VISUAL COGNITION, 1996, 3 (2), 119–164
Inhibition and Interference in Selective Attention: Some Tests of a Neural Network Model George Houghton University College London, UK
Steven P. Tipper University of Wales, Bangor, UK
Bruce Weaver and David I. Shore McMaster University, Hamilton, Ontario, Canada We describe a neural network system that models selective action—that is, how an organism selectively responds to an object when other objects evoke competing responses. Performance of the model during simulations of various selective action situations reveals a number of interesting patterns of data. Specifically, the model shows a complex relationship between how much a distractor interferes with response to a target and how much inhibition is associated with the distractor. Subsequent experiments with human subjects reveal that the paradoxical behaviour of the model is also observed in human behaviour. We conclude that the similar performance characteristics of the model and human subjects in a variety of situations suggest that the model has captured some of the essential properties of mammalian selective action mechanisms.
Imagine a fictitious organism whose only aim in life is to consume peanuts placed in front of it by reaching out with a single effector, capable of holding one peanut at a time, and carrying that peanut to its mouth. The organism is equipped with a sensory system that reliably detects, and internally represents, the positions of any peanuts placed within some finite area in front of it. Information derived from this sensory system is used by the organism to plan and execute targeted movements of its effector, so that it can successfully grasp a peanut placed anywhere within reach. We may ask how this organism should behave if it is presented with two peanuts simultaneously, both of which are analyzed in parallel. To maximize its benefit, the organism must consume both peanuts. However, it can only reach and grasp one at a time. To do this, the organism’s Requests for reprints should be sent to Steven Tipper, Department of Psychology, University of Wales, Bangor, Gwynedd, LL57 2DG, UK. (E-mail:
[email protected]). © 1996 Psychology Press, an imprint of Erlbaum (UK) Taylor & Francis Ltd.
120
HOUGHTON ET AL.
response must be guided each time by information from a single, selected, peanut. Otherwise, the combination of information from both peanuts will lead the effector to miss both, ending up, say, somewhere in-between them. At the same time, information from the non-selected peanut should continue to register, or the organism might omit to make a second response at all. Such a selection situation is not at all fanciful and has been described in the frog (Ingle, 1972, 1975). When a frog is presented with two prey objects, response latencies are substantially longer than when there is only one prey object. In this paper we consider the problem faced by such an organism to be the problem of selective attention. In short, we ask: How is it that individual responses can be selectively controlled by particular stimuli in the presence of competing inputs? A major part of our answer to this question involves what happens to the information derived from the competing, or “distractor”, stimulus during execution of a response to the selected, or “target”, item. It is our contention that active inhibitory mechanisms, globally distributed throughout the processing substrate, selectively act on representations activated by the distractor input. This inhibition is responsive to the state of activation of these representations and sufficiently attenuates their influence on response systems that actions can be coherently guided by information from the target stimulus alone. In this paper we develop a detailed model, implemented as a neural network, of the architecture and dynamics of these putative inhibitory mechanisms and describe a series of experiments aimed at testing some specific predictions derived from the model. The plan of the rest of the paper is as follows: First we consider the immediate empirical background to the model, focusing on the issue of the relationship between inhibition and the control of interference. We next describe a neural network model of selective attention (Houghton & Tipper, 1994), which employs a combined excitatory/inhibitory architecture. A series of predictions regarding the phenomena of interference and negative priming (Tipper, 1985) are then derived from a simplified version of the model, which focuses on the operation of one particular aspect of it. Finally, we describe a series of experiments aimed at testing these predictions.
INHIBITION AS A MECHANISM OF SELECTION: EMPIRICAL BACKGROUND AND THEORETICAL ISSUES The basic idea that the brain depends for its self-control on inhibitory mechanisms, with selective attention being one aspect of endogenous control, has a long history within psychology and the neurosciences. In the earlier part of the twentieth century the concept of specific (selective) inhibition played a central role in the work of such psychological theorists as Pavlov (1927) and Hull (1943). For instance, Pavlov (1927) developed the notion of “internal inhibi-
INTERFERENCE AND INHIBITION
121
tion”, whereby processes initiated in the brain by a conditioned stimulus would come under the control of inhibition, leading to a selective reduction of responsivity (see Kimble, 1968, for discussion). These ideas were developed further in the 1960s, notably by Douglas (1967) and Kimble (1968), in the context of hippocampal functioning. Based on data from a large number of hippocampal lesion studies (mainly in rats), Douglas (1967) developed a model of selective attention based on inhibition of goal-irrelevant (or non-reinforced) inputs. Animals faced with stimuli evoking conflicting response tendencies were able to generate coherent, directed behaviour by inhibiting the sensory representation of distractor stimuli. Hippocampal lesions reduced the efficacy of this goal-directed inhibitory filtering, leading to disruption of many simple behaviours (e.g. habituation, extinction, discrimination reversal, spontaneous alternation: Douglas, 1967; Kimble, 1968). Similar ideas can be found in the literature on attentional deficits in schizophrenia (Beech, Powell, McWilliams, & Claridge, 1989; Frith, 1979; Gray, Feldon, Rawlins, Hemsley, & Smith, 1991; Peters, Pickering, & Helmsley, 1994) and ageing (Hasher & Zacks, 1988; Tipper, 1991), where it is suggested, for example, that schizophrenic subjects suffer a deficit in preventing “irrelevant” information from gaining access to awareness and controlling action. This loss of inhibition leads to an inability to maintain coherent trains of thought and action, as irrelevant environmental inputs and internally generated associations compete with goal-relevant stimuli for high-level processing. Data from subjects with certain types of prefrontal brain lesions similarly lend themselves to discussion in terms of inhibitory mechanisms. For instance, Fuster (1980) develops a theory of prefrontal lobe function in which one of its functions is the control of interference of ongoing behaviours by irrelevant inputs, generated either internally or externally (see also Diamond, 1991). This control is postulated to be inhibitory and has its functional centre in the orbital region of prefrontal cortex. These functional links between schizophrenic and prefrontal syndromes receive further support from the work of Shallice, Burgess, and Frith (1991). These authors tested five chronic schizophrenic patients (of varying overall ability levels) on tasks known to be sensitive to frontal-lobe lesions. All five subjects performed poorly. Anatomical evidence for the link comes from recent PET scan studies by Liddle et al. (1992) and Friston, Liddle, Frith, Hirsch, and Frackowiak (1992). Data from 30 schizophrenics, grouped according to symptomatology, provided clear evidence for an association between symptoms of behavioural disorganization and altered functioning of distinct prefrontal regions. Despite the widespread discussion of inhibitory mechanisms of selective control in such neuropsychologically oriented work, the “informationprocessing” approach to selective attention prevalent in experimental psychology has shown a relative reluctance to appeal to inhibitory mechanisms (see Neill & Westberry, 1987, for discussion). Rather, the major emphasis has
122
HOUGHTON ET AL.
been that the selection of target from distractor objects is achieved by amplificatory, or excitatory, mechanisms. The most popular analogy is that of a spotlight (e.g. Broadbent, 1982; Posner, 1980) or zoom lens (e.g. Eriksen & St. James, 1986). The spotlight is conceived of as moving through space and alighting on particular objects. Objects within the beam of the spotlight receive processing beyond that of initial perceptual analysis. In contrast, the internal representations of distractors are assumed to decay passively (see van der Heijden, 1981). This situation has been somewhat altered in recent years, with a number of authors working in the information-processing tradition arguing for the involvement of active inhibitory mechanisms in a variety of tasks involving “cognitive control”. (For a wide range of examples, see the collections of articles edited by Dagenbach & Carr, 1994, and Dempster & Brainerd, 1995. Also, Arbuthnott, 1995, reviews both data and models from a number of relevant areas.) With respect to the use of inhibition in selective attention, the negative priming paradigm has been especially influential. In this paradigm, it is argued that if the internal representations of an ignored object are subject to inhibition during selection and execution of a response to a target object, then processing of a subsequent object requiring the inhibited representations should be impaired. For example, in the procedure developed by Tipper (1985), subjects responded to a red drawing while ignoring a green distractor in a prime display. On some trials, the distractor (e.g. a dog) from the immediately preceding prime trial could be presented as the target to be identified in a subsequent probe trial. It was predicted that in these cases responses would be slower than to the same target in a baseline condition, in which it did not previously appear as a distractor. This result was observed, and has been replicated in numerous studies since, using a variety of stimuli and response types. Indeed, an analogous phenomenon was observed as far back as 1966 by Dalrymple-Alford and Budayr (1966) in a Stroop colour naming task. These results have been taken by Tipper and his colleagues to support the idea of active inhibitory processes in selective attention. In earlier work (Tipper & Driver, 1988; Tipper, MacQueen, & Brehaut, 1988), it was argued that the inhibitory mechanism had a “central” locus, acting on semantic representations mediating between perception and response. More recently, it has been argued that the locus of inhibition may be rather more flexible, reflecting the behavioural requirements of the current task (Tipper, Lortie, & Baylis, 1992; Tipper, Weaver & Houghton, 1994). In the present context it is important to note that there have been recent interpretations of negative priming that do not invoke inhibition mechanisms. Neill and Valdes (1992) and Park and Kanwisher (1994) have argued that negative priming may reflect a memorial retrieval process. Basically, onset of the probe triggers retrieval of the previous distractor. Mismatching information between the prime distractor and probe target, such as identity (Park & Kanwisher, 1994)
INTERFERENCE AND INHIBITION
123
or response association (Neill & Valdes, 1992), slows response to the probe. For example, the prime distractor may be tagged as “do not respond”, but when the same object appears in the probe, it requires a response. We feel that the discovery of such retrieval processes is important. Indeed, we (Milliken, Tipper & Weaver, 1994) have also observed that retrieval of the prime distractor can produce negative priming effects. However, such retrieval accounts of negative priming, although interesting, do not affect our current discussions, for two reasons: (1) Discovery of a new process, such as retrieval of previously ignored information, does not necessarily imply that some other process, such as inhibition, no longer exists. (2) There are many experimental results that are compatible with inhibition mechanisms, but which cannot be accounted for by retrieval models (see Milliken et al, 1994; Tipper & Milliken, 1996). If selection is indeed achieved by inhibiting distractors, then there should be some kind of predictable relationship between difficulty of selection and the amount of inhibition required to select efficiently. Unfortunately, studies examining the relationship between the interference caused by a distractor and the amount of inhibition associated with that same distractor have provided very mixed results. Across a variety of studies, all possible relationships between interference and negative priming have been observed. For example, a negative relationship, where increases in distractor interference are accompanied by declines in negative priming (inhibition), has been obtained (see Fox, 1995; McDowd & Oseas-Kreger, 1991; Tipper, 1991; Tipper & Baylis, 1987). On the other hand, a positive relationship, where increases in distractor interference are accompanied by increases in negative priming, has also been observed (see Neill, Valdes, & Terry, 1995; Tipper et al., 1992; Meegan & Tipper, in preparation; Milliken et al., 1994). Finally, other studies have, in fact, suggested that there is no relationship between interference and negative priming (Driver & Tipper, 1989; Kane, Hasher, Stolzfus, Zacks, & Connelly, in press; Tipper, Weaver, Cameron, Brehaut, & Bastedo, 1991; Tipper, Weaver, Kirkpatrick, & Lewis, 1991). Thus a simple relationship between interference and negative priming can be very elusive. The central purpose of the current paper is to show how different predicted relationships between these variables emerge from a model of selective attention (Houghton & Tipper, 1994). These predictions are then tested in a series of experiments. Any model that aims to elucidate the relationship between inhibition and interference must, of course, provide a detailed account of the operation of inhibitory processes. We are bound to note that this is almost never done. In the majority of the works cited above, which postulate the existence of inhibitory control mechanisms, hardly any specific properties of these mechanisms are adduced. Generally little more is said than that some form of inhibitory control exists and that it acts to reduce the effect of competing or interfering inputs or response tendencies. In order to advance theoretical discussion beyond this
124
HOUGHTON ET AL.
point, we propose that any plausible model of inhibitory control must provide answers to at least the following questions: 1 1. What causes the activation of inhibitory processes, and how do they know what to act on? For instance, in selection tasks, how does the selection mechanism register the presence of distractors, and how does the inhibitory process selectively access their representations? 2. What effect do inhibitory processes actually have on distractor representations? For instance, does the claim that “distractors are actively inhibited” mean that information from non-target stimuli is simply not available to the organism? 3. Are inhibitory mechanisms entirely top-down, simply radiating out from some “central inhibitor” towards activated representations, or can they adaptively respond, in a bottom-up fashion, to the state of activation of distractor representations? We believe that how one answers these questions is important for understanding phenomena of the type discussed above. For instance, in relation to the negative priming effect, one would anticipate that the answer to Question 2 would be that distractor representations must be suppressed below background levels of activation, even when the distractor is present. Otherwise, how would responding to a later probe be impaired? However, this idea is problematic. Intuitively, obliteration of the unattended field would generate a “tunnel vision” effect. Targeted actions in complex environments could not adapt to the presence of multiple objects (lying in the path of a movement trajectory, for instance). This suggests that inhibition of distractors may be more graded, attenuating their influence sufficiently to allow coherent responding to a target but not suppressing them below background. In this case, though, there is no obvious reason to anticipate future impaired processing of these inputs, as although “inhibited”, they were still more active than the representations of stimuli not actually present. In the following section we describe a neural network model (Houghton & Tipper, 1994), which represents our first attempts at answering these questions. The architecture and dynamics of the inhibitory processes in the model are somewhat more complex than those suggested by informal discussions of inhibitory mechanisms. However, we are able to derive a number of predictions from these mechanisms. In this paper we describe a series of experiments testing these predictions. 1 We do not intend to suggest that these are the only important issues. Other questions of importance involve the top-down modifiability of inhibitory processes, for instance by the organism’s current goals (Tipper, Lortie, & Baylis, 1992; Tipper, Weaver, & Houghton, 1994) or by more general internal variables, such as state of arousal or vigilance .
INTERFERENCE AND INHIBITION
125
A MODEL OF INHIBITORY MECHANISMS IN SELECTIVE ATTENTION The model discussed here is the Houghton and Tipper (1994) selective attention model. The first part of the discussion is largely informal, concentrating on the overall architecture of the model and the way in which it proposes that selection takes place. Following this, we focus on one particular aspect of the model architecture—the use of local excitatory/inhibitory feedback mechanisms to modulate the activation levels of perceptual representations. A “reduced model” is derived, which allows us to study these components independently of other parts of the model. Particular properties of these mechanisms are illustrated by computer simulation.
An Architecture for Selective Attention The basic premise of the model is that attention operates in co-ordination with the animal’s current behavioural goal to control the interaction of perception and action; in particular, it operates at the interface between largely parallel perceptual processes and serial behavioural processes, selecting out information from the perceptual array relevant for guidance of current actions. In the absence of effective selection, an animal becomes highly distractible, manifesting debilitating interference from irrelevant environmental stimuli (e.g. Fuster, 1980). The overall organization of the model is shown in Fig. 1. The model proposes that selection involves the interaction of a number of functionally (and anatomically) separate systems, referred to as “fields”. Visual perceptual information, considered to be grouped into assemblies corresponding to objects, is represented in the object field. This information is processed in posterior cortical areas (occipital/parietal/temporal cortex). Information from this system feeds forward via learned or innate connections to response systems such as the Supplementary Motor and primary motor cortex, where one central function is to bind the parameters of variable action schemas—e.g. to specify the target location of a reaching action (Arbib, 1990). Selection is postulated to involve the interaction of internally activated target descriptions, represented in the “target field” of the Prefrontal cortex (Diamond, 1991; Goldman-Rakic, 1988), with the representations of external stimuli activated in the object field of the posterior cortex. The internal target representations are assumed to be derived from internally generated goals and plans, which, in parallel, are responsible for the activation of appropriate action schemata. Selection of objects that fit target specifications is achieved via a matching process (match/mismatch detector) possibly taking place in the Pulvinar (LaBerge, Brown, Carter, Bach, & Hartley, 1991) or Basal ganglia (Jackson & Houghton, 1994), which takes input from both the posterior object fields and the anterior target field. Matching is made on the basis of particular features (e.g. colour, location) and objects containing features that
126
HOUGHTON ET AL.
FIG. 1. A schematic view of a model of selective attention (Houghton & Tipper, 1994). Controlled interactions between perception (“object field”) and action (“response system”) require both response set (maintenance of currently active schema) and perceptual set (selective gating of perceptual information).
match the target become selectively enhanced (their activation level is increased), whereas mismatching objects have their activation levels suppressed. The opening up of this activation gap between targets and distractors allows task-relevant information from target inputs to bind coherently the parameters of the currently active action schema (variable binding system, Fig. 1). In the model, each object encoded in the object field is represented as a connected set of processing elements (nodes), which represent all aspects of the object’s automatic, pre-attentive encoding, which is taken to include categoryrelevant (“semantic”) features as well as contingent properties such as current location and details of physical form. These “property nodes” are linked to a gain-control subsystem, consisting of two feedback loops, one excitatory and
INTERFERENCE AND INHIBITION
127
FIG. 2. In the model, “gain-control” circuits are distributed in topographic fashion throughout the representational substrate. Activity in a property unit (representing the presence of some feature in the input) activates two balanced feedback circuits, one excitatory (the On-channel), one inhibitory (the Off-channel). The activity level of the property cell can be suppressed or enhanced by changing the balance of activity in these feedback channels. The symbols w1, w2, in the figure are free parameters representing the strength of the feedback connections.
one inhibitory (Fig. 2). These feedback loops act via excitatory and inhibitory units, to which we refer as on-cells and off-cells, respectively. When a property node is activated, it equally activates both the on- and offcells shown in Fig. 2. In the general case, the weights w+, w- on the feedback pathways are set to be of equal magnitude but opposite in sign (though simulations involving asymmetrical values are reported further on). Summed together, the net effect of this feedback is zero. However, as described later, the activity of the on- and off-cells can be independently modulated, leading to the feedback having a net positive or negative effect. Because of the proposed independence between these two feedback systems, selection efficiency cannot always be predicted from the activity of either one of them. It is only by knowing how both systems are functioning that clear predictions can be made regarding the relationship between inhibition and interference. Simulations demonstrating how changes in levels of distractor inhibition can appear to be uncorrelated with selection efficiency are described further on. The model is capable of selecting or ignoring whole “objects” on the basis of partial descriptions of to-be-attended items (e.g. the target’s colour or location). This is made possible by assuming that object “assemblies” consist of linked sets of the circuits, as discussed earlier. Links are formed, so that, within an assembly, all on-nodes are excitatorily linked, and all off-nodes are excitatorily linked. Nodes of opposing sign within an assembly inhibit each other. The exist-
128
HOUGHTON ET AL.
ence of links between the on-nodes in an assembly permits the usual process of spread of activation, so that increased on-activation in some component of an object or perceptual group will spread to other linked components. Conversely, the use of excitatorily linked opponent (off) pathways means that increased activity in one off-channel can spread throughout the other off-cells in the assembly. This spread of activation in opponent channels will register on linked property nodes as spread of inhibition. Thus, if some component of an object representation becomes inhibited due to heightened activation of its off-cell, this suppression will tend to spread, via the linked off-channels, to the whole assembly. 2 Houghton and Tipper (1994) point out that although it could be claimed that a dual mechanism of selection is unparsimonious, there are at least two reasons why dual mechanisms are essential: First, any amplificatory or inhibitory mechanism functioning in biological (neural) hardware can only operate within finite limits, and thus the rate at which one signal can be boosted relative to another stable signal must have some finite upper bound. In comparison, a dual mechanism that can boost a target and inhibit a distractor signal in parallel can double the rate at which target and distractor can be separated. Such rapid selection is clearly vital for interactions with complex, and often dangerous, environments. The second reason for dual mechanisms is that any signals in a biological information processing system must have a limited dynamic range—maximum and minimum amplitudes. Therefore, if two intense stimuli were present, neural responses to them would be close to the maximum firing rates for those cells. Thus selection with only an excitation mechanism would be impossible. On the other hand, two very weak stimuli close to threshold may not be separated by a single inhibition mechanism. Two mechanisms (excitation and inhibition) working in parallel would always be able to separate target and distractor signals, whatever the neural firing rate. We note that Posner and Dehaene (1994) have also recently made this point. They claim that an inhibitory component to selection is crucial because neurons in early visual areas have been found to be activated to a near-optimal level by visual stimuli. This is the case not only during passive viewing, but even in the anaesthetized animal. Therefore there is little room for enhancing the firing rate of these cells, whereas selection via inhibition is very effective. 2 Subsequent research has supported this prediction by providing evidence that the whole object assembly is associated with inhibition. Thus, the more properties in common between an ignored object and subsequent probe, the greater is the inhibition (Tipper et al, 1994, Experiment 3). However, it is important to note that other data suggest that this spread of inhibition does not always take place. Rather, inhibition can be selectively associated with particular properties of an object (e.g. its location or identity), depending on task demands (Milliken et al, 1994; Tipper et al, 1994). Future work will modify the model to account for this flexible control of inhibition. In the current context, however, the possibility of inhibition being associated with particular object dimensions does not affect the empirical testing of the model, as the reported experiments involve aspects of distractors that would be inhibited under either theory.
INTERFERENCE AND INHIBITION
129
Further support for the idea of two independent mechanisms acting in selective attention is provided by recent studies of evoked potentials. Luck et al. (1994) have provided evidence that selection of visual targets can be achieved by excitation and inhibition mechanisms similar to those we describe in this article. They show that in tasks where subjects have to detect the luminance onset of a target that may be validly or invalidly cued, two mechanisms can be observed in extrastriate cortex: One is excitation associated with cued locations, and the other is inhibition associated with invalidly cued locations (compared to neutral cueing conditions). Of greatest relevance here, these two mechanisms appear to be separate and dissociable. Thus, the posterior N1 component of the evoked potential is facilitated for the attended location, whereas the P1 component is unaffected; the P1 component is inhibited for the ignored (i.e. invalidly cued) location, but the N1 component is unaffected (see also Luck, Heinze, Mangun, & Hillyard, 1990). Luck et al. (1994) argue that these independent P1 and N1 components demonstrate increases in “neural responses for stimuli presented inside the focus of attention and decreases in these responses for stimuli outside the focus of attention” and that “separate mechanisms {are} responsible for suppressing information arising from the unattended locations, and for enhancing the processing of information at the attended location” (pp. 897–902). These conclusions are entirely in accordance with the predictions of the current model. This aspect of the model, with separate excitation and inhibition mechanisms, is crucial for understanding the relationship between measures of inhibition and interference considered in this paper. To see a strong relationship between interference and negative priming, one must assume that inhibition is the primary— and perhaps only—mechanism of selection. Most models propose at least one other mechanism, which is excitation directed to the internal representations of the target (see the spotlight or zoom lens models of Eriksen & St. James, 1988, and Posner, 1980). Thus, even though inhibition, as observed by negative priming, may appear to be lower in some subjects, selection efficiency could be maintained by boosting the intact excitation system. There are, of course, drawbacks to the study of such dual mechanism systems. Unless one has complete experimental control over both mechanisms, it is impossible to make unequivocal statements about the relationship between one of them and the efficiency of selection. Indeed, the simulations to be described explicitly show how observing a clear relationship between interference and negative priming, for example, is very difficult. The excitatory and inhibitory feedback systems in the model are independent and are postulated to reflect different anatomical systems utilizing different neurotransmitters (see Houghton & Tipper, 1995, and Jackson & Houghton, 1994, for further discussion of this point). These systems are also likely to be differentially affected by changes in global system parameters, such as degree of arousal or vigilance (Steriade, 1984).
130
HOUGHTON ET AL.
FIG. 3. Typical time course of activation for the internal representations of a simultaneously presented attended stimulus (dashed line) and ignored distractor (full line) in the Houghton and Tipper (1994) model. The time of onset and offset of the stimuli is shown at the bottom of the figure.
An important consequence of the way in which inhibition is generated in the model is that, although property nodes in the object field representing distractor inputs continue to receive external input (i.e. while distractors are still in view), their activation levels do not drop below resting levels (Figure 3). This continued activation allows information from the background to be still available, either to influence ongoing processing or to be monitored for significant events. It may appear from this that, if inhibited distractor representations are allowed to settle at activation levels above resting-level, then we cannot account for negative priming. If an ignored distractor in a priming trial still achieves active internal representation, why should responses to the same item as a probe target be retarded? The model resolves this paradox by exhibiting rebound behaviour at stimulus offset. When the distractor input is terminated, the balance between the external excitation and internal inhibition of the property nodes (represented by the equilibrium activation level) is upset, there now being only the internal inhibition. This pushes the activation level of these nodes into the
INTERFERENCE AND INHIBITION
131
negative region, representing a suppressed, or sub-baseline, level of activation. If the previous distractor (or a similar item) is re-presented as probe target while in this rebound phase, selection is impaired (in the presence of a distractor) relative to a novel probe. On the assumption that similar objects share property nodes, the model predicts greater negative priming the more similar are ignored prime and probe (see note 2). In the previous section, we set out three basic issues that we felt any model of inhibitory selection mechanisms must address. The answers the model gives to the questions raised there are as follows: 1. What causes the activation of inhibitory processes, and how do they know what to act on? Inhibitory processes are activated in the first instance by the activation of what they (may) oppose—i.e. it is postulated that any externally driven activation of perceptual representations generates a “topographically organized” inhibitory response, which precisely mirrors the initial activation pattern. However, the inhibition may be masked by excitatory feedback. A matching process involving an internally generated selection template can remove this excitatory masking for non-matching feature dimensions, leading to selective suppression of distractors. Thus, in this model, selection is affected by changing the balance of the positive and negative feedbacks, basically by inhibiting one of them. When template and perceptual input match, inhibition is turned off; when they mismatch, excitation is turned off. So selection is, on one level, a purely inhibitory process, where different feedback systems are turned off. It simply manipulates processes in perceptual networks that have taken place before selection begins. 2. What effect do inhibitory processes have on distractor representations? While the distractor itself is present, the inhibition acts to dampen the activation of its representation, but it is not suppressed below background (information from the distractor is still potentially available to support actions such as reaching around the distractor). Given abrupt offset of an unattended item, the inhibitory feedback continues to act for some time after offset, causing the distractor representation to suffer an inhibitory rebound. 3. Is inhibition generated by a top-down “central inhibitor”? No, inhibition is generated locally in a largely bottom-up fashion. Top-down processes act to modulate and focus the inhibition. In the following section we derive some testable predictions from these model properties.
132
HOUGHTON ET AL.
The Relationship Between Inhibition and Interference in the Model We now consider some particular properties that emerge from this model. As discussed, the relationship between the amount a distractor interferes with response to a target and the amount of inhibition associated with the distractor is extremely complex. Because consistent associations between the two phenomena have not always been found, it has generally been assumed that they are caused by separate/dissociated processes. That is, because inhibition does not always seem to be linked to the level of interference, then negative priming is unlikely to reflect a mechanism of selection. Such a conclusion can only be reached, however, if a particular, unrealistic, view of selective attention is held. That is, a clear and consistent relationship between interference and inhibition (negative priming) can only be predicted if inhibition is the primary—perhaps the only—mechanism of selection. As discussed earlier, we feel that this is highly unlikely in biological informationprocessing systems where complex interactions between excitation and inhibition are the norm (see also Driver, McLeod, & Dienes, 1992; Posner & Dehaene, 1994; Treisman & Sato, 1990). What we intend to demonstrate is that our formal model can capture and explain many of the diverse patterns of data observed from human subjects in a variety of experiments. Specifically, the model predicts that consistent correlations between interference and negative priming can be very elusive, even though inhibition is a mechanism of selection that influences levels of distractor interference. In the model, the “selection state”, whereby a given perceptual input is allowed privileged control over action, is distributed over a number of separate, interacting systems (Fig. 1). Disturbances to the selection state and its efficacy could be due to disruption of any of these subsystems. For instance, failure to maintain the selection template active in “working memory” properly would lead to failures of selection due to the matching process receiving degraded topdown input. Disruption of the feedback loops in the perceptual systems would lead to slower and less effective separation of competing inputs. We believe that such a distributed architecture for attention is justified on computational grounds and that it also has neurophysiological support (Houghton & Tipper, 1994, 1995; Jackson & Houghton, 1994; Posner & Peterson, 1990). However, it clearly complicates understanding and testing of the model. To do this, it is useful to break the model into its major components and treat each separately, assuming the proper functioning of other components (even using approximate simplifications where necessary). A distinction can be made in the model between its “top-down” and “bottomup” components. By its top-down components we principally mean the activation of the selection template or target description and the role of this input in the matching process (Fig. 1; top-down components also include the activation
INTERFERENCE AND INHIBITION
133
of the relevant response and possible influences of this “response set” on the selection process, Tipper et al., 1994). By bottom-up components we mean the postulate that all activated representations give rise to the activation of local excitatory and inhibitory feedback loops (Fig. 2). In other work, we have attempted to test the top-down components of the model, investigating what happens to negative priming and interference effects in selection tasks as a function of whether subjects are or are not provided with explicit target descriptions (Baylis, Tipper, & Houghton, submitted). In this paper, we concentrate on testing the bottom-up aspects of the model, again looking at interference and negative priming in selection tasks. In particular, we aim to investigate the idea that attention depends on the combined influence of both inhibitory and excitatory mechanisms, and that these mechanisms are responsive to the state of activation of the perceptual representations on which they act. One way to derive predictions from the model regarding the behaviour of the components in question is to use full simulations of the model while varying relevant components in experimentally interpretable ways. A variety of such simulations are reported in Houghton and Tipper (1994). However, as noted, the complex architecture of the model means that it is not always obvious to what aspect of the model one should attribute a particular simulated property. This can lead to a situation in which one effectively has simulation without explanation. The alternative to this is to use a reduced model in which the components of interest are allowed to vary while their interactions with other components are either held constant or simplified in a way that does not significantly alter the behaviour of the relevant parts. This is the strategy adopted in the current paper, in which we isolate and analyse the “bottom-up” excitatory-inhibitory feedback loops in the model. The behaviour of this subsystem, with the top-down inputs “assumed”, is sufficiently close to the behaviour of the full model in the respects that concern us that it can be used on its own to generate predictions. This has the advantage that, in the reduced model, it is relatively easy to see why the model has the properties it does. Indeed, most of the properties we deduce herein can be derived analytically, given a suitable choice of basic equations (see Houghton, 1994). We develop the reduced model from the full model by making two simplifications.
Simplification (1). In Houghton and Tipper (1994), objects are represented as distributed patterns of activation over sets of feature nodes. Nodes representing features belonging to the same object are “grouped” in that they come linked together. The consequence of this linkage is that nodes within a group become correlated in terms of their activation values and their rate of change, so that if one node starts to increase its activation, so do the others. Conversely, due to a “spreading inhibition” mechanism, if one node in an object assembly becomes inhibited due to increased activation of its associated off-node, then
134
HOUGHTON ET AL.
other features of the object follow suit. In many of the simulations in Houghton and Tipper (1994), this grouping effect means that the effect of attention on object representations is best shown by taking the mean activation value of all the nodes representing a given object, as they all follow essentially the same time course. One simplification that can, therefore, be made is simply to represent each object by a single node, considered to represent the mean activation value of an object group. This simplification abstracts away from the complications due to the spread of excitation/inhibition through distributed assemblies. However, this feature of the model is not important to the properties investigated here, where only the mean activation values of the features of competing representations are important. In the simplified model, we shall therefore represent each object by a single node. Where there are only two objects in a display, there will only be two representational nodes active in the object field (Fig. 1).
Simplification (2). In Houghton and Tipper, the basic balance in the excitatory/inhibitory feedback system is perturbed by input from the match/mismatch system, in which properties of perceptual inputs are compared with the selection template. Items containing matching features are enhanced by removal of inhibitory feedback. Items failing to match are inhibited by removal of excitatory feedback (unmasking of inhibition). When we represent a basic selection task in the simplified model, we shall assume that for one of the two activated object field nodes (see Simplification 1) the inhibitory feedback loop has been completely inhibited. This represents the attended item. For the other input, the excitatory feedback loop has been completely inhibited. This represents the distractor. The combination of these simplifications means that, for a simple selection task in which one input must be selected over another, we reduce the behaviour of the model to two simple circuits, one for the attended item, and one for the
FIG. 4. In many cases, the time course of activation of attended and ignored inputs in the Houghton and Tipper model can be formally generated by two simple circuits - a positive feedback loop (attended item) and a negative feedback loop (ignored item). This simplified model abstracts away from complications due to such factors as the use of distributed representations and the precise nature of the interaction between current perceptual inputs and “working memory” representations of to-be-attended items.
INTERFERENCE AND INHIBITION
135
distractor (Fig. 4). Which item is to be attended and which ignored is arbitrary, and we simply assume that the “attended” input has matched the selection template, leading to the suppression of its inhibitory feedback. Thus the attended item comprises a positive feedback loop from an “object” unit via an “on-cell” and back. Likewise, the “distractor” input is taken to mismatch the template, leading to suppression of its excitatory feedback. The distractor item, therefore, comprises a negative feedback loop from an object unit via an “off-cell” and back. The dynamical behaviour of these two simple circuits closely mimics the behaviour of whole object assemblies when embedded in the full model. This is because object assemblies are simply groups of these circuits, each element of which obeys the same dynamics.
Formal Description of the Simplified Model In the following simulations, the activations of all nodes are governed by a “leaky integrator” activation rule, with activations allowed to vary continuously in the range {2 1, 1} and with a resting activation of 0 (as in Houghton & Tipper, 1994). The particular form of the activation rule is one used in Houghton, 1994. This rule can be expressed as a differential equation:
dai 5 dt
2 Dai 1
(1 2
ai)I + 2
(1 1
ai)I –
(1)
where ai is the activation level of unit ui, dai/dt is its rate of change of activation, D is the passive decay rate, I + is the summed excitatory input to ui, and I – is the summed inhibitory input to ui. The terms 1 6 ai act as “gain control” terms keeping the activation levels of the nodes bounded in the region {2 1, 1}. For instance, as the activation level of a unit approaches its maximum value, excitatory inputs have less effect and inhibitory inputs have more effect. The function has a resting level activation of 0, and, in the absence of input, will return to that level at a rate governed by D. The separation (and separate modulation) of excitatory and inhibitory inputs to a unit (rather than conflating them into a “net input”, as is more usually done) aids analysis of the model and is especially useful in the current context, where we are concerned with the relationship between independent excitatory and inhibitory mechanisms. The overall form of equation (1), in particular the separation of the excitatory and inhibitory inputs, is based on proposals by Grossberg (e.g. Grossberg, 1980). In the simplified model, the activation function for an attended stimulus (Figure 4) is given by, datar ex t + (2) 5 2 Datar 1 (1 2 atar){I 1 w aon } dt
136
HOUGHTON ET AL. ext
where atar is the activation of the attended (target) node, I is its external input, + aon is the activation of the associated on-cell, and w is the excitatory feedback weight. Note that it is assumed that for the attended item the off-cell is completely inhibited. The activation function for the distractor is given by,
dadis 5 dt
2 Dadis 1
(1 2
adis)I
ex t
2
(1 1
(3)
–
adis)w aoff ext
where adis is the activation of the unit representing the distractor, I is its – external input, w is the inhibitory feedback weight and aoff is the activation of the associated off-cell. Note that the distractor receives no excitatory feedback, i.e. its associated off-cell is taken to be completely suppressed. The activation functions for the on- and off-cells shown in Fig. 4 are given by equations (4) and (5) respectively. daon (4) 5 2 Daon 1 (1 2 aon)w1 {atar}+ dt
daoff 5 dt
2 Daoff 1
(1 2
(5)
aoff)w1{adis}+ +
where w1 is the (excitatory) feedforward weight. The notation {x} means max(0, x) i.e. negative activation values do not propagate and are treated for these purposes as equivalent to 0. The parameters in these simulations are w1 5 feedforward weight, w+ 5 positive feedback weight, w– 5 negative feedback weight, D 5 decay, and Iex t 5 external input. In the simulations only parameters w+ , w– and Iex t are varied. The others have constant values, w1 5 1.0, D 5 0.2. Where it is not ex t varied I 5 0.75.
Simulations with the Simplified Model The existence of two independent systems, one excitatory and the other inhibitory, may make simple correlations between interference and negative priming difficult to observe. This is because interference effects may be modulated by either (or both) of the excitatory and inhibitory systems, whereas negative priming will reflect inhibition alone. This is shown by a simple simulation in Figure 5, in which the strengths of the excitatory and inhibitory components of the selection mechanism of the model are varied independently. This is done in this case by allowing the two feedback weights shown in Figs. 2 and 4 to vary in magnitude. (There are other ways in which the strength of these two systems could be modulated, but varying the feedback weights is the most straightforward.) Figure 5a shows the activation profiles of 4 different target and distractor pairs, each pair presented separately, for 4 different strengths of the excitatory + feedback weight w (Fig. 4). Only one distractor curve is apparent, as the
(a)
(b)
FIG. 5. Two simulations with the simplified model showing how relationships between degree of interference and negative priming can change depending on whether changes in selection efficiency are due to excitatory or inhibitory systems. (a) Increases in the strength of excitatory systems lead to more effective selection, but do not effect inhibitory rebound. (b) Increases in the strength of inhibitory systems increase both selection efficiency and inhibitory rebound.
137
138
HOUGHTON ET AL.
changes in the excitatory feedback do not affect what happens to the distractor. As all other factors have been kept constant (and no random noise has been added), the distractor activation is the same on each of the 4 trials, showing no difference in inhibitory rebound (negative priming effect). The target activation, on the other hand, reaches higher levels, with a more rapid rise, as the strength + of w increases. This increasing activation gap should be manifest as reduced target-distractor interference. Thus we find changes in interference without changes in negative priming. The converse situation is shown in Fig. 5b. In this case the inhibitory feed– + back weight w has been varied while w has been kept constant. Again, four separate target–distractor pairs are presented to the model. This time no effect is seen in the activation profile of the target, whereas the distractor activation clearly changes as the inhibitory weight increases. This time there are two effects: (1) There is again an increase in the target-distractor-activation gap, leading to a reduction in interference. This is caused by the distractor being more efficiently suppressed. (2) As the inhibitory weight increases, the magnitude of the inhibitory rebound increases, which may be manifest as more robust negative priming. In this case there will be a correlation between degree of interference and negative priming. Now consider what would happen to the relationship between level of inhibition and interference observed in Figure 5b when excitation varies as in Figure 5a. Clearly the relationship between interference and level of distractor inhibition would be masked. Indeed, on some trials where excitation is low, increases in inhibition may not produce more efficient selection. Thus the dynamic interplay between excitatory and inhibitory feedback, where strength of feedback may fluctuate in biological systems or may vary due to changes in subjects’ selection strategies, makes simple correlations difficult to observe. We now develop some other predictions regarding negative priming and interference derived from the inhibitory mechanisms implemented in the neural network model. These predictions fall into 3 groups: (1) Time course of negative priming; (2) effects of number of distractors; and (3) effects of distractor salience. We consider each of these in turn.
1. Time Course of Negative Priming. A clear prediction of the model, obvious from Figs. 3 and 5, is that, if negative priming can occur as a result of the inhibitory rebound of distractor representations at stimulus offset, then the NP effect should build up for a short time following offset, as the distractor representation does not become maximally inhibited straight away. In particular, if the probe follows the prime with minimal lag, little or no NP may be observed (see Lowe, 1985; Yee, 1991). In this situation the model predicts a negative relationship between interference and negative priming. When the distractor continues to stimulate the perceptual system until onset of the probe, interference will be large, and negative priming will be small.
INTERFERENCE AND INHIBITION
139
2. Number of Distractors. In the model we propose that all stimuli in the object field compete for representation, via lateral inhibition. Due to this (and possibly due to other processing limits in the perceptual system), we propose that as the number of distractors increases, the degree of activation of any one decreases. Interference increases with number of distractors, because, although each one achieves a lower activation level, the summed activation of all distractors still increases. The total interference is thus due to effects from all distractors. However, more weakly activated distractors lead to weaker activation of their associated off-cells, as the off-cells receive excitatory input from object units (Equation 5). This leads to a smaller inhibitory rebound at stimulus offset and smaller predicted NP (Houghton, 1994, provides an analytic derivation of this property of the model). An example is shown in Fig. 6, where two distractors of different saliency are compared. In this situation the model generates a
FIG. 6. The time course of activation for two distractor stimuli differing in intensity or salience (Experiment 3). The weaker stimulus is represented by the dashed line, the stronger by the solid line. The stronger stimulus, though still suppressed, achieves a higher activation and is likely to produce greater interference. At stimulus offset the stronger stimulus suffers a greater inhibitory rebound, due to the inhibitory feedback system also operating at higher intensity. If we assume that when a number of distractors are present each achieves a lower internal activation level (due, say, to competitive processes in perceptual modules) than when only one distractor is present, then the solid line can also be taken to represent the situation with only one distractor, and the dashed line the activation profile of a number of competing distractors (Experiment 2).
140
HOUGHTON ET AL.
negative relationship between interference and negative priming. That is, as interference increases with number of distractors, inhibition associated with each distractor declines. Following the principles of limited capacity models of attention (e.g. Kahneman, 1973), Neumann and DeSchepper (1992) have proposed that inhibition has a limited capacity. However, such limited-capacity models have always remained somewhat underdefined, because it has never been made clear exactly why resources are limited. In contrast, the model described here can account for effects typically described as being due to limited resources with more explicitly defined mechanisms. In this case inhibition is not limited, but it is determined by the activation level of the object to be inhibited.
3. Distractor Salience. The model postulates that inhibitory systems may be energized by that which they inhibit (Fig. 2). If property nodes in the object field become more active for some reason, then they send more activation through to their associated off-cells. Assuming that more salient (e.g. brighter) distractors generate more highly activated internal representations, then their associated off-cells will be more strongly driven, generating greater inhibitory feedback. It turns out, however, that the equilibrium activation level achieved by distractors is still an increasing function of the strength of their external excitatory input (Houghton, 1994). This is illustrated in Fig. 6, which shows the characteristic activation profiles for distractor inputs of two intensities (N.B. the distractors are not presented at the same time—each curve represents a different distractor on a different occasion, presented for the same length of time). As the stimulus intensity increases, both the peak activation level and the adapted (inhibited) activation (Point A in Fig. 6) increase. This higher activation of more salient distractors leads to greater interference in the response binding process, which, we predict, will be manifest in longer RTs to target stimuli with more salient distractors. In addition to this, the model shows an effect of salience on inhibitory rebound. At distractor offset, there is the usual precipitous decline in activation. However, at Point B in Fig. 6, the various curves “cross over ”, resulting in an inversion of the order of activation levels found at Point A. This leads, at Point C (maximum suppression), to visibly greater inhibition of the previously more active inputs. This greater suppression of more intense inputs may be used to predict greater negative priming (NP). However, it should be noted that the differential NP effect is (1) small in absolute terms, reaching a transitory maximum at Point C and thereafter declining, and (2) much smaller than the interference effect, which is stable. We decided, therefore, to see whether any such small NP difference was detectable, and whether it was much less than any observed change in interference. In this kind of experiment, therefore, the model predicts a positive relationship between interference and negative priming. That is, the more salient stimulus will produce greater interference due to its higher activation level during
INTERFERENCE AND INHIBITION
141
selection of the target, and greater negative priming because of a larger inhibitory rebound. Note, however, as discussed above, that the model produces a surprising asymmetry: The difference in interference (activation level) is substantial between salient and less salient stimuli, but the difference in inhibitory rebound is small.
EXPERIMENTAL TESTS OF THE MODEL As noted previously, recent experimental observations have suggested changes to the model (e.g. Tipper et al., 1994). For example, inhibition may not spread to all object properties in some circumstances. However, even though simulations with a new version of the model will provide better data to compare machine and human performance, we felt it was worthwhile beginning an experimental analysis of the model’s performance, because these changes do not affect many of the fundamental properties of the model. Figures 3 and 6 simulate situations that have received little experimental investigation, and thus they provide a powerful way to test the legitimacy of the model. Simply put: Does human behaviour conform to what the model predicts?
EXPERIMENT 1 Figure 3 demonstrates the activation profiles of an unattended stimulus. When attentional selection begins, inputs to the object field are matched with target field representations in the target–object comparison field. Inhibition associated with a mismatch between an unattended stimulus and the representation in the target field is then transferred to the attentional control field and feeds back on to representations in the object field. The differential feedback to the attended (excitation) and unattended (inhibition) representations in the object field results in an activation gap that enables selection of target from distractor. Importantly, however, the activation of the unattended stimulus remains above resting levels while there is perceptual input. Only after perceptual input into the object field ceases does activity drop below resting levels and result in negative priming. Experiment 1 tests the prediction that representations of unattended objects only drop below resting levels after the stimulus is no longer present. One condition replicates prior research (Tipper, Brehaut, & Driver, 1990). In this situation the target and distractor stimuli are briefly presented for 150 msec. After response to the target there is a 350-msec response stimulus interval (RSI) before the probe is presented. Therefore, when testing the nature of the internal representation of the distractor stimulus via response to the subsequent probe, input to the object field has been terminated for some time. Robust negative priming is normally obtained, as predicted by the model. However, if the target and distractor in the prime display were to be continuously presented until presenta-
142
HOUGHTON ET AL.
tion of the probe, then minimal decay of activation is predicted (although we assume one time interval to enable the distinction between the prime and probe). Therefore little or no negative priming is predicted. Further predictions can be made concerning the interference effects the distractor has upon the target during selection. When the distractor input to the object field is terminated shortly after its presentation, its level of activity rapidly drops. Therefore there should be less interference with selection of the target. Conversely, when the distractor receives continuous perceptual input into the object field, the attention-induced activation gap may be achieved more slowly. Thus, in summary: Continuous presentation of a distracting stimulus results in the activation of its representations remaining above resting levels. Such a situation will result in reduced negative priming and possibly a trend towards increased interference.
Method Subjects Twenty undergraduate psychology students from McMaster University (8 males) participated for course credit.
Apparatus and Stimuli The experiment was carried out on an Apple IIe computer with a monochrome monitor. The approximate viewing distance in all cases was 45 cm. At the start of each trial, the four locations where targets and distractors could appear on the screen were occupied by question marks (“?” in 40-column text mode). This set of locations was centred on the screen such that the horizontal visual angle between the two outside positions was 8.29° and 4.30° between the two lower inside positions. The vertical visual angle between the upper outside and lower inside positions was approximately 1.30°. The stimuli themselves (“O” and “ 1 ” in 40-column text mode) subtended 0.80° and 0.60° vertical and 0.60° and 0.60° horizontal, respectively. This method of marking the spatial locations in which targets and distractors could appear differs from the procedure used by Tipper et al. (1990), Tipper and McLaren (1990), and from Experiments 2 and 3 to be reported here. This change in procedure was necessary to remove a confound between short and long prime conditions. In the long prime conditions, the prime stimuli remain on the screen until they are replaced by the probe stimuli. In the short prime conditions, on the other hand, the prime stimuli are presented for only 150 msec, and then the screen is empty (except for location markers) until the probe appears. Therefore, when the usual procedure is followed, it is expected that long prime conditions will be impaired due to forward masking when prime and probe stimuli appear in the same positions with no interval between them (e.g. the ignored repetition condition). The use of question marks (as opposed to tape markers or
INTERFERENCE AND INHIBITION
143
FIG. 7. Display sequence in Experiment 1 for Short and Long prime displays. Panel A is the start of the trial; and Panel F is the end of the trial.
“underline” characters) removes this confound, as the probe stimuli in both short and long prime conditions are presented in positions that were previously occupied by a target, a distractor, or a question mark. Forward masking effects should therefore be equivalent for all conditions in the probe display (see Fig. 7).
Procedure Subjects were told that each trial would consist of two displays, and that in each display, the target character (O) would appear in one of the four locations marked by a question mark. They were also told that in many displays a distractor ( 1 ) would also appear in one of the four locations. The subjects’ task
144
HOUGHTON ET AL.
was to ignore the distractor and indicate the location of the target as quickly as possible by pressing a spatially compatible key. The keyboard keys that correspond to the four locations ({D},{C},{M}, and {K}) were marked with white tape. Subjects used the index and middle fingers of each hand to make responses. Thus, if the “O” appeared on the far right, the correct response would be pressing {K} with the right middle finger. Speed and accuracy of response were both emphasized. Incorrect responses were signalled by a beep. Responses that took longer than 1000 msec were considered incorrect. Each block of trials proceeded as follows: Once the subject had depressed the spacebar to begin a block, the question marks that indicated potential target positions were presented for 1500 msec before the prime display was presented. The primes were either presented for 150 msec (short) or remained on the screen until they were replaced by the probe (long presentation). After the subject responded to the prime display, there was 350-msec RSI before the probe display appeared for 150 msec. Once subjects had responded to the probe display, the screen was cleared, and a prompt to begin the next trial was presented, along with feedback that consisted of RTs for prime and probe, a beep if either was incorrect, and the percentage of trials correct to that point. Subjects had a practice session in which they made at least 10 correct responses per condition, followed by a test session of at least 30 correct per condition. The ordering of conditions was random within blocks of trials.
FIG. 8. Examples of the prime and probe displays in Experiment 1. The prime display consists of no-distractor and distractor conditions. The difference between these conditions reflects the interference effect. The probe display contains control and ignored-repetition conditions. The contrast between these two conditions reflects negative priming. The probe condition following the nodistractor condition was of no theoretical interest and so was not analyzed.
INTERFERENCE AND INHIBITION
145
Design In this experiment, each trial consisted of two displays—a prime and a probe. The prime displays differed on two dimensions: Whether or not a distractor was present (see Fig. 8), and how long the stimuli remained on the screen. In distractor conditions, the target (O) appeared in one of the four marked locations and the distractor (+) in one of the other four. In no-distractor conditions, only a target was presented. Interference would be indicated by longer RTs in distractor than in no-distractor conditions. In short-prime displays, the target and distractor were presented for 150 msec and then replaced with question marks. In long prime displays, the target and distractor remained on until the onset of the probe display. Thus, in short prime displays, the stimuli were no longer present when the subject responded to the prime, whereas in long primes the stimuli remained on even after the subject’s response. The probe displays of interest were those that followed short- and longdistractor condition primes. There were two types of such probe displays. In the control (CTRL) condition, the positions of the target and distractor in the prime and probe displays were all different. The ignored repetition (IR) condition differed from CTRL only in that the location of the “O” in the probe display was identical to that of the “1 ” in the prime display. Thus, during IR probe displays, subjects were required to respond to the location they had just ignored. A negative priming effect would be indicated by longer RTs in the IR than in the CTRL condition. Interference and negative priming effects and their interactions with short and long prime displays were all examined within-subjects. The dependent measures were RT (in msec) and percentage of errors.
Results and Discussion Interference. The means of median RTs and error percentages for the prime display are shown in Table 1. A 2 3 2 within-subjects analysis of variance (ANOVA) carried out on the median RTs revealed no effect of prime length, F (1, 19) < MS e 5 146.53. However, the interference effect (distractor . nodistractor) was significant, F (1, 19) 5 41.95, MSe 5 400.62, p , 0.001. Interference did not interact with prime length, F (1, 19) 5 2.05, MSe 5 215.01, n.s. Planned orthogonal contrasts revealed significant interference for both short and long prime conditions, p , 0.001 in both cases. The interference effects were also highly consistent: in both short and long prime conditions, 95% of the subjects exhibited interference. Analysis of the error percentages for the prime display revealed that subjects made more errors in short than in long prime conditions, F(1, 19) 5 7.63, MSe 5 4.25, p , 0.05. Also, more errors occurred in distractor than in nodistractor trials, F(1, 19) 5 6.92, MSe 5 5.82, p , 0.05. Prime length and interference did not interact, F(1, 19) , 1, MSe 5 7.15.
146
HOUGHTON ET AL. TABLE 1 Means of Median RTs and Percentage of Errors for the Short and Long PRIME conditions in Experiment 1 Short PRIME
Long PRIME
Display
Condition
RT
% Errors
RT
% Errors
PRIME
No Distractor
420
3.2
417
1.9
Distractor
444
4.6
451
3.3
CTRL
423
4.5
446
5.6
IR
442
6.6
449
4.4
PROBE
Note:
C5
Median RTs in milliseconds.
Negative priming. Means of median RTs and error percentages for the probe display are shown in Table 1. A 2 3 2 ANOVA carried out on the median RTs revealed an overall negative priming effect (I.R. . Control), F (1, 19) 5 4.41, MSe 5 518.86, p , 0.05. The main effect of prime length was also significant, F(1, 19) 5 10.00, MSe 5 435.00, p , 0.01. RTs were longer for probe displays that followed long primes, perhaps because long primes provide less temporal information to prepare for the probe display (as suggested by an anonymous reviewer). The a priori predicted interaction of negative priming and prime length was marginally significant, F(1, 19) 5 4.16, MS e 5 316.05, p , 0.06. Planned contrasts revealed significant negative priming following short primes (19 msec; p , 0.01), but not following long primes (3 msec; F , 1). The proportions of subjects exhibiting negative priming are consistent with the results of these planned contrasts. Following short primes, 80% of subjects exhibited negative priming, whereas following long primes, only 50% exhibited negative priming. Analysis of the probe display error rates revealed no significant effects. In general, therefore, the results of Experiment 1 are similar to those of the simulations and previous observations by Lowe (1985)—that is, there is a trend towards increased interference and decreased negative priming when the prime distractor remains on until the probe is presented. Such results support the notion that continuous perceptual input maintains the activation level of the distractor above resting levels. This has two effects: (1) The greater activity of the distractor interferes more with processing of the target relative to the situation where the distractor is terminated. (2) When the prime remains on until presentation of the probe, the prime distractor’s activation has not dropped below resting level, and hence no negative priming is observed. Figure 9A summarizes the interference and negative priming effects. The inverse relationship between them is confirmed by the observation of a significant interaction between interference/negative priming and short/long prime duration in a two-way ANOVA, F (1, 19) 5 7.03, MSe 5 465.87, p , 0.025.
(A) Interference Neg Priming
(B) Interference Neg Priming
(C) Interference Neg Priming
FIG. 9. Panel A represents the results from Experiment 1. In all the panels, interference effects are the difference between no-distractor and distractor conditions, whereas negative priming effects are the difference between control and ignored-repetition conditions. The negative relationship between interference and negative priming effects can clearly be seen from the crossover interaction in this figure. Panel B represents the data from Experiment 2. Again the negative relationship between interference and negative priming can easily be seen. Panel C represents the data from Experiment 3, and, in contrast to Panels A and B, a positive relationship between interference and negative priming can be seen.
147
148
HOUGHTON ET AL.
EXPERIMENT 2 Selection of a target can still be achieved, even when there are multiple distractors. Formal comparison of target selection when one, as opposed to two, distractors are present reveals two results: (1) The attentional activation gap between target and distractor is less differentiated when there are two distractors than when there is only one. Thus greater interference is predicted. (2) The level of inhibition associated with each object (the contrast between background and unattended stimulus activity) is less when two, as opposed to one, distractors are present. Therefore less negative priming is predicted when two distractors are ignored. As in Experiment 1, therefore, a negative relationship between interference and negative priming is predicted: Two distractors should produce greater interference and less negative priming relative to one distractor. Such a result would be predicted from limited-capacity models of attention, which may propose that inhibition is a limited resource. Such models would argue that one distractor can receive the maximum amount of inhibition, thus producing less interference and maximal negative priming (see Neumann & DeSchepper, 1992). Alternatively, when inhibition has to be spread between two distractors, each distractor receives less inhibition, resulting in greater interference and less negative priming. As is discussed further on, however, the current model predicts this result without recourse to concepts of limited quantities of inhibition. Also examined in this experiment are attended repetition effects (AR). Houghton and Tipper (1994) do not explicitly model AR effects, but any attempt to model positive priming with the model is likely to involve strengthening of the perception-action link established by the selectional state. Any such facilitation would not be affected by the number of distractors, because it is made possible by the excitation feedback to the target, which is separate and independent from the distractor inhibitory feedback. Therefore, in the current experiment we predict that, in sharp contrast to distractor interference and negative priming effects, attended repetition effects will not interact with the number of distracting stimuli. 4
4 It should be noted that attended-repetition conditions can produce a variety of results, from positive to negative priming. We believe this is because subsequent behaviour is determined by a number of levels of representation. Thus, after processing the prime, perceptual representations remain active and can facilitate future perceptual processes. However, to prevent response perseveration, after overt action, inhibition prevents the response from being produced immediately afterwards, an important feature of a number of models of sequential action (Mackay, 1987; Houghton, 1990; Houghton & Hartley, in press). Thus AR effects are the product of a number of different opposing states, and the balance between these states determines observable behaviour.
INTERFERENCE AND INHIBITION
149
Method Subjects Twenty-five undergraduate psychology students from McMaster University (6 males) participated for course credit.
Apparatus and Stimuli Again, the experiment was conducted on an Apple IIe microcomputer interfaced to a monochrome monitor. The stimuli were generally as in Experiment 1, but somewhat more complicated. The number of possible target locations was increased from four to six. At the start of each trial, these six locations were marked by the “underline” character (“_”) in 40-column text mode. The locations were centred on the screen such that the horizontal visual angle between positions was 7° for the top pair, 4.4° for the middle pair, and 1.9° for the bottom pair. The vertical visual angle was 1.9° between top and middle positions, and 1.9° between middle and bottom positions. The target and two distractors (“@”, “$”, and “&”) all subtended 0.8° vertical, and 0.6° horizontal.
Procedure Subjects were seated in front of the computer and asked to read instructions from the screen. The instructions were also summarized by the experimenter to ensure that subjects understood them. Each trial consisted of two displays, a prime and a probe. During each display, the target (@) would appear in one of the 6 marked locations. During prime displays, either no distractor would be present, or one or two distractors ($ and &) would also appear; during probe displays, there was always one distractor (either $ or &) present. Subjects were told that their task was to ignore distractors and to indicate the target location as quickly as possible by pressing the spatially compatible key. The keyboard keys that corresponded to the six locations ({E}, {D}, {C}, {M}, {K}, and {O}) were marked with white tape. Subjects used the index, middle, and ring fingers of each hand to make responses. Thus, if the “@” appeared on the far right, the correct response would be pressing {O} with the right ring finger. Speed and accuracy of response were both emphasized. Incorrect responses were signalled by a beep. Each block of trials proceeded as follows: Once the subject had depressed the space bar to begin a block, the location markers appeared, and there was a pause of 1500 msec before the prime display was presented for 150 msec. After the subject responded to the prime display, there was a RSI of 500 msec before the probe display was presented for 150 msec. After the subject’s response to the probe, feedback was displayed, and a prompt to begin the next trial given. The feedback consisted of RTs for prime and probe, a beep if either response was in error, and the percentage of correct trials to that point.
150
HOUGHTON ET AL.
Subjects had a practice session in which they made at least 10 correct responses per condition, followed by a test session of at least 30 correct per condition. The ordering of conditions was random within blocks of trials.
Design There were three types of prime displays in this experiment: no distractor, one distractor, and two distractors (see Figure 10). In all three conditions, the target (@) appeared in one of the six marked locations. In the no-distractor condition,
FIG. 10. The conditions tested in Experiment 2. The response to the target in the prime display is analysed when the target is alone (no-distractor) and when one and two distractors are present. The probe displays contain control and ignored-repetition conditions, as in Experiment 1, and attendedrepetition conditions where the target appears in the same location in the prime and probe displays.
INTERFERENCE AND INHIBITION
151
no distractors were present. One and two distractors were also present in the oneand two-distractor conditions, respectively. In the one-distractor condition, one of the two distractors ($ and &) was randomly selected. In the two-distractor condition, both distractors were present (i.e. there were two unique distractors). The probe displays of interest were those that followed one- and twodistractor primes. There were three types of probe displays: CTRL, IR, and AR. For all probe conditions, one target (@) and one distractor (either $ or &) were present. Following one-distractor primes, the prime distractor also appeared in the probe. Following two-distractor primes, one of the two distractors was randomly selected for presentation in the probe. For the two CTRL conditions (i.e. CTRL1 following a one-distractor prime and CTRL2 following a twodistractor prime), the probe target and distractor(s) filled positions that were unoccupied during the prime display. The IR1 and IR2 conditions differed from the CTRL conditions only in that the probe target appeared in a position that had been occupied by a prime display distractor. The AR1 and AR2 conditions differed from the CTRL conditions in that the probe target appeared in the same location as the prime target. Interference (no distractor vs. one and two distractors), negative priming (CTRL1 vs. IR1 and CTRL2 vs. IR2), and positive priming effects (CTRL1 vs. AR1 and CTRL2 vs. AR2) were all examined within-subjects. The dependent measures were RT and percentage of errors.
Results and Discussion Interference. The means of median RTs for the prime display (and percentage of errors) are shown in Table 2. A one-factor within-subjects ANOVA revealed significant differences amongst the three distractor conditions, F (2, 48) 5 51.09, MSe 5 241.375, p , 0.001. Planned orthogonal contrasts revealed that RTs were higher in the one-distractor than in the no-distractor condition, F(1,48) 5 40.60, p , 0.001, and that RTs were higher in the twodistractor than in the one-distractor condition, F(1, 48) 5 13.26, p , 0.001. A large proportion of subjects manifested effects consistent with the results of these parametric analyses. RTs were higher in the one- than in the no-distractor condition for 88% of subjects, and higher in the two- than in the one-distractor condition for 92% of subjects. Analysis of the prime display error rates revealed a similar pattern of results. A one-factor ANOVA revealed differences amongst the conditions, F (1, 48) 5 8.32, MSe 5 4.01, p , 0.01. Planned contrasts revealed a significant difference between no-distractor and one-distractor, F (1, 48) 5 4.37, p , 0.05, and a marginally significant difference between one and two distractors, F (1, 48) 5 3.95, p , 0.07.
152
HOUGHTON ET AL. TABLE 2 Means of Median RTs and Percentage of Errors for Experiment 2 Number of PRIME Distractors 0
Display
Condition
Note:
C5
2
RT
% Errors
RT
% Errors
RT
% Errors
479
1.9
507
3.1
523
4.2
CTRL
487
3.1
486
4.3
IR
514
4.6
495
3.7
AR
436
0.6
432
1.3
PRIME PROBE
1
Median RTs in milliseconds.
Negative Priming and Attended Priming. Mean of median RTs and error percentages for the probe display are shown in Table 2. A 2 3 3 (number of prime distractors 3 condition) within-subjects ANOVA carried out on the median RTs revealed that the number of distractors in the preceding prime display did influence overall RT, F(1, 24) 5 9.57, MSe 5 255.14, p , 0.01. Responses were faster when the preceding display had two distractors. There were also differences amongst the CTRL, IR, and AR conditions, F (2, 48) 5 89.84, MS e 5 737.08, p , 0.001. Finally, the interaction of these two main effects was also significant, F (2, 48) 5 4.59, MSe 5 263.35, p , 0.05. In order to unpack this interaction, two 2 x 2 ANOVAs were carried out to look at the negative priming (CTRL vs. IR) and attended priming (CTRL vs. AR) effects separately. Analysis of the negative priming effects again revealed that responses were faster following two-distractor than one-distractor primes, F (1, 24) 5 8.37, MSe 5 312.06, p , 0.01. The overall negative priming effect was also reliable, F (1, 24) 5 56.67, MSe 5 137.63, p , 0.001. More importantly, however, the negative priming effect also interacted with the number of distractors in the prime display, F (1, 24) 5 7.85, MSe 5 263.66, p , 0.01. Planned contrasts revealed significant negative priming following both one- and two-distractor primes, F(1, 48) 5 34.6, MSe 5 263.35, p < 0.001, and F(1, 48) 5 3.84, MSe 5 263.35, p , 0.05, respectively. The proportions of subjects exhibiting negative priming were also consistent with the interaction: 92% following onedistractor, and 68% following two-distractor primes. In the analysis of attended priming effects (CTRL vs. AR), only the overall priming effect was significant, F (1, 24) 5 63.98, MSe 5 1069.69, p , 0.001. The number of distractors in the preceding prime had no effect on overall RTs, F (1, 24) 5 0.78, MSe 5 190.88, n.s.; nor did the priming effect interact with the number of prime distractors, F (1, 24) 5 0.15, MSe 5 295.50. Planned contrasts revealed significant attended priming following both one- and two-distractor
INTERFERENCE AND INHIBITION
153
primes, F(1, 48) 5 123.46, MSe = 263.35, p , 0.001, and F(1, 48) 5 136.57, MSe 5 263.35, p , 0.001. Again, the proportion of subjects showing attended priming effects supported the parametric analyses: 92% were facilitated following one-distractor primes, and 100% following two-distractor primes. The probe display error rates were submitted to a 2 x 3 (number of prime distractors 3 condition) within-subjects ANOVA. The number of prime distractors had no effect, F (1, 24) 5 0.51, MSe 5 8.55, n.s. There were differences amongst the CTRL, IR, and AR conditions, F(2, 48) 5 17.02, MSe 5 8.90, p, 0.001. However, the interaction of the two main effects was marginally significant, F (2, 48) 5 3.07, MSe 5 4.63, p , 0.06. Therefore, two 2 3 2 ANOVAs were performed on the error rates as well, one to look at the negative priming and one to look at the attended priming effects. Analysis of the negative priming error rates (CTRL vs. IR) revealed no effect of number of prime distractors, and no overall negative priming effect, F (1, 24) < 1 in both cases. However, the negative priming effect did interact with the number of prime distractors, F(1, 24) 5 5.93, MSe 5 4.22, p < 0.05. As Table 2 indicates, more errors occurred in IR1 than CTRL1, but fewer errors occurred in IR2 than in CTRL2. These findings support the RT data, where the negative priming effect was smaller following two-distractor primes. Analysis of the attended priming error rates (CTRL vs. AR) revealed a marginal effect of number of prime distractors, F (1, 24) 5 3.96, MSe 5 5.66, p < 0.06. More errors occurred following two- than one-distractor primes (2.8% vs. 1.8%). The overall priming effect was also significant, F (1, 24) 5 22.88, MSe 5 8.26, p < 0.001. Fewer errors occurred in the AR than in the CTRL conditions (0.9% vs. 3.7%). The interaction was not reliable, F(1, 24) < 1, MSe 5 4.26. The results of Experiment 2 provide substantial support for the simulation. Both interference and negative priming are influenced by the number of distractors in the prime display, whereas attended priming effects are not. The interference and negative priming effects are summarized in Fig. 9B. As in Experiment 1, there is a clear inverse relationship: As the number of distractors increases, interference increases and negative priming declines. Analysis of the data in Fig. 9B with a two-factor within-subjects ANOVA yields a significant interaction between number of prime display distractors and interference/negative priming, F (1, 24) 5 21.78, MSe 5 328.87, p , 0.001.
EXPERIMENT 3 Experiments 1 and 2 have provided some support for the model. By manipulating distractors in terms of their presentation time and number, interference and negative priming have varied. Generally an inverse relationship has been observed, as predicted by the model: Increases in interference have been accompanied by declines in negative priming.
154
HOUGHTON ET AL.
In the current experiment, stimulus strength is manipulated, and, interestingly, the model now predicts a positive relationship between interference and negative priming—that is, as stimulus input increases in strength, interference should increase. However, because the initial activity level produced in the object field by perceptual input feeds into the attentional control field, this activity level influences the inhibition that feeds back into the object field. Therefore inhibition feeding back into the object field should also increase, although in the simulations this is quite a small increase (see Fig. 6). Stimulus strength is manipulated by intensity contrast between a stimulus and its background. Brighter stimuli are known to elicit faster eye movements than dim stimuli (Reuter-Lorenz, Hughes, & Fendrich, 1991). Further, neural activity is strongly dependent on stimulus intensity through at least some of the visual system, from photoreceptors to the primary visual cortex (e.g. Baylor & Hodgkin, 1973; Lennie & Perry, 1981; Miller & Glickstein, 1967). Therefore we contrasted distractors that were white on a black background (high contrast) with dark grey distractors on a black background (low contrast). We predicted greater interference and negative priming from white distractors. It should also be noted that the model assumes that inhibition mechanisms have a ceiling to their efficacy, and thus interference may increase substantially for high-intensity stimuli. However, the more interesting result that may emerge from such a property is that inhibition rises less than interference, if it is already close to ceiling. Previous data have suggested this possibility (Milliken et al., 1994).
Method Subjects Twenty introductory psychology students from McMaster University (5 males) participated for course credit.
Apparatus and Stimuli The experiment was carried out on an IBM-compatible computer (Packard Bell 386/33) with a colour monitor. The approximate viewing distance in all cases was 65 cm. The four locations where targets and distractors could appear on the screen were marked with the “underline” character (_). This set of locations was centred on the screen such that the horizontal visual angle between the two outside positions was 1.92°, and 0.62° between the two lower inside positions. The vertical visual angle between the upper outside and lower inside positions was approximately 0.71°. The stimuli themselves (“O” and “X”) subtended 0.62° vertical, and 0.53° horizontal. Response keys spatially compatible with these 4 locations were interfaced to the computer via an I/O port. RTs were computed using Bovens and Brysbaert’s (1990) TIMEX function.
INTERFERENCE AND INHIBITION
155
Procedure Subjects were seated in front of the computer and told that their task was to indicate the location of the target (O) as quickly as possible, while ignoring the distractor (X). Subjects used the index and middle fingers of each hand to make responses on keys spatially compatible with the four possible target locations. Speed and accuracy of response were both emphasized. Incorrect responses were signalled by a beep. Each trial proceeded as follows: Once the subject had depressed the start key, the four location markers and a central fixation dot appeared. (These markers and the fixation dot remained on throughout the trial.) The first display (the PRIME) was presented 1498 msec after the onset of the markers and remained on for 157 msec. After the subject responded to the first display, there was a 357-msec RSI, and then the second display (the PROBE) was presented for 157 msec. After the subject responded to the PROBE, the screen was cleared, and a prompt to press the start key reappeared. Subjects had a practice session in which they made at least 2 correct responses per condition, followed by a test session of at least 40 correct per condition. The ordering of conditions was random within blocks of trials.
Design In this experiment, each trial consisted of two displays, a prime and a probe. There were three types of prime displays, no distractor, dark gray distractor, and white distractor. In the two distractor conditions, the target (O) appeared in one of the four marked locations, and the distractor (either a dark gray or white X) in one of the other four. In the no-distractor condition, only a target (a light-gray O for all three conditions) was presented. The probe displays of interest were those that followed the distractor condition primes. There were two types of such probe displays. In the CTRL condition, the positions of the target and distractor in the prime and probe displays were all different. The IR condition differed from CTRL only in that the location of the “O” in the probe display was identical to that of the “X” in the prime display. Interference and negative priming effects were both examined withinsubjects. The dependent measures were RT (in msec) and percentage of errors.
Results and discussion Interference. The means of median RTs for the no-distractor condition and the two distractor conditions of the prime display (and percentage of errors) are reported in Table 3. A repeated-measures ANOVA revealed significant differences amongst the conditions, F(2, 38) 5 21.72, MSe 5 111.15, p , 0.001. Further analysis with planned orthogonal contrasts revealed the following: (1) RTs in the dark-gray-distractor condition were higher than in the no-
156
HOUGHTON ET AL. TABLE 3 Means of Median RTs and Percentage of Errors for Experiment 3 PRIME Distractor Colour Condition No Distractor
Display
Condition
Note:
White
RT
% Error
RT
% Error
RT
% Error
443
2.3
450
2.5
464
2.4
CTRL
433
2.0
432
1.3
IR
460
2.9
466
2.0
PRIME PROBE
Dark Gray
Median RTs in milliseconds.
distractor condition, F(1, 38) 5 4.41, p , 0.05; and (2) RTs in the whitedistractor condition were higher than in the dark-gray-distractor condition, F (1, 38) 5 17.63, p , 0.001. Analysis of the prime display error rates revealed no significant differences amongst the three conditions, F , 1. Thus, both distractor conditions produced significant interference, and the white distractor produced more interference than did the dark-gray distractor.
Negative priming. The means of median RTs for the CTRL and IR conditions (and percentage of errors) are also shown in Table 3. The median RTs were analyzed with a two-factor repeated-measures ANOVA. The overall negative priming effect (control , ignored repetition) was highly significant, F (1, 19) 5 33.64, MSe 5 564.04, p , 0.001. The colour of the distractor in the preceding prime display (dark-gray vs. white) had no effect, F (1, 19) 5 0.71, MSe 5 261.58. Finally, negative priming did not interact with prime display distractor colour, F (1, 19) 5 0.64, MSe 5 318.24. Planned contrasts revealed significant negative priming following both dark-gray distractors, F(1, 19) 5 22.91, and white distractors, F (1, 19) 5 36.33, p , 0.001 in the prime display. Analysis of the probe display error rates revealed no significant effects. Figure 9C summarizes the interference and negative priming effects for Experiment 3. Analysis of the data in Fig. 9C with a two-way within-subjects ANOVA provides no evidence for an interaction between prime display distractor intensity and interference/negative priming, F (1, 19) , 1, MSe 5 435.04. As predicted, in sharp contrast to what was seen in Experiments 1 and 2 (Figs. 9A and 9B), there is now a positive relationship between interference and negative priming: Like the model, interference increases substantially, whereas inhibition, as reflected by negative priming, increases more modestly. This positive relationship between interference and negative priming has been observed in previous studies. Tipper et al. (1992), Milliken et al. (1994), and Neill et al. (1994) have all shown that in some circumstances, as interfer-
INTERFERENCE AND INHIBITION
157
ence increases, negative priming can also increase. Importantly, the latter two studies also produced the pattern of data predicted by the model and produced in the present study. That is, the increase in interference was substantially larger than the increase in negative priming. Thus, although we do not observe a statistically different level of negative priming between bright and dim distractors, the results closely replicate a number of other studies that used quite different procedures. Milliken et al. suggested that such small increases in negative priming may be because of some ceiling on the quantity of inhibition. Such an explanation appeals to the obvious, as the alternative is that inhibition is potentially infinite. It is also far from clear how it actually explains the quantitative difference in the two effects. The explanation offered by the current model is somewhat more complex, being due to the fact that the strength of inhibition continually adapts in a bottom-up fashion to the activation level of the distractor. Somewhat unexpectedly, ignored inputs achieving higher levels of activation may actually be subject to greater levels of inhibition.5
CONCLUSIONS There has been increasing debate concerning what neural processes negative priming actually reflects. As discussed above, the dominant view has been that negative priming effects are produced by inhibition mechanisms of selective attention acting upon the internal representations of distracting objects (e.g. Neill & Westberry, 1987; Tipper, 1985). Others have recently proposed, however, that negative priming may reflect memory retrieval processes. Park and Kanwisher (1994) suggested that the probe and ignored prime are associated with the same object-file representation. However, there is mismatching perceptual information, where, for example, the prime may be green and the probe red. Such perceptual mismatching slows responses. Neill and Valdes (1992) also suggest that negative priming can be produced by the probe automatically retrieving prior episodes of that probe object. Again, when the prime episode is retrieved, there is mismatching response information where the ignored prime is tagged with “do not respond”, whereas the probe requires a response. Such mismatching response tags slow responses to the probe. As discussed in the introduction, such memory-retrieval accounts of negative priming are both intriguing and most probably correct. However, the existence of memory retrieval processes does not exclude a role for inhibition mechanisms 5 As we have emphasized here and elsewhere (Tipper & Milliken, 1996) the relationship between distractor interference and negative priming is extremely complex, and much further research into this matter is required. For example, the large changes in interference and smaller changes in NP reported here are not always observed in other manipulations, which may be taken to affect distractor salience. Thus Valdes (1993), Fuentes and Tudela (1992), and Fox (1994) report changes in interference and NP that are of similar magnitude when the distance of the distractor from fixation is manipulated.
158
HOUGHTON ET AL.
(see Milliken et al., 1994; Tipper & Milliken, 1996). Furthermore, in the current context, another major problem with these various memory retrieval models of negative priming is that they are underdefined. For example, it is not clear what processes are able to assign “do not respond” tags, and the informal verbal descriptions are flexible enough to cope with a range of new observations by including new assumptions. What is clearly required, therefore, are more formal computational models that can be tested directly. This paper has described such a model (Houghton & Tipper, 1994). One of the major controversies surrounding the hypothesis that there are inhibitory mechanisms of selective attention, and that negative priming reflects the performance of these mechanisms, is the relationship between distractor interference and negative priming effects. If distractor inhibition is a selection mechanism, then one might expect that intrusions from distractors should be reduced by higher levels of inhibition. That is, a negative relationship between interference and negative priming is predicted: As interference declines, negative priming should increase. Although such negative relationships have sometimes been observed, a variety of other results have also been obtained. For example, both a positive relationship, and no relationship at all, have been observed between interference and negative priming. As described above, the Houghton and Tipper (1994) model is actually able to account for these diverse data. It should be stressed that the model was not originally developed to address this issue. Rather, the model was an attempt to investigate a basic computational problem faced by many biological systems— namely, how is the parallel perceptual system linked to the serial motor system? Some mechanism must be able to select particular perceptual representations to guide specific actions for organisms to achieve behavioural goals. As is often the case, the properties that emerge from a model with interactive mechanisms cannot easily be predicted beforehand. Rather, only by running simulations did it become clear that the model produced the paradoxical results observed in so many experiments. This model may be viewed as a first step towards a unifying theory. It demonstrates quite clearly that a selective attention system that contains both excitation and inhibition mechanisms, which can be separately modulated by different neurotransmitter systems, is very difficult to understand and predict. In fact, it will only be possible to make strong testable predictions concerning the relationship between particular selection mechanisms (such as excitation and inhibition) and selection efficiency (distractor interference) when both mechanisms can be simultaneously measured in one experiment. Perhaps the most striking aspect of the model is that it shows that the same mechanisms can produce quite different results, depending on the experimental manipulations. Consider Fig. 6. This figure shows the change in interference effects (the activation level of the distractor during initial perceptual analysis) and negative priming (the rebound of the distractor representation after stimulus
INTERFERENCE AND INHIBITION
159
offset). However, when the number of distractors is increased, this system predicts an increase in interference and a decrease in negative priming (confirmed in Experiment 2), whereas when stimulus intensity is manipulated, an increase in both interference and negative priming is predicted (confirmed in Experiment 3). Furthermore, Fig. 6 reveals a puzzling pattern of data that had been obtained in previous experiments. That is, it shows that there can be large changes in activation level of the distractor (which determines interference), which are accompanied by smaller changes in inhibitory rebound. This result has been observed in human subjects previously (e.g. Milliken et al., 1994), but it was only during simulations that it was discovered that the model contained this same property. It should also be noted that the model can account for other negative priming results. For example, simulations also show that negative priming is larger when a probe has to be selected from a distractor than when it is presented alone (Lower, 1979; Moore, 1994; Tipper & Cranston, 1985; Tipper et al., 1990; but see also Neill, Terry, & Valdes, 1994), and negative priming declines over short periods (Banks, Roberts, & Ciranni, 1995; Neill & Valdes, 1992; Neill & Westberry, 1987; Neumann & Hood, 1993). (See Houghton & Tipper, 1994, for further discussions.) Thus, perhaps the major achievement of the model is to demonstrate the enormous difficulties we face in attempting to understand selective attention. Although it is acknowledged that the human brain is a massively complex dynamically interacting system, there are often somewhat naive assumptions concerning what can be discovered from experimental results. A single measure, such as the failure to observe a particular correlation, or that a distractor did not interfere with processing of a target (Francolini & Egeth, 1980; Kahneman & Henik, 1981; LaBerge et al., 1991; Yantis & Jonides, 1990), cannot prove that a distractor was not processed or that inhibition mechanisms were not utilized. Complex networks dynamically interacting will never be understood from such simple experimental approaches. Rather, it is necessary to combine formal models with experimental techniques emphasizing converging operations (Garner, Hake, & Eriksen, 1956). As an example, the results of the long prime condition of Experiment 1 could have led one to assume that the failure to observe negative priming effects in some circumstances occurred because inhibition of the distractor did not take place. The model clearly questions such an assumption by showing that inhibition can take place but need not be observable in certain experimental situations. In conclusion, one network of information-processing systems responds in a way that is as diverse as that of human subjects. The interplay between the performance of computer simulations and human subjects can produce valuable insights. The computer model has motivated the examination of original experimental situations; the experimental results obtained have been similar to those predicted by the model; and, most importantly, the model produces some of the
160
HOUGHTON ET AL.
controversial experimental results observed in the literature—namely, both positive and negative relationships between interference and negative priming, and the fact that null results in either effect are not adequate for the conclusion that ignored stimuli were not analyzed. Such a model may be a first step towards creating a theory that can unify the diverse and often contradictory results encountered in studies of attention.
REFERENCES Arbib, M.A. (1990). Programs, schemas, and neural networks for control of hand movements: Beyond the RS framework. In M. Jeannerod (Ed.), Attention and performance, XIII (pp. 111–138). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Arbuthnott, K.D. (1995). Inhibitory mechanisms in cognition: Phenomena and models. Cahiers de Psychologie Cognitive, 14 (1), 3–45. Banks, W.P., Roberts, D., & Ciranni, (1995). Negative priming in auditory attention. Journal of Experimental Psychology: Human Perception and Performance, 21, 1354– 1361. Baylis, G.C., Tipper, S.P., & Houghton, G. (submitted). Externally cued and internally generated selection: Differences in distractor analysis and inhibition. Journal of Experimental Psychology: Human Perception and Performance. Baylor, D.A., & Hodgkin, A.L. (1973). Detection and resolution of visual stimuli by turtle photoreceptors. Journal of Physiology, 234, 163–198. Beech, A.R., Powell, T.J., McWilliams, J., & Claridge, G.S. (1989). Evidence of reduced “cognitive inhibition” in schizophrenia. British Journal of Clinical Psychology, 28, 110–116. Bovens, N., & Brysbaert, M. (1990). IBM PC/XT/AT and PS/2 Turbo Pascal timing with extended resolution. Behavior Research Methods, Instruments, & Computers, 22, 332–334. Broadbent, D.E. (1982). Task combination and selective intake of information. Acta Psychologica, 50, 253–290. Dagenbach, D., & Carr, T.H. (Eds.) (1994). Inhibitory processes in attention, memory, and language. San Diego, CA: Academic Press. Dalyrymple-Alford, E.C., & Budayr, B. (1966). Examination of some aspects of the Stroop colour-word test. Perceptual and Motor Skills, 23, 1211– 1214. Dempster, F.N., & Brainerd, C.J., (Eds.). (1995). Interference and inhibition in cognition. San Diego, CA: Academic Press. DeSchepper, B.G., & Treisman, A. (in preparation). Novel visual shapes in negative priming. Diamond, A. (1991). Neuropsychological insights into the meaning of object concept development. In S. Carey & R. Gelman (Eds.), The epigenesis of mind: Essays in biology and cognition (pp. 67–110). Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Douglas, R.J. (1967). The hippocampus and behaviour. Psychological Bulletin, 67, 416–442. Driver, J., McLeod, P., & Dienes, Z. (1992). Motion coherence and conjunction search: Implications for guided search theory. Perception and Psychophysics, 51 (1), 79–85.
INTERFERENCE AND INHIBITION
161
Driver, J., & Tipper, S.P. (1989). On the nonselectivity of “selective” seeing: Contrasts between interference and priming in selective attention. Journal of Experimental Psychology: Human Perception and Performance, 15, 304–314. Eriksen, C.W., & St. James, J.D. (1986). Visual attention within and around the field of focal attention: A zoom lens model. Perception & Psychophysics, 40, 225–240. Fox, E., (1994). Interference and negative priming from ignored distractors: The role of selection difficulty. Perception and Psychophysics, 56, 565–574. Fox, E. (1995). Precueing target location reduces interference but not negative priming from visual distractors. Quarterly Journal of Experimental Psychology. 48A, 26–40. Francolini, C.M., & Egeth, H. (1980). On the nonautomaticity of “automatic” activation: Evidence of selective seeing. Perception & Psychophysics, 27, 331–342. Friston, K.J., Liddle, P.F., Frith, C.D., Hirsch, S.R., & Frackowiak, R.S.J. (1992). The left medial temporal region and schizophrenia: A PET study. Brain, 115, 367–382. Frith, C.D. (1979). Consciousness, information processing, and schizophrenia. British Journal of Psychiatry, 134, 225–235. Fuentes, L.J., & Tudela, P. (1992). Semantic processing of foveally presented words in a lexical decision task. Quarterly Journal of Experimental Psychology, 45A, 299–322. Fuster, J.M. (1980). The prefrontal cortex. New York, NY: Raven Press (2nd edition, 1989). Garner, W.R., Hake, H.W., & Eriksen, C.W. (1956). Operationism and the concept of perception. Psychological Review, 63, 149–159. Goldman-Rakic, P.S. (1988). Topography of cognition: Parallel distributed networks in primate association cortex. Annual Review of Neuroscience, 11, 137–156. Gray, J.A., Feldon, J., Rawlins, J.N.P., Hemsley, D.R., & Smith, A.D. (1991). The neuropsychology of schizophrenia. Behavioral and Brain Sciences, 14, 1–84. Grossberg, S. (1980). How does a brain build a cognitive code? Psychological Review, 87, 1–51. Hasher, L., & Zacks, R.L. (1988). Working memory, comprehension and aging: A review and a new view. In G. Bower (Ed.), The psychology of learning and motivation, Vol. 22 (pp. 193–225). San Diego, CA: Academic Press. Houghton, G. (1990). The problem of serial order: A neural network model of sequence learning and recall. In R. Dale, C. Mellish, & M. Zock (Eds.), Current research in natural language generation. London: Academic Press. Houghton, G. (1994). Inhibitory control of neurodynamics: Opponent mechanisms in sequencing and selective attention. In M. Oaksford & G.D.A. Brown (Eds.), Neurodynamics and psychology. London: Academic Press. Houghton, G., & Hartley, T. (in press). Parallel models of serial behaviour: Lashley Revisited. Psyche. Houghton, G., & Tipper, S.P. (1994). A model of inhibitory mechanisms in selective attention. In D. Dagenbach & T. Carr (Eds.), Inhibitory processes in attention, memory, and language. San Diego, CA: Academic Press. Houghton, G., & Tipper, S.P. (1995). Inhibitory mechanisms of neural and cognitive control: Applications to selective attention and sequential action. Brain and Cognition, 30, 20–43. Hull, I. (1943). The Principles of Behavior. New York, NY: Appleton-Century. Ingle, D. (1972). Selective choice between double prey objects by frogs. Brain Behaviour Evolution, 7, 127–144.
162
HOUGHTON ET AL.
Ingle, D. (1975). Focal attention in the frog: Behavioral and physiological correlates. Science, 188, 1033–1035. Jackson, S., & Houghton, G. (1994). Sensorimotor selection and the basal ganglia: A neural-network model. In J.C. Houk & J. Davis (Eds.), Information processing models of basal ganglia. Cambridge, MA: MIT Press. Kahneman, D. (1973). Attention and effort. Englewood Cliffs, NJ: Prentice-Hall. Kahneman, D., & Henik, A. (1981). Perceptual organization and attention. In M. Kubovy & J. R. Pomerantz (Eds.), Perceptual organization. Hillsdale, NJ: Lawrence Erlbaum Associates, Inc. Kane, M. J., Hasher, L., Stoltzfus, E. R., Zacks, R., & Connelly, S. L. (in press). Inhibitory attentional mechanisms and aging. Psychology and Aging. Kimble, D.P. (1968). Hippocampus and internal inhibition. Psychological Bulletin, 70, 285–295. LaBerge, D., Brown, V., Carter, M., Bach, D., & Hartley, A. (1991). Reducing the effects of adjacent distractors by narrowing attention. Journal of Experimental Psychology: Human Perception and Performance, 17, 65–76. Lennie, P. (1980). Parallel visual pathways: A review. Vision Research, 20, 561–594. Lennie, P., & Perry, V. H. (1981). Spatial contrast sensitivity of cells in the lateral geniculate nucleus of the rat. Journal of Physiology, 315, 69–79. Liddle, P.F., Friston, K.J., Frith, C.D., Hirsch, S.R., Jones, T., & Frackowiak, R.S.J. (1992). Patterns of cerebral blood flow in schizophrenia. British Journal of Psychiatry, 160, 179–186. Lowe, D.G. (1979). Strategies, content, and the mechanisms of response inhibition. Memory & Cognition, 7, 382–389. Lower, D.G. (1985). Further investigations of inhibitory mechanisms in attention. Memory & Cognition, 13, 74–80. Luck, S.J., Heinze, H.J., Mangun, G.R., & Hillyard, S. A. (1990). Visual event related potentials index focussed attention within bilateral stimulus arrays. II. Functional dissociation of P1 and N1 components. Electroencephalography and Clinical Neurophysiology, 75, 528–542. Luck, S.J., Hillyard, S.A., Mouloua, M., Woldorff, M.G., Clark, V.P., & Hawkins, H.L. (1994). Effects of spatial cuing on luminance detectability: Psychophysical and electrophysiological evidence for early selection. Journal of Experimental Psychology: Human Perception and Performance, 20, 887–904. Mackay D. G. (1987). The organisation of perception and action. New York: SpringerVerlag. McDowd, J. M., & Oseas-Kreger, D. M. (1991). Aging, inhibitory processes, and negative priming. Journal of Gerontology, 46, 340–345. Meegan, D., & Tipper, S. P. (in preparation). Reaching into cluttered visual environments: Spatial and temporal influences of distracting objects. Miller, J.M., & Glickstein, M. (1967). Neural circuits involved in visuomotor reaction time in monkeys. Journal of Neurophysiology, 30, 399–414. Milliken, B., Tipper, S.P., & Weaver, B. (1994). Negative priming in a spatial localization task: Feature mismatching and distractor inhibition. Journal of Experimental Psychology: Human Perception and Performance, 20, 624–646. Moore, C. (1994). Negative priming effects depend on probe-trial conflict: Where has all the inhibition gone? Memory & Cognition, 56, 133–147.
INTERFERENCE AND INHIBITION
163
Neill, W.T., Terry, K., & Valdes, L.A. (1994). Negative priming without probe selection. Psychonomic Bulletin and Review, 1, 119–121. Neill, W.T., & Valdes, L.A. (1992). Persistence of negative priming: Steady-state or decay? Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 565–576. Neill, W.T., Valdes, L.A., & Terry, K. (1995). Selective attention and inhibitory control of cognition. In Frank N. Dempster & Charles J. Brainerd (Eds.), Interference and inhibition in cognition. San Diego, CA: Academic Press. Neill, W.T., & Westberry, R.L. (1987). Selective attention and the suppression of cognitive noise. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13, 327–334. Neumann, E., & DeSchepper, B.G. (1992). An inhibition based fan effect: Evidence for an active suppression mechanism in selective attention. Canadian Journal of Psychology, 46, 1–40. Neumann, E., & Hood, K. (1993). Decay of distractor inhibition in a local-global letter matching task. Paper presented at the Thirty-fourth Annual Meeting of the Psychonomic Society, Washington, D.C. Park, J., & Kanwisher, N. (1994). Negative priming for spatial locations: Identity mismatching, not distractor inhibition. Journal of Experimental Psychology: Human Perception and Performance, 20, 613–623. Pavlov, I.P. (1927). Conditioned reflexes. Oxford: Oxford University Press. Peters, E.R., Pickering, A.D., & Hemsley, D.R. (1994). “Cognitive inhibition” and positive symptomatology in schizotypy. British Journal of Clinical Psychology, 33, 33–48. Posner, M.I. (1980). Orienting of attention. The VIIth Sir Frederick Bartlett Lecture. Quarterly Journal of Experimental Psychology, 32A, 3–25. Posner, M.I., & Dehaene, S. (1994). Attentional networks. Trends in Neuroscience, 17, 75–79. Posner, M.I., & Peterson, S.E. (1990). The attention system of the human brain. Annual review of neuroscience, 13, 25–41. Reuter-Lorenz, P.A., Hughes, H.C., & Fendrich, R. (1991). The reduction of saccadic latency by prior offset of the fixation point: An analysis of the gap effect. Perception & Psychophysics, 49, 167–175. Shallice, T., Burgess, P.W., & Frith, C.D. (1991). Can the neuropsychological case study approach be applied to schizophrenia? Psychological Medicine, 159, 661–673. Shapiro, K.L., & Loughlin, C. (1993). The locus of inhibition in the priming of static objects: Object token versus location. Journal of Experimental Psychology: Human Perception and Performance, 19, 352–363. Steriade, M. (1984). The excitatory-inhibitory response sequence in thalamic and neocortical cells: State related changes and regulatory systems. In G.M. Edelman, W.E. Gall, & W.M. Cowan (Eds.), Dynamical aspects of neocortical function. New York: Wiley. Tipper, S.P. (1985). The negative priming effect: Inhibitory effects of ignored primes. Quarterly Journal of Experimental Psychology, 37A, 571–590. Tipper, S.P. (1991). Less attentional selectivity as a result of declining inhibition in older adults. Bulletin of the Psychonomic Society, 29, 45–47. Tipper, S.P., & Baylis, G.C. (1987). Individual differences in selective attention: The relation of priming and interference to cognitive failure. Personality and Individual Differences, 8, 667–675.
164
HOUGHTON ET AL.
Tipper, S.P., Brehaut, J.C., & Driver, J. (1990). Selection of moving and static objects for the control of spatially-directed action. Journal of Experimental Psychology: Human Perception and Performance, 16, 492–504. Tipper, S.P., & Cranston, M. (1985). Selective attention and priming: Inhibitory and facilitatory effects of ignored primes. Quarterly Journal of Experimental Psychology, 37A, 591–611. Tipper, S.P., & Driver, J. (1988). Negative priming between pictures and words: Evidence for semantic analysis of ignored stimuli. Memory & Cognition, 16, 64–70. Tipper, S.P., Lortie, C., & Baylis, G.C. (1992). Selective reaching: Evidence for actioncentred attention. Journal of Experimental Psychology: Human Perception and Performance, 18, 891–905. Tipper, S.P., MacQueen, G.M., & Brehaut, J.C. (1988). Negative priming between response modalities: Evidence for the central locus of inhibition in selective attention. Perception & Psychophysics, 43, 45–52. Tipper, S.P., & McLaren, J. (1990). Evidence for efficient visual selective attention in children. In J.T. Enns (Ed.), The development of attention: Research and theory. Amsterdam: Elsevier. Tipper, S.P., Weaver, B., Kirkpatrick, J., & Lewis, S. (1991). Inhibitory mechanisms of attention: Locus, stability, and relationship with distractor interference effects. British Journal of Paychology, 82, 507–520. Tipper, S.P., & Milliken, B. (1996). Distinguishing between inhibition and episodic retrieval based accounts of negative priming. In A. Kramer, M. Coles, & G.D. Logan (Eds.), Proceedings of the Conference on Converging Operations in Visual Selective Attention. May 1994, Champaign, IL. (pp. 337–364). Tipper, S.P., Weaver, B., Cameron, S., Brehaut, J.C., & Bastedo, J. (1991). Inhibitory mechanisms of attention identification and localization tasks: Time course and disruption. Journal of Experimental Psychology: Learning, Memory, & Cognition, 17, 681–692. Tipper, S.P., Weaver, B., & Houghton, G. (1994) . Behavioural goals determine inhibitory mechanisms of selective attention. Quarterly Journal of Experimental Psychology, 47A, 809–840. Treisman, A. & Sato, S. (1990). Conjunction search revisited. Journal of Experimental Psychology: Human Perception and Performance, 16, 459–478. Valdes, L.A. (1993). The relation of negative priming to interference. Doctoral dissertation, Adelphi University, Garden City, N.Y. Van der Heijden, A.H.C. (1981). Short-term visual information forgetting. London: Routledge and Kegan Paul. Warren, C., & Morton, J. (1982). The effect of priming on picture recognition. British Journal of Psychology, 73, 117–129. Yantis, S., & Jonides, J. (1990). Abrupt visual onset and selective attention: Voluntary versus automatic allocation. Journal of Experimental Psychology: Human Perception and Performance, 16, 121–134. Yee, P.L. (1991). Semantic inhibition of ignored words during a figure classificatio n task. Quarterly Journal of Experimental Psychology, 43A, 127–153. Revised manuscript received 13 September 1995.
