Respectfully, I’m surprised you consider CLIP to be SSL. SSL implies that the supervision guiding the image reps comes from the images (ex. masked input, full image target). That’s not the case with CLIP, as the text supervision isn’t directly derived from the images, but is collected with them

Have you ever asked yourself how much your favorite vision model knows about image capture parameters (e.g., the amount of JPEG compression, the camera model, etc.)? Furthermore, could these parameters influence its semantic recognition abilities?

(Thought I hit post but it didn’t send earlier)

My poster’s up at #MIRU2025 !

If you want to talk about labeling sparse autoencoders with VLMs, feel free to swing by IS1-082!

Work done with my advisors Yuta Nakashima and @noagarciad.bsky.social !

Poster’s up at #WACV2025 !

Merry Christmas! My gift to myself is being mentored by my awesome advisor @noagarciad.bsky.social

Also featuring @ryan-ramos.bsky.social

The message makes sense in context of the other pictures we took that day lol

Geospatial folks, what are your dreams about remote sensing data ? What would you like to do that your model can't for the moment ?
I am looking for hard problems with AI on remote sensing data.

I am very pleased to see the publication of this project, which is a continuation of my PhD work on weakly supervised object detection in artworks.
Going a step further by proposing an approach for zero-shot object detection in artworks.

bsky.app/profile/noag...

My first PhD paper at Osaka University! Very very very thankful for my wonderful collaborators and advisors!

What better time to announce a new paper than during NeurIPS and ACCV?

happy happy happy to introduce NADA, our latest work on object detection in art! 🎨

with amazing collaborators:
@patrick-ramos.bsky.social, @nicaogr.bsky.social, Selina Khan, Yuta Nakashima