Speaker: Dr. Marat Freytsis of Rutgers University

Abstract: Most measurements of interest in fundamental physics involve analyzing data where signals are impossible to fully disentangle from backgrounds with complicated underlying dynamics. This situation stands in contrast to conventional machine learning applications where data is assumed to be cleanly separable but theoretical properties are intractable. I discuss two venues where present experimental data has become sufficiently complex that machine learning methods can reap significant benefits, but the close interplay of signal and background dynamics require building physics knowledge into the algorithms. In the treatment of QCD jets at hadron colliders, I present techniques to allow both well-understood pertrubative aspects of collisions and more complex jet formation to be treated together while maintaining knowledge of their relative unceratinties. I also examine additional complications in astrophysical searches for dark matter and gravitational waves where the availability of only a single set of observations requires a careful use of necessarily incomplete simulations to make connections between theory and experiment.