Neuronal Importance Ascertaining And Efficient Spiking Activity Decoding For Motor Brain Machine Interfaces

Brain machine interfaces (BMIs) build a direct communication pathway between the brain and an external device without the involvement of the spinal or the peripheral nervous system. Invasive BMIs record the activity of single neuron by implanting a micro-electrode array into the brain cortex. The signal carries rich information about the movement and can be decoded to control external devices. However, the neural signal is generally high dimensional and redundant as well as highly nonlinear and nonstationary. In this study, we analyze the recorded signals and focus on the accurate ascertaining of neural importance and the efficient decoding of spiking activity.We propose to use a local-learning-based method to decompose the nonlinear neural signal into a set of linear components, and estimate the neural importance without any encoding/decoding model. Compared with other methods, the experimental results show that the local-learning-based method can identify the task-irrelevant neurons more accurately. In addition, the important neurons can provide a comparable decoding accuracy compared to full ensemble. We also develop better decoding algorithms based on the physiological characteristic of neurons in the framework of Se-quential Monte Carlo Point Process estimation(SMCPP). The spike trains are described as point processes and the distributions of movement are no longer restricted to be Gaussian. Based on the spatio-temporal correlations in neuronal population, we design a better tuning function in SMCPP to improve the decoding performance. Since the tuning property is nonstationary, we combine the change-point detection algorithm and the static parameter estimation algorithm with SMCPP to form a dynamic decoding algorithm which can update the parameters when necessary. Though the SMCPP algorithm can interpret the neural activity accurately, it is not widely adopted in online applications due to the high computational complexity. In this paper, we implement a massively parallel version of the algorithm on GPU to speed up the decoding with the help of hundreds of cores in GPU.The innovations of this study are:1)using the local-learning-based method to estimate the neural importance which is independent on any encoding/decoding models, and only 10 out of over 70 neurons can achieve over 95% of the full recording’s decoding performance; 2)designing a better tuning function based on the spatio-temporal correlations in neuronal population which significantly improves the decoding accuracy, especially the decoding error on position decreases 23%; combining the change-point detection algorithm and the static parameter estimation algo-rithm with the decoding algorithm, which reduces the prediction error in simulation and pilot study on real data; 3)implementing the decoding algorithm in a massively parallel way on GPU, which accelerates the decoding speed by 10 times. The work lays a solid foundation for the development of long-term practical BMI systems.