An iterative loss correction method for deep learning with noisy labels

Abstract

Noisy labelling is a prevalent issue in real-world data, often causing deep neural networks (DNNs) to overfit. Prior research in this area primarily relies on the accurate estimation of noise transition matrices, which is contingent on identifying anchor points in the clean data domain. However, current methods typically estimate the anchor points using information from the noisy labels, potentially resulting in poor estimation. In contrast, our novel method aims to enhance precision by developing an estimator that jointly learns the transition matrices and anchor points through iterative learning. Our approach is validated on the IMDB and MNIST datasets, proving to be more precise and effective than previous methods.