基于输出误差与偏导数误差信息融合的神经网络训练

文章首先提出了表示前向神经网的泛化能力的一种度量,分析了提高网络泛化能力的主要途径,进而提出了基于网络输出误差与输出对输入偏导数误差信息融合的网络训练策略,给出了两者信息融合的有效方法和相应网络训练算法。具体应用结果表明所提出算法可显著提高网络的泛化能力。

A New Approach for the Training of Feedforward Neural Network Based on the Information Merging of Output Errors and Derivative Errors

In this paper,a new measurement of the generalization capacity of feedforward neural networks is proposed.The important approaches to improve the generalization capacity are discussed.The efficient computational techniques for training of feedforward neural networks based on information merging of output errors and derivative errors are given.The effectiveness of the proposed algorithm is demonstrated by several examples.The results of the experiments proves that the algorithm can improve the generalization capacity of feedforward neural networks remarkably.