Research On Residual Neural Network Based On Extreme Learning Machine

Neural networks are an important branch of artificial intelligence.They are good at processing a variety of data.There are various algorithms and models for neural networks.The extreme learning machine is a typical shallow neural network that is good at processing small classification and regression tasks.Convolutional neural network is a typical deep neural network,which is good at extracting the features of images,and can complete the tasks of image classification and recognition.At present,these neural networks have achieved great success in many fields.But most neural networks are imperfect and have some defects when used.Therefore,this paper proposes two improved neural network algorithms and combines the two improved networks.The combined network is tested on the image classification data set and shows good results.The specific research results and practical significance of this article are described as follows:(1)Residual neural network is currently a widely used convolutional neural network model.It adds a lot of skills to the basic convolutional neural network to have better generalization performance.However,there are still plenty of room for improvement in this network.In order to further improve the transfer of information and gradients,this paper proposes a Fully Convolutional Multi-parallel Residual Neural Network(FCM-Resnet).And it is applied to three classical image classification data sets,compared with the network of InceptionV3 Densenet121 and Resnet,FCM-Resnet has higher test accuracy.(2)Although Fully Convolutional Multi-parallel Residual Neural Network further improves the network performance,there is still room for improvement.This paper proposes a new algorithm to replace the classification layer of the network.This new algorithm is aimed at improving the problem that extreme learning machines is greatly affected by randomly initialized input weights and hidden layer thresholds.It is called Bidirectional Learning Machine(BLM).Compared with algorithms such as ELM,IELM and FLN,BLM has obvious advantages in classification and regression datasets.Finally,in order to further verify the effectiveness of the network,the BLM is applied to the actualdata collected in the field,and it shows better generalization performance compared with ELM and LSFLN.(3)On the basis of FCM-Resnet and BLM,this paper presents a combined algorithm—— FCM-Resnet-BLM.In view of the superiority of BLM,the combined algorithm uses BLM to replace the classification layer of FCM-Resnet,so as to improve the network performance.Finally,the classification effect of the network is tested on four image classification datasets.The combined network has a higher classification accuracy than networks such as BLM.