Elsevier

Neurocomputing

Volume 171, 1 January 2016, Pages 482-491
Neurocomputing

Multiplicity of almost periodic solutions for multidirectional associative memory neural network with distributed delays

https://doi.org/10.1016/j.neucom.2015.06.080Get rights and content

Abstract

In this paper, the multiplicity of almost periodic solutions is studied for a multidirectional associative memory (MAM) neural network with almost periodic coefficients and continuously distributed delays. Under some assumptions on activation functions, some invariant subsets of the MAM neural network are constructed. The existence of multiple almost periodic solutions are obtained by using the theory of exponential dichotomy and Schauder׳s fixed point theorem. Furthermore, a sufficient condition is derived for the local exponential stability of some almost periodic solutions and their exponential attracting domains are also given. An example is given to illustrate the effectiveness of the results.

Introduction

Generally, the multistability of neural networks, that is the existence and stability of multiple equilibrium points, periodic solutions or almost periodic solutions, is prerequisite when the neural network is applied to solve some problems of many-to-many association memory. It has attracted much attention in the recent years [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16]. According to the different characteristics of active function, such as piecewise linear nondecreasing, Mexican-hat-type, discontinuous or concave–convex, the coexistence and local stability of multiple equilibria have been extensively investigated [1], [2], [3], [4], [5], [6], [7], [8], [9], [10]. As is known to all, the cyclic behaviors or rhythmic activities are widespread in nature. Therefore, the existence and stability of multiple periodic or almost periodic solutions have also been analyzed [11], [12], [13], [14], [15], [16], [17]. From a phenomenological point of view, the almost periodic functions can represent more exactly rhythmic activities [15]. However, to the best of our knowledge, the investigation of multiple almost periodic solutions of neural networks is seldom reported [15], [17], [16]. In [15], a general methodology that involves geometric configuration of the n-neuron network structure for studying multistability was developed. Through applying the contraction mapping principle, some sufficient conditions guaranteeing the existence of 2n exponentially stable almost periodic solutions were given, when the connection strengths, time lags, and external bias are almost periodic functions of time. In [16], by considering two classes of activation functions, the authors revealed that under some conditions, there are 2n locally exponentially stable almost-periodic solutions of a delayed n-neuron neural network. The authors of [17] investigated the dynamics of 2n almost periodic attractors for cellular neural networks with variable and distributed delays, they obtained exponential attracting domains.

The multidirectional associative memory (MAM) neural network, which was proposed by the Japanese scholar M. Hagiwara in 1990 [18], is an extension of BAM neural network model [19]. In an MAM neural network, its neurons are arranged in three or more fields. The neurons in the same field of an MAM neural network are not connected to each other, but the neurons between every two different fields are fully interconnected. By using of MAM neural networks, one can achieve the many-to-many association memory. Therefore, this class of networks possesses wide application fields such as pattern recognition, image denoising and intelligent information processing [18], [20], [21], [22], [23]. In order to achieve the many-to-many association memory by MAM neural networks, it is necessary to ensure their multistability. In [24], we studied the multistability issue for a delayed MAM neural network with m fields and nk neurons in the field m as follows.dxkidt=akixki(t)+p=1,pkmj=1npwpjkifpj(xpj(tτpjki))+Ikiwhere k=1,2,,m, i=1,2,,nk, xki(t) denotes the membrane voltage of the ith neuron in the field k at time t, aki>0 denotes the decay rate of the ith neuron in the field k, fpj(·) is a neuronal activation function of the jth neuron in the field p, wpjki is the connection weight from the jth neuron in the field p to the ith neuron in the field k,Iki is the external input of the ith neuron in the field k,τpjki is the time delay of the synapse from the j neuron in the field p to the ith neuron in the field k. By using the Brouwer fixed point theorem and Dini upper right derivative, we obtained that some sufficient conditions which ensure the multidirectional associative memory neural network has 3l equilibria and 2l equilibria of them are stable, where l is a parameter associated with the number of neurons. System (1) is with discrete delay. It can only reflect the relationship between the state of the neuron xki at time t and the state of any other neuron xpj at the time ahead of it by a fixed time length τpjki.

Since a neural network usually has a spatial extent due to the presence of a multitude of parallel pathways of a variety of axon sizes and lengths, there is a distribution of propagation delays over a duration of time [25]. That is to say, the state of the neural network in the past every time affects its current state. Obviously, the distributed delays can reflect the relationship more really than discrete delays. Hence, we also studied the MAM neural network with the continuously distributed delays [26] as follows.dxkidt=akixki(t)+p=1,pkmj=1npwpjkifpj(0+gpjki(s)xpj(ts)ds)+Iki(t),where gpjki(s) is the delay kernel function of the synapse from the jth neuron in field p to the ith neuron in field k. By constructing a suitable Liapunov function and a Poincaré mapping, we proved the existence and the exponential stability of multiple periodic solutions of the MAM neural network (2).

From the viewpoint of biological nervous system, there exists memory vibration in human׳s cerebrum. And from the viewpoint of networks implementation, the neural network׳s external inputs Iki, the neuronal signal decay rates aki, and the connection weights wpjki often almost periodically vary with time because there are almost periodic phenomenons in the electrical power system. Therefore, the discussion on existence and stability of multiple almost periodic solutions of MAM neural networks is also very meaningful. Motivated by the above, in this paper, we study the existence and the exponential stability of multiple almost periodic solutions of an MAM neural network with almost periodic coefficients and continuously distributed delays as follows.dxkidt=aki(t)xki(t)+p=1,pkmj=1npwpjki(t)fpj(0+gpjki(s)xpj(ts)ds)+Iki(t).

The initial conditions associated with (3) are of the formxki(t)=θki(t),where k=1,2,,m, i=1,2,,nk and θki:(,0]R are continuous functions.

In this paper, our main object is to obtain the sufficient conditions ensuring the existence and the exponential stability of multiple almost periodic solutions of the MAM neural network (3) with initial condition (4). After constructing an invariant basin of system (3), we split it to multiple subsets. In every subset, there exists a almost periodic solution. And in the invariant subset, the almost periodic solution is exponentially stable. This paper is organized as follows. In the next section, we introduce some lemmas of exponential dichotomy method and almost periodicity. In Section 3, we construct an invariant basin of MAM neural network (3) and split it into multiple invariant subsets. In Section 4, we investigate the existence of multiple almost periodic solutions of (3) by using exponential dichotomy and Schauder׳s fixed point theorem. Meanwhile, we construct exponential attracting domain of each almost periodic solution. In Section 5, we investigate the exponential stability of multiple almost periodic solutions of (3) by constructing exponential attracting domain of each almost periodic solution. In Section 6, an example is given to illustrate the effectiveness of our results.

Section snippets

Preliminary

We firstly introduce two lemmas on exponential dichotomy [27].

Set the vector x(t)=(x11(t),,x1n1(t),,xm1(t),,xmnm)Tcol{xki(t)}.

Lemma1

Let B(t)=diag(a11(t),,a1n1(t),,am1(t),,amnm(t)). If aki(t)(k=1,2,,m,i=1,2,,nk) are almost periodic functions and limT+(1/T)tt+Taki(s)ds>0, then systemdxdt=B(t)x(t)admits an exponential dichotomy.

Lemma 2

If system (5) admits an exponential dichotomy, the almost periodic system dxdt=B(t)x(t)+η(t)has a unique almost periodic solution φ(t), and φ(t)=tX(t)PX1(s)η(s

Invariant basins

In this section, we will prove that the set Φ as following is an invariant basin of MAM neural network (3). Under some assumptions on activation functions and system, we construct 2N0 invariant subsets, which methods used in this section are similar to the literatures [24], [26].

Set Φ={col{xki(t)}C||xki(t)|xki0+forkI[m],iI[nk]}.

Theorem 1

If conditions (H1)(H3) hold, then Φ is an invariant basin of (3).

proof

Let x(t;φ)=col{xki(t;φ)} be the solution of (3) with initial condition φ=col{φki(t)}Φ. From (3),we

The existence of 3N0 almost periodic solutions

In this section, we discuss the existence of 3N0 almost periodic solution of system (3). We firstly define a vectorq=col{qki},where qki is defined as follows. If kIM and iI[nˇk], then qki=q(k+1)i=1,2 or 3; if k>M or i>nˇk then qki=0. And define the following 3N0 subsets of ΨΨq={col{ψki(t)}ΩΨ|ψki(t)R˜kiqkiforiI[nˇk],kIM,tR},where the vector q is defined as (16).

We make further the assumption (H7).

(H7) There exist constants Lkiqki>0 such that |fki(u)fki(v)|Lkiqki|uv| for any u,vR˜kiqki

The exponential stability of 2N0 almost periodic solutions

In this section, in order to discuss the stability of almost periodic solutions of MAM neural network (3), we restrict qki2 in vector (16), and thus there exist 2N0 vectors q=col{qki}. Set ηki=qki where kIM, iI[nˇk], and define 2N0 vectors η=(η11,,η1nˇ1, η31,,η3nˇ3,,η(M1)1,,η(M1)nˇM1).

We relax the condition (21) to the following assumption (H8).

(H8) There exist constants γki>0(kI[m],iI[nk]) such thata̲kiγkip=1,pkmj=1npw¯pjkiLpjqpjγpj>0.

Theorem 5

Suppose that conditions (H1)~(H8) hold.

An example

Consider an MAM neural network with four fields as follows:dxkidt=aki(t)xki(t)+p=1,pk4j=1npwpjki(t)fpj(0+gpjki(s)xpj(ts)ds)+Iki(t),where k=1,2,3,4,n1=n2=1,n3=n4=2, the neuronal signal decay rates a11(t)=1 , a21(t)=1,a31(t)=1,a32(t)=1,a41(t)=1,a42(t)=1, the external inputs I11(t)=0.5sin(2t),I21(t)=0.3sin(5t),I31(t)=0.3cos(2t),I32(t)=0.2sin(3t),I41(t)=0.3cos(5t),I42(t)=0.2cos(3t), the connection weights w2111(t)=2,w3111(t)=0.01,w3211(t)=0.01,w4111(t)=0.2,w4211(t)=0.001,w1121(t)=3,w3121(t)=

Conclusions

In this paper, the multistability has been studied for the MAM neural network (3) with almost periodic coefficients and continuously distributed delays. Sufficient conditions are obtained which ensure the existence of 3N0 almost periodic solutions. It is proved that the 2N0 almost periodic solutions among them are exponentially stable.

We know that constants and periodic functions are all special almost periodic functions. The discrete time delay can be obtained from distributed delays with

Tiejun Zhou is a professor of mathematics at the College of Science, Hunan Agricultural University, Changsha, China. He received the M.Sc. degree in applied mathematics from the University of Hunan, Changsha, China, in 2003, and received the Ph.D. degree in mathematics from Center South University, Changsha, China, in 2007. His research interests focus on the theory of differential and difference equations, neural networks and population dynamical systems.

References (27)

Cited by (10)

  • Fixed-time synchronization of fractional order memristive MAM neural networks by sliding mode control

    2020, Neurocomputing
    Citation Excerpt :

    Associative memory, as part of the representative nerve activities of brain intelligence, has aroused the attentions of numerous researchers. Studying the neural networks with associative memory function, such as bidirectional associative memory neural networks [1], multidirectional associative memory neural networks (MAMNNs) [2], cellular neural networks [3] and hopfield neural networks [4], is an effective way to study associative memory. Among these, MAMNNs can achieve a wider range of associative memories, such as many to many associative memories.

  • Fixed-time synchronization control of memristive MAM neural networks with mixed delays and application in chaotic secure communication

    2019, Chaos, Solitons and Fractals
    Citation Excerpt :

    In all associative memory models, multidirectional associative memory neural networks (MAMNNs), proposed by M.Hagiwara [1] in 1990, are more conformed to simulate many to many associations of human brain than the other methods. Although the dynamic behaviors of this neural networks have been widely studied [2,3]. But the weight of this model is still immutable, this obviously does not match the variable synaptic weights of the brain.

  • Double almost periodicity for high-order Hopfield neural networks with slight vibration in time variables

    2018, Neurocomputing
    Citation Excerpt :

    In various fields such as stable electronic circuits, ecological systems, neural networks, and so forth, almost periodic behavior is widely exhibited and analyzed, and is applied in practical applications to establish more accurate results. A vast amount of researches have been directed toward studying almost periodic phenomena (see [14–17]). As a matter of fact, although there were many works about neural networks on time scales, the study of these neural networks had been restricted to periodic time scales until the concept of almost periodic time scales was introduced (see [28–33]).

  • Multidirectional Associative Memory Neural Network Circuit Based on Memristor

    2023, IEEE Transactions on Biomedical Circuits and Systems
View all citing articles on Scopus

Tiejun Zhou is a professor of mathematics at the College of Science, Hunan Agricultural University, Changsha, China. He received the M.Sc. degree in applied mathematics from the University of Hunan, Changsha, China, in 2003, and received the Ph.D. degree in mathematics from Center South University, Changsha, China, in 2007. His research interests focus on the theory of differential and difference equations, neural networks and population dynamical systems.

Yi Wang received the master degree in computer application technology from Central South University, Changsha, China,in 2006. Since 1995, she has been in the College of Information Science and Technology at Hunan Agricultural University, Changsha, China, where she is currently an associate professor of computer science and technology. Her research interests include crop information acquisition and neural network dynamical systems.

Min Wang received the M.Sc. in mathematics from the Hunan Normal University of Changsha, Hunan, China, in 2006, and received the Ph.D. degree in agriculture from Hunan Agriculture University of Changsha, Hunan, China, in 2014. Since 2006, she has been working in the College of Science, Hunan Agriculture University. Her research interests include dynamical systems and their applications.

View full text