The what, where and how of delay activity

Abstract

Working memory is characterized by neural activity that persists during the retention interval of delay tasks. Despite the ubiquity of this ‘delay activity’ across tasks, species and experimental techniques, our understanding of this phenomenon remains incomplete. Although initially there was a narrow focus on sustained activation in a small number of brain regions, methodological and analytical advances have allowed researchers to uncover previously unobserved forms of delay activity across the entire brain. In light of these new findings, this Review reconsiders what delay activity is, where in the brain it is found, what roles it serves and how it may be generated.

Introduction

To follow a conversation, you must mentally represent the overall topic, what was said in the last sentence and what you intend to say next. Critically, these representations need to be connected despite occurring seconds apart. This ability to link multiple events over brief intervals is essential for cognition and is the core feature of working memory (WM) — the set of operations that support the temporary retention of behaviourally relevant information. One of the enduring aims of neuroscience is to understand the neurobiology underlying WM¹.

In 1971, Fuster and Alexander² and Kubota and Niki³ described the activity of individual neurons in the lateral prefrontal cortex (lPFC) of macaque monkeys performing a task that required storing information over a delay of several seconds and using this memory to guide a response. Strikingly, lPFC neurons remained active during the memory delay when no stimulus was present, bridging the temporal gap between perception and the contingent motor response. This phenomenon was later termed ‘delay activity’⁴, and is thought to reflect the sustained representation of WM content or WM-related goals⁵. Delay activity is seen more generally in contexts that require an organism to link a sensory stimulus to a delayed behaviour, such as during tasks of sustained attention⁶ or decision making⁷; here, however, we examine delay activity in tasks explicitly testing WM, where it has been characterized most extensively.

In recent years, a wealth of theoretical and experimental work has prompted an expanded consideration of delay activity, its function and the mechanisms that generate it. This Review integrates findings across subfields of neuroscience to critically evaluate the what, where and how of delay activity. There are two key areas of emphasis. First, in contrast to discussions of delay activity that typically focus on increases in spike rate or functional MRI (fMRI) signal that persist throughout a WM delay, we consider delay activity to be any task-related change in neural activity that spans the interval between a stimulus and the behavioural response in a WM delay task. This expanded definition allows for a more inclusive and comprehensive evaluation of delay activity, but at the same time calls attention to the challenge of comparing data across different experimental methods, brain regions and species. We address this issue by highlighting studies that record delay activity across multiple brain regions in the same experiment. Second, we emphasize points of connection between empirical studies and neural models of delay activity.

What is delay activity?

Delay activity can be observed in many forms at various scales — from individual neurons, to cell assemblies comprised of tens of neurons, meso-scale interregional circuits or macro-level networks that recruit regions throughout the brain. Recording delay activity at these different scales requires multiple techniques, each of which is sensitive to different aspects of the underlying neuronal signals (see Box 1 for a discussion of non-neuronal contributions to delay activity) and has a unique set of limitations. By considering these many forms of delay activity (Fig. 1), we aim to provide a more complete picture of its purpose and origin.

Box 1 |. Do astrocytes contribute to delay activity?

Astrocytes, the most abundant glial cell in the brain, are comparable in number to neurons²⁵¹. Classical views of astrocytic function proposed supportive roles, including nutritional support of neurons, the maintenance of ion concentrations in the extracellular space, support of the blood–brain barrier and the repair of injured brain tissue²⁵². In the 1990s, astrocytic function was reconsidered, with the idea of the ‘tripartite synapse’ put forth based on empirical evidence that astrocytes can integrate neuronal inputs and modulate synaptic activity²⁵³. Particularly in light of recent findings linking blood-oxygen-level-dependent responses and astrocytic function⁶⁵, as well as increased appreciation for the role of astrocytes in cognition²⁵⁴, one must consider how astrocytes might contribute to delay activity. Although there is currently no published empirical data linking astrocytes to delay activity, three features make astrocytes an intriguing candidate for participating in delay activity.

First, and most importantly, although astrocytes respond to fast neural dynamics, they process information over a much slower time scale than do neurons (for example, over tens of milliseconds to seconds versus sub-millisecond to milliseconds)²⁵⁵, suggesting that they may promote prolonged brain states such as sleep, potentially by modulating neuronal activity via gliotransmission²⁵⁶. Indeed, slow astrocytic dynamics may overcome some of the challenges for sustaining information over seconds that are imposed by the short time constants of neurons.

Second, as astrocytes enwrap nerve terminals, they are perfectly positioned to sense and emit informative signals from and to synapses²⁵⁷. They can sense neural activity and are activated during synaptic transmission and can in turn modulate neuronal activity by releasing transmitters such as glutamate²⁵⁸. For example, cystolic calcium levels in astrocytes of the mouse barrel cortex increase following whisker stimulation, in line with the frequency of stimulation²⁵⁹. Likewise, in ferret visual cortex, astrocytes respond to visual stimuli, and show tuning to stimulus features such as orientation and spatial frequency²⁶⁰. This modulation of activity in astrocytes as a function of stimulus properties suggests that they may be able to encode information about memoranda.

Third, in addition to being involved in localized neuronal activity, astrocytes also regulate network-level activity. Inhibition of glutamate release by astrocytes significantly reduced the duration of gamma oscillations in hippocampal slices, as well as gamma power in awake behaving mice, which corresponded with behavioural performance on a recognition memory task²⁶¹. Thus, gliotransmission may potentially similarly contribute to gamma oscillations in WM.

These new insights into astrocytic function have led to the proposal that astrocytes may have a role in WM. One study provided a computational model that simulated individual cortical neurons in a delay task. Bistable delay spiking emerged from short-term (on the order of several seconds) synaptic facilitation that was initiated by the cue stimulus and mediated by astrocytes²⁶². This model also recapitulated the finding of irregular patterns of delay spiking by individual neurons (see main text). Thus, astrocytes may have an essential role in WM by promoting cellular bistability via synaptic plasticity. Future models of WM should take astrocytic function into account.

Fig. 1 |. Schematic examples of different types of delay activity.

Here we use schematics to highlight two properties of delay activity. First, delay activity can be stable in time (top row), or it can be temporally irregular or dynamic (bottom row). Second, delay activity can be measured within or averaged over individual neurons, voxels or electrodes (left column), or measured as the combined response across populations of neurons/voxels/electrodes (right column). a | Temporally stable delay activity in individual neurons or blood-oxygen-level-dependent (BOLD) response averaged over functional MRI voxels. Activity remains elevated above baseline throughout the delay period, signifying the sustained representation of information. b | Temporally irregular responses in individual neurons or electrodes. Individual neurons can display spiking activity that varies over the course of the delay (schematicized at the top of this panel). Recent results demonstrate intermittent bursts in the LFP signal throughout the delay (schematicized at the bottom)⁴¹. c | Stable population coding. WM information (for example, WM content, task rules or planned responses) can be decoded from the combined activity of populations of neurons or voxels. A pattern classifier can be trained and tested on independent data sets recorded during a WM task. Above-chance classification accuracy indicates that a representation of the information being classified exists in those neurons or voxels. The classifier can be trained on data from a specific time (for example, during the early delay) in the trial, and tested (on independent data) from the same time point or different time points, allowing one to measure the stability of the representation over the course of WM maintenance. In stable population codes, the pattern of activity that encodes specific information at any given time point is the same as the pattern of activity that encodes that same information at any other time point; thus, a classifier trained at one time point will be more accurate than chance at other time points throughout the trial. This is indicated by above chance classification at each time point regardless of which time point or period (for example, those represented in light blue, red or green) the classifier is trained on. d | Dynamic population coding. In contrast to stable population coding, information encoded in the population activity at a certain time point is encoded in different forms of activity at other time points. A classifier trained on a time point will therefore perform successfully at that time point, but not at other time points in the trial. This time-dependent classification is indicated by above-chance classification accuracy along the diagonal of the training-by-testing matrix (top; compare to the matrix in c), but chance classification elsewhere in the matrix, and above-chance classification accuracy that is limited to brief periods of time for classifiers trained on the sample, delay and response periods (bottom; light blue, red and green, respectively).

Neuronal spike rate

Extracellular recordings have documented increases in spike rate (relative to a pre-trial baseline) that persist throughout WM delays; for brevity, this finding is referred to as ‘delay spiking’ (Fig. 1a). Delay spiking is found in rodent PFC^8,9; in various regions in the brains of non-human primates (NHPs), including prefrontal, parietal and sensory cortices^2,3,10–14; and in human medial temporal lobe (MTL)^15,16. Notably, delay spiking persists beyond the tens of milliseconds over which a neuron integrates its synaptic inputs¹⁷, indicating that it may involve complex network dynamics^18,19 or intrinsic cellular properties²⁰ that can sustain spiking in the absence of external stimulation (see below). Studying spike rates enables specific hypotheses about the cellular basis of delay activity to be tested (Box 2) and can provide insights into local network properties that support WM.

Box 2 |. The cellular basis of delay activity.

Computational models have long theorized that NMDA receptors (NMDARs) have a prominent role in delay spiking^216,247,263. The relatively slow excitatory dynamics of NMDAR activation can provide a stabilizing influence on attractor networks, allowing these networks to sustain a representation through persistent spiking over several seconds²⁶³. These models accord well with behavioural evidence that systemic administration of NMDAR antagonists disrupts WM in rats^264,265, non-human primates (NHPs)²⁶⁶ and humans^267–269, whereas perfusion of NMDA in monkey visual cortex improves WM performance²⁷⁰. This hypothesized role of NMDARs in delay spiking was recently confirmed empirically: blockade or antagonism of prefrontal NMDARs eliminated delay spiking in NHPs²⁷¹. This effect was specific to the NR2B subunit, presumably owing to its slow kinetics^17,272 (although see Ref.²⁷³ for a discussion of the importance of NR2A versus NR2B subunits).

NMDARs have also been implicated in fMRI and EFP measures of delay activity. Systemic administration of the NMDAR antagonist ketamine in humans reduced BOLD responses in lPFC during the early delay period in a WM task²⁶⁹. Moreover, NMDAR dynamics, in concert with the contributions of AMPA receptors to recurrent excitation, have been found to give rise to sustained gamma frequency oscillations during the delay period in an attractor network model¹⁹, consistent with the role for NMDARs in generating gamma oscillations²⁷⁴. NMDARs may also help to mediate neuromodulatory effects on WM. Stimulation of dopamine D1 receptors facilitates the activity of NMDARs in prefrontal neurons in vitro²⁷⁵, which may be the mechanism by which dopamine stabilizes or enhances delay spiking in lPFC^276,277, thus promoting more stable WM²⁷⁸.

In addition to NMDARs, hyperpolarization-activated cyclic nucleotide-gated (HCN) channels — specifically, HCN1 channels — have been implicated in delay spiking, although their exact role is debated. One study in NHP lPFC demonstrated that α2A-adrenoreceptor antagonism indirectly potentiates HCN1 channels, leading to an attenuation of delay spiking, whereas HCN1 blockade enhances delay spiking²⁷⁹. By contrast, a study of mouse PFC neurons in vitro showed that HCN1 channels were crucial for delay activity; facilitation of the HCN1 channel-mediated current resulted in intrinsic (that is, connectivity-independent) persistent spiking^280,281 (see Ref.²⁸² for a model that examines the influence of HCN channels on network and cell-intrinsic mechanisms of persistent activity). HCN1 channel activity is influenced by dopamine, noradrenaline and acetylcholine^{280,281–283}, making these channels another important target for understanding neuromodulatory effects on delay activity.

There are two caveats to this approach. First, although spiking studies can elucidate how individual neurons encode WM information, analyses of networks of neurons are increasingly considered to be more informative for understanding brain function^21,22. Important information is also encoded in large-scale network activity²³. Unfortunately, concurrent recording of spiking of meso- and macro-scale populations remains relatively rare (but see below). Second, the extensive task and stimulus training that animals undergo in such studies increases delay spiking in NHPs^24,25 and reduces the correlation between delay spiking and behaviour in mice²⁶. Thus, despite some evidence for sustained delay spiking in the absence of training or familiarity in the human MTL^15,16, the potential influence of task and stimulus familiarity should be considered when interpreting studies of delay spiking.

Extracellular field potentials

Extracellular field potentials (EFPs) represent neural activity summed over many neurons and can be measured intracranially with microelectrodes as local field potentials (LFPs); with electrodes on the surface of the brain using electrocorticography (ECoG); or with electrodes or sensors on the scalp using electroencephalography (EEG) or magnetoencephalography (MEG). In the delay period of WM tasks, sustained increases in EFP amplitude are found in human EEG^27,28 and NHP LFP²⁹ recordings. Moreover, enhancements in the oscillatory power of the EFP in the gamma band (40–100 Hz)^30–32, beta band (15–30 Hz)³³ and theta band (4–8 Hz)³⁴ are also sustained during WM delays. Interactions between oscillatory power in the beta and gamma frequency bands have been proposed to underlie top-down control of WM content³⁵. Human studies have also reported delay-period increases^36,37 and decreases^38,39 in oscillatory power in the alpha band (8–12 Hz) over posterior electrodes. Alpha attenuation may be related to the maintenance of task-relevant stimuli, whereas alpha enhancement may reflect the inhibition of the processing of task-irrelevant stimuli⁴⁰. In addition to sustained modulation, discrete bursts of gamma oscillations in NHP lPFC are thought facilitate the storage of WM information^41,42.

Despite controversy over the relative contributions of spiking and synaptic activity to the LFP signal⁴³ and the difficulty of localizing EEG or MEG signal⁴⁴, there are several features of EFPs that make them a valuable measure of delay activity. First, EFPs are straightforward to measure and compare across species. For example, one study noted that delay period EFP signals from NHP lPFC and human parietal EEG electrodes have very similar features²⁹. Second, LFP power often predicts behaviour in NHPs as well as or better than single-neuron spike rates^45,46. Third, EFPs measure micro-, meso- and macro-level phenomena, and thus may help to bridge our understanding of detailed stimulus-coding properties of neurons and the activity of large-scale brain networks. At the micro-level, LFPs coordinate local spiking activity^47–50, and different oscillatory frequencies are linked to unique cellular and local circuit mechanisms^51–53. At the meso- and macro-levels, EFPs may coordinate activity over local and brain-wide networks⁵⁴ through cross-frequency synchrony^55,56.

BOLD signal amplitude

The blood-oxygenation-level-dependent (BOLD) signal measured by fMRI remains elevated above pre-trial baseline throughout the delay period of WM tasks in several brain regions^57–60 (Fig. 1a). The BOLD signal reflects influences from spiking and LFPs⁶¹, but is most closely correlated with LFP power^62–64 (however, see Ref.⁶⁵). Indeed, one study found a positive relationship between the amplitude of the BOLD signal and ECoG gamma-band power throughout the human cortex during WM delays⁶⁶. Uncertainty surrounding the source of the BOLD signal limits its usefulness for examining local circuit mechanisms, and its sluggish temporal resolution (on the order of seconds) imposes constraints on experimental design (Box 3). However, the meso-level spatial scale and broad coverage of BOLD fMRI can be leveraged to uncover otherwise unobservable features of brain representations⁶⁷. For example, one study found that attention shifted, rather than enhanced, the population tuning of voxels in human visual cortex⁶⁸ — an outcome that could not have been anticipated from single-unit data. Further, fMRI allows for comparisons of activity across brain regions and enables the investigation of large-scale brain networks²³ that are crucial for our understanding of WM^69,70.

Box 3 |. Measuring delay activity with functional MRI.

The temporal resolution of functional MRI presents a challenge for studying delay activity in humans. The blood-oxygen-level-dependent (BOLD) signal represents the transformation of neural activity to a haemodynamic response, and therefore acts as a low-pass filter with a response that peaks several seconds after an isolated neural event. Consequently, in a typical working-memory delay task consisting of a to-be-remembered sample stimulus, an empty delay period and a probe and/or response, accurately estimating the magnitude of delay activity with fMRI is nontrivial. Although analytical methods for estimating delay-period activity from the BOLD signal were introduced more than 20 years ago^57,58, there does not seem to be a consensus among strategies used in contemporary empirical studies. Here, we review the three most common approaches.

Fixed delays modelled with a boxcar function

The most widely used approach uses a long delay period (typically 10–15 s) that is fixed in length across trials. To estimate the magnitude of delay activity, the BOLD response during the delay is modelled with a general linear model (GLM) as a single event that spans the entire delay period (that is, a boxcar function, usually convolved with the haemodynamic response function). In part a of the figure, the left column shows the three predictors — sample, delay and response. The right column shows the data (navy line) and the three predictors convolved with the haemodynamic response function. This method accurately estimates the magnitude of activity during the delay but is unable to independently estimate the contributions of encoding-, maintenance- and response-related activity. As a consequence, any estimate of delay activity is necessarily contaminated by activity during encoding and response, and caution is required when interpreting the results.

Fixed delays modelled as an impulse

Some GLMs model the BOLD signal during the delay period as a single impulse response function (or ‘event’) centred in the middle of the delay period and convolved with the haemodynamic response function (part b of the figure). Separating the delay event from the sample presentation and probe or response by 4 s or more improves the ability to estimate the independent contribution of each event to the BOLD signal⁵⁸. Therefore, designing experiments with delay periods longer than 8 s and analysing the data with a GLM that places the delay-related impulse response function at least 4 s after the offset of the sample stimulus provides an estimation of the magnitude of delay activity that is reasonably independent of the activity associated with encoding or response. However, this method tends to sacrifice accuracy of the estimation.

Variable delays and/or partial trials

A more recent strategy is to use variable delay lengths (part c). Varying the length of the delay helps to reduce the statistical dependencies between the different trial components, and allows for an independent and accurate estimation of delay activity magnitude. The use of partial trials, which terminate after the sample or the delay period, has similar benefits^284,285. The cost of this method is that it complicates efforts to estimate the magnitude of delay period activity for single trials, as is typically required for population analyses (although see Ref.¹⁷⁵ for a notable exception).

Population coding

In contrast to early studies that hypothesized that memories are sustained via the persistent activation of small populations of neurons that are highly tuned to the maintained stimulus¹⁰, recent results suggest that WM relies on population coding (Fig. 1c). Population coding distributes information across neurons or neural populations with diverse tuning preferences, potentially rendering representations more resistant to noise in individual neurons by taking into account correlations between neurons. Population activity in the lPFC of NHPs during a delay represents task variables such as stimulus–stimulus associations⁷¹ in addition to WM content⁷². In humans, population analyses of fMRI data have revealed sustained representations of WM content in the pattern of activity of voxels distributed within parietal and sensory cortices, even when the mean BOLD signal does not differ from baseline across the delay^73–77. Similarly, location information maintained in WM can be read out from the pattern of EEG-recorded alpha-band power⁷⁸. Importantly, many of these findings revealed by population analyses were undetectable using traditional univariate analysis techniques.

Dynamic delay activity

Delay activity was traditionally depicted as stable over a delay, but closer analyses reveal that its magnitude often varies over time. These temporal dynamics were largely overlooked owing to analyses that averaged delay activity over the delay or over trials^35,79 (but see Refs^8,80–82). A thorough understanding of these temporal patterns may provide valuable insight into how delay activity encodes WM information^35,83,84.

Variations in spike rate.

Single-trial spike data from NHPs indicates that some neurons exhibit monotonic increases or decreases in spike rate over the delay (Fig. 1b). These patterns have been hypothesized to represent the transformation from retrospective WM stimulus representations to prospective action plans^85,86, the anticipation of an upcoming response^86,87, or a computation of elapsed time since encoding⁸⁸. Other neurons have spike rates that vary widely without a clear pattern over the delay^81,82. Indeed, temporal fluctuations are the rule, rather than the exception; one study of lPFC estimated that only 3% of delay-active neurons spike at a stable rate throughout the delay⁸¹. It is therefore important that neural models of WM can capture this temporal variation.

Sequential delay spiking.

In contrast to individual cells showing temporal variations in delay spiking, other cells show sequential patterns of activation, in which each cell in a population spikes briefly at a specific point during the delay^89,90. If the population has diverse temporal tuning preferences, then the population activity can ‘tile’ the entire delay period. Such sequential delay period spiking is robustly observed in area CA1 of the hippocampus^91,92, but a similar phenomenon — sequential, delay-spiking ‘relay-race’ cells — has also been documented in the PFC⁹³.

Dynamic population codes.

WM information, including stimulus identity and task variables, can also be stored in a dynamic population code (reviewed elsewhere⁹⁴). In a dynamic code, the same piece of information can be encoded by different patterns of population activity at different times. These dynamics can be revealed by a cross-temporal analysis in which a classifier is trained on patterns of activity at each time point during the delay and tested separately on every other time point of the delay, resulting in a training time-by-testing time matrix of classification accuracy (Fig. 1d). Successful classification along the diagonal, where training and testing time points are the same, combined with chance performance off the diagonal, indicates that distinct activity patterns encode memory content at different points during the delay.

What is the purpose of a time-varying population code? One intuitive account is that the temporal dynamics facilitate the computation of elapsed time, which may be useful in anticipating an upcoming response⁹⁵. The strongest evidence supporting this idea is the finding that neural populations in NHP lPFC that dynamically encoded the category of the maintained stimulus simultaneously encoded elapsed maintenance time⁹⁶. Another intriguing possibility is that dynamic WM codes could be less prone to interference from external input than static codes, rendering WM more resistant to distraction⁹⁷.

Despite their potential advantages, dynamic codes pose an obstacle to the faithful readout of WM information by downstream brain regions because they preclude a unique mapping from a pattern of activity to the information stored in that pattern. This challenge may be overcome in several ways. First, downstream regions may dynamically modulate their readout weights to match the dynamic changes in coding in the upstream region. Although this is theoretically plausible, empirical support for this argument is scant. A second possibility is that neurons in downstream regions maintain static readout weights and integrate evidence over the entire WM delay⁹⁴. In this case, the mapping from activity pattern to information would only be valid for a single point in time, meaning that the WM information would only be available to a subpopulation of downstream neurons at one point during the delay. The activity at other points during the delay would therefore be noise. A third alternative was suggested by a recent NHP study that found that dynamic population codes in lPFC contained a low-dimensional subspace that encoded information stably over time, enabling downstream neurons to assign a single set of readout weights to read out WM content throughout the delay⁹⁸ (see also Refs^84,99,100). It is important to note that arguments for a stable subspace typically rely on analyses of WM tasks in which a single item is to be maintained, whereas dynamic population codes are more frequently noted in complex tasks with multiple memory items. How task complexity or other factors, such as cognitive state or the connectivity of neurons in a network, influence whether information is encoded in a dynamic or static population code is unclear.

LFP bursts.

Single-trial analyses of NHP LFP data have revealed irregular patterns of LFPs that contribute to WM. Lundqvist et al. isolated narrowband LFP bursts on individual trials: gamma bursts were associated with increased information or more items stored in WM, whereas beta bursts seemed to inhibit gamma bursts and reduced the information content of spiking activity⁴¹ (Fig. 1b). Moreover, the frequency of gamma bursts also increased when the monkey accessed WM information to compare a memory item with a visually presented probe⁴². Only after averaging activity over trials did the classic patterns of sustained gamma and beta enhancement emerge. A possible advantage of LFP bursts is that intermittent bursting activity may be metabolically more efficient and less prone to disruption than persistent spiking³⁵. Some have argued against this interpretation of LFP burst data, suggesting that gamma bursts may actually reflect spiking¹⁰¹ (see Ref.¹⁰² for a response). Nevertheless, the idea that information may be sustained by discrete bursting activity seriously challenges the long-held notion that WM information is encoded through persistent neural firing.

Where and why?

Delay activity is found throughout the brain, and probably represents many functions in service of WM¹⁰³ (as well as other cognitive processes not discussed here). We therefore consider location and function together in this section, using a representative set of brain regions to illustrate the anatomical breadth and functional diversity of delay activity (Fig. 2; see Ref.¹⁰⁴ for a comprehensive list of regions displaying delay activity).

Fig. 2 |. Schematic depictions of the properties of delay activity in different brain regions.

In each panel, simplified schematic diagrams are provided to illustrate the type of information represented by delay activity in different parts of the brain. Note that not all of the findings were shown in humans, but the regions have been depicted on a human brain for illustrative purposes. a | Delay activity in the lateral prefrontal cortex (lPFC) represents task rules. Population analyses involving pattern classification approaches have demonstrated that classification of the task rule (for example, which feature of a memorandum is relevant for response⁷⁵) is above chance during the memory delay, suggesting that lPFC delay activity represents aspects of task rules. b | lPFC delay activity represents working memory (WM) content. Neurons in the lPFC of non-human primates (NHPs) exhibit delay-period activity that varies in spike frequency with the properties of the memory stimulus, such as vibrotactile frequency¹¹⁴. c | Electroencephalogram (EEG) electrodes over PPC reveal delay activity that increases in magnitude and oscillatory power with WM load, and plateaus at an individual’s WM capacity²⁷. This is consistent with the notions that PPC delay activity encodes WM content and that PPC delay activity represents internal attention directed to items in WM. Challenges in localizing EEG activity make it unclear where in the brain these signals originate. d | Basal ganglia (BG) delay activity is associated with the upcoming behavioural response. BG delay activity is greater in magnitude when participants can anticipate the upcoming response (prospective representation) than when they cannot (retrospective representation)¹⁹¹. e | Neurons in motor association cortex show preference for specific response rules. Tasks that require NHPs to maintain response rules (for example, which memory item to indicate first with a behavioural response¹⁵⁸), suggest that this region encodes the anticipation of specific planned responses. f | Population activity in visual sensory regions is tuned to features of WM content. Functional MRI studies can identify meso-scale responses to features (such as orientation) of memory items, and have found that populations that represent features of WM content are preferentially active during the delay¹⁷⁴. g | Delay activity in the medial temporal lobe (MTL) may be involved in storing complex information in WM. The magnitude of MTL delay activity is larger for novel items than for familiar items¹⁵³, and MTL delay activity is often observed during WM for complex items¹⁵. h | The thalamus exhibits delay activity that seems to drive delay responses in PFC. Experimental disruption of thalamic delay activity in mice resulted in reduced or abolished delay activity in PFC^202,203.

Multimodal association cortex

Lateral prefrontal cortex.

A longstanding debate is whether lPFC delay activity preferentially represents WM content, or goal-related information such as response rules and WM operations (for more targeted reviews on PFC delay activity, see Refs^104–108). Extensive evidence for both points of view exists, primarily from NHP single-unit and human fMRI studies.

In support of the claim that lPFC stores WM content, lPFC delay activity is selectively tuned to specific low-level features of WM memoranda¹⁰⁹, such as spatial location^10,110,111 or direction of motion^112,113. Another form of content-selective lPFC activity was revealed by a study showing that NHP lPFC neurons monotonically vary their spike rate in proportion to the frequency of a vibrotactile stimulus held in WM^81,114. lPFC delay activity can also be selective for abstract information about WM content, such as associations between stimuli^71,115,116 or stimulus category information^116–118.

By contrast, the view that lPFC encodes goal information is supported by evidence that lPFC delay activity is sensitive to response rules^119,120, reward expectation^121,122 and elapsed maintenance time⁹⁶, and varies in magnitude with demands on cognitive operations, such as manipulation of WM information¹²³. The findings that lPFC neurons can change their tuning mid-delay^71,84 and switch from encoding WM content to encoding reward expectation¹²² are also more compatible with this latter proposal.

Attempts to reconcile these perspectives have given rise to a more nuanced view: that lPFC delay activity simultaneously represents multiple dimensions of WM content and goals^124,125. The property of nonlinear mixed selectivity enables individual neurons to respond to combinations of task and stimulus features such that feature combinations cannot be read out from the responses of the individual neuron alone¹²⁴ and instead are encoded in the combined activity of several neurons¹²⁴. The lPFC uses population coding by nonlinear mixed-selectivity neurons to store high-dimensional representations that can be decoded as low-dimensional representations of goals or content by hierarchically lower brain regions. Nonlinear mixed selectivity may even allow lPFC neurons to recode information in response to intervening input, potentially increasing the robustness of WM representations¹²⁶.

A noteworthy caveat to this debate over the role of lPFC in WM is that it largely ignores the possibility that frontal cortex may be adapted to perform different WM functions in humans and non-human animals. For example, recent work has challenged the long-assumed homology between the frontal eye fields in NHPs and the superior precentral sulcus in humans, complicating direct comparison between these two regions^127,128. Such findings underline the need for systematic comparisons of delay activity across species.

Posterior parietal cortex.

Posterior parietal cortex (PPC) delay activity has been observed in single-unit^11,12,129 and LFP⁴⁵ recordings in NHPs, and using fMRI^76,130–132 and ECoG¹³³ in humans. Human fMRI studies find that PPC delay activity increases with WM load¹³⁰, and NHP studies show that PPC delay spiking is selective for the specific content of WM¹². Both sets of findings are consistent with the idea that PPC delay activity represents low-level stimulus information. Alternatively, evidence that PPC delay activity preferentially encodes visual category information (for example, cat versus dog) over visual item information (for example, which specific cat or dog) when category information is task-relevant suggests that PPC delay activity can represent abstract features of WM content¹³⁴. A third possibility — that PPC delay activity represents attention directed internally to individual WM representations — has been suggested by human fMRI and lesion studies of verbal WM^135,136. This function is reminiscent of the well-documented role of the PPC in shifts of external attention¹³⁷. Interestingly, all three of these putative roles of PPC echo those attributed to lPFC, which is consistent with the high degree of reciprocal connectivity between the two regions. However, observed differences in morphology¹³⁸ and effective connectivity¹³⁹ between the lPFC and PPC suggest that these regions have distinct roles (reviewed in¹⁴⁰). Direct attempts to disentangle the functions of delay activity in PFC and PPC have yielded mixed results^141–143, highlighting the need for experiments that disrupt processing in one of these regions while recording delay activity in the other.

Medial temporal lobe.

Although the MTL is closely associated with long-term memory, it also exhibits delay activity during WM¹⁴⁴, particularly under conditions that can be described as ‘high demand’¹⁴⁵. Indeed, delay-related BOLD signal in the MTL is most robust at the upper limits of individuals’ WM capacity¹⁴⁶. In addition, patterns of delay spiking in the human and NHP MTL encode information about complex images^15,16,147, which are more difficult to maintain than simple features^148,149. A related possibility is that, rather than encoding individual memory items, MTL delay activity might encode associations between features or items during WM for complex items¹⁵⁰. This possible function is analogous to the well-known role of the MTL in relational binding in long-term memory¹⁵¹.

The MTL may also support WM for challenging novel information¹⁵²; MTL delay activity is modulated by the novelty of WM content, exhibiting a greater BOLD signal during WM for novel versus familiar items, even when novelty is task-irrelevant¹⁵³. Understanding the role of the MTL delay activity in the storage of simple, non-novel features will be crucial for determining whether the MTL supports WM maintenance in general or has an auxiliary role only when demand is high. Another alternative, motivated by the hippocampus’s purported role in computing sequential change¹⁵⁴, is that hippocampal populations within the MTL compute elapsed time during WM delays. Consistent with this hypothesis, rodent CA1 cells exhibit stereotyped patterns of sequential firing during delay tasks^91,92.

Unimodal cortices

Motor association cortex.

Motor association areas in frontal cortex are increasingly thought to contribute to cognition¹⁵⁵. Indeed, these regions exhibit delay activity that may help maintain the rules required to convert sensory information to a behavioural response. In line with this notion, delay activity in NHP dorsal premotor cortex is selective for prospective motor plans^156–158 or, when the NHP cannot anticipate a specific motor plan, an abstract response rule¹⁵⁹. Moreover, a human fMRI study revealed that delay activity in pre-supplementary motor area (pre-SMA) is largely independent of the sensory content of WM information¹⁶⁰. Premotor and pre-SMA delay activity may also support WM for verbal information by organizing subvocal rehearsal processes¹⁶¹.

Other fMRI and ECoG studies in humans have found that dorsal premotor regions are instead sensitive to features of WM content¹²³ and are preferentially activated during the maintenance of simple sensory information rather than complex motor sequences^162,163. For example, pre-SMA delay activity scales with the frequency of an oscillating visual, auditory or tactile stimulus¹⁶⁴. However, the fact that this activity is agnostic to the modality of the frequency information indicates abstract supramodal coding for WM memoranda. It will be important to determine whether the anatomical locus of rule- and content-based WM representations in these regions is consistent with the purported hierarchical organization of frontal cortex¹⁶⁵.

Sensory cortices.

Given that primary sensory cortex and sensory association cortex show largely similar delay-activity properties, we discuss them together. Higher-order sensory regions exhibit sustained spiking^13,166,167, BOLD signal modulation^168,169, and modulation of EFP alpha power¹⁷⁰ during WM delays. In NHPs, delay spiking has also been reported in primary auditory¹⁷¹ and somatosensory cortex¹⁷², but not visual cortex¹⁷³. However, human fMRI studies find clear evidence for sustained population coding of information in both early and higher-order visual areas^{73,74,77,174,175}.

Importantly, the delay activity in sensory cortices is highly specific for individual WM items or features^73–75,176. These findings support the ‘sensory recruitment model’ of WM^177,178, which proposes that content-selective WM representations recruit the same sensory regions that initially encode the information. However, this view has been challenged^104,108,179 on the grounds that, unlike in PFC^126,180 or PPC^132,142,181, content-selective delay activity in sensory cortices is often interrupted by intervening input^132,180,181. Distractor resistance is thought to be an essential property of delay activity that allows WM to function despite interference. Crucially, however, the sensory recruitment model does not claim that sensory cortex is sufficient for storing WM content; rather, it posits that sensory delay activity, together with influence from other regions, has a pivotal role in representing detailed information in WM^178,182. Viewed through this lens, distractor-resistant delay activity is not a critical test of the sensory recruitment model. The fact that content-selective activity in sensory cortex is reinstated following the presentation of intervening stimuli indicates that these representations can still be used to guide behaviour^176,183.

Subcortical nuclei

Basal ganglia.

The basal ganglia (BG) transiently gate information during WM encoding¹⁸⁴ and influence behavioural responses based on currently relevant WM contents¹⁸⁵; however, the BG also exhibit sustained activity during WM delays. NHP studies have found delay spiking in the putamen^186,187, caudate^188,189 and globus pallidus¹⁸⁸ that takes several trials to stabilize into persistent spiking¹⁸⁷ and ramps up over the delay¹⁸⁹ and that does not necessarily scale with the accuracy of the behavioural response¹⁸⁷. Thus, the BG may encode the anticipation of an upcoming movement rather than the motor plan itself. Consistent with these results, human fMRI studies have revealed that delay activity in the caudate during WM for spatial location was enhanced^190,191 when an upcoming response could be prospectively encoded¹⁹¹. However, fMRI observations of delay activity in the caudate and putamen during tasks in which specific behavioural responses could not be anticipated, such as during verbal WM¹⁹² or during WM for temporal duration¹⁹³, indicate that, at least in humans, BG delay activity is not limited to response anticipation.

Thalamus.

Elevated delay spiking in the thalamus was originally identified concurrently with lPFC delay activity in 1971², although the thalamic result received less attention (see also Refs^194,195). Subsequent work characterized delay activity in the mediodorsal nucleus of the thalamus^196,197 (reviewed elsewhere^198,199), which has dense reciprocal connections with the dorsolateral PFC in humans²⁰⁰ and NHPs²⁰¹. Similar to the lPFC, delay spiking in neurons of the mediodorsal nucleus was spatially specific, representing either the location of the visual cue or the target of the upcoming response¹⁹⁶.

New evidence about the role of thalamic delay activity comes from studies that optogenetically inhibited thalamocortical communication in mice. Photoinhibition of thalamic delay activity disrupted delay activity in frontal regions (including anterior lateral motor cortex and medial PFC, the putative rodent homologue of dorsolateral PFC and the recipient of projections from mediodorsal thalamus) and behaviour on simple delayed-discrimination tasks^202–204. Interestingly, thalamic delay activity did not simply share the same properties as frontal delay activity; thalamic delay activity was less content-selective than frontal delay activity and causally modulated the selectivity²⁰² and magnitude²⁰³ of frontal delay activity.

There are three key takeaway points from these studies. First, there is a tight causal relationship between thalamic delay activity, cortical delay spiking and behaviour (see also Ref.²⁰⁵). Second, local circuit mechanisms, which have long been implicated in delay activity, may be insufficient to sustain delay activity in the absence of long-range input — in this case, from the thalamus. Third, thalamic delay activity does not merely relay information to and from cortex, but instead may directly shape sustained representations and the coordination of information storage required for WM^206,207. This work is part of a growing appreciation of the role of the thalamus in cognition^208,209.

Comparing across brain regions

Differences in tasks, methodology and species in the above studies pose a significant obstacle to uncovering the location and function of delay activity. Here we consider studies that directly compare delay activity across regions. Our discussion centres on the storage of content-selective WM representations (rather than representations of WM operations) because nearly every delay-active region has been implicated in WM storage, making content-selective delay activity particularly amenable to this type of comparative analysis.

With its broad spatial coverage, fMRI is optimally suited to compare delay activity across brain regions. Human studies have started to capitalize on the power of encoding models²¹⁰ to directly contrast stimulus-encoding properties across the brain during WM for simple features. These studies have documented content-selective delay representations throughout the dorsal visual hierarchy, from early visual cortex to PPC and lPFC^211,212. Subsequent work demonstrated that WM representations throughout the cortex are flexible, shifting from retrospectively representing stimulus position to prospectively representing the motor response, and that the amplitude of delay activity in lPFC and PPC correlates with the precision of encoding in visual regions, indicating that lPFC and PPC provide a top-down signal that tunes delay activity in sensory cortices¹⁷⁵.

Owing to limited spatial coverage, comparisons of delay spiking across regions are relatively rare. Nevertheless, the few NHP studies that have sufficient coverage to investigate multiple regions describe results that are largely compatible with the fMRI studies outlined above. One study recorded from lPFC, the motion-sensitive middle temporal area (MT), and the multimodal medial superior temporal area (MST) during a motion WM task¹¹³. Delay spiking sensitive to the maintained direction of motion was identified in lPFC and MST, but not MT; however, MT exhibited persistent, stimulus-selective increases in oscillatory power in multiple frequency bands. The results were interpreted as reflecting the maintenance of motion information in higher-order regions that biased synaptic activity in visual areas. Another study recorded from an impressive 42 cortical areas in NHPs and identified multiple regions showing content-selective delay spiking, including intraparietal and visual areas, with the most robust selectivity observed in ventrolateral PFC²¹³.

These results raise a key question: what is the benefit of storing WM information in multiple regions at once? A potential answer comes from a study that used calcium imaging to measure neural activity across mouse cortex during a delayed discrimination task²¹⁴. Sustained delay activity was found in medial PFC and posterior perceptual regions, but the magnitude of delay activity in these areas was modulated by the mouse’s strategy: a prospective motor-based strategy increased PFC delay activity, whereas a retrospective, sensory-based strategy increased delay activity in posterior regions. Importantly, experimental perturbation of delay activity in either region prompted strategy switches. Thus, although content-selective delay activity can be observed across the brain, it may reflect different representational formats that are preferentially recruited depending on strategy or task demands.

What are the underlying mechanisms?

Models based on biophysical properties of neurons and neural circuits have provided valuable insights into how delay activity is generated. These models can be divided into three main classes (see Refs^20,215–218 for detailed reviews). Below, we describe these models and evaluate their ability to reproduce the experimental findings discussed above. As the relationship between the biophysical properties of neurons or micro-scale networks and the EFP or BOLD signals is incompletely understood, biophysical models almost exclusively attempt to describe spiking activity; accordingly, we focus our discussion here on delay spiking. The story that emerges is that no single model can account for all features of delay activity and that many results can be explained by multiple classes of models²⁰ (Table 1).

Table 1 |.

Features and limitations of models of delay activity

Class of model	Notable features	Limitations	Refs
Spike-based network models
Feedforward networks	Do not require recurrent connections (although similar activity can also be achieved in recurrent networks227,227) Exhibit both sequential and stable patterns of delay spiking	Timing precision required for feedforward signals may render these models more susceptible to noise	225, 228
Attractor networks	Can describe network delay activity tuned to features of WM memoranda Can encode discrete or continuous quantities and mimic parametric WM114 Disruptions in attractor network activity may account for behavioural errors230,233	Slight changes in parameters lead to instability Precise network connectivity required for these models makes it difficult to account for WM for novel items	18,19
Other
Synaptic mechanisms	Do not require persistent neural activity to sustain WM representations Can explain absence or recovery of delay activity	Unable to explain WM for novel items without STP241 Difficult to confirm absence of delay activity experimentally	235–237
Cellular bistability	Can achieve two or more levels of above-baseline delay firing that can encode features of WM stimuli Do not require complex network connectivity to generate delay activity Can recapitulate delay activity for novel items Can potentially encode parametric WM114	Limited ability to exhibit complex temporal dynamics Can represent discrete but not continuous quantities	222, 223
Hybrid models
Cellular bistability plus recurrent network architecture	Network can be more stable to initial conditions and perturbation Bistable cells can help network encode novel items	Unable to display complex single-cell spiking dynamics	232, 245, 247
Attractor networks plus STP	Can increase resistance of the network to perturbation and drift Firing patterns of individual cells capture complex temporal dynamics, including network oscillations and LFP bursts Allows for encoding of multiple items and novel items	Many of these networks still require specific architecture that involves extensive learning	231, 234, 241, 246

Intrinsic cellular properties

Under certain conditions, individual neurons can respond to a brief stimulation with persistently above-baseline firing²¹⁹. This ability to settle into a stable, above-baseline state is known as bistability (or multistability, in the case of neurons able to achieve multiple above-baseline firing states) and can be induced in vitro in rodent MTL slices by activating muscarinic cholinergic receptors^220,221. Importantly, models featuring intrinsic cellular bistability^222,223 can recapitulate persistent delay spiking, typically by incorporating features that mimic a variety of cell-intrinsic mechanisms observed in vitro (such as changes in calcium currents; reviewed in Ref²⁰).

A critical feature of cellular bistability is that it does not require complex network connectivity (see below) to generate content-selective delay activity. Cellular models may therefore be well-suited to explain delay activity in situations where the rich connectivity and tuning for memoranda required for network-based persistent spiking is probably absent, such as in regions that lack complex networks²²⁴ or during WM for novel items²¹⁶. Models involving multistable neurons can also mimic the behaviour of lPFC cells that vary their firing rates according to the rate of vibrotactile stimulation being stored in WM⁸¹.

There are some aspects of delay activity that models based on cellular bistability struggle to capture. For example, these models cannot reproduce complex temporal dynamics associated with delay spiking²¹⁸. In addition, although neurons in these models can clearly switch between discrete states and therefore represent discrete information (for example, an integer from 1–12 on a clock face), how they might represent continuous quantities (for instance, an exact position in degrees on a clock face) in the absence of complex networks is unclear.

Spiking in local cell assemblies

Feedforward models.

Cell assembly models are based on connectivity between multiple neurons. One group of such models of WM focuses on feedforward connections between cells that share similar selectivity^225–227. This connectivity enables the network to store information through brief periods of above-baseline spiking that are passed between cells²²⁸ without requiring sustained firing in any one cell, mimicking the sequential activation of neurons observed in the hippocampus^91,92 and PFC⁹³. Interestingly, feedforward networks can also reproduce stable delay spiking; a neuron that integrates activity across the feedforward chain of activity can sustain its spiking over several seconds²²⁷. However, the precise timing required to propagate activity in these networks may render them more susceptible to noise than networks that rely on population coding^216,225.

Attractor networks.

Another group of cell assembly models features recurrent excitatory connections between neurons in the network. When the connectivity between neurons in the network is properly structured, and when the balance between excitation and inhibition is well-tuned, spiking dynamics settle into a stable attractor state in the absence of any input²²⁹. Attractor networks can be discrete — with basins of attraction that encode discrete states — or continuous — encoding analogue values using a single continuous basin of attraction, a continuum of population firing rates along which network activity can stably traverse.

Attractor models have several appealing properties. First, and most critically, they replicate various empirical phenomena, including content-selective delay spiking and elevated gamma power¹⁹. Second, attractor networks inherently rely on a population code, which accords with several empirical observations described above. Third, continuous attractor models generate clear predictions about how WM representations are affected by internal neural noise; noise in model networks results in random drift along the basin of attraction and thus imprecise readout of the information encoded by the network¹⁹. One study tested this prediction in NHP lPFC and found that, in line with the model, drifts in population delay spiking predicted the degree of error in behavioural responses²³⁰.

Attractor network models also have some notable drawbacks. The primary limitation of continuous attractor models is their fragility, which manifests itself in a few ways. Activity in continuous attractor networks is susceptible to drift and interference from external stimuli, meaning that encoded information is vulnerable to noise²³¹. In addition, continuous attractor networks struggle to maintain multiple memory items simultaneously¹⁰². These issues can be overcome by approximating the behaviour of continuous attractors using discrete attractors with several basins of attraction²³². Discrete attractor models have received recent empirical support; one study measured the response of delay-active neurons in mouse anterior lateral motor cortex and found that optogenetic perturbation resulted in neural and behavioral responses that were consistent with the predictions of a discrete – but not continuous – attractor model²³³. However, even these discrete approximations of continuous attractors require a very specific architecture, with precisely tuned and highly symmetric connectivity between individual neurons, as well as a specific balance of excitation and inhibition¹⁸. In practice, this architecture is probably present in networks that have been shaped by extensive learning. As a result, attractors are valuable for understanding WM for well-learned stimuli, such as spatial position¹⁹, but are less useful for explaining WM for novel stimuli or novel combinations of features. An additional challenge for attractor models is that they have trouble reproducing complex spiking dynamics²³⁴.

Synaptic weights in local assemblies

A third class of models is based on the provocative claim that sustained activation is not necessary for WM. Instead, these ‘activity silent’ models propose that information can be maintained through rapid shifts in synaptic weights within local neural networks that encode WM information^235–237. Short-term plasticity (STP), which has been observed in PFC²³⁸, has been proposed to underlie these changes in synaptic weights^235,239,240. According to activity silent models, shifts in synaptic weights during stimulus encoding cause the network to act as a matched filter, responding to noisy input with activity that reveals the underlying state of the network (that is, the memory)²³⁶. As an analogy, a submarine uses sonar (that is, noisy input) to ‘ping’ the sea floor, resulting in an image of the otherwise unobservable state of the ocean floor (that is, the memory encoded within the network)¹⁸³. In these models, brief changes in synaptic weights can store a memory for up to one second without spiking²³⁵; further maintenance requires the synaptic changes to be periodically ‘refreshed’ by the spiking of memory-encoding neurons²⁴¹. Interestingly, activity-silent mechanisms raise the possibility that content-selective delay spiking, EFPs or BOLD responses may simply result from a non-specific input (that is, noise) that is delivered to a neuronal population that silently maintains a memory^183,242, rather than the mechanism of memory maintenance itself²³⁶.

On one hand, models based on synaptic mechanisms are appealing for several reasons. First, unlike in many models (including spiking-based models) that struggle to explain how a network can simultaneously represent WM information and incoming sensory stimuli, in synaptic models spiking input does not interfere with the memory in the same way. Second, activity-silent models may explain how neurons with short time constants, such as those in sensory cortices, maintain information over longer intervals²⁴³; if the memory is stored in synaptic weights, sparse intermittent spiking may be sufficient to store it over tens of seconds. Third, activity-silent models may be more energy-efficient than spiking models^97,235, although a countervailing view from an analysis of NHP PFC neurons found that increased metabolic demand due to delay spiking was offset by decreases in the activity in other neurons, resulting in a network-wide metabolic cost that was virtually unchanged from baseline to delay⁹⁸.

On the other hand, most models of synaptic mechanisms are unable to recreate WM for novel items. A caveat when evaluating arguments for activity-silent mechanisms is that empirical support for these models is often based on the absence of delay activity²⁴⁴. One cannot rule out the possibility that the apparent absence of delay activity instead reflects very low levels of delay activity or an alternative form of delay activity to which the method is insensitive. Attempts to clarify how activity-silent and spiking mechanisms coexist are a logical next step towards an understanding of the synaptic mechanisms of WM.

Hybrid models

Elements of the models described above have been combined to overcome some of their individual limitations. For example, several models have been formulated to increase the stability of continuous attractor networks in the face of drift and external input, using either STP²³¹ or cellular bistability^232,245. These models are able to stabilize the information encoded by delay firing over physiologically realistic intervals, thus improving memory readout. Hybrid models that blend elements of STP with traditional attractor network models are also able to reproduce irregular patterns of delay firing^234,241 that are consistent with experimental outcomes⁸¹. In addition, recurrent network models that incorporate STP can account for regular network-level oscillations in the gamma and beta frequencies, as well as intermittent bursting²⁴⁶. Two predictions of this latter model — that the gamma burst rate should increase with the anticipation of a behavioural response, and that the frequency of gamma bursts should increase with WM load — were confirmed empirically^41,42.

Hybrid models are also able to address another limitation of most traditional models — the inability to store novel information. One model embedded bistable cells in a network architecture that allowed the network to encode and maintain novel items through spiking²⁴⁷, whereas a more recent study used Hebbian STP in an attractor model to enable the encoding of novel items²⁴¹. Further refinement of hybrid models, and their use in generating network-level predictions in NHPs and humans, will be essential to fully elucidate the mechanisms that give rise to delay activity.

Concluding remarks

The work reviewed above suggests three key points about delay activity. First, although delay activity has classically been described as sustained activation of highly tuned neurons or neural populations, the abundance of evidence for time-varying forms of delay activity and population coding indicates that a broader view of delay activity is necessary. In particular, analytical advances have revealed WM information in LFP bursts⁴¹ and dynamic coding⁹⁷ — information to which standard analyses were insensitive. An important avenue for future research will be to reconcile dynamic and stable modes of coding for WM information.

Second, the evidence suggests that widespread brain regions are able to exhibit delay activity in the service of WM¹⁷⁸. Distinguishing between regions that generate delay activity and those that inherit delay activity as a result of connections with delay-active regions will be crucial in disentangling the many functions of delay activity. Even if a region does not generate delay activity itself, delay activity induced by another region may nevertheless have an important functional role. The ubiquity of delay activity across a diverse set of brain regions further calls for a significant emphasis on how information is integrated across the brain during WM maintenance. Accordingly, the study of meso- and macro-level networks using EFP and fMRI will be essential in providing data that complements single-neuron studies. The broad recruitment of brain regions during WM also highlights the need for methods that disrupt one region while recording delay activity from another, including the use of concurrent transcranial magnetic stimulation (TMS) and fMRI, and/or optogenetics with spike recordings.

Third, despite many attempts to draw links between empirical findings and the mechanisms hypothesized by biophysical models, the data are often consistent with several models. One strategy to overcome this limitation is to directly assess how different features of these models reproduce realistic data²⁴⁸. A second strategy is to identify differential predictions made by some of these models²³³. For example, synaptic mechanisms and continuous attractors seem to make opposite predictions about the effects of noise; noise in attractor networks results in poorer readout by downstream neurons, whereas noise delivered to a population of neurons that maintains information through shifts in synaptic weights could transiently enhance readout. Although some elements of this prediction have been explored empirically^183,249, these two models have not been directly pitted against one another. Connection between modelling and experimental work will also be necessary to explore whether distinct representational states within WM²⁵⁰ correspond to different cellular and network mechanisms that produce delay activity. Further crosstalk between neuron- and circuit-level modelling work and systems-level empirical work will be crucial to advance our understanding of delay activity.

Glossary

WM delay task

Tasks that temporally segregate working memory encoding, maintenance and response by introducing an unfilled memory delay between a memory stimulus and the contingent behavioural response

Population coding

A coding scheme wherein information is encoded in the combined activity of a population of neurons (or electrodes, or voxels) as opposed to the activity of individual neurons (or electrodes, or voxels)

Voxel

The volumetric unit of functional MRI (fMRI) measurement. A 3D fMRI brain image contains ~100,000 voxels, each of which represents the activity of tens of thousands of neurons

Nonlinear mixed selectivity

A property that allows neurons to respond to combinations of stimulus or task features with nonlinear changes in firing rates

WM load

The amount of information that is held in working memory. Working memory load can be manipulated by varying the number or complexity of memory items

WM capacity

The upper bound on the amount of information that an individual can store at once in working memory

Encoding model

A model that forms a prediction of brain activity for given set of experimental features (for example, specific memory items during a working memory delay task)

Attractor state

A stable state of the activity of a network of (usually, recurrently connected) neurons that persists in the absence of input

Short-term plasticity

(STP). Synaptic plasticity in response to brief (~1 s) stimulation. Hebbian forms, (involving presynaptic and postsynaptic changes) and non-Hebbian forms (involving only presynaptic changes) of STP have been proposed to underlie working memory

Matched filter

A linear filter that can help detect the presence of a known stimulus in a noisy observed signal by correlating the known stimulus with the observed signal

Time constant

A value that describes the time required for a neuron to return to a baseline state following an input

Haemodynamic response

The temporal pattern of blood-oxygen-level-dependent signal observed by functional MRI in response to a brief impulse of neural activity. It takes ~20 s to return to baseline

General linear model

(GLM). A model that describes the output of a system as a linear combination of predictors. GLMs are used to estimate BOLD responses to features of an experimental task

Impulse response function

The output of a dynamic system in response to a brief input