Machine Learning is having a major impact in jet physics. It is empowering powerful taggers for boosted (W, Z, H, top) jets as well as flavor tagging.
TreeNiNThe QCD-aware recursive neural networks devleoped as part of DIANA/HEP (Louppe et al. 2017 which leverage an analogy to natural language processing were extended to include a network-in-network. The TreeNiN method (in the table below), achieves excellent performance with orders of magnitude fewer parameters than the other top performing techniques. This pytorch implementation can be found in this repository
Top Tagging ComparisonAt ML4Jets 2018 a top-tagging comparison was initaited, resulting in ``The Machine Learning Landscape of Top Taggers''. The table shows a comparison of the methods. Irina Espejo, Sebastian Macaluso, and Heiko Mueller are using docker containers, yadage workflows, and REANA to automate and streamline such benchmark studies.
Ginkgo: growing convergent researchMore recenlty, Sebastian Macaluso and Kyle Cranmer have worked to reframe many problems in jet physics in statistical terms and connect the parton shower to latent variable models. In order to ease collaboration with computer scientists, they have developed Ginkgo, a simplified parton shower model written in pyro.
Visualizing Jet ClusteringIn order to aid research into jet physics, Sebastian Macaluso, Kyle Cranmer, and Duccio Pappadopulo have developepd new visualization tools to compare various jet clustering algorithms. See the VisualizeBinaryTrees repository in GitHub.
Hierarchical Clustering: use-inspired researchHierarchical clustering is a common clustering approach for gene expression data. Within particle physics hierarchical clusterirng appears in the context of jets -- the most copiously produced objects at the Large Hadron Collider. One challenge is that the number of hierarchical clusterings grows very quickly with the number of objects being clustered. IRIS-HEP researchers Sebastian Macaluso and Kyle Cranmer connected with computer scientists at U. Mass Amherst to extend a clustering algorithm they had developed for the hierarchical case. This algorithm was applied to both particle physics and cancer genomics studies in Compact Representation of Uncertainty in Hierarchical Clustering.
- Kyle Cranmer
- Sebastian Macaluso
- Irina Espejo
- Compact Representation of Uncertainty in Hierarchical Clustering, C. Greenberg, S. Macaluso, N. Monath, J. Lee, P. Flaherty et. al., arXiv 2002.11661 (Submitted to NeurIPS2020) (26 Feb 2020).
- Set2Graph: Learning Graphs From Sets, H. Serviansky, N. Segol, J. Shlomi, K. Cranmer, E. Gross et. al., arXiv 2002.08772 (Submitted to NeurIPS2020) (20 Feb 2020) [1 citation].
- Hamiltonian Graph Networks with ODE Integrators, A. Sanchez-Gonzalez, V. Bapst, K. Cranmer and P. Battaglia, arXiv 1909.12790 (Submitted to Machine Learning For the Physical Sciences NeurIPS2019 Workshop) (27 Sep 2019) [2 citations].
- The Machine Learning Landscape of Top Taggers, G. Kasieczka, T. Plehn, A. Butter, K. Cranmer, D. Debnath et. al., SciPost Phys. 7 014 (2019) (26 Feb 2019) [40 citations] [NSF PAR].
- QCD-Aware Recursive Neural Networks for Jet Physics, G. Louppe, K. Cho, C. Becot and K. Cranmer, JHEP 01 057 (2019) (02 Feb 2017) [96 citations].