🤖 AI Summary
This work studies noisy partial-label learning (noisy PLL), a more realistic weakly supervised setting where candidate label sets may exclude the ground-truth label—relaxing the strong assumption in conventional PLL that the true label must reside in the candidate set. To address this challenge, we propose IRNet, an iterative refinement network featuring two synergistic modules: noise sample detection and dynamic label correction, which jointly purify training data over iterations. We formally define the noisy PLL problem for the first time and theoretically prove that IRNet’s iterative framework converges to the Bayes-optimal classifier. To enhance robustness, we incorporate warm-start training, consistency regularization, and data augmentation. Extensive experiments on multiple benchmark datasets demonstrate that IRNet significantly outperforms existing state-of-the-art methods. Both theoretical analysis and empirical results validate its effectiveness in denoising and generalization.
📝 Abstract
Partial label learning (PLL) is a typical weakly supervised learning, where each sample is associated with a set of candidate labels. The basic assumption of PLL is that the ground-truth label must reside in the candidate set. However, this assumption may not be satisfied due to the unprofessional judgment of the annotators, thus limiting the practical application of PLL. In this paper, we relax this assumption and focus on a more general problem, noisy PLL, where the ground-truth label may not exist in the candidate set. To address this challenging problem, we propose a novel framework called"Iterative Refinement Network (IRNet)". It aims to purify the noisy samples by two key modules, i.e., noisy sample detection and label correction. Ideally, we can convert noisy PLL into traditional PLL if all noisy samples are corrected. To guarantee the performance of these modules, we start with warm-up training and exploit data augmentation to reduce prediction errors. Through theoretical analysis, we prove that IRNet is able to reduce the noise level of the dataset and eventually approximate the Bayes optimal classifier. Experimental results on multiple benchmark datasets demonstrate the effectiveness of our method. IRNet is superior to existing state-of-the-art approaches on noisy PLL.
