前馈神经网络结构自删除算法的研究

Self-deleting Algorithm of Feedforward Neural Networks

综述了利用删除法进行前馈神经网络设计的研究现状 ,并在重点分析根据隐节点输出相关性进行自删除的几种算法的基础上 ,在一个较高层次上提出了一种新的隐节点自删除算法 .算例说明了这种算法不仅可以压缩线性相关隐节点 ,而且可以删除不重要的隐节点 ,其重新计算量也大大减小 .

In this paper, recent development of self deleting algorithm of feedforward neural networks is surveyed. On the basis of the analysis of the self deleting algorithm according to the similarity of hidden nodes output, a new self deleting algorithm of hidden nodes, which synthesized the similarity of hidden nodes and the importance of them,is put forward. The example proves that this algorithm not only reduces linearly related hidden nodes ,but also deletes unimportant nodes,reducing computation greatly.