A variety of tasks and experimental paradigms have been used for studying motor skill learning, including juggling, visuomotor tracking and isometric force-production tasks, to name a few. Of particular relevance to the current review are studies of tasks that require practice of sequential movements, tapping skills like typing or playing various musical instruments. Here, our main focus is on learning sequential motor skills that show lasting improvements beyond baseline performance over lengthy periods of time. Another model for studying motor learning, that does not necessarily involve the acquisition of a new skill has been adaptation to externally induced perturbations, such as those induced by a force-field (dynamic adaptation) or by visuomotor rotations (visuomotor adaptation). These perturbations are more commonly introduced while subjects execute simple motor tasks, for instance point to point ballistic reaching movements (Krakauer, 2009; Shadmehr et al., 2010; Seidler, 2010; Lalazar and Vaadia, 2008). Yet, these paradigms characteristically evaluate the return to baseline levels of performance following perturbation over relatively short time periods (Krakauer and Mazzoni, 2011). However, it should be noted that repetitive practice of adaptation tasks could lead to performance improvements over time as in the form of “savings”, expressed as faster readaptation to external perturbations, relative to the initial rate of adaptation (e.g., Landi et al., 2011). Moreover, skill learning tasks, where lasting improvements are seen over time, for instance whole-body balancing (Taubert et al., 2010), may involve an adaptation component. Motor skills are typically learned slowly over multiple training sessions until performance reaches nearly asymptotic levels. Across different experimental paradigms, skill acquisition develops (Figure 1A) initially relatively fast (i.e., rapid improvements measured over the course of a single training session) and later more slowly, when further gains develop incrementally over multiple sessions of practice (Doyon and Benali, 2005; Doyon and Ungerleider, 2002). Of note, the relative duration of what can be defined as fast and slow learning is highly task specific. For example, the fast stage of learning a simple four component keypress sequence could last minutes (e.g., Karni et al., 1995), while the fast stage of learning to play a complex musical piece may last months (Figure 1B). Similarly, nearly asymptotic levels in end-point measures of skill can be acquired very rapidly when learning a keypress sequence, but much more slowly when learning to play a complex musical piece. Skill changes can occur during training (online) but also after training ended (offline; Figure 1C). Offline processes, including skill stabilization and improvement (Fischer et al., 2005; Korman et al., 2003; Walker et al., 2002), reflect motor memory consolidation (Doyon and Benali, 2005; Muellbacher et al., 2002; Robertson et al., 2004a), an intermediate stage between fast and slow learning (Doyon and Benali, 2005; Doyon et al., 2009a). Online and offline skill gains can be maintained over time, resulting in long-term retention (Romano et al., 2010).

A valuable framework for interpreting the role of this complex pattern of recruitment has been proposed by Hikosaka and colleagues (Hikosaka et al., 2002a), in a model describing the mechanisms for sequential motor skill learning. According to this model, two parallel loop circuits operate in learning spatial and motor features of sequences. While learning spatial coordinates is supported by a frontoparietal-associative striatum-cerebellar circuit, learning motor coordinates is supported by an M1-sensorimotor striatum -cerebellar circuit. Transformations between the two coordinate systems rely, according to this model, on the contribution of the SMA, pre-SMA and premotor cortices. Importantly, it was argued that learning spatial coordinates is faster, yet requires additional attentional and executive resources, putatively provided by prefrontal cortical regions (Miller and Cohen, 2001). Similarly, In a another model, Doyon and Ungerleider (2002) proposed that during fast learning a cortico-striato-thalamo-cortical loop and a cortico-cerebello-thalamo-cortical loop are both recruited, operating in parallel. Further, interactions between the two systems were believed to be crucial for establishing the motor routines necessary for learning new motor skills. (Doyon and Ungerleider, 2002; Doyon and Benali, 2005). Both models share the view that motor skill learning involves interactions between distinct cortical and subcortical circuits, crucial for the unique cognitive and control demands associated with this stage of skill acquisition (Hikosaka et al., 2002a; Doyon and Ungerleider, 2002).

One of the key brain regions involved in fast learning is M1. Fast motor skill learning is associated with substantial recruitment of neurons in M1 in behaving mice during the initial stages of learning an accelerating rotarod task (Costa et al., 2004) and with modulation of synaptic efficacy through long-term potentiation (LTP) and long-term depression (LTD) in rodents (Rioult-Pedotti et al., 1998, 2000). Consistently, it was shown in humans that learning a motor task modulates LTP-like plasticity (Ziemann et al., 2004; Stefan et al., 2006; Rosenkranz et al., 2007). BOLD activity in M1 progressively decreases as motor skill learning progresses over a single training session (Karni et al., 1995), yet it should be noted that the magnitude of engagement of M1 in fast learning is highly influenced by the specific task and by attentional demands (Hazeltine et al., 1997, Stefan et al., 2004). Consistent reorganizational changes in M1 have been described using TMS. For example the fast stage of implicit motor skill learning, as assessed with the serial reaction time task is accompanied by increased motor map size of the fingers engaged in the task. Interestingly, when the sequence becomes explicitly known, the M1 motor map size returns to baseline (Pascual-Leone et al., 1994). The cellular mechanisms behind learning-related plasticity in M1 appear to depend on protein synthesis within this structure, and may specifically involve brain-derived neurotrophic factor (BDNF)(Kleim et al., 2003). In both humans and animal models BDNF influences synaptic plasticity (Akaneya et al., 1997; Lu, 2003). Injection of protein synthesis inhibitors, targeting BDNF into the rat M1 induces a lasting loss of motor map representation (Kleim et al., 2003). Moreover, training-dependent increases in motor cortical excitability (Antal et al., 2010; Cheeran et al., 2009) and fMRI signal (McHughen et al., 2010) are reduced in healthy humans with a valine-to-methionine substitution at codon 66 (val66met) in the BDNF gene, when compared to subjects without this polymorphism (Kleim et al., 2006). These findings led to the hypothesis that the presence of this particular polymorphism could influence motor skill learning (Fritsch et al., 2010).

Although earlier imaging studies clearly established that the fast stage of motor skill learning is sustained by activity across a distributed set of brain regions, conventional univariate fMRI analysis, where brain activity is analyzed in a voxel-wise manner as if each anatomically distinguishable region is independent (Marrelec et al., 2006; Tamas Kinces et al., 2008) does not provide information on inter-regional interactions as required to properly test these models. The most widely used and straightforward approach for assessing inter-regional interactions in neuroimaging data is based on analysis of functional connectivity (Friston, 1994), which refers to the statistical dependence defined in terms of correlation or covariance between the activation in spatially remote regions. Using this approach, it was shown that M1, the premotor cortex, and the SMA have significantly greater inter- and intrahemispheric coupling during early as compared to late within-session explicit sequence learning (Sun et al., 2007). Interactions between M1, SMA, and premotor cortices are likely to reflect transformations between spatial and motor features of motor sequences, required for fast motor skill learning (Hikosaka et al., 2002a). Additionally, fast motor skill learning is characterized by increased functional connectivity between the DLPFC and premotor cortex (Sun et al., 2007), relating to the heightened attentional demands required at this stage of skill acquisition (Hikosaka et al., 2002a; Peteresen et al., 1998).

Studies that examined the neuronal mechanisms involved in the slow stage of motor skill learning, typically had subjects learn a motor skill over several weeks and scanned them on different occasions throughout the training period (Karni et al., 1995; Floyer-Lea and Matthews, 2005; Coynel et al., 2010; Lehéricy et al., 2005). Slow learning is associated with increased activation in M1 (Floyer-Lea and Matthews, 2005), primary somatosensory cortex (Floyer-Lea and Matthews, 2005), SMA (Lehéricy et al., 2005) and putamen (Lehéricy et al., 2005; Floyer-Lea and Matthews, 2005), as well as decreased activation in lobule VI of the cerebellum (Lehéricy et al., 2005). Thus, progress from early to late stages of motor skill learning is characterized by a shift in fMRI activation from anterior to more posterior regions of the brain (Floyer-Lea and Matthews, 2005), a pattern also reported when learning non-motor tasks that is thought to reflect a progressive decrease in reliance on attentional resources and executive function (Kelly and Garavan, 2005). Progressing from fast to slow motor skill learning is also associated with a shift in fMRI activation from associative to sensorimotor striatum (Coynel et al., 2010; Lehéricy et al., 2005), thought to contribute to slow learning of the motor component of sequences (Hikosaka et al., 2002a).

Slow learning has been linked with larger-scale functional reorganization as well. A recent study tracked functional connectivity using fMRI over a period of 4 weeks of training on an explicit motor sequence task (Coynel et al., 2010). Early learning was associated with increased integration, a metric reflecting functional interactions among several brain regions, of a premotor-associative striatum-cerebellar network. During slow learning, on the other hand, the authors reported decreased integration in this premotor-associative striatum-cerebellar network but stable connectivity within the M1-sensorimotor striatum -cerebellar network, largely consistent with data emerging from regional fMRI analysis (Floyer-Lea and Matthews, 2005; Lehéricy et al., 2005).

Engagement of neurons in the sensorimotor striatum during later stages of learning has been well documented in animal models (Miyachi et al., 2002; Yin et al., 2009), and has been proposed as a substrate for the acquisition of habitual and automatic behavior (Yin et al., 2004, 2009). For example in vivo recordings in behaving rodents revealed that the sensorimotor striatum is engaged later in training, when performance in an accelerated rotarod task asymptoted (Yin et al., 2009). Consistently, ex-vivo recordings from medium spiny neurons in sensorimotor striatum following training revealed long lasting changes in glutamatergic neurotransmition (Yin et al., 2009). The involvement of the striatum in the stages where motor skills become automatic has been confirmed in human neuroimaging studies (Ashby et al., 2010; Lehéricy et al., 2005; Poldrack et al., 2005). For example, using a dual-task design, where a sequence of finger movements was learned while assessing the influence of a secondary interfering task, it was found that automaticity was accompanied by a decrease in activation in the associative striatum (Poldrack et al., 2005; Lehéricy et al., 2005).

Of note, studies in both humans and animals have consistently reported the involvement of M1 in slow learning, a key brain region in other stages of motor skill learning as well. Training to perform an explicit sequence of finger movements over several weeks showed progressively increasing BOLD activity in M1 (Karni et al., 1995, 1998; Floyer-Lea and Matthews, 2005; Lehéricy et al., 2005; but see Xiong et al., 2009; Ma et al., 2010), interpreted as reflecting recruitment of additional M1 units into the local network that represents the acquired sequence of movements (Ungerlieder et al., 2002). Learning a motor sequence over several days is also accompanied by an increase in the size of motor maps and corticomotoneuronal excitability of the digits involved in the task, both measured with TMS (Pascual-Leone et al., 1995). This particular reorganization within M1 is related to learning since simple repetition of movements in the absence of a sequential order did not induce such effect. Consistently, facilitatory stimulation of M1 over 5 days with anodal transcranial direct current stimulation (tDCS) improved learning of a sequential visuomotor task. Of note, the advantage in skill of the stimulated group relative to the sham control group was still present 90 days later (Reis et al., 2009). These results support a causal link between M1 function and motor skill learning when training over multiple sessions.

A more direct evidence for learning-induced changes in gray matter emerges from studies that utilized longitudinal designs, evaluating the same individuals learning a particular skill over relatively long time periods. In one key study (Draganski et al., 2004) subjects trained over 3 months to learn a three-ball juggling routine. Structural MRI scans were acquired at baseline (before training), at the end of training and three-months later in the absence of additional practice. The authors documented at the end of training an expansion of gray matter in area MT/V5 and in the left posterior intraparietal sulcus, both involved in perception of motion and visuomotor processing. Yet, regional gray matter decreased to near baseline three months following the end of training, paralleling the decrease of skill. Similar expansion in gray- matter in area MT/V5 was reported in a group of elderly volunteers learning the same task, suggesting that reorganization in gray-matter can also occur in the ageing human brain (Boyke et al., 2008).

Later studies examined more closely the time-scales of gray-matter changes with slow motor skill learning (Driemeyer et al., 2008; Scholz et al., 2009; Taubert et al., 2010). Consistent with previous results, gray-matter expansions were documented in the medial occipital and parietal lobes after 6 weeks of juggling practice (Scholz et al., 2009) and in bilateral occipito-temporal cortex as early as following 7 days of practice (Driemeyer et al., 2008). In another study, gray-matter volume expansion was identified in parieto-frontal regions as early as following two weekly practice sessions, in a whole-body balancing task (Taubert et al., 2010). These findings are indicative of relatively rapid structural gray-matter plasticity associated with motor skill learning. What mechanisms could contribute to these striking findings? While speculative at this point, it has been proposed that processes occurring both at the synapse level and at larger scales, including rapid intracortical remodeling of dendritic spines and axonal terminals, glial hypertrophy and synaptogenesis might play a contributory role (Draganski and May, 2008; May and Gaser, 2006; Anderson et al., 1994). Consistently, rapid (within an hour) formation of postsynaptic dendritic spines has been detected in vivo in the pyramidal neurons of the mouse motor cortex following motor training (Xu et al., 2009) and the extent of spine remodeling has been shown to correlate with behavioral improvements after learning, suggesting that this mechanism of synaptic plasticity may contribute to motor memory formation (Yang et al., 2009). On the other hand, it should be noted that other animal studies demonstrated significant increases in synapse numbers in the rat M1 only after extensive training (Kleim et al., 1996, 2004).

Slow learning has been linked with structural plasticity in white-matter architecture as well. Diffusion MRI-based measures, such as fractional anisotropy (FA) are believed to reflect white-matter integrity (Fields, 2008), providing a distinctive insight into the microstructural properties of white-matter in-vivo (Le Bihan et al., 2001; Mori and Zhang, 2006). Cross sectional studies, primarily with highly trained musicians, examined white-matter correlates of skilled behavior (Bengtsson et al., 2005; Han et al., 2009; Schmithorst and Wilke, 2002). Fractional anisotropy in the posterior limb of the internal capsule, which contains descending corticospinal fibers from the primary sensorimotor and premotor cortices, correlated with number of practice hours during childhood in skilled musicians (Bengtsson et al., 2005). It has been proposed that these results may reflect experience-induced plasticity during a critical developmental period (Bengtsson et al., 2005). A recent pioneering study provided more direct evidence for experience-induced changes in white matter architecture, resulting from a relatively short period of practice (Scholz et al., 2009). In this study, it was shown that 6 weeks of juggling practice resulted in increased FA in a region of white matter underlying the intraparietal sulcus. Localized increases in gray-matter were detected in close proximity to these white matter regions. Yet the magnitude of changes of gray and white matter showed no correlation and developed over markedly different time courses. Interestingly, individual differences in white-matter mictrostructure appears to be related to variation in learning (Johansen-Berg, 2010; Della-Maggiore et al., 2009; Tomassini et al., 2011). For example, individual differences in learning an isometric visuomotor tracking task are associated with variability in FA in the premotor cortex and the cerebellum (Tomassini et al., 2010). The cellular mechanisms underlying learning-dependent changes in white-matter microstructure remain to be established, and so do the links between these changes and measures of functional plasticity. It has been proposed that changes in white matter properties, indexed by FA could affect the velocity and synchronicity of impulse conduction between distant cortical regions and thus contribute to the optimization of information flow required for skill acquisition (see Fields, 2008, 2011), a hypothesis that requires specific testing.

In humans, offline skill improvements may be affected by sleep (e.g., Diekelmann and Born, 2010; Fischer et al., 2002; Korman et al., 2003). Sleep-dependent motor memory consolidation, which correlates with the amount of stage II non-rapid eye movement sleep (Walker et al., 2002), has been mostly demonstrated for explicit motor sequence learning (Fischer et al., 2005; Korman et al., 2003; Walker et al., 2002; but see, Brawn et al., 2010; Rickard et al., 2008). Other forms of procedural motor learning are not necessarily sleep-dependent (Debas et al., 2010; Doyon et al., 2009b; Song et al., 2007). Notably, sleep does not benefit implicit forms of sequence learning (Robertson et al., 2004b; Song et al., 2007). In such circumstances, similar memory gains were reported after sleep and over an equivalent period of wakefulness (see also, Hotermans et al., 2008).

Different brain regions are involved in consolidation of motor memories. Sleep dependent improvements in learning a sequential finger movement task were linked to reduced BOLD activity in M1, as measured with fMRI (Fischer et al., 2005). Furthermore, downregulating excitability of M1 by low frequency TMS (virtual lesion) results in reduced motor memory consolidation (Muellbacher et al., 2002; Robertson et al., 2005), a time-specific effect since it was not observed when TMS was applied 6 hours post training (Muellbacher et al., 2002). The finding of differential effects of facilitatory anodal tDCS applied over M1 on online and offline learning of a sequential motor task, namely enhancement of offline learning, supports the existence of relatively different neuronal networks involved in the two processes (Reis et al., 2009). Another key contributor to consolidation of sequential motor skills is the striatum (Debas et al., 2010; Fischer et al., 2005; Albouy et al., 2008; Doyon and Ungerlieder, 2002). Recent work showed increased striatal activity in human subjects in whom offline consolidation was tested following a night of sleep, as compared to those in whom it was tested after an equivalent period of wakefulness (Debas et al., 2010; Fischer et al., 2005). Interestingly, BOLD activity in the ventral striatum and the hippocampus during the initial stages of oculomotor sequence learning predicted the magnitude of sleep-dependent behavioral improvements (Albouy et al., 2008). Additional evidence for the involvement of these two regions emerged from animal studies demonstrating that local injections of protein synthesis inhibitors disrupts consolidation of motor memories (Buitrago et al., 2004). This effect was present when injections were applied to M1 (Luft et al., 2004) and to a lesser extent the dorsal striatum (Wachter et al., 2010), but was absent after injections of control regions (Luft et al., 2004).

The neural processes leading to successful consolidation tested post training are likely to start operating during practice and evolve over time after training ended. Typically, evaluation of changes in BOLD signal induced by task performance assesses the consequences of these processes as tested a few hours or the day after practice was completed. Thus, the neuronal mechanisms that operate during and early after practice and during sleep to support motor memory consolidation remain to a large extent uncertain. It was recently suggested that a possible way of closing this gap in knowledge is through measurement of intrinsic resting-state functional connectivity (Albert et al., 2009; Ma et al., 2011, Taubert et al., 2011). Spontaneous low frequency fluctuations in the BOLD signal, in the absence of any overt input or behavior, have been widely reported in the past 15 years (for a review see, Fox and Raichle, 2007; Cole et al., 2010) and can be recorded at different times before or after training without affecting subsequent behavioral testing. Temporally coherent spontaneous fluctuations at rest have been found between spatially remote brain regions in areas known to be involved in motor, visual, and auditory processing, attention and language (Cole et al., 2010; Fox and Raichle, 2007). Thus, resting-state functional connectivity, which may be sampled multiple times during the period leading to the behavioral measurement of consolidation, may provide a unique window for examining neural networks activity along the entire course of motor skill acquisition. Available data are supportive of this contention. Learning a visuomotor tracking task over one session increased resting functional connectivity in a network that includes the prefrontal, superior and inferior parietal cortices, and Crus II of the cerebellum (Albert et al., 2009). Learning a whole-body dynamic balancing task over multiple sessions showed increased resting state connectivity between SMA/preSMA and medial parietal cortex that correlated with performance improvements (Taubert et al., 2011). Modulation of resting-state connectivity in parietal circuits was also observed along 4 weeks of daily training of an explicit sequence learning task (Ma et al., 2011). Overall, these studies suggest that functional connectivity in fronto-parietal networks support consolidation after fast (Albert et al., 2009) and slow learning (Taubert et al., 2011; Ma et al., 2011). Comparison among these studies, however, should be done with caution, since they involved different motor skill tasks. Notwithstanding, published studies have yet to identify modulation of connectivity within striatal regions, believed to play a key role in consolidation of skills (Doyon and Benali, 2005; Doyon and Ungerlieder, 2002), but preliminary findings indeed appear to support this hypothesis (Debas et al., HBM 2011; unpublished data).

Various task attributes have a profound influence on long-term retention of skill learning. For instance, reward during practice improves long-term retention of a sequential motor skill (Abe et al., 2011). A reward-related enhancement of long-term memory has been demonstrated for other forms of memory as well (Wittmann et al., 2010), and is linked with fMRI activation in the striatum, ventral tegmental area and hippocampus (Wittmann et al., 2005; Adcock et al., 2006). It has been proposed that dopaminergic modulation within these circuits, specifically through dopamine-dependent LTP in the hippocampus may contribute to this effect (Calabresi et al., 2007). In the future, it will be of interest to identify the influence of reward attributes such as predictability, magnitude and outcome uncertainty on long-term retention of motor skills. For instance, a recent study found that reward predictability and to some extent reward magnitude modulate long term episodic memory, an effect which was absent for outcome uncertainty by itself (Wittmann et al., 2010).