Victor Hannak


Artificial Neural Networks (ANNs) are commonly used in both academia and industry as a solution to challenges in the pattern recognition domain. However, there are two problems that must be addressed before an ANN can be successfully applied to a given recognition task: ANN customization and data pre-processing. First, ANNs require customization for each specific application. Although the underlying mathematics of ANNs is well understood, customization based on theoretical analysis is impractical because of the complex interrelationship between ANN behavior and the problem domain. On the other hand, an empirical approach to the task of customization can be successful with the selection of an appropriate test domain. However, this latter approach is computationally intensive, especially due to the many variables that can be adjusted within the system. Additionally, it is subject to the limitations of the selected search algorithm used to find the optimal solution. Second, data pre-processing (feature extraction) is almost always necessary in order to organize and minimize the input data, thereby optimizing ANN performance. Not only is it difficult to know what and how many features to extract from the data, but it is also challenging to find the right balance between the computational requirements for the preprocessing algorithm versus the ANN itself. Furthermore, the task of developing an appropriate pre-processing algorithm usually requires expert knowledge of the problem domain, which may not always be available. This paper contends that the concurrent evolution of ANNs and data pre-processors allows the design of highly accurate recognition networks without the need for expert knowledge in the application domain. To this end, a novel method for evolving customized ANNs with correlated feature extractors was designed and tested. This method involves the use of concurrent evolutionary processes (CEPs) as a mechanism to search the space of recognition networks. In a series of controlled experiments the CEP was applied to the digit recognition domain to show that the efficacy of this method is in-line with results seen in other digit recognition research, but without the need for expert knowledge in image processing techniques for digit recognition.

Library of Congress Subject Headings

Neural networks (Computer science); Evolutionary computation; Pattern recognition systems

Publication Date


Document Type


Student Type


Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)


Peter Anderson

Advisor/Committee Member

Andreas Savakis

Advisor/Committee Member

Shanchieh Jay Yang


Physical copy available from RIT's Wallace Library at QA76.87 .H347 2004


RIT – Main Campus