Abstract-Recently, we proposed a Kalman filter method to model the probabilistic relationship between neural firing in motor cortex and hand kinematics. In this paper, we demonstrate on-line, closed-loop, neural control of cursor motion using the Kalman filter. In this task a monkey moves a cursor on a computer monitor using either a manipulandum or their neural activity recorded with a chronically implanted micro-electrode array. A number of advantages of the Kalman filter were explored during the on-line tasks and we found that the Kalman filter had superior performance to previously reported linear regression methods. While the results suggest the applicability of the Kalman filter for neural prosthesis applications, we observed the decoded cursor position was nosier under brain control as compared with manual control using the manipulandum. To smooth the cursor motion without decreasing accuracy we propose a method that smoothes the neural firing rates. This smoothing method is described and its validity is quantitatively evaluated with recorded data.Index Terms-neural prosthesis, neural control, brainmachine interface, closed-loop control, primary motor cortex, Kalman filter.

Abstract-While various automated spike sorting techniques have been developed, their impact on neural decoding has not been investigated. In this paper we extend previous Gaussian mixture models and Expectation Maximization (EM) techniques for automatic spike sorting [1]. We suggest that good initialization of EM is critical and can be achieved via spectral clustering. To account for noise we extend the mixture model to include a uniform outlier process. Automatically determining the number of neurons recorded per electrode is a challenging problem which we solve using a greedy optimization algorithm that selects models with different numbers of neurons according to their decoding accuracy. We focus on data recorded from motor cortex and evaluate performance with respect to the decoding of hand kinematics from firing rates. We found that spike trains obtained by our automated technique result in more accurate neural decoding than those obtained by human experts.