一种能简化复杂分类问题的矩阵模块神经网络分类器

主要针对大训练集和类别非对称训练集等复杂分类问题提出一种基于新的任务分解技术的矩阵模块神经网络分类系统,它将一个复杂分类任务分解为多个简单的子任务来解决,每个子任务只是在两个子空间内进行,且由一个具有简单结构的神经网络模块来完成;所有网络模块将组成一个神经网络矩阵,最终将该神经网络矩阵的输出矩阵集成得到最终分类结果.本文通过理论分析和模拟实验证明,该矩阵模块神经网络能节省神经网络的学习时间,提高泛化能力和分类精度.

Classifier of Matrix Modular Neural Network to Simplify Complex Tasks

A modular neural network called matrix modular neural network(MMNN),adopting a novel task decomposition technique to solve the complex problem,has been developed in this paper.A complex problem could be decomposed into many easier problems in subspaces,each of which can be also solved by a single perceptron.All of these perceptron modules form a perceptron matrix structure,which produces a matrix of outputs that will be fed to an integration machine so that finally a classification decision result could be efficiently made.It was shown that the MMNN could reduce time consumption and improve generalization capability by our theoretic analysis and experiments.