Poster B82, Sunday, March 26, 8:00 – 10:00 am, Pacific Concourse
A neurocomputational model of lexical-semantic memory based on feature representation of concepts
Cristiano Cuppini1, Eleonora Catricalà2, Elisa Magosso1, Stefano Cappa2, Mauro Ursino1; 1University of Bologna, Italy, 2IUSS, Pavia, Italy
The lexical-semantic representation of concepts is object of active research in cognitive neuroscience, but the fundamental neural mechanisms are still debated. Several frameworks conceive concepts as sets of semantic features. However, the contribution of the different features in concept identification remains still to be identified. Aim of this work is to investigate, through an attractor neural network, the role of different semantic features in concept identification, simulating both normal and neurodegenerative conditions. The model includes semantic and lexical layers, coding for object features and word-forms respectively. Synapses are created using Hebb rules of potentiation and depotentiation. The main novelty consists in the use of a fixed presynaptic threshold and a post-synaptic threshold that increases with the frequency of features (linked to its saliency). This allows the formation of Semantic networks (auto-association) and lexical-semantic networks (hetero-association) with asymmetrical synapses, able to store for any given concept the effective saliency of each feature. The model was tested with two taxonomies: animals and tools, taken from a database of concept features, including shared and distinctive features with different saliency. The trained network solves object-recognition and object-naming tasks, providing a different role for salient vs. marginal features in concept identification. In case of damage, superordinate concepts are preserved better than the subordinate ones. Interestingly, the degradation of salient features, but not of the marginal ones, prevents object identification. The model suggests that Hebb rules, with adjustable post-synaptic thresholds, can provide a reliable semantic representation of objects, exploiting the statistics of input features.
Topic Area: LANGUAGE: Semantic