反标记相关Paper (complementary label)

Sugiyama 组理论范式文章：

• NIPS-17 Learning from Complementary Labels

• ICML-19 Complementary-Label Learning for Arbitrary Losses and Models（推荐)

• ICML-20 Unbiased risk estimators can mislead: A case study of learning with complementary labels

• ECCV-19 Learning with Biased Complementary Labels

• ICML-20 Learning with multiple complementary labels （推荐)

• AAAI-20 Generative-discriminative complementary learning

Negative Learning:

• ICCV-19 NLNL: Negative Learning for Noisy Labels (从 noisy label 生成反标记） （推荐)

Loss function

• ICCV-19 Symmetric Cross Entropy for Robust Learning With Noisy Labels (提出了 RCE， SCE = CE + RCE) （推荐)
• ICML-20 Normalized Loss Functions for Deep Learning with Noisy Labels （推荐)

ps: 两文来自同一组作者，理论推导 follow AAAI-17 Robust Loss Functions under Label Noise for Deep Neural Networks

Reverse Cross Entropy (THU jun zhu 学生）

• NIPS-18 Towards Robust Detection of Adversarial Examples （推荐)
  (提出了 Reverse Cross Entropy， 重名于 ICCV-19 的 RCE , 本质是 label smoothing 取参数 s=1.0。给了一些理论分析)