A dynamic Hebbian-style model of configural learning

The purpose of this dissertation is to develop a new dynamic systems model of configural learning. Configural learning is the process of perceptually chunking small pieces of information into fewer, larger perceptual units, which results in fast and highly accurate expert-like performance on the trained modality. In the present research, I formalized an information processing definition of configural learning: the development of mechanisms processing the learned features in a facilitatory interactive parallel information processing architecture under a mandatory exhaustive stopping rule, which exhibits a shift from limited to super capacity work-load efficiency over training. In a series of experiments, I utilized the methodologies and measures of Systems Factorial Technology (SFT) to assay the information processing characteristics of configural learning in experiments of unitization learning. In the first experiment, I replicated and extended a categorization design for unitization learning which incorporated a work-load manipulation necessary for measuring capacity. A series of new results were derived for the existent exhaustive processing capacity coefficient, and a new coefficient is developed for single-target self-terminating processing. Results demonstrated that configural learning is characterized by a shift from consistent limited capacity to consistent super capacity. Capacity results together with new model predictions suggest the architecture involved is parallel. A second experiment explicitly incorporated the SFT test of architecture: the survivor interaction contrast function. Results from the second experiment suggest that stimulus characteristics may modulate the changes in capacity over training. While some limited evidence was obtained for a parallel processing architecture supporting configural learning, the specific tests of architecture were largely inconclusive, due to the difficulty of selectively influencing objects in a task encouraging interactions between the processing channels. I then developed a dynamic systems model of configural learning, which combines a simple linear accumulator model of parallel processing with new Hebbian-style recurrent learning rules. The learning rule slowly modifies the level of interaction between the parallel channels. The empirical results were best captured by an initially inhibitory model, wherein the parallel channels compete for processing resources, which changes to a facilitatory model via the Hebbian-style learning rule. This model is consistent with the formal definition of configural learning and captured the empirical observations, particularly the changes in processing capacity, within a single parallel information processing architecture.