High Energy Physics Seminars

Parameterised learning: From quantum interference to systematic uncertainties

by Dr. Aishik Ghosh (University of California, Irvine, US. also at Lawrence Berkeley National Laboratory, US.)

Thursday, December 16, 2021 from to (Asia/Kolkata)
Description
Machine learning is allowing us to optimize analyses over entire phases spaces in a way we could not do before. I will discuss the power of parameterized learning with two important use cases. The first is uncertainty-aware networks, a new technique that allows us to optimally treat systematic uncertainties that come from mismodelling in simulations and can readily be applied to an existing LHC analysis.  The second is a likelihood-free inference technique that is for the first time able to overcome the non-linear effects that plague the off-shell Higgs couplings measurements in the four leptons channel. This is a network that allows to compute the likelihood in an unbinned, multi-dimensional way, taking into account the non-linear effects that arise due to the distractive interference between signal and background processes.  Finally, I will end with a cautionary tale about the dangers of attempting to use machine learning to decorrelate systematic uncertainties in an LHC analysis. After the talk I am happy to discuss ML ideas in development for theory, phenomenology, experimental physics and beyond.