Learning from Noisy Labeled Samples Using Prediction Norm for Image Classification
Learning from Noisy Labeled Samples Using Prediction Norm for Image Classification
复制标题
DOI：
10.1109/smc52423.2021.9659001
复制
发表时间：
2021-10
期刊：
影响因子：
--
通讯作者：
Daiki Okamura;Ryosuke Harakawa;M. Iwahashi
中科院分区：
文献类型：
--
作者：
Daiki Okamura;Ryosuke Harakawa;M. Iwahashi
卷积神经网络（CNN）从有噪声标记的样本中学习是一个关键问题，已经进行了许多研究。尽管称为联合训练与协正则化（JoCoR）的最先进方法已经取得了高性能，但对具有非对称噪声的样本进行准确分类仍然是一个具有挑战性的问题，即相似类之间的错误标签（例如CAT↔DOG）。在本文中，我们新发现干净标记样本的预测范数与有噪声标记样本的预测范数存在差异。此外，我们发现CIFAR-10数据集的预测范数差与分类精度之间存在正相关关系。因此，我们假设增加预测范数差可以提高判别能力。基于这一假设，提出了一种从有噪声标记样本中学习CNN的新方法。具体而言，我们以JoCoR为基础架构，并对JoCoR的损失函数进行加权，使预测范数差更大。CIFAR-10数据集的实验结果表明我们的假设是正确的。对于非对称噪声和对称噪声（即其他类别之间的错误标签）的样本，所提出的方法的分类精度高于包括JoCoR在内的一些最先进的方法。
Learning of the convolutional neural network (CNN) from noisy labeled samples is a crucial problem, and many studies have been conducted. Although the state-of-the-art method called Joint training with Co-Regularization (JoCoR) has achieved high performance, it is a still challenging problem to accurately classify samples with asymmetric noise, i.e., wrong labels between similar classes (e.g., CAT ↔ DOG). In this paper, we newly found that there is a difference between the prediction norm for clean labeled samples and that for noisy labeled samples. In addition, we found that there is a positive correlation between the prediction norm difference and the classification accuracy for the CIFAR-10 dataset. Therefore, we hypothesize that the discriminative power would be improved if we increase the prediction norm difference. Based on this hypothesis, a novel method for learning CNN from noisy labeled samples is proposed. Specifically, we take JoCoR as the base architecture and weight the loss function of JoCoR to make the prediction norm difference large. Experimental results for the CIFAR-10 dataset suggest that our hypothesis is correct. The classification accuracy by the proposed method is higher than some state-of-the-art methods including JoCoR for samples with asymmetric noise as well as those with symmetric noise (i.e., wrong labels between other classes).