Towards Safe Weakly Supervised Learning

In this paper, we study weakly supervised learning where a large amount of label information is not accessible. This includes incomplete supervision such as semi-supervised learning and domain adaptation; inexact supervision, such as multi-instance learning and inaccurate supervision, such as label noise learning. Unlike supervised learning, weakly supervised learning, however, may sometimes even degenerate performance. Such deficiency definitely hinders the deployment of weakly supervised learning to real applications. For this reason, it is desired to study safe weakly supervised learning. In this paper we present a generic ensemble learning scheme to derive the safe prediction. We consider optimizing the worst-case performance gain which leads to a maximin optimization. Our resultant formulation brings multiple advantages. Firstly, for many commonly used convex loss functions in classification and regression tasks, our formulation is guaranteed to derive a safe prediction under a mild condition. Secondly, prior knowledge related to the weight of the base weakly supervised learners can be flexibly embedded. Thirdly, our formulation can be globally and efficiently addressed. Finally, it is in an intuitive geometric interpretation. Extensive experiments on multiple weakly supervised learning tasks clearly demonstrate the effectiveness of our proposal algorithms.