Sparse Identification of Motor Learning Using Proxy Process Models

Enhanced neurorehabilitation using robotic and virtual-reality technologies requires a computational framework that can readily assess the time course of motor learning in order to recommend optimal training conditions. Error-feedback plays an important role in the acquisition of motor skills for goal-directed movements by facilitating the learning of internal models. In this study, we investigated changes in movement errors during sparse and intermittent "catch" (no-vision) trials, which served as a "proxy" of the underlying process of internal model formations. We trained 15 healthy subjects to reach for visual targets under eight distinct visuomotor distortions, and we removed visual feedback (no vision) intermittently. We tested their learning data from novision trials against our so-called proxy process models, which assumed linear, affine, and second-order model structures. In order to handle sparse (no-vision) observations, we allowed the proxy process models to either update trial-to-trial, predicting across voids of sparse samples or update sample-to-sample, disregarding the trial gaps. We exhaustively cross-validated our models across subjects and across learning tasks. The results revealed that the second-order model with trial-to-trial update best predicted the proxy process of visuomotor learning.