Abstract:

ELM（Extreme Learning Machines） always works inefficiently on large-scale and high-dimension 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 Layer Feed-forward Neural Network) in parallel.Theoretical analysis proves that the new algorithm converges to the best least-square solution monotonously with finite steps.Preliminary experimental results show that the new algorithm  has good generalization ability, excellent speedup ratio and parallel efficiency.

Key words: single hidden layer feed-forward neural network, extreme learning machines, parallel extreme learning machines, binary cascade architecture