Abstract: To ensure?the?classification accuracy?of?the?neural network?of?high-dimensional data, it proposes to firstly?reduce its dimension and then to do classification. And it in fact achieves?the dimension reduction of high-dimensional?data by Principal Component?Analysis（PCA）. By?analysis of?the?traditional BP algorithm, the?proposed?disturbance?BP?learning method is divided into two steps?to update?the network weights. It analyzes the?classification accuracy and error?convergence rate?of the algorithm through MATLAB. The simulation results show that firstly reducing its dimension and then doing?classification?of high dimensional data employing disturbance?BP network?can greatly improve the?classification accuracy and?training speed?of data.

Key words: high-dimension data, neural network, Back Propagation（BP） algorithm, high-order differential, perturbed Back Propagation（BP）

摘要： 为确保高维数据的神经网络分类精度，提出了先降维后分类的方法。采用主成分分析（PCA）法实现高维数据的降维。通过分析传统BP算法，提出分两步来更新网络权值的扰动BP学习方法。采用MATLAB对降维分类算法的分类精度和误差收敛速度进行分析。仿真结果显示：先降维再采用扰动BP网络进行高维数据分类可大大提高数据的分类精度和训练速度。

关键词: 高维数据, 神经网络, 反向传播（BP）算法, 高阶微分, 扰动反向传播（BP）