Automatic learning mechanisms for flexible human locomotion

Reviewer #1 (Public review):

Summary:

Rossi et al. asked whether gait adaptation is solely a matter of slow perceptual realignment or if it also involves fast/flexible stimulus-response mapping mechanisms. To test this, they conducted a series of split-belt treadmill experiments with ramped perturbations, revealing behavior indicative of a flexible, automatic stimulus-response mapping mechanism.

Strengths:

(1) The study includes a perceptual test of leg speed, which correlates with the perceptual realignment component of motor aftereffects. This indicates that there are motor performances that are not accounted for by perceptual re-alignment.

(2) They study incorporates qualitatively distinct, hypothesis-driven models of adaptation and proposes a new framework that integrates these various mechanisms.

Weaknesses:

(1) The study could benefit from considering other alternative models. As the authors noted in their discussion, while the descriptive models explain some patterns of behaviour/aftereffects, they don't currently account for how these mechanisms influence the initial learning process itself.

a. For example, the pattern of gait asymmetric might differ for perceptual realignment (a smooth, gradual process), structural learning (more erratic, involving hypothesis testing/reasoning to understand the perturbation, see (Tsay et al. 2024) for a recent review on Reasoning), and stimulus-response mapping (possibly through a reinforcement based trial-and-error approach). If not formally doing a model comparison, the manuscript might benefit from clearly laying out the behavioural predictions for how these different processes shape initial learning.

b. Related to the above, the authors noted that the absence of difference during initial learning suggests that the differences in Experiment 2 in the ramp-up phase are driven by two distinct processes: structural learning and memory-based processes. If the assumptions about initial learning are not clear, this logic of this conclusion is hard to follow.

c. The authors could also test a variant of the dual-rate state-space model with two perceptual realignment processes where the constraints on retention and learning rate are relaxed. This model would be a stronger test for two perceptual re-alignment processes: one that is flexible and another that is rigid, without mandating that one be fast learning and fast forgetting, and the other be slow learning and slow forgetting.

(2) The authors claim that stimulus-response mapping operates outside of explicit/deliberate control. While this could be true, the survey questions may have limitations that could be more clearly acknowledged.

a. Specifically, asking participants at the end of the experiments to recall their strategies may suffer from memory biases (e.g., participants may be biased by recent events, and forget about the explicit strategies early in the experiment), be susceptible to the framing of the questions (e.g., participants not being sure what the experimenter is asking and how to verbalize their own strategy), and moreover, not clear what is the category of explicit strategies one might enact here which dictates what might be considered "relevant" and "accurate".

b. The concept of perceptual realignment also suggests that participants are somewhat aware of the treadmill's changing conditions; therefore, as a thought experiment, if the authors have asked participants throughout/during the experiment whether they are trying different strategies, would they predict that some behaviour is under deliberate control?

(3) The distinction between structural and memory-based differences in the two subgroups was based on the notion that memory-based strategies increase asymmetry. However, an alternative explanation could be that unfamiliar perturbations, due to the ramping up, trigger a surprise signal that leads to greater asymmetry due to reactive corrections to prevent one's fall - not because participants are generalizing from previously learned representations (e.g., (Iturralde & Torres-Oviedo, 2019)).

Further contextualization:

Recognizing the differences in dependent variables (reaching position vs. leg speed/symmetry in walking), could the Proprioceptive/Perceptual Re-alignment model also apply to gait adaptation (Tsay et al., 2022; Zhang et al., 2024)? Recent reaching studies show a similar link between perception and action during motor adaptation (Tsay et al., 2021) and have proposed a model aligning with the authors' correlations between perception and action. The core signal driving implicit adaptation is the discrepancy between perceived and desired limb position, integrating forward model predictions with proprioceptive/visual feedback.

References

Iturralde, P. A., & Torres-Oviedo, G. (2019). Corrective Muscle Activity Reveals Subject-Specific Sensorimotor Recalibration. eNeuro, 6(2). https://doi.org/10.1523/ENEURO.0358-18.2019

Tsay, Jonathan S., Hyosub E. Kim, Samuel D. McDougle, Jordan A. Taylor, Adrian Haith, Guy Avraham, John W. Krakauer, Anne G. E. Collins, and Richard B. Ivry. 2024. "Fundamental Processes in Sensorimotor Learning: Reasoning, Refinement, and Retrieval." ELife 13 (August). https://doi.org/10.7554/eLife.91839.

Tsay, Jonathan S., Hyosub E. Kim, Darius E. Parvin, Alissa R. Stover, and Richard B. Ivry. 2021. "Individual Differences in Proprioception Predict the Extent of Implicit Sensorimotor Adaptation." Journal of Neurophysiology, March. https://doi.org/10.1152/jn.00585.2020.

Tsay, Jonathan S., Hyosub Kim, Adrian M. Haith, and Richard B. Ivry. 2022. "Understanding Implicit Sensorimotor Adaptation as a Process of Proprioceptive Re-Alignment." ELife 11 (August). https://doi.org/10.7554/eLife.76639.

Zhang, Zhaoran, Huijun Wang, Tianyang Zhang, Zixuan Nie, and Kunlin Wei. 2024. "Perceptual Error Based on Bayesian Cue Combination Drives Implicit Motor Adaptation." ELife. https://doi.org/10.7554/elife.94608.1.

Reviewer #2 (Public review):

Recent findings in the field of motor learning have pointed to the combined action of multiple mechanisms that potentially contribute to changes in motor output during adaptation. A nearly ubiquitous motor learning process occurs via the trial-by-trial compensation of motor errors, often attributed to cerebellar-dependent updating. This error-based learning process is slow and largely unconscious. Additional learning processes that are rapid (e.g., explicit strategy-based compensation) have been described in discrete movements like goal-directed reaching adaptation. However, the role of rapid motor updating during continuous movements such as walking has been either under-explored or inconsistent with those found during the adaptation of discrete movements. Indeed, previous results have largely discounted the role of explicit strategy-based mechanisms for locomotor learning. In the current manuscript, Rossi et al. provide convincing evidence for a previously unknown rapid updating mechanism for locomotor adaptation. Unlike the now well-studied explicit strategies employed during reaching movements, the authors demonstrate that this stimulus-response mapping process is largely unconscious. The authors show that in approximately half of subjects, the mapping process appears to be memory-based while the remainder of subjects appear to perform structural learning of the task design. The participants that learned using a structural approach had the capability to rapidly generalize to previously unexplored regions of the perturbation space.

One result that will likely be particularly important to the field of motor learning is the authors' quite convincing correlation between the magnitude of proprioceptive recalibration and the magnitude error-based updating. This result beautifully parallels results in other motor learning tasks and appears to provide a robust marker for the magnitude of the mapping process (by means of subtracting off the contribution of error-based motor learning). This is a fascinating result with implications for the motor learning field well beyond the current study.

A major strength of this manuscript is the large sample size across experiments and the extent of replication performed by the authors in multiple control experiments.

Finally, I commend the authors on extending their original observations via Experiment 2. While it seems that participants use a range of mapping mechanisms (or indeed a combination of multiple mapping mechanisms), future experiments may be able to tease apart why some subjects use memory versus structural mapping. A future ability to push subjects to learn structurally-based mapping rules has the potential to inform rehabilitation strategies.

Overall, the manuscript is well written, the results are clear, and the data and analyses are convincing. The manuscript's weaknesses are minor, mostly related to the presentation of the results and modeling.

Weaknesses:

The overall weaknesses in the manuscript are minor and can likely be addressed with textual changes.

(1) A key aspect of the experimental design is the speed of the "ramp down" following the adaptation period. If the ramp-down is too slow, then no after-effects would be expected even in the alternative recalibration-only/error-based only hypothesis. How did the authors determine the appropriate rate of ramp-down? Do alternative choices of ramp-down rates result in step length asymmetry measures that are consistent with the mapping hypothesis?

(2) Overall, the modeling as presented in Figure 3 (Equation 1-3) is a bit convoluted. To my mind, it would be far more useful if the authors reworked Equations 1-3 and Figure 3 (with potential changes to Figure 2) so that the motor output (u) is related to the stride rather than the magnitude of the perturbation. There should be an equation relating the forward model recalibration (i.e., Equation 1) to the fraction of the motor error on a given stride, something akin to u(k+1) = r * (u(k) - p(k)). This formulation is easier to understand and commonplace in other motor learning tasks (and likely what the authors actually fit given the Smith & Shadmehr citation and the derivations in the Supplemental Materials). Such a change would require that Figure 3's independent axes be changed to "stride," but this has the benefit of complementing the presentation that is already in Figure 5.

Reviewer #3 (Public review):

Summary:

In this work, Rossi et al. use a novel split-belt treadmill learning task to reveal distinct sub-components of gait adaptation. The task involved following a standard adaptation phase with a "ramp-down" phase that helped them dissociate implicit recalibration and more deliberate SR map learning. Combined with modeling and re-analysis of previous studies, the authors show multiple lines of evidence that both processes run simultaneously, with implicit learning saturating based on intrinsic learning constraints and SR learning showing sensitivity to a "perceptual" error. These results offer a parallel with work in reaching adaptation showing both explicit and implicit processes contributing to behavior; however, in the case of gait adaptation the deliberate learning component does not appear to be strategic but is instead a more implicit SR learning processes.

Strengths:

(1) The task design is very clever and the "ramp down" phase offers a novel way to attempt to dissociate competing models of multiple processes in gait adaptation.

(2) The analyses are thorough, as is the re-analysis of multiple previous data sets.

(3) The querying of perception of the different relative belt speeds is a very nice addition, allowing the authors to connect different learning components with error perception.

(4) The conceptual framework is compelling, highlighting parallels with work in reaching but also emphasizing differences, especially w/r/t SR learning versus strategic behaviors. Thus the discovery of an SR learning process in gait adaptation would be both novel and also help conjoin different siloed subfields of motor learning research.

Weaknesses:

(1) The behavior in the ramp-down phase does indeed appear to support multiple learning processes. However, I may have missed something, but I have a fundamental worry about the specific modeling and framing of the "SR" learning process. If I correctly understand, the SR process learns by adjusting to perceived L/R belt speed differences (Figure 7). What is bugging me is why that process would not cause the SR system to still learn something in the later parts of the ramp-down phase when the perceived speed differences flip (Figure 4). I do believe this "blunted learning" is what the SR component is actually modeled with, given this quote in the caption to Figure 7: "When the perturbation is perceived to be opposite than adaptation, even if it is not, mapping is zero and the Δ motor output is constant, reflecting recalibration adjustments only." It seems a priori odd and perhaps a little arbitrary to me that a SR learning system would just stop working (go to zero) just because the perception flipped sign. Or for that matter "generalize" to a ramp-up (i.e., just learn a new SR mapping just like the system did at the beginning of the first perturbation). What am I missing that justifies this key assumption? Or is the model doing something else? (if so that should be more clearly described).

(2) A more minor point, but given the sample size it is hard to be convinced about the individual difference analysis for structure learning (Figure 5). How clear is it that these two groups of subjects are fully separable and not on a continuum? The lack of clusters in another data set seems like a somewhat less than convincing control here.