Fig. 5 |. Uniform-shift progression correlates with motor memory retention.

a, c, In the force-predictive TDR subspace, preparatory states for the trained target gradually reverted to the before-learning state during washout. One example session visualized in a and four sessions quantified in c. b, d, Preparatory washout states remained separated from before-learning states along the uniform-shift axis. One example session visualized in b and four sessions quantified in d. e, Preparatory washout states shifted uniformly along a dimension orthogonal to the uniform-shift learning axis. f, Cross-validated accuracy for classifying before-learning, late-learning and late-washout trials from single-trial preparatory states. g, Behavioural relearning progressed faster than initial learning. h, Neural trajectories approached late-learning trajectories faster during relearning than initial learning. MSE, mean squared error. In g, h, shaded area shows s.e.m. across sessions. i, Centroids of late-washout and relearning states from individual sessions projected onto normalized uniform-shift learning and washout axes. Inset shows means across seven sessions. A significant shift along the washout axis and a small but significant regression along the learning axis accompanied washout, and reversed during relearning. j, Minimum trial count required for behavioural adaptation during relearning was significantly smaller than for initial learning. Normalized progression along the uniform-shift learning axis was significantly smaller through relearning. k, The magnitude of preparatory uniform shift correlated strongly with the number of trials required for adaptation. *P < 0.05, **P < 0.01, ***P < 0.001. See Supplementary Table 1 for statistics.