Adaptive Kalman filtering for closed-loop Brain-Machine Interface systems

Brain-Machine Interface (BMI) decoding algorithms are often trained offline, but this paradigm ignores both the non-stationarity of neural signals and the feedback that exists in online, closed-loop control. To address these problems, we have developed an Adaptive Kalman Filter (AKF), a Kalman filter variant that adaptively updates its model parameters during training. For a Kalman filter decoder, batch retraining methods require completely re-estimating the parameter matrices from sufficient data to perform regression accurately, even if only small changes are necessary. Conversely, the AKF is designed to update the decoder parameters continuously and more intelligently. We simulated a population of 41 neurons learning to control a 2D computer cursor. The AKF yielded significantly faster skill acquisition and better robustness to perturbation and neuron loss than a standard Kalman filter with periodic batch retraining.