Reproduce “The geometry of neural network decision boundaries” 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

Dylan M Paiton, David L Schultheiss, Matthias Kümmerer, Zac Cranko, Matthias Bethge

The authors map classification boundaries around correctly classified images by finding up to 50 orthogonal directions toward class-changing inputs and sampling additional directions between them. On tested CIFAR-10 models, boundary distance grows sublinearly with subspace dimension, with strongest growth in the first few dimensions. This suggests boundaries remain relatively close, but applies only to sampled, pixel-valid points.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.