Figure 2.

User interaction with the algorithm in offline and real-time scenarios. (a) Offline classification presents no user interaction with the algorithm. The user records a dataset, generating muscle contractions corresponding to target gestures. Data is then sent to the neural network with its output being compared to the recorded target label for training purposes. (b) The real-time inference loop allows for the user to react to the prediction feedback. Voluntary muscle contractions can be adjusted in order to maintain or change the predicted output. (c) Example of predictions during a gesture transition. Subsequent initial state predictions are considered true negatives (TN), and false negatives (FN) if the motion is engaged, but the prediction has not yet changed. Once a new gesture is predicted during the transition, the new prediction is considered a false positive (FP) if incorrect, or a true positive (TP) if correct.