New eGRU Model Enables Efficient Neural Decoding for Neuroprosthetics

A team led by Khaleelulla Khan Nazeer (TU Dresden) has introduced an event-based neural decoding method for neuroprosthetic motor control. Their key innovation is an event-based gated recurrent unit (eGRU) that generates sparse communication patterns using graded spikes—unlike classical spiking neural networks (SNNs) that rely on binary spikes. This design allows the model to surpass traditional SNNs in task performance while keeping inference efficient. The work aims to overcome major barriers in prosthetic adoption: high latency, energy demands, and the need for wired connections to bulky external processors. By enabling on-device decoding, the eGRU reduces the volume of data transmitted wirelessly, making brain-controlled prosthetics more practical and mobile.

The researchers employed an efficient training method tailored for sparse inference, further optimizing the model for embedded hardware. The approach was validated on neural decoding tasks relevant to motor control prostheses. Published at the 2025 IEEE Biomedical Circuits and Systems Conference (BioCAS), the paper demonstrates that event-driven architectures can achieve competitive accuracy while dramatically lowering power consumption and latency. This opens the door to next-generation neuroprosthetics that operate entirely on a wearable chip, freeing users from external computers. As the field moves toward implantable and wearable brain-computer interfaces, the eGRU’s combination of performance and efficiency marks a significant step forward.