Reproduce “Interfaces and Modules: A Scalable Representation for Weight-Space Learning across Architectures” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Michał Kozyra, Michael Bronstein
The authors introduce UWE, a shared encoder representing neural networks through coordinate interfaces and connecting operations. A compiler builds this representation for supported PyTorch models; UWE combines symmetry-respecting updates with global attention over interface summaries. On held-out architectures in a CIFAR-100 vision-transformer zoo, UWE reaches Kendall’s tau-b 0.7538, above the baseline, but this ranking result is limited to that zoo.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.