Quantum Neural Networks: Architecture Design and Quantum Training

Abstract

Quantum Neural Networks (QNNs) are promising machine learning models with potential quantum advantages over classical neural networks. This thesis focuses on their architecture design, training methodologies, and certain applications, addressing three challenges in QNN research: overcoming barren plateaus in training, designing problem-specific QNN models, and tackling state-of-the-art classical machine learning models. The thesis is divided into three main parts, each targeting a specific challenge. The first part proposes quantum-optimization-powered training methods that exploit hidden structures in the QNN optimization problem to mitigate the barren plateau issue. The second part designs problem-tailored QNNs for graph-structured data, incorporating inductive biases into their architectures to enhance trainability and generalization. The third part explores the quantum implementation of Generative Pre-trained Transformers (GPT) — the original version of ChatGPT. By addressing these challenges, this thesis contributes to advancing the field of Quantum Machine Learning, offering new insights and methodologies for designing and training QNNs.