This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and

Transcription

1 This article appeared in a journal published by Elsevier. The attached copy is furnished to the author for internal non-commercial research and education use, including for instruction at the authors institution and sharing with colleagues. Other uses, including reproduction and distribution, or selling or licensing copies, or posting to personal, institutional or third party websites are prohibited. In most cases authors are permitted to post their version of the article (e.g. in Word or Tex form) to their personal website or institutional repository. Authors requiring further information regarding Elsevier s archiving and manuscript policies are encouraged to visit:

2 NeuroImage 47 (2009) Contents lists available at ScienceDirect NeuroImage journal homepage: www. elsevier. com/ locate/ yni mg The neural representation of extensively trained ordered sequences Filip Van Opstal a,, Wim Fias a,b, Philippe Peigneux c, Tom Verguts a a Department of Experimental Psychology, Ghent University, H. Dunantlaan 2, B-9000 Ghent, Belgium b Ghent Institute for Functional and Metabolic Imaging, B-9000 Ghent, Belgium c UR2NF - Neuropsychology and Functional Neuroimaging Research Unit, Université Libre de Bruxelles, B-1050 Bruxelles, Belgium a r t i c l e i n f o abstract Article history: Received 20 October 2008 Revised 2 April 2009 Accepted 8 April 2009 Available online 17 April 2009 Keywords: Ordinal sequences Transitive inference Intraparietal sulcus Inferior frontal gyrus The role of the intraparietal sulcus (IPS) in number processing is largely agreed on. A current debate however concerns the specificity of the involvement of the IPS in representing numbers or ordinal sequences more generally. To test this specificity, we investigated whether the IPS would be activated by extensive training on an arbitrary ordered sequence. We found that the hippocampal-angular gyrus activation initially involved in learning the ordered sequences extends with extensive training to the left inferior frontal gyrus (left IFG), but not to the IPS. These results suggest that left IFG can be involved in processing ordinal information, and that there is no need for an IPS area specifically dedicated to the representation of all ordinal sequences. Instead, we propose that the locus of the representation might be determined by the nature of the stimuli rather than its ordinal nature per se Elsevier Inc. All rights reserved. The neural basis of numerical cognition has been the subject of much research. Besides a large agreement on the involvement of the intraparietal sulcus (IPS) in number processing (Dehaene et al., 2003; Cohen Kadosh et al., 2008), the specificity of the IPS to numbers is debated. Close links between space, time, and quantity in behavioral (Church and Meck, 1984) and neurological studies (Zorzi et al., 2002) have raised the suggestion that there exists a generalized magnitude system rather than a specific number system in the IPS (Walsh, 2003). Although there is some evidence for a common encoding of physical dimensions and numbers (Fias et al., 2003; Cohen Kadosh et al., 2005), a recent review proposed that both shared and distinct representations could exist (Cohen Kadosh et al., 2008). Likewise, the relationship between brain areas that process numerical and non-numerical ordinal information is unclear. Some studies found a stronger IPS response to number processing compared to non-numerical order processing (Le Clec'H et al., 2000). For example, in an experiment by Thioux et al. (2005), participants had to compare and categorize numbers and animal names. When animals were compared, only left temporal lobe activation was observed, although this comparison task relied on ordinal information. Parameter estimates in left and right IPS showed strong activation in the number tasks, but no activation in the animal tasks. Others, however, have suggested shared representations of ordinal knowledge in the IPS (Fias et al., 2007; Ischebeck et al., 2008). Marshuetz et al. found parietal activation when participants had to recall the order of two probe letters after storing five letters for a brief period (Marshuetz et al., 2000, 2006). They observed parietal Corresponding author. Fax: address: Filip.VanOpstal@UGent.be (F. Van Opstal). activation that was modulated by the distance between the two probe letters (Marshuetz et al., 2006). It was suggested that order information is represented via magnitude codes, and that these magnitude codes are processed in parietal cortex. A different approach to settle the debate on the specific role of the IPS in ordinal processing is to use a transitive inference (TI) task in which participants gradually acquire the ordinal relations between a set of novel elements (Van Opstal et al., 2008b). In a typical TI task participants learn the ordinal structure of a sequence of stimulus elements (e.g., ABCDE) on the basis of premises consisting of adjacent elements in the sequence (e.g., AB, BC, CD, DE). After an initial learning phase participants are then tested on non-adjacent elements in a test phase to investigate if they acquired the ordinal relations between the different elements in the sequence. In fact, in such a design, a test phase is identical to a simple comparison task that has been frequently used to study ordinal stimuli like numbers (Fias et al., 2003, 2007; Pinel et al., 2004), or letters (Fias et al., 2007). Neuroimaging data revealed activation in a hippocampal-angular gyrus network when participants learned a new sequence, and identified the left angular gyrus (ANG) as the parietal area for storing the ordinal sequence knowledge, rather than the IPS (Van Opstal et al., 2008b). However, it might be argued that more extensive training would eventually lead to IPS activation. To clarify this issue, the present experiment was designed to investigate whether extensive training would change the neural network involved in learning a novel sequence of arbitrary elements. As in Van Opstal et al. (2008b), participants learned a sequence of arbitrary figures using a TI task, but were now trained for seven sessions on seven consecutive days instead of one single session. Participants were scanned in sessions 1, 2, 4, and 7, so that the neural /$ see front matter 2009 Elsevier Inc. All rights reserved. doi: /j.neuroimage

3 368 F. Van Opstal et al. / NeuroImage 47 (2009) network initially involved in acquiring the ordinal knowledge (changes within session 1), and how this network evolves with training (changes between sessions 2, 4, and 7), could be studied. Our working hypothesis was that if extensive training could alter the initial representation into a more abstract representation, it would thereby shift the loci of activation involved in the TI task from ANG to IPS. Alternatively, if ordinal representations are not necessarily processed in IPS, the loci of activation might shift to cortical regions other than IPS or remain in the cortical areas initially involved in the acquisition of the ordinal knowledge. Materials and methods Participants Fourteen participants (all male, right-handed, aged between 19 and 26 years) from Ghent University were studied after they gave written informed consent in a manner approved by the ethical committee and the trial bureau of the Medical Department of Ghent University. None of the participants reported any neurological, psychiatric or medical history, or had any bodily ferromagnetic materials. Brain morphology, as assessed with T1-weighted MR images did not reveal any abnormalities. Participants were rewarded 130 Euros after full completion of the study. fmri task design Prior to scanning, participants received instructions on the details of the experimental procedure. They were explicitly informed that the goal of the study was to investigate the effect of training on the neural correlates of TI. Before acquiring the fmri time series, a high resolution structural image (see Data acquisition for details) was obtained. The experiment consisted of seven sessions. Each session took place at the same time of the day on seven consecutive days. The task was the same throughout the seven sessions, except for the first session. The design of the first session was nearly identical as the one used in our previous study (Van Opstal et al., 2008b), and is illustrated in Fig. 1A. This session was divided into six blocks, each block Fig. 1. Experimental design. (A) An outline of the scanning sessions (Fam = familiarization, Fix = fixation, Learn = learn phase, Test = test phase, Contr = control phase; e.g., Fam1 = first familiarization phase, Learn18 = 18th learn phase). Participants were trained to learn a sequence of six arbitrary figures using a TI task. An experimental session consisted of a continuous repetition of a fixation, learn, test, and control phase. An illustration of the learn, test, and control phase is shown in (B). In a learn phase, participants were presented with adjacent stimuli only. Feedback was provided after each trial in the learn phase by a colored circle. In the test phase, participants were presented with non-adjacent stimuli only, and no feedback was presented. In order to accurately perform in the test phase, participants need to infer the non-adjacent relations between the stimuli from the adjacent relations presented in the learn phase. The control phase was a one-back task with a similar sensory input and motor response as the test phase. The different sets of figures used are presented in (C).

4 F. Van Opstal et al. / NeuroImage 47 (2009) containing four repetitions of a fixation, learn, test, and control phase (see below). The experimental session began with a familiarization phase (120 s) in which the six figures to be learned in the first five blocks (training set) were shown at a random position on the screen to allow participants to become accustomed with the different figures. Before the sixth block, a new set of figures (novel set) was shown in a second familiarization phase. This novel set of figures was introduced to break the linearity in the learning curve, so that it would disentangle learning effects from simple linear temporal effects, for example changes in arousal or habituation. Participants were instructed to keep their eyes focused on a fixation cross that remained on the screen throughout the whole session. Each block consisted of four repetitions of the sequential presentation of a fixation phase, a learn phase, a test phase, and a control phase. During fixation phases, a white fixation cross was presented in the center of a black screen for 15 s. Prior to each of the learn, test, and control phases, a summary of the instructions for that phase was presented for the duration of a trial (i.e., 2.8 s). Fig. 1B depicts examples of the learn, test, and control phase. Trial duration was the same in the learn, test and control phases (2.8 s). A learn phase consisted of 10 learning trials. Each learning trial began with a fixation cross in the center of the screen for 200 ms. This was followed by simultaneous presentation of a figure left and right of the fixation cross for 200 ms. Participants then had to press the left or right button depending on whether they thought the left or right figure was the hindmost in the order respectively. The response time (RT) limit was 2.4 s. After a response, or after the RT limit, feedback was presented in the form of a colored circle in the center of the screen on top of the fixation point that remained visible. Because the trial duration was fixed at 2.8 s, the duration of the feedback was dependent on the RT, and was set to a minimum of 200 ms. There were three different types of feedback: A green circle indicated a correct answer, a red circle indicated an incorrect answer, and a blue circle indicated that no answer was given before the RT limit. Importantly, stimuli in the learn phase were adjacent pairs only. Each possible combination was presented once in each order (e.g., AB and BA), resulting in 10 different trials per learn phase (i.e., the pairs AB, BC, CD, DE, EF, BA, CB, DC, ED, and EF). A learn phase thus lasted 28 s. Each learn phase was followed by a test phase, which was similar to a learn phase but for two exceptions. First, no feedback was provided during the test phase. For this reason, the RT limit was extended to 2.6 s. Second, only non-adjacent stimulus pairs were presented in a test phase, and trials containing anchor items (the first and the last figure in the order) were excluded from the test phase. The remaining six different stimulus pairs (i.e., BD, CE, BE, DB, EC, and EB) were presented twice in each test phase, which thus consisted of 12 trials; each test phase lasted 33.6 s. Because a test phase could potentially be solved by a memory strategy different from TI (e.g., always press on the side of stimulus D or E), we included a trial in the test session that asked for the relation between the fourth and the sixth stimuli (DF, or FD). This trial was presented in two randomly chosen test phases and randomly replaced one other trial within the test phase. In a post scanning questionnaire, none of the participants reported having used the alternate memory strategy. After each test phase, a sensori-motor control phase was presented. This control phase consisted of a one-back task. As in the learn and test phase, two figures were presented for 200 ms. The task was to remember both figures until the next trial, and to press the button corresponding to the side on which a figure of the previous trial appeared. The sensori-motor demands for this task match those of the test phase. Furthermore, a one-back task is expected to involve working memory. Using a one-back task as a control thus removes any activation related to sensori-motor, as well as working memory processes. The stimuli used in the control task were from a different figure set (control set) than the learn and test phases. A control phase also consisted of 12 trials and thus lasted 33.6 s. After five blocks with the training set, the novel set of figures was introduced during the second familiarization phase. Then participants had to learn the sequence of these new figures through four repetitions (6th block) of the sequence of fixation, learn, test and control phase. Participants were informed of this change of figure set before entering the scanner, but were not told when it would take place. The set of figures used in the one-back task remained the same throughout the entire experiment. The other six sessions (sessions 2 to 7) consisted of five blocks only during which the ordered sequence of the first session was repeated (Fig. 1A). Furthermore, in sessions 2 to 7 the familiarization phase was limited to 20 s. Participants were scanned only in sessions 1, 2, 4 and 7. The other sessions (3, 5, and 6) took place outside the scanner. Stimulus presentation and response collection Participants were placed head first and supine in the scanner. Stimuli were presented in white on a black background through dual display MRI compatible LCD displays, mounted in a lightweight headset (VisuaStim XGA, Resonance technology Inc, resolution , refresh rate 60 Hz). Three different stimulus sets were used (Fig. 1C). Stimulus set assigned to the main task and to the control task, and order of the figures in the main task, were randomized between participants. Stimulus delivery and the recording of behavioral data were controlled by E-Prime (Psychology Software Tools) running on a Pentium 4 laptop positioned outside the magnetic room. Responses were collected through magnetic-compatible two-key response boxes. Participants responded with their left and right index finger. Data acquisition Scanning was performed at 3 T on a Siemens Trio MRI system, using a standard eight-channel head coil for radio frequency transmission and signal reception. Participants were reminded to keep their head as still as possible. Foam padding was used to further limit head motion within the coil. After automatic shimming of the magnetic field on each participant, anatomical images were collected. For each participant a series of 3D high resolution T1-weighted anatomical images (3D MPRAGE, 176 slices, slice thickness=.90 mm, in-plane resolution=.9.9 mm 2, TR=1550 ms, TE=2.89) were obtained to serve as a reference for anatomical correlates and coregistration with the functional images. Next, for functional MRI, a single-shot multiple slice T2 -weighted echo planar imaging (EPI) sequence was used with the following parameters: TR 2500 ms, TE 33 ms, flip angle 90, in-plane resolution=3 3 mm, FOV=192 mm 2, matrix dimensions 64 64, slice thickness = 3 mm, interslice gap=1.5 mm. Forty slices covered the whole brain. The whole experiment consisted of four scanning sessions (days 1, 2, 4, and 7). As noted before, only the session on the first day had 6 blocks and thus contained 1248 volumes. The other sessions had 5 blocks and contained 965 images each. Data analysis Data processing and analysis were performed using Matlab and SPM5 (Wellcome department of Imaging Neuroscience, Institute of Neurology, UCL; Motion parameters were estimated for each session separately. The functional images from all sessions were coregistered with the participants' corresponding anatomical (T1-weighted) image from the first session. The resulting images were normalized using 12- parameter affine transformation into SPM5's MNI EPI mm template using the corresponding anatomical image as a reference, smoothed using a 7 mm full width at half-maximum Gaussian kernel, and temporally filtered at a 128 s cutoff.

5 370 F. Van Opstal et al. / NeuroImage 47 (2009) SNARC experiment after session 7 An additional behavioral task was administered after the end of session 7 in order to investigate a possible SNARC (Spatial-Numerical Association of Response Codes) effect; this refers to the finding that it is easier to respond to stimuli from the beginning of a sequence with the left hand and to stimuli from the end of a sequence with the right hand (Dehaene et al., 1993). A SNARC effect would indicate that the ordered sequence is associated with spatial codes and suggest a strong internalization of the sequence. Apparatus and stimuli In the SNARC experiment, responses were collected through a response box attached to a Pentium 4 laptop. All characters were presented in white on a black screen. All participants completed both an order-relevant and an order-irrelevant task. In the order-relevant task, participants were instructed to judge the position of the stimuli as coming before or after the middle stimuli (C and D) of their learned sequence. Only the first two and the last two stimuli were presented (A, B, E, and F). The stimuli in the order-irrelevant task were the same as in the order-relevant task, but they could be presented either straight or tilted 10 clockwise. In the order-irrelevant task, participants had to judge whether the presented stimulus was tilted or not. A trial started with the presentation of a fixation cross in the middle of the screen for 1000 ms, followed by the stimulus. The stimulus was presented until a response was given. The intertrial interval was 500 ms. analysis. ROI definition for the first session was thus performed using data from these six participants. Behavioral results Fig. 2A illustrates behavioral performance in the first session. Linear trend analysis on the mean accuracy data showed a significant linear increase in the test phase from block 1 to block 5, F(1, 5)= 22.75, pb.01. When a new set of stimuli was presented, accuracy decreased significantly from 99% in block 5 to 54% in block 6, F(1, 5)= 28.02, pb.01. Analysis of the RT revealed no significant effect of block (p=.16). Accuracy data and RTs of the first session are presented in Fig. 2A. The comparison distance effect was analyzed by a 6 (Block: 1 to 6) 2 (Comparison Distance: 2 or 3) repeated measures ANOVA on the mean error rate of trials with ordinal size 5 (pairs including figure E: BE and CE). However, because there were very few errors in the fifth block, the analysis on the mean accuracy was performed on five blocks only (blocks 1 to 4 and block 6). This revealed a significant effect of block, F(4, 20)=5.65, pb.005, and a significant effect of distance, F(1, 5) =13.01, pb.05. The same analysis was performed on the correct mean RTs for all six blocks. This revealed an effect of comparison distance close to significance, F(1, 5) =4.86, p=.079: Mean RTs to trials with comparison distance 2 tended to be slower (1077 ms) than RTs to trials with comparison distance 3 (1011 ms). No interaction between block and comparison distance was observed. The presence of a distance effect points towards a representational integration of the ordinal stimuli. Procedure Both the order-relevant and the order-irrelevant task consisted of two blocks. Each block consisted of 96 trials. Half of the participants started with the two blocks of the orderrelevant task, the other half with the two blocks of the orderirrelevant task. Response hand mappings were changed between two blocks of the same task: Participants responding with the right/left hand assigned to before/after in the first block of the order-relevant task changed their responses to left/right assigned to before/after in the second block, and vice versa. A similar counterbalancing procedure was implemented for the order-irrelevant task. The order of the tasks and the response hand mapping were counterbalanced between participants. Each block was preceded by a short practice block of eight trials where feedback on accuracy was given. Participants were instructed to respond as fast as possible but to avoid errors. The SNARC experiment consisted of 4 96=386 trials and lasted about 20 min. Results One participant was excluded from all analyses because he still performed at chance level in the test phase by the end of the seventh session. Because the experiment was not designed to investigate feedback-based learning of simple associations, data from the learn phases were not analyzed further. Acquisition of ordinal knowledge (changes within session 1) To study the neural network involved in the gradual acquisition of ordinal knowledge, we first analyzed the data of the first session (as in Van Opstal et al., 2008b), in order to determine regions of interests (ROI) where changes could take place over training. Accuracy of six participants reached less than 60% on average in the fourth and fifth block of the learn or test phase, or they failed to perform the control task (i.e., average performance about 60%). Eight other participants succeeded in learning the ordered sequence in the first session. Two of these eight participants moved excessively (N5 mm) during the session and were therefore removed from the Fig. 2. Behavioral performance of six participants in session 1 and ten participants in sessions 2, 4, and 7. (A) Percent correct responses in the test blocks in session 1. (B) Mean RTs on trials with distance 1, distance 2 and distance 3 in the learn and test phase in sessions 2, 4, and 7. Error bars indicate one standard error of measurement.

6 F. Van Opstal et al. / NeuroImage 47 (2009) Table 1 Peak activations contrasting block 5 to blocks 1 and 6 in the test phase, masked exclusively with the same contrast in the control phase. Stereotaxic coordinates # Voxels T value Anatomical region X Y Z R caudate nucleus L caudate nucleus L hippocampus L superior temporal gyrus R hippocampus R superior temporal gyrus L hippocampus L pallidum L angular gyrus L inferior parietal cortex R cuneus L precentral cortex R precentral cortex Coordinates are in MNI space. Clusters are labeled according to AAL (Anatomical Automatic Labeling, Tzourio-Mazoyer et al., 2002). L = left, R = right. Imaging results To investigate the areas involved in learning the transitive structure of the sequence of stimuli, we studied the time course of brain activation related to performance in the test phase of the TI task. As in Van Opstal et al. (2008b), we therefore contrasted performance on the test phase in the block where performance was highest (block 5) with performance on the test phase in those blocks where performance was lowest (blocks 1 and 6). This contrast was masked exclusively (p=.01) with the same contrast in the control phase to eliminate the possibility that the pattern of activation was generic rather than specific to the test phase. Because of the small sample of participants that could be used in this analysis, the voxel-level threshold was set to pb.005, and the cluster size was set to a minimum of 30 contiguous voxels. Results are shown in Table 1, and Fig. 3. Because our main interest in this first analysis was to replicate the findings of our previous study, analyses were focused on the activations that were also found in Van Opstal et al. (2008b). Linear trend analysis on the psc data of the significant clusters extracted from left and right hippocampal areas and from left parietal lobe (see Table 1) showed a significant linear increase from block 1 to block 5 in for the clusters which involved left HC and left superior temporal gyrus, the cluster with right hippocampus, and the cluster containing left ANG (all ps b.05). A trend towards a significant increase was observed in the cluster consisting of left HC and left pallidum, (p=.087). A significant decrease in activation was also observed from block 5 to block 6 in all these clusters (all psb.01), except for the cluster with left HC and left superior temporal gyrus where the decrease was close to significance (p=.06). In sum, although these analyses were based on six participants only, the behavioral and imaging data of the first session replicated the Fig. 3. Imaging results for sessions 2, 4 and 7. (A) Random effects analysis localizing areas in the contrast (Test session7 Control session7 ) (Test session2 Control session2 )). The color bar denotes the t-value. The psc data for these areas are plotted in (B). The posterior probability maps in left and right IPS are shown in (C). The color bar denotes the posterior probability for the presence of an effect in this contrast (in %). The psc data for these clusters are plotted in (D).

7 372 F. Van Opstal et al. / NeuroImage 47 (2009) involvement of a HC-ANG network as reported in Van Opstal et al. (2008b). We next turned to the analysis of changes in transitive inference-related neuronal activity after extended training. Changes related to training (changes between session 2, 4, and 7) In addition to the six participants used in the analysis of the first session, six additional participants were able to solve the task during the first block in the second session. Of these, two participants were removed because of excessive motion (N4.5 mm) in one or more sessions. The following analyses to investigate the effects of extensive training were thus performed on 10 participants. Behavioral results As can be seen in Fig. 2B, performance in both the test and control phase was close to perfect throughout sessions 2, 4, and 7. A 2 (Phase: Test and Control) 3 (Session: 2, 4, and 7) ANOVA on the mean percentage correct trials only revealed a main effect of phase, F(1, 9)=7.17, pb.05: Participants scored better in the test phase (96% correct) than in the control phase (93% correct). Learning, however, proceeded until session 7 as indicated by a decrease in RTs in the test phase from session 2 to session 7 (mean RTs of 604 ms, 530 ms, and 486 ms for sessions 2, 4 and 7 respectively, see Fig. 2B). We compared the decrease in RTs in the test phase to the RTs in the control phase to exclude the possibility that the decrease was solely caused by motor learning (in which case a similar decrease should be observed in both phases). An analysis on the RTs revealed a significant interaction between phase (control versus test) and session (2, 4, or 7) (pb.05), with a larger decrease of RTs in the test phase. However, this analysis may be problematic in that RTs are slower for the test phase, which naturally leads to larger effects (e.g., Hale and Jansen, 1994). We therefore performed the same analysis after normalizing the RTs by dividing them by the mean of each task to rule out generic task differences. Again, this revealed a significant interaction (pb.05) indicating that learning effects in the transitive task over sessions 2 7 reflect more than generic processes. A 3 (Session: 2, 4, and 7) 2 (Comparison Distance: 2 and 3) ANOVA on the mean correct RTs of trials with ordinal size 5 (BE and CE) in the test phase revealed a significant main effect of session, F(2, 18)=21.33, pb.001) with slower RTs in the second session (818 ms) compared to the fourth (586 ms) and the seventh (523 ms) session. A significant interaction between session and distance, F(2, 18) =3.90, pb.05, revealed an effect of comparison distance tending towards significance in the second session, F(1, 9) =3.74, p=.085, but not in the fourth and the seventh session (p=.22, and p=.27, respectively; see Fig. 2B). To test whether this absence of a significant comparison distance effect is caused by stimulus response learning, we also looked at the RTs on distance 1 trials from the learn phase. If participants had simply learned stimulus response associations (leading to the disappearance of a comparison distance effect), the distance effect for trials with distance 1 versus trials with larger distances would certainly disappear or even reverse because participants received feedback on distance 1 trials only. As can be seen in Fig. 2B, however, RTs to these trials were slower than distance 2 and distance 3 trials from the test phase (all psb.05). The distance effect was also present in the analysis of the error rates. An ANOVA on the error rates on distance 2 and distance 3 trials in the test phases of sessions 2, 4, and 7 revealed a main effect of distance, F(1, 9)=6.12, pb.05. The presence of a distance effect in both RTs and error rates confirms that the sequence was well internalized. Imaging results To examine how extensive training affects BOLD response in the brain regions involved in solving the TI task, we first contrasted the test phase to the control phase in the last session and the activation from the same contrast in the second session; these two differences were then again contrasted (i.e., (Test session7 Control session7 ) (Test session2 Control session2 )). This interaction contrast thus points to the brain regions that are more active in the last session compared to the second session in the test phase. The activation maps were thresholded with a t-value of 4.29 (pb.001), with a cluster extent of N97 voxels (corrected pb.05). As can be seen in Fig. 4A and Table 2, this revealed two clusters in left frontal gyrus. Increased activation was observed in the psc data for both clusters across sessions 2, 4, and 7 (Fig. 4B). Remarkably, no parietal activation was found in this analysis. To further investigate the absence of IPS activation, we first lowered the threshold and the cluster size to the more liberal values used in the analysis of the first session (voxel threshold =.005, cluster extent=30). As shown in Table 2, lowering the threshold did not reveal IPS activation. Because an inability to reject the null hypothesis (i.e., the absence of an effect in IPS) cannot be interpreted as a proof of an absence of activation in a given area, we used Bayesian statistics to determine the probability that there is any condition-related activation on ROIs in bilateral IPS areas (e.g., Friston and Penny, 2003). The centers of the ROIs were defined by the coordinates of a meta-analysis about the role of the IPS in non-numerical sequence processing (Talairach coordinates 31, 51, 47, and 38, 42, 44 for the left and right hemisphere respectively, see Cohen Kadosh et al., 2008). Each Fig. 4. Median RTs and error rates on the congruent and incongruent trials in the order-relevant and order-irrelevant task. Error bars denote the standard of the mean. pb.05. The O and X denote the mean RTs in the order-relevant letter and month task respectively in Gevers et al. (2003).

8 F. Van Opstal et al. / NeuroImage 47 (2009) Table 2 Peak activations contrasting session 2 to 7 in the test phase to session 2 to 7 in the control phase. Stereotaxic coordinates X Y Z # voxels T value Anatomical region High threshold L inferior frontal gyrus L inferior frontal gyrus Low threshold L superior frontal gyrus L middle temporal gyrus L supramarginal gyrus R middle temporal gyrus R middle occipital gyrus L middle temporal gyrus The high threshold indicates a voxel-level threshold of b.001 with a cluster size N97. In the low threshold the voxel-level threshold was b.005 with a cluster size N30. Coordinates are in MNI space. Clusters are labeled according to AAL (Anatomical Automatic Labeling, Tzourio-Mazoyer et al., 2002). L = left, R = right. ROI was a sphere (diameter=10 mm) around these centers. This analysis revealed that the maximum probability that an effect is present in the contrast of interest in these ROIs was no more than 9% (Fig. 4C, see Orban et al., 2006, for the same procedure), strongly suggesting the lack of implication of the IPS, without needing to deal with the thorny issue of accepting the null hypothesis inherent in classical statistics (for discussion of these issues, see Edwards et al., 1963; Wagenmakers et al., in press). The lack of an effect in IPS was confirmed by psc analysis where no effect of session was found (pn.16; Fig. 4D). Analysis of the brain areas where a decrease was observed from session 2 to session 7 (i.e., (Test session2 Control session2 ) (Test session7 Control session7 )) did not reveal any significant voxel at the strict or lowered threshold (p=.005). Furthermore, none of the activation patterns of the ROIs identified in the first session showed a reliable increase or decrease in activation, showing that these regions remained involved in the task. An alternative explanation for the absence of IPS activation in our analysis is that increasing task-related activation was compensated by decreasing attention-related activation. Indeed, even though accuracies are matched across the different conditions and sessions, the RT decrease from session 2 to 7 is steeper for the test phase than for the control phase. From this one might argue that there was a steeper decrease in attentional resources in the test phase than in the control phase, which compensated the steeper increase of task-related processes in IPS, thus leading to a net effect of no IPS activation. If this account is true IPS activation should be observed in the interaction between session and condition for those sessions where no interaction is observed in the behavioral data. Because of the high accuracies in all sessions, these accuracies were matched for all different sessions. Next, we further explored the interaction observed in RTs between condition (test or control) and session. Planned comparisons on sessions 4 and 7 revealed that there was no significant main effect of condition (p=.1) for these sessions and no significant interaction between sessions 4 and 7 (p=.2), although a main effect of session was observed (pb.0005): There was thus no difference in the decrease in RTs between the test and control condition. If the absence of IPS activation in the previous analysis was indeed due to a masking by attentional demands, we should now find IPS activation if we perform the interaction analysis between session and condition again, this time using these sessions where no behavioral interaction was found, i.e. sessions 4 and 7. Results still did not reveal any activation in IPS. Only at the low threshold (pb.005; cluster size=30) active clusters extending to parietal cortex were found: One cluster in right temporal superior gyrus (52, 50, 22) extending to right angular gyrus, and one cluster in left temporal middle gyrus ( 38, 60, 18) extending to left angular gyrus. No activation of IPS was observed. Bayesian analyses on the same IPS clusters as reported before revealed that the maximum probability of finding an effect was 11% in the right IPS cluster and 9% in the left IPS cluster for this contrast. We also tested whether IPS was consistently more active in the test phase compared to the control phase over sessions 2, 4, and 7. Although this would mean that IPS activation is not directly related to changes in the representation of ordered sequences, it might still play a continuous role in processing ordinal information. Just as in the previous contrasts, however, no IPS activation was found. Remarkably, apart from two clusters in the frontal cortex ( 48, 36, 8, and 12, 50, 32), in cerebellum (24, 80, 28), and precuneus ( 6, 64, 22), there was a significant cluster in left Angular gyrus ( 52, 66, 38). This result indicates that the parietal cluster, i.e. left ANG that was initially involved in learning the ordered sequence remained involved in the task. Similar results were obtained when we looked at each of the sessions individually. SNARC experiment after session 7 After the last scanning session, participants took part in a short SNARC experiment to test how well the ordered sequence was internalized. Trials with stimuli A and B (left side of the sequence) that were responded to with the left hand, and trials with stimuli E and F (right side of the sequence) that were responded to with the right hand were labeled as congruent trials. Trials with the reversed response hand mapping were labeled as incongruent trials. To match the stimuli in the order-irrelevant and order-relevant task, all trials with tilted stimuli in the order-irrelevant task were removed from the analysis. Inclusion of these trials did not change the results. A 2 (Task: order relevant/order irrelevant) 2 (Congruency: congruent/incongruent) repeated measures ANOVA with both task and congruency as withinsubject variables was performed on the median of the correct trials. This revealed main effects of task and congruency, F(1, 12) =9.34, pb.01, and F(1, 12) =8.01, pb.05, respectively: Responses were faster in the order-irrelevant task (445 ms) than in the order-relevant task (520 ms), and to congruent trials (468 ms) than in incongruent trials (497 ms). Planned comparisons on the significant interaction between task and congruency, F(1, 12) =5.20, pb.05, revealed a significant effect of congruency in the order-relevant task, t(13) =2.87, pb.05. No significant congruency effect was observed in the order-irrelevant task (tb1) (see Fig. 4). Analysis on the error data revealed the same pattern. The mean error rate was 7.5%. More errors were made in the order-relevant task (8.3%) than in the order-irrelevant task (2.9%), F(1, 12) =18.2, pb.001. There was also a significant congruency effect, F(1, 12) =11.39, pb.01: More errors were made on incongruent trials (6.5%) than on congruent trials (4.7%). The interaction between task and congruency was close to significance, F(1, 12) =4.6, p=.053. Planned comparisons revealed a significant effect of congruency in the order-relevant task, t (13) =2.99, pb.05, but not in the order-irrelevant task, t(13) =1.1, p=.27. The congruency effect in the order-relevant task again shows that the order was well internalized. As further evidence for this internalization, we note that the RTs in the order-relevant task with letters as stimuli in Gevers et al. (2003) are comparable to the RTs in the present experiment (and also for months; see Fig. 4). Discussion The goal of the present study was to investigate the effects of training on the neural network involved in learning and representing an ordered sequence. Because the absence of IPS activation in the representation of a new learned ordinal sequence could be due to insufficient training in a previous study (Van Opstal et al., 2008b), participants were trained on a TI task for seven sessions in the present study.

9 374 F. Van Opstal et al. / NeuroImage 47 (2009) A first important finding is that both the behavioral and the imaging results of the first session of the present study mirrored those of Van Opstal et al. (2008b): A network consisting of HC and ANG was found again to be involved in learning TI. From the second session onwards, accuracies in the test and control phase were stable until the seventh session. A comparison distance effect was observed in all sessions, although learning was not complete as revealed by decreasing RTs from session 2 to session 7. The presence of a distance effect together with the present of a SNARC effect after the last session indicates that the sequence was strongly internalized. We next studied how extensive training affected the neural network involved in solving the TI task. The largest increase in activation from session 2 to session 7 was observed in frontal sites: Two clusters in left IFG, BA44 and BA45, showed a strong increase in activation from the first to the last session. Percent signal change analysis revealed that there was continuous learning from session 2 onwards to session 7, similar to the increase in activation observed in left IFG when participants learn an artificial grammar (Opitz and Friederici, 2004). No IPS activation was observed in this contrast, also not when the threshold and cluster extent were lowered to a more liberal value. Furthermore, Bayesian analysis revealed that there was only a very small probability of an effect in IPS. Because of its involvement in number and letter processing, it is often assumed that IPS stores all ordered sequences; from this perspective, the absence of IPS activation is surprising. At a computational level, however, very basic one-layered neural networks are capable of storing ordered sequences; when they do so, they naturally produce a comparison distance effect as proposed in both the numerical cognition (Verguts et al., 2005; Van Opstal et al., 2008a), and the animal learning (Frank et al., 2003; von Fersen et al., 1991) literature. From this alternative point of view, an ordered sequence (with ensuing comparison distance effects) can be stored in many different neocortical areas depending on the stimulus type at hand (e.g., letters and numbers in IPS, Fias et al., 2007; animals in temporal cortex, Thioux et al., 2005). A similar move away from dedicated cortical mechanisms can be observed in the domain of time perception. Earlier models of timing (e.g., Church et al.,1994) proposed there are oscillators to keep track of time residing in specific brain areas (e.g., cerebellum or parietal cortex; see Ivry and Spencer, 2004, for a review). In contrast, recent timing models emphasize that accurate timing is a property that naturally follows from the dynamic evolution of activation in neural networks (e.g., Mauk and Buonomano, 2004; Karmarkar and Buonomano, 2007). In this way, timing may not be processed in dedicated timing areas, but rather emerge in different areas depending on different situational characteristics. The IPS activation observed when comparing letters may be attributed to the fact that letters constitute an abstract ordinal sequence that is applied to express the order of a large variety of concrete lists. Letters can, for example, express items in texts, thereby making abstraction of the specific content of the list. This stands in contrast with the stimulus sequence in the present study that was only applied to the present context. Hence, it is reasonable that only with high levels of abstraction the IPS becomes involved, but future research is needed to settle this issue. The only increase in activation following training on an ordered sequence of arbitrary stimuli was observed in left IFG. Left IFG is well known for its involvement in language processing. More specifically, left IFG activation seems to support syntactic memory (Friederici et al., 2000): An increase in left IFG was, for example, observed when participants were learning an artificial grammar (Opitz and Friederici, 2004). The involvement of the left IFG in this study might therefore be attributed to an increase in verbal processes. However, although subjective reports from the participants have revealed that they initially use verbal strategies to learn the sequence in the first session (Van Opstal et al., 2008b), they don't report the use of these strategies after the seventh session when they were asked for response strategies. One possibility is that the verbal labels that are initially used are consolidated in left IFG after extensive training. Alternatively, the left IFG could be activated here because of its role in sequence processing more generally. Other tasks that involve an aspect of sequencing, such as the processing of musical sequences (Maess et al., 2001), the imagery of motion (Binkofski et al., 2000), and processing (temporally extended) vibrotactile stimulation (Preuschof et al., 2006) have also shown left IFG activation. Altogether, these results suggest that left IFG supports the processing of sequences in both language and nonlanguage domains (Friederici, 2000). In addition, our own results suggest that IFG is involved in processing at least some ordinal sequences. Acknowledgments This work was supported by Grant P6/29 from the Interuniversity Attraction Poles program of the Belgian federal government and by Grant BOF08/GOA/011 from the Ghent University Research Council. FVO is a Postdoctoral Fellow of the Research Foundation Flanders (FWO-Vlaanderen). References Binkofski, F., Amunts, K., Stephan, K.M., Posse, S., Schormann, T., Freund, H.J., Zilles, K., Seitz, R.J., Broca's region subserves imagery of motion: a combined cytoarchectonic and fmri study. Hum. Brain Mapp. 11, Church, RM., Meck, WH, In: Roitblat, H.L., Beaver, T.G., Terrace, H.S. (Eds.), Animal Cognition (Erlbaum), pp Church, R.M., Meck, W.H., Gibbon, J., Application of scalar timing theory to individual trials. J. Exp. Psychol. Anim. Behav. Process. 20, Cohen Kadosh, R., Henik, A., Rubinsten, O., Mohr, H., Dori, H., van de Ven, V., Zorzi, M., Hendler, T., Goebel, R., Linden, D.E.J., Are numbers special? The comparison systems of the human brain investigated by fmri. Neuropsychologia 43, Cohen Kadosh, R., Lammertyn, J., Izard, V., Are numbers special? An overview of chronometric, neuroimaging, developmental and comparative studies of magnitude representation. Prog. Neurobiol. 84, Dehaene, S., Bossini, S., Giraux, P., The mental representation of parity and numerical magnitude. J. Exp. Psychol. Gen. 122, Dehaene, S., Piazza, M., Pinel, P., Cohen, L., Three parietal circuits for number processing. Cognit. Neuropsychol. 20, Edwards, W., Lindmann, H., Savage, L.J., Bayesian statistical inference for psychological research. Psychol. Rev. 70, Fias, W., Lammertyn, J., Caessens, B., Orban, G.A., Processing of abstract ordinal knowledge in the horizontal segment of the intraparietal sulcus. J. Neurosci. 27, Fias, W., Lammertyn, J., Reynvoet, B., Dupont, P., Orban, G.A., Parietal representation of symbolic and nonsymbolic magnitude. J. Cognit. Neurosci. 15, Frank, M.J., Rudy, J.W., O'Reilly, R.C., Transitivity, flexibility, conjunctive representations, and the hippocampus: II. A computational analysis. Hippocampus 13, Friederici, A.D., Towards a neural basis of auditory sentence processing. Trends Cogn. Sci. 6, Friederici, A.D., Meyer, M., von Cramon, D.Y., Auditory language comprehension: an event-related fmri study on the processing of syntactic and lexical information. Brain Lang. 75, Friston, K.J., Penny, W., Posterior probability maps and SPMs. NeuroImage 19, Gevers, W., Reynvoet, B., Fias, W., The mental representation of ordinal sequences is spatially organized. Cognition 87, B87 B95. Hale, S., Jansen, J., Global processing-time coefficients characterize individual and group differences in cognitive speed. Psych. Sci. 5, Ischebeck, A., Heim, S., Siedentopf, C., Zamarian, L., Schocke, M., Kremser, C., Egger, K., Strenge, H., Scheperjans, F., Delazer, M., Are numbers special? Comparing the generation of verbal materials from ordered categories (months) to numbers and other categories (animals) in an fmri study. Hum. Brain Mapp. 29, Ivry, R.B., Spencer, R.M.C., The neural representation of time. Curr. Opin. Neurobiol. 14, Karmarkar, U.R., Buonomano, D.V., Timing in the absence of clocks: encoding time in neural network states. Neuron 53, Le Clec'H, G., Dehaene, S., Cohen, L., Mehler, J., Dupoux, E., Poline, J.B., Lehericy, S., van de Moortele, P.F., Le Bihan, D., Distinct cortical areas for names of numbers and body parts independent of language and input modality. NeuroImage 12, Maess, B., Koelsch, S., Gunter, T.C., Friederici, A.D., Musical syntax is processed in the area of Broca: an MEG study. Nat. Neurosci. 4, Marshuetz, C., Smith, E.E., Jonides, J., DeGutis, J., Chenevert, T.L., Order information in working memory: fmri evidence for parietal and prefrontal mechanisms. J. Cogn. Neurosci. 2, Marshuetz, C., Reuter-Lorenz, P.A., Smith, E.E., Jonides, J., Noll, D.C., Working memory for order and the parietal cortex: an event-related functional magnetic resonance imaging study. Neuroscience 139,