General classification model using cooperative neural networks

Other Title(s)

نموذج تصنيف عام باستخدام وحدات الشبكات العصبية المتعاونة

Dissertant

Mashjil, Hana Muhammad

Thesis advisor

Dhannun, Ban Nadim

University

University of Baghdad

Faculty

College of Science

Department

Department of Computer Science

University Country

Iraq

Degree

Master

Degree Date

2014

English Abstract

Modular Neural Network (MNN) technology took a significant role in solving complex problems in real world, and because many of these problems whether in business, science, industry, or in medicine treated as classification problems , then this work suggests an efficient general pattern classification model using Cooperative Neural Network (CNN), where a Modular Neural Network used to improve the efficiency of the unitary neural network to solve difficult classification problems by partitioning the classification task into several sub-tasks which are individually simpler to solve and can be combined to solve the full problem, Each sub task is implemented by a simple, fast and efficient classifier module; and the settings of modules structures were chosen experimentally , also to choose the suitable training algorithm several Back Propagation (BP) algorithms were tested and it founded that BP in pattern mode with adaptive learning rate gave the best result in the proposed method .

Then the model ability to choose the best useful sub-set of discriminating features given to the system through the input layer was improved, the selection of best features is a complex process because it requires extensive experience and a deep understanding of the problem domain in case of very complex data set or when the number of distinctive features are few, so neural networks were able to detect the features those have no role in CNN modules, then making decisions by decreasing their weights or even applying pruning procedure to delete them, such that the performance of the classification process is improved in terms of accuracy and time.

In this work a simulator has been developed to generate data sets that have controllable statistical behaviors with various degree of complexity to estimate the efficiency of the proposed model, where different cases of data were simulated for classification with different setting of network factors.

The generated dataset is partitioned in two sets.

One set with 40% of the input/output pairs, are unseen samples for the network used in test only, and the remaining 60% in another set which is the training set, then both the training set and testing set are used for testing.

The tests results indicated a success rate nearly 100% when simulated complex patterns are fed to the improved model.

Also, the test results indicate that the raise of complexity degree does not cause dramatic effect on the system performance; except the model execution time becomes more than few seconds which considered promising

Main Subjects

Information Technology and Computer Science

Topics

• Artificial intelligence
• Simulation methods
• Data processing
• Logic machines
• Neural networks(Computer science)
• Bionics

No. of Pages

115

Table of Contents

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-605949