Lecture Notes in Computer Science
Download 12.42 Mb. Pdf ko'rish
|
3 − y − z + I ˙ y = (a + α)x 2 − y
˙z = μ(bx + c − z)
where x represents the membrane potential y and z are associated with fast and slow currents, respectively. I is an applied current, and a, α, μ, b and c are constant parameters. We rewrite the single HR neuron as a vectorized form: (S 0 ) : ˙ w = h(w) + Ξ(x, z)θ where w =
⎡ ⎣ x y z ⎤ ⎦ T , h(w) = ⎡ ⎣ −(x 3 + y + z)
−y 0 ⎤ ⎦, Ξ(x, z) = ⎡ ⎣ x 2 1 0 0 0 0 0 0 x 2 0 0 0 0 0 0 x 1 −z ⎤ ⎦, θ = θ
1 , θ
2 , θ
3 , θ
4 , θ
5 , θ
6 T = a, I, a + α, μb, μc, μ T . 3.2 Numerical Examples The HR model shows a large variety of behaviors with respect to the parameter values in the differential equations[12]. Thus, we can characterize the dynamic behaviors with respect to different values of the parameters. We focus on the parameters a and I. The parameter a is an internal parameter in the single neuron and I is an external depolarizing current. For the fixed I = 0.05, the HR model shows a tonic bursting with a ∈ [1.8, 2.85] and a tonic spiking with a ≥ 2.9. On the other hand, for the fixed a = 2.8, the HR model shows a tonic bursting with I ∈ [0, 0.18] and a tonic spiking with a ∈ [0.2, 5].
86 K. Mitsunaga, Y. Totoki, and T. Matsuo 0 200
400 600
800 1000
−1.5 −1 −0.5 0 0.5
1 1.5
2 time
x Fig. 1. The response of x in the tonic bursting −2 −1 0 1 2 0 2 4 6 8 −1 −0.9 −0.8
−0.7 −0.6
−0.5 −0.4
−0.3 x y z Fig. 2. 3-D surface of x, y, z in the tonic bursting 0 200 400 600
800 1000
−1.5 −1 −0.5 0 0.5
1 1.5
2 time
x Fig. 3. The response of x in the tonic spiking −2
0 1 2 0 2 4 6 8 −1 −0.9 −0.8
−0.7 −0.6
−0.5 x y z Fig. 4. 3-D surface of x, y, z in the tonic spiking 0 200 400 600
800 1000
−1 −0.5
0 0.5
1 1.5
2 time
x Fig. 5. The response of x 1 in the in- trinsic bursting neuron −1 0 1 2 0 2 4 6 8 −1 −0.8 −0.6 −0.4
−0.2 x y z Fig. 6. 3-D surface of x, y, z in the in- trinsic bursting neuron The parameters of the HR model in the tonic bursting (TB) are given by a = 2.8, α = 1.6, c = 5, b = 9, μ = 0.001, I = 0.05. Figure 1 shows the response of x. Figure 2 shows the 3 dimensional surface of x, y, z. We call this neuron the intrinsic bursting neuron(IBN).
Firing Pattern Estimation of Biological Neuron Models 87 The parameters of the HR model in the tonic spiking (TS) are given by a = 3.0, α = 1.6, c = 5, b = 9, μ = 0.001, I = 0.05. The difference between the tonic bursting and the tonic spiking is only the value of the parameter a. Figure 3 shows the responses of x. Figure 4 shows the 3 dimensional surface of x, y, z. We also call this neuron the intrinsic spiking neuron(ISN). When the external current change from I = 0.05 to I = 0.2, the IBN shows a tonic spiking. Figure 5 shows the responses of x. Figure 6 shows the 3 dimensional surface of x, y, z. 4 Synaptically Coupled Model of HR Neuron 4.1 Dynamical Equations Consider the following synaptically coupled two HR neurons[1]: ˙ x 1 = a
1 x 2 1 − x
3 1 − y 1 − z
1 − g
s (x 1 − V s1 )Γ (x 2 ) ˙ y 1 = (a 1 + α
1 )x 2 1 − y
1 , ˙ z 1 = μ 1 (b 1 x 1 + c 1 − z
1 ) ˙ x 2 = a 2 x 2 2 − x
3 2 − y 2 − z
2 − g
s (x 2 − V s2 )Γ (x 1 ) ˙ y 2 = (a 2 + α
2 )x 2 2 − y
2 , ˙ z 2 = μ 2 (b 2 x 2 + c 2 − z
2 ) where Γ (x) is the sigmoid function given by Γ (x) = 1 1 + exp( −λ(x − θ s )) . 4.2
Numerical Examples Consider the IBN neuron with a = 2.8 and the ISN neuron with a = 10.8 whose other parameters are as follows: α i = 1.6, c i = 5, b i = 9, μ
i = 0.001, V si = 2, θ
s = −0.25, λ = 10. Figures 7,8 show the responses of the membrane potentials in the coupling of IBN neuron and ISN neuron with the coupling strength g s = 0.05, respectively. Each neuron behaves as an intrinsic single neuron. As increasing the coupling strength, however, the IBN neuron shows a chaotic behavior. Figures 9,10 show the responses of the membrane potentials in the coupling of IBN neuron and ISN neuron with the coupling strength g s = 1, respectively. Figure 11 shows the response of the membrane potentials in the coupling of two same IBNs with 88 K. Mitsunaga, Y. Totoki, and T. Matsuo 0 200
400 600
800 1000
−1.5 −1 −0.5 0 0.5
1 1.5
2 time
x Fig. 7. The response of x 1 of the IBN 0 200
400 600
800 1000
−4 −2 0 2 4 6 8 10 12 time x Fig. 8. The response of x 2 of the ISN 0 200
400 600
800 1000
−1.5 −1 −0.5 0 0.5
1 1.5
2 time
x Fig. 9. The response of x 1 of the IBN 0 200
400 600
800 1000
−4 −2 0 2 4 6 8 10 12 time x Fig. 10. The response of x 2 of the ISN 0 200
400 600
800 1000
−1.5 −1 −0.5 0 0.5
1 1.5
2 time
x Fig. 11. The response of x 1 of the IBN- IBN coupling with g s = 0.05 0 200
400 600
800 1000
−1 −0.5
0 0.5
1 1.5
2 time
x Fig. 12. The response of x 1 of the IBN- IBN coupling with g s = 1 the coupling strength g s = 0.05. In this case, two IBNs synchronize as bursting neurons. Figure 12 shows the response of the membrane potentials in the coupling of two same IBNs with the coupling strength g s = 1. Two IBNs synchronize as spiking neurons. Firing Pattern Estimation of Biological Neuron Models 89 5 Adaptive Observer with Full States We present the parameter estimation problem to distinguish the firing patterns by using early-time dynamic behaviors. In this section, assuming that the full states are measurable, we present an adaptive observer to estimate all parameters in the single HR neuron. 5.1
Construction of Adaptive Observer We present an adaptive observer as (O 0
w = W ( ˆ w − w) + h(w) + Ξˆθ where ˆ w = [ˆx,
ˆ y, ˆ z] is an estimate of the states, ˆθ is an estimate of the unknown parameters and W is selected as a stable matrix. Using the standard adaptive control theory[11], the parameter update law is given by ˙ˆθ = Γ Ξ T P (w − ˆ
w). where P is a positive definite solution of the following Lyapunov equation for a positive definite matrix Q: W T P + P W = −Q. 5.2
Numerical Examples We will show the simulation results of single IBN case. The parameters in the tonic bursting are given by a = 2.8, α = 1.6, c = 5, b = 9, μ = 0.001, I = 0.05. The parameters of the adaptive observers are selected as W = −10I
3 , Γ = diag{100, 50, 300}. Figure 13 shows the estimation behavior of a (solid line) and I (dotted line). The estimates ˆ a and ˆ I converge to the true values of a and I. 6 Adaptive Observer with a Partial State We assume that the membrane potential x is available, but the others are im- measurable. In this case, we consider following problems: – Estimate y and z using the available signal x; – Estimate the parameter a or I to distinguish the firing patterns by using early-time dynamic behaviors.
90 K. Mitsunaga, Y. Totoki, and T. Matsuo 6.1 Construction of Adaptive Observer The parameters a and I are key parameters that determine the firing pattern. The HR model can be rewritten by the following three forms[9]: (S 1
w = Aw + h 1 (x) + b 1 (x 2 a) (1)
(S 2 ) : ˙ w = Aw + h 2 (x) + b 2 (I)
(2) (S 3 ) : ˙ w = Aw + h 3 (x) + b
2 (θ T ξ) (3)
where A =
⎡ ⎣ 0 −1 −1 0 −1 0 μb 0 −μ ⎤ ⎦ , h 1 = ⎡ ⎣ −x 3 + I αx 2 μc ⎤ ⎦ , b 1 = ⎡ ⎣ 1 1 0 ⎤ ⎦ , b 2 = ⎡ ⎣ 1 0 0 ⎤ ⎦ , h 2 = ⎡ ⎣ −x 3 + ax
2 (a + α)x
2 μc ⎤ ⎦ , h 3 = ⎡ ⎣ −x 3 δx 2 μc ⎤ ⎦ , θ = a I , ξ =
x 2 1 . In (S
1 ) and (S
2 ), the unknown parameters are assumed to be a and I, respec- tively. In (S 3 ), we assume that the parameter δ = a + α is known, and a and I are unknown. Since the measurable signal is x, the output equation is given by x = cw = 1 0 0 w We present adaptive observers that estimate the parameter for each system (S i
(O 1 ) : ˙ˆ w 1 = A ˆ w 1 + h 1 (x) + b
1 (x 2 ˆ a) + g(x − ˆx) (4) (O
) : ˙ˆ w 2 = A ˆ w 2 + h 2 (x) + b 2 ( ˆ
I) + g(x − ˆx) (5)
(O 3 ) : ˙ˆ w 2 = A ˆ w 2 + h 3 (x) + b
2 (ˆ θ T ξ) + g(x − ˆx) (6) where g is selected such that A − gc is a stable. Since (A, b 1 , c) and (A, b 2 , c)
are strictly positive real, the parameter estimation laws are given as ˙ˆa = γ
1 x 2 (x − ˆx),
˙ˆ I = γ
2 (x − ˆx). (7) Using the Kalman-Yakubovich (KY) lemma, we can show the asymptotic sta- bility of the error system based on the standard adaptive control theory[11]. 6.2
Numerical Examples We will show the simulation results of single IBN case. The parameters in the tonic spiking are same as in the previous simulation. Figures 14 and 15 show the estimated parameters by the adaptive observers (O 1 ) and (O
2 ), respectively. Figures 16 and 17 show the responses of y (solid line) and its estimate ˆ y (dot-
ted line) the adaptive observers (O 1 ) for t ≤ 500 and for t ≤ 20, respectively. Figure 18 shows the responses of z (solid line) and its estimate ˆ z (dotted line) by the adaptive observers (O 1 ). The simulation results of other cases are omitted. The states and parameters can be asymptotically estimated. Firing Pattern Estimation of Biological Neuron Models 91 0 50 100
150 200
0 0.5
1 1.5
2 2.5
3 3.5
4 4.5
5 time
ˆ a ˆ I ˆ a ˆ I Fig. 13. ˆa(solid line) and ˆ I(dotted line) in the adaptive observer (O 0 ) with full states 0 20 40 60 80 100 1 1.5
2 2.5
3 3.5
4 4.5
5 time
ˆ a Fig. 14. ˆa(solid line) in the adaptive ob- server (O 1 ) with x 0 20 40 60 80 100 −0.5 −0.4
−0.3 −0.2
−0.1 0 0.1 time ˆ I Fig. 15. ˆ I(solid line) in the adaptive ob- server (O 2 ) with x 0 100
200 300
400 500
0 1 2 3 4 5 6 7 time y ˆ y y ˆ y Fig. 16. y(solid line) and ˆ y(dottedline) in the adaptive observer (O 1 ) (t ≤ 500) 0 5 10 15 20 0 1 2 3 4 5 6 7 time y ˆ y y ˆ y Fig. 17. y(solid line) and ˆ y(dottedline) in the adaptive observer (O 1 ) (t ≤ 20) 0 5 10 15 20 −1.5 −1 −0.5 0 time
z ˆ z z ˆ z Fig. 18. z(solid line) and ˆz(dottedline) in the adaptive observer (O 1 ) (t
≤ 20) 92 K. Mitsunaga, Y. Totoki, and T. Matsuo 7 Conclusion We presented estimators of the parameters of the HR model using the adaptive observer technique with the output measurement data such as the membrane potential. The proposed observers allow us to distinguish the firing pattern in early time and to recover the immeasurable internal states. References 1. Belykh, I., de Lange, E., Hasler, M.: Synchronization of Bursting Neurons: What Matters in the Network Topology. Phys. Rev. Lett. 94, 101–188 (2005) 2. Izhikevich, E.M.: Simple Model of Spiking Neurons. IEEE Trans on Neural Net- works 14(6), 1569–1572 (2003) 3. Izhikevich, E.M.: Which model to use for cortical spiking neurons? IEEE Trans on Neural Networks 15(5), 1063–1070 (2004) 4. Watts, L.: A Tour of NeuraLOG and Spike - Tools for Simulating Networks of Spiking Neurons (1993), http://www.lloydwatts.com/SpikeBrochure.pdf 5. Hindmarsh, J.L., Rose, R.M.: A model of the nerve impulse using two first order differential equations. Nature 296, 162–164 (1982) 6. Hindmarsh, J.L., Rose, R.M.: A model of neuronal bursting using three coupled first order differential equations. Proc. R. Soc. Lond. B. 221, 87–102 (1984) 7. Carroll, T.L.: Chaotic systems that are robust to added noise, CHAOS, 15, 013901 (2005) 8. Meunier, N., Narion-Poll, R., Lansky, P., Rospars, J.O.: Estimation of the Individ- ual Firing Frequencies of Two Neurons Recorded with a Single Electrode. Chem. Senses 28, 671–679 (2003) 9. Yu, H., Liu, Y.: Chaotic synchronization based on stability criterion of linear sys- tems. Physics Letters A 314, 292–298 (2003) 10. Marino, R.: Adaptive Observers for Single Output Nonlinear Systems. IEEE Trans. on Automatic Control 35(9), 1054–1058 (1990) 11. Narendra, K.S., Annaswamy, A.M.: Stable Adaptive Systems. Prentice Hall Inc., Englewood Cliffs (1989) 12. Arena, P., Fortuna, L., Frasca, M., Rosa, M.L.: Locally active Hindmarsh-Rose neurons. Chaos, Soliton and Fractals 27, 405–412 (2006) 13. Tokuda, I., Parlitz, U., Illing, L., Kennel, M., Abarbanel, H.: Parameter estimation for neuron models. In: Proc. of the 7th Experimental Chaos Conference (2002), http://www.physik3.gwdg.de/ ∼ ulli/pdf/TPIKA02 pre.pdf 14. Steur, E.: Parameter Estimation in Hindmarsh-Rose Neurons (2006), http://alexandria.tue.nl/repository/books/626834.pdf 15. Fujikawa, H., Mitsunaga, K., Suemitsu, H., Matsuo, T.: Parameter Estimation of Biological Neuron Models with Bursting and Spiking. In: Proc. of SICE-ICASE International Joint Conference 2006 CD-ROM, pp. 4487–4492 (2006)
Thouless-Anderson-Palmer Equation for Associative Memory Neural Network Models with Fluctuating Couplings Akihisa Ichiki and Masatoshi Shiino Department of Applied Physics, Faculty of Science, Tokyo Institute of Technology, 2-12-2 Ohokayama Meguro-ku Tokyo, Japan Abstract. We derive Thouless-Anderson-Palmer (TAP) equations and order parameter equations for stochastic analog neural network models with fluctuating synaptic couplings. Such systems with finite number of neurons originally have no energy concept. Thus they defy the use of the replica method or the cavity method, which require the energy concept. However for some realizations of synaptic noise, the systems have the effective Hamiltonian and the cavity method becomes applicable to derive the TAP equations. 1 Introduction The replica method [1] for random spin systems has been successfully employed in neural network models of associative memory to have the order parameters and the storage capacity [2] and the cavity method [3] has been employed to de- rive the Thouless-Anderson-Palmer (TAP) equations [4,5]. However these tech- niques require the energy concept. On the other hand, various types of neural network models which have no energy concept, such as networks with tempo- rally fluctuating synaptic couplings, may exist. The alternative approach to the replica method to derive the order parameter equations, called the self-consistent signal-to-noise analysis (SCSNA), is closely related to the cavity concept in the case where networks have free energy [6,7]. An advantage to apply the SCSNA to neural networks is that the energy concept is not required to derive the order parameter equations once the TAP equations are obtained. The SCSNA, which was originally proposed for deriving a set of order parameter equations for deter- ministic analog neural networks, becomes applicable to stochastic networks by noting that the TAP equations define the deterministic networks. Furthermore, the coefficients of the Onsager reaction terms characteristic to the TAP equations which determine the form of the transfer functions in analog networks are self- consistently obtained through the concept shared by the cavity method and the SCSNA. Thus the TAP equations as well as the order parameter equations are derived self-consistently by the hybrid use of the cavity method and the SCSNA in the case where the energy concept exists. However the networks with synaptic noise, which have no energy concept, defy the use of the cavity method to have the TAP equations. On the other hand, as in [8], the network with a specific Download 12.42 Mb. Do'stlaringiz bilan baham: |
ma'muriyatiga murojaat qiling