Distraction in Visual Working Memory: Resistance is Not Futile

Abstract

Over half a century of research focused on understanding how working memory is capacity constrained has overshadowed the fact that it is also remarkably resistant to interference. Protecting goal-relevant information from distraction is a cornerstone of cognitive function that involves a multifaceted collection of control processes and storage mechanisms. Here, we discuss recent advances in cognitive psychology and neuroscience that have produced new insights into the nature of visual working memory and its ability to resist distraction. We propose that distraction resistance should be an explicit component in any model of working memory and that understanding its behavioral and neural correlates is essential for building a comprehensive understanding of real-world memory function.

Resisting Distraction Makes Working Memory ‘Work’

The temporary storage, manipulation, and use of goal-relevant information is known as working memory [1]. For working memory to ‘work’, it must be robust to distraction from irrelevant perceptual input and any internal sources of interference. Due to an abundance of research on its limitations, it may be tempting to overlook the incredible resilience of working memory in the face of distraction. Distraction resistance is a key determinant of individual working memory capacity [2], which itself is linked to fluid intelligence [3]. Working memory impairments in normal aging [4] and in clinical disorders, such as attention deficit hyperactivity disorder (ADHD) [5] and schizophrenia [6] are characterized by a specific deficit in distraction resistance. The past decade has seen an explosion of research on the neural mechanisms supporting information retention in working memory [7–12], from active recruitment of sensory regions, to representational codes varying in activation level, format, and location [9]. In this review, we highlight the resilience of working memory by characterizing the diverse mechanisms that protect it from distraction.

To provide a focused discussion, we concentrate on the impacts of delay-period perceptual distraction (see Glossary) in visual working memory. Distractor filtering during working memory encoding has a separable impact on performance [2] and has recently been detailed elsewhere [13]. While we focus on paradigms that primarily use either task-irrelevant distraction or distraction with minimal control demands, future discussions should incorporate findings from dual-task paradigms, in which distractor interference may additionally arise at the level of shared control processes. We begin by exploring the behavioral limits of distraction resistance in working memory, and outline the circumstances under which task performance is affected by distraction. Then, we discuss the current diversity of proposed mechanisms for information storage and cognitive control processes that may contribute to our robust memory abilities. By integrating multiple sources of behavioral and neuroscientific evidence for the mechanisms of distraction resistance, we hope to accelerate the refinement of theoretical models of memory and to illuminate useful hypotheses for future empirical studies of memory resilience.

Behavioral Consequences of Distraction

When, how, and to what degree distraction influences working memory are fundamental questions for understanding goal-directed behavior. Some of the earliest psychological studies of working memory focused on the behavioral consequences of distraction [14], and the finding that delay-period distraction can impair performance is so ubiquitous that it is considered a ‘benchmark’ that modern working memory models must include [15]. In fact, the obligatory encoding of distractors during maintenance is a fundamental determinant of working memory capacity in one prominent model [16]. However, the behavioral effects of distraction in the laboratory are often modest, which has stimulated ongoing debate over the neural implementation(s) of memory storage [12,17–19]. Here, we discuss some of the key factors governing whether, and how, distractors can influence working memory performance.

Distractor Congruency Impacts Forgetting

Perhaps the earliest and most consistent finding about the behavioral impact of distraction is that working memory performance is most impaired when there is high featural overlap between a memory target and distractor [15] (the ‘congruency effect’). The seminal Baddeley model of working memory, which includes separate visual and phonological stores, was based on dual-task studies that found impaired verbal working memory retention during simultaneous performance of a verbal task as opposed to a visual task, and vice versa [20]. The congruency effect, reflected in both the speed and accuracy of responses, has been replicated extensively at the level of modality(i.e., audio/visual) [21] and it also extends to the category level [22–24]. For example, working memory for faces is more impaired by other face distractors than by scene distractors [24]. A similar relationship is also present for low-level features (e.g., memory for spatial frequency is impaired by distractors differing in spatial frequency but not in orientation) [25,26]. This congruency effect has been used as behavioral evidence in support of a sensory recruitment model of working memory storage [27], which proposes that perception and memory draw upon a shared neural substrate (see ‘Sensory Recruitment’ section). This model predicts that competition over such a shared substrate and, consequently, behavioral interference, would be greatest when memoranda and distractors are most similar [28,29].

Subtle Biases from Distractors

The use of continuous report paradigms has propelled working memory research by offering a more sensitive way to assess memory quality, compared with traditional methods requiring detection of a discrete change between a sample and a test probe (i.e., ‘change detection’) [30]. This approach can reveal more subtle effects of distraction, such as systematic biases in memory reports, rather than total memory loss. A large body of psychophysical studies, many using similarly sensitive fine discrimination tasks, has established a consistent effect in which memories for a variety of low-level visual features are subtly degraded by perceptual distractors (reviewed in [12,31]). A series of more recent investigations extended these findings using mixture modeling of response error distributions from continuous report paradigms, allowing memory biases to be quantified separately from changes in memory precision or the rate of complete forgetting (random guesses) (reviewed in [30]). Such modeling has revealed that the presentation of a task-irrelevant orientation shifts a memory representation toward the distractor a few degrees in orientation space, resulting in a small attractive bias in memory (Figure 1A) [31–33], and sometimes a reduction in memory precision (e.g., [31,33]). Attractive biases have also been observed for more complex stimuli, such as human faces [34]. These biases typically increase in magnitude with increasing distance in feature space between the target and distractor, within a limited range of distances [31], beyond which biases plateau [31,34] or return to baseline [35,36]. At large target–distractor distances, increased rates of random guessing have been observed, indicating that distractors can sometimes cause items to be forgotten entirely (Figure 1B) [33]. Furthermore, biases tend to occur only when distractors differ from memory items along a task-relevant feature dimension [26], which suggests an interaction at the level of feature-selective channels in visual brain regions [35] and provides further behavioral support for a shared neural resource between perception and memory [18]. Finally, biases can also arise from internal sources, such as another memory item maintained on the current or previous trial (Box 1).

Figure 1. Consequences of Distraction in Working Memory.

The diagram shows an item (orange disc) under the focus of attention (yellow spotlight) in working memory along with an incoming perceptual distractor (blue disc). The distractor can distort the memory, leading to biased responding (A) or degrade the memory, leading to forgetting (B).

Box 1. Internal Sources of Distraction.

Working memory resilience is also shaped by at least three internal sources of interference: lingering working memory representations from the immediate past; concurrently maintained working memory representations; and mind wandering. Trial history effects, wherein previously relevant working memory representations impact current ones [92], include proactive interference [121] and serial dependence [122]. Serial dependence has recently been framed as a potentially adaptive version of proactive interference, and may reflect processes supporting stability and continuity in perception, thought, and behavior [123]. Computational modeling [124,125] and recent neural evidence [126] suggest that serial dependence arises from a combination of sustained firing and synaptic augmentation supporting working memory maintenance. Task demands may shape the engagement of control processes and the resulting serial dependence effects, for example based on needs for continuity or differentiation [123]. Furthermore, active suppression [86] or unbinding [127] of irrelevant representations may mitigate proactive interference effects.

Competitive interactions between multiple items held in working memory is another source of distraction [128] and has a prominent role in models of capacity limitations [46,129]. Whereas it is common for a perceptual distractor to attract a memory item, interactions among memory items can lead to either attraction or repulsion effects, depending on the distance in feature space between the items. Repulsion tends to be seen when memory items are similar and need to be differentiated, while dissimilar memory items lead to attractive biases [130], potentially due to an influence of ensemble statistics on memory responses [131]. Recent proposals have highlighted this complementary system of biases, arguing that these memory ‘errors’ may arise from largely adaptive processes [123].

A final source of internal distraction can arise from task-unrelated mind wandering. A recent study found that participants reported mind wandering ~25% of the time in a working memory task, and mind wandering frequency was negatively correlated with working memory capacity [132]. Interestingly, participants still achieved ~80% change detection accuracy, even when they reported mind wandering. This provides further evidence that sustained attention is not always required for successful working memory maintenance [83]. Nonetheless, indulging task-unrelated thoughts may impair working memory by consuming attentional resources that would otherwise support maintenance processes [133]. To elucidate the neural processes supporting working memory in the presence of internal distraction, it would be useful to better characterize the impact of mind wandering, for example to assess whether participants show increased categorical biases [21] or reductions in memory precision.

Attentional Prioritization Influences Distractibility

A ‘spotlight’ of internal attention can be directed toward a subset of remembered information, resulting in memory representations that vary in attentional priority. Priority can be manipulated experimentally with a cue indicating the relative value of an item or the likelihood that it will be tested [37]. Such priority manipulations have yielded mixed effects on distractor susceptibility.

When a postencoding retrocue indicates with complete certainty which item will be tested, allowing extraneous information to be removed from memory, the cued item benefits from improved resilience to both delay-period distraction [38–42] and test interference [43–45]. Such findings accord with prominent working memory models that include a limited but protective ‘focus of attention’ that shields memories from distraction [46]. However, when multiple items are assigned graded priorities, either with a pre-cue [47–49] or based on recency in sequential encoding [48–50], higher-priority items do not experience the same benefit. Instead, distraction reduces performance more for higher-priority than lower-priority items [47–50]. Although this could reflect a floor effect in which low-priority information does not receive a cueing benefit that can then be reduced by distraction, a review of these findings suggests that this simple interpretation is unlikely [51]. Instead, it argues that prioritized memories are vulnerable to interference from perceptual input that enters and competes with information in the limited focus of attention (Box 2).

Box 2. Interactions between Perception and Memory.

We have reviewed ample evidence that task-irrelevant perceptual input can impact working memory. This is one of many pieces of evidence suggesting a shared resource between perception and memory, as proposed by the sensory recruitment model. Furthermore, this interaction is bi-directional; a large body of work has found that the contents of working memory also influence perception [134,135]. Perhaps the clearest example is attentional capture, whereby the contents of working memory bias perceptual attention towards matching stimuli in the environment [136]. This bias from memory contents even mirrors the inhibitory surround pattern of visual attention, in that perception is biased towards memory-related features in the environment but shows a small bias away from items just nearby in feature space [137]. Furthermore, the classic Stroop effect (color words presented in an incongruently colored font take longer to read) is also present when the incongruent color is instead held in memory [138,139].

Finally, the contents of working memory can change how visual features are perceived [140,141] through an alteration of neural response profiles in visual areas [67,70,142]. Notably, the influence of working memory on perception is not equal for all memory items, and neither is working memory equally influenced by all items in the visual field. The attentional capture and related literature provides support for a bi-directional relationship that is restricted to interactions between supraliminal perception and prioritized working memory items. Attentional biases induced by working memory are largely restricted to a single prioritized memory item [94,143,144]. Working memory content influences what is consciously perceived in the environment [145] and, analogously, only consciously perceived distractors influence working memory [31,146]. These findings help to clarify that distraction, therefore, is also bidirectional. Memory can be disrupted by the concurrent processing of perceptual inputs and perceptual attention can be disrupted by the active maintenance of information in working memory.

These seemingly divergent findings might result from differences in maintenance processes induced by the varied cueing procedures [52,53]. One of the most crucial differences may be whether the cue allows some information to be removed from memory entirely, or whether low-priority information must be retained. Removing low-priority information would free memory capacity and allow attentional resources to be redirected toward supporting persistent memory representations (see ‘Sensory Recruitment’ section) and inhibiting distractors (see ‘Control Processes’ section). Therefore, a unified account of these findings could be that the vulnerability of a memory representation depends both on its level of activation (higher priority yields higher activation and more interactions with ongoing perception; Box 2) and the availability of control processes to reduce distractor influence (higher-priority information receives more protection if lower-priority information can be discarded [41]). Testing these predictions is an important direction for future research, which should examine the neural impact of distraction on both high- and low-priority memories, and the specific relationship between neural activation and distraction susceptibility (see ‘Activity-Silent Storage’ section).

Mechanisms of Storage and Protection

The past decade of visual working memory research has produced evidence for a dizzying array of storage substrates and coding mechanisms. Here, we detail the empirical evidence for a few of these mechanisms and discuss how each may uniquely support distraction resistance. Given that much of this research has utilized paradigms without an explicit distraction component, we highlight potential opportunities for future research to more directly link the behavioral and neural consequences of distraction.

Persistent Neural Activity

Sustained delay-period firing of prefrontal cortical (PFC) neurons has long been considered a hallmark of working memory [54]. The PFC has been implicated in stabilizing persistent activity in the face of distraction [55], leading to claims that such activity is both necessary and sufficient for memory maintenance [56] (Figure 2A, Key Figure). This claim draws upon the observation in nonhuman primates that persistent spiking in PFC tends to survive distraction more successfully than spiking in sensory neurons (i.e., those in inferior temporal cortex) [57]. More recent work suggests that robust persistent activity is not exclusive to PFC, because neurons with this property have also been identified in the parietal cortex [58]. Similarly, in humans, distraction-resistant delay activity has been identified in a superior portion of the intraparietal sulcus that is also sensitive to working memory load [59], inspiring an argument that this region might be the primary site of behaviorally meaningful working memory storage. Differences between studies in whether persistent activity is most distraction resistant in PFC or parietal cortex may depend on the type of information being remembered (e.g., spatial versus nonspatial) [12].

Sensory Recruitment

However, the past two decades have witnessed the rise of an expanded account, the sensory recruitment hypothesis [27,60], which argues that persistent activity in the PFC does not always reflect working memory storage per se. Instead, PFC delay activity often lacks fine stimulus specificity and, therefore, likely reflects top-down control processes supporting active information storage in posterior sensory regions [61]. Despite an argument that active mnemonic representations in sensory regions would be unlikely to survive subsequent perceptual input [12], simultaneous representations of both working memories and sensory inputs can be observed in early visual cortex during distraction [59,62]. Such distractor-resistant sensory representations may specifically depend upon sustained top-down support from the PFC [63]; early fMRI work found that middle frontal gyrus activity before distraction predicted whether memory performance would be impaired [64]. More recently, nonhuman primate multiunit recordings in primary visual area V1 revealed that top-down inputs evoked persistent delay-period spiking, providing support for the sensory recruitment model [65]. A brief visual distractor temporarily interrupted this V1 activity, but only modestly reduced behavioral accuracy. This suggests that, while persistent spiking in V1 may support maximal performance, it is unlikely to be the sole currency of working memory storage.

In addition to supporting active firing in sensory regions, persistent top-down influences may also sustain subthreshold changes in membrane potentials [66] and/or modify the response properties of visual neurons [67]. Given that fMRI detects both spiking activity and synaptic input [68], the signal giving rise to decodable working memory representations in sensory regions may in part reflect these top-down modulations [69]. Representations maintained via subthreshold modulations may be less susceptible to distractor interference than spiking-based sensory codes [69], while still allowing the system to detect memory-matching input [70]. Interference may be further reduced by the partial segregation of mnemonic and perceptual representations to different cortical layers; in line with findings in nonhuman primates [65], high-resolution layer-specific fMRI of human V1 identified stimulus-specific mnemonic representations in deep and superficial layers only, while bottom-up perceptual input initially enters at middle layers [71]. Given that this segregation is not observed in area V2 and beyond [71], distractor-induced behavioral biases [31,34,72] likely arise from neural interactions at these sites of overlap.

Key Figure

Storage Mechanisms for Resisting Distraction

Figure 2.

Different coding schemes support distraction-resistant representations in working memory: (A) persistent; (B) dynamic; (C) parallel; and (D) silent. The diagrams have the same conventions as described in Figure 1 in the main text.

Dynamic but Stable Coding

Working memory representations need not be static; they may evolve over time, with contributions from different subpopulations of neurons [10] that together provide stable coding at the population level [73,74]. There is evidence for such dynamic coding in numerous regions, including the lateral PFC (e.g., [75,76]) and sensory regions (e.g., [77]). Importantly, mnemonic representations can persist in lateral PFC even when delay-period distractors cause changes in population codes [75], potentially because PFC neurons display mixed selectivity that allows them to respond to distractors while still maintaining remembered information in a stable lowdimensional representational ‘subspace’ [76]. Thus, distraction resistance in working memory could be provided by population codes that bend but don’t break (Figure 2B). In addition to supporting distraction-resistant storage within PFC, such a stable prefrontal code could provide consistent top-down influences to help representations in sensory regions bridge distracting input (see the ‘Sensory Recruitment’ section).

There is growing evidence that mnemonic codes also morph systematically as a function of task priority [78]. Recent EEG work observed such ‘priority-based remapping’ in which active coding for a memory item is transformed into an alternate representational code when it has low priority, and then back again when it becomes task relevant [79]. Such remapping could be useful for protecting the low-priority information from perceptual distractors [80] and interference from other items in memory (Box 1). Therefore, the degree to which a given task is conducive to priority-based remapping may dictate whether low-priority information is more or less susceptible to distractor interference (see the ‘Attentional Prioritization’ section).

Parallel Coding in Distributed Regions

In addition to persistent stable or dynamic coding within a single neural substrate, distraction-resistant working memory may also be supported by parallel coding of memoranda in multiple brain regions (Figure 2C). Much prior work found evidence that working memory storage is distributed throughout the brain (reviewed in [7]), although these representations may vary in format and quality. For instance, sensory representations may be analog and more precise, while representations in parietal and prefrontal regions may be coarser, more abstract/categorical [81] or transformed into action-based codes [82]. This division of labor may have special relevance for distraction resistance [59]. For example, parallel orientation codes were recently found in early visual cortex and the intraparietal sulcus in anticipation of potential distraction [32]. While early visual representations were substantially biased toward delay-period distractors, behavioral reports reflected a more modest bias. The parietal representations, which were unbiased and persisted only when distraction occurred, appeared to mitigate its behavioral impact [32]. Although these and other recent findings (e.g., [62]) suggest that both parietal and visual cortex representations support working memory for orientations, other researchers have argued that visual areas are unnecessary for visual working memory storage [12,17]. This argument is motivated by their finding that parietal representations were more consistently insensitive to distractors [59]. A productive path forward in reconciling these perspectives will be to systematically evaluate how parallel mnemonic representations in multiple regions are integrated to give rise to a single behavioral output, and how anticipating or responding to distractor interference may flexibly shape the utilization of these multiple mnemonic codes.

Activity-Silent Storage

As previewed earlier, stimulus-specific mnemonic representations may not always require active neural spiking. Indeed, the decodability of memory contents can vary as a function of relevance, with currently irrelevant representations temporarily assuming a ‘latent’ state that may reflect the withdrawal of internal attention [83]. Such functionally latent representations may correspond to ‘activity-silent’ storage mechanisms [84], such as short-term synaptic plasticity [85]. Importantly, these latent representations can be fully reinstated in active coding when they become task relevant [86,87], even following a distracting probe of another memory item [83,88,89]. Top-down inputs may specifically rely on the existence of activity-silent synaptic traces to reinstate active coding post distraction [90]. Although direct empirical evidence of activity-silent storage is still limited, indirect evidence is provided by the finding that a brief visual impulse [91] can evoke stimulus-specific activity corresponding to unattended, putatively activity-silent, memory representations. One potential reason for the success of this approach is that the perturbation may act like a distractor, triggering a reflexive top-down reorienting that reactivates the latent memory representation (see also [92]).

A recent theoretical model of working memory control also includes a role for synaptic weightbased memory representations; in this model, top-down oscillatory influences allow rhythmic bursts of spiking to refresh synaptic representations [10]. Compared with sustained spiking, this temporal sparsity is hypothesized to yield improved resistance to interference. However, further experimental work is needed to determine whether these cortical dynamics and synaptic mechanisms operate at the level of top-down influences on sensory regions, or whether they are a more limited local characteristic of PFC [93].

There are theoretical reasons to expect that activity-silent mnemonic representations would be particularly robust to distractor interference (Figure 2D). Silent memories do not exert biases on perception [94] nor can they be manipulated without prior reactivation [95,96]. Thus, silent coding appears to provide passive storage, but not other key properties commonly ascribed to working memory. If information cannot be intentionally manipulated while in a silent state, it also likely cannot be unintentionally disrupted by perceptual input. Initial empirical support for this idea includes observations that low-priority items in working memory can be less vulnerable than high-priority items to behavioral interference [51] and show more robust neural recovery from distraction [97] (see ‘Attentional Prioritization’ section). Future work is needed to provide direct evidence for activity-silent coding, to establish where in the brain such codes may exist, and to test whether it confers protection from distractor interference. It will be crucial to differentiate silent coding from other processes that could yield similar reductions in the ability to decode working memory representations in visual cortex, such as recoding unattended information into different regions [98] or into alternate active coding schemes (such as priority-based remapping discussed earlier [79]), or recruitment of long-term memory mechanisms [99].

Control Processes for Distraction Resistance

There is considerable evidence that control processes directly limit the detrimental effects of perceptual distractors, in addition to supporting memory maintenance (see the ‘Sensory Recruitment’ section). Here, we outline evidence for control mechanisms that actively inhibit distractor processing and prevent unwanted information from being encoded into memory. We close with a brief discussion of the ways in which top-down control processes can flexibly shape storage processes in preparation for anticipated distraction.

Top-Down Inhibition of Distractors

One role of frontoparietal control processes is in actively inhibiting sensory processing of distractors [100] (Figure 3A). For example, in nonhuman primates performing a delayed saccade task with distraction, the lateral PFC shows both anticipatory firing rate suppression and reduced distractor-evoked visual responses [101]. Furthermore, pharmacological inactivation of this region impairs behavioral distraction resistance, implicating its causal role [101]. Top-down distractor inhibition may be accomplished through coordinated modulations of frequencyspecific oscillations [102]. For instance, preparing to ignore an anticipated distractor increases midfrontal delta/theta (~2–7 Hz) [103] and posterior alpha power (8–12 Hz) [103–105], likely reflecting a top-down control process that reduces sensory processing of distracting information [106]. Recent evidence suggests that there is an intrinsic rhythmic (~2.5 Hz) fluctuation in behavioral susceptibility to delay-period distractors [107], which may result from such a rhythmic top-down control process. Furthermore, PFC beta (13–30 Hz) bursts have also been associated with protecting working memory contents from distraction [103]. Finally, many of these control mechanisms have also been shown to support distractor filtering during working memory encoding (see [13] for a comprehensive review).

Figure 3. Control Processes for Resisting Distraction.

Top-down control processes support distraction resistance in working memory by limiting the perceptual processing and encoding of distractors. (A) Inhibition; (b) gating. The diagrams have the same conventions as described in Figure 1 in the main text.

Blocking Distractor Encoding into Memory

Beyond inhibiting sensory processing of distractors, control processes may block task-irrelevant input from being encoded into working memory and interfering with (or displacing) current representations (Figure 3B). The basal ganglia (particularly the caudate nucleus of the striatum) are thought to have an integral role in this process, termed ‘input gating’ [108]. Here, the striatum is conceptualized as a gate that can switch between open and closed states, controlling access to working memory via its reciprocal connections with the lateral PFC [109]. Distractors are prohibited in the default striatal ‘closed’ state, but when the gate is opened by activation of striatal dopamine D2 receptors, new sensory input can update or modify the contents of working memory [110]. Consistent with this model, pharmacological modulation of striatal D2 receptors affects distraction susceptibility [111]. Furthermore, patients with ADHD [112] and schizophrenia [113] frequently exhibit striatal dysfunction and correspondingly impaired distraction resistance.

Prefrontal dopamine has also been implicated in distraction resistance; the relative balance of dopamine D1/D2 receptor stimulation may toggle the PFC between modes of stable distractor-resistant maintenance and malleable working memory updating [114]. In nonhuman primates, stimulating PFC dopamine D1 receptors modulates postdistraction recovery of PFC memory coding, with the direction of the effect dependent on the cell type being stimulated (pyramidal cells versus interneurons) [115]. In humans, pharmacologically activating dopamine D2 receptors increases the behavioral distractor congruency effect, corresponding with decreased PFC activity and reduced functional connectivity between the dorsolateral PFC and stimulus-selective extrastriate regions [116].

Together, these results suggest that coordinated, dopamine-dependent fronto-striatal activity determines the nature of top-down control inputs to sensory regions. These inputs, in turn, govern whether existing stimulus-specific representations persist through distraction, or whether they risk disruption or competition from the encoding of new information into memory.

Anticipating Distraction Mitigates Its Impacts

Successful anticipation of distraction should allow optimal engagement of these control processes to mitigate the effects of interference. While predictable distractors can still capture spatial attention, their anticipation allows active maintenance of memoranda (as indexed by EEG contralateral delay activity) to persist longer through distraction [117]. This suggests that anticipating distraction allows the recruitment of maintenance processes that do not rely on continued spatial attention, such as activity-silent synaptic mechanisms or contributions from long-term memory [118]. Previous work found that a predictable increase in task difficulty (including the addition of distraction demands) boosts the incidental encoding of information into episodic memory [119]. Furthermore, anticipating an intervening task during a memory delay increases categorical biases in memory for precise, continuous features, suggesting a shift toward a coarser, potentially more abstract, coding strategy [21]. Future work should investigate whether this preparatory shift in coding corresponds to greater use of categorical codes in regions such as the intraparietal sulcus [32,59,98], and/or contributions from verbal working memory or episodic long-term memory. Finally, these control processes unfold over time; the timing of distraction relative to memory encoding and the response probe [120] likely provides an important constraint on the degree to which interference can be mitigated.

Concluding Remarks and Future Directions

Protecting goal-relevant information from distraction is critical for successful cognitive function. Indeed, the limited-capacity working memory system is remarkably robust; interference from perceptual input typically incurs only subtle degradation, and more rarely catastrophic loss. The neural mechanisms by which working memory resists disruption have been studied for over half a century, but methodological and analytical advances in cognitive neuroscience over the past decade have produced new insights into the nature of working memory storage and its protection from interference. These include observations of dynamic morphing of memory representations within an ensemble of neurons in the PFC, anticipatory and parallel coding of memoranda in frontoparietal and sensory regions, transformations between different functional states of memory representation in sensory cortex, and active inhibition of distractors by frontoparietal control regions. It is clear that ‘distraction resistance’ is not a monolith, but a multifaceted collection of processes that protect goal-relevant information from interference. Therefore, it will be most fruitful to embrace this diversity and seek to understand how memory resilience may derive from multiple sources, with control processes that orchestrate their flexible engagement in response to situational demands (see Outstanding Questions). More broadly, memory resilience should be an explicit component in any model of working memory, and efforts to link the behavioral and neural correlates of distraction will be essential for building a comprehensive understanding of real-world human working memory function.

Outstanding Questions.

What are the clear predictors of the extent to which a memory representation will be biased by a distractor?

How does distraction affect working memories maintained via activity-silent mechanisms? Are they biased by perceptual distractors, such as those maintained with persistent firing?

What are the contributions of top-down modulatory processes (e.g., shifts in receptive fields or subthreshold modulations) to the observed tuning in distractor-induced behavioral biases?

How does the predictability of distraction impact the utilization of different neural mechanisms of memory storage and/or distractor inhibition?

Does categorical coding in anticipation of distraction correspond to increased reliance on more abstract representations in frontoparietal regions?

How do control processes that prepare for potential distraction relate to those that orchestrate recovery from unanticipated interference? What are the temporal dynamics of these processes?

Is distraction resistance a trainable skill? Is any one mechanism (e.g., parallel coding) more amenable to training?

To what extent does mind-wandering influence working memory representations?

If parallel mnemonic codes are used to support distractor resistance, what determines which representation should ultimately guide behavior?

Highlights.

A complete understanding of working memory function in a world full of distraction requires a careful examination of the neural evidence for mechanisms of distraction resistance.

Recent advances in testing and modeling behavioral distraction effects in working memory have led to a more nuanced understanding of the ways in which representations can be affected by task-irrelevant information.

Sophisticated measurement methods and analyses that can track the contents of working memory have revealed a diverse array of storage mechanisms and neural substrates, which may differ meaningfully in their distraction resistance properties.

Control processes that actively inhibit distractor interference and flexibly adapt storage processes are key contributors to the remarkable resilience of working memory performance.

Glossary

Activity-silent storage

working memory maintenance hypothesized to occur without continuous active firing, potentially via short-term synaptic plasticity that yields temporary item-specific changes in the configuration of a network of neurons

Attractive bias

a systematic shift in a memory response toward the feature value of a distractor. For example, a green distractor would cause memory for the color yellow to be reported as greenish-yellow.

Congruency effect

memory disruption is more profound when distractors are more similar (e.g., from the same modality) to memory items; distinct from attractive biases, which operate within a single feature dimension (e.g., color).

Continuous report paradigm

a class of working memory tests that offers a more detailed measure of memory quality, compared with discrete ‘change detection’ tasks. A remembered feature value (e.g., reddish-orange) is reported by choosing it from a continuous space that spans all possible values (e.g., all 360° of color space); also called ‘delayed-estimation’.

Delay-period perceptual distraction

distractors presented during a working memory retention interval that provide visual input (e.g., pictures of faces), but do not require the completion of a secondary task.

Dual-task paradigm

a primary working memory task must be completed along with an interposed task. To the extent that shared control processes are required for both tasks, distractor interference can arise at the level of memory representations, and/or at the control level.

Input gating

process by which new information is selectively encoded (or blocked from encoding) into working memory, so that limited capacity can be preserved for goal-relevant information. Gating policies are thought to be set via reinforcement learning mechanisms

Mixture modeling

response errors in a continuous report task are modeled as deriving from a mixture of multiple distributions. In the simplest case, one distribution reflects responses to items that were successfully remembered, and another reflects random guesses about items that were forgotten.

Proactive interference

interference of previously learned information in the acquisition and retrieval of newer information.

Representational subspace

population-level memory representations are very high dimensional if each neuron carries unique information. However, if the same information is represented across multiple neurons, the representation can be described with fewer dimensions, that is, a ‘subspace’ that still contains the remembered information.

Retrocue

a ‘retrospective cue’ that directs internal attention to a subset of items in memory. Given that the cue appears during the retention interval, any differences between cued and uncued items cannot reflect differences in encoding quality.

Sensory recruitment

a theory about the neural substrate of working memory storage: the same cortical regions responsible for perceptual processing are recruited for high-fidelity memory maintenance

Serial dependence

perception and maintenance of novel stimuli are influenced by recently perceived stimuli, that is, consecutive trials are serially dependent on one another.