6.2 Future work
6.2.1 The need for neural support
In Chapters 2 and 3 we used behavioural evidence of concurrent field learning across the same physical states, to suggest that planning different follow-throughs results in different neural states during the execution of the identical kinematic trajectories. While we were able to use controlled behaviours to isolate neural, over physical, states as the critical factor associated with motor adaptation, the suggestion that planning different motor sequences produces distinct neural trajectories needs electrophysiological verification. It is also worth acknowledging that the evidence for dynamical neural trajectories on which our interpretation is built, lie entirely within premotor and primary motor cortices (Churchland et al., 2012;
Pandarinath et al., 2015), while other motor regions such as the supplementary motor area do not appear to show these population-level rotations (Lara et al., 2018). A dynamical systems view of motor cortex ascribes motor preparation to seeding the system with an initial condition that gives rise to the subsequent neural trajectory. While it has been shown that these states are re-established albeit briefly under delayed, self-paced, and pressured reaction time movements (Ames et al., 2014; Lara et al., 2018), it remains to be seen whether this preparatory activity has a causal impact on the movement that follows. Establishing this causal link is necessary for our hypothesized connection between preparatory activity and motor adaptation.
Many studies have suggested that the cerebellum is important for supervised error-based learning such as incremental motor adaptation (Albus, 1971; Criscimagna-Hemminger et al., 2010; Galea et al., 2011; Herzfeld et al., 2015; Ito, 2006; Marr, 1969; Smith and Shadmehr, 2005; Tseng et al., 2007; Wolpert et al., 1998), while in Chapters 2 and 3 we have speculated that the dynamics of motor cortex may facilitate the specificity of such motor adaptation.
These views are complementary, as in theory the different motor cortical dynamics resulting from different plans may very well be paralleled by different patterns of activation in the cerebellum (although it is unlikely that they display similar dynamical motifs). Subsequently, in Chapter 4 our results suggested that deadaptation may be plan-independent, appearing to be specific to particular motor memories only upon their execution. It is difficult at this point to reconcile these later behavioural results with the relative contributions of the cerebellum and cortex to the specificity of adaptation and de-adaptation, but it is hoped that further evidence, both behavioural and neural may help to delineate their contributions more clearly.
Related to this, a dynamical systems perspective of motor cortex may still provide a frame-work for interpreting the temporal de-adaptation observed in Chapter 4. Previously, Church-land et al. (2006a) demonstrated that small variations in motor cortical preparatory activity were predictive of variations in the output of the upcoming reach, suggesting that fluctuations in the motor cortical initial state have a causal effect on motor output. A similar theory might support the gradual deadaptation we observe in time, such that as endogenous neural states drift with the passage of time, the same input to motor cortex may result in either slightly different preparatory activity (input, X), or different dynamics resulting from this input (f(X)), to produce a slightly different subsequent movement. That motor de-adaptation was not specific to the planned action (the input, X) might suggest that such a neural drift was in the dynamics themselves (f(X)). This is speculative, but yields testable neural predictions.
It is also worth clarifying that in Chapter 5, the assumption that chunking would affect adaptation does not imply that movements are chunked in M1. Rather, our theory is based upon the assumption that the different ways of chunking movements will ultimately yield different preparatory activity, and hence different subsequent motor cortical dynamics that may affect the specificity of force-field adaptation. The idea that sequence-related activity is reflected in areas associated with motor preparation, such as supplementary motor, parietal and premotor areas has been supported by imaging studies (Abe et al., 2007; Yokoi et al., 2018) and single-cell recordings in macaques (Tanji and Shima, 1994), but this chunking may have origins elsewhere such as in the basal ganglia (Graybiel, 1998; Jin et al., 2014). Our theory makes no claims about the locus of chunking computation, only that the downstream impact of chunked motor plans may affect the specificity of adaptation. A multi-day pre-post training human fMRI study, which extends the duration of the training protocol we put forward in Chapter 5 may help to provide support for such a hypothesis.
Our hypotheses for neural mechanisms that support motor adaptation have centred on the neural dynamics of motor cortex, however this does not preclude the involvement of more cognitive processes and upstream areas such as the prefrontal cortex. In cognitive planning tasks, functional neuroimaging studies have shown that the prefrontal cortex is highly involved in planning (Balaguer et al., 2016; Schacter and Addis, 2007; Unterrainer and Owen, 2006), and patients with prefrontal lesions exhibit unstructured actions that either fail to achieve the task goal or result in participants taking significantly longer to achieve it (Shallice, 1982; Tim Shallice, 1991). Recently, the involvement of such cognitive or strategic processes in motor adaptation has risen to the fore (McDougle et al., 2016, 2015;
Taylor and Ivry, 2014). McDougle et al. (2015) has suggested that the canonical learning curves observed in both force-field and visuomotor adaptation experiments may result from an interaction between explicit (cognitive) and implicit learning processes, paralleling the traditional dual-rate fast (explicit) and slow (implicit) process view of motor adaptation (Smith et al., 2006). By requiring participants to verbally report their reaching direction during adaptation, McDougle et al. (2015) demonstrated significant explicit strategy use in motor adaptation during blocked training. However this same work also suggests that an additional, much slower learning process is present which is unaccounted for by the explicit reports of strategy. Interestingly, when comparing to our own results, the learning we observe in all studies presented in this thesis is considerably slower (around 30-40 times slower) than the single-field learning examined by (McDougle et al., 2015), with participants adapting across 1200 trials to around 40% of the perturbations in our studies, compared to (approximately) 30 trials it took the participants to completely adapt in the study by McDougle and colleagues, and much slower than would typically be associated with a
fast-learning process. Interpreted within the explicit/implicit framework, this implicates the role of more implicit over explicit forms learning in our results.
While the extent to which explicit processes are involved in the follow-through or lead-in tasks presented in this thesis was not directly tested, the development and usage of cognitive strategies with adaptation remains possible, but may be coherent with the neural mechanisms of learning we have proposed. Tracing studies have shown that dorsolateral prefrontal cortex projects strongly to premotor areas, suggesting the possibility that changes in cognitive plans or strategy use may result in significant changes in premotor activity, such as leading to different initial preparatory neural states (Lu et al., 1994). Indeed under this hypothesis, the resultant slow learning we observe could still be motor cortical, such that with learning, different physical movements are generated across the course of adaptation for the same cognitive strategy (plan), so long as this plan is different for each context.
In the absence of neural data, one way to dissociate the roles of explicit and implicit processes in the adaptation we observe may be to force participants to initiate their movements at very low preparation times, a technique used recently by (Haith et al., 2015) in a visuomotor adap-tation paradigm. When participants were forced to initiate action on these low preparation time trials, responses produced a learning curve which corresponded closely to the slow, or more implicit component of motor adaptation. In theory, if the effects we observe are indeed an implicit form of adaptation, requiring participants to respond at reduced reaction times should result in similar rates of learning to the net learning curves we have demonstrated.
An interesting control for the work on imagery in Chapter 3 would be to establish whether imagery must be motor-related to enable distinct motor memories to form for the same physical states. That is, would associating each direction of force perturbation with a more abstract image, such as a car or house, be sufficient to drive concurrent adaptation. A negative result would support the interpretation that the learning we observe is not the consequence of cognitive associations between the follow-through and perturbation directions, but due to the similarity between motor imagery and overt movement. Previous behavioural work suggesting that passive visual cues are insufficient contextual cues for such tasks would strongly suggest such images would be insufficient (Howard et al., 2013). Neural recordings from motor cortex during the preparation of these images, versus motor imagery would help to verify our claims.
Initial neural support
Initial neural support for our hypothesis that planning is key to the representation of motor adaptation has, however, already begun to emerge. Recently neural recordings from monkeys covertly controlling a BMI-cursor showed that the proximity of motor preparatory states associated with covert and overt movements significantly correlated with the magnitude of subsequent motor adaptation transfer between the two (Vyas et al., 2018). Moreover, trial-by-trial kinematic errors significantly correlated with changes to the preparatory state with adaptation, together suggesting that 1) adaptation may involve finding the optimal set of initial neural conditions for producing the subsequent movement, and 2) the overlap of preparatory states may be a good neural correlate for motor learning transfer or interference in our force-field adaptation tasks. Together these results provide promising indications that our hypotheses on the neural underpinnings of motor memory separation may have milage.