ACD-U: Asymmetric co-teaching with machine unlearning for robust learning with noisy labels

Deep neural networks are prone to memorizing incorrect labels during training, which degrades their generalizability. Although recent methods have combined sample selection with semi-supervised learning (SSL) to exploit the memorization effect — where networks learn from clean data before noisy data — they cannot correct selection errors once a sample is misclassified. We propose asymmetric co-teaching with different architectures and machine unlearning (ACD-U), a robust learning framework that integrates different model architectures with machine unlearning. ACD-U addresses the aforementioned limitation through two core mechanisms. First, it pairs a Contrastive Language-Image Pretraining (CLIP)-pretrained vision Transformer (ViT) with a convolutional neural network (CNN), leveraging their complementary learning behaviors: the ViT provides stable predictions, whereas the CNN remains adaptive throughout training. This asymmetry, with the ViT trained only on clean samples and the CNN trained through SSL, effectively mitigates confirmation bias. Second, selective unlearning enables post hoc error correction by identifying incorrectly memorized samples through loss trajectory analysis and CLIP consistency checks, and removing their influence via Kullback–Leibler divergence-based forgetting. This approach shifts the learning paradigm from passive error avoidance to active error correction. ACD-U achieves a 35% relative improvement over DivideMix on CIFAR-100 under 90% symmetric noise, the best results across all noise rates on Red Mini-ImageNet, and top accuracy on both Clothing1M and WebVision. Ablation analyses confirm that the pretrained ViT reduces critical sample misclassifications to approximately one-sixth of those of DivideMix, and that the unlearning and asymmetric co-teaching contribute complementarily across different noise regimes. These results suggest that ACD-U can improve the reliability of deep learning models trained on large-scale web-collected datasets, where label noise is unavoidable. The code is publicly available at https://github.com/meruemon/ACD-U .