IMPROVED GENERALIZATION IN A RECURRENT HIERARCHICAL NETWORK USING A MIXTURE OF ERROR-DRIVEN AND HEBBIAN LEARNING