W. Wu1; M. Black2,4*; Y. Gao1; E. Bienenstock1,3,4; M. Serruya3; J. Donoghue3,4
1. Applied Mathematics, Brown University, Providence, RI, USA
2. Computer Science, Brown University, Providence, RI, USA
3. Neuroscience, Brown University, Providence, RI, USA
4. Brain Sciences, Brown University, Providence, RI, USA
We apply Kalman filtering, a recursive Bayesian estimation technique, to the problem of decoding neural activity in motor cortex. Our goal is to infer the position and velocity of a subject's hand from the spiking activity of 25 neurons recorded in the arm area of its primary motor cortex (MI). This simple Kalman filter has two parts. The first is a measurement model, which assumes that the firing rate of a cell within 50ms is a stochastic (Gaussian) linear function of hand position, velocity, and acceleration. This model is learned from training data, consisting of neural activity and movement recorded during a smooth manual tracking task. The second part is a system model, encoding regularities in hand motion; it too is Gaussian and linear. The model's ability to predict hand motion is tested off-line and compared to the more common linear-filtering technique. We show that Kalman-filter reconstruction is superior to linear-filter reconstruction. Analysis of the learned measurement model further suggests that, while the neural firing is closely related to hand position and velocity, acceleration is redundant. Finally, the Kalman filter framework allows us to compute the time lag between hand movement and firing rate that is optimal for trajectory reconstruction.
Supported by: NIH #R01 NS25074 & #N01-NS-9-2322, NSF ITR #0113679, DARPA MDA972-00-1-0026
Inferring hand motion from multi-cell recordings in motor cortex using a Kalman filter, Wu, W., Black, M. J., Gao, Y., Bienenstock, E., Serruya, M., and Donoghue, J. P.,
to appear: SAB'02-Workshop on Motor Control in Humans and Robots: On the Interplay of Real Brains and Artificial Devices, August 10, 2002, Edinburgh, Scotland (UK).
(abstract), (postscript), (pdf).