An Uncertainty-Aware Ensemble Approach to Modeling Utility of Pseudolabels for Semisupervised Learning

Semisupervised learning (SSL) typically filters out low-confidence predictions when generating pseudolabels. This paradigm suffers from two critical limitations: 1) the lack of an effective strategy for determining a confidence threshold and 2) the inherent overconfidence of deep neural networks (DNNs). To address these issues, this work introduces an uncertainty-aware ensemble structure (UES), which jointly models prediction uncertainty and confidence to assess the utility of pseudolabels. In particular, UES dynamically converts the utility of pseudolabels into sample weights (SWs). Furthermore, UES is equipped with architecture-agnostic metrics and can be seamlessly integrated into various computer vision tasks. Extensive experiments demonstrate that UES improves DualPose by 3.47% in the percentage of correct keypoints (PCK) on the Sniffing dataset (100 samples with 30 labeled), 7.29% in PCK on FLIC (100 samples with 50 labeled), and 3.91% in PCK on LSP (200 samples with 100 labeled). UES also increases FixMatch accuracy by 0.2% on CIFAR-10 (40 labels), 0.21% on CIFAR-100 (400 labels), and 1.05% on SVHN@250, demonstrating consistent performance gains across both small- and large-scale benchmarks. The code is publicly available at: https://github.com/Qi2019KB/UES.