From time to time I write about the need to do interdisciplinary research when AI is used to solve problems from other disciplines. Yesterday, our paper on “Implicit modeling of equivariant tensor basis with Euclidean turbulence closure neural network” was published pubs.aip.org/aip/pof/arti...
Posts by Mateusz Gajewski
🚨 Image AutoRegressive Models Leak More Training Data Than Diffusion Models🚨
IARs — like the #NeurIPS2024 Best Paper — now lead in AI image generation. But at what risk?
IARs:
🔍 Are more likely than DMs to reveal training data
🖼️ Leak entire training images verbatim
🧵 1/
Love question 2 from @tenanatc.bsky.social
Ciaran Lee from Spotify worked on quantum causal models and it seems we have a pretty good ways to go from the quantum to deterministic structural models.
We discuss it briefly in the upcoming episode of the Causal Bandits Podcast.
#CausalSky
To były dwa bardzo intensywne dni w Davos.
- W czwartek rano panel w Poland Biznes Hub o dekarbonizacji
- W czwartek w południe, to panel w Belgium House, gdzie mowę wstępną wygłosił premier Belgii
- W piątek rano "polski" panel w AI House, o znaczeniu lokalnych talentóe dla globalnych innowacji
500 miliardów dolarów - tyle Stany Zjednoczone planują wydać na centra obliczeniowe dla AI w najbliższych 4 latach. Wczoraj ten plan ogłosił Donald Trump razem z Masayoshi Sonem - CEO SoftBank, Samem Altmanem - CEO OpenAI oraz Larrym Ellisonem - prezesem Oracle -
www.theverge.com/2025/1/21/24....
Polecam moją rozmowę z
Beatą Mońką w Interii biznes.interia.pl/gospodarka/n... między innymi o tym gdzie jesteśmy aktualnie w Polsce w rozwoju AI.
That was @neuripsconf.bsky.social 2024 seen through the eyes of IDEAS NCBR researchers
See more ➡️ ideas-ncbr.pl/en/ideas-ncb... #neurips
I am totally in the FOMO mode, seeing #NeurIPS24 messages. I’m not there :( this year.
However, I recommend chatting with my co-authors and students :) @bartoszpiotrowski.bsky.social @albertqjiang.bsky.social @michalnauman @mateuszostaszewski.bsky.social
2 papers on the main track and spotlight
Duże konferencje są dobre bo są dobre i duże ;). NeurIPS 2024 to 16500 uczestników, co pokazuje dlaczego warto tutaj publikować, bo każda opublikowana praca z definicji dostaje kilkunastotysięczne audytorium. O naszej widoczności na NeurIPS pisał też Kuba Pachocki (główny naukowiec w OpenAI).
Cool news. Our paper "Accurate estimation of feature importance faithfulness for tree models" (arxiv.org/pdf/2404.03426) with Mateusz Gajewski, Adam Karczmarz, Mateusz Rapicki has just been accepted to AAAI conference (aaai.org/conference/a...).
Finally, describe in detail available real-world datasets and advocate their usage. The study is a mine of ideas 💡 for evaluation of causal discovery methods. What are your thoughts?
🚨 The paper challenges the reliance on structural metrics (SHD) and propose to use interventional metrics as “they assess the outcomes we truly care about — the effects of unseen interventions”. 🌍
The authors review a large body of work on causal discovery, analyzing how methods are evaluated. They provide a detailed summary of both datasets and metrics currently in use. 📚
How would you evaluate a new causal discovery method? A new paper by Brouillard et al. challenges the common approaches and suggests a rethink. Here’s what they found 🧵👇
arxiv.org/abs/2412.01953
#CausalSky
I have created a starter pack for the Polish ML community. Let's grow it together! Let me know if you know somebody who should be included.
go.bsky.app/TSkgFjK