Abstract:

When large-scale data is labeled artificially, labeling errors are easy to occur, which leads to the existence of noise labels in data sets, and further affects the generalization of deep neural network models. The sample selection mechanism in the existing research methods such as Co-teaching makes the noise samples easy to flow into the selected clean label sample subset, and it is difficult to control the deep neural network model's fitting to the selected clean sample subset in training. Therefore, this study presents a novel algorithm that improves upon Co-teaching. In this method, two regularization losses are added to prevent the model from placing too much trust in a single class and falling into a local optimal solution respectively. Additionally, the introduction of high learning rate attenuation training method makes the model more inclined to learn clean label sample features in the initial training to get better model parameters. Compared with the results of Co-teaching, the performance of the proposed model is improved on MNIST, CIFAR-10 synthetic noise data set and Animal10N realistic data set under 20% and 50% symmetric noise and 45% asymmetric noise environment.

Key words: deep learning, convolutional neural network, image classification, noisy-label data, label noise learning, Co-teaching training, learning rate, robust loss function