Methods of Combining Multiple Classiiers Based on Diierent Representations for Pen-based Handwritten Digit Recognition

Pen-based handwriting recognition has enormous practical utility. It is diierent from optical recognition in that the input is a temporal signal of pen movements as opposed to a static spatial pattern. We examine various ways of combining multiple learners which are trained with diierent representations of the same input signal: dynamic (pen movements) and static ((nal 2D image). We notice that the classiiers based on diierent representations fail for diierent patterns and investigate ways to combine the two representations. We benchmark voting, stacking, mixture of experts and cascading. In voting and stacking, the two are always used together. In the mixture of experts, the gating network chooses one of the two. In cascading, the static is used only when the dynamic is not \certain". On a handwritten digit database signiicant increase in accuracy has been obtained.