RRAC: Reciprocal Robustness via Asymmetric Co-training for Label Noise Resistance in Deep Learning

Learning from noisy labels remains a significant challenge in deep learning. Although recent methods typically propose single-model strategies, practical implementations often employ dual-model training, particularly in realistic scenarios where no clean validation data is available for identifying noisy labels or selecting clean samples. Existing dual-model methods, including symmetric and asymmetric co-training, have limitations in handling label noise. Symmetric co-training suffers from model convergence in the later stage of training. In contrast, asymmetric co-training lacks explicit mutual updating, limiting its ability to mutually correct or reinforce each other's predictions. To address these limitations, we propose Reciprocal Robustness via Asymmetric Co-training (RRAC). This novel framework leverages two robust models, the Temporal Weighting Model (TWM) and the Early Learning Regularization Model (ELRM), trained simultaneously and exchanging predictive information through cross-updating. TWM utilizes a dynamic instance weighting strategy (CSIW), while ELRM employs early learning-based regularization. Our experiments on synthetic and real-world noisy-label datasets demonstrate RRAC's competitive performance compared to state-of-the-art methods, offering a promising solution for robust learning from noisy labels.