论文部分内容阅读
在前馈网络中，不同的权值组合可逼近同一映射。网络的灵敏度取决于权值的变化。文中提出了计算网络灵敏度的方法和一种降低网络灵敏度的学习算法。网络的灵敏度分析包括单输出、多输出及输入变化、权值变化等情况。学习算法是在网络训练过程中加入随机噪声。次种学习算法与传统学习算法相比，可降低网络的灵敏度，但学习收敛速度基本相同。
In feedforward networks, different weight combinations can approximate the same map. The sensitivity of the network depends on the change of the weight. In this paper, a method of calculating network sensitivity and a learning algorithm of reducing network sensitivity are proposed. Network sensitivity analysis includes single output, multiple output and input changes, weight changes and so on. Learning algorithm is to add random noise in the network training process. Compared with the traditional learning algorithm, the secondary learning algorithm can reduce the sensitivity of the network, but the learning convergence rate is basically the same.