Addressing the Constraints of Active Learning on the Edge

Abstract

The design of machine learning methodology often does not take into account the limitations of edge computing. In particular, active learning approaches have not considered the constraints of the edge, such as separate data locations (labeled data is on the cloud whereas unlabeled data is on the edge), cold starting or low initial model performance, limited budget sizes due to bandwidth constraints, and computational constraints due to edge hardware. Active learning on the edge could help decide what data to cache on the edge and what data to prioritize for offloading, facilitating efficient use of memory and bandwidth resources. Active learning on the edge would also allow for a machine learning model to be trained using a minimal amount of data. In this work, we examine the constraints of performing active learning on the edge, propose an active learning method that seeks to address these constraints, and discuss advances needed at large to improve active learning on the edge.