Prominent AI researchers are resurfacing old internal disagreements about where self-supervised learning ends and unsupervised learning begins, with no clean line emerging from the recollections. The conversation centers on how teams once argued over definitions without reaching black-and-white conclusions, especially as newer frameworks prompt fresh rounds of naming and positioning.
@suchenzang @gabriberton Indeed i remember in our team we had discussions/arguments about "whats the diff between SSL and Unsupervised Learning" and it wasn't clearly black and white.
@giffmana @gabriberton even ssl was...
Researchers note that distinctions were never sharply defined in practice, leaving room for ongoing reinterpretation whenever new methods appear.
The thread shows how earlier terminology decisions continue to influence how approaches are described and differentiated today.
Many users dismissed JEPA discussions as irrelevant since few actually use it and criticized them as name-dropping arguments over credit assignment, while a few recalled fond memories of earlier self-supervised learning times.
No Digg Deeper questions have been answered for this story yet.
@gabriberton JEPA is literally the only new idea in the SSL space in the past 3 years... what other SSL approaches are there? DINO? iBOT? MAE? these are all older than 3 years!
@giffmana @AggieInCA Hey is this JEPA? /s
@gabriberton Yeah
@giffmana @gabriberton based on this level of flag planting, unsupervised has been relegated to the corner of "absolutely no learning"
@suchenzang @gabriberton Indeed i remember in our team we had discussions/arguments about "whats the diff between SSL and Unsupervised Learning" and it wasn't clearly black and white.
@giffmana @gabriberton even ssl was...
@iScienceLuvr I can think of a few new ideas but... this is really not my point
My point is that people are starting to call JEPA everything that is SSL
And the lines are very blurry because they're both pretty vague terms
@AggieInCA @gabriberton I have fond memories of these times
@gabriberton There was a time that these methods were called JE-SSL. Personally, I don’t care.i just miss the OG SSL community and OG SSL Workshops we used to have at NeurICMLRs
@gabriberton i actually saw a post call SSL "just vibes-based pattern matching" and the replies were half mad half in agreement
@gabriberton Yep, sad, and shows that the ML community is really bad at credit assignment. And the usual narrative is usually, "yeah but these guys who did that 10 years ago didn't really make it work, we managed to make it work at larger scale so we rightly deserve 95% of the credit"
@gabriberton not really. most people aren’t even using jepa. most retweet or fight over these ideas. very few use these. most are still using off the shelf stuff.
@gabriberton @AggieInCA Followed by a long discussion of "what Yann really meant" packed with name-dropping arguments.
@iScienceLuvr @gabriberton you're calling to be schmidhubered
@suchenzang @giffmana @gabriberton It is enough to make you a godfather figure tho
@giffmana @gabriberton based on this level of flag planting, unsupervised has been relegated to the corner of "absolutely no learning"
@iScienceLuvr @gabriberton I think the point is that JEPA is too general, even DINO/iBoT can be thought of as kind of JEPA
@gabriberton unsupervised methods aren't dead. they're handling fraud detection, anomaly detection, clustering at scale in production right now.
they just don't trend like supervised does.
I had a similar reaction that the name was just rebranding SSL, but now I think there’s some merit to the distinction. For one, it excludes contrastive SSL methods which I think are the methods most strongly linked to SSL. But second, and the biggest issue, is that next token prediction is also technically self supervised learning.