Hybrid Query Strategy with Diversity-Weighted Metropolis–Adjusted Langevin Algorithm

Min Young Baeg, Yoon-Yeong Kim

Keywords: Active Learning, Langevin Dynamics, Representativeness, Diversity

Abstract: Although deep learning has achieved remarkable success in various fields, most of these advances typically rely on a large-scale well-annotated dataset. However, in real-world applications, collecting labeled data is often expensive and timeconsuming. Active Learning (AL) has emerged as a promising solution to mitigate labeling costs by selectively querying the most informative instances for annotation. In particular, hybrid AL methods have been gaining attention by integrating multiple acquisition criteria such as uncertainty, diversity, or representativeness as a joint function. In this paper, we propose a novel hybrid active learning method named Diversity-Weighted Metropolis-Adjusted Langevin Algorithm (DW-MALA). Our method precisely approximates the data distribution by leveraging gradient-based Langevin dynamics, and selects instances from high-density regions using a representativeness score derived from density estimates. Simultaneously, a diversity score is incorporated by measuring the distance to the nearest labeled instance, which also ensures coverage of low-density regions. The quantitative and qualitative analyses demonstrate the effectiveness of DW-MALA in selecting diverse and representative samples under a limited labeling budget, compared to the baselines.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 2298