PDP Learnability and Innate Knowledge of Language
David Kirsh · 1992
Abstract It is widely assumed that PDP learnability has some bearing on questions of innateness. If a PDP network could be trained to make correct judgements of grammaticality, for instance, it seems to follow that innate knowledge of grammar is not necessary for language acquisition. The reason, quite simply, is that the learning rules used in PDP learning -whether backpropagation or related gradient descent methods-are general, domain independent methods. They are what AI theorists call weak methods. Hence in teaching a system to make correct judgements, we seem to have an existence proof that there is enough information in the stimulus to permit learning by inductive means alone. It is this idea, and the methodological implications that flow from believing it, that I wish to explore here. The problem I have with this argument is that to discover a network that will learn successfully, designers must choose with care the network’s architecture, the initial values the weights are set to, the learning rule, and the number of times the data set is to be presented to the network-this latter parameter effects the smoothness of the estimated function. If such parameters are not controlled for, successful learning is extremely improbable. In thoughtful modelling, these parameters are chosen on the basis of assumptions about the nature of the function the system is to learn. That is, on the basis of assumptions about the task and the task domain. Prima fade, then, although the learning mechanism operating on data is a general one, the success of this mechanism depends equally on a set of antecedent choices that seem to be domain specific.