Partial Label Learning via GANs With Multiclass SVMs and Information Maximization

Abstract

Partial label learning (PLL), an important branch of weakly supervised learning, addresses the problem that each instance is associated with a set of candidate labels and only one is correct. In this paper, a novel adversarial model PL-MIGAN is proposed to simultaneously mitigate two fundamental issues of generative adversarial networks (GANs) in PLL: label disambiguation performance of discriminator and instance synthesis quality of generator. First of all, multi-class support vector machines (SVMs) applied in discriminator to disambiguate the candidate labels and identify fake instances. This strategy not only improves the disadvantage that traditional supervised loss is unable to perform disambiguation but also reduces the influence of cumulative error caused by noise label propagation. Furthermore, a partial contrastive loss is constructed to extend the self-supervised contrastive approach to PLL, allowing us to effectively leverage ambiguous labels information. Finally, the generator jointly employ mutual information (MI) and partial contrastive loss to estimate the latent distribution of each class label. In addition, in order to reduce the impact of ambiguous information, an iteratively optimization procedure is designed to update the label confidence matrix as conditional information guides the generation of instance classes. As adversarial learning proceeds, both the discriminator and the generator alternately and iteratively boost their performance. Simulation results reveal the overwhelming performance of PL-MIGAN.