基于低维参数空间的神经网络结构设计方法

Neural network design approach based on parameter space in low-dimension manifold

本文研究了流形学习的分类器网络结构优化设计问题,针对利用神经网络对同一对象的非线性结构样本集进行分类和识别时,如何合理地设计网络结构的问题,提出了一个新颖的基于低维参数空间估计的神经网络结构设计的方法。该方法以流形学习为基础,结合Sammon系数有效估计出低维参数空间大小,并将此对应到神经网络结构分组设计的隐节点分组数目上,从而设计出具有一定泛化能力的网络结构。实验结果表明了本文所提方法的有效性。

Optimization design of classifier structure was studied based on manifold learning. A novel approach of neural network design based on parameter space in low-dimension manifold was proposed to solve the problems of designing neural network rationally, which is used in recognition and classification of congener samples with non-linear configuration. This method is based on manifold learning and combines Sammon stress in order to estimate the size of parameter space in low-dimension, furthermore this size corresponds with the number of hidden nodes in neural networks. Experimental results clearly demonstrate that the proposed method is effective.