Research On Extreme Learning Machine Theory And Algorithms

Extreme Learning Machine(ELM)algorithm is a new generalized single-hidden layer feed-forward networks(SLFNs)algorithm in recent years.Compared with the traditional BP gradient learning method,it has the advantages of simple structure,fast learning speed and good global optimization ability.ELM as emergent technology which becomes a frontier direction in machine learning field has recently attracted the attention from more and more researchers.As a variant of ELM,Kernel extreme learning machine(KELM)which applies the kernel functions to ELM algorithm greatly reduces computational complexity and the least square optimal solution can be obtained.Hence,it can provide more stable and better generalization performance.Based on the principles and existing researches of ELM and KELM,a new KELM algorithm is proposed in this paper and is applied in human action recognition in video sequence.The main work is as follows:Firstly,from the perspective of optimization,the inherent relationship among Support Vector Machine(SVM),Least Squares Support Vector Machine(LS-SVM)and ELM is analyzed.It is concluded that LS-SVM,a variant of basic SVM,is actually a simplified implementation structure of the KELM.The KELM combines SVM and LS-SVM and serves a unified solution of generalized SLFNs,which provides a guidance for the later new algorithms.Secondly,from the perspective of parameter optimization time,the triangular Hermit kernel extreme learning machine(Tri_H-KELM)methodology is presented based on Hemite polynomial.It introduces the triangular Hermite function which has been proved as a valid kernel function into extreme learning machine as kernel function.The most significant advantages of proposed kernel are that it has only one parameter chosen from a small set of natural numbers,thus the parameter optimization is facilitated greatly,and more structure information of sample data is retained.Experiments were performed on bi-spiral benchmarkdata set as well as a number of binary classification,multi-classification and regression datasets from the UCI benchmark repository.Similar or better robustness and generalization performance of the proposed method in comparison to other extreme learning machine with different kernels and SVM methods demonstrates its effectiveness and usefulness.Finally,Tri_H-KELM is used in human action recognition in video sequence to construct the suitable human action classifier.The fusion GIST feature representation of depth images and RGB images is proposed.The Tri_H-KELM method is verified for classification ability on the challenging MSR Action3 D and DHA datasets.The experiment results show that,compared with other kernel ELM and SVM algorithms,Tri_H-KELM can have a better recognition accuracy.What's more,it has obvious time advantage for the video image dates with high dimension and large data volume.