Novel deep learning methods for track reconstruction

• DOI: 10.48550/arxiv.1810.06111
• Publication Date: 2018-01-01
• Authors: Steven Farrell; P. Calafiura; Mayur Mudigonda; Prabhat; Dustin Anderson; Jean-Roch Vlimant; Stephan Zheng; J. Bendavid; M. Spiropulu; G. B. Cerati; L. Gray; Jim Kowalkowski; Panagiotis Spentz...; Aristeidis Tsaris
• Keywords: Spurious relationship; Representation

For the past year, HEP.TrkX project has been investigating machine learning solutions to LHC particle track reconstruction problems. A variety of models were studied that drew inspiration from computer vision applications and operated on an image-like representation tracking detector data. While these approaches have shown some promise, image-based methods face challenges in scaling up realistic HL-LHC data due high dimensionality sparsity. In contrast, can operate spacepoint measurements ("hits") exploit structure solve tasks efficiently. this paper we will show two sets new deep for reconstructing tracks using space-point arranged as sequences or connected graphs. first set models, Recurrent Neural Networks (RNNs) are used extrapolate, build, evaluate candidates akin Kalman Filter algorithms. Such express their own uncertainty when trained with appropriate likelihood loss function. The second use Graph (GNNs) hit classification segment classification. These read a graph hits compute features nodes edges. They adaptively learn which connections important spurious. scaleable simple architecture relatively few parameters. Results all be presented ACTS generic simulated