模式变换与前向网络训练

Pattern Transformation for Training of Feedforward Neural Networks

前向网络在用于模式分类时 ,其网络的有效训练一直是一个受到关注的问题。本文首先提出模式的可逆线性变换不改变网络的结构 ,而组成可逆线性变换的错切变换会在一定程度上改变网络训练的难度 ,从而提出了对模式进行适当错切变换可以有效改变网络训练难度提高网络训练效率的方法。文末对所提出方法的实验结果也证明了这一点。

It has payed great attention to effective training of feedforward neural networks when they are used for pattern classification. This paper deals with the structure invariance of the network when the reversible linear transformation is conducted on it. It is shown that the cutting transform, which is a kind of reversible linear transform, can decrease the training difficulty of the network, and then could be used as an effective method for improving the training speed of the network. Finally, an experimental result is presented indicating that the method presented here is effective.