Weakly Supervised Crowdsourcing Learning Based on Adversarial Consensus

Abstract

Crowdsourcing provides an efficient way to obtain labels for large datasets in the deep learning era. However, due to the non-expert workers, the annotations are usually noisy.Besides, concerning the labeling cost, sparse annotations are common. To face this challenge, we propose an approach based on adversarial consensus, which trains one classifier for each worker, and enforces their predictions over the ground-truth labels to be maximally consistent by exploiting the generative adversarial learning idea. We give two implementations respectively for the light and heavy noise cases. Extensive experiments on real-world and synthetic datasets demonstrate the effectiveness of our approach.