VJE adds uncertainty to joint embeddings
Joint Embedding Variational Bayes, or VJE, is a TMLR paper that adds variational inference to joint-embedding methods for non-contrastive representation learning. The core idea is to model embeddings with a normalized latent-variable likelihood instead of a pointwise similarity loss, using a directional-radial factorization, tied posterior/likelihood uncertainty, and a heavy-tailed Student-t form to avoid the instability that shows up as the likelihood becomes too Gaussian. The result is a self-supervised framework that still performs competitively on standard representation benchmarks while also producing feature-wise uncertainty signals that are useful for OOD detection.
Strong idea, and the paper’s main value is that it makes “probabilistic embeddings” feel operational rather than decorative.
- –The directional/radial split is the right kind of fix: it addresses the norm-angle coupling that often makes embedding objectives numerically brittle.
- –Tying posterior variance to likelihood scale is a clean way to make uncertainty affect both inference and the representation model, not just appear as an add-on head.
- –The Student-t choice is not cosmetic; the paper’s reported collapse near the Gaussian limit suggests the heavy tail is doing real stability work.
- –The downstream OOD results are the most compelling part of the pitch, since they show the uncertainty is actually usable rather than merely well-parameterized.
- –The main caveat is that this is still a mathematically dense research paper, so the practical adoption bar will be higher than for a simpler SimSiam-style baseline.
DISCOVERED
45d ago
2026-04-30
PUBLISHED
45d ago
2026-04-30
RELEVANCE
AUTHOR
ISwallow5Gum