Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Dynamical conditions for distributed semantic memory retrieval in language-related cortex
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
Agata Feledyn1,4, Tilo Schwalger2,4, Thomas Wennekers3; 1Freie Universität Berlin, 2Technische Universität Berlin, 3University of Plymouth, 4Bernstein Center for Computational Neuroscience Berlin
Introduction. It has been proposed that word meanings are grounded in the sensorimotor cortices that process their referents, with semantic representations taking the form of distributed cell assemblies spanning multiple cortical areas (Pulvermüller, 1999). Neuroimaging and electrophysiological evidence supports this view: action words preferentially recruit frontocentral motor cortex in a somatotopic fashion, while words with strong visual associations activate occipitotemporal cortex (Hauk et al., 2004; Pulvermüller et al., 2013). A central open question is what circuit-level conditions allow a sensory stimulus to trigger retrieval of a full distributed semantic representation — recruiting the entire cell assembly encoding a word's meaning. A brain-constrained 12-area neural field model of language-related cortex, in which word meanings are encoded via associative Hebbian learning across neuroanatomically constrained areas, provides a natural platform for this question. Successfully applied to symbol grounding, language-attention interactions, and visual cortex recruitment in blind individuals (Garagnani et al., 2008; Tomasello et al., 2018, 2019), the model has nonetheless been explored exclusively through simulation, leaving local circuit properties and large-scale dynamics inaccessible to direct mechanistic analysis. Methods. We map the local excitatory–inhibitory (E–I) circuit of a single model area onto Wilson–Cowan mean-field equations with piecewise-linear rate functions, deriving closed-form saddle-node (SN) and Hopf-like (H) bifurcation conditions in terms of effective excitatory and inhibitory loop gains, partitioning parameter space into input-driven, oscillatory, and bistable regimes. For the multi-area analysis, excitatory input is normalised separately for within-area and inter-areal projections at 80% and 20% respectively, consistent with empirical estimates of local versus long-range synaptic balance (Markov et al., 2013). Results. Under the original connectivity, neurons within a single cell assembly scatter across multiple dynamical regimes, reflecting heterogeneity from non-uniform excitatory drive — likely a general feature of cortical circuits rather than a modelling artefact. Normalising input weights collapses the assembly to a single operating point, making the mean-field approximation valid. The bifurcation analysis then accurately predicts single-area dynamics across the full parameter range: input-driven responses, damped oscillations near the H boundary, sustained oscillations beyond it, and persistent reverberant activity beyond SN. In the full 12-area network, inter-areal propagation is critically governed by SN proximity. Far from it, stimulation of a primary auditory area produces transient, localised activity. Approaching the boundary, recurrent amplification progressively recruits secondary and associative areas across all four functional systems — visual, auditory, articulatory, and hand-motor — mirroring the distributed structure of a semantic cell assembly. Conclusion. The conditions for distributed semantic memory retrieval are governed by identifiable dynamical principles: proximity to the saddle-node boundary is the circuit-level prerequisite for recruiting a full distributed cortical language representation. This framework renders a complex, biologically constrained model analytically tractable, replacing heuristic parameter searches with principled predictions and opening a mechanistic path toward understanding how the brain encodes and retrieves word meanings.
Topic Areas: Computational Approaches, Multisensory or Sensorimotor Integration