论文部分内容阅读
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.
Aiming at the complex classification problems such as large training set and asymmetrical training set, this paper proposes a matrix module neural network classification system based on the new task decomposition technique, which decomposes a complex classification task into a plurality of simple sub-tasks to solve each The sub-tasks are only performed in two subspaces and are completed by a neural network module with a simple structure. All network modules will form a neural network matrix, and finally the output matrix of the neural network matrix will be integrated to obtain the final classification result.Through this paper, Theoretical analysis and simulation experiments show that this matrix module neural network can save learning time of neural network and improve generalization ability and classification accuracy.