Examination of the Neural and Semantic Structure of Abstract Concepts

CNBC Brain Bag
Center for the Neural Basis of Cognition (CNBC)

Examination of the Neural and Semantic Structure of Abstract Concepts

Robert Vargas
Graduate Student
Carnegie Mellon University
February 18, 2019 - 6:00pm
Mellon Institute Social Room

The abstractness of concepts is typically defined indirectly as lacking concreteness, thus failing to specify their semantic basis and providing little insight into their cognitive or neural basis. Multivariate pattern analytic (MVPA) techniques applied to fMRI data were used to more directly characterize the neural representations of 28 individual abstract concepts. These representations were defined as the activation levels across 120 voxels with a stable semantic tuning curve across the 28 concepts. A classifier (Gaussian Naive Bayes), trained on the neural signatures of these concepts in a subset of the data for each participant, decoded the concepts in an independent subset (with a mean rank accuracy of 0.82). There was considerable commonality of the neural representations across participants as indicated by accurate classification of concepts based on their neural signatures obtained in other participants (mean rank accuracy was 0.74). Group-level factor analysis revealed 3 semantic dimensions underlying the activation patterns corresponding to the 28 concepts, suggesting a brain-based ontology for this set of abstract concepts. A cross-validated predictive model, based on independent behavioral ratings of the 28 concepts along the 3 dimensions, provided converging evidence for the interpretations of the factors (mean rank accuracy was 0.71). The location and semantic content of these 3 semantic dimensions suggest that the representations of abstract concepts rely on higher-level cognitive functions, namely language, self-representation, and social interaction.