Abstract

A novel neural network classifier, called associative memory classifier, is developed in this thesis modeled upon associative memory network. It is studied for its application potential in the recognition problem of large number categories, such as that of Chinese characters. The major findings of this work are in three aspects. First of all, it is found that a feed forward associative memory network can become a suitable pattern classifier by an appropriate selection of its output vectors, called inner codes. The classification ability of an associative memory classifier is determined ultimately by the distinctiveness of its input patterns, but a set of properly selected inner codes may help the classifier to approach the limit of its capability. In the next place, seeking for better inner coding schemes should be in connection with each specific case on the basis of input patterns' characteristics. Attempts to resolve this problem by enumeration will lead to a non-polynomial complexity which is computationally infeasible. On the other hand, real-life data are usually mathematically undescribable. Hence, a practical way to find such schemes is to work out an optimization strategy first under some ideal conditions, then apply it to the real data with remedial measures. Thirdly, when an associative memory network is used as a pattern classifier, if the feature patterns are transformed into a form more suitable to it, the performance of the entire system can be improved significantly. Also, due to the parallel computation mode of neural networks, data reduction is not a problem any longer. Therefore in addition to stable feature detection, the other objective in feature extraction here has been changed to the improvement of feature patterns suitability to the currently used neural classifier. All these findings are verified by computer simulation with sets of common and similar multi-font Chinese characters. Our experiments are conducted presently using more than two hundred and fifty categories for a single-level classifier. To test the robustness of the system and make it meet the needs of practical use, similarities are introduced into the testing data by characters which look alike and printed in different fonts.