IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, VOL. 29, NO. 7, JULY 2018 2769
Stability Analysis of Continuous-Time and
Discrete-Time Quaternion-Valued
Neural Networks With Linear
Threshold Neurons
Xiaofeng Chen, Qiankun Song, Zhongshan Li, Zhenjiang Zhao, and Yurong Liu
Abstract—This paper addresses the problem of stability
for continuous-time and discrete-time quaternion-valued neural
networks (QVNNs) with linear threshold neurons. Applying
the semidiscretization technique to the continuous-time QVNNs,
the discrete-time analogs are obtained, which preserve the
dynamical characteristics of their continuous-time counterparts.
Via the plural decomposition method of quaternion, homeomor-
phic mapping theorem, as well as Lyapunov theorem, some
sufficient conditions on the existence, uniqueness, and global
asymptotical stability of the equilibrium point are derived for
the continuous-time QVNNs and their discrete-time analogs,
respectively. Furthermore, a uniform sufficient condition on the
existence, uniqueness, and global asymptotical stability of the
equilibrium point is obtained for both continuous-time QVNNs
and their discrete-time version. Finally, two numerical examples
are provided to substantiate the effectiveness of the proposed
results.
Index Terms—Global asymptotical stability, plural decom-
position method, quaternion-valued neural networks (QVNNs),
semidiscretization technique.
I. INTRODUCTION
O
VER the past few decades, real-valued neural net-
works (RVNNs) and complex-valued NNs (CVNNs)
have been extensively investigated because of their widespread
applications in various areas, such as associative memory,
Manuscript received September 13, 2016; revised February 11, 2017;
accepted May 10, 2017. Date of publication June 5, 2017; date of current
version June 21, 2018. This work was supported in part by the National Nat-
ural Science Foundation of China under Grant 61273021 and Grant 61473332,
in part by the Program of Chongqing Innovation Team Project in University
under Grant CXTDX201601022, in part by the Natural Science Foundation
of Chongqing Municipal Education Commission under Grant KJ1705138,
and in part by the Natural Science Foundation of Chongqing under
Grant cstc2017jcyjA1353. (Corresponding author: Qiankun Song.)
X. Chen and Q. Song are with the Department of Mathematics, Chongqing
Jiaotong University, Chongqing 400074, China (e-mail: xxffch@126.com;
qiankunsong@163.com).
Z. Li is with the Department of Mathematics and Statistics, Georgia State
University, Atlanta, GA 30302, USA (e-mail: zli@gsu.edu).
Z. Zhao is with the Department of Mathematics, Huzhou University,
Huzhou 313000, China (e-mail: zhaozjcn@163.com).
Y. Liu is with the Department of Mathematics, Yangzhou University,
Yangzhou 225002, China, and also with the Communication Systems and
Networks Research Group, Faculty of Engineering, King Abdulaziz Univer-
sity, Jeddah 21589, Saudi Arabia (e-mail: yrliu_66@126.com).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TNNLS.2017.2704286
pattern recognition, parallel computation, combinatorial opti-
mization, and quantum communication [1]–[10]. In such appli-
cations, it is of prime importance to ensure that the designed
NNs are stable. Therefore, the stability of RVNNs and CVNNs
has received much attention, and there are numerical and
fruitful results [11]–[22]. Recently, due to some interesting
applications found in 3-D geometrical affine transformation,
image compression, color night vision, and so on [23]–[25],
quaternion-valued NNs (QVNNs), which are an extension of
CVNNs, have begun to receive initial research interest and
many results have been obtained [26]–[29]. Therefore, QVNNs
deserve further investigation for their broad applications in
various fields.
It is quite important to choose activation functions for the
design of NNs, because the activation functions are signif-
icant factors that affect the dynamics of the designed NNs.
Various types of activation functions, such as step functions,
sigmoid functions, piecewise linear functions, Mexican-hat-
type functions, and so on, have been used for NNs [30]–[34].
However, there are some restrictions on choosing complex-
valued activation functions. For example, the complex-valued
activation functions cannot be adopted by bounded and ana-
lytic functions in complex domain. This is because of the fact
that every bounded entire function must be constant according
to Liouville’s theorem. Naturally, there exist greater challenges
with choosing appropriate quaternion-valued activation func-
tions for QVNNs, not just because of much more complexities
of QVNNs but also for the insufficiency of quaternion theory
compared with abundant and mature complex theory. Recently,
the studies reported in [35] focused on a class of NNs
with complex-valued linear threshold activation functions.
This class of NNs has potential in many applications, such
as decision making, digital selection, or analogy amplifica-
tion [36], [37]. Therefore, in this paper, we adopted the
quaternion-valued linear threshold functions inspired by the
complex-valued ones employed in [35].
Additionally, in implementations and applications of NNs
including QVNNs for computer simulations and computational
purposes, it is usual to discretize continuous-time NNs to
discrete-time analogs [38]–[40]. In the discretization process,
we need to consider two crucial issues: how to formulate
discrete-time analogs of the continuous-time NNs, and how to
2162-237X © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
评论0
最新资源