Strategic Application of Feedforward Neural Networks to Large-Scale Classification

Feedforward neural networks have been successfully applied to a variety of classification problems, but th e number of classes used for experiments was too small to app ly the results directly to large-scale problems. This pap er presents several st rategies for applying feedforward neur al networks to large-scale, complex classificat ion problems: a two-stage classification scheme, a rapid learn ing met hod, a t raining schedul e called select ive reinforcement learning, a t raining scheme including syst ematic noise, and a weight matrix redu ction scheme. These st rategies have been applied to the design of a printed Hangul (Korean script) recognition system. Exp erim ent s with the 990 most frequently used pri nt ed Hangul syllables confirm the usefulness of th e present ed str at egies.