Learning from Noisy Complementary Labels with Robust Loss Functions

抄録

It has been demonstrated that large-scale labeled datasets facilitate the success of machine learning. However, collecting labeled data is often very costly and error-prone in practice. To cope with this problem, previous studies have considered the use of a complementary label, which specifies a class that an instance does not belong to and can be collected more easily than ordinary labels. However, complementary labels could also be error-prone and thus mitigating the influence of label noise is an important challenge to make complementary-label learning more useful in practice. In this paper, we derive conditions for the loss function such that the learning algorithm is not affected by noise in complementary labels. Experiments on benchmark datasets with noisy complementary labels demonstrate that the loss functions that satisfy our conditions significantly improve the classification performance.