Dortmund 2021 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 4: AKPIK III: Simulation & Application

AKPIK 4.8: Talk

Thursday, March 18, 2021, 17:45–18:00, AKPIKa

Deep Continuum Suppression with Predictive Uncertainties — •Lars Sowa, James Kahn, and Pablo Goldenzweig — Karlsruhe Institute of Technology (KIT)

The Belle II collaboration works on precision measurements using data collected from the SuperKEKB collider. This requires a high purity of signal candidates, therefore it is necessary to suppress e⁺ e⁻ → q q (q=u,d,c,s) continuum events effectively. To do so, the Belle II analysis framework uses traditional machine learning methods. In recent years, deep learning techniques have shown to be very powerful, outperforming these traditional methods in many fields of research. While deep learning techniques are promising for continuum suppression, an ongoing problem is that they traditionally don*t provide meaningful uncertainties to their predictions, a key requirement for physics analyses.

Recent work has shown that deep ensemble methods solve this problem by providing a measure of prediction uncertainty making them a promising candidate for use in continuum suppression. This talk presents the current status of a study into deep ensemble continuum suppression with predictive uncertainty estimation for the Belle II experiment. Additionally, a decorrelation mechanism to prevent biasing features of interest is presented.