A modern approach to restoring mobility in physically disabled patients involves neural implants, allowing for control of robotic prosthetics via recorded neural activity. An outstanding issue with this approach is the need for a wired set-up for real-time movement inference. Existing wireless solutions face the challenge of on-chip computation in a power-constrained environment, necessitating battery replacement surgeries, and increasing the risk of complication. Low-power, brain-inspired neuromorphic hardware addresses the power demands of on-chip machine learning, although they are incompatible with traditional neural networks. Likewise, a novel approach to motor imagery brain-computer interfaces native to neuromorphic hardware was developed using spiking neural networks. A 70x power consumption reduction and 4x parameter count decrease were realized while maintaining state-of-the-art accuracy on EEG datasets. These results show that the proposed pipeline has tremendous potential in increasing the portability of neural implants and providing a more attractive solution to the market.
anmutechsupport / enerspike Goto Github PK
View Code? Open in Web Editor NEWEnerspike: Pioneering Wireless Mind-Control Implants Using Energy-Efficient Neuromorphic Computing