一种前置RBF核的混合ELM网络研究

Research on a hybrid ELM network with pre-RBF kernels

极限学习机(extreme learning machine,ELM)是一种广泛使用的浅层神经网络模型。针对ELM网络模型的优化问题,提出一种前置径向基函数(radial basis function,RBF)核的混合ELM网络模型。通过在原有的ELM网络前增加一个新的RBF映射层用于对原始样本空间进行核映射可以提取原始样本空间的局部特征,以改善原始样本空间的可分性;随后所级联的ELM网络用于RBF核映射空间样本的学习。基准数据集上的实验表明,所提方法将RBF核局部响应能力强的优点以及ELM网络泛化能力强的优势结合在一起,有效改善了单一ELM网络的输出性能。

Extreme learning machine(ELM) is a widely used shallow neural network model. Aiming at the optimization problem of ELM network model, a hybrid ELM network model with pre-radial basis function(RBF) kernels is proposed. By adding a new RBF mapping layer in front of the original ELM network for kernel mapping of the original sample space, the local features of the original sample space can be extracted to improve the divisibility of original sample space. Then, the cascaded ELM network is used to learn the RBF kernel mapping space samples. The experiments on the benchmark data sets show that the method proposed combines the advantages of strong local response capabilities of the RBF kernel with the advantages of strong generalization ability of the ELM network, and effectively improves the output performance of a single ELM network.