Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Uncovering Linguistic Representations in MEG Data Using Deep Learning

Nikola Kölbl¹, Abhinav Singh¹, Patrick Krauss², Achim Schilling³; ¹Friedrich-Alexander-Universität Erlangen-Nürnberg, ²University Erlangen-Nuremberg, ³Friedrich-Alexander Universität Erlangen-Nürnberg

Presenter: Nikola Kölbl

The ability to use complex language is uniquely human and underpins abstract thought, cultural transmission, and the structure of society. Understanding its neural basis in naturalistic settings remains a major challenge. In this study, we investigate whether word classes can be decoded from low-dimensional MEG data recorded during audio book listening. Using a minimalist modeling approach, we trained neural networks on individual MEG channels and identified peak classification performance over left frontal sensors, consistent with the involvement of Broca’s area in grammar and predictive processing. As a proof of concept, we applied sequential deep dreaming to reveal prototypical neural patterns for nouns and verbs. While the results demonstrate feasibility, limitations due to data sparsity, class imbalance and single-subject design highlight the need for broader validation. Our approach represents a first step towards interpretable decoding of linguistic structure from MEG during natural, continuous speech comprehension.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF