An Improved Method of Identifying Mislabeled Data and the Mislabeled Data in MNIST and CIFAR-10

Abstract

Objects classification is an important part of machine learning and the quality of the training data plays an important role. Some datasets, such as MNIST and CIFAR-10 are regarded as ground true data, and the accuracy on the two datasets is an important criterion for a machine learning models or algorithms. A number of mislabeled data detection techniques have been proposed; however, there is no reproduction work on MNIST and CIFAR-10. In this paper I use an improved method to identify mislabeled data in MNIST and CIFAR-10 and find 675 errors in MNIST, 118 errors in CIFAR-10. After correcting mislabeled instances, the accuracy increases. And the list of the current state of art of different datasets needs to be reproduced again with new dataset.