research report announcement

[KRUSCHKE,JOHN,PSY]

    *** PLEASE DO NOT FORWARD TO OTHER BULLETIN BOARDS ***

Research Report announcement:

   ALCOVE: A Connectionist Model of Category Learning

                   John K. Kruschke
            Psychology and Cognitive Science
                  Indiana University

This report should interest cognitive scientists studying category *learning*, 
especially those familiar with the work of psychologists such as Nosofsky, 
Medin, Gluck & Bower, and Estes.  The report should also interest 
connectionists studying the abilities of back-prop architectures that use 
radial basis function nodes, and new architectures for selective attention.

                       ABSTRACT

ALCOVE is a new connectionist model of category learning that models the course
of learning in humans and their asymptotic performance.  The model is a variant
of back propagation, using Gaussian (radial basis function) hidden nodes, and
*adaptive attentional strengths* on the input dimensions.   Unlike standard
back propagation networks, ALCOVE cannot develop completely new dimensions for
representing the stimuli, but it does learn to differentially attend to the
given input dimensions. This constraint is an accurate reflection of human
performance. ALCOVE is successfully applied to several category learning
phenomena: (1)~It correctly orders the difficulty of the six category types
from the classic work of Shepard, Hovland and Jenkins (1961).  (2)~It
accurately fits trial-by-trial learning data and mimics the base-rate neglect
observed by Gluck and Bower (1988b).  In preliminary work, it is also shown
that ALCOVE can: (3)~exhibit three-stage learning of high-frequency exceptions
to rules (\cf\ Rumelhart \& McClelland 1986), (4)~show emergent graded internal
structure in categories, \ie, typicality ratings, (5)~produce asymmetries of
similarities between typical and atypical exemplars, (6)~show selective
sensitivity to correlated dimensions, and (7)~learn  non-linearly separable
categories faster than linearly separable categories, in those cases that
humans do.  It is also suggested that ALCOVE could serve as the input to a rule
generating system, so that the dimensions most attended are the ones first used
for rules. Moreover, it is shown that ALCOVE is falsifiable, in principle, and
that there are some phenomena in category learning that ALCOVE cannot capture. 
Nevertheless, ALCOVE is attractive because of the broad range of phenomena it
does model. 


If you are truly interested (supplies are limited), you are welcome to a free
copy by e-mailing your physical address to the Cognitive Science Program
secretary, Cathy Barnes, at 

  iucogsci at ucs.indiana.edu

Be sure to mention "Research Report #19 by John Kruschke".
(As usual, don't use the "reply" command to make your request.)

         *** PLEASE DO NOT FORWARD TO OTHER BULLETIN BOARDS ***