencies among the output variables. In this talk I will survey intersections between deep learning and structured prediction, explore the relationships between inference in graphical models and prediction in feed-forward neural networks, and finally introduce our recent work in "structured prediction energy networks" (Belanger and McCallum 2016), which uses a deep architecture to learn rich representations of output dependencies -- essentially replacing the factors in the factor graph with a neural network yielding a scalar energy.
Andrew McCallum is a Professor and Director of the Center for Data Science at the University of Massachusetts Amherst. He has published over 250 papers in many areas of AI, including natural language processing, machine learning, data mining and reinforcement learning, and his work has received over 45,000 citations. He obtained his PhD from University of Rochester in 1995 with Dana Ballard and a postdoctoral fellowship from CMU with Tom Mitchell and Sebastian Thrun. In the early 2000's he was Vice President of Research and Development at at WhizBang Labs, a 170-person start-up company that used machine learning for information extraction from the Web. He is a AAAI Fellow, the recipient of the UMass Chancellor's Award for Research and Creative Activity, the UMass NSM Distinguished Research Award, the UMass Lilly Teaching Fellowship, and research awards from Google, IBM, Yahoo and Microsoft. He was the General Chair for the International Conference on Machine Learning (ICML) 2012, and is the current president of the International Machine Learning Society, as well as member of the editorial board of the Journal of Machine Learning Research. For the past twenty years, McCallum has been active in research on statistical machine learning applied to text, especially information extraction, entity resolution, semi-supervised learning, topic models, and social network analysis. His work on open peer review can be found at http://openreview.net.
University of Copenhagen
Total positivity and Markov structures
Positive associations between random variables can be described in a variety of ways. One of the strongest is multivariate total positivity of order two (MTP2) introduced by Karlin and Rinott (1980). The MTP2 property is stable under a number of operations, including marginalization and conditioning. The lecture investigates how the property interacts with conditional independence structures, implying, for example, that any strictly positive MTP2 distribution becomes faithful to its pairwise independence graph. In addition we shall study how this property is manifested in Gaussian, discrete, and conditional Gaussian distributions, and give a number of examples of such distributions. I shall also discuss implications of MTP2 for issues of structure learning in graphical models. The lecture is based on joint work with S. Fallat, K. Sadeghi, C. Uhler, N. Wermuth, and Piotr Zwiernik; arXiv:1510.01290.
Steffen Lauritzen studied Statistics at the University of Copenhagen, Denmark, completing the degree of lic.stat. (PhD level) in 1975 He was appointed there as Lecturer until 1981 where he took up a Professorship of Mathematics and Statistics at Aalborg University, Denmark. In 2004 he became Professor of Statistics at the University of Oxford and in 2014 he returned to Copenhagen as Professor of Statistics there. His main research interests are graphical models and their applications. He has received numerous awards and academic recognitions, including the Award for "Outstanding Statistical Application" from the American Statistical Association in 1989; the Guy Medal in Silver 1996; and the DeGroot Prize 2002. He was elected to the Royal Danish Academy of Sciences and Letters 2008, the Royal Society 2011.