Residual Quantum Graph Recurrent Neural Networks

Quantum computing is a growing field that harnesses the peculiar quantum mechanical properties to perform computations. Quantum computers have shown huge potential in solving classically intractable problems. However, we are far from their practical use. Today, only small-scale fault-intolerant quantum computers (i.e., susceptible to noise and er- rors) exist. This era is called the Noisy Intermediate-Scale Quantum (NISQ) era. Despite their hardware limitations, developing algorithms on NISQ devices allows researchers to explore the potential of quantum computers. A potentially promising area of research in the NISQ era is Quantum Machine Learning (QML).

QML combines the fields of machine learning and quantum computing. Generally, QML can be divided into two categories. The first involves using classical machine learning to enhance quantum technologies. The second category focuses on leveraging quantum computing to enhance classical machine learning. In this thesis, we use QML defined in the second category, which usually involves algorithms that use Quantum Neural Networks (QNNs). QNNs can potentially provide exponential speedup compared to classical neural networks. However, the noise from NISQ devices significantly affects the trainability of algorithms developed in this era, which challenges their accuracy.

Motivated by the success of classical deep residual learning, we propose a novel residual model for Quantum Graph Recurrent Neural Networks (QGRNNs). To evaluate the effectiveness of our model, we conduct various experiments to compare the performance of non-residual QGRNNs with those incorporating residual connections. The experimental results have shown that our proposed design yields promising outcomes, where the cost function value outperforms the non-residual QGRNNs in almost all scenarios.