Trainability of Parameterized Quantum Circuits

Abstract: Variational quantum algorithms (VQAs) and Quantum Neural Networks (QNNs) have emerged as promising strategies to achieve quantum advantage on near-term quantum devices. The success of VQAs/QNNs depends on several factors, including the trainability and expressibility of parameterized quantum circuits (PQCs). Along with numerical experiments, rigorous analytical results are necessary to guarantee the scalability of these algorithms. In this seminar, I will first summarize well-known results on barren plateaus, where certain circumstances lead to exponentially vanishing gradients. Then I will present our recent results establishing a fundamental relationship between expressibility and trainability of PQCs. Next, I will outline our analysis on the trainability of perceptron-based QNNs. One common assumption to avoid barren plateaus is to employ problem-inspired PQCs. I will present that for problem-inspired PQCs, such as Quantum Alternating Operator Ansatz (QAOA) and Hamiltonian Variational Ansatz (HVA), trainability depends on the controllability of the system and is not always guaranteed. Finally, I will describe the effect of hardware noise on the training landscape for a generic PQC and summarize potential strategies to avoid trainability issues.

quantum computing and information

Audience: researchers in the topic

Comments: To request the zoom link, please send a message cqsiadmin@uts.edu.au using your business/organisation/institution email address.

Centre for Quantum Software and Information Seminar Series

Series comments: To request the zoom link, please send a message to: cqsiadmin@uts.edu.au using your business/organisation/institution email address. Watch previous seminars on YouTube: - QSI Seminar Series 2021 (https://youtube.com/playlist?list=PLux7B14QYkPbDDOpqKSWScHXHodiBwr48) - QSI Seminar Series 2020 (https://youtube.com/playlist?list=PLux7B14QYkPZREUXReOq01ewLl02QXBXa)