Extended Data Fig. 5: Model performance analysis and ablation studies for components in our learning models.

a, Confusion matrix for numpad typing data for each typing stroke after 20 transfer training. b, Confusion matrix for object recognition tasks for individual signal frame after 20 transfer training. c, More detailed comparison between TD-C Learning and supervised learning with last layer modification. For more precise comparison, we additionally trained TD-C learning model with labelled data used to train supervised model by removing labels. Even with the same number of training samples, our learning framework significantly outperform normal supervised learning when the model is transferred to predict different tasks. With more easily collectable unlabeled training samples, TD-C learning model pretrained with large random motion data shows higher accuracies in all transfer training epoch than other models. d, UMAP projection of latent vectors of labelled keypad typing data projected by our model pretrained with TD-C learning method. e, Ablation study for transfer accuracy comparison between applying timewise dependency loss and original contrastive learning loss. f, Ablation study for applying phase variable by comparing transfer accuracy trends for models with and without phase discrimination when inferencing different gestures in MFS.