Computational intelligence (CI) is gaining more attention as its applications in various areas grow. Soft computing is one of the popular CI facets because of its ability to handle imprecision and uncertainty. Artificial Neural Network (ANN), neuro-juzzy system,juzzy expert system, and statistical method are the prominent tools within this discipline. However, these systems infer mainly from the class of interest (positive class). The information about the differences among classes are not fully utilized. Some systems, on the other hand, do not consider the class information at all. Although these systems perform well, their performance could be further enhanced if the contribution from' negative class is taken into account. Moreover, constructing knowledge based on single class alone may cause the system to under-perform when the data is imbalanced. Viewing from the other end, most of these systems focus on boosting the accuracy, but disregard the psychological needs of user. They lack the reasoning, inference, and validation processes that user can identify with. CI system based on ANN or statistical method provides no means of understanding the system, while some expert system requires manual construction of knowledge. Furthermore, most of them are independent of biological or psychological principle, which hinder the user acceptance and trust towards the system. The debut of high-dimensional and ultra-huge databases exacerbates the situation. Thus, it is crucial to take into account the interpretability, the tractability, and the high-dimensionality, on top of the performance. To this end, a novel class of learning paradigm-Complementary Learning is proposed. Complementary learning functionally models after the pattern recognition of human being. Since most of the CI task involves pattern recognition, and the fact that human being is effective in learning patterns, mimicking the human pattern recognition could bring forth fruitful outcome. Complementary learning not only has more common ground with user, but also good performance in pattern recognition. Complementary learning embraces learning from positive and negative classes. It also segregates the positive and negative knowledge, and then exploits the lateral inhibiting relationship between the two. These, minimize confusion in inference, in which lesser susceptibility to imbalanced dataset may be offered. In order to improve the tractability and b_~A TTENTION: The Singapore Copyright Act applies to the use of this document. Nanyang Technological University Library Acknowledgement I would like to thank my supervisor Dr. Ng Geok See, for his valuable guidance, encouragement and help throughout this research. Without his patience and understanding, this research would not able to reach this stage. I would like to express my gratitude to Dr. Quek Hiok Chai as well for his motivation, encouragement, guidance, and in-time support. Discussion with him is always very inspiring. I also wish to bring my sincere thanks to Dr. Tung Whye Loon for his guidance, his cl...
