Reproduce “The balance between feature learning and collapse in dynamical systems-based generative 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
Julian Brandon, Bruno Loureiro, N Alex Cayco-Gajic, Arthur Pellegrino
The authors analyze how low-rank learning updates relate to mode coverage in controlled generative dynamics. They bound gradient stable rank using state and adjoint operators, then compare reverse-KL linear flows with flow matching. Reverse-KL runs show balanced-mixture collapse; flow matching delays small modes. A one-hidden-layer ReLU sweep reports a 0.72 ± 0.01 learning-time exponent, with a conflicting figure label; these findings are limited to controlled settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.