Dissertation > Excellent graduate degree dissertation topics show

Research of Classification Ability and Training Algorithm of Feedforward Neural Network

Author: ChenSenLin
Tutor: ZhangJunYing
School: Xi'an University of Electronic Science and Technology
Course: Applied Computer Technology
Keywords: Multilayer feedforward artificial neural network Ability to classify Training algorithm Robust sensor Continued penetration learning Support Vector
CLC: TP183
Type: Master's thesis
Year: 2003
Downloads: 352
Quote: 7
Read: Download Dissertation

Abstract

The subject to get the National Natural Science Foundation of China (No.: 60071026), the pre-research of national defense science and technology cross-industry fund projects (ID: OOJ1.4.4.DZ0106). Funded. Network of artificial neural networks as an emerging discipline, from the produce has already made considerable development, and is widely used pattern recognition, optimization problems, graphics, image processing, signal processing, automatic control, forecasting, and so many areas, provides a new and effective means to solve the problem related to these fields. As an emerging discipline, the theoretical basis of the artificial neural network itself is not perfect, the classification capabilities of neural networks, the network architecture of choice, the choice of network learning algorithm, the network's physical implementation, etc., so that the neural network in the real The application is subject to many limitations, but also limit the further development of the neural network disciplines. To artificial neural networks as the basis of a variety of other neural network, has been the people to study, the most widely used network model, the study of the theory and learning algorithm helps further study of other types of neural networks, pattern recognition and classification forward the most important applications of neural networks, so this will be before the ability and training to the artificial neural network classification algorithm as the main object of study. This paper first reviews the history, development and present situation of artificial neural networks, artificial neural network pattern recognition applications. Then from the point of view of the experimental simulation, by constructing the common types of classification surface discussed prior to the classification of the neural network and the network structure, these conclusions will help further studied theoretically before the mechanism of neural network classification, the classification ability , and also contribute to the selection of the network structure of the actual application. And then put forward through a linear transformation of the mode will change the mode of classification difficulty, thus speeding up the training of the neural network; subsequent forward neural network of multi-class classification proposed by neurons appropriate reversal operator to initialize method, so that the neural network training and try to avoid local minima. After further discussion of the the robust classification problems, the training sample set conditions based on the complexity of the base set of robust sensor sound continued penetration of the perceptron learning algorithm Finally, combined with support vector machine based on the soundness of the base set Perceptron linearly separable problem based on the robust support vector perceptron the geometric training algorithm, experimental simulation show that the sensor has better robustness, faster training speed. The key technologies used: experimental simulation; ② continued penetration learning algorithm; ③ statistical learning theory; ④ linear programming; ⑤ BP algorithm; ⑥ linear space theory. In this paper, the experimental simulation method to construct the common types of classification surface, to study the lack of theoretical guidance to the classification of the network capacity, not enough structure classification of surface types, we can consider to construct more types of classification surface, combined with theoretical derivation is given to determine the network classification ability to select the network structure, theoretical guidance. Finally, based mode linearly separable robust support vector sensor is only applicable to the case, how will it extended to nonlinear separable problem, and further improve the efficiency of the algorithm is the future research directions.

Related Dissertations

CLC: > Industrial Technology > Automation technology,computer technology > Automated basic theory > Artificial intelligence theory > Artificial Neural Networks and Computing
© 2012 www.DissertationTopic.Net Mobile