Extreme Learning Machine With Sparse Optimization And Its Application

Extreme learning machine?ELM?,which is based on the single-hidden layer feedforward network?SLFN?,has been popular in machine learning fields recently.ELM has some advantages:fast speed and simple implementation.However,there are still some problems in the conventional ELM:the randomness of the hidden nodes,the redundancy of the network architecture and the extension version for the clustering tasks.For these three problems,the work of this paper is summarized as follows:?1?In order to solve the randomness problem of ELM and obtain powerful hidden nodes,this paper proposed a method called sparse and heuristic extreme learning machine?SHELM?.This method can fuse the SLFNs in the aspect of the hidden layer to obtain powerful and diverse feature space.Then,the sparse representation method was proposed to select the best hidden nodes.In this paper,the fusion of the support vector machine and ELM was shown in details.Besides,this paper proposed a method based on the separable surrogate function?SSF?to solve the sparse representation model efficiently.Lastly,the experiment results shown that SHELM not only had better generalized performance but also had fewer hidden nodes.?2?In order to solve the redundancy of ELM,this paper proposed a novel sparse extreme leaning machine based on the L₀ norm and L₁ norm regularization from the aspect of the Bayes'theorem.Then,this paper proposed a method based on the SSF technique to solve this optimization problem and proofed the convergence.Lastly,the results of the experiments verified the sparseness of the proposed method.?3?Though ELM can be successfully apllied in supervised tasks,how to extend it into unsupervised tasks successfully is a meaningful problem.This work proposed a novel clustering method based on the extreme learning machine,which would like to find labels to obtain the optimal ELM classifier.This paper firstly proposed a clustering model based on ELM,and then adopted the alternative optimization method to solve this model.Lastly,the results of the experiments shown that the performance of the proposed method was competitive with that of other state-of-the-art methods.