This paper introduces a novel deep learning paradigm called "Co-teaching" to address the challenge of training deep neural networks with noisy labels. The authors propose training two deep neural networks simultaneously, allowing them to "teach" each other using mini-batches. Each network selects a subset of data with small losses as useful instances and shares these instances with its peer network for further training. This approach leverages the memorization effect of deep models, which initially learns clean and easy patterns before adapting to noisy labels. The empirical results on noisy versions of MNIST, CIFAR-10, and CIFAR-100 datasets demonstrate that Co-teaching significantly outperforms state-of-the-art methods in robustness, even under extremely noisy conditions (45% noisy labels). The paper also discusses the theoretical and practical implications of Co-teaching, including its potential extensions to other weak supervisions and theoretical guarantees.This paper introduces a novel deep learning paradigm called "Co-teaching" to address the challenge of training deep neural networks with noisy labels. The authors propose training two deep neural networks simultaneously, allowing them to "teach" each other using mini-batches. Each network selects a subset of data with small losses as useful instances and shares these instances with its peer network for further training. This approach leverages the memorization effect of deep models, which initially learns clean and easy patterns before adapting to noisy labels. The empirical results on noisy versions of MNIST, CIFAR-10, and CIFAR-100 datasets demonstrate that Co-teaching significantly outperforms state-of-the-art methods in robustness, even under extremely noisy conditions (45% noisy labels). The paper also discusses the theoretical and practical implications of Co-teaching, including its potential extensions to other weak supervisions and theoretical guarantees.