一种提高神经网络泛化性能的罚项最优脑外科模型  

Penalty optimal brain surgeon model to improve neural network generalization

作　　者：吴翠娟[1] 李冬[1] 嵩天[2] 

机构地区：[1]苏州经贸职业技术学院机电系,江苏苏州215009 [2]北京理工大学计算机学院,北京100081

出　　处：《计算机应用研究》2010年第4期1280-1283,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(60803002);江苏省"青蓝工程"青年骨干教师资(2007)

摘　　要：最优脑外科过程是一种训练后网络剪枝算法,计算的复杂度非常高,通过把剪枝条件以惩罚项的形式纳入神经网络的训练目标函数中,把正则化方法的结构优化蕴涵于网络训练过程,构建面向最优脑外科过程的计算模型,实现网络训练过程和最优脑外科过程并行剪枝,既保持了最优脑外科过程的准确性,又具有正则化的高效性,提高了神经网络模型的泛化性能。该模型在理论上具有收敛性,其有效性和可行性通过给出的Leven-berg-Marquardt方案仿真实验也得到了说明。The penalty optimal brain surgeon （OBS） is a post-training algorithm and it has extreme high complexity.The OBS oriented compute model implemented neural network training and OBS pruning simultaneously,by taking the OBS pruning case as a penalty term of neural network objective functions based on optimized structure of the regularization method included in the neural network training process.It both maintained the OBS＇s accuracy and had regularization method＇s high efficiency.Raised the generalization of neural network model.It is theoretically convergent,the validity and feasibility are proved by given Levenberg-Marquardt solution and experiments.

关 键 词：神经网络 罚项最优脑外科模型 Levenberg-Marquardt方案 泛化能力 收敛性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]