结合 KL 熵的改进 Mix-up 用于从噪声标签中学习
计算机视觉与模式识别
2019-08-19 v2
摘要
尽管深度神经网络（DNN）在图像分类研究中已取得优异性能，DNN 的训练需要大量带准确标注的干净数据。收集数据集容易，但标注所收集的数据困难。网络上存在着许多含不准确标注的图像数据，但在这些数据集上训练可能使网络更易过拟合噪声标签并导致性能下降。在本工作中，我们提出一种改进的联合优化框架，将 mix-up 熵与 Kullback-Leibler（KL）熵混合作为损失函数。该新损失函数在框架更新完标签标注后能提供更好的微调。我们在 CIFAR-10 数据集与 Clothing1M 数据集上开展实验。结果表明相较其他 SOTA 方法，我们的方法具有优势性能。
引用
@article{arxiv.1908.05488,
title = {Improved Mix-up with KL-Entropy for Learning From Noisy Labels},
author = {Qian Zhang and Feifei Lee and Ya-Gang Wang and Qiu Chen},
journal= {arXiv preprint arXiv:1908.05488},
year = {2019}
}
备注
The research in this paper we think is not enough to publish, so we will continue to research it. Due to we need to add more experiments and change the whole structure of this paper, it will spend a lot of time. We want to publish our research after all the change have been done, so we apply to withdraw this version