Optimizing Tree Tensor Networks for classification on hardware accelerators

This thesis focuses on the development and optimisation of classifiers utilizing Tree Tensor Networks (TTNs). TTNs are quantum-inspired hierarchical tensor structures capable of describing the ground state of multi-body systems, and widely adopted in quantum-inspired machine learning for classification tasks. The primary objective of this study is to develop binary TTN classifiers and optimise the code for their efficient deployment on computing accelerators, such as General Purpose Graphics Processing Units (GPGPU) or Field-Programmable Gate-Arrays. Specifically, we aim to leverage CUDA-based libraries like cuTENSOR for deploying TTN classifiers on GPGPUs. To evaluate the effectiveness of the implementation, benchmarking tests using synthetic datasets will be conducted. Performance metrics, including accuracy and speed, will be analysed to assess the efficiency of the GPGPU-based TTN classifiers. The thesis include the application of TTN classifiers to tag hadronic jets. The TTN tagger implementation will be based on simulations and serves as a demonstrator for a jet tagging algorithm running at the Level-1 Trigger level of the Compact Muon Solenoid (CMS) experiment, which would utilise the trigger objects produced at the Level-1 stage following the forthcoming upgrade phase of the CMS experiment.