Learning from Noisy Labeled Data via Sharpen Prediction Loss and Re-Correction

Deep learning has achieved excellent results in many applications with a large number of high-quality annotation datasets. However, it is difficult to obtain numerous high-quality labeled samples, since it may generate noisy labeled data during manually annotating. Deep neural network with noisy labeled data leads to overfitting a nd greatly affects t he performance. In order to overcome the issue, a deep model with sharpen prediction loss and re-correction is proposed for learning from noisy labeled data, which aims to modify the loss distributions of noise samples and clean samples, and eliminate noise samples by unsupervised clustering. In the proposed framework, the deep model is warmed up through several epochs, sharpen prediction loss is proposed to effectively measure the sample loss, and recorrection is utilized to separate noise and clean samples as well as predict the final l abels. Experimental results on two datasets with noisy labeled data demonstrate that our proposed model yields superior classification accuracies.