神经网络敏感性的定义分类研究

研究了神经网络的输出变量对输入变量和权重变量的敏感性定义,通过分析这些敏感性定义的本质,把神经网络的敏感性定义分为3类:分析学,几何学,物理学.在这3种敏感性定义分类中,对分析学和物理学定义的敏感性进行了模拟实验,实验结果表明,物理学定义的敏感性更适合应用在径向基函数神经网络删除冗余属性上.在解决其他网络的冗余属性问题上,需要进一步研究.

Research on the Classification of Definitions of Sensitivity for Neural Network

The definitions of sensitivity of the output variables to input variables and weight variables for neural network are analyzed in this paper.The definitions of sensitivity for neural network can be classified into three classes:analytics,geometry and physics.Some simulative experiments on the sensitivity defined by analytics and physics have been dene,experimental results indicate that the sensitivity defined by physics is more suitable for deleting redundant attributes on radial basis function neural netwo...