An efficient approximation for the real-time implementation of the Mixture of Trajectory Models decoder

Abstract

Abstract-The Mixture of Trajectory Models (MTM) decoder has been used to reconstruct arm trajectories from neural activity. While it produces reasonable results, the computational demands of previously published versions may be too high for many real-time systems. We have developed a novel method of approximating the MTM state posteriors that does not require the use of Newton's method. We show that this method results in only a small decrease in decoding performance yet reduces computational cost by 56.4%. Additionally, an MTM algorithm using this method of approximating the state posteriors produces more accurate decoded trajectories when using small bin sizes than an MTM algorithm using a Gaussian observation model. The more efficient formulation of the MTM algorithm presented here provides an alternative approximation of this algorithm for use on resource constrained embedded systems.