Karsten Roth

Also very thankful for the research environment provided by @ellis.eu and @mpi-is.bsky.social, which made this PhD such an inter-european experience!

Huge thanks also to my thesis committee Peter Gehler, Matthias Bethge, @wielandbrendel.bsky.social and @phillipisola.bsky.social, and of course all the wonderful people and collaborators I had the pleasure of spending time and working with these past years!

💫 After four PhD years on all things multimodal, pre- and post-training, I’m super excited for a new research chapter at Google DeepMind 🇨🇭!

Biggest thanks to @zeynepakata.bsky.social and Oriol Vinyals for all the guidance, support, and incredibly eventful and defining research years ♥️!

How does lifelong knowledge editing currently hold up in the real world? Fun new work probing where we are at these days with injecting new knowledge into LLMs!

📄 Disentangled Representation Learning with the Gromov-Monge Gap

with Théo Uscidda, Luca Eyring, @confusezius.bsky.social, Fabian J Theis, Marco Cuturi

📄 Decoupling Angles and Strength in Low-rank Adaptation

with Massimo Bini, Leander Girrbach

Our EML team has 4 #ICLR25 Papers accepted! I am proud of my students and grateful to be a part of many successful collaborations. More details will appear on our website (www.eml-munich.de) but here are the snapshots.

The Practitioner's Guide to Continual Multimodal Pretraining @dziadzio.bsky.social @confusezius.bsky.social @vishaalurao.bsky.social @bayesiankitten.bsky.social

Can we enhance the performance of T2I models without any fine-tuning?

We show that with our ReNO, Reward-based Noise Optimization, one-step models consistently surpass the performance of all current open-source Text-to-Image models within the computational budget of 20-50 sec!
#NeurIPS2024

How far can you push model merging over time, as more experts and options to model-merge arise?

We comprehensively and systematically investigate this in our new work, check it out!

Co-led with @vishaalurao.bsky.social, and with a wonderful team: @dziadzio.bsky.social @bayesiankitten.bsky.social Mehdi Cherti, Oriol Vinyals @olivierhenaff.bsky.social, Samuel Albanie, @bethgelab.bsky.social and @zeynepakata.bsky.social!

We will present on Wednesday - East Exhibit Hall A-C #3703 ☺️. We've also released the entire codebase with all the methods and 60+ dataloaders that can be mixed and matched in any fashion to study continual pretraining!

😵‍💫 Continually pretraining large multimodal models to keep them up-to-date all-the-time is tough, covering everything from adapters, merging, meta-scheduling to data design and more!

So I'm really happy to present our large-scale study at #NeurIPS2024!

Come drop by to talk about all that and more!

161 outstanding machine learning researchers accepted as new ELLIS Fellows & Scholars The ELLIS mission is to create a diverse European network that promotes research excellence and advances breakthroughs in AI, as well as a pan-European PhD program to educate the next generation of AI...

🎉 Congratulations to our newly accepted ELLIS Fellows & Scholars in 2024! Top researchers in #MachineLearning join the network to advance science & mentor the next generation. #ELLISforEurope #AI

🌍 Know someone on the list? bit.ly/3ZJd9Cz
Tag them in a reply with congratulations.

🚀New Paper: Active Data Curation Effectively Distills Multimodal Models
arxiv.org/abs/2411.18674

Smol models are all the rage these days & knowledge distillation (KD) is key for model compression!

We show how data curation can effectively distill to yield SoTA FLOP-efficient {C/Sig}LIPs!!
🧵👇

Read our paper:
Context-Aware Multimodal Pretraining

Now on ArXiv

Can you turn vision-language models into strong any-shot models?

Go beyond zero-shot performance in SigLixP (x for context)

Read @confusezius.bsky.social thread below…

And follow Karsten … a rising star!

Oh that's a really cool paper! Thanks for the pointer!

Beautiful paper! 😍😍😍

Captions go above the tables, but otherwise aesthetically very pleasing.

Oh neat, do you have a link? 😁

More than zero-shot generalization, few-shot *adaptation* is critical for many applications.

We find simple changes to multimodal pretraining are sufficient to yield outsized gains on a wide range of few-shot tasks.

Congratulations @confusezius.bsky.social on a very successful internship!

We maintain strong zero-shot transfer of CLIP / SigLIP across model size and data scale, while achieving up to 4x few-shot sample efficiency and up to +16% performance gains!

Fun project with @confusezius.bsky.social, @zeynepakata.bsky.social, @dimadamen.bsky.social and
@olivierhenaff.bsky.social.

LIxP was carefully designed and tested for scalability!

LIxP also maintains the strong zero-shot transfer of CLIP and SigLIP backbones across model sizes (S to L) and data (up to 15B), and allows up to 4x sample efficiency at test time, and up to +16% performance gains!

In LIxP, we utilize a learnable temperature separation and a simple cross-attention-based formalism to augment existing contrastive vision-language training.

We teach models what to expect at test-time in few-shot scenarios.

They can struggle with applications that require operating on new context, e.g. few-shot adaptation.

Why? They do not explicitly train for that!

We find a surrogate objective to optimize for -- context-aware language-image pretraining (LIxP)

This was an insightful project I worked on at Google DeepMind alongside the amazing @zeynepakata.bsky.social , @dimadamen.bsky.social , @ibalazevic.bsky.social and @olivierhenaff.bsky.social:

👉Language-image pretraining with CLIP or SigLIP is widely used due to strong zero-shot transfer, but ....

🤔 Can you turn your vision-language model from a great zero-shot model into a great-at-any-shot generalist?

Turns out you can, and here is how: arxiv.org/abs/2411.15099

Really excited to this work on multimodal pretraining for my first bluesky entry!

🧵 A short and hopefully informative thread:

Hi 👋 We're glad to be here on @bsky.app and looking forward to engaging in this community. But first, learn a little more about us...

#ELLISforEurope #AI #ML #CrossBorderCollab #PhD

Cool starterpack, would love to be added as well :)