"Novelty detection"
학습 분포와 벗어나는 샘플을 찾는 것
Hyperspherical embeddings for OOD detection
- von Mises-Fisher (vMF) distribution 활용
: a classical and important distribution in directional statistics (Mardia et al., 2000)
2 losses to promote strong ID-OOD separability
- a dispersion loss (that promotes large angular distances among different class prototypes)
- a compactness loss (that encourages samples to be close to their class prototypes)