Models of WM treat prioritization through strategic and automatic attention in the same manner. In this review, we discuss how these mechanisms may be similar or different from each other and from more effortful strategies. The evidence suggests that physical salience, statistical learning, and reward learning all increase the probability that information enters WM. We suggest that only strategic attention triggers processes such as refreshing that maintain precise memory traces. Moreover, physical salience recruits different neural mechanisms and is faster at prioritizing information than either form of implicit learning, but it comes at a cost to memory for other information.

“A good memory is one trained to forget the trivial”—Clifton Fadiman (1955)

A good working memory (WM)—the ability to retain and manipulate information that is held in mind—is prized by the aged and forgetful alike. A common complaint, for example, is getting up to perform a task and then forgetting why one got up. Failing to prioritize the task goal by allocating attentional resources toward it allows other thoughts to creep into working memory, resulting in aimless wandering until the task is recalled. Such strategic attentional control, however, is not always possible, and, thus, more automatic attentional processes may influence WM prioritization instead (Theeuwes, 2018). For example, physically salient information, such as a word presented in a brightly colored font, draws attentional resources, even if there was no intention to prioritize this information. Here, we discuss several mechanisms through which information is prioritized in working memory from an automatic attentional bias. These include (1) physical salience, (2) statistical learning, and (3) reward learning.

Although these automatic attentional biases are not under our control, they serve to direct our limited cognitive resources toward information that is potentially relevant and informative. In some cases, this information should take precedence, as in the case of product warnings that appear in a distinctive font (POISON). In other cases, this information is no more important than the rest or may be totally irrelevant to task goals. Our WM system allows such automatic prioritization to efficiently process information that is potentially relevant, whether due to changes in the environment or by taking advantage of our previous but unconscious experiences. The current paper will review the literature on automatic mechanisms of WM prioritization. Understanding how information is prioritized automatically from competing sensory and internal information may help to use limited working memory resources more effectively.

Strategic prioritization

The contents of WM can be prioritized by strategically deploying attention to information at encoding or during maintenance. For example, information is better recalled when a cue informs participants that a feature, object, or location is more likely to be probed and, thus, allows for strategic prioritization. Cued information not only is more likely to be recalled (Ravizza et al., 2016; Schmidt et al., 2002) but it is also recalled with better representational quality, or precision (Dube et al., 2017; Zhang & Luck, 2008). Researchers have posited several explanations for this effect, which highlight the interrelated nature of WM and attention. Attention is thought to act as a gatekeeper that assists in prioritizing task-relevant information (Bays & Husain, 2008) and leads to more effective binding of features of an object at a location (Myers et al., 2017). Selective attention undertaken by the higher-order cortices is thought to prime downstream sensory regions so that task-relevant information has an advantage from efficient perception (Gazzaley & Nobre, 2012). In this manner, attended information in the environment gains priority access to limited resources and can thus be encoded into WM for further processing.

In addition to selection benefits at encoding, attention is thought to play a role in the effective maintenance of information currently held in WM. One popular idea is that attention can be directed to internal representations to refresh memory representations, and thereby protect them from decay or interference (Cowan, 1999; Gazzaley & Nobre, 2012; although see Oberauer, 2019). For example, WM performance is improved when an informative cue is presented after information has been encoded (i.e., retro-cue) so that attention can be used to prioritize internal memory representations (Gunseli et al., 2015). Attention to internal memory representations may reactivate neurons coding the item and put them into a more readily accessible state (Myers et al., 2017).

Figure 1 (top row) represents how strategic prioritization could result in better WM for some information. A central arrow cue indicates that the yellow cross is more likely to be probed than are the other colored shapes. Thus, the prioritized item is more likely to enter WM because of the deployment of attention resources at encoding (top row, left side). Continued prioritization during maintenance may result in the item remaining longer in the focus of attention or being attended more frequently (top row, center). Consequently, the quality of the representation of the prioritized item remains high compared with the other items in WM. This model reflects both the increased probability of selection and higher representational quality observed in studies of WM prioritization through strategic attention.

Fig. 1
figure 1

Model of prioritization. Shapes are either sensory or memory representations of the stimuli. In all cases, the yellow plus (+) sign is the prioritized item, through either strategic attention, physical salience, or implicit learning. Strategic attention gives an advantage for selection at encoding and representational quality remains high. Physical salience also increases the probability of selection because of the capture of bottom-up attention, but has no advantage to representational quality. Implicit learning increases the probability of selection which does not come at a cost to other items in the memory set. Similar to salience, implicit learning has little effect on the quality of the representation. During maintenance, strategic attention remains on the prioritized item, keeping it in the focus of attention, but alternates in the other conditions. (Color figure online)

Full size image

While the strategic placement of attention has proven effective, it relies on compliance from the individual to direct attention in an effective manner. Without compliance, information will not be prioritized and, thus, will fail to show a recall advantage. Information, however, can also be prioritized in WM in an automatic fashion that does not rely on the willingness of an individual to expend conscious effort. Physical salience, statistical and reward learning are all automatic and/or implicit mechanisms that have been shown to improve WM recall. It is unclear, however, whether these mechanisms work similarly to more effortful strategies or to each other. In the remainder of this paper, we will review evidence for automatic mechanisms of prioritization and discuss how these mechanisms may be similar or different from each other and to more effortful strategies.

Prioritization from automatic attentional biases

The strategic deployment of attention is not always possible either because which information is relevant beforehand is unknown or because mental resources are not available to prioritize relevant information. In these situations, memory for some information may become prioritized by an automatic attentional bias. For example, implicit knowledge that a VIP always sits in the same spot may prioritize information associated with that location, such as the name of the person sitting there or what they say during the meeting. The current paper discusses three mechanisms—physical salience, statistical learning, and reward learning—that prioritize information because of an automatic deployment of attentional resources. While we focus on three types of automatic prioritization, we are not claiming that these are the only ways that WM is prioritized automatically.

Importantly, this paper reviews evidence of WM prioritization via an attentional bias when information is task relevant and top-down set is ambiguous. In contrast, most previous work has focused on how automatic attention to distracting information competes with strategic attention for prioritizing the contents of WM. In the example above, the person sitting in the VIP chair may not be an expert on the topic of the meeting, and prioritizing that person’s opinions is detracting from encoding the real expert’s opinion, who is sitting across the table. Accordingly, many studies have reported that physical salience, reward learning, and statistical learning reduce the benefit of strategic prioritization of task-relevant information (Anderson et al., 2011; Anticevic et al., 2010; Britton & Anderson, 2020; Della Libera & Chelazzi, 2006; Gaspelin et al., 2015; Majerus et al., 2012; Makovski & Jiang, 2007; Olesen et al., 2007; Sali et al., 2014; Wang & Theeuwes, 2018; West, 1999). Thus, prioritization via automatic mechanisms is often treated as a nuisance that interferes and competes with strategic attention for prioritizing the contents of WM. Information processing via automatic attentional mechanisms, however, likely accounts for a substantial portion of selection (Anderson, 2018; Constant & Liesefeld, 2020; Theeuwes, 2018). Furthermore, as stated above, there are many situations in which automatically attending to some information is important and should take priority in WM. The current review focuses on situations in which prioritized information is task-relevant in order to understand the mechanisms underlying benefits to WM performance.

Moreover, in each case, we target evidence that argues against strategic processing as the underlying cause of memory prioritization. If, for example, participants are aware that accurate memory performance for some information is associated with a reward, then they may use effortful strategies to enhance processing of that information (Allen & Ueno, 2018). This review focuses on studies in which reward associations are learned implicitly in order to observe the properties of an automatic attentional bias from reward without the influence of strategic effects. For the same reason, we also will not discuss the effects of emotional salience on WM; that is, emotion is thought to trigger controlled processing because of its inherent relevance (Kensinger & Corkin, 2003) and, in fact, this bias does not always result in prioritization effects in which emotional information is better recalled (see Mikels & Reuter-Lorenz, 2019, for a review). Instead, we are interested in cases in which it is unlikely that performance benefits are due to effortful strategies rather than automatic processes.

Physical salience

Information that is physically salient stands out in some way based on its perceptual attributes (Uddin, 2015). Physical salience can be due to differences in low-level features, such as brightness or contrast relative to surrounding stimuli (Constant & Liesefeld, 2020). Information can also be physically salient indirectly based on its location; for example, a sudden onset can imbue physical salience onto information that is subsequently presented at that location regardless of its features. Several studies have shown that physical salience improves WM performance for information that is task relevant.

One line of work has investigated how physical salience affects working memory for the location of objects in maps or naturalistic scenes (Fine & Minnery, 2009; Pedale & Santangelo, 2015;Santangelo et al., 2015 ; Santangelo & Macaluso, 2013). In these studies, physical salience of objects in the image is not manipulated, but is instead derived from a salience map calculated from discontinuities of orientation, color opponency, and contrast (Itti et al., 1998). All objects within the scene are potentially relevant such that any of the locations might be probed; thus, physical salience is task-relevant rather than being associated with distractors in this context. Memory is then assessed for objects located at areas of high or low salience. These studies have shown that accuracy for salient objects is higher and that they are more likely to be recalled (Pedale & Santangelo, 2015) than objects of lower salience (Fine & Minnery, 2009). Salient objects have a WM advantage even when accounting for the greater time spent looking at these objects (Santangelo et al., 2015; Santangelo & Macaluso, 2013). The advantage of physically salient objects is observed when recalling their location (Fine & Minnery, 2009; Santangelo et al., 2015; Santangelo & Macaluso, 2013) or their identity (Pedale & Santangelo, 2015).

Better memory for salient items is also observed when the degree of physical salience is manipulated. The orientation of salient items was recalled more accurately than other items when physical salience was manipulated by increasing the thickness (Gong & Li, 2014; short SOA), contrast (Melcher & Piazza, 2011), or relative orientation (Constant & Liesefeld, 2020) of the stimulus. Moreover, WM performance shows a parametric relationship with the degree of salience; that is, the performance advantage for salient items increases with greater physical salience (Constant & Liesefeld, 2020; Melcher & Piazza, 2011).

It is important to demonstrate that the advantage for salient objects is unlikely to be due to strategic effects. In some of these studies, the locations of highly salient objects were more likely to be probed than low-salient objects simply because there were more objects of low salience in the memory set (Santangelo et al., 2015; Santangelo & Macaluso, 2013). This might lead to strategic prioritization of salient objects because they are more likely to have to be recalled. To control for this potential confound, Pedale and Santangelo et al. (2015) used a free recall procedure. Free recall diminishes the likelihood of using salience strategically as it is important to recall all items, not just those that are physically salient. In other studies, all objects were probed (Fine & Minnery, 2009) or had an equal likelihood of being probed (Gong & Li, 2014; short SOA) so that it would not be advantageous to voluntarily prioritize salient objects above the rest. Thus, the benefit of physical salience is observed even when it is not advantageous to strategically prioritize these items.

The memory advantage for physical salience is thought to be due to the capture of attentional resources in a bottom-up fashion (Santangelo, 2015). This automatic capture of attention may determine which information is given priority for encoding and subsequent storage into WM (Bays & Husain, 2008). Alternatively, the WM advantage for physically salient items might be due to better retrieval or storage rather than an attentional bias. For example, salient items might be placed in their own category (Bruce & Gaines, 1976; Endress & Potter, 2014; S. R. Schmidt, 1991), which results in reduced interference in storage or at retrieval. Physically salient or distinctive items are also associated with more cues for recall which would give them an advantage at retrieval (Kelley & Nairne, 2001). When physical salience is inherent to the stimulus itself, it is difficult to tease apart these explanations. Studies, however, in which salience is imbued indirectly to a stimulus are informative in this regard.

Several studies have adapted the Posner cueing paradigm (Posner, 1980) to assess the benefits of physical salience to a location rather than manipulating the distinctiveness of the item itself. In this type of experiment, a sudden onset cue appears at one of several locations, and memory performance is compared between items that appear at the cued and uncued locations. The cue is not informative about which stimulus will be probed, so that any benefit is likely due to an automatic shift of attention to that location rather than due to strategic prioritization. Importantly, nonspatial features of all items, cued and uncued, are equally salient, which rules out explanations attributed to differences in categorization or the retrieval of distinctive items. These studies show that WM performance for colors (Ravizza et al., 2016; Schmidt et al., 2002), orientation (Bays & Husain, 2008), and letters (Ravizza et al., 2016) is better at the cued location compared with uncued locations. This suggests that an attentional bias from physical salience is sufficient to produce a benefit to WM.

These studies demonstrate that physically salient information draws attention in an automatic, bottom-up fashion. Indeed, the effects of bottom-up attention in perceptual tasks resemble the effects of physical salience in WM performance. For example, the effects of physical salience on WM are transitory, as are the effects of bottom-up attention in discrimination tasks (Hopfinger & Mangun, 1998). Cue–target intervals in studies showing the effects of salience on WM are short (<200 ms; Bays & Husain, 2008; Ravizza et al., 2016; Schmidt et al., 2002) and are in line with the optimal interval necessary to demonstrate cuing effects in perceptual tasks (see Prinzmetal & Landau, 2008, for a review). Accordingly, a salience effect in WM performance was not observed in a change-detection task when the interval between the two displays was long (>500 ms; Gong & Li, 2014). Thus, physical salience has a transitory effect on performance in WM that is characteristic of bottom-up attention.

Given the transitory effects of bottom-up attention, the benefit of physical salience most likely affects the selection and transfer of information into WM rather than influencing the quality of the representation itself. Evidence for this claim comes from studies showing the benefit of a nonpredictive cue presented shortly after the offset of the memoranda (i.e., retro-cue). Studies have shown better performance for cued items even when the retro-cue is not predictive of future recall (Schmidt et al., 2002; Zokaei et al., 2014).Footnote 1 Indeed, the benefit of a nonpredictive retro-cue was equal to a nonpredictive precue suggesting that a precue does not provide an additional advantage from better representational quality (Schmidt et al., 2002). Further supporting this idea, a retro-cue did not affect the quality (i.e., precision) of the representation in a continuous report task compared with a noncued baseline task (Zokaei et al., 2014). Schmidt et al. (2002) argue that bottom-up attention affects the transfer of the perceptual representation into WM rather than strengthening representational quality; namely, it is unlikely that a retro-cue could strengthen perceptual quality once the stimuli have disappeared.

Further support for salience affecting selection rather than representational quality is reported in a study in which sources of error from these factors were estimated (Constant & Liesefeld, 2020). Participants were asked to remember the color of three tilted lines among a field of vertically oriented distractors and were asked to match the color in memory by clicking on a color wheel. The target lines could be tilted such that there was a low, moderate, or high degree of salience compared with the distractor stimuli. The degree of salience parametrically affected the estimate of selection, the probability than an item entered memory, such that greater salience increased the likelihood that the color was encoded into WM. In contrast, Bayes factors indicated moderate or indecisive evidence that precision was unaffected by salience.

Rather than increasing representational quality, an attentional bias for salience may provide a selection advantage via bottom-up attentional capture at encoding. The performance benefit of physical salience may result from such items being encoded before others, thereby increasing the probability that the item is encoded into WM. If so, this benefit should be observed only when there is competition for selection, such as when memoranda are presented simultaneously. To test this prediction, we investigated the effects of a nonpredictive location cue when the memory set was presented simultaneously or sequentially. We showed that accuracy was higher for cued than for uncued items when they were presented simultaneously, but not sequentially (Ravizza et al., 2016). In comparison, strategic attention manipulated by a predictive cue was beneficial regardless of whether presentation was simultaneous or sequential. Items at physically salient locations were also more likely to be recalled first in a free recall task (Ravizza et al., 2016), which is characteristic of items that are encoded first (Tan et al., 2016). These results support the idea that salient items are automatically prioritized in WM because of an attentional bias that benefits the order of encoding. This is reflected in Fig. 1 (center row, left side), where salience directs attention to the prioritized item first.

It is unlikely that salient information is prioritized once items have been transferred into WM. Maintaining information over a delay calls for strategic efforts including attentional refreshing or articulatory rehearsal that do not occur automatically (Hasher & Zacks, 1979); for example, children have to learn to rehearse information in order to maintain it in WM because it is not an automatic process (Gathercole & Adams, 1994; Tam et al., 2010). When all items are equally important, there is no benefit to directing more attentional resources to physically salient information and, indeed, attention should shift to other items in the memory set (Fig. 1, center row, center panel). The lack of sustained processing of salient information during maintenance will thus result in a representation whose quality is just as vulnerable to decay and interference as other items in WM. This supposition remains to be tested in WM, although there is some evidence that bottom-up and strategic attention affect different stages of processing in the perceptual domain (Riggio & Kirsner, 1997); that is, salient information captures early and reflexive attention while strategic attention has effects later in processing (Muller & Rabbitt, 1989).

This automatic capture of attentional resources is thought to be difficult to overcome (Santangelo, 2015) and to diminish recall of nonsalient information (Melcher & Piazza, 2011; Pedale & Santangelo, 2015). In fact, the WM advantage due to physical salience comes at a cost to other items in the memory set, and thereby produces an overall decrement of WM capacity. When items of the memory set have equal salience, estimated WM capacity is higher than if one item is more salient than the others (Melcher & Piazza, 2011). Moreover, a greater advantage of physical salience predicts lower numbers of overall items recalled (Pedale & Santangelo, 2015). Thus, while salient information is more likely to be recalled, it comes at the cost of an overall reduction in WM performance.

If bottom-up attention is affected by physical salience, neural networks important for salience processing or bottom-up attention should be engaged in the performance of these WM tasks. For example, the anterior insula, temporoparietal junction (TPJ), and anterior cingulate are associated with processing physically salient stimuli (Corbetta et al., 2000; Hahn et al., 2006; Menon & Uddin, 2010). Surprisingly, no evidence of these regions has been observed in response to salient items that are task-relevant in fMRI studies of WM (Santangelo et al., 2015; Santangelo & Macaluso, 2013; Wills et al., 2017). In fact, one region, the right TPJ, tended to have a greater response to stimuli with lower physical salience (Santangelo et al., 2015; Wills et al., 2017). One possibility is that fMRI is unable to capture transitory effects associated with physical salience. Using a method with higher temporal resolution (i.e., EEG), however, also failed to provide evidence of a bottom-up attention effect for salient items (Wills-Conn et al., 2019).

Note, however, that the experimental design used in these imaging studies was not appropriate for isolating bottom-up attention effects. For example, high-salience items were more likely to have to be recalled, possibly invoking strategic effects (Santangelo et al., 2015; Santangelo & Macaluso, 2013). In our studies, participants had to inhibit a motor response to salient items (Wills et al., 2017; Wills-Conn et al., 2019). Accordingly, markers of strategic control rather than bottom-up attention were observed in all these studies.

Alternatively, better memory for physically salient information may be due to a triggering of strategic prioritization rather than attentional capture per se. Strategic prioritization may be necessary to reduce interference from salient features and maintain task goals. For example, physically salient information, even if it is task-relevant, should not receive more resources in WM than the rest of the memory set and strategic prioritization may be necessary to allocate attentional resources more fairly. This enhancement of strategic prioritization may then benefit WM representations incidentally (Krebs et al., 2015; Rosner et al., 2015; Rosner & Milliken, 2014) because such stimuli receive additional processing. If so, this would explain why the cognitive control network, rather than the salience network, is associated with better memory for physically salient information.

In the perceptual domain, however, physical salience effects are dissociable from the neural effects of strategic prioritization. For example, neural responses to strategic prioritization in an orientation discrimination task increased further upstream in higher levels of the visual pathway while the effects of physical salience were similar at each stage of visual processing (Dugué et al., 2020). In addition, strategic control affected the amplitude of a marker of selective attention (i.e., N2pc; Kiss et al., 2008), whereas physical salience affected the latency of this component (Bachman et al., 2020). The latency of the N2pc was also related to response times to salient targets (Bachman et al., 2020). It remains to be seen whether neural markers associated with better perceptual processing of salient information are also related to better memory of this information.

In sum, physically salient information is more likely to be recalled even in situations in which it is unlikely that effortful strategies are deployed. A benefit to WM performance is observed if the item itself is physically salient or is salient because of its location. The latter result suggests that an inherent attentional bias to physically salient information is sufficient to prioritize this information in WM apart from differences in how distinctive information is stored or the ease of retrieval. An attentional bias to salient information results in such information being encoded first, and thereby increases the likelihood that it enters WM rather than affecting the quality of the representation.

Statistical learning

In addition to automatic prioritization of physically salient items, the human visual system is also equipped to apply efficient automatic processing to prioritize information that has been imbued with potential relevance due to its repeated selection or predictable configuration. Although this idea has only recently been integrated into models of attentional selection (Awh et al., 2012; Theeuwes, 2019), early vision scientists acknowledged the preposterous nature of a theory of perception that assumes a random selection process, asserting that the visual system must account for statistical regularities and reduce redundancies in order to process information efficiently (Field, 1987; Helmholtz, 1910/Helmholtz, 1925). Accordingly, repeated selection or detection of regularities in stimulus properties or configurations (i.e., selection history) can produce an attentional bias based on previous experience even when selection is contrary to goals and the selected features or stimuli are not inherently salient (Awh et al., 2012; Theeuwes, 2019).

Jiang et al. (2015) suggested that implicit learning creates an attentional bias that results from habitually shifting to locations or features. This habitual shifting occurs involuntarily, or as a “procedural” form of attention (Jiang et al., 2013). For example, search is faster for targets that are presented more frequently in the same color or location (Conn et al., 2020; Jiang et al., 2013; Jiang et al., 2015; Sha et al., 2017). This attention bias is maintained even when statistical regularities have been removed (Conn et al., 2020; Geng & Behrmann, 2005; Jiang et al., 2013; Jiang et al., 2015; Qu et al., 2017; Sha et al., 2017), suggesting that these effects are not due to mere trial-based priming effects. Moreover, these biases are observed in the absence of awareness (Conn et al., 2020; Jiang et al., 2018) and, therefore, are not due to the strategic deployment of attention. Thus, when specific features or locations are selected repeatedly or targets are placed in regular, predictable configurations, attention can be biased automatically toward them. These studies demonstrate that an attentional bias is created from the implicit learning and exploiting of environmental regularities in perceptual tasks.

Learning how to move attention (Jiang et al., 2015) may also result in better WM performance by creating an “encoding bias” for some information (Umemoto et al., 2010). There are currently only a handful of studies investigating how an implicitly learned attentional bias affects WM prioritization (Olson et al., 2005;Umemoto et al., 2010 ; Won & Leber, 2017). In all these studies, statistical regularities were manipulated by increasing the probability that an item at a location would have to be recalled, and there are currently no studies investigating a feature-based attentional bias in the WM domain (Won & Leber, 2017). In general, these studies report that WM performance is better for items at locations that are frequently probed even after the probabilities are subsequently changed to be equal (Umemoto et al., 2010; Won & Leber, 2017). Moreover, the attentional bias is not related to explicit awareness that items are more likely to be probed at one location (Olson et al., 2005; Umemoto et al., 2010; Won & Leber, 2017). Taken together, these studies indicate that statistical regularities in the location of important information can be learned implicitly and influence WM prioritization.

There is some evidence that the advantage from statistical learning is due to the selection of information into WM rather than improving the quality of the representation. To distinguish between these two possibilities, Umemoto et al. (2010) used a change-detection task, in which the change could be large (e.g., rectangle ➔ circle) or small (e.g., oval1➔oval2). Big changes that are easy to observe should reflect whether the object was selected for encoding into WM, whereas the quality of the representation will determine how well small changes are detected. The results showed that the detection of large changes was affected by statistical regularities in the location of probed items. In contrast, there was no benefit of location in the detection of small changes. As with physical salience, these results suggest that an implicit attention bias confers an advantage to selection rather than a change to qualitative attributes of the representation (see Fig. 1c, bottom row).

While statistical learning increases the probability that an item is selected for encoding, it is unknown whether an attentional bias continues after encoding. It may be that a habit of deploying external attention during encoding transfers to the deployment of internal attention during rehearsal or mental refreshing. Such a transfer has been shown between external and internal search tasks; that is, implicitly learning whether it is better to exploit or explore for items in a visual search task influenced how participants mentally searched for solutions to anagram problems (Hills et al., 2010). More likely, goal-directed processes will overcome a habitual attention bias in order to ensure that all items are remembered equally well. It is thought that attentional refreshing and articulatory rehearsal require conscious awareness or they will not be employed (Hasher & Zacks, 1979). Accordingly, the finding that statistical learning has no observable effect on representation quality (Umemoto et al., 2010) may be evidence for the latter idea; namely, the precision of the representation is not protected from decay or interference by prioritizing the item during maintenance.

In sum, statistical regularities in the environment can create an attentional bias that influences prioritization in WM, although this bias has only been shown for space. This bias is learned implicitly and does not rely on awareness of statistical regularities. Habitual shifts of attention to the prioritized location may result in an attentional bias that persists even when the advantage of processing items at that location is removed. The performance advantage seems primarily due to the increased probability of the item entering WM rather than to the quality of the representation, similar to the advantage observed for physical salience. As there are only a few studies of statistical learning in WM, there are still many unanswered questions, such as whether performance is worse overall due to an implicit attentional bias, whether learning results in an advantage to encoding order, and whether internal attention shows a similar bias.

Reward learning

Reward is a powerful tool to shape behavior and influence task motivation. It has been associated with enhanced attention, perhaps due to functional coupling between reward-sensitive brain regions and attentional networks (Mohanty et al., 2008). Recent evidence has shown that learning stimulus–reward associations can alter attentional priority such that, when specific locations or features are associated with stronger reward value, these stimuli compete more effectively for limited attentional resources independent from goals and salience (Anderson et al., 2011). While detrimental effects from previously rewarded distractors have been well documented (Infanti et al., 2015; Ward et al., 2019; see Anderson, 2016, for a review), less studied are the benefits to performance from implicit learning when rewarded stimuli are task relevant (Failing & Theeuwes, 2014).

To date, relatively few studies have investigated whether implicit learning enhances prioritization of rewarded stimuli in WM (Gong & Li, 2014; Wallis et al., 2015). In one study, participants learned reward values associated with abstract shapes, and then performed a working memory task with the same stimuli (Wallis et al., 2015). Performance was observed to be a function of the value of the memory array of shapes; that is, d' was highest when the memory array consisted of all high-reward shapes, followed by mixed arrays of high- and low-reward shapes, with the worst performance for arrays composed of low-reward shapes. Interestingly, the benefit of reward to WM performance was not specific to the shapes associated with a high reward; for example, memory for high- and low-reward shapes was equal when they were presented in mixed arrays. Wallis et al. (2015) argued that reward modulates the gating of items into WM; that is, reward boosts the encoding of all items presented concurrently rather than reward creating an attentional bias to an individual item (Wallis et al., 2015). This finding is reflected in Fig. 1 (bottom row, left) in which more items are transferred to WM than other modes of prioritization. While Wallis and colleagues did not find evidence of an attention bias, they note that their data do not rule out this mechanism, nor is it mutually exclusive with the idea that reward modulates a gating mechanism. Moreover, the findings might be explained by a general boost to attention that is temporally rather than spatially defined (Wallis et al., 2015).

Other investigators, however, have found evidence that reward learning creates an item-specific advantage for stimuli associated with high reward (Gong & Li, 2014). Gong and Li (2014) used a training task wherein participants identified a canonically oriented line and responded to its orientation. The line was surrounded by a red or green circle, with one color associated with a greater probability of receiving a high reward and the other associated with a greater probability of receiving a low reward. Participants were not informed of these contingencies, but learned them during the training stage. In the test portion of this experiment, participants performed a WM change-detection task for the orientation of colored lines. Accuracy was higher for stimuli presented in the high-reward color than the low-reward or unrewarded colors, even when participants were unaware of the reward contingencies. Therefore, the attentional bias was not due to the strategic deployment of attention but was observed when stimulus–reward associations were implicit. Thus, an item-specific advantage was observed for the high-reward color.

A remaining question is whether attentional bias from reward learning, as with physical salience and statistical learning, influences selection without a concomitantly affecting the quality of the representation. Evidence for this hypothesis comes from Wallis et al. (2015); after performing a stimulus tracking task with high- and low-rewarded colors, participants performed a WM task where they had to remember the orientation of colored lines. The orientations of previously high-rewarded colors were more likely to be encoded into WM; however, the precision of the representation was not affected by learned reward values. This suggests that reward learning creates an attentional bias that benefits selection, but does not affect the quality of the representation (see Fig. 1c, bottom row). This may imply that the attentional bias does not trigger control mechanisms that would protect the representation from decay or interference or that enhances perceptual quality during encoding.

In contrast, when reward associations are made explicit, prioritization effects are observed during maintenance rather than selection at encoding. Thomas et al. (2016) found that eye gaze during WM encoding was not contingent upon reward value, but instead was driven by physical salience. Similarly, explicit reward feedback from the previous trial had no effect on attentional selection at encoding (measured using N2pc ERPs), but better representation in WM, as indexed by higher contralateral delay activity (Infanti et al., 2017). As reward associations were available to consciousness, it is possible that participants used strategic attention to mitigate the influence of reward when it was irrelevant to the WM task. Furthermore, the reward priming effects demonstrated by Infanti et al. (2017) may directly enhance memory because of recent occupation in WM from the previous trial rather than forming an attentional bias from repeated experiences with stimulus–reward pairings. Thus, the limited evidence to date supports a benefit to selection rather than more representational benefits when implicit learning and strategic attention are dissociable.

Evidence for a WM advantage from implicit reward learning has primarily been acquired from studies in which an attentional bias is formed from associating reward with nonspatial features—generally, color. In fact, relative to feature-based reward learning, spatial reward learning has not been consistently demonstrated; some studies of visual search failed to observe performance benefits for search targets appearing at the rewarded location (Jiang et al., 2015; Sisk et al., 2020; Won & Leber, 2016). This was not due to participants failing to learn the association, because performance was better for the rewarded location during the training session (Jiang et al., 2015). This suggests that the effect of reward increased the size of local priming effects rather than creating a long-term attentional bias toward a region of space. While some studies have found a spatial bias for rewarded locations in perceptual tasks (Chelazzi et al., 2014), this may be due to awareness of the reward contingencies that recruit strategic attention (Sisk et al., 2020). Indeed, a spatial bias for reward was shown to be entirely due to conscious awareness of the location-reward contingency (Sisk et al., 2020), and, thus, be a product of strategic attention.

Alternatively, the inconsistency in the literature regarding the influence of reward on spatial attention might suggest that reward learning for feature-based and space-based attentional biases rely on distinct mechanisms. In an effort to reconcile these findings, Anderson and Kim (2018) proposed that value-driven attention in the spatial domain relies on reinforcement learning rather than Pavlovian associative learning, which can bias attention toward rewarded features. The authors demonstrated across multiple experiments that reward learning could reliably shape spatial attention, but that these effects were context specific rather than reflecting a general attentional habit to attend to a particular region of space (Anderson & Kim, 2018). If so, it may be that reward learning requires a tighter mapping between space and reward; that is, a quadrant or side of space might be too large of an area to associate with a reward in a context-specific manner (Miller & Murphy, 1964). Similarly, some evidence suggests that reward learning shapes attentional deployment through its influence on eye movements (Liao & Anderson, 2020) rather than space per se.

The neural mechanisms underlying reward benefits to WM are largely unknown; however, some insights can be gleaned from studying the neural mechanisms underlying reward-related learning and attentional capture. Value-driven attention studies using neuroimaging techniques generally observe brain activations in visual cortex, including object-selective ventral cortex and early visual cortex, and reward-related regions in the basal ganglia, such as the striatum (for a review, see Anderson, 2019). Thus, Anderson (2019) argued that reward value affects attentional priority via two mechanisms: a feed-forward pathway, through which value-modulated signals from early visual cortex influence the parietal priority map, and a dopaminergic reward pathway that modulates reflexive attention.

Reward signals from dopamine neurons in the striatum, and particularly the tail of the caudate nucleus in humans, have been implicated in the development of persistent attentional biases to toward stimuli associated with high reward (Anderson et al., 2014; Anderson et al., 2017). The striatum is involved in the control of habitual action and the integration of learning with motivational processes and reward-related decision-making (Balleine et al., 2007), and Anderson and colleagues (2017) found that reward-related dopamine release in this region during learning was predictive of the strength of later attentional biases for stimuli associated with high reward. Thus, dopamine may serve as a mechanism for altering attentional priority, perhaps by signaling changes in synaptic weights based on information from reward signals, expectations, and errors (Suri & Schultz, 1999). Furthermore, the caudate nucleus may influence reflexive attention because of its projections to visual pathways such as the superior colliculus, which is involved in reflexive eye movements. Through this pathway, value-modulated responses in the caudate nucleus may influence attentional priority outside of goal-driven influence (Anderson, 2019).

In sum, implicit learning can create an automatic attentional bias toward information associated with reward that is outside of awareness. This bias exists when reward contingencies are removed, suggesting that it is not merely a priming effect from repeated exposure. Limited evidence suggests that implicit learning of the features associated with a reward increases the likelihood that they are encoded into WM rather than increasing representational quality. In contrast, neither the strategic use of reward (Thomas et al., 2016) nor the priming of reward associations (Infanti et al., 2017) seems to have a strong effect on creating an encoding bias. The modulation of attention from reward learning may be due to dopaminergic processes that are important for habit formation and value-based decision making.

Differences between strategic and automatic prioritization in WM

Models of WM assume that strategic and automatic attention prioritize information in the same manner; for example, both strategic attention and physical salience are claimed to increase the activation level of items in long-term memory (Brown et al., 2000; Cowan, 1988, 1999; Farrell & Lewandowsky, 2002). Thus, both ways of prioritizing information are thought to increase the likelihood that information enters an activated state in WM. After attentional selection at encoding, these models are agnostic as to the fate of information that has been prioritized through different selection mechanisms. Information that is strongly activated, either by strategic attention or automatic capture to physical salience, is placed in the focus of attention (FOA) and is maintained in an active state by control process such as attentional refreshing and articulatory rehearsal (Cowan, 1999; although see, Oberauer, 2019); thus, maintenance processes are presumed to be similar regardless of how information is prioritized. Moreover, no model has addressed the effect of prioritization through a learned attentional bias.

There are some differences in behavior, however, when prioritization is accomplished by strategic or automatic attention. While all forms of attention enhance processing of the prioritized item, larger benefits are observed for strategic attention when directly compared with physical salience (Berryhill et al., 2012; Landau et al., 2007; Ravizza et al., 2016; Schmidt et al., 2002), statistical learning (Jiang et al., 2015), and reward learning (Jiang et al., 2015) across perceptual and WM tasks. In WM, larger effects from strategic attention may be due to the persistence of attention to the prioritized item until it is recalled. Attentional refreshing and articulatory rehearsal are effortful strategies that protect the memory representation from decay or interference (Camos et al., 2018; Hasher & Zacks, 1979). Items that are prioritized strategically may be rehearsed or refreshed more often than other items. In contrast, items prioritized through automatic processes may not be given priority during maintenance because they are not necessarily more important than the other items to be remembered (see Fig. 1). Thus, strategies to prioritize items during maintenance might not be biased by physically salient information or from implicit learning.

A weaker attentional bias during maintenance may also explain why the quality of the representation is not affected by automatic prioritization. Strategic attention has marked effects on the precision of the representation (Bays & Husain, 2008; Dube et al., 2017; Zhang & Luck, 2008), but no such effect has been observed from automatic attention (Constant & Liesefeld, 2020; Umemoto et al., 2010; Wallis et al., 2015). Strategic processes during maintenance may keep the precision of the representation high whereas resolution is lost without effortful control. In contrast, precision may decrease at the same rate as unprioritized information when prioritization occurs automatically. Alternatively, or additionally, strategic processes at encoding may result in a higher-resolution memory trace, perhaps through deeper encoding rather than better maintenance processes. Future research is necessary to understand whether physical salience and implicit learning bias attention during encoding and/or maintenance. For example, articulatory suppression during maintenance may reduce the effect of strategic attention, but not prioritization from physical salience or implicit learning. Moreover, fMRI or EEG studies could directly assess how brain activity changes during encoding and maintenance from strategic and automatic prioritization.

Strategic and automatic prioritization also show opposite relationships to working memory capacity. It is well known that individuals with a greater working memory capacity are better able to resist distraction from physical salience or reward-related cues (Anderson et al., 2011; Anderson et al., 2013; Anderson & Yantis, 2012) when automatic processing competes with strategic attention to task-relevant stimuli. WM capacity is used more effectively when attention is directed to task-relevant information in the face of competing distractors. Interestingly, these results suggest a counterintuitive prediction when task-relevant stimuli are automatically prioritized. Individuals with lower WM capacity may show greater effects of automatic prioritization, and thus better WM performance for prioritized information than those with higher WM capacity. Alternatively, better memory for automatically prioritized items may come at a greater cost to the rest of the memory set for those with lower WM capacity. It is currently unknown how WM capacity or, relatedly, general intelligence relates to benefits from automatic prioritization.

Differences between automatic forms of WM prioritization

Each form of automatic prioritization creates an attentional bias that increases the probability of selection into WM but does not confer an additional advantage upon the quality of the representation. The mechanism by which this attentional bias is formed may differ between them, however. Physical salience is thought to rely on innate processing advantages for distinctive or changing information. In contrast, detecting statistical regularities and forming stimulus-reward associations are learned over time. In the following section, we will discuss how differences in the mechanism of prioritization may have different consequences on the timing of attentional selection, competition for selection, and type of attentional bias, spatial or non-spatial.

Physical salience versus reward learningFootnote 2

Physical salience is thought to draw upon bottom-up attentional mechanisms (Santangelo, 2015) that interrupt current processing in order to orient to changing or distinctive stimuli (Corbetta et al., 2000). Indeed, the perceptual system is thought to have an innate preference for physically salient information because novel or changing stimuli typically carry more information about the environment than does nonsalient information. In contrast, attentional effects from learned associations may require more time to implement because value or expectancies have to be retrieved. In consequence, a learned attentional bias may affect WM performance at a later stage of processing than that shown for physically salient information. This difference in the timing of attentional effects was observed in an ERP study of visual search in which the physical salience and reward value of the targets were manipulated (Bachman et al., 2020). Physical salience shortened the latency of the N2pc component, a marker of selective attention, whereas reward value was reflected in a greater amplitude of this component. Thus, there is some evidence that attentional effects from reward learning occur later than the effects of physical salience.

The fast orienting of attention to physically salient information may be the cause of the advantage to encoding order such that salient information is more likely to enter WM first (Ravizza et al., 2016). This encoding advantage is observed when simultaneously presenting all items of the memory set such that selection demands are high, but not when presentation is sequential (Ravizza et al., 2016). In contrast, reward value affects WM performance both when there is high (Gong & Li, 2014) and low selection demands at encoding (Wallis et al., 2015). This is consistent with the idea that physical salience primarily benefits the speed of allocating attentional resources, whereas reward modulates the strength of the attentional bias.

Reward learning produces different effects on competitors than does physical salience. Information associated with high reward does not come at a cost to other items in the memory set (Gong & Li, 2014; Wallis et al., 2015), whereas items that are physically salient pull attentional resources to their location. In Fig. 1 (middle and bottom rows), this is reflected in the fate of the purple square which is encoded when prioritization is through implicit learning, but not from physical salience. This capture of spatial attention from physical salience decreases the recall of the features of nonsalient items (Melcher & Piazza, 2011). Moreover, WM performance shows a spatial gradient such that memory for nearby items is enhanced compared with distant items (Schmidt et al., 2002). In contrast, an item-specific attentional bias from reward does not appear to pull attentional resources to its location. For example, Gong and Li (2014) examined WM performance for stimuli with no value when high- or low-reward items were part of the memory set. Recall of the orientation of no-value stimuli was affected equally by high- and low-reward items in the memory set; that is, the WM advantage from reward did not come at a cost to other items in the display. Therefore, the item-specific benefit from reward seems to be due to a modulation of feature-based attention rather than a product of spatially based attentional capture (Gong & Li, 2014).

Physical salience and reward increase the likelihood that items enter WM, but do so through different mechanisms. Physical salience evokes a rapid and automatic capture of spatial attention that increases the likelihood of encoding features at that location. In turn, this causes an overall decrement to WM capacity such that other stimuli are less likely to be encoded. The attentional bias from reward learning does not appear to be spatially based but, instead, modulates WM through feature-based (Gong & Li, 2014) or temporally based (Wallis et al., 2015) attention at a later stage of processing (Bachman et al., 2020). Taken together, these results argue against the idea that the memory advantage for stimuli associated with high reward is merely due to the recruitment of the bottom-up attention system (Anderson, 2019).

Statistical versus reward learning

Statistical and reward learning are both examples of an attention bias formed from experience, but it is unclear whether they derive from the same process or are independent. One possibility is that the successful execution of goals, which can be enhanced by predictive environmental factors, generates internal reward signals that reinforce attention to particular features. Levy and Glimcher (2012) suggested that neurons represent various types of value as a “common neural currency” by which attention is influenced, suggesting that value may be represented in a domain-general manner. Thus, a common mechanism for the development of long-term biases in attention to rewarded stimuli and stimuli that are frequently selected may result from the generation of reward signals from the successful identification of targets (Anderson, 2016). The brain may represent value for these different types of encounters in a uniform way.

Indeed, there is evidence that both statistical and reward learning recruit similar neural regions in the striatum and hippocampus. Wittmann et al. (2005) found that pictures that were predictive of reward were remembered more accurately in a later test phase and showed increased activation in both the reward-sensitive basal ganglia and the hippocampus. Similarly, Turk-Browne et al. (2009) found activation in the medial temporal lobe, including the hippocampus, and the striatum associated with implicit learning of patterns and environmental regularities (although see Kim & Anderson, 2019b, who did not find caudate tail activation in statistical learning). Other investigators have supported this claim, suggesting that the hippocampus and striatum are adept at detecting regularities in the environment, including those in temporal (Schapiro et al., 2016) and visuospatial structure (Fiser & Aslin, 2001). Furthermore, both statistical and reward learning activate sensory areas, including visual cortex (Anderson et al., 2014; Kim & Anderson, 2019b). Thus, these regions support both reward and statistical learning and are likely involved in learning associations and regularities that can later bias attention to continue to select the learned features or stimuli.

In addition to shared neural substrates, there are similarities in the attentional effects of statistical and reward learning. As described above, both affect the probability of selection into WM without a concomitant effect on representational quality. Both also show a similar boundary condition for observing an attentional bias; namely, the object or feature has to be diagnostic such that it differentially predicts a favorable outcome (i.e., more likely to predict a target or reward value). In statistical learning, a lasting attentional bias will not be formed if features have a 100% chance of being a target, whereas, if one color becomes more predictive of a target or distractor, the attentional bias persists after the statistical regularity is removed (Sha et al., 2017). Similar findings are observed in reward learning: In the absence of reward feedback, which may serve as a diagnostic feature that a specific color is more predictive of a favorable outcome, participants fail to show persistent attentional biases toward colors that contained the target during the training phase (Anderson et al., 2011; Gong & Li, 2014).

Despite these similarities, there is also evidence that attention from reward and statistical learning are independent. For example, Anderson, Chiu, et al. (2017) demonstrated that although value-driven attention capture was blunted in depressed individuals, attentional biases related to unrewarded selection history remained robust. The authors interpret this as further evidence that value-driven attention and attention driven by statistical regularities are best characterized as separate influences on attention. In a subsequent study, Kim and Anderson (2019b) found that activation in dopamine-sensitive regions such as the caudate nucleus was not elicited in an unrewarded statistical learning task. Conversely, statistical learning recruited visual areas, suggesting that these effects were due to plasticity of visual neurons as regularities were learned rather than internal reward responses.

Furthermore, recent work by Anderson and colleagues (Anderson & Kim, 2018; Kim & Anderson, 2019a) has begun to identify potential distinct learning mechanisms underlying statistical and reward learning. When participants learned to make a saccade away from a particular color, which could be paired with a reward or simply presented more frequently, their response patterns in a subsequent test phase differed based on the training method. Surprisingly, participants in the reward training condition showed an attentional bias toward the high-reward color despite being rewarded for ignoring it during training; response times were faster when the target (now defined by shape) matched the high-reward color from the training task and slower when the distractor matched the high-reward color. However, participants who completed the version of the antisaccade training task that presented one color more frequently showed the opposite pattern; they responded more slowly and made more errors when targets were presented in the high-probability color and faster when the distractor was presented in the high-probability color (Kim & Anderson, 2019a).

Kim and Anderson (2019a) argue that this pattern of results reflects underlying mechanistic differences between statistical and reward learning. While reward learning is reliant on associative learning, whereby learning a predictive relationship between reward and a particular feature results in changes to attentional priority, statistical learning is thought to rely on habitual attention shifts developed through instrumental conditioning. Similarly, Anderson and Kim (2018) propose a mechanistic explanation to account for another apparent difference between the two forms of implicit learning: the lack of consistent evidence for spatial biases in reward learning (Jiang et al., 2015; Won & Leber, 2016). Rather than relying on associative learning, Anderson and Kim (2018) argued that spatial reward learning (similar to statistical/frequency-based learning) relies on reinforcement learning related to behavior. In support of this theory, the authors demonstrated similar spatial attention shifts following both monetary reward and corrective feedback (unlike feature-based value-driven attention, which does not emerge in the absence of reward; see Anderson et al., 2011).

In summary, current research suggests a number of potential cognitive and neural mechanisms by which attention is automatically biased toward information that has been previously selected, and these mechanisms are likely enacted by a combination of common and distinct underlying neural substrates. Mechanisms of associative learning, including the influence of learning on expectations, inherent stimulus value, and stimulus–outcome associations, can enact long-term changes in attention that influence the likelihood that an item will enter WM in future encounters. In the brain, dopamine-sensitive neurons in the basal ganglia and hippocampal learning and memory systems may be enacted toward structures that are diagnostic of preferred outcomes and produce these changes in attentional biases. Future studies should continue to compare these two types of learning to determine the best method to incorporate past experience into models of attention, whether that is under a unitary mechanism, or as distinct contributions from statistical and reward history. Studying the impact of these mechanisms on WM might also provide greater insights into their common or distinct influence on information processing.

Conclusion

Prioritization in WM results from both strategic and automatic attentional biases, and a complete model of WM will include both as mechanisms by which information becomes prioritized. This literature review suggests that there are important differences between strategic and automatic prioritization, as well as between automatic biases from physical salience and implicit learning. Strategic deployment of attention is the most effective way to prioritize information, resulting in both increased likelihood of selection and a high-quality memory trace. It is likely that prioritization through automatic attentional biases will provide an advantage for selection into WM, but that strategic attention will be necessary for sustained benefits to memory quality.

Many questions remain unanswered including basic questions such as whether learning feature-based statistical contingencies improves WM, whether an implicit attentional bias for space can be observed from reward, and the neural mechanisms underlying enhanced performance for physically salient information. The answers to these questions will improve our understanding of prioritization in WM and suggest practical avenues to use automatic methods to improve memory in others. This knowledge will be useful to those wishing to increase the likelihood that information is remembered with little conscious effort on the part of the receiver.