Machine Learning in Fundamental Physics

Machine learning techniques became more and more popular also in the context of particle physics. Our group develops new object and event classificators based on deep neural networks, but also investigate the potential of transfer learning and adverserial attacks.

Open Data

There is lots of open data from the LHC Experiments like ATLAS and CMS already published. That is of course great. However, we noticed that most computer scientists have a hard time to get used to our ROOT data format, which is typically used within high energy physics. Not only the data format itself is a problem, but also the missing understanding of what all individual variables mean. Hence we invested quite some time to transform CMS Open Data into Panda data frames, which are well known on the ML community, along with an introduction to the meaning of the different variables.

OpenData
© Tima Saale, Matthias Schott
Eine Wissenschaftlerin und ein Wissenschaftler arbeiten hinter einer Glasfassade und mischen Chemikalien mit Großgeräten.
© I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessingadversarial examples,” 2014

Adversarial Attacks

Adversarial attacks for deep neural network classifiers try to construct examples where only marginal changes of the input variables yield to a completely different result of the classification. In a typical ML context, those attacks can be used to train more robust classifiers. We try to evaluate the possibility to use adversarial attacks to estimate intrinsic systematic uncertainties of classifiers, which are used for signal/background separation in rare event searches. This work is done in the context of the AISafety-Project, funded by the federal ministry of research and education (BMBF) of Germany.

Wird geladen