A Simple Framework for Contrastive Learning of Visual Representations How do they prevent mode collapse?
