Selective attention in visual search: A neural network of phase oscillators

doi:10.1016/S0925-2312(01)00483-0

Neurocomputing

Volumes 38–40, June 2001, Pages 1151-1160

https://doi.org/10.1016/S0925-2312(01)00483-0 Get rights and content

Abstract

We study the selective attention mechanism in the context of visual search and formulate a neurodynamical model consisting of different pools of interconnected phase oscillators. Each oscillator is described by an integrate-and-fire-type equation. Numerical calculations for the cases of feature and conjunction searches are presented. The neural system is build up with parallel process only. Independent competition mechanisms along each feature dimension are taken into account. Visual attention appears as an emergent property of the dynamic of the system resulting from the temporal synchronization of the pools which bind the features of the searched target.

Introduction

Understanding the mechanisms of the visual attention process has become the subject of active research nowadays. The psychological theories of human vision usually consider two different levels in the visual processing: the preattentive (or early) and the attentive (or focal) stages. In order to study the visual processing in the laboratory, visual-search tasks are performed. That is, a subject is asked to find a target item in the midst of other distractors items. In general, there are two conditions [8]: the feature-search (where the target differs from the distractors only by a single feature) and the conjunction-search condition (where the target is defined by the conjunction of information from two or more properties). Treisman and Gelade [8] assumed that, if the preattentive processing occurs automatically and across the visual field, a target that is distinct from the distractors in its preattentive representation in the brain, should pop out of the display. In the case of feature-search they found that the target is detected equally fast independently of the display size. On the contrary, for the conjunction-search experiments they found that the time taken to find the target increases linearly with the number of distractors. The authors summarized these results in the feature integration theory (FIT) claiming that the first parallel stage identifies the basic visual features that are present and that the slower serial stage then combines these features to produce complex representations of visual objects. In this way they describe the preattentive stage mentioned above as a parallel mechanism whereas the attentive stage is related to the serial mechanism. In the case of feature-search, FIT states that the parallel stage should be able to detect the target and predicts an almost flat slope for the reaction time-display size curve. In the conjunction-search case the serial stage must process each element individually until it finds the target conjunction, requiring more time as the number of elements in the display increases. Working within FIT, priority maps are used for each feature dimension to finally obtain an overall activation map where focused attention selects and integrates the features present at particular locations, indicating how likely the stimulus at each location is to be the target. Therefore, selective visual attention has been thought as a spotlight which enhances information within a selected region in the scene and filters out information outside it [8]; such processing of the information done iteratively overall the visual field.

An alternative to the spotlight metaphor used to describe selective visual attention is that it is an emergent property of many neural mechanisms working to resolve competition for visual processing [3]. Two basic phenomena that define the problem of visual attention are considered: the limited capacity of processing information and the ability to filter out unwanted information. The authors suggest that at several points between input and response, objects in the visual field compete for limited processing capacity. This competition is biased in part by bottom–up neural mechanisms that separate figures from their background and in part by top–down mechanisms that select objects of relevance to current behaviour.

Deco and Zihl [2] studied also the visual search tasks. They wondered if the linear increasing reaction time observed in the conjunction search experiments was necessarily due to a serial mechanism or if there is only parallel processing followed by a dynamical time-consuming latency. In order to clarify this point, they formulated a neurodynamical model consisting of interconnected populations of biological-inspired neurons. They showed that a neural system for visual search can be built, which works across the visual field in parallel but due to its intrinsic dynamics can show the two experimentally observed modes of visual attention; neither explicit serial focal search nor saliency maps need to be assumed. The model takes into account independent competition mechanisms along each feature dimension to explain the experimental data, implying that the parallel competition is performed before binding the features. The competitive mechanisms were implemented at the neural rate activity level and the neural population dynamics was described by a system of coupled differential equations derived in the framework of the mean-field approximation.

From the experimental point of view, the measurements of Chelazzi et al. [1] provide evidence for parallel processing across space in tasks where monkeys have to search for a target.

On the other hand, let us recall that there exist several works in the literature indicating that coherent oscillations in a population of neurons might be the basic mechanism to ensure feature binding in the visual system [7], [10]. The neurophysiological evidence for employing oscillators in the theoretical description of the processes comes from a large variety of oscillations observed in neural systems. Different studies of the visual cortex of cats and monkeys found oscillatory responses which are induced by external stimuli [4], [5]. We can mention the discovery of oscillations within the beta frequency range found in the firing of individual neurons of the visual cortex in response to moving light bars [4].

In this paper we present a neurodynamical model consisting of an assembly of phase oscillators to be used to describe the measurements of Chelazzi et al. as well as the visual-search tasks experiments. The oscillators are interconnected and organized in different pools. The neural system is build up with parallel processes only. Independent competition mechanisms along each feature dimension are also taken into account. Visual attention will be interpreted as the result of the synchronous oscillation of certain pools (the ones that characterize the features of the target). Binding is thus achieved by the synchronized oscillatory activity of those pools.

Section snippets

Theory

The basic functional unit of our approach will be a population of N oscillators with all-to-all coupling (we will refer to such unit as a column), each of the oscillators described by an integrate-and-fire-type equation as given by Kuramoto [6]: $d φ_{n} d t =a(t)−φ_{n} +η_{n} (t)+ K N ∑ m≠n ∑ r δ(t−t_{m}^{(r)}), φ_{n} (t_{n})=0,$ where φ_n represents the phase of nth oscillator (n=1,…,N), a(t) is the external visual input, K is the strength coupling and η_n is the Gaussian noise component. The last term in Eq. (1) represents the

Numerical simulations

We make numerical calculations for the cases of feature and standard conjunction-searches. In the standard conjunction-search the target shares only one feature with each distractor group. Let us assume that the items are defined by two feature dimensions (K=2, e.g. size and colour), each one having two values (L(k)=2 for $k=1, 2$ , e.g. size: big/small, colour: white/black).

At each display size we repeat the experiment 20 times, each time with different randomly generated target definition (i.e.

Acknowledgements

S.C. acknowledges financial support from Consejo Nacional de Investigaciones Cientı́ficas y Técnicas and Universidad Nacional de Rosario, Argentina

References (10)

Y Kuramoto
Collective synchronization of pulse-coupled oscillators and excitable units
Physica D
(1991)
A Treisman et al.
A feature-integration theory of attention
Cognitive Psychol.
(1980)
L Chelazzi et al.
A neural basis for visual search in inferior temporal cortex
Nature
(1993)
G. Deco, J. Zihl, Top–down selective visual attention: a neurodynamical approach, Visual Cognition 8 (2001)...
R Desimone et al.
Neural mechanisms of selective visual attention
Annu. Rev. Neurosci.
(1995)

There are more references available in the full text version of this article.

Cited by (2)

An attention transfer entropy based causality analysis with applications in rooting short-term disturbances for chemical processes
2023, ISA Transactions
Citation Excerpt :
The attention refers to a cognition behavior of humans that enables them to concern certain information while simultaneously ignoring the other ones [42,43]. Recently, the attention mechanism is proliferating to the field of machine learning in reducing algorithmic parameters and increasing calculating speeds [44–47]. Nevertheless, it is still hard for practical engineering applications in the absence of prior knowledge.
The transfer entropy (TE) based causality analysis is able to provide a typical solution for fault rooting of industrial processes. However, short-term disturbances that occur during nominal operations of chemical processes are usually neglected because of the fixed time window of TE for global data distributions. Inspired by the selective attention idea, we propose attention transfer entropy (ATE) that helps to locate prominent targets. Concerning temporal features of industrial time series, prior knowledge is employed for constructing an interpretable model. We verify the reliability and effectiveness of the method with coal gasification process data. Additionally, the algorithm is compared to conventional causality analysis methods, proving that ATE enjoys excellent performances in rooting short-term disturbances with lower calculation burden.
Competition for synchronization in a phase oscillator system
2013, Physica D: Nonlinear Phenomena
Citation Excerpt :
Our interest in system (1)–(2) and its stable states is related to the application of this system (or its generalization) as a model of selective attention [23–28]. The systems with relaxation oscillators and a central inhibitory neuron, the systems with Hodgkin–Huxley type neurons and star-like connections, as well as phase oscillator systems of other architectures were used as attention models [29–33]. Selective attention is a built-in mechanism of perception that allows the animals to extract from the large amount of simultaneously obtained information a smaller part (which is usually of significant importance) that should be processed in more detail.
A system of phase oscillators with a Central Oscillator (CO) and a set of n Peripheral Oscillators (POs) is considered. Feed-forward and feedback connections between the CO and POs are determined by two interaction functions which are assumed to be smooth, odd, and periodic. To describe the competition of POs for synchronization with the CO, we study the asymptotic stability of fixed points corresponding to in-phase synchronization of a group of k POs, while other POs are in anti-phase with the CO. It is shown that stability conditions can be formulated in terms of four parameters that describe the slopes of the interaction functions at zero and half-period points. Analytical description of stability in terms of the regions in 4-dimensional parameter space is given. Combining stability analysis with the detailed study of geometry of invariant manifolds, the bifurcations of fixed points are investigated. We show that various dynamical regimes such as multistability, heteroclinic orbits, and chaos are possible. Analytical stability conditions for global synchronization of POs with the CO are formulated for the systems with local connections between POs. It is shown that synchronization in a large system with local connections becomes unstable even under weak desynchronizing influence from the CO. The application of the results to modeling in neuroscience and, in particular, for modeling visual attention is discussed.

¹: Permanent address: Instituto de Fı́sica Rosario, CONICET-UNR, Argentina.

View full text

Selective attention in visual search: A neural network of phase oscillators

Abstract

Introduction

Section snippets

Theory

Numerical simulations

Acknowledgements

Physica D

Cognitive Psychol.

A neural basis for visual search in inferior temporal cortex

Nature

Neural mechanisms of selective visual attention

Annu. Rev. Neurosci.