fpsyg-10-01103 May 22, 2019 Time: 17:2 # 1
ORIGINAL RESEARCH
published: 24 May 2019
doi: 10.3389/fpsyg.2019.01103
Edited by:
Motonori Yamaguchi,
Edge Hill University, United Kingdom
Reviewed by:
Sophie Portrat,
Université Grenoble Alpes, France
Alexandre Schaefer,
Monash University, Malaysia
*Correspondence:
Joni Holmes
joni.holmes@mrc-cbu.cam.ac.uk
Specialty section:
This article was submitted to
Cognition,
a section of the journal
Frontiers in Psychology
Received: 13 November 2018
Accepted: 26 April 2019
Published: 24 May 2019
Citation:
Holmes J, Woolgar F,
Hampshire A and Gathercole SE
(2019) Are Working Memory Training
Effects Paradigm-Specific?
Front. Psychol. 10:1103.
doi: 10.3389/fpsyg.2019.01103
Are Working Memory Training Effects
Paradigm-Specific?
Joni Holmes
1
*
, Francesca Woolgar
1
, Adam Hampshire
2
and Susan E. Gathercole
1,3
1
MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, United Kingdom,
2
Department of Medicine,
Imperial College London, London, United Kingdom,
3
Developmental Psychiatry, University of Cambridge, Cambridge,
United Kingdom
A randomized controlled trial compared complex span and n-back training regimes
to investigate the generality of training benefits across materials and paradigms. The
memory items and training intensities were equated across programs, providing the
first like-with-like comparison of transfer in these two widely used training paradigms.
The stimuli in transfer tests of verbal and visuo-spatial n-back and complex span
differed from the trained tasks, but were matched across the untrained paradigms.
Participants were randomly assigned to one of three training groups: complex span
training, n-back training, or no training. Pre- to- post training changes were observed
for untrained n-back tasks following n-back training. Following complex span training
there was equivocal evidence for improvements on a verbal complex span task, but
no evidence for changes on an untrained visuo-spatial complex span activity. Relative
to a no intervention group, the evidence supported no change on an untrained verbal
complex span task following either n-back or complex span training. Equivocal evidence
was found for improvements on visuo-spatial complex span and verbal and visuo-spatial
n-back tasks following both training regimes. Evidence for selective transfer (comparing
the two active training groups) was only found for an untrained visuo-spatial n-back
task following n-back training. There was no evidence for cross-paradigm transfer.
Thus transfer is constrained by working memory paradigm and the nature of individual
processes executed within complex span tasks. However, within-paradigm transfer can
occur when the change is limited to stimulus category, at least for n-back.
Keywords: working memory, cognitive training, transfer, intervention, memory
INTRODUCTION
Research investigating transfer within and across working memory paradigms following training
has largely concluded that transfer is confined to untrained working memory tests that are highly
similar to the trained activities (e.g., Harrison et al., 2013; Sprenger et al., 2013; von Bastian et al.,
2013; Minear et al., 2016; Blacker et al., 2017; Soveri et al., 2017). This has led to speculation that
training induces changes in the processes and strategies tied to particular training experiences
rather than fundamental improvements in the more general capacity of the working memory
system (see Dahlin et al., 2008; Sprenger et al., 2013; von Bastian and Oberauer, 2013, 2014;
Dunning and Holmes, 2014). Progress in understanding the boundary conditions for transfer and
the nature of the cognitive training changes induced by training has been hampered by the absence
of robust theories of transfer and, consequently, a lack of hypothesis-driven research. Experimental
Frontiers in Psychology | www.frontiersin.org 1 May 2019 | Volume 10 | Article 1103
fpsyg-10-01103 May 22, 2019 Time: 17:2 # 2
Holmes et al. Working Memory Training and Transfer
studies manipulating the overlap in key task features between
training and transfer measures are needed to identify the precise
conditions under which transfer does and does not occur (e.g.,
von Bastian et al., 2013; Minear et al., 2016). The current study
addressed this by examining the degree to which near transfer
across working memory tasks is limited by two properties of both
the trained and untrained tasks: the stimulus category and the
working memory paradigm.
Several process- and task-specific theories of transfer have
been advanced. One broad view is that the efficiency of particular
working memory activities may be enhanced through training.
For example, Dahlin et al. (2008) proposed that training
that both involves continuous updating of the contents of
working memory (in running span) and activates the striatum
enhances performance in n-back, another working memory
paradigm involving updating. They concluded that “transfer
can occur if the criterion and transfer tasks engage specific
overlapping processing components and brain regions” (p. 1510).
An alternative position advanced by Gathercole et al. (2019) is
that training on complex working memory tasks such as these
leads to the development of highly specific skills and not to the
expansion of existing capacities. By this account, new cognitive
routines coordinating the execution of the individual processes
are developed for unfamiliar tasks that are not readily served
by existing mechanisms. These routines can only be successfully
applied, and hence generate transfer, to untrained tasks with
similar higher-level cognitive structures that require the same
flow of information across the task. Thus although running
span and n-back both involve updating the contents of working
memory, their other distinctive task requirements (serial recall
for running span as opposed to single item recognition for
n-back) would predict weak or absent transfer across the tasks
due to their distinct task structures.
A further possibility is that training may lead to the
development of explicit mnemonic strategies such as learning to
group individual memory items into larger chunks, or recoding
items into alternative and more distinctive forms (von Bastian
and Oberauer, 2013; Dunning and Holmes, 2014; Minear et al.,
2016). Finally, transfer may not have a single origin but may
instead be the product of multiple training-induced changes
in processes specific to individual task features. Examples of
these features are the stimulus category, the response modality,
and the timing parameters of the task, as well as the broader
paradigms in which they are embedded (Sprenger et al., 2013;
von Bastian and Oberauer, 2013; Minear et al., 2016).
The current study examined the limits to transfer in two
commonly used training regimes, complex span and n-back.
Transfer to untrained n-back tasks following n-back training is
common (e.g., Soveri et al., 2017). This holds true when the
stimulus items are either from both the same (e.g., letters to
digits) or different domain (e.g., spatial locations to digits) to
the trained items (Li et al., 2008; Jaeggi et al., 2010; Anguera
et al., 2012; Bürki et al., 2014; Waris et al., 2015; Küper and
Karbach, 2016; Minear et al., 2016; Blacker et al., 2017). All
complex span training studies to date have changed both the
storage items and the nature of the interpolated distraction (e.g.,
symmetry judgments vs. spatial orientation decisions, or sentence
judgments vs. mathematical operations) between trained and
untrained tasks. Significant transfer has been reported across the
majority of these studies (Harrison et al., 2013; Henry et al.,
2014; Richmond et al., 2014). However, transfer across stimulus
categories is not always consistent. Minear et al. (2016) reported
mixed patterns of transfer within and across domains in different
conditions. Following training on a verbal complex span task
transfer was found for operation and rotation span, but not for
symmetry or alignment span. Blacker et al. (2017) reported no
transfer between two verbal variants of complex span.
The evidence to date largely indicates little transfer across
n-back and complex span paradigms. Although a few studies
report positive transfer following n-back training to a variety of
complex span tasks (Anguera et al., 2012; Sprenger et al., 2013;
Minear et al., 2016; Schwarb et al., 2016), the majority have failed
to observe transfer (Jaeggi et al., 2008, 2010; Li et al., 2008; Chooi
and Thompson, 2012; Lilienthal et al., 2013; Redick et al., 2013;
Sprenger et al., 2013; Thompson et al., 2013; Bürki et al., 2014;
Minear et al., 2016; Schwarb et al., 2016; Blacker et al., 2017). Only
three studies have tested transfer from complex span training to
n-back. Minear et al. (2016) reported transfer to an object n-back
task but not to a letter n-back task following training on a verbal
complex span task. von Bastian et al. (2013) failed to find transfer
to a letter n-back task following training on a verbal complex span
task. Blacker et al. (2017) found that training on a symmetry span
task did not improve performance on an object n-back task.
The primary purpose of the present study was to test whether
transfer occurs within a particular paradigm when the stimuli
between training and transfer (for both paradigms) and the
interpolated distractor activity (for complex span) differ. The
study is the first to compare the two training approaches while
matching potentially key features for transfer across paradigms:
the stimuli in the transfer tests within and across paradigm, and
the duration of training of the two training regimes. Holding
these features constant provides a direct test of whether paradigm
is the critical factor constraining transfer.
It was predicted that transfer following both complex span
and n-back training will be limited to new tasks employing the
same working memory paradigm, consistent with the weight of
evidence reviewed above. Transfer is therefore not expected to
cross the two paradigms, in line with process- and paradigm-
specific theories of transfer (Dahlin et al., 2008; Gathercole et al.,
2019). Minear et al. (2016) is one of two studies to compare
directly these two different forms of training. They used spatial
n-back and verbal complex span training tasks, and found no
evidence for cross-paradigm transfer. However, as both the
stimulus category and domain of the memory items differed
between the training regimes it is impossible to determine from
this study whether transfer was restricted by working memory
paradigm or differences in the training stimuli. A second study
comparing dual n-back training with locations and letters to
a symmetry span training regime, which also failed to find
cross-paradigm transfer, is limited by the same confounds; one
training regime targeted only visuo-spatial abilities while the
other included both verbal and visuo-spatial materials (Blacker
et al., 2017). By using the same stimuli in the two training
paradigms and transfer tests in the present study it was possible
Frontiers in Psychology | www.frontiersin.org 2 May 2019 | Volume 10 | Article 1103
fpsyg-10-01103 May 22, 2019 Time: 17:2 # 3
Holmes et al. Working Memory Training and Transfer
to test the extent to which it is training paradigm or stimulus
material (or both) that is the boundary condition to transfer.
Transfer across n-back tasks that differ only in the stimulus
category within a common domain was expected on the basis of
previous findings (e.g., Jaeggi et al., 2010; Küper and Karbach,
2016; Minear et al., 2016; Soveri et al., 2017). A final question
addressed in this study was whether transfer occurs across
complex span tasks with different storage items and distractor
activities. We tested this by employing trained and untrained
complex span tasks with distinct distractor activities (lexical
decision and rhyme judgment; symmetry judgment and mental
rotation) and memory items (digits and letters; spatial locations
and visual objects). Previous evidence has pointed to transfer
across tasks sharing only the paradigm and not either the
stimuli or distractor activity (Harrison et al., 2013; von Bastian
et al., 2013; Henry et al., 2014; Richmond et al., 2014; Minear
et al., 2016). In its present form the cognitive routine theory
(Gathercole et al., 2019) is not sufficiently well specified to
generate strong predictions about the levels of task structure
in a complex span routine that can and cannot be adapted to
an untrained task and hence to generate transfer. If only the
higher-level task structure of alternating stimulus presentation
and distractor processing episodes needs to be preserved, there
should be transfer to other complex span tasks with different
stimuli and distractor processing activities. Alternatively, if the
specific subroutines guiding the distractor activities must match,
transfer across different complex span tasks should not occur. In
this way, the study will provide key information regarding the
boundary conditions to transfer to inform new theory.
The trained and untrained tasks are summarized in
Table 1, with images of the tasks shown in Figure 1.
Bayesian inference was used to evaluate the strength of the
evidence of training and transfer effects. Bayes factors (BF)
quantify the evidence for both null hypotheses (the absence
of training and transfer) and alternative hypotheses (the
presence of training and transfer). BFs are increasingly popular
in cognitive training research as a means of quantifying
positive evidence for the null hypothesis of no transfer (e.g.,
Sprenger et al., 2013; De Simoni and von Bastian, 2018).
MATERIALS AND METHODS
Participants
Target recruitment was 16 participants per group (total N = 48),
yielding power of 0.86 to detect a large effect size, f
2
= 0.35 or
Cohen’s d = 0.8, with linear regression. Data were excluded for
eight participants who failed to complete all training sessions
and four participants who did not attend for the Time 2
assessment. The training program failed for one participant.
Further participants were recruited to replace those with
incomplete data.
The final sample included 48 native-English speaking adults
aged between 18 and 35 years (12 males, mean age 28 years
and 9 months). Participants were recruited through the MRC
Cognition and Brain Sciences Unit research participation system
and provided written informed consent prior to participation.
Ethical approval was obtained from the University of Cambridge’s
Psychology Research Ethics Committee (PRE.2012.86).
Materials
Transfer Tasks
Complex span
Participants completed two complex span tasks, one verbal, and
one visuo-spatial. For both tasks participants were presented
with a series of storage items interpolated with a same-
domain processing task performed for a 6000 ms period
intervening between the presentation of successive memory
items. Participants were required to recall the storage items in
serial order at the end of the trial. Two practice trials were
presented at a list length of one item. Test trials were presented
in blocks of 3. The first block started at a list length of one (a
single memory item followed by a processing episode prior to
recall) and increased by one item (additional storage item and an
additional processing episode) if two or more trials were correct
in any block. Trials were scored as correct if all storage items were
recalled in the correct serial order and >66% of the processing
items were correct. The tasks discontinued if two of the three
trials in a block were incorrect. A trial was incorrect if the storage
items were recalled incorrectly, accuracy for the processing tasks
was <66%, or if there were no responses for the processing tasks.
The maximum span reached was scored. This was counted as the
span level at which the task discontinued.
The auditory stimulus items in the verbal complex span task
were the digits 1 through to 9. Digital recordings of each item
spoken in a female voice were p-centered – a process of aligning
the waveforms of the recordings so the digits sound regular
(Morton et al., 1976). Individual lists were compiled by sampling
the digits drawn randomly without replacement. Presentation
rate was 1000 ms; each spoken item had a duration of 750 ms and
was followed by an ISI of 250 ms, which was silence with a blank
screen. The duration of the interpolated processing task was
6000 ms, starting after the presentation of the first list item. The
processing activity required participants to judge whether pairs of
spoken letters rhymed. The letter pairs consisted of monosyllabic
English alphabet letter names. Pairs were constrained to avoid
successive letters in the alphabet (e.g., J,K), highly confusable
fricative letter names (e.g., F, S), and familiar acronyms (e.g.,
PC, IT, GB). Each letter sound was presented within a 800 ms
window, followed by a 200 ms window of silence before the onset
of the second letter sound. The task was participant-paced with
an inter-stimulus interval (ISI) of 200 ms between a response
and the onset of the next letter pair. New letter pairs were
not presented if there was <500 ms remaining of the 6000 ms
window. Participants were able to respond at the onset of the
second letter in a pair by clicking on an on-screen “rhyme” or
“non-rhyme” button.
For the visuo-spatial complex span task, participants were
required to remember a series of static, visually presented abstract
line figures for serial recall. Each stimulus was presented on
screen for 1000 ms (750 ms followed by a 250 ms ISI). The
stimuli set consisted of 9 line figures presented in random order.
A dynamic visual processing task was interpolated between each
Frontiers in Psychology | www.frontiersin.org 3 May 2019 | Volume 10 | Article 1103
fpsyg-10-01103 May 22, 2019 Time: 17:2 # 4
Holmes et al. Working Memory Training and Transfer
TABLE 1 | Trained and untrained tasks.
Category of stimulus Interpolated distractor
materials activity
Paradigm Domain Training Transfer Training Transfer
n-back Verbal Letters Digits
Visuo-spatial Spatial locations Visual objects
Complex span Verbal Letters Digits Lexical decision Rhyme judgment
Visuo-spatial Spatial locations Visual objects Symmetry judgment Mental rotation
FIGURE 1 | Training and transfer tasks.
line figure. This required participants to mentally rotate two
identically shaped polygons to an upright position to decide
whether they were pointing in the same or opposite (mirror
image) direction. The polygon on the left was always presented
in an upright position. The one on the right could appear at
one of 7 rotated positions (45
◦
, 90
◦
, 135
◦
, 180
◦
, 225
◦
, 270
◦
, and
315
◦
). Pairs of polygons were presented in a random order from
a choice of 5 shapes (total stimuli set size of 70). The shapes
remained on screen until a response was made. An ISI of 200 ms
followed a response before the onset of the next polygon pair.
New pairs were not presented if there was <500 ms remaining
of the 6000 ms window. Participants could respond by clicking
an on screen “same” or “mirror” button as soon as each pair of
shapes was presented.
N-back
Two variants of the n-back task were used – one verbal and one
visuo-spatial. For both tasks, participants were presented with
a series of stimuli one at a time. The stimuli presented were
identical to the storage items used in the complex span transfer
tasks (auditory p-centered digits for the verbal n-back task, and
static visually presented abstract line figures for the visuo-spatial
task). Participants were to judge whether each stimuli matched
the one presented n items previously in the sequence by pressing
the down arrow on the keyboard if it was a match (target) and
by not responding if it was not a match (non-target). The ISI
was a blank screen lasting 2500 ms. Participants could respond as
soon as the stimulus had been presented. Stimuli were presented
in blocks of 20+n trials, with 6 targets and 14+n non-targets
presented randomly in each block. Errors included misses (failing
to respond to a target) and false alarms (responding to a non-
target). Both tasks started with two blocks of practice trials at
1-back. Test trials began with a block at 1-back. If there were
fewer than five errors (sum of misses and false alarms) the level
of n increased by one in the next block. If there were five or more
errors the tasks discontinued and the maximum n-level reached
to this point was scored.
Training Tasks
Complex span
Participants in the complex span training group completed 32
trials per training session, 16 trials each on the verbal and visuo-
spatial tasks. The training tasks had the same task structure as
the transfer tasks with storage items interpolated with a same-
domain processing task. The presentation rate of the storage
items and overall length of each processing episode (6000 ms)
was identical to the transfer tasks. To account for learning of
the processing items across training sessions, the presentation
rate of the stimuli within each processing window was titrated
to individual performance for the training tasks. To determine
these presentation rates, participants completed 2 min of the
processing task before each training session. Average RTs for
correct items were calculated and 50% was added to calculate the
rate of presentation during training.
Training started with a list length of one (one storage item,
one processing episode) in the first session. Task difficulty was
adapted across the 20 training sessions based on individual
performance. There was an increase of 1 storage item and
corresponding processing episode following 2 consecutive
correct trials (a correct trial being the storage item(s) recalled in
serial order and accuracy >75% across the processing episodes).
Task difficulty decreased by 1 storage item and processing
episode if there were 2 consecutive incorrect trials (storage
items recalled incorrectly and processing performance <75%).
Otherwise, difficulty remained the same. Each new training
Frontiers in Psychology | www.frontiersin.org 4 May 2019 | Volume 10 | Article 1103
fpsyg-10-01103 May 22, 2019 Time: 17:2 # 5
Holmes et al. Working Memory Training and Transfer
session started at one span level lower than the level reached at
the end of the previous session. The average span level reached in
each training session was recorded.
The storage items for the verbal complex span training task
were consonants presented visually on-screen. The interpolated
processing task was a lexical decision task. Words and non-words
were presented one at a time visually on-screen. Participants were
required to judge whether each stimulus was a real word by
clicking on “word” or “non-word” buttons. They could respond
as soon as each stimulus was presented. The word stimuli were
drawn at random from a pool of 187 items generated by searching
the MRC Psycholinguistic database (Coltheart, 1981) for single
syllable words of 4–6 letters with Kucera-Francis written word
frequency >50 per million (Ku
ˇ
cera and Francis, 1967). An
equivalent numbers of non-words were constructed by blending
words from the real word stimuli set (e.g., swapping the initial
consonant or consonant cluster) to form non-words with a
similar phonotactic composition.
For the visuo-spatial complex span training task, participants
were required to recall a series of spatial locations presented
dynamically in a 4 × 4 grid (akin to a Corsi task). The
presentation of each storage item was interleaved with a
symmetry decision task that involved judging whether patterns
presented on screen were symmetrical. Participants could
respond as soon as each pattern was presented by clicking on-
screen buttons labeled “symmetrical” and “asymmetrical.”
N-back
Participants completed 10 blocks of verbal n-back and 10 blocks
of visuo-spatial n-back in each training session, totalling 20
blocks per session. The n-back training tasks differed from the
transfer tasks only in the stimuli – all other features including the
presentation rate of the stimuli were the same as in the transfer
tasks. For the verbal n-back training, consonants were presented
visually on-screen. For the visuo-spatial n-back training, spatial
locations were presented sequentially one at a time as highlighted
green squares in a 4 × 4 grid. Training started with a block of
trials at 1-back and adapted up and down to match participants’
current performance throughout each training session. If there
were fewer than three errors (sum of misses and false alarms) in
block, the level of n-back increased by one in the next block. If
there were five or more errors in a block the level of n decreased
by one in the subsequent block. In all other cases the level of n
remained the same. Each new training session started at n-1 from
the end of the previous training session. The average level of n
reached in each session was scored.
Analysis Plan
Bayesian ANOVAs were conducted to analyze on-task training
gains across the four training tasks with session (2–20) and
training task (four training tasks) entered as factors. To
investigate whether transfer effects differed by group, Bayesian
general linear regression models (GLMs) were performed
separately for each transfer task. As there were three groups, three
dummy variables were used to run two regression models for each
outcome variable. In the first, dummy variables one (D1) and two
(D2) were entered to compare the effects of n-back training to
complex span training (D1 result) and to compare complex span
training to no intervention (D2 result). To compare the effects of
n-back training to the no intervention group, a second regression
model was run for each measure in which D2 and dummy
variable three (D3) were entered (D2 result). The outcomes of
these models also replicated the comparison between the n-back
and complex span training groups (D3 result) obtained in the first
regression models.
Three questions were addressed in the analyses. The first is
whether the outcome measures improve after training for each
group. The second is whether there is evidence for transfer to
each outcome measure for the n-back and complex span training
groups relative to the no intervention group. The final question is
whether there is evidence for each training condition relative to
the other paradigm (complex span or n-back). This is the most
stringent comparison that allows us to examine the paradigm-
specificity of transfer.
All results are reported as BF. Bayesian methods were
conducted in JASP (Love et al., 2015) with default prior scales.
Inverse BF
10
are used to express the odds in favor of the
alternative hypothesis compared to the null (Jeffreys, 1961).
Values lower than 0.33 provide evidence for the null hypothesis
(no training or transfer), values 0.33–3 provide equivocal
evidence for both hypotheses, and those higher than 3 provide
evidence in favor of the alternative hypothesis (training and
transfer effects).
Procedure
Participants were pseudo-randomly assigned to complex span
training, n-back training or no training at recruitment (n = 16 per
group). Stratified randomization was used to ensure the groups
were matched at baseline in terms of age and gender: Complex
span M
age
= 31.61 (7.27), males n = 4; n-back M
age
= 26.97 (6.08),
males n = 3; no training M
age
= 27.75 (7.40), males n = 5.
All participants completed four tasks during a Time 1
assessment at the MRC Cognition and Brain Sciences Unit
lasting approximately 1 h. Participants assigned to either of the
active training groups also completed their first training session
during this visit, adding another 30 min to the session. Trainees
completed the subsequent 18 online training sessions remotely
(e.g., at home) before returning to the MRC Cognition and
Brain Sciences Unit to complete their final training session and
a Time 2 assessment that was identical to the Time 1 assessment.
Participants in the no training group were contacted and asked
to return to the MRC Cognition and Brain Sciences Unit for
their Time 2 assessment at intervals equivalent to those taken
for the trainees. There was substantial evidence for a null effect
for differences in the intervals between Time 1 and Time 2
between groups, BF
10
= 0.185: Complex span training, M = 60.31,
SD = 26.95 days; n-back training, M = 62.25, SD = 29.39 days;
no training, M = 66.50, SD = 22.98 days. Participants were paid
£6 per hour for their time plus a contribution toward travel
costs. All tasks were hosted by the online platform Cambridge
Brain Sciences
1
.
1
www.cambridgebrainsciences.com
Frontiers in Psychology | www.frontiersin.org 5 May 2019 | Volume 10 | Article 1103
![FIGURE 2 | Average difficulty level attained on each session was converted to standard deviation units relative to session 1 [e.g., (session 2 – session 1)/SD session 1; (session 3 – session 1)/session 1 SD]. Note that Averages are based on the maximum number of data points available for that session.](/figures/figure-2-average-difficulty-level-attained-on-each-session-2c42ne0x.webp)