Review

Visual working memory capacity:

from psychophysics and neurobiology
to individual differences
Steven J. Luck1 and Edward K. Vogel2
1
Center for Mind and Brain and Department of Psychology, University of California, Davis, 267 Cousteau Place, Davis, CA 95618, USA
2
Department of Psychology, University of Oregon, Straub Hall, Eugene, OR 97403, USA

Visual working memory capacity is of great interest problem of maintaining multiple active representations in
because it is strongly correlated with overall cognitive networks of interconnected neurons. This problem can be
ability, can be understood at the level of neural circuits, solved by maintaining a limited number of discrete repre-
and is easily measured. Recent studies have shown that sentations, which then impacts almost every aspect of
capacity influences tasks ranging from saccade targeting cognitive function.
to analogical reasoning. A debate has arisen over wheth-
er capacity is constrained by a limited number of discrete Why study visual working memory?
representations or by an infinitely divisible resource, but There are at least four major reasons for the explosion of
the empirical evidence and neural network models cur- research on VWM capacity. First, studies of change blind-
rently favor a discrete item limit. Capacity differs ness in the 1990s (Figure 1A) provided striking examples of
markedly across individuals and groups, and recent re- the limitations of VWM capacity in both the laboratory and
search indicates that some of these differences reflect the real world [5,6].
true differences in storage capacity whereas others re- Second, the change detection paradigm (Figure 1B–D)
flect variations in the ability to use memory capacity was popularized to provide a means of studying the same
efficiently. basic phenomenon with more precisely controlled visual
displays [7–9]. This paradigm made it possible to quantify
The rise of visual working memory VWM capacity and to link VWM to the enormous body of
For many decades, the concept of working memory capaci-
ty has played a central role in large-scale theories of
cognition [1,2] and in explaining individual differences Glossary
in cognitive ability [3]. Originally, research on working Cell assembly: set of neurons that together represent a single item (e.g., a set
memory was dominated by verbal paradigms, such as digit of red-selective neurons and a set of vertical-selective neurons that together
represent a red vertical bar).
span tasks, which require participants to repeat back a Complex span tasks: tasks that assess working memory capacity by requiring
series of digits, and complex span tasks, in which partici- subjects to switch back and forth between a memory encoding task and a
pants must alternate between two tasks that require processing task. In the operation span task, for example, subjects see a simple
mathematics problem followed by a word that is to be stored in memory. In
processing of information and storing of items in memory each trial, a sequence of these pairs is presented and subjects are asked to
[4]. However, the past 15-year period has seen an explo- recall the words at the end of the trial. Memory span is quantified in terms of
sion of research on visual working memory (VWM, as the number of words that can be recalled at the end of the trial.
Contralateral delay activity: sustained ERP response during the delay period of
defined in Box 1; see Glossary). a visual working memory task that is observed over the hemisphere
In this article, we review some of the key findings of this contralateral to the items being maintained in memory.
research, focusing on the cognitive and neural mechanisms Event-related potential: ERPs are specific event-related brain responses that
are embedded within the electroencephalogram (EEG). They arise from the
of VWM capacity and on individual and group differences summed postsynaptic potentials of many thousands of neurons and are
in VWM capacity. We begin by asking why vision needs a conducted through the brain and skull to the scalp, where they can be recorded
noninvasively via surface electrodes.
working memory system. We then discuss whether capaci-
Functional magnetic resonance imaging: this technique takes advantage of the
ty is constrained by a limit on the number of discrete items different magnetic properties of oxygenated and deoxygenated hemoglobin to
that can be represented or by a limit on a resource that can localize changes in blood flow that are triggered by changes in neural activity.
K: number of items stored in working memory on a given trial type.
be divided among large numbers of items. We then discuss Kmax: maximum number of items that a given individual can store in working
how and why capacity varies among individuals and be- memory. This is a measure of working memory capacity.
tween groups. Finally, we discuss the neural mechanisms Pmem: probability that a given item is present in memory at the time of a test.
Set size: number of items in an array.
that may determine VWM capacity. Our overall perspec- Single-unit recordings: recordings of the action potentials of individual
tive is that limits on VWM capacity reflect the broader neurons via the tip of an electrode placed just outside the cell body of a neuron.
Standard deviation: measure of the spread of a distribution of values. In VWM
research, standard deviation is used to quantify the distribution of memory
Corresponding author: Luck, S.J. ([email protected]). errors and is inversely related to the precision of the memory representation.
Visual working memory: active maintenance of visual information to serve the
1364-6613/$ – see front matter needs of ongoing tasks.
! 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.tics.2013.06.006

Trends in Cognitive Sciences, August 2013, Vol. 17, No. 8 391

 Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

Box 1. What is visual working memory? 43% of individual differences in a global fluid intelligence
The term working memory is used in many different ways and it is
[14] and 46% of individual differences in overall perfor-
therefore important for researchers to define exactly what they mance on a broad battery of cognitive tasks (Figure 1E)
mean when they use this term. In this article, we define VWM as the [13]. These high correlations are particularly striking given
active maintenance of visual information to serve the needs of that the change detection paradigm provides a relatively
ongoing tasks. We are not suggesting that this is the only valid simple measure of VWM capacity, with little or no impact of
definition of VWM. Instead, this definition is intended to provide a
clear statement of the scope of the memory system that we address
long-term memory when canonical task parameters are
in this review. We also believe that it reflects the typical usage of the used. Specifically, there is little or no effect of proactive
term by researchers who come from a vision science perspective. interference [15] (but see [16,17]) or medial temporal lobe
Our definition has three key components. First, to qualify as VWM, lesions [18] (but see [19]) with canonical parameters.
it is not sufficient that the information was acquired through the Finally, researchers have discovered neural correlates
visual modality; the representation of the information must be
visual in nature. If the observer stores a verbal or amodal conceptual
of VWM maintenance that are strongly correlated with
representation of the sensory input, we no longer consider it to be a individual differences in VWM capacity. In studies of VWM
visual memory. Second, VWM is based on active maintenance. That in non-human primates, neurons in several brain areas
is, a VWM representation is maintained by a change in sustained, exhibit elevated firing rates and increased synchrony dur-
energy-requiring neural activity rather than by a change in synaptic
ing the delay interval [20,21]. In human event-related
strength (which can be verified with physiological recordings). This
distinguishes VWM representations from passively stored, longer- potential (ERP) studies, an analogous sustained change
term memories. Third, the representations must be used in the in voltage is observed during the delay interval in change
service of broader cognitive tasks. This is the ‘‘workin’’ part of VWM. detection tasks [22,23]. This effect is called contralateral
delay activity (CDA) because it is found in the hemisphere
contralateral to a set of lateralized objects that are being
research on vision [10]. Moreover, this task is so simple remembered over a delay period (Figure 2A). CDA ampli-
that even pigeons can do it [11]. tude increases as the set size increases, reaching an as-
Third, estimates of VWM capacity have excellent psycho- ymptote at the capacity limit (typically three or four items).
metric properties when optimal methods are used [12,13]. This is true both at the group level and the single-subject
For example, a 10-min change localization task yielded test– level, with very strong correlation between an individual’s
retest reliability of 0.77 for testing episodes separated by 1.5 behaviorally measured VWM capacity and that individua-
years [13]. In addition, VWM capacity is highly correlated l’s CDA asymptote point (Figure 2D). An analogous effect
with measures of broad cognitive function, accounting for can be seen in functional magnetic resonance imaging

(A) (B)

Version 1 Blank Version 2 Blank Sample array Delay Test array

240 ms 80 ms 240 ms 80 ms
(100 ms) (900 ms) (Un!l response)

(C) (D) (E) 80

100 5 100 5
%Correct
Overall cogni!vve ability

%Correct
60
(MATRICS T score)

4 4 N=176
Items in memory (K)
Items in memory (K)
% Correct

% Correct

3 3 40
75 75
2 2
20
K 1
K 1 r2 = .46
(age-adjusted)
50 0 50 0 0
0 4 8 12 0 4 8 12 1 2 3 4
Set size (N) Set size (N) Storage capacity (Kmax)
TRENDS in Cognitive Sciences

Figure 1. (A) Example of a change blindness task. Many cycles are required before an observer notices the difference between the two images. Reprinted from [10] by
permission of Oxford University Press, USA. (B) Example of a change detection task [8]. A brief sample array is followed by a blank delay and then a test array. The test array
is either identical to the sample array or differs in one feature of one of the objects, and the observer indicates whether a change is present. In the change localization
variant, a change is always present and the subject indicates which item has changed [47,60,79]. (C) Hypothetical results for an observer with a capacity (Kmax) of four items,
assuming a slot model. Accuracy (% correct) is perfect when the set size (N) is less than Kmax (assuming that changes in color are very large, when present). When N > Kmax,
the changed item will be present in memory for N/Kmax trials, and subjects will fail to detect the change when the changed item is not in memory. Accuracy will therefore
decrease systematically as N increases above Kmax. By taking into account guessing, it is possible to estimate the number of items that the observer must have had in
memory (K) for each set size [80–83]. (D) Data from an actual experiment with college student subjects [8]. (E) Scatter plot of the relationship between storage capacity
(Kmax) measured in a 10-min change localization task and a measure of broad cognitive function (the T score from the MATRICS battery) in a sample of subjects including
both schizophrenia patients and matched controls [13]. The correlations were similar in both groups, justifying an aggregated analysis.

392
 Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

(A) (B)
300-400 ms Memory Reten!on Test
SOA array interval array -3

-200 500 1000

ms

+3 Key:
200 ms 100 ms 900 ms 2000 ms Contralateral
Ipsilateral
(C) -3 (D) -1
Mean
-2.5 capacity
Mean amplitude (in microvolts)

-0.75
-2
Increase
-1.5
from -0.5
2 to 4
items
-1
-0.25

-0.5

0
0 1 2 3 4
1 2 3 4 6 8 10
Memory capacity
Memory array size
TRENDS in Cognitive Sciences

Figure 2. (A) Event-related potential (ERP) paradigm for recording contralateral delay activity (CDA) [22]. Subjects are instructed to remember the colors of the items on the
side indicated by the arrow and report whether a color has changed on that side in the test array. (B) ERP waveforms from ipsilateral versus contralateral electrode sites
relative to the side of the array that was encoded into memory. Time zero is the onset of the test array, and the CDA is the difference in voltage between the ipsilateral and
contralateral waveforms during the delay period. Note the negative direction of the y-axis. (C) CDA amplitude as a function of the number of items on the side to be
remembered, averaged over subjects. Note that CDA amplitude reaches an asymptote near the average working memory capacity limit. (D) Scatter plot for individual
subjects, showing that individual differences in working memory capacity (Kmax) are correlated with differences in the CDA asymptote (quantified as the difference in CDA
amplitude between set sizes of N = 2 and N = 4).

(fMRI) studies, where the signal in the intraparietal sulcus VWM also plays a key role in higher-level visual tasks,
(IPS) during the delay period increases as the set size demonstrating that it is truly a working memory. For
increases, reaching an asymptote at the individual sub- example, when the target for a visual search task is cued
ject’s VWM capacity [24,25]. in a trial-by-trial manner, the cue is stored in VWM,
leading to a CDA in the interval between the cue and
The role of working memory in vision the search array [32]. In this situation, search performance
Visual information is typically acquired during short per- is impaired if VWM is filled to capacity by a secondary
iods of fixation (usually 200–500 ms in duration) separated object memory task [33]. However, after several trials of
by saccadic eye movements that temporarily suppress searching for the same target, the CDA disappears [32].
processing and then shift the retinal image. Some kind Moreover, search performance is no longer impaired by a
of memory is needed to bridge the temporal gaps and concurrent VWM load when the target remains the same
spatial shifts created by eye movements [26]. Seminal on trial after trial [34]. These results indicate that the
research by Irwin and colleagues [27,28] demonstrated ‘search template’ is transferred from VWM into a longer-
that iconic memory does not survive eye movements but term memory store when the target remains constant over
that VWM can be used to link the pre-saccade representa- several trials.
tion of an object at one retinal location with the post- It is natural to assume that the sole purpose of working
saccade representation of that object in a different retinal memory is to store items that are no longer present, but
location. More recent research has shown that the target of recent research indicates that the same system is also used
an upcoming eye movement is automatically stored in to represent information that is currently visible. For
VWM, and after the eye movement this VWM representa- example, both VWM capacity limits and neural indices
tion is compared with the newly fixated object [29]. In of VWM activation have been observed in tasks in which
addition, eye movements may be biased toward objects the items remain visible throughout the trial, such as
that match the current contents of VWM [30], and even the visual search [35,36] and multiple object tracking
simplest saccades are faster if the saccade target matches [37,38]. A recent study took this a step further and showed
the current contents of VWM [31]. that when observers were asked to remember the colors of
393
 Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

items in a static array that remained visible for the entire However, complexity is not a very well-defined term
trial, their VWM capacity was indistinguishable from [10,43], and it is much easier to understand capacity limits
trials in which the items disappeared during a 1-s reten- for simple, unidimensional features.
tion period [39]. Moreover, the CDA was the same wheth- Two main classes of theories of VWM capacity have been
er or not the stimuli were visible during the retention proposed, a discrete slots class and a continuous resource
period, indicating that the same neural mechanisms are class [10,44–46]. Figure 3A shows how a display of six
used to represent visual objects whether or not they are colored squares would be represented in VWM according to
currently visible. Together, these findings suggest that these two theory classes. Slot-based theories assume that a
VWM may not really be a memory system per se, but may limited number of items, Kmax, can be stored in VWM; if the
instead be a general-purpose visual representation sys- number of items in the sensory input is greater than Kmax,
tem that can, when necessary, maintain information over then Kmax of the items are stored in VWM and no informa-
short delays. tion about the other items is stored in VWM. Note, howev-
er, that internal and external sources of variability will
The nature of VWM capacity limitations: slots versus cause each representation to be imperfect and may cause
resources Kmax to vary from trial to trial.
When memory for simple, highly discriminable colored Resource-based theories assume that VWM capacity is a
squares is tested, the typical college student has a capacity flexibly divisible resource that can be spread among all the
of only three to four objects’ worth of information [40]. items in the display, but with fewer resources per item and
What is the nature of this limit? Object complexity plays a therefore reduced precision as the set size increases. These
clear role: task performance is less accurate for complex theories can also be framed in terms of an increase neural
objects than for simple objects in most cases [41,42]. noise as the set size increases.

(A) You see this Do you remember this? Or this?

Con!nuous resource Discrete slots

(B) Sample array Delay Cue/response display (C) 0.020 (D) 0.020
(100 ms) (900 ms) (un!l response) In memory
Set size 3
Probability density

Probability density

0.015 0.015
Pmem = 0.84
Kmax = 2.52
0.010 0.010 Set size 6 SD = 20.8°
Pmem = 0.41
Not in Kmax = 2.46
0.005 memory 0.005
SD = 23.3°
Mixture
0.000 0.000
-180 -120 -60 0 60 120 180 -180 -120 -60 0 60 120 180
Response error(°) Response error(°)

(E) Sample array Delay Cue/response display (F) (G) 6

(200 ms) (1000 ms) (un!l response)
20
5
SD inflec!on point

18 R2 = .657
4
16
SD

3
14
2
12
1
10
1 2 3 4 5 6 7 8 0 1 2 3 4
Kmax
TRENDS in Cognitive Sciences

Figure 3. (A) Essence of the continuous resource and discrete slot model classes. (B) Example of a continuous report task with color stimuli. The cue (thicker box) indicates
which item should be reported by clicking on the color wheel. (C) Hypothetical distribution of response errors (difference between actual color and reported color) according
to the slot model [48]. If the cued item is present in memory (violet line), the errors will be normally distributed around the correct value (the Von Mises distribution is used
for circular dimensions such as hue). If the cued item is not remembered (brown line), errors will be random (a uniform distribution). The actual data consist of a weighted
sum of these two distributions (black line). (D) Data observed for set sizes of N = 3 and N = 6, and estimates of the parameters of the underlying distributions [48]. (E)
Continuous report task for orientation [49]. The sample array contains circles with gaps; when the test display appears, the subject reports the orientation remembered for
the gap in the item that is cued by the thicker circle; the orientation is reported by clicking on the corresponding location on the cue circle. (F) Standard deviation (SD) of the
distribution of response errors in the task shown in (E) as a function of set size. The group data fit well to a function that increases linearly, has an inflection point at the
average Kmax, and is then flat. (G) Inflection point as a function of Kmax for individual subjects, showing that the point at which the SD reaches an asymptote for a given
subject is predicted by that subject’s visual working memory capacity.

394
Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

Evidence of discrete slots from continuous report Figure 4A. A sample array of colored squares was pre-
experiments sented, followed after a delay by a probe stimulus, and the
The most obvious way to distinguish between these classes task was to report whether the probe stimulus was dis-
of theories is to determine whether VWM representations placed to the left or to the right of the corresponding item in
become less precise as the set size increases, and this is most the sample array. Slot-based models predict that observers
easily assessed with the continuous report paradigm shown should make errors for large set sizes, even when the
in Figure 3B [47,48]. Each trial begins with a sample array, displacements are very large (when no information about
followed by a retention interval and then a test display. The a given item is present in VWM). By contrast, resource-
test display includes a cue at one location, and the observer based models predict that performance should remain near
clicks on a color wheel to report the color remembered for the perfect for sufficiently large displacements. The latter
cued item. If the cued item is present in memory, the color pattern was found (Figure 4B). Moreover, the precision
reported should be close to the actual color (i.e., the magni- of the representations declined monotonically as the set
tude of the error will be small), and the distribution of error size increased.
magnitudes over many trials would be expected to be ap- Although this appears to be strong evidence of continu-
proximately normal (Figure 3C). The width of this distribu- ous resources, it appears to reflect a guessing strategy. If,
tion (quantified as the standard deviation, SD) is inversely as shown in Figure 4A, the probe is near the left edge of the
related to the precision of the memory representation. The display, the subject can guess that it was a leftward shift
same method can be used for other feature dimensions, such even if the corresponding sample item was not stored in
as orientation [49] (but see [50]) and shape [48]. memory. Indeed, a subsequent study [52] showed that
When the set size exceeds the number of items that can be near-perfect performance can be obtained for large displa-
stored in VWM (Kmax), the cued item will not be present in cements when this guessing strategy is possible
memory on a subset of trials. When this happens, the (Figure 4B), but performance does not reach a ceiling when
observer will guess randomly, leading to a uniform distri- the task is modified to prevent this strategy (Figure 4C).
bution of errors (Figure 3C). Because the data from a set of A more compelling challenge to slot-based models was
trials may contain a mixture of in-memory and out-of-mem- provided by van den Berg et al. [44], who proposed a new
ory trials, the observed distribution of errors will consist of a resource-based model in which the precision varies ran-
mixture of a normal distribution (from in-memory trials) domly from trial to trial. According to this model, the
and a uniform distribution (from out-of-memory trials). This distribution of errors in continuous report tasks
is equivalent to a normal distribution that has been shifted (Figure 3D) reflects the average of many different normal
vertically by an amount that is related to the probability that distributions with different widths (SDs). A careful analy-
the cued item was absent from memory. From the observed sis showed beyond doubt that a single normal distribution
mixture, it is possible to derive two key VWM parameters: systematically misfits the actual distribution of responses
Pmem, the probability that the cued item was present in and that a mixture of multiple normal distributions more
memory; and SD, the width of the normal distribution. The accurately fits the observed distribution.
number of items being held in memory for a given set size (K) This model also proposes that the amount of variation in
is simply Pmem ! set size. the SD increases with the set size. For large set sizes,
Figure 3D shows that increasing the set size from N = 3 memories will sometimes be so imprecise that very large
to N = 6 leads to a large vertical shift in the distribution of errors will occur, making it seem as if observers are gues-
responses, with no substantial change in the width of the sing randomly. Consequently, the frequent occurrence of
normal portion of the distribution [48]. Thus, an increase in extreme errors for a set size of N = 6 in Figure 3D may
the set size appears to decrease the probability that the reflect very poor memory precision for a subset of trials,
cued item is present in memory (Pmem) without changing and not the complete absence of a representation of the
the precision (SD) of the representation. These data are item tested. However, it remains to be seen whether this
consistent with discrete slot models and are incompatible variable-precision resource model fits the data better than
with most resource-based models. a slot-based model in which precision is allowed to vary
Although precision did not change significantly between from trial to trial (as would be expected in any imperfect
set sizes of N = 3 and N = 6, precision did improve when the storage system) but does not increase with set size.
set size was reduced to less than three items. This is best Sims et al. proposed a very different resource-based
illustrated by a study in which subjects remembered orien- model in which VWM capacity can be conceived in terms
tation information rather than color information (Figure 3E) of classic information theory [53]. In this model, the sensory
[49]. SD increased linearly as the set size increased up to a input is optimally recoded so that it can be represented in
point and then reached an asymptote (Figure 3F). Interest- terms of a specific number of bits of information. The model
ingly, the inflection point in this bilinear function was predicts that because of optimal recoding, observers will be
closely related to VWM capacity (Figure 3G). These results able to retain more precise information when the range of
suggest that resources can be shared among items until a possible values is small than when it is large, and this
maximum number of items (Kmax) is reached [49,51]. prediction was confirmed. If visual information can be arbi-
trarily recoded in abstract bits, this naturally brings up the
Evidence of continuous resources question of whether the representations are still visual. It
Evidence against discrete slots and in favor of continuous remains to be seen whether the storage of this recoded
resources was provided by Bays and Husain [46] using information occurs in visual cortex or instead occurs in a
variations on the spatial memory paradigm shown in more generic, amodal working memory system [54].
395
 Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

(A) Sample array Delay Probe Response

(1000 ms) (500 ms) (250 ms)

Le"ward
or
rightward
shi"?

(B) (C)
1.0

1.0
8−items 8−items
Probability of rightward response

Probability of rightward response

0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0
−10 −5 0 5 10 −10 −5 0 5 10
Displacement in degrees Displacement in degrees
TRENDS in Cognitive Sciences

Figure 4. (A) Spatial visual working memory (VWM) paradigm of Bays and Husain [46]. Observers report whether the probe is displaced leftward or rightward relative to the
corresponding sample item. (B) Results from a set size of N = 8 in a replication experiment [52]. The x-axis shows the displacement of the probe relative to the original item,
with negative values indicating leftward and positive values indicating rightward displacement. The y-axis shows the probability that the subject reports rightward
displacement. When the displacement was large, subjects were nearly perfect: they nearly always reported rightward displacement for a large rightward displacement and
almost never reported rightward displacement for a large leftward displacement. Bays and Husain argued that this nearly perfect memory for large displacements for a set
size of N = 8 is strong evidence against the slot model and in favor of the resource model, but later research showed that these results could be explained by a guessing
strategy [52]. (C) Results when the task was changed slightly to eliminate the guessing strategy. Observers were no longer nearly perfect for large displacements. Panels (B)
and (C) are reprinted from [52] with kind permission from Springer Science and Business Media.

Another possibility is that VWM capacity limits are a Resource-based models would predict that precision would
by-product of competition between similar representations be very high on valid trials, intermediate on neutral trials,
[55,56]. However, this would predict that capacity would be and very low on invalid trials (because the invalidly cued
lower when the items to be remembered are similar to each objects should have ‘just a few drops’ of resources). Slot-
other, which is inconsistent with the available evidence based models, however, predict that precision should be
[41,57]. the same on neutral and invalid trials(because it is not
possible to have ‘just a few drops’ of resources in these
Additional evidence of discrete slots models). Instead, the probability of having any represen-
Converging evidence from multiple experimental para- tation at all should decline for invalid trials. This is exactly
digms will be needed to distinguish between the broad what was found.
classes of slot-based and resource-based models. The data Third, a more recent study tested whether observers
from set size manipulations such as those shown in Figures could trade precision for capacity, increasing the number of
3 and 4 are not yet conclusive, but three additional items stored in VWM beyond the typical Kmax by decreas-
approaches have provided evidence in favor of discrete ing the precision of the representations [58]. Observers
slots. were never able to increase Kmax by reducing precision,
First, Rouder et al. recorded confidence judgments from even when given monetary incentives to do so. This is
observers in a change detection task so that receiver strong evidence against the idea that resources can be
operating characteristic (ROC) curves could be constructed allocated flexibly to increase the number of VWM repre-
[45]. When large change magnitudes are used, the all-or- sentations.
none storage posited by slot models should lead to a linear Fourth, if observers devoted all of their resources to the
ROC curve, whereas low-resolution memory representa- items in the display, regardless of whether one or 20 items
tions would lead to a bowed ROC curve. The ROC curves were present, then it is difficult to explain why ERP and
observed were close to linear, supporting the slot model fMRI measures of VWM delay activity increase as the set
assumption of all-or-none memory encoding. size increases from one up to the individual observer’s Kmax
Second, Zhang and Luck combined the color wheel and then reach an asymptote [22,24]. Resource models
paradigm shown in Figure 3B with a spatial cuing manip- would instead predict that delay period activity should
ulation [48]. The sample array contained two items along be constant as long as observers are devoting all their
with a spatial cue, which could be valid, invalid, or neutral. resources to the task.
396
Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

Multiple sources of evidence therefore support the idea basal ganglia, and parietal cortex [61]. In addition, recovery
of discrete slots. However, state-of-the-art resource models is slower in low-capacity individuals than in high-capacity
[44,53] have not yet had an opportunity to explain these individuals after attention is captured by irrelevant infor-
other results. Thus, significant progress is being made and mation [62]. Thus, differences in Kmax among healthy young
many specific models have been ruled out, but additional adults may not reflect the capacity of VWM per se, but may
research will be needed to provide definitive evidence instead reflect variations in filtering processes that control
distinguishing between these two broad classes of model. the encoding of information into VWM.
Although attentional differences explain the bulk of the
Individual and group differences in visual working between-subject variability in VWM capacity among
memory capacity healthy college students, other factors appear to contribute
As discussed earlier, VWM capacity is a stable individual to group differences. For example, Kmax is lower in patients
difference and is impressively correlated with measures of with Parkinson’s disease than in healthy control subjects,
higher cognitive function (Figure 1E). In addition, substan- but only part of this can be explained by impaired filtering;
tial differences in VWM capacity can be observed across these patients also appear to have lower VWM storage
groups. For example, Kmax is lower in people with schizo- capacity per se [63]. Similarly, Kmax declines in aging, as
phrenia than in healthy control subjects [59,60], with a does filtering efficiency, but the timing of the filtering
very large effect size (Cohen’s d) of 1.11 in a study of 99 efficiency differences between younger and older adults
patients and 77 controls [13]. Moreover, the reduction in is not the same as that for differences between low- and
Kmax accounted for approximately 40% of the impairment high-Kmax younger adults[64,65].
for a measure of broad intellectual function in the patient In addition, the lower Kmax exhibited by schizophrenia
group. Assuming that reduced capacity actually causes patients does not appear to reflect impaired filtering at all.
reduced intellectual function (Box 2), this finding suggests First, schizophrenia patients show the same ability to
that 40% of the cognitive deficit in people with schizophre- exclude irrelevant distractor items as control subjects do
nia could be eliminated by a treatment that normalizes [60]. Second, CDA amplitude is actually greater in schizo-
their VWM capacity. phrenia patients than in control subjects when a single
Several studies have attempted to determine the causes object is stored in memory, but is lower in patients com-
underlying individual and group differences in VWM capac- pared to controls when three or five items must be stored
ity. Among healthy young adults, Vogel and colleagues [66]. This difference was observed even in subsets of
found that differences in VWM capacity can be attributed patients and controls with equivalent Kmax values. The
to differences in attentional processes that are responsible patient impairment may reflect a tendency to hyperfocus
for filtering out irrelevant information. For example, when on a small number of items, directly reducing VWM capac-
given arrays containing both relevant and irrelevant ity. Schizophrenia patients may also tend to hyperfocus on
objects, low-capacity individuals tend to encode irrelevant perceptually salient information: they exhibit impaired
information into VWM to a greater extent than high-capaci- filtering [67] and slowed disengagement [68] when faced
ty individuals do [23]. An fMRI study indicated that this with high-salience distractors.
reflects impaired connectivity among prefrontal cortex, the
Neural mechanisms of visual working memory
The simplicity of the change detection task makes VWM
Box 2. Outstanding questions amenable to neural network modeling and electrophysio-
logical recordings. Most neural network models assume
" Debate is continuing about whether VWM is best conceived as a that VWM representations are maintained by recurrent
set of discrete slot-like representations or as a flexible continuous
feedback loops, in which information flows from one set of
resource. Substantial progress has been made and many specific
models have been ruled out by the data. However, creative new neurons to another and then back again (Figure 5A) [69].
experimental designs and analytical procedures are needed This explains the persistence of memories, the increased
before we can definitively distinguish between these broad theory neural activity observed during the retention interval, and
classes. the finding of oscillations as the activity bounces back and
" Most models assume that trial-by-trial variations in memory (e.g.,
the distribution of errors shown in Figure 2D, in main text) reflect
forth among neural populations.
neural noise. However, much of this variance could instead reflect A recurrent feedback loop can easily maintain a single
systematic differences across trials (e.g., differences in the specific item, but it is more difficult to keep representations of
stimuli being remembered). The source of this variance is multiple different items from collapsing into a single re-
therefore an important issue for future research. presentation. To solve this problem, all the neurons that
" Studies of individual differences typically assume that having
more memory capacity causes people to perform better on
represent a given item are linked together in a synchro-
broader tests of cognitive ability. However, the direction of nously firing cell assembly, and only one cell assembly fires
causation may actually be in the opposite direction. That is, at a given moment in time (Figure 5B). Synchrony within a
smarter people may figure out better ways to perform working cell assembly helps to maintain recurrent activation, and
memory tasks. Determining the actual direction of causality will
asynchrony between different cell assemblies avoids inter-
be vitally important in future research.
" If variations in VWM capacity actually cause variations in overall ference between the representations of different items. A
cognitive ability, then the next obvious question is whether it is synchronous cell assembly effectively serves as a slot in
possible to improve VWM capacity and thereby improve overall VWM. The cell assemblies are formed dynamically, com-
cognitive ability. This is currently a hot topic among working bining whatever set of neurons is necessary to represent a
memory researchers, but no clear answer has yet emerged.
given object.
397
 Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

(A)
Area 1 Area 2

Neurons Neurons Neurons Neurons Neurons Neurons

coding coding coding coding coding coding
object object object object object object
#1 #2 #3 #1 #2 #3

(B) Synchronized
spiking

Neurons
coding Passi
ve de
object cay
#1

Neurons
coding
object
#2

Neurons
coding
object
#3
Time
TRENDS in Cognitive Sciences

Figure 5. Neural representation of three cell assemblies (groups of neurons coding separate objects in visual working memory, VWM). Each cell assembly consists of a
group of neurons from one or more cortical areas. In some models, neurons are recruited to a specific cell assembly at the moment of encoding to represent the features of
the object being encoded, and a given neuron may be allocated to different cell assemblies depending on the information being stored in memory. (A) Groups of neurons
coding a given object form local recurrent loops within an area (small U-shaped arrows) and long-range recurrent loops between areas (large arrows). The recurrent
connections cause the activity to be maintained over time, and the activity oscillates as it bounces back and forth between neurons (both within and between cortical areas).
Most models include only one or two cortical areas (e.g., inferotemporal and prefrontal cortex), but many different areas are likely synchronized in this manner. (B) The
neurons in a given cell assembly spike together briefly (represented by vertical lines) and then the activity decays. The different cell assemblies spike at different times,
minimizing interference between them. However, a given cell assembly must spike again before it decays too far (in which case the cell assembly stops firing and the VWM
representation is lost). This limits the number of cell assemblies that can be simultaneously active without either interfering with each other or decaying into oblivion.

In these models, a cell assembly passively decays after and gamma-band oscillations represent the individual cell
each time it fires, and the representation will be lost if too assemblies [73], which are then sequenced by means of
much time passes before it fires again. Consequently, the coupling to theta-band oscillations [74–77]. Alpha-band
number of items that can be maintained is limited by the oscillations may also play an important role. For example,
need to keep multiple cell assemblies from firing at the asymmetric modulations of alpha amplitude may contrib-
same time, while also preventing long delays between ute to sustained slow waves such as the CDA [78].
successive firings of a given cell assembly so that it does
not decay too far [70,71]. Realistic biophysical parameters Putting it all together
lead to an average capacity of three or four discrete objects, Differences in VWM capacity among healthy individuals
with some stochastic variation in the number of items are strongly predictive of broad cognitive abilities [13,14],
stored on each trial [71]. This model can also explain the and impairments in VWM capacity in patient groups may
fact that multidimensional objects can be remembered as provide an important key to understanding their real-
easily as single-dimension objects [8], because the neurons world cognitive impairments [13]. It is remarkable that
coding different dimensions can be synchronized into a memory for simple stimuli such as colored squares is so
single cell assembly [71,72]. In general, models of this strongly predictive of broader measures of cognitive ability
nature can explain how slot-like behavior can arise from and so clearly impaired in a variety of groups. However,
the dynamics of a continuous neural network [56]. this is fortunate, because VWM for simple colored squares
It is difficult to test these models from neural recordings is amenable to rigorous psychophysical measurement,
because of the difficulty of recording from dozens of indi- neural network modeling, ERP and fMRI experiments in
vidual neurons at the same time and determining how humans, and invasive measures of neural activity in ani-
they are linked together. Nonetheless, electrophysiologi- mals. Consequently, we are rapidly gaining a detailed
cal recordings from both humans and non-human pri- mechanistic understanding of the factors that determine
mates have provided evidence that synchronized spikes VWM capacity, and this may in turn lead to major
398
Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

advances in understanding individual and group differ- 26 Henderson, J.M. (2008) Eye movements and visual memory. In Visual
Memory (Luck, S.J. and Hollingworth, A., eds), pp. 87–121, Oxford
ences in broader cognitive function.
University Press
27 Irwin, D.E. (1992) Memory for position and identity across eye
Acknowledgments movements. J. Exp. Psychol. Learn. Mem. Cogn. 18, 307–317
This study was supported by grants from the National Institute of Mental 28 Irwin, D.E. and Andrews, R.V. (1996) Integration and accumulation of
Health (R01MH076226 and R01MH065034 to S.J.L. and R01MH087214 information across saccadic eye movements. In Attention and
to E.K.V.) and the Office of Naval Research (N000141210972 to E.K.V.). Performance XVI (Inui, T. and McClelland, J.L., eds), pp. 125–155,
We thank Ed Awh and Wei Ji Ma for many interesting discussions of MIT Press
these issues and Andy Yonelinas for comments on the manuscript. 29 Hollingworth, A. et al. (2008) Understanding the function of visual
short-term memory: transsaccadic memory, object correspondence,
References and gaze correction. J. Exp. Psychol. Gen. 137, 163–181
1 Anderson, J.R. et al. (1997) ACT-R: a theory of higher level cognition and 30 Mannan, S.K. et al. (2010) Early oculomotor capture by new onsets
its relation to visual attention. Hum. Comput. Interact. 12, 439–462 driven by the contents of working memory. Vis. Res. 50, 1590–1597
2 Meyer, D.E. and Kieras, D.E. (1997) A computational theory of 31 Hollingworth, A. et al. (2013) Visual working memory modulates rapid
executive cognitive processes and multiple-task performance: part 1. eye movements to simple onset targets. Psychol. Sci. 24, 790–796
Basic mechanisms. Psychol. Rev. 104, 3–65 32 Carlisle, N.B. et al. (2011) Attentional templates in visual working
3 Kyllonen, P.C. and Christal, R.E. (1990) Reasoning ability is (little memory. J. Neurosci. 31, 9315–9322
more than) working-memory capacity?! Intelligence 14, 389–433 33 Woodman, G.F. et al. (2007) The role of working memory
4 Engle, R.W. et al. (1999) Working memory, short-term memory, and representations in the control of attention. Cereb. Cortex 17, i118–i124
general fluid intelligence: a latent-variable approach. J. Exp. Psychol. 34 Woodman, G.F. et al. (2001) Visual search remains efficient when
Gen. 128, 309–331 visual working memory is full. Psychol. Sci. 12, 219–224
5 Simons, D.J. and Levin, D.T. (1997) Change blindness. Trends Cogn. 35 Luria, R. and Vogel, E.K. (2011) Visual search demands dictate
Sci. 1, 261–267 reliance on working memory storage. J. Neurosci. 31, 6199–6207
6 Simons, D.J. and Rensink, R.A. (2005) Change blindness: past, present, 36 Anderson, D.E. et al. (2013) A common discrete resource for visual
and future. Trends Cogn. Sci. 9, 16–20 working memory and visual search. Psychol. Sci. 24, 929–938
7 Phillips, W.A. (1974) On the distinction between sensory storage and 37 Drew, T. et al. (2011) Delineating the neural signatures of tracking
short-term visual memory. Percept. Psychophys. 16, 283–290 spatial position and working memory during attentive tracking. J.
8 Luck, S.J. and Vogel, E.K. (1997) The capacity of visual working Neurosci. 31, 659–668
memory for features and conjunctions. Nature 390, 279–281 38 Drew, T. and Vogel, E.K. (2008) Neural measures of individual
9 Irwin, D.E. (1991) Information integration across saccadic eye differences in selecting and tracking multiple moving objects. J.
movements. Cogn. Psychol. 23, 420–456 Neurosci. 28, 4183–4191
10 Luck, S.J. (2008) Visual short-term memory. In Visual Memory (Luck, 39 Tsubomi, H. et al. (2013) Neural limits to representing objects still
S.J. and Hollingworth, A., eds), pp. 43–85, Oxford University Press within view. J. Neurosci. 33, 8257–8263
11 Gibson, B. et al. (2011) Qualitative similarities in the visual short-term 40 Vogel, E.K. and Awh, E. (2008) How to exploit diversity for scientific
memory of pigeons and people. Psychonom. Bull. Rev. 18, 979–984 gain: using individual differences to constrain cognitive theory. Curr.
12 Kyllingsbaek, S. and Bundesen, C. (2009) Changing change detection: Dir. Psychol. Sci. 17, 171–176
improving the reliability of measures of visual short-term memory 41 Awh, E. et al. (2007) Visual working memory represents a fixed number
capacity. Psychonom. Bull. Rev. 16, 1000–1010 of items regardless of complexity. Psychol. Sci. 18, 622–628
13 Johnson, M.K. et al. (2013) The relationship between working memory 42 Alvarez, G.A. and Cavanagh, P. (2004) The capacity of visual short-
capacity and broad measures of cognitive ability in healthy adults and term memory is set both by information load and by number of objects.
people with schizophrenia. Neuropsychology 27, 220–229 Psychol. Sci. 15, 106–111
14 Fukuda, K. et al. (2010) Quantity, not quality: the relationship between 43 Brady, T.F. et al. (2011) A review of visual memory capacity: beyond
fluid intelligence and working memory capacity. Psychonom. Bull. Rev. individual items and toward structured representations. J. Vis. 11, 4
17, 673–679 44 van den Berg, R. et al. (2012) Variability in encoding precision accounts
15 Lin, P-H. and Luck, S.J. (2012) Proactive interference does not for visual short-term memory limitations. Proc. Natl. Acad. Sci. U.S.A.
meaningfully distort visual working memory capacity estimates in 109, 8780–8785
the canonical change detection task. Front. Psychol. 3, 42 45 Rouder, J.N. et al. (2008) An assessment of fixed-capacity models
16 Hartshorne, J.K. (2008) Visual working memory capacity and proactive of visual working memory. Proc. Natl. Acad. Sci. U.S.A. 105,
interference. PLoS ONE 3, e2716 5975–5979
17 Shipstead, Z. and Engle, R.W. (2013) Interference within the focus of 46 Bays, P.M. and Husain, M. (2008) Dynamic shifts of limited working
attention: working memory tasks reflect more than temporary memory resources in human vision. Science 321, 851–854
maintenance. J. Exp. Psychol. Learn. Mem. Cogn. 39, 277–289 47 Wilken, P. and Ma, W.J. (2004) A detection theory account of change
18 Jeneson, A. et al. (2012) Visual working memory capacity and the detection. J. Vis. 4, 1120–1135
medial temporal lobe. J. Neurosci. 32, 3584–3589 48 Zhang, W. and Luck, S.J. (2008) Discrete fixed-resolution
19 Olson, I.R. et al. (2006) Working memory for conjunctions relies on the representations in visual working memory. Nature 453, 233–235
medial temporal lobe. J. Neurosci. 26, 4596–4601 49 Anderson, D.E. et al. (2011) Precision in visual working memory
20 Fuster, J.M. and Jervey, J.P. (1982) Neuronal firing in the reaches a stable plateau when individual item limits are exceeded.
inferotemporal cortex of the monkey in a visual memory task. J. Neurosci. 31, 1128–1138
J. Neurosci. 2, 361–375 50 Anderson, D.E. and Awh, E. (2012) The plateau in mnemonic
21 Miller, E.K. and Desimone, R. (1991) A neural mechanism for working resolution across large set sizes indicates discrete resource limits in
and recognition memory in inferior temporal cortex. Science 254, visual working memory. Atten. Percept. Psychophys. 74, 891–910
1377–1379 51 Xu, Y. and Chun, M.M. (2006) Dissociable neural mechanisms
22 Vogel, E.K. and Machizawa, M.G. (2004) Neural activity predicts supporting visual short-term memory for objects. Nature 440, 91–95
individual differences in visual working memory capacity. Nature 52 Thiele, J.E. et al. (2011) On perfect working-memory performance with
428, 748–751 large numbers of items. Psychonom. Bull. Rev. 18, 958–963
23 Vogel, E.K. et al. (2005) Neural measures reveal individual differences 53 Sims, C.R. et al. (2012) An ideal observer analysis of visual working
in controlling access to working memory. Nature 438, 500–503 memory. Psychol. Rev. 119, 807–830
24 Todd, J.J. and Marois, R. (2005) Posterior parietal cortex activity 54 Saults, J.S. and Cowan, N. (2007) A central capacity limit to the
predicts individual differences in visual short-term memory simultaneous storage of visual and auditory arrays in working
capacity. Cogn. Affect. Behav. Neurosci. 5, 144–155 memory. J. Exp. Psychol. Gen. 136, 663–684
25 Todd, J.J. and Marois, R. (2004) Capacity limit of visual short-term 55 Franconeri, S.L. et al. (2013) Flexible cognitive resources: competitive
memory in human posterior parietal cortex. Nature 428, 751–754 content maps for attention and memory. Trends Cogn. Sci. 17, 134–141

399
Review Trends in Cognitive Sciences August 2013, Vol. 17, No. 8

56 Wei, Z. et al. (2012) From distributed resources to limited slots in 70 Lisman, J.E. and Idiart, M.A.P. (1995) Storage of 7#2 short-term
multiple-item working memory: a spiking network model with memories in oscillatory subcycles. Science 267, 1512–1515
normalization. J. Neurosci. 32, 11228–11240 71 Raffone, A. and Wolters, G. (2001) A cortical mechanism for binding in
57 Lin, P-H. and Luck, S.J. (2008) The influence of similarity on visual visual working memory. J. Cogn. Neurosci. 13, 766–785
working memory representations. Vis. Cogn. 17, 356–372 72 Luck, S.J. and Vogel, E.K. (1998) Response from Luck and Vogel.
58 Zhang, W. and Luck, S.J. (2011) The number and quality of Trends Cogn. Sci. 2, 78–80
representations in working memory. Psychol. Sci. 22, 1434–1441 73 Pesaran, B. et al. (2002) Temporal structure in neuronal activity during
59 Gold, J.M. et al. (2003) Working memory for visual features and working memory in macaque parietal cortex. Nat. Neurosci. 5, 805–811
conjunctions in schizophrenia. J. Abnorm. Psychol. 112, 61–71 74 Liebe, S. et al. (2012) Theta coupling between V4 and prefrontal cortex
60 Gold, J.M. et al. (2006) Intact attentional control of working memory predicts visual short-term memory performance. Nat. Neurosci. 15,
encoding in schizophrenia. J. Abnorm. Psychol. 115, 658–673 456–462 S451–452
61 McNab, F. and Klingberg, T. (2008) Prefrontal cortex and basal ganglia 75 Lee, H. et al. (2005) Phase locking of single neuron activity to theta
control access to working memory. Nat. Neurosci. 11, 103–107 oscillations during working memory in monkey extrastriate visual
62 Fukuda, K. and Vogel, E.K. (2009) Human variation in overriding cortex. Neuron 45, 147–156
attentional capture. J. Neurosci. 29, 8726–8733 76 Sauseng, P. et al. (2009) Brain oscillatory substrates of visual short-
63 Lee, E-Y. et al. (2010) Visual working memory deficits in patients with term memory capacity. Curr. Biol. 19, 1846–1852
Parkinson’s disease are due to both reduced storage capacity and 77 Siegel, M. et al. (2009) Phase-dependent neuronal coding of objects in
impaired ability to filter out irrelevant information. Brain 133, short-term memory. Proc. Natl. Acad. Sci. U.S.A. 106, 21341–21346
2677–2689 78 van Dijk, H. et al. (2013) Modulations in oscillatory activity with
64 Jost, K. et al. (2011) Are old adults just like low working memory young amplitude asymmetry can produce cognitively relevant event-
adults? Filtering efficiency and age differences in visual working related responses. Proc. Natl. Acad. Sci. U.S.A. 107, 900–905
memory. Cereb. Cortex 21, 1147–1154 10.1073/pnas.0908821107
65 Cashdollar, N. et al. (2013) Prolonged disengagement from attentional 79 Hyun, J-S. et al. (2009) The comparison of visual working memory
capture in normal aging. Psychol. Aging 28, 77–86 representations with perceptual inputs. J. Exp. Psychol. Hum. Percept.
66 Leonard, C.J. et al. (2013) Toward the neural mechanisms of reduced Perform. 35, 1140–1160
working memory capacity in schizophrenia. Cereb. Cortex 23, 1582–1592 80 Pashler, H. (1988) Familiarity and visual change detection. Percept.
67 Hahn, B. et al. (2010) Failure of schizophrenia patients to overcome Psychophys. 44, 369–378
salient distractors during working memory encoding. Biol. Psychiatry 81 Cowan, N. et al. (2005) On the capacity of attention: its estimation and
68, 603–609 its role in working memory and cognitive aptitudes. Cogn. Psychol. 51,
68 Mayer, J.S. et al. (2012) Impaired contingent attentional capture 42–100
predicts reduced working memory capacity in schizophrenia. PLoS 82 Rouder, J.N. et al. (2011) How to measure working memory capacity in
ONE 7, e48586 the change detection paradigm. Psychonom. Bull. Rev. 18, 324–330
69 Deco, G. and Rolls, E.T. (2008) Neural mechanisms of visual memory: a 83 Morey, R.D. and Morey, C.C. (2011) WoMMBAT: A user interface for
neurocomputational perspective. In Visual Memory (Luck, S.J. and hierarchical Bayesian estimation of working memory capacity. Behav.
Hollingworth, A., eds), pp. 247–289, Oxford University Press Res. Methods 43, 1044–1065

400