Brain machine interfaces incorporating neural population dynamics
A brain-machine interface is provided that incorporates a neural dynamical structure in the control of a prosthetic device to restore motor function and is able to significantly enhance the control performance compared to existing technologies. In one example, a neural dynamical state is inferred from neural observations, which are obtained from a neural implant. In another example, the neural dynamical state can be inferred from both the obtained neural observations and from the kinematics. A controller interfaced with the prosthetic device uses the inferred neural dynamical state as input to the controller to control kinematic variables of the prosthetic device.
1. A brain-machine interface for restoring motor function, comprising:
(a) a neural implant for obtaining neural observations y k , wherein the neural observations are defined by a spike frequency of one or more neurons;
(b) a computer-implemented inference system for inferring a neural dynamical state s k from the obtained neural observations (y k ), wherein the inferred neural dynamical state is a state of a dynamical system which is defined by:
s k+1 =f ( s k )+ g ( u k )+ n k
y k =h ( s k )+ l ( u k )+ r k
where f(s k ) is a function describing how the neural dynamical state evolves over time from s k to s k+1 ,
where h(s k ) is a function mapping the neural dynamical state s k to the neural observations y k ,
where (u k ) is an input to the dynamical system at time k,
where g(u k ) is a function mapping the input u k to the dynamical state S k+1 ,
where l(u k ) is a function mapping the input u k to the neural observations y k ,
where n k and r k are noise variables, and
where k denotes time;
(c) a prosthetic device; and
(d) a controller interfaced with the prosthetic device, wherein the inferred neural dynamical state (s k ) is input to the controller to control kinematic variables (x k ) of the prosthetic device.
2. The brain-machine interface as set forth in claim 1 , wherein the kinematic variables (x k ) include position and velocity of the prosthetic device.
3. The brain-machine interface as set forth in claim 1 , wherein the state of the dynamical system is updated by a state update matrix M, and an observation mapping P, so that:
s k+1 =Ms k +n k
y k =Ps k +r k
where n k has noise covariance N and r k has noise covariance R.
4. The brain-machine interface as set forth in claim 1 , wherein the neural dynamical state is inferred from the obtained neural observations (y k ) and from the kinematics (x k ) according to:
s k+1 =f ( s k )+ g ( u k )+ n k
a ( y k ,x k )= h ( s k )+ l ( u k )+ r k
where a(y k , x k ) is a function of the kinematics x k and the neural observations y k .
5. The brain-machine interface as set forth in claim 1 , wherein the inferred neural dynamical state (s k ) and the neural observations (y k ) are input to the controller to control the kinematic variables (x k ) of the prosthetic device.
6. A method of controlling a prosthetic device for restoring motor function, comprising:
(a) obtaining neural observations y k , wherein the neural observations are defined by a spike frequency of one or more neurons;
(b) inferring through a computer-implemented method a neural dynamical state s k from the obtained neural observations y k , wherein the inferred neural dynamical state is a state of a dynamical system which is defined by:
s k+1 =f ( s k )+ g ( u k )+ n k
y k =h ( s k )+ l ( u k )+ r k
where f(s k ) is a function mapping how the neural dynamical state evolves over time from s k to s k+1 ,
where h(s k ) is a function mapping the neural dynamical state s k to the neural observations y k ,
where (u k ) is an input to the dynamical system at time k,
where g(u k ) is a function mapping the input u k to the dynamical state s k+1 ,
where l(u k ) is a function mapping the input u k to the neural observations y k ,
where n k and r k are noise variables, and
where k denotes time; and
(c) controlling the prosthetic device, wherein the inferred neural dynamical state (s k ) is input to the controller to control kinematic variables (x k ) of the prosthetic device.
7. The method as set forth in claim 6 , wherein the kinematic variables (x k ) include position and velocity of the prosthetic device.
8. The method as set forth in claim 6 , wherein the state of the dynamical system is updated by a state update matrix M, and an observation mapping P, so that:
s k+1 =Ms k +n k
y k =Ps k +r k
where n k has noise covariance N and r k has noise covariance R.
9. The method as set forth in claim 6 , wherein the inferring comprises inferring the neural dynamical state (s k ) from the obtained neural observations y k and from the kinematics x k according to:
s k+1 =f ( s k )+ g ( u k )+ n k
a ( y k ,x k )= h ( s k )+ l ( u k )+ r k
where a(y k , x k ) is a function of the kinematics x k and the neural observations y k .
10. The method as set forth in claim 6 , wherein the inferred neural dynamical state (s k ) and the neural observations (y k ) are input to the controller to control the kinematic variables (x k ) of the prosthetic device.