Particles and Fields Seminar — HEP seminar
Beauty Fragmentation Tagging: Leveraging Primary Vertex for Weakly Supervised Learning
Mr. Matan Yosef
BGU
Abstract
Many processes at the LHC produce energetic b-quarks that hadronize into b-hadrons. Identifying which b-hadron is produced is important for many analyses, yet most infer it from a single exclusive decay channel, usually a decay into fully charged final particles. This approach is limited by small branching ratios and an irreducible background from competing decays with the same final state. We instead propose an inclusive fragmentation tagger based on machine learning.
The central challenge is the training data. A model trained on simulation inherits the poor modelling of low-energy, non-perturbative QCD. Training on real collisions removes that dependence but introduces a bias toward the small subset of reconstructible decay topologies.
We address this with BLIND (Bias-free Labeling for INclusive Decays), a two-stage algorithm. It builds on the fact that the particles produced during fragmentation at the primary vertex have some correlation with the b-hadron's identity, regardless of the decay channel. A first network infers the b-hadron species from primary-vertex information alone, assigning pseudo-labels to a fully inclusive sample. A second, weakly supervised network then trains directly on this sample, learning to separate species from the decay. The two outputs may be combined to exploit both the fragmentation and decay views for the final prediction. The result is a tagger trained entirely on data, free of simulation and of decay-selection bias, that leverages the full jet information.
The central challenge is the training data. A model trained on simulation inherits the poor modelling of low-energy, non-perturbative QCD. Training on real collisions removes that dependence but introduces a bias toward the small subset of reconstructible decay topologies.
We address this with BLIND (Bias-free Labeling for INclusive Decays), a two-stage algorithm. It builds on the fact that the particles produced during fragmentation at the primary vertex have some correlation with the b-hadron's identity, regardless of the decay channel. A first network infers the b-hadron species from primary-vertex information alone, assigning pseudo-labels to a fully inclusive sample. A second, weakly supervised network then trains directly on this sample, learning to separate species from the decay. The two outputs may be combined to exploit both the fragmentation and decay views for the final prediction. The result is a tagger trained entirely on data, free of simulation and of decay-selection bias, that leverages the full jet information.