Vision Transformers Need Registers (arxiv.org)

Ameo 14 hours ago

Extremely cool!

It's interesting and honestly encouraging that this kind of thing can be discovered and understood using just "simple linear methods" and high-level analysis of patterns in layer activations.

refulgentis 22 hours ago
Scene_Cast2 21 hours ago

So basically multiple CLS tokens.

Fwiw, I tried multiple global tokens in my chess neural net and didn't see any uplift compared to my baseline of just having one.

ninja3925 21 hours ago

Note that it's not done for performance reason but rather to generate clear feature maps.

bigdict 21 hours ago

Has this been used widely since?

ashvardanian 21 hours ago

I can't speak for the whole industry, but we used it in older UForm <https://github.com/unum-cloud/uform> and saw good adoption, especially among those deploying on the Edge, where every little trick counts. It's hard to pin down exact numbers since most deployments didn't go through Hugging Face, but at the time, these models were likely among the more widely deployed by device count.

kombine 19 hours ago

For example, it is used here https://github.com/facebookresearch/vggt/

heyitsguay 21 hours ago

I ran a comparison of DINOv2 with and without registers on some image embedding tasks for work; DINOv2+registers saw a performance metric bump of 2-3%. Not nothing, not transformative, worth using when the only difference for inference is the model name string you're loading.

godelski 21 hours ago

yes