Gaussian Process Based Filtering for Neural Decoding Karthik Lakshmanan, Humphrey Hu, Arun Venkatraman April 24, 2013 University of Pittsburgh Sony Pictures http://cs.cmu.edu/~arunvenk/academics/neural/
Setup & Motivation
Proposed Method Model non-linear observation mapping with Gaussian Processes (GPs) Need to use Unscented Kalman Filter (UKF) However, this can be slow to evaluate…
Dimensionality Reduction
Trajectory Reconstruction Neural Reconstruction Results & Conclusion Improved decoding & produced a higher fidelity generative (observation) model Trajectory Reconstruction Final Cursor Position Neural Reconstruction % Improvement of GP-UKF over KF (both non-dim-reduced) 33.6% 42.8% 43.4% % Improvement of GP-UKF (w/PCA) over KF (non-dim-reduced) 22.2% 16.8% - % Improvement of GP-UKF (w/FA) over KF (non-dim-reduced) -0.80% -7.20% (trained on 1/5 of training data)