论文部分内容阅读
严谨的线性判别函数与判别面的理论 ,适用于线性阈值 (MP模型 )神经元分类行为分析。本文将此理论扩展到非线性 sigmoid神经元 ,分析了用来解决模式分类问题的、由 sigmoid神经元构成的单隐层 MLP(多层感知机 )的内部行为 ;并通过一系列近似推理与实验验证 ,提出了将隐层权重矢量初始值均匀地分布在权重空间的一个圆 (超球面 )上的方法。针对几个困难的分类问题的实验表明 ,该方法抓住了 MLP分类器内部行为的重要特征 ,它使 MLP分类器跨越了可能存在的学习难点 ,把学习起点放在达到目标较简便的路经上。此方法在理论上简单直接 ,应用上方便有效 ,具有一定的普遍性。
Rigorous linear discriminant function and discriminant theory, suitable for linear threshold (MP model) neuron classification behavior analysis. This paper extends this theory to non-linear sigmoid neurons and analyzes the internal behavior of a single hidden layer MLP (multi-layer perceptron) that is composed of sigmoid neurons to solve the pattern classification problem. Through a series of approximate reasoning and experiment The method of uniformly distributing the initial value of hidden layer weight vector to a circle (hypersphere) in weight space is proposed. Experiments on several difficult classification problems show that this method grasps the important characteristics of the internal behavior of MLP classifier. It makes the MLP classifier straddle the possible learning difficulties and places the learning starting point on the simpler path to achieve the goal on. This method is simple and direct in theory, convenient and effective in application, and has certain universality.