Research Of DBN Classification Network Based On T-RBM Algorithm

Using machine learning algorithms to process data classification task has been studiedfor many years, a number of classic and effectively algorithms has been emerged during thisperiod, such as K-means, SVM, and gradient descent algorithm. These algorithms are widelyused in data classification, data mining, image recognition, medical imaging, and radarcommercial, industrial, medical, military and other fields. Due to the extremely complexenvironment in applications, data classification tasks still are one of the most challengingtasks in the field of computer science and technology.Since the complexity of data classification task, deep learning network (DBN) basedclassification algorithms began to be used widely for large data mining, image processing,pattern recognition tasks, compared to traditional classical classification algorithm, thesealgorithms can found some internal data structures or features of the data automatically, andaccomplish the classification task due to these structures and feature. This paper focuses onvarious learning algorithm and did further research of them, especially for DBN (deep beliefnetwork) which is the most important learning algorithm currently, a new learning algorithmis proposed in this paper which is T-RBM based DBN classification algorithm, the maincontribution of this paper are as follows:(1) The current main machine learning algorithms have been studied and analyzed,based on the characteristics of these learning algorithms, this paper classified them into twokinds, they are supervised and unsupervised learning algorithms, this paper analyzes boththeir advantages and their disadvantages, according to the analysis of the advantages anddisadvantages,this paper choose unsupervised classification algorithm in-depth study.(2) On the basis of unsupervised learning, depth learning algorithm which is a kind ofunsupervised learning algorithms has been further analyzed, including the principle analysisform the aspect of mathematics, the disadvantages of it, thus propose the main purpose of theresearch of this paper.(3) Focusing on the disadvantage that it is difficult to adjust the parameters for thenetwork DBN, this paper proposed a new classification algorithm—T-RBM based DBNalgorithm using calculate feature value entropy, commend the number of the hide units, andthe data abstraction of each layer to improve the performances of data feature. (4)Simulations classical DBN algorithm and simulation-based T-RBM algorithms havebeen done in Matlab2013b, simulation results show that T-RBM based DBN classificationalgorithm has an average2%lower classification error rate and an around10second lesstraining time compare with classical DBN classification algorithm under the same samplesand certain assumptions.Mathematical derivation and experimental analysis shows that, T-RBM based DBNclassification algorithm can effectively reduce unnecessary computation time by reducingnoise and improving the accuracy of the classification of the network. Simulation resultsproved that T-RBM based DBN classification algorithm is optimizer that classical DBNclassification algorithm in classification task in various applications.