An investigation into quantum machine learning based image classification

Easom-McCaldin, Philip (2021) An investigation into quantum machine learning based image classification. Doctoral thesis, Northumbria University.

Preview

Text (Doctoral thesis) easom-mccaldin.philip_phd(15002750).pdf - Submitted Version Download (1MB) | Preview

Abstract

Image classification is dominated by high-performing deep learning methods, that typically take on the form of large scale convolutional neural networks. Over the past decade, the magnitude of these systems has grown to the point where a moderately-sized system can consist of millions of trainable parameters. The subsequent impact on the computational load during optimization and inference is massive, therefore a change in approach is needed to reduce the scale of this problem. Quantum computing is a modern development that utilises natural quantum-mechanical principles, in an effort to efficiently process data in a classically-intractable manner. Furthermore, quantum machine learning, the application of quantum computing for machine learning tasks, has seen a surge in interest and development. This makes quantum computing an appealing area of research to consider for solutions to the problems faced. Firstly, this thesis investigates the classical-based solution of transfer learning to reduce the computational load of optimizing deep learning algorithms. Following this, quantum-based methods are analysed to determine their effectiveness for the task of image classification. This culminates in the proposal of a novel quantum image classification algorithm. This thesis makes several contributions of knowledge to the working area. Firstly, it is demonstrated that a computational speedup can be gained via quantum routines over classical algorithms for image classification. However, if a classical approach is preferred, then it is presented that transfer learning can maintain classification performance whilst negating the costly optimization of a large proportion of parameters. Secondly, an enhanced understanding of single-qubit encoding is gained. Experimental results show that substantial classification accuracy improvements can be made as data encodings are repeated. In addition, results support that an element of robustness to environmental noise can be gained for repeated encodings, which is important to consider in the NISQ era. Finally, a novel quantum image classification algorithm is proposed, which demonstrates that a lone qubit is a capable image classifier. Results determine that classification accuracies in the 90th percentile can be achieved using a minimum of 6 working parameters. Overall, this research may have a large impact towards the development of quantum image classification algorithms, where a plethora of options for future development are opened as well.

Actions (login required)

View Item

Downloads