Chen, L. & Rogers, T.
University of Wisconsin-Madison
Several recent papers have reported a category-specific pattern of functional activation in the posterior fusiform cortex, with lateral regions responding more for animals than for tools and medial regions showing the reverse pattern. One interpretation of this finding is that lateral and medial fusiform regions encode visual properties common to animals or tools, respectively. This idea was recently challenged, however, by reports of a similar pattern in congenitally blind participants listening to the spoken names of animals or tools (Mahon et al., 2009). These authors advocated an alternative view, in which lateral and medial fusiform regions contribute to different cortical networks innately constrained to process animals versus tools. In simulations with a neural network model, we show how category-specific patterns can emerge in both sighted and blind individuals without innate domain-specific knowledge. Visual, lexical, and action representations in the model interact both directly and through a shared cross-modal “hub.” The direct pathway from vision to action is assumed to be graded in effectiveness, with medial visual units more strongly connected than lateral units. After learning with all modalities intact, a graded category-specific pattern emerged in the visual hidden layer, with “lateral” units responding more to animals and “medial” units responding more to tools. When the model was trained without input to visual units, the same selectivity nevertheless emerged, albeit somewhat weaker. When the architecture was adapted so that action representations did not connect directly to the cross-modal “hub,” the category-specific pattern emerged equally strongly in the “blind” and “sighted” model, just as observed by Mahon et al. (2009). The results thus suggest that apparent category-specificity may reflect very general initial constraints on effective connectivity between brain regions rather than innate domain-specific knowledge.