Distinguished Lecture Series
On Gradient-Based Optimization: Accelerated, Stochastic and Nonconvex
Professor Michael I. Jordan, University of California, Berkeley
30 August 2018
Glasgow 109 at 1500
Many new theoretical challenges have arisen in the area of gradient-based optimization for large-scale statistical data analysis, driven by the needs of applications and the opportunities provided by new hardware and software platforms. I discuss several recent, related results in this area: (1) a new framework for understanding Nesterov acceleration, obtained by taking a continuous-time, Lagrangian/Hamiltonian/symplectic perspective, (2) a discussion of how to escape saddle points efficiently in nonconvex optimization, and (3) the acceleration of Langevin diffusion.
Michael I. Jordan is the Pehong Chen Distinguished Professor in the Department of Electrical Engineering and Computer Science and the Department of Statistics at the University of California, Berkeley. His research interests bridge the computational, statistical, cognitive and biological sciences. Prof. Jordan is a member of the National Academy of Sciences and a member of the National Academy of Engineering. He has been named a Neyman Lecturer and a Medallion Lecturer by the Institute of Mathematical Statistics. He received the IJCAI Research Excellence Award in 2016, the David E. Rumelhart Prize in 2015 and the ACM/AAAI Allen Newell Award in 2009.
Perspectives on Stochastic Modeling
Professor Peter Glynn, Stanford University
Uncertainty is present in almost every decision-making environment. In many applications settings, the explicit quantitative modeling of uncertainty clearly improves decision-making. In this talk, I will discuss some perspectives on such stochastic models and their application. Specifically, I will talk about the interplay between modeling, data, and computation, and some of the lessons learned that are relevant to building models that can add value and insight.
Peter W. Glynn is the Thomas Ford Professor in the Department of Management Science and Engineering (MS&E) at Stanford University. He is a Fellow of INFORMS and of the Institute of Mathematical Statistics, and has been co-winner of Best Publication Awards from the INFORMS Simulation Society in 1993, 2008, and 2016 and the INFORMS Applied Probability Society in 2009. He was the co-winner of the John von Neumann Theory Prize from INFORMS in 2010 and in 2012, he was elected to the National Academy of Engineering. His research interests lie in stochastic simulation, queueing theory, and statistical inference for stochastic processes.
The lecture will be recorded and posted.
Bayesian Search for Missing Aircraft
Lawrence D. Stone
In recent years there have been a number of highly publicized searches for missing aircraft such as the ones for Air France flight AF 447 and Malaysia Airlines flight MH 370.
Bayesian search theory provides a well-developed method for planning searches for missing aircraft, ships lost at sea, or people missing on land. The theory has been applied successfully to searches for the missing US nuclear submarine Scorpion, the SS Central America (ship of gold), and the wreck of AF 447. It is used routinely the by U. S. Coast Guard to find people and ships missing at sea.
This talk presents the basic elements of the theory. It describes how Bayesian search theory was used to locate the wreck of AF 447 after two-years of unsuccessful search and discusses how it was applied to the search for MH 370. A crucial feature of Bayesian search theory is that it provides a principled method of combining all the available information about the location of a search object. This is particularly important in one-of-a-kind searches such as the one for AF 447 where there is little or no statistical data to rely upon.
Applied Risk Analytics: Making Advanced Analytics More Useful
Dr. Tony Cox
Traditional operations research emphasizes finding a feasible decision that maximizes an objective function. In practice, how decisions affect the objective function, and even what decisions are feasible, are often initially unknown. Managing risks effectively usually requires using available data, however limited, to answer the following questions, and then improve the answers in light of experience:
- DESCRIPTIVE ANALYTICS: What’s happening? What has changed recently? What should we be worrying about?
- PREDICTIVE ANALYTICS: What will (probably) happen if if we do nothing?
- CAUSAL ANALYTICS: What will (probably) happen if we take different actions or implement different policies? How soon are the consequences likely to occur, and how sure can we be?
- PRESCRIPTIVE ANALYTICS: What should we do next? How should we allocate available resources to explore, evaluate, and implement different actions or policies in different locations?
- EVALUATION ANALYTICS: How well are our risk management policies and decisions working? Are they producing (only) their intended effects? For what conditions or sub-populations do they work or fail?
- LEARNING ANALYTICS: How might we do better, taking into account value of information and opportunities to learn from small trials before scaling up?
- COLLABORATIVE ANALYTICS: How can we manage uncertain risks more effectively together?
This talk discusses recent advances in these areas and suggests how they might be integrated into a single decision support framework, which we call risk analytics, and applied to important policy questions such as whether, when, and how to revise risk management regulations or policies. Current technical methods of risk analytics, including change point analysis, quasi-experimental design and analysis, causal graph modeling, Bayesian Networks and influence diagrams, Granger causality and transfer entropy methods for time series, causal analysis and modeling, and low-regret learning provide a valuable toolkit for using data to assess and improve the performance of risk management decisions and policies by actively discovering what works, what does not, and how to improve over time.