Poster C75, Sunday, March 26, 5:00 – 7:00 pm, Pacific Concourse
Semantic grounding in a neurocomputational model including realistic connectivity and spiking neurons
Rosario Tomasello1,2, Max Garagnani1,4, Thomas Wennekers3, Friedemann Pulvermüller1,2; 1Freie Universität Berlin, Brain Language Laboratory, 2Humboldt-Universität zu Berlin, Berlin School of Mind and Brain, 3University of Plymouth, Centre for Robotics and Neural Systems (CRNS), 4Goldsmiths, University Of London
Previous neurocomputational work has addressed the question why and how many cortical areas contribute to semantic processing and, specifically, why semantic hubs involved in all types of semantics contrast with category-specific areas preferentially processing certain meaning subtypes. However, much of the pre-existing work used either basic neuron models or much-simplified connectivity so that a more sophisticated and biologically-realistic model would be desirable. Here, we applied a neural-network model replicating anatomical and physiological features of a range of cortical areas in the temporal-occipital and frontal lobes to simulate the learning of semantic relationships between word-forms and specific object perceptions and motor movements of the own body. The two neuronal architecture differed in the level of detail with which cortico-cortical connectivity was implemented. Furthermore, one model adopted a mean-field approach by using graded-response neurons, whereas the other implemented leaky integrate-and-fire neurons. Equipped with correlation-based learning rules and under the impact of repeated sensorimotor pattern presentations, both models showed spontaneous emergence of specific tightly interlinked cell assemblies within the larger networks, interlinking the processing of word-form information to that of sensorimotor semantic information. Both models also showed category-specificity in the cortical distribution of word-related circuits, with high-degree connection hub areas central to the network architecture exhibiting involvement in all types of semantic processing and only moderate category-specificity. The present simulations account for the emergence of both category-specific and general-semantic hub areas in the human brain and show that realistic neurocomputational models at different levels of detail consistently provide such explanation.
Topic Area: LANGUAGE: Semantic