Learning from Noisy Labeled Data Using Symmetric Cross-Entropy Loss for Image Classification
Hiroshi Takeda, Soh Yoshida, Mitsuji Muneyasu · 2020
While image classification with deep neural networks (DNNs) has made remarkable progress, learning from noisy labeled data degrades performance, and this problem remains challenging. Some previous methods, such as Co-teaching and MentorNet, use the memorization effects, which are characteristics of DNN training that first learn easy (most likely clean) samples and then difficult (most likely noisy) samples. These methods treat samples with small training loss as clean, and we call this idea the “small-loss criterion”. However, when there is a difference in learning speed between samples, a class including easy samples tends towards overfitting. In this paper, we propose a robust learning method by introducing a symmetric cross-entropy (SCE) loss to control the difference in learning speed between all classes. Experimental results on the CIFAR-100 image classification dataset containing 45% and 50% of noisy labels show the robustness of the proposed method.