摘要
为克服神经网络受噪声和冗余特征的影响而出现过拟合,提出一种自适应级联神经网络(ACNN)及学习算法。ACNN从少量特征开始学习,在学习过程中根据特征对分类的有效性增加新特征,用映射递归算法调节权值,逐步确定网络结构,使其含有最少数目的输入和隐层神经元。此方法应用于区分两种思维状态下的脑电信号(EEG),经训练的网络对测试段的分类正确率为83.1%,与文献[1]中采用BP网络的结果相比,显示了ACNN较好的分类能力。
An Adaptive Cascade Neural Network ( ACNN ) and its learning algorithm are described to avoid over-fitting caused by noise and redundant features. The ACNN starts to learn with few features and then add new ones according to their classifying validity . The neuron weights are fitted by using a recurrent algorithm based on a projection method. The ACNN topology is decided with a minimal number of input and hidden neurons. It is applied to classify electroencephalogram (EEG) between two mental tasks. The trained ACNN has correctly classified 83. 1% of the test segments. It shows a better result compared with a standard BP network.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2005年第6期713-716,共4页
Pattern Recognition and Artificial Intelligence
基金
国家自然科学基金(No.60375017)
关键词
级联结构
脑电
自适应
过拟合
特征提取
Cascade Architecture, Electroencephalogram, Adaptive, Over-Fitting, Feature Selection