Triton kernels for the G(3, 0, 1) geometric product (as used by geometric algebra transformer) are now available on @maxxxzdn.bsky.social's great repository Flash Clifford.
🔗 github.com/maxxxzdn/flash-clifford
⚡ Triton implementation of Clifford algebra neural networks. - maxxxzdn/flash-clifford
Julian Suk
We have a competition @gram-org.bsky.social!
- MCML award: 500€
- Deadline: April 22nd
- Publication of the results in our workshop proceedings (PMLR)
✨Thank you to @sukjulian.bsky.social, Gavin and BeyondMath for making this happen. ✨
gram-competition.github.io
Alison Pouplin
Contrary to what Big Scale has you believe, equivariant models are alive and thriving 😉
Jokes aside, there is increasing evidence that enforcing symmetry is beneficial in certain applications.
arxiv.org/abs/2510.01349
arxiv.org/abs/2605.18816
Symmetry-aware methods for machine learning, such as data augmentation and equivariant architectures, encourage correct model behavior on all transformations (e.g. rotations or permutations) of the or...
DM me to discuss if this new addition can speed up your model!
Results for the GRaM Competition are online. Congratulations to the winner and runners-up 🎉
🥇 github.com/julka01 with SmoothSplatNet
🥈 github.com/jorgesarrato with CDFDoubleGridNet
🥉 github.com/v3ctr0id with VRTEnsemble
🔗 gram-competition.github.io
Julian Suk
Overall, submissions were really high-quality 🚀 The first place came down to an extremely thin margin at effectively identical accuracy. We were particularly excited to see the high-frequency flow components captured quite well!
We have published the test split on Hugging Face and revealed the test metric. Feel free to continue developing and benchmarking your models!
If you know your way around neural operators or have other good ideas, check out our website!
