Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Asymmetric dynamical regimes underlie bilingual language control in intracranial recordings from Dutch-English speakers
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Lauren M. Ostrowski1,2, Paul Weger3, Sophia Gimple3, Maxime Verwoert3, Stéphanie K. Riès4, Vikash Gilja1,5, Timothy Q. Gentner1,6,7, Pieter L. Kubben3,8, Christian Herff3; 1Neurosciences Graduate Program, University of California San Diego, 2Medical Scientist Training Program, University of California San Diego, 3Department of Neurosurgery, Mental Health and Neuroscience Research Institute, Maastricht University, 4School of Speech, Language and Hearing Sciences, San Diego State University, 5Department of Electrical and Computer Engineering, University of California San Diego, 6Department of Psychology, University of California San Diego, 7Kavli Institute for Brain and Mind, University of California San Diego, 8Department of Neurosurgery, Maastricht University Medical Center
Bilingual speakers fluidly navigate production of their languages, supported by a well-documented network of brain regions. Here, we investigate the internal processes that govern the activity of that network. Employing a dynamical systems framework, we ask not only when and where the target language can be decoded from neural activity, but also how L1 (Dutch) and L2 (English) production differ in their underlying neural dynamics. We hypothesize that L1 and L2 are supported by asymmetric dynamical regimes, with L1 nearer the brain’s autonomous dynamics and requiring less internal drive to engage. We recorded intracranial EEG (iEEG) from six bilingual Dutch-English patients (2 female, age 19-52) undergoing pre-surgical epilepsy monitoring at Maastricht University Medical Center. Patients performed a picture-naming task in which images were preceded by a national flag (Dutch or British) cueing the target language, counterbalanced across languages and switch contexts to control for switch-cost effects. We extracted normalized high-gamma power (70-150 Hz) from all contacts and epoched around flag, image, and speech onset. A channel × timepoint searchlight classifier revealed single-contact language decodability across a distributed network spanning the prefrontal cortex, inferior frontal gyrus, middle and superior temporal gyri, anterior cingulate cortex, and hippocampus (label-permutation test, α=0.05). We then applied FINDR (Flow field Inference from Neural Data using deep Recurrent networks; Kim et al. 2025), adapted here for continuous iEEG, to learn the low-dimensional latent neural dynamics underlying speech production and to quantify the latent flow or “push” displacing the population from its baseline configuration into L1/L2 regimes. Language identity was encoded as a sustained context variable from flag onset, parameterizing the latent vector field and partitioning neural trajectories into autonomous and language-conditioned components. FINDR identified a low-dimensional latent space in which neural trajectories unfolded. L1 and L2 trials separated into distinct regions of this latent space along a Linear Discriminant Analysis (LDA) axis (centroid Mahalanobis distance, p<0.001). Gradient-based attribution of the LDA axis recovered the language-selective contacts identified by the searchlight analysis, along with others below single-contact detection thresholds. Across participants, the autonomous fixed point of the system lay closer to L1 than L2 (Driscoll et al., 2024; Sussillo & Barak, 2013). The language-conditioned component of the latent flow was larger for L2 than L1 (Wilcoxon signed-rank tests; p<0.001), indicating a stronger “push” to reach L2. L1 occupied more locally stable regions of latent space than L2 in four of six participants (more contractive Jacobian spectra; paired permutation tests; p<0.001), consistent with an L1-default regime. In one participant with focal coverage (ECoG over left speech motor and superior temporal cortex), both languages occupied locally stable regions on their respective trials. A sixth participant, who responded faster on L2 trials, showed the opposite asymmetry, with more stability under L2. By modeling the neural dynamics underlying observed iEEG activity, we reveal process-level mechanisms of bilingual language control. The asymmetric regimes we identify offer a neurobiological basis for L1 as the preferred state of the bilingual language system, while accommodating speakers for whom high L2 proficiency reduces or reverses that asymmetry.
Topic Areas: Multilingualism, Computational Approaches