Incremental Semi-Supervised Tri-Training: A framework for model personalization

The need for real-time, tailored, and autonomous decision-making applications driven by machine learning and deep learning models has propelled the development and use of edge computing devices. Edge computing devices are characterized by the generation of large volumes of data and minimal computational resources (memory, processing capacity, etc.). In this work, we present a novel and integrated framework, called Incremental Semi-Supervised Tri-Training, for incremental update as well as personalization of models residing on edge computing devices in the absence of domain specific labelled data. The proposed framework encompasses three heavy weight neural networks working collaboratively to generate high quality pseudo-labels in a self-supervised manner to train a teacher model. The knowledge from the teacher model is then distilled into the student model residing on the edge device. The developed models are incrementally updated resulting in personalization of the edge model over time without requiring labels/annotations. We demonstrate the efficacy of the proposed framework in the context of a sleep staging application using the Physionet Sleep EDF-Expanded data repository.