An active-symbol connectionist model of concept representation and concept learning.

Many authors have emphasized the role that concepts play as basic building blocks of cognition. This suggests two things. First, if one wishes to understand cognition, one must understand how concepts are acquired and represented. Second, if an understanding of concepts is critical primarily because it serves as the basis for a general understanding of cognition, then a theory of concepts is most meaningful when developed in the context of a general cognitive theory and when connected to as many psychological phenomena as possible. The first point is widely appreciated--models of concept learning and representation are ubiquitious. The second point, though, is widely and unfortunately neglected. Because this second point is taken seriously, this dissertation represents a departure from the usual practices of connectionist modelling. An initial set of constraints is imposed by virute of working within the framework of an general cognitive theory known as "active symbol connectionism". Additional constraints are drawn from the domains of neuropsychology, object recognition, the psychology of categories and concepts, evolution, and connectionist modelling. Taking a "constraint-rich" approach has resulted in two classes of useful results. The first class comprises technical innovations in connectionist modelling. These include: (1) a method of using coarse-coding to improve the resolution of a parts-based models of object recognition and representation, (2) novel learning and activation rules based on an economic metaphor of diminishing returns, and (3) neurophysiologically plausible methods of implementing connectionist map learning. The struggle to accomodate multiple constraints has also suggested new perspectives for a number of issues. These include: (1) a recasting of the distributed vs. local representation debate in connectionism by breaking the issue down into its subcomponents, (2) an evolutionary perspective for the relative roles of statistical and theoretical mechanisms of categorization, and (3) the implications of separate cortical visual systems for connectionist models of concept representation.PhDComputer and Communication SciencesUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/103512/1/9319653.pdfDescription of 9319653.pdf : Restricted to UM users only