Towards a Real-Time, Interactive, Incremental Learning Algorithm for Prosthetic Myocontrol

Myocontrol based upon machine learning usually requires a large initial labeled dataset to build its model and prescribes that said model will never change in the course of time. This approach has several limitations, e.g., the difficulty of building such a proper dataset and the low generalization power of the obtained model against the variety of situations encountered by the user in daily life. In this work we introduce an interactive, incremental learning method for intention detection in real-time, along with some optimization techniques. The method enables users to quickly update the myocontrol system on demand, reducing the impact of the non-stationarity of input signals, improving the generalization power and providing an interactive experience to the user.