Introduction

Working memory (WM) is a key concept for understanding human cognition. Originally conceived as a storage system for temporary retention of information while other tasks are being performed (Baddeley 1986), current theories suggest that this form of temporary storage system is critical for imaging information when we recall it from long-term memory (see Chun and Kuhl, this volume) and that it also plays a more general role in the control of goal-directed behaviour, for example by keeping task-priorities and relevant feature information ‘online’. This chapter is concerned with the role of WM in behavioural control, and we focus on the somewhat counterintuitive possibility that, although WM is usually helpful for guiding ongoing behaviour, it can also be disruptive in particular circumstances. Those circumstances are informative about the interplay between WM and task control.

To illustrate a case where WM may be disruptive rather than helpful for performance, consider the following armchair example. Imagine driving to the market and thinking of the list of items you need to buy; you may rehearse the items to be used in cooking dinner, and this could involve, for example, holding a representation of a hamburger in WM. This representation of a hamburger could be disruptive, however, if you happen to be passing a Burger King advertisement at the time, as your attention might stray to this rather than the task at hand (driving). Having items active in memory may inadvertently affect the stimuli you select for action. Under such conditions, the contents of WM may control action but the action is not integrated with the greater context of ongoing behaviour. This would suggest that WM can operate ‘locally’ to control attention, and in some conditions it can be divorced from broader aspects of task control. Here we will review the evidence for this contention, and consider the broader implications for understanding the relations between WM, attention, and task control. (p. 754)

An important initial conceptualization of WM was developed by Baddeley (see Baddeley and Hitch 1974; Baddeley 2000). Baddeley has proposed the operation of different ‘maintenance’ components (phonological, visuospatial, and episodic buffers), plus an ‘executive’ component which acts to control the operation of the maintenance mechanisms. The executive component of WM is similar to the idea of an attentional supervisory system (Norman and Shallice 1986), which acts to modulate processing in routine cognitive processes (driven bottom-up from the environment) to bias activity to favour the goals of the current task. However, although central to the conceptualization of WM, the executive component of WM remains poorly understood. The work we will discuss, on how WM interacts with visual attention, speaks to this issue by showing how components of this ‘WM executive’ can be fractionated, according to which aspects of WM affect stimulus selection. In particular, and in line with our everyday example, the evidence indicates that WM representations can affect stimulus selection even when they are irrelevant, or even misleading, for the primary task at hand. The results, as we shall review below, call for a re-evaluation of the relations between executive functions, attentional control, and WM.

Models of Visual Selection

Our knowledge of how we select from vision objects for action has made enormous strides over the past 30 years, drawing on converging evidence from behavioural studies, functional brain imaging, studies of attentional disorders, and electrophysiological studies of attention in animal models. These empirical developments have been matched by the generation of explicit models of selection. Models such as the ‘guided search’ account (Wolfe et al. 1989), the ‘biased competition’ framework (Desimone and Duncan 1995), the ‘selective attention for identification model’ (Heinke and Humphreys 2003), and the ‘theory of visual attention’ (TVA) (Bundesen 1990) postulate the notion of ‘attentional templates’, which act to guide the selection of visual stimuli. These templates hold an internal representation set-up for behaviourally relevant targets and, though not always described in these terms, they can be thought akin to representations held in WM. In neural accounts, it is often assumed that cognitive control operates through cell assemblies, coded in the prefrontal cortex, which maintain and manipulate goal-related information (Miller and Cohen 2001). In the context of a visual selection task, the prefrontal activity may reflect the ‘template’, which regulates responses in visual processing pathways, selectively enhancing neural pathways associated, for example, with the most relevant location in the visual field (Kastner et al. 1999). Classic neurophysiological evidence comes from single-unit recording studies within the inferior temporal (IT) cortex of monkeys performing a memory-guided search task. In Chelazzi et al. (1998), a memory cue instructed the monkeys about the target for a forthcoming eye movement. This target had to be kept ‘online’ during a delay period of several seconds until a search display containing the saccade target and a distractor appeared. During the delay interval (p. 755) the sustained firing rate of IT neurons reflected maintenance and/or an expectancy of target-related information. Critically, after the onset of the search array, the response of the pre-activated neuron came to reflect only information related to the relevant target object in its receptive field and the neural responses to irrelevant distractors in the search display were suppressed. These data suggest that WM-based feedback helped IT neurons respond selectively to the target which ‘won’ the competition for representation with the distractor. Similar patterns of results for spatially defined targets have been documented in extrastriate cortex (V4) (Luck et al. 1997) and in prefrontal neurons (Rao et al. 1997; Rainer et al. 1998). Such results support the general contention that attentional templates, established as a WM representation in prefrontal cortex, bias selection towards expected targets.

Now, in these single-unit studies, the contents of WM overlap with the target of attention, making it difficult to judge if WM is necessarily linked to task-based guidance of behaviour. In an attempt to pull WM and attentional task factors apart, a recent body of research has used new paradigms where the contents of WM and selective attention for a given target are varied in an orthogonal fashion—creating conditions where the information in WM is irrelevant or even directly disruptive of performance in another ongoing (primary) task. Does WM still determine selection? If it does, what are the conditions generating such effects? By asking such questions, we can learn about the relations between WM, task control, and attentional selection. Here we review this research, highlighting the psychological and the neural mechanisms that support WM biases in visual selective attention and the extent to which they are subjected to cognitive control. We subsequently consider the consequences for theories of visual selection.

Commonalities between WM and Attention

Machizawa and Driver (2011) examined individual differences in the efficiency with which participants showed effects of (i) an alerting signal, (ii) cues to orient attention to the spatial location of a target, and (iii) response competition between a target and distractors (Fan et al. 2002). The same participants also performed a battery of tests measuring visual WM capacity, precision, and vulnerability to interference (from irrelevant items). Using principal component analysis to pull out common factors underlying performance in the attention and WM tests, Machizawa and Driver showed that these aspects of attention (alerting, orienting, and response conflict) mapped onto a common component that also reflected WM capacity, precision, and vulnerability to interference (Vogel et al. 2005; Bays and Husain 2008; Zhang and Luck 2008). These results point to a functional overlap between WM and attention.

There is also ample evidence of functional interactions between WM and attention. For example, our ability to ignore distractors during visual selection tasks depends on (p. 756) the availability of WM capacity; loading that capacity by giving participants a secondary task leads to a reduced ability to withstand distractor competition (de Fockert et al. 2001). The study by de Fockert and colleagues presented observers with a visual selection task (attend to a target whilst ignoring distractors) that had to be performed under a concurrent WM task of either high or low processing load. The visual target of the attention task could be surrounded by distractors, which could be either congruent or incongruent with the response associated with the target. When the capacity of WM was highly taxed, performance in attentional selection decreased (i.e. response interference by incongruent distractors increased). This finding indicates that stressing the capacity of WM can lead to impairments in control processes that help to filter out distractors and to focus attention on a task-relevant target. This provides converging evidence to the results using principal component analysis (Machizawa and Driver 2011), in this case by showing that a variable that affects WM (increasing load) also disrupts attentional selection.

However, not all ways of stressing the capacity of WM are equal, as indicated by evidence of crossover effects based upon the type of mental representation in WM and the representations on which selection is carried out. There are instances where high processing loads in WM can either disrupt or boost selection, depending on whether the load in WM extenuates target or distractor information processing. For example, when the representations filling WM capacity match with distractors rather than targets in a task, then distractor processing is reduced and hence target processing is facilitated (Kim et al. 2005; Park et al. 2007). Such evidence indicates WM can de-couple from (and override) task control. We now discuss the conditions under which this happens.

Coupling WM and Attention and De-coupling Task Control

The paradigm depicted in Fig. 26.1a illustrates a scenario used to assess whether the biasing effects of WM on visual selection are under task control. In this paradigm, search is defined as the primary task which must be undertaken whilst participants hold an irrelevant item in WM. This irrelevant item can reappear in the search task, aligned either with the search target or a distractor. The primary goal of visual selection (prioritize the target in the search task) can be orthogonally varied with the contents of WM, so that WM can be set against the task goals (e.g. when the item in WM reappears in the search display alongside a search distractor). What happens under these conditions?

Click to view larger

Figure 26.1 (a) Example of the display sequences used to assess the automatic influence of the contents of WM on visual selection. Observers are presented with a cue and subsequently the task is to search for a tilted line target. In the invalid condition, the pre-cued object again reappears but this time it contains a distractor. In the neutral condition the memory item does not reappear in the search display. (b) Search is slowed on invalid relative to neutral trials when the irrelevant cue is kept in WM, and this effect is stronger under WM conditions relative to when the cue is merely attended but not held in WM for later report. (c) The WM bias modulates the N2pc. Waveforms were obtained by subtracting activity in electrodes ipsilateral to the target from those contralateral to the target. Reproduced from Sanjay Kumar, David Soto, Glyn W. Humphreys, Electrophysiological evidence for attentional guidance by the contents of working memory, European Journal of Neuroscience, 30(2) pp. 307–317, Copyright © 2003, The Authors (2009). Journal Compilation © Federation of European Neuroscience Societies and Blackwell Publishing Ltd.

In this paradigm search performance can be indexed by indirect measures of attentional selection such as the response latencies to find the target or via direct measures of overt attentional orienting such as the direction of the first eye movement made in the search display. The typical finding is that, when the WM item reappears in the search display alongside a distractor, search is slower and less accurate when compared with a neutral baseline, where the search display contains the same items but they no longer (p. 757) (p. 758) match the memory cue (Fig. 26.1b) (Downing 2000; Soto et al. 2005, 2006; Olivers et al. 2006; Pan and Soto 2010). In these studies the working memory item is at best irrelevant to the search task and at worst it can be consistently detrimental (e.g. in experiments where the memory item only ever reappears at a distractor location rather than at the location of the target; Soto et al. 2005, 2007); nevertheless the re-presented WM stimulus still affects selection of the search target.

These effects of holding an item in memory can be contrasted with performance in another baseline condition (the ‘mere repetition’ condition), when the same visual events precede the search display but participants do not have to hold the cue in memory (the cue can be ‘merely repeated’ in the search display). This baseline condition controls for the effect of the mere presentation of an initial cue, in the WM condition. Despite this, biases from the cue are attenuated under mere repetition conditions (Soto et al. 2005; Olivers et al. 2006). These data indicate that the biases on selection reflect the presence of the cue in WM and not mere bottom-up priming from the initial presentation of the cue itself. Furthermore, the biasing effects from WM cannot be easily modulated by task-based attention to search targets; in such cases the strong association between WM and attention trumps task-based control. The data suggest that WM can operate in a relatively modular way and is not necessarily part of the structures that generate task control. We now go on to discuss the parameters of these effects.

WM Biasing: Early or Late?

The capture of attention by irrelevant items in WM happens rapidly. This is indicated by several pieces of data. One finding is based on the measurement of visual evoked electrophysiological responses based on electrical activity measures on the scalp. Spatial selection of a target within one visual field has an electrophysiological marker known as the N2pc, which reflects a greater negative going waveform over the hemisphere opposite to the selected location in the visual field and arises about 250 ms after the onset of a display (Hopf et al. 2000; Luck and Hillyard 1994). It has been shown that the amplitude of the N2pc is modulated if a search display contains an item matching another irrelevant stimulus being held in WM, and this effect is much stronger than that found when a stimulus is repeated but not initially held in WM (Kumar et al. 2009; Mazza et al. 2011). Fig. 26.1c illustrates the N2pc profiles for the search bias driven by WM and by (mere) repetition priming. As can be seen, the amplitude of the N2pc is larger when the irrelevant WM item falls on the same side of space as the search target (i.e. ipsilateral cueing) relative to when the cue does not reappear in the search display (on neutral cueing trials) or when the irrelevant WM cue appears on the opposite side of space (and presumably competes for selection with) the search target. These results are consistent with the WM item affecting early selection of items in the search display. (p. 759)

Matching these electrophysiological data are results of studies measuring eye movements. Here it has been found that the first eye movement made in search is affected if the search display repeats an item currently held in WM (Soto et al. 2005). A recent study by Mannan and colleagues (Mannan, Kennard, Potter, Pan, and Soto 2010) investigated saccadic performance in a variation of the combined WM–Attention paradigm of Fig. 26.1. In this study, a search distractor appeared abruptly in the search display. There is much research indicating that the abrupt onset of a stimulus can capture attention (e.g. Yantis and Jonides 1984). In addition to the abrupt onset distractor being a potent bottom-up cue to attention, Mannan found that initial saccades were still modulated by whether the new onset distractor matched the contents of WM (the effect of the onset distractor was greater when the same stimulus was held in WM; Fig. 26.2a and b). Interestingly, the timing of this effect matched the time course of the N2pc, affecting eye movements initiated around 200–250 ms after the appearance of the search display. Mannan et al. also measured the trajectory of the saccade to the search target. They found that saccades directed to the target curved away from the location of the irrelevant onset distractor (Fig. 26.2c), and this effect was strongest when the irrelevant distractor matched the item in WM. This result indicates that even salient distractors (distractor onsets), otherwise thought to capture attention, can be affected by matching an irrelevant item held in WM. Interestingly, the curving of saccades away from distractors has been taken as evidence for the distractors being inhibited (Van der Stigchel et al. 2006), suggesting that, at least on trials where targets are selected rather than the item repeated from WM, there is rapid inhibition of the repeated distractor. On such occasions, there is evidence of attentional control affecting performance but in a manner that is independent of WM, which still biases selection to matching items. We will return to this point.

A third piece of evidence for WM exerting an early effect on visual selection comes from studies demonstrating that stimuli in WM can modulate perceptual sensitivity to visual targets (Soto et al. 2010 in the same manner as spatial selective attention) (Downing 1988; Carrasco et al. 2000; Blanco and Soto 2002) (but see Theeuwes and Van der Burg 2007; Cosman and Vecera 2011). Perceptual sensitivity can be measured when the search display is briefly presented and masked, when report accuracy can be used to derive the d′ index. d′ has been shown to be enhanced when an item in WM is re-presented at the location of the search target, compared with the re-presented WM item falling at the location of a search distractor (Soto et al. 2010). Note that, in these studies, participants do not know the location where the repeated WM item will fall, so attentional guidance is based on a match between the features of stimuli in the search display and the WM item. This suggests that stimuli held in WM may produce neural ‘baseline’ shifts in the activation of perceptual channels for corresponding features, so that the subsequent processing of stimuli along those channels is automatically enhanced. Search targets falling at such locations also benefit.

Click to view larger

Figure 26.2 The WM bias modulates early eye movement performance in the presence of salient stimulus-onset signals. (a) Cumulative distribution of initial latencies of eye movements directed to the target and an onset distractor. (b) Probability of the initial saccade being directed to the onset as a function of saccade latency. (c) Curvature of initial target-directed saccades as function of saccade latency. Reprinted from Vision Research, 50(16), Sabira K. Mannan, Christopher Kennard, Daniela Potter, Yi Pan, and David Soto, Early oculomotor capture by new onsets driven by the contents of working memory, pp. 1590–1597, Copyright (2010), with permission from Elsevier.

Click to view larger

Figure 26.3 Example displays used to assess semantic interference in visual search. N2pc waveforms at electrodes contralateral and ipsilateral to the target as a function of the position of the distractor in the search display. Reproduced from Anna L. Telling, Sanjay Kumar, Antje S. Meyer and Glyn W. Humphreys, ‘Electrophysiological Evidence of Semantic Interference in Visual Search’, Journal of Cognitive Neuroscience, 22:10 (October, 2010), pp. 2212–2225. © 2010 by the Massachusetts Institute of Technology.

These results from items being held in WM are mirrored by other results in visual search where the match between a stimulus and a ‘template’ can direct attention even when it is detrimental to the task. For example, Moores, Laiti, and Chelazzi (2003) found that telling people to search for a particular target (e.g. for a motor bike) led to them sometimes selecting an associated distractor (e.g. a crash helmet), even though (p. 760) (p. 761) such a distractor could never be the target. One account of this result is that the attentional template does not only specify a proposed target but also associates of that target. The associated information, held in the template, can direct search to the distractor outside of task control. Telling et al. (2010) further showed that interference from associated distractors modulated the amplitude of the N2pc amplitude (see Fig. 26.3) in the same manner as found by Kumar et al. (2009) using an irrelevant WM cue. The N2pc was higher when the distractor and the search target fell in the same hemifield relative to when they appeared in different hemifields (Telling et al. 2010). Taken together, these findings indicate that information held in memory can modulate visual selection even when it is irrelevant to the primary task (visual search).

Not all results support this contention, however. Carlisle and Woodman (2011a) used a similar WM–Attention paradigm but found no modulation of the N2pc when they re-presented as distractors items that matched the contents of WM. Caution should be exerted, however, as WM biases on attention were found in search latencies (i.e. search was slower on invalid-WM trials relative to the neutral condition, when the WM cue did not reappear).

Features in WM that Bias Attention

The studies of Moores et al. (2003) and Telling et al. (2010), noted above, indicate that attention may be biased by distractors that are semantically associated with target information coded in an attentional template (Moores and Maxwell 2008; Telling et al. 2010). Several studies have additionally shown that abstract cues in WM (e.g. words or even labelled stimulus dimensions) can be sufficient to bias subsequent visual selection (Pan et al. 2009; Balani et al. 2011). For example, having participants hold the words ‘red (p. 762) square’ in WM is sufficient for attention subsequently to go to a red square when it is presented as a distractor in a search display (Soto and Humphreys 2007). Thus WM biases can be driven not only by visual but also by more abstract information held in WM. These effects suggest that WM can bias selection through multiple levels of representation, which may include automatic linkages between semantic and visual codes. Some studies have even reported similar magnitudes of bias from visual and verbal information held in WM (Soto and Humphreys 2007; Soto et al. 2010). Interestingly, visual WM biases have been documented to occur with both short and long time intervals between the memory cue and the search task (Olivers et al. 2006; Soto and Humphreys 2008), as long as the WM cue is actively maintained. However, the effects of verbal WM cues may be less sustained than the visual counterpart (Dombrowe et al. 2010). This last result could arise if the verbal information is initially held in an executive representation in WM but is subsequently coded into a phonological ‘slave’ system (cf. Baddeley 1986). Once in the ‘slave’ representation, the verbal cue is less effective at directing attention. Here we begin to de-couple attention from some forms of WM, complementing our earlier distinction between WM and task control.

WM Biasing: Intentional Memory Resampling or True Capture?

Though the evidence we have reviewed fits with the idea that items held in WM automatically bias attention to matching stimuli, counter-arguments can be made. As these alternatives would change any interpretation of the relation between WM, task goals, and attention, they need to be considered. One proposal is that the critical data reflect strategic rather than automatic linkages between WM and attention. For example, possibly participants try to refresh their memory trace when the WM reappears in a search display and so they deliberately attend to re-presented WM cues (Woodman and Luck 2007) (see Fig. 26.1). The evidence, however, indicates that this is not the case. First, WM biases of selection are greatly attenuated or even abolished as the capacity of WM is taxed by larger memory loads (Soto and Humphreys 2008; Soto et al. 2012a; Zhang et al. 2011). This finding goes against any ‘strategic’ memory-refreshing account because memory resampling would be most beneficial to sustain the memory representation when the capacity of WM is taxed, yet this is clearly not the case. Second, Bahrami-Balani, Soto, and Humphreys (2010) had participants hold a specific object exemplar in WM and re-presented in the search display a different exemplar of that search item (e.g. a labrador dog might be held in WM and a setter might appear in the search display). In the subsequent memory test, participants had to discriminate the exemplar they had originally seen in WM and to reject distractors that were different exemplars. Balani et al. found that search was affected when a different exemplar of the WM item appeared, even though it would have been detrimental to memory performance to strategically attend to this different exemplar. Finally, Kiyonaga et al. (2012) (p. 763) used surprise memory tests which replaced the search display, abolishing any potential strategy to use the reappearance of the WM item in search to boost the memory trace for a subsequent test. Re-presentation of the WM item still affected visual selection. We conclude that a strategic resampling account is not viable; rather, information held in at least some components of WM (not the phonological slave system) can bias attention outside of the bounds of task control.

The Bounds of Task Control

Do items in the appropriate WM state always escape task control? Above we discussed evidence that the effects of items held in WM weaken as the WM load increases (Soto and Humphreys 2008). This is interesting, because it is typically harder to implement task control as the load on working memory increases (de Fockert et al. 2001). Yet, here, an increasing WM load is associated with greater task control in the sense that there are weaker distraction effects from WM. One way to explain this is to suggest that, when there are multiple items held in WM, the representation of each item is weakened. The weakened WM representation has a less potent effect on directing attention.

Other evidence points to additional conditions where WM effects weaken and task-based control over selection is increased. One factor is whether the search target remains the same or changes across trials. Several studies have shown that strong biases from WM are apparent when the search target remains the same from trial to trial and the WM stimuli are updated on a trial-by-trial basis. In contrast, when the search target also changes on a trial-by-trial-basis, biasing effects from WM decrease (Downing and Dodds 2004; Houtkamp and Roelfsema 2006). This is again counterintuitive. Varying the search target on a trial-by-trial basis makes search harder; hence you might expect task control to then be reduced (de Fockert et al. 2001). However, what seems critical here is how the WM cue and the (primary) search target are represented in WM. The result tells us about the nature of WM in relation to attention and task control.

Olivers and colleagues (2011) have argued that the interplay between WM and visual selection is determined by distinct states in WM. Although it has typically been argued that, for example, visual WM may hold three or four items (Bundesen 1990), Olivers proposed that only a single item is represented in at the ‘forefront’ of WM at any one time (Oberauer 2002). It is this item that is intimately linked to attention. Hence, attention is directed to the WM cue if this cue is in the forefront of WM; however, if the search target is at the forefront, search will be directed to that and be less affected by the (irrelevant) WM cue. How does this fit with the data on keeping the search target constant or changing it on a trial-by-trial basis? The proposal is that changing the target on a trial-by-trial basis highlights the search target in WM rather than the irrelevant WM item, and this reduces the effectiveness of the WM cue for the subsequent guidance of attention (Olivers et al. 2011). This is supported by other data. Peters and colleagues (2009) asked observers to keep two items in WM, each of which changed on a trial-by-trial basis; one item had to be detected in a visual stream at central fixation, the other item had to be held for a (p. 764) subsequent memory test after the search response. The WM item could sometimes appear in the central stream, when they should have been treated as distractors. The authors measured the P3 event-related potential, and they interpreted this ERP component with matching stimuli to memory contents (Duncan-Johnson and Donchin 1977). There was a change in the P3 response when the search target occurred, while there was no change in the P3 response when WM stimuli appeared, when compared with the responses to non-matching distractors (Peters et al. 2009). These data indicate that, when the search item was constantly updated, it was immediately available to modulate the P3 response. There is a final caveat, however, which is how general these effects are. In a recent study using a similar paradigm to Peters et al., Zhang and colleagues (2011) had participants hold a colour rather than a letter shape in WM and found that the colour in WM did affect responses to a matching colour in the central stream of items. Apparently, some properties of items in WM may still permeate attentional guidance even if the representation in WM is held in a ‘background’ rather than a ‘foreground’ state. Nevertheless, at least for complex shape stimuli, it appears that WM can be compartmentalized into foreground and background components, and that the coupling of the WM state to attention is most effective for the foreground representation. This foreground representation can be dissociated from task-based control, though, given the evidence so far presented in this chapter that the foreground stimulus in WM directs attention to matching distractors.

There is evidence that frontal lobe structures are important for this compartmentalization of WM into foreground and background components. Soto et al. (2006) tested patients with frontal lobe lesions on the WM–Attention paradigm. They expected that frontal lobe patients might show weaker effects of WM-based guidance of attention, given that the frontal lobes may be critical for representing stimuli in WM (Miller, this volume). This is not what was found. Instead Soto and colleagues reported exaggerated WM biases. To account for this, Soto et al. proposed that the frontal lobe patients had problems in ‘compartmentalizing’ the WM cue and the search target in WM. We suggest that normal participants do compartmentalize the representations of these items to some degree, even if the WM item is held in the foreground of WM under appropriate conditions. Having the items compartmentalized enables participants to switch attention rapidly from an irrelevant WM item to the search target. However, if the search item is not separated from the WM cue, then WM biases may be exacerbated.

Aside from cases where WM stimuli are held in the ‘background’ of WM, are there other conditions in which behaviour can be controlled by the primary search task, and not by items in WM? The evidence here is that task control can be exerted under the appropriate circumstances. For example, Han and Kim (2009) examined conditions in which the WM cue was always invalid (falling at the location of a distractor) when it reappeared in a search display and they varied the time between when the WM cue was given and when the same item reappeared in the search display. They found that the cost from the invalid WM cue changed to a facilitation effect as the interval between the cue and the search display increased. The change from the standard cost from re-presenting the WM item to a facilitation effect (faster RTs on trials when the WM cue appeared (p. 765) and was invalid relative to when it did not reappear, on neutral trials) is striking. The result suggests that participants were able to inhibit the cue in WM over time, so that search was more likely to be directed away from this item and towards the search target (see also Woodman and Luck 2007). Apparently, even if an item is initially placed in the forefront of WM, it can be suppressed over time when participants know that it is always going to be irrelevant for the primary task (here visual search).

Further evidence demonstrates that cognitive control over WM biases can operate through modulation of the memory trace itself (Kiyonaga et al. 2012). Kiyonaga and colleagues orthogonally varied the probability of whether WM cues were valid or invalid. In some blocks of the task participants could reliably anticipate that the WM item would be irrelevant. The WM bias was weaker in this condition relative to when the WM cue could sometimes be valid in the search task (see also Carlisle and Woodman 2011b). Interestingly, Kiyonaga’s study showed that, when WM items were expected to be invalid, responses were slowed to the items on a surprise memory test. This result is consistent with the memory representation itself being weaker, perhaps due to memory suppression taking place. WM biases can be also ‘controlled’ by the demands imposed by the attentional task, for example by training observers in visual search tasks under conditions of high perceptual demand (e.g. by using brief exposures of the search items), which arguably enhances the prioritization of search goals over memory goals (Dalvit and Eimer 2011). Also, by spatially pre-cueing the location of the upcoming search target, the influence of an irrelevant WM-matching item in the search display can be eliminated (Pan and Soto 2010; Soto et al. 2011b).

Taken together, the above findings indicate that WM biases of attention can be subject to task-based control; specifically, WM biases can be partly enhanced or inhibited based on expectations about either the distractor (its likely validity), the spatial location of the target, and the time available for control processes to be recruited.

Role of Conscious Awareness in WM Biases

As well as being linked to attentional guidance, information in WM has also been linked to awareness (Baddeley 1986). However, just as the link between WM and attention can be questioned under some conditions (when task-based control is maximized and when items are held in the background of WM; see above), so the link between WM and awareness can also be challenged. Several questions are relevant to this issue. One is whether WM biases in visual search are dependent on conscious awareness of the search display items. Consider work on patients with ‘visual extinction’ following a brain lesion. Patients with extinction are typically unaware of an item presented in the contralesional field if a competing stimulus is shown simultaneously on the ipsilesional side. Despite this problem in awareness, the patients can have improved report of (p. 766) contralesional targets if those targets match the contents of WM (Soto and Humphreys 2006). This evidence indicates that WM biases may boost access of non-conscious information into awareness. Research on the phenomenon of ‘inattentional blindness’ also supports this notion. Inattentional blindness is the failure to be aware of the presence of stimuli that are not attended. Studies have demonstrated that inattentional blindness may be reduced when participants are directed to look for a particular stimulus and the unattended stimulus is either visually or semantically related to the behaviourally relevant stimulus being held in memory (Most et al. 2005; Koivisto and Revonsuo 2007). Here we have a result reminiscent of the findings for semantic guidance of attention in visual search (selecting the distractor ‘crash helmet’ when searching for a motorbike; Moores et al. 2003; Telling et al. 2010), but even more striking since participants appear not to be aware of the stimulus unless this match is made to the representation in memory. Memory-driven biases appear therefore capable of increasing the signal strength of items that would otherwise remain non-conscious, thus determining which information gains access into conscious awareness.

We can also ask whether awareness of the memory cues themselves is necessary to observe a WM bias. Intuitively you would think this would be the case but recent work suggests otherwise. Soto et al. (2012c) presented participants with an oriented Gabor patch acting as a cue, which was followed by a further oriented Gabor as a test stimulus. The task was to report whether the test Gabor was tilted clockwise or anti-clockwise relative to the cued Gabor. The first cue was masked so that participants were not aware of its orientation. Despite this, participants could perform the delayed discrimination task well above chance in the absence of awareness of the initial Gabor cue and even in the presence of intervening distractor Gabors across a 4 second cue–target delay. This effect contrasts with what are often reported to be short-lived effects from merely repeating non-conscious items (typically found with prime-target intervals of around 0.2–0.3 seconds (Kouider and Dehaene 2007; though see Kunst-Wilson and Zajonc 1980; Bar and Biederman 1998) which have also been shown to capture the deployment of spatial attention (Astle et al. 2009). This finding indicates that unseen items may be maintained in WM in a non-conscious state and still affect the subsequent deployment of attention and awareness (see Pan, Lin, Zhao and Soto, in press). This will be an interesting avenue for future research.

The Neural Basis of WM-based Guidance of Attention

Our understanding of the relations between WM, attention, and task control has also been enhanced by studies of the neural basis of WM-based attentional guidance. Soto and colleagues (2007) used functional MRI in conjunction with the WM–Attention paradigm (see Fig. 26.1). They compared activity when the WM item reappeared in the search display compared with a neutral condition, when the WM stimulus did not (p. 767) reappear. Relative to this neutral condition, the reappearance of the WM item in the search display led to increased activity in frontal, temporal, and visual cortical regions. In contrast, when cues were attended but not held in WM (in a ‘mere repetition’ condition), the same regions responded in the opposite manner—in this case there was a reduction in the neural response to the reappearance of the initial cue in the search display. This pattern of results is illustrated in Fig. 26.4a. This last effect is reminiscent of ‘neural adaptation’ effects, reflective of a reduced neural response to repeated stimulation, which may in turn be associated with facilitated perceptual processing (‘priming’) of previously experienced items (Wiggs and Martin 1998; Grill-Spector et al. 2006). In contrast, the enhanced neural response to the reappearance of the WM cue in search appears to reflect the WM bias, namely, the capture of attention by the WM stimulus (again note that this is not due just to carrying an item in WM, as this arises in the neutral baseline condition too). This dissociation in neural repetition effects by the memory context has now been replicated in several studies (Soto et al. 2011b, 2012a; Greene and Soto 2012). The data confirm that holding an item in WM engages distinct neural processes to effects of mere repetition.

Brain interference methods, for example using transcranial magnetic stimulation (TMS), have provided added evidence that holding an item in WM and merely repeating the same item reflect distinct functional states in the cortex. Soto and colleagues (Soto et al. 2012c) used a similar paradigm to the one depicted in Fig. 26.1 and included trials where the WM item was valid (and cued the target) as well as when it was invalid. TMS was applied to early visual cortex at the onset of the search display. This modulated the impact of the memory cue on search. When a feature cue was held in WM and re-presented in the search task, then TMS to visual cortex enhanced search performance on valid relative to invalid cueing trials (and relative to a TMS control condition). When the cue was merely repeated without being held in WM, however, TMS to visual cortex produced the opposite pattern of results. Now occipital TMS impaired search performance on valid relative to invalid trials. This dissociation of the occipital TMS influence on selection may be accounted for by short-term changes in the neural state of the visual cortex under WM and mere repetition conditions. TMS may enhance the capture of attention by items in WM because the WM stimuli contents are represented in a heightened state in visual cortex; TMS may increase the strength of these visual representations (Silvanto and Cattaneo 2010). In the mere-repetition condition, however, TMS may interfere with the process of neural adaptation, observed in fMRI studies (Soto et al. 2007), lessening and even reversing the cue-validity effect. Fig. 26.4b illustrates this finding.

Click to view larger

Figure 26.4 (a) Example of the brain network linked with WM biases. Superior frontal, parahippocampal regions, and the lingual gyrus were modulated by the reappearance in the search display of a previously experience feature cue. When the cue was kept in WM, reappearance of the WM item in the search display enhanced activity in these areas. In contrast, when the initial cue had to be attended but not held in WM, the same regions showed decreased responses to the reappearance of the cue in the search display (relative to the neutral condition, when the cue did not reappear in the search display). Reproduced from Soto, D., Humphreys, G. W., and Rotshtein, P., Dissociating the neural mechanisms of memory-based guidance of visual selection, Proceedings of the National Academy of Sciences USA, 104, pp. 17186–17191. Copyright, (c) 2007 National Academy of Sciences, USA <http://www.nasonline.org>. (b) Target discrimination following TMS to the occipital cortex (right) or sham TMS, showing opposite effects of stimulation under WM and mere identification conditions. Reproduced from Soto, D., Llewelyn, D., and Silvanto, J., Distinct causal mechanisms of attentional guidance by working memory and repetition priming in early visual cortex, Journal of Neuroscience, 32, pp. 3447–3452 © 2012, The Society for Neuroscience. <www.jneurosci.org>.

Several studies have also shown that WM biases modulate activity in the thalamus and prefrontal cortex (Grecucci et al. 2010; Soto et al. 2007, 2012a, 2012b). However, in these studies the thalamus and prefrontal cortex did not merely respond to the reappearance of the WM cue in search display but rather activity reflected the congruity of the WM cue in relation to the search goal—that is, the regions were differentially active according to whether the WM item was valid or invalid (see Fig. 26.4a). Taken together the data suggest that a fronto-thalamic pathway may be important to integrate the contents of WM and perceptual input based on the goals imposed by the selection task. (p. 768) (p. 769)

Are the neural regions the same for different types of information in WM? Soto et al. (2012d) assessed the neural regions involved in biasing visual selection by verbal and visual items in WM. In the verbal condition, the visual cue and the memory probe tests in Fig. 26.1 were replaced by coloured words (i.e. ‘Red’, ‘Green’). Soto et al. found that biases of visual attention linked to visual cues in WM were associated with increased activity in the left superior frontal gyrus (SFG), while biases of visual attention driven by verbal information in WM modulated responses in the lateral occipital cortex (LOC) (Soto et al. 2012d). This finding suggests two things. One is that biases may come from distinct sources, determined by whether visual or verbal information is represented in WM. The second is that WM biases may operate through neuroanatomical substrates that provide top-down support of earlier visual processing regions; the SFG is a good potential source for biasing visual coding from WM biasing signals because it contains regions (e.g. around the frontal eye fields) which are densely connected to visual cortex (Schall et al. 1995; Moore and Armstrong 2003). The LOC is a visual processing area associated with object recognition (Grill-Spector et al. 1998; Kourtzi and Kanwisher 2000) and visual WM (Xu and Chun 2006) plus also cross-modal sensory processing (Amedi et al. 2007) and semantic processing (Mechelli et al. 2007). Verbal WM may bias visual selection by modulating neural regions associated with semantic analysis of perceptual input.

fMRI studies are also informative for understanding why WM biases are reduced at high memory loads. Soto et al. (2012d) varied the load in WM before participants undertook visual search. Confirming previous results, Soto et al. (2012d) found a robust attentional bias when a single item was held in WM and then repeated in the search display (as in Fig. 26.1). This WM bias then virtually disappeared when WM capacity was taxed with three items. We have argued above that this reduction in WM bias was unlikely to be due to cognitive control structures becoming more engaged under high load conditions. Consistent with this, Soto et al. (2012a) found no increase in activity in prefrontal regions associated with cognitive control as the memory load increased. It could also not simply be the case that representations of the memory cues are lost due to inter-item competition at high WM loads because the presence of WM biases was examined only when participants correctly remembered the information retained on every trial. In addition, even though the WM biases of behaviour were eliminated, there was increased activation in visual cortical regions when WM cues were re-presented in search at high memory loads. This did not reflect the greater load in WM as the increase in visual activation was greater when the WM item was re-presented than when it did not reappear (in a neutral condition). However, the functional coupling between prefrontal regions (around the ventral anterior inferior prefrontal cortex) and the visual cortex reduced as the load increased. These data suggest that WM biases on attention are brought about through functional cross-talk between frontal and visual brain regions. Items held in the foreground of WM generate stronger coupling even when irrelevant to the task, while increases in memory load decrease this coupling.

In contrast to the neural regions implicated in WM-based guidance of attention, the operation of task-based control of behaviour is typically thought to operate through a (p. 770) network of areas within parietal and frontal brain regions (Corbetta and Shulman 2002). Task-based control of visual attention, for example, may reflect the modulation of visual cortical regions through this fronto-parietal network. For example, this network shows enhanced activity prior to the onset of expected stimuli reflecting task-based prioritization of upcoming stimuli (Corbetta and Shulman 2002). Intriguingly, involvement of the posterior parietal cortex (PPC) has not been associated with the presence of WM biases on visual selection (Soto et al. 2007, 2011b). For instance, PPC activity has not been found to be sensitive either to the reappearance of the WM cue in search, or to the validity of the cue in imaging studies of WM biases on attention. Also, visual extinction is strongly associated with damage to the PPC (e.g. Chechlacz et al. 2013) yet, as we have noted, visual ‘extinction’ patients can nevertheless display WM biases in visual selection tasks (Soto and Humphreys 2006; Soto et al. 2011a). These last data indicate that the PPC is not necessary for WM to guide attention. However, as we have noted, there are conditions where there is task control over WM. A cautious interpretation would be that, in these imaging studies, the conditions did not strongly promote strategic cognitive control over WM—for example, the validity of the WM stimulus for search has typically varied randomly on a trial-by-trial basis and therefore cue validity could not be actively anticipated by observers (Soto et al. 2007, 2011b; Grecucci et al. 2010).

The behavioural work indicates that task-based control processes are recruited more strongly when the relation between the WM cue and the subsequent task can be predicted (e.g. when the cue is always invalid; Kiyonaga et al., 2012; Han and Kim 2009). Studies need to assess whether the PPC is recruited to modulate the interaction between WM and attention under such conditions before we conclude that the PPC plays no role. Some evidence that is consistent with this possibility has been recently reported (Soto et al. 2012b). Specifically, Soto and colleagues (2012b) delineated a novel parieto-medial temporal pathway, involving the posterior parietal cortex, the hippocampus proper, and also the posterior cingulate cortex, which may be critical for the wilful regulation of WM biases on visual selection. In particular, the PPC showed anticipatory activity based on foreknowledge of the validity of WM for search, with activity in the PPC enhanced when the WM items were highly predictive of the target or the search distractor. In a similar manner, the hippocampus along with the posterior cingulate were associated with the individual’s ability to either enhance or inhibit the influence of WM on the search task based on foreknowledge of whether the WM cue was predictive (see Fig. 26.5 for more details). The data provide some indication that parietal areas concerned with task-based control are recruited in conjunction with memory regions to modulate WM biases on attention, which may otherwise be established through fronto-thalamic links with visual cortex.

Conclusion

Click to view larger

Figure 26.5 Regions involved in strategic cognitive control over WM biases of selection based on foreknowledge of WM validity for search. Middle temporal regions, including the hippocampus proper and also the posterior cingulate cortex, showed increased activation when the WM cues reliably coincided on a trial-by-trial basis with targets or distractors than when cue–target relations could not be anticipated. We also depict the correlation between the hippocampal response and the benefits in search efficiency gained from suppressing expected WM distractors or by enhancing WM items that are known to surround the forthcoming target. Interestingly, posterior parietal cortices were modulated by the cue period only, showing anticipatory signals when observers have foreknowledge that the WM cue is a valid predictor of the search target (in green) or when it reliably predicted an invalid search distractor (in red), relative to when there was no such foreknowledge. Reproduced from Soto et al. (2012b) <http://www.jneurosci.org> Reproduced from Soto, D., Greene, C. M., Kiyonaga, A., Rosenthal, C. R., and Egner, T., A parieto-medial temporal pathway for the strategic control over working memory biases in human visual attention, Journal of Neuroscience, 32, pp. 17563–17571 © 2012, The Society for Neuroscience.

Our review has highlighted the relations between WM, attention, and task-based control of behaviour. Traditionally, WM has been assigned a key role in cognition in which it is thought to modulate task-based biases on attentional selection. We have argued that (p. 771) the situation is more complex than this. We have suggested that WM can operate in a relatively modular manner, divorced from the effects of task-based control—though task-based control can be brought to bear under suitable conditions and, under everyday conditions, WM is often recruited to help implement task-based control. We also propose that WM itself can be fractionated not only into modality-specific components (visual and verbal components) but also into foreground and background representations. Items held in the foreground of WM modulate attentional selection in a relatively automatic way that can be distinct from effects of task control. This compartmentalization of WM into foreground and background components may have additional benefits; notably this compartmentalization can facilitate the switching of attention between task-specific and task-irrelevant stimuli by separating task-specific representations from the representations of other stimuli. In addition, there are intriguing suggestions that non-conscious items may nevertheless permeate into WM systems and be maintained and used to control subsequent behaviour. Further stimuli can be brought into consciousness by contact with WM representations. The results indicate how, by studying how WM interacts with attention and task-based control processes, we can learn about the structure and operation of WM itself.

(p. 772) Acknowledgements