In this thesis, we propose several strategies for applying neural network to real world problems. Proposed are a rapid learning method, a training scheme including systematic noise, a training schedule called selective reinforcement learning, and a weight matrix reduction scheme.
The rapid learning method accelerates the learning speed by applying the Aitken's $\Lambda^2$ process, which is developed for solving nonlinear optimization problems. The noise included training scheme adds noises systematically to the given training patterns. This results in the same effect as expanding the number of training patterns and, therefore, improves its generalization capability. Since a neural network tends to waste most of its time to learn a few hard patterns, the selective reinforcement learning focuses its attention to the hard patterns. A large number of link weights do not contribute to its decision making process because of small magnitude. The weight matrix reduction scheme cuts off insignificant links for improving recognition speed.
The above strategies have been applied to the design of a printed Hangul recognition system. The system is composed of a type classification network and six recognition networks. The former classifies input character images into one of the six types by their overall structure, and the latter further classify them into character code.
Experiments are conducted with most frequently used 990 printed Hangul characters. By the noise included training, the recognition rate amounts up to 98.28＼%, which is superior to other neural network methods. With the selective reinforcement learning, the network learns two times fast. The network keeps on producing acceptable results until pruning more than half of the total links.
