Title: Assistant Professor
Track: Advanced Technologies
Presentation Title: “Big Data Is Not Enough: Thoughts on Casuality, Physics-Informed Machine Learning, and Decision Making Under Uncertainty"
Big data is not enough: Thoughts on causality, physics-informed machine learning, and decision making under uncertainty. Suppose that you have collected a lot of data from your business/engineering/scientific activities and that you would like to use them to improve your position. However, you have a finite budget to dedicate in research and development. How do you decide what to do next? The naïve answer is “I will build a model out of the data and I will use it to predict the most valuable intervention.” However, analyses based on purely data-driven models should be taken with a grain of salt, because the trends or correlations captured therein do not necessarily represent causal relationships. A costly intervention to change the value of a variable is not logically followed by the expected effect on the important quantities of interest. Furthermore, even the best model has inherent uncertainties when we use it to extrapolate beyond the data used to build it. How do we take this uncertainty into account? In this talk, I will present a framework for making such decisions with the goal of maximizing a value function. The core of the methodology builds upon the causality research of Judea Pearl (structural equations framework), but it expands it to allow for the quantification of model-form uncertainty (Bayesian inference over the space of possible models). Decision making takes place by maximizing sequentially the expected marginal value of information or the expected information gain. I will provide motivating examples from research projects funded by NSF, NASA, and the industry.
Bio: Bilionis obtained his Diploma in Applied Mathematics and Physical Sciences from the National Technical University of Athens in 2008. In 2013, he obtained his Ph.D. in Applied Mathematics from Cornell University. After graduation, he spent a year working as a postdoctoral researcher at the Mathematics and Computer Science Division of Argonne National Laboratory. In August 2014, he became an Assistant Professor of Mechanical Engineering at Purdue University, where he established the Predictive Science Laboratory (PSL). The mission of PSL is to understand how to optimally design engineering systems (ES) under uncertainty. The applications of PSL span the range between technical (e.g., oil production, power system operation, electric machines) and sociotechnical systems (e.g., efficient office and residential buildings, extra-terrestrial habitats). His research has been funded by NSF, NASA, DARPA, Ford, Facebook, Purdue University, and the University of Illinois. His collaborative research programs have been awarded a total of $23.7M.
Bilionis’ research develops communication channels between theories and between theories and data. The communication protocol is based on probability theory (thought of as an extension of logic under uncertainty) with an additional layer of causality (expressed through physical laws in the form of differential equations and graphical models). Due to the high cost of information acquisition, engineering applications are characterized by small numbers of observations, e.g., how many times can one test an aircraft in a wind tunnel? In this regime, standard statistical/machine learning techniques fail because they cannot extrapolate beyond observations. This is where PSL operates. It develops methods that exploit physical knowledge to predict beyond observations with limited data.
Bilionis’ has published 34 journal papers, 2 book chapters, and 15 conference papers. He has chaired/co-chaired 19 Master and Ph.D. student committees and was presented with the “Outstanding Faculty Mentor of Mechanical Engineering Graduate Students” award. He has developed a graduate course on uncertainty quantification which seeks to introduce engineers to probabilistic thinking and machine learning directly on engineering applications. Bilionis provides code for all concepts discussed in class in the form of incomplete Python notebooks. Video lectures are available through nanoHUB. Based on his teaching evaluations, he has received the “Outstanding Engineering Teacher Recognition” three times.