Abstract

In understanding the brain’s response to extensive practice and development of high-level, expert skill, a key question is whether the same brain structures remain involved throughout the different stages of learning and a form of adaptation occurs, or a new functional circuit is formed with some structures dropping off and others joining. After training subjects on a set of complex motor tasks (tying knots), we utilized fMRI to observe that in subjects who learned the task well new regional activity emerged in posterior medial structures, i.e. the posterior cingulate gyrus. Activation associated with weak learning of the knots involved areas that mediate visual spatial computations. Brain activity associated with no substantive learning indicated involvement of areas dedicated to the declarative aspects learning such as the anterior cingulate and prefrontal cortex. The new activation for the pattern of strong learning has alternate interpretations involving either retrieval during episodic memory or a shift toward non-executive cognitive control of the task. While these interpretations are not resolved, the study makes clear that single time-point images of motor skill can be misleading because the brain structures that implement action can change following practice.

Introduction

Neural regions crucial to transforming action goals into signals that support complex sequenced movements have been studied extensively (Houk and Miller, 1995; Hikosaka et al., 1999). The way in which these brain systems respond to extensive practice and come to implement high-level, expert motor skill is less well known. Single time-point brain imaging studies encourage the assumption that the neural circuit involved in a task remains invariant with additional learning. Several investigations have established that plasticity occurs within the same region of cortex as a function of practice with both simple motor movements (Merzenich et al., 1988; Karni et al., 1995; Tracy et al., 2001a) and more complex tasks (Jansma et al., 2001). A key question is whether various brain structures ‘turn off or on’ and a new circuit of function is formed as a result of factors such as automaticity (e.g. reduced resource utilization), attention (e.g. reduced error monitoring, shift to non-focal attention) or compositional changes in the algorithym used to implement the task (e.g. collapsing/reducing the action steps).

Recent evidence suggests there is no shift in the brain regions involved after practice on tasks such as the well-known Sternberg short-term memory task, with alterations only occurring within the same foci across the practice period (Jansma et al., 2001). In contrast, Petersson and colleagues used PET and found a shift (decrease) in the medial temporal lobe following repeated recall trials of a visual abstract design (Petersson et al., 1997). Based on this and a second set of studies involving object location and pseudoword recall, the authors concluded that a consistent pattern of learning effects could be observed involving activation reductions in prefrontal, anterior cingulate, and posterior parietal regions, and increases in auditory and posterior insular–opercular/perisylvian supramarginal cortices. The decreases were taken to reflect the development of automaticity and less reliance on attentional and working memory resources, while the increases were considered to reflect a reduction in the suppression of task irrelevant processing. Sakai and colleagues (Sakai et al., 1998) studied keypress motor learning in humans and found an increase in the inferior parietal sulcus was best associated with late stage performance. These and other studies of practice-related shifts in brain foci mediating task performance have been completed (Petersen et al., 1998; Poldrack et al., 1998; Wiser et al., 2000) but none have specifically examined the influence of different patterns of learning success to determine if they initiate distinct areas of brain activation. A depiction of regional change or cortical network differences as a function of learning success should help us understand why individuals fail to learn, and how one might foster learning success both on a cognitive and biological level. If any reliable change can be identified in individuals who become highly skilled, either in cortical regions or networks, this should help us articulate the principles governing the neuroanatomical dynamics behind optimum learning.

We specifically investigated the issue of learning success by training subjects on a set of complex motor tasks (tying knots) and then determined the brain activation change associated with different patterns of success or failure. In association with strong learning, we expected activation to shift away from dependence on structures such as the medial temporal lobe which implements declarative knowledge formation, the prefrontal cortex which implements working memory, and the anterior cingulate cortex which responds to increases in effort and error monitoring (Squires et al., 1992; D’Esposito et al., 1995; Carter et al., 1998). We also expected reductions in pre-supplementary and premotor areas that are considered important to the initial generation and formation of motor procedures (Hikosaka et al., 2000).

Materials and Methods

Subjects

Subjects were 15 healthy normal adults between the ages of 18 and 45 (mean = 28.9, SD = 7.9), comprising eight males and seven females. All were strongly right-handed as verified by the Edinburgh Handedness Inventory (mean laterality score = 92.7, SD = 11.6). All subjects were free of medical, neurologic, or psychiatric complications as established by an interview with a Ph.D. neuropsychologist. All were recruited from the Thomas Jefferson University community, provided written informed consent, and were paid for participation. All subjects were recruited according to Institutional Review Board guidelines.

Procedure

A knot tying task served as stimulation at 1.5 T in a repeated measures design (pre-and post-training fMRI scans; 15 normal healthy controls as participants) with five 1 h sessions of training that took place during a 2 week span between the two fMRI scanning sessions. A set of eight knots was utilized. During the training sessions, each of eight knots was performed five times with time to completion recorded in tenths of a second.

During the pr-e and post-training sets of functional images, procedures were identical. Prior to scanning, each knot was tied once with a picture of the knots and steps in view. Subjects were instructed to tie the knot accurately and as rapidly as possible. During scanning a slide depicting the identical picture of the knot and steps was shown to the subject and viewed via the headcoil mirror and a rear projection screen at the subject’s feet. Rope to tie the knots hung from a plastic bar in the bore of the magnet. Subjects were allowed up to 60 s per knot, and upon completion or expiration of the time allotted (60 s) subjects were instructed to drop the rope. There were eight knots with between three and six steps for completion (see Fig. 1). There were four experimental conditions: knot tying, visual maze control, baseline rest, and intertrial rest. The knot tying condition included whole brain volumes acquired solely during the process of knot tying. The time devoted to knot tying varied by knot and subject and session in accordance with actual performance, i.e. time-to-completion. At each fMRI session each of eight knots was presented once yielding eight knot tying epochs of up to 60 s. The visual maze control (eight epochs, 20 s each) provided the visual stimulus of a rope (see Figure 1) and subjects were instructed to follow the length of the rope from one end to the other with their eyes, and to do this for the time the maze was on the screen. This maze allowed subtraction of the visual stimulation, visual spatial processing, and eye movement elements of the knot tying task that were related to visually examining the reference knots. A baseline rest condition involved two separate periods of rest at the beginning (30 s) and end (60 s) of the scanning run during which subjects viewed a fixation cross. A period of rest of 8 s immediately prior to and after the knot tying epochs was necessary so that the rope could be delivered to the subject and then removed upon completion of the knot or expiration of the 60 second epoch (referred to as the intertrial rest; note, volumes collected during these periods were ‘binned’ and are included in the design matrix). Time-to-completion of the knots in the scanner was recorded in tenths of a second. Total scanning time was 14 min 30 s.

Imaging Parameters and Postprocessing

Whole-brain functional magnetic resonance imaging (fMRI) scans were conducted involving 21 parallel axial slices. Subjects were scanned with a GE LX 1.5 T clinical system using a quadrature RF coil. A single-shot echoplanar gradient imaging sequence acquiring T2* signal was used with the following parameters: TE = 54 ms, TR = 5.0 s (interleaved collection, contiguous slices), FOV = 21 mm, 128 ′ 128 ′ 6 data matrix, flip angle = 900, bandwidth = 1470Hz/pixel. The in-plane resolution was 1.64 ′ 1.64 mm providing slices 6 mm in thickness. T1-weighted images (21 slices) were collected using a standard spin-echo pulse sequence in positions identical to the functional scans to provide an anatomical reference to determine slice location of the echoplanar images (TE = 9 ms, TR = 450 ms, 256 ′ 256 mm). At each session subjects were placed in the headcoil using a stereotaxic crosshair light beam and positional markers on the headcoil to align their head identically at each session. Subjects lay in the same foam pad at each session to comfortably stabilize the head. Shimming was conducted at each location prior to acquisition to reduce system inhomogeneities and artifactual fluctuations in signal across images. Each EPI imaging series started with four discarded scans to allow for T1 signal stabilization.

SPM ’99 was utilized for all post-processing image analyses. All volumes within an fMRI run were co-registered or computationally aligned to correct for interscan movement with the first volume of the first fMRI session used as the reference. All volumes were then co-registered with the subject’s own T1 image and then transformed into standard anatomical space using the subject’s T1 image and the Statistical Parametric Mapping normalization procedure. Next, all volumes underwent spatial smoothing by convolution with a Gaussian kernel of two times the voxel size (FWHM) to increase signal-to-noise and account for residual inter-subject differences in anatomy. Low frequency fluctuations were removed by high-pass filtering (whitening procedures to remove artifacts such as aliased biorhythms). Temporal filtering (attenuate high frequency components to correct for temporal autocorrelation) was also applied. Both filtering processes utilized the standard procedure provided in SPM ’99 (Friston et al., 1994).

Statistical Analysis

The general linear model (GLM) procedure of the Statistical Parametric Mapping software (SPM ’99, http//:www.fil.ion.ucl.ac.uk/spm) was used to create a statistical model containing boxcar waveforms representing four different learning patterns (Strong, Weak and No Learning, and Consistent Performance). Scans that occurred after the knot was correctly tied were identified or binned as intertrial period scans and not included with the volumes dedicated to knot tying. Volumes collected during these intertrial periods were included in the model and design matrix. Using knot tying accuracy and speed at each fMRI session as a criterion variable, four distinct patterns of performance across the two fMRI sessions were identified: Strong Learning (a knot incorrectly tied at session one but correct at session two with a significant, 33%, reduction in time-to-completion across fMRI sessions); Weak Learning (incorrect then correct at session two with no substantive time-to-completion improvement); No Learning (correct/incorrect, or incorrect at both fMRI sessions); and Consistent Performance (correct at both sessions). These patterns were constructed on a knot-by-knot basis so a given subject typically would have each of their eight knots classified differently. Boxcar (sinusoid) waveforms identifying the knots falling into these four learning patterns were constructed. In addition, waveforms for the following other experimental events at each of the fMRI sessions were generated: visual maze control session 1, visual maze control session 2, baseline rest at session 1, baseline rest at session 2, intertrial periods during session 1, intertrial periods during session 2. As mentioned, classification was assigned on a knot specific basis, so subjects could be ‘Strong Learners’ on some knots and ‘Weak Learners’ for others. A total of 34.6% of the knots (eight knots by the 15 subjects) fit the Strong Learning pattern, 15.8% the Weak, 33.7% the No Learning pattern and 15.9% the Consistent Performance pattern.

All comparisons were specified by utilizing volumes associated with the relevant epochs to form the appropriate contrast or linear compound of parameter estimates with T statistics computed at every voxel to produce an SPM{T}. Six motion parameters for each session were also included in the model. The key contrasts determined activation increases or decreases associated with each learning pattern through an interaction term directly comparing the relevant knots at each session. For instance, to identify activation increases across the fMRI sessions that occurred for knots where Strong Learning occurred the following contrast was specified: (Strong Learning knots at session 2 – visual maze control, session 2) – (Strong Learning knots at session 1 – visual maze control, session 1). This contrast would reveal activation significant at the second fMRI session relative to the first. Note, the visual maze served as the active control condition and activation differences relative to this were determined separately for each session prior to the comparison between sessions. This strategy allowed us to account for session effects and the non-specific effects of time by comparing performance to a control task collected at the same time period as the key experimental condition. The strategy assumes that both time points are similarly influenced by non-specific time effects. It also accounts for temporally autocorrelated low-frequency noise such as biorhythms that differ at the two time points, variations in bold signal related to the physical performance of the scanner, and motion. The interaction does not make assumptions about the form of any nonspecific effect of time and is therefore sensitive to both linear and nonlinear effects, but will not be sensitive to effects that are not present equally in both the experimental and control tasks. This interaction approach also is insensitive to learning effects that occur in both the control and the experimental tasks. However, we assumed no learning effects of interest were present in the control task and, of course, subjects were not trained on that task and saw them only once previously during pre-scan instructions.

Additionally, each learning pattern was analyzed for activation decreases across the sessions using the same type of interaction contrast noted above, this time highlighting activation greater in the first fMRI session relative to the second for each of the patterns. In all cases regressors were formed using a sinusoid stimulus function representing the occurrence of each relevant epoch convolved with a canonical hemo-dynamic response function. The subject-specific contrasts associated with each learning pattern were thresholded at P < 0.05 and then each was entered into a random effects statistical model using a one-sample t-test to determine significant activation against the null hypothesis. Alpha values were corrected for multiple non-independent comparisons and analyzed using the theory of Gaussian fields. Results used for statistical inference and interpretation were corrected at a cluster-level specificity of at least P < 0.05 (although two trends are reported) with the spatial extent threshold reflecting the expected voxels per cluster given the smoothness of the image (Strong Learning, k = 22; Weak Learning, No Learning and Consistent Performance, k = 13).

Results

Behavioral data on knot tying training (time-to-completion) revealed that all 8 knots showed a significantly decreasing slope across the training sessions (P < 0.001; see Fig. 2a). Time-to-completion means for the knots from the two fMRI sessions differed significantly with the second session faster (df = 18, t = 4.2, P < 0.001; see Fig. 2a). Also, a higher proportion of knots were correctly tied at the second fMRI session (65.5% compared to 16.5%; McNemar test, P < 0.0001).

To determine if time on task differed across the four learning patterns we conducted an analysis of variance on our time-to-completion variable using a repeated measures MANOVA with Learning pattern as a between-subject factor and fMRI session as a within-subject factor. As would be expected given that the patterns were constructed with time-to-completion in mind, the patterns did differ both at session one (df = 3,116, F = 15.08, P < 0.001) and session two (df = 3,116, F = 50.8, P < 0.001). At the first fMRI session, the Strong and No Learning patterns significantly differed from the Consistent and Weak patterns (Scheffe post-hoc tests, P < 0.01) but not from each other (see Figure 2b). At the second fMRI session, the No Learning differed from all the others (P < 0.01) and the Consistent pattern differed from the Weak (P < 0.01). At this session there was statistical equivalence between the Consistent and Strong, and the Weak and Strong Learning patterns in terms of time-to-completion scores. When one examines overall time-to-completion for each pattern by combining across the sessions (mean times-to-completion: Strong = 38.5 s; Weak = 35.6s; Consistent = 27.2 s; No Learning = 53.5 s) only the No Learning pattern significantly differed (P < 0.01) from the others in terms of total time on task. Thus, overall total time on tasks was equivalent for three of the four Learning patterns. As our statistical comparisons focused on differences across sessions within each of the patterns, differences in time between the patterns was not an issue. By definition, the Strong Learning pattern had a different time-on-task at each session and this was verified through a paired t-test (df = 37, t = 17.6, P < 0.001). Regarding the other patterns, the Strong, Weak, and No Learning patterns showed no statistical difference between the sessions (paired t-test), however, the times for the Consistent Learning pattern did differ by session (df = 17, t = 3.9, P < 0.01).

The activation increases associated with the Strong Learning pattern included increases in the posterior cingulate gyrus region (see Table 1A and Fig. 3). No significant decreases in activation were evident across the sessions. The activation increases associated with the Weak Learning pattern involved an inferior parietal and precuneus region (see Table 1B and Fig. 3). This cluster was more medial and superior to the activation during Strong Learning. Weak Learning also showed a significant voxel activated in the precentral gyrus. No significant activation decreases were observed for the Weak Learning pattern.

The No Learning pattern showed a significant increase in medial anterior cingulate gyrus/middle frontal gyrus and a separate middle/dorsolateral frontal region (see Table 1c and Fig. 3). Additionally, an activation increase was observed in a voxel in premotor cortex. Again, no significant activation decreases were observed across sessions. Regarding the Consistent learning pattern, no significant activation increases or decreases were observed.

As learning effects were of most interest, we further examined the activation associated with the Strong Learning knots. A comparison of knot tying to the maze control task at fMRI session one, revealed two significant clusters of activation. A very large cluster in the extrastriate area (BA 18; k = 1748, P-corrected < 0.0001, maxima t = 8.1, x = 2, y = −65,z = 3) that was inferior and posterior to the posterior cingulate and precuneus changes observed for Strong Learning knots across the fMRI sessions. The second cluster observed was in prefrontal/medial frontal cortex (BA 6; P < 0.0003, k = 24, maximum t = 2.75, talairach coordinates: x = 38, y = −2, z = 50) suggesting that this anterior region was prominent during early learning at the first fMRI session.

At the second fMRI session, a region of precuneus/posterior cingulate gyrus activation was observed (BA 31; k = 861, P-corrected < 0.0001, maxima t = 5.3, x = 20, y = −68, z = 20) that was similar to the area reported for the Strong Learning knots across the sessions. Lastly, using the data from the Strong Learning knots we conducted a region-of-interest analysis directly examining whether the posterior cingulate region which changed across the session was observable at the first session, and it was not (we used a 20 mm sphere centered around the voxel located at x = −13, y = −11, z = −34, the maxima of the posterior cingulate cluster found to have emerged by session two for Strong Learning knots, see Table 1a). We also examined whether any anterior cingulate activation was present for the Strong Learning knots at the first fMRI session using the anterior cingulate location observed in the No Learning pattern as our guide (20 mm sphere centered around x = −13, y = 27, z = 32, see Table 1c) and there was none. These data go beyond our initial analyses and strengthen our claim that high skill and strong learning of the knots uniquely engaged specific medial posterior brain structures.

Discussion

The role of anterior cingulate cortex in new learning (e.g. error monitoring, inhibition of ‘prepotent’ responses, response selection from a large, unconstrained pool) and its responsiveness to effortful, resource demanding activities is well demonstrated through neuroimaging (Carter et al., 1998; Petersen et al., 1998). Also, anterior cingulate and dorsolateral prefrontal cortex are known to be crucial for early stage learning and, in fact, are areas where we expected a decrease in the context of learning success (Schneider et al., 1994; Petersson et al., 1999). The role of the posterior cingulate during learning, however, is not well established. The posterior cingulate has been associated with monitoring sensory events and orientation (Vogt et al., 1992), but this is the first study to show it to be activated during skilled, well-developed performance of a complex motor task relative to the novice, unmastered state.

In previous studies of skill learning, the posterior cingulate region has been associated with changes in the task-related role of working memory or retrieval processes. For instance, Sakai and colleagues (Sakai et al., 1998) found increased activation in the precuneus at the intermediate stage of visuo-motor sequence learning. They attributed this to retrieval of the sequence from visual working memory suggesting that this region plays a role in a visual-spatial memory buffer. Note in our study, for the Strong Learning pattern, we saw no signs of anterior brain activation that might indicate a working memory system was at work.

There is also data connecting the retrosplenial region of the posterior cingulate to acquisition and the precuneus to retrieval processes during episodic memory (e.g. verbal paired associate recall) (Shallice et al., 1994). The former is considered to emerge out of the posterior cingulate’s connection with the entorhinal cortex. In the context of our task, however, the acquisition demands would be much greater at the pre-training fMRI session and we see no signs of posterior cingulate activation at that point. Regarding retrieval processes, there is a large literature involving animal (Wyss and Van Groen, 1992; Aggleton and Pearce, 2001), human lesional (Valenstein et al., 1987), and functional imaging studies reporting retrosplenial activity in association with a whole range of forms of episodic retrieval such as spatial context (Burgess et al., 2001), object (Wiggs et al., 1999), word (Clemens von Zerssen et al., 2001) and autobiographical memory (Maddock, 1999; Maguire and Mummery, 1999). Evidence showing anatomical connections between retrosplenial cortex and the parahippocampus (Wyss and Van Groen, 1992) help support these findings. We did, indeed, observe such retrosplenial (precuneus) activity in association with Strong Learning. Thus, one potential interpretation of our data is that the retrosplenial cortex is active during practiced performance because it is involved in retrieving memories of the previous knot tying trials and practice episodes. However, none of the above studies involved episodic memory for an acquired motor skill, and the posterior activation we report here for knot tying is anterior and superior to the retrosplenial regions many of these studies reported (Wiggs et al., 1999; Burgess et al., 2001; Clemens von Zerssen et al., 2001). There is no prior evidence to suggest that the posterior cingulate or precuneus are specifically involved in episodic retrieval for well-formed motor production units, and most workers suggest the long term storage sites for such material are likely to be in supplementary or premotor areas of frontal cortex (Rizzolatti et al., 1996; Hikosaka et al., 2000) Nevertheless, given the strong evidence for retrosplenial involvement in episodic retrieval our data clearly suggest that such posterior medial regions may also be involved in motor skill retrieval and specifically raise the possibility that this area is a site for long-term storage for motor production engrams of the complex kind studied here.

A second interpretation of the data is that the posterior medial structures observed for the Strong Learning pattern reflect a shift in cognitive control whereby resources allocated to the task and any attentional monitoring that occurs becomes relegated to posterior, not anterior, brain structures. If such a shift occurs the activation in the posterior cingulate could reflect the presence of a ‘non-executive’ cognitive control system that implements and monitors well-learned motor routines. This system may utilize a less effortful, more passive mode of attention (Smith et al., 1993; Tracy, 1999; Tracy et al., 2001b) that has greatly reduced attention resource needs, requires little in the way of strategy and decisions, and has low monitoring demands as errors during skilled performance are rare (Fitts and Posner, 1967). In support of this, cognitive models have long emphasized that a shift in cognitive resources and modes of attention is a likely consequence of the development of expertise and automaticity (Posner and Snyder, 1975; Schneider, 1985; Anderson, 1993). Also, there is a growing body of neuroimaging results that strongly suggest the functional neuroanatomy of novice and expert performances are likely to be dissociable (Petersen et al., 1998; Petersson, 2001; Tracy et al., 2001a). This notion of ‘non-executive’ cognitive control differs from prior conceptions of a posterior attention system which emphasized that such a system is involved in aspects of executive control such as focal attention [e.g. disengaging from targets or moving focal attention to relevant locations in visual space, (Posner and Petersen, 1990)]. It is important to note that the absence of activation in areas that implement the automatic ‘read out’ of motor sequences (supplementary motor cortex, anterior cerebellum or posterior striatum) (Rizzolatti et al., 1996; Hikosaka et al., 2000) indicates that the Strong Learning pattern reflected highly skilled but not yet automatized performance. Therefore, the association between posterior medial regions and Strong Learning cannot be said to link these areas to automaticity, although such a possibility is worth pursuing.

Another line of interpretation is that the anterior and posterior cingulate worked in interaction during the development of knot tying skill. Consider the fact that for the Strong Learning pattern, at the second session, there was a noticeable absence of activation in the areas typically associated with early learning such as the anterior cingulate, medial temporal lobe, and dorsolateral prefrontal cortex, suggesting that early encoding operations including working memory and error monitoring (Goldman-Rakic, 1987; Shadmehr and Holcomb, 1997; Carter et al., 1998) were not crucial to the implementation of skilled knot tying. There was, however, prefrontal/medial region activation observed at the first session for the Strong Learners and this is consistent with work showing that premotor regions such as Brodmann’s area 6 are important to refining the order of movements and settling the muscle group control issues, etc., that go into a final motor algorithm (Hikosaka et al., 1999, 2000). With skill development, these components and steps drop out or become organized in a very efficient hierarchy to run off as a single production to guide action (Anderson, 1982). Thus, one possible reason for the opposing effects in anterior and posterior brain regions may be that anterior regions are no longer needed after a complete and effective motor engram has been formed. Note, our data provide no evidence that the activation in posterior cingulate arose in conjunction with a simultaneous decrease in the anterior cingulate. Yet, the anatomical connections between anterior and posterior cingulate areas would certainly allow for the necessary communication between regions to generate such a linked set of activity. One speculative implication is that such an interaction between the posterior and anterior cingulate would set the stage for important activities such as dual-task processing by ‘freeing up’ anterior regions such as the anterior cingulate and prefrontal cortex to take on a new, more demanding concurrent task (Posner and Petersen, 1990).

Regarding the No Learning pattern, there were activation increases at the second session in areas strongly correlated with the declarative aspects of learning such as the anterior cingulate, middle/dorsolateral prefrontal cortex, and supplementary cortex. This suggests that the learning was still at an ‘early’, novice level where attention resource demands are high, monitoring of errors was prevalent, and the learning remained deliberative, serial, and highly piecemeal (componential) (Schneider, 1985; Anderson, 1993). This dorsolateral and premotor activation is fully consistent with work in motor learning by Hikosaka and colleagues who show that learning new but not old (practiced) motor sequences activates these areas (Hikosaka et al., 1999, 2000). In this sense, the No Learning pattern behaved in a well-predicted fashion involving areas known to mediate the initial stages of declarative learning. It appears that in response to the continued flawed performance and poor learning, these areas which mediate the acquisition of skill actually had to increase their activation.

Activation increases associated with the Weak Learning pattern occurred in regions known to mediate visual spatial processing and attention such as the inferior parietal cortex (Mesulam, 1981; LaBerge, 2000). This suggested that less accomplished knot tying performance heavily relied on visual– spatial computations. There was, however, also some medial posterior (precuneus) activation, perhaps suggesting that the mechanisms and effects observed for Strong Learning were beginning to set in. Note that the cognitive and behavioral state for the Strong and No Learning patterns is readily definable because of their clear performance properties. One can state that the mental structures that encode the information for implementing the knots either is or is not in place. In contrast, the Weak Learning pattern remains inherently ambiguous as it is unclear whether these individuals are on their way to Strong Learning, i.e. a kind of intermediate stage of learning, or reached a plateau of inferior performance where the mental structures are not adequately constructed.

A statistical power difference did exist for the Weak and Consistent Learning patterns compared to that of Strong and No Learning. Only 15.8% of the knots were identified as Weak, and only 15.9% as Consistent Learning. In part, because of such power differences, we did not directly compare the knot patterns with each other and focused instead on identifying activation regions that could help us understand the neuroanatomical change that emerges in concert with the cognitively distinct learning patterns. We do acknowledge, however, that the findings for the Weak and Consistently knots were no doubt affected by this low power and are in need of replication. It is intriguing that no change in activation accompanied the Consistent Learning pattern, which did show a significant reduction in time-to-completion across the fMRI sessions. This may suggest that the changes in activation we observed between the sessions for the other patterns were more closely tied to the cognitive aspects of the knot tying task such as retrieval or ‘non-executive’ monitoring, rather than to skilled, more purely motor output mechanisms.

In conclusion, while expert and highly skilled performance of tasks such as knot tying possess many features that are not addressed by these data (e.g. why performance is faster, knowledge less readily verbalized) the results clearly show that single-time point images of motor skill can be misleading as the brain structures that implement action can change following practice. Ultimately, our data cannot dissociate the lines of interpretation described above. The data provide strong evidence that posterior, medial structures such as the posterior cingulate and precuneus are active during highly skilled motor performance though it is unclear whether this is because it involves retrieval of episodic memories of particular knots, or a shift toward non-executive cognitive control associated with the reduced requirements for attention resource utilization and error monitoring that come with acquired skill. The former possibility raises the important possibility that this area is a site for long-term storage for the newly learned motor skill. The latter is consistent with the notion that a posterior attention system exists which mediates performance at a very different point on the learning curve than does the anterior attention system well described by several authors (Mesulam, 1981; Posner and Petersen, 1990). The present study makes clear that the brain activation which emerges with high skill and mastery is represented at both a cognitive and neuroanatomical level. Cognitively the nature of the shift is unclear, but neuroanatomically new activity emerges in medial posterior brain structures.

Notes

The project was supported, in part, by a grant from the Department of Defense, Office of Naval Research (N00014-00-1-0859).

Address correspondence to Joseph Tracy, Associate Professor of Neurology and Radiology, Thomas Jefferson University/Jefferson Medical College, 111 South 11th Street, Suite 4150, Philadelphia, PA 19107, USA. Email: joseph.tracy@mail.tju.edu.

Table 1

Activation increases associated with learning patterns

 Cluster/voxel level Maxima (mm) Talairach coordinates Brain region 
 kE P-corrected T x y z   
(A) Strong Learning pattern 
Cluster 321 <0.012 4.22 −13 −11 34 posterior cingulate gyrus 24, 31 
 184 <0.109 3.59 20 −53 36 precuneus 31 
(B) Weak Learning pattern 
Cluster 203 <0.015 5.63 −33 −35 46 inferior parietal 40 
 140 <0.08 5.55 16 −30 46 precuneus 
Voxel  <0.004 3.38 −34 −15 45 precentral gyrus 
(C) No Learning pattern 
Cluster 929 <0.0001 5.46 −13 27 32 medial anterior cingulated gyrus / middle frontal gyrus 32:9 
 192 <0.021 3.74 31 19 32 middle/dorsolateral frontal region 
Voxel  <0.002 3.39 −46 −1 39 premotor cortex 
 Cluster/voxel level Maxima (mm) Talairach coordinates Brain region 
 kE P-corrected T x y z   
(A) Strong Learning pattern 
Cluster 321 <0.012 4.22 −13 −11 34 posterior cingulate gyrus 24, 31 
 184 <0.109 3.59 20 −53 36 precuneus 31 
(B) Weak Learning pattern 
Cluster 203 <0.015 5.63 −33 −35 46 inferior parietal 40 
 140 <0.08 5.55 16 −30 46 precuneus 
Voxel  <0.004 3.38 −34 −15 45 precentral gyrus 
(C) No Learning pattern 
Cluster 929 <0.0001 5.46 −13 27 32 medial anterior cingulated gyrus / middle frontal gyrus 32:9 
 192 <0.021 3.74 31 19 32 middle/dorsolateral frontal region 
Voxel  <0.002 3.39 −46 −1 39 premotor cortex 
Figure 1.

Example of knot tying and visual maze control stimuli.

Figure 1.

Example of knot tying and visual maze control stimuli.

Figure 2.

Time-to-completion data during training and fMRI sessions.

Figure 2.

Time-to-completion data during training and fMRI sessions.

Figure 3.

Activation increases associated with the Strong, Weak, and No Learning patterns.

Figure 3.

Activation increases associated with the Strong, Weak, and No Learning patterns.

References

Aggleton J, Pearce J (
2001
) Neural systems underlying episodic memory: insights from animal research.
Trans R Soc Lond Biol Sci
 
356
(1413):
1467
–1482.
Anderson JR (
1982
) Acquisition of cognitive skill.
Psychol Rev
 
89
:
369
–406.
Anderson JR (
1993
) Psychological science 10. Rules of the mind. Pittsburgh, PA: Carnegie Mellon University.
Burgess N, Maguire E, Spiers H, O’Keefe J (
2001
) A temporal and prefontal network for retrieving the spatial context of lifelike events.
Neuroimage
 
14
:
439
–453.
Carter CS, Braver TS, Barch DM, Botvinick MM, Noll DC, Cohen JD, Schooler JW (
1998
) Anterior cingulate cortex, error detection, and the online monitoring of performance.
Science
 
280
:
747
–749.
Clemens von Zerssen, G, Mecklinger, A, Opitz, B, Yves von Cramon D (
2001
) Conscious recollection and illusory recognition: an event-related fMRI study.
Eur J Neurosci
 
13
:
2148
–2156.
D’Esposito, M, Detre, J, Alsop, D C, Shin, R K, Atlas, S, Grossman, M (
1995
) The neural basis of the central executive system of working memory
Nature
 
378
:
279
–281.
Fitts PM, Posner MI (
1967
) Human Performance. Belmont, CA: Brooks/Cole.
Friston KJ, Jezzard P, Turner R (
1994
) Analysis of functional MRI time-series.
Hum Brain Mapp
 
1
:
153
–171.
Goldman-Rakic PS (
1987
) Circuitry of primate prefrontal cortex and regulation of behavior by representational memory. In: The nervous system. Vol. V. Higher functions of the brain (Plum F, Mountcastle VB, eds), pp. 373–417. Bethesda, MD: American Physiological Society.
Hikosaka, O, Nakahara, H, Rand, M, Sakai, K, Lu, X, Nakamura, K, Miyachi, S, Doya, K (
1999
) Parallel neural networks for learning sequential procedures Trends in
Neurosciences
 
22
:
464
–471.
Hikosaka O, Sakai K, Nakahara H, Lu X, Miyachi S, Nakamura K, Rand MK (
2000
) Neural mechanisms for learning of sequential procedures. In: The new cognitive neurosciences (Gazzaniga MS, ed.), pp. 553–572. Cambridge, MA: MIT Press.
Houk JC, Miller LE (
1995
) Neural mechanisms for the generation of limb motor programs.
J Physiol
 
487
:
P225
–P226.
Jansma JM, Ramsey NF, Slagter HA, Kahn RS (
2001
) Functional anatomical correlates of controlled and automatic processing.
J Cogn Neurosci
 
13
:
730
–743.
Karni, A, Meyer, G, Jezzard, P, Adams, M, Turner, R, Ungerleider, L (
1995
) Functional MRI evidence for adult motor cortex plasticity during motor skill learning
Nature
 
377
:
155
–158.
LaBerge D (
2000
) Networks of attention. In: The new cognitive neurosciences (Gazzaniga MS, ed.), pp. 711–724. Cambridge, MA: MIT Press.
Maddock RJ (
1999
) The retrosplenial cortex and emotion: new insights from functional neuroimaging of the human brain.
Trends Neurosci
 
22
:
310
–316.
Maguire E, Mummery C (
1999
) Differential modulation of a common memory retrieval network revealed by position emission tomography.
Hippocampus
 
9
:
54
–61.
Merzenich MM, Recanzone G, Jenkins WM, Allard TT, Nudo RT (
1988
) Cortical representational plasticity. In: Neurobiology of neocortex (Rakic W, ed.), pp. 41–67. Chichester: John Wiley.
Mesulam M (
1981
) A cortical network for directed attention and unilateral neglect.
Ann Neurol
 
4
:
309
–325.
Petersen S, Van Mier H, Fiez J, Raichle M (
1998
) The effects of practice on the functional anatomy of task performance.
Proc Natl Acad Sci USA
 
95
:
853
–860.
Petersson KM (
2001
) Learning related modulation of functional retrieval networks in man.
Scand J Psychol
 
42
:
197
–216.
Petersson KM, Elfgren C, Ingvar DH (
1999
) Learning related effects and functional neuroimaging.
Hum Brain Mapp
 
7
:
234
–243.
Petersson KM, Elfgren C, Ingvar M (
1997
) A dynamic role of the medial temporal lobe during retrieval of declarative memory in man.
Neuroimage
 
6
:
1
–11.
Poldrack RA, Desmond JE, Glover GH, Gabrieli JD (
1998
) The neural basis of visual skill learning; an fMRI study of mirror reading.
Cereb Cortex
 
8
:
1
–10.
Posner MI, Snyder C (
1975
) Attention and cognitive control. Paper presented at Information processing and cognition: The Loyola Symposium, Hillsdale, NJ.
Posner MI, Petersen SE (
1990
) The attention system of the human brain.
Annu Rev Neurosci
 
13
:
25
–42.
Rizzolatti G, Fadiga L, Gallese V, Fogassi L (
1996
) Premotor cortex and the recognition of motor actions.
Cogn Brain Res
 
3
:
131
–141.
Sakai K, Hikosaka O, Miyauchi S, Takino R, Sasaki Y, Putz B (
1998
) Transition of brain activation from frontal to parietal areas in visumotor sequence learning.
J Neurosci
 
18
:
1827
–1840.
Schneider W (
1985
) Toward a model of attention and the development of automatic processing. In: Attention and performance XI (Posner MI, ed.), pp. 475–492. Hillsdale, NJ: Lawrence Erlbaum Associates.
Schneider W, Pimm-Smith M, Worden M (
1994
) Neurobiology of attention and automaticity.
Curr Opin Neurobiol
 
4
:
177
–182.
Shadmehr R, Holcomb HH (
1997
) Neural correlates of motor memory consolidation.
Science
 
277
:
821
–825.
Shallice T, Fletcher P, Frith CD, Grasby P, Frackowiak RS, Dolan RJ (
1994
) Brain regions associated with acquisition and retrieval of verbal episodic memory.
Nature
 
368
:
633
–635.
Smith JD, Tracy J, Murray MJ (
1993
) Depression and category learning.
J Exp Psychol Gen
 
122
:
331
–346.
Squires LR, Ojemann JG, Miezin FM, Petersen SE, Fox PT, Snyder AZ, Raichle ME (
1992
) Activation of the hippocampus in normal humans: a functional anatomical study of memory.
Proc Natl Acad Sci USA
 
89
:
1837
–1841.
Tracy J, Faro SS, Mohammed F, Pinus A, Christensen H, Burkland D (
2001
) A comparison of ‘Early’ and ‘Late’ stage brain activation during brief practice of a simple motor task.
Cogn Brain Res
 
10
:
303
–316.
Tracy J, Pinsk M, Helverson J, Urban G, Dietz T, Smith DJ (
2001
) Test of a potential link between analytic and nonanalytic category learning and automatic, effortful processing.
Brain Cogn
 
46
:
326
–341.
Tracy JI (
1999
) Learning complex categories two ways at once: the link between automatic processing and non analytic learning.
Percept Motor Skills
 
88
:
563
–578.
Valenstein E, Bowers D, Verfaellie M, Heilman K, Day A, Watson RT (
1987
) Retrosplenial amnesia.
Brain
 
110
:
1631
–1646.
Vogt BA, Finch DM, Olson CR (
1992
) Functional heterogeneity in cingulate cortex: the anterior executive and posterior evaluative regions.
Cereb Cortex
 
2
:
435
–443.
Wiggs C, Weisberg J, Martin A (
1999
) Neural correlates of semantic and episodic memory retrieval.
Neuropsychologia
 
37
:
103
–118.
Wiser A, Andreasen NC, O’Leary DS, Cizaldo T, Arndt S, Rezai K, Ponto LL, Watkins GL, Hichwa RD (
2000
) Novel vs well-learned memory for faces: a positron emission tomography study.
J Cogn Neurosci
 
12
:
255
–266.
Wyss J, Van Groen T (
1992
) Connections between the retrosplenial cortex and the hippocampal formulation in the rat: a review.
Hippocampus
 
2
:
1
–12.