Speaker
Description
Anomaly detection offers a way to search for new physics without assuming a specific signal model: instead of designing cuts for a particular scenario, one selects events that look unusual compared with the Standard Model expectation. Two questions arise immediately: how to quantify "unusual," and how to faithfully estimate the Standard Model background among the selected events. In this talk, I present a method that addresses both questions. A neural network called a normalizing flow is trained to learn a smooth, invertible change of variables that turns the correlated kinematic observables of events into new variables that behave like independent standard Gaussians. In these new variables, an event's distance from the origin measures how atypical it is. Dividing the new variables into two groups gives two independent measures of how unusual each event is, which is exactly the condition needed for the well-established ABCD method of data-driven background estimation. As a proof of concept, I apply the method to an R-parity-violating SUSY benchmark against a t-tbar background at the HL-LHC. When the network is trained on a background-only sample, the method raises the S/√B from about 1.5 to about 15, and the ABCD background prediction agrees well with the true background. Finally, I discuss the prospects and current limitations of applying the method directly to data.