基于卡尔曼滤波的神经网络修剪算法研究
首发时间:2009-02-20
摘要:传统的BP神经网络在应用过程中,经常面临无法确定合适的网络节点问题。网络规模小,则运算时间长;而网络规模过大,容易产生过学习现象,影响泛化能力。本文在传统的BP神经网络学习的基础上,采用Kalman滤波算法对神经网络中的权值向量进行修剪,实现对神经网络结构的简化,提高泛化能力。不同与其他修剪算法的边修剪、边训练,该算法在完成学习之后,一次性进行修剪。在入侵检测数据集中的测试表明,该方法修剪比例高,精确度好,修剪后的网络能够很好的保持原始网络的识别率,对学习速度和泛化能力的提高是有效的。
For information in English, please click here
Research of pruning algorithm in Neural Networks based on Kalman Filter
Abstract:In the application of traditional BP neural networks, we usually in face of the question that how to determine the appropriate numbers of the neurons. If the scale of the network is too small, it will cause long training time. On the contrary, if the scale is too big, the networks will lead to over fitting which plays an important role to generalization ability of NN. In this paper, Kalman Filter algorithm is applied to prune the weights of Neural Networks in order to improve the speed of learning and the generalization ability of Neural Networks. Compared with the traditional pruning algorithm, this method is different. While the traditional pruning algorithm prunes while training, this method prunes the weights after a complete training. Tests in IDS dataset show that Kalman Filter algorithm can prune with a higher rate and accuracy. Furthermore, the pruned Neural Networks can keep the detection rate of unpruned ones.
Keywords: Neural Networks Pruning Generalization ability Kalman Filter
论文图表:
引用
No.2940741645812351****
同行评议
共计0人参与
勘误表
基于卡尔曼滤波的神经网络修剪算法研究
评论
全部评论0/1000