PaperReading20200403

CanChen ggchen@mail.ustc.edu.cn

---

 

Circle Loss

• Motivation: This paper wants to provide a unified view of two elemental learning approaches, i.e., learning with class-level labels and pair-wise labels.
• Method: The paper proposes a unified formula to show that learning with class-level labels and pair-wise labels can all be seen as minimizing the between-class similarity while maxmizing the within-class similarity. Given the unified formula, the author shows that the formula degenerates to triplet loss or classification loss, through slight modifications.
• Contribution: I am not quite familiar with metric learning. Anyway this paper gives me a high view before seeing any papers about specific metrics.