The hierarchical structure of neural network for the complicated data set classification problems /
The area of the neural networks is getting more and more attention in the hope that it will narrow the gap between a human being's intelligence and intelligence built into a computer. Especially in the classification of patterns, the area of neural networks has already had many applications and...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
1996.
|
| Subjects: | |
| Online Access: | http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=739667361&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | The area of the neural networks is getting more and more attention in the hope that it will narrow the gap between a human being's intelligence and intelligence built into a computer. Especially in the classification of patterns, the area of neural networks has already had many applications and good results. However for the complicated data set classification, neural networks still remain with many weak points compared to a human being's intelligence. The main reason for this big difference between a human and a neural network can be said to be the analysis ability for patterns in a hierarchical manner. This dissertation proposes two different styles of Hierarchical Multilayer Perceptron (HMLP) neural network as the improved classifiers. The complicated data set classification problems can be represented as two different types, 1) the large sized classification and 2) the highly overlapped classification problems. In the case of large sized classification problem, the conventional single stage of a MLP classifier is insufficient and causing the overtraining problem to build up complicated decision boundaries for all classes at one stage. HMLP is built up by hierarchically stacking the small MLP subnetworks that work as local classifiers in the feature space. Each subnetwork is trained so as not to cause overtraining. The generalization and overtraining problems of neural networks were reduced by this method. Also in the case of a highly overlapped data set, the Bayesian classification rule is implemented with several subMLP neural networks. This scheme and its relation to overtraining and generalization problems are discussed. The test results for both classification problems showed that the HMLP had better classification ability with the smaller network size and computation burden. These two different styles of HMLP classifiers were implemented for the automobile license plate number recognition system and the cork quality classification system. |
|---|---|
| Item Description: | Vita. "Major Subject: Electrical Engineering". |
| Physical Description: | xii, 113 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilms Inc. |
| Bibliography: | Includes bibliographical references: pages 108-112. |