Cox, C. 1 , Seidenberg, M. 1 , Binder, J. 2 , Desai, R. 2 & Rogers, T. 1
1 University of Wisconsin - Madison
2 Medical College of Wisconsin
Neural network models of semantic memory often propose that semantic representations are highly distributed--the very same units can encode information about even quite disparate domains (such as animals and manmade objects), with many such units involved in the representation of any individual item (e.g. Rogers and McClelland, 2004; Patterson, Nestor, & Rogers, 2007). Much of the neuroimaging literature, however, supports the alternative hypothesis that knowledge about different conceptual domains is functionally localized in cortex, with different cortical areas dedicated to representing different kinds of concepts. We consider the possibility that such results may reflect a localizationist bias in standard fMRI methods. To reduce the likelihood of Type 1 error, for instance, it is common to spatially blur signal, to focus on contiguous regions of interest, to average data over multiple participants, and to adopt cluster-wise significance criteria---steps which assume that different functions are localized to different regions of cortex in the same way across individuals. We investigate a new approach that avoids these assumptions and so may be better suited to testing the hypothesis that semantic representations are highly distributed. Voxels of interest are identified by training a pattern classifier with sparse logistic regression, which jointly minimizes logistic loss the magnitude of the regression coefficients, effectively finding the best possible classifier with the smallest number of voxels. Simulation results show that this method matches or exceeds standard univariate approaches at finding signal-carrying voxels under many different assumptions about how these are anatomically distributed. Application of the method to fMRI data from a word-reading task revealed a set of voxels distributed along the ventral temporal and prefrontal cortices that carry information about both animal and artifact domains--suggesting that such representations are not localized in cortex but are highly distributed, as neural network models of semantic task performance have suggested.