Theory Seminar: Identification of top quarks using Deep Neural Networks.
Speaker: Wojciech Fedorko, UBC
Location: Theory Room
Time: 13:00
The identification of "fat jets" originating from hadronically decaying top quarks is increasing in importance in searches for physics beyond the Standard Model as well as in precision Standard Model measurements at the LHC. In this talk I will introduce briefly the approaches to this problem currently employed by ATLAS and CMS as well as recently developed methods treating jets as an image. In contrast, the approach developed by our group uses a Deep Neural Network, processing the momenta of jets constituents. This strategy does not result in a loss of information during pixelisation or the calculation of high level features. I will discuss lossless processing of the jet information, the dense network architecture and show the performance of the network. We further develop this approach utilizing Long Short-Term Memory (LSTM) layers which are capable of learning long-range dependencies in sequential data. In an analogy to natural language processing, I will discuss the approaches to jet constituent sequence ordering and their impact on the performance of the classifier.