Proceedings of the Annual Meeting of the Cognitive Science Society

Neural network models of reading provide a good account of many aspects of normal and disordered performance, butthe training procedures are unrealistically comprehensive. Unlike models, child learners have limited instructional timeand are explicitly taught only a small subset of the words they will regularly encounter. To address this discrepancy weinvestigated alternative learning procedures in a standard orthography-to-phonology multilayer network, to identify oneswhich involve a small, teachable subset of words that facilitate learning untrained words with less effort. We also asked,for any such set, whether the procedure can be improved by optimizing the training sequence. Candidate training setsand sequences were derived using a model-based procedure and from elementary reading curricula. The results indicatetradeoffs between the size and composition of the training set and generalization. These procedures suggest pedagogicalsolutions to the problem of learning more than there is time to teach.