Motor sequence learning
|Edwin Robertson and Ádám Takács (2018), Scholarpedia, 13(5):12319.||doi:10.4249/scholarpedia.12319||revision #186421 [link to/cite this article]|
A broad range of human behaviour is dependent upon sequences. Organizing actions, events, words, memories and thoughts into a sequence is a critical part of everyday behaviour (Ashe et al, 2006). Performing these sequences can initially be challenging; yet, with sufficient practice they can come to be performed almost effortlessly. A failure to organize actions into the correct sequence is a feature of neurological impairments, such as the movement disorder apraxia. Equally, our cognitive lives depend upon sequencing for organizing our thoughts, for instance, into clear and intelligible language constrained by grammatical rules.
A sequence is a list of elements, such as events, cues, or actions, which follow an order. The order may be determined by a rule; such as no repetitions, and as a consequence -A-A- would be illegal. A rule may operate at the level of the specific items determining what specific item appears; for example, 2nd or 7th within a sequence. Alternatively, a rule may operate at a more abstract level with particular sets of items preceding a different class of items. Regardless, learning occurs over multiple repetitions of the sequence. Motor sequence learning broadly refers to the process by which a sequence of movements comes to be performed faster and more accurately than before. It falls into the broad category of procedural knowledge, as opposed to declarative knowledge, and refers to knowing how to perform certain activities (Cohen et al, 1980; Willingham, 1997). An individual can have little or no awareness of learning a new skill (implicit learning), or alternatively, it is also possible for them to be aware of learning a new skill (explicit learning). Many experimental approaches have been developed to better understand sequence learning as it potentially holds the promise of providing fundamental insights into human cognition and into a diverse range of human diseases. One of the most popular tasks designed specifically to address issues of sequential learning is the Serial Reaction Time Task.
The Serial Reaction Time TaskThe Serial Reaction Time Task (SRTT) has been widely used to understand human motor sequence learning (Figure 1Nissan and Bullemer, 1987, Willingham et al, 1989). It is a deceptively simple task (for a review please see Robertson, 2007). In its original form participants perform a choice reaction time task involving visual stimuli: a visual target can appear at any one of four positions arranged horizontally on a computer screen. Each screen location, designated 1–4, corresponds to a button on a response pad. At the start of each trial, a visual target appears and a participant selects the appropriate response button, which ends the trial (Figure 1). The interval between the presentation of the stimulus and selection of the correct response defines the response time, which is the primary measure of performance in this task. At the end of each trial, there is a short fixed delay called a response stimulus interval (RSI), which often varies between 200 and 500ms before another target is presented. The visual targets can play out a repeating sequence of positions, typically 6-12 items long (for example, a 12-item sequence such as 2-3-1-4-3-2-4-1-3-4-2-1 that is repeated multiple times) or the visual targets can be presented in a random order. During practice, participants’ response time during the sequential trials decreases, while response time during the subsequent random trials remains elevated.
The SRTT has provided insights into various aspects of learning and a wide variety of cognitive processes. For instance, it has provided insight into the circuits involved in sequence learning, and how these are affected by awareness for learning. Moreover, it has been used to study the development trajectory of sequence learning from children, to adolescents, to older adults (Howard and Howard, 1997), and learning in neuropsychological patients and even animals (Conwar and Christiansen, 2001; Christie and Dalrymple-Alford, 2004). Therefore, the SRTT is simple, yet powerful, giving insights into a wide array of processes underlying a wide range of behaviours, including but not limited to cognitive and biological principles of learning and memory. Here we initially focus on methodological aspects of the task, and then subsequently discuss the innovative changes that have been made to the task since its introduction over 20 years ago.
Measures of motor sequence learning
A simple and intuitive measure of motor sequence learning is a reduction in response time during training. Response times can, however, be reduced for independent reasons. In the SRTT, response time gradually decreases across the sequenced trials due to participants’ growing expertise in performing the sequence but also in learning the visuomotor association, or “mapping”, between the position of the visual cue and the required response.
To factor out the influence of contaminating factors (such as visuomotor associations, fatigue and motivation) a widely used approach is to contrast the sequential response times against response times for random trials that occur after the sequential trials. This analysis provides a skill measure specific to the sequence. There is also a subtle, but nonetheless important advantage of using the difference between sequential and random response times over using sequential response times alone. When the sequence is unexpectedly removed and replaced with random trials, the participant initially continues to inappropriately play out the sequence. This mistake inflates the random response times, increasing the difference between the sequential and random response times. Thus, the difference between sequential and random response times provides both a specific and sensitive measure of skill acquisition in the SRTT.
Measures of Awareness
Individuals can learn a sequence of movements whilst being unaware of the underlying sequence, so called implicit learning, or alternatively, they can be aware of the sequence, so called explicit learning. Awareness for the sequence can be acquired simply by telling participants about the sequence prior to practice. Frequently, awareness for the sequence develops as participants practice the task (Willingham et al, 1989; Honda et al, 1998; Willingham et al, 2002; Robertson et al, 2004). An important challenge is to identify when awareness develops during practice, and define the extent of that awareness.
One of the strengths of the SRTT is that many complementary techniques have been devised for measuring awareness in this task. For example, explicit learning can be defined as recalling four or more items of the sequence during a free recall test (see for example; Willingham et al, 2002; Robertson et al, 2004). Other tests require participants to successfully recognize either a segment of the sequence, or the sequence itself. A flaw inherent in many of these tests is that they may lack the necessary sensitivity or specificity to detect awareness (Hannula et al, 2005). Awareness tests should tap into exactly the same knowledge upon which performance and learning are based but they also have to be sensitive to all of the relevant conscious knowledge. Verbal reports (such as, a free recall test) may not be sensitive enough: subjects might fail to report verbally knowledge that they hold with low confidence. Thus, tests of awareness may be neither specific, nor sufficiently sensitive to identify when an individual has become aware of the learning a sequence.
An elegant approach to overcoming sensitivity and specificity challenges in detecting awareness was devised almost two decades ago. This process dissociation procedure (PDP) was developed to provide separate estimates of implicit and explicit knowledge by setting those in opposition to one another (Destrebecqz and Cleeremans, 2001; Curran, 2001). Participants are given inclusion and exclusion tasks to assess their explicit knowledge of the sequence. In the inclusion task the instructions are to recreate the sequence they have just learned, in the exclusion task they are asked to generate a new sequence that does not share similarities with the learned sequence. Under the inclusion instructions, participants should generate a sequence that they recollect or that is merely familiar. In contrast under explicit instructions, if participants reproduce a sequence despite being instructed not to, they are likely unconsciously accessing their implicit knowledge of the sequence. The PDP was introduced as a general-purpose tool for assessing how the performance at cognitive tasks is mediated by consciously-accessed versus unconsciously-controlled processes. Rather than relying on two distinct tests, the PDP relies on a single task that is performed under two different instructions. The limitation of this method lies in the complexity of the instructions, making it unsuitable for some populations (e.g. children, amnesic patients).
Another novel method introduced to measure the awareness is post-decision wagering. Subjects after making a decision on a task are required to bet on its correctness by making either a high or a low wager (Persaud et al, 2007). If the decision is correct, the subjects earn the wagered amount of money otherwise they lose the money. The design offers an opportunity for the subjects to turn their performance into financial profit. Post-decision wagering may be able to distinguish between implicit and explicit learning with greater sensitivity and specificity than confidence judgements.
Awareness is frequently treated as a binary state. Participants are assumed to be either aware or unaware of a sequence, and it is expected that there is a specific moment at which an individual is transformed from being unaware to being aware of the sequence. However, awareness may be better conceptualized as a continuum. Overall, seeking to define or identify when awareness is achieved is fraught with challenges. Potentially, it may be more realistic to seek to compare the awareness achieved by participants. For example, comparing what scores participants achieve in particular tests may reliably identify those with greater awareness; although, not be as reliable as tests about the absolute awareness achieved by any participant or group of participants.
Overall, measuring an individual’s awareness for a sequence is challenging. Elegant techniques including the process dissociation procedure and wagering have been developed to provide both a sensitive and specific measure of awareness. A complementary approach is to directly manipulate awareness, for example, through the instructions provided to participants. Using these different approaches either alone or together are likely to provide new insights into the contribution of awareness to learning.
The Structure of Sequences
The structure of sequences has crucial effect on learning. The most important elements of the structure are: frequency-based statistical structure (how often each elements occurs in the sequence), relational structure (the occurrence of a specific element after or before a different element), temporal structure (the occurrence of elements in time), and spatial structure (the position of the different elements in space; Figure 2).Elements can appear at different frequencies within a sequence (Figure 2aCleeremans and McClelland; Stadler, 1992; Boyer et al). For a 12-item sequence with 4-elements (2-3-1-4-3-2-4-1-3-4-2-1) each element appears 3 times, and so frequency learning cannot play a role. By contrast, 10-item sequences with 4-elements, which are often used in studies, inevitably have elements appearing at different rates. In those cases, frequency learning could account for some of the performance improvement.
Related to item frequency is the concept of “repetition distance”, or the number of items that have passed since the last appearance of an item (Figure 2b; e.g., the repetition distance between the first and subsequent instance of 1 in the sequence 1-3-2-4-3-2-4-1 is 7). Repetition distance can impact sequence learning based on expectations of when a given element will appear. Participants have a priori expectations that elements will appear equally often. Attention is directed toward the items that are most likely to occur next, in SRTT these items will be ones that have not occurred recently.
Sequences can further be differentiated by ordinal and relational components (Figure 2c). Relational information does not depend on the absolute order of the elements but rather, the relationship between one item and a subsequent item (e.g. in A-B-C-D, C follows B in the sequence; (Schuck et al). Alternatively, what is learned are the ordinal positions of items: each item it associated with its linear position in the sequence (e.g. in the sequence 4-2-3-1, 4 is the first item to appear; (Schuck et al). Some recent work nicely demonstrates the importance of the distinction between ordinal and relational components. Relational skill was gained during training but not offline over a night of sleep; whereas, ordinal knowledge was not gained during training but did improve offline over a night of sleep (Song and Cohen, 2014; for a discussion on offline processing (Robertson et al; Robertson, 2009; King et al, 2014)). Moreover, single-cell electrophysiological studies with primates have shown that different brain areas are involved in ordinal and relational processing of sequence learning (Tanji and Shima, 1994; Isoda and Tanji, 2004). The presence of differential anatomical representations of relational and ordinal components, and the double dissociation of online and overnight gains serves to demonstrate the biological and cognitive importance of the relational and ordinal components of a sequence.
The formation of associations between the elements within a sequence can extend beyond adjacent items. Predicting the next event (n) within so-called high order sequences requires knowledge of the two immediately preceding events (i.e. (n-2) and (n-1); Figure 2). For instance, in the sequence 1-2-1-4-2-3-4-1-3-2-4-3 a 2 (i.e., n-1) can predict a subsequent (i.e., n) 1 (2-1), 3 (2-3) or a 4 (2-4). It is only by looking further back in the sequence (i.e., to n-2) can a unique prediction be made. By contrast, predicting the next event in so-called low order sequences only requires knowledge only of the preceding event (i.e., n-1 successfully predicts n; Figure 2d). This distinction between high and low order sequence has an important anatomical basis. Learning higher order sequences is dependent upon the medial temporal lobe (MTL); whereas learning low order sequence occurs independently of the MTL (Curran, 1997, Schendan et al, 2003). Thus, the engagement of the MTL is related to the computational requirements of that task (high order vs. low order).
Other types of Motor Sequence Learning Tasks
There are a rich variety of different types of sequence learning. Some of these have been designed to examine particular components of learning while others have been designed to examine the effects of being distracted during learning. Equally, the SRTT has served to inspire the creation of substantially different types of sequence learning task.One example is a probabilistic variant of the SRTT that aims to minimize the contamination of implicit learning by explicit knowledge of the sequence. Probabilistic variants of the SRTT change the surface elements of the sequence at each iteration while the abstract grammatical rule is kept constant throughout the experiment (Figure 3). In one version, the order of the stimuli is determined by a set of rules that describe allowable transitions between stimuli (Cleeremans and McCLelland, 1991, Peigneux et al, 2000). The sequence may be a simple repeat of X, Y, and Z, with each letter representing a category, and each category containing specific sequence items (Figure 3aSong et al, 2007).
One type of probabilistic task is the Alternating Serial Response Time Task (ASRTT) in which sequenced trials are interleaved with random trials (Howard and Howard, 1997; Song et al, 2008). For example, the sequence is 2-R-4-R-3-R-1 where R represents any of the four elements chosen at random. Triplets beginning with 2 and ending with 4 are high frequency whereas those starting with 4 and ending with 2 are of low frequency (Figure 3b). Learning is frequently measured as an increasing difference in response time between the high and low frequent triplets. The ASRTT provides a continuous learning measure (less predictable trials can be compared with more predictable ones throughout the learning). Due to the higher complexity of the sequence it also reduces the likelihood that participants will identify, recognize or recall the sequence (Song et al; Song et al, 2007). Thus, participants are typically unaware of the alternating sequence or of the varying triplet frequencies.Another type of learning task loosely based on SRTT is the Triplet Learning Task (TLT; Figure 4; Howard et al, 2008). Unlike in the SRTT, in the TLT the trial consists of three visual cues (triplet) appearing one after the other. The first two cues are red and the third one is green. Participants are asked to observe the red cues and respond only to the green cue, called the target. The structure, which the participants implicitly learn is as follows: the location of the first green cue predicts the location of the third cue, while the location of the second cue is random (e.g. 3-R-1, 2-R-2, 1-R-3 where R is a random location of the cue; Figure 4Willingham, 1999). In TLT statistical dependencies are introduced among the three visual cues. By varying the frequency with which the triplets occur during the training, it is possible to manipulate how predictable the target is. Sequence learning is calculated by comparing performance on high and low probability triplets throughout the training. The advantage of the triplet task is that it possesses a minimized motor requirement that enables studying sequence learning in special populations (e.g. aging adults). Another commonly used type of sequence learning is the finger-tapping task (Figure 5; Kami et al, 1995; Walker et al, 2003). In this task, participants repeatedly tap out a known fixed sequence. The number of sequences correctly completed in 30 seconds provides a measure of skill (Figure 5a
Serial reaction time tasks apart from exploring motor sequence learning have also been used to gain insight into the learning of temporal sequences, in which the time between sequence elements can follow a repeating pattern. Manipulation of temporal structure allows sequential timing of the elements’ appearance, in addition to the order of elements, to be studied (Shin and Ivry, 2002). Learning of temporal sequences is measured by comparing response times in blocks of sequences that are sequentially and temporally matched versus those that are only sequentially matched.
With an understanding of the basic types of SRTT and the factors that affect motor sequence learning, we can now look at the more complex variants of SRTT. There are a number of task components that may be manipulated to investigate how the successful learning of a sequence is achieved. One of them is the adding a secondary task while learning SRTT.
Dual Task Sequence Learning
In everyday life, learning is often distracted by other tasks. Nonetheless, individuals are still able to learn and so improve their performance. Understanding how learning can still occur despite distraction has been explored using dual task experiments. An additional task, frequently, a tone counting task can be performed simultaneously with motor sequence learning (Figure 5b; Seidler et al, 2008; Lehericy et al, 2005). For example, a participant may learn a motor sequence while being asked to keep track of the number of high pitched audio tones played during the experiment and must report the count at the end of the task. Performance on a dual task can be used to identify that aspect of a learnt skill that has become automated (Lehericy et al, 2005). What has remained poorly understood is how the additional secondary task impairs performance.
Over the years, various hypotheses have emerged to explain the impaired performance in dual tasks (Nissen and Bullemer, 1987; Schmidtke and Heuer, 1997). Impaired performance may be due to the increased attentional demands. In a dual task, attention needs to be directed to performing sequences while simultaneously listening for the tones and maintaining an accurate cumulative count of the high or low-pitched tones (Curran and Keele, 1993). However there are alternative explanations for the decreased performance during dual task learning. Tones may increase the number of uncorrelated events within the task, which may make it more difficult to perform the sequence (Keele et al, 2003). In essence, the tones are increasing the “noise” within the task making it more difficult to perform the sequence. Another explanation is that performing the sequence and tone counting are integrated together (Rah et al, 2000). Rather than there being two discrete tasks there is really only one single but highly complex single task. Participants attempt to integrate the visual and auditory stimuli into one sequence, which lowers performance (Schmidtke and Heuer, 1997; Rah et al, 2000). In sum, performance may be impaired under dual task conditions due to the increased attentional demands of the task, an increase in noise due to the addition of the tones, or because the tones become integrated within the sequence to create a complex multidimensional task. Yet, whilst performance is impaired sequence learning continues, with tone counting impairing the expression of that learning.
The dual-task provides an opportunity where learning and performance changes can be dissociated. In an elegant set of fMRI experiments, participants performed an SRTT with and without a visual distractor task. With the distractor, performance did not change between sequenced and random parts of the SRTT, in contrast, when the distractor was not present, sequenced trials were performed significantly faster than randomized trials. It allowed participants to encode the motor sequence without demonstrating an improvement in motor performance (Seidler et al, 2002). The distractor task served to suppress performance change but did not prevent learning. Thus, using a dual task design it became possible to distinguish the neural process of motor skill acquisition from its expression.
Overall, the dual task design has given powerful insights into the learning process. For example the design allows the processes associated with learning to be distinguished from those associated with performance of a motor sequence. The mechanisms responsible for the impaired performance of a motor sequence during dual task conditions are still not completely understood. Nonetheless, the dual task sequence learning has provided the basis for elegant work that has allowed performance to be distinguished from actual learning.
Chunks in Sequence Learning
The human brain clusters information into chunks. When we recall a sequence of numbers, such as a phone number 070470078313, we might divide the sequence into a series of chunks such as 070 4700 78313, or into 07047 007 8313. By segmenting a sequence of elements into chunks information becomes easier to learn, retain and recall in the correct order (Chase and Ericsson; Bo et al, 2012; Bo et al, 2012). The strategy of breaking down sequences into chunks may be used for those sequences that we can verbally recall; such as a phone number, and also in a wide array of other tasks. Chunking offers a powerful mechanism for individuals to escape the limitations imposed by short-term memory. A single sequence is initially learned as several short segments or chunks. With continued practice, these chunks may become concatenated together, so that eventually, in motor sequence learning, a string of movements is produced as a single unit. Thus, chunking is a potentially powerful mechanism for enhancing learning. A number of different approaches have been used to identify chunks.
Traditionally, chunks in a movement sequence have been identified by increased pauses between successive actions or an increase in error at the start and end of chunks (Lashley, 1951; Diedrichsen and Kornsheva, 2015; Sakai et al, 2003). Such pauses and errors are usually found when there is a change in the pattern of movements (Koch and Hoffmann, 2000). For example, the sequence 1-2-3-4-3-2-1 can be decomposed into two main chunks. The 1-2-3 chunk creates an ascending “run”, while the 3-2-1 creates a descending run. The transition from one chunk, 1-2-3, to the other, 3-2-1, is marked by an increase in error and response time. Thus, at least in principle, chunks may be identified by simple changes in performance. Yet, chunks are rarely as universally obvious and consistent across participants as ascending or descending runs. Instead, the chunks created from a sequence can frequently depend upon the idiosyncrasies of an individual. Even a relatively simple sequence can be divided into distinct chunks in different ways. One approach to address this challenge is to design specific sequences; while, another approach is to use analytical techniques, not to identify chunks per se; but the consequences of chunks being formed on the temporal structure of responses during sequence learning.Chunks have been consistently detected in “hyperset” sequence learning designs. In this design, a hyperset is built up from five components “sets” each of which consists of two fixed button presses (Figure 6). Only two stimuli are presented as a set each time, and subjects are required to press the two corresponding keys successively. Learning is accomplished by trial-and-error of the correct order of the two buttons. Upon successful completion of one set, subsequent sets are presented in succession. Participants continuously practice the same sequence, or hyperset throughout the experiment for a fixed amount of time; say for 30 minutes. The number of chunks is identified by testing for a significant pause between elements of the hyperset sequence (Figure 6Hikosaka et al, 1995; Hikosaka et al, 1999). A clear strength of the hyperset is that the chunks are dictated by the design with any chunk being composed of a combination of sets. Overall, the hyperset design encourages the emergence of chunks, provides a clear and reliable way to identify those chunks, and so provides a platform to understand the mechanisms supporting their creation.
The creation of chunks alters the temporal structure of a movement sequence. These temporal features can be measured using time series analysis, and specifically autocorrelation (Verstynen et al, 2012). Autocorrelation performs a comparison between a signal, and a delayed copy of that signal. A signal in this case could be made of either response time or errors. Comparing a set of response times, against a delayed copy identifies similarities within the response times as a function of the lag between them. For example, a tight correlation between an element n, and the successive element, n+1 would be identified as a lag 1 within an autocorrelation. Small chunks of say 3-4 items within a sequence would be anticipated to give rise to a longer lag or lag 3 or lag 4. As learning continued the chunks might reasonably be expected to become larger, which can be observed as an increasing lag detected by an autocorrelation (Verstynen et al, 2012). For instance, after prolonged training over multiple successive days the lag can become as high a lag 7. Overall, a time series analysis such as an autocorrelation provides a straightforward way to detect changes in the temporal structure of responses that are consistent with the formation of chunks.
Despite its theoretical appeal, the creation of chunks may not be inextricably linked to sequence learning (Song and Cohen, 2014). Chunking does not change or improve during sequence learning, and nor does it differ depending upon whether participants are or are not aware of learning. In other words motor learning was observed to progress independently of any chunking process. As a consequence, although chunking may be found within many sequences, the creation of chunks is not linked, and it may not even be critical to sequence learning.
Overall, chunking may be a fundamental strategy to acquire long, complex sequences. Simple measures of performance can identify chunks, and time series analysis can demonstrate the effect of chunking upon the temporal structure of responses during learning (Hikosaka et al, 1995; Verstynen et al, 2012; Verwey, 2010; Wymbs et al, 2012). Nonetheless, some work has demonstrated that chunking may not be inextricably linked to sequence learning.
Dimensions of motor sequence learning
Motor learning versus perceptual learning
The importance of perceptual and motor processes in sequence learning is not well understood. It has been suggested, that perceptual processes only, such as observing the moving targets on the screen are enough to learn contingencies presented in the SRT task, without asking the participants to press corresponding buttons during training (Dennis et al, 2006; Howard et al, 1992). However, other studies showed that without a motor component, only explicit sequence-learning can be achieved in the SRTT (Willingham et al, 1989; Willingham, 1999). Potentially, perceptual information is sufficient for explicit learning, whereas, a motor response is necessary for implicit learning. Unravelling the relative importance of perceptual and motor information taps into the broader question of what information is being acquired, and stored during sequence learning.
Simply watching a sequence on a screen leads to a level of performance that does not differ substantially from undergoing traditional motor training (Howard et al, 1992). Perceptual and motor training both yielded higher sequence-knowledge than simply watching random cue changes. Thus, a sequence can be acquired simply through observation. Yet, perceptual or random cue training lead to a performance on the subsequent sequence that were no longer distinguishable after removing participants who gained conscious knowledge of the task structure. Overall, a sequence can be acquired through perceptual learning, and this may only be effective when participants are aware or become aware of the sequence.
Directly comparing groups with motor training against those with that only observed the task has limitations. Participants in the perceptual learning groups may not have attended to each stimulus. As a consequence, lack of implicit sequence learning in the perceptual condition may result from low motivation to engage in the task (Willingham et al, 1989; Willingham, 1999). Perhaps of greater concern is that to observe the sequence, participants may have simultaneously been making a sequence of eye movements. Any skill acquired at performing a sequence of eye movements could, at least in principle be transferred to the hand (i.e., oculomanual transfer). Thus, the observation group was not purely perceptual because it was contaminated by covert motor learning due to eye movements.
Many of these limitations have been overcome by recent elegant work. A study used a visuo-auditory word categorization task analogue to the SRTT in young and elderly populations (Dennis et al, 2006). First, the response mapping was shown on the screen, indicating which response was associated with which word category. Then, a target word from the four possible categories was presented auditory, and participants had to press the corresponding button. In this way, the stimulus-response relationship was remapped on a trial-to-trial basis to eliminate the motor aspect of learning. That is, the executed motor response was irrelevant to the sequence. At the same time, participants were required to actively attend to the stimuli, and make decisions, leaving out the possibility of low task engagement. Learning was present even without explicit knowledge of the task structure. Thus, sequence learning is possible through perceptual cues only, and it does not require motor learning (Dennis et al, 2006). However, elderly participants do not show significant learning in this task perhaps because removal of the motor cues diminishes the contextual support for learning. That is, individuals with limited learning capacities, or those who face complex sequence learning tasks, could rely more on the information provided by motor learning. Thus, sequence learning can be guided through purely perceptual cues, and also benefits from motor cues.
Perceptual learning of a sequence can be achieved without being aware of the underlying sequence. As a consequence, perceptual learning is not tied to explicit knowledge. Yet, the converse relationship the link between explicit learning and perceptual learning is less clear, and may depend upon how explicit learning was achieved. For example, achieving explicit knowledge through practice would imply that some perceptual learning had already occurred. Conversely, achieving explicit knowledge by being informed of the underlying sequence would imply that initially there was little or no perceptual learning despite sequence learning being explicit. Thus, explicit learning may not be tied to perceptual learning.
Goal component versus movement component
Classically, a distinction has been made between how a movement is performed and the goal of the movement (Robertson, 2009; Brooks, 1986; Robertson and Cohen, 2006). Learning can occur as a set of finger movements, a so-called egocentric co-ordinate frame; alternatively, learning can occur as a set of goals or targets, a so-called allocentric co-ordinate frame. Learning occurs concurrently within both these co-ordinate frames not only during sequence learning, but also for reaching movements in novel force fields, and spatial navigation through mazes (Robertson, 2009; Robertson and Cohen, 2006).Intermanual transfer (shown in Figure 7Willingham et al, 2000; Verway and Clegg, 2005; Verwey and Wright, 2004; Cohen et al, 2005). Switching hands makes it possible to distinguish between these skill components: (a) maintaining the goal (for example, -2-4-3-) but altering the order of finger movements (goal configuration) measures the skill derived from knowledge of the goal (i.e., knowledge of the sequence, independent of the fingers used (Verwey and Wright, 2004; Cohen et al, 2005)), whereas (b) maintaining the order of finger movements (for example, -middle-little-ring-) but altering the goal (movement configuration) measures the skill derived from the finger movements (i.e., knowledge of the specific finger movements, independent of the sequence of response buttons (Verwey and Clegg, 2005; Cohen et al, 2005)).
Isolating the movement from the goal component of learning can also be achieved by switching from performing the sequence with the hand to the whole arm (Grafton et al, 1998). Imaging work using this approach has shown that the primary motor cortex (M1) is associated with movement-based processing; while, part of the parietal cortex is associated with goal-based processing (Grafton et al, 1998). Thus, the distinction between goal and movement based processing is found within the organization of human brain circuits.
The goal and movement components of a motor skill memory are differentially affected during consolidation over sleep and wakefulness (Cohen et al, 2005). Following its formation, a memory continues to be processed during consolidation leading for instance to its stabilization, reorganization, or its enhancement (Robertson, 2009; King et al, 2017; for a brief review of mechanism Genzel and Robertson, 2015). The goal component is enhanced over sleep, but not over wakefulness. Conversely, the movement component is enhanced over wakefulness, but not over sleep. Thus, distinct offline mechanisms are responsible for enhancing different aspects of a procedural memory.
The amount of practice can modify the relative proportion of sequence learning within each co-ordinate frame. Initially, learning is predominately goal-based, and with increasing practice the skill acquired becomes increasingly movement-based (Robertson, 2009; Hikosaka et al, 1999; Hikosaka et al, 2002). Potentially, this qualitative change may provide a means to identify the transition from early to later stages of learning. Currently, there is no clear criterion to define objectively the transition between learning stages (Diedrichsen and Kornysheva, 2015). Overall, sequence learning occurs within distinct co-ordinate frames, which map onto distinct neural circuits, and in turn are consolidated over different brain states (i.e., wakefulness vs. sleep).
While only a few of us can play Rachmaninoff, we all learn to perform complex sequences on a daily basis from our morning routine to obeying the grammatical rules of language. Sequence learning is the cornerstone of a wide array of behaviours from learning athletic to musical skills, to understanding language, and onto social interactions. Important insights into sequence learning have been achieved through apparently simply tasks, such as the SRT and finger tapping tasks.
A compelling question not only for sequence learning, but also even more broadly for psychology and neuroscience, is the role of awareness in learning. Learning can occur with little or no awareness, or complete awareness for the sequence. These studies have brought into sharp focus the challenges that need to be overcome for detecting when, how and to what extent awareness for learning a sequence develops. New assessment techniques, such as post-decision wagering, are rising to these challenges and seem likely to give valuable insights in this area of research.
An equally important question is the content of memories. Distinguishing between memories being formed of perceptual or motor information has provided new insights into sequence learning. That information content is represented within different co-ordinates frames (i.e., allocentric vs. egocentric). Initially, one co-ordinate frame predominates at the early stages of learning, while later, the other co-ordinate frame becomes dominant. Perhaps these qualitative shifts can provide a robust way to identify and discriminate between early, and later learning. In sum, ingenious approaches have been successfully applied to understand sequence memory content. Further progress is likely to be made not only about memory content, but how different content is represented, and how these components interact.
Sequence learning is not limited to simply learning that one item (i.e., n) follows another (i.e., n-1). Instead, humans are capable of learning first, and second-order sequence regularities, and probabilistic associations in a seemingly effortless fashion. Part of this capacity may stem from clustering parts of the sequence into chunks. Converging evidence demonstrates that motor sequences are divided into chunks. Yet, what remains less clear is whether chunks are inextricably linked to, and so critical for learning. Chunking may only be necessary for the performance of a sequence; rather than necessary for learning. Elegant work has distinguished between performance and learning using dual task designs. In these designs participants concurrently perform a sequence, and another task such as listening to tones. Removal of the tone counting aspect of the task allows the full expression, or performance of the learnt sequence, which is otherwise prevented by the tone. Dual task designs have also provided insight into the role of attention in learning, and the integration of information across the different aspects of the task. Overall, sequence learning is a critical cognitive function, which is related to, and has provided important insight into a wide array of cognitive processes from awareness, to memory content, to representation, and onto clustering of information.
We are grateful to Aysha Keisler for her contributions to the initial version of this work, Iga Nowak for help co-authoring the current version, and to Kevin Caulfied, Adam Steel and Martina Bracco for their thoughtful comments.
Ashe, J., Lungu, O. V., Basford, A. T., & Lu, X. (2006). Cortical control of motor sequences. Curr Opin Neurobiol, 16(2), 213-221.
Bapi, R. S., Doya, K., & Harner, A. M. (2000). Evidence for effector independent and dependent representations and their differential time course of acquisition during motor sequence learning. Exp Brain Res, 132(2), 149-162.
Bapi, R. S., Miyapuram, K. P., Graydon, F. X., & Doya, K. (2006). fMRI investigation of cortical and subcortical networks in the learning of abstract and effector-specific representations of motor sequences. Neuroimage, 32(2), 714-727.
Bo, J., Peltier, S. J., Noll, D. C., & Seidler, R. D. (2011). Symbolic representations in motor sequence learning. Neuroimage, 54(1), 417-426.
Boyer, M., Destrebecqz, A., & Cleeremans, A. (2005). Processing abstract sequence structure: learning without knowing, or knowing without learning? Psychol Res, 69(5-6), 383-398.
Cleeremans, A. (1993). Attention and awareness in sequence learning.
Cleeremans, A., & McClelland, J. L. (1991). Learning the structure of event sequences. J Exp Psychol Gen, 120(3), 235-253.
Cohen, A., Ivry, R. I., & Keele, S. W. (1990). Attention and structure in sequence learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(1), 17.
Cohen, D. A., Pascual-Leone, A., Press, D. Z., & Robertson, E. M. (2005). Off-line learning of motor skill memory: a double dissociation of goal and movement. Proc Natl Acad Sci U S A, 102(50), 18237-18241.
Cohen, D. A., & Robertson, E. M. (2007). Motor sequence consolidation: constrained by critical time windows or competing components. Exp Brain Res, 177(4), 440-446.
Curran, T. (2001). Implicit learning revealed by the method of opposition. Trends Cogn Sci, 5(12), 503-504.
Destrebecqz, A., & Cleeremans, A. (2001). Can sequence learning be implicit? New evidence with the process dissociation procedure. Psychon Bull Rev, 8(2), 343-350.
Destrebecqz, A., Peigneux, P., Laureys, S., Degueldre, C., Del Fiore, G., Aerts, J., . . . Maquet, P. (2005). The neural correlates of implicit and explicit sequence learning: Interacting networks revealed by the process dissociation procedure. Learn Mem, 12(5), 480-490.
Frensch, P. A., Buchner, A., & Lin, J. (1994). Implicit learning of unique and ambiguous serial transitions in the presence and absence of a distractor task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(3), 567.
Grafton, S. T., Hazeltine, E., & Ivry, R. B. (1998). Abstract and effector-specific representations of motor sequences identified with PET. J Neurosci, 18(22), 9420-9428.
Hazeltine, E., Grafton, S. T., & Ivry, R. (1997). Attention and stimulus characteristics determine the locus of motor-sequence encoding. A PET study. Brain, 120 ( Pt 1), 123-140.
Hikosaka, O., Nakamura, K., Sakai, K., & Nakahara, H. (2002). Central mechanisms of motor skill learning. Curr Opin Neurobiol, 12(2), 217-222.
Honda, M., Deiber, M. P., Ibáñez, V., Pascual-Leone, A., Zhuang, P., & Hallett, M. (1998). Dynamic cortical involvement in implicit and explicit motor sequence learning. A PET study. Brain, 121(11), 2159-2173.
Howard, J. H. J., & Howard, D. V. (1997). Age differences in implicit learning of higher order dependencies in serial patterns. Psychology & Aging, 12.
Hsiao, A. T., & Reber, A. S. (2001). The dual-task SRT procedure: fine-tuning the timing. Psychon Bull Rev, 8(2), 336-342.
Jacoby, L. L. (1991). A process dissociation framework: Separating automatic from intentional uses of memory. Journal of Memory and Language, 30(5), 513-541.
Japikse, K. C., Negash, S., Howard, J. H., Jr., & Howard, D. V. (2003). Intermanual transfer of procedural learning after extended practice of probabilistic sequences. Exp Brain Res, 148(1), 38-49.
Jimenez, L. (2008). Taking patterns for chunks: is there any evidence of chunk learning in continuous serial reaction-time tasks? Psychol Res, 72(4), 387-396.
Jimenez, L., & Vazquez, G. A. (2005). Sequence learning under dual-task conditions: Alternatives to a resource-based account. Psychol Res, 69(5), 352-368.
Karni, A., Meyer, G., Jezzard, P., Adams, M. M., Turner, R., & Ungerleider, L. G. (1995). Functional MRI evidence for adult motor cortex plasticity during motor skill learning. Nature, 377(6545), 155-158.
Keele, S. W., Ivry, R., Mayr, U., Hazeltine, E., & Heuer, H. (2003). The cognitive and neural architecture of sequence representation. Psychol Rev, 110(2), 316-339.
Keele, S. W., & Jennings, P. J. (1992). Attention in the representation of sequence: Experiment and theory. Hum Mov Sci, 11(1-2), 125-138.
Knowlton, B. J., Ramus, S. J., & Squire, L. R. (1992). Intact artificial grammar learning in amnesia: Dissociation of classification learning and explicit memory for specific instances. Psychological Science, 3(3), 172-179.
Knowlton, B. J., & Squire, L. R. (1996). Artificial grammar learning depends on implicit acquisition of both abstract and exemplar-specific information. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(1), 169-181.
Koch, I., & Hoffmann, J. (2000). Patterns, chunks, and hierarchies in serial reaction-time tasks. Psychol Res, 63(1), 22-35.
Liu, T., Lungu, O. V., Waechter, T., Willingham, D. T., & Ashe, J. (2007). Frames of reference during implicit and explicit learning. Exp Brain Res, 180(2), 273-280.
Mayr, U. (1996). Spatial attention and implicit sequence learning: Evidence for independent learning of spatial and nonspatial sequences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(2), 350.
Nissen, M. J., & Bullemer, P. (1987). Attentional requirements of learning: Evidence from performance measures. Cognitive psychology, 19(1), 1-32.
O'Reilly, J. X., McCarthy, K. J., Capizzi, M., & Nobre, A. C. (2008). Acquisition of the temporal and ordinal structure of movement sequences in incidental learning. J Neurophysiol, 99(5), 2731-2735.
P, P., Maquet, P., Meulemans, T., Destrebecqz, A., S, L., C, D., . . . Cleeremans, A. (2000). Striatum forever, despite sequence learning variability: a random effect analysis of PET data. 10(4), 1065-9471.
Park, J. H., & Shea, C. H. (2005). Sequence learning: response structure and effector transfer. Q J Exp Psychol A, 58(3), 387-419.
Penhune, V. B., & Doyon, J. (2002). Dynamic cortical and subcortical networks in learning and delayed recall of timed motor sequences. J Neurosci, 22(4), 1397-1406.
Perlman, A., Pothos, E. M., Edwards, D. J., & Tzelgov, J. (2010). Task-relevant chunking in sequence learning. J Exp Psychol Hum Percept Perform, 36(3), 649-661.
Persaud, N., McLeod, P., & Cowey, A. (2007). Post-decision wagering objectively measures awareness. Nat Neurosci, 10(2), 257-261.
Rah, S. K., Reber, A. S., & Hsiao, A. T. (2000). Another wrinkle on the dual-task SRT experiment: it's probably not dual task. Psychon Bull Rev, 7(2), 309-313.
Reber, A. S. (1967). Implicit learning of articifial grammars. Journal of Verbal Learning & Verbal Behavior, 6(6), 855-863.
Reber, A. S., Allen, R., & Reber, P. J. (1999). Implicit versus explicit learning. The nature of cognition, 475-513.
Reber, P. J., & Squire, L. R. (1994). Parallel brain systems for learning with and without awareness. Learn Mem, 1(4), 217-229.
Reed, J., & Johnson, P. (1994). Assessing implicit learning with indirect tests: Determining what is learned about sequence structure. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(3), 585.
Remillard, G. (2008). Implicit learning of second-, third-, and fourth-order adjacent and nonadjacent sequential dependencies. Q J Exp Psychol (Hove), 61(3), 400-424.
Robertson, E., Tormos, J., Maeda, F., & Pascual-Leone, A. (2001). The role of the dorsolateral prefrontal cortex during sequence learning is specific for spatial information. Cerebral Cortex, 11(7), 628-635.
Robertson, E. M., Pascual-Leone, A., & Press, D. Z. (2004). Awareness modifies the skill-learning benefits of sleep. Curr Biol, 14(3), 208-212.
Robertson, E. M. (2007). Robertson, E. M. (2007). The serial reaction time task: Implicit motor skill learning? The Journal of Neuroscience, 27(38), 10073-10075.
Rosenthal, C. R., Roche-Kelly, E. E., Husain, M., & Kennard, C. (2009). Response-dependent contributions of human primary motor cortex and angular gyrus to manual and perceptual sequence learning. J Neurosci, 29(48), 15115-15125.
Rowland, L. A., & Shanks, D. R. (2006). Attention modulates the learning of multiple contingencies. Psychon Bull Rev, 13(4), 643-648.
Rüsseler, J., & Rösler, F. (2000). Implicit and explicit learning of event sequences: Evidence for distinct coding of perceptual and motor representations. Acta Psychol (Amst), 104(1), 45-67.
Sakai, K., Kitaguchi, K., & Hikosaka, O. (2003). Chunking during human visuomotor sequence learning. Exp Brain Res, 152(2), 229-242.
Schuck, N. W., Gaschler, R., Keisler, A., & Frensch, P. A. (2012). Position-item associations play a role in the acquisition of order knowledge in an implicit serial reaction time task. J Exp Psychol Learn Mem Cogn, 38(2), 440-456.
Shin, J. C., & Ivry, R. B. (2002). Concurrent learning of temporal and spatial sequences. J Exp Psychol Learn Mem Cogn, 28(3), 445-457.
Spencer, R. M., Sunm, M., & Ivry, R. B. (2006). Sleep-dependent consolidation of contextual learning. Curr Biol, 16(10), 1001-1005.
Stadler, M. A. (1992). Statistical structure and implicit serial learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18(2), 318.
Stadler, M. A. (1993). Implicit serial learning: questions inspired by Hebb (1961). Mem Cognit, 21(6), 819-827.
Stadler, M. A. (1995). Role of attention in implicit learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(3), 674.
Stefaniak, N., Willems, S., Adam, S., & Meulemans, T. (2008). What is the impact of the explicit knowledge of sequence regularities on both deterministic and probabilistic serial reaction time task performance? Mem Cognit, 36(7), 1283-1298.
Verwey, W. B. (2003). Effect of sequence length on the execution of familiar keying sequences: lasting segmentation and preparation? J Mot Behav, 35(4), 343-354.
Verwey, W. B., & Eikelboom, T. (2003). Evidence for lasting sequence segmentation in the discrete sequence-production task. J Mot Behav, 35(2), 171-181.
Walker, C. M., Walker, M. F., & Sunderland, A. (2003). Dressing after a Stroke: a Survey of Current Occupational Therapy Practice. The British Journal of Occupational Therapy, 66(6), 263-268.
Wilkinson, L., & Shanks, D. R. (2004). Intentional control and implicit sequence learning. J Exp Psychol Learn Mem Cogn, 30(2), 354-369.
Willingham, D. B. (1998). A neuropsychological theory of motor skill learning. Psychol Rev, 105(3), 558-584.
Willingham, D. B. (1999). Implicit motor sequence learning is not purely perceptual. Mem Cognit, 27(3), 561-572.
Willingham, D. B., & Goedert-Eschmann, K. (1999). The relation between implicit and explicit learning: Evidence for parallel development. Psychological Science, 10(6), 531-534.
Willingham, D. B., Nissen, M. J., & Bullemer, P. (1989). On the development of procedural knowledge. J Exp Psychol Learn Mem Cogn, 15(6), 1047-1060.
Willingham, D. B., Salidis, J., & Gabrieli, J. D. (2002). Direct comparison of neural systems mediating conscious and unconscious skill learning. J Neurophysiol, 88(3), 1451-1460.
Willingham, D. B., Wells, L. A., Farrell, J. M., & Stemwedel, M. E. (2000). Implicit motor sequence learning is represented in response locations. Mem Cognit, 28(3), 366-375.
Witt, J. K., Ashe, J., & Willingham, D. T. (2008). An egocentric frame of reference in implicit motor sequence learning. Psychol Res, 72(5), 542-552.
Ziessler, M., Hanel, K., & Sachse, D. (1990). The programming of structural properties of movement sequences. Psychol Res, 52(4), 347-358.