Currently, the construction of a large-scale manual annotation databases is still a prerequisite for the success of DNN. Although there is no shortage of data, there is a lack of clean label data in many fields, because it takes a lot of time and huge labor costs to build such a database. As many studies have shown that noisy label will seriously affect the stability and performance of the DNN. Learning from noisy labels has become more and more important, and many methods have been proposed by scholars. The purpose of this paper is to systematically summarize the different ideas for solving the noisy label learning problem, analyze the problems with existing methods, and try to analyze how to solve these problems. First, we will describe the problem of learning with label noise from the perspective of supervised learning. And then we will summarize the existing methods from the perspective of dataset usage. Subsequently, we will analyze the problems with the data and existing methods. Finally we will give some possible solution ideas.