Matthias Kümmerer

Neuroscience, Machine Learning and Classical Music. Postdoc @bethgelab. Founding Member of @grammofy.

[Matthias Kümmerer @matthiaskue]

poster session right now at @eccvconf!

[Matthias Kümmerer @matthiaskue]

Now out at ECCV 2026: our new paper on using VLMs for building substantially improved scanpath prediction models that can easily be extended towards tasks, subjects, fixation durations and other aspects. Great work by @AgrawalSus35924! https://t.co/Sc41hw8MM2

[Matthias Kümmerer @matthiaskue]

https://t.co/lWB4oa8acA

[Matthias Kümmerer @matthiaskue]

Now out at ECCV 2026: our new paper on using VLMs for building substantially improved scanpath prediction models that can easily be extended towards tasks, subjects, fixation durations and other aspects. Great work by @AgrawalSus35924! https://t.co/Sc41hw8MM2

[Matthias Kümmerer @matthiaskue]

I'm very excited to present our paper together with @fededagos here at #NeurIPS2025 today at 4:30pm, poster no 2103. Come by if you want to see how DNNs can help build better mechanistic models of human behaviour!

[Federico D'Agostino @fededagos]

🚨 New paper at #NeurIPS2025!

A systematic fixation-level comparison of a performance-optimized DNN scanpath model and a mechanistic cognitive model reveals behaviourally relevant mechanisms that can be added to the mechanistic model to substantially improve performance.

🧵👇

[Matthias Kümmerer @matthiaskue]

Check out our new paper at #NeurIPS2025! I'm excited to be here!

[Federico D'Agostino @fededagos]

🚨 New paper at #NeurIPS2025!

A systematic fixation-level comparison of a performance-optimized DNN scanpath model and a mechanistic cognitive model reveals behaviourally relevant mechanisms that can be added to the mechanistic model to substantially improve performance.

🧵👇

[Federico D'Agostino @fededagos]

🚨 New paper at #NeurIPS2025!

A systematic fixation-level comparison of a performance-optimized DNN scanpath model and a mechanistic cognitive model reveals behaviourally relevant mechanisms that can be added to the mechanistic model to substantially improve performance.

🧵👇

[Vishaal Udandarao @vishaal_urao]

🚀New Paper
https://t.co/9TDiyZbOaw

Most data-curation for contrastive VLMs is offline and concept-agnostic!

We introduce online, concept-aware batch sampling (CABS), enabling flexible and targeted curation, yielding ⬆️boosts over IID sampling (+7-9% on ImageNet + ret)!

🧵👇

[Matthias Kümmerer @matthiaskue]

Our model sets a new SOTA on multiple benchmark datasets for generalization, adapation and full training.

[Matthias Kümmerer @matthiaskue]

I'm happy to be here at @ICCVConference and looking forward to meeting many interesting people! I will present our new paper on Modeling Saliency Dataset Bias for better predictions on new datasets: https://t.co/92qjSxUdZW

[Matthias Kümmerer @matthiaskue]

We found that less than 20 interpretable parameters are responsible for most of the performance drop on unseen datasets and that they can be adapted in a very data efficient way.

[Matthias Kümmerer @matthiaskue]

@brialong Sounds exciting! Do you happen to have a timecode? The video is quite long and I couldn't find a table of content.

[Vishaal Udandarao @vishaal_urao]

🚀New Preprint Alert!

📊Exploring the notion of "Zero-Shot" Generalization in Foundation Models. Is it all just a myth? Our latest preprint dives deep. Check it out!🔍

https://t.co/4PKwBGzr6K

[Matthias Kümmerer @matthiaskue]

@TrackingActions @stes_io @ELLISforEurope @bethgelab @EDNE_EPFL @mwmathislab @HelmholtzMunich Congratulations!

[Matthias Kümmerer @matthiaskue]

@Enemy_Anemone Amazing, congratulations!!!

[Ori Press @ori_press]

Interested in deploying computer vision models in the real world? You're probably going to want to continuously adapt to your data.

We show that a simple baseline approach beats all current continuously adapting methods.
⬇️🧵 1/6

[Matthias Kümmerer @matthiaskue]

I'm looking forward to giving my talk "Modeling Internal State Changes in Free-Viewing and Visual Search Scanpaths With Gain Control in DeepGaze III" at 12:00pm @VSSMtg . Come by if you're interested!

[Matthias Kümmerer @matthiaskue]

@Enemy_Anemone Congratulations!!!

[Wieland Brendel @wielandbr]

Yeah! Our work on an intelligent #robot for #sustainable #farming (POLYBOT) made it into the opening of #COP27 World Leader Summit in the talk of @VeronikaEyring. Interested in bold climate tech solutions? Talk to me or @clmich at #COP27 or DM us.@Cyber_Valley @MPI_IS @ml4science

[Matthias Kümmerer @matthiaskue]

@hidoba_tarazi @MatthiasBethge @tsawallis @CentreForCogSci @uni_tue @ARVOJOV Sorry, I forgot to answer you... No, the model so far only incorporates fixations and saccades, but doesn't handle microsaccades. But that might be interesting to do in the future.

[Matthias Kümmerer @matthiaskue]

Just published: DeepGaze, now with scanpath prediction! Joint work with @MatthiasBethge and @tsawallis (@CentreForCogSci, @uni_tue) @ARVOJOV 1/8
https://t.co/aC4xivUiK0

[Matthias Kümmerer @matthiaskue]

@covertpickle @hidoba_tarazi @MatthiasBethge @tsawallis @CentreForCogSci @uni_tue @ARVOJOV The code for DeepGaze III is (finally...) online at https://t.co/ISrpsKGfIA