Deep Learning with noisy labels is a practically challenging problem in weakly supervised learning. The state-of-the-art approaches "Decoupling" and "Co-teaching+" claim that the "disagreement" strategy is crucial for alleviating the problem of learning with noisy labels. In this paper, we start from a different perspective and propose a robust learning paradigm called JoCoR, which aims to reduce the diversity of two networks during training. Specifically, we first use two networks to make predictions on the same mini-batch data and calculate a joint loss with Co-Regularization for each training example. Then we select small-loss examples to update the parameters of both two networks simultaneously. Trained by the joint loss, these two networks would be more and more similar due to the effect of Co-Regularization. Extensive experimental results on corrupted data from benchmark datasets including MNIST, CIFAR-10, CIFAR-100 and Clothing1M demonstrate that JoCoR is superior to many state-of-the-art approaches for learning with noisy labels.
@article{arxiv.2003.02752,
title = {Combating noisy labels by agreement: A joint training method with co-regularization},
author = {Hongxin Wei and Lei Feng and Xiangyu Chen and Bo An},
journal= {arXiv preprint arXiv:2003.02752},
year = {2020}
}
Comments
Accepted by CVPR 2020; Code is available at: https://github.com/hongxin001/JoCoR. arXiv admin note: text overlap with arXiv:1901.04215 by other authors