Abstract: In order to explore the possibility of cross fusion between quantum computing and neural networks, a digital recognition model based on hybrid quantum classical neural networks is proposed. In this model, the parametric quantum circuit is added to the convolution model to realize the mixing of embedding and direct output. The convolution full connection layer information is taken as the control parameter of the quantum layer, and the feature is processed through a series of unitary transformations. The potential of this structure is demonstrated by digit recognition experiments on MNIST datasets. Combined with transfer learning, convolution networks and hybrid networks are used to extract deep-seated features of images and perform feature fusion to improve the overfitting problem of single convolution networks on small sample datasets. The experimental results show that the model has validity, practicability and high generalization performance. On MNIST datasets, the recognition rate of direct output hybrid network and embedded hybrid network is 0.039 8 and 0.021 1 higher than that of convolution network, respectively. After transfer learning and feature fusion, the recognition rate of the hybrid model in small sample datasets is better than that of the single network, reaching 83%, which verifies the learning and generalization ability of the model.

Key words: quantum computing, hybrid quantum neural network, transfer learning, feature fusion

摘要： 为探索量子计算与神经网络交叉融合的可能性，提出一种基于混合量子经典神经网络的数字识别模型。该模型将参数化量子电路加入卷积模型实现嵌入与直接输出混合，将卷积全连接层信息作为量子层控制参数，经过系列酉变换进行特征处理。在MNIST数据集上通过数字识别实验展示该结构潜力。结合迁移学习，利用卷积网络与混合网络提取图像深层次特征并进行特征融合，改善单一卷积网络在小样本数据集上的过拟合问题。实验结果表明，该模型具有有效性、实用性和高泛化性能，直接输出混合网络和嵌入式混合网络在MNIST数据集上的识别率分别较卷积网络提高0.039?8和0.021?1。经过迁移学习和特征融合后的混合模型在小样本数据集上的识别率均优于单一网络，达到83%，验证了模型的学习和泛化能力。

关键词: 量子计算, 混合量子神经网络, 迁移学习, 特征融合