Strategic Application of Feedforward Neural Networks to Large-Scale Classification
Sung-Bae Cho
Jin H. Kim
Center for Artificial Intelligence Research
and
Computer Science Department,
Korea Advanced Institute of Science and Technology,
373-1, Koosung-dong, Yoosung-ku, Taejeon 305-701, Republic of Korea
Abstract
Feedforward neural networks have been successfully applied to a variety of classification problems, but the number of classes used for experiments was too small to apply the results directly to large-scale problems. This paper presents several strategies for applying feedforward neural networks to large-scale, complex classification problems: a two-stage classification scheme, a rapid learning method, a training schedule called selective reinforcement learning, a training scheme including systematic noise, and a weight matrix reduction scheme. These strategies have been applied to the design of a printed Hangul (Korean script) recognition system. Experiments with the 990 most frequently used printed Hangul syllables confirm the usefulness of the presented strategies.
