Fig. 5: Classification of condition (object, sign) based on different subspaces of kinematics.

a Mean grasp classification performance after progressively removing PCs, from high variance to low. Results are averaged across eight subjects with five repetitions each (random allocation of training and testing data in each repetition, see “Methods” section for more details). Error bars denote the SEM. b Mean ASL classification performance based on reduced kinematic subspaces, averaged across three subjects with five repetitions each. Vertical red lines denote the PC beyond which the subspace accounts for <1% of the variance; error bars denote the standard error of the mean (SEM). Note that objects and letters can be classified accurately based on just a handful of high-variance PCs (Supplementary Fig. 4). Classification performance is highly consistent both across repetitions of the same subject and across subjects (note the small SEM).