Biomimetic and Non-biomimetic Extraction of Motor Control Signals Through Matched Filtering of Neural Population Dynamics

Brain-machine interfaces rely on extracting motor control signals from brain activity in real time to actuate external devices such as robotic limbs. Whereas biomimetic approaches to neural decoding use motor imagery/observation signals, non-biomimetic approaches assign an arbirary transformation that maps neural activity to motor control signals. In this work, we present a unified framework for the design of both biomimetic and non-biomimetic decoders based on kernel-based system identification. This framework seamlessly incorporates the neural population dynamics in the decoder design, is particularly robust even with short training data records, and results in decoders with small filter delays. The theory and results presented here provide a new formulation of optimal linear decoding, a formal method for designing non-biomimetic decoders, and a set of proposed metrics for assessing decoding performance from an online control perspective. The theoretical framework is also applicable to the design of closed-loop neural control schemes.