Hybrid Quantum-Classical Generative Models Based On Parametrized Quantum Circuit

With the rapid development of domestic and international quantum computing hardware technology,how to solve practical problems based on the current quantum computing hardware has become a research focus in quantum machine learning.The hybrid quantum-classical model combines quantum computing resources with classical computing resources,and it can fully exploit the advantages of quantum computing as well as make use of the current classical computing resources.The hybrid quantum-classical model is considered as the most suitable quantum machine learning model for the recent quantum computing hardware.Based on the current research on quantum generative models and quantum generative adversarial networks,two hybrid quantum-classical generative models were designed.The main contents of the research are as follows:The noise on the current noisy intermediate-scale quantum computer can cause gradient disappearance in the parameter optimization process of the parametrized quantum circuit.To reduce the impact of noise on the performance of the hybrid quantum-classical generative model,a quantum architecture search algorithm for improving the structure of the hybrid quantum-classical generative model was proposed.According to the types of quantum gates,the algorithm can search for the optimal circuit architecture under the specified circuit depth.The search process is accelerated by sharing parameters between the candidate quantum circuits of the same structure of parametrized quantum gates.Simulation results on the BAS dataset and the chessboard dataset indicate that the generation ability of the quantum generation model searched by the quantum architecture algorithm is better when the number of quantum gates and optimization parameters is lower than those of the quantum generation model designed based on the rotational and entanglement layers.Inspired by the existing quantum generative adversarial network model,a hybrid quantum-classical generator was designed to generate quantum distribution and Gaussian distribution.The architecture of the hybrid quantum-classical generator consists of a parametrized quantum circuit and a classical neural network.The parameters of the hybrid quantum-classical generator are updated by the adversarial learning with the parameters of the classical neural network discriminator.The quantum generative adversarial network usually suffers from mode collapse in the generation training for Gaussian distribution.It would result in the generator can only approximately generate partially correct data.To solve the problem of mode collapse,an unrolled quantum generative adversarial network model was constructed.The unrolled quantum generative adversarial network can match the generator with a better discriminator by separately training the discriminator prior to the training on the generator.Comparative experiments were performed between the unrolled quantum generative adversarial network and the quantum generative adversarial network.It is shown that the generation effect of the unrolled quantum generative adversarial network is substantially enhanced compared with that of the quantum generative adversarial network.