基于卷积神经网络的反向传播算法改进

摘要

在真实音识别系统中, 针对卷积神经网络的超大规模模型参数和海量训练数据导致的训练效率问题, 提出一种缩小权值范围反向传播 (NWBP) 算法, 围绕网络参数训练后期寻找误差极小值时易出现的振荡现象, 采用K-MEANS算法获取逼近误差极小值的种子节点, 通过迭代计算过程缩小权值变化范围避免振荡现象, 使训练结果的网络误差收敛, 提高训练效率.通过仿真实验, NWBP算法在复杂卷积神经网络的权值训练过程中相比可变学习速率反向传播算法收敛效果得到提升, 一定程度上减少了冗余计算, 缩短了训练时间, 算法效果相比在简单网络中更显著.%In the real speech recognition system, in view of the problem of training efficiency caused by large-scale model parameters of convolutional neural network and mass training data, a narrowed weight range back propagation (NWBP) algorithm was proposed, which solved the oscillation phenomenon which was prone to error at the end of the network parameter training.The K-MEANS algorithm was used to obtain the seed node with the minimum error value.Through the iterative calculation process, the range of weight was reduced for avoiding the oscillation phenomenon, making the network error of training results converge, thereby, training efficiency was improved.Through the simulation experiment, the NWBP algorithm has an increase in the convergence effects comparing with the variable learning rate back propagation algorithm in the process of weight training of the convolution neural network, which reduces the redundancy calculation and shortens the training time to a certain extent.The effect of the algorithm is more significant than that in the simple network.