Brain localization of memory chunks in chessplayers.

Abstract

Chess experts store domain-specific representations in their long-term memory; due to the activation of such representations, they perform with high accuracy in tasks that require the maintenance of previously seen information. Chunk-based theories of expertise (chunking theory: Chase & Simon, 1973; template theory: Gobet & Simon, 1996) state that expertise is acquired mainly by the acquisition and storage in long-term memory of familiar chunks that allow quick recognition. This study tested some predictions of these theories by using fMRI while chessplayers performed a recognition memory task. These theories predict that chessplayers access long-term memory chunks of domain-specific information, which are presumably stored in the temporal lobes. It was also predicted that the recognition memory tasks would activate working memory areas in the frontal and parietal lobes. These predictions were supported by the data.

Full text

Page 1

1
Brain Localisation of Memory Chunks in Chessplayers

Guillermo Campitelli
Fernand Gobet
Brunel University

Kay Head
University of Nottingham

Mark Buckley
University of Oxford

Amanda Parker
University of Newcastle


Correspondence should be addressed to:
Fernand Gobet
Centre for Cognition and Neuroimaging, Brunel University, West London, Uxbridge,
Middlesex, UB8 3PH, United Kingdom
Tel: ++ 44 1895 265484, e-mail: Fernand.Gobet@brunel.ac.uk






Running head: Brain Localisation of Chunks

Page 2

2

Abstract
Chess experts store domain-specific representations in their long-term
memory; due to the activation of such representations, they perform with high
accuracy in tasks that require the maintenance of previously seen information. Chunk-
based theories of expertise (chunking theory: Chase & Simon, 1973; template theory:
Gobet & Simon, 1996) state that expertise is acquired mainly by the acquisition and
storage in long-term memory of familiar chunks that allow quick recognition. We
tested some predictions of these theories by using fMRI while chessplayers performed
a recognition memory task. These theories predict that chessplayers access long-term
memory chunks of domain-specific information, which are presumably stored in the
temporal lobes. We also predicted that the recognition memory tasks would activate
working memory areas in the frontal and parietal lobes. These predictions were
supported by the data.








Acknowledgements
We would like to thank Chris Vincent and Gareth Williams for their collaboration.

Page 3

3
Brain Localisation of Memory Chunks in Chessplayers
Chase and Simon’s (1973) seminal study of cognitive processes in
chessplayers has had a strong impact in cognitive psychology. Their research spawned
several studies in which chess was used successfully as a research tool for studying
cognitive processes such as perception, memory and decision making (see Charness,
1992; Gobet, De Voogt, & Retschitzki, 2004, and Saariluoma, 1995 for reviews).
One of the paradigmatic results obtained in this line of research is that chess
experts are able to reconstruct with high accuracy a game position that had been
shown for only a few seconds (Chase & Simon, 1973; Gobet & Simon, 2000).
However, when the pieces are randomly placed throughout the board, experts perform
only slightly better than novices (Gobet & Simon, 2000). Chunk-based theories of
expertise (chunking theory: Chase & Simon, 1973; template theory: Gobet & Simon,
1996) account for these results by proposing that, during study and practice, experts
store domain-specific “chunks” (perceptual patterns that can be used as units of
meaning) in their long-term memory. When this practice becomes serious and
continuous, some of the chunks evolve into more complex structures called
“templates,” consisting of core information supplemented with slots in which more
information (perceptual or abstract) can be added (Gobet & Simon, 1996). Evidence
for chunks and templates stored in long-term memory as the building blocks of chess
expertise originates from behavioural studies (Chase & Simon, 1973; Gobet & Simon,
1996) and computational simulations (Gobet & Simon, 2000). Chunk-based theories
of expertise are generalizable to other domains of expertise such as computer
programming, medical diagnosis and engineering (Gobet et al., 20001; Simon &
Gobet, 2000).

Page 4

4
Since chess has proven to be successful as a research tool to study cognitive
processes, it is surprising that only a handful of studies have been done in cognitive
neuroimaging with this game. Nichelli et al. (1994) asked chessplayers to perform
simple chess tasks while brain activity was monitored with a PET scanner; Onofrj
(1995) performed a SPECT study in which they asked players to solve a chess
problem; Amizdic, Riehle, Fehr, Wienbruch and Elbert (2001) asked participants,
ranging from strong grandmasters to class B players (see Methods for an explanation
of chess skill levels), to play a game against a computer while they were scanned
using a gamma burst technique; and Atherton, Zhuang, Bart, Hu and Sheng (2003)
carried out an fMRI study in which players had to solve a chess problem. The results
of these studies suggest that, when players had to solve a problem or play a game,
they activated frontal and parietal areas. Finally, Campitelli, Gobet and Parker (2005)
used fMRI in order to compare chessplayers with non-chessplayers in a memory task
using chess positions and displays made of geometrical shapes.
In the present study, we used chess as a tool to investigate the neural
substrates of working and long-term memory, and, more specifically, the localisation
of domain-specific chunks in long-term memory. We did this by performing two
types of comparisons. First, we compared the brain activity of chessplayers
performing a recognition memory task with chess stimuli to that of the same players
performing a similar task with unfamiliar non-chess stimuli (in order to identify long-
term domain-specific chunks). Second, we compared the brain activity of players
performing a recognition memory task with chess stimuli to that of the same players
performing a perceptual-motor control task with the same chess stimuli (in order to
identify brain activity related to working memory). It is well established (Britton &
Tesser, 1982; Chase & Simon, 1973; Gobet, 1998) that chessplayers automatically

Page 5

5
activate chunks in long-term memory when perceiving familiar positions. Therefore,
the added value of using chessplayers to study memory is that we can be confident
that the players are using their long-term memory chunks, whereas in normal studies
the use of long-term memory or working memory is assumed by the researcher as a
function of the task used. For example, chessplayers can perform a so-called “working
memory task” such as a delayed-match-to-sample task and a so-called “long-term
memory task” such as a recall task, and in both they would access chunks that have
been stored for a long time (i.e., long-term memory chunks). This is well established
because their performance diminishes when non-domain-specific stimuli are used,
regardless of the task. To our knowledge, this is the first study that investigates the
brain localisation of long-term memory chunks in experts.
Previous studies both in cognitive neuroimaging in humans and single-cell
recording in non-human primates have investigated the brain areas involved in
working memory and long-term memory. There is evidence supporting the hypothesis
that the ventral areas of the temporal lobe (fusiform gyrus and parahippocampal
gyrus) are involved in the storage of familiar patterns in long-term memory
(Desimone, Albright, Gross, & Bruce, 1984; Gross, 1992; Logothetis, Pauls &
Poggio, 1995; Stark & Squire, 2000; Tanaka, 1993) whilst substantial evidence
supports the role of the dorso-lateral prefrontal cortex and parietal lobes in working
memory (Fuster, 1998, 2000; Goldman-Rakic, 1998). (But see Cabeza & Nyberg,
2000, for studies relating the dorso-lateral prefrontal cortex and parietal lobes to other
memory processes, including episodic memory processes such as storage and
retrieval.)

Page 6

6
We submitted 5 skilled chessplayers to a recognition memory task while brain
images were obtained using fMRI.1 We hypothesized that players would access their
long-term memory chunks automatically when chess stimuli were presented, and that
this would activate long-term memory areas in the fusiform gyrus and
parahippocampal gyrus.
Methods
Participants
Five right-handed healthy chessplayers with normal vision participated in the
experiment. Informed consent was obtained from each subject. The mean age of the
participants was 24.6 years (sd = 8.0). The players were recruited from
Nottinghamshire chess clubs and from the University of Nottingham chess club. The
mean in Elo rating points (Elo, 1978) was 1971 (range 1750-2200). The Elo system
rates chessplayers from intermediate to world-champion level. Players with more than
2200 points are considered masters; players between 2000 and 2200 are classified as
candidate masters; players between 1800 and 2000 are considered class A level
players; chessplayers between 1600 and 1800 are considered class B level players,
etc. The World Chess Federation awards players with titles; usually players with more
than 2400 are international masters and players with more than 2500 are
grandmasters.
Originally, seven players took part in the experiment; the data of two of the
players were discarded because of failure to fulfil the motion criterion (see below).
Ethical approval was obtained from the School of Psychology, University of

1
We did not include a control group made of non-chess players, as their
lack of knowledge of the game implies that they do not have chess chunks,
which would make comparisons with chess players difficult to interpret.

Page 7

7
Nottingham ethics committee. Regulations of the Sir Peter Mansfield Magnetic
Resonance Centre, University of Nottingham, were followed.
Task and stimuli
All the blocks of all the conditions had the same structure (see Figure 1). Each
block started with a fixation cross which remained on screen for 12.5 s. Following the
fixation cross, a sample stimulus was presented for 6.5 s. After a 2-s delay, a series of
7 test stimuli of the same type appeared sequentially (i.e., if the sample stimulus was a
scene, all the test stimuli were scenes). Each of the seven test stimuli was presented
for 3 s with an inter-stimulus interval of 1 s. The task consisted of deciding whether
the test stimuli matched the sample stimulus or not, by pressing the left and right
buttons, respectively. The participants had to respond within the 3 seconds during
which the test stimuli were on the screen.

INSERT FIGURE 1 ABOUT HERE
INSERT FIGURE 2 ABOUT HERE

There were 4 conditions: game, random, scene and dot (see Figure 2). In the
first three conditions, the task was the one explained above, the only difference
between the conditions being the type of stimuli. The dot condition was a perceptual-
motor control for the game condition. In the dot condition the block structure was the
same as that of the other conditions, but the task was different. The sample stimulus
consisted of a chess position with a black dot in one of the middle squares of the
board. The test stimuli were also chess positions; some of them had a black dot and
some others did not. The task consisted of pressing the left button of the button box if

Page 8

8
there was a black dot and the right button if there was no black dot. Note that we
deliberately chose simple tasks so that the conditions are of similar difficulty.
In the game and dot conditions, the stimuli were the right half of a grey-scale
chessboard (4 x 8 squares) with black and white chess pieces forming a pattern
resembling that of a game position. Half boards, instead of full boards, were used in
order to restrict the number of eye movements. Saariluoma (1994) has shown that
players can recognize chunks within a quadrant of the board even when the location
of the pieces on the adjacent quadrants has been randomised; thus, it is unlikely that
the use of half boards negatively affected the access to LTM chunks. In the random
condition, the pieces were haphazardly distributed throughout the board. All the
positions in these conditions contained 16 pieces (8 ± 1 white and 8 ± 1 black). In the
scene condition, the stimuli were a grey-scale background with ellipses and different
types of black and white shapes (2 triangles, 2 squares, 2 rhombuses, and 2 circles).
Forty-four blocks (11 blocks of each condition) were pseudo-randomly
presented. The number of test stimuli matching the sample within each block varied
from 2 to 4, with an average of 3. The test stimuli that did not match the sample had
the same number of pieces (or shapes) as the sample stimulus, the only difference
being their location. Twelve stimuli were used for each condition, each taking the role
of sample stimulus only once. The chess positions consisted of middle game positions
with familiar configurations of pieces.
fMRI imaging parameters and analyses
The experiment was carried out in a 3-Tesla scanner with a TEM Nova
medical head-coil at the University of Nottingham Sir Peter Mansfield Magnetic
Resonance Centre. The stimuli had a vertical visual angle of 16º and a horizontal
visual angle of 8º. Participants wore prism glasses in order to see the stimuli.

Page 9

9
Images of the whole brain were obtained. Twenty-two coronal slices were
obtained at a rate of 136 ms each, hence the TR was 2,992 ms. The images were T2*
weighted Echo-Planar images (EPIs). The size of the images was 64 x 64 voxels. The
voxel size was 3 mm x 3 mm in-plane, and the slice thickness was 9 mm. The
experimental paradigm started 12 s after the scanner started recording, in order to
allow for magnetic saturation effects. The 4 volumes obtained during these 12 s were
discarded.
The processing of the data was carried out with Statistical Parametric Mapping
(SPM99) developed by Friston et al. (1995). The software Talairach Daemon
(Lancaster, Summerln, Rainey, Freitas, & Fox, 1997) was used to obtain Brodmann
areas given Talairach coordinates as input. Realignment, spatial normalization and
spatial smoothing were performed in SPM99. The data from subjects who translated
their heads more than 5 mm in any direction or rotated their heads more than 5º in any
of the axes were discarded from further analysis. For the spatial smoothing, an 8 mm
x 8 mm x 8 mm Gaussian kernel was utilized. A high-pass temporal filter of 131 s.
was used and the hemodynamic response function was chosen as a low-pass temporal
filter. Since the accuracy rate was above 90% we used all the epochs for the analysis.
We performed a fixed-effects analysis. A box-car function convolved with a
hemodynamic response function was used to model the data. Three contrasts were
planned: game > dot, game > scene and game > random. SPM maps of t values were
obtained after correction for multiple comparisons, showing p < .05. Only the clusters
of more than 5 voxels were reported.
Game was the key condition, and the other three conditions controlled for
different aspects of the task in the game condition. In the principal task, participants
were supposed to perceive the stimulus; recognise it as a chess pattern; encode it;

Page 10

10
maintain a representation of it in memory during the delay and during the presentation
of the test stimuli; match the test stimuli to this representation; decide whether they
are the same or not; and finally press a button. In the dot condition, participants were
required to carry out only the perceptual and motor aspects of the task explained
above; therefore, the subtraction game > dot captures the working-memory
components of the task. In the scene condition, participants were required to carry out
the same processes as in the game condition, but the type of stimuli differed. The
subtraction of brain activity in the scene condition from that in the game condition
would show the brain activity due to access to long-term memory areas. The random
condition included the same type of task and stimuli as the game condition, but the
meaningfulness and the typicality of the stimuli varied. Thus, the contrast game >
random afforded us the possibility of investigating the brain location of familiar
perceptual patterns. Because we had specific theoretical predictions, we limited the
analysis to uni-directional contrasts; this also had the advantage of keeping the
number of comparisons reasonably low.
Results
Behavioural data
The players performed very accurately (above 90% correct). The mean
percentage correct and standard error (±) were: 97.1 ± 2.6. The mean percentage
correct and the standard error for the four conditions were the following: game, 94.4 ±
1.3; random, 98.4 ± 0.8; scene, 97.6 ± 0.8; and dot, 98.2 ± 0.9. We carried out a one-
way ANOVA for related samples and there was a significant effect (F(3,12) = 4.25; p
< .05). The only significant post-hoc difference was that between game and random.
The overall mean reaction time (in milliseconds) and standard error were the
following: 1164 ± 40. The mean reaction time and standard error for the four

Page 11

11
conditions were the following: game, 1274 ± 80; random, 1202 ± 77; scene, 1149 ±
69; and dot, 1031 ± 75. A one-way ANOVA for related samples showed a significant
effect (F(3,12) = 12.1; p < .001). The post-hoc comparisons indicated that players
were faster in the dot condition than in both the game and random conditions, and that
they were slower in the game condition than in the scene condition.
The effect of condition in accuracy, although significant, is quite small (a
difference of only 4% between the game and random conditions). However, the
differences in reaction time were considerable. It is not surprising that the dot
condition was performed faster because it was computationally simpler than the other
conditions. The difference between the game and scene conditions may be explained
by the fact that players may automatically think of moves in the chess positions,
moves that are not relevant for the task at hand, which may have slowed down the
response. In support for this hypothesis, De Groot and Gobet (1996) report that, in a
memory recall task, players often mentioned possible moves and plans, although the
task was specifically presented as a memory task. In addition, Britton and Tesser
(1982) have shown that engagemenf of chess knowledge slows down performance in
a simultaneous reaction task.
Brain imaging data
Statistical parametrical maps were obtained of the contrasts game > dot, game
> scene, and game > random. The first contrast assessed differences in working
memory processes and the last two contrasts showed the activations due to different
aspects of long-term memory.
Contrast game > dot: Working Memory.
The contrast game > dot (see Figure 3 and Table 1) reveals the brain activity
due to the working-memory aspects of the game condition. A strong prediction of

Page 12

12
chunk-based theories for this contrast is that it would activate working memory areas
(i.e., dorso-lateral prefrontal and parietal areas) but not temporal areas. The access to
domain-specific long-term memory patterns in the temporal lobe would occur in the
game but also in the dot condition, due to the automatic character of pattern
recognition (Britton & Tesser, 1982; Chase & Simon, 1973; Gobet, 1998); therefore,
no activation should be found in the temporal lobe. However, the game condition
requires the maintenance of the sample stimulus that would cause the activation of
working-memory areas. The results support this prediction by showing the main
activation in the dorso-lateral prefrontal and parietal lobes and little activation in the
temporal lobes.

INSERT FIGURE 3 ABOUT HERE
INSERT TABLE 1 ABOUT HERE

There were 576 voxels activated in the players. Activations were found in
BA37 (left occipital and bilateral temporal fusiform gyri), BA19 (bilateral superior
occipital gyri), and BA7 (right superior parietal lobule and precunei). In the prefrontal
lobes the activations were in BA46 (left middle frontal gyrus), BA45 (right middle
frontal gyrus) and BA9 (inferior frontal gyrus). (In all cases, p < .05 with correction
for multiple comparisons.)
Contrast game > scene: Long-term memory.
The contrast game > scene (see Table 2 and Figure 4) reveals the activations
due to chess-specific features (i.e., chess pieces and board, and typical patterns of
pieces). Chunk-based theories predict activation in areas that store long-term memory
patterns (i.e., fusiform gyrus and parahippocampal gyrus) in chessplayers (because the

Page 13

13
representations of the typical chess configurations will be activated) but not in
working-memory areas (because in both conditions working-memory areas would be
activated). Following our prediction, players showed activity in temporal lobe areas
including the fusiform gyrus, the parahippocampal gyrus, and inferior temporal gyrus.
On the other hand, they also showed unexpected activations in the precunei, the
posterior cingulate, and the supramarginal gyrus.

INSERT FIGURE 4 ABOUT HERE
INSERT TABLE 2 ABOUT HERE
Contrast game > random: Long-term memory.
In the contrast game > random, chunk-based theories predict activity in areas
that store long-term memory patterns (hence, in fusiform and parahippocampal gyri),
but to a lesser extent than the contrast game > scene. This is because in the present
contrast, only the typicality of the positions differs between conditions. We found
activation only in BA37 (left parahippocampal gyrus, -28 -44 -7, t = 5.31).
Discussion
We investigated the neural substrates of working and long-term memory in a
recognition memory task with chessplayers. The rationale of our study was that, given
that chessplayers possess domain-specific chunks stored in long-term memory, the
comparison of brain activity between a condition with chess stimuli and another
condition with non-chess stimuli (but the same task), would be in the area where these
chunks are stored. Following chunk-based theories of expertise and previous
neuroscientific literature, we proposed that chessplayers would activate long-term
memory chunks in the temporal lobe (i.e., fusiform and parahippocampal gyri) when
presented with chess stimuli. Moreover, following the literature showing the

Page 14

14
engagement of the dorso-lateral prefrontal and parietal lobes in working memory
tasks, we expected activation in those areas in the conditions with a memory task. In
terms of the contrasts of interest, we predicted brain activity in the fusiform and
parahippocampal gyri in the game > scene contrast and in the game > random contrast
(though, to a lesser degree in the latter), as well as activity in the dorso-lateral
prefrontal, ventro-lateral prefrontal and parietal (but not temporal) lobes in the game >
dot contrast.
The results supported our hypotheses. The game > dot contrast showed brain
activity in the dorso-lateral prefrontal and parietal lobes and very little activation in
the temporal lobes. Furthermore, the game > scene contrast showed activation in the
temporal lobe and the game > random contrast showed a small cluster of brain
activity in the temporal lobe. However, some results were not expected. First, given
the large differences in performance found in memory recall tasks between the game
and random conditions (e.g., Gobet & Simon, 2000), we expected more activation in
the game > random contrast. However, we have to remember that the differences in
performance between the game and random conditions were typically observed in
recall tasks, whereas our experiment used a recognition task. In fact, previous
experiments have shown that differences are much smaller in recognition tasks,
Saariluoma (1984) finding a difference of 10% in favour of game positions and
Goldin (1979) even finding a 2% advantage with random positions. Moreover, it is
well known from the literature on verbal memory that, in recognition tasks, better
performance is obtained with atypical words than with typical words (Kintsch, 1970).
A natural continuation of our study would be to use the traditional recall task,
although this may be difficult in a brain-imaging setting due to the necessity of

Page 15

15
moving the computer mouse for placing pieces on the board, which requires
numerous finger and hand movements.
Second, in the game > scene contrast, we found unexpected bilateral activation
in the precunei and posterior cingulate, as well as activation in the right supramarginal
gyrus. We suggest the following explanation. It is well known that there exist two
visual pathways in the cortex: both start in the occipital lobe and then divide in the
“where” dorsal pathway in the parietal lobe and the “what” ventral pathway in the
temporal lobe (Ungerleider & Mishkin, 1982). We hypothesize that all conditions
activated these pathways because of the visual nature of the task, and that they also
activated motor areas due to the necessary motor response. Since all these areas were
common for all the conditions, no activity was expected in these areas in the
contrasts. We also hypothesize that in the game and the random conditions the
activation of domain-specific long-term memory chunks in the temporal lobes was
followed by the activation of the working memory areas of the dorso-lateral prefrontal
and parietal lobes that, in turn, generated a feedback activation to the temporal lobes.
This is similar to a proposal by Fuster (1998, 2000), who explained working memory
by the feedback activation from the prefrontal lobes to the temporal lobes. This would
allow a link to be generated between the domain-specific chunks already stored in
long-term memory and the information held in working memory. On the other hand,
we hypothesize that in the scene condition players activated the long-term memory
patterns related to geometrical figures but they did not need to make a link between
these patterns and areas involved in working memory; hence, the feedback
mechanism just described did not take place in this case. This explanation accounts
for the activation of parietal areas in the game > scene contrast, because in the game