摘要
为解决因庞大的矩阵存储和计算,ELM(Extreme Learning Machines)难以应用到大规模、高维数据集的问题,提出一种基于"分而治之"策略的并行极速学习机算法。该算法利用二叉级联结构,将大规模数据集分派到多个计算节点上,并行地更新单隐层前馈网络的输出权值,且能有限步地单调收敛到最小二乘解。实验结果表明,该算法不仅泛化性能优异,并且具有非常高的加速比和并行效率。
ELM(Extreme Learning Machines) always works inefficiently on large-scale and high-dimen- sion datasets for huge memory and computation costs. We prepose a novel parallel algorithm for training ELM quickly based on "divide and conquer" strategy. It dispatches large-scale datasets to a cluster of computing nodes by utilizing special binary cascade architecture,and then updates weights of SLFN(Single Hidden I.ayer Feed-forward Neural Network) in parallel. Theoretical analysis proves that the new al gorithm converges to the best least-square solution monotonously with finite steps. Preliminary experi mental results show that the new algorithm has good generalization ability, excellent speedup ratio and parallel efficiency.
出处
《吉林大学学报(信息科学版)》
CAS
2012年第4期418-425,共8页
Journal of Jilin University(Information Science Edition)
基金
中央高校基本科研业务费专项基金资助项目(JBK120126)
教育部人文社会科学研究基金资助项目(10YJCZH153)
关键词
单隐层前馈神经网络
极速学习机
并行极速学习机
二叉级联结构
single hidden layer feed-forward neural network extreme learning machines parallel extremelearning machines binary cascade architecture