A synthetic averaging method on point process for motor Brain Machine Interfaces

Previous decoding algorithms for brain machine interfaces (BMIs) are normally utilized to estimate animalpsilas movement from binned spike rates, which loses spike timing resolution and may exclude rich neural dynamics due to single spikes. The Monte Carlo sequential estimation on point process enables the decoding directly from the multi-neuron spike trains without constraints on posterior density estimation. However, it poses the problem to bridge the time scale gap between the spike events and the behaviors, and generates stochastic kinematic estimation due to the randomness embedded in the time structure of the spike trains. We propose here a synthetic averaging method to mimic the neural population effects by generating synthetic spike trains from neural recordings and utilizing them as extra model observation. The decoding performance by Monte Carlo SE is averaged in the kinematics domain, which preserves the resolution of the neuron activity and bypasses the possible distortion by nonlinear tuning function due to the binning in spike domain. Tested in simulation and real data, synthetic averaging provides better reconstructions with less NMSE.