The authors start from the multi-view assumption: if two augmented views contain enough information for a downstream task, a useful representation should retain what those views have in common.
Patch-level self-supervision scales to seven billion parameters with simpler objectives
JEM aligns image patches across augmented views while preserving information and spatial structure, producing strong representations for segmentation without combining separate image-level and patch-level objectives.
Big Tech
Maximilian Seitzer · Gabriele Trivigno · Antonín Vobecký · Seungeun Yi · Maxime Oquab · Huy V. Vo · +2 more
Meta FAIR
Research Digest··3 min read
Seitzer and colleagues derive a student-teacher learning objective from the idea that useful visual information should be shared across different views of an image.
Why this paper
From Meta FAIR
In one line
JEM uses a principled patch-level objective to train stable visual encoders up to 7B parameters while outperforming DINOv2 on dense prediction tasks.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 7B)
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§