
Research Scientist at @GoogleDeepMind my personal account is @zdanielz if you like baked goods and can optionally read Hebrew.
London, England
Joined July 2021
- Tweets20
- Following138
- Followers276
- Likes32
Come see our poster at #CVPR2026 today! Morning session 11:45am, poster 326!
Popular self-supervised models like V-JEPA v2 avoid pixel reconstruction, claiming it's "irrelevant" for high-level semantics.
But what if pixels were never a problem?
Introducing Recurrent Video Masked Autoencoders (RVM) to be presented at #CVPR2026
arxiv.org/abs/2512.13684
🧵
Daniel Zoran retweeted
Looped transformers are all the rage, but the value of recurrence is fundamental, not just hype- after all, the brain is highly recurrent. In that spirit, I'm thrilled to share RVM (to be presented at #CVPR2026): a new class of SoTA recurrent self-supervised video encoders! 🧵
Popular self-supervised models like V-JEPA v2 avoid pixel reconstruction, claiming it's "irrelevant" for high-level semantics.
But what if pixels were never a problem?
Introducing Recurrent Video Masked Autoencoders (RVM) to be presented at #CVPR2026
arxiv.org/abs/2512.13684
🧵
Popular self-supervised models like V-JEPA v2 avoid pixel reconstruction, claiming it's "irrelevant" for high-level semantics.
But what if pixels were never a problem?
Introducing Recurrent Video Masked Autoencoders (RVM) to be presented at #CVPR2026
arxiv.org/abs/2512.13684
🧵
The results are promising:
✅ Competitive with V-JEPA v2 on spatio-temporal tasks.
✅ Competitive with DINOv2 on spatial tasks.
✅ Strong performance in the small-model regime WITHOUT distillation.
Read the full paper here:
arxiv.org/abs/2512.13684
Code and checkpoints available here:
rvm-paper.github.io/
This is joint work with my wonderful collaborators @nikparth1, Yi Yang, @drewAhudson, @joaocarreira, and Andrew Zisserman
Daniel Zoran retweeted
Excited to announce MooG for learning video representations. MooG allows tokens to move “off-the-grid” enabling better representation of scene elements, even as they move across the image plane through time.
📜arxiv.org/abs/2411.05927
🌐moog-paper.github.io/
That was easily one of the best talks I've been to at a conference, fantastic work! at @CVPR
Daniel Zoran retweeted
Today @ICLR_conf: Perceiver IO, an any modal attention architecture that scales gracefully to large inputs and outputs, allowing it to produce excellent results on tasks in vision, language, multimodal domains, and beyond: dpmd.ai/ICLR-Perceiver #ICLR2022 1/
It’s that time of year again…@ICLR_conf! Here’s a brief glimpse into our upcoming oral, spotlight, and poster presentations this week: dpmd.ai/ICLR2022
Don’t miss our 5+ sessions kicking off this morning at 10:30AM BST. See you there! #ICLR2022







