Reproduce “Fisher-IRG: Fisher-Induced Local Invariant Representation Geometry across Language and Vision Models” 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
Abdullah All Tanvir, Xin Zhong
The authors introduce Fisher-IRG to identify representation directions more predictively sensitive to meaning-changing than meaning-preserving variants. The method estimates local Fisher information from output gradients and contrasts the two neighborhoods to form subspaces. Across ten model families, it has a positive mean selectivity advantage over Cov-IRG; on ViT-B/16, the difference is 0.628. This supports local behavioral analysis, not improved task performance or universal invariance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.