Masters Thesis

Development of spiking neural network for electromyographical signal classification

Major challenges of implementing pattern recognition based multi-gesture myoelectric prosthetic control systems on embedded platforms have been the reduction of latency, power consumption, and system size. To circumvent these challenges, state of the art prosthetic systems offload computationally expensive processes to co-processors. Hardware implementations of learning algorithms may help resolve the need for codesigns but have been historically limited by lack of suitable algorithms and electronic devices. Recent developments in memristor technology and Spiking Neural Network (SNN) training theory have opened the door for purely analog hardware implementations of supervised neural network classifier algorithms. In this work a SNN simulator is used to study network topologies and novel training algorithms for classification of myoelectric signals. Empirical design methods are employed to gain insight into a variety of derivativefree error feedback strategies. A modular network topology with direct error feedback is demonstrated to be a suitable candidate for memristor based neuromorphic processor.

Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.