Quantum neural network for MNIST classification

Abstract

The field of quantum computing is a relatively new and intriguing area that many scientists and academics believe has great potential, and this field is continually developing. Quantum computers and quantum bits are the main core of quantum computing. These computers are capable of operating several tasks at once. One remarkable method in quantum processing and computation is Quantum Neural Network (QNN), which is inspired from classical neural network models in machine learning. Quantum Neural Networks (QNNs) are usually composed of three layers of encoder, data processor, and measurement. This computation method is suggested to be much faster and more efficient compared to the classical one. While quantum neural networks offer numerous advantages, they also face challenges similar to any other system, which require careful assessment and innovative approaches to overcome. One of these challenges is quantum noise, which is in the nature of quantum systems and decreases the accuracy of QNN models. Moreover, the performance of QNNs can be improved further by combining with other predictive models or considering some performance enhancement approaches.

In this master thesis, first the recent research on quantum computing, quantum machine learning, quantum neural networks, hybrid quantum-classical models, and noise reduction in QNNs is explored to provide background knowledge on this subject. Then, two Hybrid Artificial-Quantum Neural Network (HAQNN) structures are suggested that benefit from both quantum and classical neural network structures. Following HAQNN models, three approaches are proposed and discussed that can enhance the performance of QNN. These approaches involve utilizing ensemble learning for HAQNN structures, applying post-measurement normalization, and employing the Mixture of Experts (MOE) technique for the QNN model. The results of all the proposed architectures and methods are evaluated through running several experiments on three datasets of MNIST, Fashion MNIST (FMNIST), and CIFAR-10.