Combining Active Learning and Self-Labeling for Deep Learning from Data Streams

Jan Wasilewski, Michal Wozniak, Bartosz Krawczyk

Abstract: The growing volume of data streams poses distinct challenges for machine learning algorithms. For instance, real-time video analysis is challenging due to the similarity of consecutive frames and the general richness of the stream. In order to apply a supervised learning algorithm, one needs labels for the frames. However, not every frame can (nor should) be annotated due to limited budget constraints. Stream active learning addresses this problem by designing an acquisition function that assigns weights to every incoming frame. In this study, we combine stream active learning with self-supervised learning to utilize frames not selected for annotation. We experimentally show that adding additional, unannotated frames to the training batch and propagating labels boosts the accuracy of the algorithm.