In recent years, deep learning has become an essential tool in LHC physics, with ML taggers playing a key role in the most sensitive measurements and searches from ATLAS and CMS. While their performance is now beyond question, these taggers are still fundamentally “black boxes” that leverage mysterious — and potentially deeply unphysical — aspects of high-dimensional input spaces. In this talk, I will discuss recent work (2207.12411) that sheds light on what physical information a simple parton-level quark/gluon tagger might be learning, and proposes a general strategy for matching the discrimination power of an ML classifier using first-principles physics calculations. I will also briefly discuss recent work (2308.XXXXX, presented at BOOST 2023) that presents a new strategy for mitigating the systematic uncertainties that come about from training ML classifiers on inaccurate Monte Carlo simulations.
ZOOM LINK: https://uci.zoom.us/j/99380793791