Learning from Noisy Labels with Complementary Loss Functions

Recent researches reveal that deep neural networks are sensitive to label noises hence leading poor generalization performance in some tasks. Although different robust loss functions have been proposed remedy this issue, they suffer from an underfitting problem, thus not sufficient learn accurate models. On the other hand, commonly used Cross Entropy (CE) loss, which shows high standard supervised learning (with clean supervision), is non-robust noise. In paper, we propose a general framework with complementary functions. our framework, CE and play roles joint objective as per their sufficiency robustness properties respectively. Specifically, find by exploiting memorization effect of networks, can easily filter out proportion hard samples generate reliable pseudo labels for easy samples, reduce noise quite low level. Then, simply on supervision original noisy supervision. procedure, guarantee optimization while be regarded supplement. Experimental results benchmark classification datasets indicate method helps achieve network training simultaneously.