Unsupervised Deep Learning And Optimization Integrated Algorithm And Application Research

Currently,deep learning and swarm intelligence optimization algorithms are relatively popular research topics.Among them,unsupervised deep learning can obtain the abstract and essential features of data.The swarm intelligence optimization algorithm can be applied to the clustering research of high-dimensional data.This dissertation chooses the deep belief network as the research unsupervised deep learning model and the bat algorithm as the clustering application algorithm.In order to obtain more intrinsic characteristics of the deep belief network model,bat algorithm clustering applications get better results.This paper proposes an unsupervised deep learning and swarm intelligence optimization algorithm,and finally obtains a streamlined deep belief network structure according to the integrated algorithm.Better bat algorithm clustering application results.The research content of this article mainly includes the following aspects:1.Improved bat algorithm.In order to make the optimized clustering application in the integrated algorithm proposed in this paper can obtain higher clustering accuracy,an improved bat algorithm is proposed.In the improved bat algorithm,individual bats move toward an individual with a lower fitness value,so as to avoid falling into a local extreme value.The updated bat individuals also have a global optimal solution,so some of the updated bat individuals keep bat populations optimized.The probability that there is a global optimal solution around an individual with a higher fitness value is lower,so after each iteration,1/3 population is randomly initialized instead of those bat individuals with a higher fitness value.Some data sets in Iris,Wine,Sonar,and MNIST in UCI were selected as data sets for improved bat algorithm and intelligent algorithm clustering applications such as original bat algorithm and differential evolution algorithm.Experimental results show that the improved bat algorithm proposed in this paper can obtain higher clustering accuracy and stronger robustness.2.The method of determining the DBN network structure based on bat algorithm.The determination of the deep belief network(DBN)structure boils down to the determination of the number and number of layers of neurons in the hidden layer.The DBN network training aims at minimizing the error between the reconstructed samples and the original samples,so the best number of neurons corresponds to the DBN structure with the smallest sample reconstruction error.In this paper,we propose bat individuals as the number of hidden layer neurons and DBN network reconstruction error as the fitness function of bat individuals to obtain the best number of hidden layer neurons,and obtain a preliminary DBN structure to extract input.Data feature data.Then,according to the proposed improved bat algorithm to cluster these features,the final increase in the accuracy of the features obtained by the hidden layer is much less than the accuracy of the clustering of records,and the increase of the hidden layer is stopped.It can be seen that the bat algorithm-based DBN network structure proposed in this paper constitutes an integrated algorithm with the improved bat algorithm.The Uris Iris data set is used as an integrated algorithm to optimize the application data set.The experimental results show that the proposed integration The algorithm can obtain a simplified DBN network structure and high accuracy of clustering application accuracy.3.Based on DBNSD algorithm DBN network structure determination method.The proposed DBNSD algorithm increases hidden layer neurons from a single sample to all sample angles.In the DBN network training process,if a certain sample's corresponding weights and biases vary greatly,consider adding a hidden layer neuron..From all sample perspectives,if the weights and biases can fit most of the samples,there is no need to add hidden layer neurons and continue training the current DBN network.Then the clustering test is performed on the features extracted by the current DBN.If the hidden layer is increased to obtain a higher clustering accuracy,the hidden layer is added,otherwise the current DBN network structure is not changed.It can be seen that the DBNSD algorithm proposed in this paper and the improved bat algorithm constitute an integrated algorithm,and both promote each other.Selecting partial data sets from the MNIST data set as the proposed clustering test data set in the integrated algorithm,finally obtains a simplified DBN structure and the improved bat algorithm can obtain higher clustering accuracy,indicating that the proposed DBNSD algorithm is effective Sexuality also illustrates the feasibility of the integrated algorithm proposed in this paper.