Enhanced positive and unlabeled learning via pseudo supervision

Chengjie Wang; Chengming Xu; Zhenye Gan; Jianlong Hu; Wenbing Zhu; Lizhuag Ma

Enhanced positive and unlabeled learning via pseudo supervision

计算机视觉与模式识别 2024-07-10 v1 机器学习

作者: Chengjie Wang , Chengming Xu , Zhenye Gan , Jianlong Hu , Wenbing Zhu , Lizhuag Ma

摘要

正负标记(PU)学习是一种仅使用正样本和未标记样本训练的二分类模型，通常因数据分布不一致而导致过拟合风险估计。为此，我们引入一种基于伪监督的PU学习框架(PSPU)，其中我们首先训练PU模型，然后利用其收集置信样本生成伪监督，进而通过非PU目标纠正PU模型的权重。我们还引入额外的一致性损失以缓解噪声样本的影响。在MNIST、CIFAR-10、CIFAR-100的平衡和不平衡设置下，PSPU显著优于最近的PU学习方法，在MVTecAD上实现竞争性性能，用于工业异常检测。

关键词

positive and unlabeled learning pseudo supervision consistency loss overfitting anomaly detection

引用

@article{arxiv.2407.06698,
  title  = {PSPU: Enhanced Positive and Unlabeled Learning by Leveraging Pseudo Supervision},
  author = {Chengjie Wang and Chengming Xu and Zhenye Gan and Jianlong Hu and Wenbing Zhu and Lizhuag Ma},
  journal= {arXiv preprint arXiv:2407.06698},
  year   = {2024}
}

备注

accepted by ICME2024

Enhanced positive and unlabeled learning via pseudo supervision

摘要

关键词

引用

备注

相关论文