Reproduce “Learning reveals invisible structure in low-rank RNNs” 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
Yoav Ger, Omri Barak
The authors derive low-dimensional learning dynamics for low-rank recurrent networks using overlaps that track function and parameter changes. In linear networks, some overlaps determine behavior while others shape learning without changing it; a parameterization-induced metric reproduces training dynamics. For nonlinear models, the reduction relies on a Gaussian assumption, with a Q–Q correlation of r ≥ 0.998 in one simulation. History decoding is simulated, not biological.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.