CCT: A Cyclic Co-Teaching Approach to Train Deep Neural Networks With Noisy Labels

Abstract

The detrimental impact of noisy labels on the generalization performance of deep neural networks has sparked research interest in learning with noisy labels (LNL). Among the various methods proposed to mitigate this effect, the Co-Teaching method, characterized by co-training with the small-loss criterion, is one of the most established approaches and is widely employed as a key component in recent LNL methods. Although Co-Teaching can mitigate the overfitting effect, it still remains, especially in scenarios with high rates of label noise in datasets. Strategies from the LNL literature to address this typically include the use of disagreement techniques and alternative loss functions. In this paper, we propose the Cyclic Co-Teaching (CCT) method, which employs cyclic variations in the learning rate and sample retention rate at the mini-batch level, along with a checkpoint mechanism that ensures that training in subsequent cycles always resumes from the best models obtained so far. For optimizing the method, we developed a framework that incorporates a pre-training phase to obtain an optimized vanilla model used to initialize CCT model weights, and a transparent univariate optimization strategy for hyperparameters that does not necessarily require a clean validation set. Experimental results on synthetic and real-world datasets, under different types and levels of noise and employing various neural network architectures, demonstrate that CCT outperforms several state-of-the-art LNL methods in most evaluated scenarios.