Investigation on Amari’s dynamical neural field with global constant inhibition☆
Introduction
Dynamical neural field models are presented during 1970s, which describe large scale activation of cortical neurons as a continuous neural field (Amari, 1977, Ermentrout and Cowan, 1979, Feldman and Cowan, 1975, Kishimoto and Amari, 1979, Wilson and Cowan, 1972, Wilson and Cowan, 1973). Among these models, an important one is Amari’s dynamical neural field model (Amari, 1977). Amari’s dynamical neural field model has successfully interpreted large amounts of important phenomena and problems (Giese, 1999, Simmering et al., 2008), so that it has been extensively applied to psychophysics, neurophysiology, machine vision and cognition (Engels and Schöner, 1995, Erlhagen and Bicho, 2006, Faubel and Schöner, 2008, Schöner et al., 1995). In these application, Amari’s dynamical neural field is usually employed to describe the fundamental interaction function among neurons in the same cortical layer such as primary visual cortex V1, while some other models, for example, the neural model of Favorov and Kursun (2011), are employed to approximate more complex functions. In recent years, Amari’s dynamical neural field has been introduced to the fields like data clustering and obtained good results (Jin, Peng, & Li, 2011).
In the research and applications of Amari’s dynamical neural field model, especially for practical applications such as data clustering, the stationary solution with local excited region to Amari’s dynamical neural field model that are usually considered as the patterns been perceived. As a result, The existence and stability of stationary solution to dynamical neural field have been discussed a lot. The existence of a local excitation pattern solution as well as its waveform stability in 1-dimensional Euclid space is proved in the absent of external input, which shows that the field can keep short-term memory (Amari, 1977, Kishimoto and Amari, 1979, Laing and Troy, 2003, Owen et al., 2007, Wennekers, 2002). Kubota et al. prove that with time-invariant external input there can be bistable local excitation solutions with different lengths (Kubota, Hamaguchi, & Aihara, 2009). The properties of stationary solution to Amari’s dynamical neural field in 2-dimensional Euclid space are studied, which show that some different conditions are required for the existence and stability for local excitation solutions (Taylor, 1999, Taylor, 2003, Werner and Richter, 2001). For high dimensional space , some results on the existence and stability of solution and stationary solution to Amari’s dynamical neural field are obtained (Jin, Liang et al., 2011, Potthast and Graben, 2010).
In these studies of the stationary solution to Amari’s dynamical neural field model, most discussions focus on dynamical neural field with typical “Mexican Hat” shape kernel. For dynamical neural field with such kind of kernel, the type of its stationary solution is sensitive to its parameters. When it is applied to missions like data clustering, sometimes it is difficult to find proper parameters, especially in some cases for data set with complex structures in high dimensional space (Jin, Peng et al., 2011). As a result, though dynamical neural field with “Mexican Hat” shape kernel has been studied a lot, it is not quite convenient for practical issues. Another kind of kernels of Amari’s dynamical neural field are those which generate global constant inhibition. Dynamical neural field with such global constant inhibitory kernel has also been employed to pattern recognition and data clustering (Faubel and Schöner, 2008, Jin and Huang, 2013), which shows good performance. However, the properties of dynamical neural field with global constant inhibitory kernel are rarely discussed.
In this paper, the existence and stability of stationary solution to dynamical neural field with global constant inhibitory kernel are discussed. Some existence conditions for three important types of stationary solution: “”-solution, “”-solution and “bubble”-solution for Heaviside Step and Sigmoid threshold functions, as well as unbounded and bounded perceive field, are presented, which provide the theoretical basis for further extensive application of dynamical neural field with global constant inhibitory kernel.
The remainder of this paper is arranged as follows. In Section 2, Amari’s dynamical neural field model is briefly described, and the clustering result using dynamical neural field with“Mexican Hat” shape kernel and global constant inhibitory kernel are given for comparison. In Section 3, some general results on stationary solution to dynamical neural field with global constant inhibitory kernel are given. In Section 4, when the threshold function is a discontinuous Heaviside Step function, the properties of dynamical neural field with global constant inhibitory kernel is discussed, in unbounded and bounded perceive field respectively. In Section 5, when the threshold function is a continuous Sigmoid function, the properties of dynamical neural field with global constant inhibitory kernel is discussed, in unbounded and bounded perceive field respectively. Section 6 is conclusion.
Section snippets
Amari’s dynamical neural field model
Typical Amari’s dynamical neural field model is usually described by is called perceive space. is a positive time constant. is the resting level of the neural field. is the input signal distribution. The region is called excited region. is a monotonically increasing nonlinear threshold function, satisfying that and , which describes the neural field’s feedback of each excited point to
General results on stationary solution to dynamical neural field with global constant inhibitory kernel
Consider the following Amari’s dynamical neural field model: where containing the origin as its inner point. is its initial state. Without losing generality, is assumed to be bounded, i.e., and continuous, where . Since in most applications, the input function is usually assumed to be time invariant, it can be written as . Suppose where is “Mexican Hat” shape and satisfies
being a Heaviside step function
Suppose that the threshold function is a Heaviside step function given by (2). Proposition 2 If there exists a stationary solution to dynamical neural field (13) andthen is a -solution.
Proof Since the threshold function for all , It follows that there is only -solution to dynamical neural field (13). □
It is a sufficient condition for the existence of -solution. It indicates
being a Sigmoid function
When the threshold function is required to be continuous, it is usually selected as a Sigmoid function (3). Different from the discontinuous threshold function defined as a Heaviside step function, since for all , the Sigmoidal threshold function would activate the lateral effect of dynamical neural field even when . It makes the properties of dynamical neural field (13) more complicated.
Conclusion
In this paper, we discuss the properties of Amari’s dynamical neural field with global constant inhibition. Some important results on the existence and stability of its stationary solution are obtained, providing a basis for further extensive application of dynamical neural field to technical fields.
When threshold function is a Heaviside step function, some sufficient conditions and necessary conditions for existence of -solution and “bubble”-solution to dynamical neural field with global
References (24)
- et al.
Dynamic fields endow behavior-based robots with representations
Robotics and Autonomous Systems
(1995) - et al.
Learning to recognize objects on the fly: A neurally based dynamic field approach
Neural Networks
(2008) - et al.
Existence and properties of stationary solution of dynamical neural field
Nonlinear Analysis-Real
(2011) - et al.
Dynamics of behavior: theory and applications for autonomous robot architectures
Robotics and Autonomous Systems
(1995) - et al.
Generalizing the dynamic field theory of spatial cognition across real and developmental time scales
Brain Research
(2008) Bubbles in the brain?
Trends in Cognitive Sciences
(2003)- et al.
Excitatory and inhibitory interactions in localized populations of model neurons
Biophysical Journal
(1972) Dynamics of pattern formation in lateral-inhibition type neural filed
Biological Cybernetics
(1977)- et al.
The dynamic neural field approach to cognitive robotics
Journal of Neural Engineering
(2006) - et al.
A mathematical theory of visual hallucination patterns
Biological Cybernetics
(1979)
Neocortical layer 4 as a pluripotent function linearizer
Journal of Neurophysiology
Large-scale activity in neural nets I: Theory with application to motoneuron pool responses
Biological Cybernetics
Cited by (1)
A novel neural model with lateral interaction for learning tasks
2021, Neural Computation
- ☆
This project is supported by National Natural Science Foundation of China (Grant No. 11301096, 11226141) and Natural Science Foundation of Guangxi under the contact no. 2013GXNSFBA019018.