Comprehensive Physiology Wiley Online Library

Motor Learning

Full Article on Wiley Online Library



ABSTRACT

Motor learning encompasses a wide range of phenomena, ranging from relatively low‐level mechanisms for maintaining calibration of our movements, to making high‐level cognitive decisions about how to act in a novel situation. We survey the major existing approaches to characterizing motor learning at both the behavioral and neural level. In particular, we critically review two long‐standing paradigms used in motor learning research—adaptation and sequence learning. We discuss the extent to which these paradigms can be considered models of motor skill acquisition, defined as the incremental improvement in our ability to rapidly select and then precisely execute appropriate actions, and conclude that they fall short of doing so. We then discuss two classes of emerging research paradigms—learning of arbitrary visuomotor mappings de novo and learning to execute movements with improved acuity—that more effectively address the acquisition of motor skill. Future work will be needed to determine the degree to which laboratory‐based studies of skill, as described in this review, will relate to true expertise, which is likely dependent on the effects of practice on multiple cognitive processes that go beyond traditional sensorimotor neural architecture. © 2019 American Physiological Society. Compr Physiol 9:613‐663, 2019.

Figure 1. Figure 1. Motor learning tasks covered in this review and their relation to the pathway from goals to actions. Being skilled in any motor task requires effective goal selection (i.e. where to move to or what to act on), effective action selection (i.e. what movement can achieve the selected goal), and accurate and precise action execution. Improvements at any stage of this pathway can be described as “motor learning”. Different motor‐learning tasks stress improvement at different stages of this pathway. For instance, tasks that involve discrete actions – either as part of a learned sequence or through a learned association with discrete stimuli – require improved action selection and goal selection, but do not require any improvements in action execution. Conversely, tasks that focus on learning at level of action execution (motor acuity paradigms) typically do not involve any learning at the level of goal selection or action selection. Other motor learning tasks (e.g. continuous sequence‐production tasks, adaptation tasks, and tasks that require de novo learning of a new controller) likely engage learning at multiple levels.
Figure 2. Figure 2. Brain regions that contribute to motor learning. Numerous regions throughout the brain have been identified as contributing in some way to motor learning at the level of goal‐setting, selection, or execution. Discussion of the contributions of these regions to specific categories of motor learning paradigms may be found within the corresponding sections of this review. Index to abbreviations: PFC (red): prefrontal cortex; SMA (yellow): supplementary motor area; pre‐SMA (orange): presupplementary motor area; PMd (bright green): dorsal premotor cortex; PMv (cyan): ventral premotor cortex; M1 (dark green): primary motor cortex; S1 (cyan): primary somatosensory cortex; PPC (blue): posterior parietal cortex; hippocampus (pink); cerebellum (purple); basal ganglia (blue). Note that the colors used here are not intended to relate to the colors used in Figure 1.
Figure 3. Figure 3. Force‐field adaptation and aftereffects. This figure illustrates behavior in a typical force‐field adaptation task (). In this study, participants held a robotic manipulandum, illustrated in panel A, and made planar reaching movements toward eight different targets spaced 45° apart. After a baseline, unperturbed period, the manipulandum applied a force proportional to the speed of the hand, and directed perpendicularly to the direction of movement, as illustrated in panel B. Whereas baseline, unperturbed movements were relatively straight (C), the introduction of the force field resulted in movement errors in the direction of the force field (D). After prolonged training, participants adapted to the force field, resulting in straight trajectories (panel E). Interspersed with the training trials were occasional “catch” trials, in which the force‐field was removed, revealing the aftereffects of adaptation, with movements exhibiting errors in the opposite direction to the perturbation as shown in panel F: note how the direction of errors for each target in F is opposite to those experienced in D (and opposite to the direction of the perturbation in B). Panel A is adapted, with permission, from (); panels C to F are adapted, with permission, from ().
Figure 4. Figure 4. Visuomotor rotation adaptation. In visuomotor adaptation studies, participants typically make reaching movements without direct vision of their hand but instead observe the movement of a cursor which represents the location of their hand. During baseline, unperturbed movement, the cursor follows the hand position (veridical visual feedback, A). In B, a visuomotor rotation makes the cursor move in a direction 30° counterclockwise relative to hand motion, resulting in error. After learning, the hand moves at a 30° angle relative to the target direction bringing the cursor directly to the target (C). Removal of the rotation leads to aftereffects (D); moving along the adapted hand direction now leads to a 30° clockwise error.
Figure 5. Figure 5. Forms of deadaptation. Panels are taken from a study () that trained participants on a 30° visuomotor rotation and then subjected them to different modes of deadaptation. A shows the experimental apparatus used. B illustrates the training schedule and the four different modes of deadaptation employed: clamped feedback in red (participants’ movement is projected onto a straight line connecting the start position and the target, leading to the impression of zero error); no feedback in green; washout in blue (participants receive veridical feedback); and time in black (participants do nothing for about 12 minutes—the time it normally takes to complete one of the other deadaptation blocks). The results, shown in C, illustrate considerably faster deadaptation in the washout condition compared to the clamp and no‐feedback conditions. This is quantified in E, which estimates the deadaptation (unlearning) rate for each condition: washout deadapts at a significantly faster rate. This is not surprising: the clamp and no‐feedback conditions remove error feedback, letting the adapted condition slowly revert back to baseline; the washout condition, on top of that, provides an error which actively drives the adapted state back to baseline. At the end of the unlearning block, these three conditions lead to deadaptation as shown in panel D. Allowing for time to pass (black) also results in deadaptation, albeit only partial.
Figure 6. Figure 6. Multiple components of motor adaptation. It has come to be appreciated that motor adaptation is not a monolithic process but instead consists of multiple components. There are multiple different ways in which adaptation can be dissected into its component processes. These panels illustrate how each component responds after the introduction of a perturbation. (A) Using model fits, adaptation can be divided into a fast process, which learns quickly from error but has weak retention, and a slow process, which learns slowly but has strong retention [the simulation shown in this panel uses the parameter values for these two processes as found in ()]. When a perturbation is first introduced, the fast process compensates rapidly. However, as learning proceeds, the slow process begins to compensate in addition, ultimately accounting for the bulk of learning as the contribution from the fast process diminishes. An adaptation coefficient of 1 indicates complete adaptation. (B) Decomposition of adaptation based on whether it can be expressed at low reaction time (). In this study, participants adapted to a 30° visuomotor rotation. In a subset of trials (Low PT) participants were forced to initiate movement with a very low preparation time (∼300 ms), unlike the majority of trials where they were allowed 1.5 s to prepare their movement (High PT trials). (C) Dissection of adaptation into an explicit and implicit component. Figure plotted using adaptation data from (). In this study, participants adapted to a 45° visuomotor rotation. To assess explicit adaptation, participants were asked to report their aiming location. Implicit adaptation was then taken as equal to the difference between the aiming direction and the hand angle.
Figure 7. Figure 7. Implicit adaptation is involuntary and driven by sensory prediction error. (A) In this study (), participants threw darts while looking through prisms which displaced their visual field, leading to errors. This panel shows data from one participant. After donning prism glasses (on trial 13) this particular subject utilized an aiming strategy which led to almost eliminating the error in the next trial. However, continuing to use the same strategy led to involuntary adaptation: while the strategy eliminated task error, sensory prediction errors continued to (involuntarily) recalibrate their throwing, as witnessed by the increasing error after trial 13. On trial 19, the subject was instructed to stop using their strategy and aim directly at the target. Figure re‐plotted from (); used by permission of Oxford University Press. (B) Mazzoni and Krakauer trained participants in a 45° visuomotor rotation task (). Top panel: Experiencing the rotation for two movements (II) led to an increase of about 45° in directional error (y‐axis). At that point, participants in this group were briefed on the nature of the perturbation and provided with a strategy to counter it: aim to a target 45° away on the other direction. This led to an immediate reduction in error close to zero (beginning of III). However, as participants continued to use this strategy, they began to display errors in the opposite direction due to involuntary adaptation as in A. When the participants were instructed to stop using the strategy and instead aim at the intended target (IV), they experienced errors that were significantly smaller than the 45° error that would be expected had there been no recalibration. After the perturbation was removed, participants displayed persistent, slowly decaying aftereffects (V), a hallmark of implicit adaptation. Bottom panel: participants in a further group were not provided with a strategy, but still displayed similar levels of implicit adaptation and aftereffects as the first group.
Figure 8. Figure 8. A forward model for the prediction of the sensory consequences of motor commands (in this case, for arm movements). In this diagram, the forward model receives a copy of the efferent motor command and predicts the sensory consequences of that motor command. If the actual sensory consequence is different than the one predicted, the resulting sensory prediction error will act as a training signal to update the forward model.
Figure 9. Figure 9. Savings in motor adaptation. (A) Savings in visuomotor rotation adaptation [with permission from ()], illustrated by the faster reduction in error during readaptation (black circles) compared to initial adaptation (white circles). (B) Savings in locomotor adaptation [with permission from ()]. In this adaptation paradigm, participants walk on a split‐belt treadmill which can impose different speeds on each leg. The introduction of this leg speed discrepancy reduces gait symmetry (y‐axis, with 0 indicating perfect symmetry). The restoration of gait symmetry is faster during relearning (red) compared to initial learning (blue). (C) Savings in saccadic adaptation [with permission from ()]. This study trained monkeys on a positive (> 1) saccadic gain, had them unlearn it by imposing a gain in the opposite direction, and then had them learn a positive saccadic gain again. The rate of relearning was faster, as indicated by the steeper slope in the learning curve. Panels (B) and (C) republished with permission of the Society for Neuroscience from () and (), respectively; permission conveyed through Copyright Clearance Center, Inc.
Figure 10. Figure 10. Forms of sequence learning. (A) At the level of the task, sequential order dictates the series of complex, potentially multimovement actions that must be performed to satisfy an overall goal, such as the individual steps required to prepare a cup of tea. Learning to perform these steps in the correct order—largely a cognitive form of learning—has little bearing on the quality of how these steps are executed. (B) At the level of controlling a single movement, muscles must be activated in a particular order to successfully change the position of the limb accurately. Although these steps do not reach the level of conscious thought, the ability for the motor system to execute each sequence element in the correct order and with the correct timing critically affects the quality of the resulting movement. Note that, although these two types of sequences are frequently cited as inspiration for studying the learning of sequences, in practice sequence‐learning paradigms largely focus on the acquisition of discrete sequential actions.
Figure 11. Figure 11. Sequence‐specific and sequence‐independent learning. (A) When learning a fixed sequence of keypresses (filled circles), response times typically decrease with practice. In contrast, rehearsal of random sequences (x's) also exhibits some improvements in response time, although these improvements are typically smaller. Figure panels reprinted from (), with permission from Elsevier. (B) Findings such as in (A) suggest that learning during sequence tasks can occur in two ways: improvement of execution of the individual elements regardless of order (sequence‐independent learning), and better performance of the elements in a specific order (sequence‐dependent learning; this includes knowledge of the sequence). In SRTT tasks, insertion of a random‐sequence block (R) toward the end of learning is thought to be a way to distinguish these two forms of learning, although in practice this assay is likely contaminated by additional cognitive influences such as changes in motivation and confidence associated with the unexpected introduction of the random sequence.
Figure 12. Figure 12. Representations of learned action sequences. (A) According to the traditional version of the simple chaining model, each element of a sequence is cued directly by the observable outcome of the prior sequence element. Thus, cuing element A automatically leads to a series of events that produces the remaining elements in the appropriate order (black arrows). This model would predict that cuing an element in the middle of a sequence, such as element C, would result in the execution of a partial sequence of all the remaining elements (blue dashed lines). More recently, it has been suggested that these chains do not necessarily link observed outcomes to action, but could be links that exist in the purely motor or stimulus domain. (B) The hierarchical model proposes that individual elements are organized into increasingly larger chunks that ultimately are assembled into a single sequence representation. While individual chunks can be cued, cuing element C in this case would not lead to a partial sequence since performance of element C does not automatically trigger Chunk 2. (C) Previous research has suggested that sequence chunks can be identified using variations in RT within a sequence. RTs between some pairs of elements tend to be consistently longer than between other pairs, suggesting a hierarchical organization in which elements within a single sequence chunk are triggered together, with additional preparation required between chunks. For example, in this figure, average response times (normalized to baseline performance) are illustrated for each of four practice blocks across a single sequence‐learning session (block 1, filled circles; block 2, open squares; block 3, x's; block 4, filled triangles). Serial positions 5 and 8 have consistently prolonged RTs compared to that of the other elements, and are suggestive of being the beginnings of two successive sequence chunks. Figure panel reprinted from (), with permission from Elsevier.
Figure 13. Figure 13. Evidence for implicit and explicit sequence learning. (A) Response times of Korsakoff patients (filled circles, who have declarative‐memory impairments) and healthy age‐matched controls (open circles) reveal that although patients are in general slower, they can still learn to improve their response times during practice of a fixed sequence (blocks 1‐4). This improvement in response time is greater than that observed during rehearsal of random sequences (dashed lines, blocks 5‐8). Neuropsychological findings such as this provided one piece of evidence supporting the assertion that sequence learning is implicit, as it does not require declarative memory. However, patients do not perform as well as controls in general, and there is no definitive evidence that patients do not rely on declarative memory for trial‐to‐trial learning. Indeed, patients with amnesia do retain the ability to report knowledge of sequence fragments (). Panel reprinted from (), with permission from Elsevier. (B) Despite the presence of a secondary task, participants are able to improve their performance of a first‐order sequence. The extent of learning under dual‐task conditions (measured by the S‐R difference in block 11) was found to be comparable to the S‐R difference when the secondary task was removed (block 15), suggesting that the secondary task had no impact on the extent of sequence learning that took place (i.e., that sequence learning does not require attentional resources). However, the majority of learning that did occur under dual‐task conditions appears to be sequence‐independent learning, as indicated by performance in the Random blocks. Panel reprinted, with permission, from (). (C) Wong and colleagues () contrasted practice of a fully explicit sequence (red diamonds; taught to participants prior to training) and movements toward randomly appearing targets (blue circles). They observed an immediate improvement in response time reflecting the explicit sequence knowledge, and gradual improvements with practice. Because the learning rate of these gradual improvements did not differ between the sequence and the random groups, these practice‐related improvements in response time reflected sequence‐independent learning; there was no evidence of any implicit sequence‐specific learning. Adapted, with permission, from ().
Figure 14. Figure 14. Arbitrary visuomotor association learning. (A) In (), participants were required to associate a set of visual stimuli with specific finger presses. (B) Although the basic mapping was learned quickly, with practice, participants were able to reduce the reaction time needed to generate the correct response. Adapted from (); republished with permission of the Society for Neuroscience, permission conveyed through Copyright Clearance Center, Inc.
Figure 15. Figure 15. Mirror reversal learning. (A) Mirror‐reversal task in which cursor motion is reflected across a mirroring axis. (B) Practice enables accurate compensation to be achieved at lower and lower reaction times. (C) Feedback responses can be assessed by displacing the cursor midway through a movement toward a target positioned on the midline. (D) Responses to target displacement early in learning (blue; when movements are accurate, but require long reaction times) are similar to baseline (nonmirrored) responses (black), rather than ideal responses (dashed black). Even after 2 days of practice, responses fail to be appropriate for the mirror‐reversal and are initially directed in the wrong direction (red). Adapted from (); republished with permission of the Society for Neuroscience, permission conveyed through Copyright Clearance Center, Inc.
Figure 16. Figure 16. A complex de novo learning task. (A) Hand posture is mapped onto a 2‐d cursor position. (B) Initially, participants are unable to effectively control the cursor but, with practice, become able to generate smooth, straight cursor trajectories. Adapted from (); republished with permission of the Society for Neuroscience, permission conveyed through Copyright Clearance Center, Inc.
Figure 17. Figure 17. Examples of prehension tasks. (A‐D) Vermicelli handling task [adapted, with permission, from (), Fig. 1]. A sequence of snapshots showing the task during one trial, from (A) the vermicelli piece being dropped into a conveniently viewed part of the cage, (B) the rat begins eating using an asymmetrical holding pattern, with the two paws designated as “grasp” and “guide” ones, to (C and D) the two paws moving together as the piece becomes shorter and digits become interposed. (E) Food pellet retrieval task [adapted, with permission, from (), Fig. 1]. Monkeys retrieve a food‐pallet from the apparatus, Plexiglas board (Klüver board) containing four small food wells with sizes ranging from 9.5 to 25 mm. (F‐J) A sequence of snapshots showing a monkey retrieving a food pellet [with permission from (), Fig. 2]: (F) finger extension, (G) finger flexion, (H) finger flexion + wrist extension, (I) wrist extension, and (J) Forearm supination. Panels (E) and (F‐J) republished with permission of the Society for Neuroscience from () and (), respectively; permission conveyed through Copyright Clearance Center, Inc.
Figure 18. Figure 18. The arc‐pointing task [adapted, with permission, from ()]. (A) A picture of experimental setup: the participant controls a cursor on the screen via wrist flexion‐extension and pronation‐supination, and try to move the cursor in a clockwise direction through a circular channel. A proreflex infrared camera tracks pointing direction of a reflective marker the participant wears on her index finger proximal interphalangeal joint, projecting it as a cursor on the screen. (B) Representative trajectories before (day 1, top panel) and after (day 5, bottom panel) training. (C) Group‐level performance before and after training. Proportion of within‐channel movements as a function of movement time (MT). The histogram shows the distribution of average MT.


Figure 1. Motor learning tasks covered in this review and their relation to the pathway from goals to actions. Being skilled in any motor task requires effective goal selection (i.e. where to move to or what to act on), effective action selection (i.e. what movement can achieve the selected goal), and accurate and precise action execution. Improvements at any stage of this pathway can be described as “motor learning”. Different motor‐learning tasks stress improvement at different stages of this pathway. For instance, tasks that involve discrete actions – either as part of a learned sequence or through a learned association with discrete stimuli – require improved action selection and goal selection, but do not require any improvements in action execution. Conversely, tasks that focus on learning at level of action execution (motor acuity paradigms) typically do not involve any learning at the level of goal selection or action selection. Other motor learning tasks (e.g. continuous sequence‐production tasks, adaptation tasks, and tasks that require de novo learning of a new controller) likely engage learning at multiple levels.


Figure 2. Brain regions that contribute to motor learning. Numerous regions throughout the brain have been identified as contributing in some way to motor learning at the level of goal‐setting, selection, or execution. Discussion of the contributions of these regions to specific categories of motor learning paradigms may be found within the corresponding sections of this review. Index to abbreviations: PFC (red): prefrontal cortex; SMA (yellow): supplementary motor area; pre‐SMA (orange): presupplementary motor area; PMd (bright green): dorsal premotor cortex; PMv (cyan): ventral premotor cortex; M1 (dark green): primary motor cortex; S1 (cyan): primary somatosensory cortex; PPC (blue): posterior parietal cortex; hippocampus (pink); cerebellum (purple); basal ganglia (blue). Note that the colors used here are not intended to relate to the colors used in Figure 1.


Figure 3. Force‐field adaptation and aftereffects. This figure illustrates behavior in a typical force‐field adaptation task (). In this study, participants held a robotic manipulandum, illustrated in panel A, and made planar reaching movements toward eight different targets spaced 45° apart. After a baseline, unperturbed period, the manipulandum applied a force proportional to the speed of the hand, and directed perpendicularly to the direction of movement, as illustrated in panel B. Whereas baseline, unperturbed movements were relatively straight (C), the introduction of the force field resulted in movement errors in the direction of the force field (D). After prolonged training, participants adapted to the force field, resulting in straight trajectories (panel E). Interspersed with the training trials were occasional “catch” trials, in which the force‐field was removed, revealing the aftereffects of adaptation, with movements exhibiting errors in the opposite direction to the perturbation as shown in panel F: note how the direction of errors for each target in F is opposite to those experienced in D (and opposite to the direction of the perturbation in B). Panel A is adapted, with permission, from (); panels C to F are adapted, with permission, from ().


Figure 4. Visuomotor rotation adaptation. In visuomotor adaptation studies, participants typically make reaching movements without direct vision of their hand but instead observe the movement of a cursor which represents the location of their hand. During baseline, unperturbed movement, the cursor follows the hand position (veridical visual feedback, A). In B, a visuomotor rotation makes the cursor move in a direction 30° counterclockwise relative to hand motion, resulting in error. After learning, the hand moves at a 30° angle relative to the target direction bringing the cursor directly to the target (C). Removal of the rotation leads to aftereffects (D); moving along the adapted hand direction now leads to a 30° clockwise error.


Figure 5. Forms of deadaptation. Panels are taken from a study () that trained participants on a 30° visuomotor rotation and then subjected them to different modes of deadaptation. A shows the experimental apparatus used. B illustrates the training schedule and the four different modes of deadaptation employed: clamped feedback in red (participants’ movement is projected onto a straight line connecting the start position and the target, leading to the impression of zero error); no feedback in green; washout in blue (participants receive veridical feedback); and time in black (participants do nothing for about 12 minutes—the time it normally takes to complete one of the other deadaptation blocks). The results, shown in C, illustrate considerably faster deadaptation in the washout condition compared to the clamp and no‐feedback conditions. This is quantified in E, which estimates the deadaptation (unlearning) rate for each condition: washout deadapts at a significantly faster rate. This is not surprising: the clamp and no‐feedback conditions remove error feedback, letting the adapted condition slowly revert back to baseline; the washout condition, on top of that, provides an error which actively drives the adapted state back to baseline. At the end of the unlearning block, these three conditions lead to deadaptation as shown in panel D. Allowing for time to pass (black) also results in deadaptation, albeit only partial.


Figure 6. Multiple components of motor adaptation. It has come to be appreciated that motor adaptation is not a monolithic process but instead consists of multiple components. There are multiple different ways in which adaptation can be dissected into its component processes. These panels illustrate how each component responds after the introduction of a perturbation. (A) Using model fits, adaptation can be divided into a fast process, which learns quickly from error but has weak retention, and a slow process, which learns slowly but has strong retention [the simulation shown in this panel uses the parameter values for these two processes as found in ()]. When a perturbation is first introduced, the fast process compensates rapidly. However, as learning proceeds, the slow process begins to compensate in addition, ultimately accounting for the bulk of learning as the contribution from the fast process diminishes. An adaptation coefficient of 1 indicates complete adaptation. (B) Decomposition of adaptation based on whether it can be expressed at low reaction time (). In this study, participants adapted to a 30° visuomotor rotation. In a subset of trials (Low PT) participants were forced to initiate movement with a very low preparation time (∼300 ms), unlike the majority of trials where they were allowed 1.5 s to prepare their movement (High PT trials). (C) Dissection of adaptation into an explicit and implicit component. Figure plotted using adaptation data from (). In this study, participants adapted to a 45° visuomotor rotation. To assess explicit adaptation, participants were asked to report their aiming location. Implicit adaptation was then taken as equal to the difference between the aiming direction and the hand angle.


Figure 7. Implicit adaptation is involuntary and driven by sensory prediction error. (A) In this study (), participants threw darts while looking through prisms which displaced their visual field, leading to errors. This panel shows data from one participant. After donning prism glasses (on trial 13) this particular subject utilized an aiming strategy which led to almost eliminating the error in the next trial. However, continuing to use the same strategy led to involuntary adaptation: while the strategy eliminated task error, sensory prediction errors continued to (involuntarily) recalibrate their throwing, as witnessed by the increasing error after trial 13. On trial 19, the subject was instructed to stop using their strategy and aim directly at the target. Figure re‐plotted from (); used by permission of Oxford University Press. (B) Mazzoni and Krakauer trained participants in a 45° visuomotor rotation task (). Top panel: Experiencing the rotation for two movements (II) led to an increase of about 45° in directional error (y‐axis). At that point, participants in this group were briefed on the nature of the perturbation and provided with a strategy to counter it: aim to a target 45° away on the other direction. This led to an immediate reduction in error close to zero (beginning of III). However, as participants continued to use this strategy, they began to display errors in the opposite direction due to involuntary adaptation as in A. When the participants were instructed to stop using the strategy and instead aim at the intended target (IV), they experienced errors that were significantly smaller than the 45° error that would be expected had there been no recalibration. After the perturbation was removed, participants displayed persistent, slowly decaying aftereffects (V), a hallmark of implicit adaptation. Bottom panel: participants in a further group were not provided with a strategy, but still displayed similar levels of implicit adaptation and aftereffects as the first group.


Figure 8. A forward model for the prediction of the sensory consequences of motor commands (in this case, for arm movements). In this diagram, the forward model receives a copy of the efferent motor command and predicts the sensory consequences of that motor command. If the actual sensory consequence is different than the one predicted, the resulting sensory prediction error will act as a training signal to update the forward model.


Figure 9. Savings in motor adaptation. (A) Savings in visuomotor rotation adaptation [with permission from ()], illustrated by the faster reduction in error during readaptation (black circles) compared to initial adaptation (white circles). (B) Savings in locomotor adaptation [with permission from ()]. In this adaptation paradigm, participants walk on a split‐belt treadmill which can impose different speeds on each leg. The introduction of this leg speed discrepancy reduces gait symmetry (y‐axis, with 0 indicating perfect symmetry). The restoration of gait symmetry is faster during relearning (red) compared to initial learning (blue). (C) Savings in saccadic adaptation [with permission from ()]. This study trained monkeys on a positive (> 1) saccadic gain, had them unlearn it by imposing a gain in the opposite direction, and then had them learn a positive saccadic gain again. The rate of relearning was faster, as indicated by the steeper slope in the learning curve. Panels (B) and (C) republished with permission of the Society for Neuroscience from () and (), respectively; permission conveyed through Copyright Clearance Center, Inc.


Figure 10. Forms of sequence learning. (A) At the level of the task, sequential order dictates the series of complex, potentially multimovement actions that must be performed to satisfy an overall goal, such as the individual steps required to prepare a cup of tea. Learning to perform these steps in the correct order—largely a cognitive form of learning—has little bearing on the quality of how these steps are executed. (B) At the level of controlling a single movement, muscles must be activated in a particular order to successfully change the position of the limb accurately. Although these steps do not reach the level of conscious thought, the ability for the motor system to execute each sequence element in the correct order and with the correct timing critically affects the quality of the resulting movement. Note that, although these two types of sequences are frequently cited as inspiration for studying the learning of sequences, in practice sequence‐learning paradigms largely focus on the acquisition of discrete sequential actions.


Figure 11. Sequence‐specific and sequence‐independent learning. (A) When learning a fixed sequence of keypresses (filled circles), response times typically decrease with practice. In contrast, rehearsal of random sequences (x's) also exhibits some improvements in response time, although these improvements are typically smaller. Figure panels reprinted from (), with permission from Elsevier. (B) Findings such as in (A) suggest that learning during sequence tasks can occur in two ways: improvement of execution of the individual elements regardless of order (sequence‐independent learning), and better performance of the elements in a specific order (sequence‐dependent learning; this includes knowledge of the sequence). In SRTT tasks, insertion of a random‐sequence block (R) toward the end of learning is thought to be a way to distinguish these two forms of learning, although in practice this assay is likely contaminated by additional cognitive influences such as changes in motivation and confidence associated with the unexpected introduction of the random sequence.


Figure 12. Representations of learned action sequences. (A) According to the traditional version of the simple chaining model, each element of a sequence is cued directly by the observable outcome of the prior sequence element. Thus, cuing element A automatically leads to a series of events that produces the remaining elements in the appropriate order (black arrows). This model would predict that cuing an element in the middle of a sequence, such as element C, would result in the execution of a partial sequence of all the remaining elements (blue dashed lines). More recently, it has been suggested that these chains do not necessarily link observed outcomes to action, but could be links that exist in the purely motor or stimulus domain. (B) The hierarchical model proposes that individual elements are organized into increasingly larger chunks that ultimately are assembled into a single sequence representation. While individual chunks can be cued, cuing element C in this case would not lead to a partial sequence since performance of element C does not automatically trigger Chunk 2. (C) Previous research has suggested that sequence chunks can be identified using variations in RT within a sequence. RTs between some pairs of elements tend to be consistently longer than between other pairs, suggesting a hierarchical organization in which elements within a single sequence chunk are triggered together, with additional preparation required between chunks. For example, in this figure, average response times (normalized to baseline performance) are illustrated for each of four practice blocks across a single sequence‐learning session (block 1, filled circles; block 2, open squares; block 3, x's; block 4, filled triangles). Serial positions 5 and 8 have consistently prolonged RTs compared to that of the other elements, and are suggestive of being the beginnings of two successive sequence chunks. Figure panel reprinted from (), with permission from Elsevier.


Figure 13. Evidence for implicit and explicit sequence learning. (A) Response times of Korsakoff patients (filled circles, who have declarative‐memory impairments) and healthy age‐matched controls (open circles) reveal that although patients are in general slower, they can still learn to improve their response times during practice of a fixed sequence (blocks 1‐4). This improvement in response time is greater than that observed during rehearsal of random sequences (dashed lines, blocks 5‐8). Neuropsychological findings such as this provided one piece of evidence supporting the assertion that sequence learning is implicit, as it does not require declarative memory. However, patients do not perform as well as controls in general, and there is no definitive evidence that patients do not rely on declarative memory for trial‐to‐trial learning. Indeed, patients with amnesia do retain the ability to report knowledge of sequence fragments (). Panel reprinted from (), with permission from Elsevier. (B) Despite the presence of a secondary task, participants are able to improve their performance of a first‐order sequence. The extent of learning under dual‐task conditions (measured by the S‐R difference in block 11) was found to be comparable to the S‐R difference when the secondary task was removed (block 15), suggesting that the secondary task had no impact on the extent of sequence learning that took place (i.e., that sequence learning does not require attentional resources). However, the majority of learning that did occur under dual‐task conditions appears to be sequence‐independent learning, as indicated by performance in the Random blocks. Panel reprinted, with permission, from (). (C) Wong and colleagues () contrasted practice of a fully explicit sequence (red diamonds; taught to participants prior to training) and movements toward randomly appearing targets (blue circles). They observed an immediate improvement in response time reflecting the explicit sequence knowledge, and gradual improvements with practice. Because the learning rate of these gradual improvements did not differ between the sequence and the random groups, these practice‐related improvements in response time reflected sequence‐independent learning; there was no evidence of any implicit sequence‐specific learning. Adapted, with permission, from ().


Figure 14. Arbitrary visuomotor association learning. (A) In (), participants were required to associate a set of visual stimuli with specific finger presses. (B) Although the basic mapping was learned quickly, with practice, participants were able to reduce the reaction time needed to generate the correct response. Adapted from (); republished with permission of the Society for Neuroscience, permission conveyed through Copyright Clearance Center, Inc.


Figure 15. Mirror reversal learning. (A) Mirror‐reversal task in which cursor motion is reflected across a mirroring axis. (B) Practice enables accurate compensation to be achieved at lower and lower reaction times. (C) Feedback responses can be assessed by displacing the cursor midway through a movement toward a target positioned on the midline. (D) Responses to target displacement early in learning (blue; when movements are accurate, but require long reaction times) are similar to baseline (nonmirrored) responses (black), rather than ideal responses (dashed black). Even after 2 days of practice, responses fail to be appropriate for the mirror‐reversal and are initially directed in the wrong direction (red). Adapted from (); republished with permission of the Society for Neuroscience, permission conveyed through Copyright Clearance Center, Inc.


Figure 16. A complex de novo learning task. (A) Hand posture is mapped onto a 2‐d cursor position. (B) Initially, participants are unable to effectively control the cursor but, with practice, become able to generate smooth, straight cursor trajectories. Adapted from (); republished with permission of the Society for Neuroscience, permission conveyed through Copyright Clearance Center, Inc.


Figure 17. Examples of prehension tasks. (A‐D) Vermicelli handling task [adapted, with permission, from (), Fig. 1]. A sequence of snapshots showing the task during one trial, from (A) the vermicelli piece being dropped into a conveniently viewed part of the cage, (B) the rat begins eating using an asymmetrical holding pattern, with the two paws designated as “grasp” and “guide” ones, to (C and D) the two paws moving together as the piece becomes shorter and digits become interposed. (E) Food pellet retrieval task [adapted, with permission, from (), Fig. 1]. Monkeys retrieve a food‐pallet from the apparatus, Plexiglas board (Klüver board) containing four small food wells with sizes ranging from 9.5 to 25 mm. (F‐J) A sequence of snapshots showing a monkey retrieving a food pellet [with permission from (), Fig. 2]: (F) finger extension, (G) finger flexion, (H) finger flexion + wrist extension, (I) wrist extension, and (J) Forearm supination. Panels (E) and (F‐J) republished with permission of the Society for Neuroscience from () and (), respectively; permission conveyed through Copyright Clearance Center, Inc.


Figure 18. The arc‐pointing task [adapted, with permission, from ()]. (A) A picture of experimental setup: the participant controls a cursor on the screen via wrist flexion‐extension and pronation‐supination, and try to move the cursor in a clockwise direction through a circular channel. A proreflex infrared camera tracks pointing direction of a reflective marker the participant wears on her index finger proximal interphalangeal joint, projecting it as a cursor on the screen. (B) Representative trajectories before (day 1, top panel) and after (day 5, bottom panel) training. (C) Group‐level performance before and after training. Proportion of within‐channel movements as a function of movement time (MT). The histogram shows the distribution of average MT.

 

Teaching Material

J. W. Krakauer, A. M. Hadjiosif, J. Xu, A. L. Wong, A. M. Haith. Motor Learning. Compr Physiol 9: 2019, 613-663.

Didactic Synopsis

Major Teaching Points:

  • Motor learning can be defined as any experience-dependent improvement in performance.
  • Explicit and implicit processes both contribute to how we learn new motor skills.
  • Implicit adaptation serves to maintain motor performance in a fluctuating environment through a sensory-prediction-error-driven learning mechanism.
  • Discrete sequence learning tasks reveal how we anticipate temporal regularities in the environment, but are not likely good models for skilled continuous sequential actions.
  • Many skills, like riding a bicycle, cannot be assembled from pre-existing skills and require building a de novo controller.
  • Motor acuity – the quality of movement execution – can be improved through practice.
  • Implicit adaptation is dependent on the cerebellum.
  • Explicit components of both adaptation and sequence tasks have been shown to have pre-frontal and hippocampal dependencies.
  • Action selection is associated with interactions between the basal ganglia and motor cortex.
  • Motor acuity is accompanied by changes in primary and premotor cortex and cerebellum.

Didactic Legends

The figures—in a freely downloadable PowerPoint format—can be found on the Images tab along with the formal legends published in the article. The following legends to the same figures are written to be useful for teaching.

Figure 1 – Motor learning tasks in the laboratory and their relation to the pathway from goal to action Teaching points: The acquisition of skilled behaviors may involve changes at one or more stages along the action-generation pathway. Common motor learning paradigms interrogate learning at a subset of these stages: goal determination, action selection given the goal, and execution of the selected action.

Figure 2 - Brain regions that contribute to motor learning. Teaching points: Each motor learning paradigm might engage multiple different brain areas. For example, motor adaptation has been shown to largely – though not exclusively – involve the cerebellum, whereas sequence learning has predominantly been associated with changes in neural activity cortically in the supplementary motor area (SMA) and the pre-SMA region just rostral to it, as well as subcortically in the basal ganglia and cerebellum. While we describe speculations about the specific roles of such regions in the text, the guiding hypotheses regarding the contributions of these regions to motor learning remain unclear.

Figure 3 Force-field adaptation and aftereffects.Teaching points: When the dynamical environment is perturbed - for example, using a special force-field as shown in this figure - reaching movements display errors. With training, the motor system adapts to the force-field, resulting in smaller errors and performance that is close to baseline. When the force-field is removed, the commands sent by the motor system are still tailored to countering it, resulting in aftereffects - errors in the opposite direction compared to the ones when the force-field was initially introduced.

Figure 4 Visuomotor rotation adaptation. Teaching points: Visuomotor rotation is a widely used paradigm in motor adaptation studies, which introduces a mismatch between the direction of hand movement and cursor movement. Adaptation to the rotation results in aftereffects after the perturbation is removed.

Figure 5 Forms of deadaptation.Teaching points: Motor adaptation decays with the removal of the perturbation (which results in errors which actively train the adapted state back to baseline), with the removal of errors, removal of visual feedback, or simply by the passage of time.

Figure 6 Multiple components of motor adaptation.Teaching points: We have come to appreciate that motor adaptation is not a monolithic process but instead consists of multiple components. There are different axes along which adaptation can be dissected: into components that are learned quickly or slowly; components that are labile or stable with time; components that are explicit or explicit; and components that require long reaction times to be expressed or not.

Figure 7 Implicit adaptation is involuntary and driven by sensory prediction error. Teaching points: Implicit adaptation is involuntary: it proceeds even in the absence of task error. For example, when participants use a reaiming strategy to counter a visuomotor rotation, they reduce the task error but they still experience sensory prediction error - the discrepancy between expected and displayed cursor position. This sensory prediction error leads to motor adaptation - even in cases where this increases task errors, illustrating that adaptation is involuntary.

Figure 8 A forward model for the prediction of the sensory consequences of motor commands. Teaching points: An important theory behind motor control is that the motor system can feed the efferent motor command to a forward model of the body and environment (e.g. arm dynamics) and predict the sensory consequences of that motor command before actual sensory feedback arrives. This forward model can alleviate the effects of delays in sensory loops and can adapt when exposed to sensory prediction errors – mismatches between the predicted and actual sensory consequence.

Figure 9 Savings in motor adaptation.Teaching points: Savings, the phenomenon whereby relearning a previously learned task is faster than initial learning, has been shown in a wide array of adaptation paradigms such as visuomotor rotation adaptation, split-treadmill adaptation, and saccadic adaptation.

Figure 10 Forms of sequence learning.Teaching points: There are typically two motivating factors cited for why sequences are studied: an ordered series of steps that allow one to complete a broad task goal (such as making a cup of tea), and the organized set of neural control signals required to execute a single movement. Unfortunately, current forays into sequence learning largely focus on the former of these factors, examining how the brain forms compact cognitive representations of an order of discrete actions.

Figure 11 Sequence-specific and sequence-independent learning.Teaching points: Evidence of learning in sequence tasks is typically assayed by decreases in response time, suggesting that individuals become faster with practice. This is often contrasted with individuals practicing non-sequential (randomly ordered) movements to argue for the presence of sequence-specific learning. However, practice even of randomly ordered movements leads to decreases in response time, suggesting that within sequence learning paradigms, practice influences both sequence-specific and sequence-independent components.

Figure 12 Representations of a learned action sequences. Teaching points: Early models of sequence learning proposed that a sequence is learned as a series of outcome-action pairings, such that each observed served as an operant trigger for the next action. In contrast, more contemporary models propose that sequence elements are instead organized into a hierarchical structure such that elements are grouped into chunks, and so on, until ultimately there exists a single representation encompassing the entire sequence. Note, however, that such representations are most likely to be cognitive, not motor; hence, sequence tasks are primarily studies of cognitive representations.

Figure 13 Evidence of implicit and explicit sequence learning. Teaching points: Motor learning in sequence tasks (i.e., implicit learning) has been claimed based on two lines of evidence. The first is that patients with amnesia tend to learn sequences similarly to controls, according to estimates of the change in performance between performing fixed and random sequences. The second is that healthy participants are able to learn sequences under dual-task conditions, suggesting that sequence learning may not require attention. However, other studies have found contradictory conclusions to these claims, questioning these findings. For example, a recent experiment examined the extent to which explicit knowledge contributes to sequence learning. This study found that all of the improvements occurring within sequence tasks arise from two sources: sequence-independent rehearsal of the individual movement elements (regardless of their order), and explicit knowledge of the sequential order. Hence there is little convincing evidence at present that supports the existence of true implicit sequence-specific learning, i.e., that sequence-learning tasks contain both the explicit and the implicit learning components that together comprise motor learning.

Figure 14 Arbitrary visuomotor association learning. Teaching points: The arbitrary visuomotor association paradigm challenges participants to generate specific responses to stimuli. For instance, in (A), participants must associate abstract visual stimuli with pressing a key with a specific finger. This learning paradigm presents two challenges. The first is initial learning of which symbols map to which actions. Second, once this mapping is identified, participants can improve the speed at which they are able to respond through practice. This is illustrated in (B), practicing this mapping from visual stimuli to symbols leads to a gradual decline in reaction times across sessions. This paradigm encapsulates the transition from a skill being initially mediated by slow, cognitive mechanisms (identifying the mapping) to becoming mediated by more rapid and automatic mechanisms following practice (111). Adapted from (22).

Figure 15 Motor acuity. Teaching points: The arc-point task is one of the few tasks designed with the focus on learning of motor acuity the quality of movement execution. This task requires participants to make continuous curved movement and navigate through a U-shaped tube. Plots in (C) depict a shift of the speed-accuracy trade-off function (SAF) before and after training, indicating a change of skill level before and after training: participants showed improved performance at all MTs.

Figure 16 Mirror reversal learning. Teaching points: Mirror reversals are more challenging to learn than a visuomotor rotation since they require learning of a de novo policy rather than recalibration of a baseline controller. (A) In this paradigm, the location of an on-screen cursor is given by the location of the hand, reflected across a mirroring axis, in this case horizontally. (B) Participants generally are able to recognize this perturbation and learn successful compensate quite quickly. At first, however, they require quite long reaction times, presumably in order to deliberate about the required movement direction. Through practice, however, they become able to generate the required movement much more rapidly. (C) Participants’ capability in the task can also be probed by testing how they respond to a displacement of the cursor midway through a movement towards a target positioned on the mirroring axis. (D) At baseline, the appropriate response to this rightward cursor displacement is to move the hand in the opposite direction to the cursor displacement, i.e. to the left (black line). Under the mirror reversal, the appropriate correction is opposite, i.e. participants must move the hand in the same direction that the cursor moved. (dashed line). Even after participants have learned to make feedforward movements that successfully compensate for the mirror reversal, their feedback responses to the cursor displacement do not reflect learning of the reversal, remaining very similar to baseline. After a day of practice, the feedback response begins to become more appropriate for the mirror-reversal, but still initially reflects the baseline pattern of compensation. Adapted from (438).

Figure 17 A complex de novo learning task. Teaching points: Many tasks challenge participants to learn entirely unfamiliar control schemes. (A) In this example, the posture of the hand (measured through a special glove that can record 19 degrees of freedom of the hand and fingers) is mapped to an on-screen cursor location (314). Initially, participants struggle to control the cursor, moving randomly until the target is acquired. (B) With practice over multiple sessions, however, they become able to move more rapidly with less error. This task illustrates how, through practice, people are able to learn to manipulate effectors under entirely arbitrary control structures, similar to the way in which we learn to drive a car, or play a video game. Adapted from (77).

Figure 18 Examples of prehension tasks. Teaching points: Prehension tasks have been extensively used to study motor skill learning in the animal model literature. Two examples are shown in this figure. In rodents, the vermicelli-handling task is a sensitive tool to measure forepaw dexterity. Rodents learn skillful forepaw and digits movements to manage thin pasta pieces. Typically, dexterity is measured by counting the number of adjustment using each paw. In monkeys, the food-pellet retrieval task prehension skill can be trained from larger to smaller wells.  Typical measures are number of successful retrieval and numbers of finger flexions per retrieval.

 


Related Articles:

Teaching Material

Contact Editor

Submit a note to the editor about this article by filling in the form below.

* Required Field

How to Cite

John W. Krakauer, Alkis M. Hadjiosif, Jing Xu, Aaron L. Wong, Adrian M. Haith. Motor Learning. Compr Physiol 2019, 9: 613-663. doi: 10.1002/cphy.c170043