Practicality of training a quantum machine in the NISQ era

Advancements in classical machine learning have ignited sparks of Artificial General Intelligence, through sophisticated neural network architectures and large datasets. However, inherent limitations in terms of energy, resource and efficiency are still prevalent on these classical devices. Quantum computing presents a tantalizing solution to transcend these barriers by leveraging a fundamentally different computational approach. Yet, this new approach also presents its own set of difficulties. Recent explorations in variational quantum computing present a promising avenue for implementing machine learning algorithms on quantum platforms. Notably, theoretical and experimental studies of the Data Re-uploading Algorithm have provided an outline for its practical application. In this experimental study, we explore the feasibility of end-to-end training for a supervised learning model on a hybrid quantum system. This system integrates an ion trap quantum computer with a classical processor. We evaluate the performance of various genetic and gradient-based (classical) optimizers as they traverse through the noisy and complex optimization landscape inherent in Noisy Intermediate-Scale Quantum (NISQ)-based quantum computers. We experimentally demonstrate the superiority and robustness of genetic-based optimizer within the scope of a binary classification problem and achieve a training accuracy of 93% and a test accuracy of 92%. These results offer valuable perspectives on the operational performance of quantum-classical hybrid systems, emphasizing the critical role of an efficient training scheme and the importance of hardware considerations in terms of practical quantum machine learning applications.