Blocking trial-by-trial error correction does not interfere with motor learning in human walking

Abstract

Movements can be learned implicitly in response to new environmental demands or explicitly through instruction and strategy. The former is often studied in an environment that perturbs movement so that people learn to correct the errors and store a new motor pattern. Here, we demonstrate in human walking that implicit learning of foot placement occurs even when an explicit strategy is used to block changes in foot placement during the learning process. We studied people learning a new walking pattern on a split-belt treadmill with and without an explicit strategy through instruction on where to step. When there is no instruction, subjects implicitly learn to place one foot in front of the other to minimize step-length asymmetry during split-belt walking, and the learned pattern is maintained when the belts are returned to the same speed, i.e., postlearning. When instruction is provided, we block expression of the new foot-placement pattern that would otherwise naturally develop from adaptation. Despite this appearance of no learning in foot placement, subjects show similar postlearning effects as those who were not given any instruction. Thus locomotor adaptation is not dependent on a change in action during learning but instead can be driven entirely by an unexpressed internal recalibration of the desired movement.

the human nervous system uses a repertoire of learning mechanisms to change walking patterns to account for a variety of environments and situations. For example, we explicitly think about where to place our feet when stepping on stones in a river but adapt more implicitly to walking in snow. We currently do not understand how these explicit and implicit processes interact when they are put in conflict or engaged toward the same learning goal. For example, if an explicit strategy is used to block the expression of implicit learning, does the latter still occur? Here, we studied the interactions between these explicit and implicit processes using a strategy via visual feedback and a split-belt treadmill, respectively.

When subjects walk on a split-belt treadmill with one belt moving faster than the other, they learn a new gait pattern over hundreds of steps by changing where and when they place their feet on the treadmill (Finley et al. 2015; Long et al. 2015; Reisman et al. 2005). When the treadmill belts are then returned to the same speed, postlearning effects are observed such that subjects retain much of the newly learned spatial pattern that then decays over a couple hundred steps (Reisman et al. 2005). This learning is largely implicit, particularly when the perturbation is gradually introduced (Sawers et al. 2013; Torres-Oviedo and Bastian 2012): subjects are often unaware that there is a perturbation for much of the learning period. Additionally, subjects are hard-pressed to describe how their walking is changing during this process, and they are surprised that they have postlearning effects once the perturbation is abruptly removed.

It is generally believed that these postlearning effects result from the updating of an implicit, cerebellar-driven internal model of how the body should move in an environment (Morton and Bastian 2006; Wolpert et al. 1995). Recently, we showed that online visual feedback of the walking pattern accelerates adaptation to an abrupt perturbation during split-belt treadmill walking (Malone and Bastian 2010), suggesting that explicit and implicit learning processes can work together to recalibrate gait to a new environment more quickly.

In this study, we asked whether explicit strategies monitored via visual feedback with a controlled learning rate would influence the postlearning effects typically observed as a result of implicit learning. We compared the effects of providing explicit strategies that were either incongruent or congruent with the goal of implicit learning (i.e., reducing step-length asymmetry). We hypothesized that explicitly blocking the expression of a new gait pattern using incongruent visual feedback should prohibit implicit learning such that no postlearning effects are observed once the belts return to the same speed. Data supporting this hypothesis have been shown previously for blocking learning in an abrupt split-belt paradigm (Malone et al. 2012). Surprisingly, we observed that blocking expression of the new pattern during gradual learning was not sufficient to block the postlearning effects, but rather the postlearning effects were completely unaffected by the clamping of the motor output during learning. Furthermore, we also observed that congruent visual feedback (i.e., a coach) neither enhances nor interferes with the postlearning effects. These striking results suggest that we can implicitly learn a new walking pattern without ever changing our motor output during the learning phase.

METHODS

Subjects.

Fifty healthy adults (thirty-two women, 18–38 yr old) gave informed, written consent before volunteering for this study. The Johns Hopkins Institutional Review Board approved all protocols.

Overall protocol.

All subjects walked on a split-belt treadmill, some with additional visual feedback provided on a television screen (1.25 × 0.69 m) placed in front of the treadmill as shown in Fig. 1A. The visual feedback and treadmill belt speeds were controlled with custom software (Vizard; WorldViz). We refer to the portions of the experiment in which the belts moved at the same speed as “tied belts” and different speeds as “split belts.” During split-belt walking, the right belt moved faster than the left. A thin board was placed between the two belts to prevent subjects from stepping from one belt to the other. To prevent falling, subjects wore a non-load-bearing safety harness attached to the ceiling and held onto the handrail of the treadmill for a few seconds at the beginning and end of each walking trial. During most portions of the experiment, subjects were instructed to avoid looking at their feet and to look instead at the display screen while they were walking. One group of subjects as described below was instructed to look at their feet during learning. Occasionally, the visual feedback was not displayed and subjects were allowed to watch a movie or show on the display.

Fig. 1.

A: experimental setup with a split-belt treadmill and a visual feedback display. The visual feedback indicated the foot position, target zones, and foot placement (end points). The target zones turned green for success and red for failure. B: the treadmill belt speeds as a function of time for the baseline, learning, and postlearning phases. The visual feedback was displayed for the gray portions. C: foot placement is defined as the distance between the hip center and the ankle at heel strike. Step length is defined as the distance between the 2 ankles at heel strike. Tracking markers were placed bilaterally on the greater trochanter (GT) and lateral malleolus (LM).

Subjects completed a basic paradigm that consisted of a baseline phase, learning phase, and postlearning phase (see Fig. 1B). The baseline phase consisted of 2 min of walking without visual feedback, 2 min of walking with visual feedback, and 2 min of walking without visual feedback. The baseline period with visual feedback was used to familiarize the subjects with the visual feedback. The belt speeds were tied at 0.5 m/s during all parts of the baseline phase.

During the learning phase, the right treadmill belt was gradually increased over a period of 10 min from 0.5 to 1.5 m/s by increasing the right speed every 3 s and then held constant for 5 min more as shown in Fig. 1B. The left treadmill belt was held constant at 0.5 m/s during this period. All subjects received a short break every 5 min. The subjects received visual feedback during the learning phase (unless specified below). We decided to use a gradual learning paradigm because 1) it allows us to control the rate of learning, and 2) previous split-belt work has demonstrated that conscious correction with feedback modulates the rate of learning and may interact with postlearning effects (Malone and Bastian 2010). By controlling the rate of learning, we can confidently compare the postlearning effects from multiple feedback experiments.

For all experiments, the postlearning phase consisted of 15 min of tied-belt walking at 0.5 m/s with no visual feedback and with no breaks. This postlearning speed was selected as it has shown to produce the largest postlearning effects (Vasudevan and Bastian 2010). The treadmill was stopped briefly between the learning phase and postlearning phase for safety reasons and to turn off any feedback. Before beginning the postlearning phase, subjects were not told whether the belt speeds would change, and they were instructed to “do whatever felt natural.” This allowed us to measure the automatic postlearning effects without a specific explicit strategy.

Data collection.

Three-dimensional kinematic data of infrared-emitting markers were recorded at 100 Hz (filtered with a fourth-order Butterworth filter) with an Optotrak motion capture system (Northern Digital). Markers were placed bilaterally over the lateral malleolus and greater trochanter as shown in Fig. 1C. Subjects walked on a split-belt treadmill (Woodway Split-Belt treadmill) that was equipped with a vertical force sensor under each belt. These forces were recorded at 1,000 Hz (smoothed by five time steps) and were synchronized with the motion capture data. The forces were used to detect heel-strike events as a threshold crossing of ∼10% of the subject's weight. The kinematic data and forces were used in real-time in calculation of the visual feedback provided to some of the subjects.

Gait parameters.

It has been shown previously that split-belt adaptation induces learning to reduce step-length asymmetry (Finley et al. 2015; Long et al. 2015; Malone and Bastian 2010; Reisman et al. 2005). Here, step length is defined as the anterior-posterior distance between the ankle markers at heel strike as shown in Fig. 1C. The fast step length refers to step length at heel strike on the fast belt, and slow step length refers to step length at heel strike on the slow belt. We have previously shown that the differences in step lengths can be decomposed into the sum of spatial, temporal, and perturbation contributions (Finley et al. 2015; Long et al. 2015).

$$\text{Step-Length Difference}=\text{Fast Step Length}-\text{Slow Step Length}$$

$$=\text{Spatial}+\text{Temporal}+\text{Perturbation}\text{.}$$

For a derivation of this decomposition, see our previous work (Finley et al. 2015; Long et al. 2015). The spatial term is based on where subjects place their feet relative to their body at heel strike:

$$\text{Spatial}={\alpha }_{\text{F}}-{\alpha }_{\text{s}}$$

where α_F is where the fast foot is placed relative to the previous slow foot placement, and α_s is where the slow foot is placed relative to the previous fast foot placement. We define foot placement as the anterior-posterior distance (in millimeters) between the ankle at heel strike and the average of the two hip markers as shown in Fig. 1C. The temporal term is based on the difference in step times:

$$\text{Temporal}=\frac{v_{\text{s}}+v_{\text{f}}}{\text{2}}\left(t_{\text{s}}-t_{\text{f}}\right)$$

where t_s is the slow step time, t_f is the fast step time, v_s is the average speed of the slow ankle relative to the body during the slow step time, and v_f is the average speed of the fast ankle relative to the body during the fast step time. Note that the step times are in time units (seconds) and therefore must be multiplied by the average speeds (millimeters per second) to compute how they contribute to the step-length asymmetry (millimeters). We define step time as the time between heel strikes on opposite sides of the body. The perturbation term is based on the difference in belt speeds.

$$\text{Perturbation}=\frac{t_{\text{s}}+t_{\text{f}}}{\text{2}}\left(v_{\text{s}}-v_{\text{f}}\right).$$

Throughout all phases of the experiment, we calculate step-length difference as well as the spatial, temporal, and perturbation contributions.

In our previous work, we showed that the spatial and temporal terms adapt to cancel an abrupt split-belt perturbation to reduce step-length differences (Finley et al. 2015; Long et al. 2015). Additionally, we observed that the postlearning step-length differences were dominated by changes in the spatial term: the temporal and perturbation terms were small and decayed rapidly (Long et al. 2015). This is also true as shown here for a gradual split-belt learning protocol: the spatial term explains a majority of the postlearning effect in step-length difference (Fig. 2). Therefore, we were interested in providing visual feedback about foot placement to add explicit strategy to the split-belt walk protocol and will focus primarily on the spatial component throughout the manuscript. Experiments with temporal feedback have been previously attempted, but prior work has suggested that subjects may not be able to adjust their gait timing consciously to a resolution similar to that of spatial parameters (Malone et al. 2012).

Fig. 2.

Stride-by-stride results (group mean ± 2 SE; in mm) for step-length difference, spatial component, temporal component, and perturbation component for the IMPLICIT_ONLY group. The number of strides in each phase is cropped to the person with the fewest strides. Note that the learning phase is broken into 3 5-min sections.

Visual feedback.

The visual feedback provided on the display showed foot position of each side of the body. The display (Fig. 1A) provides feedback to the subject about the real-time foot position relative to the body, foot placement at heel strike (end points), and target zones (horizontal bars). The color of target zones briefly showed success (green) or failure (red) at heel strike. The width of each target zone was set at 4 cm for all subjects. The subjects were able to see where they stepped relative to the targets, so they were easily able to determine directional information from the end-point feedback. The feedback display focused on the heel strike instead of the toe-off because previous work has suggested that parameters associated with heel strike adapt in a feedforward manner, whereas parameters associated with toe-off do not exhibit typical motor output behavior (Malone et al. 2012).

Experimental groups.

Subjects were randomly divided into five different groups. The first group, IMPLICIT_ONLY (n = 10), performed the gradual split-belt learning protocol shown in Fig. 1B while watching television (i.e., no visual feedback). This group served as a no-feedback control for comparison in all experiments in this study. Since this group did not receive visual feedback, the baseline phase consisted of only 2 min of walking.

A second group of subjects, INCONGRUENT (n = 10), was tested with the split-belt protocol shown in Fig. 1B and with visual feedback displayed on the screen. Both the left and right target locations did not change during the learning phase and were fixed at the average foot placements for the baseline of the IMPLICIT_ONLY group. Another group, LOOK_DOWN (n = 10), performed a similar experiment but looked at their feet to step on symmetric targets located next to the treadmill. We tested the LOOK_DOWN group to understand whether direct observation of the feet and treadmill belts rather than a virtual representation of the feet influenced learning. In both INCONGRUENT and LOOK_DOWN, subjects were instructed to walk with nearly equal foot placements, even when the belt speeds became asymmetric. In both of these groups, there was a conflict between the natural adaptation pattern (where the spatial component grows over time) and adherence to the strategy (where the spatial component is near 0). We predicted that this clamping of the spatial component near 0 would block the postlearning effects. We also tested whether the subjects implicitly learned the spatial component despite adherence to the strategy with a 10-s catch trial after 12 min of learning. During the catch trial, subjects did not look at their feet, were not given any feedback on the screen, and walked with the 3:1 split in the belt speeds. Following the catch trial, the feedback (either on the screen or by looking at their feet) was restored. Note that the treadmill was stopped briefly between each of these transitions.

Two additional groups (CONGRUENT: n = 10; EXPLICIT_ONLY: n = 10) were tested with visual feedback on the screen that was congruent with implicit learning to investigate whether the visual feedback alone induced any postlearning effects when placed either in concert with implicit learning or in isolation. In these groups, the target locations during the learning phase were based on the foot placements of IMPLICIT_ONLY's learning phase (averaged across subjects). These target locations were used to ensure that the foot placements in CONGRUENT and EXPLICIT_ONLY changed similarly to IMPLICIT_ONLY over the 15-min learning phase. CONGRUENT performed the split-belt protocol shown in Fig. 1B, whereas EXPLICIT_ONLY performed tied-belt walking at 0.5 m/s. This meant that, by the end of learning, both CONGRUENT and EXPLICIT_ONLY demonstrated the same asymmetric foot-placement patterns, but only CONGRUENT experienced asymmetric belt speeds.

Statistical analysis.

For all groups, we calculated the spatial, temporal, and perturbation contributions as well as the step-length difference during late learning (mean of last 30 strides) and initial postlearning (mean of 1st 5 strides). For INCONGRUENT and LOOK_DOWN, we also measured the mean of the 30 strides leading up to the catch trial and the mean of the strides in the catch trial to measure whether the subjects were still learning where to place their feet despite incongruent instruction. ANOVAs or 2-sample t-tests (when only 2 groups) were used to compare these measures between groups. Post hoc tests were conducted when appropriate. Decay rates for the spatial component were quantified using 20 epochs of 10 strides for the 1st 200 strides (with 5 strides in the 1st bin) similar to Malone and Bastian (2014). Mixed-model repeated-measures ANOVAs (rmANOVAs) with epoch as the within-subjects factor and group as the between-subjects factor were used to compare decay rates over time and among groups. All statistical analysis was conducted in MATLAB and with an α value = 0.05.

RESULTS

Postlearning effects in step-length difference arise from the spatial component for gradual split-belt learning.

All subjects in this study walked on a split-belt treadmill (Fig. 1A). One group of subjects (IMPLICIT_ONLY) walked without visual feedback as the right belt speed was gradually increased according to the protocol shown in Fig. 1B. The step length and foot placement (see Fig. 1C) was recorded using tracking markers throughout the experiment. Step-length difference (fast minus slow) during this walking task can be broken down into spatial, temporal, and perturbation components as shown in Fig. 2 (Finley et al. 2015; Long et al. 2015). To minimize the step-length difference during learning, the spatial and temporal components increase to cancel the perturbation component. The spatial and temporal components cancel most of the perturbation, keeping the step-length difference within 100 mm on average throughout learning. As in our previous results (Long et al. 2015), we observed that the postlearning effects in step-length difference in IMPLICIT_ONLY were due largely to the learning of the spatial component, which only depends on where subjects were placing their feet.

In this experiment, we see aftereffects in the difference in step times as seen previously by Malone and colleagues (2012). However, when the difference in step times is multiplied by the average belt speed in the temporal component term of the model, the postlearning effects of the temporal component have very small contribution to the step-length difference aftereffect. Since the temporal and perturbation components contribute very little to the step-length difference postlearning effect, we focus our analysis largely on the spatial component (i.e., where subjects are placing their feet).

The spatial component can be clamped through late learning.

The step-length difference model for the INCONGRUENT and LOOK_DOWN groups are shown with respect to IMPLICIT_ONLY in Figs. 3 and 4, respectively. The subjects in IMPLICIT_ONLY did not receive visual feedback, and their foot placements (i.e., spatial component) gradually changed during learning. In contrast, subjects in the INCONGRUENT and LOOK_DOWN groups were instructed to walk with asymmetric step lengths due to symmetric foot placements throughout the learning phase via a feedback display screen (see Fig. 1A) or by looking at their feet, respectively. Subjects in these groups adhered to the strategy well (74.4 ± 2.4% success for INCONGRUENT), and the spatial component as shown in Figs. 3B and 4B did not grow in late learning compared with IMPLICIT_ONLY [ANOVA: F(2,27) = 52, P < 0.001; post hoc tests: P < 0.001 for each group compared with IMPLICIT_ONLY]. As a result, the step-length difference error grew throughout the learning period and was much greater compared with the IMPLICIT_ONLY group by the end [ANOVA: F(2,27) = 24, P < 0.001; post hoc tests: P < 0.001 for each group compared with IMPLICIT_ONLY] as shown in Figs. 3A and 4A.

Fig. 3.

Stride-by-stride results (group mean ± 2 SE; in mm) for step-length difference (A), spatial component (B), temporal component (C), and perturbation component (D) for the INCONGRUENT and IMPLICIT_ONLY groups. The number of strides in each phase is cropped to the person with the fewest strides. Note that the learning phase is broken into 3 5-min sections.

Fig. 4.

Stride-by-stride results (group mean ± 2 SE; in mm) for step-length difference (A), spatial component (B), temporal component (C), and perturbation component (D) for the LOOK_DOWN and IMPLICIT_ONLY groups. The number of strides in each phase is cropped to the person with the fewest strides. Note that the learning phase is broken into 3 5-min sections.

During late learning, we also observed significant differences between groups in the temporal [ANOVA: F(2,27) = 6.6, P < 0.01] and perturbation [ANOVA: F(2,27) = 4.2, P = 0.03] contributions. Post hoc tests showed that LOOK_DOWN had a higher temporal contribution than INCONGRUENT (P < 0.01) and IMPLICIT_ONLY (P = 0.02) but no difference between the INCONGRUENT and IMPLICIT_ONLY groups (P = 0.6). We also observed that IMPLICIT_ONLY's perturbation was not different from either INCONGRUENT (P = 0.1) or LOOK_DOWN (P = 0.13), but the perturbation was larger in LOOK_DOWN compared with INCONGRUENT (P = 0.02). As a result, it appears that the subjects in LOOK_DOWN increased the temporal contribution as they increased the perturbation contribution, whereas the INCONGRUENT and IMPLICIT_ONLY groups performed both similarly.

Clamping the spatial component does not block postlearning effects.

We predicted that blocking the expression of spatial learning might eliminate the postlearning step-length difference effects. This was not the case: despite showing little change in foot placement during learning, INCONGRUENT and LOOK_DOWN both showed initial postlearning effects similar to IMPLICIT_ONLY [step-length difference: F(2,27) = 1.2, P = 0.32; spatial: F(2,27) = 1.8, P = 0.19; temporal: F(2,27) = 0.24, P = 0.8; perturbation: F(2,27) = 1.6, P = 0.23]. As a result, it appears that forcing subjects to walk with nearly symmetric foot placements and a large step-length asymmetry during learning is not sufficient to block the postlearning effects. Similarly, we observed no differences between IMPLICIT_ONLY and INCONGRUENT in the decay of the spatial component [rmANOVA: group effect, F(1,18) = 0.02, P = 0.9; interaction effect, F(10,180) = 0.7, P = 0.7]. However, we did observe that subjects in LOOK_DOWN had a faster decay compared with IMPLICIT_ONLY [rmANOVA: group effect, F(1,18) = 4.9, P = 0.04; interaction effect, F(10,180) = 1.3, P = 0.2] and compared with INCONGRUENT [rmANOVA: group effect, F(1,18) = 0.08; interaction effect, F(10,180) = 2.39, P = 0.01]. Thus it appears that the method of feedback can influence the rate of decay.

To examine whether subjects were still learning the spatial pattern during the learning phase, we introduced a catch trial that momentarily removed the feedback while maintaining the split belts. The spatial component immediately jumped to a higher value in this catch trial for both groups (P < 0.001) as shown in Figs. 3B and 4B, but the INCONGRUENT value was higher than the LOOK_DOWN value (P = 0.03). Note that these values are lower than the IMPLICIT_ONLY spatial component (both P < 0.001), but this may be due to the fact that the treadmill was stopped briefly to turn off the feedback. The immediate jump in the spatial components indicates that the subjects were still learning how to walk on the split-belt treadmill despite the maladaptive incongruent instruction.

Congruent explicit strategy neither enhances nor interferes with implicit learning.

The step-length difference components for the CONGRUENT and IMPLICIT_ONLY groups are shown in Fig. 5. Subjects in the CONGRUENT group experienced gradual splitting of the treadmill belt speeds while receiving congruent visual feedback based on IMPLICIT_ONLY's foot placements. This was designed so the explicit strategy was consistent with implicit learning of foot placement (i.e., the instruction guided the subjects to step with the same foot placements that IMPLICIT_ONLY developed naturally during the learning phase). Subjects in CONGRUENT learned a similar gait pattern as IMPLICIT_ONLY by the end of the learning phase (step-length difference, P = 0.18; spatial, P = 0.39; temporal, P = 0.98; perturbation, P = 0.49), indicating that subjects adhered to the instruction (82.6 ± 1.9% success) and the visual feedback did not interfere with the learning phase. Furthermore, subjects in CONGRUENT and IMPLICIT_ONLY were similar during the initial postlearning effect (step-length difference, P = 0.68; spatial component, P = 0.87; temporal component, P = 0.56; perturbation component, P = 0.85). Similarly, the spatial contribution decayed at a similar rate for both groups [rmANOVA: group effect, F(1,18) = 0.03, P = 0.9; interaction effect, F(10,180) = 0.78, P = 0.6]. Overall, this suggests that explicit instruction in the form of congruent visual feedback does not enhance nor interfere with the automatic postlearning behavior of implicit learning.

Fig. 5.

Stride-by-stride results (group mean ± 2 SE; in mm) for step-length difference (A), spatial component (B), temporal component (C), and perturbation component (D) for the CONGRUENT, EXPLICIT_ONLY, and IMPLICIT_ONLY groups. The number of strides in each phase is cropped to the person with the fewest strides. Note that the learning phase is broken into 3 5-min sections.

Congruent explicit strategy in isolation does not produce postlearning effects.

To isolate whether the feedback alone had any effect on postlearning effects, we tested another group of subjects (EXPLICIT_ONLY) with the same visual feedback as CONGRUENT but kept the belt speeds tied at 0.5 m/s. This tied-belt walking produced a slower stride time than IMPLICIT_ONLY and thus reduced the total number of strides taken during the learning phase as can be seen in the stride-by-stride plots in Fig. 5. Subjects in EXPLICIT_ONLY followed the visual feedback (75.7 ± 1.8% success) by expressing asymmetric foot placements (i.e., non-0 spatial component) by late learning (P < 0.001) despite the tied belts but showed no initial postlearning effects (for non-0 mean: step-length difference, P = 0.48; spatial component, P = 0.35; temporal component, P = 0.13; perturbation component, P = 0.22). To verify that the number of strides in the learning phase did not influence this result, we conducted a regression with independent variables of group (IMPLICIT_ONLY vs. EXPLICIT_ONLY) and number of strides and dependent variable of the initial postlearning effect in the spatial component. This regression revealed that group (P < 0.001) was a significant predictor, but number of strides was not a significant predictor (P = 0.6). In other words, the fewer strides observed in EXPLICIT_ONLY did not influence this result. More importantly, this result indicates that subjects immediately switched off the explicit strategy when the feedback was removed and they were informed to “do whatever felt natural” in the postlearning phase. This suggests that explicit instruction of new foot placements cannot substitute for the immediate effects of split-belt walking adaptation and has little influence on the postlearning effects observed in INCONGRUENT and CONGRUENT.

DISCUSSION

Explicit and implicit processes are constantly being used by the nervous system to adjust the walking pattern to changes in the environment. Here, we demonstrate that these processes can be used successfully in parallel and that they show little interference. Perhaps the most remarkable finding from this work is that the changes in motor output of foot placement during implicit learning need not occur for learning to proceed. Specifically, postlearning effects are surprisingly normal even when an explicit strategy prohibits any change in the motor output during implicit learning. This means that the error (i.e., step-length asymmetry) previously thought to drive implicit learning need not be corrected for learning to occur.

Explicit and implicit contributions to motor adaptation have recently been studied in work using visuomotor reaching paradigms. These findings have shown that adaptation can be decomposed into an implicit component and an explicit aiming strategy (Taylor and Ivry 2011; Taylor et al. 2014). An aiming strategy can be acquired with or without instruction and helps subjects to counter the perturbation more rapidly (Mazzoni and Krakauer 2006; Taylor et al. 2014). However, subjects voluntarily change this explicit aiming strategy over time as they simultaneously learn implicitly from sensory-prediction errors (i.e., mismatches between subjects' actual and perceived hand position) even after they have completely counteracted the perturbation (Mazzoni and Krakauer 2006; Taylor et al. 2014). Thus, in the visuomotor reaching paradigm, explicit and implicit contributions to adaptation interact with one another and work together to recalibrate the motor system. The same visual signal drives these processes, although this signal is compared to different references (e.g., target versus prediction of hand position).

In our work, we were able to investigate the independence of these processes by providing explicit strategy errors based on visual feedback and implicit sensory-prediction errors based on proprioceptive signals (Torres-Oviedo and Bastian 2010). Here, subjects can theoretically minimize both error types independently. The INCONGRUENT group showed this behavior as they successfully stepped in the visual target regions (i.e., reduce visual error) and still showed postlearning effects similar to those of IMPLICIT_ONLY. This demonstrates that the implicit processes can proceed independently of explicit strategy. We show that they operate on different errors since the nervous system is able to build a new predictive model even in the presence of the maladaptive explicit strategy. Similar results have also been seen for subjects who were distracted during a sequence learning task (Seidler et al. 2002, 2005) as well as with visuomotor rotations (Taylor and Ivry 2011). In these studies, there were no performance changes during the learning phase, but aftereffects were revealed during a postlearning phase. Our results combined with the sequence learning and reaching experiments provide clear evidence that the nervous system can learn implicitly (i.e., minimize sensory-prediction errors) even if the motor output is not altered during learning.

We also observed a postlearning effect when the spatial component was clamped while subjects looked at their feet, but this effect decayed away more quickly than that of IMPLICIT_ONLY and had a much smaller catch trial when the feedback was removed. This effect is similar to that seen when subjects look at their feet during an abrupt split-belt paradigm (Malone et al. 2012). It is likely that the nervous system is able to predict based on visually seeing the belts moving at different speeds, which likely limits the construction of a robust internal model. Note that in the INCONGRUENT group, subjects are likely paying more attention to their success rather than the speed of the moving markers.

In this study, we were also interested in understanding whether learning shifted to the explicit process when the visual feedback was congruent with implicit learning. Previous work in our laboratory has compared simultaneous implicit and explicit learning during walking, although with a different type of paradigm (Malone and Bastian 2010). Those subjects were exposed to an abrupt split in the treadmill belts and were instructed to equalize their step lengths using video feedback of their legs. Video feedback induced faster learning but no difference in the postlearning effects (i.e., size and rate of washout) compared to a control group. Since the learning rates differed between groups, the effect of the explicit process on the postlearning phase is unclear, as previous work in reaching showed that the rate of learning also influences the rate of washout (Huang and Shadmehr 2009).

Here, we hypothesized that providing visual feedback congruent with the rate of implicit learning could interfere with postlearning effects (Benson et al. 2011; Malone and Bastian 2010; Wulf and Prinz 2001). Subjects may rely more on the visual feedback and thus learn less through the implicit process (Schmidt 1991). Contrary to our hypothesis, subjects with congruent visual feedback during gradual learning showed similar postlearning effects (i.e., size and rate of washout) compared to a control group with no visual feedback. This suggests that explicit control of walking neither interferes with nor enhances the postlearning effects of implicit learning.

As a control experiment, we were also interested in understanding whether explicit instruction in isolation influences postlearning effects. We had hypothesized that we may see a small effect due to use-dependent plasticity (Diedrichsen et al. 2010; Verstynen and Sabes 2011), but this was not the case following 15 min of isolated explicit instruction. This result is similar to a reaching experiment that had subjects use an explicit strategy when aiming at a target with veridical feedback. This work showed that subjects had no difficulty abruptly switching between explicit strategies, which resulted in no postlearning effects (Mazzoni and Krakauer 2006). As a result, it is possible that subjects were relying on the feedback to produce the asymmetric foot placements in our explicit instruction experiment and therefore did not have to recalibrate their automatic walking patterns.

The cerebellum has long been considered the site of implicit error-based learning as patients with cerebellar damage show impairments in adaptation to split-belt walking (Morton and Bastian 2006), throwing with prisms (Martin et al. 1996; Weiner et al. 1983), force-field reaching (Rabe et al. 2009; Smith and Shadmehr 2005), and visuomotor rotation reaching (Gibo et al. 2013; Izawa et al. 2012; Rabe et al. 2009; Taylor et al. 2010). However, neural circuits involved in explicit motor learning are still unknown. Recently, Taylor et al. (2010) instructed patients with cerebellar damage to use an explicit strategy to cancel a visuomotor rotation successfully; they were able to do this but did not show postlearning effects. This same experimental paradigm was then repeated with patients with prefrontal cortex lesions, who had difficulty following the explicit strategy but showed postlearning effects (Taylor and Ivry 2014). This double dissociation suggests that the explicit strategy depends on prefrontal cortex. Other studies of patients with prefrontal cortex lesions demonstrated that these patients have reduced explicit awareness and impairments in learning (Gómez Beldarrain et al. 1999; Slachevsky et al. 2001, 2003; Taylor and Ivry 2014). This idea of the prefrontal cortex as an important site involved in explicit learning is further supported by work in neuroimaging, which indicates that the prefrontal cortex is active during sensorimotor adaptation (Floyer-Lea and Matthews 2004, 2005; Sakai et al. 1998; Shadmehr and Holcomb 1997). Therefore, it is possible in our walking experiments that the prefrontal cortex uses visual error to update an explicit strategy, whereas the cerebellum uses sensory-prediction error to update the implicit internal model.

In summary, our experiments with explicit learning via visual feedback and implicit learning via a split-belt treadmill provide evidence that the implicit learning processes can act independently of explicit strategies, depend on different error signals, and may operate in parallel in the central nervous system.

GRANTS

This work was supported by National Institutes of Health Grants HD-048741, NS-092241, and NS-090751.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the author(s).

AUTHOR CONTRIBUTIONS

A.W.L. and A.J.B. conception and design of research; A.W.L. performed experiments; A.W.L. analyzed data; A.W.L. and A.J.B. interpreted results of experiments; A.W.L. prepared figures; A.W.L. drafted manuscript; A.W.L., R.T.R., and A.J.B. edited and revised manuscript; A.W.L., R.T.R., and A.J.B. approved final version of manuscript.