神经网络过学习问题的统计学分析及改进算法

Statistics Analysis of Over-Fitting of Neural Networks and New Algorithm

着重分析了前向神经网络在学习过程中过学习问题产生的统计学机理 ,并根据统计学习理论提出了一种新的准则函数·该准则函数综合考虑了网络的逼近性、容错性和归纳性 ,并且算法实现起来简单 ,易于在线选择·同时还提出一种新的网络结构学习算法·最后用该学习算法进行网络训练 ,仿真结果显示了算法的有效性和实用性·

The over fitting problem of feed forward neural network was analyzed in statistics. A new criterion function was presented according to the statistical learning theory. The criterion concerns the natures of approach, fault tolerance and prediction of the neural network. The algorithm is simple and easy to be realized on line. A new learning algorithm of the network structure was developed and the particular steps of the algorithm were given. A neural network training was made by using the new algorithm. The efficiency and practicability of the algorithm were pvoved by simulation.