快速自底向上构造神经网络的方法

Fast Approach for Cascade-Correlation Learning

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 for constructing feedforward neural networks is introduced. The standard Cascade-Correlation Learning is to start with a minimal network (no hidden units), and then train and add new hidden units one by one to the network until desired performance is reached. We propose a fast approach, which corresponds to constructive algorithm, starts with an appropriate network architecture and then grows additional hidden units and weight until a satisfactory solution is found. Experimental result demonstrates that with the fast approach, considerable performance gains are obtained compared to the standard Cascade-Correlation Learning. This includes better classification, smaller network size, and faster learning.