神经网络的学习误差函数及泛化能力

The Learning Error Function of Neural Network and Its Generalization

用于训练神经网络的样本点集不可避免地会受到噪声污染。利用神经网络的概率描述，通过研究Ｋ—Ｌ信息距离和神经网络泛化能力的关系，构造一个新的神经网络学习误差函数。泛化能力分析和仿真结果表明了该学习误差函数的合理性。

The training data is inevitably contaminated with noise. At first the neural network is represented with a probability relation. Then the K_L information distance is adopted to be a learning error function, which keeps the consistence with the generalization. Finally the effectiveness of the new learning functions is proved through generalization analysis and the simulation example.