
Research Fellow in Machine Learning. AI Group at Universitat Pompeu Fabra (UPF), Barcelona. Junior Leader Grant with @FundlaCaixa. Before postdoc @DTU_Compute.
Barcelona
Joined March 2012
- Tweets336
- Following573
- Followers538
- Likes2.7K
Pablo Moreno-Muñoz retweeted
The purpose of the AI industry should be to produce tools that, in the human hand, will improve human prosperity and welfare.
It should not be to create a "successor species" to the human race. Entertaining such a thought makes you, de facto, the enemy of all present and future humans.
Self-regulation in such an opaque industry seems naive regarding its ability to mitigate future risks or prevent damage.
This one is a constant reminder: caltech.edu/about/news/getti…
as if 'self-regulation' would have ever saved us from 100y of toxic gasoline additives.
1
109
“While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.”
I’m speechless
We congratulate Levent Alpöge and Tristan Buckmaster on their remarkable mathematical work.
We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem.
While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.
However, our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs. unforced).
3
261
Pablo Moreno-Muñoz retweeted
value-driven transport!
a new framework for generative modeling, combining elements of
* optimal control / RL
* optimal transport
* stochastic primal-dual optimization
thread about our new work with @pablorenoz (@UPFBarcelona) & @adrianlmueller (@ETH_en)
1/
Some news here! I’m excited to share that I’ve started a new position as Research Fellow @ Universitat Pompeu Fabra (UPF) in Barcelona 🌞
For the next 3 years, I will be doing research in Machine Learning fully funded by @FundlaCaixa via one JL Fellowship Grant.
Reading this Bertrand Russell’s 📖 jewel these last days. Published in 1938, it’s shocking to explore his thoughts on propaganda and technical advances. How that inflated the power of some.
He didn’t know yet what was coming, he refers to them as politicians, not yet criminals.
6
462
This is so cool! I am a bit jealous of PhD candidates nowadays 🌝 — and I would really enjoy to be part of such program if I was a student today. It is really a huge opportunity for EU and international applicants. What a pity that we did not have such opportunities in 2016 👀
🇨🇦 🇪🇺 ELLIS and @CIFAR are joining forces to expand the ELLIS PhD & Postdoctoral Program in AI
CIFAR Fellows and Canada AI Chairs will join ELLIS advisors, connecting PhD students with top AI researchers across Europe & North America.
Learn more here: bit.ly/4hFvurq
5
626
Pablo Moreno-Muñoz retweeted
Ever thought about using matrix-exp's or log-det's in large-scale ML (think PDEs/GGN matrices with >1M rows)? I have, and maybe you have, too—but gradients? That's where it gets tricky.
Not anymore! We have a #NeurIPS2024 spotlight: "Gradients of Functions of Large Matrices."🧵
Pablo Moreno-Muñoz retweeted
So a H100 uses as much power as a German.
nvidia.com/en-us/data-center…
en.wikipedia.org/wiki/List_o…
Pablo Moreno-Muñoz retweeted
Never ask a woman her age, a man his salary or a Cambridge machine learning department why waste taxpayer funds on frameworks that neither work nor scale like on Gaussian processes or Bayesian deep nets.
#bayesianism
Riemannian geometry meets the Laplace approximation in New Orleans @ #NeurIPS2023 ⚡️⚡️⚡️
On my way to #NeurIPS23 for the first time. I’ll be there presenting our work "Riemannian Laplace approximations for Bayesian Neural Networks". Work done together with @pablorenoz, Søren Hauberg, and @arvanitg. I’ll be at poster #1223 on Wednesday from 17pm (Session 4).
We enjoyed quite a lot the poster session #4!! Grateful for all the people who came by to chat and make insightful questions on our work.
If you could not attend but still interested, contact us or check out our paper (arxiv.org/abs/2306.07158) 📑 – at New Orleans, LA
We’ll be presenting our work in poster session 5 (#1214, Thursday 10:45) @ #NeurIPS2023 — Come to chat with us if you want to know more about the role of masking in SSL and its connection with Bayesian principles! ⚖️🔮 – at New Orleans, LA
Not my first NeurIPS, but actually this #NeurIPS2023 conference is getting to a new order of magnitude 👀👀👀 – at New Orleans, LA
13
684
Pablo Moreno-Muñoz retweeted
We're hiring a Postdoc and two Ph.D. students to work together on privacy problems in ML at DTU. (and yes, I moved again, lol)
efzu.fa.em2.oraclecloud.com/…
efzu.fa.em2.oraclecloud.com/…
Super excited about the progress of this work! — which is now one of our two papers accepted at #NeurIPS2023 🌝
I am really looking forward to be in New Orleans, sharing ideas and discussions on this beautiful link between masked pre-training and marginal likelihood 📈
Glad to share our new work on the connection between masked pre-training and Bayesian principles!
We formally prove and provide empirical results that the widely used pre-training method in LLMs maximises the marginal likelihood of the model 📈
arxiv.org/abs/2306.00520
Exciting news in Copenhagen! Super glad to know that @IgnacioPeis will be joining to our research family in DTU very soon 👀
This Post is from an account that no longer exists. Learn more
2
9
1,025
Also happy for seeing such strong researchers joining the AI/ML research community in Denmark, which now will include @luigigres as well!
5
399
Glad to share our new work on the connection between masked pre-training and Bayesian principles!
We formally prove and provide empirical results that the widely used pre-training method in LLMs maximises the marginal likelihood of the model 📈
arxiv.org/abs/2306.00520
Beyond shedding light on the learning mechanism behind masked pre-training, we also suggests that Bayesian models can be trained with appropriate self-supervision.
This could be important for the probabilistic ML community in the future 🔎
1
2
607
I want also to remark the contribution of @pgarec — whose knowledge in LLMs and transformers was fundamental for the development of the theory, together with Søren.
Tomorrow, I will be presenting this work in Tübingen. Slides and extra details will also come soon!
1
542
























