Check out our new paper, fresh on arXiv! Along with @MolecularRobot, @MichalTomza, @adauphin4, @marcinplodzien, and Maciek Lewenstein, we explored the OOD generalization of NNs in the disordered SSH model, leading to some curious results! [1/7] https://t.co/lefZoBFx40

Last week, our outstanding MSc student @KacperCybinski participated in @APSphysics Joint March Meeting and April Meeting: Global Physics Summit 2025 in Anaheim, California, and presented results of his work on machine learning for quantum physics as a poster and contributed talk!

You can check out the final published version below 👇🏻 https://t.co/Gs28wJpgD4

🎉 We got published! 📜👀 Check out the paper, to see our thoughts on the problems with OOD generalisation of NNs in a topological system and the ideas we propose to improve the trust in networks’ predictions and their interpretations.

📚 2024 was quite a productive year! 🌟 🚀11 submitted papers, 7 published papers, 8 papers under review or in press for 2025 ✨ 🥂 Cheers to an even brighter 2025! 🎉

Our outstanding MSc student Kacper Cybiński @KacperCybinski presented his results on interpretable machine learning in Melbourne! He worked on this project with Prof. Anna Dawid @MolecularRobot during his internship at @FlatironInst in NYC last year. Fot. Prof. @EliskaGreplova

BREAKING NEWS The Royal Swedish Academy of Sciences has decided to award the 2024 #NobelPrize in Physics to John J. Hopfield and Geoffrey E. Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”

Przede wszystkim chcę zacząć od słowa: dziękuję. Tym, którzy tak licznie zaangażowali się we wsparcie dla mnie, mojego zespołu, tego co robiliśmy i bardzo chcemy robić nadal. @IDEAS_NCBR to projekt, w który zaangażowałem się aby budować AI w Polsce. Po to aby kluczowe dla

IDEAS has been a great partner at pushing AI for sciences and it has developed into an exceptional powerhouse of European innovation. Now, I find it very quite troubling that there was no transparency about the recent leadership selection process and I have mostly heared

Polska nauka tonie.

Wczoraj doszło do bezprecedensowej sytuacji, w której Rada Nadzorcza IDEAS NCBR nie przedłużyła kadencji prezesowi i twórcy tej instytucji prof. Piotrowi Sankowskiemu. IDEAS NCBR to najlepszy polski ośrodek badawczo-rozwojowy działający w obszarze sztucznej inteligencji. Zebrał

In summary, you can learn a lot of cool things by interpeting your networks! Adding such analysis to your pipeline gives you great insight into the NN’s decision process and helps you trust its predictions more - all this at a low computational cost! [7/7]

We find that CNN tends to have good OOD generalization when: 1. CNN represents the data with slight disorder similar to its disorderless training data. 2. The representation gets smoothly disconnected from initial clusters with increasing disorder strength. [6/7]

We succeeded in distinguishing the well- and poorly-generalizing networks by analyzing their latent space representation. To do that, we used PCA to visualize in 2D the high-dimensional data representation learned by networks. [5/7]

Beware! CAM gave unreliable explanations for the SSH model with disorder. The work by Ghorbani et al. (2019) helped us understand this observation. https://t.co/3Yyh2nOoTc [4/7]

We used CAM to see which parts of the data are important to the class predictions. Surprisingly, the CNNs rarely looked at the edge states—the system's known topological invariant. The ones that looked at edge states were more likely to generalize well OOD. [3/7]

We successfully trained many CNNs to classify phases in the SSH model. To our surprise, almost all of them failed to generalize out of distribution (OOD), that is, the SSH model with disorder. 1. Why do they fail? 2. When can we expect OOD generalization from the network? [2/7]

Our excellent MSc student, Kacper Cybiński @KacperCybinski, was awarded the scholarship for outstanding undergraduate students by the Polish Minister of Science and Higher Education @MNiSW_GOV__PL. Congratulations!

Los tej grupy bardzo zdolnych naukowców, oraz setek podobnych, jest w rękach @wieczorekdarek @MarekGzik @m_gdula @szeptycki. Nie zmarnujcie ich potencjału oraz szans rozwojowych dla polskiej nauki i gospodarki! #NCNtoTlen

[beware, Tweet contains paper placement] It was for sure interesting to understand better latent variable energy-based models and Yann's proposal. Check out our notes following @ylecun's three lectures at the Summer School on ML&Stat Phys! #LesHouches2022 https://t.co/f2H5sV7RRc