The Improvements And Applications Of Deep Learning Model Based On Restricted Boltzmann Machine

Recently,deep learning has made breakthroughs in the fields of computer vision,speech understanding,intelligent control,biomedicine,intelligent medical diagnosis,intelligent driving,etc.,with attracting more and more scholars’ attention.The deep learning models based on restricted boltzmann machine are becoming a research hotspot because of its strong model expression ability and its suitability for many kinds of tasks.Currently,the deep learning models based on the restricted Boltzmann machine mainly have two models:deep belief network and deep Boltzmann machine.Nevertheless the two deep models have their own advantages,both of them have the problems of extracting low quality features,poor feature selection and poor robustness.The core reason of the deficiency is that such submodule constructing them does not have the ability of high-order feature extraction,feature selection,and feature robustness,so that this thesis deeply studies this topic.Simultaneously along with the development of intelligent medical diagnosis,such as early diagnosis of alzheimer’s disease becoming an urgent,although the deep learning model based on the restricted Boltzmann machine has shown great potential,there is still much room about early diagnosis of alzheimer’s disease.Therefore,it is of great theoretical significance and application value to study the deep learning model based on the restricted Boltzmann machine with stronger representation capacity.Aiming at the above-mentioned problems,this thesis focuses on solving the problem that the deep learning models based on the restricted Boltzmann machine have low feature quality,poor feature selection and insufficient robustness when analyzing different types of data,and attempts to adopt an improved model to solve a medical diagnosis problem.The main content of this thesis includes the following three aspects.(1)Aiming at the problem that deep belief network cannot select useful original features and cannot extract high-order features when analyzing low-resolution images with complex backgrounds or noise,this paper proposes an enhanced high-order Boltzmann machine(EHBM)and its deep model DBN-EHBM.By using a point-by-point gating mechanism,the useful patterns are separated from the useless patterns,and then the high-order features of the useful patterns are extracted using the double hidden variable mechanism.Simultaneously,the high-order features are used to further improve the feature selection,and improve the accuracy of the point-by-point gating mechanism,and better separate the useful patterns and the useless patterns.Finally,SEHBM and Semi-SEHBM are proposed based on EHBM to solve the supervision and Semi-supervision problems.The simulation results show that EHBM,DBN-EHBM and its extension models improve the quality of feature extraction and feature extraction on diverse dataset.(2)Aiming at the problem that the convolutional deep belief network cannot extract high-order features and insufficient feature robustness when analyzing high-resolution image,the contractive spike and slab convolutional restricted Boltzmann machine(CssCRBM)and it deep model-contractive spike and slab deep convolutional belief network(CssCDBN)are proposed.The proposed models extract high-order features through bivariate mechanism.The contraction penalty term is then used to improve the robustness of the feature.The experimental results show that CssCRBM and CssCDBN extract high-order features with good robustness.(3)Aiming at the urgent need for early diagnosis of Alzheimer’s disease,combining with the advantages of deep boltzmann machine in processing eeg spectral fuzzy data sets,a discriminative contractive spike and slab deep convolutional Boltzmann(DCssCDBM)based on multi-task framework is proposed.Firstly,DCssCDBM is proposed to complete high-order feature extraction and classification in a unified context,which makes the feature extraction more discriminant.At the same time,in order to avoid the model falling into overfitting,identity authentication and identification are introduced as auxiliary tasks,so as to improve the recognition performance of the model.The experimental results show that the proposed algorithm achieves the best diagnostic rate.