Check👇out! New paper by Natasha Butt from AMLAB with Qualcomm collaborators on models that learn to discover programs to solve complex tasks. Congrats Natasha & Co!

Excited to share that our paper “CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay” was accepted into ICML!
Blaze(j) Manczak 🇵🇱🇱🇺🇪🇺 Auke Wiggers Corrado Rainone David Zhang Michaël Defferrard Taco Cohen 1/5

CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay

Natasha Butt, Blaze(j) Manczak 🇵🇱🇱🇺🇪🇺, Auke Wiggers, Corrado Rainone, David Zhang, Michaël Defferrard, Taco Cohen

tl;dr: sample a program, try it, add to the replay pool.
New sota on ARC
arxiv.org/abs/2402.04858…

For the free online Learning on Graphs Conference 2023 we have a bunch of free local meetups!
Find out if there is one close to you.
Links to some of them are in the reply 👇
1/3

A lesser-known fact about ML open source contributors: About 50% of code contributors to ML paper implementations are practitioners collaborating with researchers. Here are the topk researchers & practitioners contributing to open source and open to sponsorship.…

We are happy to announce the first Learning on Graphs Conference 2023 Meetup in Lausanne, supported by VantAI! 🤗
Join us on Nov 22nd at EPFL to hear Charlotte Bunne, @ClementVignac, Dorina Thanou and Michaël Defferrard. Bring your posters!
Registration: forms.gle/c4HjeuDoBTpXN9… Webpage: sites.google.com/view/log-meetu…

Layer-wise equivariance symmetries (e.g. conv layers) allow neural nets to generalise effectively. But can we learn them automatically using gradients? We show we can! Excited to share that our method ELLA has been accepted as a spotlight paper at #neurips2023 . A thread.👇1/11🧵

Had an incredible time hosting the inaugural Paris Learning-on-Graphs Meetup! 🎉 Thank you to everyone who attended and made it such a fantastic hybrid event! Special thanks to our inspiring speakers for their outstanding talks and to the poster presenters for sharing their work.

Here is the schedule ! Thanks to our amazing line of speakers for accepting to share their research and insights about Graph Machine Learning.
Xavier Bresson Michael Bronstein Johannes Lutzeyer Michaël Defferrard Dominic Masters Graphcore and Vassilis N. Ioannidis from Amazon Science.

The schedule of the LoG-Paris-Meetup is out: sites.google.com/view/learning-… !! :)

Big thanks to our speakers Michael Bronstein Xavier Bresson Johannes Lutzeyer Michaël Defferrard vassilis ioannidis Dominic Masters.

Registrations and poster submissions are still open !
Learning on Graphs Conference 2023 CentraleSupélec Inria

The perpetually undervalued least-squares:

minₓ‖y−Ax‖²

can teach a lot about some complex ideas in modern machine learning including overfitting & double-descent.

Let's assume A is n-by-p. So we have n data points and p parameters

1/n

The supremacy of search as a paradigm—and the associated attention-selling business—is coming to an end.

Virtual assistants will disrupt search like search disrupted web portals.

Note, a similar observation was made by Michaël Defferrard on spherical data, showing that with isotropic filters (only way to get equivariance without g-convs! cf. Lecture 1.7 uvagedl.github.io) one can beat specialized equiv layers like group convs
openreview.net/forum?id=B1e3O…
4/4

After a hectic month off the #100DaysOfCode challenge, I'm back. Day 52!💥🦾

👉Working on Michaël Defferrard 's GCNN implementation for a signal classification task in #Keras

✅Loaded + numpyed features, adj matrix & labels
🚧Coarsening the graph

#GraphML #AI #WomenWhoCode #WomenInSTEM

Really excited to be introducing our new framework for de novo protein design, tomorrow at the MLDD_Workshop #ICLR_2022 at 10:30-10:50am ET!

Paper: openreview.net/forum?id=DwN81…

For those (virtually) at SIAM Imaging Science '22 #SIAMIS22 : There's this amazing session on 'Learning from Vision'. I'll be speaking at 18.10 CET (times below in UTC-4) about equivariant graph NNs using sub-Riemannian geometry😎(cf arxiv.org/abs/2111.12139 w Michaël Defferrard & Aguettaz)

The essential nature of convolutions to space is the backbone of the thesis I'll defend today.

My contribution: generalized convolutions. They enable parameter sharing for non-transitive and unknown symmetry groups to efficiently learn on arbitrary domains.

Looking forward!

To make the implicit explicit:

I strongly believe we should ask students (and all researchers) to optimize quality over quantity.

But we can't honestly do that at the same time our orgs tweet 'we have 1000 NeurIPS papers.'

Let's normalize not caring about counts *EVERYWHERE*.

A classic! I love sub-Riemannian geometry and've been trying to get it back in my research. Inspired by the math, Aguettaz+Michaël Defferrard+I just put
“ChebLieNet: Invariant Spectral Graph NNs Turned Equivariant by Riemannian Geometry on Lie Groups” on arXiv: arxiv.org/abs/2111.12139 [1/4]