ICML 2021 is going to show us a new self-supervised embedding algorithm called Barlow Twins. It's name is ofc a reference to the siamese networks, but the key novelty is lack of contrastive samples.
Preprint by Zbontar et al.:
https://arxiv.org/abs/2103.03230
Follow
There is a similar, less known method called Invariant Information Clustering which also uses augmented images as a positive anchor
