Dual adaptation supports a parallel architecture of motor memory

Abstract

Although our understanding of the mechanisms underlying motor adaptation has greatly benefited from previous computational models, the architecture of motor memory is still uncertain. On one hand, two-state models that contain both a fast-learning-fast-forgetting process and a slow-learning-slow-forgetting process explain a wide range of data on motor adaptation, but cannot differentiate whether the fast and slow processes are arranged serially or in parallel and cannot account for learning multiple tasks simultaneously. On the other hand, multiple parallel-state models learn multiple tasks simultaneously but cannot account for a number of motor adaptation data. Here, we investigated the architecture of human motor memory by systematically testing possible architectures via a combination of simulations and a dual visuomotor adaptation experimental paradigm. We found that only one parsimonious model can account for both previous motor adaptation data and our dual-task adaptation data: a fast process that contains a single state is arranged in parallel with a slow process that contains multiple states switched via contextual cues. Our result suggests that during motor adaptation, fast and slow processes are updated simultaneously from the same motor learning errors.

Figure 1.

Ten possible motor adaptation models that address the following three questions: (1) Are there slow and fast timescales? (2) Is there one state or more than one for each timescale? (3) Are the fast and slow processes arranged serially or in parallel?

Figure 2.

The two possible architectures of the 1-fast n-slow model: parallel (A) and serial (B). e is a motor error, c is a contextual cue, and x is a motor output. Multiple boxes in the slow process represent internal states switched by the contextual cue input.

Figure 3.

Simulation of spontaneous recovery for all 10 models considered: 1-state model, parallel n-state model, serial and parallel 1-fast 1-slow models, serial and parallel n-fast 1-slow models, serial and parallel n-fast n-slow models, and serial and parallel 1-fast n-slow models. A, Schedule of the error-clamping paradigm used to induce spontaneous recovery, which consists of 180 trials for adaptation to one stimulus (1), 20 trials for adaptation to the opposite stimulus (−1, deadaptation), and 50 error-clamping trials, during which errors are clamped to zero. B, Model predictions of adaptation performance for all models. The parallel and serial models are superimposed in all panels except for the 1-state and parallel n-state models. Check marks or crosses are used to show which models account or do not account for the data, respectively. All models except the 1-state and parallel n-state model can reproduce the characteristic of spontaneous recovery: the output for the first error-clamping trial starts near the baseline (zero), increases trial by trial, and decays slowly.

Figure 4.

Simulation of anterograde interference for the eight remaining candidate models: serial and parallel n-fast n-slow models, serial and parallel 1-fast 1-slow models, serial and parallel n-fast 1-slow models, and serial and parallel 1-fast n-slow models. The parallel and serial models are superimposed in all panels. A, Schedule of the A–B–A paradigm used to induce anterograde interference, which consists of 100 trials for adaptation to one stimulus (1), 100 trials for adaptation to the opposite stimulus (−1, deadaptation), and 100 trials for readaptation to 1. B, Model predictions of adaptation performance in the A–B–A paradigm. C, Comparisons of initial errors in each session. As in Figure 3, check marks or crosses are used to show which models account or do not account for the data, respectively. All models except the serial and parallel n-fast n-slow models can reproduce the characteristic of anterograde interference: the initial errors of both deadaptation and readaptation are greater than the initial error of the first adaptation.

Figure 5.

Simulations of two dual-adaptation experiments for the remaining six models considered: serial and parallel 1-fast 1-slow models, serial and parallel n-fast 1-slow models, and serial and parallel 1-fast n-slow models. The parallel and serial models are superimposed in all panels. A, B, Intermittent alternation of two tasks (A) and random alternation between two tasks (B). For each model, the same parameters are used in A and B. Only the serial and parallel 1-fast n-slow models can reproduce dual adaptations in both intermitted and random conditions. Note that the parallel and serial 1-fast n-slow models behave identically in A but differently in B: the parallel 1-fast n-slow model shows faster adaptation rates than the serial 1-fast n-slow model in random dual adaptation.

Figure 6.

Average performance data across subjects during learning (black dots) and predictions of serial (red stars) and parallel (blue crosses) 1-fast n-slow models. Red- and blue-shaded areas show the ranges of ±SEs of the serial and parallel model predictions, respectively. Model parameter estimation was performed using the data in the massed schedules. The models were then used to predict the data in the random schedules. Both models fitted subject data well during the A–B–A massed schedule. However, during the random schedule, the parallel model predicted the data better than the serial model, as shown by smaller MSE between the data and the parallel model predictions compared with the MSE between the data and the serial model predictions (p = 0.0001). The estimated parameters (with 95% confidence intervals) are as follows: for the parallel model, A_f = 0.8251 (0.6338–0.9767), A_s = 0.9901 (0.9876–0.9986), B_f = 0.3096 (0.1585–0.5118), and B_s = 0.2147 (0.0582–0.2729); for the serial model, A_f = 0.8749 (0.7082–0.9643), A_s = 0.9917 (0.9894–0.9984), B_f = 0.4831 (0.2923–0.6655), and B_s = 0.0456 (0.0077–0.1178).

Figure 7.

Simulation of the washout paradigm with the two-state model, the varying-parameter model, and the parallel 1-fast n-slow model. A, Schedule of the washout paradigm, which consists of 10 null trials, 80 learning trials, 40 washout trials, and 30 relearning trials. In learning and relearning trials, there was 45° of disturbance, and in null and washout trials, no disturbance. B, Model predictions of errors in the relearning-after-washout paradigm. We used the same parameters as Zarahn et al. (2008) for the varying-parameter model and two-state model and chose parameters for the parallel 1-fast 1-slow model to reproduce the results of the varying-parameter model. C, Comparison of model predictions during the initial learning and relearning. First 30 trials in learning and relearning trials are superimposed. The error traces of the two-state model in learning and relearning trials are identical and cannot reproduce savings after washout trials. In contrast, the parallel 1-fast n-slow model can predict savings after washout trials with fewer parameters than the varying-parameter model.