Training Deep Neural Networks for Image Applications with Noisy Labels by Complementary Learning

In recent years, deep neural networks (DNNs) have made great progress in many fields such as image recognition, speech recognition and natural language processing, etc. The rapid development of the Internet and mobile devices promotes the popularity of image applications and provides a large amount of data to be used for training DNNs. Also, the manually annotated data is the key of training DNNs. However, with the rapid growth of data scale, the cost of manual annotation is getting higher and the quality is hard to be guaranteed, which will damage the performance of DNNs. Combining the idea of easy example mining and transfer learning, we propose a method called complementary learning to train DNNs on large-scale noisy labels for image applications. With a small number of clean labels and a large number of noisy labels, we jointly train two DNNs with complementary strategies and meanwhile transfer the knowledge from the auxiliary model to the main model. Through experiments we show that this method can efficiently train DNNs on noisy labels. Compared with current approaches, this method can handle more complicated noise labels, which demonstrates its value for image applications.