前馈网络灵敏度：分析及降低算法

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。

Sensitivity of multilayer feedforward network:analysis and reduced learning algorithm

In multilayer feedforward network, different combination of weights can give approximately the same mapping. The sensitivity of the network varies with the weights. This paper proposes a method for measuring the sensitivity of the network and a learning algorithm for reducing the network sensitiyity. The sensitivity analysis covers cases such as with single output or multiple outputs, and with weight perturbation or input perturbation. The learning algorithm involves the injection of random noise into the network during its training phase. The proposed learning algorithm can reduce the network sensitivity and keep the same convergence speed comparing with traditional learning algorithms.