On the Utility of Representation Learning Algorithms for Myoelectric Interfacing

Electrical activity produced by muscles during voluntary movement is a reflection of the firing patterns of relevant motor neurons and, by extension, the latent motor intent driving the movement. Once transduced via electromyography (EMG) and converted into digital form, this activity can be processed to provide an estimate of the original motor intent and is as such a feasible basis for non-invasive efferent neural interfacing. EMG-based motor intent decoding has so far received the most attention in the field of upper-limb prosthetics, where alternative means of interfacing are scarce and the utility of better control apparent. Whereas myoelectric prostheses have been available since the 1960s, available EMG control interfaces still lag behind the mechanical capabilities of the artificial limbs they are intended to steer—a gap at least partially due to limitations in current methods for translating EMG into appropriate motion commands. As the relationship between EMG signals and concurrent effector kinematics is highly non-linear and apparently stochastic, finding ways to accurately extract and combine relevant information from across electrode sites is still an active area of inquiry.This dissertation comprises an introduction and eight papers that explore issues afflicting the status quo of myoelectric decoding and possible solutions, all related through their use of learning algorithms and deep Artificial Neural Network (ANN) models. Paper I presents a Convolutional Neural Network (CNN) for multi-label movement decoding of high-density surface EMG (HD-sEMG) signals. Inspired by the successful use of CNNs in Paper I and the work of others, Paper II presents a method for automatic design of CNN architectures for use in myocontrol. Paper III introduces an ANN architecture with an appertaining training framework from which simultaneous and proportional control emerges. Paper Iv introduce a dataset of HD-sEMG signals for use with learning algorithms. Paper v applies