Purchase this article with an account.
Leslie Blaha, James Townsend; A Hebbian-style dynamic systems model of configural learning. Journal of Vision 2008;8(6):1126. doi: 10.1167/8.6.1126.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Configural learning is the process by which configural perceptual representations and processing mechanisms develop. Blaha and colleagues (Blaha & Busey, VSS 2007; Blaha & Townsend, Under Revision) characterized this process by a qualitative shift in processing capacity, reflecting a fundamental change in the efficiency of the visual information processing system. To capture these empirical findings, we developed a neurologically-plausible dynamic processing model of configural learning. We began with an interactive parallel processing model defined by simultaneously-operating linear accumulator channels (Townsend & Wenger, 2004). In this system, learning is defined by a non-linear, recursive function operating on both the within- and across-channel activation weight parameters of the linear accumulators. The activation weight parameters define the system's cross-talk, or the ways in which information from each channel interacts to facilitate or inhibit information accumulation in other channels. Over the course of simulated training, the learning function incrementally changes the activation weights, effectively adjusting the levels of interchannel cross-talk. We applied this model to Blaha and Townsend's (Under Revision; VSS 2006) configural learning data, in which participants unitized novel objects requiring conjunctive categorization of all object features. Unitization training of adult participants revealed a shift from extreme limited capacity to the super capacity processing efficiency characteristic of configural processing. Within the Hebbian-style model, the capacity limitations observed early in training were best captured by setting the interchannel cross-talk to be inhibitory (negative activation weights). As recursive feedback learning transformed the activation weight parameters to positive values, the model behavior mimicked the empirically observed shift to super capacity processing. The resulting trained model was a facilitatory parallel system exhibiting super capacity efficiency, in accordance with Wenger and Townsend's (2001) working definition of configural information processing.
This PDF is available to Subscribers Only