Bartos, Paul D.
Connectionist modelling of category learning.
The Open University.
Full text available as:
A shortcoming is identified with respect to the ability of exemplar-based connectionist models of category learning to offer accounts of learning about stimuli with variable dimensionality. Models which may simulate these tasks, such as the configural-cue network (Gluck & Bower, 1988b), appear to be unable to accurately simulate certain data well simulated by exemplar-based models such as ALCOVE (Kruschke, 1992). A task in which the advantage of ALCOVE is exemplified is the prediction of human learning rates on the six category structures tested by Shepard, Hovland, and Jenkins (1961). The ability of ALCOVE to simulate the observed order of difficulty depends on its incorporation of selective attention processes (Nosofsky, Gluck, Palmeri, McKinley, & Glauthier, 1994). This thesis focuses on developing configural-cue network models which incorporate these processes. Informed by an information-theoretic approach to modelling the implementation of selective attention using a configural-cue representation, five connectionist models are developed. Each is capable of predicting the order of difficulty reported by Shepard et al. (1961). Two models employ a modular structure, but analysis suggests that these may lack much of the functionality of the basic configural-cue network. The remaining three incorporate dimensional attention schemes. These models appear to offer superior generalisability in relation to the simulation of learning about variable dimensionality stimuli. This generalisability is examined by applying a variant of one of these dimensional attention models, to data collected by Kruschke (1996a) on the inverse base-rate effect and base-rate neglect. The model provides a qualitative fit to this data. The success of these configural-cue models on these two tasks, only successfully modelled previously using two distinct types of representation, indicates that the approach has some potential for further applications. Differences between the models applied, however, indicates that more sophisticated conceptions of the attention process may be required to allow further generalisability..
Actions (login may be required)