A control model of the movement of attention

Abstract

A control model of the movement of the focus of attention is developed and applied to explain its observed effects on single cell activity and to various quantitative features of the Posner benefit paradigm. This supports the presence of an inverse controller and a rules component in the control model. The ability of the control model to explain a range of deficits is then analyzed, as is its relation to other modeling approaches.

Introduction

There is presently great interest in attention from both a psychological and a neuro-scientific point of view. In the psychological approach, a number of paradigms have been used to measure reaction time and success or error levels in human response to determine the nature of attention and how it is controlled (Pashler, 1998a, Pashler, 1998b, Wright, 1998, Parasuraman, 1999). Neuro-scientific techniques have been developed more recently, associated with brain imaging (PET, fMRI, MEG and EEG) or single cell measurements, to find sites in the brain and their dynamical flow of activity as attention is controlled (Kastner and Unegerleider, 2000, Reynolds et al., 1999). The use of brain imaging and single cell recording has now become ever more closely linked to psychological paradigms. At the same time a better understanding of brain processing, including attention, has been obtained by analysis of the deficits brought about by stroke or other brain damage, as well as in cases of mental disability of a neuro-chemical nature. Use of brain imaging for these cases, and comparison with normal subjects, has led to a better understanding of attention.

Attention is seen as composed of a number of separate parts: alerting and preparation for response, vigilance and the holding of attention, and movement of the attention focus. Brain sites active during these various processes are being uncovered by brain imaging, and the detailed topography of the attention movement system, in particular, is now becoming clear, it being composed of at least two separate systems:

• The ‘plant’ (in control terms), consisting of early cortical sites (and associated sub-cortical thalamic and nuclearis-reticularis thalami regions) in which feature analysis and object coding occur. Attention control is now reasonably well accepted as achieving the process of relative amplification of neural activity at the attended sites and reduction of it at unattended places (Reynolds et al., 1999). How extensive is this modulation in cortex is still being determined, both for unimodal input sites and those for output response (Romo & Salinas, 2001).

• The controller of the movement of attention, determining where this relative amplification/reduction occurs (and how), is in separate sites to the plant (Vidyasagar, 1999, Kastner and Unegerleider, 2000). There has been controversy about the site of this function: components are now recognized as in the parietal and frontal lobes (Wright, 1998, Kastner and Unegerleider, 2000). However how control is divided amongst modalities is still undetermined.

It is also controversial that a possible third component is present, acting as a state estimator, in other words as an ‘observer’ or forward model in the control sense (Jacobs, 1993, Franklin et al., 1989). There is a set of prefrontal modules (anterior cingulate, SEF, dorsolateral prefrontal cortex, SMA) which have been noted (Kastner & Ungerleider, 2000 and references therein) as part of the attention control system acting in a manner independent of input. This network could act as an observer in the attention control system to achieve more rapid attention control than by waiting for afferent feedback to arrive at the control signal generator from the amplified signal. At the same time sites in parietal, playing the role of working memory buffers, could also be important components of any such observer system. They would presumably function as part of state updating. However, the existence of an observer is controversial, and ultimately is to be tackled by attention modeling, combined with further experimentation. A preliminary discussion is given by Taylor, 2000a, Taylor, 2000b (see also Taylor, 2001b).

The division of attention into a separate automatic ‘exogenous’ and a controlled ‘endogenous’ part is also now well established (Pashler, 1998b, Wright, 1998, Parasuraman, 1999, Kastner and Unegerleider, 2000). The automatic control component is based on modules sited in the parietal and sub-cortical regions (especially the superior colliculus), whilst the prefrontal and parietal modules mentioned earlier (Kastner & Ungerleider, 2000) are employed for endogenous attention control to be achieved. It is supposed that top-down control is governed by components in working memory, since attention and working memory are realized to be closely linked (Coull & Frith, 1998). To understand attention, we must also improve our comprehension of working memory; as noted earlier it plays an important role in any putative observer system for more efficient control. Evidence has recently been presented for identification of the controllers for endogenous and exogenous attention (Pashler, 1998b); these will be considered in more detail in a later section.

There have been numerous attempts to model attention by neural networks (Mozer and Sitton, 1999, Koch and Ullman, 1985, Lee et al., 1999, Fellenz, 1994, Deco, 1998). Some neural models use a specific control signal, generated by a competitive process on a separate module, to relatively amplify the inputs to earlier sites. There are also attempts to include prefrontal sites as template holders; as such memory-based rule processing is one of the functions of that area (Jackson, Marrocco, & Posner, 1994). Here we will systematically use engineering control theory as a framework for analysis of the attention process as a whole. The purpose of the present paper is thus to present the simulation of a neural model, based on a control framework, for the movement of attention under either exogenous or endogenous control. It will be tested on a range of known single cell, brain imaging and deficit data, with the results of simulations being compared with experimental data.

A control approach to attention is justified by the following.

• Such a framework is important in understanding how the brain enables motor control to be so effective, in particular, so fast (Miall & Wolpert, 1996). Engineering models of control allow functionality to be assigned to brain regions in a manner difficult to understand using the standard neural modeling approach in which it is hoped that distinct functionality would arise automatically by learning in a set of modules initially without specific connections.

• Higher cognitive functions should not be considered separately from other bodily functions. If motor actions on ‘muscular’ parts of the body can be more efficiently viewed under the framework of engineering control theory then the same should be true for other processing performed by the brain in which transformations are performed by one neural system on another—the crux of thinking (Chomsky, 1975).

• The most influential psychological models of attention are those of Posner et al., 1987, LaBerge and Brown, 1989, Desimone and Duncan, 1995. None of these models explicitly uses an engineering control framework, but we will see later that these models are compatible with the control approach we will develop.

Let us consider that last point more fully. We start with the model of Posner et al. (1987), who proposed a dissociation of attention movement into various subcomponents. The sequence of processes involved was suggested as:

$$\text{alert}\text{→}\text{interrupt}\text{→}\text{localize}\text{→}\text{disengage}\text{→}\text{move}\text{→}\text{engage}\text{→}\text{inhibit}$$

Evidence has been presented in a series of papers by Posner and colleagues to support such a dissociation with various brain sites being put forward to support the separate components of the process I (Posner, 1988, Posner et al., 1995). All of these involve control of neural activity by various modules: alerting, interrupting, localizing, moving, and so on. Thus the Posner model incorporates attention movement as controlled by signals from specific modules, consistent with an engineering control approach. The model of LaBerge and Brown (1989) is based on the selection of an area in visual space, using both excitatory and inhibitory mechanisms in a filter module to enable object identification. Subsequent movement of attention is handled by the presence of a processing gradient. The location filter is guided in its selection by a higher order goal system, and thereby biases feature flow to an object module. Yet again the model of LaBerge and Brown (1989) is of control form. Finally the influential model of Desimone and Duncan (1995) appears on the surface not to involve any control structures. As they say (p. 194) “Instead the model we develop is that attention is an emergent property of many neural mechanisms working together to resolve competition for visual processing and control of behavior”. The model uses biased competition between various object representations, as based on related single cell observations in anterior temporal cortex in monkeys performing saccades to object shapes matching previously presented ones among a set of distracters. The bias may be bottom-up, such as brought about in exogenous attention control by sudden onset, or can arise from a top-down goal to observe some target feature such as a specific color or to observe a particular object such as a letter X, or from a spatial location bias. This approach is yet again an example of a control system, contrary to what would appear from the quotation, with attention control signals either being bottom-up or top-down. We will turn to a more complete discussion of the control aspects of these approaches later in the paper.

From the earlier comments, especially in association with the above three influential psychological models of attention, it is reasonable to use engineering control as a general framework inside which to construct a model of attention. This will especially allow quantitative testing of the control approach to attention movement, and more specifically justify the presence or absence of specific components of the overall control model. We start in Section 2 with a description of the basis of the model and its mathematical formulation. In Section 2, we turn to describe single cell data, which strongly support the feedback modulation hypothesis for the basic action of attention on sensory cortical areas. Analysis of the higher order control aspects are then tackled, starting in Section 4 with the discussion of the Posner paradigm for the dependence of valid and invalid cues for direction of attention on values of the inter-stimulus interval. The Posner paradigm allows the mechanisms involved in the shifting of attention to be probed. Detailed simulation results of both the single cell effects and the Posner benefit paradigm are presented in Section 5. The effects on these benefit results of altered inhibition in monkeys as well as certain human deficits are compared with experimental data in Section 6. In Section 7, there is a comparison with other models of attention. The paper finishes with a brief summary of results and a discussion of extension of the model to further paradigms.