Shift Guided Active Learning

Abstract

Active learning is a pivotal machine learning paradigm where the algorithm queries data iteratively from an information source and updates itself accordingly. Active learning provides an instrument to investigate data selection and has been proven effective in reducing annotation costs. In a typical active learning framework, the query step only takes information from the current learning cycle and the information between cycles is usually ignored. It turns out that both inner-cycle and inter-cycle information provide crucial insights for learning progression. In this study, we identify the existence of distribution shifts that include both inner-cycle and inter-cycle information. This shift negatively impacts stability and model performance. To counter the impact of such a shift, we propose to integrate them into an active learning framework with specialized models. Our framework, Shift Adaptation via Guided Enquiry (SAGE), is founded on a set of dedicated query strategies guided by the distribution shift. We show that this new framework mitigates distribution shifts and outperforms previous studies on multiple computer vision benchmarks. With extensive experiments, we conclude that SAGE improves the state-of-the-art, with a significant 3.28% absolute accuracy improvement over the previous methods in the field of active learning. This framework is also compatible with semi-supervised (SSL) settings, allowing state-of-the-art SSL methods to attain higher performance.