Cognition 191 (2019) 103984

Contents lists available at ScienceDirect

Cognition
journal homepage: www.elsevier.com/locate/cognit

Original Articles

What can half a million change detection trials tell us about visual working T
memory?
⁎
Halely Balabana, , Keisuke Fukudab, Roy Luriaa
a
Sagol School of Neuroscience and The School of Psychological Sciences, Tel Aviv University, Tel Aviv 6997801, Israel
b
Department of Psychology, University of Toronto Mississauga, Mississauga, ON L5L 1C6, Canada

A R T I C LE I N FO A B S T R A C T

Keywords: Visual working memory (VWM) represents the surrounding world in an active and accessible state, but its
Working memory capacity capacity is severely limited. To better understand VWM and its limits, we collected data from over 3,800 par-
Individual differences ticipants in the canonical change detection task. This unique population-level data-set sheds new light on classic
Change detection debates regarding VWM capacity. First, the result supported a view of VWM as an active process, as manifested
by the fact that capacity estimates were not stable across set-sizes, but rather lower for the larger set-size.
Another support for this notion came from the tight connection capacity estimates had with a measure of at-
tentional control. Together, the data suggested that individual differences in capacity do not reflect only dif-
ferences in storage-size, but differences in the efficiency of using this storage. Second, we found a response bias
such that subjects are more likely to respond that the probed item changed, and this criterion bias was further
shifted as the set-size increased. These findings are naturally explained by a slot-like theory arguing that when
load exceeds capacity, certain items remain completely outside of VWM (instead of all items being represented in
lower resolution), therefore causing subjects to perceive them as different from VWM contents even when they
are unchanged. Additionally, we found that the pattern of d' also confirmed the predictions of a slot-like view of
VWM, such that some items are represented with high fixed resolution and others are not represented at all,
although this finding is based on two measures with very different underlying assumptions. We also discuss how
flexible-resource views can accommodate these results. Moreover, comparing performance between the first and
last trials demonstrated no evidence for proactive interference as the driving factor of capacity limitations. We
provide further details regarding the distribution of individual capacity, the relations between capacity and
demographic variables, and the spatial prioritization of the items.

1. Introduction existence of a severe limitation on this capacity is vastly acceptable and

can be considered as one of the defining characteristics of VWM.
Visual working memory (VWM) holds relevant visual information in Despite average capacity limits being quite low (∼3 simple items’
an active state, ready to be accessed and manipulated by higher cog- worth of information), a great deal of variability exists between in-
nitive functions (Cowan, 2001). However, only a very limited amount dividuals. Capacity is highly stable at the individual level (e.g., across
of information can be retained in this privileged state, creating a bot- set-sizes or blocks, Cronbach’s alphas > 0.9; Xu, Adam, Fang, & Vogel,
tleneck for how we process incoming information. Corroborating the 2018), and is tightly correlated with measures of fluid intelligence,
importance of VWM in everyday behavior, it is specifically damaged in attentional control, and many aptitude measures (e.g., Cowan et al.,
a range of conditions such as Alzheimer’s disease, attention deficit 2005; Fukuda, Vogel, Mayr, & Awh, 2010; Vogel, McCollough, &
hyperactivity disorder (ADHD), old age, and schizophrenia (e.g., Machizawa, 2005). Understanding the nature of these capacity differ-
Johnson et al., 2013; Jost, Bryck, Vogel, & Mayr, 2011; Martinussen, ences and further exploring how VWM is related to other important
Hayden, Hogg-Johnson, & Tannock, 2005; Parra et al., 2011). The cognitive constructs is the focus of numerous ongoing research projects.
nature of VWM capacity is heavily debated (e.g., whether it is better A prominent method for quantifying VWM performance is the
described as a continuous or discrete resource, cf. Brady, Konkle, & change detection paradigm (Luck & Vogel, 1997; Pashler, 1988;
Alvarez, 2011; Luck & Vogel, 2013; Ma, Husain, & Bays, 2014), but the Phillips, 1974). In the canonical form of the task, several simple items

⁎
Corresponding author.
E-mail address: halelyba@mail.tau.ac.il (H. Balaban).

https://doi.org/10.1016/j.cognition.2019.05.021
Received 9 November 2018; Received in revised form 16 May 2019; Accepted 25 May 2019
Available online 21 June 2019
0010-0277/ © 2019 Elsevier B.V. All rights reserved.
H. Balaban, et al. Cognition 191 (2019) 103984

are briefly presented, and after a short retention interval the test array
appears, with an item in one of the previously-occupied locations. The
subjects' task is to indicate whether the probed item, i.e., the item in the
test array, is the same or different (typically with an equal probability)
from the item appearing in the same place in the memory array. Per-
formance is usually transformed into a capacity estimate (K), i.e., the
number of items that can be successfully retained in VWM (Cowan,
2001; Pashler, 1988). Even a ∼10 min long version of this task is highly
reliable, with capacity measures being stable between testing sessions
more than a year apart (r = 0.77; Johnson et al., 2013). This makes the Fig. 1. Trial sequence in the change detection task employed in the present
paradigm ideal for examining individual differences in capacity. study. A memory array of colored squares appeared for a short duration, fol-
In the present study, we report the results of over 3,800 subjects that lowed by a blank retention interval, and then a test array with only one probe
completed a short change detection task in our lab, before participating that can be either the same as the item in the same location in the memory
in one of many different experiments. Usually, each experiment (even array, or different. This is an example of a set-size 4 trial, including a change.
when sample-size is adequate to examine individual differences) pro-
vides a small number of trials overall. Pooling together trials from an keys (i.e., pressing the “same” key for a “different” response and vice
extremely large number of subjects, for a total of over 460,000 trials of versa). Additionally, due to an error in the code, some participants
the same task, allowed us to analyze the data in ways that are beyond completed more than 120 trials. We included data-sets with 121–130
the scope of ordinary studies. Our main goal was to shed new light on trials (54 data-sets, 1.4%), and excluded data-sets with over 130 trials
VWM capacity limitations. Additionally, the large sample size allowed (28 data-sets, 0.7%). This left a total of 3,849 analyzed data-sets with
us to better characterize the change detection task, in terms of perfor- 462,186 trials.
mance biases, both at the individual and at the “population” level.
An important advantage of the canonical change detection para-
2.2. Stimuli and procedure
digm is that task performance is only minimally influenced by processes
external to VWM, such as verbal working memory, iconic memory,
Subjects performed a ∼10 min color change detection task (see
long-term memory, and response-related processes (Cowan, 2001). This
Fig. 1). A memory array of colored squares (approximately 1.3° × 1.3°
had made change detection performance almost synonym to VWM ca-
of visual angle, from a viewing distance of approximately 60 cm) ap-
pacity. However, it is important to remember that the change detection
peared for 150 ms around a black fixation cross (0.4° × 0.4°) at the
task is not a cognitive process per se, but a paradigm, and the ability of
center of a grey (RGB values: 125,125,125) screen. Items were ran-
this paradigm to adequately measure VWM processes depends on many
domly selected, without replacement, from a set of 9 highly dis-
carefully-thought parameters of the experimental setup (Vogel,
criminable colors: red, magenta, blue, cyan, green, yellow, orange,
Woodman, & Luck, 2001). Specifically, isolating VWM processes re-
brown, and black (RGB values, respectively: 254,0,0; 255,0,254;
quires using short presentation times or adding articulatory suppression
0,0,254; 0,255,255; 0,255,1; 255,255,0; 255,128,65; 128,64,0; 0,0,0;
to prevent verbal coding (Luck & Vogel, 1997; Vogel et al., 2001), in-
note that the colors would render differently depending on the video
cluding masking or a long-enough (∼1 s) retention interval to delete
card and monitor). Participants were instructed to memorize the
the retinal after-image (Phillips, 1974), and using simple and highly
squares’ colors. The squares disappeared for a 900 ms retention interval
discriminable items to minimize comparison-process errors that arise
(leaving only the fixation cross), and then the probe appeared: a single
for complex stimuli (Awh, Barton, & Vogel, 2007). Adhering to these
square in one of the previously occupied locations (randomly de-
conventions allow strong conclusions about VWM to be drawn directly
termined). Participants’ task was to indicate in a non-speeded manner,
from change detection performance.
via button press, whether the probe was the same color as the square in
that location, or a different color (using the “Z” and “/” keys on a
2. Materials and methods
standard keyboard, respectively; a subset of the subjects did a coun-
terbalanced version, see the Results section).
All analyzed data-sets can be found at the Open Science Framework:
Half of the trials included 4 colors in the memory array, and the
http://dx.doi.org/10.17605/OSF.IO/MZS9E.
other half included 8 colors (randomly intermixed). Half of the trials
included a change, and half did not (randomly intermixed). After about
2.1. Participants
6 practice trials, participants completed one block of 120 experimental
trials, with 30 trials for each combination of set-size (4 or 8) and trial
Subjects were mostly Tel Aviv University students and several in-
type (change or no-change).
dividuals from the Tel Aviv University community, who completed the
On each trial, one quarter of the items (1 in set-size 4 trials and 2 in
task before taking part in a longer experiment (a small subset of the K
set-size 8 trials) were presented on each quadrant. Items’ locations were
estimates data was previously published in Allon & Luria, 2017; Allon,
selected randomly from 9 possible locations in each quadrant: 2.7°,
Vixman, & Luria, 2018; Vaskevich & Luria, 2018). Subjects participated
5.3°, or 8° vertically from the center of the screen, and 2.7°, 5.3°, or 8°
for either payment (approximately 40 NIS, or $10, per hour), partial
horizontally from the center of the screen (e.g., in the top left quadrant,
course credit, or voluntarily without compensation. All had normal or
items could appear centered at 2.7°, 5.3°, or 8° to the left of the fixation,
corrected-to-normal visual acuity and normal color vision. All partici-
and 2.7°, 5.3°, or 8° above the fixation).
pants gave informed consent following the procedures of protocols
approved by the Ethics Committee at Tel Aviv University. Age ranged
between 18 and 39. 2.3. Individual-level analyses
A total of 3,923 data-sets were collected for the following analyses.
Some participants took part in more than one experiment at the lab, and We extracted from each data-set a measure of K, using a standard
thus contributed more than one data-set to the current analysis (ap- formula (Cowan, 2001; Pashler, 1988): K = N × (H − FA), where K is
proximately 1,000 data sets came from such repeated participants). We the capacity estimate, N is the set-size, H is hit rate, and FA is the false
excluded data-sets with below-chance performance at set-size 4 or a K alarm rate (meaning the individual proportion of correct responses in
estimate of less than 0 (46 data-sets, 1.2%), because this results from change trials, and of incorrect responses in no-change trials, respec-
below-chance performance which likely reflects switching the response tively). In the present situation, of a single-probe task with a 50%

2
H. Balaban, et al. Cognition 191 (2019) 103984

probability of change, an equivalent formula is K = N × (2 × Accu- Xu et al., 2018). Second, we found large individual differences, as
racy − 1), where Accuracy is the proportion correct across all trials of manifested by the 0.8 items standard deviation. Fig. 2 depicts the dis-
that set-size (regardless of whether there was or wasn’t a change). tribution of K values across participants.
Following standard practice, we averaged K across the two set-sizes. As can be seen in Fig. 2, the observed K values appear to follow a
Additionally, we calculated the signal detection theory measures of normal distribution. To qualitatively estimate this, we present a Q-Q
d', which in the present context is an index of memory signal strength, plot in Fig. 3. This plot compares our observed quantiles to the theo-
and criterion, following standard practices for these kinds of “yes/no” retical quantiles expected from a normal distribution. The more similar
tasks (Macmillan & Creelman, 2004). Note that in the change detection the two compared distributions, the closer the individual data points
paradigm, since the task is to indicate whether a change occurred, a will be to a straight line. As can be seen from the figure, the observed
“yes” response corresponds to a changed or new item, unlike in classical data indeed closely matches the expected quantiles, suggesting that the
“old/new” recognition task in which a “yes” response corresponds to a data is normally distributed. To quantitively support this observation,
previously-shown item (e.g., Keshvari, van den Berg, & Ma, 2013). we performed the D’Agostino K2 test, which suggested that the dis-
Memory signal strength, i.e., d', was calculated as: d′ = Z(H) − Z(FA), tribution is not significantly different from a normal distribution
where Z(H) is the Z-score of the hit rate, and Z(FA) is the Z-score of the (K2 = 1.23, p = 0.54). It is noteworthy that the results of an extremely
false alarm rate. Criterion, or bias, was calculated as C = −0.5 × (Z simple 10-min color memory task are normally distributed.
(H) + Z(FA)), where C is the criterion, and Z(H) and Z(FA) are as in the It might be argued that this normality does not reflect the shape of
d′ calculation. the K construct itself, but rather the normality of the measurement error
in this task, resulting from a sum of binary choices. However, K esti-
2.4. Group-level analyses mates for set-size 8 separately was not normally distributed
(K2 = 31.08, p < 0.001; we tested this for set-size 8 because although
To examine several fine-grained characteristics of the change de- K at set-size 4 was also not normally distributed, it has a relatively low
tection task, we pooled together all data-sets, for a total of 462,186 upper bound), despite also involving similar measurement errors. In
trials. We then performed two analyses on the aggregate set of trials. fact, K estimates for 8 items were positively skewed, and K estimates for
First, we computed K separately for each serial position of the trial 4 items were negatively skewed, which suggests a different alternative
within the experiment (1–120; 3,849 trials per serial position number). explanation for the normality of average K estimates: the normal dis-
This means we treated all trials from a single position as if they came tribution might be an averaging artifact. To rule this out, we averaged
from a single subject that performed 3,849 trials and then extracted K in for each individual data-set two other oppositely-skewed measures,
the usual way. Second, we computed accuracy separately for each of the namely d' which was positively skewed and criterion which was nega-
36 probed locations (12,625–13,039 trials per location). For this ana- tively skewed (we used the measures from trials of both set-sizes
lysis, we also report a subset of 905 participants who did a counter- combined). We found that the average measure was not normally dis-
balanced version of the task in terms of the response keys mapping, in tributed (K2 = 26.91, p < 0.001), meaning that not every two oppo-
which we break down trials also by the type of trial (change vs. no- sitely-skewed measures will create a normal distribution when aver-
change). aged together. Finally, it might be argued that it is the variance in
individual estimates of K that follows a normal distribution (i.e., mea-
2.5. Statistical analyses suring each participant numerous times would have produced a normal
distribution of K estimates for that individual), thus producing a normal
Due to the very large sample-size in most of our analyses, traditional distribution across participants. If this is the case, however, it should
statistical tests are not informative, because they’re likely to be sig- also hold for other measures. We tested this for d' and criterion, and
nificant even for trivial effects. For the same reason, traditional con- found that they were not normally distributed, either when considering
fidence intervals (CIs) will be extremely narrow and hence not in- each set-sizes separately or when aggregating both set-sizes (all
dicative. Instead, for our main analyses we relied on the large sample- K2s > 18, all p’s < 0.001). Thus, the normal distribution of K esti-
size and treated the data-sets as a population. We report the observed mates is an important finding, which we suggest reflects the underlying
2.5 and 97.5 percentiles, i.e., the values that mark the range holding the normal distribution of VWM capacity in the population, demonstrating
central 95% of the “population” distribution. We refer to these observed that K is a meaningful construct.
percentiles as the population 95% range. We additionally report 95% Next, we examined K estimates at the different set-sizes. Capacity
Bayesian credibility intervals, calculated using the JASP software (with estimates at set-size 4 trials were higher than in set-size 8 trials (mean
default priors). We report traditional 95% CIs for correlations measures, K: 2.79 vs. 2.54, respectively), and varied less (SD: 0.66 vs. 1.21, re-
because for these measures a single value is computed for all subjects. spectively). Since both set-sizes are beyond the average capacity limit
To test whether the data is normally distributed, we performed (i.e., most participants cannot hold even 4 simple items in VWM), the
D’Agostino’s K2 test (D’agostino, Belanger, & D’agostino, 1990), based fact that the larger set-size produces a lower capacity estimates is in-
on skewness and kurtosis, which is suitable for our large sample-size. teresting. A view of VWM as a passive storage would lead to a predic-
tion that capacity should be stable across set-size (e.g., if one can only
3. Results hold 2 items, they should always hold 2 items, regardless of the pre-
sented set-size). This is not the case, however, and increasing the set-
3.1. Individual VWM capacity estimates size further decreases performance, which strongly indicates that VWM
is a dynamic process. Thus, capacity estimates do not solely reflect the
For each individual data-set, we calculated a measure of K, ac- size of the storage, but also the ability to use it well under different
cording to the above-mentioned formula. Average K was 2.66, with a circumstances, for example varying memory loads.
standard deviation of 0.83. Table 1 summarizes descriptive statistics for If that is the case, we should expect that average capacity estimates
all “population-level” reported measures. mainly reflect the larger set-size, which poses a greater difficulty. To
This replicates two of the key characteristics regarding VWM ca- test this, we examined the correlations between K estimates at each set-
pacity. First, the highly limited nature of this workspace (less than 3 size separately, and overall K. K estimates at set-size 4 and 8 had a
simple items) is very similar to previous reports, including the findings correlation of r = 0.53 (95% CI: [0.51, 0.56]), meaning they shared
of a recent large-scale study conducted in the US (average K: 2.55, 28% of their variance. The correlation between overall K and set-size
N = 495; Fukuda, Woodman, & Vogel, 2015), though somewhat higher 4 K was r = 0.79 (95% CI: [0.77, 0.80]), while the correlation between
than the findings of one conducted in China (average K: 2.14, N = 135; overall K and set-size 8 K was r = 0.94 (95% CI: [0.94, 0.95]), see

3
H. Balaban, et al. Cognition 191 (2019) 103984

Table 1
Descriptive statistics for the different measures: mean (standard deviation in parentheses), population 95% range (2.5 and 97.5 percentiles, see Methods), and lower
and higher boundaries of the 95% Bayesian credibility interval (CI).
Measure Mean (SD) Population 95% range 95% Bayesian CI

2.5 percentile 97.5 percentile Lower boundary Higher boundary

K (averaged across set-sizes) 2.66 (0.83) 1.07 4.27 2.64 2.69

K: set-size 4 2.79 (0.66) 1.33 3.87 2.77 2.81
K: set-size 8 2.54 (1.21) 0.27 5.07 2.50 2.58
4–8 drop 0.25 (1.03) −1.87 2.13 0.21 0.28
Hit rate 0.90 (0.08) 0.70 1.00 0.89 0.90
Correct rejection 0.61 (0.13) 0.33 0.85 0.61 0.62
Hit rate: set-size 4 0.93 (0.07) 0.77 1.00 0.93 0.93
Correct rejection: set-size 4 0.77 (0.14) 0.43 1.00 0.76 0.77
Hit rate: set-size 8 0.86 (0.11) 0.60 1.00 0.86 0.87
Correct rejection: set-size 8 0.46 (0.16) 0.17 0.77 0.45 0.46
Criterion: set-size 4 −0.41 (0.36) −1.11 −0.28 −0.42 −0.40
Criterion: set-size 8 −0.69 (0.44) −1.56 0.10 −0.71 −0.68
d': set-size 4 2.48 1.03 4.23 2.45 2.50
d': set-size 8 1.14 0.11 2.34 1.13 1.16
K: females 2.75 (0.81) 1.23 4.33 2.71 2.79
K: males 2.64 (0.82) 1.07 4.27 2.58 2.70

Fig. 4. Thus, the overall capacity estimate shared 62% of its variance
with the set-size 4 estimate, and as much as 89% of its variance with the
set-size 8 estimate. This tight relationship suggests that K mainly re-
flects the highest set-size (notably, because the variance is larger at set-
size 8, its correlations are expected to be higher, but here it completely
dominates the explained variance). This might lead one to argue that it
could be enough to measure only set-size 8, because of its high corre-
lation with overall K. Nevertheless, we argue that including only arrays
that are far beyond average capacity is problematic, because their ex-
treme difficulty might seriously damage subjects’ motivation. We do
note, however, that it is vital to include these set-sizes to properly es-
timate capacity.
The idea of VWM capacity as reflecting an active storage ability,
instead of storage size, could explain why VWM capacity is correlated
with the ability to filter out irrelevant distractors (Vogel et al., 2005): in
Fig. 2. A histogram of K (capacity estimate) values for our 3,849 individual
tasks involving only relevant items, when the set-size exceeds the in-
data-sets (density on the Y axis). The dashed black line depicts a normal dis- dividual capacity limit, over-capacity targets are just like distractors.
tribution with the same mean (μ) and standard deviation (σ) as our observed This means that set-size 8 is more difficult than set-size 4 partially
distribution, and these values are also presented on the histogram. because of the need to select a subset of items among more possibilities.
For this reason, the difference between K at set-sizes 4 and 8 (termed
“4–8 drop”) was previously used as a measure of attentional control,
argued to be conceptually separated from the size of VWM-storage
(Unsworth, Fukuda, Awh, & Vogel, 2014, 2015). We found that the 4–8
drop was on average greater than zero (mean: 0.25), with substantial
variance (SD: 1.03). While the 4–8 drop did not correlate with set-size
4 K (r = 0.01, 95% CI: [−0.02, 0.04]), it strongly correlated with set-
size 8 K (r = −0.84, 95% CI: [−0.85, −0.83]) and with overall K
(r = −0.61, 95% CI: [−0.63, −0.59]), such that high-capacity in-
dividuals had a smaller 4–8 drop, i.e., less of a difference between the
two set-sizes (see Fig. 5). Because the 4–8 drop shares 37% of its var-
iance with overall K and 71% of its variance with set-size 8 K, we argue
that the 4–8 drop does not reflect an ability independent from capacity.
Instead, both the 4–8 drop and classical capacity estimates seem to
reflect, at least to some degree, the ability to flexibly use limited VWM
resources. Another interesting possibility along this line is that the 4–8
drop reflects some sort of a strategy shift, which high-capacity in-
Fig. 3. A Q-Q plot of K values. Observed quantiles appear on the Y axis, and dividuals should be more likely to successfully utilize (as suggested by
theoretical quantiles, derived from a normal distribution, appear on the X axis. the correlations between VWM capacity and fluid intelligence, e.g.,
The diagonal line depicts the standardized x = y line (i.e., the line on which the Fukuda et al., 2010). This also goes with a view of capacity as the ef-
percentiles exactly match). The fewer deviations from the line, the more nor- ficient use of VWM.
mally distributed the data is. Notably, the correlations were calculated between different mea-
sures derived from the same task, using the same small set of trials. This
means that random factors that drive performance (e.g., attentional
fluctuations or guesses) might affect several different measures in a

4
H. Balaban, et al. Cognition 191 (2019) 103984

Fig. 4. The correlations between capacity estimates (K) from (a) set-sizes 4 and 8, r = 0.53; (b) set-size 4 and the average of both set-sizes, r = 0.79; (c) set-size 8 and
the average of both set-sizes, r = 0.94. To better illustrate the distribution, the dots are semi-transparent, hence the darkness of an area shows the frequency of this
combination of values.

similar way, thus inflating the observed correlations. To test this, we Thus, our analysis raised the possibilities that capacity estimates
divided each data-set into two parts, each constructed of only odd- or mostly reflect the larger set-size, and that the 4–8 drop is not in-
even-numbered trials. We then calculated the correlations between the dependent from the overall capacity estimates and from set-size 8, but
different measures from these mutually-exclusive subsets of trials, and more research is needed for a stronger conclusion to be drawn. If more
averaged across the two correlations (e.g., the correlation between K at direct evidence will support our observations, it will be in line with a
set-size 4 and averaged K was calculated as the average of the corre- view of capacity that diverges from its interpretation as simply the size
lation between K at set-size 4 in odd trials and average K in even trials, of VWM-storage. Instead, as is suggested by our finding that capacity
and the correlation between K at size 4 in even trials and average K in estimates are lower in the larger set-size (despite both set-sizes being
odd trials). Since this results in a very small number of trials con- above average capacity), we argue that capacity should be regarded as
tributing to each measure, we used the Spearman-Brown correction the ability to efficiently use VWM (Mance & Vogel, 2013). In this view,
(Brown, 1910; Spearman, 1910) to estimate the correlations for double individual differences in capacity do not solely manifest a different
the number of trials. storage-size, but instead largely reflect the differential capability to
In line with the hypothesis that the strong correlations stem at least flexibly and adaptably use a similarly-limited storage.
partially from the reliance on the same trials, when different trials were We next turned to compare trials that included a change in the
used to calculate each measure, the pattern of correlations was different probed color to trials in which the color was the same as in the memory
from what we originally observed. We found correlations of r = 0.41 array. Accuracy in change trials (i.e., hit rate) was much higher than in
between K at set-size 4 and K at set-size 8, r = 0.54 between K at set- no-change trials (i.e., correct rejection): 0.90 compared with only 0.61.
size 4 and overall K, and r = 0.58 between K at set-size 8 and overall K, This pattern was observed both for set-size 4 (0.93 vs. 0.77, respec-
lower than the correlations from all trials together. Thus, set-size 8 no tively) and for set-size 8 (0.86 vs. 0.46, respectively). This suggests a
longer dominated the overall K estimates. The correlation between the shift of criterion, and indeed we found that the criterion was −0.41 for
4–8 drop and K at set-size 4 remained around zero (r = -0.01), and the set-size 4 trials, and −0.69 for set-size 8 trials. This indicates that
correlations between the 4–8 drop and K at set-size 8 or overall K were subjects had a bias to respond “different”, and this tendency was more
lower than those based on all trials: r = −0.37 for K at set-size 8, and evident in the larger set-size.
r = −0.28 for overall K. Critically, however, these results are difficult The shift in criterion is an interesting finding, which can be natu-
to interpret, because the reliability of our measures was quite low, even rally accounted for in a slot-like view of VWM (Zhang & Luck, 2008),
after the Spearman-Brown Correction was applied (r = 0.25 for the 4–8 which posits that only a handful of items are represented with high
drop, r = 0.51 for K at set-size 4, r = 0.53 for K at set-size 8, and resolution, and other items simply remain outside of VWM (it is also
r = 0.64 for overall K). This is expected based on previous reports that possible that these items are represented with extremely low resolution,
using a small number of trials will lead to a low reliability for K esti- as was previously suggested (van den Berg, Shin, Chou, George, & Ma,
mates, whether overall or by set-size, even using the Spearman-Brown 2012), but for conciseness we focus on the simpler zero-resolution
correction (Xu et al., 2018), a finding which remains stable regardless framework and return to the more elaborated alternative in the Dis-
of the number of participants. cussion). If some items are not represented, they should be perceived as

5
H. Balaban, et al. Cognition 191 (2019) 103984

Fig. 5. The correlations between the 4–8 drop (the difference in K between set-size 4 and set-size 8; Y axis) and the three K measures (X axis): (a) K averaged across
both set-sizes, r = −0.61; (b) K at set-size 4 trials, r = 0.01, (c) K at set-size 8 trials, r = −0.84. To better illustrate the distribution, the dots are semi-transparent,
hence the darkness of an area shows the frequency of this combination of values.

different from the current contents of VWM even when they are un- Finally, another possibility is that the observed shift of the criterion
changed, resulting in a greater tendency to respond “different”, i.e., is the result of factors external to the VWM capacity debate, such as
causing a shift of criterion in the observed direction. The slot-view in- attentional load or task difficulty. For example, subjects might utilize a
terpretation of our observed average K is that subjects only hold an different strategy in the larger set-size, causing them to shift the cri-
average of 2–3 items of the 4 or 8 presented colors, while all other items terion. However, this notion does not explain why the criterion was
are completely unrepresented. Therefore, when the test array happens shifted to begin with (in set-size 4), or why the shift when the set-size
to present one of the unrepresented items in its original (i.e., un- increased was in the observed direction. Therefore, more work, perhaps
changed) color, subjective experience should be that this item is dif- involving formal model comparison, is needed to better understand the
ferent from those held in memory, leading to more “different” responses origin of the observed pattern of results. We conclude that the shifted
than there should be, i.e., the observed response bias. A slot view also criterion is predicted by a discrete-slot view of VWM but does not rule
predicts that the bias should be larger when more items are presented, out a flexible-resource view of VWM and might also be the result of
because there are more un-stored items that subjectively resemble new factors orthogonal to the structure of VWM.
items once presented in the test array. To complement this signal detection analysis, we also report d', i.e.,
The alternative view of VWM is of a flexible resource, such that all memory signal strength: d′ was 2.48 for set-size 4, and 1.14 for set-size
items enter VWM, and as more items are maintained, the resolution of 8, which is lower as expected. Notably, the exact amount of decrease in
the representations deteriorates (Bays & Husain, 2008). The greater d′ from set-size 4 to set-size 8 might shed light on the source of this
representational noise that is associated with an increased set-size in decrease. Specifically, according to a simple version of the flexible-re-
this account can result in several possibilities with regards to the re- source view of VWM, all of the presented items are maintained in VWM,
sponse bias. While accuracy should decrease as set-size increases, the because capacity is distributed across all items. Consequently, as set-
predictions for the criterion are less clear. As set-size increases the re- size increases, the resolution of each representation should decrease. In
presentations become noisier, which might lead participants to perceive other words, capacity remains constant, and is divided across all N
unchanged items as different from those in memory, shifting the cri- items. Therefore, on average d′ at set-size 8 should be half of d′ at set-
terion in the observed direction. However, noisier representations size 4, because there are twice as many items that share the same
might necessitate stronger evidence to report a change, exactly because overall capacity. Here, we found that d′ at set-size 4 was 2.48, meaning
the presented color is always different from the perceived one, shifting that according to the flexible resource view d′ at set-size 8 should be
the criterion in a positive direction, the opposite from what was ob- 1.24, which is quite close to the observed d′ of 1.14. In fact, for each
served here. Thus, a continuous-resource account of VWM does not participant d′ at set-size 8 should be half of d′ at set-size 4, and we can
necessarily predict any criterion shift, but it can accommodate it. The test the correlation between the actual d′ at set-size 8 and the d′ pre-
observed direction of the criterion shift suggests that the factors that dicted based on d′ at set-size 4 (note that the predicted d′ at set-size 8 is
cause a negative criterion shift are stronger than those causing a posi- a simple linear transformation of d′ at set-size 4), which we found to be
tive criterion shift, an interesting direction that can be empirically r = 0.47 (95% CI: [0.44, 0.49]). Finding this correlation suggests that a
tested in future studies. simple version of the flexible-resource model gives a good prediction of

6
H. Balaban, et al. Cognition 191 (2019) 103984

Fig. 6. K estimates, for all 3,849 data-sets pooled together, by serial position within the task. Green dots are the 3 highest K trials, with the green line showing their
average. Pink dots are the 3 lowest K trials, with the pink line showing their average. The dashed grey line shows the linear relationship between K and serial
position. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

the decrease in d′. Versions of the flexible-resource model that in- found for the simple flexible-resource prediction of d′.
corporate other factors, such as a variable resolution, could probably Thus, we found that the pattern of d′, a signal detection theory
produce better predictions, and in fact, in the present setting (with only measure of memory strength, was well accounted for in a resource-like
two set-sizes and no additional manipulations), some sort of a flexible- view of VWM, but it followed even more closely the predictions of a
resource view could have predicted any amount of a decrease in d'. slot-like VWM which holds a limited set of items in high resolution,
Interestingly, the alternative view, of VWM as a slot-like storage, while other presented items remain outside of the storage. Moreover, it
makes very strong predictions regarding the exact level of d′ across set- supports the slot model’s assumption that the same memory strength is
sizes as well. According to this view, at each set-size some of the items used for each represented item, regardless of the set-size. Indeed, the
are fully represented (i.e., with a very high resolution), and the rest of average d′ across set-sizes followed what is expected from the capacity
the items remain completely outside of VWM. This means that the estimates of K, and this was true even at the individual level. These
observed d' at each set-size is a mixture of two d' measures: one for the somewhat surprising findings give novel support for the notion of a
fully-remembered items, and a d' of 0 for the items that were not re- discrete, instead of resource-like, capacity limit of VWM. However, it is
presented. Furthermore, the slot model assumes that the d′ for the re- important to note that this analysis relies on mixing K and d′, two
presented items should be identical between set-size 4 and set-size 8, measures that have very different underlying assumptions (high-
due to the all-or-none nature of the slots. threshold versus signal detection). Indeed, the slot model assumes that
To test this prediction, we can extract the individual d' for fully some items will produce maximal d′ and others will have d′ = 0, which
remembered items from the observed d' at set-size 4: the observed d' is different from the usual analysis of d′. Moreover, the current slot-
(2.48) times 4, divided by K at set-size 4, which gives d' = 3.56. This model predictions of d’ at set-size 8 were based on K at set-size 8, which
means that according to the slot model, the observed d′ at set-size 4 involve the same trials. If instead we use K at set-size 4 to predict d’ at
(2.48) is an average of the 2.79 items that were represented in VWM set-size 8, we go back to a correlation of r = 0.47 (95% CI: [0.44, 0.49],
with a d′ of 3.56 and 1.21 items with d′ of 0. Because the slot model exactly as strong as using d’ at set-size 4 to predict d’ at set-size 8. It is
assumes that the individual-item d′ for the represented items should be noteworthy that measures that are not based on the high-threshold
identical in set-size 8, we can calculate the predicted average d' at set- assumptions of K, namely d′ and criterion, follow closely the predictions
size 8, because the slot model predicts that this d′ is an average of K of the slot model, but more work is needed to clarify this point, and
items with d′ = 3.56 and 8-K items with d′ = 0. Thus, according to the using formal model comparison could allow a future definite conclu-
slow model, the average d′ should be 3.56 times 2.54 (K at set-size 8), sion.
divided by 8. This gives a predicted d' at set-size 8 of 1.13, which is Finally, we found a correlation of r = −0.31 (95% CI: [−0.34,
remarkably close to the observed d' at set-size 8, which was 1.14. This −0.29] between overall response bias and d′ (across the different set-
suggests that on average, the pattern of d′ across set-sizes is very well sizes).
predicted by the slot model, which assumes all-or-none fixed re-
presentations.
While these results strongly support the slot model at the average 3.2. Group-level capacity estimates
level, if indeed a slot-like view of VWM can account for the pattern of d′
across set-sizes, this should be true also at the individual level. Thus, an We next turned to examine the dynamics of capacity throughout the
even stronger test for the predictions of the slot model would be ap- span of the experiment. Comparing K values from the different serial
plying the same calculation for each data set separately and examining positions of the trials (1–120) can uncover the influence of either
how well the predicted d′ correlates with the observed d′. Corroborating practice or proactive interference on capacity estimates: practice should
and extending the group-level finding, we found a very strong corre- manifest in larger K values as the experiment progresses, while proac-
lation of r = 0.83 (95% CI: [0.82, 0.84]) between the d' at set-size 8 that tive interference will produce smaller K values in later trials. Notably,
was actually observed and the d' predicted by the slot model. Notably, the ability to overcome proactive interference has been argued to un-
this correlation is not merely a byproduct of a tight relationship be- derly at least some part of individual differences in VWM capacity, e.g.,
tween the observed d' at set-size 4 (which was used to calculate the because of the need to suppress previously active items (Hartshorne,
predicted d' at set-size 8) and the observed d' at set-size 8, because their 2008; Unsworth & Engle, 2007). K estimates for each of the serial po-
correlation was much lower, r = 0.47 (95% CI: [0.44, 0.49]), as was sitions in the task are presented in Fig. 6.
Contrary to the interpretation of capacity as stemming from

7
H. Balaban, et al. Cognition 191 (2019) 103984

proactive interference, we found that K estimates were extremely stable task. Applying the spatial analysis to this subset allowed us to disen-
throughout the experiment, with an SD of only 0.10 between the dif- tangle the previously discussed bias for “different” responses from the
ferent serial positions. As can be seen from Fig. 6, the highest K trials effect of response side, because in the regular version of our task “dif-
and lowest K trials were relatively distributed across the entire ex- ferent” was associated with the right-hand side. For 905 counter-
periment, and the difference between them was not very large (mean of balanced data-sets, accuracy by the type of trial (change vs. no-change)
the 3 best serial positions: 2.89, worst: 2.47). Additionally, we found and the response-mapping (“different” on the right vs. left) is shown in
that K on the first trial of the task, where proactive interference is Fig. 8 (note that this analysis is much noisier than the previous one,
minimal, was already quite low (2.63), and was even slightly lower including only about 6% of the amount of data). The results, presented
than K on the last trial, where proactive interference is maximal (2.74). in Fig. 8, suggest no clear effect of response-mapping, and importantly,
To quantify the contribution of proactive interference to capacity, we both groups have a much higher hit rate than a correct rejection rate.
calculated the Spearman correlation between serial position and K, This suggests that the response bias doesn’t stem from a right-hand
which was negative but relatively weak, r = −0.23 (95% CI: [−0.39, preference, and instead reflects a genuine shift of criterion, in line with
−0.05]). This suggests that while some proactive interference might an item-limit on VWM (see above).
occur, it accounts for only 5% of the variance in capacity, in line with
previous claims that it is fairly negligible in the canonical change de- 3.3. Large-scale analysis of capacity
tection task (Lin & Luck, 2012).
We found no clear evidence of improvement beyond the first ∼5 Our final set of analyses examined the relationship between K es-
trials, suggesting that after some minimal practice took place, capacity timates and several factors, as a characterization of the “population-
is very stable. This improvement is likely due to a slightly smaller bias level” parameters of VWM capacity. We first examined capacity by the
(criterion was −0.55 in the first trial, and −0.54 in the fifth trial), and hour of day and month of year in which the experiment took place, and
after this adjustment of criterion participants did not continue to im- then by demographic parameters: gender, age, and field of study (since
prove, suggesting practice does not play a large role in capacity esti- the data was not collected for the current analysis but for a range of
mate in this short task. different experiments conducted through the course of several years,
As mentioned above, since VWM capacity is limited, a key question the number of data-sets for which a specific measure was available
is what happens in the face of many to-be-encoded items. A simple differed between the reported factors).
flexible resource model should predict that capacity is evenly dis- Our data was collected throughout the year, enabling us to examine
tributed between all items, because they are equally important and all whether capacity is, for example, higher at the beginning of a semester
of them can enter VWM (though in lower resolution, Bays & Husain, (November and March for the winter and spring semester, respectively)
2008). Contrary to this view, most present theories of VWM posit that than at the end of a semester (January and June), during the exams
when facing supra-capacity arrays, participants must prioritize some period (February and July-August), or during summer break
items at the expense of others. The unprioritized items either won’t (September-October). Similarly, we could examine whether capacity
enter VWM at all (Zhang & Luck, 2008), or will enter VWM in lower drops or rises throughout the day. As can be seen in Fig. 9, K estimates
resolution (van den Berg et al., 2012). Our data-set allowed us to ex- were quite stable across the year, and throughout the day, although
amine whether there is systematic prioritization based on items’ loca- some variation exists. The highest K month was January (the end of
tions, reflected in systematic differences in accuracy between the winter semester; mean: 2.75, SD: 0.87), and the lowest K month was
probed locations. As can be seen from Fig. 7, accuracy is not uniformly June (the end of spring semester; mean: 2.51; SD: 0.85), but the range
distributed across space, and instead a handful locations are privileged. was very narrow, and the different months, both during the school year
The two top-central locations produce the best accuracy, and the lo- and during vacations, produced highly similar capacity estimates (SD of
cations around them follow. The top part of the screen is preferred to the different months: 0.07). As for the time of day (for this analysis we
the bottom part, while there is no systematic difference between the left focused on 8am-5 pm, because earlier and later hours had less than 30
and right sides of the screen. The source of the advantage for the two data-sets; this resulted in a total of 3,813 data-sets for this analysis), we
top-central locations could be attributed to different stages involved in found that K was slightly higher in the morning than in the afternoon,
the change detection task: privileged attention during encoding or less although it had a relatively restricted range. The highest K hour was 9
encoding noise, more stable maintenance, or easier retrieval. am (mean: 2.76, SD: 0.82) and the lowest K hour was 5 pm (mean: 2.47,
A subset of participants performed a counterbalanced version of the SD: 0.81), with a very small range (SD of the different hours: 0.07).
As another way to characterize VWM capacity, we examined the
relationship between capacity and demographic factors, starting with
gender (N = 2,079, 1,456 females, 623 males). We found that females’
VWM capacity was slightly higher than that of males (females’ mean:
2.75, SD: 0.81; males’ mean: 2.64, SD: 0.82; BF10 = 2.79). To the best of
our knowledge, this interesting, though delicate, finding has not been
previously reported.
Next, we examined the correlation between capacity and age
(N = 1,833), because it has been shown that VWM deteriorates in old
age (e.g., Jost et al., 2011). However, in our data-set of healthy young
adults, with a restricted range of ages (18–39, mean 23.75), we found
absolutely no relationship between capacity and age (r = 0.01, 95% CI:
[−0.03, 0.06]). This shows another aspect of stability in capacity es-
timates using the canonical change detection task.
Finally, we examined the distribution of capacity across the dif-
ferent fields of study among our student participants. We divided par-
ticipants into 6 groups: arts and humanities (N = 132), education and
therapeutic professions (N = 150), exact sciences and engineering
(N = 227), law and management (N = 150), life sciences and medicine
Fig. 7. A heat map of accuracy, for all 3,849 data-sets pooled together, by (N = 222), and social sciences (including psychology; N = 703). We
spatial location of the probed item. found the lowest capacity among law and management students (mean:

8
H. Balaban, et al. Cognition 191 (2019) 103984

Fig. 8. A heat map of accuracy, for a subset of 905 data-sets with counterbalancing, by spatial location of the probed item, broken down by the type of trial (change
vs. no-change) and the response-mapping. The left two panels depict change trials, with the “different” key on the (a) right-hand side or (b) left-hand side. The two
right panels depict no-change trials, with the “different” key on the (c) left-hand side or (d) right-hand side. Note that different color-scales were used for change and
no-change trials, because of the large difference in accuracy between them.

2.61, SD: 0.89, 95% CI: [2.46, 2.75]) and arts and humanities students 4. Discussion
(mean: 2.62, SD: 0.85, 95% CI: [2.48, 2.77]), medium among education
and therapeutic professions students (mean: 2.69, SD: 0.84, 95% CI: By analyzing over 460,000 trials of the canonical change detection
[2.56, 2.83]) and exact sciences and engineering students (mean: 2.71, task, the present study had two goals: first, exploring the nature of
SD: 0.80, 95% CI: [2.61, 2.82]), and highest among social sciences VWM capacity, and second, better characterizing the change detection
students (mean: 2.75, SD: 0.81, 95% CI: [2.69, 2.81]) and life sciences task itself. Our unique data-set draws a “population-level” picture of
and medicine students (mean: 2.84, SD: 0.86, 95% CI: [2.73, 2.95]). capacity limitations, which we describe below. By making our data
Thus, there are some differences among the different fields of study, but publicly available, we hope other researchers will be able to analyze it
overall K estimates are relatively similar for all fields, at least for uni- in other ways, reaching new interesting conclusions regarding VWM.
versity students. The extremely large sample-size allowed us to uncover the fact that

Fig. 9. Mean K estimates by (a) the month of year, and (b) the time of day in which the experiment took place. Error bars depict standard deviation.

9
H. Balaban, et al. Cognition 191 (2019) 103984

individual K values are naturally normally distributed, which comple- that the additional items were held in VWM with extremely low re-
ments previous reports of tight connections between VWM capacity and solution (van den Berg et al., 2012), if their representations are so
factors such as fluid intelligence (Fukuda et al., 2010). Because capacity corrupted they are unusable even when changes are very large as in the
is easily quantified in a stable and meaningful way using the change present paradigm, it might be sufficient to treat them as effectively
detection task in its present form, we believe that measuring capacity absent from VWM under circumstances similar to the change detection
for each participant should be the standard practice in VWM research task employed here. It may be that in the real world, regardless of
(Luck & Vogel, 2013). We believe this will produce new insights on the whether VWM operates in a slot-like or resource-like manner, observers
sources of capacity limits, as those found in the present investigation. act as if the representations are discrete given specific constraints. In
Several findings support recent claims that capacity does not simply further support for a slot-like view of VWM, such that some items are
reflect the size of storage space (Fukuda et al., 2015; Mance & Vogel, represented with very high resolution while the others are left com-
2013) and that a fundamental attribute of VWM is its active nature. K pletely outside of VWM, we found a decrease in d' (the signal detection
estimates were lower for set-size 4 than for set-size 8, suggesting that theory measure of memory signal strength) with increased set-size in a
not only storage-size but other factors (e.g., selecting which items to manner compatible with the predictions drawn from this discrete ca-
encode) influence VWM processes. We found initial evidence that ca- pacity assumption. Namely, d' at set-size 8 could be well predicted, at
pacity estimates are affected by these factors, as manifested in the both the average and the individual levels, from the observed d' at set-
stronger effect set-size 8 trials have on overall K than set-size 4 trials. size 4 and the K estimates at both set-sizes. This suggests that d′ at each
Furthermore, the 4–8 drop, previously used to asses attentional control set-size N reflects a mixture of K items that have high resolution that is
independently from storage-size, significantly correlated with K, sug- fixed across set-sizes, and N-K items that are not represented at all in
gesting that similar processes dominate the two measures. While the VWM. Notably, this analysis was based on mixing two measures with
findings based on correlations need more direct evidence to be con- very different underlying assumptions: K, which is a high-threshold
sidered convincing, they are in line with the capacity reduction in the measure, and d′, which is a signal detection measure. Future formal
larger set-size, both suggesting that there’s more to capacity than mere model comparisons could help clarify the source of the decrease in d′.
storage-size. We therefore argue that a central cause for individual Importantly, while the pattern of the signal detection measures, i.e.,
differences in VWM capacity is the ability to flexibly and adaptably use the criterion shift and d′, is naturally predicted by a slot-like view of
the VWM storage space, that might be similar in size among individuals. VWM, it can also be accounted for in a flexible-resource view, and
It could be that most individuals can store about 3 simple items, but might even be the result of factors external to the structure of VWM.
low-K individuals store only 1 task-relevant item and 2 completely ir- Thus, more research is needed to explore the criterion and d′ pattern in
relevant items, while high-K individuals efficiently select 3 task-re- the classic change detection task, but we do note that these measures,
levant items, and perhaps even more by chunking several items to- which are not based on high-threshold assumptions, seem to offer initial
gether or by making better use of ensemble representations (Brady & novel support for the predictions of a discrete slot-like structure of
Alvarez, 2011; Haberman, Brady, & Alvarez, 2015). These abilities VWM.
would become critical in complex real-life situations, which include The present capacity estimations agree with numerous findings
vast amounts of meaningful information from which only a small subset (Cowan, 2001) that place the average capacity limit at around 3 simple
should be carefully selected to enter VWM. This likely contributes to the items, with vast individual differences (Luck & Vogel, 2013). Our ex-
tight connections between VWM and attentional control in a range of tremely large sample-size allowed us to uncover two novel character-
situations (Vogel et al., 2005). istics of K estimates at the “population-level”: the normal shape of the
Despite the central role attentional control plays in individual dif- distribution of individual K estimates (see above), and the fact that
ferences of VWM capacity, it is not likely the source of capacity limit females have a slightly higher capacity than males.
itself. Namely, we found that proactive interference, i.e., the ability to Aside from new insights on VWM capacity, our data-set allowed us
prevent previously relevant items from disrupting the maintenance of to better characterize the canonical change detection task itself, high-
currently relevant items, doesn’t explain the low average capacity, lighting important issues that should be kept in mind when using the
unlike previous claims (Unsworth & Engle, 2007). The correlation be- paradigm to investigate capacity. First, it is critical to include set-sizes
tween K and the serial position within the task was quite low (ac- that are well above average capacity limits, as we found that K is
counting for ∼5% of K variability across the task), and K was low al- dominated by estimates from set-size 8 trials, although using only very
ready in the first trial (in fact, slightly lower than in the last trial). Thus, large set-sizes is not recommended because of the expected motiva-
VWM is genuinely limited in capacity, even without proactive inter- tional effects. Second, to remove the effect of practice, the task should
ference (Lin & Luck, 2012). include about 10–12 practice trials to allow participants to adjust their
A central issue in VWM research is how to best describe its capacity criterion (we had ∼6 practice trials and found improvement for an-
limits. One leading approach views capacity as a continuous resource other 5 trials). After this phase, practice has no impact, at least when
that can be shared among an arbitrary number of representations (Ma the task is as short as here. Third, our heat map revealed several lo-
et al., 2014), trading quantity for quality as more items are added or as cations that benefit from greater attentional allocation, specifically the
their complexity increases (Alvarez & Cavanagh, 2004). The contrasting ones closest to fixation on the top of the screen, followed by the loca-
view is that VWM has a discrete set of place-holders that can each tions surrounding them. Performance was comparable at the right and
maintain one item (regardless of its complexity), and once all slots are left side, and better above than below fixation. These preferences might
allocated additional items are left completely outside of VWM (Vogel & arise at one of several possible processing stages, and perhaps at mul-
Machizawa, 2004; Zhang & Luck, 2008). This is admittedly a simplified tiple stages together. The privileged items might have more attention or
description of a complicated debate, and many subtle intermediate less noise at encoding, better maintenance once in VWM, or easier re-
positions have been proposed (Brady et al., 2011). Regardless, the trieval, and these options are not mutually exclusive. The source of the
question of whether it is fruitful to conceptualize VWM as holding a spatial preference we found could be the target of future research.
very small set of items remains an important one. Our results confirmed Finally, counterbalancing the response keys doesn’t seem to be neces-
one prediction of a strict item limit, in the form of a criterion bias to- sary.
wards “different” responses. If some of the items simply do not enter Importantly, all our data and conclusions are relevant to the classic
VWM, it is natural that subjects will treat them as not matching the same-different change detection paradigm. In a continuous-report recall
contents of VWM when they are presented again (unchanged) at test. version of this task (Wilken & Ma, 2004; Zhang & Luck, 2008), one can
Accordingly, the bias was larger for 8 items than for 4, meaning when disentangle the probability that an item enters VWM from the resolu-
more items are presumably left outside of VWM. While it is possible tion with which it is represented. It is not yet known whether this

10
H. Balaban, et al. Cognition 191 (2019) 103984

version correlates with aptitude measure as the classical version does, working memory capacity: Contributions of attentional control to storage.
and if so which of the two factors of capacity drives this correlation Mechanisms of Sensory Working Memory: Attention and Perfomance, XXV, 105.
Fukuda, K., Vogel, E., Mayr, U., & Awh, E. (2010). Quantity, not quality: The relationship
(though evidence from the canonical change detection task suggested between fluid intelligence and working memory capacity. Psychonomic Bulletin &
that it is storage-size and not resolution that correlates with in- Review, 17(5), 673–679. https://doi.org/10.3758/17.5.673.
telligence; Fukuda et al., 2010), which calls for future research. It is Haberman, J., Brady, T. F., & Alvarez, G. (2015). Individual differences in ensemble
perception reveal multiple, independent levels of ensemble representation. Journal of
perfectly plausible that analyzing the recall change detection task will Experimental Psychology: General, 144(2), 432.
reveal support for a more resource-like view of VWM. Notably, K from Hartshorne, J. K. (2008). Visual working memory capacity and proactive interference.
change detection performance is tightly correlated with the probability PLoS One, 3(7), e2716.
Johnson, M. K., McMahon, R. P., Robinson, B. M., Harvey, A. N., Hahn, B., Leonard, C. J.,
of remembering an item derived from continuous report (Zhang & Luck, ... Gold, J. M. (2013). The relationship between working memory capacity and broad
2008), which is another way to quantify the number of items held in measures of cognitive ability in healthy adults and people with schizophrenia.
VWM. This suggests that at least some of the conclusions drawn on K Neuropsychology, 27(2), 220–229. https://doi.org/10.1037/a0032060.
Jost, K., Bryck, R. L., Vogel, E. K., & Mayr, U. (2011). Are old adults just like low working
will be generalized to other methods and ways of measuring capacity.
memory young adults? Filtering efficiency and age differences in visual working
To conclude, the change detection task is an excellent way to memory. Cerebral Cortex, 21(5), 1147–1154. https://doi.org/10.1093/cercor/
measure capacity, being stable not only within an individual, but also bhq185.
across the population: K was highly similar between different months of Lin, P., & Luck, S. J. (2012). Proactive interference does not meaningfully distort visual
working memory capacity estimates in the canonical change detection task. Frontiers
the year, hours of the day, and, in our sample of normal students, even in Psychology, 3, 42.
ages. While it is critical to understand the task’s limitations and not Luck, S. J., & Vogel, E. K. (1997). The capacity of visual working memory for features and
confuse it with the theoretical construct of VWM capacity, within these conjunctions. Nature, 390(6657), 279–281. https://doi.org/10.1038/36846.
Luck, S. J., & Vogel, E. K. (2013). Visual working memory capacity: From psychophysics
limitations (e.g., paying attention to the use of complex items; Awh and neurobiology to individual differences. Trends in Cognitive Sciences, 17(8),
et al., 2007, see the Introduction) the now-classic change detection 391–400. https://doi.org/10.1016/j.tics.2013.06.006.
paradigm indeed deserves its canonical status. Keshvari, S., van den Berg, R., & Ma, W. J. (2013). No evidence for an item limit in change
detection. PLoS Computational Biology, 9(2), e100297.
Ma, W. J., Husain, M., & Bays, P. M. (2014). Changing concepts of working memory.
Declaration of Competing Interest Nature Neuroscience, 17(3), 347–356. https://doi.org/10.1038/nn.3655.
Macmillan, N. A., & Creelman, C. D. (2004). Detection theory: A user's. Guide Psychology
Press.
None. Mance, I., & Vogel, E. K. (2013). Visual working memory. Wiley Interdisciplinary Reviews:
Cognitive Science, 4(2), 179–190.
Acknowledgments Martinussen, R., Hayden, J., Hogg-Johnson, S., & Tannock, R. (2005). A meta-analysis of
working memory impairments in children with attention-deficit/hyperactivity dis-
order. Journal of the American Academy of Child and Adolescent Psychiatry, 44(4),
This work was supported by the Israel Science Foundation (grant 377–384 S0890-8567(09)61489-1 [pii].
number 862/17 awarded to Roy Luria) and the Azrieli Fellowship Parra, M. A., Sala, S. D., Abrahams, S., Logie, R. H., Mendez, L. G., & Lopera, F. (2011).
(awarded to Halely Balaban). We wish to thank all past and present lab Specific deficit of colour-colour short-term memory binding in sporadic and familial
alzheimer's disease. Neuropsychologia, 49(7), 1943–1952. https://doi.org/10.1016/j.
members, especially Ayala Allon, Maya Ankaoua, Hagar Cohen, Britt neuropsychologia.2011.03.022.
Hadar, Shiri Tascesme, and Anna Vaskevich, for data collection. Pashler, H. (1988). Familiarity and visual change detection. Perception & Psychophysics,
44(4), 369–378.
Phillips, W. (1974). On the distinction between sensory storage and short-term visual
References memory. Perception & Psychophysics, 16(2), 283–290.
Spearman, C. (1910). Correlation calculated from faulty data. British Journal of
Allon, A. S., & Luria, R. (2017). Compensation mechanisms that improve distractor fil- Psychology, 3(3), 271–295.
tering are short-lived. Cognition, 164, 74–86. Unsworth, N., & Engle, R. W. (2007). The nature of individual differences in working
Allon, A. S., Vixman, G., & Luria, R. (2018). Gestalt grouping cues can improve filtering memory capacity: Active maintenance in primary memory and controlled search
performance in visual working memory. Psychological Research, 1–17. from secondary memory. Psychological Review, 114(1), 104.
Alvarez, G. A., & Cavanagh, P. (2004). The capacity of visual short-term memory is set Unsworth, N., Fukuda, K., Awh, E., & Vogel, E. K. (2015). Working memory delay activity
both by visual information load and by number of objects. Psychological Science, predicts individual differences in cognitive abilities. Journal of Cognitive Neuroscience,
15(2), 106–111 01502006 [pii]. 27(5), 853–865.
Awh, E., Barton, B., & Vogel, E. K. (2007). Visual working memory represents a fixed Unsworth, N., Fukuda, K., Awh, E., & Vogel, E. K. (2014). Working memory and fluid
number of items regardless of complexity. Psychological Science, 18(7), 622–628 intelligence: Capacity, attention control, and secondary memory retrieval. Cognitive
PSCI1949 [pii]. Psychology, 71, 1–26. https://doi.org/10.1016/j.cogpsych.2014.01.003.
Bays, P. M., & Husain, M. (2008). Dynamic shifts of limited working memory resources in van den Berg, R., Shin, H., Chou, W. C., George, R., & Ma, W. J. (2012). Variability in
human vision. Science, 321(5890), 851–854. https://doi.org/10.1126/science. encoding precision accounts for visual short-term memory limitations. Proceedings of
1158023. the National Academy of Sciences of the United States of America, 109(22), 8780–8785.
Brady, T. F., & Alvarez, G. A. (2011). Hierarchical encoding in visual working memory: https://doi.org/10.1073/pnas.1117465109.
Ensemble statistics bias memory for individual items. Psychological Science, 22(3), Vaskevich, A., & Luria, R. (2018). Adding statistical regularity results in a global slow-
384–392. down in visual search. Cognition, 174, 19–27.
Brady, T. F., Konkle, T., & Alvarez, G. A. (2011). A review of visual memory capacity: Vogel, E. K., & Machizawa, M. G. (2004). Neural activity predicts individual differences in
Beyond individual items and toward structured representations. Journal of Vision, visual working memory capacity. Nature, 428(6984), 748–751. https://doi.org/10.
11(5), 4. https://doi.org/10.1167/11.5.4. 1038/nature02447.
Brown, W. (1910). Some experimental results in the correlation of mental abilities 1. Vogel, E. K., McCollough, A. W., & Machizawa, M. G. (2005). Neural measures reveal
British Journal of Psychology, 3(3), 296–322. individual differences in controlling access to working memory. Nature, 438(7067),
Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of 500–503 nature04171 [pii].
mental storage capacity. The Behavioral and Brain Sciences, 24(1), 87–114 discussion Vogel, E. K., Woodman, G. F., & Luck, S. J. (2001). Storage of features, conjunctions and
114–85. objects in visual working memory. Journal of Experimental Psychology: Human
Cowan, N., Elliott, E. M., Scott Saults, J., Morey, C. C., Mattox, S., Hismjatullina, A., & Perception and Performance, 27(1), 92–114.
Conway, A. R. (2005). On the capacity of attention: Its estimation and its role in Wilken, P., & Ma, W. J. (2004). A detection theory account of change detection. Journal of
working memory and cognitive aptitudes. Cognitive Psychology, 51(1), 42–100 S0010- Vision, 4(12) 11 11.
0285(05)00002-2 [pii]. Xu, Z., Adam, K., Fang, X., & Vogel, E. (2018). The reliability and stability of visual
D’agostino, R. B., Belanger, A., & D’agostino, R. B., Jr (1990). A suggestion for using working memory capacity. Behavior Research Methods, 50(2), 576–588.
powerful and informative tests of normality. The American Statistician, 44(4), Zhang, W., & Luck, S. J. (2008). Discrete fixed-resolution representations in visual
316–321. working memory. Nature, 453(7192), 233–235. https://doi.org/10.1038/
Fukuda, K., Woodman, G. F., & Vogel, E. K. (2015). Individual differences in visual nature06860.

11