Connectionist generalization for production: an example from GridFont

The authors designed and trained a connectionist network so that it could generate letterforms in a new font given just a few exemplars from that font. During learning, the network constructed a distributed internal representation of different fonts and letters, even though each training instance had both font characteristics and letter characteristics. For successful generation of letterforms, it was found that it was necessary to have separate but interconnected hidden units for `letter´ and `font´ representations. The limitations of the network can be attributed, in part, to limited training data