😅
08.11.2025 22:55
👍 0
🔁 0
💬 0
📌 0
😅
and of course a good base policy!
Training another musculoskeletal system @xiaotiansu.bsky.social for tennis, after having trained other musculoskeletal systems for soccer and pingpong.
With cognitive function, reward learning, and some sort of teacher distillation, this system learns a quite robust forehand within 36,000 steps👍
Something exciting just left the gym -- meet Arnold the first generalist musculoskeletal control policy!
The human body has 600+ muscles, yet we seamlessly coordinate them for everything from whispering a syllable to felling a forrest. How? www.arxiv.org/abs/2508.18066
Interesting study, do check this out 💪
The more you use JAX, the more you feel its power.. like wielding the sharpest knife in your toolbox 🗡️
glad to see distill.pub revived at anthropic :)
It’s pretty strange they don’t have this in zurich. At EPFL, all stem master’s students have a mandatory internship, and we get the chance to intern at big tech companies
Quick life update: I recently started my PhD back in the beautiful Lac Léman region at EPFL Campus Biotech.
Just wanted to show off the office a bit! 😎
More nature views coming soon…
Ep 2 of #InsideAI is out! 🎙️🎉
We sat with Google Fellow Urs Hölzle. From optimizing data centers to AI efficiency, AGI, autonomous vehicles & Europe’s AI outlook—don’t miss this insightful conversation.
On Spotify: open.spotify.com/episode/5Tu9...
or Apple Podcast: podcasts.apple.com/us/podcast/c...
Stopped by BAAI (Beijing Academy of AI) to visit a friend! ☕️ The coffee was great, and it’s right between China’s top two universities. Sadly, the hanging garden is closed for now. Overall a quite nice place to do research
#Beijing
Awesome work, congrats!
Picture of the SusTec research group at ETH Zurich with all its members standing in front of the Geneva lake on a sunny summer afternoon. There is text saying: We're hiring! 1 PhD student: Modelling net-zero supply chains.
We are hiring 1 PhD in energy modeling!
The PhD investigates policies to accelerate supply chain development for net-zero technologies like hydrogen through agent-based and energy system modeling.
Details: emea2.softfactors.com/job-opening/...
Apply until 31.01.2025.
Please share!
I've put together a starter pack of EPFL researchers across all labs and domains! 🇨🇭 Would love to expand this list and showcase more amazing work happening at EPFL. Drop a reply to be added!
#EPFL #academicsky
go.bsky.app/73zdbtp
🚀 Introducing PICLe: a framework for in-context named-entity detection (NED) using pseudo-annotated demonstrations.
🎯 No human labeling needed—yet it outperforms few-shot learning with human annotations!
#AI #NLProc #LLMs #ICL #NER
Super interesting example! The trained model seems to understand the colliding better than rolling (the colliding part looks more natural)
the benchmark section looks amazing! congrats! 👏
Goodbye sunny Barcelona
Hello again Zürich
#photography #barcelona
Another interesting project I worked on at @icepfl.bsky.social Kudos to @sjaved.bsky.social and all the co-authors!
The findings in MQAT highlight the potential of exploiting modularity in neural networks for efficient and performant compression/adaptation.
Check out the #TMLR paper for details!
Congrats Saqib! Enjoyed working with you!
Ilya’s full talk at neurips 2024 "pre-training as we know it will end" at #NeurIPS
youtu.be/6gTjxQLwK4o?...
source: Vincent Weisser
Yeah, back in the day interns at bd had some serious access permissions, but this is taking it too far;
still imo retracting the paper is unlikely, I wonder how neurips will react
Last chance! #AMLDEPFL2025 early bird ticket sales end tomorrow December 11. Get yours👉https://buff.ly/3UYv8Cr. Join us for inspiring speakers, comprehensive tracks, cutting-edge exhibition, hands-on workshops, and unmatched networking. February 11-14, 2025 SwissTech Convention Center EPFL Lausanne
Perhaps instead of solely focusing on preventing forgetting in AI systems, we should be asking: How can we implement more selective, adaptive forgetting mechanisms similar in a continual learning framework? [4/4]
Thoughts?
#continual
So probably, the real challenge isn't preventing ALL forgetting - it's developing mechanisms to identify which patterns should be preserved vs. discarded. This is where biological systems could excel through social feedback, interactions, expert guidance, and iterative refinement. [3/4]
In both natural and artificial learning systems, selective forgetting (perturbation/regularization) could help escape local optima and unlearn inefficient or corrupted patterns (especially relevant given concerns about poisoned training data) [2/4]
💭When reading continual learning literature, there's always the term 'catastrophic forgetting' being discussed as a major challenge. But looking at some efficient learning systems (e.g. biological-based ones), it's hard to not notice forgetting isn't necessarily a bug - it's sometimes a merit. [1/4]
TMLR + Bluesky
Introducing the @tmlr-pub.bsky.social account, that is now posting the latest TMLR published papers! Check it out!
Three days left to apply for a PhD in my group! ⬇️⬇️
Come join us to work on the forefront of adaptive AI & lifelong machine learning. We’ve got lots of exciting ideas in store 🤗
