Abstract

This work proposes and theoretically justifies a model that fuses weak supervision and generative adversarial networks to improve the estimate of unobserved labels and data augmentation, outperforming baseline weak supervision models on multiclass image classification datasets.