Research On Deep Belief Network And Its Application

Deep Learning is a feature learning method,which transforms the original data into a higher-level and more abstract expression through some simple nonlinear models.Deep Learning method has been applied in many fields,such as image recognition,speech recognition,object detection and so on.Deep belief networks(DBN)is a kind of Deep Learning,which can find complex structures and probability distributions in large data through multiple mapping.Feature engineering is an extremely important process in the process of data mining,directly affecting the results of the model.The traditional feature engineering find the suitable characteristic rely on the business experience,several attempts.In the classical probability statistics,there are not only linear feature extraction methods such as principal component analysis and factor analysis,but also non-linear feature extraction methods such as support vector machine,Logistic regression.However,traditional methods can only extract the most intuitive and most obvious features of the data.The deep belief network can not only extract nonlinear and linear features from the data automatically,but also can extract more in-depth features,through supervised learning to achieve automatic variable screening.Depth belief network can be used for feature selection of various algorithms preprocessing process,and then it use stochastic gradient descent algorithm to realize on-line calculation and large data learning,and solving the problem that all the data need full loaded into memory.In this paper,the topology structure and training algorithm of DBN model in Deep Learning are improved,and the effectiveness of the algorithm is validated by the application of DBN model in exchange rate forecasting.The main contents of this paper are as follows: Firstly,it reviews the development of Deep Learning,introduces the basic concepts of shallow learning and Deep Learning,and summarizes the research progress at home and abroad.Secondly,we study the topological structure and the main theory of the depth belief network,proving that the classical linear correlation analysis is also a kind of shallow learning.Then,the learning process of depth belief network is studied,and a new learning algorithm based on probability distribution is proposed.Finally,we use multivariate analysis of variance to determine the topological structure of DBN,predict the exchange rate of RMB and Indian lu against US dollar respectively,and verify that the improved DBN model has been greatly improved in accuracy and robustness.