Self Error Detection and Correction for Noisy Labels Based on Error Correcting Output Code in Convolutional Neural Networks

Abstract

When using convolutional neural networks in different applications, human errors may occur in labeling the data samples. To solve this problem, a self error detection and correction based on Error Correcting Output Code (SEDC-ECOC) method is proposed in this paper. The SEDC-ECOC method works in two stages. In the first stage, the distance between each sample and each class is measured using ECOC, which provides the base of error detection and correction. In the second stage, SVM ECOC conducts further correction as well as plays the role of classification layer in deep networks. Having the advantages of simple construction and independence from deep networks, the SEDC-ECOC method can be applied with different convolutional neural networks. The experimental results show that the proposed method achieves high correction performance for MNIST and CIFAR-10 datasets. Up to 56.09% and 92.11% erroneous sample labels are corrected by applying the proposed method once and twice respectively to noisy labels.