PMAL: Progressive Multi-Label Active Learning via Dynamic Diversity Reweighting

Yuyang Cai, Yan XIAO, Ranao Liu, Wenqiang Wang, Jingwen Yan

Keywords: Active Learning, Data Subset Selection, Deep Neural Network, Facial Attribute Recognition

TL;DR: We propose PMAL, a dynamic active learning framework for multi-label classification that leverages Riemannian geometry and submodular optimization, achieving state-of-the-art performance on facial recognition and multi-class datasets.

Abstract: Facial attribute recognition under limited annotation budgets poses significant challenges due to strong label correlations, costly annotations, and the lack of principled active learning strategies for multi-label settings. To address these challenges, we propose Progressive Multi-Label Active Learning (PMAL), a framework that efficiently identifies informative samples under tight labeling budgets. PMAL decomposes multi-label joint entropy into independent components for uncertainty estimation, evaluates feature diversity through similarity computations on Riemannian manifolds, and employs a greedy batch-adaptive scoring mechanism that dynamically updates selection priorities. We further extend submodular optimization theory to dynamic multi-label selection and, for the first time, establish an $\mathcal{O}(\log n)$ stability bound under label relevance matrix perturbations. Extensive experiments on CelebA and LFWA demonstrate that PMAL consistently outperforms eight state-of-the-art baselines. Beyond multi-label tasks, we also introduce Progressive Active Learning (PAL) for multi-class settings, achieving superior results on CIFAR-10, CIFAR-100, and SVHN benchmarks. Beyond multi-label tasks, we also introduce Progressive Active Learning (PAL) for multi-class settings, achieving superior results on CIFAR-10, CIFAR-100, and SVHN benchmarks. Meanwhile, PAL is designed with scalable computational complexity, remaining efficient even on large-scale datasets, and achieves substantial runtime savings through lightweight diversity updates without compromising selection quality.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 2707