A Robust Quantum Neural Network Design

With the advent of the era of big data,machine learning in classical fields,especially neural networks,is a research hotspot in computer science.However,in the face of the exponential growth of data generated every year around the world,the computing power of classical computers is approaching the bottleneck,and machine learning algorithms are facing enormous challenges.Quantum computing has shown a significant speedup for some specific problems(such as factorization of large numbers,unstructured database search,etc.).Therefore,combining the advantages of neural networks with quantum computing to propose new quantum machine learning algorithms has received more and more attention.However,current medium-scale noisy quantum computers severely limit the application of quantum neural networks.Based on the above research background,this paper conducts related research on the design of the quantum neural network.The main work is as follows:First is the structural design scheme of the quantum neural network.In designing a quantum neural network,three parameters to evaluate the circuit performance are used to expression ability: entanglement ability,and dispersion.To improve the robustness,avoid a barren plateau,and optimize the circuit performance,the construction scheme of the quantum circuit is studied.A circuit optimization strategy is proposed.The four circuit structures of the quantum circuit,the spin layer,entanglement layer,the connection method of the entanglement layer,and circuit depth,are experimentally verified by the control variable method by assuming the factors that affect the circuit’s performance.Finally,the quantum circuit model with better comprehensive performance is selected and applied to the following algorithm in the experimental circuit.Second,the regularization quantum Wasserstein distance is designed.Due to the unique properties of quantum information,such as superposition and entanglement in quantum data,the classical measurement method is no longer suitable when measuring quantum data.This paper draws on the Wasserstein distance,which can solve the gradient vanishing and gradient exploding problems of classical generative adversarial networks,and proposes a quantum Wasserstein distance that can measure quantum data according to symmetric subspace.At the same time,a regularization quantum Wasserstein distance with quantum relative entropy added is designed to solve the problem that the Lipschitz parameter needs to be restricted by the quantum Wasserstein distance.Finally,a regularization quantum Wasserstein generative adversarial network-based on optimistic mirror descent is designed.This paper creates a regularization quantum Wasserstein generative adversarial network based on the proposed regularization Wasserstein distance.The optimistic mirror descent algorithm is used for the gradient optimization algorithm to optimize the limit cycle problem in quantum generative adversarial networks.In the pure quantum state and mixed state convergence experiments,the designed quantum generative adversarial network optimization for limit cycle problems is verified.In the anti-noise experiment,the robustness of the designed quantum generative adversarial network is demonstrated.