Poster D29, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Controlled Semantic Cognition Necessitates a Deep Multimodal Hub
Rebecca Jackson1, Timothy T. Rogers2, Matthew A. Lambon Ralph1;1MRC Cognition & Brain Sciences Unit, Cambridge University, 2University of Wisconsin-Madison
Semantic cognition, or the controlled representation and use of conceptual knowledge, is a core process underlying language. The semantic system must satisfy a number of essential properties. Principally, it must 1) learn to form coherent context-invariant conceptual representations by abstracting over episodes across time and by learning the complex non-linear relationships between features across different sensory modalities, and 2) dynamically use subsets of features to create a context-appropriate similarity space and produce context-dependent behaviours. These performance criteria are non-trivial to achieve, particularly because they necessitate the presence of and interaction between context-variant and context-invariant processes. A variety of different architectures can and have been theorised to subserve the semantic system, however, the ability of these architectures to synthesise context-invariant representations and task-specific outputs have never been formally tested. We designed a framework in which to test different possible architectures of a semantic network, to inform on the plausibility of various theorised candidate cortical architectures. We investigated the importance of five architectural features: a hub, a multimodal hub, depth, hierarchical convergence across modalities and the inclusion of sparse long-range connections. An architecture employing a single, deep multimodal hub with sparse long-range connections from modality-specific inputs, was identified as optimal. We also explored where the control signal should connect into the network, and the consequences of lesioning control and representation regions of the model. Implications for the neurobiology of the cortical semantic system, in health and disorder, are considered.
Themes: Meaning: Lexical Semantics, Computational Approaches
Method: Computational Modeling