一种基于逐维扩展的前向网络训练方法

A Training Algorithm for Feedforward Neural Networks by Gradually Increasing Demensionality

前向网络的快速训练问题是前向网络研究的一个非常重要的课题。本文针对一类n-维超立方体的分类问题（当为二分类问题时，这实际上是一个n-维Boole函数的神经网络实现问题），提出了一种基于逐维扩展的前向网络快速训练方法，将一个n个输入的大网络的各权训练问题转化为小网络逐维递归的扩展部分的参数训练问题，提高了网络训练的速度，实验结果表明了这种训练方法的有效性和可行性。

The trainning of feedforward neural networks (FNNs) is one of most important problern in the research of FNNs. This paper, towards to calssification of n-dimensional hypercube, actually, the realization of n-dimensional Boole an functions with FNNs, suggests a fast trainning algorithm based on the gradual increase of the dimensionality, which transform the weight trainning of a large network into the trainning of the weight of the increased part of the network, with the network trained in such a way recursively. Finally, the computer simulation results show the effectiveness and attractiveness of the method.