In the past several decades, the dynamics of BAM neural networks has been widely investigated for their essential applications in classification, pattern recognition, optimization, signal and image processing, and so on [
1‐
41]. In 1987, Kosko [
42] proposed the following BAM neural network:
$$\begin{aligned} \textstyle\begin{cases} \frac{du_{i}(t)}{dt}=-a_{i}u_{i}(t)+\sum_{j=1}^{n}a_{ij}f_{j}(v_{j}(t-\sigma _{j}(t)))+I_{i},\\ \frac{dv_{i}(t)}{dt}=-b_{i}v_{i}(t)+\sum_{j=1}^{n}b_{ij}g_{j}(u_{j}(t-\tau_{j}(t)))+J_{i}, \end{cases}\displaystyle \end{aligned}$$
(1.1)
where
\(i=1,2,\ldots,n,t>0\). Here,
\(a_{i}>0,b_{i}>0\) denote the time scales of the respective layers of the network;
\(-a_{i}u_{i}(t)\) and
\(-b_{i}v_{i}(t)\) stand for the stabilizing negative feedback of the model. Noticing that the leakage delay often appears in the negative feedback term of neural networks (see [
43‐
47]), Gopalsmay [
48] studied the stability of the equilibrium and periodic solutions for the following BAM neural network:
$$\begin{aligned} \textstyle\begin{cases} \frac{dx_{i}(t)}{dt}=-a_{i}x_{i}(t-\tau_{i}^{(i)})+\sum_{j=1}^{n}a_{ij}f_{j}(y_{j}(t-\sigma_{j}^{(2)}))+I_{i},\\ \frac{dy_{i}(t)}{dt}=-b_{i}y_{i}(t-\tau_{i}^{(2)})+\sum_{j=1}^{n}b_{ij}g_{j}(x_{j}(t-\sigma_{j}^{(1)}))+J_{i}, \end{cases}\displaystyle \end{aligned}$$
(1.2)
where
\(i=1,2,\ldots,n,t>0\). Since the delays in neural networks are usually time-varying in the real world, Liu [
49] discussed the global exponential stability for the following general BAM neural network with time-varying leakage delays:
$$\begin{aligned} \textstyle\begin{cases} \frac{dx_{i}(t)}{dt}=-a_{i}x_{i}(t-\delta_{i}(t))+\sum_{j=1}^{n}a_{ij}f_{j}(y_{j}(t-\sigma_{ij}(t))+I_{i},\\ \frac{dy_{i}(t)}{dt}=-b_{i}y_{i}(t-\eta_{i}(t)+\sum_{j=1}^{n}b_{ij}g_{j}(x_{j}(t-\tau _{ij}(t))+J_{i}. \end{cases}\displaystyle \end{aligned}$$
(1.3)
However, so far, there have been rare reports on the existence and exponential stability of anti-periodic solutions of neural networks, especially for neural networks with leakage delays. Furthermore, the existence of anti-periodic solutions can be applied to help better describe the dynamical properties of nonlinear systems [
49‐
65]. So we think that the investigation on the existence and stability of anti-periodic solutions for neural networks with leakage delays has significant value. Inspired by the ideas and considering the change of system parameters in time, we can modify neural network model (
1.3) as follows:
$$\begin{aligned} \textstyle\begin{cases} \frac{dx_{i}(t)}{dt}=-a_{i}x_{i}(t-\delta_{i}(t))+\sum_{j=1}^{n}a_{ij}(t)f_{j}(y_{j}(t-\sigma_{ij}(t)))+I_{i}(t),\\ \frac{dy_{i}(t)}{dt}=-b_{i}y_{i}(t-\eta_{i}(t))+\sum_{j=1}^{n}b_{ij}(t)g_{j}(x_{j}(t-\tau_{ij}(t)))+J_{i}(t). \end{cases}\displaystyle \end{aligned}$$
(1.4)
The main objective of this article is to analyze the exponential stability behavior of anti-periodic oscillations of (
1.4). Based on the fundamental solution matrix, Lyapunov function, and fundamental function sequences, we establish a sufficient condition ensuring the existence and global exponential stability of anti-periodic solutions of (
1.4). The derived findings can be used directly to numerous specific networks. Besides, computer simulations are performed to support the obtained predictions. Our findings are a good complement to the work of Gopalsmay [
48] and Liu [
49].
The paper is planned as follows. In Sect.
2, several notations and preliminary results are prepared. In Sect.
3, we give a sufficient condition for the existence and global exponential stability of anti-periodic solution of (
1.4). In Sect.
4, we present an example to show the correctness of the obtained analytic findings.