co-learning:learning from noisy labels with self-supervision 解读

摘要

受到co-training的启发，作者将supervised learning和self-supervised learning 结合起来。
The constraints of intrinsic similarity with the self-supervised module and the structural similarity with the noisily-supervised module are imposed on a shared common feature encoder to regularize the network to maximize the agreement between the two constraints。

novelties

训练了一个单一的共享的encoder，有两个头，分别是self-supervised 和 noisily-supervised, 这两个互相约束，maximizes the agreement between them in latent space：

1. self-supervised learning 作为一个feature-independent view来辅助supervised learning。
2. 有一个共享的feature encoder，但是有两个不一样的头。projection head 负责self-supervised learning，探索和特征相关的信息。classifier head执行监督学习，学习和标签相关的信息。
   作者提出的unsupervised learning 隐式的帮助从noisy labels中学习。作者尝试证明了不同的views，可以给出更多可靠的信息，尤其是面对noisy training data。作者提出，传统的方法想要在两个模型的预测上取得最