Research On The Model Construction And Learning Algorithms Of Deep Learning

As a frontier in Machine Learning, deep learning is of an unprecedented importance on the development of artificial intelligence technology. Based on the ability and algorithm of cloud computing parallel processing the big data, deep learning builds a learning network whose mode of thinking is much closer to the human being, which arms the computer with the ability to handle the abstract conception so as to make the computer more intellective. As a kind of method that based on unsupervised feature learning and feature hierarchical structure learning, deep network aims at interpreting data by simulating the thinking mechanism of human brain. It builds a deep neural network model simulating the analysis and study process of human brain through assembling low- level feature into high- level feature which is more abstract to present the property of object so as to detect the distributed feature expression of data.Deep network building and the research of learning algorithm are particularly important because the two core contents of deep lear ning network are reasonable construction of network and effective learning of the network. For this meaningful task, the paper carried on a thorough analysis and research, then applied the deep network on the polarimetric SAR image classification and leaded to a benign effect. The main research work and research results are listed as follow:(1)This paper establish chaotic simulated annealing network and chaotic simulated annealing deep wavelet network, and proposed a learning algorithm based on chaotic simulated annealing algorithm. In the construction of the networks, the full application of auto encoding principle and the theory of wavelet analysis are used to build the two networks. In the study of network learning algorithm, we proposed a deep network learning algorithm based on simulated annealing against the problem that traditional network algorithm usually plunges into local optima. The algorithm introduced the idea of simulated annealing into the weights optimization part of deep network, solving t he local optima in weights optimization. The paper introducing chaos model based on the above algorithm to solving the slow convergence speed of simulated annealing, realizing the quick finding of global optimal solution. The algorithm shortened the training time of network with high classification accuracy, which made a breakthrough progress.(2)Based on the wavelet analysis, the paper built a deep network on the wavelet domain analysis. We introduced the concept of knowledge representation and applicatio n on the basis of existing models to build the deep network based on knowledge. The network automatically extracted the advanced features of the original data, and introduced the specific Wishart concept of Pol SAR data into itself. Through the principle of former type priority and significant priority, the proposed method improved the classification precision of the network making full use of the prior knowledge from knowledge level. The algorithm solved the low precision of traditional method for polarimetric SAR image classification, realizing a reasonable build of deep network.(3) The paper introduced the thoughts of global searching of evolutionary algorithms PSO into The deep network weights optimization process. Based on PSO, adaptive PSO Orthogonal PSO Quantum PSO and simulated annealing PSO, the paper built a deep network evolutionary learning algorithm, then applied the method on the classification of Pol SAR images to analyze and evaluate the result. The proposed method is flexible in category selection, strong in universality and splendid in transferability. It enriches the study method of deep network with il ustriousness robustness.