Efficient cortical spike train decoding for brain-machine interface implants with recurrent spiking neural networks

Invasive cortical brain-machine interfaces (BMIs) can significantly improve life quality for motor-impaired patients. However, externally mounted BMIs carry the risk of wound infection, necessitating fully implanted systems with integrated signal processing. Such systems must ensure reliable and accurate data processing while meeting tight latency and energy constraints. To meet these processing requirements, spiking neural networks (SNNs) seem an attractive solution due to their minimal preprocessing requirements of cortical spike trains (CSTs) and their suitability for ultra-low-power, low-latency neuromorphic hardware. While explorative studies give a glimpse of their capabilities (Taeckens and Shah, 2023; Liao et al., 2024) in strictly feed-forward SNNs (fSNNs), competitive solutions for BMI applications remain largely unexplored.
In this study, we investigate the capability of recurrent SNNs for real-time regression in decoding motor outputs directly from CSTs. We used surrogate gradients to train small recurrent SNNs (tinySRNNs) composed of a single hidden layer of 64 adaptive leaky integrate-and-fire (ALIF) neurons to decode finger velocity from CSTs collected from the primary motor cortex of two macaque monkeys (Indy & Loco) performing a reaching task. After training (5-fold cross-validation, 4 different initializations for each fold), tinySRNNs significantly outperformed conventional CST decoding approaches, such as Kalman filtering (KF) and long-short-term-memory (LSTM) architectures as well as fSNNs in terms of correlation coefficients (CCs) and mean square error between the ground truth and decoded finger velocity. Architectural ablation experiments revealed that besides adding recurrent connections, trainable time constants (TCs) in particular contribute to the decoding performance of tinySRNNs, while the adaptation mechanism of the ALIF neurons provides further small performance improvements. Furthermore, by maintaining a low average population firing rate, tinySRNNs require only about 1.25% of the theoretical energy consumption of LSTMs of comparable size. Our results thus underscore the immense potential to revolutionize patient care by integrating tinySRNNs into fully implanted ultra-low-power neuromorphic BMIs.