Poster Presentation

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Dynamical systems analysis of a large-scale cortical language model enables mechanistic interpretation

Agata Feledyn¹, Tilo Schwalger², Thomas Wennekers³; ¹Freie Universität Berlin, ²Bernstein Center for Computational Neuroscience Berlin, Technische Universität Berlin, ³University of Plymouth

Presenter: Agata Feledyn

The meanings of words and concepts arise in the brain as distributed assemblies of cortical populations spanning sensory-motor and associative areas. Understanding how such representations form and propagate across the brain during comprehension processes pose a key challenge to cognitive and language neuroscience. Large-scale cortical models grounded in neuroanatomy provide a natural platform for this question, yet their complexity makes it difficult to relate observed dynamics to underlying circuit properties. We apply Wilson-Cowan mean-field theory and dynamical systems analysis to the MatCo model, a 12-area neural field model of language-related cortex, and derive analytical stability boundaries that partition parameter space into distinct dynamical regimes. We find the individual assembly excitatory-inhibitory units – representing cortical columns, exhibit heterogeneous firing rates arising from variability in local connectivity and synaptic strength, producing complex within-assembly dynamics. To facilitate further analysis, we normalise synaptic inputs to each column, yielding a cell assembly population that is homogeneous with respect to firing rates while preserving heterogeneity in the underlying structural connectivity. This normalisation enables a clearer characterisation of dynamical regimes – from transient stimulus-driven responses to sustained reverberant activity – which are accurately predicted by the theoretical analysis. Inter-areal propagation is governed by proximity to a saddle-node stability boundary: only near this boundary does recurrent amplification successfully recruit secondary and associative areas across different modalities, linking local excitatory–inhibitory balance to the conditions for distributed semantic memory retrieval. These results ground a complex cortical language model in dynamical systems theory, opening a principled path toward mechanistic accounts of language and meaning, and offering a transferable framework for related large-scale models.

Topic Area: Methods, Tools, Theory & Neural Coding