Mode Connectivity Beyond Classifiers: Evidence from Generative and Contrastive Models
This work demonstrates for the first time that mode connectivity between independently trained DDPM and NanoCLIP modes is discovered, and provides a novel perspective for understanding the geometric properties of the loss landscapes in modern generative and contrastive models.
Chengzheyi Yao, Yong-Zhao Zhang, Yong Tian
· 0 citations