Deep Sets based Neural Networks for Impact Parameter Flavour Tagging in ATLAS

This work introduces a new architecture for Flavour Tagging based on Deep Sets, which models the jet as a set of tracks, in order to identify the experimental signatures of jets containing heavy flavour hadrons using the impact parameters and kinematics of the tracks. This approach is an evolution with respect to the Recurrent Neural Network (RNN) currently adopted in the ATLAS experiment, which treats track collections as a sequence. The Deep Sets model comprises a permutation-invariant and highly parallelisable architecture, leading to a significant decrease in training and evaluation time, and thus allowing for much faster turn-around times for optimisation. Additionally, this permutation invariance encoded in the model is more physically motivated than the sequence-based RNN. We compare the Deep Sets algorithm with the RNN benchmark, probe the model to interpret the information learned, and provide studies optimising the Deep Sets algorithm by loosening the track selection and including additional inputs.