Evaluating User and Machine Learning in Short-and Long-term Pattern Recognition-based Myoelectric Control
Abstract : ? Proper training is essential to achieve reliable pattern recognition (PR) based myoelectric control.
? The amount of training is commonly determined by experience.
? The purpose of this study is to provide an offline validation method that makes the offline performance transferable to online control and find the proper amount of training that achieves good online performance.
? In the offline experiment, eight able-bodied subjects and three amputees participated in a ten-day training.
? Repeatability index (RI) and classification error (CE) were used to evaluate user learning and machine learning, respectively.
? The performance of cross-validation (CV) and time serial related validation (TSV) was compared.
? Learning curves were established with different training trials by TSV.
? In the online experiment, sixteen ablebodied subjects were randomly divided into two groups with oneor five-trial training, respectively, followed by participating in the test with and without classifier-output feedback.
? The correlation between offline and online tests was analyzed. Results indicated that five-trial training was proper to train the user and the classifier.
? The long-term retention of skills could not shorten the learning process. The correlation between CEs of TSV and the online test was strong (r = 0.87) with five-trial training, while the correlation between CEs of CV and the online test was weak (r = 0.30).
? Outcomes demonstrate that offline performance evaluated by TSV is transferable to online performance and the learning process can guide the user to achieve good online myoelectric control with minimum training
? It consists of training a model on a source domain (abundance of labeled data) and using the trained weights as a starting point when presented with a new task. However, fine-tuning can suffer from catastrophic forgetting, where relevant and important features learned during pre-training are lost on the target domain (i.e. new task). Moreover, by design, fine-tuning is ill-suited when significant differences exist between the source and the target, as it can bias the network into poorly adapted features for the task at hand. Progressive Neural Networks (PNN)attempt to address these issues by pre-training a model on the source domain and freezing its weights.
The participants took part in a ten-day experiment. In each day, the subjects were asked to perform one contraction of each motion randomly displayed on the screen in one trial. Each contraction lasted 5 s and there was a 5-s rest between contractions and a 60-s rest between trials to avoid fatigue. Each subject completed 20 trials per day and the electrode location was marked. The experiment usually took about 1 hour. Able-bodied subjects performed all the thirteen motions while the amputees only performed eleven motions because they reported difficulties performing ball grasp and cylindrical grasp
? Training one network per source-task (i.e. per participant) for the PNN is not scalable in the present context.
? However, by training a Source Network (presented in Sec. V) shared across all participants of the pre-training dataset with the multi-stream AdaBatch and adding only a second network for the target task using the PNN architecture, the scaling problem in the current context vanishes.
? This second network will hereafter be referred to as the Second Network.
? The architecture of the Second Network is almost identical to the Source Network.
? The difference being in the activation functions employed.
In CV, the original data is randomly divided into training and testing sets, and this process is conducted many times to evaluate the averaged classification error. However, EMG patterns are non-stationary because of user learning.The effect of user learning cannot be evaluated effectively due to the iterations in CV. By contrast, in a few studies, the first few trials were regarded as the training set and the rest trials were the testing set, which was called time serial related validation (TSV) in this study. The training set of TSV is in time sequence, which has the potential to evaluate the effect of user learning. Moreover, whether the offline performance evaluated by TSV is transferable to online control has not been studied. The correlation between offline and online performance is an open question.
|