Beruflich Dokumente
Kultur Dokumente
Ketan Savla
Munther A. Dahleh
I. I NTRODUCTION
Since the seminal work of Hodgkin and Huxley, there has
been a continued interest in the dynamical systems viewpoint
of a neuron. Hodgkin-Huxley (HH), Hindmarsh-Rose (HR),
and FitzHugh-Nagumo (FHN) models are among the most
successful dynamical models in computational neuroscience
for capturing neural firing behaviors. A detailed explanation
of these and several other models can be found in [6]. The
HH model consists of four differential equations with a high
number of coefficients. Although this model is capable of
generating all the behaviors of neuron spiking, it is a highly
nonlinear model. The HR model, on the other hand, consists
of three differential equations, which are highly coupled, and
it can exhibit all the firing modes obtained from the HH
model except for biophysically meaningful behaviors [5].
Finally, the FHN model consists of two differential equations,
and is simpler than the HH and HR models, though it is
unable to exhibit important firing behaviors such as bursting.
In fact, it has been noted [5] that without using a reset
or adding noise, the FHN model can not exhibit bursting.
The focus of this paper is on low complexity dynamical
models that can exhibit most of the neural firing activities of
interest that are possible by well-known dynamical systems
models (such as the HH model) but are nevertheless of low
complexity. Here, we use the term complexity to refer to
the presence of redundancies in the model on top of its
capability to capture neural firing behaviors and difficulty
in parameter estimation. With this in mind, we propose an
extension of the FHN model with time-varying parameters,
Rose T. Faghih, Ketan Savla and Munther A. Dahleh are with the
Laboratory for Information and Decision Systems, Massachusetts Institute
of Technology. {rfaghih,ksavla,dahleh}@mit.edu
Emery N. Brown is with the Department of Brain and Cognitive Sciences,
Massachusetts Institute of Technology, Massachusetts General Hospital and
Harvard Medical School. enbrown1@mit.edu
This work was supported in part by NIH DP1 OD003646 and NSF
0836720. Rose T. Faghihs work was also supported in part by MIT
Fellowship in Control and NSF Graduate Fellowship.
Emery N. Brown
Fig. 2.
Fig. 1.
Tonic Bursting
1
1
v0 = (b + 1 b2 b + 1), v2 = (b + 1 + b2 b + 1)
3
3
As observed in Fig. 3, w2 and w0 do not exactly correspond
to maximum and minimum values of f; however, we can
approximate w2 and w0 by plugging in the values of v2 and
v0 in f (v, b, I), as shown below:
2
2
1
1
2
w0
(b2 b + 1)3 + b3 b2 b +
+I
27
27
9
9
27
2
2
1
1
2
our method, a b estimate of 0.5079 (1.58% error) was
w2
(b2 b + 1)3 + b3 b2 b +
+I
27
27
9
9
27
obtained. Then, EKF was implemented on the data,
and the estimated value of b after one limit cycle was
Let h0 (b) = w0 I and h2 (b) = w2 I. From the time series
0.5001 (0.02% error). In this simulation, the sampling
data of membrane potentials, it is possible to obtain the
rate was very high and the sensor noise was very low,
maximum and minimum values of v, v1 and v3 , respectively.
putting EKF in an advantage. However, by increasing
As observed in Figure 3, f (v1 , b, I) w0 and f (v3 , b, I) w2 .
the sensor noise or decreasing the sampling rate, our
Hence, we obtain:
method performed better than EKF.
v1 (v1 1)(v1 b) h0 (b), v3 (v3 1)(v3 b) h2 (b)
2) By adding zero-mean sensor noise with s =0.01 to
the simulated action potentials, and using =105 , a b
Since there are two equations and one unknown, the system
estimate of 0.5253 (5.06% error) was obtained using
might not have a solution; in cases that the system has a
our method while b estimates for EKF did not converge
solution, the solution can be obtained using equation (1):
)
+
, we found the period
T
by looking at the time series plot of the b estimates, and then
using the trigonometric identity, we rewrote this problem
2
as sin( 2
T t)cos( )+cos( T t)sin( )+, and implemented
multiple regression to find the coefficients , , and . In
order to implement EKF, we used the actual initial condition
values as the initial guess, and an initial covariance estimate
of zero.
1) After adding a zero-mean sensor noise with s =0.001
to data with =105 , we implemented FSD and EKF.
The parameter estimates and the corresponding percent
error for each of the parameters is reported in Table I.
In reporting the EKF estimate, we ignored the estimates for which the error covariance matrix becomes
very high.
TABLE I
C OMPARISON OF PARAMETER ESTIMATES OBTAINED BY FSD AND EKF
FOR EXAMPLE 1 FOR TIME - VARYING b
=0.5
T =12
cos( )=1
=0.5
FSD estimate
0.4926
11.933
0.9804
0.4937
FSD error
1.48%
0.5583%
1.9558%
1.26%
EKFs estimate
0.5
12
1
0.5
EKFs error
0%
0%
0%
0%
FSD estimate
0.5708
11.933
0.9944
0.5602
FSD error
4.72%
0.5583%
0.56%
12.04%
EKFs estimate
0.5045
12.1461
1
0.5062
EKFs error
0.9%
1.2176%
0%
1.24%
TABLE III
C OMPARISON OF PARAMETER ESTIMATES OBTAINED BY FSD AND EKF
FOR EXAMPLE 3 FOR TIME - VARYING b
=0.5
T=12
cos( )=1
=0.5
FSD estimate
0.4638
12.029
0.9215
0.4996
FSD error
7.24%
0.2417%
7.8511%
0.08%
EKFs estimate
Diverged
Diverged
Diverged
Diverged
EKFs error
NA
NA
NA
NA