Research On Conditional Generative Adversarial Networks Based On Parameterized Quantum Circuits

Quantum machine learning is an emerging sub-topic in the field of quantum information.The study combines the potential acceleration capabilities of quantum computing with the learning and adaptability of classical machine learning models,and proposes new quantum machine learning algorithms or the corresponding quantum version of the classical machine learning algorithms.With the breakthrough of quantum computer in computing scale and stability,the research of quantum machine learning is also deepening.The main research object of this paper is the quantum generative adversarial network(QGAN)based on parameterized quantum circuits(PQC),which is an extension of the classical generative adversarial network in the quantum field.Through the adversarial training of quantum generators and discriminators,the model learns to fit sample sets probability distribution.The influence of parameterized quantum circuits on the performance of the algorithm is a point neglected in the current related work,and there is still room for improvement in the function and application of the algorithm.Therefore,on the basis of indepth research on the PQC structure,this paper proposes a quantum conditional generative adversarial network algorithm combined with the conditional constraint model.The main work is as follows:(1)Aiming at avoiding barren plateaus,improving the stability of PQC training and optimizing circuit performance,three indicators for evaluating the performance of parameterized quantum circuits are proposed and the structural design of quantum circuits is studied.We propose the circuit optimization strategies and assume factors that may affect the circuit performance.The controlled variable method is used to conduct experiments to verify the correctness of the optimization strategies.Finally,structures with better comprehensive performance are selected from the experimental circuits and used in the subsequent algorithm design.(2)Inspired by the classical conditional generation adversarial network,a pure quantum scheme of the conditional generation adversarial network that is suitable for both generating quantum data and classical data is proposed.By adding conditional information to the model input,the generator can learn the target distribution and has the ability to generate specific data based on the conditional information,thus realizing the expansion of algorithm functions.The preparation of the quantum state of the conditional information,the PQC design of the discriminator and generator,and the estimation method of parameters' gradient of the quantum circuit are elaborated in detail.Through the classify generation experiments of classical data and quantum mixed states,the correctness and effectiveness of the algorithm are proved.(3)Based on the quantum conditional generation adversarial network,a hybrid quantumclassical conditional generation adversarial network that is more in line with practical application requirements is proposed.This design uses the classical discriminator to complete the classification task,avoiding the "input bottleneck" of quantum machine learning.Through the classify generation experiment of BAS data,the effectiveness of the algorithm is proved.The classification and generation experiments of custom image data set and public data set(MNIST)prove that the algorithm has the potential to fit the distribution of high-dimensional classical data.