Nonlinear point-process estimation of neural spiking activity based on variational Bayesian inference.

Abstract

<i>Objective.</i>Understanding neural encoding and decoding processes are crucial to the development of brain-machine interfaces (BMI). Higher decoding speed of neural signals is required for the large-scale neural data and the extremely low detection delay of closed-loop feedback experiment.<i>Approach.</i>To achieve higher neural decoding speed, we proposed a novel adaptive higher-order nonlinear point-process filter based on the variational Bayesian inference (VBI) framework, called the HON-VBI. This algorithm avoids the complex Monte Carlo random sampling in the traditional method. Using the VBI method, it can quickly implement inferences of state posterior distribution and the tuning parameters.<i>Main results.</i>Our result demonstrates the effectiveness and advantages of the HON-VBI by application for decoding the multichannel neural spike trains of the simulation data and real data. Compared with traditional methods, the HON-VBI greatly reduces the decoding time of large-scale neural spike trains. Through capturing the nonlinear evolution of system state and accurate estimating of time-varying tuning parameters, the decoding accuracy is improved.<i>Significance.</i>Our work can be applied to rapidly decode large-scale multichannel neural spike trains in BMIs.