.png){kind=icon}
Return
Temporal Evolution-Aware Adaptive Sampling for Open-Set Active Learning
.png){kind=icon}
PRE
AI
DOI：10.1109/tcsvt.2026.3727187.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Traditional informativeness-based sample selection strategies exhibit poor generalization performance in open-set scenarios. To address this issue, existing studies attempt to mitigate Out-of-Distribution (OOD) interference by prioritizing the sampling of pure In-Distribution (ID) samples. However, such purity-oriented approaches are constrained by limited model representation capacity and overlook the ID purity-informativeness dilemma. Furthermore, reliance on random sampling to construct the initial base-class labeled pool leads to substantially increased annotation costs. To address these challenges, we propose a novel adaptive sampling framework that jointly considers ID likelihood and informativeness. Specifically, we introduce a temporal evolution-aware strategy for the ID sampling branch that quantifies epistemic and aleatoric uncertainty via Spearman correlation and perturbation-sensitive feedback, effectively compensating for limited representation capacity in early cycles. In the informativeness evaluation brach, we leverage distribution attribution entropy to decouple information gain from OOD influences. Additionally, we incorporate a meta-confidence evaluator to calibrate the reliability of ID likelihood, achieving a dynamic balance between ID purity and informativeness. Building upon this framework, we initialize annotation of base-class labeled pool through a one-shot mechanism. Experimental results on multiple benchmark datasets demonstrate that our proposed method achieves state-of-the-art performance.
Keywords:
Open-set active learning
adaptive sampling
uncertainty quantification
temporal evolution
meta-confidence calibration
Journal
IF:
11.1
Papers:
845
Citations:
3.1W
Organization
View more organizations
Cited Papers
No cited papers available