The Improvement Of Online Sequence Extreme Learning Machine And Its Application

In recent years,with the explosive growth of digital information,how to collect,process,and analyze effective information efficiently and quickly has become a hot topic in the society.Artificial neural network has become the research object of many scholars for its powerful data processing capabilities.At the same time,a variety of algorithms have emerged.Online sequential extreme learning machine is a fast and accurate online sequential learning algorithm,which belongs to a single hidden layer feedforward neural network.Its input weights and biases are randomly generated and do not require iteration.Therefore,compared with the neural network based on the gradient descent method,the training time is greatly shortened,and it can be used for batch learning.It can learn data one by one,or by blocks with fixed or variable size.Its input weights and biases are randomly generated,do not need iteration,and can only learn new data that has not been trained,so it is greatly shortened compared with gradient descent-based algorithms,such as back propagation algorithm,support vector machine,etc.The training time has been improved,and the accuracy has also been improved.However,online sequential extreme learning machine still has deficiencies in some aspects.In order to solve the problems such as: a single online sequential extreme learning machine does not perform well on data classification problems,the hardware circuit is difficult to implement,the randomly generated input weights and bias distribution are uneven,and the characteristics of the experimental object are insufficiently obtained.This paper goes deep into studying of the online sequential extreme learning machine,firstly,an ensembled online sequential extreme learning machine based on cross-validation is proposed to make up for the deficiencies of a single network in handling classification problems.Then,through calculation,a new type of activation function based on memristor is proposed and applied to the online sequential extreme learning machine to provide the possibility of hardware implementation,again,the original direct generation of learning parameters is changed to piecewise random generation,and the randomness of the learning parameters is enhanced,finally,the approximation ability of the online sequential overrun learning machine is used to construct a sparse automatic encoder to solve the problem of incomplete acquisition of complex features due to the simple structure.The main work and innovations of this article are as follows:(1)Due to the simple structure and poor stability of the online sequential extreme learning machine,two ideas of cross-validation and ensemble learning are introduced into the online sequential extreme learning machine,and the cross-validation method is used to select the learning parameters and ensemble learning is used to increase system stability.And based on the K-fold cross-validation,a hierarchical cross-validation is proposed,which can better solve the problem of imbalanced data classification,reduce the occurrence of overfitting and underfitting,and obtain better results.(2)A new activation function is realized by using the relationship between the memristive value and the charge of the memristor,that is,the memristive activation function.At the same time,change the random input mode of input weights and biases,segment these learning parameters first and then randomly,and randomly combine several small matrices and then combine the matrices as the final input weights and offset matrix.These two methods make it possible to realize the hardware circuit of the algorithm,and improve the stability of the system by increasing the randomness of the learning parameters.(3)Because of its shallow architecture,even if the system has a large number of hidden nodes,feature learning using an online sequential extreme learning machine may not be effective for natural signals,so a new hierarchical structure based on online sequential extreme learning machine is proposed multilayer perceptron learning framework.The proposed hidden layer framework trains in a forward-looking way.Once the previous layer is established,the weights of the current layer can be fixed without fine-tuning.Therefore,it has better feature learning ability.