Neural Correlates of Motor Memory Consolidation

See allHide authors and affiliations

Science  08 Aug 1997:
Vol. 277, Issue 5327, pp. 821-825
DOI: 10.1126/science.277.5327.821


Computational studies suggest that acquisition of a motor skill involves learning an internal model of the dynamics of the task, which enables the brain to predict and compensate for mechanical behavior. During the hours that follow completion of practice, representation of the internal model gradually changes, becoming less fragile with respect to behavioral interference. Here, functional imaging of the brain demonstrates that within 6 hours after completion of practice, while performance remains unchanged, the brain engages new regions to perform the task; there is a shift from prefrontal regions of the cortex to the premotor, posterior parietal, and cerebellar cortex structures. This shift is specific to recall of an established motor skill and suggests that with the passage of time, there is a change in the neural representation of the internal model and that this change may underlie its increased functional stability.

As one practices a motor task, stiffness of the limbs decreases (1), movements become smoother (2), and the muscle activations reflect a reliance of the motor output on an internal model (IM) that anticipates the force requirements of the task (3, 4). In a computational framework, the IM for arm movements may be characterized, in part (5), as a map from a desired trajectory for the hand to a set of muscle torques (6). Because we routinely use our hands to interact with a diverse variety of objects and systems, we rely on visual and haptic properties of the task to act as cues that facilitate recall of an appropriate IM from motor memory (7). Attempting to pick up an empty bottle of milk that has been painted white readily illustrates the consequences of visually cued recall of an inappropriate IM.

A single session of practice with a novel mechanical system may lead to long-term storage of an IM in the brain (8). However, when practice ends, a functional property of the IM continues to develop. Within 5 hours, the recently acquired IM gradually becomes resistant to behavioral interference (8, 9), that is, it consolidates. Although the mechanisms of motor memory consolidation are unknown, examples from other memory systems of the brain show that a change in the neural representation of memory may contribute to consolidation (10). There is also evidence that neural representation of motor function is dynamic (11) and that motor areas of the primate brain are differentially associated with the performance of either a new or well-practiced motor task (12). Here we ask whether with the passage of time, as the IM becomes less fragile, there is a change in the neural representation of its motor memory.

We used positron emission tomography (PET) to monitor changes in regional cerebral blood flow (rCBF), an indirect marker of neural activity, mainly around the synapses (13), as participants (n = 16) learned an IM of a novel mechanical system (Fig. 1A). The dynamics of the novel system were represented as a force field and were produced by the torque motors of a robotic arm (6). The task was to make rapid reaching movements to a series of targets while holding the handle of the robot (14). Participants initially practiced the task with the robot motors turned off (300 targets, during which no rCBF measures were taken). They made accurate, straight movements, similar to that shown in Fig. 1B. In session 1, we acquired rCBF measures (15) as participants performed the task during two repetitions of four successive conditions: (i) during a null field condition in which the robot's motors were off (Fig. 1B); (ii) during a random field condition in which the robot produced a random, nonstationary, velocity-dependent force field representing an unlearnable mechanical system (Fig. 1C); (iii) during early learning of a force field (16) (Fig. 1D), in which the robot produced a stationary force field we labeled “A” and which represented a learnable mechanical system (Fig. 1E); and (iv) during late learning of force field A, in which participants performed the task skillfully (Fig. 1F) after further practice in field A. When participants were first exposed to the forces, movements deviated from the straight-line trajectories (Fig. 1G). In the random field, movements did not significantly improve with practice. However, rapid improvements occurred when the field was held stationary. With practice, the movements gradually converged to those recorded in the null field condition (Fig. 1H). Participants then returned 5.5 hours later for session 2, in which we acquired rCBF measures during two repetitions of only one of the following conditions: (i) recall of the IM of field A (n = 9 participants) or (ii) early learning of field B (n = 7 participants) (16).

Figure 1

The motor learning task. (A) Participants gripped the handle of a robot manipulandum and moved it to targets that appeared on the monitor in one of eight directions: 0 to 315° in 45° increments, randomly selected (8). Participants were provided with continous visual feedback. The task was to reach the target in a precise time (14). We acquired PET scans during five conditions: null field (robot motors not engaged); random force field (robot motors producing a nonstationary field); early learning of field A (16); late learning of field A; and, at +5.5 hours, recall of field A or early learning of field B. (B) Hand trajectories (mean ±SD) for a typical participant during the null field condition. (C) Typical hand trajectories during a random field condition. The robot's motors produced a velocity-dependent force field that randomly varied from target to target. This substantially disturbed the hand trajectories and required corrective movements while precluding the possiblity of learning an IM. (D). The force field A (16). (E) Trajectories (mean ± SD) during the early learning stage of field A (first 100 movements) for a typical participant. (F) Trajectories (mean ± SD) during the late learning stage of field A (last 100 movements) for a typical participant. (G) Length (mean ± SE) of reaching movements during task performance. Each point is an average of eight movements. Gray bars indicate periods of brain image acquisition. There was no significant improvement during the random field condition. However, participants were skillfully controlling their arms during the late stage of learning of A and were able to recall the appropriate IM at +5.5 hours. Performances were not significantly different at recall versus late learning. (H) Hand trajectories for each participant during each condition were correlated with that participant's typical trajectory during the null condition (6). Shown here are the population mean ± 95% confidence intervals (CI). With practice, movements converge to the trajectories recorded in the null field.

We initially asked whether during session 1 there were brain regions where rCBF correlated with measures of total motor output. The average length of a movement was selected as an indicator of motor output (Fig.1G) (17). Statistical parametric maps were generated (18), and we found three regions where activations significantly correlated with motor output: the left sensorimotor cortex (SMC) (–58, –32, 52; Z = +4.81; Fig.2A), with the peak corresponding to Brodmann's area (BA) 4; the right SMC (48, –40, 52; Z = +3.41); and the right putamen (30, –6, –4; Z = +3.49). Changes in rCBF in the SMC have been shown to correlate with arm and finger force production in a task that precluded motor learning (19). Given the significant projections from the SMC to the putamen, it is likely that changes observed in these regions are associated with large-scale reductions in motor output from the random to the late learning condition rather than with acquisition of an IM.

Figure 2

Relative changes in rCBF (with respect to the null field condition) for two areas of the brain that showed increased activation during motor execution or learning. The regions are shown on the normalized MRI of a typical participant (18). Regions of activation are also localized on the rendered brain of the same participant. (A) We looked for regions of the brain where rCBF changes correlated with total motor output (18). The region with the highest correlation was located in the left sensorimotor cortex (peak location at –58, –32, 52; BA 4; Z = +4.81), shown here on a transverse view. The figure also shows the percentage change of the rCBF for all participants at the peak location (mean ± 95% CI). (B) Sagittal view of an area in the right prefrontal cortex (middle frontal gyrus) showing an increase in activity during early learning with respect to the random condition (peak location at 42, 40, 10; BA 46; Z = +3.78). Also shown are the rCBF changes at this location.

Because learning of the IM has components associated with visual perception, force production, attention, and error-reduction processes, a comparison of the adaptation condition with a rest condition does not imply learning-related activity. To test for learning-specific changes, we compared the rCBFs during the random condition, where every component of the task but learning was present, with that of early learning of A. The only significant change was an increase in a region encompassing the dorsomedial and medial pulvinar thalamus (peak at 4, –24, 10; Z = +4.47). This increase was accompanied with increases in the medial occipital gyrus (–14, –94, –12; BA 18; Z = +3.84) and dorsolateral prefrontal cortex (42, 40, 10; BA 46; Z = +3.78; Fig. 2B). There were no significant differences in comparisons of the early and late learning conditions, other than the decreases observed in the SMC (Fig. 2A) and putamen. This suggests that the improvement in performance from the random to the learning condition during session 1 was at least in part due to an increase in activation of visuomotor association areas of BA 46 in the prefrontal cortex (20).

We found that with the passage of time, however, significant changes took place in the representation of the IM. Participants returned 5.5 hours after completion of session 1 and were presented with either field A or a novel field B. Motor performances during the late learning stage of A and the recall stage of A were not significantly different (Fig. 1, G and H). However, there were significant increases in rCBF in three structures: in the left posterior parietal cortex (–30, –78, 54; BA 7; Z = +4.95; late A versus recall A; Fig.3A); the left dorsal premotor cortex (PMC) (–50, –16, 60; BA 6; Z = +3.74; late A versus recall A; Fig. 3B); and the right anterior cerebellar cortex (18, –60, –20; Z = +4.93; early A versus recall A; Fig. 3C). These changes were specific to recall of the previously learned field. The group that was presented with field B during session 2 did not show similar changes, despite the fact that they had an increased motor output (Fig. 1G). We also found that recall of A involved significantly decreased levels of rCBF (with respect to late A) in the left (–46, 32, 28; BA 46; Z = –4.97) and right (42, 28, 20; BA 46; Z = –4.39) middle frontal gyri of the prefrontal cortex. The decreases in rCBF observed in these two regions were 7.0 ± 1.9% and 5.1 ± 2.0% (mean ± 95% confidence interval) for the left and right prefrontal regions, respectively. In comparison, no significant decreases were observed in the prefrontal cortex when participants were presented with field B.

Figure 3

Brain regions showing a significant increase in rCBF from the learning stage of field A to recall of field A at +5.5 hours. The regions are shown on the normalized MRI of a typical participant (18). (A) Transverse view of an area in the left posterior parietal cortex (peak location at –30, –78, 54; BA 7; Z = +4.95; late A versus recall A) and the peak location's changes in rCBF with respect to the null field (mean 95% CI). Shown also are the changes in rCBF from the group of participants that learned a novel field B at +5.5 hours. (B) Transverse view of an area of activation in the left dorsal premotor cortex (–50, –16, 60; BA 6; Z = +3.74; late A versus recall A). (C) Sagittal view of an area of activation in the right anterior cerebellar cortex (18, –60, –20; Z = +4.93; early A versus recall A). The increased activations in these regions were specific to recall of the recently acquired IM of field A.

It has been hypothesized that acquisition of a skilled movement is mediated principally through structures in the prefrontal cortex, and that with time or practice, as the task becomes “automatic,” motor structures such as the cerebellum assume a greater role and possibly become the site of the motor memory (21). There is evidence that in humans, disruption of the prefrontal cortex prevents motor learning without disrupting motor execution (22). In our experiment, acquisition of the IM was associated with increased activation in the dorsolateral prefrontal cortex. Although the performance of our participants neared asymptotic levels during the late stage of learning (23), we did not observe an increased role for the anterior regions of the cerebellum or other motor structures with respect to random or early learning. This is in agreement with a number of other PET studies of motor learning (24, 25). However, we cannot rule out the influence of the cerebellum in initial acquisition of the IM, because posterior regions of the cerebellum were not sampled (18). When the participants were retested at +5.5 hours, there was no significant change in motor performance. However, comparison of rCBFs between recall and late learning stages of A revealed that there was a significant reorganization of the representation of the memory of the IM. With the passage of time, recall of the IM engaged areas of the contralateral dorsal premotor, contralateral posterior parietal, and ipsilateral anterior cerebellar cortex structures. This was coincident with a reduction in activations of the bilateral middle frontal gyri of the prefrontal cortex. The decreased role of the prefrontal cortex has been observed in other studies in which a previously learned motor skill was recalled (25, 26).

A function of the prefrontal cortex is temporary storage of arbitrary sensorimotor information for use in the near term (27). Inherent in this faculty is the transient nature of the associations (28). Previous results on learning control of novel mechanical systems suggest that the representation of an IM in humans is most fragile soon after it has been acquired (8, 9). Within 5 hours after initial practice, the IM's representation becomes resistant to behavioral interference. We have shown here that this change in the functional stability of the acquired memory coincides with a reduced activation in the prefrontal structures and an increase in regions of the brain where long-term motor memory storage has been hypothesized (29).

Recordings of electromyographic activity from the arm during practice of this task suggest that participants gradually learn to recruit new arm muscles and precisely control the timing of activations of these muscles in order to compensate for the force field (30). Studies of similar tasks in highly trained monkeys suggest that the cerebellum is likely to play a critical role in generating this response (3). In humans, cerebellar malfunction results in the loss of ability to anticipate and compensate for interaction torques that are generated in multijoint arm movements (31). Although the role of the cerebellar cortex in initial acquisition of the IM is unclear (32), it has been shown that within an hour after completion of motor learning, biochemical processes that are involved in the synaptic remodeling of Purkinje cells are initiated (33). Therefore, it seems likely that the cerebellum is part of the system that maintains long-term motor memories. On the other hand, lesion, inactivation, and recording studies of the PMC suggest that it is primarily involved in retrieval of a motor response as cued by a visual or auditory stimulus (34). Neuronal recordings show a phase lag between increased activity in some of the cells in the dorsal PMC and behavioral improvement (35). This has suggested that PMC cells function in the retrieval processes of an established visuomotor association, rather than in learning of the association (36). A major input to the dorsal PMC is from the posterior parietal cortex (37). The architecture of this network has been proposed to code reaching movements as the result of a combination of visual and somatic information (38). Indeed, parietal lobe lesions produce apraxia, an impairment of skilled movements in the absence of elementary sensory or motor deficits. Motor memory deficits in apraxic patients suggest a loss of a component of the IM (39).

The results presented here suggest that the representation of a motor skill is reorganized in the brain shortly after an IM has been acquired. Although this reorganization does not affect task performance, it may contribute to increased stability of the representation of the motor skill.


Stay Connected to Science

Navigate This Article