Megi Dervishi

PostDoc @amilabs | PhD @MetaAI @psl_univ

[TimDarcet @TimDarcet]

@hugues_va Yeah LeJEPA is not a JEPA, that's some bullshit and I'm annoyed about it

[Megi Dervishi @megdrv]

🇰🇷 #ICML2026 Alert! 🇰🇷

Come check out our new work with @mathuvu_ and @ylecun 👀
📍 Starting 2PM - Poster #1606

💡 Representation learning for text done efficiently - clean scaling, 2× retrieval, ~100× less compute.

🚨 Spoiler Alert: BERT does not scale

[Saining Xie @sainingxie]

see you at ICML in Seoul and the AMI × SBVA mixer on thursday evening! looking forward to meeting everyone :)
https://t.co/O8vjcGweEv

[Yann LeCun @ylecun]

V-JEPA: a step towards getting machines to understand how the world works by watching.
The Joint embedding Predictive Architecture (JEPA) is a non-generative architecture that predicts the representation of a signal from a corrupted or transformed version of that signal.
In V-JEPA the signal is a short video that is corrupted by masking a large portion of each frame.
After training, the learned representation is used as input to a simple classifier tested on action recognition from videos.
The system gives excellent results on action recognition on SSv2 and K400 with frozen backbone and fine-tuning.
blog post: https://t.co/mfLvtvk8jj
paper: https://t.co/UuABuXRsXp
code (CC-BY-NC): https://t.co/2LuxnfPGH6

[Dieuwke Hupkes @_dieuwke_]

Have a look at the paper for more details!

https://t.co/Q6pZ9KdOyP

Work done at @AIatMeta

Core WorldSense team: @megdrv @marksibrahim @pascal20100 @benchek_youssef @_dieuwke_ & Emmanuel Dupoux