@nicolaskeriven.bsky.social
CNRS researcher. Hates genAI with passion. Food and music ๐น https://linktr.ee/bluecurlmusic
Here you go good sir :)
distrokid.com/hyperfollow/...
I'm BEYOND THRILLED to tell you that our first original song is now available on all platforms!
Plz share if you like it ๐๐ผ๐๐ผ๐
Cheers
New cover! A fun and funky OST
youtu.be/mupu15xDCuk
New cover! We continue the third album, with a bit of stompbox on the chorus. Cheers!
youtu.be/jlKAQjlnpEg
New cover with Blue Curl! Take it with me is an astoundingly beautiful love song by legendary songwriter Tow Waits. We've been playing this one since the very first gig!
youtu.be/qwa95TBJw4Q
Some more chill piano improv for you :)
youtu.be/uhjG4Sxa9Q8
New cover!
AURORA is one of my fav artist, the river is the absolute banger that opens the album A Different Kind of Human. Here is a chill piano cover. Hope you enjoy!
youtu.be/4fTCr__hWJQ
Coldplay 022 Square One
Let's start the third album !
youtu.be/RoTHZ5dZ9QA
Colon, question mark, ticking all the boxes
23.05.2025 08:32 โ ๐ 0 ๐ 0 ๐ฌ 0 ๐ 0Also note that we limit ourselves to vanilla GNNs: no skip connections, normalization, etc classical anti-oversmoothing strategies. That's for the future!
7/7
Note that, while many optimization results are adapted from MLP to GNNs, here this is *specific* to GNNs: we show that this does not happen for MLPs.
6/7
...and that's bad! As soon as the last layer is trained (that happens immediately!), this average vanishes. The entire GNN is quickly stuck in a terrible near-critical point!
(*at every layers*, due to further mechanisms).
That's the main result of this paper.
5/7
...the "oversmoothing" of the backward signal is very particular: *when the forward is oversmoothed*, the updates to the backward are almost "linear-GNN" with no non-linearity.
So we *can* compute its oversmoothed limit! It is the **average** of the prediction error...
4/7
... but at the *middle* layers! Due to the non-linearity, it oversmoothes *when the forward is also oversmoothed* ๐ฒ
Why is it bad for training? Well...
3/7
The gradients are formed of two parts: the forward signal, and the backward "signal", initialized at the prediction error at the output then backpropagated.
Like the forward, the backward is multiplied by the graph matrix, so it will oversmooth...
2/7
New preprint! ๐ค
"Backward Oversmoothing: why is it hard to train deep Graph Neural Networks?"
GNNs could learn to anti-oversmooth. But they generally don't. Why is that? We uncover many interesting stuff ๐ค 1/7
arxiv.org/abs/2505.16736
New cover ! A sweet song by the mighty Pomme
youtu.be/vnZpxgWv5-E
Je sais pas si รงa servira ร quelque chose, mais je partage ici la rรฉponse de John Tolan, l'un des pilotes du projet ERC sur le Coran europรฉen, ร propos de la polรฉmique et des rรฉactions sur le projet de recherche en question.
12.05.2025 09:52 โ ๐ 182 ๐ 71 ๐ฌ 6 ๐ 9New cover! Like everyone I'm a bit obsessed with Expedition 33 right now, here is my version of the main track of the incredible OST :)
Cheers!
youtu.be/2aEGaOojEKA?...
Hello! A small excerpt from our last rehearsal. Too Fast is a fantastic song by Camรฉlia Jordana with a minimalist instrumentation. Hope you enjoy! :)
youtu.be/z7mjcU9tb_o
Blue in Green has got to be one of the greatest jazz standard of all time, by the greatest jazz pianist, on the greatest jazz album. Definitely a fav of mine. Here is my humble, chillest version.
youtu.be/JjkUdIgW6CQ
Si vous passez en Bretagne, prochain concert avec Blue Curl le 24 mai !
En extรฉrieur (abritรฉ, c'est la Bretagne) avec buvette et food truck, viendez รงa va รชtre chouette !
Le directeur de la plateforme nous annonce une forte augmentation des chansons gรฉnรฉrรฉes par lโIA. ยซ Ces contenus sont dรฉsormais clairement signalรฉs aux utilisateurs et sortent de lโalgorithme ยป, complรจte-t-il, se rรฉjouissant que les auditeurs prรฉfรจrent les ยซ vrais ยป artistes.
16.04.2025 08:47 โ ๐ 30 ๐ 22 ๐ฌ 3 ๐ 3Playlist for my covers for the full second album of Coldplay!
Remember, we aim to cover *everything*
www.youtube.com/playlist?lis...
Aaand that's it for the second #coldplay album! Next up, we start X&Y...
youtu.be/iYF9WUUuN0w
New rehearsal excerpt!
youtu.be/R-3hmPPapC4?...
Would you present your next NeurIPS paper in Europe instead of traveling to San Diego (US) if this was an option? Sรธren Hauberg (DTU) and I would love to hear the answer through this poll: (1/6)
30.03.2025 18:04 โ ๐ 280 ๐ 159 ๐ฌ 6 ๐ 14"generative AI knows no bounds, it will unleash creativity on a scale never seen before ๐คฉ"
*copies over and over whatever trend the CEO of a multibillion dollar company say is cool*
Know that if you work or interact with OpenAI in any way, shape or form, I have nothing but utter contempt for you right now. Utter contempt.
29.03.2025 10:35 โ ๐ 0 ๐ 0 ๐ฌ 0 ๐ 0