Recalibration of myoelectric control with active learning

Figure 1. The generic framework of active learning. The feedback loop includes a small set of labeled data L that is used to train a machine learning model. Based on that, the active learner explores a large pool of unlabeled data U and selects new samples. These samples are then passed to an oracle, which would be a prosthesis user in this paper, for labeling. These samples join the training set for model retraining.

Figure 2. Averaged test accuracy across all able-bodied participants (A) and people with limb difference (B). Initial training sets for subjects in both groups contained 50 samples from each class. Initial accuracy without active learning is presented at query step zero. Offline baseline: 82% (A) and 63% (B) for able-bodied and people with limb loss, respectively.

Figure 3. Accuracy plots to compare all investigated query strategies at two stages: After the first 200 and after all 1,500 queried samples. Each marker represents results from one participant. Random sampling (in A) compared against the remaining three approaches shows a visible tendency to achieve lower accuracy. Irrespective of the choice of query sampling, active learning outperformed the conventional random sampling, because results from all participants lie above the unity line, i.e., in the blue zone, when 1,500 samples were selected. A similar conclusion can be drawn when 200 samples wee drawn although for a very few participants, random sampling was more successful. When different active learning query strategies were compared (B), different query sampling were comparable but margin sampling seemed to offer a lightly higher performance.

Figure 4. Grip instances acquisition during 1,500 querying process across all the subjects: able-bodied participants (A) and people with limb difference (B). Query strategies are compared against class. In (A), the highest demand for new instances within 1,500 queries was for the lateral grip. The pointer grip and hand open classes needed the fewest samples. In contrast to able-bodied participants, active learning requested fewest samples for the rest class. In all cases, random sampling drew 250 samples on average.

Figure 5. Comparison of single instance (A), batch mode (B) and ranked batch mode (C) sampling methods in one representative participant.

Figure 6. The envisaged real-time operation of a user-in-the loop setting with an active learning. Movement intention generates new signal data each second (w.r.t. the sampling rate). Unlabeled samples after being preprocessed are classified by pre-trained decoder. Active learner using query sampling criteria decides whether to request the label or not. Confident samples pass certainty check criteria to be sent to the prosthesis actuator. If a sample is queried by the active learner, the oracle dis-/approves the decoder prediction. Incorrect predictions are filtered with R&I (representativeness and informativeness criteria) for outlier detection and interesting samples to be weakly-labeled. These combined with ground truth samples are used to update the decoder. *R&I, representativeness and informativeness.