.png){kind=icon}
.png){kind=icon}
Return
MD2PLL: a meta-learning framework for robust disambiguation in partial-label visual recognition
.png){kind=icon}
PRE
AI
X
X
J
L
K
Q
J
DOI：10.1007/s00371-026-04686-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Partial-label learning (PLL) is a weakly supervised learning framework that aims to identify the ground-truth label from a set of candidate labels for each instance. Existing methods often treat all training samples uniformly, making them vulnerable to early-stage noisy supervision and unreliable label disambiguation. To address these issues, this paper proposes meta-learning-based dual-model disambiguation for partial-label learning (MD2PLL), a robust framework for candidate-label disambiguation under ambiguous annotations. MD2PLL uses a Gaussian mixture model (GMM) to model training losses, identify low-loss and high-confidence samples, and dynamically construct a pseudo-validation set during training. Meanwhile, a dual-model architecture is designed, where the main network learns task representations and the meta-network generates corrected candidate-label distributions through pseudo-validation-guided optimization. Experiments are conducted on five real-world PLL datasets covering visual, audio, and textual scenarios, three benchmark image datasets under different ambiguity levels, and additional CIFAR-10 experiments for visual benchmark scalability. On the real-world datasets, MD2PLL achieves the best accuracy on Lost, MSRCv2, and BirdSong, improving over the second-best methods by 3.07, 16.79, and 2.70 percentage points, respectively. It also achieves competitive performance on Yahoo! News and Soccer Player, although it is not the top-performing method on these two datasets. Ablation studies further confirm the contributions of dynamic pseudo-validation updating, GMM-based screening, and label correction. We also analyze convergence behavior, feature clustering, GMM-based sample selection, and confidence evolution to better understand the proposed framework. The source code is publicly available at https://github.com/yuwanw/MD2PLL .
Keywords:
Partial-label learning
Weakly supervised learning
Partial-label visual recognition
Label disambiguation
Meta-learning
Dynamic sample selection
Label correction
Journal
IF:
2.9
Papers:
4.5K
Citations:
6.5K
Organization
Cited Papers