Gradient Agreement Hinders the Memorization of Noisy Labels

The performance of deep neural networks (DNNs) critically relies on high-quality annotations, while training DNNs with noisy labels remains challenging owing to their incredible capacity memorize the entire set. In this work, we use two synchronously trained reveal that may result in more divergent gradients when updating parameters. To overcome this, propose a novel co-training framework named gradient agreement learning (GAL). By dynamically evaluating coefficient every pair parameters from identical determine whether update them process. GAL can effectively hinder memorization labels. Furthermore, utilize pseudo produced by as supervision for another network, thereby gaining further improvement correcting some overcoming confirmation bias. Extensive experiments various benchmark datasets demonstrate superiority proposed GAL.