Martin Hebart

martinhebart.bsky.social

25 September 2026 12:42

What is the nature of universal representations in AI models, and what determines whether they emerge? Our paper accepted at #neurips2026 led by @florianmahner.bsky.social & @rothj.bsky.social addressed these questions comparing 162 vision models, with intriguing results. arxiv.org/abs/2605.13675 🧵

re-vision-init.bsky.social

03 August 2026 12:41

(6/6) To learn more, browse studies, and sign up visit re-vision-initiative.org or come to our Poster at #CCN2026 next week on Wednesday, August 5, 2:00 – 3:45 pm We'd love to have you on board!

re-vision-init.bsky.social

03 August 2026 12:41

(5/6) How you can participate: teams of 1–3 pick a published finding and replicate it using LAION-fMRI. No new data collection. Everything is provided via our Python package. Every valid report gets the chance at co-authorship on the consortium paper and to win one of our cash prices of $2,500

re-vision-init.bsky.social

03 August 2026 12:41

(4/6) Enter LAION-fMRI: densely-sampled 7T fMRI responses to 25,000+ natural images across 5 subjects, designed to cover the space of natural images at exceptional scale and diversity.

re-vision-init.bsky.social

03 August 2026 12:41

(3/6) They're rarely tested directly. And recent work (e.g. by @ykamit.bsky.social) suggests common datasets may lack image diversity and have high train/test similarity, opening the door to leakage. doi.org/10.1016/j.ne...