ABSTRACT
This chapter shows how a highly structured neural network can be combined with a dynamic programming lattice to construct a system that learns to recognize multicharacter images. The normalization that is appropriate for an single-character recognizer throws away information which is needed for proper segmentation. Recognition arcs (rec-arcs) are daytime arcs, from a morning to an evening. Glue-arcs are nighttime arcs, from an evening to the next morning. If all images were naturally divided into segments containing one character apiece, the multicharacter recognizer would be straightforward. The normalized score for a path gives the joint probability of a given interpretation and segmentation. Instead, the problem is formalized in terms of a probability measure; the learning algorithm must then be arranged to make this probability conform to the customer's needs. Also note that during the computation of the runnerup, no nodes in the gold plane need be updated.
