Learning noisy transition matrix using a neural network

In the field of deep learning, it is crucial to know the accurate distribution of dataset. However, to obtain a high quality of dataset using traditional methods is prone to be both costly and inefficient. Using deep learning methods to estimate the noisy transition matrix provides a feasible way, as the result of its essential function of learning to denote the relationship between clean labels and noisy labels and building a statistically consistent classifier. The major difficulty of learning the noisy transition matrix stems from the unavailability of the distribution of clean data and noisy data. In this paper, we propose a practical and convenient method to study a combination of augumentation and a novel loss function, only leveraging the already known clean labels to aid in learning the noisy transition matrix in the whole dataset. Finally, Through the experiment, the result demonstrates a superior performance and generalization capabilities of the proposed method.