Available online at www.sciencedirect.com

Procedia Computer Science 7 (2011) 244–245

The European Future Technologies Conference and Exhibition 2011

Computational Modeling of Visual Selective Attention

Kleanthis C. Neokleous, Christos N. SchizasUniversity of Cyprus, Department of Computer Science 75, kallipoleos str., P.O, Box 20537, Nicosia, Cyprus

Abstract

An overview of a nerocomputational model of visual selective attention that has been properly implemented is presented in thisabstract.

Visual selective attention is a fundamental function of human cognition and a highly important brain mechanism, essential forthe functioning of the human brain as a system. A comprehensive example of the role of human attention can be seen by noting thatat each instant of conscious life, each person receives millions of external stimulations from his/her sensory systems, while only alimited amount is selected by attention for further processing that leads to conscious perception. If every stimulus was allowed topass into perception, one would have been soon overflowed and in constant distraction. Adding to external stimulation all internalstimuli (e.g., thoughts), a person would end up in a totally unstable state. Selective attention is thus regarded as the main controlmechanism, necessary for keeping the brain system in stability. It does so by filtering out any irrelevant information while at thesame time advancing any vital stimulation to higher cortical areas for further processing.

Attention can be oriented towards object or empty space either in a voluntary or an automatic manner. That is, attention can beguided by top-down and bottom-up processing as cognition can be regarded as a balance between internal motivations and externalstimulations. Top-down or endogenous attention refers to the volitional modulation of neural activity that corresponds to an objector a location in space, and it functions in response to signals initiated by internal goals, that most likely originate in the parietaland frontal lobes of the brain (Buschman & Miller, 2007). Bottom-up or exogenous attention on the other hand is a faster and moreautomatic process that relies on the sensory saliency of stimuli registered by sub cortical structures and the primary sensory cortices(Corbetta & Shulman, 2002).

Studying the brain from the computer scientists’ perspective has always being a great challenge, and is usually divided undertwo main paths within the computational intelligence (CI) field. On one, to understand and mimic in a sense the functionality of thehuman brain has triggered the design and implementation of artificial intelligent systems such as robotics, expert systems etc. Onthe other, the understanding of certain brain functions can be facilitated with the implementation of relevant cognitive computationalmodels.

Our objective is to develop a plausible and biologically realistic computational model of visual selective attention using toolsfrom the field of computational intelligence and use it in engineering and other applications. In recent years, an increased interestin developing cognitive models for a variety of technology and engineering applications has been observed and more specifically,there has been much interest in the development of systems capable of simulating users’ attention and how these systems could bepractically and effectively used.

For example, a tendency towards practical systems that are based on human attention has been observed (Horvitz et al., 2003)while research on computer vision is as well heavily dependent on the principles of human attention (Sun & Fisher, 2003). To applyideas and concepts of human attention in the Computational Intelligence area, it is necessary to develop relevant computationalmodels that will allow pinpoint the functional details of this brain mechanism.

In line with the above, we present here a brief description of a computational model of visual selective attention that we havedesigned and implemented.

© Selection and peer-review under responsibility of FET11 conference organizers and published by Elsevier B.V.

Keywords: Attention; Computational Modeling; Spiking Neural Networks

1877-0509/$ – see front matter © Selection and peer-review under responsibility of FET11 conference organizers and published by Elsevier B.V.doi:10.1016/j.procs.2011.09.030

1

iiiociptis

bsmc&nac2

iaN

A

R

J

MSE

CK

K

N

K

A

K.C. Neokleous, C.N. Schizas / Procedia Computer Science 7 (2011) 244–245 245

. Main

The proposed computational model was built through an appropriate system of dynamical equations that weremplemented and simulated in the MATLAB/SIMULINK environment. The model has two stages of processingmplemented with spiking neural networks (SNN). The first stage simulates the initial bottom-up competitive neuralnteractions among visual stimuli, while the second stage involves modulations of neural activity based on the semanticsf the stimulus. During the progression of the neural activity in the two stages of processing, the encoded stimuli willompete for access to working memory (WM) through forward, backward and lateral inhibitory interactions whichnfluence the strength of their neural response. At the same time, top down interactions can influence the overallrocessing in both stages, depending on their nature. For instance if the top-down signals contain information regardinghe spatial location of a brief visual stimulus (i.e. spatial cues), they will influence the first stage of processing, whilef perceptual cues contain information about the semantics of a stimulus, they will manipulate the processing in theecond stage.

The basic functionality of the model relies on the assumption that an incoming visual stimulus will be processedy the model based on the rate and temporal coding of its associated neural activity. The rate associated with a visualtimulus is crucial in the case of exogenous attention as this has been originally proposed by relevant models. Theseodels are based on the presence of a saliency map, according to which, an image is initially analyzed by distinct

haracteristic maps and then is processed by specific operands inspired by the functionality of the brain (e.g. TreismanGelade, 1980; Koch & Ullman, 1985). The overall process is completed in the saliency map and a winner-take-all

eural network selects the area of the image to which attention is oriented. Endogenous attention is believed to beffected by the synchronization of incoming stimuli with the goals that guide the execution of a task. The presence of alosed link between endogenous attention and synchronization is supported by many recent studies (Gregoriou et al.,009).

The developed model was used for simulating the findings from several behavioral experiments that are well-knownn the scientific literature of visual selective attention. More details about the mathematical representation of the modelnd the specific experiments that have been simulated, along with the model’s fit to experimental data can be found ineokleous et al., 2009a, 2009b, 2009c, 2010.

cknowledgements

This research is supported by grant 0308(BE)/16 from the Cyprus Research Promotion Foundation.

eferences

. Buschman, K. Miller, Top-Down Versus Bottom-Up Control of Attention in the Prefrontal and Posterior Parietal Cortices, Science 30 315 (5820)(2007) 1860–1862.

. Corbetta, G.L. Shulman, Control of goal-directed and stimulus-driven attention in the brain, Nat. Rev. Neurosci. 3 (2002) 201–215.un, Fisher, Object-based visual attention for computer vision, Artificial Intelligence 20 (11) (2003) 77–123.. Horvitz, K. Carl, P. Tim, H. David, Models of attention in computing and communication: from principles to applications, Communications of

the ACM 46 (3) (2003).. Koch, S. Ullman, Shifts in selective visual attention: towards the underlying neural circuitry, Hum. Neurobiol. 4 (4) (1985) 219–227.. Neokleous, M. Avraamides, Marios, C. Neocleous, C. Schizas, A Neural Network Computational model of visual selective attention, Engineering

Intelligent Systems journal (2009a).. Neokleous, M. Avraamides, C. Schizas, Computational modeling of visual selective attention based on correlation and synchronization of neural

activity, in: L. Iliadis, I. Vlahavas, M. Bramer (Eds.), Artificial Intelligence Applications and Innovations III, Springer, Boston, 2009b, pp.215–223.

eokleous, K., Avraamides, M., Neokelous, C, & Schizas, C. 2010 (in press). Selective attention and consciousness: investigating their relation

through computational modelling. Cognitive Computation.

. Neokleous, M. Koushiou, M. Avraamides, C. Schizas, A coincidence detector neural network model of selective attention, Proceedings of the31st Annual Meeting of the Cognitive Science Society, Amsterdam, the Netherlands (2009c).

. Treisman, G. Gelade, A feature integration theory of attention, Cognitive Psychology 12 (1) (1980) 97–136.