Different mechanisms of contextual inference govern associatively learned and sensory-evoked postural responses

Significance

Central to human movement control is the accurate estimation of our body and environment’s current and future states. These contextual inferences allow our motor system to adapt to varying sensorimotor experiences. However, whether these inferences are made centrally and shared across different movements or operate in parallel is unknown. We used a robotic balance simulator to subtly or explicitly withdraw the active control over whole-body movement while standing. The (lack of) control over motion attenuated classically conditioned balance responses differently than the known suppression of sensory-evoked responses, demonstrating that different forms of movement control are not governed by a single, central contextual inference process. This outcome raises important considerations for unified explanations of how different context-driven movement systems make inferences.

Abstract

Human standing balance relies on the continuous monitoring and integration of sensory signals to infer our body’s motion and orientation within the environment. However, when sensory information is no longer contextually relevant to balancing the body (e.g., when sensory and motor signals are incongruent), sensory-evoked balance responses are rapidly suppressed, much earlier than any conscious perception of changes in balance control. Here, we used a robotic balance simulator to assess whether associatively learned postural responses are similarly modulated by sensorimotor incongruence and contextual relevance to postural control. Twenty-nine participants in three groups were classically conditioned to generate postural responses to whole-body perturbations when presented with an initially neutral sound cue. During catch and extinction trials, participants received only the auditory stimulus but in different sensorimotor states corresponding to their group: 1) during normal active balance, 2) while immobilized, and 3) throughout periods where the computer subtly removed active control over balance. In the balancing and immobilized states, conditioned responses were either evoked or suppressed, respectively, according to the (in)ability to control movement. Following the immobilized state, conditioned responses were renewed when balance was restored, indicating that conditioning was retained but only expressed when contextually relevant. In contrast, conditioned responses persisted in the computer-controlled state even though there was no causal relationship between motor and sensory signals. These findings suggest that mechanisms responsible for sensory-evoked and conditioned postural responses do not share a single, central contextual inference and assessment of their relevance to postural control, and may instead operate in parallel.

Human standing relies on the integration of sensory and motor signals to form probabilistic estimates of whole-body self-motion within the world (1–4). These estimates are compared to actual sensory feedback to detect unexpected movement, trigger corrective responses, and adapt ongoing control (5–8). Predictions arising from these processes can also be formed by associating desired motor behavior with external events. Classical conditioning can be used to train participants to produce automatic postural responses to sensory stimuli that are initially unrelated to normal balance control (9–12). This classical conditioning of postural responses relies on the pairing of an a priori neutral or irrelevant sensory stimulus (e.g., a tone) termed the conditioned stimulus (CS) with a balance perturbation as an unconditioned stimulus (US). With sufficient repetition of trials pairing these stimuli, the CS will come to elicit a postural response termed the conditioned response (CR) when the CS is presented in isolation. Subsequently, after training, the repetitive presentation of a CS without a US progressively reduces the emergence of CRs through a process called extinction, which is thought to reflect an additional learning process that establishes the irrelevance of the CS in this new context (13).

Critically, the emergence of conditioned responses can be suppressed through extinction in one context and maintained in another, showing that the brain can simultaneously retain multiple learned associations (13–16). Context then serves as an occasion setter to disambiguate the meaning of the CS and select the appropriate response (13). Similarly, motor adaptation during upper-limb movements follows contextual modulation where different motor memories can be maintained and disambiguated by several different contextual cues (17, 18), including reinforcement of reward (19), spatial location of feedback, target, or movement (20, 21), planned lead-in and follow-through of a motion (22), auditory cues (23), or even a given control point on an object (24). In standing balance, the recruitment of muscles for postural responses is also known to be context-dependent (25–27). For example, stereotypical lower limb muscle responses that are evoked by electrical activation of the vestibular system are completely absent when standing participants are externally supported by a fixed backboard (28, 29), or using only the arm to control balance (30). More notably, Luu et al. (29) showed that this contextual suppression of continuous vestibular contributions depends on the unconscious congruence between sensory and motor signals of ongoing balance. When participants stood in a robotic balance simulator that subtly took over control of balancing the platform, vestibular input to lower-limb muscles was rapidly disengaged (<200 ms). Yet, participants detected only 50% of the transitions and took over 2 s to do so. This implies that the nervous system can rapidly assess the contextual relevance of responding to a vestibular input without requiring a conscious perception that the postural control task has been interrupted. Here, our aim was to determine whether associatively learned postural responses share a similar assessment of contextual relevance through the subconscious integration of sensory and motor signals.

Human participants stood in a robotic balance simulator (Fig. 1A) that replicated the neural and physical characteristics of natural standing. Participants performed a classical conditioning paradigm that paired a neutral auditory stimulus (CS) with a postural perturbation (US) in the form of an added moment that accelerated the participant’s whole body in the posterior direction. In subsequent catch and extinction trials, participants in three different groups were presented a CS without a US with different control states depending on their group designation (Fig. 1 B and C). Participants in the Balance group (N = 10) received a CS while balancing, replicating previous reports that conditioned responses could be formed without the presence of a perturbation (9–12). Participants in the Immobile group (N = 10) were externally supported before the CS was delivered and were explicitly informed about their lack of control, similar to vestibular responses as shown by Fitzpatrick, Burke, and Gandevia (28). Last, for participants in the Computer-controlled group (N = 9), the CS was delivered 4 s after switching to a condition that provided the illusion of balance control, similar to the condition that Luu et al. (29) used to assess vestibular-evoked responses. The balance simulator would, unbeknownst to the participants, transition to a Computer-controlled state that replayed a segment of each participant’s own whole-body movement from a baseline trial, thereby breaking the congruence between currently applied balancing forces and experienced movement. We hypothesized that the continuous sensorimotor integration process that suppressed irrelevant vestibular signals in a Computer-controlled state (29) could similarly suppress conditioned responses, which are known to depend on their contextual relevance (13–16). Our results demonstrate that, similar to vestibular-evoked balance responses, associatively learned postural responses are appropriately evoked or suppressed when participants are consciously aware of their (in)ability to control movement. However, unlike the vestibular control of balance, the continuous, subconscious monitoring of sensorimotor incongruence was not sufficient, and therefore not responsible, for suppressing irrelevant associatively learned postural responses when a conditioning stimulus was provided after the control of balance was removed. This difference in subconscious modulation shows that contextual cues of control are not centrally inferred and generalized across these responses from a single process of sensorimotor integration. Instead, different mechanisms or inferences govern vestibular-evoked and associatively learned postural responses.

Fig. 1.

Overview of experimental setup and protocol. (A) Control loop and setup of the robotic balance simulator. The human central nervous system (CNS) forms an estimated state through sensorimotor integration. The balance controller forms motor commands that are sent to the muscles following a control policy. Computational models of balance control (left panel) propose that the predicted state formed by the forward model is compared to the estimated state (i.e., a comparator), allowing the brain to distinguish self-generated from imposed movement. Participant-generated ground reaction forces and moments in the robot are measured by the force plate and conveyed to the simulation of body mechanics. A moment perturbation (i.e., the US) can be added to the participant-generated ankle moment to disrupt the ongoing control of upright standing. The whole-body angle θ was actuated differently depending on the trial conditions in our three separate groups: using the normal outcome of the balance simulation (Balance group), at an immobile position (Immobile group), or through a computer-controlled signal (Computer-controlled group). The Balance, Immobile, and Computer-controlled conditions are marked with a balancing figure, lock, and computer, respectively. (B) Overview of the experimental protocol. The blue speaker indicates the presence of a CS, and the red perturbed figure signifies the presence of a US. The solid or dashed CS or US line indicates the presence or absence of the stimulus, respectively. Five catch trials were interspersed within the conditioning block, indicated by the dark gray area, following at least 30 conditioning trials. During the catch and extinction trials of the Immobile and Computer-controlled groups (black dashed lines), movement is not dependent on ground reaction forces and moments. Instead, movement is dictated by a constant angular position or a prerecorded motion profile, respectively (see Protocol for more information). Participants in the Immobile and Computer-controlled groups experienced a second extinction block in the Balance condition to test whether conditioned responses would re-emerge when returning to the balancing context. (C) Representative data of the whole-body angle θ and the filtered, rectified, and normalized right tibialis anterior (TA) EMG signal for single trials in each block. Catch and extinction trials show representative data from each group. The extinction traces are from a first trial within that block.

Results

Postural Responses can be Associatively Learned Using Sound Cues.

All participants (N = 29) first balanced in the robotic simulator at a target angle of −2° anterior (actual angle −1.99 ± 0.11°) using visual feedback, while being exposed to balance perturbations during twelve US-only trials (see Fig. 1 and Materials and Methods). To maintain balance in response to the perturbation, participants had to generate a dorsiflexion moment to accelerate themselves in the anterior direction. Electromyography (EMG) recordings were made of the tibialis anterior (TA), soleus (SOL), and medial gastrocnemius (mGAS) muscles, and changes in plantarflexion moment were used to represent the net contribution of all muscles’ activity to the anteroposterior control of posture at the ankle. In trials where balance was maintained (data presented from last five valid trials), participant responses were characterized by i) a large burst of TA muscle activity in all trials occurring on average at 85 ± 20 ms after the onset of the US (Fig. 2A, US-only), ii) a decrease in plantarflexion moment in all trials initiated 94 ± 34 ms after the US onset (Fig. 2B), and iii) the posterior displacement of the body to an average peak angle of 1.69 ± 0.56° that occurred 836 ± 167 ms after the perturbation (Fig. 2C). EMG activity and ankle moment responses during each trial were identified when these signals exceeded predefined thresholds based on balance periods preceding the CS- or US-onset (see Materials and Methods).

Fig. 2.

Tibialis anterior EMG activation, ankle moment, and whole-body angle responses prior to conditioning (US-only), as well as Early and Late in the conditioning phase prior to any catch trials. US-only trials are the last 5 valid recoveries during the US-only block. Trials of the conditioning block are split into Early and Late phases, which are trials 6 to 10 and the last 5 trials of the conditioning block prior to the first catch trial, respectively. (A–C) Group means (black line) with the bootstrapped 95% CI of the mean (gray area) in the top row, and the per-group data from this same period below. The CS (solid blue line when present, dashed when not present) occurs 400 ms before the US (red line). (D) Advancement in time of the initial EMG activity between the US-only and Late conditioning trials, where negative onset times indicate activity before the US was delivered. Dots are used for participant means and horizontal lines for group means. (E) Change in ankle moment in the CS-to-US interval. The decrease in plantarflexion moment (i.e., a negative delta M_AP) indicates a preemptive change in moment to the US that minimized the postural instability caused by the posterior moment perturbation. (F) Mean peak posterior angle a participant reached during the trial. A lower peak angle is indicative of a better postural recovery after the US. Gray lines indicate the posterior limit and the target the participant maintained prior to the perturbation. Bal: Balance, Imm: Immobile, Com: Computer-controlled, US: Unconditioned stimulus (perturbation), CS: Conditioning stimulus (auditory cue), TA: Tibialis anterior, M_AP: Ankle moment for AP movement, θ: Whole-body angle.

Participants then underwent a block of CS-only trials, where a sound cue was presented in five consecutive trials to test whether the presentation of the CS itself elicited any responses. In the majority of trials during the CS-only block, neither the TA EMG activity (4/145) nor the plantarflexor EMG inhibition (SOL 12/145, mGAS 7/145) exceeded the response thresholds. Furthermore, normalized plantarflexion moment underwent almost no change (0.01 ± 0.11, responses exceeding threshold in 7/145 trials) in the 400 ms after the CS was presented.

During the subsequent conditioning block, participants of all groups balanced in the robot while receiving at least 30 paired trials consisting of perturbations (US) preceded by a sound cue (CS) exactly 400 ms earlier. To assess whether conditioned responses emerged early in the learning process, the conditioning trials were separated into “Early”and “Late” phases, corresponding to conditioning trials 6 to 10, and the last 5 conditioning trials preceding the first catch trial, respectively. Comparison of the dependent variables (TA activity onset time, CS-to-US difference of moment, max whole-body angle; see Fig. 2 D–F) across phase (US-only, Early, Late) and group (Balance, Immobile, Computer-controlled) indicated that while responses significantly differed with phase (H₂ = 38.91, P < 0.001, H₂ = 38.23, P < 0.001, H₂ = 38.45, P < 0.001, respectively), they remained similar across groups (H₂ = 1.37, P = 0.505, H₂ = 2.32, P = 0.314, H₂ = 2.75, P = 0.253, respectively).

In the Early phase (Fig. 2 A–C, “Early”), the presence of the CS allowed participants to rapidly infer the timing of the US onset. This resulted in i) TA activity that preceded the US-onset by 47 ± 96 ms, ii) a decrease in normalized plantarflexion moment of −0.48 ± 0.46 in the CS-to-US interval, and iii) a maximal posterior whole-body angle of 0.40 ± 0.93°. By the end of training, during the last 5 trials before the first catch trial (Fig. 2A, “Late”), participants exhibited i) increased TA activity 62 ± 86 ms prior to the US (Fig. 2B), ii) a decrease in normalized plantarflexion moment of −0.68 ± 0.71 in the CS-to-US interval (Fig. 2C; bottom row), and iii) a peak posterior whole-body angle of −0.06 ± 1.00°. Dunn’s post hoc comparisons across the different phases showed that the US-only trials differed significantly from both the Early and the Late conditioning trials for all three variables (SI Appendix, Table S2 for P-values). Furthermore, the three variables did not differ significantly between the Early and Late trials, suggesting that significant conditioned responses were already formed in the Early trials. Overall, these results suggest that all groups similarly conditioned to the CS by generating preemptive ankle dorsiflexion moments through activation of the TA muscle prior to the US onset, and that the CS allows participants to reduce the extent to which they are displaced by a perturbation.

Conditioned Responses Emerge in Catch Trials of Balance and Computer-Controlled Conditions but Not in the Immobile Condition.

During catch trials, participants received a CS in the absence of a US under conditions that matched their group designation; i.e., Balance, Immobile, and Computer-controlled (Fig. 1B). In the Balance group, the majority of single catch trials resulted in muscle responses that exceeded the TA EMG threshold (32/50; Fig. 3) and/or crossed the moment derivative threshold (46/50; Fig. 4). Participants in this group reduced their normalized plantarflexion moment by −0.30 ± 0.32 between the CS and expected US, and as a result swayed in the opposite direction of the expected perturbation (i.e., anteriorly), reaching on average a peak anterior angle of −2.63 ± 0.37° (SI Appendix, Fig. S1). A Wilcoxon signed-rank test indicated that the change of the participants’ mean plantarflexion moment in the CS-to-US interval was significantly different from zero (z = 2.80, P < 0.001). In addition, the SOL and mGAS showed inhibition exceeding the response threshold in approximately half of the trials (SOL 24/50, mGAS 20/50, SI Appendix, Fig. S2 and S3). Campbell, Dakin, and Carpenter (11) reported a similar rate of SOL inhibition in 8/20 participants during CS-only trials after the conditioning block.

Fig. 3.

Normalized TA EMG responses during catch trials. (A) Group mean TA activity (black line) with the bootstrapped 95% CI of the mean (gray surface) for each of the groups’ catch trials. (B) Individual responses that constitute the group mean. Single trials that exceed the EMG threshold for 20 consecutive milliseconds are colored black, and trials that do not are colored gray. The majority of trials in the Balance (32/50) and Computer-controlled (31/45) groups produced EMG responses that exceeded the threshold, while only a limited number of trials in the Immobile group (4/50) surpassed this level.

Fig. 4.

Normalized plantarflexion moment during catch trials. (A) Mean plantarflexion moment (black line) with the bootstrapped 95% CI of the mean (gray surface) for each of the groups’ catch trials. (B) Individual responses that constitute the group mean. Single trials that exceed the moment derivative threshold are colored black, and trials that do not are colored gray. The majority of trials in the Balance (46/50) and Computer-controlled (38/45) groups produced moment responses that exceeded the threshold, while only a limited number of trials in the Immobile group (9/50) surpassed this level.

In the Immobile group, participants were made immobile during catch trials and were aware that their ankle moment could not influence their motion. The aim of the immobile catch trial was to test whether conditioned responses were absent in a context where the control of balance movement was explicitly removed. Only a few trials crossed the TA EMG (4/50) or moment derivative thresholds (9/50; see Figs. 3 and 4). Furthermore, a limited number of trials showed EMG inhibition (SOL 4/50, mGAS 7/50; SI Appendix, Fig. S2) and participants only changed their normalized plantarflexion moment between the CS and expected US by −0.02 ± 0.04, which was not significantly different from zero (z = 0.663, P < 0.278).

Finally, participants in the Computer-controlled group were exposed to a CS exactly 4 s after their ongoing balance movement was replaced with an 8-s-long prerecorded trajectory of unperturbed baseline standing. In the majority of trials, participants generated TA EMG activity (31/45) and moment changes (38/45) in response to the CS that exceeded the thresholds, despite their inability to control their movement (Figs. 3 and 4). Similarly, around half of the trials showed EMG inhibition (SOL 29/45, mGAS 17/45; SI Appendix, Fig. S2) and the normalized plantarflexion moment in the CS-to-US interval decreased significantly by −0.78 ± 1.03 (z = 2.55, P = 0.004). Notably, in comparison to the Balance group, the minimum normalized ankle moments in the Computer-controlled group reached lower values (0.12 ± 0.43 Balance vs −0.48 ± 1.02 Computer-controlled) and remained below 1 for a longer period of time after the CS (Fig. 4A), likely because the change in ankle moment had no causal effect on ongoing whole-body motion. Overall, these results indicate that incongruence between sensorimotor signals is not sufficient for the suppression of conditioned responses, as participants did inhibit conditioned responses to the CS during immobility (Immobile group), but not during segments of prerecorded movement (Computer-controlled group).

Conditioned Responses Decay with Consecutive Conditioned Stimuli, but this Decay is Reduced in the Computer-Controlled Condition.

After the catch trials, participants were exposed to consecutive CSs without a US to test the extinction of the conditioned responses. In the Balance group, most participants exceeded the TA EMG threshold (7/10; Fig. 5A) and moment derivative threshold (10/10) during the first extinction trial. By their final extinction trial, fewer participants produced TA EMG (2/10, note that the first participant only performed 4 extinction trials) and moment responses (4/10) that exceeded the response tresholds. The CS-to-US change in normalized plantarflexion moment decreased significantly from −0.16 ± 0.28 in the first extinction trial to 0.01 ± 0.08 in the last trial (z = −1.99, P = 0.024). These results suggest that conditioned postural responses show rapid extinction within roughly five CS-only trials (Fig. 5A), matching the results reported by Campbell, Dakin, and Carpenter (11).

Fig. 5.

Extinction of postural responses on a trial-by-trial basis after consecutive presentation of sound cues without perturbations. (A) Extinction block of the Balance group. The left column is the mean normalized TA EMG signal (black) and the bootstrapped 95% CI of the mean (gray surface). The right column is the mean normalized plantarflexion moment (black) with all individual participant responses (gray). The number of single trials within each group exceeding the EMG threshold or moment derivative threshold, respectively, are shown below the traces. (B) The two extinction blocks of the Immobile group. The Top row shows responses during the Immobile block, where the participants were explicitly aware that they were not in control of their movement. The Bottom row plots responses in the balancing block, where the participants actively controlled their balance while another five CSs were presented. (C) The two extinction blocks of the Computer-controlled group. The Top row plots the Computer-controlled block, where similar to the catch trials in this group control was taken away before presenting a CS. The Bottom row depicts the Balance block, where participants were balancing when CSs were presented.

The Immobile group first performed an extinction block in the Immobile condition, and then a second extinction block while balancing (Fig. 5B). Throughout all immobile extinction trials, no trials crossed the TA EMG threshold (0/50). Only four of fifty trials, all originating from participant 5, crossed the moment derivative threshold. Furthermore, the CS-to-US change in plantarflexion moment was not significantly different between the first extinction trial (−0.02 ± 0.04) and the last extinction trial (0.00 ± 0.04; z = −1.48, P = 0.080). In contrast, when these participants were then presented with CSs while balancing, almost all participants generated TA EMG (8/10) and moment (9/10) responses to the first CS and moved in the anterior direction (peak at −3.71 ± 1.37°; see SI Appendix, Fig. S3). The CS-to-US change in normalized plantarflexion moment was −0.62 ± 0.68 in the first trial, and significantly decreased over subsequent trials (see Fig. 5B), such that the CS-to-US moment change in the last trial was 0.06 ± 0.08 (z = −2.80, P < 0.001). In the last trial, no participant generated a TA EMG response and only one participant exceeded the moment derivative threshold (1/10). These results suggest that a renewal of conditioned responses occurred when returning to the Balance condition (e.g., the conditioning context) following the absence of these responses in the Immobile condition.

The extinction block for the Computer-controlled group was performed with active balance control taken away without informing the participant on five separate trials (Fig. 5C). In between each period of computer-control, participants balanced normally while maintaining their target posture using the visual feedback for approximately 15 to 20 s. During the first trial, most participants generated TA muscle activity (6/9) and moment (7/9) responses, which persisted for 2 to 3 trials but eventually decreased such that EMG and moment responses were only observed in four participants in the final trial. The CS-to-US change in plantarflexion was −0.27 ± 0.37 in the first and −0.43 ± 0.90 in the last trial, indicating that the average response magnitude to the CS increased over the course of extinction. This is notably different from the Balance group, where the change became less negative over the subsequent extinction trials, indicating a reduced response to the CS over its repeated presentation. Despite this change in the average response, no significant effect was found (z = −0.30, P = 0.410) with a one-sided (first < last) Wilcoxon signed-rank test. When these participants then performed a second extinction block while actively balancing, changes in TA muscle activity (3/8) and plantarflexion moment (6/8) persisted for some of the participants during the first trial. The last trial evoked no EMG responses and only one plantarflexion moment response. The CS-to-US change in normalized plantarflexion moment varied from −0.33 ± 0.92 in the first trial to −0.09 ± 0.16 in the last trial, though this difference was not significant (z = 0.280, P = 0.629). Overall, these results suggest that the rate of extinction of conditioned responses is diminished when participants are in the Computer-controlled context when compared to the Balance group. When returning to the Balance context, the conditioned responses are less apparent but do decrease in their frequency and magnitude following repeated exposure to the CS.

Participants Show Limited Perception in Changes of Control during Computer-Controlled States.

Following all conditioning and extinction trials, participants from the Computer-controlled group then performed an additional control experiment to estimate how aware participants were of any change in control leading up to the time of the CS (i.e., 4 s after transition). Participants (N = 8, one participant was excluded due to a data collection error) underwent 14 transitions of 8 s long during a single trial lasting 438 ± 104 s and were tasked to press a handheld button whenever they detected any change in balance control. On average, participants detected a change in control during the Computer-controlled state in only 49.1 ± 30.4% of the trials at a time before when the US would have occurred in the catch trials, i.e., 4 s after the onset of Computer-control. The average latency of these successful detections was 2.04 ± 0.53 s. Both the response performance and the latencies were similar to previous reports by Luu et al. (29), where a 4-s transition was used. When considering the whole duration of the 8-s Computer-controlled segment, the mean detection rate was 72.3 ± 33.6% and occurred on average after 3.35 ± 0.78 s. Notably, participants also made on average 3.3 ± 2.2 false detections, the majority of which (2.3 ± 1.8) occurred within 4 s immediately following the end of computer control. One participant (Participant 9) never correctly detected a transition despite reporting four false positives. An overview of all individual participants’ perception results can be found in SI Appendix, Table S3. Overall, these results indicate that although participants may have been aware that control of the platform had been modified during half of the Computer-controlled conditions, they still generated postural responses to the CS during the majority of catch trials and numerous extinction trials.

Discussion

In the current study, we used a robotic balance simulator to examine whether associatively learned postural responses are modulated by a conscious perception of balancing the body or through an unconscious integration of sensory and motor signals. The results first confirmed that when a neutral sound cue was repeatedly paired with a postural disturbance, the postural responses were conditioned to occur when the auditory stimulus is presented in isolation, aligning with previous experiments (9–12). In “Early” conditioning trials (trials 6 to 10), significant conditioned responses were already formed, indicating that future conditioning protocols could be performed with fewer conditioning trials. Our results then showed that when participants were immobilized and consciously aware that they no longer needed to balance, the conditioned responses to the sound cue were nearly absent. Nevertheless, these participants showed renewed conditioned responses to the auditory stimulus when they returned to the Balancing context. In contrast, when the control of balance was subtly withdrawn using Computer-control conditions, participants repeatedly generated conditioned postural responses even though they had no causal relevance to ongoing body movement. Furthermore, during an extinction block under Computer-control conditions, participants generated repeated postural responses even in the final extinction trial and when returning to the normal Balancing context. Overall, our results indicate that the state of control can modulate conditioned postural responses. However, in contrast to the vestibular contributions to standing, an unconscious integration of sensory and motor signals (29) was not sufficient to suppress conditioned responses when control over balance movement was removed, suggesting that these two types of responses do not share a single mechanism for inferring the control state.

An Immobile State is Sufficient Context for the Suppression of Conditioned Responses.

When participants in the Immobile group performed catch trials, associatively learned postural responses were absent in most trials for both TA EMG (4/50) and moment responses (9/50, Figs. 3 and 4). Furthermore, almost no conditioned responses were produced (0/50 TA EMG and 4/50 moment) throughout the Immobile extinction block, where five CSs were presented consecutively. During these trials, participants were both physically immobile and explicitly aware that the applied ankle moments did not influence the control of their movement. These results are similar to vestibular-evoked balance responses, which are suppressed in (equivalent) conditions where both the vestibular signals (28, 29) and/or the activated muscle groups (30, 31) cannot contribute to balance. Interestingly, when participants of the Immobile group were returned to the Balance context and experienced another CS, a renewal of conditioned responses occurred. Most participants (8/10 EMG, 9/10 moment) showed a conditioned response during the first trial of the subsequent Balance extinction block (Fig. 5B, second row), the emergence of which quickly diminished over subsequent CSs similar to the extinction process in the Balance group. Renewal driven by revisiting the training context (e.g., condition A) after experiencing another context (e.g., condition B), also called ABA renewal (16), suggests that participants can use the balance control context (i.e., presence or lack of causal control over movement) as a state for generating or suppressing conditioned responses. Campbell, Dakin, and Carpenter (11) demonstrated that, without washout, the memory of conditioned responses persists for at least 15 min, suggesting that the elapsed time during the first Immobile (and Computer-controlled) extinction block is unlikely to have affected the magnitude of the renewed response. ABA renewal is thought to occur through a state classification process, where different states are maintained and associated with learned behavior (32). By perceiving changes in the control state, the correct memory can be expressed, and different learned behaviors can be maintained simultaneously. Such a classification could be responsible for attenuating associatively learned postural responses if participants were made aware of their Immobile state. However, from the results of the Immobile condition alone, it is not clear whether the conscious experience of an immobile context, or an unconscious integration of congruent sensory and motor signals, modulated the conditioned responses.

The Removal of Control over Bodily Movement is not Sufficient to Suppress Associatively Learned Postural Responses.

The Computer-controlled condition decoupled the causal relationship between actions (motor commands) and their consequence (body movement), introducing an incongruency between these signals while still experiencing balance-like whole-body motion. Luu et al. (29) used a similar condition in conjunction with a continuous stochastic electrical vestibular stimulus to demonstrate that incongruence between sensory and motor signals is sufficient to rapidly (>200 ms) and unconsciously suppress ongoing postural responses to vestibular error signals. The authors attributed this rapid suppression to an unconscious integration of congruent sensory and motor signals by the balance controller. In this interpretation, a comparison is made between the sensory information and the expected sensory consequences produced by the balancing motor command (i.e., through a forward model, Fig. 1) to check whether they are congruent (33–35). However, when participants in our Computer-controlled group were transitioned, without being made aware, into this incongruent context, conditioned responses were still formed (31/45 EMG and 38/45 moment). The presence of these responses implies that an unconscious perception of sensorimotor incongruence caused by the removal of control over bodily movement, which suppresses continuous vestibular-evoked responses (29), does not suppress discrete conditioned postural responses to a sound cue. Since associative learning and automatic postural responses both seem to be modulated mostly unconsciously by contextual relevance (29, 36), we hypothesized that their modulation by control context would be similar. Instead, the persistence of conditioned responses in the Computer-controlled catch trials suggests that different mechanisms govern vestibular-evoked and associatively learned postural responses. This further implies that the inference required for their modulation cannot be unified within a single simplified representation such as the comparator, as this model cannot account for the differences in expression between the responses.

To test whether participants were able to consciously perceive the transition, an additional perception experiment was performed with participants of the Computer-controlled group (N = 8). Participants were asked to press and hold a button for as long as they perceived any alteration in control and were not informed about the nature of the alternation. They perceived half of the transitions into computer control within 4 s when the CS would have occurred in catch trials at an average latency of 2,044 ms. Luu et al. (29) reported similar detection rates (i.e., 50%) at nearly equivalent latencies of 2,247 ms. Despite the low detection rates and slow perception of changes in control, participants in the Luu et al. (29) experiment quickly modulated vestibular responses ~175 ms after the transition into the Computer-controlled state. The similar detection performance in our study implies that participants may have consciously perceived a change in control during Computer-controlled trials prior to the CS. Regardless, this possible perception of changes in control together with the introduced incongruence between motor and sensory signals was not sufficient to suppress postural responses to the CS throughout the catch and extinction trials in the Computer-controlled condition (Figs. 4 and 5C). Future work could pair a conditioning paradigm with vestibular stimulation to assess how both mechanisms function simultaneously.

One possible explanation for the difference between the modulation of continuous vestibular control and associatively learned postural responses is that they rely on different error signals and neural structures. The sensory prediction error (i.e., the deviation between the actual and predicted sensory state) is thought to drive the continuous control and implicit recalibration of movement (18, 37). Through processes of learning, this sensory prediction error is minimized through networks in the cerebellum in order to mediate changes in vestibular control (38–41). In contrast, the reward prediction error encodes the expectation of positive or negative outcomes. This error signal drives classical conditioning and is produced primarily by the striatum and amygdala, with additional contributions of the cerebellum (42). The Computer-controlled condition continuously evokes a sensory prediction error as the balance-like movement of the system is not causally related to the participant’s current input. The reward prediction error, on the other hand, is driven by the expected outcome of a discrete event and may only be updated when an expected perturbation is, or is not, delivered. The emergence of conditioning is modulated by the current state or context (13, 16), so to correctly update the expected outcome, the participant would have to continuously and correctly assess their degree of control (e.g., Balance or Computer-controlled). As the neural structures responsible for minimizing these errors are distinct and consist of parallel circuits (43), perhaps separate assessments of the control state are made for each system (Context-Dependent Learning under Sensorimotor Incongruence).

Lastly, it is possible that participants did not generalize the incongruency experienced in continuous motor control to the inability to control movement during a learned perturbation. The electrical vestibular stimulus used by Luu et al. (29) continuously perturbed balance control, and the accompanying vestibular-evoked responses were rapidly suppressed when participants transitioned to computer control. While a continuous assessment of context likely gates conditioned responses, participants in the current study may need to learn about how perturbations, and their own actions, affect their balance in the current context to respond appropriately. In the Immobile catch trials, it was demonstrated that the lack of control can be a sufficient cue to suppress conditioned responses. However, participants in this condition had explicit knowledge of their immobility and could use this information to change their expectations for the perturbation period. In the Computer-controlled condition, participants may have falsely perceived themselves to still be in control over their movement, the implications of which are discussed further below.

Misattribution of Agency to Self-Motion.

The persistence of conditioned responses while in a Computer-controlled state could also be interpreted through the concept of agency. The sense of agency is our internal assessment of whether we influence an outcome, which is often argued to be constructed by comparing our predicted outcome of a desired motor behavior to the perceived one, as similarly argued in the comparator model (7, 44). Humans, however, are susceptible to illusions of control, in which we have an exaggerated belief of the extent to which we influence the world around us (45). In such an illusion, we see ourselves as the actor or initiator of events around us while we do not have any causal control over them. This may have impaired the capacity to generalize the absence of continuous control to the irrelevance of responding to a discrete perturbation. An optimistic sense of agency might explain why participants generated postural responses during the Computer-controlled catch trials and continued to do so in the extinction block, but it does not explain why Luu et al. (29) observed a rapid, unconscious suppression of the vestibular-evoked responses in this same condition.

A possible explanation may be found in the two-factor account of agency by Synofzik, Vosgerau, and Newen (46), where a subconscious, implicit sense of agency and a conscious, explicit judgment of agency were proposed. The implicit sense of agency is formed by low-level integration of sensorimotor signals such as the efference copy (47, 48), for example with a comparator [(33, 34); see Fig. 1]. Following the low-level comparisons, an explicit judgment of agency can be formed based on past sensorimotor signals as well as contextual cues (46, 49, 50). The explicit judgment of agency is often tested by asking participants directly whether outcomes result from their intentional actions (7). On the other hand, the implicit sense of agency is subconscious and usually quantified using a proxy such as intentional binding, which is a measurable reduction in the perceived elapsed time between a voluntarily initiated action and its outcome (7, 51). Perhaps associatively learned and vestibular-evoked postural responses are both modulated by agency, but each driven by its explicit and implicit components, respectively. Vestibular contributions to continuous postural control are known to be automatically modulated by their relevance (6, 52), long before conscious perception of changes in control (29). In contrast, although classical conditioning also occurs mostly implicitly (36), it is strongly modulated by an explicit, presumably conscious, judgment of the current context or state (13, 53), similar to the explicit judgment of agency proposed by Synofzik, Vosgerau, and Newen (46). In our study, neither the explicit nor implicit formation of agency was specifically tested. Future work may use intentional binding (51) or neural correlates of agency (54, 55) during standing balance experiments to assess whether subconscious perceptions of sensorimotor incongruency reduce the sense of agency and whether this reduction of agency predicts the suppression of postural responses.

Context-Dependent Learning under Sensorimotor Incongruence.

The extinction of conditioned responses is thought to represent additional learning processes, where the CS is learned to have no relevant relationship to the (lack of a) perturbation in the current context (13, 16). Processes of learning observed during extinction and conditioning may also be described by contextual inference models of learning (17, 32, 56). In this interpretation, motor memories are updated and expressed based on the current state, which is inferred from contextual cues in the environment. Crucially, the updating of existing memories is proportional to how likely each memory is inferred to be responsible for the currently perceived contextual cues (17). In our study, the Immobile condition provided explicit evidence for the removal of balance control, but in the Computer-controlled context, the balance controller may lack sufficient evidence to decisively conclude that there is no control over movement (i.e., that the state is Computer-control). Although there is an incongruence between the motor actions and the experienced movement, the statistics of the movement continued to resemble natural human balance. The balance controller may still have attributed some responsibility to the Balance context (where conditioning occurred), evident by the continued expression of conditioned responses to the CS. However, because the conditioned responses did not affect movement during the Computer-controlled extinction trials, a low responsibility may be attributed to the Balance context and the updating of memories (e.g., extinction) would be reduced (see Fig. 6A).

Fig. 6.

Contextual inference as a framework to explain the differential expression of vestibular-evoked and associatively learned postural responses. (A) Abstracted example of how predicted probabilities could drive the formation of postural responses (PR) in different control contexts. The predicted probabilities (colored bars) encode our belief of the environment and are computed from the prior knowledge and current sensory information through Bayesian inference. Memories are expressed (i.e., used to generate an action) based on the predicted probability of their state, for example through a weighted average. After memory expression, the inferred responsibilities of each state for the observed outcome are used to update the prior knowledge (not depicted; see work by Heald, Lengyel, and Wolpert (17) for the mathematical description). In the Balance and Immobile states, the balance controller should have sufficient information to quickly determine the correct state. In the Computer-controlled state the participant’s assessment of their control over movement may be indecisive, which would lead to slower updating of motor memories (such as extinction) in this condition. (B) Alternate hypotheses for contextual inference in postural control for vestibular-evoked and associatively learned responses. Hypothesis 0 is the traditional view of contextual inference, where a central inference operation determines how learned behavior is expressed as shown in panel A. Hypothesis 1 resolves the conflict between continuous vestibular and classically conditioned balance responses by using the same predicted probabilities but with different functions for memory expression. Hypothesis 2 instead poses that both systems could use contextual inference, but in parallel with separate predicted probabilities.

This hypothesis could explain why conditioned responses persist in conditions where the balance controller cannot decisively conclude the state of the system, and why extinction occurred rapidly when these participants returned to the Balance condition in the second extinction block. However, a single (or central) inference of a control state (Hypothesis 0 in Fig. 6B) does not explain why conditioned responses are generated in the Computer-control condition while vestibular-evoked responses are suppressed. Either these responses use a different method to derive an action from centrally estimated context probabilities (e.g., weighted average or winner-take-all; Hypothesis 1 in Fig. 6B), or multiple contextual inference integrations are performed separately and in parallel for low- and high-level responses (Hypothesis 2 in Fig. 6B). While this last hypothesis challenges the view that we maintain a single, coherent generative model to assess our control over our environment and ourself, it is remarkably similar to how Synofzik, Vosgerau, and Newen (46) proposed agency to be organized. In that interpretation different neural structures could form separate low and high-level inferences of agency over bodily movement in parallel, which may prioritize different error signals (i.e., reward prediction error or sensory prediction error). Continuous movement control and conditioned responses partially rely on different neural systems, and as a result might not share central assessments of the environment, as posed in Hypothesis 2. We caution, however, that careful experimental design will be required to disentangle these hypotheses because the predicted probabilities are not directly observable. Further research is required to examine how responsibilities of contexts are expressed under sensorimotor incongruence and whether contextual inference theories can describe continuous control and associatively learned responses in a unified framework.

Conclusion.

In summary, the present study demonstrates the interaction between associative learning, postural responses, and the contextual framework in which they operate. Our results show that postural responses can be conditioned to auditory cues and modulated by control context. In contrast to vestibular-evoked responses, incongruence between sensory and motor signals introduced by the subtle interruption of human balance control was not sufficient to suppress associatively learned postural responses. This suggests that different mechanisms govern these responses and that their modulation cannot be described using a single comparator-like process. Theories of agency and contextual inference show promise as a unifying framework for associatively learned postural responses and vestibular contributions to balance.

Materials and Methods

Participants.

A total of 32 healthy adult participants aged 18 to 50 with no history of balance, neurological, or musculoskeletal disorders were recruited for this study. Data from two participants were excluded due to high degrees of cocontraction observed during the experiments. All other participants performed either a Balance (N = 10), Immobile (N = 10), or Computer-controlled (N = 10) experimental protocol; the exact experimental differences between these groups are further described in the Protocol subsection and SI Appendix. Data from one participant in the Computer-controlled group were excluded due to the premature delivery of a catch trial during the conditioning period. Characteristics of the included participants are summarized in SI Appendix, Table S1. All experiments were conducted following the declaration of Helsinki and under the approval of the Erasmus Medical Center Ethics Committee. All participants gave their written informed consent before the experiment was conducted.

Experimental Setup.

A custom-designed robotic balance simulator (Fig. 1A) was used to emulate natural human standing balance. During the Balance condition, participants control their motion through the ground reaction forces and moments exerted on the force plate. An inverted pendulum simulation computes the whole-body motion using participant parameters of their body mechanics, and motors actuate the backboard to follow this movement. Alternatively, the system could be made Immobile, or switch to a Computer-controlled state where motion is replaced with previously recorded balancing movement. More details on the equipment and computations are provided in SI Appendix.

Stimuli.

A delayed conditioning protocol was performed where the US was a moment perturbation (M_pert) added inside the inverted pendulum simulation on top of the participant’s self-generated moment to disrupt the ongoing control of upright balance (Fig. 1A). This moment accelerated the participants in the posterior direction for a duration of 100 ms and required an increase in dorsiflexion moment to maintain balance. The perturbation moment was normalized to the anthropometrics of each participant to achieve an angular acceleration (α) of 80°/s² in the posterior direction irrespective of their length or weight. The acceleration was multiplied by the mass moment of inertia (I) to determine the moment required for this acceleration.

The conditioning stimulus (CS) was delivered as a sound tone of 1,000 Hz (<70 dB) that started playing 400 ms before the onset of the moment perturbation (i.e., the US) and lasted a duration of 1,000 ms. Participants wore noise-canceling headphones (WH-1000XM3 Noise-Canceling Headphones, Sony Japan) to deliver the tone and eliminate other auditory cues of motion provided by the motors. Participants were not given any information regarding the timing or relation between the tone and perturbation.

When perturbed, participants generate a postural response, which in the context of conditioning is the unconditioned response (UR) and is initially unrelated to the sound cue (CS). A postural response to a posterior perturbation involves contraction of dorsiflexor muscles such as the TA, and generally a suppression of the plantarflexor muscles such as the SOL and mGAS (11). Electromyography (EMG) measurements were therefore performed for the right TA, SOL, and mGAS muscles. Electrodes were placed according to anatomical landmarks described by Perotto (57). EMG signals were amplified with a gain of 1000× (NL844 preamplifier; Digitimer Ltd., UK), low-pass filtered at 500 Hz, amplified a second time with a gain of 2× (Neurolog NL900D Amplifier, NL820 Isolator, NL135 filter; Digitimer Ltd., UK), and recorded at 1,000 Hz for all participants on the real-time controller.

Protocol.

Representative single-participant data over the experimental protocol are shown in Fig. 1 B and C, and a detailed description is given in SI Appendix. After a period of familiarization, participants used a visual target to maintain a posture at −2° (anterior) and performed 5 CS-only trials and 12 US-only trials, where all participants successfully maintained balance during at least 5 of the US-only trials. Next, all participants completed a conditioning block of 50 trials. In the initial 30 trials, a CS was delivered 400 ms before a US, allowing the participants to learn the relationship between the sound cue and the perturbation. During the last 20 trials, 5 catch trials were pseudorandomly interspersed between the other 15 paired trials such that no catch trials occurred consecutively. The condition of the catch trial corresponded to the group designation (e.g., Balance, Immobile, or Computer-controlled), meaning that a CS was delivered while the participant was actively balancing, immobilized (and made aware of their immobility), or experienced a computer-control transition, respectively. In the Computer-controlled condition, a CS was delivered 4 s after movement was replaced with 8 s of previously recorded motion, such that the participant had no causal influence over their motion during this period. During a subsequent extinction block, 5 CS-only trials were performed, first in the group condition, and then another 5 while balancing for the Immobile and Computer-controlled group to test for the re-emergence of conditioned responses.

Data Analysis and Statistics.

All data analysis was performed using MATLAB (2021b, MathWorks, MA, USA). Conditioned responses were assessed using the onset time of TA activation, the change in ankle plantarflexion moments in the CS-to-US interval, and the inhibition of SOL and mGAS muscle activity. All group data in text, tables, and figures are presented as mean ± SD unless mentioned otherwise. Statistical analyses were performed using JASP (58), and P-values are reported with a significance level of 0.05 for all analyses. All details regarding data processing, analysis, and statistical tests are reported in SI Appendix.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The data that support the findings of this study and the software used to generate Figs. 2–5 and SI Appendix, Figs. S1–S4 are provided in “Data and code for ‘Different mechanisms of contextual inference govern associatively learned and sensory-evoked postural responses’”, https://doi.org/10.34894/ITHTAI, DataverseNL (59).