Reproduce “Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize” 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
Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino, Canlin Zhang, Xiuwen Liu
Neural Expansion links training examples to test predictions by measuring generalization intervals, distances predictions remain stable along input directions, and ray coverage, the fraction of paths from training examples reaching a test input. It relates longer intervals to directions with small penultimate-layer Jacobian singular values. On MNIST, expansion raised test accuracy from 87.38% to 89.04%; the intervention changed training losses, so it does not isolate coverage as the cause.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.