前向神经网络泛化问题研究

针对前向神经网络泛化问题,从函数论的角度分析了影响前向神经网络泛化性能的因素。为了提高网络的泛化性能,从理论上分析指出,在网络学习过程中通过增加隐含层神经元来降低网络最大固有误差和最大样本误差的同时,要求确保网络泛化定义域尽可能接近网络输入定义域,否则将有可能降低网络的泛化性能。通过数值试验验证了上述结论。

Research on generalization capability of feedforward neural networks

To investigate generalization capability of feedforward neural networks,the influencing factors of generalization capability of feedforward neural networks are analyzed according to function theory.In order to improve generalization capability of feedforward neural networks,the convinced networks generalization domain should be guaranteed to be close to networks input domain as maximal intrinsic error of networks output and maximal samples error are reduced by increasing hidden neurons in number in the progress of networks learning,otherwise generalization capability of feedforward neural networks is likely to be decreased.The above conclusions are illustrated through concrete test.