一种特征压缩及分类神经网络的研究

Research of A Feature Compression and Classification Neural Network

由于对多类问题的高维数据无法直接观察其聚类和分布特性，本文采用神网络法实现自适应主元特征提取（ＡＰＥＸ），以压缩特征空间的维数，并保持足够的信息来鉴别事物之间的类别。它可有效地提取信号的主要特征和抑制噪声。我们将高维数据压缩影射到２或３维，从而实现特征数据的可视性分析，显示物体对象间的类似程度和关系结构。并采用高阶函数的神经网络对其进行非线性分类，同时与ＢＰ网络的非线性分类能力进行了实验比较。结果表明高阶函数神经网络较ＢＰ网络分类能力强，训练速度快。

:Because the high-demension feature data of multiclasses problem are unable to be directly observed by their classes and distribution characterisitics,the paper achieves the Adaptive Principal Component Extraction (APEX) by using neural networks to compress the feature size and keep enough imformation to distinguish different classes.It can effective extract principal features of the signal and repress the noise. When the high demension feature data are mapped to 2 or 3 demensions, feature classes can be observed and analyzed directly and thus the relation among the feature data can be shown.A high-rank function neural network(HRFNN)is used to non-linear classification of the feature data. The non-linear classification abiltiy of the HRFNN and BP network is compared and experiment results show that the HRFNN has stronger classification ability and faster training speed.