Bonn 2020 – scientific programme

The DPG Spring Meeting in Bonn had to be cancelled! Read more ...

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

T: Fachverband Teilchenphysik

T 47: Neural networks and systematic uncertainties

T 47.5: Talk

Wednesday, April 1, 2020, 17:30–17:45, H-HS IV

Runtime optimisation of Adversarial Neural Networks in the tW dilepton channel using the ATLAS detector — •Nicolas Boeing, Ian C. Brock, and Christian Kirfel — Physikalisches Institut, Bonn, Deutschland

Neural networks have proven effective for signal to background separation in high energy physics. These classifier networks can be highly sensitive to systematic uncertainties. A possible solution is the use of an adversarial neural network, a technique that pits two networks against each other. The first network has the classic task of separating signal and background, while a second adversarial network attempts to separate nominal from systematic samples, based on the output of the first network. By minimising the separation of the adversarial network, the classifier can be made more robust with respect to systematic uncertainties. This type of network structure has been shown to work for training on Monte Carlo simulated tW dilepton signal events and tt background events using the ATLAS detector, but a significant downside of training in this channel has been computation time. In this talk, we introduce methods to reduce training time using GPUs. Based on this improved performance, further improvements to the network are presented.