What is Priming and Why?

Abstract

and Specific Visual-Form Recognition and Priming In this section, two additional subsystems are briefly described to exemplify theory development within the approach of theorizing about functions of neural subsystems. Stemming from the mechanistic theory of priming above, two critical subsystems of consolidated memory in the neocortex may be an abstract-category subsystem and a specific-exemplar subsystem. Both accomplish visual-form recognition and priming, but in different ways to subserve contradictory transformations and goals. Computational Analysis Fundamental to visual form recognition is the ability to recognize the abstract category to which an input shape corresponds (e.g., cup versus pen, etc.) as well as the ability to recognize the specific exemplar to which that same input shape corresponds (e.g., an individual pen). Postvisual feedback can help a visual subsystem to learn that multiple input shapes (even dissimilar ones; e.g., an upright piano and a grand piano) should be categorized together because they are associated with the same post-visual information. Post-visual feedback also can help to learn that multiple input shapes (even similar ones; e.g., two highly similar upright pianos) should be distinguished because they correspond to different individual object entities in the world. Interestingly, mapping an input shape to its category representation and mapping that input to its exemplar representation involve contradictory computations when real-world stimuli are considered. Figure 3 helps to convey the computational analysis that leads to the theory of a two subsystems architecture. Object recognition can be conceptualized as instantiating a mapping from points in image space (retinotopically-mapped input representations for a visual-form recognition subsystem) to points in a long-term memory space (output representations from a visual-form recognition subsystem). First, dissimilar exemplars in a category reside in relatively distant points in image space, and they are mapped together for category recognition versus apart for exemplar recognition (Figure 3A). Such mappings are not contradictory; they can take place effectively in a common neural network model (Marsolek, 1992; Marsolek & Burgund, 1997). Second, similar exemplars in a category reside in relatively nearby points in image space, and they are mapped together for category recognition versus apart for exemplar recognition (Figure 3B). Such mappings also are not contradictory; they can take place effectively in a common neural network model (e.g., Hummel & Stankiewicz, 1998; Knapp & Anderson, 1984; Marsolek, 1992; Marsolek & Burgund, 1997; McClelland & Rumelhart, 1985). However, contradictory mapping solutions are demanded when categories contain both dissimilar exemplars and similar exemplars (as in most real world visual form categories; e.g., pianos). Assuming a common internal representation for the mappings, the transformations useful for bringing together dissimilar exemplars contradict the transformations useful for separating similar exemplars (Figure 3C; Marsolek, 1994; Marsolek & Burgund, 1997). A computationally useful solution is to separate the mappings across different sets of weights and internal representations (i.e., implement separate, parallel subnetworks for the two mappings; see Figure 3D). Note that the same argument applies to learning of visual word-form categories and exemplars (e.g., the same word printed in different letter cases or fonts). The internal representations that are most useful for abstract-category and specificexemplar mappings may be qualitatively different, suggesting important mechanistic differences between the two subsystems. Both processors should utilize sparse distributed activations, of the sort hypothesized above, but the two may differ in the degree of sparseness that is most useful. In J. S. Bowers & C. J. Marsolek (Eds.), What is Priming and Why? 15 Rethinking implicit memory. Oxford: Oxford University Press. Figure 3. Abstract-category mappings and specific-exemplar mappings are not contradictory when categories contain either dissimilar exemplars (A) or similar exemplars (B), but they are contradictory when categories contain both dissimilar and similar exemplars (C). A computationally useful solution to the latter problem is to implement separate subnetworks for the two mappings (D). See text for explanation. Specific Exemplar 2 Specific Exemplar 1 Specific Exemplar 1 A Specific Exemplar 2 Abstract Category 1 B Specific Exemplar 1 Specific Exemplar 2 Abstract Category 1

3 Figures and Tables

Cite this paper

@inproceedings{Marsolek2003WhatIP, title={What is Priming and Why?}, author={Chad J. Marsolek}, year={2003} }