神经网络设计的特征空间序贯划分算法

A Geometrical Strategy of Constructive Initial Neural Networks

1引言 神经网络已广泛地用于处理实际问题,如语音处理、图像处理和计算机视觉、模式分类和识别等.面对越来越复杂的应用,传统的神经网络学习算法变得不能适应.与许多其他的有效算法相比,神经网络的学习速度慢和固定拓扑结构的不适应性两个缺陷显得异常突出.

A geometrical strategy for constructive neural networks is proposed in the paper. Firstly it can acquire quickly initial input weight parameters and topolgy by sequentially partioning feature space with the presented geometrical method. Secondly with SVD,its initial output weights are obtained very quickly. Finally these weights are retuned with BP algorithms. Its distinctive features are that it can construct quickly an initial neural networks using geometrical method other than backpropagation algorithms so that overtraining and undertraining are avoided automatically,and experimentally it performs better on two-spiral classification than Cascade-Correlation Algorithm.