iiinova — where your digital self keeps evolving

Multiscale AI @ ICLR 2025

@multiscaleai.bsky.social

ICLR 2025 workshop working building AI to answer a single question: Given low-level theory and computationally-expensive simulation code, how can we model complex systems on a useful time scale?

Posts Replies Media

Multiscale AI @ ICLR 2025 @multiscaleai.bsky.social · 05/05/2025

The ICLR 2025 MLMP Best Paper Award, along with 2k GPU-hours from Nebius, goes to "On Incorporating Scale into Graph Networks"! Congratulations, Christian Koke, Yuesong Shen, @abhi-rf.bsky.social, Marvin Eisenberger, @pseudomanifold.topology.rocks, Michael M. Bronstein, @dcremers.bsky.social!

Standard graph neural networks assign vastly different latent embeddings to graphs describing the same physical system at different resolution scales. This precludes consistency in applications and prevents generalization between scales as would fundamentally be needed in many scientific applications. We uncover the underlying obstruction, investigate its origin and show how to overcome it.

162

Multiscale AI @ ICLR 2025 @multiscaleai.bsky.social · 05/05/2025

ICLR 2025 MLMP best poster award goes to "ViNE-GATr: scaling geometric algebra transformers with virtual nodes embeddings"! Congratulations @sukjulian.bsky.social, Thomas Hehn, @arashbehboodi.bsky.social, Gabriele Cesa!

Equivariant neural networks can effectively model physical systems by naturally handling the underlying geometric quantities and preserving their symmetries, but scaling them to large geometric data remains challenging. Naive downsampling typically disrupts features’ transformation laws, limiting their applicability in large scale settings. In this work, we propose a scalable equivariant transformer that efficiently processes geometric data in a coarse-grained latent space while preserving E(3) symmetries of the problem. In particular, by building on the Geometric Algebra Transformer (GATr) and PerceiverIO architectures, our method learns equivariant latent tokens which allow us to decouple the processing complexity from the input data representation while maintaining global equivariance.

141