Abstract:
Machine learning is a powerful technique which is increasingly important in everyday life as well as scientific research. Despite enormous recent progress, in particular involving deep neural networks, it remains a challenge to understand what the machine is "learning."
I will first give a brief review of neural networks. Then, focusing on classification problems (e.g. signal vs. background), I will describe a technique called "data planing" which allows one to test the importance of any given physics variable in a classification scenario. Thus, this allows one to reverse engineer the relevant information that the neural network has "discovered."
As a demonstration, I will use a simple particle physics search to illustrate the procedure in action. At the end of the talk, I will discuss future directions in data planing as well as other ideas on how to better understand/utilize machine learning in physics.
Brief Bio:
Education and Academic Positions
1995-1999
Stanford University - B.S. in Physics
1999-2004
Harvard University - PhD in Theoretical Particle Physics (Advisor: Howard Georgi)
2004-2008
New York University - Postdoc
2008-2011
University of California at Davis - Postdoc
2011-2017
University of Oregon, Assistant Professor of Physics
2017-current
University of Oregon, Associate Professor of Physics
2018-2019
National Taiwan University, Courtesy Associate Professor of Physics