具有混合量子门结构的Elman神经网络及其学习算法.docVIP

  具有混合量子门结构的 Elman 神经网络及 其学习算法# 李鹏华1，熊庆宇2，柴毅3** 5 10 15 20 （1. 重庆大学自动化学院， 重庆 400044； 2. 重庆大学软件工程学院， 重庆 400044； 3. 重庆大学自动化学院，重庆 400044） 摘要：本文提出一种具有混合量子门结构的 Elman 神经网络模型及其梯度扩展的反向传播 学习算法。该网络由量子比特神经元和经典神经元构成，通过增加量子映射层以确保来自上 下文单元的局部反馈与量子映射层输出之间的模式一致。学习算法采用搜索然后收敛的学习 进度表自适应地调度学习率参数并将上下文单元的权值扩展到隐藏层的权值矩阵中，使上下 文单元权值在更新过程中获取额外的梯度信息，从而提高网络上下文单元输出与隐藏层输入 之间的匹配程度。以峰值检波为例的数值实验结果表明，在量子反向传播学习过程中，量子 门 Elman 神经网络具有较快的学习速度和良好的泛化性能。 关键词：量子比特神经元；Elman 神经网络；梯度扩展；反向传播 中图分类号：TP181 Elman neural network with hybrid quantum gate structure and it’s learning algorithm LI Penghua1, XIONG Qingyu2, CHAI Yi3 (1. College of Automation, Chongqing University, Chongqing, 400044, China; 2. School of Software Engineering, Chongqing University, Chongqing 400044, China; 3. College of Automation, Chongqing University, Chongqing 400044, China) 25 Abstract: In this paper, a new Elman neural model with quantum gate and it’s quantized extended-gradient backpropagation training algorithm are proposed for improving the performance of the conventional Elman networks. The novel model is comprised of qubit neurons and classical neurons, and the quantum map layer is employed to address the pattern mismatch between the context layer and the input laye.The learning rate is adaptive scheduling by the 30 35 40 searching and convergent learning schedule, and the weights of the context layer are extended to the weight matrix of the hidden layer for obtaining the extra gradient information, which makes the pattern of the context layer matching the pattern of the input layer with high level. The numerical experiment is carried out to verify the theoretical results and clearly show that the quantum gate Elman networks has good performance in term of both weight convergence and generalization ability. Key words: qubit neuron; Elman neural network; extended-gradient; back-propagation 0 引言 Elman 神经网络[1]是一种动态递归神经网络，本质上属于单隐藏层前馈类型的网络，其 结构同 Jordan 提出的结构[2]相似。它具有从隐藏层神经元的输出到网络输入的局部反馈连 接，该反馈连接称为上下文单元，这些单元的作用是在前时间步复制出隐藏层输出信号，以