We consider the nonlinear dynamical behavior of a three-dimensional recurrent neural network with time delay. By choosing the time delay as a bifurcation parameter, we prove that Hopf bifurcation occurs when the delay passes through a sequence of critical values. Applying the nor- mal form method and center manifold theory, we obtain some local bifurcation results and derive formulas for determining the bifurcation direction and the stability of the bifurcated periodic solution. Some numerical examples are also presented to verify the theoretical analysis.

1. Introduction

Starting with the work of Hopfield [1] on neural networks, recurrent neural networks including Hopfield neural networks, Cohen-Grossberg neural networks, and cellular neural networks have been used extensively in different areas such as signal processing, pattern recognition, optimization, and associative memories. Many researchers studied the dynamical behavior of Recurrent neural network systems, and most of papers are devoted to the stability of equilibrium, existence and stability of periodic solutions, bifurcation, and chaos [2–5]. In [5], Ruiz et al. considered a particular configuration of a recurrent neural network, illustrated in Figure 1. In Figure 1, u(t) is the input and y(t) is the output of the network. This recurrent neural network can be described by the following system:

()

Here, x(t) ∈ Rⁿ is the state, w_i ∈ R, i = 1, …, n − 1 are the network parameters or weights, u(t) is a smooth input, and y(t) is the output. The transfer function of the neurons is taken as f(·) = tanh(·). A three-node network of the form (1.1) in feedback configuration, with u(t) = y(t), has been studied in [5]; that is,

()

Ruiz et al. found and analyzed the Hopf bifurcation behavior in system (1.2). In [4], Maleki et al. considered system (1.1) with a transfer function

, where α_2i−1 > 0 for i odd and α_2i−1 < 0 for i even. The authors analyzed the Bogdanov-Takens bifurcation in the system.

Details are in the caption following the image — **Figure 1**
Open in figure viewer PowerPoint

Class of recurrent neural networks.

It is well known that there exist time delays in the information processing of neurons. The delayed axonal signal transmissions in the neural network models make the dynamical behaviors become more complicated and may destabilize the stable equilibria and admit periodic oscillation, bifurcation, and chaos. Therefore, the delay is an important control parameter in living nervous system: different ranges of delays correspond to different patterns of neural activities (see, e.g., [6–11]).

In the present paper, we consider the following three-dimensional recurrent neural network model with time delay

()

By choosing the time delay as a bifurcation parameter, we prove that Hopf bifurcation occurs in the neuron and study the properties of periodic solutions of this model.

The organization of this paper is as follows. In Section 2, by analyzing the characteristic equation of the linearized system of system (1.3) at the equilibrium, we discuss the stability of the equilibrium and the existence of the Hopf bifurcation occurring at the equilibrium. In Section 3, the formulae determining the direction of the Hopf bifurcations and the stability of bifurcating periodic solutions on the center manifold are obtained by using the normal form theory and the center manifold theorem due to Hassard et al. [12]. We do some computer observations to validate our theoretical results in Section 4.

2. Stability and Existence of Hopf Bifurcation

For most of the models in the literature, including the ones in [5, 7, 8], the activation function f is f(u) = tanh(cu). However, we only make the following assumption on function f:

(H)
f ∈ C³(R), f(0) = 0, and f′(0) ≠ 0.

Clearly, (x₁, x₂, x₃) ^T = (0,0, 0) ^T is equilibrium of system (1.3). Linearization of (1.3) at the zero equilibrium yields

()

whose characteristic equation is

()

that is,

()

The zero equilibrium is stable if all roots of (2.3) have negative real parts and unstable if at least one root has positive real part. Therefore, in order to study the local stability of the zero equilibrium of system (2.3), we need to investigate the distribution of the roots of (2.3).

When τ = 0, characteristic equation (2.3) yields

()

Let y = λ + 1, (2.4) reduces to

()

Denote

()

From Cardano formula for the third-degree algebra equation, we have the following lemma.

Lemma 2.1. (1) If Δ > 0, then (2.5) has a real root α + β and a pair of conjugate complex roots . Furthermore, the roots of (2.4) are given by λ₁ = −1 + α + β and .

(2) If Δ = 0, then (2.5) has a simple root 2α and a multiple root −α with the multiplicity of 2. Furthermore, the roots of (2.4) are given by λ₁ = −1 + 2α and λ_2,3 = −1 − α. Meanwhile, if w₁ = w₂ = 0, that is, α = 0, then (2.4) has a multiple root −1 with the multiplicity of 3.

(3) If Δ < 0, then (2.5) has three real roots 2 Re{α}, 2 Re{αε}, and . Furthermore, the roots of (2.4) are given by λ₁ = −1 + 2 Re{α}, λ₂ = −1 + 2 Re{αε}, and .

Applying Lemma 2.1, we have the following lemma.

Lemma 2.2. All roots of (2.4) have negative real parts if one of the following holds.

(1) Δ > 0 and −2 < α + β < 1.

(2) Δ = 0 and −1 < α < 1/2.

(3) Δ < 0 and .

Let z = (λ + 1)e^λτ, then (2.3) becomes

()

We notice that (2.7) and (2.5) have the same coefficients.

Denote the three roots of (2.7) by z_n = R_n + iI_n (n = 1,2, 3). Hence, (2.3) is equivalent to

()

Clearly, iω (ω > 0) is a root of (2.3) if and only if ω satisfies

()

which implies that

()

This yields

()

Obviously, we have the following lemma.

Lemma 2.3. Equation (2.11) is meaningless when . If (2.7) has a root, denoted by z_j, satisfying |z_n | > 1, then (2.11) has a positive root given by

()

Summarizing the discussion above, we have the following.

Lemma 2.4. If Δ ≥ 0, then (2.11) has at most two positive roots. If Δ < 0, then (2.11) has at most three positive roots.

Without loss of generality, one assumes that (2.11) has three positive roots ω_n (n = 1,2, 3).

Define

()

Then, ±iω_n is a pair of purely imaginary roots of (2.3) with

Let λ(τ) = σ(τ) + iω(τ) be the root of (2.3) near

satisfying

()

We have the following.

Lemma 2.5. One has

()

Proof. Differentiating both sides of (2.8) with respect to τ, we have

()

Note that

, therefore

()

We have

()

This completes the proof.

For convenience, we let

, such that

()

where

()

and

is defined in (2.13).

Applying Lemmas 2.1–2.5 and Corollary 2.4 of Ruan and Wei [13], we have the following results.

Lemma 2.6. All roots of (2.3) have negative real parts if one of the following holds:

(H₁) Δ > 0, −1 < α + β < 1 and (1/4)(α+β)² + (3/4)(α−β)² < 1,
(H₂) Δ = 0 and −1 < α < 1/2,
(H₃) Δ < 0 and .

Lemma 2.7. Suppose that one of the following hypothesis is satisfied:

(H₄) Δ > 0, −2 < α + β < −1,
(H₅) Δ > 0, −1 < α + β < 1 and (1/4)(α+β)² + (3/4)(α−β)² > 1.

Then, there exists a sequence values of τ defined by (2.19) such that all roots of (2.3) have negative real parts for all τ ∈ [0, τ₀), and (2.3) has at least one root with positive real part when τ > τ₀, and (2.3) exactly has a pair of purely imaginary roots ±iω_n (n = 1,2, 3) when

, where ω_n and

are defined by (2.12) and (2.13), respectively.

From Lemmas 2.5–2.7 and the Hopf bifurcation theorem for functional differential equations in [14], we have the theorem.

Theorem 2.8. (1) If one of the hypothesis (H₁), (H₂), (H₃) is satisfied, then the zero solution of system (1.3) is asymptotically stable for all τ ≥ 0.

(2) If (H₄) or (H₅) is satisfied, then the zero solution of system (1.3) is asymptotically stable for τ ∈ [0, τ₀) and unstable for τ > τ₀, and system (1.3) undergoes a Hopf bifurcation at the origin when τ = τ_j (j = 0,1, 2,3, …).

3. Direction of Hopf Bifurcations and Stability of the Bifurcating Periodic Orbits

In this section, we will study the direction of the Hopf bifurcation and stability of bifurcating periodic solutions by using the normal theory and the center manifold theorem due to Hassard et al. [12].

Let u₁(t) = x₁(τt), u₂(t) = x₂(τt), u₃(t) = x₃(τt), then system (1.3) becomes functional differential equation in 𝒞 = 𝒞([−1,0], ℝ³) as

()

where

()

Setting τ = ν + τ_j, we know that ν = 0 is Hopf bifurcation value of system (3.1).

For

, let

()

By the Riesz representation theorem, there exists a function η(θ, ν) of bounded variation for θ ∈ [−1,0], such that

()

In fact, we can choose

()

where δ denotes the Dirac delta function. For ϕ ∈ 𝒞([−1,0], ℝ³), define

()

where

()

Then system (3.1) is equivalent to

()

where

, u_t = u(t + θ) for θ ∈ [−1,0].

For ψ ∈ 𝒞¹([0,1], (ℝ³) ^*), define

()

and a bilinear inner product

()

where η(θ) = η(θ, 0). Then, A(0) and A^* are adjoint operators. By the discussion in Section 2, we know that ±iω₀τ_j are eigenvalues of A(0) and A^* corresponding to iω₀τ_j and −iω₀τ_j, respectively.

Suppose

is the eigenvectors of A(0) corresponding to iω₀τ_j, then A(0)q(θ) = iω₀τ_jq(θ). Then from the definition of A(0) and (3.3)–(3.5), we have

()

For

, then we obtain

()

Similarly, we can obtain the eigenvector

of A^* corresponding to −iω₀τ_j, where

()

In order to assure 〈q^*(s), q(θ)〉 = 1, we need to determine the value of D. By (3.10), we have

()

Therefore, we can choose D as

()

Following the algorithms given in [12] and using similar computation process in [7], we can get that the coefficients which will be used to determine the important quantities:

()

where

()

moreover E₁, E₂ satisfy the following equations, respectively,

()

Therefore, all g_ij in (3.16) can be expressed in terms of parameters. And we can compute the following values:

()

which determine the qualities of bifurcating periodic solution in the center manifold at the critical values τ_j, that is, μ₂ determines the directions of the Hopf bifurcation: if μ₂ > 0(μ₂ < 0), then the Hopf bifurcation is supercritical (subcritical) and the bifurcating periodic solutions exist for τ > τ_j(τ < τ_j); β₂ determines the stability of the bifurcating periodic solutions: the bifurcating periodic solutions are stable (unstable) if β₂ < 0(β₂ > 0); T₂ determines the period of the bifurcating periodic solutions: the period increases (decreases) if T₂ > 0(T₂ < 0).

4. Computer Simulation

In this section, we will confirm our theoretical analysis by numerical simulation. We give an example of system (3.1) with w₁ = 1, w₂ = −1, and f(·) = tanh(·). Then, f(0) = 0 and f′(0) = 1.

Equation (1.3) becomes

()

From (2.6), we have Δ = 0.2870, α = 1.0118, β = −0.3295. By Lemma 2.1, we know (2.7) has roots z₁ = α + β = 0.6823 and

. Clearly, |z₁ | = 0.6823 < 1, |z_2,3 | = 1.4655 > 1. From (2.12), it follows that

, and, from (2.13), we get τ₀ = 1.8434. Thus, the zero equilibrium is asymptotically stable when τ ∈ [0, τ₀) as is illustrated by the computer simulations (see Figures 2(a)–2(d)). When τ passes through the critical value τ₀, zero equilibrium loses its stability and a Hopf bifurcation occurs, that is, a family of periodic solutions bifurcates from the origins (0,0, 0), which are depicted in Figures 3(a)–3(d).

Acknowledgment

The author is very grateful to the referees for their valuable suggestions, which help to improve the paper significantly.

References

1 Hopfield J. J., Neurons with graded response have collective computational properties like those of two-state neurons, Proceedings of the National Academy of Sciences of the United States of America. (1984) 81, no. 10, 3088–3092.
10.1073/pnas.81.10.3088
CAS PubMed Web of Science® Google Scholar
2 Gao B. and Zhang W., Equilibria and their bifurcations in a recurrent neural network involving iterates of a transcendental function, IEEE Transactions on Neural Networks. (2008) 19, no. 5, 782–794, https://doi.org/10.1109/TNN.2007.912321.
10.1109/TNN.2007.912321
PubMed Google Scholar
3 Haschke R. and Steil J. J., Input space bifurcation manifolds of recurrent neural networks, Neurocomputing. (2005) 64, no. 1–4, 25–38, https://doi.org/10.1016/j.neucom.2004.11.030.
10.1016/j.neucom.2004.11.030
Web of Science® Google Scholar
4 Maleki F., Beheshti B., Hajihosseini A., and Lamooki G. R.R., The Bogdanov-Takens bifurcation analysis on a three dimensional recurrent neural network, Neurocomputing. (2010) 73, no. 16–18, 3066–3078, https://doi.org/10.1016/j.neucom.2010.06.023.
10.1016/j.neucom.2010.06.023
Web of Science® Google Scholar
5 Ruiz A. C., Owens D. H., and Townley S., Existence, learning, and replication of periodic motions in recurrent neural networks, IEEE Transactions on Neural Networks. (1998) 9, no. 4, 651–661.
10.1109/72.701178
CAS PubMed Web of Science® Google Scholar
6 Liao X., Wong K. W., Leung C. S., and Wu Z., Hopf bifurcation and chaos in a single delayed neuron equation with non-monotonic activation function, Chaos, Solitons and Fractals. (2001) 12, no. 8, 1535–1547, https://doi.org/10.1016/S0960%2D0779(00)00132%2D6, 1839914, ZBL1012.92005.
10.1016/S0960-0779(00)00132-6
Web of Science® Google Scholar
7 Wei J. and Ruan S., Stability and bifurcation in a neural network model with two delays, Physica D. Nonlinear Phenomena. (1999) 130, no. 3-4, 255–272, https://doi.org/10.1016/S0167%2D2789(99)00009%2D3, 1692866, ZBL1066.34511.
10.1016/S0167-2789(99)00009-3
Web of Science® Google Scholar
8 Gopalsamy K. and Leung I., Delay induced periodicity in a neural netlet of excitation and inhibition, Physica D. Nonlinear Phenomena. (1996) 89, no. 3-4, 395–426, https://doi.org/10.1016/0167%2D2789(95)00203%2D0, 1369242, ZBL0883.68108.
10.1016/0167-2789(95)00203-0
Web of Science® Google Scholar
9 Huang C., He Y., Huang L., and Zhaohui Y., Hopf bifurcation analysis of two neurons with three delays, Nonlinear Analysis. Real World Applications. (2007) 8, no. 3, 903–921, https://doi.org/10.1016/j.nonrwa.2006.03.014, 2307759, ZBL1149.34046.
10.1016/j.nonrwa.2006.03.014
Web of Science® Google Scholar
10 Yan X.-P., Hopf bifurcation and stability for a delayed tri-neuron network model, Journal of Computational and Applied Mathematics. (2006) 196, no. 2, 579–595, https://doi.org/10.1016/j.cam.2005.10.012, 2249447, ZBL1175.37086.
10.1016/j.cam.2005.10.012
Web of Science® Google Scholar
11 Fan D. and Wei J., Hopf bifurcation analysis in a tri-neuron network with time delay, Nonlinear Analysis. Real World Applications. (2008) 9, no. 1, 9–25, https://doi.org/10.1016/j.nonrwa.2006.08.008, 2370159, ZBL1149.34044.
10.1016/j.nonrwa.2006.08.008
Web of Science® Google Scholar
12 Hassard B., Kazarinoff D., and Wan Y., Theory and Applications of Hopf Bifurcation, 1981, Cambridge University Press, Cambridge, Mass, USA, 603442.
Google Scholar
13 Ruan S. and Wei J., On the zeros of transcendental functions with applications to stability of delay differential equations with two delays, Dynamics of Continuous, Discrete & Impulsive Systems. Series A. Mathematical Analysis. (2003) 10, no. 6, 863–874, 2008751, ZBL1068.34072.
Web of Science® Google Scholar
14 Hale J. K., Theory of Functional Differential Equations, 1977, Springer, New York, NY, USA.
10.1007/978-1-4612-9892-2
Web of Science® Google Scholar

Citing Literature

All articles

Stability and Bifurcation Analysis of a Three-Dimensional Recurrent Neural Network with Time Delay

Abstract

1. Introduction

2. Stability and Existence of Hopf Bifurcation

3. Direction of Hopf Bifurcations and Stability of the Bifurcating Periodic Orbits

4. Computer Simulation

Acknowledgment

References

Citing Literature

Figures

References

Information

About Wiley Online Library

Help & Support

Opportunities

Connect with Wiley

Stability and Bifurcation Analysis of a Three-Dimensional Recurrent Neural Network with Time Delay

Abstract

1. Introduction

2. Stability and Existence of Hopf Bifurcation

3. Direction of Hopf Bifurcations and Stability of the Bifurcating Periodic Orbits

4. Computer Simulation

Acknowledgment

References

Citing Literature

Figures

References

Related

Information