 |
|
A third cluster, which cuts across the first two areas and also encompasses other theoretical fields in which RuCCS has particular strength, concerns such issues as the architecture of the cognitive system. This includes the nature of the underlying (computational) mechanisms of cognition, as well as the way in which these mechanisms encode information and the way that both the information and the mechanisms may change over time through learning or development. These are among basic theoretical and foundational questions on which RuCCS members have written extensively. For example, Fodor, Pylyshyn, Lepore, Stich, McLaughlin and Matthews have written on constraints on Cognitive Architecture, including its modularity and the requirements placed on it by the empirical facts of systematicity and productivity. These writers have argued, among other things, that the representational systems provided by the architecture must meet certain general but powerful constraints which exclude, for example, "prototypes" or analogs, or "pictorial" pics, or any other noncompositional representational system, from being by itself adequate to the task. Areas in which RuCCS faculty have been leaders involve an analysis of the nature of concepts and their mental representation, the nature of mental pics, and the fundamental problems faced by Connectionist models as models of cognitive processes (on which Fodor, Pylyshyn, McLaughlin, Stich, and Prince have written extensively).
Some RuCCS members or associates are concerned with the problem of representing knowledge in general (McCarty, Hirsh, Borgida, Amarel) or in particular domains, such as vision or language or planning and problem-solving (see the description of relevant research clusters). Some are concerned with understanding how representations change in response to relevant information -- i.e., learning in persons and machines (Hirsh, Amarel, Ellman, and Schmidt in the case of general knowledge; Fodor, Grimshaw, Matthews, Prince, Stevenson, Stromswold in the case of language; Julesz, Leslie, Dickinson, Feldman in the case of visual perception). Others are concerned to show how certain mechanisms are innate and/or universal or how they develop ontogenetically. For example, Leslie has shown that sophisticated cognitive mechanisms are present at very early stages of infancy (e.g. perception of causation and objecthood), whereas other mechanisms take a number of years to develop (e.g. ascribing mental states, such as beliefs and desires, to others). The study of universals of language and perception also sheds light on the nature of the cognitive architecture. In this vein Grimshaw and Prince have argued that both phonological and syntactic differences among languages can be accounted for by the ordering which they place on a finite set of universal but violable constraints. Leyton, Feldman, Dickinson and Julesz have shown that certain primitive form properties may be the (possibly innate and universal) basis for the encoding of shape in general. Stromswold has analyzed empirical evidence showing that certain structures are acquired despite mpoverished evidence that lacks negative feedback. The development of computational models of cognition raises both philosophical and methodological questions that many RuCCS members have addressed (e.g. Stich, Fodor, Pylyshyn). From the particular perspective of RuCCS researchers, the attempt to build specific computational models in domains such as language comprehension and visual perception goes hand in hand with designing an underlying architecture in terms of converging computational, linguistic, geometrical and psychological constraints. When this is done then the model's match to the observed human behavior is an nescapable consequence of that architecture, rather than based on mere mimicry. For example, Stevenson and Grimshaw are investigating the architectural requirements of a model of human parsing based on principles developed from Prince et al's Optimality Theory.