An Investigation of Balance Scale Success

Decreasing the number of training presentations in all models caused slower convergence, as did increasing training bias. The precise effects of segregating hidden units depended on the method of weight updating, but the general principle was that psychological realism followed slow convergence. One ramification of the current findings is that models, like humans, need not have access to all of the information about a problem in order to succeed in finding a solution. Indeed, if models are supplied with complete information, realistic effects do not occur. Shultz (1991) suggested that stages would emerge whenever network models solve part of the overall problem before solving the range of possible problem types. Among the techniques he listed for encouraging partial solutions (and thus stages) were hidden unit herding, over-generalization, training pattern bias, and hidden unit recruitment. Working on too much of the problem at once may encourage overly rapid convergence on a general solution and thus preclude the appearance of stages. The present findings would suggest that all of these methods slow network convergence. A useful heuristic to apply in the creation of connectionist models of cognitive development may be to consider possible convergence slowing assumptions that bear on the problem domain. A phenomenon that may be related to the current findings is Elman's (1991) "starting small" effect. Elman reported that recurrent networks had difficulty learning a small grammar unless there was a gradual increase in either the complexity of the training instances or the "working memory capacity" of the network. Here we find analogously that models trained with a reduced number of exemplars perform more realistically, and in the case of batch updating under a heavy bias, often fail to discover a solution unless trained with a small subset of training examples. An important issue for future investigation is whether the staging of the child's environment and her developing cognitive resources work in concert to selectively filter information accessible to learning. A second result of the current work is that McClelland's (1988) specific set of assumptions is one of several sufficiently capable of producing realistic psychological performance. Although the incremental nature of connectionist learning is crucial for the success of balance scale models (Schmidt, 1991), so is convergence slowing. It would appear that some assumptions of McClelland's original model are replaceable. The current findings demonstrate balance scale performance without the architectural assumption of a segregated network topology. The bias and subset …