Implicit Adaptation via Visual Error Clamp J. Ryan Morehead, Jordan A. Taylor, Darius Parvin, Elizabeth Marrone, Richard B. Ivry Sensorimotor adaptation is thought to be driven by an error signal based on the difference between the actual and predicted outcome of a motor command. Experiments that study this process often confound variables such as extrinsic reward and task goals with the error that adaptation seeks to minimize. Therefore changes in behavior during these tasks cannot always be attributed to adaptation alone. Previous work has attempting to isolate adaptation from these confounds by pitting task success against adaptation (Mazzoni & Krakauer, 2006; Taylor & Ivry, 2011). We extend this line of research by using a visual error clamp to induce adaptation. Error clamps are typically used to assay the current state of learning without inducing errors related to task goals. Here we bias the feedback of the clamp to introduce a sensory prediction error. This approach eliminates the influence of task error, reward and cognitive factors by removing the contingency between behavior and task success, allowing us to assess adaptation in relative isolation. Experiment 1 combined the use of a cognitive aiming strategy to compensate for a rotation with clamped visual feedback centered on the goal target. Participants (n=10/group) were randomly assigned to three groups: a veridical rotated feedback group (NormRot), a naïve error clamp group (NaïveClamp), and a group that saw clamped feedback and was informed about the error clamp (AwareClamp). All groups were instructed to aim at a landmark 45° clockwise from the goal target (Fig. 1A). The AwareClamp group was further instructed to ignore the cursor and focus on executing the aiming strategy. All groups showed very similar drift rates, asymptotes and aftereffects in response to these different task conditions (Fig. 1B). Thus, removal of the contingency between motor behavior and task error did not lead to differences in peak drift. Moreover, implicit adaptation appeared to saturate independently of task error, contrary to the two-process model of Taylor and Ivry (2011). In Experiment 2, we investigated adaptation with error clamps by comparing learning and generalization of a 22.5° rotation group (n=20) with that of an error clamp group (n=20) whose heading was offset from the target by 22.5° (Fig. 2A). The error clamp group was told to move to the target and that they should ignore the cursor since it was not under their control. Performance during the perturbation block was different between the two groups, with the error clamp group showing reduced change in hand angle compared to the rotation group (Fig. 2B). Despite this discrepancy in performance during learning, the aftereffect and generalization functions were remarkably similar for the two groups (Fig. 2C). Thus, measures that isolate the effects of adaptation appear to be similar when the “error” is signaled by an error clamp in comparison to standard, rotated feedback. Experiment 3 measured adaptation to error clamp offsets of increasing magnitudes: 7.5°, 15°, 30°, 45°, 60°, 95°, 135° and 175° (n=10/group). With a traditional visuomotor rotation, the magnitude of learning and sensorimotor aftereffect is positively correlated with the size of the imposed error. Here we observed that the rate and amount of change in behavior during the error clamp block, as well as the size of the aftereffect did not linearly scale with the error size. Instead, the groups with clamps between 7.5° to 60° displayed changes in reach direction of similar magnitude. It is particularly notable that the 7.5° group adjusts to nearly twice the perturbation error size. At 95°, the magnitude of the change in reach direction was reduced, and negligible changes were observed at 135° and 175°. We fit Experiment 3 individual hand angle data with a linear state-space model, SSM, that ignores hand position in the calculation of error when the clamp is present (Fig. 3, Eq. 1). We fixed the forgetting rate, A, to a value of .01, leaving only the learning rate, B, as a free parameter. If the median B parameter estimate is used to predict aftereffects, SSM underpredicts learning at small error sizes, and vastly overpredicts learning for large errors (Fig. 3D, red). However, individual best-fit parameters had very different B values that declined with error size (Fig. 3C). We found that these parameters were well characterized by an exponential decay function, 𝐵 =. 029−𝑒/.073, with 𝑒 denoting error size. This led us to consider a second model, EDM, where B is replaced with the above exponential error discounting function (Fig. 3, Eq. 3). When EDM is used to simulate learning (Fig. 3D, blue), it underpredicts learning in the 45° and 65° groups. We therefore we also considered a third model, QAM, that makes quantal adjustments of a fixed size to oppose errors (Fig. 3, Eq. 2), again with only B as a free parameter. The predictions of QAM, using the median value of the B parameter, .125, are similar to the real aftereffects observed in groups below 90° (Fig. 3D, green). If the QAM model included discounting of errors greater than 90°, it may accurately predict learning at all sizes of error clamp rotations in this task. We have shown that visuomotor adaptation can be induced by an error clamp if the heading is offset from the reach target. Excepting one group in Exp. 1, all participants exposed to clamped feedback were fully informed that the cursor did not correspond to their hand, was not under their control, and that the reaching task did not involve the cursor during the clamp block. Despite the irrelevance of the visual feedback to the participant’s task and full explicit knowledge of the relationship between behavior and feedback from the clamp, we reliably observed adaptation. While there are some differences between adaptation from visual error clamps and rotations, many features that are considered hallmarks of adaptation are present, including fully implicit learning, sign-dependent corrections, Gaussian generalization and a robust sensorimotor aftereffect. Depending on the circumstances, the adaptation reported here may partially or fully constitute visuomotor adaptation measured with other paradigms. This technique may help illuminate features of adaptation that are sometimes obscured by other learning processes in conventional adaptation tasks.

Experiment 1 – Aiming Strategy + Error Clamp A

Strategy + Rotation - Early

Strategy + Rotation - Late

Strategy + Clamp - Early

Strategy + Clamp - Late

A. Top – When instructed to use a strategy, participants initially reach directly to the blue target. However, with more trials, implicit adaptation causes the hand to shift opposite the direction of the rotation.

B

Bottom – In the error clamp condition, the cursor moved to the target. Here, too, the hand shifts opposite the rotation across trials.

Hand Angle for all Groups

B. Hand angle for all groups rises and asymptotes at the same level, despite large differences in visual feedback and variables like task success. The rotation group begins to reduce task performance error after many trials, but had not yet returned to good performance by the end of the block.

Experiment 2 – 22.5° Rotation/Error Clamp Generalization Train Rotation

A

or

Train Clamp

C

Test Generalization

Generalization Following Perturbation

A. Groups were either exposed to a 22.5° rotation or 22.5° error clamp. Generalization was tested in following blocks.

B

Hand Angle at Trained Target

B. For both groups, the sensory discrepancy led to a change in hand angle. While the change was greater for the rotation group during learning, the aftereffect and washout measures were similar for the two groups, suggesting similar adaptation.

C. Generalization was tested by having participants reach to 24 target locations around 360° during the aftereffect phase. No feedback was provided on these trials.

1 x(n + 1) = (1 − A) * x(n) − B * e(n) 2 x(n + 1) = (1 − A) * x(n) − B * sign(e(n)) 3 x(n + 1) = (1 − A) * x(n) − e(n) * .029− abs (e( n))*.073 Exponential Fit for SSM B Values

C

Experiment 3 – Insensitivity to Error Clamp Size A

7.5°/30°/60°/135° Groups

B

Hand Angle at End of Clamp Block C. Individual Model 1 B values with exponential fit.

D

Simulated Aftereffect

15°/45°/95°/175° Groups

A. Hand angle across trials for all groups. Behavior was very similar when the error clamp was below 90°. The 135° and 175° groups showed negligible evidence of adaptation.

B. Average hand angle for last 8 trials of the error clamp block. Lines indicate group mean. Dots denote individuals.

D. Simulated aftereffect for three models superimposed over real data.