Disclosure: first author. The paper was just published in TMLR, and I figured it might be of interest to some people here. It is fairly dense mathematically, but straightforward conceptually: to add operational variational semantics to joint-embedding architectures for non-contrastive representation
