Academic Jia-Bin Huang has posted a 36-minute YouTube explainer that walks through three decades of self-supervised learning ideas, starting with 1992 random-dot stereograms and building up to Joint-Embedding Predictive Architectures such as I-JEPA and V-JEPA. The video focuses on the persistent problem of representation collapse and the successive fixes developed to keep models from ignoring input variety.
This video by @jbhuang0604 is a compact but very informative dive into the progress of self-supervised learning over the past few decades.
from IMAX in 1992
covering methods like MoCo, SimCLR, DINO, BYOL, MAE
all the way up to LeJEPA in 2025
Highly recommend watching!
Viewers see how contrastive methods, distillation tricks, and masked modeling each try to stop collapse, with the later JEPA family shifting prediction into latent space instead of pixels.
The video ends at latent-space planning and world-model hints, yet leaves open how far these architectures scale beyond the examples shown.
video link: https://www.youtube.com/watch?v=gVEr2cnDE_8
This video by @jbhuang0604 is a compact but very informative dive into the progress of self-supervised learning over the past few decades.
from IMAX in 1992
covering methods like MoCo, SimCLR, DINO, BYOL, MAE
all the way up to LeJEPA in 2025
Highly recommend watching!
I didn't know what JEPA is, and at this point I am too afraid to ask ...😬
so I made a video tracing the last 30+ years of self-supervised learning, covering ideas from contrastive learning, distillation, masked modeling, JEPA, and world models.
https://youtu.be/gVEr2cnDE_8
@jbhuang0604 Please help me with $1000 for medical emergency sir
No Digg Deeper questions have been answered for this story yet.
