Several machine learning methods, including artificial neural networks (ANN), lineardiscriminant analysis (LDA), support vector machines (SVM),
and Gaussian mixture models havebeen previously used in myoelectric-based motion control using EMG data, with variable but generally favorable accuracy.
To the best of our knowledge, the only methods previously applied specifically to synergy-based gesture classification are SVMs and a variant of ANN’s called
Extreme Learning Machines (ELMs). However as noted above, they have only been tested in settings where the low dimensional structure of muscle activations
was derived from the same dataset that was used for gesture classification. Here we investigate the ability to transfer synergy patterns across three different
task domains, as described below. We compare SVM vs. ELM classifiers for this purpose and report on the sensitivity of ELM to the particular set of random initialization
weights.
