This project recapitulates Wu et al.'s work on neural decoding in a sequential reaching task using a switching Kalman filter [1] and a linear Kalman filter [2]. Decoding performance is characterized in terms of position prediction MSE. Performance improvement of the switching Kalman filter compared to the linear Kalman filter is comparable to that reported in the original paper. The dataset I use is prepared by Perich et al. [3], originally appearing in the paper by Lawlor et al. [4], and now published on CRCNS.org.
Run kalman_filter_demo script in MATLAB to check the result.
There is another nonlinear filter implemented in this repository, which uses a kinematic state model with a nonlinear transform function and a Poisson observation model. This filter is based on the work of Yu et al.[5]. It is NOT working currently, and is still under development.
For an overview of the project, see my slides; for more details, refer to my project report.