Biomedical Engineering Reference
In-Depth Information
Neuronal
Behavior
Plant to be identified
activity
Compare
Adaptive system
Kinematic error
FIgURE 3.1:
System identification framework for BMIs.
2.
The system is MIMO, with a very large input (hundred of neurons), and the data are lim-
ited; therefore, there is a need for proper regularization of the solution.
Although the input dimensionality is high, it still coarsely samples in unknown ways the
large pool of neurons involved in the motor task (missing data).
Not all the channels are relevant for the task, and because of the spike sorting and other
difficulties, the data are noisy.
Because of the way that neurons are involved in behavior, the data are nonstationary.
3.
4.
5.
With these remarks in mind, one has to choose the class of functions and model topologies that
best match the data while being sufficiently powerful to create a mapping from neuronal activity
to a variety of behaviors. There are several candidate models available, and based on the amount of
neurophysiological information that is utilized about the system, an appropriate modeling approach
shall be chosen. Three types of I/O models based on the amount of prior knowledge exist in the
literature [
3
]:
•
“White box”: The physical system principles of operation are perfectly known, except for
the parameters that are learned from the data.
“Gray box”: Some physical insight is available but model is not totally specified, and other
parameters need to be determined from the data.
“Black box”: Little or no physical insight is available, so the chosen model is picked on
other grounds (robustness, easy implementation, etc).
•
•
The choice of white, gray, or black box is dependent upon our ability to access and measure signals
at various levels of the motor system, the type of questions one is interested to answer, as well as the
computational cost of implementing the model in current computing hardware.
