Implementing a robust anti-QCD tagger with mass de-correlated jet image data
Researcher: Rapetsoa Kokotla, University of Venda
Supervisors: Prof. Deepak Kar, University of the Witwatersrand, Johannesburg and Dr Eric Maluta, University of Venda
This project studies a robust anti-QCD tagger with mass de-correlating jet image data produced using the pre-processing method introduced in arXiv: 1903.02032. A semi-supervised (where data is only trained on background) learning anomaly detection approach using convolutional autoencoder neural networks is explored as an anti-QCD tagger in this study. Jet image data is used to train the algorithm instead of conventional high level multivariate observables. The pre-processing steps first perform momentum re-scaling followed by a Lorentz boost transformation to find a frame of reference where any given jet is characterised by the same mass and energy, and remove the residual rotation by applying the Gram-Schmidt method on the transverse plane to the jet axis. This is expected to increase the sensitivity of the autoencoder to non-hypothesised resonance and particles as it will not experience non-linear correlation of the jet-mass with other jet observables. A negative result shows that contrary to the convolutional autoencoder outperforming autoencoder in every case where image data is used, it failed to do so in this project. The pre-processing method results in jet images data that the convolutional layer cannot extract information or features from.