Isabella Yu

Dataset Distillation for Pre-Trained Self-Supervised Vision Models Distilling image datasets to a single sample per class to train classifiers on top of pre-trained self-supervised vision models.

Happy to finally share our latest work on Dataset Distillation!

"Dataset Distillation for Pre-Trained Self-Supervised Vision Models," set to appear at #NeurIPS 2025!

We learn 1 image per class to train linear heads for pre-trained models.

linear-gradient-matching.github.io

More in thread 🔽

21.11.2025 04:01 — 👍 6    🔁 2    💬 1    📌 0

A young child in a red dress examines the flowers of a drooping saguaro cactus

Fan art (?) for The Glass Castle by Jeannette Walls, anyone else read this in high school #procreate

16.11.2025 10:41 — 👍 1    🔁 0    💬 0    📌 0

DUSt3R et al. are impressive, but how do they actually work? We investigate this in our project 𝘜𝘯𝘥𝘦𝘳𝘴𝘵𝘢𝘯𝘥𝘪𝘯𝘨 𝘔𝘶𝘭𝘵𝘪-𝘝𝘪𝘦𝘸 𝘛𝘳𝘢𝘯𝘴𝘧𝘰𝘳𝘮𝘦𝘳𝘴!⁣

We share findings on the iterative nature of reconstruction, the roles of cross and self-attention, and the emergence of correspondences across the network [1/8] ⬇️

04.11.2025 19:40 — 👍 6    🔁 4    💬 1    📌 0

An explorer from a distant utopia visits the ruins of a mass surveillance state, where a decommissioned guard robot stands by a deserted castle. The explorer naively follows the instructions on its body, unknowingly triggering one of its kinder behaviors it learned by watching the world become kinder over millennia

Machines of loving grace #procreate #illustration

20.10.2025 09:59 — 👍 2    🔁 0    💬 0    📌 0

Wow!

So while I still wish this case would’ve gone to trial, and believe the amount given per work should have been EVEN higher, $1.5 billion in settlement AND the destruction of infringing datasets is still a fantastic start! Congratulations to all those involved as these things are not easy!

05.09.2025 21:00 — 👍 1563    🔁 559    💬 20    📌 1

Really excited to share this with you all! The website is already live at anadodik.github.io/publication/...

05.08.2025 15:59 — 👍 37    🔁 14    💬 2    📌 0

MIT tool visualizes and edits “physically impossible” objects The “Meschers” tool from MIT CSAIL represents “physically impossible” objects commonly found in M.C. Escher’s illustrations by converting both images and 3D models in 2.5-dimensional objects. The tool...

MIT News today features a fun SIGGRAPH paper by GDP student Ana (@geometry.gay) collaboarting w/ students @yukaryote.bsky.social @kartikchandra.bsky.social plus us old folks. This work explores what it means to do geometry processing on Escher-like "impossible" shapes! news.mit.edu/2025/mit-mes...

05.08.2025 13:07 — 👍 28    🔁 8    💬 1    📌 2

A succubus wakes up to protesting men who want her back, while her partner stares in incredulity

Merry pride #yuri #clipstudiopaint

25.06.2025 05:19 — 👍 2    🔁 0    💬 0    📌 0

it took me like 10 hours lol I'm so out of practice. I used Academy Watercolor paints my friend gave me :)

11.04.2025 05:31 — 👍 1    🔁 0    💬 0    📌 0

Trying watercolor for the first time in a while :) #pleinairpril #pleinair #watercolor

11.04.2025 04:09 — 👍 30    🔁 1    💬 2    📌 0