人工神经网络变量选取与隐藏单元数的确定

根据多隐藏层所有训练样本误差平方和最小设计优化问题,求解并绘出计算流程图。Trevor等人认为隐藏单元过多比过少好,交叉验证估计（隐藏单元）正则化参数没有必要。还有一种通常做法是常常利用分类树挑选变量作为输入变量进行人工神经网络建模。而从人工神经网络与多元统计、传统回归和其他数据挖掘工具的区别和联系出发,认为这些观点和做法值得商酌;用ZIP编码实例说明隐藏单元过多不一定比过少好,实际数据分析中所需隐藏单元数的确定可以用交叉验证结合经验判断来实现,利用分类树选择的变量对于人工神经网络没有太大的效果;通过分类树删节变量以降低计算量的效果不如通过压缩隐藏单元个数降低计算量来得好;非完全问题“从简单到一般”思想与完全问题中选择所有变量的思想不矛盾。在总结了Le Cun等人的局部联结以有效降低权数思想的基础上,提出通过随机选择人工变量建立人工神经网络分布式模型系统的设想。

The Determination of the Number of the Hidden Cells and the Variable Selection of the Neural Network

This article gives the solution and the flow chart by way of a optimization model that minimize total sum of error.According to some masters such as Trevor,it's better to take hidden cells,there is no need to use CR to achieve it.And CART is frequently used before ANN to delete some variables.With the example of ZIP code,you will see in this article that too more hidden cells may be worse than less,CR may be useful,using CART to delete some variable do little help to ANN model,along with the viewpoint that there is no paradox between the idea that "from simple to general" in non-complete problem and the idea that choosing all variable in complete problem.In the end,this article proposed some thought to improve.