Modelling divided visual attention with a winner-take-all network

Abstract

Experimental evidence on the distribution of visual attention supports the idea of a spatial saliency map, whereby bottom-up and top-down influences on attention are integrated by a winner-take-all mechanism. We implement this map with a continuous attractor neural network, and test the ability of our model to explain experimental evidence on the distribution of spatial attention. The majority of evidence supports the view that attention is unitary, but recent experiments provide evidence for split attentional foci. We simulate two such experiments. Our results suggest that the ability to divide attention depends on sustained endogenous signals from short term memory to the saliency map, stressing the interplay between working memory mechanisms and attention.

Introduction

Attention is an old concept in psychology correlated with enhanced processing of objects or regions in space (Posner, Snyder, & Davidson, 1980). While attention is a multi-modal phenomenon (Cherry, 1953; Zelano et al., 2004), the majority of research has focused on selective visual attention (SVA). The limited capacity of the visual system necessitates a mechanism to select stimuli from the visual field, and Tsotsos pointed out that attention solves the complexity problem of sensory processing (Tsotsos, 1992).

A distinction can be drawn between pre-attentive and attentive visual processing (Neisser, 1967). Pre-attentive processing refers to bottom-up (BU) feature saliency of visual stimuli whereby items that differ from their surroundings ‘pop out’ to the viewer. Attentive processing refers to top-down (TD) influences on perception of stimuli determined by object and locational bias such as task instructions or foreknowledge of stimulus characteristics. Determining saliency, then, is both a BU and TD requirement, and computational models of SVA include maps that integrate BU salience across object features (Koch & Ullman, 1985), TD bias (Treisman, 1998), and the interplay of both (Deco, Pollatos, & Zihl, 2002; Wolfe, 1994).

Koch and Ullman (1985) provide a neural network model of SVA in which topographic feature maps are integrated by a winner-take-all (WTA) saliency map of BU stimuli. In their model, inhibiting the selected location causes a shift to the next most salient location. Wolfe (1994) builds on Neisser's pre-attentive/attentive distinction (Neisser, 1967), integrating BU and TD saliency criteria in his Guided Search model. Treisman (1998) provides a model of spatial attention to solve the Binding Problem, in which a TD saliency map determines object features selected for further processing, and suggests parietal cortex as the biological correlate of her ‘master’ map. Deco et al. (2002) use inhibition to mediate BU and TD influences in an instantiation of Duncan and Humphreys' biased competition model (Duncan & Humphreys, 2002), simulating saliency in posterior parietal cortex (PP) with a Continuous Attractor Neural Network (CANN). Spatial saliency in PP interacts with BU feature maps to converge on a winning location. See Shipp (2004) and Itti & Koch (2001) for a review of these and other models.

There is long-standing debate about the distribution of SVA. Many cognitive models propose a unitary focus of attention, likened to a roving spotlight over the visual field (Posner et al., 1980). Variants of the spotlight metaphor include gradient (Downing & Pinker, 1985; LaBerge & Brown, 1989) and zoom lens (Eriksen & James, 1986) models, suggesting that attention may be a graded phenomenon, attenuated around a central focus. A large body of evidence supports such unitary models (McCormick, Klein, & Johnston, 1998; Posner et al., 1980), but several more recent experiments have provided evidence for non-contiguous allocation of SVA (Awh & Pashler, 2000; Hahn & Kramer, 1998; Muller, Malinowski, Gruber, & Hillyard, 2003).

Here, we study how split attention can be achieved by a dynamic implementation of a WTA map. Despite their WTA nature, CANNs are able to account for split attention when network dynamics facilitate long transition states between regimes (Trappenberg & Standage, 2005) and when dominated by sustained inputs (Standage, Trappenberg, & Klein, 2005). We simulate the experiments of Muller et al. (2003) with a 1-dimensional (1D) CANN model. We build on simulations presented in Standage et al. (2005) that use a narrow weight profile, facilitating steeply sloped regions of activity that occupy a small portion of the network. Because we do not know the size of the active region of PP and its relation to coordinates in the visual field, we run similar experiments with a wide weight profile, resulting in activity that spans the majority of the network. We demonstrate that the ability of the model to account for divided attention does not depend on fine tuning this network parameter.

We simulate two experiments by Awh and Pashler (2000) with a 2-dimensional (2D) CANN model, demonstrating how the model accounts for their finding divided attention in one experiment and unitary attention in the other. Preliminary simulations in 1D are reported in Standage et al. (2005). Our simulations are consistent with their experimental findings, but our model offers an alternative conclusion.