Dynamic Selection for Reconstructing Instance-Dependent Noisy Labels

As an inevitable issue in annotating large-scale datasets, instance-dependent label noise (IDN) can cause serious overfitting in neural networks. To combat IDN, label reconstruction methods have been developed with noise transition matrices or DNNs to simulate the transition from clean labels to noisy labels. Nevertheless, the absence of correct supervisions will lead to learning wrong noise transitions. This motivates us to select samples with clean labels to fetch the correct supervisions. However, the difficulty in obtaining prior knowledge of the noise rate prohibits the use of existing sample selection methods. To this end, we propose a dynamic sample selection method, namely Identity Mapping (IdMap), to overcome this limitation. Inspired by the feature-dependent characteristic of IDN, we first introduce the extracted instance features and pseudo-ground-truth labels to reconstruct noisy labels. A partial identity mapping between two labels is then established and samples with consistent identity mapping output are selected as clean data to update the classifier. Extensive experiments on both artificial and real-world noisy datasets demonstrate the superiority of IdMap compared with other state-of-the-art methods.