Bonn 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Bonn musste abgesagt werden! Lesen Sie mehr ...

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 47: Neural networks and systematic uncertainties

T 47.6: Vortrag

Mittwoch, 1. April 2020, 17:45–18:00, H-HS IV

Evaluation of performance gains in the training of neural networks for HEP analysis applications using GPUs — •Michael Holzbock, Günter Duckeck, and Klaus Dolag — Ludwig-Maximilians-Universität München

It has become more and more popular to tackle typical analysis task in high-energy physics (HEP) with approaches based on machine learning (ML) techniques. The training of such algorithms is computationally demanding but can be accelerated by the usage of GPUs, which architectures are better suited to perform the underlying calculations.

Studies are presented which evaluate the speed-up in case the training of ML algorithms is accelerated by GPUs. Several test cases are considered, that involve ML techniques commonly used in HEP data analysis such as deep and convolutional neural networks. To study whether benefits of GPUs depend on a network’s layout, several of its hyperparameters such as the number of neurons were systematically varied. Finally, the results of these to some extend artificial test cases are compared with the acceleration observed for a neural network configuration used in a search for physics beyond the Standard Model.