Generali:ation .lot" production is a diJfi¢ult problem soh,ed hy intelligent systems. One such problem might be "paint a portrait of George Bush in the style ~f lTn¢ent van Gogh." In these tasks, a hm' dimensional representation ("George Bush "' and "styh, of van Gogh ") is exT~an&'d into a (nonunique) output ~?f extremely high dimension../or instance represented by color, line. shading, etc. g ? des~ened a ¢omwctionist network.6," generalization of production in such a way--to generate lettel'/brms in a new/imt given just a./ew evemplarL/kom that font. During learning, our network constructs distributed internal representations of d~[.]t, rent./imts attd letters', even thottgh each training instance had both /ont ¢haracteristkw and h, tter characteristics. Separate hidden representations~or "letter" and ']bnt" were ne¢essary./br Slt¢¢t, ss o/'the network. The integration ql'it!/brmation.from limt attd letter representations had to be at a proper, intermediate level of abstraction. The limitations o[the ttt, Iwork can be attributed, in part. to limited training corpus, and hick of translation and scale invariances.