Intelligent Intracortical Brain-Machine Interfaces

Abstract

Intracortical brain machine interfaces (iBMIs) are essential in restoring the quality of life of severely motor impaired patients. They take neural activity as an input, which is then subject to signal processing and neural decoding, in order to drive prosthetics which enable paralyzed individuals perform activities of daily living. Significant technological advances and simple proof of concept demonstrations have been made in the recent past. Furthermore, in order to increase the dexterity of iBMI systems, the number of simultaneously recorded neurons is exponentially rising mimicking a Moore-like law with doubling observed after every 7 years. However, state-of-the-art iBMI systems involve bulky, wired devices that limit mobility and pose infection risks. The answer to these challenges lies in building wireless fully implantable iBMI systems that (a) adhere to the power dissipation constraint of 80 mW/cm² in cortical territory, and (b) are able to scale to an exponential increase in input data emanating from simultaneously recorded neurons. To this end, it is shown that embedding signal processing and machine learning capabilities in the implant lends itself virtually independent to the scaling of input neuronal data. Since machine learning-based inference is implemented in the implant itself, such a system is known as an intelligent iBMI (i²BMI). Reported i²BMI systems resort to unconventional forms of computing architectures in order to respect the power dissipation constraints dictated by the cortical territory. Prominent proof-of-concept bio-chips include spiking neural network (SNN)-based and random projection-based neural network (RPNN) implementations. In particular, bioinspired RPNN chips have reported a 10X increase in power efficiency, while yielding a superior performance over state-of-the-art algorithms such as Kalman filter.