Co-adaptive Human-Machine Learning

Recent advancements in the field of prosthetics and biomedical engineering introduced potential in applying machine learning to create more adaptable and integrable hand prosthesis. Today research is directed towards adopting and replicating the biological functionality of the human hand. Yet, about 40% of the amputees reject bionics hand. We believe that by creating a co-adaptive environment with advanced machine learning approaches users will be enabled to their previous lives.

Our approach focuses on decoding motor control via pattern recognition. Using for the first time active learning within the prosthetics field, we aim to personalise the adaptation process via the user as an oracle during the training. By defining a suitable query sampling strategy for the annotation process, the model will learn the feature space of the muscle contraction signals. Based on the initial insights and literature review, stream-based sampling with explored query rules will be investigated. Due to the nature of active learning in high dimensional feature space, combining deep learning and active learning, known as DAL, in the pipeline will be considered.

To ensure human adaptation within the learning process, this Ph.D. work will also establish an interface for sensory feedback. As noted little to no feedback is one of the crucial flaws for the cognitive load in the closed-loop communication between the bionic arm and user. Interpretable and transparent features derived from dimensionality reduction algorithms will provide a crucial component for a co-adaptive environment in the learning process. Research within this programme will present novel paradigms that will be applied within control groups and in real-life settings.