Inferring Hand Motion from Multi-Cell Recordings in Motor Cortex using a Kalman Filter Wei Wu*, Michael Black †, Yun Gao*, Elie Bienenstock* §, Mijail Serruya §, Ammar Shaikhouni §, Carlos Vargas §†, John Donoghue § *Applied Mathematics † Computer Science § Neuroscience Brown University
Outline Introduction Kalman Filter Model and its Algorithm Experimental Result Analysis Optimal Time Lag Comparison with Linear Filter Conclusion
off-line data processing Goals neural signals neural reconstruction mathematical algorithm hand position Kalman Filter Firing rates observations inference/decoding
on-line direct neural control Goals neural reconstruction visual feedback Kalman Filter Firing rates observations inference/decoding
Related Work Georgopoulos et al. (1986) Taylor et al. (2002) Warland et al. (1997) Linear filter, ANN Wessberg et al.(2000) Linear filter, ANN Brown et al. (1998)Kalman filter Serruya et al.(2002)Linear filter Gao et al. (2002)Particle filter Population Vector
spike wave form Multi-electrode Array Implant for Spike Timing Recordings 1 ms 80µV Utah Array (4x4 mm) 100 electrodes, 400 m separation
Target Tracking Task Motions: fast, unconstrained Data (training 3.5 min, testing 1 min) : Position (Velocity, Acceleration) Firing rate (42 cells, non- overlapping 70ms bins)
has a sound probabilistic framework makes explicit assumptions about the data and noise indicates the uncertainty of the estimate requires a small amount of “training” data provides on-line estimation of hand position with short delay(within 200ms) has more accurate estimation than the standard linear filter does Mathematical Model
42 X 42 matrix 42 X 6 matrix system state vector firing rate vector 6 X 6 matrix Kalman Filter Model Measurement Equation: 6 X 6 matrix System Equation:
System Encoding by Training Data Centralize the training data, such that
Time Update Measurement Update Welch and Bishop 2002 Kalman Filter Algorithm Prior estimate Error covariance Posterior estimate Kalman gain Error covariance
Reconstruction on Test Data
Uniform: lag j time steps (1 time step = 70ms) Optimal Lag Non-uniform: lag time steps Changing it in two ways: Measurement Equation
Methods CC MSE ( x, y ) Kalman(0ms lag) (0.77, 0.91) 6.96 Kalman(70ms lag) (0.79, 0.93) 6.67 Kalman(140ms lag) (0.81, 0.93) 6.09 Kalman(210ms lag) (0.81, 0.89) 6.98 Kalman(280ms lag) (0.76, 0.82) 8.91 Kalman(non-uniform) (0.82, 0.93) 5.24 Optimal Lag on Test Data
Linear Filter hand position vector of firing rates for 42 cells over 20 bins (1.4sec) learned “filter” Simple regression model, fast decoding, reasonable reconstruction No explicitly probabilistic model, No uncertainty estimation, slow encoding constant offset
Linear Reconstruction Methods CC MSE ( x, y ) Kalman(140ms lag) (0.81, 0.93) 6.09 Linear filter (0.76, 0.92) 8.30
Conclusion Kalman Filter: has sound probabilistic framework, explicit assumptions, and uncertainty in estimation is more accurate than linear filter in estimation provides efficient filtering algorithm shows better reconstruction with time lag analysis
Future Work Exploring Poisson model for spiking activity instead of Gaussian Exploring the non-linear system model Further comparison with population vector methods (Taylor et al, 2002) and particle filtering techniques (Gao et al, 2002) on-line experiment of direct neural control using the Kalman filter
Thanks David Mumford Applied Mathematics Juliana Dushanova Neuroscience Lauren Lennox Neuroscience Matthew Fellows Neuroscience Liam Paninski NYU Neuroscience and Mathematics Nicholas Hatsopoulos U. Chicago Comp. Neuroscience Support: National Science Foundation Keck Foundation National Institutes of Health
Firing rate gives better estimation
Linear filters built on-line Mijail Serruya targetNeural control
(off-line) reconstruct monkey’s hand trajectory from its neural activity (on-line) control cursor movement from monkey’s neural activity (ultimate) provide control of prosthetic devices for severely disabled humans Goals