Reproduce “Affine Geometry of Gaussian ReLU Networks via Conditional Kac-Rice Formulas” 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
Recep Özkan, Christian Hirsch
The authors derive an exact finite-width method for counting hidden activation switches in randomly initialized Gaussian ReLU networks and use it to predict scalar-output boundaries. Conditioning on earlier layers enables a cell-by-cell Kac–Rice calculation, with leading intensities as widths grow. On breast-cancer paths, training reduced switches within classes and increased them between classes; this finding is empirical for one dataset.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.