Deep Learning on a Microcontroller Unit in a Hand-Prosthesis Control Context

Abstract: This master's thesis investigates in what capacity a typical hand-prosthesis classification ANN (Artificial Neural Network) can be deployed on a microcontroller unit. A working, implemented interface is presented along with extra scripts, applications and functions for easy training and testing. Performance of different architectures as well as model-conversion tools are evaluated and compared. Furthermore, potential drawbacks, advantages of ANN architectures and tools are discussed. A dense neural network, albeit primitive, implementation on current readily available microcontrollers might be feasible as a classifier in a real time hand-prosthesis context. Finally, possible solutions and improvements are proposed.