Rethinking Label Refurbishment: Model Robustness under Label Noise

A family of methods that generate soft labels by mixing the hard with a certain distribution, namely label refurbishment, are widely used to train deep neural networks. However, some these still poorly understood in presence noise. In this paper, we revisit four refurbishment and reveal strong connection between them. We find they affect network models different manners. Two them smooth estimated posterior for regularization effects, other two force model produce high-confidence predictions. conduct extensive experiments evaluate related observe both effects improve generalization under Furthermore, theoretically show lead guarantees on clean distribution despite being trained noisy labels.