Research On Removing Redundancy And Feature Mapping Method In Deep Learning

In recent years,the theory and methods of deep learning based on neural networks have attracted wide attention.Compared with the traditional machine learning method,the deep learning has more powerful feature learning ability.Among them,Restricted Boltzmann Machine(RBM),which can extract abstract features from the data layer by layer,is one of the most important models in deep learning.In recent years,RBM has been widely studied and applied in the field of visual,voice and text.In terms of image feature mapping,Convolutional Neural Network(CNN)has attracted wide attention due to its powerful ability of image feature learning and classification.In this paper,the theory and application of some algorithm models in deep learning are studied.In this paper,the redundant hidden units produced in the training process of RBM are studied.A method for evaluating the redundant hidden units in RBM is proposed.And then put forward a Removing Redundancy Restricted Boltzmann Machine.In this paper,I also study the clustering convolution model and convolutional neural network model.I propose a multi-scale feature mapping method based on clustering convolution and Restricted Boltzmann Machine and apply it to medical image classification.The main contents of this paper are as follows:(1)Put forward a Removing Redundancy Restricted Boltzmann Machine.In this paper,the training process of RBM and the data distribution of the hidden unit are analyzed.A measurement method of redundant hidden units in RBM is proposed,so as to add eliminates redundant mechanism in the training process of RBM,and remove redundant hidden units in RBM and Optimizing the structure of RBM by eliminating redundancy.In this way,the number of hidden units of RBM is greatly reduced,the training efficiency is improved,the training time is shortened,and the convergence speed of the model is improved(2)A multi-scale feature mapping method based on clustering convolution and Restricted Boltzmann Machine is proposed.In this paper,we combine restricted Boltzmann machine and clustering convolution method,and use multi-scale feature mapping method to improve the method of clustering learning convolution feature.The new method does not have to preprocess the data,and improves the generalization ability of the convolution kernel.On the basis of high efficiency and stability,the quality of the extracted features is further improved.Based on experiments on CIFAR-10 dataset,the accuracy of image classification is achieved by mapping features.The accuracy is 83.53%.Compared with the original convolution method and general convolutional neural network,the algorithm efficiency is improved.(3)The improved clustering convolution method is applied to medical image classification.In this paper,the improved clustering convolution method is applied to the diagnosis and classification of lung X-ray images.In this paper,a multi-layer clustering convolution mapping model is constructed for the X-ray images of the lung.The advantages of fast training speed of clustering convolution network and the mapping of layer by layer abstract feature in deep model are given full play.Compared with the CNN model,the new method has a stronger ability of feature mapping compared with the original clustering convolution method.The new method has a certain degree of improvement in training speed and feature mapping ability.