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

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

Generalizable, real-time neural decoding with hybrid state-space models

Avery Hee-Woon Ryoo¹, Nanda H Krishna², Ximeng Mao³, Matthew G Perich², Guillaume Lajoie⁴; ¹Mila - Québec AI Institute, Université de Montréal, ²Université de Montréal, ³University of Montreal, ⁴Mila - Quebec Artificial Intelligence Institute

Presenter: Avery Hee-Woon Ryoo

Brain-computer interfaces (BCIs) offer a promising approach for restoring mobility and communication in individuals with paralysis, motor impairments, or neurodegenerative diseases. This is achieved by learning a "decoder", which translates neural activity to an intended behaviour. Traditional decoding approaches often rely on simple statistical methods and recurrent neural networks that are highly specific to individual sessions of data collection, resulting in models that struggle to generalize to new data. On the other hand, empowered by the availability of large-scale multi-session neural datasets, recently developed transformer-based decoders demonstrate strong generalization performance, but are ill-suited for real-time inference due to their high computational complexity. To bridge the gap between these two approaches, we propose POSSM, a hybrid architecture that combines attention with a state-space model backbone. Our trained models demonstrate efficient real-time inference on intervals that are a fraction of a second long, while also demonstrating strong generalization to unseen sessions and individuals, thereby preserving the strengths of both the traditional and modern approaches to neural decoding.

Topic Area: Brain Networks & Neural Dynamics

Extended Abstract: Full Text PDF