Recurrent neural network decoder

Recurrent neural network decoders — most often long short-term memory (LSTM) or gated recurrent unit (GRU) networks — process neural signals as a sequence, maintaining a hidden state that carries temporal context across time steps. In intracortical motor and speech BCIs they replaced the linear Kalman filter's fixed first-order dynamics with a learned nonlinear temporal model, capturing longer-range dependencies (the slow evolution of a reach, or coarticulation between phonemes) that a linear state-space model cannot.

RNNs dominated speech and handwriting BCI results around 2019 to 2021 and remain competitive when latency budgets forbid the large context windows transformers prefer. Their weaknesses are the classic ones: sequential computation limits parallelism, gradients over long sequences are fragile, and — critically for BCI — they can drift silently when the input distribution shifts across days, so they are usually paired with recalibration or online normalization.