Karlsruhe 2011 – scientific programme

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

T: Fachverband Teilchenphysik

T 39: Top-Produktion IV

T 39.3: Talk

Thursday, March 31, 2011, 17:15–17:30, 30.22: 130

Application of unsupervised learning methods in high energy physics — •Peter Kövesarki, Adriana Elizabeth Nuncio Quiroz, and Ian C. Brock — Physikalisches Institut Universitaet Bonn, Bonn, Germany

High energy physics is a home for a variety of multivariate techniques, mainly due to the fundamentally probabilistic behaviour of nature. These methods generally require training based on some theory, in order to discriminate a known signal from a background. Nevertheless, new physics can show itself in ways that previously no one thought about, and in these cases conventional methods give little or no help. A possible way to discriminate between known processes (like vector bosons or top-quark production) or look for new physics is using unsupervised machine learning to extract the features of the data. A technique was developed, based on the combination of neural networks and the method of principal curves, to find a parametrisation of the non-linear correlations of the data. The feasibility of the method is shown on ATLAS data.