Abstract: Training deep models with a large amount of label noise can mislead model training and significantly reduce model accuracy. Therefore, it is important to train high-precision and high-robustness models on low-cost, high-noise labeled data. We proposed a "data self-augmentation" technique to address the problem of neural network overfitting to noisy labels. We proposed a noise filtering algorithm called automatic collaborative differential teaching, which used neural network architecture search design schedules to assist noise label filtering and better utilize the network memory effect to simulate the optimal selection process of small loss samples. Experimental results show that at a high noise rate of 0.8, the accuracy of MNIST, CIFAR-10, CIFAR-100 datasets are increased by 14.49%, 10.86%, 4.36% respectively.