ArticlePDF Available

How chunks, long-term working memory and templates offer a cognitive explanation for neuroimaging data on expertise acquisition: A two-stage framework

How chunks, long-term working memory and templates offer a cognitive
explanation for neuroimaging data on expertise acquisition: A two-stage framework
Alessandro Guida
, Fernand Gobet
, Hubert Tardieu
, Serge Nicolas
Institut de Psychologie, Université Paris Descartes, Boulogne Billancourt, France
CRPCC, Université Rennes 2, Rennes, France
Centre for Cognition and Neuroimaging, Brunel University, Uxbridge, United Kingdom
Institut de Psychologie, Université Paris Descartes, EPHE, Boulogne Billancourt, France
article info
Article history:
Accepted 19 January 2012
Long-term working memory
Template theory
Brain functional reorganization
Working memory
Our review of research on PET and fMRI neuroimaging of experts and expertise acquisition reveals two
apparently discordant patterns in working-memory-related tasks. When experts are involved, studies
show activations in brain regions typically activated during long-term memory tasks that are not
observed with novices, a result that is compatible with functional brain reorganization. By contrast, when
involving novices and training programs, studies show a decrease in brain regions typically activated dur-
ing working memory tasks, with no functional reorganization. We suggest that the latter result is a con-
sequence of practice periods that do not allow important structures to be completely acquired:
knowledge structures (i.e., Ericsson and Kintsch’s retrieval structures; Gobet and Simon’s templates)
and in a lesser way, chunks. These structures allow individuals to improve performance on working-
memory tasks, by enabling them to use part of long-term memory as working memory, causing a cerebral
functional reorganization. Our hypothesis is that the two brain activation patterns observed in the liter-
ature are not discordant, but involve the same process of expertise acquisition in two stages: from
decreased activation to brain functional reorganization. The dynamic of these two physiological stages
depend on the two above-mentioned psychological constructs: chunks and knowledge structures.
Ó2012 Elsevier Inc. All rights reserved.
1. Introduction
One of the fundamental questions about working memory
(WM) concerns its limit. Miller (1956) famously proposed that
the amount of information that can be kept in mind at one time
is about seven chunks or meaningful units of information. More re-
cently, Ericsson and Kintsch (1995), with their long-term working
memory theory (LT-WMT), and Gobet and Simon (1996a), with
their template theory (TT), have proposed that, in the case of
expertise, part of long-term memory (LTM) can be used during
WM tasks in order to circumvent the limit imposed by the magical
number 7 (Miller, 1956); this would explain, for example, the per-
formance of experts who are able to recall more than 100 digits
(e.g., Chase & Ericsson, 1981). Recent developments in brain imag-
ing via functional magnetic resonance imaging (fMRI) and positron
emission tomography (PET) have provided partial support for this
idea. Two patterns of results can be observed during WM-related
tasks. When the experiments involve expert individuals, these do
indeed tend to show an activation of brain regions typically acti-
vated during LTM tasks (LTM areas hereafter). This pattern of re-
sults could correspond to a cerebral functional reorganization
related to the acquisition of expertise, where functional reorgani-
zation is seen as the recruitment of new activation areas and a shift
in the cognitive process underlying task performances (Poldrack,
2000), which in this case means that resources previously (before
expertise acquisition) allocated to WM are later (after expertise
acquisition) allocated to LTM.
By contrast, when the experiments involve novices who under-
go extended practice with WM-related tasks, the results tend to
show a decrease in activation of brain regions typically activated
during WM tasks (WM areas hereafter), with no cerebral functional
reorganization. The latter result could be a consequence of practice
periods that do not allow important cognitive structures to be
completely acquired, mainly knowledge structures
(i.e., retrieval
structures, Ericsson & Kintsch, 1995; templates, Gobet & Simon,
1996a) and in a less important way chunks (Chase & Simon, 1973a).
In this paper, after presenting chunking theory and two
more recent theories of expertise (LT-WMT and TT), we will first
0278-2626/$ - see front matter Ó2012 Elsevier Inc. All rights reserved.
Corresponding author at: Centre de Recherche en Psychologie, Cognition et
Communication, Université Rennes 2 – Haute Bretagne, Place du Recteur Henri Le
Moal, CS 24 307, Bâtiment S, 35 043 Rennes Cedex, France.
E-mail addresses:,alessandro.guida.psychology@ (A. Guida).
We use ‘‘knowledge structures’’ as an umbrella term for retrieval structures
(Ericsson & Kintsch, 1995) and templates (Gobet & Simon, 1996a).
Brain and Cognition 79 (2012) 221–244
Contents lists available at SciVerse ScienceDirect
Brain and Cognition
journal homepage:
establish using the Bayes Factor that it is possible to separate epi-
sodic long-term memory activation from working memory activa-
tion and then review the studies that report the two different brain
activation patterns mentioned above. We will then propose an
explanation that makes these results more coherent by suggesting
that the two brain patterns are two stages of the same process
occurring during expertise acquisition. We will conclude by dis-
cussing the psychological constructs – chunks and knowledge
structures – that underpin expertise in this two-stage view, linking
them with physiological processes.
2. Chunking theory
One important element for understanding WM limits is chunk-
ing. This mechanism is important not only in understanding stan-
dard cognitive performance, but also in explaining the differences
between novices and experts. The chunking mechanism was ini-
tially described by de Groot (1946/1978) and Miller (1956), and
then theorized by Chase and Simon (1973a). A current definition
is given by Gobet et al. (2001, p. 236): a chunk refers to ‘‘...a col-
lection of elements having strong associations with one another,
but weak associations with elements within other chunks.’’ To ex-
plain chunks, Miller (1956) and subsequently Cowan (2001) used
the same example. When the letters ‘‘fbiibm’’ are presented, if
one knows the acronyms ‘‘FBI’’ and ‘‘IBM,’’ then it is possible to
simplify the information by forming two chunks
(‘‘FBI’’ and
‘‘IBM’’) in WM. Because these familiar patterns exist in LTM – FBI
is the Federal Bureau of Investigation and IBM is a well-known com-
puter company – the letters ‘‘f,’’ ‘‘b,’’ ‘‘i,’’ ‘‘i,’’ ‘‘b’’ and ‘‘m’’ can be en-
coded as two elements in WM instead of six.
Chunking theory provides an explanation for the superiority of
experts over novices. For example, in the field of chess, chunking
has been used to explain how chess experts are able to recall more
chess pieces on a board than novices (Chase & Simon, 1973b; de
Groot, 1946/1978). Thanks to their greater knowledge of chess
positions – in terms of chunks in LTM – the experts are able to
encode the presented chess positions in fewer chunks in WM,
thereby gaining storage space in WM. For example, a chess master
can encode 15 pieces presented on a board as one chunk (Gobet &
Clarkson, 2004). This process is statistically less likely for novice
players since they have less knowledge of chess positions and
therefore possess fewer chunks in LTM.
Chase and Simon (1973a) confirmed de Groot’s (1946/1978) re-
sults, but also found that when the chess pieces to be recalled were
placed in random locations, the superiority of experts over novices
disappeared. They argued that experts could not use their LTM
chunks to encode the random positions, since these positions did
not contain any of the chunks the experts had in LTM. Thus, the
positions were mostly as new to them as to the novices. The ex-
perts’ advantage was significant only when they could actually
use their knowledge. With their MAPP (Memory Aided Pattern Per-
ceiver) computer simulation, Simon and Gilmartin (1973) esti-
mated that the number of chunks in LTM to reach a master level
ranged from 10,000 to 100,000.
2.1. Weaknesses of chunking theory
Since 1973, new data have led to a revision of some of the re-
sults presented above. For example, Gobet and Simon (1996b, p.
159) established, through a review of chess experiments where
random positions served as control material, that ‘‘strong players
generally maintain some superiority over weak players even with
random positions, although the relative difference between skill
levels is much smaller than with game positions.’’ Using the
CHREST (Chunk Hierarchy and REtrieval STructures) computer simu-
lation, the same authors estimated that the number of chunks re-
quired in LTM to reach a master level was about 300,000 (Gobet
& Simon, 2000).
The new studies have helped clarify the chunking mechanisms
but have also cast doubt on one crucial point of the theory. Chunk-
ing theory, as developed by Chase and Simon (1973a), hypothe-
sizes that information, once encoded as chunks (also called
pointers), is stored in WM. But, as Gobet (2000a) has noted, (a) it
is difficult to understand how very complex tasks can be solved
by juggling chunks in a limited space like WM, and (b) the informa-
tion encoded as chunks (or at least part of these chunks) may be
stored in LTM.
Concerning this last point, Charness showed as early as in 1976
that an interfering task between the presentation of a chess posi-
tion and the test phase reduced the recall performance of experts
only marginally. If the chunks were stored in WM, as theorized
by Chase and Simon (1973a), then at least some of them should
have been erased by the interfering task. But, since this did not oc-
cur, Charness (1976) suggested that at least some chunks were rap-
idly transferred to LTM. In the same year, Frey and Adesman (1976)
observed a similar result; more recently, Cooke, Atlas, Lane, and
Berger’s (1993) study confirmed it. In addition, similar results have
been found in text comprehension using an interfering task
(Glanzer, Dorfman, & Kaplan, 1981; Glanzer, Fisher, & Dorfman,
1984). The absence of the effect of the interfering task has been
interpreted as being due to LTM storage (Ericsson & Kintsch,
1995; Gobet & Simon, 1996a; Kintsch, 1998). The idea that part
of LTM can be used to store information during WM tasks also
explains the performance of experts who are able to recall more
than 100 briefly presented digits (Chase & Ericsson, 1981; for a
complete review, see Ericsson & Kintsch, 1995).
To sum up, it seems that processes supposed to occur solely in
WM also involve, in the case of expertise, LTM storage. As empha-
sized by Gobet (2000a, p. 552), the information is ‘‘transferred in
LTM more rapidly than suggested by the chunking theory.’’ It is
following this idea that Ericsson and Kintsch (1995) and Gobet
and Simon (1996a) proposed their respective theories.
3. Long-term working memory and template theory: a same
core idea
LT-WMT and TT revolve around the same fundamental core
idea: with expertise, part of LTM can be used as WM, thus expand-
ing an individual’s memory storage and processing capacities. In
both theories, this is possible only when knowledge structures
have been built. These structures have been called ‘‘retrieval struc-
tures’’ in Ericsson and Kintsch’s (1995) theory and ‘‘templates’’ in
Gobet and Simon’s theory (1996a). We briefly introduce the princi-
pal features of these two structures.
3.1. Template theory
Templates (Gobet & Simon, 1996a) can be seen as schemas with
slots that can be filled with variable information, including chunks.
An example would be the schema of a train station that may (or may
not) be filled with a train. Chunks and templates are both typical
configurations, but, while chunks are simple patterns, templates
are configurations that can be filled with additional information,
such as pieces in chess. In other words, templates are high-level
patterns (prototypes) that can change their aspect slightly – via
The term ‘‘chunk’’ can be employed to mean both ‘‘the elements that are in WM’’
and ‘‘the elements that are in LTM.’’ To distinguish between these two meanings,
some researchers prefer to use the term ‘‘pointers’’ to refer to the content of WM that
is linked (that points) to LTM knowledge. When the distinction is required, we will
use the terms ‘‘chunks in WM’’ and ‘‘chunks in LTM.’’
222 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
the filling of slots – but that are still categorized in the same way.
Thus, the same template can be used by an expert to encode differ-
ent chess positions, assuming that they belong to the same category.
Gobet and Simon (1996a) presented an example of a template
(see Fig. 1). If a complete novice looks at a typical position from
the King’s Indian Defense – of which the chess board in Fig. 1 de-
picts some of the key pieces – the pattern will not really be mean-
ingful. However, if a fairly strong chess player looks at such a
position, then two templates might be perceived: a first template
for the white pieces and a second template for the black pieces.
If an even stronger player – say a grandmaster – looks at the posi-
tion, then the two templates are likely to be perceived as a single
template. A template has a core (the non-variable part of the tem-
plate) and slots (the variable part). On the chessboard in Fig. 1, the
pieces on the board indicate the core, while the crosses indicate the
positions where chess pieces can be added without changing the
template. This means that a player will categorize different but re-
lated positions with the same template. For example, a white pawn
can be added on h2 or h3, and a black knight can be added on d7. A
template can also be linked to other types of information: open-
ings, moves, plans, and other templates. TT has since been partly
implemented with CHREST. Based on simulations with this model,
Gobet and Simon (2000) have estimated that it takes 250 ms to fill
a template slot.
3.2. Long-term working memory theory
LT-WMT (Ericsson & Kintsch, 1995) differs from TT in that the
storage of information by an expert is not considered as the partic-
ularization of a schema, but as an association between the encoded
information and a set of ‘‘retrieval’’ LTM cues. In order to retrieve the
encoded information, the expert must reinstate the encoding condi-
tions by using the same set of retrieval cues. Long-Term Working
Memory becomes available (but restricted to the field of expertise),
when a set of cues becomes a stable structure in LTM: a retrieval
structure. Ericsson and Kintsch’s theory is a generalization of Chase
and Ericsson’s (1981) skilled memory theory and applies the same
three principles: meaningful encoding, structured retrieval, and
acceleration of encoding and retrieval. The first principle states that,
to be easily stored in LTM, the incoming information needs to be
transformed into meaningful units. The second principle concerns
the retrieval structure (see below for an example). And the last prin-
ciple states that, with extensive practice, experts can speed up the
encoding and retrieval of information in LTM.
Fig. 2 gives an example – inspired by S.F.’s (a runner who used
his knowledge of footraces to encode and recall digits as running
times) performance and verbal protocols analyzed and described
by Chase and Ericsson (1981; see also Ericsson, 1985) – of a retrie-
val structure combined with supplementary associations called
elaborated memory structure. These supplementary associations
are ‘‘knowledge-based associations relating units of encoded infor-
mation to each other along with patterns and schemas establishing
an integrated memory representation...’’ (Ericsson & Kintsch,
1995, p. 221).
In Fig. 2, ‘146732981417’ is the sequence that must be re-
called. If one has a good knowledge of times for footraces (like S.F.
had), the sequence can be encoded by coding the digits as running
times for races (the principle of meaningful encoding). Each running
time is associated with one specific cue: ‘‘first’’ for the first running
time, ‘‘middle’’ for the second running time, and ‘‘last’’ for the third
time. Therefore, ‘‘1 46732981417becomes ‘‘1 min 46.7 s’’ linked
to ‘‘first,’’ ‘‘3 min 29.8 s’’ linked to ‘‘middle,’’ and ‘‘1 min 41.7 s’’
linked to ‘‘last.’’ To recall the information, one must just activate
the cues: first, middle and last (the principle of structured retrieval).
Moreover, supplementary associations can also be created as the
information is encoded following the retrieval structure. In this
example, the encoded stimuli can be associated by means of knowl-
edge of running times; hence, the first and last times can be associ-
ated since they are both 800-m times, while the last two times can
be linked together since they are both very good running times.
These supplementary associations vary according to the incoming
information and help the individual to have ‘‘an integrated memory
representation’’ (Ericsson & Kintsch, 1995).
3.3. Long-term working memory and template theory specificity
compared to working-memory models and theories
Even though, as we have just explained, some differences exist
between LT-WMT and TT (for an analysis, see Cowan, 2005;
Ericsson & Kintsch, 2000; Gobet, 1998a, 2000a, 2000b), both
Fig. 1. Example of a template from Gobet and Simon (1996a). On the left, the attributes of the template: the non-variable part of a template (the template-core) and the
variable part of a template (slots for pieces, squares, openings, plans, moves, and links to other templates). On the right, the diagrammatic representation of the same
template: pieces on the board indicate the core pieces in the template and crosses indicate values contained in piece or square slots.
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 223
theories propose that, when an individual has acquired expertise in
a domain, part of LTM can be accessed rapidly and reliably and
therefore used as WM. This proposal can be compared with WM
models or theories that include immediate access to LTM elements
that are available for processing but are not actively maintained by
attention processes (e.g., Cowan, 1995; Just & Carpenter, 1992;
Oberauer, 2002). But it differs from these models concerning the
description and the influence of the LTM structures that are ac-
quired with expertise: knowledge structures. This central feature
makes LT-WMT and TT different from models based on the classi-
cal definition of WM, that is, ‘‘temporary storage of information
that is being processed in any of a range of cognitive tasks’’
(Baddeley, 1986, p. 34) particularly if by ‘‘temporary storage’’ one
intends ‘‘active maintenance’’ or ‘‘rehearsal.’’
In the domain of WM models, the dominant point of view –
known as the ‘‘resource-sharing model’’ (Hitch, Towse, & Hutton,
2001) – considers that WM depends on cognitive resources, re-
sources that are supposed to be a pool shared by short-term stor-
age and processing. This resource-sharing hypothesis has
dominated WM literature at least since Daneman and Carpenter
(1980). Even if other viewpoints have been advanced (e.g., Engle,
Kane, & Tuholski, 1999; Hasher & Zacks, 1988; MacDonald &
Christiansen, 2002; Maehara & Saito, 2007; Towse, Hitch, & Hutton,
1998, 2000; Waters & Caplan, 1996; for a review, see Miyake,
2001) that do not appeal to the idea of resource sharing, they
‘‘are still based on a core assumption, which is also at the root of
the resource sharing hypothesis, that memory items are ‘‘actively’’
maintained during performance on working memory span tests’’
(Saito, 2006, p. 54, italics added).
LT-WMT and TT clearly assume that WM tasks do not only mea-
sure WM in the form of a temporary storage, but also as a capacity
to massively retrieve information from LTM via knowledge struc-
tures. This allows LT-WMT and TT to explain data that are difficult
to interpret otherwise, for instance, how skilled activities can be
interrupted and resumed later with no detrimental effects on per-
formance (e.g., Charness, 1976; Cooke et al., 1993; Frey & Adesman,
1976; Glanzer et al., 1981, 1984). If one assumes that the elements
in WM can be associated to elements in LTM (i.e., knowledge struc-
tures) and can thus be rapidly transferred to LTM in order to later
be retrieved, the absence of a disruption effect becomes
3.4. Physiological consequences of long-term working memory and
template theory
In theory, the capacity to use LTM structures during WM tasks
has also physiological consequences. In fact, in the case of experts,
if the incoming information instead of being stored simply in
WM is rapidly linked to knowledge structures – which depend
on semantic memory – and if the result of the association is stored
as an episode in episodic memory, then one can assume a consid-
erable involvement by both WM and LTM brain areas during WM-
related tasks. This should contrast with the involvement of almost
only WM brain areas in the case of novices. In other words, from the
point of view of both Ericsson and Kintsch’s and Gobet and Simon’s
theories, a cerebral functional reorganization (Kelly & Garavan,
2005) implicating LTM areas is expected with expertise acquisi-
tion: Here, cerebral functional reorganization does not simply
mean that experts do the same mental operations as novices but
with different brain areas. Rather, it means that experts do the
tasks differently, that is, they execute WM tasks using different
mental operations based on LTM areas.
However, while behavioral data in various domains seem to
corroborate the idea of LTM storage during WM-related tasks, with
the involvement of both WM and LTM in the case of experts (for
examples in air piloting, Sohn & Doane, 2003; in chess, Gobet &
Clarkson, 2004; Gobet & Jackson, 2002; Gobet & Simon, 1996a,
1998; Saariluoma & Kalakoski, 1997; in the game of Go, Masunaga
& Horn, 2000; in gymnastic floor routines, Tenenbaum, Tehan,
Stewart, & Christensen, 1999; in mnemonists, Ericsson, Delaney,
Weaver, & Mahadevan, 2004; in reading span, Guida, Tardieu, &
Nicolas, 2009; in text comprehension, Guida & Tardieu, 2005;
Hupet, Schelstraete,Demanet, & Pourtois, 2000;inverbalknowledge
of football, Postal, 2004; and in written production, Kellogg, 2001),
neuroimaging results in PET and fMRI studies provide only partial
confirmation. In fact, in WM-related tasks, the neuroimaging liter-
ature on experts and expertise acquisition exhibits two contrasted
patterns of results. While experts’ neuroimaging studies are fairly
congruent with Ericsson and Kintsch’s and Gobet and Simon’s pre-
dictions of a cerebral functional reorganization implicating LTM
areas, the studies involving trained novices are mostly consistent
with a decrease in cerebral activity related to practice, with no
cerebral functional reorganization.
First Middle Last Retrieval
distance times
An average
800 m time
A very good
1500 m time
A very good
800 m time Associatio ns
via patterns
Two 800 m times
Two very
good times
Fig. 2. Example of a retrieval structure and an elaborated memory structure, adapted from Ericsson and Kintsch (1995) and Ericsson and Delaney (1999).
224 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
Before reviewing the studies having shown these two different
patterns of results, we need to put forward the assumptions that
will guide us throughout this paper and some of the issues that
stem from these assumptions.
4. Assumptions, limitations and physiological predictions of the
4.1. Separating episodic long-term memory activation from working
memory activation
Our main assumption is that it is possible to separate brain acti-
vations that are from episodic LTM and brain activations that are
from WM. In the literature, there seems to be evidence supporting
the idea that WM and episodic LTM are different mental functions
underpinned by specific and different neural pathways (for re-
views, see for example Baddeley, 2003; Gazzaniga, Ivry, & Mangun,
2009; Kandel, Schwartz, & Jessell, 2000; Smith & Kosslyn, 2007;
Squire & Wixted, 2011). However, the main difficulty when
defending this point of view is the overlap that does exist between
WM and LTM structures. For example, a meta-analysis that com-
pared regions activated during verbal LTM and short-term memory
(STM) tasks clearly showed overlapping regions in frontal and pari-
etal lobes (Cabeza, Dolcos, Graham, & Nyberg, 2002; Cabeza &
Nyberg, 2000). Nevertheless, we would like to argue that there is
enough evidence showing that the medial temporal lobe is the
key element to separate episodic LTM from STM/WM. For over
50 years, data from neuropsychology have clearly linked the med-
ial temporal with LTM and not with STM/WM (e.g., Cave & Squire,
1992; Müller & Knight, 2006; Scoville & Milner, 1957; Shallice &
Warrington, 1970). A large number of neuroimaging studies have
also confirmed this, exhibiting a link between episodic memory
and the medial temporal lobe (for reviews, Eichenbaum, Yonelinas,
& Ranganath, 2007; Squire, Stark, & Clark, 2004) and an absence of
link between the STM/WM and the medial temporal lobe (for re-
views, Baddeley, 2003; Collette, Hogge, Salmon, & Van Der Linden,
2006; D’Esposito, 2001).
Several authors have cast doubt on this 50 year old link by putt-
ing forward data that show that the medial temporal lobe could
also be engaged in STM tasks (e.g., Ranganath & Blumenfeld,
2005; Ranganath & D’Esposito, 2005; and Nichols, Kao, Verfaellie,
& Gabrieli, 2006). However, recent reviews have proposed a con-
vincing explanation concerning this contradiction between
50 years of neuropsychology and some neuroimaging studies.
Jonides et al. (2008) made the observation that the neuroimag-
ing studies in question have used tasks with longer retention inter-
vals than those employed in neuropsychological studies, making
STM tasks look more similar to LTM tasks. This crucial and basic
idea that tasks must tap STM to uncover STM activity has also been
put forward in a recent review by Squire and Wixted (2011). These
authors claim that the medial temporal lobe does not play any kind
of role in WM, and that these results are only an artifact. For these
authors, it is only when WM capacity has been exceeded, and
therefore when performance must also rely on LTM, that the med-
ial temporal lobe becomes active. Squire and Wixted (2011) sug-
gest that if methodological precautions are taken, then the
results are straightforward as shown in recent studies (Jeneson,
Mauldin, Hopkins, & Squire, 2011; Jeneson, Mauldin, & Squire,
2010 ;Shrager, Levy, Hopkins, & Squire, 2008).
Of course, an alternative possibility is that neurons subtending
different cognitive functions are more closely interleaved than nor-
mally assumed (Smith & Kosslyn, 2007); this could for example be
the case for the neural tissues devoted to WM and LTM in the med-
ial temporal lobe. If the latter hypothesis turns out to be correct,
more sophisticated techniques (such as high-resolution fMRI)
would be needed to identify the two stages that we propose. But
it should also be noted that chunk-based theories, by assuming
that pointers to LTM chunks are placed in WM, to some extent cap-
ture the close interaction between these two sets of structures.
However, building on Jonides et al.’s (2008) and Squire and
Wixted’s (2011) arguments and basing ourselves on the substantial
neuropsychological and neuroimaging literature (see also Sec-
tion 5.1.), we feel that the medial temporal lobe can be used as a
signature for LTM. Doing so, we assume that it is possible to ‘‘re-
verse engineer’’ the mapping between brain structures and cogni-
tive functions, arguing that reverse inference is the name of the
game of much brain imaging. While techniques have been devel-
oped to facilitate such inferences (Schyns, Gosselin, & Smith,
2009), the logical and technical difficulties of doing so are consid-
erable (Poldrack, 2006), and some authors have questioned the
assumption that WM and LTM cognitive constructs can be mapped
to specific regions in the brain (e.g., Postle, 2006). Nonetheless,
building on Poldrack’s (2006) argumentation, we will show that
this issue is mitigated in our review.
4.2. Reverse inference and Bayes factor
Reverse inference can be translated into probabilistic terms
(Poldrack, 2006; Sarter, Berntson, & Cacioppo, 1996) using Bayes’
theorem. Doing so, it appears that the degree of belief in a reverse
inference (in order to know if the activation of an area increases
the probability of a process to have occurred, here LTM or WM) de-
pends upon the selectivity of the neural response (viz., the ratio of
process-specific activation to the overall likelihood of activation in
that area across all tasks) and the prior belief in the engagement of
cognitive process (here LTM or WM) given the task manipulation.
It follows that if one wants to increase confidence into reverse
inference, one needs to have greater selectivity of response in the
brain region of interest, or/and a higher prior probability of the
cognitive process in question.
Concerning the latter, a clear and independently supported the-
ory seems a way to increase the prior probability of a cognitive
process to be engaged. In our case, and taking experts undergoing
WM-related tasks, LT-WMT (Ericsson & Kintsch, 1995) and TT
(Gobet & Simon, 1996a) strongly predict the involvement of LTM
processes. Predictions, it is to be noted, that were made before
the data themselves were collected.
Concerning selectivity in reverse inference, if a region (e.g., the
hippocampus) is activated relatively selectively by a specific pro-
cess of interest (e.g., an LTM task), then one can infer with substan-
tial confidence that the process is engaged given activation in the
region. Poldrack (2006) sketches out a method to estimate selectiv-
ity using internet databases and the Bayes factor (i.e., the ratio of
the posterior odds to the prior odds), in order to calculate if the
activation in the area of interest increases the odds of engagement
of the cognitive process. We followed the procedure using the
same database (;Laird, Lancaster, &
Fox, 2005)asPoldrack (2006). Before taking a look at the results,
it is important to mention that, according to Jeffreys (1961) and
Kass and Raftery (1995), a Bayes factor between 1 and 3 offers
weak evidence, between 3 and 10 offers substantial evidence,
and from 10 offers strong evidence.
Taking the medial temporal lobe (the hippocampus and para-
hippocampal gyrus) as Location and WM as the Behavioral Domain
as noted in the database, the Bayes factor is 0.4 (see for data,
Appendix A,Table 1A); conversely, for explicit LTM as Behavioral
Domain and for the same Location, the Bayes factor is 6 (see for
data, Appendix A,Table 2A; the factor is above 10 for the hippo-
campus taken alone). This pattern is reversed, when one takes
the prefrontal lobe (inferior, middle and superior frontal gyri)
as Location. That is, the Bayes factor is above 10 (see for data,
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 225
Appendix A,Table 3A) when WM is the Behavioral Domain and
1.88 (see for data, Appendix A,Table 4A) when it is explicit LTM.
From this, two important things can be derived. First, since the
degree of belief in a reverse inference depends partly upon the
selectivity of the neural response and the prior belief in the
engagement of a cognitive process (LTM or WM) given the task
manipulation, we believe that in our case reverse inference can
be used. Second, given the Bayes factors computed above, it ap-
pears that reverse inference gives us quantitative arguments to
postulate that the medial temporal lobe is clearly linked with
LTM while the medial temporal lobe is not linked with WM.
4.3. Physiological predictions
Putting all these remarks together, we believe that we can feel
confident about the following two predictions. First, based on the
Bayes factors computed above and the neuropsychological and neu-
roimaging literature (presented above and in Section 5.1.), we pre-
dict that medial temporal activation, which is supposed to
represent episodic LTM activation, should be absent in novices in
WM-related tasks while prefrontal and parietal activations should
be found. Second, based on TT and LT-WMT, the Bayes factors, and
the neuropsychological and neuroimaging literature (presented
above and in the Section 5.1.), we predict that experts’ brain activa-
tion will include a medial temporal activation in WM-related tasks
and prefrontal and parietal activations. Therefore, the key element
for contrasting both groups should be the medial temporal lobe.
Concerning semantic memory and the areas involved (mainly
lateral and ventral areas of the temporal lobe, see for more preci-
sions Section 5.1.), unfortunately the difference between experts
and novices cannot be predicted in binary terms, absence vs. pres-
ence. In fact, in both groups, one can expect semantic knowledge to
be activated during WM-related tasks, the areas depending on the
type of information manipulated in the task (e.g., knowledge of
words, knowledge of objects, knowledge of faces, and so on). The
only prediction that could be made concerns the richness of the
semantic knowledge involved, which should be more important
for the experts in their domain of expertise; this could have a con-
sequence in terms of brain activation clusters, but not in a clear-cut
way as for the episodic memory.
5. Neuroimaging of experts
Chronologically, the first study using experts that exhibited re-
sults consistent with a cerebral functional reorganization implicat-
ing LTM areas was the one by Pesenti et al. (2001). The authors
contrasted an expert prodigy (R. Gamm, 6 year of practice) with
a group of non-experts, asking all the participants to carry out sim-
ple multiplications (the scan base-line condition) and complex
ones. While simple multiplications were supposed to be just re-
trieved, complex ones necessitated retrieval and multi-step calcu-
lations through WM. Due to R. Gamm’s expertise, the step
dynamics used in his multi-step calculations were highly struc-
tured and practiced (see Pesenti, Seron, Samson, & Duroux, 1999;
Pesenti et al., 2001), taking the form of schemas or retrieval struc-
tures, LTM knowledge that allowed him to encode information in
LTM as suggested by LT-WMT and TT.
The PET results (see Table 1 for results in extenso) showed that
the non-expert group and the expert used some of the same brain
areas: mainly prefrontal areas and parietal areas, with also occipi-
tal areas involved and the left inferior temporal gyrus. By contrast
the areas activated only by the expert were the following: the right
medial frontal gyrus, the upper part of the right anterior cingulate,
the left paracentral lobule, the right middle temporal gyrus, and
the right parahippocampal gyrus.
In the field of chess, Saariluoma, Karlsson, Lyytinen, Teräs, and
Geisler (2004) asked six chess (37 years of practice) players to play
blindfold chess and to execute blindfold chess tasks during a PET
investigation. Saariluoma et al. (2004) did not employ a control
group, because contrary to a normal chess game, a novice cannot
play blindfold chess without being quickly overwhelmed. In order
to play blindfold, experts must keep track of the location of each
piece on the board, and therefore they have to continuously update
their representation of the game via WM while thinking and plan-
ning the subsequent moves in WM. To do so, experts use also their
knowledge of the game under the form of LTM representations:
chunks (Chase & Simon, 1973a), templates (Gobet & Simon,
1996a) and retrieval structures (Ericsson & Kintsch, 1995).
The authors suggested that contrasting a memory condition
(participants had to follow a chess game auditorally, memorizing
the moves of a game and recalling them when prompted) and an
attention condition (participants had to identify a specific piece
when its name was pronounced) would provide information about
mental imagery and storage in LTM (Ericsson & Kintsch, 1995;
Gobet & Simon, 1996a). This is because none of these processes
were required in the attention condition, while the memory
condition necessitated storing the locations of the pieces and their
moves (in addition to the identification process necessary in the
attention condition).
PET results (see Table 1) showed mainly an activation pattern
that concerned bilaterally the middle frontal gyrus, the angular
gyrus bilaterally, and the inferior temporal cortex. There was also
activation in the middle cingulate, and in the right middle tempo-
ral cortex.
The same kind of pattern of brain activation was also obtained
with chess players in another study, but in a less constrained situ-
ation. Campitelli, Gobet, Head, Buckley, and Parker (2007) asked
five chess experts to recognize previously-presented chess patterns
during functional magnetic resonance imaging (fMRI) with a
forced-choice procedure. Every sample stimulus was displayed
for 6.5 s; this stimulus could belong to one of three experimental
conditions: a ‘‘game’’ condition or a ‘‘random’’ condition (16 pieces
displayed as a game position or randomly on a 4 8 matrix), a
‘‘scene’’ condition (16 geometrical figures displayed on a board
the same size as the 4 8 matrix). After a 2-s delay, seven test
stimuli of the same type appeared sequentially. For the first three
conditions, the participant had to decide whether each one of the 7
test stimuli matched the sample stimulus. This is a standard WM
task. However, experts can typically use simple patterns (chunks)
and slotted schemas (templates) to encode and retrieve the incom-
ing information using both WM and LTM (Ericsson & Kintsch,
1995; Gobet & Simon, 1996a).
Subtracting the scene condition (which was supposed to elicit
WM but not chess LTM knowledge) from the game condition,
and subtracting the random condition (which was supposed to eli-
cit WM but not chess LTM knowledge) from the game condition
had for result activation clusters in the following areas: posterior
cingulate, both precunei, right supramarginal gyrus, left lingual
gyrus, temporal lobe areas including right fusiform gyrus, as well
as inferior temporal gyrus and parahippocampal gyrus bilaterally.
Maguire, Valentine, Wilding, and Kapur (2003) used also fMRI
to compare a group of 10 control participants to a group of 10
memory experts who had been using mnemonics techniques for
11 years. During the fMRI session, the experimenters asked the
two groups to remember sequences of six items presented visually.
Each item was presented alone for 4 s. It is during this period of
presentation that the scans of interest were made. Three types of
items were employed: three-digit numbers, faces and snowflakes.
Just after the six items, participants saw three pairs of items from
the last sequence and, had to indicate which item of each pair was
presented first in the last sequence. This task typically elicits WM.
226 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
Table 1
Neuroimaging studies of experts and novices during working memory related tasks key (ranked in order of appearance from left to right in the table).
Fontal Cin-
PTemporal Sub-
1expert mental
calculator (6 years)
Med.R Ant.R L Mid.R R Controls-experts
differences are
congruent with a
functional brain
reorganization that
involves LTM
Mid.L Inf.Bi L Bi Bi L Inf.Bi
Maguire et
al. (2003)
6 digits,
faces and
serial order
10 adult mnemonists
(11 years)
10 novice adults
Sup.L LbR Pa.Bic Controls-experts
differences are
congruent with a
functional brain
reorganization that
involves LTM
L Mid.Ld
additions of
1 digit
and of 4
9 abacus expert
children (6 years)
differences are
congruent with a
functional brain
reorganization that
involves general
LTM areas and
numerical LTM
Si: L
et al. (2004)
5 chess expertse
(37 years)
Mid. Bi Mid Bi Inf.Bi
Experts used both
WM and LTM
Sup. R
Ant R Bi CB.R
et al. (2007)
(16 pieces)
5 chess expertse
(not provided)
Pos.R Bi R L Inf.R R Bi Experts used both
WM and LTM
Inf. R
Sup.R Bi Sup.Bi Bi
LTM ar.
LTM ar.
6 novice adults
8 novice children
WM ar.
WM ar.
Pesenti et
al. (2001)
Mental mul-
(2 digit
Tanaka et
al. (2002)
Digit span
task using
max. digit
span - 2 g
10 abacus experts
(12 years)abus
(12 years)
Sup.Bi Mid.L
differences are
congruent with a
functional brain
reorganization that
involves numerical
LTM areas
et al.
(1, 3, 6
6 abacus experts
(17 years)
8 novice adults
Inf.R R R Controls-experts
differences are
congruent with a
functional brain
reorganization that
involves numerical
LTM areas
Inf.L Sup.
Med.L Ant.R Inf.L
Chen, Wu et
al. (2006)
(1, 2 digit
6 abacus expertsh
(not provided)
6 novice adults
R R Inf.Bi Controls-experts
differences are
congruent with a
functional brain
reorganization that
involves numerical
LTM areas
L Sup.Bi Bi
Inf.L Ant L
13 novice adults
Chen, Hu et
al. (2006)
E: brain areas activated exclusively by the experts, E&C: brain areas activated both by the experts and the controls, C: brain areas activated exclusively by the controls, LTM ar:
brain areas obtained with a contrast supposed to elicit long-term memory areas during a working memory related task, WM ar: brain areas obtained with a contrast supposed
to elicit working memory areas during a working memory related task, Si: simple additions, Co: complex additions, R: right, L: left, Bi: bilateral, inf: inferior, sup: superior,
mid: middle, med: medial, ant: anterior, pos: posterior, FG: Frontal Gyrus, PrG: Precentral Gyrus, FrS: Frontal Sulcus, PrS: Precentral Sulcus, PoG: Postcentral Gyrus, Pcl:
Paracentral Lobule, PL: Parietal Lobule, IPS: Intra-Parietal Sulcus, Pc: Precuneus, SmG: Supramarginal Gyrus, AnG: Angular Gyrus, IOS: Intra-Occipital Sulcus, OG: Occipital
Gyrus, LgG: Lingual Gyrus, Cun: Cuneus, Ins: Insula, TG: Temporal Gyrus, FuG: Fusiform Gyrus, Hi: Hippocampus, PhG: Parahippocampal Gyrus, Cd: Caudate, P: Palladum, Am:
Amygdala, CB: Cerebellum, LT-WM: Long-Term Working Memory, TT: Template Theory, LTM: Long-Term Memory, WM: Working Memory.
We have not reported in the table the behavioral results because, as expected the experts always outperformed the controls. However it is to mention that no inferential test
was reported by Chen, Hu et al. (2006).
This activation was detected only when the order recognition task concerned digits.
Concerning the activation in the pallidum, the activation was on the right hemisphere for faces and on the left hemisphere for snowflakes.
This activation was detected only when the order recognition task concerned snowflakes.
Since Saariluoma et al. (2004) and Campitelli et al. (2007) did not include a control group, we have included only the data from experts. For the first authors we have
presented the areas activated when subtracting the ‘‘attention’’ condition from the ‘‘memory’’ condition, the result is supposed to show the LTM areas activated during
blindfolded chess. For the latter authors we have presented the areas activated when subtracting the ‘‘scene’’ condition (which putatively elicits WM) to the ‘‘game’’ condition
(which putatively elicits LTM knowledge and WM), the result is mainly supposed to be an activation pattern of LTM areas during the chess positions recognition task.
Since Saariluoma et al. (2004) and Campitelli et al. (2007) did not include a control group, we have included only the data from experts. For the first authors we have
presented the areas activated when subtracting the ‘‘memory’’ condition from the ‘‘problem solving’’ condition, the result is supposed to show the WM areas activated during
blindfolded chess. For the latter authors we have presented the areas activated when subtracting the ‘‘perceptual-motor control’’ condition (which putatively elicits LTM
knowledge but not WM) from the ‘‘game’’ condition (which putatively elicits LTM knowledge and WM), the result is supposed to show the WM areas activated during the
chess positions recognition task.
Before the fMRI scanning, participants had their digit span tested, the maximum digit span score minus two was used as a criterion in the fMRI scanning task.
Chen, Wu et al. (2006) did not report the Talairach coordinates for the activations, they mainly reported Brodmann areas. In order to compare Chen, Wu et al.’s (2006) results
to the rest of the studies, we have transformed the Brodmann areas into gyri.
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 227
However, the experts, instead of simply storing the information in
WM, used the method of loci.
Therefore, they used a mental route
as a LTM slotted schema associating the items to be remembered
with the slots (the salient points or the familiar locations). The neu-
roimaging data showed, in both groups and during all stimuli (see
Table 1 for results in extenso), an activation in frontal areas, in the left
angular gyrus, in the left middle occipital gyrus, and in the caudate
and cerebellum bilaterally. By contrast, three brain regions were ac-
tive only for the mnemonist group for all stimuli types: the left med-
ial superior parietal cortex, the bilateral retrosplenial cortex and the
right posterior hippocampus.
We conclude this first part of our review by discussing expertise
in mental abacus calculation. Just like Maguire et al. (2003) did
with mnemonists, Chen, Hu et al. (2006) used fMRI to compare a
group of (eight children) novices to a group of (nine children) ex-
perts. The experts had practiced operations and mental abacus cal-
culations for 30–60 min a day, for about 5.5 years on average.
Simple (single-digit numbers) and complex additions (four-digit
numbers for the experts and two-digit numbers for the controls)
were used. This task is demanding in terms of WM resources, espe-
cially with complex calculations. With experts, it is proposed that
they use a mental representation of an abacus that is stored in LTM,
which constitutes a retrieval structure (see Ericsson & Kintsch,
The fMRI results showed that in the case of novices, the activa-
tion pattern for both simple and complex calculations concerned
the left inferior frontal gyrus, the anterior cingulate, the postcen-
tral gyrus, and the inferior parietal lobule bilaterally. For novices
and experts, activations were observed in the right superior frontal
gyrus for simple calculations and in the left middle frontal gyrus
and the left parahippocampal gyrus for complex calculations. In
the case of experts and for simple calculations, activations were
observed in the right lateral premotor cortex and bilaterally in
the posterior temporal areas, and for complex calculations the acti-
vation occurred bilaterally in the lateral premotor cortex, the pos-
terior superior parietal lobule, and in the parahippocampal gyrus.
5.1. Summarizing the cerebral activation in experts and providing
evidence of a LTM involvement in WM-related tasks
The first part of Table 1 presents the neuroimaging studies we
have just reviewed. When one examines the columns of Table 1
presenting the areas activated by novice controls and experts, it
seems clear that prefrontal and parietal areas are involved to a great
extent. This result was expected given the fact that the tasks used
were WM-related and given the importance of these areas for
WM (for prefrontal areas see for example, D’Esposito et al., 1995;
Postle, Berger, & D’Esposito, 1999; Prabhakaran, Narayanan, Zhao,
& Gabrieli, 2000; Shimamura, 1995; Ungerleider, Courtney, &
Haxby, 1998; and for the importance of parietal areas see for exam-
ple, Cowan, 2001; Cowan, 2005; Dehaene & Cohen, 1994; Postle &
D’Esposito, 1999; Rodriguez et al., 1999; Todd & Marois, 2004).
By contrast and noticeably, the areas activated only in experts
showed activations in medial temporal regions, such as the para-
hippocampus for Pesenti et al. (2001), Campitelli et al. (2007)
and Chen, Hu et al. (2006) and in the right hippocampus for
Maguire et al. (2003). As emphasized earlier, we believe that this
is a marker of experts’ strong reliance on episodic LTM during
WM-related tasks, which is also the point of view put forward by
the authors of the studies reviewed.
Concerning the lateral and ventral temporal regions that we
relate to semantic memory, experts showed more clusters of acti-
vation distributed across these areas. Activations were observed
in the inferior temporal gyrus (right, Campitelli et al., 2007;or
bilaterally, Chen, Hu et al., 2006; Saariluoma et al., 2004), in the
middle temporal gyrus (bilaterally, Chen, Hu et al., 2006; or right
Saariluoma et al., 2004), in the superior temporal gyri (bilaterally,
Chen, Hu et al., 2006), and in the fusiform gyrus (right, Campitelli
et al., 2007, or bilaterally, Chen, Hu et al., 2006).
In the following paragraphs, we provide evidence from the liter-
ature in favor of the involvement of the above-mentioned areas
concerning LTM, starting with the medial temporal lobe and epi-
sodic memory. Activations in medial temporal structures (hippo-
campus and parahippocampus) are often observed during
episodic encoding and retrieval (sometimes bilaterally), regardless
of the verbal or non-verbal nature of the materials tested (for re-
views, Eichenbaum et al., 2007; Squire et al., 2004). More specifi-
cally, the parahippocampal region in the right hemisphere seems
to control the storage and maintenance of stimuli representations
for long delays (Young, Otto, Fox, & Eichenbaum, 1997), and seems
predominantly dedicated to the visuospatial aspects of these pro-
cesses. It has also been proposed that this area plays an important
role in a kind of expertise for local visual environment. Epstein and
colleagues (Epstein & Kanwisher, 1998; Epstein, Stanley, Harris, &
Kanwisher, 2000) have thus proposed the name of parahippocam-
pal place area, but it is possible that in other kinds of expertise (for
example in chess, Campitelli et al., 2007) the utilization of this area
could slightly be transformed in order to be used in other kind of
visual expertise (e.g., from local visual environment to chess).
The right posterior hippocampus has been shown to be very
important for LTM and specifically for episodic memory. Several
important models of memory are built around the hippocampus,
making it play an essential role in the encoding and retrieval pro-
cesses (the ‘‘standard’’ consolidation model, e.g., Squire & Alvarez,
1995; the multiple trace theory, e.g., Nadel & Moscovitch, 1997;
the HIPER model, Lepage, Habib, & Tulving, 1998). The posterior
hippocampus has been said to be more involved in the encoding
of unfamiliar information (Gabrieli, Brewer, Desmond, & Glover,
1997) but conversely also in the retrieval process (Lepage et al.,
1998). The right posterior hippocampus has also a clear involve-
ment in spatial memory and navigation (Burgess, Maguire, &
O’Keefe, 2002; O’Keefe & Nadel, 1978). In the case of Maguire
et al.’s (2003) experts, this activation could be due to the use of
the method of loci.
While these regions seem linked to episodic LTM, we believe
that they are poorly linked to WM. And even if some researchers
have exhibited a link between WM and the medial temporal lobe
(e.g., Ranganath & Blumenfeld, 2005; Ranganath & D’Esposito,
2005; and Nichols et al., 2006), we believe that this evidence would
have to be compared with the huge amount of neuropsychological
and neuroimaging literature (for reviews, see Baddeley, 2003; Cave
& Squire, 1992; Collette et al., 2006; D’Esposito, 2001; Gazzaniga
et al., 2009; Müller & Knight, 2006; Squire & Wixted, 2011), where
such a link has not been found. As seen in Section 4.2, one way of
approaching this question is to compute Bayes factors in order to
estimate the selectivity of the medial temporal lobe for each pro-
cess (LTM vs. WM). It appears that there is no evidence of a link be-
tween WM and the temporal medial lobe while there is substantial
evidence for a link between explicit LTM and the medial temporal
Concerning the lateral temporal cortex (superior, middle and
inferior temporal gyri), it seems clearly involved in semantic mem-
ory, as shown since the late 1970s in clinical research in several
pathologies (Alzheimer’s disease, e.g., Martin & Fedio, 1983; Pick’s
This famous mnemonic technique has been known for at least 2500 years. The
ancient Greeks used it to remember important words during speeches. Before a
speech, an orator would visualize a familiar route – which can be considered as a
schema – with salient points along it or familiar locations (loci means ‘‘locations’’ in
Latin) and use them to mentally store or associate important words. Later, during the
speech, the orator could retrieve these words just by mentally walking down the
route again.
228 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
disease, e.g., Warrington, 1975; semantic dementia, e.g., Hodges,
Patterson, Oxbury, & Funnell, 1992; and strokes, e.g., Cloutman
et al., 2009). More recently the link between temporal atrophy
(especially in the anterior part) and semantic memory deficits
has been confirmed by several authors (Galton et al., 2001; Garrard
& Hodges, 2000; Gorno-Tempini et al., 2004; Mummery et al.,
2000). Neuroimaging studies have also confirmed the involvement
of the lateral temporal cortex in semantic memory, especially in
the left hemisphere (e.g., Binder et al., 1997, 1999; Chee, O’Craven,
Bergida, Rosen, & Savoy, 1999; Démonet et al., 1992; Mummery,
Patterson, Hodges, & Price, 1998; Vandenberghe, Price, Wise,
Josephs, & Frackowiak, 1996). Even if the majority of the studies
favor the hypothesis that semantic information is stored or
processed mainly in the left side of the brain (for a review, see
Martin, 2006), the exact organization of semantic memory does
not seem to be perfectly known, within each hemisphere (for an
overview, see Caramazza & Mahon, 2006) or even when comparing
one hemisphere to the other (e.g., Damasio, Tranel, Grabowski,
Adolphs, & Damasio, 2004).
As far as the ventral areas of the temporal lobes (the inferior
temporal and the fusiform gyri in the studies we reviewed) are
concerned, cognitive neuroimaging studies in humans and single-
cell recording in non-human primates seem to indicate that
they are involved in the storage of familiar patterns in LTM (e.g.,
Damasio, Grabowski, Tranel, Hichwa, & Damasio, 1996; Desimone,
Albright, Gross, & Bruce, 1984; Gross, 1992; Logothetis, Pauls, &
Poggio, 1995; Stark & Squire, 2000; Tanaka, 1993). In the right side
of the brain ventral areas, the fusiform gyrus and the inferior
temporal gyrus are known to be involved in object recognition,
and more generally in the encoding and retrieval of the figurative
properties of visual representations, even when the objects are just
imagined (e.g., Mellet et al., 1996).
As for the fusiform gyrus in particular – most often the right
mid-fusiform, but sometimes bilaterally – a subset of it has been
named the Fusiform Face Area (FFA), because (a) its activity in-
creases when faces are seen or generated by mental imagery
(Kanwisher, Tong, & Nakayama, 1998; McCarthy, Puce, Gore, &
Allison, 1997; O’Craven & Kanwisher, 2000); and (b) because its
response is greater for faces than for objects, animals or scenes, or
other parts of the human body (Aguirre, Singh, & D’Esposito,
1999; Haxby, Hoffman, & Gobbini, 2000; Ishai, Ungerleider, Martin,
Schouten, & Haxby, 1999; Kanwisher, McDermott, & Chun, 1997;
Kanwisher et al., 1998; Yovel & Kanwisher, 2004). However, the
FFA may not be specifically a face area. Gauthier, Tarr, and Ander-
son (1999) suggested that the FFA may be involved in visual pro-
cessing for items individuals have expertise in: faces, of course,
but also cars and birds for car and bird experts (e.g., Gauthier,
Skudlarski, Gore, & Anderson, 2000). Therefore, taking these results
into account, it is possible that in the case of Campitelli et al.’s
(2007) chess experts and Chen, Hu et al.’s (2006) child abacus ex-
perts, the activation of the right fusiform (bilaterally for Chen, Hu
et al., 2006) could be due to a form of visual processing and mental
imagery for items for which these individuals have expertise.
Taken together, these results seem to indicate that the areas
used by experts in WM-related tasks tend to be, in addition to
the WM areas used by novices, also LTM areas. This is clearly in line
with LT-WMT (Ericsson & Kintsch, 1995) and TT (Gobet & Simon,
1996a), which, in the case of experts executing WM-related tasks,
both hypothesize the use of LTM via knowledge structures.
5.2. Neuroimaging of experts: strong hypothesis vs. weak hypothesis
If the first part of Table 1 presents data that is directly compat-
ible with LT-WMT (Ericsson & Kintsch, 1995) and TT (Gobet &
Simon, 1996a), the three studies presented in the second part
of Table 1 (Chen, Wu et al., 2006; Hanakawa et al., 2003; Tanaka,
Michimata, Kaminaga, Honda, & Sadato, 2002) do not exactly fit
in. In fact, instead of having data compatible with a functional reor-
ganization via an involvement of episodic LTM areas (medial tem-
poral lobe) or semantic LTM areas (ventral and lateral temporal
areas), these studies are only compatible with a functional reorga-
nization non-specific to LTM. By ‘‘functional reorganization non-
specific to LTM,’’ we mean that when an individual passes from
the status of novice to that of expert, there is indeed a change
concerning the pattern of areas involved with the activation of
new brain areas or the deactivation of old brain areas (viz., func-
tional reorganization) but this change does not concern the
involvement of more LTM areas (LTM non-specific).
Ericsson (2003, p. 235, text in brackets added) has clearly stated
that both types of results (a functional reorganization with
involvement of LTM areas and a non-specific functional reorgani-
zation) are compatible with LT-WMT, and we agree with him:
‘‘Most important, the differential brain activation (between experts
and a control group) is consistent with cognitive processes pre-
dicted by long-term working memory accounts. For example,
exceptional mental calculators rely on storage in long-term mem-
ory (here Ericsson refers to Pesenti et al. (2001)) and expert mental
abacus calculators encode numbers in a manner qualitatively dif-
ferent from controls (here Ericsson refers to Tanaka et al.’s
(2002) study, which we present below).
Even if we agree with Ericsson about the compatibility of both re-
sults, we feel that the latter result (a non-specific functional reorga-
nization) is qualitatively more distant from LT-WMT and TT
compared with a brain functional reorganization with involvement
of LTM areas, since expertise is supposed to be reached through the
building of LTM structures (chunks and knowledge structures).
Therefore we have decided to name the non-specific functional
reorganization the ‘‘weak physiological hypothesis’’ (‘‘weak
hypothesis’’ from here on) and the brain functional reorganization
with involvement of LTM areas ‘‘the strong physiological hypothe-
sis’’ (‘‘strong hypothesis’’ from here on). We next review and explain
the studies showing results compatible with the weak hypothesis.
5.3. Three abacus studies in favor of the weak hypothesis
To our knowledge, Tanaka et al. (2002) was the first study
examining abacus experts with fMRI neuroimaging. The authors
compared ten abacus experts that had from 8 to 16 years of train-
ing to thirteen non-expert controls. A delayed match-to-sample
task was utilized during fMRI. A target sequence was presented
during 3 s; for each participant, the length of the digit sequence
was two digits shorter than their digit span score (measured previ-
ously). After a 15-s delay participants had to decide whether a test
sequence was the exact same sequence as in the sample. This task
is typically supposed to tap WM. However, as in Chen, Hu et al.
(2006), experts were also supposed to use their LTM through their
mental abacus, which constitutes a retrieval structure that helps
them encode information in LTM (see Ericsson & Kintsch, 1995).
Regarding the fMRI data of the control group, the areas engaged
were located mostly in the left hemisphere (Broca’s area, insula,
medial and precentral gyrus), plus bilaterally the parietal lobules
and the cerebellum. For the experts, the areas engaged were sym-
metrical, mainly in the superior frontal sulcus (bilaterally) and in
the superior and inferior parietal lobules (bilaterally).
Similar results were observed by Hanakawa et al. (2003), which
have also tested abacus experts in an fMRI experiment, using WM-
related tasks. They compared six experts that had more than
17 years of almost daily training to eight controls. All participants
performed a numeral mental-operation task as in Chen, Wu et al.
(2006). Sixteen numbers were presented, one every 2 s. Partici-
pants had to sequentially add the numbers and keep in mind the
final result, in order to perform a recognition test.
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 229
For the controls, except for a bilateral activation in the superior
precentral sulcus and the cerebellum, all the other activations were
in the left hemisphere: in the prefrontal cortex, in Broca’s area, in
the medial frontal gyrus, in the inferior parietal lobule, in the pre-
cuneus, in the intraparietal sulcus, in the fusiform gyrus. For the
experts, except for a left activity in the fusiform gyrus, results in
the numeral task showed a bilateral activity: in the inferior frontal
gyrus, in the superior precentral sulcus, in the precuneus, in the
intraparietal sulcus, and in the cerebellum.
The final study in this section is Chen, Wu et al.’s (2006) exper-
iment. Like in the two studies we have just reviewed, Chen, Wu
et al. (2006) carried out an fMRI investigation using abacus experts.
The authors asked six experts and six controls to execute two types
of calculations: one-digit additions or two-digit additions. Covert
reading (of the numbers to add) was used as the control task.
The fMRI results were obtained subtracting the covert reading
from the calculation condition.
For the non-experts, one-digit and two-digit calculations en-
gaged mainly the same left hemisphere areas with an exclusive
activation of prefrontal and parietal areas. The activations were
in the left Brodmann areas (BA) (Broca’s area) 44 and 45
and in
the left BA 39. In addition the left BA 6 and BA 7 and 40 were also
activated (on both sides of the brain during two-digit calculations).
Finally, the BA 24 and 32 were also activated.
In the case of the experts, the pattern of activation was symmet-
rical, exactly like in Tanaka et al.’s (2002) and Hanakawa et al.’s
(2003) studies. In the two-digit condition, there was a bilateral
activation of the precentral gyrus and in parietal areas (BA 7, 40
and 19). The right postcentral gyrus (BA 2) was also activated. In
the two-digit condition, the same pattern was found, with an acti-
vation in the postcentral gyrus (BA 3 and 4).
5.4. Explaining the weak hypothesis in abacus studies
The three abacus studies we have just reviewed suggest that the
cerebral areas used by abacus experts – during additions in Chen,
Wu et al. (2006) and Hanakawa et al. (2003), or during a digit span
in Tanaka et al. (2002) – are different from those used by control
novices in these same tasks (see also Table 1). Even if frontal and
parietal areas are involved in both cases, the pattern is symmetrical
for experts.
All the authors have suggested that in the case of experts, the
pattern is compatible with visuospatial/visuomotor imagery pro-
cessing (e.g., Courtney, Petit, Maisog, Ungerleider, & Haxby, 1998;
Rowe, Toni, Josephs, Frackowiak, & Passingham, 2000), whereas
for controls, the pattern is more compatible with verbal processing
of information (e.g., Burbaud et al., 1999; Dehaene et al., 1996;
Pesenti et al., 2001; Zago et al., 2001).
This result is clearly congruent with a functional brain reorgani-
zation that would occur during expertise acquisition – the weak
physiological hypothesis derived from LT-WMT and TT – but does
not seem to be compatible with a functional brain reorganization
involving LTM areas, the strong hypothesis.
The question that stems from these results is as follows: why
are all studies using abacus experts only compatible with the weak
hypothesis (with the exception of Chen, Hu et al. (2006), while the
studies involving other tasks (mental multiplication, blindfold
chess, chess positions recognition, digits, faces and snowflakes
serial order recognition) are consistent with the strong hypothesis?
We think that an important part of the answer is to be found in the
dissociation between the areas involved in general semantic LTM
and the areas involved in semantic LTM concerned specifically
with the representation of numbers.
As seen previously, neuroimaging studies have shown that gen-
eral semantic LTM tasks tend to activate mostly a network of areas
where the left temporal cortex plays a crucial role (e.g., Martin,
Haxby, Lalonde, Wiggs, & Ungerleider, 1995; Martin, Wiggs,
Underleider, & Haxby, 1996; Mummery et al., 1998; Tyrrell,
Warrington, Frackowiak, & Rossor, 1990). By contrast, the sense of
number seems to activate parietal areas, near the intra-parietal sul-
cus and neighbor areas (like the inferior parietal lobe). Some studies
pinpoint the sense of number only in the left parietal cortex
(Chochon, Cohen, Van de Moortele, & Dehaene, 1999; Cowell, Egan,
Code, Harasty, & Watson, 2000) while other studies also localize it in
the right parietal cortex (Chochon et al., 1999; Dehaene et al., 1996;
Kazui, Kitagaki, & Mori, 2000). Studies using single neuropsycholog-
ical cases have yielded the same kind of results showing a double
dissociation between general semantic knowledge and numerical
semantic knowledge. For example, in cases of semantic dementia
(Warrington, 1975) where there is an atrophy of the temporal lobes
(e.g., Garrard & Hodges, 2000; Hodges et al., 1992; Mummery et al.,
2000), general semantic knowledge is gradually lost while
numerical capacities are spared (e.g., Cappelletti, Kopelman,
Morton, & Butterworth, 2005; Crutch & Warrington, 2002). Inver-
sely, general semantic knowledge can be spared while numerical
capacities are lost in Gertsman syndrome (Gerstmann, 1940)orin
posterior cortical atrophy (e.g., Delazer, Karner, Zamarian,
Donnemiller, & Benke, 2006; Tang-Wai et al., 2004), which both con-
cern mainly parietal areas.
Therefore, it is possible that, in the case of abacus experts, there
is indeed a functional brain reorganization involving LTM areas,
like in the other studies we have reviewed but, because of the pe-
culiar cerebral localization of numerical semantic knowledge in the
parietal lobes, no physiological reorganization is apparent. In fact,
the parietal areas (mainly the posterior areas) are important for
both WM (e.g., Postle & D’Esposito, 1999; Rodriguez et al., 1999;
Todd & Marois, 2004; for the importance of parietal areas in WM,
see also Cowan, 2001; Cowan, 2005) and for numerical semantic
knowledge (e.g., Chochon et al., 1999; Cowell et al., 2000; Dehaene
et al., 1996; Kazui et al., 2000). Therefore, even though with exper-
tise acquisition, in the case of abacus experts, there is a gradual
activation shift from WM areas (prefrontal and parietal) to WM
areas (prefrontal and parietal) and LTM areas (parietal in the case
of mental abacus calculations), it is understandable that this shift
goes undetected, since parietal areas are crucially involved in both
cases (WM and LTM).
The results of the three above-mentioned studies (Chen, Wu
et al., 2006; Hanakawa et al., 2003; Tanaka et al., 2002) clearly con-
firm the involvement of parietal areas, since it appears that the
only areas that are common to the experts in all three studies
are the intra-parietal sulcus and neighboring areas. More precisely,
these areas are the intra-parietal sulcus (bilaterally) in Hanakawa
et al. (2003), the two parietal lobes (bilaterally, superior and infe-
rior) that surround the intra-parietal sulcus in Tanaka et al. (2002),
and BA 40, 7 and 19 bilaterally in Chen, Wu et al. (2006). Moreover,
in the three studies when the experts - controls contrast was sta-
tistically tested, the only areas (plus the superior frontal sulcus
in Tanaka et al., 2002) that showed a significant activation differ-
ence in favor of the experts were in the parietal region, and more
precisely the superior parietal lobule in Tanaka et al. (2002) and
Hanakawa et al. (2003), and the postcentral gyrus in Chen, Wu
et al. (2006).
Given these findings, we suggest that the results observed by
Tanaka et al. (2002), Hanakawa et al. (2003) and Chen, Wu et al.
In their study, Chen, Wu et al. (2006) did not always provide the stereotaxic
coordinates and the names of the gyri that were activated; instead they always
indicated the Brodmann areas involved. In our review, we have always indicated the
gyri and/or sulci, but since the overlap between Brodmann areas and gyri is not
perfect and since the stereotaxic coordinates have not been provided, when it was not
possible to do otherwise we have just indicated the Brodmann areas concerning Chen,
Wu et al.’s (2006) study.
230 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
(2006) can also be consistent with a brain functional reorganiza-
tion involving LTM areas, although these are numerical semantic
LTM areas.
6. Neuroimaging of trained novices
The results from the fMRI and PET studies that we have re-
viewed above in different tasks and in different domains clearly
suggest that experts and novices do not utilize the same brain
areas in tasks related to WM. We have argued that LTM areas are
strongly involved in the case of experts (for a summary of the re-
sults, see Table 1) and not in the case of novices. We consider this
finding as evidence in favor of a functional cerebral reorganization
involving LTM areas related to expertise acquisition. However, we
acknowledge that the evidence is only indirect because two differ-
ent populations are being compared. In principle, a better approach
would be to follow the evolution of expertise acquisition from the
start in a group of novices, and look for brain functional reorgani-
zation. Some studies have followed this approach. But, surpris-
ingly, when novices have been trained to gain expertise in tasks
related to WM, no cerebral functional reorganization has been ob-
served, the results mostly showing a decrease in activation.
We will next review these studies on novice training, using two
dimensions to group them. The first dimension concerns the exper-
imental paradigm employed and the second concerns the duration
of practice. With regard to the experimental paradigm, the great
majority of the authors (eight studies out of eleven) have used
what we have called the Scanning–Training–Scanning (S–T–S) par-
adigm, where the brain activity is measured before training and
after training, or additionally in between, but the main idea is to
compare the situation before training to the situation after train-
ing. By contrast, in three studies (Andreasen, O’Leary, Arndt,
et al., 1995; Andreasen, O’Leary, Cizadlo, et al., 1995; Moore, Co-
hen, & Ranganath, 2006), the authors have used what we have
called the Training-Scanning (T–S) paradigm, where participants
are first trained concerning a task and a type of material and
scanned only after. We first present the studies using the S–T–S
paradigm; then, we describe the T–S paradigm, presenting the
three studies that are concerned.
With regard to the practice duration dimension, short training
regimens (in the order of hours) were employed in six studies,
while medium-length training regimens
(in the order of weeks)
were employed in five studies. We next review these studies, begin-
ning with short-training regimens. It is to note that we have taken
into consideration for the short-length training only experiments
that encompassed a practice of at least 1 h, which seemed to be an
adequate length in order to start speaking of training.
6.1. Scanning–Training–Scanning paradigm experiments with short-
length training
Chronologically, the first study using short training was Garavan,
Kelley, Rosen, Rao, and Stein’s (2000). They used a visuospatial
delayed-match-to-sample task that was used originally by Jonides
et al. (1993). During the presentation phase, participants saw three
dots briefly. After a 3-s delay, three dots were again displayed with
a probe circle surrounding one of those dots. Participants had to de-
cide, by pressing a yes or no button, if the surrounded dot was at the
same location as in the presentation phase. Participants had to
complete 880 trials (over 8 h). Response times improved whereas
accuracy did not. The decrease in response time corresponded to
a decrease in activation in the prefrontal and parietal areas that
are involved in WM (Cabeza & Nyberg, 2000; Carpenter, Just, &
Reichle, 2000) and in the occipital cortex; no functional reorganiza-
tion was detected (see Table 2 for more details).
Landau and colleagues (Landau, Garavan, Schumacher, &
D’Esposito, 2007; Landau, Schumacher, Garavan, Druzgal, &
D’Esposito, 2004) also showed an activation decrease in delayed-
match-to-sample tasks, using event-related fMRI, which allowed
them to differentiate the encoding process from the retrieval and
isolate the delay period. In the first study (Landau et al., 2004), a
stimulus set composed of four intact or scrambled faces was shown
to the participants, who were asked to remember all the intact
faces after an 8-s delay period; there could be one, two, three or
four intact faces (the load factor). During the scanning session,
the first 18 min were compared to the last 18 min, for a total of
48 min of scanning (plus 30 min of unscanned practice). The re-
sults showed a decrease in activation in the frontal, temporal, pari-
etal, and occipital areas related to practice, and no functional
reorganization. Concerning the load factor and the periods of im-
age acquisition (encoding, delay, and retrieval), a decrease in acti-
vation and effects of memory load both occurred primarily during
encoding. The authors concluded that practice improved encoding
efficiency, especially at higher memory loads. It is also important
to note that the decreases in activation did not correspond to
improvements in the behavioral task.
More recently, Landau et al. (2007) have confirmed the practice-
related decreasing pattern but with a visual and a spatial WM task.
Participants practiced approximately 60 min. Like in their previous
study, the authors used an event-related fMRI investigation, but
unlike the previous studies, they used irregular polygons, designed
to be difficult to encode verbally. For spatial runs, participants
were asked to remember the spatial position of polygons whereas
for object runs, participants were asked remember the shape.
Behavioral results showed a linear decrease of the reaction times
during scanning. The accuracy did not change across the runs. With
respect to the neuroimaging data, practice-effects were only ob-
served during WM encoding and retrieval, but not during mainte-
nance. The decrease of activation concerned mostly WM frontal
and parietal areas (see Table 2 for more details).
Like Landau et al. (2007), Sayala, Sala, and Courtney (2006) have
investigated the effect of practice on WM using spatial and object
instructions, but they utilized faces as stimuli. In the sample phase
a sequence of three stimuli was presented. After a memory delay of
3, 4.5 or 9 s, the test phase began; a face was displayed during 3 s.
In object trials, participants had to decide if the face was part of the
sample phase; in spatial trials, participants had to decide if a face
in the sample had been displayed in the same location. The behav-
ioral results showed no significant changes in terms of reaction
times or errors across the scanning. Moreover, no correlations be-
tween fMRI activity and behavioral measures were observed.
The fMRI results were analyzed separately for the sample, the
delay and the test phases. Across the three phases, and collapsing
data across object and spatial trials, a decrease of brain activation
was observed. The practice-related effect was more important dur-
ing the sample and delay phases. The main network of regions
common to the sample phase and the delay phase is congruent
with WM activities: it mainly comprised frontal and parietal areas,
plus the cingulate, and the insula. During the test phase, WM areas
also were elicited: the left inferior frontal gyrus showed a practice-
related decrease in activity (for more details see Table 2).
A decrease in activation in WM areas was also observed with a
more complex task, which is mainly supposed to tap into the plan-
ning function. With a modified, computerized version of the tower
of London, Beauchamp, Dagher, Aston, and Doyon (2003) wanted
to observe the effect of repeated practice, that is, solving problems
for 2 h. The training resulted in an improvement in behavioral
We consider a delay of weeks as a medium-length training, because the experts
that are usually compared to novices in standard experiments (e.g., in the studies we
reviewed in the previous section) go through much longer training periods, in the
order of years.
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 231
performances: a decrease in the time required for planning and
execution, as well as in the total time to solve the problems. Like
in the previous studies, a decrease in frontal regions was detected.
The authors observed a decrease mainly in the orbitofrontal and
frontopolar regions (for more details see Table 2), which are partic-
ularly important for guessing, decision-making and use of feedback
(Bechara, Damasio, & Damasio, 2000; Elliott et al., 1997; O’Doherty,
Kringelbach, Rolls, Hornak, & Andrews, 2001; Rogers et al., 1999).
The results also showed a positive correlation between activity in
these areas and the decrease in time taken to solve the problems.
Jansma, Ramsey, Slagter, and Kahn (2001) used a verbal Stern-
berg task. Participants had to remember a target set composed of
five consonants; then a consonant was presented and the partici-
pants had to decide whether it belonged to the target set; for each
target set, 10 consonants were presented. Participants were asked
to practice the task for 45 min and were then scanned for approx-
imately 40 min while performing the same task. Behavioral data
revealed that the answers became faster and more accurate. The
fMRI data showed a decrease in activity during the scanning,
mainly in areas related to WM: the bilateral dorsolateral prefrontal
cortex, the right superior frontal cortex, the right frontopolar area
and the supplementary motor area.
6.2. Scanning–Training–Scanning paradigm experiments with
medium-length training
We review three studies that used medium-length sessions: 4–
5 weeks. Using a spatial n-back task, Hempel et al. (2004) observed
first an increase in activation in fMRI signals and then a decrease in
WM areas. Participants were scanned over three periods: before
training, after 2 weeks of training and after 4 weeks of training. A
load factor was used by modifying the n-value of the n-back task
(0-back, 1-back, and 2-back). Participants trained independently
at home, twice a day. The fMRI results mainly showed an activation
involving the right inferior and medial frontal gyri, the right intra-
parietal sulcus and the right superior parietal lobe. The load factor
was associated with activation increases in these regions. The
authors suggested that the increase–decrease pattern of activa-
tions followed an inverse U-shaped quadratic function.
Olesen, Westerberg, and Klingberg (2004) also observed an in-
crease in brain activation in WM areas. They used a visuospatial
WM task, training their participants for 5 weeks during 35–45 min
of training per day. In their first experiment, brain activity was mea-
sured with fMRI twice, before and after training. In a second exper-
iment, subjects were scanned once a week during a 5-week training
period. Both experiments showed an increase in activity in the mid-
dle frontal gyrus and in the superior and inferior parietal cortices,
correlated with an increase in WM performance due to training.
The important point concerning training was that the difficulty
of the training tasks was automatically adjusted to performance by
changing the number of stimuli to be remembered. This may ex-
plain, in different ways, the peculiar pattern observed, that is, brain
increase activation. In fact, as pointed out by Kelly and Garavan
(2005, p. 1097), ‘‘a number of studies have demonstrated how in-
creased task difficulty or load is associated with increased BOLD
activation (Rypma & D’Esposito, 1999; see also e.g., Druzgal &
D’Esposito, 2003; Rypma, Berger, & D’Esposito, 2002).’’ The same
authors also proposed a second explanation. As discussed above,
Hempel et al. (2004) observed first an increase (0–2 weeks) and la-
ter a decrease (2–4 weeks) in activation in an n-back task. It is pos-
sible that Olesen et al.’s (2004) result could also be described by a
quadratic function, since the final scanning session activation ap-
peared to be decreasing (see Fig. 3dinOlesen et al., 2004).
Dahlin, Stigsdotter-Neely, Larsson, Bäckman, and Nyberg (2008)
used the same training procedure, adjusting the difficulty of the
training tasks to participants’ performance. However, they adapted
it to an updating task. One group of 15 participants was trained dur-
ing 5 weeks and compared with a group of 7 participants that did not
receive any kind of training. Letters were employed in the updating
task during fMRI acquisition. It is important to note that unlike all
the previous studies, the authors did not only analyze a simple effect
(pre-training vs. post-training), but also an interaction that incorpo-
rated the control group vs. training group contrast, comparing the
training group pre-training vs. post-training difference with the
control group pre-training vs. post-training difference.
Behavioral results showed that the training group exhibited
considerably larger gains in both tasks than the control group. Con-
cerning the fMRI data, an increase of activation was observed in the
temporal lobe, in the striatum, and in the occipital lobe. Conversely
a decrease was observed in the frontal and parietal lobe. These re-
sults could be compatible with a functional reorganization as de-
fined earlier (Kelly & Garavan, 2005; Poldrack, 2000). However
no behavioral indicators was available in the papers to confirm
that with the increase of activation of new areas there also was a
shift in the cognitive process underlying task performances. Finally
it is interesting to note that among the regions that increased, a
large cluster in the temporal lobe was found, which could be com-
patible with LT-WMT and TT predictions.
6.3. Training-Scanning paradigm experiments with medium-length
We finish this review with three studies (Andreasen, O’Leary,
Arndt, et al., 1995; Andreasen, O’Leary, Cizadlo, et al., 1995; Moore
et al., 2006) that also used medium-length training, but in a differ-
ent way compared with the studies we have just reviewed. In the
eight previous studies, the authors have used the S–T–S paradigm
comparing brain activation before training and after training. In
contrast, in the three following studies, participants were first
trained on a task and a type of material, and only after scanned
(that is, these studies used the T–S paradigm). In order to have a
contrast, the participants’ brain activation was measured in two
ways: (a) while the participants executed the task for which they
had been trained for, and with the same material used during
the training, the practiced condition, and (b) while the participants
executed the task for which they had been trained for, but with a
novel kind of material for which they had no expertise, the
(pseudo)novice condition.
In a PET study, Andreasen and colleagues used the T–S para-
digm to investigate the effect of practice with a narratives recall
task (Andreasen, O’Leary, Arndt, et al., 1995) and with a word span
memory task (Andreasen, O’Leary, Cizadlo, et al., 1995; in both
cases, a practiced condition was contrasted with a novice condi-
tion). In Andreasen, O’Leary, Arndt, et al. (1995), two stories (Story
A and Story B) were taken from the Wechsler Memory Scale-Re-
vised (Wechsler, 1987). Participants heard Story A and recalled it
continuously until they knew it perfectly; this was done 1 week
before the scanning and one day before the scanning. Finally, just
before the scanning, participants heard Story A again, then they
had to recall Story A during the scanning process. Conversely, Story
B was heard only once, just before the scanning, and then the par-
ticipants had to recall it during the scanning. Behavioral data
showed that participants recalled 99% of Story A and 56% of Story
B. The structure of Andreasen, O’Leary, Cizadlo, et al.’s (1995) study
was the same, except that they used two lists words (List A and List
B) from the Rey Auditory Verbal Learning task (Rey, 1964). Partic-
ipants scored a mean of 14.2 hits out of 15 words for List A and 6.6
hits for the 15 words of List B.
Concerning the PET signals, results from the two studies
showed that the practiced condition produced smaller and fewer
activations. When subtracting the practiced condition to the
unpracticed condition, Andreasen, O’Leary, Arndt, et al. (1995) ob-
232 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
served activation in the left frontal operculum, in the cingulate
gyrus, and in the cerebellum. In Andreasen, O’Leary, Cizadlo,
et al.’s (1995) study, the same subtraction showed an activation
of superior and inferior frontal areas, in the parietal region, and
in the cerebellum. Results, especially from the latter study, are
comparable with the main pattern observed in the studies using
the S–T–S paradigm: a brain decrease activation pattern mainly
in WM prefrontal and parietal areas, with no functional
Moore et al. (2006) have also employed the S–T paradigm, in a
match-to-sample task with complex visual objects (polygons).
Participants were first trained during 10.5 h with two categories
of objects. Half the participants were trained with one category
and the other half with the second category. Once the training
phase finished (an average of 10 days), participants were scanned
during a match-to-sample task. In each trial, a cue (a complex ob-
ject) was displayed followed by a delay period that ranged from
6.75 to 12.75 s. Finally a probe object was displayed for 1.25 s. Par-
ticipants had to indicate if the cue and the probe were the same.
Two types of objects were displayed: those that belonged to the
category for which the participants had become experts (the prac-
ticed condition) and those that belonged to the category for which
the participants were not trained (the novice condition).
Results showed that participants had a better accuracy in the
match-to-sample task in the practiced condition, that is, for the ob-
jects they were trained with. The fMRI results showed mostly an
increase of activation in WM frontal and parietal areas and in the
middle occipital gyrus (for more detail, see Table 2). These activa-
tions areas are comparable with the main activation pattern ob-
served in the studies using an S–T–S paradigm; however, in this
last study, no activation decrease was observed, but instead an in-
crease of activation.
6.4. Summary
In this section, we have reviewed twelve PET and fMRI studies
in which novices were trained in order to observe brain activation
changes during WM-related tasks (see Table 2). Among the studies
using the S–T–S paradigm, six showed a decrease in activation in
WM areas, one showed an initial increase followed, after 2 weeks,
by a decrease in activation (Hempel et al., 2004), one showed only
an increase of activation (Olesen et al., 2004), and one was
compatible with a functional reorganization (Dahlin et al., 2008).
Concerning the S–T studies, the two studies of Andreasen, O’Leary,
Arndt, et al. (1995) and Andreasen, O’Leary, Cizadlo, et al. (1995)
showed a decrease of activation, and Moore et al. (2006) observed
an increase mainly in WM areas. We stress the fact that these three
S–T studies cannot be directly compared to the studies using the
S–T–S paradigm.
From the target studies we have reviewed in this section, it first
appears that, with the exception of Dahlin et al.’s (2008) study, no
data consistent with functional reorganization has been detected
with WM-related tasks. Second, the main pattern observed is a
practice-related decrease of brain activation, that is, eight among
twelve. And third, the main areas engaged and showing a decrease
of activation are – as expected given the tasks employed, that is
WM – prefrontal and parietal areas. Table 2 presents all the studies
we have reviewed in this section with all the areas showing prac-
tice-related changes.
7. Bridging the two patterns of neuroimaging
7.1. A two-stage view of expertise acquisition
Based on the description of the state of the art we have provided
above, there seems to be a clear discrepancy between (a) the data
collected with novices who are trained to become experts, which
mainly show a decrease of cerebral activity in WM areas, and (b)
the data collected with experts who are compared to novices,
which are compatible with a cerebral functional reorganization
from WM areas to WM and LTM areas. We think that an important
factor explaining this discrepancy could be the duration of practice
and its content.
With regard to duration, several training periods in the former
studies (showing mainly a decrease in cerebral activity) are prob-
ably too short – experiments using short-training regimens, in or-
der of hours – to allow novices to build crucial cognitive structures
for expertise, that is, chunks but especially knowledge structures
(Ericsson and Kintsch’s retrieval structures and Gobet and Simon’s
templates). Among the six studies using a medium-length training,
which should be more favorable to functional reorganization, only
one exhibited this pattern.
With regard to the content of the training sessions in the former
studies (showing mainly a decrease in cerebral activity), and espe-
cially in the case of longer practice durations – experiments using
medium-length training regimens, in order of weeks – the training
does not seem enough oriented towards the acquisitions knowl-
edge structures. Concerning this last point, it is to be noted that
specific training programs (e.g., Ericsson & Harris, 1990; Gobet &
Jackson, 2002; Saariluoma & Laine, 2001) or suggestions of mne-
monics to participants (Chase & Ericsson, 1981, see Fig. 5.13) can
speed up the process of acquiring chunks and especially knowledge
structures. These last structures are crucial for expertise, because
they allow individuals to expand their WM, enabling them to use
part of their LTM as WM, which is theoretically supposed to cause
a cerebral functional reorganization.
Even if we argue that the duration of practice and its content are
crucial factors for explaining the difference between the data col-
lected with novices and the data collected with experts, we do
not deny the possibility that part of the difference between the
two patterns could be due to the type of task employed. In fact
when considering the tasks used in the neuroimaging studies of
trained novices, they appeared to be often abstract, unlike the tasks
used in the neuroimaging studies of experts. Some are constructed
so that it is very difficult to use any semantic knowledge structures
that may help out with doing the task. However, this is not the
case for all the studies. In fact, Landau et al. (2004) and Sayala
et al. (2006) used faces; Jansma et al. (2001) used letters; and
Andreasen, O’Leary, Arndt, et al. (1995) and Andreasen, O’Leary,
Cizadlo, et al. (1995) used words and sentences. That is why our
claim is that the length and the content of the training period are
the crucial factors explaining the discrepancy.
Our hypothesis is that the two brain-imaging patterns of results
observed in the literature are not contradictory but are two stages
of the same process: the acquisition of expertise. We propose that,
in terms of brain activation, expertise acquisition in WM-related
tasks is a two-stage process that starts with a decrease in activity
and ends with a functional brain reorganization.
As far as the first stage – brain decrease activation – is con-
cerned, the only three studies that are not congruent with this
view, among the twelve reviewed, are Hempel et al.’s (2004),
Olesen et al.’s (2004), and Moore et al.’s (2006). As argued above,
we considered the latter one with caution because of the paradigm
used (T–S), which makes delicate the comparison with the
majority of the studies that involve the S–T–S paradigm.
Kelly and Garavan give the following description of functional reorganization
(2005, p. 1089, italics added): ‘‘Three main patterns of practice-related activation
change can be distinguished. Practice may result in an increase or a decrease in
activation in the brain areas involved in task performance, or it may produce a
functional reorganization of brain activity which is a combined pattern of activation
increases and decreases across a number of brain areas.’’
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 233
Concerning Hempel et al.’s (2004), Olesen et al.’s (2004) studies,
they tend to show that the first stage of expertise could be an
increase in brain activation, which is compatible with a three-stage
process view of expertise: increase, decrease, and then brain func-
tional reorganization. Nevertheless, we do not propose a three-stage
view because, as we have already noted, Olesen et al.’s (2004) meth-
odology is peculiar and could be misleading. As for Hempel et al.’s
(2004) results, the increase in activity seems genuine but we think
that more results of this kind are necessary before we should take
into consideration an increase in activity with WM-related tasks.
In 2004, Jonides already hypothesized that the two patterns of
activity described above could be two different consequences of
practice. In his view, the decrease in activity could correspond to
greater skill at applying the initial strategy and the functional reor-
ganization could correspond to the development of new and more
efficient strategies. However, Jonides (2004) did not specify when
these patterns should appear during expertise acquisition, nor
whether they should follow a sequential order.
The interpretation of the review we have provided is that the
patterns occur in a sequence, with functional brain reorganization
occurring after activation decrease. However, we do not think that
this means that novices in the early stages of expertise acquisition
do not develop new and more efficient strategies, as suggested by
Jonides. And, indeed, in the only study (Olesen et al., 2004, exper-
iment 2) that reported the content of participants’ interviews
about their strategy, two out of eight participants said that they
had changed strategies. What we suggest, and what the data seem
to indicate, is that a simple change of strategy is insufficient to in-
duce functional brain reorganization in a context of practice: it
may be a necessary but not a sufficient condition.
From the data we reviewed in the ‘‘Neuroimaging of experts’’
section, it seems that functional reorganization corresponds to a
profound cognitive modification that occurs only at a late phase
of expertise acquisition when important cognitive structures
(chunks and especially knowledge structures) have been built
and become efficient – this point of view is consistent with what
Kelly and Garavan wrote in 2005, when they suggested that ‘‘prac-
tice on cognitive tasks...tends to be brief (hours) (in the case of
novice trained), which may carry the implication that the full ex-
tent of practice related changes has not yet been examined in a
cognitive task’’ (p. 1095) – making it possible for experts to use
part of LTM as WM. The consequence of this last process is a func-
tional reorganization.
7.2. Alternative explanation of the two patterns of neuroimaging
Before taking our explanation forward, it is important to
acknowledge an alternative explanation of the two different pat-
terns of neuroimaging results observed in the literature: the tasks
used in the Neuroimaging of Experts section are of a peculiar kind
compared to those used in the Neuroimaging of trained novices
section. In fact, in the studies reviewed in the Neuroimaging of Ex-
perts section, knowledge is of core importance and the tasks em-
ployed are typically adapted to the domain of knowledge in
order for the experts to use their knowledge. By contrast, in the
Neuroimaging of trained novices section, the study tasks are stan-
dard STM/WM one’s. In other words, they are very abstract tasks,
built so that the participants can use their knowledge only mini-
mally. This difference could have the consequence of making it
very difficult for the participants, when trained, to build up any rel-
evant knowledge structures, even after a very long period. This
could be why no functional reorganization involving LTM areas is
observed. If this explanation is correct, then there would be no
need to hypothesize that the two patterns occur in a succession
and that they are related to a same sequence. To rule out this pos-
sibility, we will first put forward our arguments to explain why this
explanation is not genuine and, secondly, we will explain why the
two stages operate in a sequence.
Concerning the first point, it is true that the tasks used with nov-
ices and the tasks employed with experts are of a different kind.
However, we feel that it is not this difference that explains the
two different neuroimaging patterns, for two reasons. Firstly, it is
not the case that, in the Neuroimaging of trained novices section,
all the study tasks are very abstract. For example, Jansma et al.
(2001) used letters, Landau et al. (2004) faces, and Andreasen,
O’Leary, Arndt, et al. (1995) short narratives, and nevertheless brain
decrease activation was observed with no functional reorganiza-
tion. Therefore, we do not think that the abstractness is a key factor.
Secondly, we believe that it is always possible to develop
semantic knowledge structures in WM related tasks. Let us take
a spatial 3-back task as prototype of a standard task. By practicing
this task, we believe that important information concerning the
elements to be remembered will be picked up and encoded in an
increasingly meaningful manner by the participants. Positions that
share features and that are included in a sequence of 3 positions
(the task is 3-back task) can potentially be chunked. Moreover,
the fact that positions can share features allows categorization
with high hierarchical elements containing low hierarchical ele-
ments. For example, one can imagine the following category of
positions: one element from a corner and two elements from
external rows. This corresponds rather directly to what Gobet
and Simon (1996a) call a template in chess (see Section 3.1). All se-
quences of 3 positions that are consistent with this category are
exemplars of this template. Therefore, as we wrote in Section 3.1
to characterize templates, ‘‘this means that a player will categorize
different but related positions with the same template.’’ Of course,
structure needs time and practice to be cognitively extracted. And
this precisely takes us to the second point, the succession of the
two stages.
Alternatives to our sequential two-stage view of the two pat-
terns of neuroimaging exist. For example, the two stages could
concern two different processes operating in different situations
and perhaps simultaneously. This could lead to a parallel view of
the two stages. However, we do not subscribe to this point of view.
First, this is because we do not think that the stages constitute two
different processes. In a way, this is what the structure of the re-
search and of the literature would make us think a priori. In fact,
there is almost not link in the neuroimaging literature between
the research domain where experts are studied and the research
domain where novices are trained and studied. However, looking
at these two separate domains through the prisms of LT-WMT
and TT allows one to bridge the two patterns in a global dynamic.
Concerning the view that the two stages could be two different
processes, a possibility could be that the kind of tasks employed
were different. However, we think that we have shown that the
tasks used are not so different and that there is always the possi-
bility to build knowledge and expertise. If the two stages are not
due to two different processes, we see no reason to hypothesize
a parallel view. We have suggested that they operate in sequence
because templates and retrieval structure (see Section 3) are highly
hierarchical structures that can incorporate chunks. They consti-
tute a phase of expertise that is superior to the use of simple
chunks. And acquiring expertise takes time. It is almost impossible
to imagine retrieval structures to be built in few hours, while
chunks can be built rapidly (e.g., Cowan, Chen, & Rouder, 2004).
This is confirmed in terms of neuroimagery since no functional
reorganization has been observed in less than 5 weeks. That is
why we hypothesize a first stage where only chunks are build
and used, which necessitates only a low level of expertise, and a
second stage that occurs only when expertise has become stronger,
which allows building templates or retrieval structures (see for a
summary of the chronology of the two-stage view, Section 10).
234 A. Guida et al. / Brain and Cognition 79 (2012) 221–244
In the next sections, we will describe and try to give an expla-
nation for each expertise acquisition stage. As it was the aim of this
article, we will attempt to show for both stages that chunks, retrie-
val structures and templates provide a legitimate explanation for
neuroimaging data on expertise acquisition.
8. The first stage in expertise acquisition: brain decrease
8.1. Explaining decreases in brain activation: from chunk creation to
chunk retrieval
Among the reviewed authors who observed a decrease of brain
activation, two explain the decrease via the acquisition of
knowledge, appealing to the chunking theory (Jansma et al.,
2001; Landau et al., 2004). We completely agree and think this
theory offers a good explanation of the decrease in activation in
WM-related tasks. We have introduced the chunking mechanism
at the beginning of the article. We would like to go further
by separating two important chunking mechanisms: (a) chunk
creation – the mechanism that involves binding different elements
in WM into a chunk that does not yet exist in LTM; and (b) chunk
retrieval – chunking mechanism that involves the perception of a
chunk in WM that already exists in LTM.
Concerning the second type of chunking mechanism, it is the
one that we described at the beginning of the article; if one has
to remember the letters ‘‘fbiibm’’ and if one knows the acronyms
‘‘FBI’’ and ‘‘IBM,’’ then it is possible to chunk the information in
WM into two elements ‘‘FBI’’ and ‘‘IBM,’’ because these chunks al-
ready exist in LTM.
Chunk creation contrasts with chunk retrieval. It mainly
occurs when individuals do not have a strong knowledge of
the information they are processing, for example in the early
stages of expertise acquisition; however, it can also take place
when experts are processing new information. For example, if
the letters ‘‘wdg’’ are presented once, and if ‘‘wd’’ or ‘‘dg’’ or
‘‘wdg’’ do not correspond to something known, one will certainly
use three chunks in WM to encode the information: w – d – g.
But if the same information is presented again, one will be able
to chunk more efficiently w – d – g together in WM, and if the
information is displayed several times, then it is almost sure that
this chunk will become an LTM chunk. At this point, if the
letters ‘‘wdg’’ are presented again, one will be able to directly
retrieve ‘‘wdg’’ from LTM as one chunk and therefore use only
one chunk in WM to encode the information. This means that,
over several presentations, a chunk that has been created can
become an LTM chunk, and that the process has shifted from
chunk creation to chunk retrieval.
Since novices are not experts concerning the information they
are processing, the main chunking mechanism that occurs during
the early stages of expertise acquisition is chunk creation. We
believe that this mechanism relies strongly on WM, thus involv-
ing prefrontal WM areas (e.g., D’Esposito et al., 1995; Postle
et al., 1999; Prabhakaran et al., 2000; Ungerleider et al., 1998)
and parietal WM areas (e.g., Dehaene & Cohen, 1994; Postle &
D’Esposito, 1999; Rodriguez et al., 1999; Todd & Marois, 2004;
for the importance of parietal areas in WM, see Cowan, 2001,
2005). As the number of chunks created grows with practice,
there is a gradual shift towards chunk retrieval since increas-
ingly more chunks (in LTM) become available. We believe that
this gradual shift co-occurs with a decrease of activation in pre-
frontal and parietal areas as less chunk creation is needed. To be
more specific, and incidentally to link the chunk creation and
chunk retrieval mechanisms to other established mechanisms,
we put forward two reasons that can explain why the decrease
of activity in the prefrontal and parietal areas can be linked with
the decrease of chunk creation and the increase of chunk
The first reason relates to the binding mechanism. Chunk crea-
tion clearly depends on binding, which can be defined as the pro-
cess of encoding the relations among stimuli that co-occur (Cohen
& Eichenbaum, 1993). It is through the binding process that the
elements in WM can be bound and form a chunk. The binding pro-
cess is thought to occur in WM and more precisely in the focus of
attention (Oakes, Ross-Sheehy, & Luck, 2006; Wheeler & Treisman,
Several theoretical models emphasize this point. Cowan and
Chen (2008) have indicated that one of the functions of the focus
of attention in Cowan’s WM model (Cowan, 1999, 2001, 2005)is
to allow items represented concurrently to be bound into a new
structure. The same idea can also be applied to the region of
direct access (that is roughly compatible with the focus of atten-
tion in Cowan’s model) in Oberauer’s WM model (Oberauer,
2002; Oberauer & Lange, 2009). And in the last version of
Baddeley’s model (Baddeley, 2000; Baddeley & Wilson, 2002),
the binding process is said to occur in the episodic buffer
(Baddeley, 2001), a WM buffer. Concerning physiology, several
studies suggest that binding occurs in prefrontal regions
(Prabhakaran et al., 2000; Raffone & Wolters, 2001) and in parietal
regions (Oakes et al., 2006; Robertson, Treisman, Friedman-Hill,
& Grabowecky, 1997; Shafritz, Gore, & Marois, 2002; Todd &
Marois, 2004).
When all these results are taken together, it is easy to under-
stand that, as more LTM chunks become available (through chunk
retrieval) during the process of acquiring expertise, the necessity
and opportunity to create chunks by binding separate elements be-
comes quantitatively less important, and therefore the activity of
regions (prefrontal and parietal areas) that undergird the process
of binding (and thus chunk creation) decreases.
The second reason for explaining the decrease of activity in the
prefrontal and parietal areas related to practice (observed in the
studies reviewed in the ‘‘Neuroimaging of Trained Novice’’
section) is the variation of the size of chunks in WM. In fact,
whereas the number of chunks in WM is not supposed to vary
with expertise (e.g., Gobet, 1998b; Lane, Gobet, & Cheng, 2001;
or just slightly, Gobet & Simon, 2000) or with practice (the
constant-capacity hypothesis, see for data, Chen & Cowan, 2005;
Cowan et al., 2004; Tulving & Patkau, 1962; see for a review,
Cowan, 2005), the chunks in WM (and therefore in LTM) can
become larger with practice (Chen & Cowan, 2005; Cowan
et al., 2004) and expertise (e.g., Chase & Simon, 1973a; Gobet &
Simon, 1996a, 1996c). Therefore, if one can represent the same
amount of information in WM with larger chunks, then the
number of chunks (or the percentage of chunks)
in WM
necessary to represent this amount of information decreases. This
could easily explain why the brain activity in WM regions
decreases. In this kind of interpretation, one supposes a direct link
between the pattern of decrease of activation and the decrease in
the number of chunks needed. This is for example consistent with
Vogel and Machizawa (2004) data. Using event-related potentials,
these authors showed that, if the number of chunks necessary to
represent visual information in WM diminishes, then the amplitude
of an event-related waveform in posterior parietal areas and in
occipital areas also decreases.
A possibility (compatible with Cowan and colleagues’ data, Chen & Cowan, 2005;
Cowan et al., 2004) when chunks become larger is that instead of using a smaller
number of chunks to represent the same amount of information, the number of
chunks used stays stable. What can vary (i.e., decreases) is the percentage of the same
number of chunks used or the amount of activation in order to activate the same
number of chunks.
A. Guida et al. / Brain and Cognition 79 (2012) 221–244 235
Even if we believe that, in the activation decreasing pattern
(in prefrontal and parietal areas) observed in the studies we
reviewed, there is a decreasing activation that directly concerns
the number of chunks, it is also likely that the link between the
decrease of the number of chunks and the decrease of activation
is done through other factors, such as cognitive resources. In fact,
as the number of chunks in WM needed to represent the processed
information in a task diminishes (because the chunks get bigger),
the cognitive resources needed to execute the task also decrease;
this could also explain part of the decreasing pattern in prefrontal
and parietal areas.
8.2. Explaining decreases in brain activation: from chunk retrieval to
knowledge structures
As mentioned above, chunks in WM (and therefore in LTM) get
larger with practice and expertise. TT clearly states it: the chunks
get also richer because more LTM knowledge is associated with
each one of them. Moreover, several LTM chunks can become
linked to knowledge. And eventually, if an individual becomes an
expert, the presence of these links between several chunks can
result in the creation of high-level hierarchical chunks (i.e.,
Ericsson and Kintsch’s retrieval structures and Gobet and Simon’s
Table 2
Practice-related brain activation decreases in novices during working-memory-related tasks key (ranked in orde