Browse Teach Curate

Robust Supervised Contrastive Loss against Label Noise

A theoretical framework that derives a general condition under which an arbitrary pairwise supervised contrastive loss is robust to label noise, providing a criterion to test any such loss. It shows the InfoNCE loss is non-robust and constructs Symmetric InfoNCE (SymNCE), a modified loss that satisfies the condition, while also explaining prior heuristics such as nearest-neighbor sample selection.

B9 Text

An Inclusive Theoretical Framework of Robust Supervised Contrastive Loss against Label Noise Jingyi

This paper builds a unified theory of when a supervised contrastive loss is robust to label noise. Starting from the pairwise contrastive paradigm, the authors derive a general robustness condition t…