←
Mathilde Papillon🦋 mathildepapillon .bsky .social524 posts
Mathilde Papillon🦋 mathildepapillon .bsky .social avatar

Mathilde Papillon🦋 mathildepapillon .bsky .social

@mathildepapillo

Non-Euclidean + Interpretable AI PhD-ing @Geometric_Intel @UCSBPhysics Incoming MTS @GoodfireAI Prev. @Dodgers⚾️ @McGillUPhysics BlueSky: @mathildepapillon

◷ Joined July 2022▧ 53 Photos and videos
Mathilde Papillon🦋 mathildepapillon .bsky .social avatar
Mathilde Papillon🦋 mathildepapillon .bsky .social@mathildepapillo·Oct 26

Interpretability is the most important scientific problem of our lifetime. But it's too slow right now! I wrote some thoughts with @banburismus_ on how we might accelerate it. (Probably a good time to share that I'll be joining @GoodfireAI post PhD. I'm over the moon about it🌝)

if any science should be revolutionised by AI, it should be interpretability. I wrote about how we can accelerate progress, and why we need a Human Genome Project for AI. coauthored with @mathildepapillo

Mathilde Papillon🦋 mathildepapillon .bsky .social avatar
Mathilde Papillon🦋 mathildepapillon .bsky .social@mathildepapillo·16h

Beyond "regular" research, I spent many hours on Adobe Illustrator making what are now some of the go-to figures of non-Euc ML. I was very kindly invited to share the foundations and research stories of our field in workshops, keynotes, and summer schools all over 🇨🇦🇬🇧🇭🇷🇵🇹🇲🇽🇺🇸

Post attachmentPost attachmentPost attachment
Mathilde Papillon🦋 mathildepapillon .bsky .social avatar
Mathilde Papillon🦋 mathildepapillon .bsky .social@mathildepapillo·16h

I focused on non-Euclidean structures in both data and models, with a special focus on spatiotemporal and topological cases🍩 I got to work with dance, baseball, molecular, social, and medical data alongside brilliant scientists, mathematicians, clinicians, and software pros.

Mathilde Papillon🦋 mathildepapillon .bsky .social avatar
Mathilde Papillon🦋 mathildepapillon .bsky .social@mathildepapillo·16h

I just defended my physics PhD thesis, "Learning with Structure: Building and Explaining Spatiotemporal and Topological Deep Learning" @ucsantabarbara🎓 When I started in 2022, AI was still a pretty niche field, so it's been amazing to grow with it and contribute my own tidbit⬇️

Congratulations, Dr. @mathildepapillo🎓 Advised by @ninamiolane, she defended her PhD on building and explaining deep learning on spatiotemporal and topological domains, and is heading to @GoodfireAI 🚀 #PhDLife #UCSBECE @geometric_intel @UCSBengineering

Post attachmentPost attachmentPost attachment
Petar Veličković avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Petar Veličković@PetarV_93·Nov 10

the full geometric deep learning book draft is now complete. all chapters available online on our website (link in thread). 🫡 @mmbronstein @TacoCohen

Post attachmentPost attachment
Tom McGrath avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Tom McGrath@banburismus_·Nov 5

we study how this kind of training affects monitorability - over relatively short training times we just get new linear representations, but under enough pressure things can go fully nonlinear! much more to do to preserve full monitorability, but I think we'll get there, and the way we'll do it is with careful analysis and algorithm design

We investigated this with both theory and experiments. Probes can break, but predictably: the concept usually moves rather than disappears. In our language model experiments, retraining a probe was enough to recover the signal.

0:08
Eric Ho avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Eric Ho@eric_ho·Nov 5

being able to shape models during training is a big step forward for alignment, moving from trial and error to designing models with intention. incredibly proud of the team - this has been cooking for a long time, and we now see a path to truly being able to predict and shape model behavior from the inside out

We cut through the mystery around interpretability in training. A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities. We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment. But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵

0:11
Chris Earls avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Chris Earls@ChrisEarls1235·Nov 5

This work is the departure point for a science of model training: we must have models that are "right for the right reasons"

We cut through the mystery around interpretability in training. A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities. We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment. But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵

0:11
Fatih Dinc avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Fatih Dinc@fatihdin4en·Nov 5

This work is now out and has been giving me hope for the future of mechanistic interpretability and alignment! I also have my own little contribution to it! You can train networks to change which concepts they rely on when making decisions - even specifying conditional rules like “use the tongue when visible, otherwise the ears.” Here’s an example of dog classification using interpretable features discovered in the model’s internal representations through unsupervised learning (SAEs)! Read the paper here: static.goodfire.ai/mechanist… Blogpost: goodfire.com/research/mechan…

Post attachment

A science of post-training can be built on interpretability. Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules. In vision, DPO & RL, and its effects on monitoring 🧵

0:14
Matthew Kowal avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Matthew Kowal@MatthewKowal9·Nov 5

This was one of the coolest and most fun projects I have ever worked on: mechanistic shaping. We introduce a new framework for shaping what concepts a model learns to rely on during training. My favorite applications from the paper: medical imaging. 🩻

Post attachment

We cut through the mystery around interpretability in training. A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities. We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment. But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵

0:11
Goodfire avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Goodfire@GoodfireAI·Nov 5

We cut through the mystery around interpretability in training. A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities. We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment. But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵

0:11
Thomas Fel avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Thomas Fel@thomas_fel_·Nov 5

A science of post-training can be built on interpretability. Models know when they reward hack, and we see it, yet training ignores it. We propose Mechanistic Shaping: using the model’s own concepts to write training rules. In vision, DPO & RL, and its effects on monitoring 🧵

0:14
Mathilde Papillon🦋 mathildepapillon .bsky .social avatar
Mathilde Papillon🦋 mathildepapillon .bsky .social@mathildepapillo·Nov 2

Going to miss my desk🥲 But hey, it could be yours! @geometric_intel is recruiting:)

From your desk you’ll see dolphins 🐬, beach volleyball courts 🏐, a great surf spot 🏄, & more importantly: your wonderful colleagues. @mathildepapillo @AdeleMyersPhD @hopfbifurcator @louisacornelis @AbbyBertics @bariskurtkaya @gbg1441 @helloyesimsarah Luis Pereira Jamie Shih!

Post attachment
Nina Miolane 🦋 @ninamiolane.bsky.social avatar
⟳ Mathilde Papillon🦋 mathildepapillon .bsky .social retweeted
Nina Miolane 🦋 @ninamiolane.bsky.social@ninamiolane·Nov 2

I am recruiting PhD students for Fall 2027!😃 You want to find the mathematical laws of intelligence in brains and machines? 🌐🧠🤖 Apply to the UCSB Geometric Intelligence Lab✨ This is the view you'll have from.... your desk 🌴

Post attachment