A Noise Rate Estimation Method for Image Classification with Label Noise

Abstract

In a dataset, the misidentified labels can be assumed as the true labels flipped with a probability. In this paper, we study a general situation in which sample labels are corrupted at random. We propose a noise rate estimation method and prove that by adopting importance reweighting, the accuracy of classification with label noise problem can rise approximately 10% through any surrogate loss function. The two classification methods we choose for robustness analysis are convolutional neural network and convolutional neural network with importance reweighting. The details of these two methods are fully illustrated in this paper. We discuss the label noise problems and solutions in the introduction part and explain how the importance reweighting method and the noise rate estimation method are combined to deal with this problem. Experiments on Fashion-MNIST0.5, Fashion-MNIST0.6, and CIFAR with noise verify our approach. In the end, we also provide the transition matrix of the flip rate for each dataset.