Environmental consistency determines the rate of motor adaptation

Abstract

Background: The motor system has the remarkable ability not only to learn but also to learn how fast it should learn. However, the mechanisms behind this ability are not well understood. Previous studies have posited that the rate of adaptation in a given environment is determined by Bayesian sensorimotor integration based on the amount of variability in the state of the environment. However, experimental results have failed to support several predictions of this theory.

Results: We show that the rate at which the motor system adapts to changes in the environment is primarily determined not by the degree to which environmental change occurs but by the degree to which the changes that do occur persist from one movement to the next, i.e., the consistency of the environment. We demonstrate a striking double dissociation whereby feedback response strength is predicted by environmental variability rather than consistency, whereas adaptation rate is predicted by environmental consistency rather than variability. We proceed to elucidate the role of stimulus repetition in speeding up adaptation and find that repetition can greatly potentiate the effect of consistency, although unlike consistency, repetition alone does not increase adaptation rate. By leveraging this understanding, we demonstrate that the rate of motor adaptation can be modulated over a range that encompasses a 20-fold increase from lowest to highest.

Conclusions: Understanding the mechanisms that determine the rate of motor adaptation could lead to the principled design of improved procedures for motor training and rehabilitation. Regimens designed to control environmental consistency and repetition during training might yield faster, more robust motor learning.

Figure 1. Experimental paradigm

(A) Schematic of experimental setup. Subjects gripped the handle of a 2D robotic manipulandum and performed 10cm reaching arm movements in 2 directions (90° and 270°) based on targets presented on a vertical screen in front of them. (B) Illustration of force-field environments and adaptation rate measurement sequence. Example sections of force field activation patterns for the anti-consistent, inconsistent, medium and high consistency learning environments are shown. In a subset of the FF cycles of each environment, we introduced error-clamp trials before and after the first force-field trial of the cycle (vertical black bars and blowout area). The adaptation rate was calculated as the difference in force field compensation between the post and pre force field error-clamp trials (see Equation 1). The statistical consistency, R(1), of each environment is shown on the right.

Figure 2. Upregulation and downregulation of motor adaptation rates by environmental consistency

(A) Average initial adaptation rate for the P1N1, P1, P7, and P20 environments and evolution of adaptation rate in these environments as a function of the number of FF cycles experienced. Circles indicate actual data from FF cycles where the adaptation rates were measured and solid lines show a 3-point moving average. Errorbars indicate SEM across subjects. (B) Lateral force profiles comprising the single-trial adaptive response for the different learning environments. The average learning-related change in lateral force is shown for the first cycle of the experiment averaged across experiments (gray) compared to the average FF compensation observed in the last half of the measurement FF cycles in each different environment (colors). (C) Average initial adaptation rate for all the different learning environments (gray) compared to the last-half adaptation rates for each environment (colors). Errorbars indicate SEM. * p < 0.05, ** p < 0.01, *** p < 0.001. (D) Average perturbed hand paths seen during the first FF trial of the experiment (gray) and in the last half of the cycles in each environment (colors). The small circles near the midway point indicate the location of the peak speed point.

Figure 3. Environmental consistency modulates next-trial motor adaptation rates but not same-trial feedback control

(A) Schematic of learning environments including the random noise (RN) environment. Note that the RN environment was designed to have a high variability like the high-consistency environment (P20) but at the same time a near zero consistency like the inconsistent learning environment (P1). (B and C) Adaptation rate and feedback response strength (% reduction in kinematic error) in the last half of the P1N1, P1, RN, P7, and P20 environments. Statistical comparisons between RN and the other environments are shown (* p < 0.05, ** p < 0.01, *** p < 0.001). Errorbars indicate SEM. (D and E) The relationships between adaptation rate and environmental consistency (D) or variability (E) across experiments. Both before (gray solid line) and after including the RN and P1L data (black solid line), environmental consistency explains a large fraction of the variance in motor adaptation rate, R2>89% in both cases. However inclusion of the RN and P1L data reduces the ability of environmental variability to explain changes in adaptation rate, reducing the R² value from 69% to only 16%. The gray dotted horizontal line indicates the initial adaptation rate before exposure to the different environments. (G and H) The relationship between feedback response strength and environmental consistency (G) or variability (H) across experiments in a format similar to panels D and E. Both before and after including the RN data, environmental variability explains a large fraction of the variance in feedback response strength, R²=82% when including the RN/P1L data and R²=95% without. However inclusion of the RN/P1L data reduces the ability of environmental consistency to explain changes in feedback response strength, reducing the R² value from 58% to 0%. Errorbars indicate SEM. (F and I) Summary of the strength of the relationships (R2) between adaptation rate (F) or feedback response strength (I) and consistency or variability. The first pair of bars summarizes the univariate regression analyses shown in panel D, E, G, and H. The 3rd and 4th bars summarize the corresponding bi-variate regressions. Here the 3rd bar shows the improvement in R² when a univariate analysis based on consistency (C, blue) is augmented by variability (V, orange). The 4th bar shows the improvement when a univariate analysis based on V is augmented by C. Full results of the bivariate analysis are shown in tables S2 and S3. Errorbars indicate SEM.

Figure 4. Upregulation of adaptation rates cannot be explained by savings

(A) Comparison of the P20 (red), P7 (light green), P7 Long (dark green), P1 (light blue) and P1L (dark blue) environments. Notice that although the number of FF cycles is the same (27) for both the P20 and P7 learning environments, the number of FF trials experienced is different, 540 vs. 189. Correspondingly, the number of FF cycles for the P1 learning environment is 45 (i.e. only 45 FF trials). The P7-Long (P7L) and P1-Long (P1L) environments were designed to match the number of FF trials of the P20 environment (540), and hence assess the influence of the number of FF trials on the observed adaptation rate increases. (B) Lateral force profiles comprising the single-trial adaptive response in the P20, P7, P7L, P1 and P1L environments. Note that the subjects in the P7L environment, despite experiencing the same number of FF trials as those in the P20 environment, compensated less than those exposed to the P20 environment. Subjects exposed to the P1L environment compensated even less. The inset shows the average adaptation rates in the P20, P7, P7L, P1 and P1L experiments in the last third of each environment. Errorbars indicate SEM. Notice that subjects in the P20 environment exhibited adaptation rates that were significantly greater than those of subjects in all other environments. (C) Mean angular error, a second measure of motor adaptation, in late FF cycles in the P20, P7 and P7L environments. The error in trials 4, 5, 6 and 7 (gray shaded region) of the FF cycles was significantly lower in the P20 environment than P7 or P7L. The inset quantifies these differences. Errorbars indicate SEM. * p < 0.05, ** p < 0.01, *** p < 0.001.

Figure 5. Upregulation of adaptation rates cannot be explained by savings

(A) Illustration of the P7L-Opposite (P7L-Opp, yellow) experiment in which a single oppositely-directed FF perturbation was presented at the end of the P7L environment. The adaptation rate for this trial was compared to the adaptation rate displayed in the last third of trials in P7L (P7L, green). (B) Comparison of the average initial single-trial adaptive response (gray) with that observed for the last third of the P7L trials (green) and the P7L-Opposite trial (yellow). Note that in the P7L-Opposite trial, subjects produce a force compensation that is largely inappropriate for the FF experienced. (C) Hand trajectories during the null trials immediately following FF blocks (P7L-Catch, blue: average trajectories for each subject, pink: trajectories of trials where each subject experienced their maximum deviation) compared to hand trajectories during the single P7L-Opposite trial (yellow). Thin lines indicate individual subject data; thick lines indicate data averaged across subjects. Note that the most deviated P7L-Catch trajectories are similar to the P7L-Opp ones, indicating that, while the FF experienced during the P7L-Opposite trial was completely novel, the errors it elicited were not. (D) Peak lateral displacements / errors experienced during the P7L-Catch and P7L-Opposite trials. While on average each subject experienced weaker errors during P7L-Catch (blue), the largest P7L-Catch errors for each subject (pink) were similar to the errors experienced during P7L-Opposite. Inner errorbars indicate SEM (inner whereas outer errorbars indicate standard deviation across subjects.

Figure 6. Synergistic interaction between repetition and consistency for learning rate upregulation

(A) Illustration of the random walk (RW) environment (orange trace). Note that this environment was designed to have a variance that is similar to that of the random noise (RN) environment (57.71 vs. 55.37), while at the same time, have a consistency similar to that of the medium-consistency (P7) environment (0.76 vs. 0.74) without having its repetitive structure. (B) Lateral force profiles and average adaptation levels (inset) for the single-trial adaptive response in the second half of training for the RW, RN, P1, P1L and P7 environments. Note that while the force compensation in the RW environment (orange) is greater than that in the low consistency environments (RN, P1, P1L) – p<0.01 in all cases – indicating that consistency, even without repetition, leads to higher adaptation rates. However, the learning in RW does not reach the level of that in P7 (p<0.01), indicating that repetition can enhance consistency-modulated learning rate increases. Errorbars indicate SEM. * p<0.05, ** p<0.01, *** p<0.001. (C) Repetition and consistency-driven responses combine to produce the observed P7L-Opposite response. A response predicted by the hypothesis that both consistency and repetition contribute distinct components to the adaptation (combined-CR transfer, pink dashed curve) matches the P7L-Opposite response (yellow curve) much better than a response predicted by the hypothesis that consistency enables repetition-based learning (R-only transfer, purple dashed curve). Errorbars indicate SEM. (D) Same as (C) but for the P20 experiment, illustrating how the P20-Opposite response matches the combined-CR hypothesis rather than the R-only hypothesis.