论文部分内容阅读
介绍了一种构造神经网络的新方法 .常规的瀑流关联 (Cascade-Correlation)算法起始于最小网络(没有隐含神经元 ) ,然后逐一地往网络里增加新隐含神经元并训练 ,结束于期望性能的获得 .我们提出一种与构造算法 (Constructive Algorithm)相关的快速算法 ,这种算法从适当的初始网络结构开始 ,然后不断地往网络里增加新的神经元和相关权值 ,直到满意的结果获得为止 .实验证明 ,这种快速方法与以往的常规瀑流关联方法相比 ,有几方面优点 :更好的分类性能 ,更小的网络结构和更快的学习速度 .
A new method to construct neural network is introduced.A conventional Cascade-Correlation algorithm starts from the smallest network (without implicit neurons) and then adds new implicit neurons to the network one by one and trains them, At the end of the desired performance gain.We propose a fast algorithm that is related to the Constructive Algorithm, starting with an appropriate initial network structure and then adding new neurons and related weights to the network, Until satisfactory results are obtained.Experiments show that this fast method has several advantages compared with the conventional waterfall cascade method: better classification performance, smaller network structure and faster learning speed.