Aachen 2019 – scientific programme

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

T: Fachverband Teilchenphysik

T 4: Deep Learning I

T 4.6: Talk

Monday, March 25, 2019, 17:15–17:30, H06

Hyperparameter optimization of Adversarial Neural Networks in the tW dilepton channel using the ATLAS detector — •Christian Kirfel, Ian Brock, and Rui Zhang — Physikalisches Institut, Bonn, Deutschland

Neural networks are widely used for signal to background separation in high energy collider physics. Neural networks trained on Monte Carlo simulations can be highly sensitive to systematic uncertainties. A proposed technique to diminish this sensitivity is an adversarial neural network consisting of two networks that are trained against each other. In our case, the first network tries to separate between signal and background, while the second network tries to separate between a nominal signal sample and a signal sample with different settings. We are using a Minimax decision rule to achieve a good signal to background separation for the first network and a poor nominal to systematics separation for the second network. In this talk an adversarial neural network trained on tW dilepton channel Monte Carlo simulations with tt background using the ATLAS detector is introduced. Testing and tuning of the hyperparameters is presented for both networks as well as a comparison to a single neural network approach. Lastly the dependence for both approaches on systematic uncertainties is investigated.